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 unsigned fgVNPassesCompleted; // Number of times fgValueNumber has been run.
3844 // Utility functions for fgValueNumber.
3846 // Perform value-numbering for the trees in "blk".
3847 void fgValueNumberBlock(BasicBlock* blk);
3849 // Requires that "entryBlock" is the entry block of loop "loopNum", and that "loopNum" is the
3850 // innermost loop of which "entryBlock" is the entry. Returns the value number that should be
3851 // assumed for the memoryKind at the start "entryBlk".
3852 ValueNum fgMemoryVNForLoopSideEffects(MemoryKind memoryKind, BasicBlock* entryBlock, unsigned loopNum);
3854 // Called when an operation (performed by "tree", described by "msg") may cause the GcHeap to be mutated.
3855 void fgMutateGcHeap(GenTreePtr tree DEBUGARG(const char* msg));
3857 // For a GC heap store at curTree, record the new curHeapVN and update curTree's MemorySsaMap.
3858 void recordGcHeapStore(GenTreePtr curTree, ValueNum gcHeapVN DEBUGARG(const char* msg));
3860 // Tree caused an update in the current memory VN. If "tree" has an associated heap SSA #, record that
3861 // value in that SSA #.
3862 void fgValueNumberRecordMemorySsa(MemoryKind memoryKind, GenTreePtr tree);
3864 // The input 'tree' is a leaf node that is a constant
3865 // Assign the proper value number to the tree
3866 void fgValueNumberTreeConst(GenTreePtr tree);
3868 // Assumes that all inputs to "tree" have had value numbers assigned; assigns a VN to tree.
3869 // (With some exceptions: the VN of the lhs of an assignment is assigned as part of the
3871 // If "evalAsgLhsInd" is true, evaluate a GT_IND node, even if it's labeled as the LHS of
3873 void fgValueNumberTree(GenTreePtr tree, bool evalAsgLhsInd = false);
3875 // Does value-numbering for a block assignment.
3876 void fgValueNumberBlockAssignment(GenTreePtr tree, bool evalAsgLhsInd);
3878 // Does value-numbering for a cast tree.
3879 void fgValueNumberCastTree(GenTreePtr tree);
3881 // Does value-numbering for an intrinsic tree.
3882 void fgValueNumberIntrinsic(GenTreePtr tree);
3884 // Does value-numbering for a call. We interpret some helper calls.
3885 void fgValueNumberCall(GenTreeCall* call);
3887 // The VN of some nodes in "args" may have changed -- reassign VNs to the arg list nodes.
3888 void fgUpdateArgListVNs(GenTreeArgList* args);
3890 // Does value-numbering for a helper "call" that has a VN function symbol "vnf".
3891 void fgValueNumberHelperCallFunc(GenTreeCall* call, VNFunc vnf, ValueNumPair vnpExc);
3893 // Requires "helpCall" to be a helper call. Assigns it a value number;
3894 // we understand the semantics of some of the calls. Returns "true" if
3895 // the call may modify the heap (we assume arbitrary memory side effects if so).
3896 bool fgValueNumberHelperCall(GenTreeCall* helpCall);
3898 // Requires "helpFunc" to be pure. Returns the corresponding VNFunc.
3899 VNFunc fgValueNumberHelperMethVNFunc(CorInfoHelpFunc helpFunc);
3901 // These are the current value number for the memory implicit variables while
3902 // doing value numbering. These are the value numbers under the "liberal" interpretation
3903 // of memory values; the "conservative" interpretation needs no VN, since every access of
3904 // memory yields an unknown value.
3905 ValueNum fgCurMemoryVN[MemoryKindCount];
3907 // Return a "pseudo"-class handle for an array element type. If "elemType" is TYP_STRUCT,
3908 // requires "elemStructType" to be non-null (and to have a low-order zero). Otherwise, low order bit
3909 // is 1, and the rest is an encoding of "elemTyp".
3910 static CORINFO_CLASS_HANDLE EncodeElemType(var_types elemTyp, CORINFO_CLASS_HANDLE elemStructType)
3912 if (elemStructType != nullptr)
3914 assert(varTypeIsStruct(elemTyp) || elemTyp == TYP_REF || elemTyp == TYP_BYREF ||
3915 varTypeIsIntegral(elemTyp));
3916 assert((size_t(elemStructType) & 0x1) == 0x0); // Make sure the encoding below is valid.
3917 return elemStructType;
3921 elemTyp = varTypeUnsignedToSigned(elemTyp);
3922 return CORINFO_CLASS_HANDLE(size_t(elemTyp) << 1 | 0x1);
3925 // If "clsHnd" is the result of an "EncodePrim" call, returns true and sets "*pPrimType" to the
3926 // var_types it represents. Otherwise, returns TYP_STRUCT (on the assumption that "clsHnd" is
3927 // the struct type of the element).
3928 static var_types DecodeElemType(CORINFO_CLASS_HANDLE clsHnd)
3930 size_t clsHndVal = size_t(clsHnd);
3931 if (clsHndVal & 0x1)
3933 return var_types(clsHndVal >> 1);
3941 // Convert a BYTE which represents the VM's CorInfoGCtype to the JIT's var_types
3942 var_types getJitGCType(BYTE gcType);
3944 enum structPassingKind
3946 SPK_Unknown, // Invalid value, never returned
3947 SPK_PrimitiveType, // The struct is passed/returned using a primitive type.
3948 SPK_ByValue, // The struct is passed/returned by value (using the ABI rules)
3949 // for ARM64 and UNIX_X64 in multiple registers. (when all of the
3950 // parameters registers are used, then the stack will be used)
3951 // for X86 passed on the stack, for ARM32 passed in registers
3952 // or the stack or split between registers and the stack.
3953 SPK_ByValueAsHfa, // The struct is passed/returned as an HFA in multiple registers.
3955 }; // The struct is passed/returned by reference to a copy/buffer.
3957 // Get the "primitive" type that is is used when we are given a struct of size 'structSize'.
3958 // For pointer sized structs the 'clsHnd' is used to determine if the struct contains GC ref.
3959 // A "primitive" type is one of the scalar types: byte, short, int, long, ref, float, double
3960 // If we can't or shouldn't use a "primitive" type then TYP_UNKNOWN is returned.
3962 var_types getPrimitiveTypeForStruct(unsigned structSize, CORINFO_CLASS_HANDLE clsHnd);
3964 // Get the type that is used to pass values of the given struct type.
3965 // If you have already retrieved the struct size then pass it as the optional third argument
3967 var_types getArgTypeForStruct(CORINFO_CLASS_HANDLE clsHnd,
3968 structPassingKind* wbPassStruct,
3969 unsigned structSize = 0);
3971 // Get the type that is used to return values of the given struct type.
3972 // If you have already retrieved the struct size then pass it as the optional third argument
3974 var_types getReturnTypeForStruct(CORINFO_CLASS_HANDLE clsHnd,
3975 structPassingKind* wbPassStruct = nullptr,
3976 unsigned structSize = 0);
3979 // Print a representation of "vnp" or "vn" on standard output.
3980 // If "level" is non-zero, we also print out a partial expansion of the value.
3981 void vnpPrint(ValueNumPair vnp, unsigned level);
3982 void vnPrint(ValueNum vn, unsigned level);
3985 // Dominator computation member functions
3986 // Not exposed outside Compiler
3988 bool fgDominate(BasicBlock* b1, BasicBlock* b2); // Return true if b1 dominates b2
3990 bool fgReachable(BasicBlock* b1, BasicBlock* b2); // Returns true if block b1 can reach block b2
3992 void fgComputeDoms(); // Computes the immediate dominators for each basic block in the
3993 // flow graph. We first assume the fields bbIDom on each
3994 // basic block are invalid. This computation is needed later
3995 // by fgBuildDomTree to build the dominance tree structure.
3996 // Based on: A Simple, Fast Dominance Algorithm
3997 // by Keith D. Cooper, Timothy J. Harvey, and Ken Kennedy
3999 BlockSet_ValRet_T fgGetDominatorSet(BasicBlock* block); // Returns a set of blocks that dominate the given block.
4000 // Note: this is relatively slow compared to calling fgDominate(),
4001 // especially if dealing with a single block versus block check.
4003 void fgComputeReachabilitySets(); // Compute bbReach sets. (Also sets BBF_GC_SAFE_POINT flag on blocks.)
4005 void fgComputeEnterBlocksSet(); // Compute the set of entry blocks, 'fgEnterBlks'.
4007 bool fgRemoveUnreachableBlocks(); // Remove blocks determined to be unreachable by the bbReach sets.
4009 void fgComputeReachability(); // Perform flow graph node reachability analysis.
4011 BasicBlock* fgIntersectDom(BasicBlock* a, BasicBlock* b); // Intersect two immediate dominator sets.
4013 void fgDfsInvPostOrder(); // In order to compute dominance using fgIntersectDom, the flow graph nodes must be
4014 // processed in topological sort, this function takes care of that.
4016 void fgDfsInvPostOrderHelper(BasicBlock* block, BlockSet& visited, unsigned* count);
4018 BlockSet_ValRet_T fgDomFindStartNodes(); // Computes which basic blocks don't have incoming edges in the flow graph.
4019 // Returns this as a set.
4021 BlockSet_ValRet_T fgDomTreeEntryNodes(BasicBlockList** domTree); // Computes which nodes in the dominance forest are
4022 // root nodes. Returns this as a set.
4025 void fgDispDomTree(BasicBlockList** domTree); // Helper that prints out the Dominator Tree in debug builds.
4028 void fgBuildDomTree(); // Once we compute all the immediate dominator sets for each node in the flow graph
4029 // (performed by fgComputeDoms), this procedure builds the dominance tree represented
4032 // In order to speed up the queries of the form 'Does A dominates B', we can perform a DFS preorder and postorder
4033 // traversal of the dominance tree and the dominance query will become A dominates B iif preOrder(A) <= preOrder(B)
4034 // && postOrder(A) >= postOrder(B) making the computation O(1).
4035 void fgTraverseDomTree(unsigned bbNum, BasicBlockList** domTree, unsigned* preNum, unsigned* postNum);
4037 // When the flow graph changes, we need to update the block numbers, predecessor lists, reachability sets, and
4039 void fgUpdateChangedFlowGraph();
4042 // Compute the predecessors of the blocks in the control flow graph.
4043 void fgComputePreds();
4045 // Remove all predecessor information.
4046 void fgRemovePreds();
4048 // Compute the cheap flow graph predecessors lists. This is used in some early phases
4049 // before the full predecessors lists are computed.
4050 void fgComputeCheapPreds();
4053 void fgAddCheapPred(BasicBlock* block, BasicBlock* blockPred);
4055 void fgRemoveCheapPred(BasicBlock* block, BasicBlock* blockPred);
4065 // Initialize the per-block variable sets (used for liveness analysis).
4066 void fgInitBlockVarSets();
4068 // true if we've gone through and created GC Poll calls.
4069 bool fgGCPollsCreated;
4070 void fgMarkGCPollBlocks();
4071 void fgCreateGCPolls();
4072 bool fgCreateGCPoll(GCPollType pollType, BasicBlock* block);
4074 // Requires that "block" is a block that returns from
4075 // a finally. Returns the number of successors (jump targets of
4076 // of blocks in the covered "try" that did a "LEAVE".)
4077 unsigned fgNSuccsOfFinallyRet(BasicBlock* block);
4079 // Requires that "block" is a block that returns (in the sense of BBJ_EHFINALLYRET) from
4080 // a finally. Returns its "i"th successor (jump targets of
4081 // of blocks in the covered "try" that did a "LEAVE".)
4082 // Requires that "i" < fgNSuccsOfFinallyRet(block).
4083 BasicBlock* fgSuccOfFinallyRet(BasicBlock* block, unsigned i);
4086 // Factor out common portions of the impls of the methods above.
4087 void fgSuccOfFinallyRetWork(BasicBlock* block, unsigned i, BasicBlock** bres, unsigned* nres);
4090 // For many purposes, it is desirable to be able to enumerate the *distinct* targets of a switch statement,
4091 // skipping duplicate targets. (E.g., in flow analyses that are only interested in the set of possible targets.)
4092 // SwitchUniqueSuccSet contains the non-duplicated switch targets.
4093 // (Code that modifies the jump table of a switch has an obligation to call Compiler::UpdateSwitchTableTarget,
4094 // which in turn will call the "UpdateTarget" method of this type if a SwitchUniqueSuccSet has already
4095 // been computed for the switch block. If a switch block is deleted or is transformed into a non-switch,
4096 // we leave the entry associated with the block, but it will no longer be accessed.)
4097 struct SwitchUniqueSuccSet
4099 unsigned numDistinctSuccs; // Number of distinct targets of the switch.
4100 BasicBlock** nonDuplicates; // Array of "numDistinctSuccs", containing all the distinct switch target
4103 // The switch block "switchBlk" just had an entry with value "from" modified to the value "to".
4104 // Update "this" as necessary: if "from" is no longer an element of the jump table of "switchBlk",
4105 // remove it from "this", and ensure that "to" is a member. Use "alloc" to do any required allocation.
4106 void UpdateTarget(IAllocator* alloc, BasicBlock* switchBlk, BasicBlock* from, BasicBlock* to);
4109 typedef SimplerHashTable<BasicBlock*, PtrKeyFuncs<BasicBlock>, SwitchUniqueSuccSet, JitSimplerHashBehavior>
4110 BlockToSwitchDescMap;
4113 // Maps BasicBlock*'s that end in switch statements to SwitchUniqueSuccSets that allow
4114 // iteration over only the distinct successors.
4115 BlockToSwitchDescMap* m_switchDescMap;
4118 BlockToSwitchDescMap* GetSwitchDescMap()
4120 if (m_switchDescMap == nullptr)
4122 m_switchDescMap = new (getAllocator()) BlockToSwitchDescMap(getAllocator());
4124 return m_switchDescMap;
4127 // Invalidate the map of unique switch block successors. For example, since the hash key of the map
4128 // depends on block numbers, we must invalidate the map when the blocks are renumbered, to ensure that
4129 // we don't accidentally look up and return the wrong switch data.
4130 void InvalidateUniqueSwitchSuccMap()
4132 m_switchDescMap = nullptr;
4135 // Requires "switchBlock" to be a block that ends in a switch. Returns
4136 // the corresponding SwitchUniqueSuccSet.
4137 SwitchUniqueSuccSet GetDescriptorForSwitch(BasicBlock* switchBlk);
4139 // The switch block "switchBlk" just had an entry with value "from" modified to the value "to".
4140 // Update "this" as necessary: if "from" is no longer an element of the jump table of "switchBlk",
4141 // remove it from "this", and ensure that "to" is a member.
4142 void UpdateSwitchTableTarget(BasicBlock* switchBlk, BasicBlock* from, BasicBlock* to);
4144 // Remove the "SwitchUniqueSuccSet" of "switchBlk" in the BlockToSwitchDescMap.
4145 void fgInvalidateSwitchDescMapEntry(BasicBlock* switchBlk);
4147 BasicBlock* fgFirstBlockOfHandler(BasicBlock* block);
4149 flowList* fgGetPredForBlock(BasicBlock* block, BasicBlock* blockPred);
4151 flowList* fgGetPredForBlock(BasicBlock* block, BasicBlock* blockPred, flowList*** ptrToPred);
4153 flowList* fgSpliceOutPred(BasicBlock* block, BasicBlock* blockPred);
4155 flowList* fgRemoveRefPred(BasicBlock* block, BasicBlock* blockPred);
4157 flowList* fgRemoveAllRefPreds(BasicBlock* block, BasicBlock* blockPred);
4159 flowList* fgRemoveAllRefPreds(BasicBlock* block, flowList** ptrToPred);
4161 void fgRemoveBlockAsPred(BasicBlock* block);
4163 void fgChangeSwitchBlock(BasicBlock* oldSwitchBlock, BasicBlock* newSwitchBlock);
4165 void fgReplaceSwitchJumpTarget(BasicBlock* blockSwitch, BasicBlock* newTarget, BasicBlock* oldTarget);
4167 void fgReplaceJumpTarget(BasicBlock* block, BasicBlock* newTarget, BasicBlock* oldTarget);
4169 void fgReplacePred(BasicBlock* block, BasicBlock* oldPred, BasicBlock* newPred);
4171 flowList* fgAddRefPred(BasicBlock* block,
4172 BasicBlock* blockPred,
4173 flowList* oldEdge = nullptr,
4174 bool initializingPreds = false); // Only set to 'true' when we are computing preds in
4177 void fgFindBasicBlocks();
4179 bool fgIsBetterFallThrough(BasicBlock* bCur, BasicBlock* bAlt);
4181 bool fgCheckEHCanInsertAfterBlock(BasicBlock* blk, unsigned regionIndex, bool putInTryRegion);
4183 BasicBlock* fgFindInsertPoint(unsigned regionIndex,
4184 bool putInTryRegion,
4185 BasicBlock* startBlk,
4187 BasicBlock* nearBlk,
4188 BasicBlock* jumpBlk,
4191 unsigned fgGetNestingLevel(BasicBlock* block, unsigned* pFinallyNesting = nullptr);
4193 void fgRemoveEmptyBlocks();
4195 void fgRemoveStmt(BasicBlock* block, GenTreePtr stmt, bool updateRefCnt = true);
4197 bool fgCheckRemoveStmt(BasicBlock* block, GenTreePtr stmt);
4199 void fgCreateLoopPreHeader(unsigned lnum);
4201 void fgUnreachableBlock(BasicBlock* block);
4203 void fgRemoveConditionalJump(BasicBlock* block);
4205 BasicBlock* fgLastBBInMainFunction();
4207 BasicBlock* fgEndBBAfterMainFunction();
4209 void fgUnlinkRange(BasicBlock* bBeg, BasicBlock* bEnd);
4211 void fgRemoveBlock(BasicBlock* block, bool unreachable);
4213 bool fgCanCompactBlocks(BasicBlock* block, BasicBlock* bNext);
4215 void fgCompactBlocks(BasicBlock* block, BasicBlock* bNext);
4217 void fgUpdateLoopsAfterCompacting(BasicBlock* block, BasicBlock* bNext);
4219 BasicBlock* fgConnectFallThrough(BasicBlock* bSrc, BasicBlock* bDst);
4221 bool fgRenumberBlocks();
4223 bool fgExpandRarelyRunBlocks();
4225 bool fgEhAllowsMoveBlock(BasicBlock* bBefore, BasicBlock* bAfter);
4227 void fgMoveBlocksAfter(BasicBlock* bStart, BasicBlock* bEnd, BasicBlock* insertAfterBlk);
4229 enum FG_RELOCATE_TYPE
4231 FG_RELOCATE_TRY, // relocate the 'try' region
4232 FG_RELOCATE_HANDLER // relocate the handler region (including the filter if necessary)
4234 BasicBlock* fgRelocateEHRange(unsigned regionIndex, FG_RELOCATE_TYPE relocateType);
4236 #if FEATURE_EH_FUNCLETS
4237 #if defined(_TARGET_ARM_)
4238 void fgClearFinallyTargetBit(BasicBlock* block);
4239 #endif // defined(_TARGET_ARM_)
4240 bool fgIsIntraHandlerPred(BasicBlock* predBlock, BasicBlock* block);
4241 bool fgAnyIntraHandlerPreds(BasicBlock* block);
4242 void fgInsertFuncletPrologBlock(BasicBlock* block);
4243 void fgCreateFuncletPrologBlocks();
4244 void fgCreateFunclets();
4245 #else // !FEATURE_EH_FUNCLETS
4246 bool fgRelocateEHRegions();
4247 #endif // !FEATURE_EH_FUNCLETS
4249 bool fgOptimizeUncondBranchToSimpleCond(BasicBlock* block, BasicBlock* target);
4251 bool fgBlockEndFavorsTailDuplication(BasicBlock* block);
4253 bool fgBlockIsGoodTailDuplicationCandidate(BasicBlock* block);
4255 bool fgOptimizeFallthroughTailDup(BasicBlock* block, BasicBlock* target);
4257 bool fgOptimizeEmptyBlock(BasicBlock* block);
4259 bool fgOptimizeBranchToEmptyUnconditional(BasicBlock* block, BasicBlock* bDest);
4261 bool fgOptimizeBranch(BasicBlock* bJump);
4263 bool fgOptimizeSwitchBranches(BasicBlock* block);
4265 bool fgOptimizeBranchToNext(BasicBlock* block, BasicBlock* bNext, BasicBlock* bPrev);
4267 bool fgOptimizeSwitchJumps();
4269 void fgPrintEdgeWeights();
4271 void fgComputeEdgeWeights();
4273 void fgReorderBlocks();
4275 void fgDetermineFirstColdBlock();
4277 bool fgIsForwardBranch(BasicBlock* bJump, BasicBlock* bSrc = nullptr);
4279 bool fgUpdateFlowGraph(bool doTailDup = false);
4281 void fgFindOperOrder();
4283 // method that returns if you should split here
4284 typedef bool(fgSplitPredicate)(GenTree* tree, GenTree* parent, fgWalkData* data);
4286 void fgSetBlockOrder();
4288 void fgRemoveReturnBlock(BasicBlock* block);
4290 /* Helper code that has been factored out */
4291 inline void fgConvertBBToThrowBB(BasicBlock* block);
4293 bool fgCastNeeded(GenTreePtr tree, var_types toType);
4294 GenTreePtr fgDoNormalizeOnStore(GenTreePtr tree);
4295 GenTreePtr fgMakeTmpArgNode(
4296 unsigned tmpVarNum FEATURE_UNIX_AMD64_STRUCT_PASSING_ONLY_ARG(const bool passedInRegisters));
4298 // The following check for loops that don't execute calls
4299 bool fgLoopCallMarked;
4301 void fgLoopCallTest(BasicBlock* srcBB, BasicBlock* dstBB);
4302 void fgLoopCallMark();
4304 void fgMarkLoopHead(BasicBlock* block);
4306 unsigned fgGetCodeEstimate(BasicBlock* block);
4309 const char* fgProcessEscapes(const char* nameIn, escapeMapping_t* map);
4310 FILE* fgOpenFlowGraphFile(bool* wbDontClose, Phases phase, LPCWSTR type);
4311 bool fgDumpFlowGraph(Phases phase);
4313 #endif // DUMP_FLOWGRAPHS
4318 void fgDispBBLiveness(BasicBlock* block);
4319 void fgDispBBLiveness();
4320 void fgTableDispBasicBlock(BasicBlock* block, int ibcColWidth = 0);
4321 void fgDispBasicBlocks(BasicBlock* firstBlock, BasicBlock* lastBlock, bool dumpTrees);
4322 void fgDispBasicBlocks(bool dumpTrees = false);
4323 void fgDumpStmtTree(GenTreePtr stmt, unsigned blkNum);
4324 void fgDumpBlock(BasicBlock* block);
4325 void fgDumpTrees(BasicBlock* firstBlock, BasicBlock* lastBlock);
4327 static fgWalkPreFn fgStress64RsltMulCB;
4328 void fgStress64RsltMul();
4329 void fgDebugCheckUpdate();
4330 void fgDebugCheckBBlist(bool checkBBNum = false, bool checkBBRefs = true);
4331 void fgDebugCheckBlockLinks();
4332 void fgDebugCheckLinks(bool morphTrees = false);
4333 void fgDebugCheckNodeLinks(BasicBlock* block, GenTreePtr stmt);
4334 void fgDebugCheckFlags(GenTreePtr tree);
4335 void fgDebugCheckFlagsHelper(GenTreePtr tree, unsigned treeFlags, unsigned chkFlags);
4336 void fgDebugCheckTryFinallyExits();
4339 #ifdef LEGACY_BACKEND
4340 static void fgOrderBlockOps(GenTreePtr tree,
4344 GenTreePtr* opsPtr, // OUT
4345 regMaskTP* regsPtr); // OUT
4346 #endif // LEGACY_BACKEND
4348 static GenTreePtr fgGetFirstNode(GenTreePtr tree);
4349 static bool fgTreeIsInStmt(GenTree* tree, GenTreeStmt* stmt);
4351 inline bool fgIsInlining()
4353 return fgExpandInline;
4356 void fgTraverseRPO();
4358 //--------------------- Walking the trees in the IR -----------------------
4363 fgWalkPreFn* wtprVisitorFn;
4364 fgWalkPostFn* wtpoVisitorFn;
4365 void* pCallbackData; // user-provided data
4366 bool wtprLclsOnly; // whether to only visit lclvar nodes
4367 GenTreePtr parent; // parent of current node, provided to callback
4368 GenTreeStack* parentStack; // stack of parent nodes, if asked for
4370 bool printModified; // callback can use this
4374 template <bool computeStack>
4375 static fgWalkResult fgWalkTreePreRec(GenTreePtr* pTree, fgWalkData* fgWalkPre);
4377 // general purpose tree-walker that is capable of doing pre- and post- order
4378 // callbacks at the same time
4379 template <bool doPreOrder, bool doPostOrder>
4380 static fgWalkResult fgWalkTreeRec(GenTreePtr* pTree, fgWalkData* fgWalkPre);
4382 fgWalkResult fgWalkTreePre(GenTreePtr* pTree,
4383 fgWalkPreFn* visitor,
4384 void* pCallBackData = nullptr,
4385 bool lclVarsOnly = false,
4386 bool computeStack = false);
4388 fgWalkResult fgWalkTree(GenTreePtr* pTree,
4389 fgWalkPreFn* preVisitor,
4390 fgWalkPostFn* postVisitor,
4391 void* pCallBackData = nullptr);
4393 void fgWalkAllTreesPre(fgWalkPreFn* visitor, void* pCallBackData);
4397 template <bool computeStack>
4398 static fgWalkResult fgWalkTreePostRec(GenTreePtr* pTree, fgWalkData* fgWalkPre);
4400 fgWalkResult fgWalkTreePost(GenTreePtr* pTree,
4401 fgWalkPostFn* visitor,
4402 void* pCallBackData = nullptr,
4403 bool computeStack = false);
4405 // An fgWalkPreFn that looks for expressions that have inline throws in
4406 // minopts mode. Basically it looks for tress with gtOverflowEx() or
4407 // GTF_IND_RNGCHK. It returns WALK_ABORT if one is found. It
4408 // returns WALK_SKIP_SUBTREES if GTF_EXCEPT is not set (assumes flags
4409 // properly propagated to parent trees). It returns WALK_CONTINUE
4411 static fgWalkResult fgChkThrowCB(GenTreePtr* pTree, Compiler::fgWalkData* data);
4412 static fgWalkResult fgChkLocAllocCB(GenTreePtr* pTree, Compiler::fgWalkData* data);
4413 static fgWalkResult fgChkQmarkCB(GenTreePtr* pTree, Compiler::fgWalkData* data);
4415 /**************************************************************************
4417 *************************************************************************/
4420 friend class SsaBuilder;
4421 friend struct ValueNumberState;
4423 //--------------------- Detect the basic blocks ---------------------------
4425 BasicBlock** fgBBs; // Table of pointers to the BBs
4427 void fgInitBBLookup();
4428 BasicBlock* fgLookupBB(unsigned addr);
4430 void fgMarkJumpTarget(BYTE* jumpTarget, IL_OFFSET offs);
4432 void fgFindJumpTargets(const BYTE* codeAddr, IL_OFFSET codeSize, BYTE* jumpTarget);
4434 void fgMarkBackwardJump(BasicBlock* startBlock, BasicBlock* endBlock);
4436 void fgLinkBasicBlocks();
4438 unsigned fgMakeBasicBlocks(const BYTE* codeAddr, IL_OFFSET codeSize, BYTE* jumpTarget);
4440 void fgCheckBasicBlockControlFlow();
4442 void fgControlFlowPermitted(BasicBlock* blkSrc,
4443 BasicBlock* blkDest,
4444 BOOL IsLeave = false /* is the src a leave block */);
4446 bool fgFlowToFirstBlockOfInnerTry(BasicBlock* blkSrc, BasicBlock* blkDest, bool sibling);
4448 void fgObserveInlineConstants(OPCODE opcode, const FgStack& stack, bool isInlining);
4450 void fgAdjustForAddressExposedOrWrittenThis();
4452 bool fgProfileData_ILSizeMismatch;
4453 ICorJitInfo::ProfileBuffer* fgProfileBuffer;
4454 ULONG fgProfileBufferCount;
4455 ULONG fgNumProfileRuns;
4457 unsigned fgStressBBProf()
4460 unsigned result = JitConfig.JitStressBBProf();
4463 if (compStressCompile(STRESS_BB_PROFILE, 15))
4474 bool fgHaveProfileData();
4475 bool fgGetProfileWeightForBasicBlock(IL_OFFSET offset, unsigned* weight);
4477 bool fgIsUsingProfileWeights()
4479 return (fgHaveProfileData() || fgStressBBProf());
4481 void fgInstrumentMethod();
4483 //-------- Insert a statement at the start or end of a basic block --------
4487 static bool fgBlockContainsStatementBounded(BasicBlock* block, GenTree* stmt, bool answerOnBoundExceeded = true);
4491 GenTreeStmt* fgInsertStmtAtEnd(BasicBlock* block, GenTreePtr node);
4493 public: // Used by linear scan register allocation
4494 GenTreeStmt* fgInsertStmtNearEnd(BasicBlock* block, GenTreePtr node);
4497 GenTreePtr fgInsertStmtAtBeg(BasicBlock* block, GenTreePtr stmt);
4498 GenTreePtr fgInsertStmtAfter(BasicBlock* block, GenTreePtr insertionPoint, GenTreePtr stmt);
4500 public: // Used by linear scan register allocation
4501 GenTreePtr fgInsertStmtBefore(BasicBlock* block, GenTreePtr insertionPoint, GenTreePtr stmt);
4504 GenTreePtr fgInsertStmtListAfter(BasicBlock* block, GenTreePtr stmtAfter, GenTreePtr stmtList);
4506 GenTreePtr fgMorphSplitTree(GenTree** splitPoint, GenTree* stmt, BasicBlock* blk);
4508 // Create a new temporary variable to hold the result of *ppTree,
4509 // and transform the graph accordingly.
