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 lvAssignOne : 1; // assigned at least once?
273 unsigned char lvAssignTwo : 1; // assigned at least twice?
276 unsigned char lvSpilled : 1; // enregistered variable was spilled
277 #ifndef _TARGET_64BIT_
278 unsigned char lvStructDoubleAlign : 1; // Must we double align this struct?
279 #endif // !_TARGET_64BIT_
280 #ifdef _TARGET_64BIT_
281 unsigned char lvQuirkToLong : 1; // Quirk to allocate this LclVar as a 64-bit long
284 unsigned char lvKeepType : 1; // Don't change the type of this variable
285 unsigned char lvNoLclFldStress : 1; // Can't apply local field stress on this one
287 unsigned char lvIsPtr : 1; // Might this be used in an address computation? (used by buffer overflow security
289 unsigned char lvIsUnsafeBuffer : 1; // Does this contain an unsafe buffer requiring buffer overflow security checks?
290 unsigned char lvPromoted : 1; // True when this local is a promoted struct, a normed struct, or a "split" long on a
292 unsigned char lvIsStructField : 1; // Is this local var a field of a promoted struct local?
293 unsigned char lvContainsFloatingFields : 1; // Does this struct contains floating point fields?
294 unsigned char lvOverlappingFields : 1; // True when we have a struct with possibly overlapping fields
295 unsigned char lvContainsHoles : 1; // True when we have a promoted struct that contains holes
296 unsigned char lvCustomLayout : 1; // True when this struct has "CustomLayout"
298 unsigned char lvIsMultiRegArg : 1; // true if this is a multireg LclVar struct used in an argument context
299 unsigned char lvIsMultiRegRet : 1; // true if this is a multireg LclVar struct assigned from a multireg call
302 unsigned char _lvIsHfa : 1; // Is this a struct variable who's class handle is an HFA type
303 unsigned char _lvIsHfaRegArg : 1; // Is this a HFA argument variable? // TODO-CLEANUP: Remove this and replace
304 // with (lvIsRegArg && lvIsHfa())
305 unsigned char _lvHfaTypeIsFloat : 1; // Is the HFA type float or double?
306 #endif // FEATURE_HFA
309 // TODO-Cleanup: See the note on lvSize() - this flag is only in use by asserts that are checking for struct
310 // types, and is needed because of cases where TYP_STRUCT is bashed to an integral type.
311 // Consider cleaning this up so this workaround is not required.
312 unsigned char lvUnusedStruct : 1; // All references to this promoted struct are through its field locals.
313 // I.e. there is no longer any reference to the struct directly.
314 // In this case we can simply remove this struct local.
316 #ifndef LEGACY_BACKEND
317 unsigned char lvLRACandidate : 1; // Tracked for linear scan register allocation purposes
318 #endif // !LEGACY_BACKEND
321 // Note that both SIMD vector args and locals are marked as lvSIMDType = true, but the
322 // type of an arg node is TYP_BYREF and a local node is TYP_SIMD*.
323 unsigned char lvSIMDType : 1; // This is a SIMD struct
324 unsigned char lvUsedInSIMDIntrinsic : 1; // This tells lclvar is used for simd intrinsic
325 #endif // FEATURE_SIMD
326 unsigned char lvRegStruct : 1; // This is a reg-sized non-field-addressed struct.
329 unsigned lvFieldLclStart; // The index of the local var representing the first field in the promoted struct
331 unsigned lvParentLcl; // The index of the local var representing the parent (i.e. the promoted struct local).
332 // Valid on promoted struct local fields.
334 var_types lvBaseType; // The base type of a SIMD local var. Valid on TYP_SIMD locals.
335 #endif // FEATURE_SIMD
338 unsigned char lvFieldCnt; // Number of fields in the promoted VarDsc.
339 unsigned char lvFldOffset;
340 unsigned char lvFldOrdinal;
342 #if FEATURE_MULTIREG_ARGS
343 regNumber lvRegNumForSlot(unsigned slotNum)
349 else if (slotNum == 1)
351 return lvOtherArgReg;
355 assert(false && "Invalid slotNum!");
360 #endif // FEATURE_MULTIREG_ARGS
378 bool lvIsHfaRegArg() const
381 return _lvIsHfaRegArg;
387 void lvSetIsHfaRegArg()
390 _lvIsHfaRegArg = true;
394 bool lvHfaTypeIsFloat() const
397 return _lvHfaTypeIsFloat;
403 void lvSetHfaTypeIsFloat(bool value)
406 _lvHfaTypeIsFloat = value;
410 // on Arm64 - Returns 1-4 indicating the number of register slots used by the HFA
411 // on Arm32 - Returns the total number of single FP register slots used by the HFA, max is 8
413 unsigned lvHfaSlots() const
416 assert(lvType == TYP_STRUCT);
418 return lvExactSize / sizeof(float);
419 #else // _TARGET_ARM64_
420 if (lvHfaTypeIsFloat())
422 return lvExactSize / sizeof(float);
426 return lvExactSize / sizeof(double);
428 #endif // _TARGET_ARM64_
431 // lvIsMultiRegArgOrRet()
432 // returns true if this is a multireg LclVar struct used in an argument context
433 // or if this is a multireg LclVar struct assigned from a multireg call
434 bool lvIsMultiRegArgOrRet()
436 return lvIsMultiRegArg || lvIsMultiRegRet;
440 regNumberSmall _lvRegNum; // Used to store the register this variable is in (or, the low register of a
441 // register pair). For LEGACY_BACKEND, this is only set if lvRegister is
442 // non-zero. For non-LEGACY_BACKEND, it is set during codegen any time the
443 // variable is enregistered (in non-LEGACY_BACKEND, lvRegister is only set
444 // to non-zero if the variable gets the same register assignment for its entire
446 #if !defined(_TARGET_64BIT_)
447 regNumberSmall _lvOtherReg; // Used for "upper half" of long var.
448 #endif // !defined(_TARGET_64BIT_)
450 regNumberSmall _lvArgReg; // The register in which this argument is passed.
452 #if FEATURE_MULTIREG_ARGS
453 regNumberSmall _lvOtherArgReg; // Used for the second part of the struct passed in a register.
454 // Note this is defined but not used by ARM32
455 #endif // FEATURE_MULTIREG_ARGS
457 #ifndef LEGACY_BACKEND
459 regNumberSmall _lvArgInitReg; // the register into which the argument is moved at entry
460 regPairNoSmall _lvArgInitRegPair; // the register pair into which the argument is moved at entry
462 #endif // !LEGACY_BACKEND
465 // The register number is stored in a small format (8 bits), but the getters return and the setters take
466 // a full-size (unsigned) format, to localize the casts here.
468 /////////////////////
470 __declspec(property(get = GetRegNum, put = SetRegNum)) regNumber lvRegNum;
472 regNumber GetRegNum() const
474 return (regNumber)_lvRegNum;
477 void SetRegNum(regNumber reg)
479 _lvRegNum = (regNumberSmall)reg;
480 assert(_lvRegNum == reg);
483 /////////////////////
485 #if defined(_TARGET_64BIT_)
486 __declspec(property(get = GetOtherReg, put = SetOtherReg)) regNumber lvOtherReg;
488 regNumber GetOtherReg() const
490 assert(!"shouldn't get here"); // can't use "unreached();" because it's NORETURN, which causes C4072
491 // "unreachable code" warnings
495 void SetOtherReg(regNumber reg)
497 assert(!"shouldn't get here"); // can't use "unreached();" because it's NORETURN, which causes C4072
498 // "unreachable code" warnings
500 #else // !_TARGET_64BIT_
501 __declspec(property(get = GetOtherReg, put = SetOtherReg)) regNumber lvOtherReg;
503 regNumber GetOtherReg() const
505 return (regNumber)_lvOtherReg;
508 void SetOtherReg(regNumber reg)
510 _lvOtherReg = (regNumberSmall)reg;
511 assert(_lvOtherReg == reg);
513 #endif // !_TARGET_64BIT_
515 /////////////////////
517 __declspec(property(get = GetArgReg, put = SetArgReg)) regNumber lvArgReg;
519 regNumber GetArgReg() const
521 return (regNumber)_lvArgReg;
524 void SetArgReg(regNumber reg)
526 _lvArgReg = (regNumberSmall)reg;
527 assert(_lvArgReg == reg);
530 #if FEATURE_MULTIREG_ARGS
531 __declspec(property(get = GetOtherArgReg, put = SetOtherArgReg)) regNumber lvOtherArgReg;
533 regNumber GetOtherArgReg() const
535 return (regNumber)_lvOtherArgReg;
538 void SetOtherArgReg(regNumber reg)
540 _lvOtherArgReg = (regNumberSmall)reg;
541 assert(_lvOtherArgReg == reg);
543 #endif // FEATURE_MULTIREG_ARGS
546 // Is this is a SIMD struct?
547 bool lvIsSIMDType() const
552 // Is this is a SIMD struct which is used for SIMD intrinsic?
553 bool lvIsUsedInSIMDIntrinsic() const
555 return lvUsedInSIMDIntrinsic;
558 // If feature_simd not enabled, return false
559 bool lvIsSIMDType() const
563 bool lvIsUsedInSIMDIntrinsic() const
569 /////////////////////
571 #ifndef LEGACY_BACKEND
572 __declspec(property(get = GetArgInitReg, put = SetArgInitReg)) regNumber lvArgInitReg;
574 regNumber GetArgInitReg() const
576 return (regNumber)_lvArgInitReg;
579 void SetArgInitReg(regNumber reg)
581 _lvArgInitReg = (regNumberSmall)reg;
582 assert(_lvArgInitReg == reg);
585 /////////////////////
587 __declspec(property(get = GetArgInitRegPair, put = SetArgInitRegPair)) regPairNo lvArgInitRegPair;
589 regPairNo GetArgInitRegPair() const
591 regPairNo regPair = (regPairNo)_lvArgInitRegPair;
592 assert(regPair >= REG_PAIR_FIRST && regPair <= REG_PAIR_LAST);
596 void SetArgInitRegPair(regPairNo regPair)
598 assert(regPair >= REG_PAIR_FIRST && regPair <= REG_PAIR_LAST);
599 _lvArgInitRegPair = (regPairNoSmall)regPair;
600 assert(_lvArgInitRegPair == regPair);
603 /////////////////////
605 bool lvIsRegCandidate() const
607 return lvLRACandidate != 0;
610 bool lvIsInReg() const
612 return lvIsRegCandidate() && (lvRegNum != REG_STK);
615 #else // LEGACY_BACKEND
617 bool lvIsRegCandidate() const
619 return lvTracked != 0;
622 bool lvIsInReg() const
624 return lvRegister != 0;
627 #endif // LEGACY_BACKEND
629 regMaskTP lvRegMask() const
631 regMaskTP regMask = RBM_NONE;
632 if (varTypeIsFloating(TypeGet()))
634 if (lvRegNum != REG_STK)
636 regMask = genRegMaskFloat(lvRegNum, TypeGet());
641 if (lvRegNum != REG_STK)
643 regMask = genRegMask(lvRegNum);
646 // For longs we may have two regs
647 if (isRegPairType(lvType) && lvOtherReg != REG_STK)
649 regMask |= genRegMask(lvOtherReg);
655 regMaskSmall lvPrefReg; // set of regs it prefers to live in
657 unsigned short lvVarIndex; // variable tracking index
658 unsigned short lvRefCnt; // unweighted (real) reference count
659 unsigned lvRefCntWtd; // weighted reference count
660 int lvStkOffs; // stack offset of home
661 unsigned lvExactSize; // (exact) size of the type in bytes
663 // Is this a promoted struct?
664 // This method returns true only for structs (including SIMD structs), not for
665 // locals that are split on a 32-bit target.
666 // It is only necessary to use this:
667 // 1) if only structs are wanted, and
668 // 2) if Lowering has already been done.
669 // Otherwise lvPromoted is valid.
670 bool lvPromotedStruct()
672 #if !defined(_TARGET_64BIT_)
673 return (lvPromoted && !varTypeIsLong(lvType));
674 #else // defined(_TARGET_64BIT_)
676 #endif // defined(_TARGET_64BIT_)
679 unsigned lvSize() // Size needed for storage representation. Only used for structs or TYP_BLK.
681 // TODO-Review: Sometimes we get called on ARM with HFA struct variables that have been promoted,
682 // where the struct itself is no longer used because all access is via its member fields.
683 // When that happens, the struct is marked as unused and its type has been changed to
684 // TYP_INT (to keep the GC tracking code from looking at it).
685 // See Compiler::raAssignVars() for details. For example:
686 // N002 ( 4, 3) [00EA067C] ------------- return struct $346
687 // N001 ( 3, 2) [00EA0628] ------------- lclVar struct(U) V03 loc2
688 // float V03.f1 (offs=0x00) -> V12 tmp7
689 // f8 (last use) (last use) $345
690 // Here, the "struct(U)" shows that the "V03 loc2" variable is unused. Not shown is that V03
691 // is now TYP_INT in the local variable table. It's not really unused, because it's in the tree.
693 assert(varTypeIsStruct(lvType) || (lvType == TYP_BLK) || (lvPromoted && lvUnusedStruct));
695 #if defined(FEATURE_SIMD) && !defined(_TARGET_64BIT_)
696 // For 32-bit architectures, we make local variable SIMD12 types 16 bytes instead of just 12. We can't do
697 // this for arguments, which must be passed according the defined ABI.