4510 GenTree* fgInsertCommaFormTemp(GenTree** ppTree, CORINFO_CLASS_HANDLE structType = nullptr);
4511 GenTree* fgMakeMultiUse(GenTree** ppTree);
4514 // Recognize a bitwise rotation pattern and convert into a GT_ROL or a GT_ROR node.
4515 GenTreePtr fgRecognizeAndMorphBitwiseRotation(GenTreePtr tree);
4516 bool fgOperIsBitwiseRotationRoot(genTreeOps oper);
4518 //-------- Determine the order in which the trees will be evaluated -------
4520 unsigned fgTreeSeqNum;
4521 GenTree* fgTreeSeqLst;
4522 GenTree* fgTreeSeqBeg;
4524 GenTree* fgSetTreeSeq(GenTree* tree, GenTree* prev = nullptr, bool isLIR = false);
4525 void fgSetTreeSeqHelper(GenTree* tree, bool isLIR);
4526 void fgSetTreeSeqFinish(GenTreePtr tree, bool isLIR);
4527 void fgSetStmtSeq(GenTree* tree);
4528 void fgSetBlockOrder(BasicBlock* block);
4530 //------------------------- Morphing --------------------------------------
4532 unsigned fgPtrArgCntCur;
4533 unsigned fgPtrArgCntMax;
4534 hashBv* fgOutgoingArgTemps;
4535 hashBv* fgCurrentlyInUseArgTemps;
4537 bool compCanEncodePtrArgCntMax();
4539 void fgSetRngChkTarget(GenTreePtr tree, bool delay = true);
4542 void fgMoveOpsLeft(GenTreePtr tree);
4545 bool fgIsCommaThrow(GenTreePtr tree, bool forFolding = false);
4547 bool fgIsThrow(GenTreePtr tree);
4549 bool fgInDifferentRegions(BasicBlock* blk1, BasicBlock* blk2);
4550 bool fgIsBlockCold(BasicBlock* block);
4552 GenTreePtr fgMorphCastIntoHelper(GenTreePtr tree, int helper, GenTreePtr oper);
4554 GenTreePtr fgMorphIntoHelperCall(GenTreePtr tree, int helper, GenTreeArgList* args);
4556 GenTreePtr fgMorphStackArgForVarArgs(unsigned lclNum, var_types varType, unsigned lclOffs);
4558 bool fgMorphRelopToQmark(GenTreePtr tree);
4560 // A "MorphAddrContext" carries information from the surrounding context. If we are evaluating a byref address,
4561 // it is useful to know whether the address will be immediately dereferenced, or whether the address value will
4562 // be used, perhaps by passing it as an argument to a called method. This affects how null checking is done:
4563 // for sufficiently small offsets, we can rely on OS page protection to implicitly null-check addresses that we
4564 // know will be dereferenced. To know that reliance on implicit null checking is sound, we must further know that
4565 // all offsets between the top-level indirection and the bottom are constant, and that their sum is sufficiently
4566 // small; hence the other fields of MorphAddrContext.
4567 enum MorphAddrContextKind
4572 struct MorphAddrContext
4574 MorphAddrContextKind m_kind;
4575 bool m_allConstantOffsets; // Valid only for "m_kind == MACK_Ind". True iff all offsets between
4576 // top-level indirection and here have been constants.
4577 size_t m_totalOffset; // Valid only for "m_kind == MACK_Ind", and if "m_allConstantOffsets" is true.
4578 // In that case, is the sum of those constant offsets.
4580 MorphAddrContext(MorphAddrContextKind kind) : m_kind(kind), m_allConstantOffsets(true), m_totalOffset(0)
4585 // A MACK_CopyBlock context is immutable, so we can just make one of these and share it.
4586 static MorphAddrContext s_CopyBlockMAC;
4589 GenTreePtr getSIMDStructFromField(GenTreePtr tree,
4590 var_types* baseTypeOut,
4592 unsigned* simdSizeOut,
4593 bool ignoreUsedInSIMDIntrinsic = false);
4594 GenTreePtr fgMorphFieldAssignToSIMDIntrinsicSet(GenTreePtr tree);
4595 GenTreePtr fgMorphFieldToSIMDIntrinsicGet(GenTreePtr tree);
4596 bool fgMorphCombineSIMDFieldAssignments(BasicBlock* block, GenTreePtr stmt);
4597 void impMarkContiguousSIMDFieldAssignments(GenTreePtr stmt);
4599 // fgPreviousCandidateSIMDFieldAsgStmt is only used for tracking previous simd field assignment
4600 // in function: Complier::impMarkContiguousSIMDFieldAssignments.
4601 GenTreePtr fgPreviousCandidateSIMDFieldAsgStmt;
4603 #endif // FEATURE_SIMD
4604 GenTreePtr fgMorphArrayIndex(GenTreePtr tree);
4605 GenTreePtr fgMorphCast(GenTreePtr tree);
4606 GenTreePtr fgUnwrapProxy(GenTreePtr objRef);
4607 GenTreeCall* fgMorphArgs(GenTreeCall* call);
4609 void fgMakeOutgoingStructArgCopy(GenTreeCall* call,
4612 CORINFO_CLASS_HANDLE copyBlkClass FEATURE_UNIX_AMD64_STRUCT_PASSING_ONLY_ARG(
4613 const SYSTEMV_AMD64_CORINFO_STRUCT_REG_PASSING_DESCRIPTOR* structDescPtr));
4615 void fgFixupStructReturn(GenTreePtr call);
4616 GenTreePtr fgMorphLocalVar(GenTreePtr tree);
4617 bool fgAddrCouldBeNull(GenTreePtr addr);
4618 GenTreePtr fgMorphField(GenTreePtr tree, MorphAddrContext* mac);
4619 bool fgCanFastTailCall(GenTreeCall* call);
4620 void fgMorphTailCall(GenTreeCall* call);
4621 void fgMorphRecursiveFastTailCallIntoLoop(BasicBlock* block, GenTreeCall* recursiveTailCall);
4622 GenTreePtr fgAssignRecursiveCallArgToCallerParam(GenTreePtr arg,
4623 fgArgTabEntryPtr argTabEntry,
4625 IL_OFFSETX callILOffset,
4626 GenTreePtr tmpAssignmentInsertionPoint,
4627 GenTreePtr paramAssignmentInsertionPoint);
4628 static int fgEstimateCallStackSize(GenTreeCall* call);
4629 GenTreePtr fgMorphCall(GenTreeCall* call);
4630 void fgMorphCallInline(GenTreeCall* call, InlineResult* result);
4631 void fgMorphCallInlineHelper(GenTreeCall* call, InlineResult* result);
4633 void fgNoteNonInlineCandidate(GenTreeStmt* stmt, GenTreeCall* call);
4634 static fgWalkPreFn fgFindNonInlineCandidate;
4636 GenTreePtr fgOptimizeDelegateConstructor(GenTreePtr call, CORINFO_CONTEXT_HANDLE* ExactContextHnd);
4637 GenTreePtr fgMorphLeaf(GenTreePtr tree);
4638 void fgAssignSetVarDef(GenTreePtr tree);
4639 GenTreePtr fgMorphOneAsgBlockOp(GenTreePtr tree);
4640 GenTreePtr fgMorphInitBlock(GenTreePtr tree);
4641 GenTreePtr fgMorphBlkToInd(GenTreeBlk* tree, var_types type);
4642 GenTreePtr fgMorphGetStructAddr(GenTreePtr* pTree, CORINFO_CLASS_HANDLE clsHnd, bool isRValue = false);
4643 GenTreePtr fgMorphBlkNode(GenTreePtr tree, bool isDest);
4644 GenTreePtr fgMorphBlockOperand(GenTreePtr tree, var_types asgType, unsigned blockWidth, bool isDest);
4645 void fgMorphUnsafeBlk(GenTreeObj* obj);
4646 GenTreePtr fgMorphCopyBlock(GenTreePtr tree);
4647 GenTreePtr fgMorphForRegisterFP(GenTreePtr tree);
4648 GenTreePtr fgMorphSmpOp(GenTreePtr tree, MorphAddrContext* mac = nullptr);
4649 GenTreePtr fgMorphSmpOpPre(GenTreePtr tree);
4650 GenTreePtr fgMorphModToSubMulDiv(GenTreeOp* tree);
4651 GenTreePtr fgMorphSmpOpOptional(GenTreeOp* tree);
4652 GenTreePtr fgMorphRecognizeBoxNullable(GenTree* compare);
4654 GenTreePtr fgMorphToEmulatedFP(GenTreePtr tree);
4655 GenTreePtr fgMorphConst(GenTreePtr tree);
4658 GenTreePtr fgMorphTree(GenTreePtr tree, MorphAddrContext* mac = nullptr);
4661 #if LOCAL_ASSERTION_PROP
4662 void fgKillDependentAssertionsSingle(unsigned lclNum DEBUGARG(GenTreePtr tree));
4663 void fgKillDependentAssertions(unsigned lclNum DEBUGARG(GenTreePtr tree));
4665 void fgMorphTreeDone(GenTreePtr tree, GenTreePtr oldTree = nullptr DEBUGARG(int morphNum = 0));
4667 GenTreeStmt* fgMorphStmt;
4669 unsigned fgGetBigOffsetMorphingTemp(var_types type); // We cache one temp per type to be
4670 // used when morphing big offset.
4672 //----------------------- Liveness analysis -------------------------------
4674 VARSET_TP fgCurUseSet; // vars used by block (before an assignment)
4675 VARSET_TP fgCurDefSet; // vars assigned by block (before a use)
4677 MemoryKindSet fgCurMemoryUse; // True iff the current basic block uses memory.
4678 MemoryKindSet fgCurMemoryDef; // True iff the current basic block modifies memory.
4679 MemoryKindSet fgCurMemoryHavoc; // True if the current basic block is known to set memory to a "havoc" value.
4681 bool byrefStatesMatchGcHeapStates; // True iff GcHeap and ByrefExposed memory have all the same def points.
4683 void fgMarkUseDef(GenTreeLclVarCommon* tree);
4685 void fgBeginScopeLife(VARSET_TP* inScope, VarScopeDsc* var);
4686 void fgEndScopeLife(VARSET_TP* inScope, VarScopeDsc* var);
4688 void fgMarkInScope(BasicBlock* block, VARSET_VALARG_TP inScope);
4689 void fgUnmarkInScope(BasicBlock* block, VARSET_VALARG_TP unmarkScope);
4691 void fgExtendDbgScopes();
4692 void fgExtendDbgLifetimes();
4695 void fgDispDebugScopes();
4698 //-------------------------------------------------------------------------
4700 // The following keeps track of any code we've added for things like array
4701 // range checking or explicit calls to enable GC, and so on.
4706 AddCodeDsc* acdNext;
4707 BasicBlock* acdDstBlk; // block to which we jump
4709 SpecialCodeKind acdKind; // what kind of a special block is this?
4710 unsigned short acdStkLvl;
4714 static unsigned acdHelper(SpecialCodeKind codeKind);
4716 AddCodeDsc* fgAddCodeList;
4718 bool fgRngChkThrowAdded;
4719 AddCodeDsc* fgExcptnTargetCache[SCK_COUNT];
4721 BasicBlock* fgRngChkTarget(BasicBlock* block, unsigned stkDepth, SpecialCodeKind kind);
4723 BasicBlock* fgAddCodeRef(BasicBlock* srcBlk, unsigned refData, SpecialCodeKind kind, unsigned stkDepth = 0);
4726 AddCodeDsc* fgFindExcptnTarget(SpecialCodeKind kind, unsigned refData);
4729 bool fgIsCodeAdded();
4731 bool fgIsThrowHlpBlk(BasicBlock* block);
4732 unsigned fgThrowHlpBlkStkLevel(BasicBlock* block);
4734 unsigned fgBigOffsetMorphingTemps[TYP_COUNT];
4736 unsigned fgCheckInlineDepthAndRecursion(InlineInfo* inlineInfo);
4737 void fgInvokeInlineeCompiler(GenTreeCall* call, InlineResult* result);
4738 void fgInsertInlineeBlocks(InlineInfo* pInlineInfo);
4739 GenTreePtr fgInlinePrependStatements(InlineInfo* inlineInfo);
4740 void fgInlineAppendStatements(InlineInfo* inlineInfo, BasicBlock* block, GenTreePtr stmt);
4742 #if FEATURE_MULTIREG_RET
4743 GenTreePtr fgGetStructAsStructPtr(GenTreePtr tree);
4744 GenTreePtr fgAssignStructInlineeToVar(GenTreePtr child, CORINFO_CLASS_HANDLE retClsHnd);
4745 void fgAttachStructInlineeToAsg(GenTreePtr tree, GenTreePtr child, CORINFO_CLASS_HANDLE retClsHnd);
4746 #endif // FEATURE_MULTIREG_RET
4748 static fgWalkPreFn fgUpdateInlineReturnExpressionPlaceHolder;
4751 static fgWalkPreFn fgDebugCheckInlineCandidates;
4754 void fgPromoteStructs();
4755 fgWalkResult fgMorphStructField(GenTreePtr tree, fgWalkData* fgWalkPre);
4756 fgWalkResult fgMorphLocalField(GenTreePtr tree, fgWalkData* fgWalkPre);
4757 void fgMarkImplicitByRefArgs();
4758 bool fgMorphImplicitByRefArgs(GenTree** pTree, fgWalkData* fgWalkPre);
4759 static fgWalkPreFn fgMarkAddrTakenLocalsPreCB;
4760 static fgWalkPostFn fgMarkAddrTakenLocalsPostCB;
4761 void fgMarkAddressExposedLocals();
4762 bool fgNodesMayInterfere(GenTree* store, GenTree* load);
4764 // Returns true if the type of tree is of size at least "width", or if "tree" is not a
4766 bool fgFitsInOrNotLoc(GenTreePtr tree, unsigned width);
4768 // The given local variable, required to be a struct variable, is being assigned via
4769 // a "lclField", to make it masquerade as an integral type in the ABI. Make sure that
4770 // the variable is not enregistered, and is therefore not promoted independently.
4771 void fgLclFldAssign(unsigned lclNum);
4773 static fgWalkPreFn gtHasLocalsWithAddrOpCB;
4774 bool gtCanOptimizeTypeEquality(GenTreePtr tree);
4775 bool gtIsTypeHandleToRuntimeTypeHelper(GenTreePtr tree);
4776 bool gtIsActiveCSE_Candidate(GenTreePtr tree);
4779 bool fgPrintInlinedMethods;
4782 bool fgIsBigOffset(size_t offset);
4784 // The following are used when morphing special cases of integer div/mod operations and also by codegen
4785 bool fgIsSignedDivOptimizable(GenTreePtr divisor);
4786 bool fgIsUnsignedDivOptimizable(GenTreePtr divisor);
4787 bool fgIsSignedModOptimizable(GenTreePtr divisor);
4788 bool fgIsUnsignedModOptimizable(GenTreePtr divisor);
4791 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
4792 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
4796 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
4797 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
4804 LclVarDsc* optIsTrackedLocal(GenTreePtr tree);
4807 void optRemoveRangeCheck(
4808 GenTreePtr tree, GenTreePtr stmt, bool updateCSEcounts, unsigned sideEffFlags = 0, bool forceRemove = false);
4809 bool optIsRangeCheckRemovable(GenTreePtr tree);
4812 static fgWalkPreFn optValidRangeCheckIndex;
4813 static fgWalkPreFn optRemoveTreeVisitor; // Helper passed to Compiler::fgWalkAllTreesPre() to decrement the LclVar
4816 void optRemoveTree(GenTreePtr deadTree, GenTreePtr keepList);
4818 /**************************************************************************
4820 *************************************************************************/
4823 // Do hoisting for all loops.
4824 void optHoistLoopCode();
4826 // To represent sets of VN's that have already been hoisted in outer loops.
4827 typedef SimplerHashTable<ValueNum, SmallPrimitiveKeyFuncs<ValueNum>, bool, JitSimplerHashBehavior> VNToBoolMap;
4828 typedef VNToBoolMap VNSet;
4830 struct LoopHoistContext
4833 // The set of variables hoisted in the current loop (or nullptr if there are none).
4834 VNSet* m_pHoistedInCurLoop;
4837 // Value numbers of expressions that have been hoisted in parent loops in the loop nest.
4838 VNSet m_hoistedInParentLoops;
4839 // Value numbers of expressions that have been hoisted in the current (or most recent) loop in the nest.
4840 // Previous decisions on loop-invariance of value numbers in the current loop.
4841 VNToBoolMap m_curLoopVnInvariantCache;
4843 VNSet* GetHoistedInCurLoop(Compiler* comp)
4845 if (m_pHoistedInCurLoop == nullptr)
4847 m_pHoistedInCurLoop = new (comp->getAllocatorLoopHoist()) VNSet(comp->getAllocatorLoopHoist());
4849 return m_pHoistedInCurLoop;
4852 VNSet* ExtractHoistedInCurLoop()
4854 VNSet* res = m_pHoistedInCurLoop;
4855 m_pHoistedInCurLoop = nullptr;
4859 LoopHoistContext(Compiler* comp)
4860 : m_pHoistedInCurLoop(nullptr)
4861 , m_hoistedInParentLoops(comp->getAllocatorLoopHoist())
4862 , m_curLoopVnInvariantCache(comp->getAllocatorLoopHoist())
4867 // Do hoisting for loop "lnum" (an index into the optLoopTable), and all loops nested within it.
4868 // Tracks the expressions that have been hoisted by containing loops by temporary recording their
4869 // value numbers in "m_hoistedInParentLoops". This set is not modified by the call.
4870 void optHoistLoopNest(unsigned lnum, LoopHoistContext* hoistCtxt);
4872 // Do hoisting for a particular loop ("lnum" is an index into the optLoopTable.)
4873 // Assumes that expressions have been hoisted in containing loops if their value numbers are in
4874 // "m_hoistedInParentLoops".
4876 void optHoistThisLoop(unsigned lnum, LoopHoistContext* hoistCtxt);
4878 // Hoist all expressions in "blk" that are invariant in loop "lnum" (an index into the optLoopTable)
4879 // outside of that loop. Exempt expressions whose value number is in "m_hoistedInParentLoops"; add VN's of hoisted
4880 // expressions to "hoistInLoop".
4881 void optHoistLoopExprsForBlock(BasicBlock* blk, unsigned lnum, LoopHoistContext* hoistCtxt);
4883 // Return true if the tree looks profitable to hoist out of loop 'lnum'.
4884 bool optIsProfitableToHoistableTree(GenTreePtr tree, unsigned lnum);
4886 // Hoist all proper sub-expressions of "tree" (which occurs in "stmt", which occurs in "blk")
4887 // that are invariant in loop "lnum" (an index into the optLoopTable)
4888 // outside of that loop. Exempt expressions whose value number is in "hoistedInParents"; add VN's of hoisted
4889 // expressions to "hoistInLoop".
4890 // Returns "true" iff "tree" is loop-invariant (wrt "lnum").
4891 // Assumes that the value of "*firstBlockAndBeforeSideEffect" indicates that we're in the first block, and before
4892 // any possible globally visible side effects. Assume is called in evaluation order, and updates this.
4893 bool optHoistLoopExprsForTree(GenTreePtr tree,
4895 LoopHoistContext* hoistCtxt,
4896 bool* firstBlockAndBeforeSideEffect,
4899 // Performs the hoisting 'tree' into the PreHeader for loop 'lnum'
4900 void optHoistCandidate(GenTreePtr tree, unsigned lnum, LoopHoistContext* hoistCtxt);
4902 // Returns true iff the ValueNum "vn" represents a value that is loop-invariant in "lnum".
4903 // Constants and init values are always loop invariant.
4904 // VNPhi's connect VN's to the SSA definition, so we can know if the SSA def occurs in the loop.
4905 bool optVNIsLoopInvariant(ValueNum vn, unsigned lnum, VNToBoolMap* recordedVNs);
4907 // Returns "true" iff "tree" is valid at the head of loop "lnum", in the context of the hoist substitution
4908 // "subst". If "tree" is a local SSA var, it is valid if its SSA definition occurs outside of the loop, or
4909 // if it is in the domain of "subst" (meaning that it's definition has been previously hoisted, with a "standin"
4910 // local.) If tree is a constant, it is valid. Otherwise, if it is an operator, it is valid iff its children are.
4911 bool optTreeIsValidAtLoopHead(GenTreePtr tree, unsigned lnum);
4913 // If "blk" is the entry block of a natural loop, returns true and sets "*pLnum" to the index of the loop
4914 // in the loop table.
4915 bool optBlockIsLoopEntry(BasicBlock* blk, unsigned* pLnum);
4917 // Records the set of "side effects" of all loops: fields (object instance and static)
4918 // written to, and SZ-array element type equivalence classes updated.
4919 void optComputeLoopSideEffects();
4922 // Requires "lnum" to be the index of an outermost loop in the loop table. Traverses the body of that loop,
4923 // including all nested loops, and records the set of "side effects" of the loop: fields (object instance and
4924 // static) written to, and SZ-array element type equivalence classes updated.
4925 void optComputeLoopNestSideEffects(unsigned lnum);
4927 // Add the side effects of "blk" (which is required to be within a loop) to all loops of which it is a part.
4928 void optComputeLoopSideEffectsOfBlock(BasicBlock* blk);
4930 // Hoist the expression "expr" out of loop "lnum".
4931 void optPerformHoistExpr(GenTreePtr expr, unsigned lnum);
4934 void optOptimizeBools();
4937 GenTree* optIsBoolCond(GenTree* condBranch, GenTree** compPtr, bool* boolPtr);
4939 void optOptimizeBoolsGcStress(BasicBlock* condBlock);
4942 void optOptimizeLayout(); // Optimize the BasicBlock layout of the method
4944 void optOptimizeLoops(); // for "while-do" loops duplicates simple loop conditions and transforms
4945 // the loop into a "do-while" loop
4946 // Also finds all natural loops and records them in the loop table
4948 // Optionally clone loops in the loop table.
4949 void optCloneLoops();
4951 // Clone loop "loopInd" in the loop table.
4952 void optCloneLoop(unsigned loopInd, LoopCloneContext* context);
4954 // Ensure that loop "loopInd" has a unique head block. (If the existing entry has
4955 // non-loop predecessors other than the head entry, create a new, empty block that goes (only) to the entry,
4956 // and redirects the preds of the entry to this new block.) Sets the weight of the newly created block to
4958 void optEnsureUniqueHead(unsigned loopInd, unsigned ambientWeight);
4960 void optUnrollLoops(); // Unrolls loops (needs to have cost info)
4963 // This enumeration describes what is killed by a call.
4967 CALLINT_NONE, // no interference (most helpers)
4968 CALLINT_REF_INDIRS, // kills GC ref indirections (SETFIELD OBJ)
4969 CALLINT_SCL_INDIRS, // kills non GC ref indirections (SETFIELD non-OBJ)
4970 CALLINT_ALL_INDIRS, // kills both GC ref and non GC ref indirections (SETFIELD STRUCT)
4971 CALLINT_ALL, // kills everything (normal method call)
4975 // A "LoopDsc" describes a ("natural") loop. We (currently) require the body of a loop to be a contiguous (in
4976 // bbNext order) sequence of basic blocks. (At times, we may require the blocks in a loop to be "properly numbered"
4977 // in bbNext order; we use comparisons on the bbNum to decide order.)
4978 // The blocks that define the body are
4979 // first <= top <= entry <= bottom .
4980 // The "head" of the loop is a block outside the loop that has "entry" as a successor. We only support loops with a
4981 // single 'head' block. The meanings of these blocks are given in the definitions below. Also see the picture at
4982 // Compiler::optFindNaturalLoops().
4985 BasicBlock* lpHead; // HEAD of the loop (not part of the looping of the loop) -- has ENTRY as a successor.
4986 BasicBlock* lpFirst; // FIRST block (in bbNext order) reachable within this loop. (May be part of a nested
4987 // loop, but not the outer loop.)
4988 BasicBlock* lpTop; // loop TOP (the back edge from lpBottom reaches here) (in most cases FIRST and TOP are the
4990 BasicBlock* lpEntry; // the ENTRY in the loop (in most cases TOP or BOTTOM)
4991 BasicBlock* lpBottom; // loop BOTTOM (from here we have a back edge to the TOP)
4992 BasicBlock* lpExit; // if a single exit loop this is the EXIT (in most cases BOTTOM)
4994 callInterf lpAsgCall; // "callInterf" for calls in the loop
4995 ALLVARSET_TP lpAsgVars; // set of vars assigned within the loop (all vars, not just tracked)
4996 varRefKinds lpAsgInds : 8; // set of inds modified within the loop
4998 unsigned short lpFlags; // Mask of the LPFLG_* constants
5000 unsigned char lpExitCnt; // number of exits from the loop
5002 unsigned char lpParent; // The index of the most-nested loop that completely contains this one,
5003 // or else BasicBlock::NOT_IN_LOOP if no such loop exists.
5004 unsigned char lpChild; // The index of a nested loop, or else BasicBlock::NOT_IN_LOOP if no child exists.
5005 // (Actually, an "immediately" nested loop --
5006 // no other child of this loop is a parent of lpChild.)
5007 unsigned char lpSibling; // The index of another loop that is an immediate child of lpParent,
5008 // or else BasicBlock::NOT_IN_LOOP. One can enumerate all the children of a loop
5009 // by following "lpChild" then "lpSibling" links.
5011 #define LPFLG_DO_WHILE 0x0001 // it's a do-while loop (i.e ENTRY is at the TOP)
5012 #define LPFLG_ONE_EXIT 0x0002 // the loop has only one exit
5014 #define LPFLG_ITER 0x0004 // for (i = icon or lclVar; test_condition(); i++)
5015 #define LPFLG_HOISTABLE 0x0008 // the loop is in a form that is suitable for hoisting expressions
5016 #define LPFLG_CONST 0x0010 // for (i=icon;i<icon;i++){ ... } - constant loop
5018 #define LPFLG_VAR_INIT 0x0020 // iterator is initialized with a local var (var # found in lpVarInit)
5019 #define LPFLG_CONST_INIT 0x0040 // iterator is initialized with a constant (found in lpConstInit)
5021 #define LPFLG_VAR_LIMIT 0x0100 // iterator is compared with a local var (var # found in lpVarLimit)
5022 #define LPFLG_CONST_LIMIT 0x0200 // iterator is compared with a constant (found in lpConstLimit)
5023 #define LPFLG_ARRLEN_LIMIT 0x0400 // iterator is compared with a.len or a[i].len (found in lpArrLenLimit)
5024 #define LPFLG_SIMD_LIMIT 0x0080 // iterator is compared with Vector<T>.Count (found in lpConstLimit)
5026 #define LPFLG_HAS_PREHEAD 0x0800 // lpHead is known to be a preHead for this loop
5027 #define LPFLG_REMOVED 0x1000 // has been removed from the loop table (unrolled or optimized away)
5028 #define LPFLG_DONT_UNROLL 0x2000 // do not unroll this loop
5030 #define LPFLG_ASGVARS_YES 0x4000 // "lpAsgVars" has been computed
5031 #define LPFLG_ASGVARS_INC 0x8000 // "lpAsgVars" is incomplete -- vars beyond those representable in an AllVarSet
5032 // type are assigned to.
5034 bool lpLoopHasMemoryHavoc[MemoryKindCount]; // The loop contains an operation that we assume has arbitrary
5035 // memory side effects. If this is set, the fields below
5036 // may not be accurate (since they become irrelevant.)