698 if ((lvType == TYP_SIMD12) && !lvIsParam)
700 assert(lvExactSize == 12);
703 #endif // defined(FEATURE_SIMD) && !defined(_TARGET_64BIT_)
705 return (unsigned)(roundUp(lvExactSize, TARGET_POINTER_SIZE));
708 unsigned lvSlotNum; // original slot # (if remapped)
710 typeInfo lvVerTypeInfo; // type info needed for verification
712 BYTE* lvGcLayout; // GC layout info for structs
715 GenTreePtr lvKnownDim; // array size if known
719 BlockSet lvRefBlks; // Set of blocks that contain refs
720 GenTreePtr lvDefStmt; // Pointer to the statement with the single definition
721 void lvaDisqualifyVar(); // Call to disqualify a local variable from use in optAddCopies
723 var_types TypeGet() const
725 return (var_types)lvType;
727 bool lvStackAligned() const
729 assert(lvIsStructField);
730 return ((lvFldOffset % sizeof(void*)) == 0);
732 bool lvNormalizeOnLoad() const
734 return varTypeIsSmall(TypeGet()) &&
735 // lvIsStructField is treated the same as the aliased local, see fgDoNormalizeOnStore.
736 (lvIsParam || lvAddrExposed || lvIsStructField);
739 bool lvNormalizeOnStore()
741 return varTypeIsSmall(TypeGet()) &&
742 // lvIsStructField is treated the same as the aliased local, see fgDoNormalizeOnStore.
743 !(lvIsParam || lvAddrExposed || lvIsStructField);
746 void lvaResetSortAgainFlag(Compiler* pComp);
747 void decRefCnts(BasicBlock::weight_t weight, Compiler* pComp, bool propagate = true);
748 void incRefCnts(BasicBlock::weight_t weight, Compiler* pComp, bool propagate = true);
749 void setPrefReg(regNumber regNum, Compiler* pComp);
750 void addPrefReg(regMaskTP regMask, Compiler* pComp);
751 bool IsFloatRegType() const
753 return isFloatRegType(lvType) || lvIsHfaRegArg();
755 var_types GetHfaType() const
757 return lvIsHfa() ? (lvHfaTypeIsFloat() ? TYP_FLOAT : TYP_DOUBLE) : TYP_UNDEF;
759 void SetHfaType(var_types type)
761 assert(varTypeIsFloating(type));
762 lvSetHfaTypeIsFloat(type == TYP_FLOAT);
765 #ifndef LEGACY_BACKEND
766 var_types lvaArgType();
769 PerSsaArray lvPerSsaData;
772 // Keep track of the # of SsaNames, for a bounds check.
773 unsigned lvNumSsaNames;
776 // Returns the address of the per-Ssa data for the given ssaNum (which is required
777 // not to be the SsaConfig::RESERVED_SSA_NUM, which indicates that the variable is
778 // not an SSA variable).
779 LclSsaVarDsc* GetPerSsaData(unsigned ssaNum)
781 assert(ssaNum != SsaConfig::RESERVED_SSA_NUM);
782 assert(SsaConfig::RESERVED_SSA_NUM == 0);
783 unsigned zeroBased = ssaNum - SsaConfig::UNINIT_SSA_NUM;
784 assert(zeroBased < lvNumSsaNames);
785 return &lvPerSsaData.GetRef(zeroBased);
790 void PrintVarReg() const
792 if (isRegPairType(TypeGet()))
794 printf("%s:%s", getRegName(lvOtherReg), // hi32
795 getRegName(lvRegNum)); // lo32
799 printf("%s", getRegName(lvRegNum));
804 }; // class LclVarDsc
807 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
808 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
812 XX The temporary lclVars allocated by the compiler for code generation XX
814 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
815 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
818 /*****************************************************************************
820 * The following keeps track of temporaries allocated in the stack frame
821 * during code-generation (after register allocation). These spill-temps are
822 * only used if we run out of registers while evaluating a tree.
824 * These are different from the more common temps allocated by lvaGrabTemp().
835 static const int BAD_TEMP_OFFSET = 0xDDDDDDDD; // used as a sentinel "bad value" for tdOffs in DEBUG
843 TempDsc(int _tdNum, unsigned _tdSize, var_types _tdType) : tdNum(_tdNum), tdSize((BYTE)_tdSize), tdType(_tdType)
847 0); // temps must have a negative number (so they have a different number from all local variables)
848 tdOffs = BAD_TEMP_OFFSET;
852 IMPL_LIMITATION("too many spill temps");
857 bool tdLegalOffset() const
859 return tdOffs != BAD_TEMP_OFFSET;
863 int tdTempOffs() const
865 assert(tdLegalOffset());
868 void tdSetTempOffs(int offs)
871 assert(tdLegalOffset());
873 void tdAdjustTempOffs(int offs)
876 assert(tdLegalOffset());
879 int tdTempNum() const
884 unsigned tdTempSize() const
888 var_types tdTempType() const
894 // interface to hide linearscan implementation from rest of compiler
895 class LinearScanInterface
898 virtual void doLinearScan() = 0;
899 virtual void recordVarLocationsAtStartOfBB(BasicBlock* bb) = 0;
902 LinearScanInterface* getLinearScanAllocator(Compiler* comp);
904 // Information about arrays: their element type and size, and the offset of the first element.
905 // We label GT_IND's that are array indices with GTF_IND_ARR_INDEX, and, for such nodes,
906 // associate an array info via the map retrieved by GetArrayInfoMap(). This information is used,
907 // for example, in value numbering of array index expressions.
910 var_types m_elemType;
911 CORINFO_CLASS_HANDLE m_elemStructType;
913 unsigned m_elemOffset;
915 ArrayInfo() : m_elemType(TYP_UNDEF), m_elemStructType(nullptr), m_elemSize(0), m_elemOffset(0)
919 ArrayInfo(var_types elemType, unsigned elemSize, unsigned elemOffset, CORINFO_CLASS_HANDLE elemStructType)
920 : m_elemType(elemType), m_elemStructType(elemStructType), m_elemSize(elemSize), m_elemOffset(elemOffset)
925 // This enumeration names the phases into which we divide compilation. The phases should completely
926 // partition a compilation.
929 #define CompPhaseNameMacro(enum_nm, string_nm, short_nm, hasChildren, parent) enum_nm,
930 #include "compphases.h"
934 extern const char* PhaseNames[];
935 extern const char* PhaseEnums[];
936 extern const LPCWSTR PhaseShortNames[];
938 // The following enum provides a simple 1:1 mapping to CLR API's
939 enum API_ICorJitInfo_Names
941 #define DEF_CLR_API(name) API_##name,
942 #include "ICorJitInfo_API_names.h"
946 //---------------------------------------------------------------
950 // A "CompTimeInfo" is a structure for tracking the compilation time of one or more methods.
951 // We divide a compilation into a sequence of contiguous phases, and track the total (per-thread) cycles
952 // of the compilation, as well as the cycles for each phase. We also track the number of bytecodes.
953 // If there is a failure in reading a timer at any point, the "CompTimeInfo" becomes invalid, as indicated
954 // by "m_timerFailure" being true.
955 // If FEATURE_JIT_METHOD_PERF is not set, we define a minimal form of this, enough to let other code compile.
958 #ifdef FEATURE_JIT_METHOD_PERF
959 // The string names of the phases.
960 static const char* PhaseNames[];
962 static bool PhaseHasChildren[];
963 static int PhaseParent[];
965 unsigned m_byteCodeBytes;
966 unsigned __int64 m_totalCycles;
967 unsigned __int64 m_invokesByPhase[PHASE_NUMBER_OF];
968 unsigned __int64 m_cyclesByPhase[PHASE_NUMBER_OF];
969 #if MEASURE_CLRAPI_CALLS
970 unsigned __int64 m_CLRinvokesByPhase[PHASE_NUMBER_OF];
971 unsigned __int64 m_CLRcyclesByPhase[PHASE_NUMBER_OF];
973 // For better documentation, we call EndPhase on
974 // non-leaf phases. We should also call EndPhase on the
975 // last leaf subphase; obviously, the elapsed cycles between the EndPhase
976 // for the last leaf subphase and the EndPhase for an ancestor should be very small.
977 // We add all such "redundant end phase" intervals to this variable below; we print
978 // it out in a report, so we can verify that it is, indeed, very small. If it ever
979 // isn't, this means that we're doing something significant between the end of the last
980 // declared subphase and the end of its parent.
981 unsigned __int64 m_parentPhaseEndSlop;
984 #if MEASURE_CLRAPI_CALLS
985 // The following measures the time spent inside each individual CLR API call.
986 unsigned m_allClrAPIcalls;
987 unsigned m_perClrAPIcalls[API_ICorJitInfo_Names::API_COUNT];
988 unsigned __int64 m_allClrAPIcycles;
989 unsigned __int64 m_perClrAPIcycles[API_ICorJitInfo_Names::API_COUNT];
990 unsigned __int32 m_maxClrAPIcycles[API_ICorJitInfo_Names::API_COUNT];
991 #endif // MEASURE_CLRAPI_CALLS
993 CompTimeInfo(unsigned byteCodeBytes);
997 #ifdef FEATURE_JIT_METHOD_PERF
999 #if MEASURE_CLRAPI_CALLS
1000 struct WrapICorJitInfo;
1003 // This class summarizes the JIT time information over the course of a run: the number of methods compiled,
1004 // and the total and maximum timings. (These are instances of the "CompTimeInfo" type described above).
1005 // The operation of adding a single method's timing to the summary may be performed concurrently by several
1006 // threads, so it is protected by a lock.
1007 // This class is intended to be used as a singleton type, with only a single instance.
1008 class CompTimeSummaryInfo
1010 // This lock protects the fields of all CompTimeSummaryInfo(s) (of which we expect there to be one).
1011 static CritSecObject s_compTimeSummaryLock;
1015 CompTimeInfo m_total;
1016 CompTimeInfo m_maximum;
1018 int m_numFilteredMethods;
1019 CompTimeInfo m_filtered;
1021 // This method computes the number of cycles/sec for the current machine. The cycles are those counted
1022 // by GetThreadCycleTime; we assume that these are of equal duration, though that is not necessarily true.
1023 // If any OS interaction fails, returns 0.0.
1024 double CyclesPerSecond();
1026 // This can use what ever data you want to determine if the value to be added
1027 // belongs in the filtered section (it's always included in the unfiltered section)
1028 bool IncludedInFilteredData(CompTimeInfo& info);
1031 // This is the unique CompTimeSummaryInfo object for this instance of the runtime.
1032 static CompTimeSummaryInfo s_compTimeSummary;
1034 CompTimeSummaryInfo()
1035 : m_numMethods(0), m_totMethods(0), m_total(0), m_maximum(0), m_numFilteredMethods(0), m_filtered(0)
1039 // Assumes that "info" is a completed CompTimeInfo for a compilation; adds it to the summary.
1040 // This is thread safe.
1041 void AddInfo(CompTimeInfo& info, bool includePhases);
1043 // Print the summary information to "f".
1044 // This is not thread-safe; assumed to be called by only one thread.
1045 void Print(FILE* f);
1048 // A JitTimer encapsulates a CompTimeInfo for a single compilation. It also tracks the start of compilation,
1049 // and when the current phase started. This is intended to be part of a Compilation object. This is
1050 // disabled (FEATURE_JIT_METHOD_PERF not defined) when FEATURE_CORECLR is set, or on non-windows platforms.
1054 unsigned __int64 m_start; // Start of the compilation.
1055 unsigned __int64 m_curPhaseStart; // Start of the current phase.
1056 #if MEASURE_CLRAPI_CALLS
1057 unsigned __int64 m_CLRcallStart; // Start of the current CLR API call (if any).
1058 unsigned __int64 m_CLRcallInvokes; // CLR API invokes under current outer so far
1059 unsigned __int64 m_CLRcallCycles; // CLR API cycles under current outer so far.
1060 int m_CLRcallAPInum; // The enum/index of the current CLR API call (or -1).
1061 static double s_cyclesPerSec; // Cached for speedier measurements
1064 Phases m_lastPhase; // The last phase that was completed (or (Phases)-1 to start).
1066 CompTimeInfo m_info; // The CompTimeInfo for this compilation.
1068 static CritSecObject s_csvLock; // Lock to protect the time log file.
1069 void PrintCsvMethodStats(Compiler* comp);
1072 void* operator new(size_t);
1073 void* operator new[](size_t);
1074 void operator delete(void*);
1075 void operator delete[](void*);
1078 // Initialized the timer instance
1079 JitTimer(unsigned byteCodeSize);
1081 static JitTimer* Create(Compiler* comp, unsigned byteCodeSize)
1083 return ::new (comp, CMK_Unknown) JitTimer(byteCodeSize);
1086 static void PrintCsvHeader();
1088 // Ends the current phase (argument is for a redundant check).
1089 void EndPhase(Phases phase);
1091 #if MEASURE_CLRAPI_CALLS
1092 // Start and end a timed CLR API call.
1093 void CLRApiCallEnter(unsigned apix);
1094 void CLRApiCallLeave(unsigned apix);
1095 #endif // MEASURE_CLRAPI_CALLS
1097 // Completes the timing of the current method, which is assumed to have "byteCodeBytes" bytes of bytecode,
1098 // and adds it to "sum".
1099 void Terminate(Compiler* comp, CompTimeSummaryInfo& sum, bool includePhases);
1101 // Attempts to query the cycle counter of the current thread. If successful, returns "true" and sets
1102 // *cycles to the cycle counter value. Otherwise, returns false and sets the "m_timerFailure" flag of
1103 // "m_info" to true.
1104 bool GetThreadCycles(unsigned __int64* cycles)
1106 bool res = CycleTimer::GetThreadCyclesS(cycles);
1109 m_info.m_timerFailure = true;
1114 #endif // FEATURE_JIT_METHOD_PERF
1116 //------------------- Function/Funclet info -------------------------------
1117 DECLARE_TYPED_ENUM(FuncKind, BYTE)
1119 FUNC_ROOT, // The main/root function (always id==0)
1120 FUNC_HANDLER, // a funclet associated with an EH handler (finally, fault, catch, filter handler)
1121 FUNC_FILTER, // a funclet associated with an EH filter
1124 END_DECLARE_TYPED_ENUM(FuncKind, BYTE)
1131 BYTE funFlags; // Currently unused, just here for padding
1132 unsigned short funEHIndex; // index, into the ebd table, of innermost EH clause corresponding to this
1133 // funclet. It is only valid if funKind field indicates this is a
1134 // EH-related funclet: FUNC_HANDLER or FUNC_FILTER
1136 #if defined(_TARGET_AMD64_)
1138 // TODO-AMD64-Throughput: make the AMD64 info more like the ARM info to avoid having this large static array.
1139 emitLocation* startLoc;
1140 emitLocation* endLoc;
1141 emitLocation* coldStartLoc; // locations for the cold section, if there is one.
1142 emitLocation* coldEndLoc;
1143 UNWIND_INFO unwindHeader;
1144 // Maximum of 255 UNWIND_CODE 'nodes' and then the unwind header. If there are an odd
1145 // number of codes, the VM or Zapper will 4-byte align the whole thing.
1146 BYTE unwindCodes[offsetof(UNWIND_INFO, UnwindCode) + (0xFF * sizeof(UNWIND_CODE))];
1147 unsigned unwindCodeSlot;
1149 #ifdef UNIX_AMD64_ABI
1150 jitstd::vector<CFI_CODE>* cfiCodes;
1151 #endif // UNIX_AMD64_ABI
1153 #elif defined(_TARGET_ARMARCH_)
1155 UnwindInfo uwi; // Unwind information for this function/funclet's hot section
1156 UnwindInfo* uwiCold; // Unwind information for this function/funclet's cold section
1157 // Note: we only have a pointer here instead of the actual object,
1158 // to save memory in the JIT case (compared to the NGEN case),
1159 // where we don't have any cold section.
1160 // Note 2: we currently don't support hot/cold splitting in functions
1161 // with EH, so uwiCold will be NULL for all funclets.
1163 #endif // _TARGET_ARMARCH_
1165 // Eventually we may want to move rsModifiedRegsMask, lvaOutgoingArgSize, and anything else
1166 // that isn't shared between the main function body and funclets.
1169 struct fgArgTabEntry
1172 #if defined(FEATURE_UNIX_AMD64_STRUCT_PASSING)
1175 otherRegNum = REG_NA;
1176 isStruct = false; // is this a struct arg
1178 #endif // defined(FEATURE_UNIX_AMD64_STRUCT_PASSING)
1180 GenTreePtr node; // Initially points at the Op1 field of 'parent', but if the argument is replaced with an GT_ASG or
1182 // it will point at the actual argument in the gtCallLateArgs list.
1183 GenTreePtr parent; // Points at the GT_LIST node in the gtCallArgs for this argument
1185 unsigned argNum; // The original argument number, also specifies the required argument evaluation order from the IL
1187 regNumber regNum; // The (first) register to use when passing this argument, set to REG_STK for arguments passed on
1189 unsigned numRegs; // Count of number of registers that this argument uses
1191 // A slot is a pointer sized region in the OutArg area.
1192 unsigned slotNum; // When an argument is passed in the OutArg area this is the slot number in the OutArg area
1193 unsigned numSlots; // Count of number of slots that this argument uses
1195 unsigned alignment; // 1 or 2 (slots/registers)
1196 unsigned lateArgInx; // index into gtCallLateArgs list
1197 unsigned tmpNum; // the LclVar number if we had to force evaluation of this arg
1199 bool isSplit : 1; // True when this argument is split between the registers and OutArg area
1200 bool needTmp : 1; // True when we force this argument's evaluation into a temp LclVar
1201 bool needPlace : 1; // True when we must replace this argument with a placeholder node
1202 bool isTmp : 1; // True when we setup a temp LclVar for this argument due to size issues with the struct
1203 bool processed : 1; // True when we have decided the evaluation order for this argument in the gtCallLateArgs
1204 bool isHfaRegArg : 1; // True when the argument is passed as a HFA in FP registers.
1205 bool isBackFilled : 1; // True when the argument fills a register slot skipped due to alignment requirements of
1206 // previous arguments.
1207 bool isNonStandard : 1; // True if it is an arg that is passed in a reg other than a standard arg reg, or is forced
1208 // to be on the stack despite its arg list position.
1210 #if defined(FEATURE_UNIX_AMD64_STRUCT_PASSING)
1211 bool isStruct : 1; // True if this is a struct arg
1213 regNumber otherRegNum; // The (second) register to use when passing this argument.
1215 SYSTEMV_AMD64_CORINFO_STRUCT_REG_PASSING_DESCRIPTOR structDesc;
1216 #elif defined(_TARGET_X86_)
1217 __declspec(property(get = getIsStruct)) bool isStruct;
1220 return varTypeIsStruct(node);
1222 #endif // _TARGET_X86_
1225 void SetIsHfaRegArg(bool hfaRegArg)
1227 isHfaRegArg = hfaRegArg;
1230 void SetIsBackFilled(bool backFilled)
1232 isBackFilled = backFilled;
1235 bool IsBackFilled() const
1237 return isBackFilled;
1239 #else // !_TARGET_ARM_
1240 // To make the callers easier, we allow these calls (and the isHfaRegArg and isBackFilled data members) for all
1242 void SetIsHfaRegArg(bool hfaRegArg)
1246 void SetIsBackFilled(bool backFilled)
1250 bool IsBackFilled() const
1254 #endif // !_TARGET_ARM_
1260 typedef struct fgArgTabEntry* fgArgTabEntryPtr;
1262 //-------------------------------------------------------------------------
1264 // The class fgArgInfo is used to handle the arguments
1265 // when morphing a GT_CALL node.
1270 Compiler* compiler; // Back pointer to the compiler instance so that we can allocate memory
1271 GenTreePtr callTree; // Back pointer to the GT_CALL node for this fgArgInfo
1272 unsigned argCount; // Updatable arg count value
1273 unsigned nextSlotNum; // Updatable slot count value
1274 unsigned stkLevel; // Stack depth when we make this call (for x86)
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();
1326 void EvalArgsToTemps();
1328 void RecordStkLevel(unsigned stkLvl);
1329 unsigned RetrieveStkLevel();
1335 fgArgTabEntryPtr* ArgTable()
1339 unsigned GetNextSlotNum()
1349 return hasStackArgs;
1351 bool AreArgsComplete() const
1353 return argsComplete;
1358 // XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1359 // We have the ability to mark source expressions with "Test Labels."
1360 // These drive assertions within the JIT, or internal JIT testing. For example, we could label expressions
1361 // that should be CSE defs, and other expressions that should uses of those defs, with a shared label.
1363 enum TestLabel // This must be kept identical to System.Runtime.CompilerServices.JitTestLabel.TestLabel.
1366 TL_VN, // Defines a "VN equivalence class". (For full VN, including exceptions thrown).
1367 TL_VNNorm, // Like above, but uses the non-exceptional value of the expression.
1368 TL_CSE_Def, // This must be identified in the JIT as a CSE def
1369 TL_CSE_Use, // This must be identified in the JIT as a CSE use
1370 TL_LoopHoist, // Expression must (or must not) be hoisted out of the loop.
1373 struct TestLabelAndNum
1378 TestLabelAndNum() : m_tl(TestLabel(0)), m_num(0)
1383 typedef SimplerHashTable<GenTreePtr, PtrKeyFuncs<GenTree>, TestLabelAndNum, JitSimplerHashBehavior> NodeToTestDataMap;
1385 // XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1388 // This class implements the "IAllocator" interface, so that we can use
1389 // utilcode collection classes in the JIT, and have them use the JIT's allocator.
1391 class CompAllocator : public IAllocator
1394 #if MEASURE_MEM_ALLOC
1398 CompAllocator(Compiler* comp, CompMemKind cmk)
1400 #if MEASURE_MEM_ALLOC
1406 inline void* Alloc(size_t sz);
1408 inline void* ArrayAlloc(size_t elems, size_t elemSize);
1410 // For the compiler's no-release allocator, free operations are no-ops.
1417 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1418 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1420 XX The big guy. The sections are currently organized as : XX
1422 XX o GenTree and BasicBlock XX
1434 XX o PrologScopeInfo XX
1435 XX o CodeGenerator XX
1440 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1441 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1446 friend class emitter;
1447 friend class UnwindInfo;
1448 friend class UnwindFragmentInfo;
1449 friend class UnwindEpilogInfo;
1450 friend class JitTimer;
1451 friend class LinearScan;
1452 friend class fgArgInfo;
1453 friend class Rationalizer;
1455 friend class Lowering;
1456 friend class CSE_DataFlow;
1457 friend class CSE_Heuristic;
1458 friend class CodeGenInterface;
1459 friend class CodeGen;
1460 friend class LclVarDsc;
1461 friend class TempDsc;
1463 friend class ObjectAllocator;
1465 #ifndef _TARGET_64BIT_
1466 friend class DecomposeLongs;
1467 #endif // !_TARGET_64BIT_
1470 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1471 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1473 XX Misc structs definitions XX
1475 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1476 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1480 hashBvGlobalData hbvGlobalData; // Used by the hashBv bitvector package.
1499 bool dumpIRDataflow;
1500 bool dumpIRBlockHeaders;
1502 LPCWSTR dumpIRPhase;
1503 LPCWSTR dumpIRFormat;
1505 bool shouldUseVerboseTrees();
1506 bool asciiTrees; // If true, dump trees using only ASCII characters
1507 bool shouldDumpASCIITrees();
1508 bool verboseSsa; // If true, produce especially verbose dump output in SSA construction.
1509 bool shouldUseVerboseSsa();
1510 bool treesBeforeAfterMorph; // If true, print trees before/after morphing (paired by an intra-compilation id:
1511 int morphNum; // This counts the the trees that have been morphed, allowing us to label each uniquely.
1513 const char* VarNameToStr(VarName name)
1518 DWORD expensiveDebugCheckLevel;
1521 #if FEATURE_MULTIREG_RET
1522 GenTreePtr impAssignMultiRegTypeToVar(GenTreePtr op, CORINFO_CLASS_HANDLE hClass);
1523 #endif // FEATURE_MULTIREG_RET
1526 bool isSingleFloat32Struct(CORINFO_CLASS_HANDLE hClass);
1527 #endif // ARM_SOFTFP
1529 //-------------------------------------------------------------------------
1530 // Functions to handle homogeneous floating-point aggregates (HFAs) in ARM.
1531 // HFAs are one to four element structs where each element is the same
1532 // type, either all float or all double. They are treated specially
1533 // in the ARM Procedure Call Standard, specifically, they are passed in
1534 // floating-point registers instead of the general purpose registers.
1537 bool IsHfa(CORINFO_CLASS_HANDLE hClass);
1538 bool IsHfa(GenTreePtr tree);
1540 var_types GetHfaType(GenTreePtr tree);
1541 unsigned GetHfaCount(GenTreePtr tree);
1543 var_types GetHfaType(CORINFO_CLASS_HANDLE hClass);
1544 unsigned GetHfaCount(CORINFO_CLASS_HANDLE hClass);
1546 bool IsMultiRegPassedType(CORINFO_CLASS_HANDLE hClass);
1547 bool IsMultiRegReturnedType(CORINFO_CLASS_HANDLE hClass);
1549 //-------------------------------------------------------------------------
1550 // The following is used for validating format of EH table
1554 typedef struct EHNodeDsc* pEHNodeDsc;
1556 EHNodeDsc* ehnTree; // root of the tree comprising the EHnodes.
1557 EHNodeDsc* ehnNext; // root of the tree comprising the EHnodes.
1570 EHBlockType ehnBlockType; // kind of EH block
1571 IL_OFFSET ehnStartOffset; // IL offset of start of the EH block
1572 IL_OFFSET ehnEndOffset; // IL offset past end of the EH block. (TODO: looks like verInsertEhNode() sets this to
1573 // the last IL offset, not "one past the last one", i.e., the range Start to End is
1575 pEHNodeDsc ehnNext; // next (non-nested) block in sequential order
1576 pEHNodeDsc ehnChild; // leftmost nested block
1578 pEHNodeDsc ehnTryNode; // for filters and handlers, the corresponding try node
1579 pEHNodeDsc ehnHandlerNode; // for a try node, the corresponding handler node
1581 pEHNodeDsc ehnFilterNode; // if this is a try node and has a filter, otherwise 0
1582 pEHNodeDsc ehnEquivalent; // if blockType=tryNode, start offset and end offset is same,
1584 inline void ehnSetTryNodeType()
1586 ehnBlockType = TryNode;
1588 inline void ehnSetFilterNodeType()
1590 ehnBlockType = FilterNode;
1592 inline void ehnSetHandlerNodeType()
1594 ehnBlockType = HandlerNode;
1596 inline void ehnSetFinallyNodeType()
1598 ehnBlockType = FinallyNode;
1600 inline void ehnSetFaultNodeType()
1602 ehnBlockType = FaultNode;
1605 inline BOOL ehnIsTryBlock()
1607 return ehnBlockType == TryNode;
1609 inline BOOL ehnIsFilterBlock()
1611 return ehnBlockType == FilterNode;
1613 inline BOOL ehnIsHandlerBlock()
1615 return ehnBlockType == HandlerNode;
1617 inline BOOL ehnIsFinallyBlock()
1619 return ehnBlockType == FinallyNode;
1621 inline BOOL ehnIsFaultBlock()
1623 return ehnBlockType == FaultNode;
1626 // returns true if there is any overlap between the two nodes
1627 static inline BOOL ehnIsOverlap(pEHNodeDsc node1, pEHNodeDsc node2)
1629 if (node1->ehnStartOffset < node2->ehnStartOffset)
1631 return (node1->ehnEndOffset >= node2->ehnStartOffset);
1635 return (node1->ehnStartOffset <= node2->ehnEndOffset);
1639 // fails with BADCODE if inner is not completely nested inside outer
1640 static inline BOOL ehnIsNested(pEHNodeDsc inner, pEHNodeDsc outer)
1642 return ((inner->ehnStartOffset >= outer->ehnStartOffset) && (inner->ehnEndOffset <= outer->ehnEndOffset));
1646 //-------------------------------------------------------------------------
1647 // Exception handling functions
1650 #if !FEATURE_EH_FUNCLETS
1652 bool ehNeedsShadowSPslots()
1654 return (info.compXcptnsCount || opts.compDbgEnC);
1657 // 0 for methods with no EH
1658 // 1 for methods with non-nested EH, or where only the try blocks are nested
1659 // 2 for a method with a catch within a catch
1661 unsigned ehMaxHndNestingCount;
1663 #endif // !FEATURE_EH_FUNCLETS
1665 static bool jitIsBetween(unsigned value, unsigned start, unsigned end);
1666 static bool jitIsBetweenInclusive(unsigned value, unsigned start, unsigned end);
1668 bool bbInCatchHandlerILRange(BasicBlock* blk);
1669 bool bbInFilterILRange(BasicBlock* blk);
1670 bool bbInTryRegions(unsigned regionIndex, BasicBlock* blk);
1671 bool bbInExnFlowRegions(unsigned regionIndex, BasicBlock* blk);
1672 bool bbInHandlerRegions(unsigned regionIndex, BasicBlock* blk);
1673 bool bbInCatchHandlerRegions(BasicBlock* tryBlk, BasicBlock* hndBlk);
1674 unsigned short bbFindInnermostCommonTryRegion(BasicBlock* bbOne, BasicBlock* bbTwo);
1676 unsigned short bbFindInnermostTryRegionContainingHandlerRegion(unsigned handlerIndex);
1677 unsigned short bbFindInnermostHandlerRegionContainingTryRegion(unsigned tryIndex);
1679 // Returns true if "block" is the start of a try region.