5037 bool lpContainsCall; // True if executing the loop body *may* execute a call
5039 VARSET_TP lpVarInOut; // The set of variables that are IN or OUT during the execution of this loop
5040 VARSET_TP lpVarUseDef; // The set of variables that are USE or DEF during the execution of this loop
5042 int lpHoistedExprCount; // The register count for the non-FP expressions from inside this loop that have been
5044 int lpLoopVarCount; // The register count for the non-FP LclVars that are read/written inside this loop
5045 int lpVarInOutCount; // The register count for the non-FP LclVars that are alive inside or accross this loop
5047 int lpHoistedFPExprCount; // The register count for the FP expressions from inside this loop that have been
5049 int lpLoopVarFPCount; // The register count for the FP LclVars that are read/written inside this loop
5050 int lpVarInOutFPCount; // The register count for the FP LclVars that are alive inside or accross this loop
5052 typedef SimplerHashTable<CORINFO_FIELD_HANDLE,
5053 PtrKeyFuncs<struct CORINFO_FIELD_STRUCT_>,
5055 JitSimplerHashBehavior>
5057 FieldHandleSet* lpFieldsModified; // This has entries (mappings to "true") for all static field and object
5058 // instance fields modified
5061 typedef SimplerHashTable<CORINFO_CLASS_HANDLE,
5062 PtrKeyFuncs<struct CORINFO_CLASS_STRUCT_>,
5064 JitSimplerHashBehavior>
5066 ClassHandleSet* lpArrayElemTypesModified; // Bits set indicate the set of sz array element types such that
5067 // arrays of that type are modified
5070 // Adds the variable liveness information for 'blk' to 'this' LoopDsc
5071 void AddVariableLiveness(Compiler* comp, BasicBlock* blk);
5073 inline void AddModifiedField(Compiler* comp, CORINFO_FIELD_HANDLE fldHnd);
5074 // This doesn't *always* take a class handle -- it can also take primitive types, encoded as class handles
5075 // (shifted left, with a low-order bit set to distinguish.)
5076 // Use the {Encode/Decode}ElemType methods to construct/destruct these.
5077 inline void AddModifiedElemType(Compiler* comp, CORINFO_CLASS_HANDLE structHnd);
5079 /* The following values are set only for iterator loops, i.e. has the flag LPFLG_ITER set */
5081 GenTreePtr lpIterTree; // The "i <op>= const" tree
5082 unsigned lpIterVar(); // iterator variable #
5083 int lpIterConst(); // the constant with which the iterator is incremented
5084 genTreeOps lpIterOper(); // the type of the operation on the iterator (ASG_ADD, ASG_SUB, etc.)
5085 void VERIFY_lpIterTree();
5087 var_types lpIterOperType(); // For overflow instructions
5090 int lpConstInit; // initial constant value of iterator : Valid if LPFLG_CONST_INIT
5091 unsigned lpVarInit; // initial local var number to which we initialize the iterator : Valid if
5095 /* The following is for LPFLG_ITER loops only (i.e. the loop condition is "i RELOP const or var" */
5097 GenTreePtr lpTestTree; // pointer to the node containing the loop test
5098 genTreeOps lpTestOper(); // the type of the comparison between the iterator and the limit (GT_LE, GT_GE, etc.)
5099 void VERIFY_lpTestTree();
5101 bool lpIsReversed(); // true if the iterator node is the second operand in the loop condition
5102 GenTreePtr lpIterator(); // the iterator node in the loop test
5103 GenTreePtr lpLimit(); // the limit node in the loop test
5105 int lpConstLimit(); // limit constant value of iterator - loop condition is "i RELOP const" : Valid if
5106 // LPFLG_CONST_LIMIT
5107 unsigned lpVarLimit(); // the lclVar # in the loop condition ( "i RELOP lclVar" ) : Valid if
5109 bool lpArrLenLimit(Compiler* comp, ArrIndex* index); // The array length in the loop condition ( "i RELOP
5110 // arr.len" or "i RELOP arr[i][j].len" ) : Valid if
5111 // LPFLG_ARRLEN_LIMIT
5113 // Returns "true" iff "*this" contains the blk.
5114 bool lpContains(BasicBlock* blk)
5116 return lpFirst->bbNum <= blk->bbNum && blk->bbNum <= lpBottom->bbNum;
5118 // Returns "true" iff "*this" (properly) contains the range [first, bottom] (allowing firsts
5119 // to be equal, but requiring bottoms to be different.)
5120 bool lpContains(BasicBlock* first, BasicBlock* bottom)
5122 return lpFirst->bbNum <= first->bbNum && bottom->bbNum < lpBottom->bbNum;
5125 // Returns "true" iff "*this" (properly) contains "lp2" (allowing firsts to be equal, but requiring
5126 // bottoms to be different.)
5127 bool lpContains(const LoopDsc& lp2)
5129 return lpContains(lp2.lpFirst, lp2.lpBottom);
5132 // Returns "true" iff "*this" is (properly) contained by the range [first, bottom]
5133 // (allowing firsts to be equal, but requiring bottoms to be different.)
5134 bool lpContainedBy(BasicBlock* first, BasicBlock* bottom)
5136 return first->bbNum <= lpFirst->bbNum && lpBottom->bbNum < bottom->bbNum;
5139 // Returns "true" iff "*this" is (properly) contained by "lp2"
5140 // (allowing firsts to be equal, but requiring bottoms to be different.)
5141 bool lpContainedBy(const LoopDsc& lp2)
5143 return lpContains(lp2.lpFirst, lp2.lpBottom);
5146 // Returns "true" iff "*this" is disjoint from the range [top, bottom].
5147 bool lpDisjoint(BasicBlock* first, BasicBlock* bottom)
5149 return bottom->bbNum < lpFirst->bbNum || lpBottom->bbNum < first->bbNum;
5151 // Returns "true" iff "*this" is disjoint from "lp2".
5152 bool lpDisjoint(const LoopDsc& lp2)
5154 return lpDisjoint(lp2.lpFirst, lp2.lpBottom);
5156 // Returns "true" iff the loop is well-formed (see code for defn).
5159 return lpFirst->bbNum <= lpTop->bbNum && lpTop->bbNum <= lpEntry->bbNum &&
5160 lpEntry->bbNum <= lpBottom->bbNum &&
5161 (lpHead->bbNum < lpTop->bbNum || lpHead->bbNum > lpBottom->bbNum);
5166 bool fgMightHaveLoop(); // returns true if there are any backedges
5167 bool fgHasLoops; // True if this method has any loops, set in fgComputeReachability
5170 LoopDsc optLoopTable[MAX_LOOP_NUM]; // loop descriptor table
5171 unsigned char optLoopCount; // number of tracked loops
5174 unsigned optCallCount; // number of calls made in the method
5175 unsigned optIndirectCallCount; // number of virtual, interface and indirect calls made in the method
5176 unsigned optNativeCallCount; // number of Pinvoke/Native calls made in the method
5177 unsigned optLoopsCloned; // number of loops cloned in the current method.
5180 unsigned optFindLoopNumberFromBeginBlock(BasicBlock* begBlk);
5181 void optPrintLoopInfo(unsigned loopNum,
5183 BasicBlock* lpFirst,
5185 BasicBlock* lpEntry,
5186 BasicBlock* lpBottom,
5187 unsigned char lpExitCnt,
5189 unsigned parentLoop = BasicBlock::NOT_IN_LOOP);
5190 void optPrintLoopInfo(unsigned lnum);
5191 void optPrintLoopRecording(unsigned lnum);
5193 void optCheckPreds();
5196 void optSetBlockWeights();
5198 void optMarkLoopBlocks(BasicBlock* begBlk, BasicBlock* endBlk, bool excludeEndBlk);
5200 void optUnmarkLoopBlocks(BasicBlock* begBlk, BasicBlock* endBlk);
5202 void optUpdateLoopsBeforeRemoveBlock(BasicBlock* block, bool skipUnmarkLoop = false);
5204 bool optIsLoopTestEvalIntoTemp(GenTreePtr test, GenTreePtr* newTest);
5205 unsigned optIsLoopIncrTree(GenTreePtr incr);
5206 bool optCheckIterInLoopTest(unsigned loopInd, GenTreePtr test, BasicBlock* from, BasicBlock* to, unsigned iterVar);
5207 bool optComputeIterInfo(GenTreePtr incr, BasicBlock* from, BasicBlock* to, unsigned* pIterVar);
5208 bool optPopulateInitInfo(unsigned loopInd, GenTreePtr init, unsigned iterVar);
5209 bool optExtractInitTestIncr(BasicBlock* head,
5214 GenTreePtr* ppIncr);
5216 void optRecordLoop(BasicBlock* head,
5222 unsigned char exitCnt);
5224 void optFindNaturalLoops();
5226 // Ensures that all the loops in the loop nest rooted at "loopInd" (an index into the loop table) are 'canonical' --
5227 // each loop has a unique "top." Returns "true" iff the flowgraph has been modified.
5228 bool optCanonicalizeLoopNest(unsigned char loopInd);
5230 // Ensures that the loop "loopInd" (an index into the loop table) is 'canonical' -- it has a unique "top,"
5231 // unshared with any other loop. Returns "true" iff the flowgraph has been modified
5232 bool optCanonicalizeLoop(unsigned char loopInd);
5234 // Requires "l1" to be a valid loop table index, and not "BasicBlock::NOT_IN_LOOP". Requires "l2" to be
5235 // a valid loop table index, or else "BasicBlock::NOT_IN_LOOP". Returns true
5236 // iff "l2" is not NOT_IN_LOOP, and "l1" contains "l2".
5237 bool optLoopContains(unsigned l1, unsigned l2);
5239 // Requires "loopInd" to be a valid index into the loop table.
5240 // Updates the loop table by changing loop "loopInd", whose head is required
5241 // to be "from", to be "to". Also performs this transformation for any
5242 // loop nested in "loopInd" that shares the same head as "loopInd".
5243 void optUpdateLoopHead(unsigned loopInd, BasicBlock* from, BasicBlock* to);
5245 // Updates the successors of "blk": if "blk2" is a successor of "blk", and there is a mapping for "blk2->blk3" in
5246 // "redirectMap", change "blk" so that "blk3" is this successor. Note that the predecessor lists are not updated.
5247 void optRedirectBlock(BasicBlock* blk, BlockToBlockMap* redirectMap);
5249 // Marks the containsCall information to "lnum" and any parent loops.
5250 void AddContainsCallAllContainingLoops(unsigned lnum);
5251 // Adds the variable liveness information from 'blk' to "lnum" and any parent loops.
5252 void AddVariableLivenessAllContainingLoops(unsigned lnum, BasicBlock* blk);
5253 // Adds "fldHnd" to the set of modified fields of "lnum" and any parent loops.
5254 void AddModifiedFieldAllContainingLoops(unsigned lnum, CORINFO_FIELD_HANDLE fldHnd);
5255 // Adds "elemType" to the set of modified array element types of "lnum" and any parent loops.
5256 void AddModifiedElemTypeAllContainingLoops(unsigned lnum, CORINFO_CLASS_HANDLE elemType);
5258 // Requires that "from" and "to" have the same "bbJumpKind" (perhaps because "to" is a clone
5259 // of "from".) Copies the jump destination from "from" to "to".
5260 void optCopyBlkDest(BasicBlock* from, BasicBlock* to);
5262 // The depth of the loop described by "lnum" (an index into the loop table.) (0 == top level)
5263 unsigned optLoopDepth(unsigned lnum)
5265 unsigned par = optLoopTable[lnum].lpParent;
5266 if (par == BasicBlock::NOT_IN_LOOP)
5272 return 1 + optLoopDepth(par);
5276 void fgOptWhileLoop(BasicBlock* block);
5278 bool optComputeLoopRep(int constInit,
5281 genTreeOps iterOper,
5283 genTreeOps testOper,
5286 unsigned* iterCount);
5287 #if FEATURE_STACK_FP_X87
5290 VARSET_TP optAllFloatVars; // mask of all tracked FP variables
5291 VARSET_TP optAllFPregVars; // mask of all enregistered FP variables
5292 VARSET_TP optAllNonFPvars; // mask of all tracked non-FP variables
5293 #endif // FEATURE_STACK_FP_X87
5296 static fgWalkPreFn optIsVarAssgCB;
5299 bool optIsVarAssigned(BasicBlock* beg, BasicBlock* end, GenTreePtr skip, unsigned var);
5301 bool optIsVarAssgLoop(unsigned lnum, unsigned var);
5303 int optIsSetAssgLoop(unsigned lnum, ALLVARSET_VALARG_TP vars, varRefKinds inds = VR_NONE);
5305 bool optNarrowTree(GenTreePtr tree, var_types srct, var_types dstt, ValueNumPair vnpNarrow, bool doit);
5307 /**************************************************************************
5308 * Optimization conditions
5309 *************************************************************************/
5311 bool optFastCodeOrBlendedLoop(BasicBlock::weight_t bbWeight);
5312 bool optPentium4(void);
5313 bool optAvoidIncDec(BasicBlock::weight_t bbWeight);
5314 bool optAvoidIntMult(void);
5319 // The following is the upper limit on how many expressions we'll keep track
5320 // of for the CSE analysis.
5322 static const unsigned MAX_CSE_CNT = EXPSET_SZ;
5324 static const int MIN_CSE_COST = 2;
5326 // Keeps tracked cse indices
5327 BitVecTraits* cseTraits;
5331 /* Generic list of nodes - used by the CSE logic */
5339 typedef struct treeLst* treeLstPtr;
5343 treeStmtLst* tslNext;
5344 GenTreePtr tslTree; // tree node
5345 GenTreePtr tslStmt; // statement containing the tree
5346 BasicBlock* tslBlock; // block containing the statement
5349 typedef struct treeStmtLst* treeStmtLstPtr;
5351 // The following logic keeps track of expressions via a simple hash table.
5355 CSEdsc* csdNextInBucket; // used by the hash table
5357 unsigned csdHashValue; // the orginal hashkey
5359 unsigned csdIndex; // 1..optCSECandidateCount
5360 char csdLiveAcrossCall; // 0 or 1
5362 unsigned short csdDefCount; // definition count
5363 unsigned short csdUseCount; // use count (excluding the implicit uses at defs)
5365 unsigned csdDefWtCnt; // weighted def count
5366 unsigned csdUseWtCnt; // weighted use count (excluding the implicit uses at defs)
5368 GenTreePtr csdTree; // treenode containing the 1st occurance
5369 GenTreePtr csdStmt; // stmt containing the 1st occurance
5370 BasicBlock* csdBlock; // block containing the 1st occurance
5372 treeStmtLstPtr csdTreeList; // list of matching tree nodes: head
5373 treeStmtLstPtr csdTreeLast; // list of matching tree nodes: tail
5376 static const size_t s_optCSEhashSize;
5377 CSEdsc** optCSEhash;
5382 CSEdsc* optCSEfindDsc(unsigned index);
5383 void optUnmarkCSE(GenTreePtr tree);
5385 // user defined callback data for the tree walk function optCSE_MaskHelper()
5386 struct optCSE_MaskData
5388 EXPSET_TP CSE_defMask;
5389 EXPSET_TP CSE_useMask;
5392 // Treewalk helper for optCSE_DefMask and optCSE_UseMask
5393 static fgWalkPreFn optCSE_MaskHelper;
5395 // This function walks all the node for an given tree
5396 // and return the mask of CSE definitions and uses for the tree
5398 void optCSE_GetMaskData(GenTreePtr tree, optCSE_MaskData* pMaskData);
5400 // Given a binary tree node return true if it is safe to swap the order of evaluation for op1 and op2.
5401 bool optCSE_canSwap(GenTree* firstNode, GenTree* secondNode);
5402 bool optCSE_canSwap(GenTree* tree);
5404 static fgWalkPostFn optPropagateNonCSE;
5405 static fgWalkPreFn optHasNonCSEChild;
5407 static fgWalkPreFn optUnmarkCSEs;
5409 static int __cdecl optCSEcostCmpEx(const void* op1, const void* op2);
5410 static int __cdecl optCSEcostCmpSz(const void* op1, const void* op2);
5412 void optCleanupCSEs();
5415 void optEnsureClearCSEInfo();
5418 #endif // FEATURE_ANYCSE
5420 #if FEATURE_VALNUM_CSE
5421 /**************************************************************************
5422 * Value Number based CSEs
5423 *************************************************************************/
5426 void optOptimizeValnumCSEs();
5429 void optValnumCSE_Init();
5430 unsigned optValnumCSE_Index(GenTreePtr tree, GenTreePtr stmt);
5431 unsigned optValnumCSE_Locate();
5432 void optValnumCSE_InitDataFlow();
5433 void optValnumCSE_DataFlow();
5434 void optValnumCSE_Availablity();
5435 void optValnumCSE_Heuristic();
5436 void optValnumCSE_UnmarkCSEs(GenTreePtr deadTree, GenTreePtr keepList);
5438 #endif // FEATURE_VALNUM_CSE
5441 bool optDoCSE; // True when we have found a duplicate CSE tree
5442 bool optValnumCSE_phase; // True when we are executing the optValnumCSE_phase
5443 unsigned optCSECandidateTotal; // Grand total of CSE candidates for both Lexical and ValNum
5444 unsigned optCSECandidateCount; // Count of CSE's candidates, reset for Lexical and ValNum CSE's
5445 unsigned optCSEstart; // The first local variable number that is a CSE
5446 unsigned optCSEcount; // The total count of CSE's introduced.
5447 unsigned optCSEweight; // The weight of the current block when we are
5448 // scanning for CSE expressions
5450 bool optIsCSEcandidate(GenTreePtr tree);
5452 // lclNumIsTrueCSE returns true if the LclVar was introduced by the CSE phase of the compiler
5454 bool lclNumIsTrueCSE(unsigned lclNum) const
5456 return ((optCSEcount > 0) && (lclNum >= optCSEstart) && (lclNum < optCSEstart + optCSEcount));
5459 // lclNumIsCSE returns true if the LclVar should be treated like a CSE with regards to constant prop.
5461 bool lclNumIsCSE(unsigned lclNum) const
5463 return lvaTable[lclNum].lvIsCSE;
5467 bool optConfigDisableCSE();
5468 bool optConfigDisableCSE2();
5470 void optOptimizeCSEs();
5472 #endif // FEATURE_ANYCSE
5480 unsigned ivaVar; // Variable we are interested in, or -1
5481 ALLVARSET_TP ivaMaskVal; // Set of variables assigned to. This is a set of all vars, not tracked vars.
5482 bool ivaMaskIncomplete; // Variables not representable in ivaMaskVal were assigned to.
5483 varRefKinds ivaMaskInd; // What kind of indirect assignments are there?
5484 callInterf ivaMaskCall; // What kind of calls are there?
5487 static callInterf optCallInterf(GenTreePtr call);
5490 // VN based copy propagation.
5491 typedef ArrayStack<GenTreePtr> GenTreePtrStack;
5492 typedef SimplerHashTable<unsigned, SmallPrimitiveKeyFuncs<unsigned>, GenTreePtrStack*, JitSimplerHashBehavior>
5493 LclNumToGenTreePtrStack;
5495 // Kill set to track variables with intervening definitions.
5496 VARSET_TP optCopyPropKillSet;
5498 // Copy propagation functions.
5499 void optCopyProp(BasicBlock* block, GenTreePtr stmt, GenTreePtr tree, LclNumToGenTreePtrStack* curSsaName);
5500 void optBlockCopyPropPopStacks(BasicBlock* block, LclNumToGenTreePtrStack* curSsaName);
5501 void optBlockCopyProp(BasicBlock* block, LclNumToGenTreePtrStack* curSsaName);
5502 bool optIsSsaLocal(GenTreePtr tree);
5503 int optCopyProp_LclVarScore(LclVarDsc* lclVarDsc, LclVarDsc* copyVarDsc, bool preferOp2);
5504 void optVnCopyProp();
5506 /**************************************************************************
5507 * Early value propagation
5508 *************************************************************************/
5514 SSAName(unsigned lvNum, unsigned ssaNum) : m_lvNum(lvNum), m_ssaNum(ssaNum)
5518 static unsigned GetHashCode(SSAName ssaNm)
5520 return (ssaNm.m_lvNum << 16) | (ssaNm.m_ssaNum);
5523 static bool Equals(SSAName ssaNm1, SSAName ssaNm2)
5525 return (ssaNm1.m_lvNum == ssaNm2.m_lvNum) && (ssaNm1.m_ssaNum == ssaNm2.m_ssaNum);
5529 #define OMF_HAS_NEWARRAY 0x00000001 // Method contains 'new' of an array
5530 #define OMF_HAS_NEWOBJ 0x00000002 // Method contains 'new' of an object type.
5531 #define OMF_HAS_ARRAYREF 0x00000004 // Method contains array element loads or stores.
5532 #define OMF_HAS_VTABLEREF 0x00000008 // Method contains method table reference.
5533 #define OMF_HAS_NULLCHECK 0x00000010 // Method contains null check.
5535 unsigned optMethodFlags;
5537 // Recursion bound controls how far we can go backwards tracking for a SSA value.
5538 // No throughput diff was found with backward walk bound between 3-8.
5539 static const int optEarlyPropRecurBound = 5;
5541 enum class optPropKind
5549 bool gtIsVtableRef(GenTreePtr tree);
5550 GenTreePtr getArrayLengthFromAllocation(GenTreePtr tree);
5551 GenTreePtr getObjectHandleNodeFromAllocation(GenTreePtr tree);
5552 GenTreePtr optPropGetValueRec(unsigned lclNum, unsigned ssaNum, optPropKind valueKind, int walkDepth);
5553 GenTreePtr optPropGetValue(unsigned lclNum, unsigned ssaNum, optPropKind valueKind);
5554 bool optEarlyPropRewriteTree(GenTreePtr tree);
5555 bool optDoEarlyPropForBlock(BasicBlock* block);
5556 bool optDoEarlyPropForFunc();
5557 void optEarlyProp();
5558 void optFoldNullCheck(GenTreePtr tree);
5559 bool optCanMoveNullCheckPastTree(GenTreePtr tree, bool isInsideTry);
5562 /**************************************************************************
5563 * Value/Assertion propagation
5564 *************************************************************************/
5566 // Data structures for assertion prop
5567 BitVecTraits* apTraits;
5571 enum optAssertionKind
5586 O1K_ARRLEN_OPER_BND,
5587 O1K_ARRLEN_LOOP_BND,
5588 O1K_CONSTANT_LOOP_BND,
5609 optAssertionKind assertionKind;
5612 unsigned lclNum; // assigned to or property of this local var number
5620 struct AssertionDscOp1
5622 optOp1Kind kind; // a normal LclVar, or Exact-type or Subtype
5629 struct AssertionDscOp2
5631 optOp2Kind kind; // a const or copy assignment
5635 ssize_t iconVal; // integer
5636 unsigned iconFlags; // gtFlags
5638 struct Range // integer subrange
5652 bool IsArrLenArithBound()
5654 return ((assertionKind == OAK_EQUAL || assertionKind == OAK_NOT_EQUAL) && op1.kind == O1K_ARRLEN_OPER_BND);
5656 bool IsArrLenBound()
5658 return ((assertionKind == OAK_EQUAL || assertionKind == OAK_NOT_EQUAL) && op1.kind == O1K_ARRLEN_LOOP_BND);
5660 bool IsConstantBound()
5662 return ((assertionKind == OAK_EQUAL || assertionKind == OAK_NOT_EQUAL) &&
5663 op1.kind == O1K_CONSTANT_LOOP_BND);
5665 bool IsBoundsCheckNoThrow()
5667 return ((assertionKind == OAK_NO_THROW) && (op1.kind == O1K_ARR_BND));
5670 bool IsCopyAssertion()
5672 return ((assertionKind == OAK_EQUAL) && (op1.kind == O1K_LCLVAR) && (op2.kind == O2K_LCLVAR_COPY));
5675 static bool SameKind(AssertionDsc* a1, AssertionDsc* a2)
5677 return a1->assertionKind == a2->assertionKind && a1->op1.kind == a2->op1.kind &&
5678 a1->op2.kind == a2->op2.kind;
5681 static bool ComplementaryKind(optAssertionKind kind, optAssertionKind kind2)
5683 if (kind == OAK_EQUAL)
5685 return kind2 == OAK_NOT_EQUAL;
5687 else if (kind == OAK_NOT_EQUAL)
5689 return kind2 == OAK_EQUAL;
5694 static ssize_t GetLowerBoundForIntegralType(var_types type)
5714 static ssize_t GetUpperBoundForIntegralType(var_types type)
5738 bool HasSameOp1(AssertionDsc* that, bool vnBased)
5740 return (op1.kind == that->op1.kind) &&
5741 ((vnBased && (op1.vn == that->op1.vn)) || (!vnBased && (op1.lcl.lclNum == that->op1.lcl.lclNum)));
5744 bool HasSameOp2(AssertionDsc* that, bool vnBased)
5746 if (op2.kind != that->op2.kind)
5752 case O2K_IND_CNS_INT:
5754 return ((op2.u1.iconVal == that->op2.u1.iconVal) && (op2.u1.iconFlags == that->op2.u1.iconFlags));
5756 case O2K_CONST_LONG:
5757 return (op2.lconVal == that->op2.lconVal);
5759 case O2K_CONST_DOUBLE:
5760 // exact match because of positive and negative zero.
5761 return (memcmp(&op2.dconVal, &that->op2.dconVal, sizeof(double)) == 0);
5763 case O2K_LCLVAR_COPY:
5765 return (op2.lcl.lclNum == that->op2.lcl.lclNum) &&
5766 (!vnBased || op2.lcl.ssaNum == that->op2.lcl.ssaNum);
5769 return ((op2.u2.loBound == that->op2.u2.loBound) && (op2.u2.hiBound == that->op2.u2.hiBound));
5772 // we will return false
5776 assert(!"Unexpected value for op2.kind in AssertionDsc.");
5782 bool Complementary(AssertionDsc* that, bool vnBased)
5784 return ComplementaryKind(assertionKind, that->assertionKind) && HasSameOp1(that, vnBased) &&
5785 HasSameOp2(that, vnBased);
5788 bool Equals(AssertionDsc* that, bool vnBased)
5790 return (assertionKind == that->assertionKind) && HasSameOp1(that, vnBased) && HasSameOp2(that, vnBased);
5794 typedef unsigned short AssertionIndex;
5797 static fgWalkPreFn optAddCopiesCallback;
5798 static fgWalkPreFn optVNAssertionPropCurStmtVisitor;
5799 unsigned optAddCopyLclNum;
5800 GenTreePtr optAddCopyAsgnNode;
5802 bool optLocalAssertionProp; // indicates that we are performing local assertion prop
5803 bool optAssertionPropagated; // set to true if we modified the trees
5804 bool optAssertionPropagatedCurrentStmt;
5806 GenTreePtr optAssertionPropCurrentTree;
5808 AssertionIndex* optComplementaryAssertionMap;
5809 ExpandArray<ASSERT_TP>* optAssertionDep; // table that holds dependent assertions (assertions
5810 // using the value of a local var) for each local var
5811 AssertionDsc* optAssertionTabPrivate; // table that holds info about value assignments
5812 AssertionIndex optAssertionCount; // total number of assertions in the assertion table
5813 AssertionIndex optMaxAssertionCount;
5816 void optVnNonNullPropCurStmt(BasicBlock* block, GenTreePtr stmt, GenTreePtr tree);
5817 fgWalkResult optVNConstantPropCurStmt(BasicBlock* block, GenTreePtr stmt, GenTreePtr tree);
5818 GenTreePtr optVNConstantPropOnRelOp(GenTreePtr tree);
5819 GenTreePtr optVNConstantPropOnJTrue(BasicBlock* block, GenTreePtr stmt, GenTreePtr test);
5820 GenTreePtr optVNConstantPropOnTree(BasicBlock* block, GenTreePtr stmt, GenTreePtr tree);
5821 GenTreePtr optPrepareTreeForReplacement(GenTreePtr extractTree, GenTreePtr replaceTree);
5823 AssertionIndex GetAssertionCount()
5825 return optAssertionCount;
5827 ASSERT_TP* bbJtrueAssertionOut;
5828 typedef SimplerHashTable<ValueNum, SmallPrimitiveKeyFuncs<ValueNum>, ASSERT_TP, JitSimplerHashBehavior>
5829 ValueNumToAssertsMap;
5830 ValueNumToAssertsMap* optValueNumToAsserts;
5832 static const AssertionIndex NO_ASSERTION_INDEX = 0;
5834 // Assertion prop helpers.
5835 ASSERT_TP& GetAssertionDep(unsigned lclNum);
5836 AssertionDsc* optGetAssertion(AssertionIndex assertIndex);
5837 void optAssertionInit(bool isLocalProp);
5838 void optAssertionTraitsInit(AssertionIndex assertionCount);
5839 #if LOCAL_ASSERTION_PROP
5840 void optAssertionReset(AssertionIndex limit);
5841 void optAssertionRemove(AssertionIndex index);
5844 // Assertion prop data flow functions.