1680 bool bbIsTryBeg(BasicBlock* block);
1682 // Returns true if "block" is the start of a handler or filter region.
1683 bool bbIsHandlerBeg(BasicBlock* block);
1685 // Returns true iff "block" is where control flows if an exception is raised in the
1686 // try region, and sets "*regionIndex" to the index of the try for the handler.
1687 // Differs from "IsHandlerBeg" in the case of filters, where this is true for the first
1688 // block of the filter, but not for the filter's handler.
1689 bool bbIsExFlowBlock(BasicBlock* block, unsigned* regionIndex);
1691 bool ehHasCallableHandlers();
1693 // Return the EH descriptor for the given region index.
1694 EHblkDsc* ehGetDsc(unsigned regionIndex);
1696 // Return the EH index given a region descriptor.
1697 unsigned ehGetIndex(EHblkDsc* ehDsc);
1699 // Return the EH descriptor index of the enclosing try, for the given region index.
1700 unsigned ehGetEnclosingTryIndex(unsigned regionIndex);
1702 // Return the EH descriptor index of the enclosing handler, for the given region index.
1703 unsigned ehGetEnclosingHndIndex(unsigned regionIndex);
1705 // Return the EH descriptor for the most nested 'try' region this BasicBlock is a member of (or nullptr if this
1706 // block is not in a 'try' region).
1707 EHblkDsc* ehGetBlockTryDsc(BasicBlock* block);
1709 // Return the EH descriptor for the most nested filter or handler region this BasicBlock is a member of (or nullptr
1710 // if this block is not in a filter or handler region).
1711 EHblkDsc* ehGetBlockHndDsc(BasicBlock* block);
1713 // Return the EH descriptor for the most nested region that may handle exceptions raised in this BasicBlock (or
1714 // nullptr if this block's exceptions propagate to caller).
1715 EHblkDsc* ehGetBlockExnFlowDsc(BasicBlock* block);
1717 EHblkDsc* ehIsBlockTryLast(BasicBlock* block);
1718 EHblkDsc* ehIsBlockHndLast(BasicBlock* block);
1719 bool ehIsBlockEHLast(BasicBlock* block);
1721 bool ehBlockHasExnFlowDsc(BasicBlock* block);
1723 // Return the region index of the most nested EH region this block is in.
1724 unsigned ehGetMostNestedRegionIndex(BasicBlock* block, bool* inTryRegion);
1726 // Find the true enclosing try index, ignoring 'mutual protect' try. Uses IL ranges to check.
1727 unsigned ehTrueEnclosingTryIndexIL(unsigned regionIndex);
1729 // Return the index of the most nested enclosing region for a particular EH region. Returns NO_ENCLOSING_INDEX
1730 // if there is no enclosing region. If the returned index is not NO_ENCLOSING_INDEX, then '*inTryRegion'
1731 // is set to 'true' if the enclosing region is a 'try', or 'false' if the enclosing region is a handler.
1732 // (It can never be a filter.)
1733 unsigned ehGetEnclosingRegionIndex(unsigned regionIndex, bool* inTryRegion);
1735 // A block has been deleted. Update the EH table appropriately.
1736 void ehUpdateForDeletedBlock(BasicBlock* block);
1738 // Determine whether a block can be deleted while preserving the EH normalization rules.
1739 bool ehCanDeleteEmptyBlock(BasicBlock* block);
1741 // Update the 'last' pointers in the EH table to reflect new or deleted blocks in an EH region.
1742 void ehUpdateLastBlocks(BasicBlock* oldLast, BasicBlock* newLast);
1744 // For a finally handler, find the region index that the BBJ_CALLFINALLY lives in that calls the handler,
1745 // or NO_ENCLOSING_INDEX if the BBJ_CALLFINALLY lives in the main function body. Normally, the index
1746 // is the same index as the handler (and the BBJ_CALLFINALLY lives in the 'try' region), but for AMD64 the
1747 // BBJ_CALLFINALLY lives in the enclosing try or handler region, whichever is more nested, or the main function
1748 // body. If the returned index is not NO_ENCLOSING_INDEX, then '*inTryRegion' is set to 'true' if the
1749 // BBJ_CALLFINALLY lives in the returned index's 'try' region, or 'false' if lives in the handler region. (It never
1750 // lives in a filter.)
1751 unsigned ehGetCallFinallyRegionIndex(unsigned finallyIndex, bool* inTryRegion);
1753 // Find the range of basic blocks in which all BBJ_CALLFINALLY will be found that target the 'finallyIndex' region's
1754 // handler. Set begBlk to the first block, and endBlk to the block after the last block of the range
1755 // (nullptr if the last block is the last block in the program).
1756 // Precondition: 'finallyIndex' is the EH region of a try/finally clause.
1757 void ehGetCallFinallyBlockRange(unsigned finallyIndex, BasicBlock** begBlk, BasicBlock** endBlk);
1760 // Given a BBJ_CALLFINALLY block and the EH region index of the finally it is calling, return
1761 // 'true' if the BBJ_CALLFINALLY is in the correct EH region.
1762 bool ehCallFinallyInCorrectRegion(BasicBlock* blockCallFinally, unsigned finallyIndex);
1765 #if FEATURE_EH_FUNCLETS
1766 // Do we need a PSPSym in the main function? For codegen purposes, we only need one
1767 // if there is a filter that protects a region with a nested EH clause (such as a
1768 // try/catch nested in the 'try' body of a try/filter/filter-handler). See
1769 // genFuncletProlog() for more details. However, the VM seems to use it for more
1770 // purposes, maybe including debugging. Until we are sure otherwise, always create
1771 // a PSPSym for functions with any EH.
1772 bool ehNeedsPSPSym() const
1774 return compHndBBtabCount > 0;
1777 bool ehAnyFunclets(); // Are there any funclets in this function?
1778 unsigned ehFuncletCount(); // Return the count of funclets in the function
1780 unsigned bbThrowIndex(BasicBlock* blk); // Get the index to use as the cache key for sharing throw blocks
1781 #else // !FEATURE_EH_FUNCLETS
1782 bool ehAnyFunclets()
1786 unsigned ehFuncletCount()
1791 unsigned bbThrowIndex(BasicBlock* blk)
1793 return blk->bbTryIndex;
1794 } // Get the index to use as the cache key for sharing throw blocks
1795 #endif // !FEATURE_EH_FUNCLETS
1797 // Returns a flowList representing the "EH predecessors" of "blk". These are the normal predecessors of
1798 // "blk", plus one special case: if "blk" is the first block of a handler, considers the predecessor(s) of the first
1799 // first block of the corresponding try region to be "EH predecessors". (If there is a single such predecessor,
1800 // for example, we want to consider that the immediate dominator of the catch clause start block, so it's
1801 // convenient to also consider it a predecessor.)
1802 flowList* BlockPredsWithEH(BasicBlock* blk);
1804 // This table is useful for memoization of the method above.
1805 typedef SimplerHashTable<BasicBlock*, PtrKeyFuncs<BasicBlock>, flowList*, JitSimplerHashBehavior>
1807 BlockToFlowListMap* m_blockToEHPreds;
1808 BlockToFlowListMap* GetBlockToEHPreds()
1810 if (m_blockToEHPreds == nullptr)
1812 m_blockToEHPreds = new (getAllocator()) BlockToFlowListMap(getAllocator());
1814 return m_blockToEHPreds;
1817 void* ehEmitCookie(BasicBlock* block);
1818 UNATIVE_OFFSET ehCodeOffset(BasicBlock* block);
1820 EHblkDsc* ehInitHndRange(BasicBlock* src, IL_OFFSET* hndBeg, IL_OFFSET* hndEnd, bool* inFilter);
1822 EHblkDsc* ehInitTryRange(BasicBlock* src, IL_OFFSET* tryBeg, IL_OFFSET* tryEnd);
1824 EHblkDsc* ehInitHndBlockRange(BasicBlock* blk, BasicBlock** hndBeg, BasicBlock** hndLast, bool* inFilter);
1826 EHblkDsc* ehInitTryBlockRange(BasicBlock* blk, BasicBlock** tryBeg, BasicBlock** tryLast);
1828 void fgSetTryEnd(EHblkDsc* handlerTab, BasicBlock* newTryLast);
1830 void fgSetHndEnd(EHblkDsc* handlerTab, BasicBlock* newHndLast);
1832 void fgSkipRmvdBlocks(EHblkDsc* handlerTab);
1834 void fgAllocEHTable();
1836 void fgRemoveEHTableEntry(unsigned XTnum);
1838 #if FEATURE_EH_FUNCLETS
1840 EHblkDsc* fgAddEHTableEntry(unsigned XTnum);
1842 #endif // FEATURE_EH_FUNCLETS
1846 #endif // !FEATURE_EH
1848 void fgSortEHTable();
1850 // Causes the EH table to obey some well-formedness conditions, by inserting
1851 // empty BB's when necessary:
1852 // * No block is both the first block of a handler and the first block of a try.
1853 // * No block is the first block of multiple 'try' regions.
1854 // * No block is the last block of multiple EH regions.
1855 void fgNormalizeEH();
1856 bool fgNormalizeEHCase1();
1857 bool fgNormalizeEHCase2();
1858 bool fgNormalizeEHCase3();
1861 void dispIncomingEHClause(unsigned num, const CORINFO_EH_CLAUSE& clause);
1862 void dispOutgoingEHClause(unsigned num, const CORINFO_EH_CLAUSE& clause);
1863 void fgVerifyHandlerTab();
1864 void fgDispHandlerTab();
1867 bool fgNeedToSortEHTable;
1869 void verInitEHTree(unsigned numEHClauses);
1870 void verInsertEhNode(CORINFO_EH_CLAUSE* clause, EHblkDsc* handlerTab);
1871 void verInsertEhNodeInTree(EHNodeDsc** ppRoot, EHNodeDsc* node);
1872 void verInsertEhNodeParent(EHNodeDsc** ppRoot, EHNodeDsc* node);
1873 void verCheckNestingLevel(EHNodeDsc* initRoot);
1876 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1877 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1879 XX GenTree and BasicBlock XX
1881 XX Functions to allocate and display the GenTrees and BasicBlocks XX
1883 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1884 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1887 // Functions to create nodes
1888 GenTreeStmt* gtNewStmt(GenTreePtr expr = nullptr, IL_OFFSETX offset = BAD_IL_OFFSET);
1891 GenTreePtr gtNewOperNode(genTreeOps oper, var_types type, GenTreePtr op1, bool doSimplifications = TRUE);
1893 // For binary opers.
1894 GenTreePtr gtNewOperNode(genTreeOps oper, var_types type, GenTreePtr op1, GenTreePtr op2);
1896 GenTreePtr gtNewQmarkNode(var_types type, GenTreePtr cond, GenTreePtr colon);
1898 GenTreePtr gtNewLargeOperNode(genTreeOps oper,
1899 var_types type = TYP_I_IMPL,
1900 GenTreePtr op1 = nullptr,
1901 GenTreePtr op2 = nullptr);
1903 GenTreeIntCon* gtNewIconNode(ssize_t value, var_types type = TYP_INT);
1905 GenTree* gtNewPhysRegNode(regNumber reg, var_types type);
1907 GenTree* gtNewPhysRegNode(regNumber reg, GenTree* src);
1909 GenTreePtr gtNewJmpTableNode();
1910 GenTreePtr gtNewIconHandleNode(
1911 size_t value, unsigned flags, FieldSeqNode* fields = nullptr, unsigned handle1 = 0, void* handle2 = nullptr);
1913 unsigned gtTokenToIconFlags(unsigned token);
1915 GenTreePtr gtNewIconEmbHndNode(void* value,
1918 unsigned handle1 = 0,
1919 void* handle2 = nullptr,
1920 void* compileTimeHandle = nullptr);
1922 GenTreePtr gtNewIconEmbScpHndNode(CORINFO_MODULE_HANDLE scpHnd, unsigned hnd1 = 0, void* hnd2 = nullptr);
1923 GenTreePtr gtNewIconEmbClsHndNode(CORINFO_CLASS_HANDLE clsHnd, unsigned hnd1 = 0, void* hnd2 = nullptr);
1924 GenTreePtr gtNewIconEmbMethHndNode(CORINFO_METHOD_HANDLE methHnd, unsigned hnd1 = 0, void* hnd2 = nullptr);
1925 GenTreePtr gtNewIconEmbFldHndNode(CORINFO_FIELD_HANDLE fldHnd, unsigned hnd1 = 0, void* hnd2 = nullptr);
1927 GenTreePtr gtNewStringLiteralNode(InfoAccessType iat, void* pValue);
1929 GenTreePtr gtNewLconNode(__int64 value);
1931 GenTreePtr gtNewDconNode(double value);
1933 GenTreePtr gtNewSconNode(int CPX, CORINFO_MODULE_HANDLE scpHandle);
1935 GenTreePtr gtNewZeroConNode(var_types type);
1937 GenTreePtr gtNewOneConNode(var_types type);
1940 GenTreePtr gtNewSIMDVectorZero(var_types simdType, var_types baseType, unsigned size);
1941 GenTreePtr gtNewSIMDVectorOne(var_types simdType, var_types baseType, unsigned size);
1944 GenTreeBlk* gtNewBlkOpNode(
1945 genTreeOps oper, GenTreePtr dst, GenTreePtr srcOrFillVal, GenTreePtr sizeOrClsTok, bool isVolatile);
1947 GenTree* gtNewBlkOpNode(GenTreePtr dst, GenTreePtr srcOrFillVal, unsigned size, bool isVolatile, bool isCopyBlock);
1950 void gtBlockOpInit(GenTreePtr result, GenTreePtr dst, GenTreePtr srcOrFillVal, bool isVolatile);
1953 GenTree* gtNewObjNode(CORINFO_CLASS_HANDLE structHnd, GenTreePtr addr);
1954 void gtSetObjGcInfo(GenTreeObj* objNode);
1955 GenTree* gtNewStructVal(CORINFO_CLASS_HANDLE structHnd, GenTreePtr addr);
1956 GenTree* gtNewBlockVal(GenTreePtr addr, unsigned size);
1958 GenTree* gtNewCpObjNode(GenTreePtr dst, GenTreePtr src, CORINFO_CLASS_HANDLE structHnd, bool isVolatile);
1960 GenTreeArgList* gtNewListNode(GenTreePtr op1, GenTreeArgList* op2);
1962 GenTreeCall* gtNewCallNode(gtCallTypes callType,
1963 CORINFO_METHOD_HANDLE handle,
1965 GenTreeArgList* args,
1966 IL_OFFSETX ilOffset = BAD_IL_OFFSET);
1968 GenTreeCall* gtNewIndCallNode(GenTreePtr addr,
1970 GenTreeArgList* args,
1971 IL_OFFSETX ilOffset = BAD_IL_OFFSET);
1973 GenTreeCall* gtNewHelperCallNode(unsigned helper,
1976 GenTreeArgList* args = nullptr);
1978 GenTreePtr gtNewLclvNode(unsigned lnum, var_types type, IL_OFFSETX ILoffs = BAD_IL_OFFSET);
1981 GenTreeSIMD* gtNewSIMDNode(
1982 var_types type, GenTreePtr op1, SIMDIntrinsicID simdIntrinsicID, var_types baseType, unsigned size);
1983 GenTreeSIMD* gtNewSIMDNode(var_types type,
1986 SIMDIntrinsicID simdIntrinsicID,
1991 GenTreePtr gtNewLclLNode(unsigned lnum, var_types type, IL_OFFSETX ILoffs = BAD_IL_OFFSET);
1992 GenTreeLclFld* gtNewLclFldNode(unsigned lnum, var_types type, unsigned offset);
1993 GenTreePtr gtNewInlineCandidateReturnExpr(GenTreePtr inlineCandidate, var_types type);
1995 GenTreePtr gtNewCodeRef(BasicBlock* block);
1997 GenTreePtr gtNewFieldRef(
1998 var_types typ, CORINFO_FIELD_HANDLE fldHnd, GenTreePtr obj = nullptr, DWORD offset = 0, bool nullcheck = false);
2000 GenTreePtr gtNewIndexRef(var_types typ, GenTreePtr arrayOp, GenTreePtr indexOp);
2002 GenTreeArgList* gtNewArgList(GenTreePtr op);
2003 GenTreeArgList* gtNewArgList(GenTreePtr op1, GenTreePtr op2);
2004 GenTreeArgList* gtNewArgList(GenTreePtr op1, GenTreePtr op2, GenTreePtr op3);
2006 static fgArgTabEntryPtr gtArgEntryByArgNum(GenTreePtr call, unsigned argNum);
2007 static fgArgTabEntryPtr gtArgEntryByNode(GenTreePtr call, GenTreePtr node);
2008 fgArgTabEntryPtr gtArgEntryByLateArgIndex(GenTreePtr call, unsigned lateArgInx);
2009 bool gtArgIsThisPtr(fgArgTabEntryPtr argEntry);
2011 GenTreePtr gtNewAssignNode(GenTreePtr dst, GenTreePtr src);
2013 GenTreePtr gtNewTempAssign(unsigned tmp, GenTreePtr val);
2015 GenTreePtr gtNewRefCOMfield(GenTreePtr objPtr,
2016 CORINFO_RESOLVED_TOKEN* pResolvedToken,
2017 CORINFO_ACCESS_FLAGS access,
2018 CORINFO_FIELD_INFO* pFieldInfo,
2020 CORINFO_CLASS_HANDLE structType,
2023 GenTreePtr gtNewNothingNode();
2025 GenTreePtr gtNewArgPlaceHolderNode(var_types type, CORINFO_CLASS_HANDLE clsHnd);
2027 GenTreePtr gtUnusedValNode(GenTreePtr expr);
2029 GenTreePtr gtNewCastNode(var_types typ, GenTreePtr op1, var_types castType);
2031 GenTreePtr gtNewCastNodeL(var_types typ, GenTreePtr op1, var_types castType);
2033 GenTreePtr gtNewAllocObjNode(unsigned int helper, CORINFO_CLASS_HANDLE clsHnd, var_types type, GenTreePtr op1);
2035 //------------------------------------------------------------------------
2036 // Other GenTree functions
2038 GenTreePtr gtClone(GenTree* tree, bool complexOK = false);
2040 // If `tree` is a lclVar with lclNum `varNum`, return an IntCns with value `varVal`; otherwise,
2041 // create a copy of `tree`, adding specified flags, replacing uses of lclVar `deepVarNum` with
2042 // IntCnses with value `deepVarVal`.
2043 GenTreePtr gtCloneExpr(
2044 GenTree* tree, unsigned addFlags, unsigned varNum, int varVal, unsigned deepVarNum, int deepVarVal);
2046 // Create a copy of `tree`, optionally adding specifed flags, and optionally mapping uses of local
2047 // `varNum` to int constants with value `varVal`.
2048 GenTreePtr gtCloneExpr(GenTree* tree, unsigned addFlags = 0, unsigned varNum = (unsigned)-1, int varVal = 0)
2050 return gtCloneExpr(tree, addFlags, varNum, varVal, varNum, varVal);
2053 GenTreePtr gtReplaceTree(GenTreePtr stmt, GenTreePtr tree, GenTreePtr replacementTree);
2055 void gtUpdateSideEffects(GenTreePtr tree, unsigned oldGtFlags, unsigned newGtFlags);
2057 // Returns "true" iff the complexity (not formally defined, but first interpretation
2058 // is #of nodes in subtree) of "tree" is greater than "limit".
2059 // (This is somewhat redundant with the "gtCostEx/gtCostSz" fields, but can be used
2060 // before they have been set.)
2061 bool gtComplexityExceeds(GenTreePtr* tree, unsigned limit);
2063 bool gtCompareTree(GenTree* op1, GenTree* op2);
2065 GenTreePtr gtReverseCond(GenTree* tree);
2067 bool gtHasRef(GenTree* tree, ssize_t lclNum, bool defOnly);
2069 bool gtHasLocalsWithAddrOp(GenTreePtr tree);
2071 unsigned gtHashValue(GenTree* tree);
2073 unsigned gtSetListOrder(GenTree* list, bool regs, bool isListCallArgs);
2075 void gtWalkOp(GenTree** op1, GenTree** op2, GenTree* adr, bool constOnly);
2078 GenTreePtr gtWalkOpEffectiveVal(GenTreePtr op);
2081 void gtPrepareCost(GenTree* tree);
2082 bool gtIsLikelyRegVar(GenTree* tree);
2084 unsigned gtSetEvalOrderAndRestoreFPstkLevel(GenTree* tree);
2086 // Returns true iff the secondNode can be swapped with firstNode.
2087 bool gtCanSwapOrder(GenTree* firstNode, GenTree* secondNode);
2089 unsigned gtSetEvalOrder(GenTree* tree);
2091 #if FEATURE_STACK_FP_X87
2093 void gtComputeFPlvls(GenTreePtr tree);
2094 #endif // FEATURE_STACK_FP_X87
2096 void gtSetStmtInfo(GenTree* stmt);
2098 // Returns "true" iff "node" has any of the side effects in "flags".
2099 bool gtNodeHasSideEffects(GenTreePtr node, unsigned flags);
2101 // Returns "true" iff "tree" or its (transitive) children have any of the side effects in "flags".
2102 bool gtTreeHasSideEffects(GenTreePtr tree, unsigned flags);
2104 // Appends 'expr' in front of 'list'
2105 // 'list' will typically start off as 'nullptr'
2106 // when 'list' is non-null a GT_COMMA node is used to insert 'expr'
2107 GenTreePtr gtBuildCommaList(GenTreePtr list, GenTreePtr expr);
2109 void gtExtractSideEffList(GenTreePtr expr,
2111 unsigned flags = GTF_SIDE_EFFECT,
2112 bool ignoreRoot = false);
2114 GenTreePtr gtGetThisArg(GenTreePtr call);
2116 // Static fields of struct types (and sometimes the types that those are reduced to) are represented by having the
2117 // static field contain an object pointer to the boxed struct. This simplifies the GC implementation...but
2118 // complicates the JIT somewhat. This predicate returns "true" iff a node with type "fieldNodeType", representing
2119 // the given "fldHnd", is such an object pointer.
2120 bool gtIsStaticFieldPtrToBoxedStruct(var_types fieldNodeType, CORINFO_FIELD_HANDLE fldHnd);
2122 // Return true if call is a recursive call; return false otherwise.
2123 bool gtIsRecursiveCall(GenTreeCall* call)
2125 return (call->gtCallMethHnd == info.compMethodHnd);
2128 //-------------------------------------------------------------------------
2130 GenTreePtr gtFoldExpr(GenTreePtr tree);
2133 // TODO-Amd64-Unix: Remove this when the clang optimizer is fixed and/or the method implementation is
2134 // refactored in a simpler code. This is a workaround for a bug in the clang-3.5 optimizer. The issue is that in
2135 // release build the optimizer is mistyping (or just wrongly decides to use 32 bit operation for a corner case
2136 // of MIN_LONG) the args of the (ltemp / lval2) to int (it does a 32 bit div operation instead of 64 bit) - see
2137 // the implementation of the method in gentree.cpp. For the case of lval1 and lval2 equal to MIN_LONG
2138 // (0x8000000000000000) this results in raising a SIGFPE. The method implementation is rather complex. Disable
2139 // optimizations for now.
2140 __attribute__((optnone))
2142 gtFoldExprConst(GenTreePtr tree);
2143 GenTreePtr gtFoldExprSpecial(GenTreePtr tree);
2144 GenTreePtr gtFoldExprCompare(GenTreePtr tree);
2146 //-------------------------------------------------------------------------
2147 // Get the handle, if any.
2148 CORINFO_CLASS_HANDLE gtGetStructHandleIfPresent(GenTreePtr tree);
2149 // Get the handle, and assert if not found.
2150 CORINFO_CLASS_HANDLE gtGetStructHandle(GenTreePtr tree);
2152 //-------------------------------------------------------------------------
2153 // Functions to display the trees
2156 void gtDispNode(GenTreePtr tree, IndentStack* indentStack, __in_z const char* msg, bool isLIR);
2158 void gtDispVN(GenTreePtr tree);
2159 void gtDispConst(GenTreePtr tree);
2160 void gtDispLeaf(GenTreePtr tree, IndentStack* indentStack);
2161 void gtDispNodeName(GenTreePtr tree);
2162 void gtDispRegVal(GenTreePtr tree);
2174 void gtDispChild(GenTreePtr child,
2175 IndentStack* indentStack,
2177 __in_opt const char* msg = nullptr,
2178 bool topOnly = false);
2179 void gtDispTree(GenTreePtr tree,
2180 IndentStack* indentStack = nullptr,
2181 __in_opt const char* msg = nullptr,
2182 bool topOnly = false,
2183 bool isLIR = false);
2184 void gtGetLclVarNameInfo(unsigned lclNum, const char** ilKindOut, const char** ilNameOut, unsigned* ilNumOut);
2185 int gtGetLclVarName(unsigned lclNum, char* buf, unsigned buf_remaining);
2186 char* gtGetLclVarName(unsigned lclNum);
2187 void gtDispLclVar(unsigned varNum, bool padForBiggestDisp = true);
2188 void gtDispTreeList(GenTreePtr tree, IndentStack* indentStack = nullptr);
2189 void gtGetArgMsg(GenTreePtr call, GenTreePtr arg, unsigned argNum, int listCount, char* bufp, unsigned bufLength);
2190 void gtGetLateArgMsg(GenTreePtr call, GenTreePtr arg, int argNum, int listCount, char* bufp, unsigned bufLength);
2191 void gtDispArgList(GenTreePtr tree, IndentStack* indentStack);
2192 void gtDispFieldSeq(FieldSeqNode* pfsn);
2194 void gtDispRange(LIR::ReadOnlyRange const& range);
2196 void gtDispTreeRange(LIR::Range& containingRange, GenTree* tree);
2198 void gtDispLIRNode(GenTree* node);
2210 typedef fgWalkResult(fgWalkPreFn)(GenTreePtr* pTree, fgWalkData* data);
2211 typedef fgWalkResult(fgWalkPostFn)(GenTreePtr* pTree, fgWalkData* data);
2214 static fgWalkPreFn gtAssertColonCond;
2216 static fgWalkPreFn gtMarkColonCond;
2217 static fgWalkPreFn gtClearColonCond;
2219 GenTreePtr* gtFindLink(GenTreePtr stmt, GenTreePtr node);
2220 bool gtHasCatchArg(GenTreePtr tree);
2221 bool gtHasUnmanagedCall(GenTreePtr tree);
2223 typedef ArrayStack<GenTree*> GenTreeStack;
2225 static bool gtHasCallOnStack(GenTreeStack* parentStack);
2226 void gtCheckQuirkAddrExposedLclVar(GenTreePtr argTree, GenTreeStack* parentStack);
2228 //=========================================================================
2229 // BasicBlock functions
2231 // This is a debug flag we will use to assert when creating block during codegen
2232 // as this interferes with procedure splitting. If you know what you're doing, set
2233 // it to true before creating the block. (DEBUG only)
2234 bool fgSafeBasicBlockCreation;
2237 BasicBlock* bbNewBasicBlock(BBjumpKinds jumpKind);
2240 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
2241 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
2245 XX The variables to be used by the code generator. XX
2247 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
2248 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
2252 // For both PROMOTION_TYPE_NONE and PROMOTION_TYPE_DEPENDENT the struct will
2253 // be placed in the stack frame and it's fields must be laid out sequentially.