5845 void optAssertionPropMain();
5846 GenTreePtr optVNAssertionPropCurStmt(BasicBlock* block, GenTreePtr stmt);
5847 bool optIsTreeKnownIntValue(bool vnBased, GenTreePtr tree, ssize_t* pConstant, unsigned* pIconFlags);
5848 ASSERT_TP* optInitAssertionDataflowFlags();
5849 ASSERT_TP* optComputeAssertionGen();
5851 // Assertion Gen functions.
5852 void optAssertionGen(GenTreePtr tree);
5853 AssertionIndex optAssertionGenPhiDefn(GenTreePtr tree);
5854 AssertionIndex optCreateJTrueBoundsAssertion(GenTreePtr tree);
5855 AssertionIndex optAssertionGenJtrue(GenTreePtr tree);
5856 AssertionIndex optCreateJtrueAssertions(GenTreePtr op1, GenTreePtr op2, Compiler::optAssertionKind assertionKind);
5857 AssertionIndex optFindComplementary(AssertionIndex assertionIndex);
5858 void optMapComplementary(AssertionIndex assertionIndex, AssertionIndex index);
5860 // Assertion creation functions.
5861 AssertionIndex optCreateAssertion(GenTreePtr op1, GenTreePtr op2, optAssertionKind assertionKind);
5862 AssertionIndex optCreateAssertion(GenTreePtr op1,
5864 optAssertionKind assertionKind,
5865 AssertionDsc* assertion);
5866 void optCreateComplementaryAssertion(AssertionIndex assertionIndex, GenTreePtr op1, GenTreePtr op2);
5868 bool optAssertionVnInvolvesNan(AssertionDsc* assertion);
5869 AssertionIndex optAddAssertion(AssertionDsc* assertion);
5870 void optAddVnAssertionMapping(ValueNum vn, AssertionIndex index);
5872 void optPrintVnAssertionMapping();
5874 ASSERT_TP optGetVnMappedAssertions(ValueNum vn);
5876 // Used for respective assertion propagations.
5877 AssertionIndex optAssertionIsSubrange(GenTreePtr tree, var_types toType, ASSERT_VALARG_TP assertions);
5878 AssertionIndex optAssertionIsSubtype(GenTreePtr tree, GenTreePtr methodTableArg, ASSERT_VALARG_TP assertions);
5879 AssertionIndex optAssertionIsNonNullInternal(GenTreePtr op, ASSERT_VALARG_TP assertions);
5880 bool optAssertionIsNonNull(GenTreePtr op,
5881 ASSERT_VALARG_TP assertions DEBUGARG(bool* pVnBased) DEBUGARG(AssertionIndex* pIndex));
5883 // Used for Relop propagation.
5884 AssertionIndex optGlobalAssertionIsEqualOrNotEqual(ASSERT_VALARG_TP assertions, GenTreePtr op1, GenTreePtr op2);
5885 AssertionIndex optLocalAssertionIsEqualOrNotEqual(
5886 optOp1Kind op1Kind, unsigned lclNum, optOp2Kind op2Kind, ssize_t cnsVal, ASSERT_VALARG_TP assertions);
5888 // Assertion prop for lcl var functions.
5889 bool optAssertionProp_LclVarTypeCheck(GenTreePtr tree, LclVarDsc* lclVarDsc, LclVarDsc* copyVarDsc);
5890 GenTreePtr optCopyAssertionProp(AssertionDsc* curAssertion,
5892 GenTreePtr stmt DEBUGARG(AssertionIndex index));
5893 GenTreePtr optConstantAssertionProp(AssertionDsc* curAssertion,
5894 const GenTreePtr tree,
5895 const GenTreePtr stmt DEBUGARG(AssertionIndex index));
5896 GenTreePtr optVnConstantAssertionProp(const GenTreePtr tree, const GenTreePtr stmt);
5898 // Assertion propagation functions.
5899 GenTreePtr optAssertionProp(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
5900 GenTreePtr optAssertionProp_LclVar(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
5901 GenTreePtr optAssertionProp_Ind(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
5902 GenTreePtr optAssertionProp_Cast(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
5903 GenTreePtr optAssertionProp_Call(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
5904 GenTreePtr optAssertionProp_RelOp(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
5905 GenTreePtr optAssertionProp_Comma(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
5906 GenTreePtr optAssertionProp_BndsChk(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
5907 GenTreePtr optAssertionPropGlobal_RelOp(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
5908 GenTreePtr optAssertionPropLocal_RelOp(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
5909 GenTreePtr optAssertionProp_Update(const GenTreePtr newTree, const GenTreePtr tree, const GenTreePtr stmt);
5910 GenTreePtr optNonNullAssertionProp_Call(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
5912 // Implied assertion functions.
5913 void optImpliedAssertions(AssertionIndex assertionIndex, ASSERT_TP& activeAssertions);
5914 void optImpliedByTypeOfAssertions(ASSERT_TP& activeAssertions);
5915 void optImpliedByCopyAssertion(AssertionDsc* copyAssertion, AssertionDsc* depAssertion, ASSERT_TP& result);
5916 void optImpliedByConstAssertion(AssertionDsc* curAssertion, ASSERT_TP& result);
5918 ASSERT_VALRET_TP optNewFullAssertSet();
5919 ASSERT_VALRET_TP optNewEmptyAssertSet();
5922 void optPrintAssertion(AssertionDsc* newAssertion, AssertionIndex assertionIndex = 0);
5923 void optDebugCheckAssertion(AssertionDsc* assertion);
5924 void optDebugCheckAssertions(AssertionIndex AssertionIndex);
5926 void optAddCopies();
5927 #endif // ASSERTION_PROP
5929 /**************************************************************************
5931 *************************************************************************/
5934 struct LoopCloneVisitorInfo
5936 LoopCloneContext* context;
5939 LoopCloneVisitorInfo(LoopCloneContext* context, unsigned loopNum, GenTreePtr stmt)
5940 : context(context), loopNum(loopNum), stmt(nullptr)
5945 bool optIsStackLocalInvariant(unsigned loopNum, unsigned lclNum);
5946 bool optExtractArrIndex(GenTreePtr tree, ArrIndex* result, unsigned lhsNum);
5947 bool optReconstructArrIndex(GenTreePtr tree, ArrIndex* result, unsigned lhsNum);
5948 bool optIdentifyLoopOptInfo(unsigned loopNum, LoopCloneContext* context);
5949 static fgWalkPreFn optCanOptimizeByLoopCloningVisitor;
5950 fgWalkResult optCanOptimizeByLoopCloning(GenTreePtr tree, LoopCloneVisitorInfo* info);
5951 void optObtainLoopCloningOpts(LoopCloneContext* context);
5952 bool optIsLoopClonable(unsigned loopInd);
5954 bool optCanCloneLoops();
5957 void optDebugLogLoopCloning(BasicBlock* block, GenTreePtr insertBefore);
5959 void optPerformStaticOptimizations(unsigned loopNum, LoopCloneContext* context DEBUGARG(bool fastPath));
5960 bool optComputeDerefConditions(unsigned loopNum, LoopCloneContext* context);
5961 bool optDeriveLoopCloningConditions(unsigned loopNum, LoopCloneContext* context);
5962 BasicBlock* optInsertLoopChoiceConditions(LoopCloneContext* context,
5966 void optInsertLoopCloningStress(BasicBlock* head);
5968 #if COUNT_RANGECHECKS
5969 static unsigned optRangeChkRmv;
5970 static unsigned optRangeChkAll;
5979 #define MAX_ARRAYS 4 // a magic max number of arrays tracked for bounds check elimination
5984 RngChkDsc* rcdNextInBucket; // used by the hash table
5986 unsigned short rcdHashValue; // to make matching faster
5987 unsigned short rcdIndex; // 0..optRngChkCount-1
5989 GenTreePtr rcdTree; // the array index tree
5992 unsigned optRngChkCount;
5993 static const size_t optRngChkHashSize;
5995 ssize_t optGetArrayRefScaleAndIndex(GenTreePtr mul, GenTreePtr* pIndex DEBUGARG(bool bRngChk));
5996 GenTreePtr optFindLocalInit(BasicBlock* block, GenTreePtr local, VARSET_TP* pKilledInOut, bool* isKilledAfterInit);
5998 bool optReachWithoutCall(BasicBlock* srcBB, BasicBlock* dstBB);
6001 bool optLoopsMarked;
6004 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6005 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6009 XX Does the register allocation and puts the remaining lclVars on the stack XX
6011 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6012 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6016 #ifndef LEGACY_BACKEND
6021 #else // LEGACY_BACKEND
6026 #endif // LEGACY_BACKEND
6028 #ifdef LEGACY_BACKEND
6030 void raAssignVars(); // register allocation
6031 #endif // LEGACY_BACKEND
6033 VARSET_TP raRegVarsMask; // Set of all enregistered variables (not including FEATURE_STACK_FP_X87 enregistered
6035 regNumber raUpdateRegStateForArg(RegState* regState, LclVarDsc* argDsc);
6037 void raMarkStkVars();
6040 // Some things are used by both LSRA and regpredict allocators.
6042 FrameType rpFrameType;
6043 bool rpMustCreateEBPCalled; // Set to true after we have called rpMustCreateEBPFrame once
6045 #ifdef LEGACY_BACKEND
6046 regMaskTP rpMaskPInvokeEpilogIntf; // pinvoke epilog trashes esi/edi holding stack args needed to setup tail call's
6048 #endif // LEGACY_BACKEND
6050 bool rpMustCreateEBPFrame(INDEBUG(const char** wbReason));
6052 #if FEATURE_FP_REGALLOC
6053 enum enumConfigRegisterFP
6055 CONFIG_REGISTER_FP_NONE = 0x0,
6056 CONFIG_REGISTER_FP_CALLEE_TRASH = 0x1,
6057 CONFIG_REGISTER_FP_CALLEE_SAVED = 0x2,
6058 CONFIG_REGISTER_FP_FULL = 0x3,
6060 enumConfigRegisterFP raConfigRegisterFP();
6061 #endif // FEATURE_FP_REGALLOC
6064 regMaskTP raConfigRestrictMaskFP();
6067 #ifndef LEGACY_BACKEND
6068 LinearScanInterface* m_pLinearScan; // Linear Scan allocator
6069 #else // LEGACY_BACKEND
6070 unsigned raAvoidArgRegMask; // Mask of incoming argument registers that we may need to avoid
6071 VARSET_TP raLclRegIntf[REG_COUNT]; // variable to register interference graph
6072 bool raNewBlocks; // True is we added killing blocks for FPU registers
6073 unsigned rpPasses; // Number of passes made by the register predicter
6074 unsigned rpPassesMax; // Maximum number of passes made by the register predicter
6075 unsigned rpPassesPessimize; // Number of passes non-pessimizing made by the register predicter
6076 unsigned rpStkPredict; // Weighted count of variables were predicted STK (lower means register allocation is better)
6077 unsigned rpPredictSpillCnt; // Predicted number of integer spill tmps for the current tree
6078 regMaskTP rpPredictAssignMask; // Mask of registers to consider in rpPredictAssignRegVars()
6079 VARSET_TP rpLastUseVars; // Set of last use variables in rpPredictTreeRegUse
6080 VARSET_TP rpUseInPlace; // Set of variables that we used in place
6081 int rpAsgVarNum; // VarNum for the target of GT_ASG node
6082 bool rpPredictAssignAgain; // Must rerun the rpPredictAssignRegVars()
6083 bool rpAddedVarIntf; // Set to true if we need to add a new var intf
6084 bool rpLostEnreg; // Set to true if we lost an enregister var that had lvDependReg set
6085 bool rpReverseEBPenreg; // Decided to reverse the enregistration of EBP
6087 bool rpRegAllocDone; // Set to true after we have completed register allocation
6089 regMaskTP rpPredictMap[PREDICT_COUNT]; // Holds the regMaskTP for each of the enum values
6091 void raSetupArgMasks(RegState* r);
6093 const regNumber* raGetRegVarOrder(var_types regType, unsigned* wbVarOrderSize);
6095 void raDumpVarIntf(); // Dump the variable to variable interference graph
6096 void raDumpRegIntf(); // Dump the variable to register interference graph
6098 void raAdjustVarIntf();
6100 regMaskTP rpPredictRegMask(rpPredictReg predictReg, var_types type);
6102 bool rpRecordRegIntf(regMaskTP regMask, VARSET_VALARG_TP life DEBUGARG(const char* msg));
6104 bool rpRecordVarIntf(unsigned varNum, VARSET_VALARG_TP intfVar DEBUGARG(const char* msg));
6105 regMaskTP rpPredictRegPick(var_types type, rpPredictReg predictReg, regMaskTP lockedRegs);
6107 regMaskTP rpPredictGrabReg(var_types type, rpPredictReg predictReg, regMaskTP lockedRegs);
6109 static fgWalkPreFn rpMarkRegIntf;
6111 regMaskTP rpPredictAddressMode(
6112 GenTreePtr tree, var_types type, regMaskTP lockedRegs, regMaskTP rsvdRegs, GenTreePtr lenCSE);
6114 void rpPredictRefAssign(unsigned lclNum);
6116 regMaskTP rpPredictBlkAsgRegUse(GenTreePtr tree, rpPredictReg predictReg, regMaskTP lockedRegs, regMaskTP rsvdRegs);
6118 regMaskTP rpPredictTreeRegUse(GenTreePtr tree, rpPredictReg predictReg, regMaskTP lockedRegs, regMaskTP rsvdRegs);
6120 regMaskTP rpPredictAssignRegVars(regMaskTP regAvail);
6122 void rpPredictRegUse(); // Entry point
6124 unsigned raPredictTreeRegUse(GenTreePtr tree);
6125 unsigned raPredictListRegUse(GenTreePtr list);
6127 void raSetRegVarOrder(var_types regType,
6128 regNumber* customVarOrder,
6129 unsigned* customVarOrderSize,
6131 regMaskTP avoidReg);
6133 // We use (unsigned)-1 as an uninitialized sentinel for rpStkPredict and
6134 // also as the maximum value of lvRefCntWtd. Don't allow overflow, and
6135 // saturate at UINT_MAX - 1, to avoid using the sentinel.
6136 void raAddToStkPredict(unsigned val)
6138 unsigned newStkPredict = rpStkPredict + val;
6139 if ((newStkPredict < rpStkPredict) || (newStkPredict == UINT_MAX))
6140 rpStkPredict = UINT_MAX - 1;
6142 rpStkPredict = newStkPredict;
6146 #if !FEATURE_FP_REGALLOC
6147 void raDispFPlifeInfo();
6151 regMaskTP genReturnRegForTree(GenTreePtr tree);
6152 #endif // LEGACY_BACKEND
6154 /* raIsVarargsStackArg is called by raMaskStkVars and by
6155 lvaSortByRefCount. It identifies the special case
6156 where a varargs function has a parameter passed on the
6157 stack, other than the special varargs handle. Such parameters
6158 require special treatment, because they cannot be tracked
6159 by the GC (their offsets in the stack are not known
6163 bool raIsVarargsStackArg(unsigned lclNum)
6167 LclVarDsc* varDsc = &lvaTable[lclNum];
6169 assert(varDsc->lvIsParam);
6171 return (info.compIsVarArgs && !varDsc->lvIsRegArg && (lclNum != lvaVarargsHandleArg));
6173 #else // _TARGET_X86_
6177 #endif // _TARGET_X86_
6180 #ifdef LEGACY_BACKEND
6181 // Records the current prediction, if it's better than any previous recorded prediction.
6182 void rpRecordPrediction();
6183 // Applies the best recorded prediction, if one exists and is better than the current prediction.
6184 void rpUseRecordedPredictionIfBetter();
6186 // Data members used in the methods above.
6187 unsigned rpBestRecordedStkPredict;
6188 struct VarRegPrediction
6190 bool m_isEnregistered;
6191 regNumberSmall m_regNum;
6192 regNumberSmall m_otherReg;
6194 VarRegPrediction* rpBestRecordedPrediction;
6195 #endif // LEGACY_BACKEND
6198 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6199 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6203 XX Get to the class and method info from the Execution Engine given XX
6204 XX tokens for the class and method XX
6206 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6207 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6211 /* These are the different addressing modes used to access a local var.
6212 * The JIT has to report the location of the locals back to the EE
6213 * for debugging purposes.
6219 VLT_REG_BYREF, // this type is currently only used for value types on X64
6222 VLT_STK_BYREF, // this type is currently only used for value types on X64
6236 siVarLocType vlType;
6239 // VLT_REG/VLT_REG_FP -- Any pointer-sized enregistered value (TYP_INT, TYP_REF, etc)
6241 // VLT_REG_BYREF -- the specified register contains the address of the variable
6249 // VLT_STK -- Any 32 bit value which is on the stack
6250 // eg. [ESP+0x20], or [EBP-0x28]
6251 // VLT_STK_BYREF -- the specified stack location contains the address of the variable
6252 // eg. mov EAX, [ESP+0x20]; [EAX]
6256 regNumber vlsBaseReg;
6257 NATIVE_OFFSET vlsOffset;
6260 // VLT_REG_REG -- TYP_LONG/TYP_DOUBLE with both DWords enregistered
6269 // VLT_REG_STK -- Partly enregistered TYP_LONG/TYP_DOUBLE
6270 // eg { LowerDWord=EAX UpperDWord=[ESP+0x8] }
6278 regNumber vlrssBaseReg;
6279 NATIVE_OFFSET vlrssOffset;
6283 // VLT_STK_REG -- Partly enregistered TYP_LONG/TYP_DOUBLE
6284 // eg { LowerDWord=[ESP+0x8] UpperDWord=EAX }
6290 regNumber vlsrsBaseReg;
6291 NATIVE_OFFSET vlsrsOffset;
6297 // VLT_STK2 -- Any 64 bit value which is on the stack, in 2 successsive DWords
6298 // eg 2 DWords at [ESP+0x10]
6302 regNumber vls2BaseReg;
6303 NATIVE_OFFSET vls2Offset;
6306 // VLT_FPSTK -- enregisterd TYP_DOUBLE (on the FP stack)
6307 // eg. ST(3). Actually it is ST("FPstkHeight - vpFpStk")
6314 // VLT_FIXED_VA -- fixed argument of a varargs function.
6315 // The argument location depends on the size of the variable
6316 // arguments (...). Inspecting the VARARGS_HANDLE indicates the
6317 // location of the first arg. This argument can then be accessed
6318 // relative to the position of the first arg
6322 unsigned vlfvOffset;
6329 void* rpValue; // pointer to the in-process
6330 // location of the value.
6336 bool vlIsInReg(regNumber reg);
6337 bool vlIsOnStk(regNumber reg, signed offset);
6340 /*************************************************************************/
6345 void eeGetCallInfo(CORINFO_RESOLVED_TOKEN* pResolvedToken,
6346 CORINFO_RESOLVED_TOKEN* pConstrainedToken,
6347 CORINFO_CALLINFO_FLAGS flags,
6348 CORINFO_CALL_INFO* pResult);
6349 inline CORINFO_CALLINFO_FLAGS addVerifyFlag(CORINFO_CALLINFO_FLAGS flags);
6351 void eeGetFieldInfo(CORINFO_RESOLVED_TOKEN* pResolvedToken,
6352 CORINFO_ACCESS_FLAGS flags,
6353 CORINFO_FIELD_INFO* pResult);
6357 BOOL eeIsValueClass(CORINFO_CLASS_HANDLE clsHnd);
6359 #if defined(DEBUG) || defined(FEATURE_JIT_METHOD_PERF) || defined(FEATURE_SIMD) || defined(TRACK_LSRA_STATS)
6361 bool IsSuperPMIException(unsigned code)
6363 // Copied from NDP\clr\src\ToolBox\SuperPMI\SuperPMI-Shared\ErrorHandling.h
6365 const unsigned EXCEPTIONCODE_DebugBreakorAV = 0xe0421000;
6366 const unsigned EXCEPTIONCODE_MC = 0xe0422000;
6367 const unsigned EXCEPTIONCODE_LWM = 0xe0423000;
6368 const unsigned EXCEPTIONCODE_SASM = 0xe0424000;
6369 const unsigned EXCEPTIONCODE_SSYM = 0xe0425000;
6370 const unsigned EXCEPTIONCODE_CALLUTILS = 0xe0426000;
6371 const unsigned EXCEPTIONCODE_TYPEUTILS = 0xe0427000;
6372 const unsigned EXCEPTIONCODE_ASSERT = 0xe0440000;
6376 case EXCEPTIONCODE_DebugBreakorAV:
6377 case EXCEPTIONCODE_MC:
6378 case EXCEPTIONCODE_LWM:
6379 case EXCEPTIONCODE_SASM:
6380 case EXCEPTIONCODE_SSYM:
6381 case EXCEPTIONCODE_CALLUTILS:
6382 case EXCEPTIONCODE_TYPEUTILS:
6383 case EXCEPTIONCODE_ASSERT:
6390 const char* eeGetMethodName(CORINFO_METHOD_HANDLE hnd, const char** className);
6391 const char* eeGetMethodFullName(CORINFO_METHOD_HANDLE hnd);
6393 bool eeIsNativeMethod(CORINFO_METHOD_HANDLE method);
6394 CORINFO_METHOD_HANDLE eeGetMethodHandleForNative(CORINFO_METHOD_HANDLE method);
6397 var_types eeGetArgType(CORINFO_ARG_LIST_HANDLE list, CORINFO_SIG_INFO* sig);
6398 var_types eeGetArgType(CORINFO_ARG_LIST_HANDLE list, CORINFO_SIG_INFO* sig, bool* isPinned);
6399 unsigned eeGetArgSize(CORINFO_ARG_LIST_HANDLE list, CORINFO_SIG_INFO* sig);
6401 // VOM info, method sigs
6403 void eeGetSig(unsigned sigTok,
6404 CORINFO_MODULE_HANDLE scope,
6405 CORINFO_CONTEXT_HANDLE context,
6406 CORINFO_SIG_INFO* retSig);
6408 void eeGetCallSiteSig(unsigned sigTok,
6409 CORINFO_MODULE_HANDLE scope,
6410 CORINFO_CONTEXT_HANDLE context,
6411 CORINFO_SIG_INFO* retSig);
6413 void eeGetMethodSig(CORINFO_METHOD_HANDLE methHnd, CORINFO_SIG_INFO* retSig, CORINFO_CLASS_HANDLE owner = nullptr);
6415 // Method entry-points, instrs
6417 void* eeGetFieldAddress(CORINFO_FIELD_HANDLE handle, void*** ppIndir);
6419 CORINFO_METHOD_HANDLE eeMarkNativeTarget(CORINFO_METHOD_HANDLE method);
6421 CORINFO_EE_INFO eeInfo;
6422 bool eeInfoInitialized;
6424 CORINFO_EE_INFO* eeGetEEInfo();
6426 // Gets the offset of a SDArray's first element
6427 unsigned eeGetArrayDataOffset(var_types type);
6428 // Gets the offset of a MDArray's first element
6429 unsigned eeGetMDArrayDataOffset(var_types type, unsigned rank);
6431 GenTreePtr eeGetPInvokeCookie(CORINFO_SIG_INFO* szMetaSig);
6433 // Returns the page size for the target machine as reported by the EE.
6434 inline size_t eeGetPageSize()
6436 #if COR_JIT_EE_VERSION > 460
6437 return eeGetEEInfo()->osPageSize;
6438 #else // COR_JIT_EE_VERSION <= 460
6439 return CORINFO_PAGE_SIZE;
6440 #endif // COR_JIT_EE_VERSION > 460
6443 // Returns the frame size at which we will generate a loop to probe the stack.
6444 inline size_t getVeryLargeFrameSize()
6447 // The looping probe code is 40 bytes, whereas the straight-line probing for
6448 // the (0x2000..0x3000) case is 44, so use looping for anything 0x2000 bytes
6449 // or greater, to generate smaller code.
6450 return 2 * eeGetPageSize();
6452 return 3 * eeGetPageSize();
6456 inline bool IsTargetAbi(CORINFO_RUNTIME_ABI abi)
6458 #if COR_JIT_EE_VERSION > 460
6459 return eeGetEEInfo()->targetAbi == abi;
6461 return CORINFO_DESKTOP_ABI == abi;
6465 inline bool generateCFIUnwindCodes()
6467 #ifdef UNIX_AMD64_ABI
6468 return IsTargetAbi(CORINFO_CORERT_ABI);
6476 unsigned eeGetEHcount(CORINFO_METHOD_HANDLE handle);
6478 // Debugging support - Line number info
6480 void eeGetStmtOffsets();
6482 unsigned eeBoundariesCount;
6484 struct boundariesDsc
6486 UNATIVE_OFFSET nativeIP;
6488 unsigned sourceReason;
6489 } * eeBoundaries; // Boundaries to report to EE
6490 void eeSetLIcount(unsigned count);
6491 void eeSetLIinfo(unsigned which, UNATIVE_OFFSET offs, unsigned srcIP, bool stkEmpty, bool callInstruction);
6495 static void eeDispILOffs(IL_OFFSET offs);
6496 static void eeDispLineInfo(const boundariesDsc* line);
6497 void eeDispLineInfos();
6500 // Debugging support - Local var info
6504 unsigned eeVarsCount;
6506 struct VarResultInfo
6508 UNATIVE_OFFSET startOffset;
6509 UNATIVE_OFFSET endOffset;
6513 void eeSetLVcount(unsigned count);
6514 void eeSetLVinfo(unsigned which,
6515 UNATIVE_OFFSET startOffs,
6516 UNATIVE_OFFSET length,
6521 const siVarLoc& loc);
6525 void eeDispVar(ICorDebugInfo::NativeVarInfo* var);
6526 void eeDispVars(CORINFO_METHOD_HANDLE ftn, ULONG32 cVars, ICorDebugInfo::NativeVarInfo* vars);
6529 // ICorJitInfo wrappers
6531 void eeReserveUnwindInfo(BOOL isFunclet, BOOL isColdCode, ULONG unwindSize);
6533 void eeAllocUnwindInfo(BYTE* pHotCode,
6539 CorJitFuncKind funcKind);
6541 void eeSetEHcount(unsigned cEH);
6543 void eeSetEHinfo(unsigned EHnumber, const CORINFO_EH_CLAUSE* clause);
6545 WORD eeGetRelocTypeHint(void* target);
6547 // ICorStaticInfo wrapper functions
6549 bool eeTryResolveToken(CORINFO_RESOLVED_TOKEN* resolvedToken);
6551 #if defined(FEATURE_UNIX_AMD64_STRUCT_PASSING)
6553 static void dumpSystemVClassificationType(SystemVClassificationType ct);
6556 void eeGetSystemVAmd64PassStructInRegisterDescriptor(
6557 /*IN*/ CORINFO_CLASS_HANDLE structHnd,
6558 /*OUT*/ SYSTEMV_AMD64_CORINFO_STRUCT_REG_PASSING_DESCRIPTOR* structPassInRegDescPtr);
6559 #endif // FEATURE_UNIX_AMD64_STRUCT_PASSING
6561 template <typename ParamType>
6562 bool eeRunWithErrorTrap(void (*function)(ParamType*), ParamType* param)
6564 return eeRunWithErrorTrapImp(reinterpret_cast<void (*)(void*)>(function), reinterpret_cast<void*>(param));
6567 bool eeRunWithErrorTrapImp(void (*function)(void*), void* param);
6569 // Utility functions
6571 const char* eeGetFieldName(CORINFO_FIELD_HANDLE fieldHnd, const char** classNamePtr = nullptr);
6574 const wchar_t* eeGetCPString(size_t stringHandle);
6577 const char* eeGetClassName(CORINFO_CLASS_HANDLE clsHnd);
6579 static CORINFO_METHOD_HANDLE eeFindHelper(unsigned helper);
6580 static CorInfoHelpFunc eeGetHelperNum(CORINFO_METHOD_HANDLE method);
6582 static fgWalkPreFn CountSharedStaticHelper;
6583 static bool IsSharedStaticHelper(GenTreePtr tree);
6584 static bool IsTreeAlwaysHoistable(GenTreePtr tree);
6586 static CORINFO_FIELD_HANDLE eeFindJitDataOffs(unsigned jitDataOffs);
6587 // returns true/false if 'field' is a Jit Data offset
6588 static bool eeIsJitDataOffs(CORINFO_FIELD_HANDLE field);
6589 // returns a number < 0 if 'field' is not a Jit Data offset, otherwise the data offset (limited to 2GB)
6590 static int eeGetJitDataOffs(CORINFO_FIELD_HANDLE field);
6592 /*****************************************************************************/
6597 enum TEMP_USAGE_TYPE
6603 static var_types tmpNormalizeType(var_types type);
6604 TempDsc* tmpGetTemp(var_types type); // get temp for the given type
6605 void tmpRlsTemp(TempDsc* temp);
6606 TempDsc* tmpFindNum(int temp, TEMP_USAGE_TYPE usageType = TEMP_USAGE_FREE) const;
6609 TempDsc* tmpListBeg(TEMP_USAGE_TYPE usageType = TEMP_USAGE_FREE) const;
6610 TempDsc* tmpListNxt(TempDsc* curTemp, TEMP_USAGE_TYPE usageType = TEMP_USAGE_FREE) const;
6614 bool tmpAllFree() const;
6617 #ifndef LEGACY_BACKEND
6618 void tmpPreAllocateTemps(var_types type, unsigned count);
6619 #endif // !LEGACY_BACKEND
6622 #ifdef LEGACY_BACKEND
6623 unsigned tmpIntSpillMax; // number of int-sized spill temps
6624 unsigned tmpDoubleSpillMax; // number of double-sized spill temps
6625 #endif // LEGACY_BACKEND
6627 unsigned tmpCount; // Number of temps
6628 unsigned tmpSize; // Size of all the temps
6631 // Used by RegSet::rsSpillChk()
6632 unsigned tmpGetCount; // Temps which haven't been released yet
6635 static unsigned tmpSlot(unsigned size); // which slot in tmpFree[] or tmpUsed[] to use
6637 TempDsc* tmpFree[TEMP_MAX_SIZE / sizeof(int)];
6638 TempDsc* tmpUsed[TEMP_MAX_SIZE / sizeof(int)];
6641 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6642 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6646 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6647 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6651 CodeGenInterface* codeGen;
6653 // The following holds information about instr offsets in terms of generated code.