2255 // For PROMOTION_TYPE_INDEPENDENT each of the struct's fields is replaced by
2256 // a local variable that can be enregistered or placed in the stack frame.
2257 // The fields do not need to be laid out sequentially
2259 enum lvaPromotionType
2261 PROMOTION_TYPE_NONE, // The struct local is not promoted
2262 PROMOTION_TYPE_INDEPENDENT, // The struct local is promoted,
2263 // and its field locals are independent of its parent struct local.
2264 PROMOTION_TYPE_DEPENDENT // The struct local is promoted,
2265 // but its field locals depend on its parent struct local.
2268 static int __cdecl RefCntCmp(const void* op1, const void* op2);
2269 static int __cdecl WtdRefCntCmp(const void* op1, const void* op2);
2271 /*****************************************************************************/
2273 enum FrameLayoutState
2276 INITIAL_FRAME_LAYOUT,
2277 PRE_REGALLOC_FRAME_LAYOUT,
2278 REGALLOC_FRAME_LAYOUT,
2279 TENTATIVE_FRAME_LAYOUT,
2284 bool lvaRefCountingStarted; // Set to true when we have started counting the local vars
2285 bool lvaLocalVarRefCounted; // Set to true after we have called lvaMarkLocalVars()
2286 bool lvaSortAgain; // true: We need to sort the lvaTable
2287 bool lvaTrackedFixed; // true: We cannot add new 'tracked' variable
2288 unsigned lvaCount; // total number of locals
2290 unsigned lvaRefCount; // total number of references to locals
2291 LclVarDsc* lvaTable; // variable descriptor table
2292 unsigned lvaTableCnt; // lvaTable size (>= lvaCount)
2294 LclVarDsc** lvaRefSorted; // table sorted by refcount
2296 unsigned short lvaTrackedCount; // actual # of locals being tracked
2297 unsigned lvaTrackedCountInSizeTUnits; // min # of size_t's sufficient to hold a bit for all the locals being tracked
2299 #ifdef FEATURE_UNIX_AMD64_STRUCT_PASSING
2300 // Only for AMD64 System V cache the first caller stack homed argument.
2301 unsigned lvaFirstStackIncomingArgNum; // First argument with stack slot in the caller.
2302 #endif // !FEATURE_UNIX_AMD64_STRUCT_PASSING
2305 VARSET_TP lvaTrackedVars; // set of tracked variables
2307 #ifndef _TARGET_64BIT_
2308 VARSET_TP lvaLongVars; // set of long (64-bit) variables
2310 VARSET_TP lvaFloatVars; // set of floating-point (32-bit and 64-bit) variables
2312 unsigned lvaCurEpoch; // VarSets are relative to a specific set of tracked var indices.
2313 // It that changes, this changes. VarSets from different epochs
2314 // cannot be meaningfully combined.
2316 unsigned GetCurLVEpoch()
2321 // reverse map of tracked number to var number
2322 unsigned lvaTrackedToVarNum[lclMAX_TRACKED];
2324 #ifdef LEGACY_BACKEND
2325 // variable interference graph
2326 VARSET_TP lvaVarIntf[lclMAX_TRACKED];
2329 // variable preference graph
2330 VARSET_TP lvaVarPref[lclMAX_TRACKED];
2334 // # of procs compiled a with double-aligned stack
2335 static unsigned s_lvaDoubleAlignedProcsCount;
2339 // Getters and setters for address-exposed and do-not-enregister local var properties.
2340 bool lvaVarAddrExposed(unsigned varNum);
2341 void lvaSetVarAddrExposed(unsigned varNum);
2342 bool lvaVarDoNotEnregister(unsigned varNum);
2344 // Reasons why we can't enregister. Some of these correspond to debug properties of local vars.
2345 enum DoNotEnregisterReason
2350 DNER_VMNeedsStackAddr,
2351 DNER_LiveInOutOfHandler,
2352 DNER_LiveAcrossUnmanagedCall,
2353 DNER_BlockOp, // Is read or written via a block operation that explicitly takes the address.
2354 DNER_IsStructArg, // Is a struct passed as an argument in a way that requires a stack location.
2355 #ifdef JIT32_GCENCODER
2360 void lvaSetVarDoNotEnregister(unsigned varNum DEBUGARG(DoNotEnregisterReason reason));
2362 unsigned lvaVarargsHandleArg;
2364 unsigned lvaVarargsBaseOfStkArgs; // Pointer (computed based on incoming varargs handle) to the start of the stack
2366 #endif // _TARGET_X86_
2368 unsigned lvaInlinedPInvokeFrameVar; // variable representing the InlinedCallFrame
2369 unsigned lvaReversePInvokeFrameVar; // variable representing the reverse PInvoke frame
2370 #if FEATURE_FIXED_OUT_ARGS
2371 unsigned lvaPInvokeFrameRegSaveVar; // variable representing the RegSave for PInvoke inlining.
2373 unsigned lvaMonAcquired; // boolean variable introduced into in synchronized methods
2374 // that tracks whether the lock has been taken
2376 unsigned lvaArg0Var; // The lclNum of arg0. Normally this will be info.compThisArg.
2377 // However, if there is a "ldarga 0" or "starg 0" in the IL,
2378 // we will redirect all "ldarg(a) 0" and "starg 0" to this temp.
2380 unsigned lvaInlineeReturnSpillTemp; // The temp to spill the non-VOID return expression
2381 // in case there are multiple BBJ_RETURN blocks in the inlinee.
2383 #if FEATURE_FIXED_OUT_ARGS
2384 unsigned lvaOutgoingArgSpaceVar; // dummy TYP_LCLBLK var for fixed outgoing argument space
2385 unsigned lvaOutgoingArgSpaceSize; // size of fixed outgoing argument space
2386 #endif // FEATURE_FIXED_OUT_ARGS
2389 // On architectures whose ABIs allow structs to be passed in registers, struct promotion will sometimes
2390 // require us to "rematerialize" a struct from it's separate constituent field variables. Packing several sub-word
2391 // field variables into an argument register is a hard problem. It's easier to reserve a word of memory into which
2392 // such field can be copied, after which the assembled memory word can be read into the register. We will allocate
2393 // this variable to be this scratch word whenever struct promotion occurs.
2394 unsigned lvaPromotedStructAssemblyScratchVar;
2395 #endif // _TARGET_ARM_
2398 unsigned lvaReturnEspCheck; // confirms ESP not corrupted on return
2399 unsigned lvaCallEspCheck; // confirms ESP not corrupted after a call
2402 bool lvaGenericsContextUsed;
2404 bool lvaKeepAliveAndReportThis(); // Synchronized instance method of a reference type, or
2405 // CORINFO_GENERICS_CTXT_FROM_THIS?
2406 bool lvaReportParamTypeArg(); // Exceptions and CORINFO_GENERICS_CTXT_FROM_PARAMTYPEARG?
2408 //-------------------------------------------------------------------------
2409 // All these frame offsets are inter-related and must be kept in sync
2411 #if !FEATURE_EH_FUNCLETS
2412 // This is used for the callable handlers
2413 unsigned lvaShadowSPslotsVar; // TYP_BLK variable for all the shadow SP slots
2414 #endif // FEATURE_EH_FUNCLETS
2416 unsigned lvaCachedGenericContextArgOffs;
2417 unsigned lvaCachedGenericContextArgOffset(); // For CORINFO_CALLCONV_PARAMTYPE and if generic context is passed as
2420 unsigned lvaLocAllocSPvar; // variable which has the result of the last alloca/localloc
2422 unsigned lvaNewObjArrayArgs; // variable with arguments for new MD array helper
2424 // TODO-Review: Prior to reg predict we reserve 24 bytes for Spill temps.
2425 // after the reg predict we will use a computed maxTmpSize
2426 // which is based upon the number of spill temps predicted by reg predict
2427 // All this is necessary because if we under-estimate the size of the spill
2428 // temps we could fail when encoding instructions that reference stack offsets for ARM.
2430 // Pre codegen max spill temp size.
2431 static const unsigned MAX_SPILL_TEMP_SIZE = 24;
2433 //-------------------------------------------------------------------------
2435 unsigned lvaGetMaxSpillTempSize();
2437 bool lvaIsPreSpilled(unsigned lclNum, regMaskTP preSpillMask);
2438 #endif // _TARGET_ARM_
2439 void lvaAssignFrameOffsets(FrameLayoutState curState);
2440 void lvaFixVirtualFrameOffsets();
2442 #ifndef LEGACY_BACKEND
2443 void lvaUpdateArgsWithInitialReg();
2444 #endif // !LEGACY_BACKEND
2446 void lvaAssignVirtualFrameOffsetsToArgs();
2447 #ifdef UNIX_AMD64_ABI
2448 int lvaAssignVirtualFrameOffsetToArg(unsigned lclNum, unsigned argSize, int argOffs, int* callerArgOffset);
2449 #else // !UNIX_AMD64_ABI
2450 int lvaAssignVirtualFrameOffsetToArg(unsigned lclNum, unsigned argSize, int argOffs);
2451 #endif // !UNIX_AMD64_ABI
2452 void lvaAssignVirtualFrameOffsetsToLocals();
2453 int lvaAllocLocalAndSetVirtualOffset(unsigned lclNum, unsigned size, int stkOffs);
2454 #ifdef _TARGET_AMD64_
2455 // Returns true if compCalleeRegsPushed (including RBP if used as frame pointer) is even.
2456 bool lvaIsCalleeSavedIntRegCountEven();
2458 void lvaAlignFrame();
2459 void lvaAssignFrameOffsetsToPromotedStructs();
2460 int lvaAllocateTemps(int stkOffs, bool mustDoubleAlign);
2463 void lvaDumpRegLocation(unsigned lclNum);
2464 void lvaDumpFrameLocation(unsigned lclNum);
2465 void lvaDumpEntry(unsigned lclNum, FrameLayoutState curState, size_t refCntWtdWidth = 6);
2466 void lvaTableDump(FrameLayoutState curState = NO_FRAME_LAYOUT); // NO_FRAME_LAYOUT means use the current frame
2467 // layout state defined by lvaDoneFrameLayout
2470 // Limit frames size to 1GB. The maximum is 2GB in theory - make it intentionally smaller
2471 // to avoid bugs from borderline cases.
2472 #define MAX_FrameSize 0x3FFFFFFF
2473 void lvaIncrementFrameSize(unsigned size);
2475 unsigned lvaFrameSize(FrameLayoutState curState);
2477 // Returns the caller-SP-relative offset for the SP/FP relative offset determined by FP based.
2478 int lvaToCallerSPRelativeOffset(int offs, bool isFpBased);
2480 // Returns the caller-SP-relative offset for the local variable "varNum."
2481 int lvaGetCallerSPRelativeOffset(unsigned varNum);
2483 // Returns the SP-relative offset for the local variable "varNum". Illegal to ask this for functions with localloc.
2484 int lvaGetSPRelativeOffset(unsigned varNum);
2486 int lvaToInitialSPRelativeOffset(unsigned offset, bool isFpBased);
2487 int lvaGetInitialSPRelativeOffset(unsigned varNum);
2489 //------------------------ For splitting types ----------------------------
2491 void lvaInitTypeRef();
2493 void lvaInitArgs(InitVarDscInfo* varDscInfo);
2494 void lvaInitThisPtr(InitVarDscInfo* varDscInfo);
2495 void lvaInitRetBuffArg(InitVarDscInfo* varDscInfo);
2496 void lvaInitUserArgs(InitVarDscInfo* varDscInfo);
2497 void lvaInitGenericsCtxt(InitVarDscInfo* varDscInfo);
2498 void lvaInitVarArgsHandle(InitVarDscInfo* varDscInfo);
2500 void lvaInitVarDsc(LclVarDsc* varDsc,
2502 CorInfoType corInfoType,
2503 CORINFO_CLASS_HANDLE typeHnd,
2504 CORINFO_ARG_LIST_HANDLE varList,
2505 CORINFO_SIG_INFO* varSig);
2507 static unsigned lvaTypeRefMask(var_types type);
2509 var_types lvaGetActualType(unsigned lclNum);
2510 var_types lvaGetRealType(unsigned lclNum);
2512 //-------------------------------------------------------------------------
2516 unsigned lvaLclSize(unsigned varNum);
2517 unsigned lvaLclExactSize(unsigned varNum);
2519 bool lvaLclVarRefs(GenTreePtr tree, GenTreePtr* findPtr, varRefKinds* refsPtr, void* result);
2521 // Call lvaLclVarRefs on "true"; accumulate "*result" into whichever of
2522 // "allVars" and "trkdVars" is indiated by the nullness of "findPtr"; return
2523 // the return result.
2524 bool lvaLclVarRefsAccum(
2525 GenTreePtr tree, GenTreePtr* findPtr, varRefKinds* refsPtr, ALLVARSET_TP* allVars, VARSET_TP* trkdVars);
2527 // If "findPtr" is non-NULL, assumes "result" is an "ALLVARSET_TP*", and
2528 // (destructively) unions "allVars" into "*result". Otherwise, assumes "result" is a "VARSET_TP*",
2529 // and (destructively) unions "trkedVars" into "*result".
2530 void lvaLclVarRefsAccumIntoRes(GenTreePtr* findPtr,
2532 ALLVARSET_VALARG_TP allVars,
2533 VARSET_VALARG_TP trkdVars);
2535 bool lvaHaveManyLocals() const;
2537 unsigned lvaGrabTemp(bool shortLifetime DEBUGARG(const char* reason));
2538 unsigned lvaGrabTemps(unsigned cnt DEBUGARG(const char* reason));
2539 unsigned lvaGrabTempWithImplicitUse(bool shortLifetime DEBUGARG(const char* reason));
2542 void lvaSortByRefCount();
2543 void lvaDumpRefCounts();
2545 void lvaMarkLocalVars(BasicBlock* block);
2547 void lvaMarkLocalVars(); // Local variable ref-counting
2549 void lvaAllocOutgoingArgSpace(); // 'Commit' lvaOutgoingArgSpaceSize and lvaOutgoingArgSpaceVar
2551 VARSET_VALRET_TP lvaStmtLclMask(GenTreePtr stmt);
2553 static fgWalkPreFn lvaIncRefCntsCB;
2554 void lvaIncRefCnts(GenTreePtr tree);
2556 static fgWalkPreFn lvaDecRefCntsCB;
2557 void lvaDecRefCnts(GenTreePtr tree);
2558 void lvaDecRefCnts(BasicBlock* basicBlock, GenTreePtr tree);
2559 void lvaRecursiveDecRefCounts(GenTreePtr tree);
2560 void lvaRecursiveIncRefCounts(GenTreePtr tree);
2563 struct lvaStressLclFldArgs
2565 Compiler* m_pCompiler;
2569 static fgWalkPreFn lvaStressLclFldCB;
2570 void lvaStressLclFld();
2572 void lvaDispVarSet(VARSET_VALARG_TP set, VARSET_VALARG_TP allVars);
2573 void lvaDispVarSet(VARSET_VALARG_TP set);
2578 int lvaFrameAddress(int varNum, bool mustBeFPBased, regNumber* pBaseReg, int addrModeOffset);
2580 int lvaFrameAddress(int varNum, bool* pFPbased);
2583 bool lvaIsParameter(unsigned varNum);
2584 bool lvaIsRegArgument(unsigned varNum);
2585 BOOL lvaIsOriginalThisArg(unsigned varNum); // Is this varNum the original this argument?
2586 BOOL lvaIsOriginalThisReadOnly(); // return TRUE if there is no place in the code
2587 // that writes to arg0
2589 // Struct parameters that are passed by reference are marked as both lvIsParam and lvIsTemp
2590 // (this is an overload of lvIsTemp because there are no temp parameters).
2591 // For x64 this is 3, 5, 6, 7, >8 byte structs that are passed by reference.
2592 // For ARM64, this is structs larger than 16 bytes that are passed by reference.
2593 bool lvaIsImplicitByRefLocal(unsigned varNum)
2595 #if defined(_TARGET_AMD64_) || defined(_TARGET_ARM64_)
2596 LclVarDsc* varDsc = &(lvaTable[varNum]);
2597 if (varDsc->lvIsParam && varDsc->lvIsTemp)
2599 assert((varDsc->lvType == TYP_STRUCT) || (varDsc->lvType == TYP_BYREF));
2602 #endif // defined(_TARGET_AMD64_) || defined(_TARGET_ARM64_)
2606 // Returns true if this local var is a multireg struct
2607 bool lvaIsMultiregStruct(LclVarDsc* varDsc);
2609 // If the class is a TYP_STRUCT, get/set a class handle describing it
2611 CORINFO_CLASS_HANDLE lvaGetStruct(unsigned varNum);
2612 void lvaSetStruct(unsigned varNum, CORINFO_CLASS_HANDLE typeHnd, bool unsafeValueClsCheck, bool setTypeInfo = true);
2614 #define MAX_NumOfFieldsInPromotableStruct 4 // Maximum number of fields in promotable struct
2616 // Info about struct fields
2617 struct lvaStructFieldInfo
2619 CORINFO_FIELD_HANDLE fldHnd;
2620 unsigned char fldOffset;
2621 unsigned char fldOrdinal;
2624 CORINFO_CLASS_HANDLE fldTypeHnd;
2627 // Info about struct to be promoted.
2628 struct lvaStructPromotionInfo
2630 CORINFO_CLASS_HANDLE typeHnd;
2632 bool requiresScratchVar;
2635 unsigned char fieldCnt;
2636 lvaStructFieldInfo fields[MAX_NumOfFieldsInPromotableStruct];
2638 lvaStructPromotionInfo()
2639 : typeHnd(nullptr), canPromote(false), requiresScratchVar(false), containsHoles(false), customLayout(false)
2644 static int __cdecl lvaFieldOffsetCmp(const void* field1, const void* field2);
2645 void lvaCanPromoteStructType(CORINFO_CLASS_HANDLE typeHnd,
2646 lvaStructPromotionInfo* StructPromotionInfo,
2648 void lvaCanPromoteStructVar(unsigned lclNum, lvaStructPromotionInfo* StructPromotionInfo);
2649 void lvaPromoteStructVar(unsigned lclNum, lvaStructPromotionInfo* StructPromotionInfo);
2650 #if !defined(_TARGET_64BIT_)
2651 void lvaPromoteLongVars();
2652 #endif // !defined(_TARGET_64BIT_)
2653 unsigned lvaGetFieldLocal(LclVarDsc* varDsc, unsigned int fldOffset);
2654 lvaPromotionType lvaGetPromotionType(const LclVarDsc* varDsc);
2655 lvaPromotionType lvaGetPromotionType(unsigned varNum);
2656 lvaPromotionType lvaGetParentPromotionType(const LclVarDsc* varDsc);
2657 lvaPromotionType lvaGetParentPromotionType(unsigned varNum);
2658 bool lvaIsFieldOfDependentlyPromotedStruct(const LclVarDsc* varDsc);
2659 bool lvaIsGCTracked(const LclVarDsc* varDsc);
2661 BYTE* lvaGetGcLayout(unsigned varNum);
2662 bool lvaTypeIsGC(unsigned varNum);
2663 unsigned lvaGSSecurityCookie; // LclVar number
2664 bool lvaTempsHaveLargerOffsetThanVars();
2666 unsigned lvaSecurityObject; // variable representing the security object on the stack
2667 unsigned lvaStubArgumentVar; // variable representing the secret stub argument coming in EAX
2669 #if FEATURE_EH_FUNCLETS
2670 unsigned lvaPSPSym; // variable representing the PSPSym
2673 InlineInfo* impInlineInfo;
2674 InlineStrategy* m_inlineStrategy;
2676 // The Compiler* that is the root of the inlining tree of which "this" is a member.
2677 Compiler* impInlineRoot();
2679 #if defined(DEBUG) || defined(INLINE_DATA)
2680 unsigned __int64 getInlineCycleCount()
2682 return m_compCycles;
2684 #endif // defined(DEBUG) || defined(INLINE_DATA)
2686 bool fgNoStructPromotion; // Set to TRUE to turn off struct promotion for this method.
2687 bool fgNoStructParamPromotion; // Set to TRUE to turn off struct promotion for parameters this method.
2689 //=========================================================================
2691 //=========================================================================
2694 //---------------- Local variable ref-counting ----------------------------
2697 BasicBlock* lvaMarkRefsCurBlock;
2698 GenTreePtr lvaMarkRefsCurStmt;
2700 BasicBlock::weight_t lvaMarkRefsWeight;
2702 static fgWalkPreFn lvaMarkLclRefsCallback;
2703 void lvaMarkLclRefs(GenTreePtr tree);
2705 // Keeps the mapping from SSA #'s to VN's for the implicit "Heap" variable.
2706 PerSsaArray lvHeapPerSsaData;
2707 unsigned lvHeapNumSsaNames;
2710 // Returns the address of the per-Ssa data for "Heap" at the given ssaNum (which is required
2711 // not to be the SsaConfig::RESERVED_SSA_NUM, which indicates that the variable is
2712 // not an SSA variable).
2713 LclSsaVarDsc* GetHeapPerSsaData(unsigned ssaNum)
2715 assert(ssaNum != SsaConfig::RESERVED_SSA_NUM);
2716 assert(SsaConfig::RESERVED_SSA_NUM == 0);
2718 assert(ssaNum < lvHeapNumSsaNames);
2719 return &lvHeapPerSsaData.GetRef(ssaNum);
2723 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
2724 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
2728 XX Imports the given method and converts it to semantic trees XX
2730 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
2731 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
2737 void impImport(BasicBlock* method);
2739 CORINFO_CLASS_HANDLE impGetRefAnyClass();
2740 CORINFO_CLASS_HANDLE impGetRuntimeArgumentHandle();
2741 CORINFO_CLASS_HANDLE impGetTypeHandleClass();
2742 CORINFO_CLASS_HANDLE impGetStringClass();
2743 CORINFO_CLASS_HANDLE impGetObjectClass();
2745 //=========================================================================
2747 //=========================================================================
2750 //-------------------- Stack manipulation ---------------------------------
2752 unsigned impStkSize; // Size of the full stack
2754 #define SMALL_STACK_SIZE 16 // number of elements in impSmallStack
2756 StackEntry impSmallStack[SMALL_STACK_SIZE]; // Use this array if possible
2758 struct SavedStack // used to save/restore stack contents.