6657 IPmappingDsc* ipmdNext; // next line# record
6658 IL_OFFSETX ipmdILoffsx; // the instr offset
6659 emitLocation ipmdNativeLoc; // the emitter location of the native code corresponding to the IL offset
6660 bool ipmdIsLabel; // Can this code be a branch label?
6663 // Record the instr offset mapping to the generated code
6665 IPmappingDsc* genIPmappingList;
6666 IPmappingDsc* genIPmappingLast;
6668 // Managed RetVal - A side hash table meant to record the mapping from a
6669 // GT_CALL node to its IL offset. This info is used to emit sequence points
6670 // that can be used by debugger to determine the native offset at which the
6671 // managed RetVal will be available.
6673 // In fact we can store IL offset in a GT_CALL node. This was ruled out in
6674 // favor of a side table for two reasons: 1) We need IL offset for only those
6675 // GT_CALL nodes (created during importation) that correspond to an IL call and
6676 // whose return type is other than TYP_VOID. 2) GT_CALL node is a frequently used
6677 // structure and IL offset is needed only when generating debuggable code. Therefore
6678 // it is desirable to avoid memory size penalty in retail scenarios.
6679 typedef SimplerHashTable<GenTreePtr, PtrKeyFuncs<GenTree>, IL_OFFSETX, JitSimplerHashBehavior>
6680 CallSiteILOffsetTable;
6681 CallSiteILOffsetTable* genCallSite2ILOffsetMap;
6683 unsigned genReturnLocal; // Local number for the return value when applicable.
6684 BasicBlock* genReturnBB; // jumped to when not optimizing for speed.
6686 // The following properties are part of CodeGenContext. Getters are provided here for
6687 // convenience and backward compatibility, but the properties can only be set by invoking
6688 // the setter on CodeGenContext directly.
6690 __declspec(property(get = getEmitter)) emitter* genEmitter;
6691 emitter* getEmitter()
6693 return codeGen->getEmitter();
6696 const bool isFramePointerUsed()
6698 return codeGen->isFramePointerUsed();
6701 __declspec(property(get = getInterruptible, put = setInterruptible)) bool genInterruptible;
6702 bool getInterruptible()
6704 return codeGen->genInterruptible;
6706 void setInterruptible(bool value)
6708 codeGen->setInterruptible(value);
6712 const bool genDoubleAlign()
6714 return codeGen->doDoubleAlign();
6716 DWORD getCanDoubleAlign();
6717 bool shouldDoubleAlign(unsigned refCntStk,
6719 unsigned refCntWtdReg,
6720 unsigned refCntStkParam,
6721 unsigned refCntWtdStkDbl);
6722 #endif // DOUBLE_ALIGN
6724 __declspec(property(get = getFullPtrRegMap, put = setFullPtrRegMap)) bool genFullPtrRegMap;
6725 bool getFullPtrRegMap()
6727 return codeGen->genFullPtrRegMap;
6729 void setFullPtrRegMap(bool value)
6731 codeGen->setFullPtrRegMap(value);
6734 // Things that MAY belong either in CodeGen or CodeGenContext
6736 #if FEATURE_EH_FUNCLETS
6737 FuncInfoDsc* compFuncInfos;
6738 unsigned short compCurrFuncIdx;
6739 unsigned short compFuncInfoCount;
6741 unsigned short compFuncCount()
6743 assert(fgFuncletsCreated);
6744 return compFuncInfoCount;
6747 #else // !FEATURE_EH_FUNCLETS
6749 // This is a no-op when there are no funclets!
6750 void genUpdateCurrentFunclet(BasicBlock* block)
6755 FuncInfoDsc compFuncInfoRoot;
6757 static const unsigned compCurrFuncIdx = 0;
6759 unsigned short compFuncCount()
6764 #endif // !FEATURE_EH_FUNCLETS
6766 FuncInfoDsc* funCurrentFunc();
6767 void funSetCurrentFunc(unsigned funcIdx);
6768 FuncInfoDsc* funGetFunc(unsigned funcIdx);
6769 unsigned int funGetFuncIdx(BasicBlock* block);
6773 VARSET_TP compCurLife; // current live variables
6774 GenTreePtr compCurLifeTree; // node after which compCurLife has been computed
6776 template <bool ForCodeGen>
6777 void compChangeLife(VARSET_VALARG_TP newLife DEBUGARG(GenTreePtr tree));
6779 void genChangeLife(VARSET_VALARG_TP newLife DEBUGARG(GenTreePtr tree))
6781 compChangeLife</*ForCodeGen*/ true>(newLife DEBUGARG(tree));
6784 template <bool ForCodeGen>
6785 void compUpdateLife(GenTreePtr tree);
6787 // Updates "compCurLife" to its state after evaluate of "true". If "pLastUseVars" is
6788 // non-null, sets "*pLastUseVars" to the set of tracked variables for which "tree" was a last
6789 // use. (Can be more than one var in the case of dependently promoted struct vars.)
6790 template <bool ForCodeGen>
6791 void compUpdateLifeVar(GenTreePtr tree, VARSET_TP* pLastUseVars = nullptr);
6793 template <bool ForCodeGen>
6794 inline void compUpdateLife(VARSET_VALARG_TP newLife);
6796 // Gets a register mask that represent the kill set for a helper call since
6797 // not all JIT Helper calls follow the standard ABI on the target architecture.
6798 regMaskTP compHelperCallKillSet(CorInfoHelpFunc helper);
6800 // Gets a register mask that represent the kill set for a NoGC helper call.
6801 regMaskTP compNoGCHelperCallKillSet(CorInfoHelpFunc helper);
6804 // Requires that "varDsc" be a promoted struct local variable being passed as an argument, beginning at
6805 // "firstArgRegNum", which is assumed to have already been aligned to the register alignment restriction of the
6806 // struct type. Adds bits to "*pArgSkippedRegMask" for any argument registers *not* used in passing "varDsc" --
6807 // i.e., internal "holes" caused by internal alignment constraints. For example, if the struct contained an int and
6808 // a double, and we at R0 (on ARM), then R1 would be skipped, and the bit for R1 would be added to the mask.
6809 void fgAddSkippedRegsInPromotedStructArg(LclVarDsc* varDsc, unsigned firstArgRegNum, regMaskTP* pArgSkippedRegMask);
6810 #endif // _TARGET_ARM_
6812 // 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
6814 static GenTreePtr fgIsIndirOfAddrOfLocal(GenTreePtr tree);
6816 // This is indexed by GT_OBJ nodes that are address of promoted struct variables, which
6817 // have been annotated with the GTF_VAR_DEATH flag. If such a node is *not* mapped in this
6818 // table, one may assume that all the (tracked) field vars die at this point. Otherwise,
6819 // the node maps to a pointer to a VARSET_TP, containing set bits for each of the tracked field
6820 // vars of the promoted struct local that go dead at the given node (the set bits are the bits
6821 // for the tracked var indices of the field vars, as in a live var set).
6822 NodeToVarsetPtrMap* m_promotedStructDeathVars;
6824 NodeToVarsetPtrMap* GetPromotedStructDeathVars()
6826 if (m_promotedStructDeathVars == nullptr)
6828 m_promotedStructDeathVars = new (getAllocator()) NodeToVarsetPtrMap(getAllocator());
6830 return m_promotedStructDeathVars;
6834 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6835 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6839 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6840 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6843 #if !defined(__GNUC__)
6844 #pragma region Unwind information
6849 // Infrastructure functions: start/stop/reserve/emit.
6852 void unwindBegProlog();
6853 void unwindEndProlog();
6854 void unwindBegEpilog();
6855 void unwindEndEpilog();
6856 void unwindReserve();
6857 void unwindEmit(void* pHotCode, void* pColdCode);
6860 // Specific unwind information functions: called by code generation to indicate a particular
6861 // prolog or epilog unwindable instruction has been generated.
6864 void unwindPush(regNumber reg);
6865 void unwindAllocStack(unsigned size);
6866 void unwindSetFrameReg(regNumber reg, unsigned offset);
6867 void unwindSaveReg(regNumber reg, unsigned offset);
6869 #if defined(_TARGET_ARM_)
6870 void unwindPushMaskInt(regMaskTP mask);
6871 void unwindPushMaskFloat(regMaskTP mask);
6872 void unwindPopMaskInt(regMaskTP mask);
6873 void unwindPopMaskFloat(regMaskTP mask);
6874 void unwindBranch16(); // The epilog terminates with a 16-bit branch (e.g., "bx lr")
6875 void unwindNop(unsigned codeSizeInBytes); // Generate unwind NOP code. 'codeSizeInBytes' is 2 or 4 bytes. Only
6876 // called via unwindPadding().
6877 void unwindPadding(); // Generate a sequence of unwind NOP codes representing instructions between the last
6878 // instruction and the current location.
6879 #endif // _TARGET_ARM_
6881 #if defined(_TARGET_ARM64_)
6883 void unwindPadding(); // Generate a sequence of unwind NOP codes representing instructions between the last
6884 // instruction and the current location.
6885 void unwindSaveReg(regNumber reg, int offset); // str reg, [sp, #offset]
6886 void unwindSaveRegPreindexed(regNumber reg, int offset); // str reg, [sp, #offset]!
6887 void unwindSaveRegPair(regNumber reg1, regNumber reg2, int offset); // stp reg1, reg2, [sp, #offset]
6888 void unwindSaveRegPairPreindexed(regNumber reg1, regNumber reg2, int offset); // stp reg1, reg2, [sp, #offset]!
6889 void unwindSaveNext(); // unwind code: save_next
6890 void unwindReturn(regNumber reg); // ret lr
6891 #endif // defined(_TARGET_ARM64_)
6894 // Private "helper" functions for the unwind implementation.
6898 #if FEATURE_EH_FUNCLETS
6899 void unwindGetFuncLocations(FuncInfoDsc* func,
6900 bool getHotSectionData,
6901 /* OUT */ emitLocation** ppStartLoc,
6902 /* OUT */ emitLocation** ppEndLoc);
6903 #endif // FEATURE_EH_FUNCLETS
6905 void unwindReserveFunc(FuncInfoDsc* func);
6906 void unwindEmitFunc(FuncInfoDsc* func, void* pHotCode, void* pColdCode);
6908 #if defined(_TARGET_AMD64_) || (defined(_TARGET_X86_) && FEATURE_EH_FUNCLETS)
6910 void unwindReserveFuncHelper(FuncInfoDsc* func, bool isHotCode);
6911 void unwindEmitFuncHelper(FuncInfoDsc* func, void* pHotCode, void* pColdCode, bool isHotCode);
6913 #endif // _TARGET_AMD64_ || (_TARGET_X86_ && FEATURE_EH_FUNCLETS)
6915 #if defined(_TARGET_AMD64_)
6917 UNATIVE_OFFSET unwindGetCurrentOffset(FuncInfoDsc* func);
6919 void unwindBegPrologWindows();
6920 void unwindPushWindows(regNumber reg);
6921 void unwindAllocStackWindows(unsigned size);
6922 void unwindSetFrameRegWindows(regNumber reg, unsigned offset);
6923 void unwindSaveRegWindows(regNumber reg, unsigned offset);
6925 #ifdef UNIX_AMD64_ABI
6926 void unwindBegPrologCFI();
6927 void unwindPushCFI(regNumber reg);
6928 void unwindAllocStackCFI(unsigned size);
6929 void unwindSetFrameRegCFI(regNumber reg, unsigned offset);
6930 void unwindSaveRegCFI(regNumber reg, unsigned offset);
6931 int mapRegNumToDwarfReg(regNumber reg);
6932 void createCfiCode(FuncInfoDsc* func, UCHAR codeOffset, UCHAR opcode, USHORT dwarfReg, INT offset = 0);
6933 #endif // UNIX_AMD64_ABI
6934 #elif defined(_TARGET_ARM_)
6936 void unwindPushPopMaskInt(regMaskTP mask, bool useOpsize16);
6937 void unwindPushPopMaskFloat(regMaskTP mask);
6938 void unwindSplit(FuncInfoDsc* func);
6940 #endif // _TARGET_ARM_
6942 #if !defined(__GNUC__)
6943 #pragma endregion // Note: region is NOT under !defined(__GNUC__)
6947 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6948 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6952 XX Info about SIMD types, methods and the SIMD assembly (i.e. the assembly XX
6953 XX that contains the distinguished, well-known SIMD type definitions). XX
6955 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6956 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6959 // Get highest available instruction set for floating point codegen
6960 InstructionSet getFloatingPointInstructionSet()
6962 #if defined(_TARGET_XARCH_) && !defined(LEGACY_BACKEND)
6965 return InstructionSet_AVX;
6970 return InstructionSet_SSE3_4;
6974 assert(canUseSSE2());
6975 return InstructionSet_SSE2;
6977 assert(!"getFPInstructionSet() is not implemented for target arch");
6979 return InstructionSet_NONE;
6983 // Get highest available instruction set for SIMD codegen
6984 InstructionSet getSIMDInstructionSet()
6986 #if defined(_TARGET_XARCH_) && !defined(LEGACY_BACKEND)
6987 return getFloatingPointInstructionSet();
6989 assert(!"Available instruction set(s) for SIMD codegen is not defined for target arch");
6991 return InstructionSet_NONE;
6997 // Should we support SIMD intrinsics?
7000 // Have we identified any SIMD types?
7001 // This is currently used by struct promotion to avoid getting type information for a struct
7002 // field to see if it is a SIMD type, if we haven't seen any SIMD types or operations in
7004 bool _usesSIMDTypes;
7005 bool usesSIMDTypes()
7007 return _usesSIMDTypes;
7009 void setUsesSIMDTypes(bool value)
7011 _usesSIMDTypes = value;
7014 // This is a temp lclVar allocated on the stack as TYP_SIMD. It is used to implement intrinsics
7015 // that require indexed access to the individual fields of the vector, which is not well supported
7016 // by the hardware. It is allocated when/if such situations are encountered during Lowering.
7017 unsigned lvaSIMDInitTempVarNum;
7020 CORINFO_CLASS_HANDLE SIMDFloatHandle;
7021 CORINFO_CLASS_HANDLE SIMDDoubleHandle;
7022 CORINFO_CLASS_HANDLE SIMDIntHandle;
7023 CORINFO_CLASS_HANDLE SIMDUShortHandle;
7024 CORINFO_CLASS_HANDLE SIMDUByteHandle;
7025 CORINFO_CLASS_HANDLE SIMDShortHandle;
7026 CORINFO_CLASS_HANDLE SIMDByteHandle;
7027 CORINFO_CLASS_HANDLE SIMDLongHandle;
7028 CORINFO_CLASS_HANDLE SIMDUIntHandle;
7029 CORINFO_CLASS_HANDLE SIMDULongHandle;
7030 CORINFO_CLASS_HANDLE SIMDVector2Handle;
7031 CORINFO_CLASS_HANDLE SIMDVector3Handle;
7032 CORINFO_CLASS_HANDLE SIMDVector4Handle;
7033 CORINFO_CLASS_HANDLE SIMDVectorHandle;
7035 // Get the handle for a SIMD type.
7036 CORINFO_CLASS_HANDLE gtGetStructHandleForSIMD(var_types simdType, var_types simdBaseType)
7038 if (simdBaseType == TYP_FLOAT)
7043 return SIMDVector2Handle;
7045 return SIMDVector3Handle;
7047 if ((getSIMDVectorType() == TYP_SIMD32) || (SIMDVector4Handle != NO_CLASS_HANDLE))
7049 return SIMDVector4Handle;
7058 assert(simdType == getSIMDVectorType());
7059 switch (simdBaseType)
7062 return SIMDFloatHandle;
7064 return SIMDDoubleHandle;
7066 return SIMDIntHandle;
7068 return SIMDUShortHandle;
7070 return SIMDUShortHandle;
7072 return SIMDUByteHandle;
7074 return SIMDShortHandle;
7076 return SIMDByteHandle;
7078 return SIMDLongHandle;
7080 return SIMDUIntHandle;
7082 return SIMDULongHandle;
7084 assert(!"Didn't find a class handle for simdType");
7086 return NO_CLASS_HANDLE;
7090 CORINFO_METHOD_HANDLE SIMDVectorFloat_set_Item;
7091 CORINFO_METHOD_HANDLE SIMDVectorFloat_get_Length;
7092 CORINFO_METHOD_HANDLE SIMDVectorFloat_op_Addition;
7094 // Returns true if the tree corresponds to a TYP_SIMD lcl var.
7095 // Note that both SIMD vector args and locals are mared as lvSIMDType = true, but
7096 // type of an arg node is TYP_BYREF and a local node is TYP_SIMD or TYP_STRUCT.
7097 bool isSIMDTypeLocal(GenTree* tree)
7099 return tree->OperIsLocal() && lvaTable[tree->AsLclVarCommon()->gtLclNum].lvSIMDType;
7102 // Returns true if the type of the tree is a byref of TYP_SIMD
7103 bool isAddrOfSIMDType(GenTree* tree)
7105 if (tree->TypeGet() == TYP_BYREF || tree->TypeGet() == TYP_I_IMPL)
7107 switch (tree->OperGet())
7110 return varTypeIsSIMD(tree->gtGetOp1());
7112 case GT_LCL_VAR_ADDR:
7113 return lvaTable[tree->AsLclVarCommon()->gtLclNum].lvSIMDType;
7116 return isSIMDTypeLocal(tree);
7123 static bool isRelOpSIMDIntrinsic(SIMDIntrinsicID intrinsicId)
7125 return (intrinsicId == SIMDIntrinsicEqual || intrinsicId == SIMDIntrinsicLessThan ||
7126 intrinsicId == SIMDIntrinsicLessThanOrEqual || intrinsicId == SIMDIntrinsicGreaterThan ||
7127 intrinsicId == SIMDIntrinsicGreaterThanOrEqual);
7130 // Returns base type of a TYP_SIMD local.
7131 // Returns TYP_UNKNOWN if the local is not TYP_SIMD.
7132 var_types getBaseTypeOfSIMDLocal(GenTree* tree)
7134 if (isSIMDTypeLocal(tree))
7136 return lvaTable[tree->AsLclVarCommon()->gtLclNum].lvBaseType;
7142 bool isSIMDClass(CORINFO_CLASS_HANDLE clsHnd)
7144 return info.compCompHnd->isInSIMDModule(clsHnd);
7147 bool isSIMDClass(typeInfo* pTypeInfo)
7149 return pTypeInfo->IsStruct() && isSIMDClass(pTypeInfo->GetClassHandleForValueClass());
7152 // Get the base (element) type and size in bytes for a SIMD type. Returns TYP_UNKNOWN
7153 // if it is not a SIMD type or is an unsupported base type.
7154 var_types getBaseTypeAndSizeOfSIMDType(CORINFO_CLASS_HANDLE typeHnd, unsigned* sizeBytes = nullptr);
7156 var_types getBaseTypeOfSIMDType(CORINFO_CLASS_HANDLE typeHnd)
7158 return getBaseTypeAndSizeOfSIMDType(typeHnd, nullptr);
7161 // Get SIMD Intrinsic info given the method handle.
7162 // Also sets typeHnd, argCount, baseType and sizeBytes out params.
7163 const SIMDIntrinsicInfo* getSIMDIntrinsicInfo(CORINFO_CLASS_HANDLE* typeHnd,
7164 CORINFO_METHOD_HANDLE methodHnd,
7165 CORINFO_SIG_INFO* sig,
7168 var_types* baseType,
7169 unsigned* sizeBytes);
7171 // Pops and returns GenTree node from importers type stack.
7172 // Normalizes TYP_STRUCT value in case of GT_CALL, GT_RET_EXPR and arg nodes.
7173 GenTreePtr impSIMDPopStack(var_types type, bool expectAddr = false);
7175 // Create a GT_SIMD tree for a Get property of SIMD vector with a fixed index.
7176 GenTreeSIMD* impSIMDGetFixed(var_types simdType, var_types baseType, unsigned simdSize, int index);
7178 // Creates a GT_SIMD tree for Select operation
7179 GenTreePtr impSIMDSelect(CORINFO_CLASS_HANDLE typeHnd,
7181 unsigned simdVectorSize,
7186 // Creates a GT_SIMD tree for Min/Max operation
7187 GenTreePtr impSIMDMinMax(SIMDIntrinsicID intrinsicId,
7188 CORINFO_CLASS_HANDLE typeHnd,
7190 unsigned simdVectorSize,
7194 // Transforms operands and returns the SIMD intrinsic to be applied on
7195 // transformed operands to obtain given relop result.
7196 SIMDIntrinsicID impSIMDRelOp(SIMDIntrinsicID relOpIntrinsicId,
7197 CORINFO_CLASS_HANDLE typeHnd,
7198 unsigned simdVectorSize,
7199 var_types* baseType,
7203 // Creates a GT_SIMD tree for Abs intrinsic.
7204 GenTreePtr impSIMDAbs(CORINFO_CLASS_HANDLE typeHnd, var_types baseType, unsigned simdVectorSize, GenTree* op1);
7206 #if defined(_TARGET_XARCH_) && !defined(LEGACY_BACKEND)
7207 // Transforms operands and returns the SIMD intrinsic to be applied on
7208 // transformed operands to obtain == comparison result.
7209 SIMDIntrinsicID impSIMDLongRelOpEqual(CORINFO_CLASS_HANDLE typeHnd,
7210 unsigned simdVectorSize,
7214 // Transforms operands and returns the SIMD intrinsic to be applied on
7215 // transformed operands to obtain > comparison result.
7216 SIMDIntrinsicID impSIMDLongRelOpGreaterThan(CORINFO_CLASS_HANDLE typeHnd,
7217 unsigned simdVectorSize,
7221 // Transforms operands and returns the SIMD intrinsic to be applied on
7222 // transformed operands to obtain >= comparison result.
7223 SIMDIntrinsicID impSIMDLongRelOpGreaterThanOrEqual(CORINFO_CLASS_HANDLE typeHnd,
7224 unsigned simdVectorSize,
7228 // Transforms operands and returns the SIMD intrinsic to be applied on
7229 // transformed operands to obtain >= comparison result in case of int32
7230 // and small int base type vectors.
7231 SIMDIntrinsicID impSIMDIntegralRelOpGreaterThanOrEqual(
7232 CORINFO_CLASS_HANDLE typeHnd, unsigned simdVectorSize, var_types baseType, GenTree** op1, GenTree** op2);
7233 #endif // defined(_TARGET_XARCH_) && !defined(LEGACY_BACKEND)
7235 void setLclRelatedToSIMDIntrinsic(GenTreePtr tree);
7236 bool areFieldsContiguous(GenTreePtr op1, GenTreePtr op2);
7237 bool areArrayElementsContiguous(GenTreePtr op1, GenTreePtr op2);
7238 bool areArgumentsContiguous(GenTreePtr op1, GenTreePtr op2);
7239 GenTreePtr createAddressNodeForSIMDInit(GenTreePtr tree, unsigned simdSize);
7241 // check methodHnd to see if it is a SIMD method that is expanded as an intrinsic in the JIT.
7242 GenTreePtr impSIMDIntrinsic(OPCODE opcode,
7243 GenTreePtr newobjThis,
7244 CORINFO_CLASS_HANDLE clsHnd,
7245 CORINFO_METHOD_HANDLE method,
7246 CORINFO_SIG_INFO* sig,
7249 GenTreePtr getOp1ForConstructor(OPCODE opcode, GenTreePtr newobjThis, CORINFO_CLASS_HANDLE clsHnd);
7251 // Whether SIMD vector occupies part of SIMD register.
7252 // SSE2: vector2f/3f are considered sub register SIMD types.
7253 // AVX: vector2f, 3f and 4f are all considered sub register SIMD types.
7254 bool isSubRegisterSIMDType(CORINFO_CLASS_HANDLE typeHnd)
7256 unsigned sizeBytes = 0;
7257 var_types baseType = getBaseTypeAndSizeOfSIMDType(typeHnd, &sizeBytes);
7258 return (baseType == TYP_FLOAT) && (sizeBytes < getSIMDVectorRegisterByteLength());
7261 bool isSubRegisterSIMDType(GenTreeSIMD* simdNode)
7263 return (simdNode->gtSIMDSize < getSIMDVectorRegisterByteLength());
7266 // Get the type for the hardware SIMD vector.
7267 // This is the maximum SIMD type supported for this target.
7268 var_types getSIMDVectorType()
7270 #if defined(_TARGET_XARCH_) && !defined(LEGACY_BACKEND)
7277 assert(canUseSSE2());
7281 assert(!"getSIMDVectorType() unimplemented on target arch");
7286 // Get the size of the SIMD type in bytes
7287 int getSIMDTypeSizeInBytes(CORINFO_CLASS_HANDLE typeHnd)
7289 unsigned sizeBytes = 0;
7290 (void)getBaseTypeAndSizeOfSIMDType(typeHnd, &sizeBytes);
7294 // Get the the number of elements of basetype of SIMD vector given by its size and baseType
7295 static int getSIMDVectorLength(unsigned simdSize, var_types baseType);
7297 // Get the the number of elements of basetype of SIMD vector given by its type handle
7298 int getSIMDVectorLength(CORINFO_CLASS_HANDLE typeHnd);
7300 // Get preferred alignment of SIMD type.
7301 int getSIMDTypeAlignment(var_types simdType);
7303 // Get the number of bytes in a SIMD Vector for the current compilation.
7304 unsigned getSIMDVectorRegisterByteLength()
7306 #if defined(_TARGET_XARCH_) && !defined(LEGACY_BACKEND)
7309 return YMM_REGSIZE_BYTES;
7313 assert(canUseSSE2());
7314 return XMM_REGSIZE_BYTES;
7317 assert(!"getSIMDVectorRegisterByteLength() unimplemented on target arch");
7322 // The minimum and maximum possible number of bytes in a SIMD vector.
7323 unsigned int maxSIMDStructBytes()
7325 return getSIMDVectorRegisterByteLength();
7327 unsigned int minSIMDStructBytes()
7329 return emitTypeSize(TYP_SIMD8);
7332 #ifdef FEATURE_AVX_SUPPORT
7333 // (maxPossibleSIMDStructBytes is for use in a context that requires a compile-time constant.)
7334 static const unsigned maxPossibleSIMDStructBytes = 32;
7335 #else // !FEATURE_AVX_SUPPORT
7336 static const unsigned maxPossibleSIMDStructBytes = 16;
7337 #endif // !FEATURE_AVX_SUPPORT
7339 // Returns the codegen type for a given SIMD size.
7340 var_types getSIMDTypeForSize(unsigned size)
7342 var_types simdType = TYP_UNDEF;
7345 simdType = TYP_SIMD8;
7347 else if (size == 12)
7349 simdType = TYP_SIMD12;
7351 else if (size == 16)
7353 simdType = TYP_SIMD16;
7355 #ifdef FEATURE_AVX_SUPPORT
7356 else if (size == 32)
7358 simdType = TYP_SIMD32;
7360 #endif // FEATURE_AVX_SUPPORT
7363 noway_assert(!"Unexpected size for SIMD type");
7368 unsigned getSIMDInitTempVarNum()
7370 if (lvaSIMDInitTempVarNum == BAD_VAR_NUM)
7372 lvaSIMDInitTempVarNum = lvaGrabTempWithImplicitUse(false DEBUGARG("SIMDInitTempVar"));
7373 lvaTable[lvaSIMDInitTempVarNum].lvType = getSIMDVectorType();
7375 return lvaSIMDInitTempVarNum;
7378 #endif // FEATURE_SIMD
7381 //------------------------------------------------------------------------
7382 // largestEnregisterableStruct: The size in bytes of the largest struct that can be enregistered.
7384 // Notes: It is not guaranteed that the struct of this size or smaller WILL be a
7385 // candidate for enregistration.
7387 unsigned largestEnregisterableStructSize()
7390 unsigned vectorRegSize = getSIMDVectorRegisterByteLength();
7391 if (vectorRegSize > TARGET_POINTER_SIZE)
7393 return vectorRegSize;
7396 #endif // FEATURE_SIMD
7398 return TARGET_POINTER_SIZE;
7403 // These routines need not be enclosed under FEATURE_SIMD since lvIsSIMDType()
7404 // is defined for both FEATURE_SIMD and !FEATURE_SIMD apropriately. The use
7405 // of this routines also avoids the need of #ifdef FEATURE_SIMD specific code.
7407 // Is this var is of type simd struct?
7408 bool lclVarIsSIMDType(unsigned varNum)
7410 LclVarDsc* varDsc = lvaTable + varNum;
7411 return varDsc->lvIsSIMDType();
7414 // Is this Local node a SIMD local?
7415 bool lclVarIsSIMDType(GenTreeLclVarCommon* lclVarTree)
7417 return lclVarIsSIMDType(lclVarTree->gtLclNum);
7420 // Returns true if the TYP_SIMD locals on stack are aligned at their
7421 // preferred byte boundary specified by getSIMDTypeAlignment().
7423 // As per the Intel manual, the preferred alignment for AVX vectors is 32-bytes. On Amd64,
7424 // RSP/EBP is aligned at 16-bytes, therefore to align SIMD types at 32-bytes we need even
7425 // RSP/EBP to be 32-byte aligned. It is not clear whether additional stack space used in
7426 // aligning stack is worth the benefit and for now will use 16-byte alignment for AVX
7427 // 256-bit vectors with unaligned load/stores to/from memory. On x86, the stack frame
7428 // is aligned to 4 bytes. We need to extend existing support for double (8-byte) alignment
7429 // to 16 or 32 byte alignment for frames with local SIMD vars, if that is determined to be
7432 bool isSIMDTypeLocalAligned(unsigned varNum)
7434 #if defined(FEATURE_SIMD) && ALIGN_SIMD_TYPES
7435 if (lclVarIsSIMDType(varNum) && lvaTable[varNum].lvType != TYP_BYREF)
7438 int off = lvaFrameAddress(varNum, &ebpBased);
7439 // TODO-Cleanup: Can't this use the lvExactSize on the varDsc?
7440 int alignment = getSIMDTypeAlignment(lvaTable[varNum].lvType);
7441 bool isAligned = (alignment <= STACK_ALIGN) && ((off % alignment) == 0);
7444 #endif // FEATURE_SIMD
7449 // Whether SSE2 is available
7450 bool canUseSSE2() const
7452 #ifdef _TARGET_XARCH_
7453 return opts.compCanUseSSE2;
7459 // Whether SSE3, SSE3, SSE4.1 and SSE4.2 is available
7460 bool CanUseSSE3_4() const
7462 #ifdef _TARGET_XARCH_
7463 return opts.compCanUseSSE3_4;
7469 bool canUseAVX() const
7471 #ifdef FEATURE_AVX_SUPPORT
7472 return opts.compCanUseAVX;
7479 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7480 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7484 XX Generic info about the compilation and the method being compiled. XX
7485 XX It is responsible for driving the other phases. XX
7486 XX It is also responsible for all the memory management. XX
7488 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7489 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7493 Compiler* InlineeCompiler; // The Compiler instance for the inlinee
7495 InlineResult* compInlineResult; // The result of importing the inlinee method.
7497 bool compDoAggressiveInlining; // If true, mark every method as CORINFO_FLG_FORCEINLINE
7498 bool compJmpOpUsed; // Does the method do a JMP
7499 bool compLongUsed; // Does the method use TYP_LONG
7500 bool compFloatingPointUsed; // Does the method use TYP_FLOAT or TYP_DOUBLE
7501 bool compTailCallUsed; // Does the method do a tailcall
7502 bool compLocallocUsed; // Does the method use localloc.
7503 bool compQmarkUsed; // Does the method use GT_QMARK/GT_COLON
7504 bool compQmarkRationalized; // Is it allowed to use a GT_QMARK/GT_COLON node.
7505 bool compUnsafeCastUsed; // Does the method use LDIND/STIND to cast between scalar/refernce types
7507 // NOTE: These values are only reliable after
7508 // the importing is completely finished.
7510 ExpandArrayStack<GenTreePtr>* compQMarks; // The set of QMark nodes created in the current compilation, so
7511 // we can iterate over these efficiently.
7513 #if CPU_USES_BLOCK_MOVE
7514 bool compBlkOpUsed; // Does the method do a COPYBLK or INITBLK
7518 // State information - which phases have completed?
7519 // These are kept together for easy discoverability
7521 bool bRangeAllowStress;
7522 bool compCodeGenDone;
7523 int64_t compNumStatementLinksTraversed; // # of links traversed while doing debug checks
7524 bool fgNormalizeEHDone; // Has the flowgraph EH normalization phase been done?
7525 size_t compSizeEstimate; // The estimated size of the method as per `gtSetEvalOrder`.
7526 size_t compCycleEstimate; // The estimated cycle count of the method as per `gtSetEvalOrder`
7529 bool fgLocalVarLivenessDone; // Note that this one is used outside of debug.
7530 bool fgLocalVarLivenessChanged;
7532 bool compStackProbePrologDone;
7534 #ifndef LEGACY_BACKEND
7536 #endif // !LEGACY_BACKEND
7537 bool compRationalIRForm;
7539 bool compUsesThrowHelper; // There is a call to a THOROW_HELPER for the compiled method.
7541 bool compGeneratingProlog;
7542 bool compGeneratingEpilog;
7543 bool compNeedsGSSecurityCookie; // There is an unsafe buffer (or localloc) on the stack.
7544 // Insert cookie on frame and code to check the cookie, like VC++ -GS.
7545 bool compGSReorderStackLayout; // There is an unsafe buffer on the stack, reorder locals and make local
7546 // copies of susceptible parameters to avoid buffer overrun attacks through locals/params
7547 bool getNeedsGSSecurityCookie() const
7549 return compNeedsGSSecurityCookie;
7551 void setNeedsGSSecurityCookie()
7553 compNeedsGSSecurityCookie = true;
7556 FrameLayoutState lvaDoneFrameLayout; // The highest frame layout state that we've completed. During
7557 // frame layout calculations, this is the level we are currently
7560 //---------------------------- JITing options -----------------------------
7573 JitFlags* jitFlags; // all flags passed from the EE
7574 unsigned compFlags; // method attributes
7576 codeOptimize compCodeOpt; // what type of code optimizations
7580 #ifdef _TARGET_XARCH_
7581 bool compCanUseSSE2; // Allow CodeGen to use "movq XMM" instructions
7582 bool compCanUseSSE3_4; // Allow CodeGen to use SSE3, SSSE3, SSE4.1 and SSE4.2 instructions
7584 #ifdef FEATURE_AVX_SUPPORT
7585 bool compCanUseAVX; // Allow CodeGen to use AVX 256-bit vectors for SIMD operations
7586 #endif // FEATURE_AVX_SUPPORT
7587 #endif // _TARGET_XARCH_
7589 // optimize maximally and/or favor speed over size?
7591 #define DEFAULT_MIN_OPTS_CODE_SIZE 60000
7592 #define DEFAULT_MIN_OPTS_INSTR_COUNT 20000
7593 #define DEFAULT_MIN_OPTS_BB_COUNT 2000
7594 #define DEFAULT_MIN_OPTS_LV_NUM_COUNT 2000
7595 #define DEFAULT_MIN_OPTS_LV_REF_COUNT 8000
7597 // Maximun number of locals before turning off the inlining
7598 #define MAX_LV_NUM_COUNT_FOR_INLINING 512
7601 unsigned instrCount;
7602 unsigned lvRefCount;
7603 bool compMinOptsIsSet;
7605 bool compMinOptsIsUsed;
7607 inline bool MinOpts()
7609 assert(compMinOptsIsSet);
7610 compMinOptsIsUsed = true;
7613 inline bool IsMinOptsSet()
7615 return compMinOptsIsSet;
7618 inline bool MinOpts()
7622 inline bool IsMinOptsSet()
7624 return compMinOptsIsSet;
7627 inline void SetMinOpts(bool val)
7629 assert(!compMinOptsIsUsed);
7630 assert(!compMinOptsIsSet || (compMinOpts == val));
7632 compMinOptsIsSet = true;
7635 // true if the CLFLG_* for an optimization is set.
7636 inline bool OptEnabled(unsigned optFlag)
7638 return !!(compFlags & optFlag);
7641 #ifdef FEATURE_READYTORUN_COMPILER
7642 inline bool IsReadyToRun()
7644 return jitFlags->IsSet(JitFlags::JIT_FLAG_READYTORUN);
7647 inline bool IsReadyToRun()
7653 // true if we should use the PINVOKE_{BEGIN,END} helpers instead of generating
7654 // PInvoke transitions inline (e.g. when targeting CoreRT).
7655 inline bool ShouldUsePInvokeHelpers()
7657 #if COR_JIT_EE_VERSION > 460
7658 return jitFlags->IsSet(JitFlags::JIT_FLAG_USE_PINVOKE_HELPERS);
7664 // true if we should use insert the REVERSE_PINVOKE_{ENTER,EXIT} helpers in the method
7666 inline bool IsReversePInvoke()
7668 #if COR_JIT_EE_VERSION > 460
7669 return jitFlags->IsSet(JitFlags::JIT_FLAG_REVERSE_PINVOKE);
7675 // true if we must generate code compatible with JIT32 quirks
7676 inline bool IsJit32Compat()
7678 #if defined(_TARGET_X86_) && COR_JIT_EE_VERSION > 460
7679 return jitFlags->IsSet(JitFlags::JIT_FLAG_DESKTOP_QUIRKS);
7685 // true if we must generate code compatible with Jit64 quirks
7686 inline bool IsJit64Compat()
7688 #if defined(_TARGET_AMD64_) && COR_JIT_EE_VERSION > 460
7689 return jitFlags->IsSet(JitFlags::JIT_FLAG_DESKTOP_QUIRKS);
7690 #elif defined(_TARGET_AMD64_) && !defined(FEATURE_CORECLR)
7697 bool compScopeInfo; // Generate the LocalVar info ?
7698 bool compDbgCode; // Generate debugger-friendly code?
7699 bool compDbgInfo; // Gather debugging info?
7702 #ifdef PROFILING_SUPPORTED
7703 bool compNoPInvokeInlineCB;
7705 static const bool compNoPInvokeInlineCB;
7709 bool compGcChecks; // Check arguments and return values to ensure they are sane
7710 bool compStackCheckOnRet; // Check ESP on return to ensure it is correct
7711 bool compStackCheckOnCall; // Check ESP after every call to ensure it is correct
7715 bool compNeedSecurityCheck; // This flag really means where or not a security object needs
7716 // to be allocated on the stack.
7717 // It will be set to true in the following cases:
7718 // 1. When the method being compiled has a declarative security
7719 // (i.e. when CORINFO_FLG_NOSECURITYWRAP is reset for the current method).
7720 // This is also the case when we inject a prolog and epilog in the method.
7722 // 2. When the method being compiled has imperative security (i.e. the method
7723 // calls into another method that has CORINFO_FLG_SECURITYCHECK flag set).
7725 // 3. When opts.compDbgEnC is true. (See also Compiler::compCompile).
7727 // When this flag is set, jit will allocate a gc-reference local variable (lvaSecurityObject),
7728 // which gets reported as a GC root to stackwalker.
7729 // (See also ICodeManager::GetAddrOfSecurityObject.)
7736 #if defined(_TARGET_XARCH_) && !defined(LEGACY_BACKEND)
7737 bool compEnablePCRelAddr; // Whether absolute addr be encoded as PC-rel offset by RyuJIT where possible
7741 #ifdef UNIX_AMD64_ABI
7742 // This flag is indicating if there is a need to align the frame.
7743 // On AMD64-Windows, if there are calls, 4 slots for the outgoing ars are allocated, except for
7744 // FastTailCall. This slots makes the frame size non-zero, so alignment logic will be called.
7745 // On AMD64-Unix, there are no such slots. There is a possibility to have calls in the method with frame size of
7746 // 0. The frame alignment logic won't kick in. This flags takes care of the AMD64-Unix case by remembering that
7747 // there are calls and making sure the frame alignment logic is executed.
7748 bool compNeedToAlignFrame;
7749 #endif // UNIX_AMD64_ABI
7751 bool compProcedureSplitting; // Separate cold code from hot code
7753 bool genFPorder; // Preserve FP order (operations are non-commutative)
7754 bool genFPopt; // Can we do frame-pointer-omission optimization?
7755 bool altJit; // True if we are an altjit and are compiling this method
7758 bool optRepeat; // Repeat optimizer phases k times
7759 bool compProcedureSplittingEH; // Separate cold code from hot code for functions with EH
7760 bool dspCode; // Display native code generated
7761 bool dspEHTable; // Display the EH table reported to the VM
7762 bool dspInstrs; // Display the IL instructions intermixed with the native code output
7763 bool dspEmit; // Display emitter output
7764 bool dspLines; // Display source-code lines intermixed with native code output
7765 bool dmpHex; // Display raw bytes in hex of native code output
7766 bool varNames; // Display variables names in native code output
7767 bool disAsm; // Display native code as it is generated
7768 bool disAsmSpilled; // Display native code when any register spilling occurs
7769 bool disDiffable; // Makes the Disassembly code 'diff-able'
7770 bool disAsm2; // Display native code after it is generated using external disassembler
7771 bool dspOrder; // Display names of each of the methods that we ngen/jit
7772 bool dspUnwind; // Display the unwind info output
7773 bool dspDiffable; // Makes the Jit Dump 'diff-able' (currently uses same COMPlus_* flag as disDiffable)
7774 bool compLongAddress; // Force using large pseudo instructions for long address
7775 // (IF_LARGEJMP/IF_LARGEADR/IF_LARGLDC)
7776 bool dspGCtbls; // Display the GC tables
7780 bool doLateDisasm; // Run the late disassembler
7781 #endif // LATE_DISASM
7783 #if DUMP_GC_TABLES && !defined(DEBUG) && defined(JIT32_GCENCODER)
7784 // Only the JIT32_GCENCODER implements GC dumping in non-DEBUG code.
7785 #pragma message("NOTE: this non-debug build has GC ptr table dumping always enabled!")
7786 static const bool dspGCtbls = true;
7789 // We need stack probes to guarantee that we won't trigger a stack overflow
7790 // when calling unmanaged code until they get a chance to set up a frame, because
7791 // the EE will have no idea where it is.
7793 // We will only be doing this currently for hosted environments. Unfortunately
7794 // we need to take care of stubs, so potentially, we will have to do the probes
7795 // for any call. We have a plan for not needing for stubs though
7796 bool compNeedStackProbes;
7798 #ifdef PROFILING_SUPPORTED
7799 // Whether to emit Enter/Leave/TailCall hooks using a dummy stub (DummyProfilerELTStub()).
7800 // This option helps make the JIT behave as if it is running under a profiler.
7801 bool compJitELTHookEnabled;
7802 #endif // PROFILING_SUPPORTED
7804 #if FEATURE_TAILCALL_OPT
7805 // Whether opportunistic or implicit tail call optimization is enabled.
7806 bool compTailCallOpt;
7807 // Whether optimization of transforming a recursive tail call into a loop is enabled.
7808 bool compTailCallLoopOpt;
7812 static const bool compUseSoftFP = true;
7813 #else // !ARM_SOFTFP
7814 static const bool compUseSoftFP = false;
7817 GCPollType compGCPollType;
7821 static bool s_pAltJitExcludeAssembliesListInitialized;
7822 static AssemblyNamesList2* s_pAltJitExcludeAssembliesList;
7827 template <typename T>
7830 return (p == ZERO) ? ZERO : (opts.dspDiffable ? T(0xD1FFAB1E) : p);
7833 template <typename T>
7836 return (o == ZERO) ? ZERO : (opts.dspDiffable ? T(0xD1FFAB1E) : o);
7839 static int dspTreeID(GenTree* tree)
7841 return tree->gtTreeID;
7843 static void printTreeID(GenTree* tree)
7845 if (tree == nullptr)
7851 printf("[%06d]", dspTreeID(tree));
7858 #define STRESS_MODES \
7862 /* "Variations" stress areas which we try to mix up with each other. */ \
7863 /* These should not be exhaustively used as they might */ \
7864 /* hide/trivialize other areas */ \
7866 STRESS_MODE(REGS) STRESS_MODE(DBL_ALN) STRESS_MODE(LCL_FLDS) STRESS_MODE(UNROLL_LOOPS) \
7867 STRESS_MODE(MAKE_CSE) STRESS_MODE(LEGACY_INLINE) STRESS_MODE(CLONE_EXPR) \
7868 STRESS_MODE(USE_FCOMI) STRESS_MODE(USE_CMOV) STRESS_MODE(FOLD) \
7869 STRESS_MODE(BB_PROFILE) STRESS_MODE(OPT_BOOLS_GC) STRESS_MODE(REMORPH_TREES) \
7870 STRESS_MODE(64RSLT_MUL) STRESS_MODE(DO_WHILE_LOOPS) STRESS_MODE(MIN_OPTS) \
7871 STRESS_MODE(REVERSE_FLAG) /* Will set GTF_REVERSE_OPS whenever we can */ \
7872 STRESS_MODE(REVERSE_COMMA) /* Will reverse commas created with gtNewCommaNode */ \
7873 STRESS_MODE(TAILCALL) /* Will make the call as a tailcall whenever legal */ \
7874 STRESS_MODE(CATCH_ARG) /* Will spill catch arg */ \
7875 STRESS_MODE(UNSAFE_BUFFER_CHECKS) \
7876 STRESS_MODE(NULL_OBJECT_CHECK) \
7877 STRESS_MODE(PINVOKE_RESTORE_ESP) \
7878 STRESS_MODE(RANDOM_INLINE) \
7880 STRESS_MODE(GENERIC_VARN) STRESS_MODE(COUNT_VARN) \
7882 /* "Check" stress areas that can be exhaustively used if we */ \
7883 /* dont care about performance at all */ \
7885 STRESS_MODE(FORCE_INLINE) /* Treat every method as AggressiveInlining */ \
7886 STRESS_MODE(CHK_FLOW_UPDATE) \
7887 STRESS_MODE(EMITTER) STRESS_MODE(CHK_REIMPORT) STRESS_MODE(FLATFP) \
7889 STRESS_MODE(GENERIC_CHECK) STRESS_MODE(COUNT) \
7893 #define STRESS_MODE(mode) STRESS_##mode,
7900 static const LPCWSTR s_compStressModeNames[STRESS_COUNT + 1];
7901 BYTE compActiveStressModes[STRESS_COUNT];
7904 #define MAX_STRESS_WEIGHT 100
7906 bool compStressCompile(compStressArea stressArea, unsigned weightPercentage);
7910 bool compInlineStress()
7912 return compStressCompile(STRESS_LEGACY_INLINE, 50);
7915 bool compRandomInlineStress()
7917 return compStressCompile(STRESS_RANDOM_INLINE, 50);
7922 bool compTailCallStress()
7925 return (JitConfig.TailcallStress() != 0 || compStressCompile(STRESS_TAILCALL, 5));
7931 codeOptimize compCodeOpt()
7934 // Switching between size & speed has measurable throughput impact
7935 // (3.5% on NGen mscorlib when measured). It used to be enabled for
7936 // DEBUG, but should generate identical code between CHK & RET builds,
7937 // so that's not acceptable.
7938 // TODO-Throughput: Figure out what to do about size vs. speed & throughput.
7939 // Investigate the cause of the throughput regression.
7941 return opts.compCodeOpt;
7943 return BLENDED_CODE;
7947 //--------------------- Info about the procedure --------------------------
7951 COMP_HANDLE compCompHnd;
7952 CORINFO_MODULE_HANDLE compScopeHnd;
7953 CORINFO_CLASS_HANDLE compClassHnd;
7954 CORINFO_METHOD_HANDLE compMethodHnd;
7955 CORINFO_METHOD_INFO* compMethodInfo;
7957 BOOL hasCircularClassConstraints;
7958 BOOL hasCircularMethodConstraints;
7960 #if defined(DEBUG) || defined(LATE_DISASM)
7961 const char* compMethodName;
7962 const char* compClassName;
7963 const char* compFullName;
7964 #endif // defined(DEBUG) || defined(LATE_DISASM)
7966 #if defined(DEBUG) || defined(INLINE_DATA)
7967 // Method hash is logcally const, but computed
7969 mutable unsigned compMethodHashPrivate;
7970 unsigned compMethodHash() const;
7971 #endif // defined(DEBUG) || defined(INLINE_DATA)
7973 #ifdef PSEUDORANDOM_NOP_INSERTION
7974 // things for pseudorandom nop insertion
7975 unsigned compChecksum;
7979 // The following holds the FLG_xxxx flags for the method we're compiling.
7982 // The following holds the class attributes for the method we're compiling.
7983 unsigned compClassAttr;
7985 const BYTE* compCode;
7986 IL_OFFSET compILCodeSize; // The IL code size
7987 UNATIVE_OFFSET compNativeCodeSize; // The native code size, after instructions are issued. This
7988 // is less than (compTotalHotCodeSize + compTotalColdCodeSize) only if:
7989 // (1) the code is not hot/cold split, and we issued less code than we expected, or
7990 // (2) the code is hot/cold split, and we issued less code than we expected
7991 // in the cold section (the hot section will always be padded out to compTotalHotCodeSize).
7993 bool compIsStatic : 1; // Is the method static (no 'this' pointer)?
7994 bool compIsVarArgs : 1; // Does the method have varargs parameters?
7995 bool compIsContextful : 1; // contextful method
7996 bool compInitMem : 1; // Is the CORINFO_OPT_INIT_LOCALS bit set in the method info options?
7997 bool compUnwrapContextful : 1; // JIT should unwrap proxies when possible
7998 bool compProfilerCallback : 1; // JIT inserted a profiler Enter callback
7999 bool compPublishStubParam : 1; // EAX captured in prolog will be available through an instrinsic
8000 bool compRetBuffDefStack : 1; // The ret buff argument definitely points into the stack.
8002 var_types compRetType; // Return type of the method as declared in IL
8003 var_types compRetNativeType; // Normalized return type as per target arch ABI
8004 unsigned compILargsCount; // Number of arguments (incl. implicit but not hidden)
8005 unsigned compArgsCount; // Number of arguments (incl. implicit and hidden)
8006 unsigned compRetBuffArg; // position of hidden return param var (0, 1) (BAD_VAR_NUM means not present);
8007 int compTypeCtxtArg; // position of hidden param for type context for generic code (CORINFO_CALLCONV_PARAMTYPE)
8008 unsigned compThisArg; // position of implicit this pointer param (not to be confused with lvaArg0Var)
8009 unsigned compILlocalsCount; // Number of vars : args + locals (incl. implicit but not hidden)
8010 unsigned compLocalsCount; // Number of vars : args + locals (incl. implicit and hidden)
8011 unsigned compMaxStack;
8012 UNATIVE_OFFSET compTotalHotCodeSize; // Total number of bytes of Hot Code in the method
8013 UNATIVE_OFFSET compTotalColdCodeSize; // Total number of bytes of Cold Code in the method
8015 unsigned compCallUnmanaged; // count of unmanaged calls
8016 unsigned compLvFrameListRoot; // lclNum for the Frame root
8017 unsigned compXcptnsCount; // Number of exception-handling clauses read in the method's IL.
8018 // You should generally use compHndBBtabCount instead: it is the
8019 // current number of EH clauses (after additions like synchronized
8020 // methods and funclets, and removals like unreachable code deletion).
8022 bool compMatchedVM; // true if the VM is "matched": either the JIT is a cross-compiler
8023 // and the VM expects that, or the JIT is a "self-host" compiler
8024 // (e.g., x86 hosted targeting x86) and the VM expects that.
8026 /* The following holds IL scope information about local variables.
8029 unsigned compVarScopesCount;
8030 VarScopeDsc* compVarScopes;
8032 /* The following holds information about instr offsets for
8033 * which we need to report IP-mappings
8036 IL_OFFSET* compStmtOffsets; // sorted
8037 unsigned compStmtOffsetsCount;
8038 ICorDebugInfo::BoundaryTypes compStmtOffsetsImplicit;
8040 #define CPU_X86 0x0100 // The generic X86 CPU
8041 #define CPU_X86_PENTIUM_4 0x0110
8043 #define CPU_X64 0x0200 // The generic x64 CPU
8044 #define CPU_AMD_X64 0x0210 // AMD x64 CPU
8045 #define CPU_INTEL_X64 0x0240 // Intel x64 CPU
8047 #define CPU_ARM 0x0300 // The generic ARM CPU
8049 unsigned genCPU; // What CPU are we running on
8052 // Returns true if the method being compiled returns a non-void and non-struct value.
8053 // Note that lvaInitTypeRef() normalizes compRetNativeType for struct returns in a
8054 // single register as per target arch ABI (e.g on Amd64 Windows structs of size 1, 2,
8055 // 4 or 8 gets normalized to TYP_BYTE/TYP_SHORT/TYP_INT/TYP_LONG; On Arm HFA structs).
8056 // Methods returning such structs are considered to return non-struct return value and
8057 // this method returns true in that case.
8058 bool compMethodReturnsNativeScalarType()
8060 return (info.compRetType != TYP_VOID) && !varTypeIsStruct(info.compRetNativeType);
8063 // Returns true if the method being compiled returns RetBuf addr as its return value
8064 bool compMethodReturnsRetBufAddr()
8066 // There are cases where implicit RetBuf argument should be explicitly returned in a register.
8067 // In such cases the return type is changed to TYP_BYREF and appropriate IR is generated.
8069 // 1. Profiler Leave calllback expects the address of retbuf as return value for
8070 // methods with hidden RetBuf argument. impReturnInstruction() when profiler
8071 // callbacks are needed creates GT_RETURN(TYP_BYREF, op1 = Addr of RetBuf) for
8072 // methods with hidden RetBufArg.
8074 // 2. As per the System V ABI, the address of RetBuf needs to be returned by
8075 // methods with hidden RetBufArg in RAX. In such case GT_RETURN is of TYP_BYREF,
8076 // returning the address of RetBuf.
8078 // 3. Windows 64-bit native calling convention also requires the address of RetBuff
8079 // to be returned in RAX.
8080 CLANG_FORMAT_COMMENT_ANCHOR;
8082 #ifdef _TARGET_AMD64_
8083 return (info.compRetBuffArg != BAD_VAR_NUM);
8084 #else // !_TARGET_AMD64_
8085 return (compIsProfilerHookNeeded()) && (info.compRetBuffArg != BAD_VAR_NUM);
8086 #endif // !_TARGET_AMD64_
8089 // Returns true if the method returns a value in more than one return register
8090 // TODO-ARM-Bug: Deal with multi-register genReturnLocaled structs?
8091 // TODO-ARM64: Does this apply for ARM64 too?
8092 bool compMethodReturnsMultiRegRetType()
8094 #if FEATURE_MULTIREG_RET
8095 #if defined(_TARGET_X86_)
8096 // On x86 only 64-bit longs are returned in multiple registers
8097 return varTypeIsLong(info.compRetNativeType);
8098 #else // targets: X64-UNIX, ARM64 or ARM32
8099 // On all other targets that support multireg return values:
8100 // Methods returning a struct in multiple registers have a return value of TYP_STRUCT.
8101 // Such method's compRetNativeType is TYP_STRUCT without a hidden RetBufArg
8102 return varTypeIsStruct(info.compRetNativeType) && (info.compRetBuffArg == BAD_VAR_NUM);
8103 #endif // TARGET_XXX
8105 #else // not FEATURE_MULTIREG_RET
8107 // For this architecture there are no multireg returns
8110 #endif // FEATURE_MULTIREG_RET
8113 #if FEATURE_MULTIREG_ARGS
8114 // Given a GenTree node of TYP_STRUCT that represents a pass by value argument
8115 // return the gcPtr layout for the pointers sized fields
8116 void getStructGcPtrsFromOp(GenTreePtr op, BYTE* gcPtrsOut);
8117 #endif // FEATURE_MULTIREG_ARGS
8119 // Returns true if the method being compiled returns a value
8120 bool compMethodHasRetVal()
8122 return compMethodReturnsNativeScalarType() || compMethodReturnsRetBufAddr() ||
8123 compMethodReturnsMultiRegRetType();
8128 void compDispLocalVars();
8132 //-------------------------- Global Compiler Data ------------------------------------
8135 static unsigned s_compMethodsCount; // to produce unique label names
8136 unsigned compGenTreeID;
8139 BasicBlock* compCurBB; // the current basic block in process
8140 GenTreePtr compCurStmt; // the current statement in process
8142 unsigned compCurStmtNum; // to give all statements an increasing StmtNum when printing dumps
8145 // The following is used to create the 'method JIT info' block.
8146 size_t compInfoBlkSize;
8147 BYTE* compInfoBlkAddr;
8149 EHblkDsc* compHndBBtab; // array of EH data
8150 unsigned compHndBBtabCount; // element count of used elements in EH data array
8151 unsigned compHndBBtabAllocCount; // element count of allocated elements in EH data array
8153 #if defined(_TARGET_X86_)
8155 //-------------------------------------------------------------------------
8156 // Tracking of region covered by the monitor in synchronized methods
8157 void* syncStartEmitCookie; // the emitter cookie for first instruction after the call to MON_ENTER
8158 void* syncEndEmitCookie; // the emitter cookie for first instruction after the call to MON_EXIT
8160 #endif // !_TARGET_X86_
8162 Phases previousCompletedPhase; // the most recently completed phase
8164 //-------------------------------------------------------------------------
8165 // The following keeps track of how many bytes of local frame space we've
8166 // grabbed so far in the current function, and how many argument bytes we
8167 // need to pop when we return.
8170 unsigned compLclFrameSize; // secObject+lclBlk+locals+temps
8172 // Count of callee-saved regs we pushed in the prolog.
8173 // Does not include EBP for isFramePointerUsed() and double-aligned frames.
8174 // In case of Amd64 this doesn't include float regs saved on stack.
8175 unsigned compCalleeRegsPushed;
8177 #if defined(_TARGET_XARCH_) && !FEATURE_STACK_FP_X87
8178 // Mask of callee saved float regs on stack.
8179 regMaskTP compCalleeFPRegsSavedMask;
8181 #ifdef _TARGET_AMD64_
8182 // Quirk for VS debug-launch scenario to work:
8183 // Bytes of padding between save-reg area and locals.
8184 #define VSQUIRK_STACK_PAD (2 * REGSIZE_BYTES)
8185 unsigned compVSQuirkStackPaddingNeeded;
8186 bool compQuirkForPPPflag;
8189 unsigned compArgSize; // total size of arguments in bytes (including register args (lvIsRegArg))
8191 unsigned compMapILargNum(unsigned ILargNum); // map accounting for hidden args
8192 unsigned compMapILvarNum(unsigned ILvarNum); // map accounting for hidden args
8193 unsigned compMap2ILvarNum(unsigned varNum); // map accounting for hidden args
8195 //-------------------------------------------------------------------------
8197 static void compStartup(); // One-time initialization
8198 static void compShutdown(); // One-time finalization
8200 void compInit(ArenaAllocator* pAlloc, InlineInfo* inlineInfo);
8203 static void compDisplayStaticSizes(FILE* fout);
8205 //------------ Some utility functions --------------
8207 void* compGetHelperFtn(CorInfoHelpFunc ftnNum, /* IN */
8208 void** ppIndirection); /* OUT */
8210 // Several JIT/EE interface functions return a CorInfoType, and also return a
8211 // class handle as an out parameter if the type is a value class. Returns the
8212 // size of the type these describe.
8213 unsigned compGetTypeSize(CorInfoType cit, CORINFO_CLASS_HANDLE clsHnd);
8216 // Components used by the compiler may write unit test suites, and
8217 // have them run within this method. They will be run only once per process, and only
8218 // in debug. (Perhaps should be under the control of a COMPlus_ flag.)
8219 // These should fail by asserting.
8220 void compDoComponentUnitTestsOnce();
8223 int compCompile(CORINFO_METHOD_HANDLE methodHnd,
8224 CORINFO_MODULE_HANDLE classPtr,
8225 COMP_HANDLE compHnd,
8226 CORINFO_METHOD_INFO* methodInfo,
8227 void** methodCodePtr,
8228 ULONG* methodCodeSize,
8229 JitFlags* compileFlags);
8230 void compCompileFinish();
8231 int compCompileHelper(CORINFO_MODULE_HANDLE classPtr,
8232 COMP_HANDLE compHnd,
8233 CORINFO_METHOD_INFO* methodInfo,
8234 void** methodCodePtr,
8235 ULONG* methodCodeSize,
8236 JitFlags* compileFlags,
8237 CorInfoInstantiationVerification instVerInfo);
8239 ArenaAllocator* compGetAllocator();
8241 #if MEASURE_MEM_ALLOC
8243 static bool s_dspMemStats; // Display per-phase memory statistics for every function
8247 unsigned allocCnt; // # of allocs
8248 UINT64 allocSz; // total size of those alloc.
8249 UINT64 allocSzMax; // Maximum single allocation.
8250 UINT64 allocSzByKind[CMK_Count]; // Classified by "kind".
8251 UINT64 nraTotalSizeAlloc;
8252 UINT64 nraTotalSizeUsed;
8254 static const char* s_CompMemKindNames[]; // Names of the kinds.
8256 MemStats() : allocCnt(0), allocSz(0), allocSzMax(0), nraTotalSizeAlloc(0), nraTotalSizeUsed(0)
8258 for (int i = 0; i < CMK_Count; i++)
8260 allocSzByKind[i] = 0;
8263 MemStats(const MemStats& ms)
8264 : allocCnt(ms.allocCnt)
8265 , allocSz(ms.allocSz)
8266 , allocSzMax(ms.allocSzMax)
8267 , nraTotalSizeAlloc(ms.nraTotalSizeAlloc)
8268 , nraTotalSizeUsed(ms.nraTotalSizeUsed)
8270 for (int i = 0; i < CMK_Count; i++)
8272 allocSzByKind[i] = ms.allocSzByKind[i];
8276 // Until we have ubiquitous constructors.
8279 this->MemStats::MemStats();
8282 void AddAlloc(size_t sz, CompMemKind cmk)
8286 if (sz > allocSzMax)
8290 allocSzByKind[cmk] += sz;
8293 void Print(FILE* f); // Print these stats to f.
8294 void PrintByKind(FILE* f); // Do just the by-kind histogram part.
8296 MemStats genMemStats;
8298 struct AggregateMemStats : public MemStats
8302 AggregateMemStats() : MemStats(), nMethods(0)
8306 void Add(const MemStats& ms)
8309 allocCnt += ms.allocCnt;
8310 allocSz += ms.allocSz;
8311 allocSzMax = max(allocSzMax, ms.allocSzMax);
8312 for (int i = 0; i < CMK_Count; i++)
8314 allocSzByKind[i] += ms.allocSzByKind[i];
8316 nraTotalSizeAlloc += ms.nraTotalSizeAlloc;
8317 nraTotalSizeUsed += ms.nraTotalSizeUsed;
8320 void Print(FILE* f); // Print these stats to jitstdout.
8323 static CritSecObject s_memStatsLock; // This lock protects the data structures below.
8324 static MemStats s_maxCompMemStats; // Stats for the compilation with the largest amount allocated.
8325 static AggregateMemStats s_aggMemStats; // Aggregates statistics for all compilations.
8327 #endif // MEASURE_MEM_ALLOC
8329 #if LOOP_HOIST_STATS
8330 unsigned m_loopsConsidered;
8331 bool m_curLoopHasHoistedExpression;
8332 unsigned m_loopsWithHoistedExpressions;
8333 unsigned m_totalHoistedExpressions;
8335 void AddLoopHoistStats();
8336 void PrintPerMethodLoopHoistStats();
8338 static CritSecObject s_loopHoistStatsLock; // This lock protects the data structures below.
8339 static unsigned s_loopsConsidered;
8340 static unsigned s_loopsWithHoistedExpressions;
8341 static unsigned s_totalHoistedExpressions;
8343 static void PrintAggregateLoopHoistStats(FILE* f);
8344 #endif // LOOP_HOIST_STATS
8346 void* compGetMemArray(size_t numElem, size_t elemSize, CompMemKind cmk = CMK_Unknown);
8347 void* compGetMemArrayA(size_t numElem, size_t elemSize, CompMemKind cmk = CMK_Unknown);
8348 void* compGetMem(size_t sz, CompMemKind cmk = CMK_Unknown);
8349 void* compGetMemA(size_t sz, CompMemKind cmk = CMK_Unknown);
8350 static void* compGetMemCallback(void*, size_t, CompMemKind cmk = CMK_Unknown);
8351 void compFreeMem(void*);
8353 bool compIsForImportOnly();
8354 bool compIsForInlining();
8355 bool compDonotInline();
8358 const char* compLocalVarName(unsigned varNum, unsigned offs);
8359 VarName compVarName(regNumber reg, bool isFloatReg = false);
8360 const char* compRegVarName(regNumber reg, bool displayVar = false, bool isFloatReg = false);
8361 const char* compRegPairName(regPairNo regPair);
8362 const char* compRegNameForSize(regNumber reg, size_t size);
8363 const char* compFPregVarName(unsigned fpReg, bool displayVar = false);
8364 void compDspSrcLinesByNativeIP(UNATIVE_OFFSET curIP);
8365 void compDspSrcLinesByLineNum(unsigned line, bool seek = false);
8368 //-------------------------------------------------------------------------
8370 typedef ListNode<VarScopeDsc*> VarScopeListNode;
8372 struct VarScopeMapInfo
8374 VarScopeListNode* head;
8375 VarScopeListNode* tail;
8376 static VarScopeMapInfo* Create(VarScopeListNode* node, IAllocator* alloc)
8378 VarScopeMapInfo* info = new (alloc) VarScopeMapInfo;
8385 // Max value of scope count for which we would use linear search; for larger values we would use hashtable lookup.
8386 static const unsigned MAX_LINEAR_FIND_LCL_SCOPELIST = 32;
8388 typedef SimplerHashTable<unsigned, SmallPrimitiveKeyFuncs<unsigned>, VarScopeMapInfo*, JitSimplerHashBehavior>
8389 VarNumToScopeDscMap;
8391 // Map to keep variables' scope indexed by varNum containing it's scope dscs at the index.
8392 VarNumToScopeDscMap* compVarScopeMap;
8394 VarScopeDsc* compFindLocalVar(unsigned varNum, unsigned lifeBeg, unsigned lifeEnd);
8396 VarScopeDsc* compFindLocalVar(unsigned varNum, unsigned offs);
8398 VarScopeDsc* compFindLocalVarLinear(unsigned varNum, unsigned offs);
8400 void compInitVarScopeMap();
8402 VarScopeDsc** compEnterScopeList; // List has the offsets where variables
8403 // enter scope, sorted by instr offset
8404 unsigned compNextEnterScope;
8406 VarScopeDsc** compExitScopeList; // List has the offsets where variables
8407 // go out of scope, sorted by instr offset
8408 unsigned compNextExitScope;
8410 void compInitScopeLists();
8412 void compResetScopeLists();
8414 VarScopeDsc* compGetNextEnterScope(unsigned offs, bool scan = false);
8416 VarScopeDsc* compGetNextExitScope(unsigned offs, bool scan = false);
8418 void compProcessScopesUntil(unsigned offset,
8420 void (Compiler::*enterScopeFn)(VARSET_TP* inScope, VarScopeDsc*),
8421 void (Compiler::*exitScopeFn)(VARSET_TP* inScope, VarScopeDsc*));
8424 void compDispScopeLists();
8427 bool compIsProfilerHookNeeded();
8429 //-------------------------------------------------------------------------
8430 /* Statistical Data Gathering */
8432 void compJitStats(); // call this function and enable
8433 // various ifdef's below for statistical data
8436 void compCallArgStats();
8437 static void compDispCallArgStats(FILE* fout);
8440 //-------------------------------------------------------------------------
8447 ArenaAllocator* compAllocator;
8450 // This one presents an implementation of the "IAllocator" abstract class that uses "compAllocator",
8451 // suitable for use by utilcode collection types.
8452 IAllocator* compAsIAllocator;
8454 #if MEASURE_MEM_ALLOC
8455 IAllocator* compAsIAllocatorBitset; // An allocator that uses the CMK_bitset tracker.
8456 IAllocator* compAsIAllocatorGC; // An allocator that uses the CMK_GC tracker.
8457 IAllocator* compAsIAllocatorLoopHoist; // An allocator that uses the CMK_LoopHoist tracker.
8459 IAllocator* compAsIAllocatorDebugOnly; // An allocator that uses the CMK_DebugOnly tracker.
8461 #endif // MEASURE_MEM_ALLOC
8463 void compFunctionTraceStart();
8464 void compFunctionTraceEnd(void* methodCodePtr, ULONG methodCodeSize, bool isNYI);
8467 size_t compMaxUncheckedOffsetForNullObject;
8469 void compInitOptions(JitFlags* compileFlags);
8471 void compSetProcessor();
8472 void compInitDebuggingInfo();
8473 void compSetOptimizationLevel();
8474 #ifdef _TARGET_ARMARCH_
8475 bool compRsvdRegCheck(FrameLayoutState curState);
8477 void compCompile(void** methodCodePtr, ULONG* methodCodeSize, JitFlags* compileFlags);
8479 // Clear annotations produced during optimizations; to be used between iterations when repeating opts.
8480 void ResetOptAnnotations();
8482 // Regenerate loop descriptors; to be used between iterations when repeating opts.
8483 void RecomputeLoopInfo();
8485 #ifdef PROFILING_SUPPORTED
8486 // Data required for generating profiler Enter/Leave/TailCall hooks
8488 bool compProfilerHookNeeded; // Whether profiler Enter/Leave/TailCall hook needs to be generated for the method
8489 void* compProfilerMethHnd; // Profiler handle of the method being compiled. Passed as param to ELT callbacks
8490 bool compProfilerMethHndIndirected; // Whether compProfilerHandle is pointer to the handle or is an actual handle
8493 #ifdef _TARGET_AMD64_
8494 bool compQuirkForPPP(); // Check if this method should be Quirked for the PPP issue
8497 // Assumes called as part of process shutdown; does any compiler-specific work associated with that.
8498 static void ProcessShutdownWork(ICorStaticInfo* statInfo);
8500 IAllocator* getAllocator()
8502 return compAsIAllocator;
8505 #if MEASURE_MEM_ALLOC
8506 IAllocator* getAllocatorBitset()
8508 return compAsIAllocatorBitset;
8510 IAllocator* getAllocatorGC()
8512 return compAsIAllocatorGC;
8514 IAllocator* getAllocatorLoopHoist()
8516 return compAsIAllocatorLoopHoist;
8518 #else // !MEASURE_MEM_ALLOC
8519 IAllocator* getAllocatorBitset()
8521 return compAsIAllocator;
8523 IAllocator* getAllocatorGC()
8525 return compAsIAllocator;
8527 IAllocator* getAllocatorLoopHoist()
8529 return compAsIAllocator;
8531 #endif // !MEASURE_MEM_ALLOC
8534 IAllocator* getAllocatorDebugOnly()
8536 #if MEASURE_MEM_ALLOC
8537 return compAsIAllocatorDebugOnly;
8538 #else // !MEASURE_MEM_ALLOC
8539 return compAsIAllocator;
8540 #endif // !MEASURE_MEM_ALLOC
8545 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8546 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8550 XX Checks for type compatibility and merges types XX
8552 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8553 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8557 // Set to TRUE if verification cannot be skipped for this method
8558 // If we detect unverifiable code, we will lazily check
8559 // canSkipMethodVerification() to see if verification is REALLY needed.
8560 BOOL tiVerificationNeeded;
8562 // It it initially TRUE, and it gets set to FALSE if we run into unverifiable code
8563 // Note that this is valid only if tiVerificationNeeded was ever TRUE.
8564 BOOL tiIsVerifiableCode;
8566 // Set to TRUE if runtime callout is needed for this method
8567 BOOL tiRuntimeCalloutNeeded;
8569 // Set to TRUE if security prolog/epilog callout is needed for this method
8570 // Note: This flag is different than compNeedSecurityCheck.
8571 // compNeedSecurityCheck means whether or not a security object needs
8572 // to be allocated on the stack, which is currently true for EnC as well.
8573 // tiSecurityCalloutNeeded means whether or not security callouts need
8574 // to be inserted in the jitted code.
8575 BOOL tiSecurityCalloutNeeded;
8577 // Returns TRUE if child is equal to or a subtype of parent for merge purposes
8578 // This support is necessary to suport attributes that are not described in
8579 // for example, signatures. For example, the permanent home byref (byref that
8580 // points to the gc heap), isn't a property of method signatures, therefore,
8581 // it is safe to have mismatches here (that tiCompatibleWith will not flag),
8582 // but when deciding if we need to reimport a block, we need to take these
8584 BOOL tiMergeCompatibleWith(const typeInfo& pChild, const typeInfo& pParent, bool normalisedForStack) const;
8586 // Returns TRUE if child is equal to or a subtype of parent.
8587 // normalisedForStack indicates that both types are normalised for the stack
8588 BOOL tiCompatibleWith(const typeInfo& pChild, const typeInfo& pParent, bool normalisedForStack) const;
8590 // Merges pDest and pSrc. Returns FALSE if merge is undefined.
8591 // *pDest is modified to represent the merged type. Sets "*changed" to true
8592 // if this changes "*pDest".
8593 BOOL tiMergeToCommonParent(typeInfo* pDest, const typeInfo* pSrc, bool* changed) const;
8595 // Set pDest from the primitive value type.
8596 // Eg. System.Int32 -> ELEMENT_TYPE_I4
8598 BOOL tiFromPrimitiveValueClass(typeInfo* pDest, const typeInfo* pVC) const;
8601 // <BUGNUM> VSW 471305
8602 // IJW allows assigning REF to BYREF. The following allows us to temporarily
8603 // bypass the assert check in gcMarkRegSetGCref and gcMarkRegSetByref
8604 // We use a "short" as we need to push/pop this scope.
8606 short compRegSetCheckLevel;
8610 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8611 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8613 XX IL verification stuff XX
8616 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8617 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8621 // The following is used to track liveness of local variables, initialization
8622 // of valueclass constructors, and type safe use of IL instructions.
8624 // dynamic state info needed for verification
8625 EntryState verCurrentState;
8627 // this ptr of object type .ctors are considered intited only after
8628 // the base class ctor is called, or an alternate ctor is called.
8629 // An uninited this ptr can be used to access fields, but cannot
8630 // be used to call a member function.
8631 BOOL verTrackObjCtorInitState;
8633 void verInitBBEntryState(BasicBlock* block, EntryState* currentState);
8635 // Requires that "tis" is not TIS_Bottom -- it's a definite init/uninit state.
8636 void verSetThisInit(BasicBlock* block, ThisInitState tis);
8637 void verInitCurrentState();
8638 void verResetCurrentState(BasicBlock* block, EntryState* currentState);
8640 // Merges the current verification state into the entry state of "block", return FALSE if that merge fails,
8641 // TRUE if it succeeds. Further sets "*changed" to true if this changes the entry state of "block".
8642 BOOL verMergeEntryStates(BasicBlock* block, bool* changed);
8644 void verConvertBBToThrowVerificationException(BasicBlock* block DEBUGARG(bool logMsg));
8645 void verHandleVerificationFailure(BasicBlock* block DEBUGARG(bool logMsg));
8646 typeInfo verMakeTypeInfo(CORINFO_CLASS_HANDLE clsHnd,
8647 bool bashStructToRef = false); // converts from jit type representation to typeInfo
8648 typeInfo verMakeTypeInfo(CorInfoType ciType,
8649 CORINFO_CLASS_HANDLE clsHnd); // converts from jit type representation to typeInfo
8650 BOOL verIsSDArray(typeInfo ti);
8651 typeInfo verGetArrayElemType(typeInfo ti);
8653 typeInfo verParseArgSigToTypeInfo(CORINFO_SIG_INFO* sig, CORINFO_ARG_LIST_HANDLE args);
8654 BOOL verNeedsVerification();
8655 BOOL verIsByRefLike(const typeInfo& ti);
8656 BOOL verIsSafeToReturnByRef(const typeInfo& ti);
8658 // generic type variables range over types that satisfy IsBoxable
8659 BOOL verIsBoxable(const typeInfo& ti);
8661 void DECLSPEC_NORETURN verRaiseVerifyException(INDEBUG(const char* reason) DEBUGARG(const char* file)
8662 DEBUGARG(unsigned line));
8663 void verRaiseVerifyExceptionIfNeeded(INDEBUG(const char* reason) DEBUGARG(const char* file)
8664 DEBUGARG(unsigned line));
8665 bool verCheckTailCallConstraint(OPCODE opcode,
8666 CORINFO_RESOLVED_TOKEN* pResolvedToken,
8667 CORINFO_RESOLVED_TOKEN* pConstrainedResolvedToken, // Is this a "constrained." call
8668 // on a type parameter?
8669 bool speculative // If true, won't throw if verificatoin fails. Instead it will
8670 // return false to the caller.
8671 // If false, it will throw.
8673 bool verIsBoxedValueType(typeInfo ti);
8675 void verVerifyCall(OPCODE opcode,
8676 CORINFO_RESOLVED_TOKEN* pResolvedToken,
8677 CORINFO_RESOLVED_TOKEN* pConstrainedResolvedToken,
8679 bool readonlyCall, // is this a "readonly." call?
8680 const BYTE* delegateCreateStart,
8681 const BYTE* codeAddr,
8682 CORINFO_CALL_INFO* callInfo DEBUGARG(const char* methodName));
8684 BOOL verCheckDelegateCreation(const BYTE* delegateCreateStart, const BYTE* codeAddr, mdMemberRef& targetMemberRef);
8686 typeInfo verVerifySTIND(const typeInfo& ptr, const typeInfo& value, const typeInfo& instrType);
8687 typeInfo verVerifyLDIND(const typeInfo& ptr, const typeInfo& instrType);
8688 void verVerifyField(CORINFO_RESOLVED_TOKEN* pResolvedToken,
8689 const CORINFO_FIELD_INFO& fieldInfo,
8690 const typeInfo* tiThis,
8692 BOOL allowPlainStructAsThis = FALSE);
8693 void verVerifyCond(const typeInfo& tiOp1, const typeInfo& tiOp2, unsigned opcode);
8694 void verVerifyThisPtrInitialised();
8695 BOOL verIsCallToInitThisPtr(CORINFO_CLASS_HANDLE context, CORINFO_CLASS_HANDLE target);
8697 // Register allocator
8698 void raInitStackFP();
8699 void raEnregisterVarsPrePassStackFP();
8700 void raSetRegLclBirthDeath(GenTreePtr tree, VARSET_VALARG_TP lastlife, bool fromLDOBJ);
8701 void raEnregisterVarsPostPassStackFP();
8702 void raGenerateFPRefCounts();
8703 void raEnregisterVarsStackFP();
8704 void raUpdateHeightsForVarsStackFP(VARSET_VALARG_TP mask);
8706 regNumber raRegForVarStackFP(unsigned varTrackedIndex);
8707 void raAddPayloadStackFP(VARSET_VALARG_TP mask, unsigned weight);
8709 // returns true if enregistering v1 would save more mem accesses than v2
8710 bool raVarIsGreaterValueStackFP(LclVarDsc* lv1, LclVarDsc* lv2);
8713 void raDumpHeightsStackFP();
8714 void raDumpVariableRegIntfFloat();
8717 #if FEATURE_STACK_FP_X87
8719 // Currently, we use FP transition blocks in only 2 situations:
8721 // -conditional jump on longs where FP stack differs with target: it's not strictly
8722 // necessary, but its low frequency and the code would get complicated if we try to
8723 // inline the FP stack adjustment, as we have a lot of special casing going on to try
8724 // minimize the way we generate the jump code.
8725 // -case statements of switch where the FP stack differs with the one of evaluating the switch () statement
8726 // We do this as we want to codegen switch as a jumptable. Again, this is low frequency.
8728 // However, transition blocks have 2 problems
8730 // - Procedure splitting: current implementation of procedure splitting requires all basic blocks to
8731 // be known at codegen time, as it generates all hot blocks first and cold blocks later. This ties
8732 // us up in codegen and is a solvable problem (we could make procedure splitting generate blocks
8733 // in the right place without preordering them), this causes us to have to generate the transition
8734 // blocks in the cold area if we want procedure splitting.
8737 // - Thread abort exceptions and transition blocks. Transition blocks were designed under the assumption
8738 // that no exceptions can happen inside them. Unfortunately Thread.Abort can happen in any instruction,
8739 // and if we have handlers we will have to try to call them. Fixing this the right way would imply
8740 // having multiple try native code regions for a single try il region. This is doable and shouldnt be
8741 // a big change in the exception.
8743 // Given the low frequency of the cases where we have transition blocks, I've decided to dumb down
8744 // optimizations. For these 2 cases:
8746 // - When there is a chance that we will have FP transition blocks, we won't do procedure splitting.
8747 // - When a method has a handler, it won't enregister any FP variables that go thru a conditional long or
8748 // a switch statement.
8750 // If at any point we find we need to optimize this, we should throw work at unblocking the restrictions our
8751 // current procedure splitting and exception code have.
8752 bool compMayHaveTransitionBlocks;
8754 VARSET_TP raMaskDontEnregFloat; // mask for additional restrictions
8756 VARSET_TP raLclRegIntfFloat[REG_FPCOUNT];
8758 unsigned raCntStkStackFP;
8759 unsigned raCntWtdStkDblStackFP;
8760 unsigned raCntStkParamDblStackFP;
8762 // Payload in mem accesses for enregistering a variable (we dont want to mix with refcounts)
8763 // TODO: Do we want to put this in LclVarDsc?
8764 unsigned raPayloadStackFP[lclMAX_TRACKED];
8765 unsigned raHeightsStackFP[lclMAX_TRACKED][FP_VIRTUALREGISTERS + 1];
8767 // Useful for debugging
8768 unsigned raHeightsNonWeightedStackFP[lclMAX_TRACKED][FP_VIRTUALREGISTERS + 1];
8770 #endif // FEATURE_STACK_FP_X87
8773 // One line log function. Default level is 0. Increasing it gives you
8774 // more log information
8776 // levels are currently unused: #define JITDUMP(level,...) ();
8777 void JitLogEE(unsigned level, const char* fmt, ...);
8779 bool compDebugBreak;
8781 bool compJitHaltMethod();
8786 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8787 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8789 XX GS Security checks for unsafe buffers XX
8791 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8792 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8795 struct ShadowParamVarInfo
8797 FixedBitVect* assignGroup; // the closure set of variables whose values depend on each other
8798 unsigned shadowCopy; // Lcl var num, valid only if not set to NO_SHADOW_COPY
8800 static bool mayNeedShadowCopy(LclVarDsc* varDsc)
8802 #if defined(_TARGET_AMD64_) && !defined(LEGACY_BACKEND)
8803 // GS cookie logic to create shadow slots, create trees to copy reg args to shadow
8804 // slots and update all trees to refer to shadow slots is done immediately after
8805 // fgMorph(). Lsra could potentially mark a param as DoNotEnregister after JIT determines
8806 // not to shadow a parameter. Also, LSRA could potentially spill a param which is passed
8807 // in register. Therefore, conservatively all params may need a shadow copy. Note that
8808 // GS cookie logic further checks whether the param is a ptr or an unsafe buffer before
8809 // creating a shadow slot even though this routine returns true.
8811 // TODO-AMD64-CQ: Revisit this conservative approach as it could create more shadow slots than
8812 // required. There are two cases under which a reg arg could potentially be used from its
8814 // a) LSRA marks it as DoNotEnregister (see LinearScan::identifyCandidates())
8815 // b) LSRA spills it
8817 // Possible solution to address case (a)
8818 // - The conditions under which LSRA marks a varDsc as DoNotEnregister could be checked
8819 // in this routine. Note that live out of exception handler is something we may not be
8820 // able to do it here since GS cookie logic is invoked ahead of liveness computation.
8821 // Therefore, for methods with exception handling and need GS cookie check we might have
8822 // to take conservative approach.
8824 // Possible solution to address case (b)
8825 // - Whenver a parameter passed in an argument register needs to be spilled by LSRA, we
8826 // create a new spill temp if the method needs GS cookie check.
8827 return varDsc->lvIsParam;
8828 #else // !(defined(_TARGET_AMD64_) && defined(LEGACY_BACKEND))
8829 return varDsc->lvIsParam && !varDsc->lvIsRegArg;
8836 printf("assignGroup [%p]; shadowCopy: [%d];\n", assignGroup, shadowCopy);
8841 GSCookie* gsGlobalSecurityCookieAddr; // Address of global cookie for unsafe buffer checks
8842 GSCookie gsGlobalSecurityCookieVal; // Value of global cookie if addr is NULL
8843 ShadowParamVarInfo* gsShadowVarInfo; // Table used by shadow param analysis code
8845 void gsGSChecksInitCookie(); // Grabs cookie variable
8846 void gsCopyShadowParams(); // Identify vulnerable params and create dhadow copies
8847 bool gsFindVulnerableParams(); // Shadow param analysis code
8848 void gsParamsToShadows(); // Insert copy code and replave param uses by shadow
8850 static fgWalkPreFn gsMarkPtrsAndAssignGroups; // Shadow param analysis tree-walk
8851 static fgWalkPreFn gsReplaceShadowParams; // Shadow param replacement tree-walk
8853 #define DEFAULT_MAX_INLINE_SIZE 100 // Methods with > DEFAULT_MAX_INLINE_SIZE IL bytes will never be inlined.
8854 // This can be overwritten by setting complus_JITInlineSize env variable.
8856 #define DEFAULT_MAX_INLINE_DEPTH 20 // Methods at more than this level deep will not be inlined
8859 #ifdef FEATURE_JIT_METHOD_PERF
8860 JitTimer* pCompJitTimer; // Timer data structure (by phases) for current compilation.
8861 static CompTimeSummaryInfo s_compJitTimerSummary; // Summary of the Timer information for the whole run.
8863 static LPCWSTR JitTimeLogCsv(); // Retrieve the file name for CSV from ConfigDWORD.
8864 static LPCWSTR compJitTimeLogFilename; // If a log file for JIT time is desired, filename to write it to.
8866 inline void EndPhase(Phases phase); // Indicate the end of the given phase.
8868 #if MEASURE_CLRAPI_CALLS
8869 // Thin wrappers that call into JitTimer (if present).
8870 inline void CLRApiCallEnter(unsigned apix);
8871 inline void CLRApiCallLeave(unsigned apix);
8874 inline void CLR_API_Enter(API_ICorJitInfo_Names ename);
8875 inline void CLR_API_Leave(API_ICorJitInfo_Names ename);
8880 #if defined(DEBUG) || defined(INLINE_DATA) || defined(FEATURE_CLRSQM)
8881 // These variables are associated with maintaining SQM data about compile time.
8882 unsigned __int64 m_compCyclesAtEndOfInlining; // The thread-virtualized cycle count at the end of the inlining phase
8883 // in the current compilation.
8884 unsigned __int64 m_compCycles; // Net cycle count for current compilation
8885 DWORD m_compTickCountAtEndOfInlining; // The result of GetTickCount() (# ms since some epoch marker) at the end of
8886 // the inlining phase in the current compilation.
8887 #endif // defined(DEBUG) || defined(INLINE_DATA) || defined(FEATURE_CLRSQM)
8889 // Records the SQM-relevant (cycles and tick count). Should be called after inlining is complete.
8890 // (We do this after inlining because this marks the last point at which the JIT is likely to cause
8891 // type-loading and class initialization).
8892 void RecordStateAtEndOfInlining();
8893 // Assumes being called at the end of compilation. Update the SQM state.
8894 void RecordStateAtEndOfCompilation();
8896 #ifdef FEATURE_CLRSQM
8897 // Does anything SQM related necessary at process shutdown time.
8898 static void ProcessShutdownSQMWork(ICorStaticInfo* statInfo);
8899 #endif // FEATURE_CLRSQM
8902 #if FUNC_INFO_LOGGING
8903 static LPCWSTR compJitFuncInfoFilename; // If a log file for per-function information is required, this is the
8904 // filename to write it to.
8905 static FILE* compJitFuncInfoFile; // And this is the actual FILE* to write to.
8906 #endif // FUNC_INFO_LOGGING
8908 Compiler* prevCompiler; // Previous compiler on stack for TLS Compiler* linked list for reentrant compilers.
8910 // Is the compilation in a full trust context?
8911 bool compIsFullTrust();
8913 #ifndef FEATURE_TRACELOGGING
8914 // Should we actually fire the noway assert body and the exception handler?
8915 bool compShouldThrowOnNoway();
8916 #else // FEATURE_TRACELOGGING
8917 // Should we actually fire the noway assert body and the exception handler?
8918 bool compShouldThrowOnNoway(const char* filename, unsigned line);
8920 // Telemetry instance to use per method compilation.
8921 JitTelemetry compJitTelemetry;
8923 // Get common parameters that have to be logged with most telemetry data.
8924 void compGetTelemetryDefaults(const char** assemblyName,
8925 const char** scopeName,
8926 const char** methodName,
8927 unsigned* methodHash);
8928 #endif // !FEATURE_TRACELOGGING
8932 NodeToTestDataMap* m_nodeTestData;
8934 static const unsigned FIRST_LOOP_HOIST_CSE_CLASS = 1000;
8935 unsigned m_loopHoistCSEClass; // LoopHoist test annotations turn into CSE requirements; we
8936 // label them with CSE Class #'s starting at FIRST_LOOP_HOIST_CSE_CLASS.
8937 // Current kept in this.
8939 NodeToTestDataMap* GetNodeTestData()
8941 Compiler* compRoot = impInlineRoot();
8942 if (compRoot->m_nodeTestData == nullptr)
8944 compRoot->m_nodeTestData = new (getAllocatorDebugOnly()) NodeToTestDataMap(getAllocatorDebugOnly());
8946 return compRoot->m_nodeTestData;
8949 typedef SimplerHashTable<GenTreePtr, PtrKeyFuncs<GenTree>, int, JitSimplerHashBehavior> NodeToIntMap;
8951 // Returns the set (i.e., the domain of the result map) of nodes that are keys in m_nodeTestData, and
8952 // currently occur in the AST graph.
8953 NodeToIntMap* FindReachableNodesInNodeTestData();
8955 // Node "from" is being eliminated, and being replaced by node "to". If "from" had any associated
8956 // test data, associate that data with "to".
8957 void TransferTestDataToNode(GenTreePtr from, GenTreePtr to);
8959 // Requires that "to" is a clone of "from". If any nodes in the "from" tree
8960 // have annotations, attach similar annotations to the corresponding nodes in "to".
8961 void CopyTestDataToCloneTree(GenTreePtr from, GenTreePtr to);
8963 // These are the methods that test that the various conditions implied by the
8964 // test attributes are satisfied.
8965 void JitTestCheckSSA(); // SSA builder tests.
8966 void JitTestCheckVN(); // Value numbering tests.
8969 // The "FieldSeqStore", for canonicalizing field sequences. See the definition of FieldSeqStore for
8971 FieldSeqStore* m_fieldSeqStore;
8973 FieldSeqStore* GetFieldSeqStore()
8975 Compiler* compRoot = impInlineRoot();
8976 if (compRoot->m_fieldSeqStore == nullptr)
8978 // Create a CompAllocator that labels sub-structure with CMK_FieldSeqStore, and use that for allocation.
8979 IAllocator* ialloc = new (this, CMK_FieldSeqStore) CompAllocator(this, CMK_FieldSeqStore);
8980 compRoot->m_fieldSeqStore = new (ialloc) FieldSeqStore(ialloc);
8982 return compRoot->m_fieldSeqStore;
8985 typedef SimplerHashTable<GenTreePtr, PtrKeyFuncs<GenTree>, FieldSeqNode*, JitSimplerHashBehavior> NodeToFieldSeqMap;
8987 // Some nodes of "TYP_BYREF" or "TYP_I_IMPL" actually represent the address of a field within a struct, but since
8988 // the offset of the field is zero, there's no "GT_ADD" node. We normally attach a field sequence to the constant
8989 // that is added, but what do we do when that constant is zero, and is thus not present? We use this mechanism to
8990 // attach the field sequence directly to the address node.
8991 NodeToFieldSeqMap* m_zeroOffsetFieldMap;
8993 NodeToFieldSeqMap* GetZeroOffsetFieldMap()
8995 // Don't need to worry about inlining here
8996 if (m_zeroOffsetFieldMap == nullptr)
8998 // Create a CompAllocator that labels sub-structure with CMK_ZeroOffsetFieldMap, and use that for
9000 IAllocator* ialloc = new (this, CMK_ZeroOffsetFieldMap) CompAllocator(this, CMK_ZeroOffsetFieldMap);
9001 m_zeroOffsetFieldMap = new (ialloc) NodeToFieldSeqMap(ialloc);
9003 return m_zeroOffsetFieldMap;
9006 // Requires that "op1" is a node of type "TYP_BYREF" or "TYP_I_IMPL". We are dereferencing this with the fields in
9007 // "fieldSeq", whose offsets are required all to be zero. Ensures that any field sequence annotation currently on
9008 // "op1" or its components is augmented by appending "fieldSeq". In practice, if "op1" is a GT_LCL_FLD, it has
9009 // a field sequence as a member; otherwise, it may be the addition of an a byref and a constant, where the const
9010 // has a field sequence -- in this case "fieldSeq" is appended to that of the constant; otherwise, we
9011 // record the the field sequence using the ZeroOffsetFieldMap described above.
9013 // One exception above is that "op1" is a node of type "TYP_REF" where "op1" is a GT_LCL_VAR.
9014 // This happens when System.Object vtable pointer is a regular field at offset 0 in System.Private.CoreLib in
9015 // CoreRT. Such case is handled same as the default case.
9016 void fgAddFieldSeqForZeroOffset(GenTreePtr op1, FieldSeqNode* fieldSeq);
9018 typedef SimplerHashTable<const GenTree*, PtrKeyFuncs<GenTree>, ArrayInfo, JitSimplerHashBehavior>
9020 NodeToArrayInfoMap* m_arrayInfoMap;
9022 NodeToArrayInfoMap* GetArrayInfoMap()
9024 Compiler* compRoot = impInlineRoot();
9025 if (compRoot->m_arrayInfoMap == nullptr)
9027 // Create a CompAllocator that labels sub-structure with CMK_ArrayInfoMap, and use that for allocation.
9028 IAllocator* ialloc = new (this, CMK_ArrayInfoMap) CompAllocator(this, CMK_ArrayInfoMap);
9029 compRoot->m_arrayInfoMap = new (ialloc) NodeToArrayInfoMap(ialloc);
9031 return compRoot->m_arrayInfoMap;
9034 NodeToUnsignedMap* m_memorySsaMap[MemoryKindCount];
9036 // In some cases, we want to assign intermediate SSA #'s to memory states, and know what nodes create those memory
9037 // states. (We do this for try blocks, where, if the try block doesn't do a call that loses track of the memory
9038 // state, all the possible memory states are possible initial states of the corresponding catch block(s).)
9039 NodeToUnsignedMap* GetMemorySsaMap(MemoryKind memoryKind)
9041 if (memoryKind == GcHeap && byrefStatesMatchGcHeapStates)
9043 // Use the same map for GCHeap and ByrefExposed when their states match.
9044 memoryKind = ByrefExposed;
9047 assert(memoryKind < MemoryKindCount);
9048 Compiler* compRoot = impInlineRoot();
9049 if (compRoot->m_memorySsaMap[memoryKind] == nullptr)
9051 // Create a CompAllocator that labels sub-structure with CMK_ArrayInfoMap, and use that for allocation.
9052 IAllocator* ialloc = new (this, CMK_ArrayInfoMap) CompAllocator(this, CMK_ArrayInfoMap);
9053 compRoot->m_memorySsaMap[memoryKind] = new (ialloc) NodeToUnsignedMap(ialloc);
9055 return compRoot->m_memorySsaMap[memoryKind];
9058 // The Refany type is the only struct type whose structure is implicitly assumed by IL. We need its fields.
9059 CORINFO_CLASS_HANDLE m_refAnyClass;
9060 CORINFO_FIELD_HANDLE GetRefanyDataField()
9062 if (m_refAnyClass == nullptr)
9064 m_refAnyClass = info.compCompHnd->getBuiltinClass(CLASSID_TYPED_BYREF);
9066 return info.compCompHnd->getFieldInClass(m_refAnyClass, 0);
9068 CORINFO_FIELD_HANDLE GetRefanyTypeField()
9070 if (m_refAnyClass == nullptr)
9072 m_refAnyClass = info.compCompHnd->getBuiltinClass(CLASSID_TYPED_BYREF);
9074 return info.compCompHnd->getFieldInClass(m_refAnyClass, 1);
9078 static BitSetSupport::BitSetOpCounter m_varsetOpCounter;
9080 #if ALLVARSET_COUNTOPS
9081 static BitSetSupport::BitSetOpCounter m_allvarsetOpCounter;
9084 static HelperCallProperties s_helperCallProperties;
9086 #ifdef FEATURE_UNIX_AMD64_STRUCT_PASSING
9087 static var_types GetTypeFromClassificationAndSizes(SystemVClassificationType classType, int size);
9088 static var_types GetEightByteType(const SYSTEMV_AMD64_CORINFO_STRUCT_REG_PASSING_DESCRIPTOR& structDesc,
9090 static void GetStructTypeOffset(const SYSTEMV_AMD64_CORINFO_STRUCT_REG_PASSING_DESCRIPTOR& structDesc,
9093 unsigned __int8* offset0,
9094 unsigned __int8* offset1);
9095 void fgMorphSystemVStructArgs(GenTreeCall* call, bool hasStructArgument);
9096 #endif // defined(FEATURE_UNIX_AMD64_STRUCT_PASSING)
9098 void fgMorphMultiregStructArgs(GenTreeCall* call);
9099 GenTreePtr fgMorphMultiregStructArg(GenTreePtr arg, fgArgTabEntryPtr fgEntryPtr);
9101 }; // end of class Compiler
9103 // Inline methods of CompAllocator.
9104 void* CompAllocator::Alloc(size_t sz)
9106 #if MEASURE_MEM_ALLOC
9107 return m_comp->compGetMem(sz, m_cmk);
9109 return m_comp->compGetMem(sz);
9113 void* CompAllocator::ArrayAlloc(size_t elems, size_t elemSize)
9115 #if MEASURE_MEM_ALLOC
9116 return m_comp->compGetMemArray(elems, elemSize, m_cmk);
9118 return m_comp->compGetMemArray(elems, elemSize);
9122 // LclVarDsc constructor. Uses Compiler, so must come after Compiler definition.
9123 inline LclVarDsc::LclVarDsc(Compiler* comp)
9124 : // Initialize the ArgRegs to REG_STK.
9125 // The morph will do the right thing to change
9126 // to the right register if passed in register.
9129 #if FEATURE_MULTIREG_ARGS
9130 _lvOtherArgReg(REG_STK)
9132 #endif // FEATURE_MULTIREG_ARGS
9134 lvRefBlks(BlockSetOps::UninitVal())
9136 #endif // ASSERTION_PROP
9137 lvPerSsaData(comp->getAllocator())
9142 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9143 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9145 XX Miscellaneous Compiler stuff XX
9147 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9148 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9151 // Values used to mark the types a stack slot is used for
9153 const unsigned TYPE_REF_INT = 0x01; // slot used as a 32-bit int
9154 const unsigned TYPE_REF_LNG = 0x02; // slot used as a 64-bit long
9155 const unsigned TYPE_REF_FLT = 0x04; // slot used as a 32-bit float
9156 const unsigned TYPE_REF_DBL = 0x08; // slot used as a 64-bit float
9157 const unsigned TYPE_REF_PTR = 0x10; // slot used as a 32-bit pointer
9158 const unsigned TYPE_REF_BYR = 0x20; // slot used as a byref pointer
9159 const unsigned TYPE_REF_STC = 0x40; // slot used as a struct
9160 const unsigned TYPE_REF_TYPEMASK = 0x7F; // bits that represent the type
9162 // const unsigned TYPE_REF_ADDR_TAKEN = 0x80; // slots address was taken
9164 /*****************************************************************************
9166 * Variables to keep track of total code amounts.
9171 extern size_t grossVMsize;
9172 extern size_t grossNCsize;
9173 extern size_t totalNCsize;
9175 extern unsigned genMethodICnt;
9176 extern unsigned genMethodNCnt;
9177 extern size_t gcHeaderISize;
9178 extern size_t gcPtrMapISize;
9179 extern size_t gcHeaderNSize;
9180 extern size_t gcPtrMapNSize;
9182 #endif // DISPLAY_SIZES
9184 /*****************************************************************************
9186 * Variables to keep track of basic block counts (more data on 1 BB methods)
9189 #if COUNT_BASIC_BLOCKS
9190 extern Histogram bbCntTable;
9191 extern Histogram bbOneBBSizeTable;
9194 /*****************************************************************************
9196 * Used by optFindNaturalLoops to gather statistical information such as
9197 * - total number of natural loops
9198 * - number of loops with 1, 2, ... exit conditions
9199 * - number of loops that have an iterator (for like)
9200 * - number of loops that have a constant iterator
9205 extern unsigned totalLoopMethods; // counts the total number of methods that have natural loops
9206 extern unsigned maxLoopsPerMethod; // counts the maximum number of loops a method has
9207 extern unsigned totalLoopOverflows; // # of methods that identified more loops than we can represent
9208 extern unsigned totalLoopCount; // counts the total number of natural loops
9209 extern unsigned totalUnnatLoopCount; // counts the total number of (not-necessarily natural) loops
9210 extern unsigned totalUnnatLoopOverflows; // # of methods that identified more unnatural loops than we can represent
9211 extern unsigned iterLoopCount; // counts the # of loops with an iterator (for like)
9212 extern unsigned simpleTestLoopCount; // counts the # of loops with an iterator and a simple loop condition (iter <
9214 extern unsigned constIterLoopCount; // counts the # of loops with a constant iterator (for like)
9215 extern bool hasMethodLoops; // flag to keep track if we already counted a method as having loops
9216 extern unsigned loopsThisMethod; // counts the number of loops in the current method
9217 extern bool loopOverflowThisMethod; // True if we exceeded the max # of loops in the method.
9218 extern Histogram loopCountTable; // Histogram of loop counts
9219 extern Histogram loopExitCountTable; // Histogram of loop exit counts
9221 #endif // COUNT_LOOPS
9223 /*****************************************************************************
9224 * variables to keep track of how many iterations we go in a dataflow pass
9229 extern unsigned CSEiterCount; // counts the # of iteration for the CSE dataflow
9230 extern unsigned CFiterCount; // counts the # of iteration for the Const Folding dataflow
9232 #endif // DATAFLOW_ITER
9234 #if MEASURE_BLOCK_SIZE
9235 extern size_t genFlowNodeSize;
9236 extern size_t genFlowNodeCnt;
9237 #endif // MEASURE_BLOCK_SIZE
9239 #if MEASURE_NODE_SIZE
9240 struct NodeSizeStats
9245 genTreeNodeSize = 0;
9246 genTreeNodeActualSize = 0;
9249 size_t genTreeNodeCnt;
9250 size_t genTreeNodeSize; // The size we allocate
9251 size_t genTreeNodeActualSize; // The actual size of the node. Note that the actual size will likely be smaller
9252 // than the allocated size, but we sometimes use SetOper()/ChangeOper() to change
9253 // a smaller node to a larger one. TODO-Cleanup: add stats on
9254 // SetOper()/ChangeOper() usage to quanitfy this.
9256 extern NodeSizeStats genNodeSizeStats; // Total node size stats
9257 extern NodeSizeStats genNodeSizeStatsPerFunc; // Per-function node size stats
9258 extern Histogram genTreeNcntHist;
9259 extern Histogram genTreeNsizHist;
9260 #endif // MEASURE_NODE_SIZE
9262 /*****************************************************************************
9263 * Count fatal errors (including noway_asserts).
9267 extern unsigned fatal_badCode;
9268 extern unsigned fatal_noWay;
9269 extern unsigned fatal_NOMEM;
9270 extern unsigned fatal_noWayAssertBody;
9272 extern unsigned fatal_noWayAssertBodyArgs;
9274 extern unsigned fatal_NYI;
9275 #endif // MEASURE_FATAL
9277 /*****************************************************************************
9281 #ifdef _TARGET_XARCH_
9283 const instruction INS_SHIFT_LEFT_LOGICAL = INS_shl;
9284 const instruction INS_SHIFT_RIGHT_LOGICAL = INS_shr;
9285 const instruction INS_SHIFT_RIGHT_ARITHM = INS_sar;
9287 const instruction INS_AND = INS_and;
9288 const instruction INS_OR = INS_or;
9289 const instruction INS_XOR = INS_xor;
9290 const instruction INS_NEG = INS_neg;
9291 const instruction INS_TEST = INS_test;
9292 const instruction INS_MUL = INS_imul;
9293 const instruction INS_SIGNED_DIVIDE = INS_idiv;
9294 const instruction INS_UNSIGNED_DIVIDE = INS_div;
9295 const instruction INS_BREAKPOINT = INS_int3;
9296 const instruction INS_ADDC = INS_adc;
9297 const instruction INS_SUBC = INS_sbb;
9298 const instruction INS_NOT = INS_not;
9304 const instruction INS_SHIFT_LEFT_LOGICAL = INS_lsl;
9305 const instruction INS_SHIFT_RIGHT_LOGICAL = INS_lsr;
9306 const instruction INS_SHIFT_RIGHT_ARITHM = INS_asr;
9308 const instruction INS_AND = INS_and;
9309 const instruction INS_OR = INS_orr;
9310 const instruction INS_XOR = INS_eor;
9311 const instruction INS_NEG = INS_rsb;
9312 const instruction INS_TEST = INS_tst;
9313 const instruction INS_MUL = INS_mul;
9314 const instruction INS_SIGNED_DIVIDE = INS_sdiv;
9315 const instruction INS_UNSIGNED_DIVIDE = INS_udiv;
9316 const instruction INS_BREAKPOINT = INS_bkpt;
9317 const instruction INS_ADDC = INS_adc;
9318 const instruction INS_SUBC = INS_sbc;
9319 const instruction INS_NOT = INS_mvn;
9323 #ifdef _TARGET_ARM64_
9325 const instruction INS_SHIFT_LEFT_LOGICAL = INS_lsl;
9326 const instruction INS_SHIFT_RIGHT_LOGICAL = INS_lsr;
9327 const instruction INS_SHIFT_RIGHT_ARITHM = INS_asr;
9329 const instruction INS_AND = INS_and;
9330 const instruction INS_OR = INS_orr;
9331 const instruction INS_XOR = INS_eor;
9332 const instruction INS_NEG = INS_neg;
9333 const instruction INS_TEST = INS_tst;
9334 const instruction INS_MUL = INS_mul;
9335 const instruction INS_SIGNED_DIVIDE = INS_sdiv;
9336 const instruction INS_UNSIGNED_DIVIDE = INS_udiv;
9337 const instruction INS_BREAKPOINT = INS_bkpt;
9338 const instruction INS_ADDC = INS_adc;
9339 const instruction INS_SUBC = INS_sbc;
9340 const instruction INS_NOT = INS_mvn;
9344 /*****************************************************************************/
9346 extern const BYTE genTypeSizes[];
9347 extern const BYTE genTypeAlignments[];
9348 extern const BYTE genTypeStSzs[];
9349 extern const BYTE genActualTypes[];
9351 /*****************************************************************************/
9353 // VERY_LARGE_FRAME_SIZE_REG_MASK is the set of registers we need to use for
9354 // the probing loop generated for very large stack frames (see `getVeryLargeFrameSize`).
9357 #define VERY_LARGE_FRAME_SIZE_REG_MASK (RBM_R4 | RBM_R5 | RBM_R6)
9358 #elif defined(_TARGET_ARM64_)
9359 #define VERY_LARGE_FRAME_SIZE_REG_MASK (RBM_R9 | RBM_R10 | RBM_R11)
9362 /*****************************************************************************/
9364 #define REG_CORRUPT regNumber(REG_NA + 1)
9365 #define RBM_CORRUPT (RBM_ILLEGAL | regMaskTP(1))
9366 #define REG_PAIR_CORRUPT regPairNo(REG_PAIR_NONE + 1)
9368 /*****************************************************************************/
9370 extern BasicBlock dummyBB;
9372 /*****************************************************************************/
9373 /*****************************************************************************/
9375 // foreach_treenode_execution_order: An iterator that iterates through all the tree
9376 // nodes of a statement in execution order.
9377 // __stmt: a GT_STMT type GenTree*
9378 // __node: a GenTree*, already declared, that gets updated with each node in the statement, in execution order
9380 #define foreach_treenode_execution_order(__node, __stmt) \
9381 for ((__node) = (__stmt)->gtStmt.gtStmtList; (__node); (__node) = (__node)->gtNext)
9383 // foreach_block: An iterator over all blocks in the function.
9384 // __compiler: the Compiler* object
9385 // __block : a BasicBlock*, already declared, that gets updated each iteration.
9387 #define foreach_block(__compiler, __block) \
9388 for ((__block) = (__compiler)->fgFirstBB; (__block); (__block) = (__block)->bbNext)
9390 /*****************************************************************************/
9391 /*****************************************************************************/
9395 void dumpConvertedVarSet(Compiler* comp, VARSET_VALARG_TP vars);
9397 /*XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9398 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9400 XX Debugging helpers XX
9402 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9403 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9406 /*****************************************************************************/
9407 /* The following functions are intended to be called from the debugger, to dump
9408 * various data structures. The can be used in the debugger Watch or Quick Watch
9409 * windows. They are designed to be short to type and take as few arguments as
9410 * possible. The 'c' versions take a Compiler*, whereas the 'd' versions use the TlsCompiler.
9411 * See the function definition comment for more details.
9414 void cBlock(Compiler* comp, BasicBlock* block);
9415 void cBlocks(Compiler* comp);
9416 void cBlocksV(Compiler* comp);
9417 void cTree(Compiler* comp, GenTree* tree);
9418 void cTrees(Compiler* comp);
9419 void cEH(Compiler* comp);
9420 void cVar(Compiler* comp, unsigned lclNum);
9421 void cVarDsc(Compiler* comp, LclVarDsc* varDsc);
9422 void cVars(Compiler* comp);
9423 void cVarsFinal(Compiler* comp);
9424 void cBlockPreds(Compiler* comp, BasicBlock* block);
9425 void cReach(Compiler* comp);
9426 void cDoms(Compiler* comp);
9427 void cLiveness(Compiler* comp);
9428 void cCVarSet(Compiler* comp, VARSET_VALARG_TP vars);
9430 void cFuncIR(Compiler* comp);
9431 void cBlockIR(Compiler* comp, BasicBlock* block);
9432 void cLoopIR(Compiler* comp, Compiler::LoopDsc* loop);
9433 void cTreeIR(Compiler* comp, GenTree* tree);
9434 int cTreeTypeIR(Compiler* comp, GenTree* tree);
9435 int cTreeKindsIR(Compiler* comp, GenTree* tree);
9436 int cTreeFlagsIR(Compiler* comp, GenTree* tree);
9437 int cOperandIR(Compiler* comp, GenTree* operand);
9438 int cLeafIR(Compiler* comp, GenTree* tree);
9439 int cIndirIR(Compiler* comp, GenTree* tree);
9440 int cListIR(Compiler* comp, GenTree* list);
9441 int cSsaNumIR(Compiler* comp, GenTree* tree);
9442 int cValNumIR(Compiler* comp, GenTree* tree);
9443 int cDependsIR(Compiler* comp, GenTree* comma, bool* first);
9445 void dBlock(BasicBlock* block);
9448 void dTree(GenTree* tree);
9451 void dVar(unsigned lclNum);
9452 void dVarDsc(LclVarDsc* varDsc);
9455 void dBlockPreds(BasicBlock* block);
9459 void dCVarSet(VARSET_VALARG_TP vars);
9461 void dVarSet(VARSET_VALARG_TP vars);
9462 void dRegMask(regMaskTP mask);
9465 void dBlockIR(BasicBlock* block);
9466 void dTreeIR(GenTree* tree);
9467 void dLoopIR(Compiler::LoopDsc* loop);
9468 void dLoopNumIR(unsigned loopNum);
9469 int dTabStopIR(int curr, int tabstop);
9470 int dTreeTypeIR(GenTree* tree);
9471 int dTreeKindsIR(GenTree* tree);
9472 int dTreeFlagsIR(GenTree* tree);
9473 int dOperandIR(GenTree* operand);
9474 int dLeafIR(GenTree* tree);
9475 int dIndirIR(GenTree* tree);
9476 int dListIR(GenTree* list);
9477 int dSsaNumIR(GenTree* tree);
9478 int dValNumIR(GenTree* tree);
9479 int dDependsIR(GenTree* comma);
9482 GenTree* dFindTree(GenTree* tree, unsigned id);
9483 GenTree* dFindTree(unsigned id);
9484 GenTreeStmt* dFindStmt(unsigned id);
9485 BasicBlock* dFindBlock(unsigned bbNum);
9489 #include "compiler.hpp" // All the shared inline functions
9491 /*****************************************************************************/
9492 #endif //_COMPILER_H_
9493 /*****************************************************************************/