2760 unsigned ssDepth; // number of values on stack
2761 StackEntry* ssTrees; // saved tree values
2764 bool impIsPrimitive(CorInfoType type);
2765 bool impILConsumesAddr(const BYTE* codeAddr, CORINFO_METHOD_HANDLE fncHandle, CORINFO_MODULE_HANDLE scpHandle);
2767 void impResolveToken(const BYTE* addr, CORINFO_RESOLVED_TOKEN* pResolvedToken, CorInfoTokenKind kind);
2768 void impPushOnStackNoType(GenTreePtr tree);
2770 void impPushOnStack(GenTreePtr tree, typeInfo ti);
2771 void impPushNullObjRefOnStack();
2772 StackEntry impPopStack();
2773 StackEntry impPopStack(CORINFO_CLASS_HANDLE& structTypeRet);
2774 GenTreePtr impPopStack(typeInfo& ti);
2775 StackEntry& impStackTop(unsigned n = 0);
2777 void impSaveStackState(SavedStack* savePtr, bool copy);
2778 void impRestoreStackState(SavedStack* savePtr);
2780 GenTreePtr impImportLdvirtftn(GenTreePtr thisPtr,
2781 CORINFO_RESOLVED_TOKEN* pResolvedToken,
2782 CORINFO_CALL_INFO* pCallInfo);
2784 void impImportAndPushBox(CORINFO_RESOLVED_TOKEN* pResolvedToken);
2786 void impImportNewObjArray(CORINFO_RESOLVED_TOKEN* pResolvedToken, CORINFO_CALL_INFO* pCallInfo);
2788 bool impCanPInvokeInline();
2789 bool impCanPInvokeInlineCallSite(BasicBlock* block);
2790 void impCheckForPInvokeCall(
2791 GenTreePtr call, CORINFO_METHOD_HANDLE methHnd, CORINFO_SIG_INFO* sig, unsigned mflags, BasicBlock* block);
2792 GenTreePtr impImportIndirectCall(CORINFO_SIG_INFO* sig, IL_OFFSETX ilOffset = BAD_IL_OFFSET);
2793 void impPopArgsForUnmanagedCall(GenTreePtr call, CORINFO_SIG_INFO* sig);
2795 void impInsertHelperCall(CORINFO_HELPER_DESC* helperCall);
2796 void impHandleAccessAllowed(CorInfoIsAccessAllowedResult result, CORINFO_HELPER_DESC* helperCall);
2797 void impHandleAccessAllowedInternal(CorInfoIsAccessAllowedResult result, CORINFO_HELPER_DESC* helperCall);
2799 void impInsertCalloutForDelegate(CORINFO_METHOD_HANDLE callerMethodHnd,
2800 CORINFO_METHOD_HANDLE calleeMethodHnd,
2801 CORINFO_CLASS_HANDLE delegateTypeHnd);
2803 var_types impImportCall(OPCODE opcode,
2804 CORINFO_RESOLVED_TOKEN* pResolvedToken,
2805 CORINFO_RESOLVED_TOKEN* pConstrainedResolvedToken, // Is this a "constrained." call on a
2807 GenTreePtr newobjThis,
2809 CORINFO_CALL_INFO* callInfo,
2810 IL_OFFSET rawILOffset);
2812 bool impMethodInfo_hasRetBuffArg(CORINFO_METHOD_INFO* methInfo);
2814 GenTreePtr impFixupCallStructReturn(GenTreePtr call, CORINFO_CLASS_HANDLE retClsHnd);
2816 GenTreePtr impFixupStructReturnType(GenTreePtr op, CORINFO_CLASS_HANDLE retClsHnd);
2819 var_types impImportJitTestLabelMark(int numArgs);
2822 GenTreePtr impInitClass(CORINFO_RESOLVED_TOKEN* pResolvedToken);
2824 GenTreePtr impImportStaticReadOnlyField(void* fldAddr, var_types lclTyp);
2826 GenTreePtr impImportStaticFieldAccess(CORINFO_RESOLVED_TOKEN* pResolvedToken,
2827 CORINFO_ACCESS_FLAGS access,
2828 CORINFO_FIELD_INFO* pFieldInfo,
2831 static void impBashVarAddrsToI(GenTreePtr tree1, GenTreePtr tree2 = nullptr);
2833 GenTreePtr impImplicitIorI4Cast(GenTreePtr tree, var_types dstTyp);
2835 GenTreePtr impImplicitR4orR8Cast(GenTreePtr tree, var_types dstTyp);
2837 void impImportLeave(BasicBlock* block);
2838 void impResetLeaveBlock(BasicBlock* block, unsigned jmpAddr);
2839 GenTreePtr impIntrinsic(GenTreePtr newobjThis,
2840 CORINFO_CLASS_HANDLE clsHnd,
2841 CORINFO_METHOD_HANDLE method,
2842 CORINFO_SIG_INFO* sig,
2846 CorInfoIntrinsics* pIntrinsicID);
2847 GenTreePtr impArrayAccessIntrinsic(CORINFO_CLASS_HANDLE clsHnd,
2848 CORINFO_SIG_INFO* sig,
2851 CorInfoIntrinsics intrinsicID);
2852 GenTreePtr impInitializeArrayIntrinsic(CORINFO_SIG_INFO* sig);
2854 GenTreePtr impMethodPointer(CORINFO_RESOLVED_TOKEN* pResolvedToken, CORINFO_CALL_INFO* pCallInfo);
2856 GenTreePtr impTransformThis(GenTreePtr thisPtr,
2857 CORINFO_RESOLVED_TOKEN* pConstrainedResolvedToken,
2858 CORINFO_THIS_TRANSFORM transform);
2860 //----------------- Manipulating the trees and stmts ----------------------
2862 GenTreePtr impTreeList; // Trees for the BB being imported
2863 GenTreePtr impTreeLast; // The last tree for the current BB
2867 CHECK_SPILL_ALL = -1,
2868 CHECK_SPILL_NONE = -2
2872 void impBeginTreeList();
2873 void impEndTreeList(BasicBlock* block, GenTreePtr firstStmt, GenTreePtr lastStmt);
2874 void impEndTreeList(BasicBlock* block);
2875 void impAppendStmtCheck(GenTreePtr stmt, unsigned chkLevel);
2876 void impAppendStmt(GenTreePtr stmt, unsigned chkLevel);
2877 void impInsertStmtBefore(GenTreePtr stmt, GenTreePtr stmtBefore);
2878 GenTreePtr impAppendTree(GenTreePtr tree, unsigned chkLevel, IL_OFFSETX offset);
2879 void impInsertTreeBefore(GenTreePtr tree, IL_OFFSETX offset, GenTreePtr stmtBefore);
2880 void impAssignTempGen(unsigned tmp,
2883 GenTreePtr* pAfterStmt = nullptr,
2884 IL_OFFSETX ilOffset = BAD_IL_OFFSET,
2885 BasicBlock* block = nullptr);
2886 void impAssignTempGen(unsigned tmpNum,
2888 CORINFO_CLASS_HANDLE structHnd,
2890 GenTreePtr* pAfterStmt = nullptr,
2891 IL_OFFSETX ilOffset = BAD_IL_OFFSET,
2892 BasicBlock* block = nullptr);
2893 GenTreePtr impCloneExpr(GenTreePtr tree,
2895 CORINFO_CLASS_HANDLE structHnd,
2897 GenTreePtr* pAfterStmt DEBUGARG(const char* reason));
2898 GenTreePtr impAssignStruct(GenTreePtr dest,
2900 CORINFO_CLASS_HANDLE structHnd,
2902 GenTreePtr* pAfterStmt = nullptr,
2903 BasicBlock* block = nullptr);
2904 GenTreePtr impAssignStructPtr(GenTreePtr dest,
2906 CORINFO_CLASS_HANDLE structHnd,
2908 GenTreePtr* pAfterStmt = nullptr,
2909 BasicBlock* block = nullptr);
2911 GenTreePtr impGetStructAddr(GenTreePtr structVal,
2912 CORINFO_CLASS_HANDLE structHnd,
2916 var_types impNormStructType(CORINFO_CLASS_HANDLE structHnd,
2917 BYTE* gcLayout = nullptr,
2918 unsigned* numGCVars = nullptr,
2919 var_types* simdBaseType = nullptr);
2921 GenTreePtr impNormStructVal(GenTreePtr structVal,
2922 CORINFO_CLASS_HANDLE structHnd,
2924 bool forceNormalization = false);
2926 GenTreePtr impTokenToHandle(CORINFO_RESOLVED_TOKEN* pResolvedToken,
2927 BOOL* pRuntimeLookup = nullptr,
2928 BOOL mustRestoreHandle = FALSE,
2929 BOOL importParent = FALSE);
2931 GenTreePtr impParentClassTokenToHandle(CORINFO_RESOLVED_TOKEN* pResolvedToken,
2932 BOOL* pRuntimeLookup = nullptr,
2933 BOOL mustRestoreHandle = FALSE)
2935 return impTokenToHandle(pResolvedToken, pRuntimeLookup, mustRestoreHandle, TRUE);
2938 GenTreePtr impLookupToTree(CORINFO_RESOLVED_TOKEN* pResolvedToken,
2939 CORINFO_LOOKUP* pLookup,
2941 void* compileTimeHandle);
2943 GenTreePtr getRuntimeContextTree(CORINFO_RUNTIME_LOOKUP_KIND kind);
2945 GenTreePtr impRuntimeLookupToTree(CORINFO_RESOLVED_TOKEN* pResolvedToken,
2946 CORINFO_LOOKUP* pLookup,
2947 void* compileTimeHandle);
2949 GenTreePtr impReadyToRunLookupToTree(CORINFO_CONST_LOOKUP* pLookup, unsigned flags, void* compileTimeHandle);
2951 GenTreePtr impReadyToRunHelperToTree(CORINFO_RESOLVED_TOKEN* pResolvedToken,
2952 CorInfoHelpFunc helper,
2954 GenTreeArgList* arg = nullptr,
2955 CORINFO_LOOKUP_KIND* pGenericLookupKind = nullptr);
2957 GenTreePtr impCastClassOrIsInstToTree(GenTreePtr op1,
2959 CORINFO_RESOLVED_TOKEN* pResolvedToken,
2962 bool VarTypeIsMultiByteAndCanEnreg(var_types type,
2963 CORINFO_CLASS_HANDLE typeClass,
2967 static bool IsIntrinsicImplementedByUserCall(CorInfoIntrinsics intrinsicId);
2968 static bool IsTargetIntrinsic(CorInfoIntrinsics intrinsicId);
2969 static bool IsMathIntrinsic(CorInfoIntrinsics intrinsicId);
2970 static bool IsMathIntrinsic(GenTreePtr tree);
2973 //----------------- Importing the method ----------------------------------
2975 CORINFO_CONTEXT_HANDLE impTokenLookupContextHandle; // The context used for looking up tokens.
2978 unsigned impCurOpcOffs;
2979 const char* impCurOpcName;
2980 bool impNestedStackSpill;
2982 // For displaying instrs with generated native code (-n:B)
2983 GenTreePtr impLastILoffsStmt; // oldest stmt added for which we did not gtStmtLastILoffs
2984 void impNoteLastILoffs();
2987 /* IL offset of the stmt currently being imported. It gets set to
2988 BAD_IL_OFFSET after it has been set in the appended trees. Then it gets
2989 updated at IL offsets for which we have to report mapping info.
2990 It also includes flag bits, so use jitGetILoffs()
2991 to get the actual IL offset value.
2994 IL_OFFSETX impCurStmtOffs;
2995 void impCurStmtOffsSet(IL_OFFSET offs);
2997 void impNoteBranchOffs();
2999 unsigned impInitBlockLineInfo();
3001 GenTreePtr impCheckForNullPointer(GenTreePtr obj);
3002 bool impIsThis(GenTreePtr obj);
3003 bool impIsLDFTN_TOKEN(const BYTE* delegateCreateStart, const BYTE* newobjCodeAddr);
3004 bool impIsDUP_LDVIRTFTN_TOKEN(const BYTE* delegateCreateStart, const BYTE* newobjCodeAddr);
3005 bool impIsAnySTLOC(OPCODE opcode)
3007 return ((opcode == CEE_STLOC) || (opcode == CEE_STLOC_S) ||
3008 ((opcode >= CEE_STLOC_0) && (opcode <= CEE_STLOC_3)));
3011 GenTreeArgList* impPopList(unsigned count,
3013 CORINFO_SIG_INFO* sig,
3014 GenTreeArgList* prefixTree = nullptr);
3016 GenTreeArgList* impPopRevList(unsigned count,
3018 CORINFO_SIG_INFO* sig,
3019 unsigned skipReverseCount = 0);
3022 * Get current IL offset with stack-empty info incoporated
3024 IL_OFFSETX impCurILOffset(IL_OFFSET offs, bool callInstruction = false);
3026 //---------------- Spilling the importer stack ----------------------------
3032 SavedStack pdSavedStack;
3033 ThisInitState pdThisPtrInit;
3036 PendingDsc* impPendingList; // list of BBs currently waiting to be imported.
3037 PendingDsc* impPendingFree; // Freed up dscs that can be reused
3039 // We keep a byte-per-block map (dynamically extended) in the top-level Compiler object of a compilation.
3040 ExpandArray<BYTE> impPendingBlockMembers;
3042 // Return the byte for "b" (allocating/extending impPendingBlockMembers if necessary.)
3043 // Operates on the map in the top-level ancestor.
3044 BYTE impGetPendingBlockMember(BasicBlock* blk)
3046 return impInlineRoot()->impPendingBlockMembers.Get(blk->bbInd());
3049 // Set the byte for "b" to "val" (allocating/extending impPendingBlockMembers if necessary.)
3050 // Operates on the map in the top-level ancestor.
3051 void impSetPendingBlockMember(BasicBlock* blk, BYTE val)
3053 impInlineRoot()->impPendingBlockMembers.Set(blk->bbInd(), val);
3056 bool impCanReimport;
3058 bool impSpillStackEntry(unsigned level,
3062 bool bAssertOnRecursion,
3067 void impSpillStackEnsure(bool spillLeaves = false);
3068 void impEvalSideEffects();
3069 void impSpillSpecialSideEff();
3070 void impSpillSideEffects(bool spillGlobEffects, unsigned chkLevel DEBUGARG(const char* reason));
3071 void impSpillValueClasses();
3072 void impSpillEvalStack();
3073 static fgWalkPreFn impFindValueClasses;
3074 void impSpillLclRefs(ssize_t lclNum);
3076 BasicBlock* impPushCatchArgOnStack(BasicBlock* hndBlk, CORINFO_CLASS_HANDLE clsHnd);
3078 void impImportBlockCode(BasicBlock* block);
3080 void impReimportMarkBlock(BasicBlock* block);
3081 void impReimportMarkSuccessors(BasicBlock* block);
3083 void impVerifyEHBlock(BasicBlock* block, bool isTryStart);
3085 void impImportBlockPending(BasicBlock* block);
3087 // Similar to impImportBlockPending, but assumes that block has already been imported once and is being
3088 // reimported for some reason. It specifically does *not* look at verCurrentState to set the EntryState
3089 // for the block, but instead, just re-uses the block's existing EntryState.
3090 void impReimportBlockPending(BasicBlock* block);
3092 var_types impGetByRefResultType(genTreeOps oper, bool fUnsigned, GenTreePtr* pOp1, GenTreePtr* pOp2);
3094 void impImportBlock(BasicBlock* block);
3096 // Assumes that "block" is a basic block that completes with a non-empty stack. We will assign the values
3097 // on the stack to local variables (the "spill temp" variables). The successor blocks will assume that
3098 // its incoming stack contents are in those locals. This requires "block" and its successors to agree on
3099 // the variables that will be used -- and for all the predecessors of those successors, and the
3100 // successors of those predecessors, etc. Call such a set of blocks closed under alternating
3101 // successor/predecessor edges a "spill clique." A block is a "predecessor" or "successor" member of the
3102 // clique (or, conceivably, both). Each block has a specified sequence of incoming and outgoing spill
3103 // temps. If "block" already has its outgoing spill temps assigned (they are always a contiguous series
3104 // of local variable numbers, so we represent them with the base local variable number), returns that.
3105 // Otherwise, picks a set of spill temps, and propagates this choice to all blocks in the spill clique of
3106 // which "block" is a member (asserting, in debug mode, that no block in this clique had its spill temps
3107 // chosen already. More precisely, that the incoming or outgoing spill temps are not chosen, depending
3108 // on which kind of member of the clique the block is).
3109 unsigned impGetSpillTmpBase(BasicBlock* block);
3111 // Assumes that "block" is a basic block that completes with a non-empty stack. We have previously
3112 // assigned the values on the stack to local variables (the "spill temp" variables). The successor blocks
3113 // will assume that its incoming stack contents are in those locals. This requires "block" and its
3114 // successors to agree on the variables and their types that will be used. The CLI spec allows implicit
3115 // conversions between 'int' and 'native int' or 'float' and 'double' stack types. So one predecessor can
3116 // push an int and another can push a native int. For 64-bit we have chosen to implement this by typing
3117 // the "spill temp" as native int, and then importing (or re-importing as needed) so that all the
3118 // predecessors in the "spill clique" push a native int (sign-extending if needed), and all the
3119 // successors receive a native int. Similarly float and double are unified to double.
3120 // This routine is called after a type-mismatch is detected, and it will walk the spill clique to mark
3121 // blocks for re-importation as appropriate (both successors, so they get the right incoming type, and
3122 // predecessors, so they insert an upcast if needed).
3123 void impReimportSpillClique(BasicBlock* block);
3125 // When we compute a "spill clique" (see above) these byte-maps are allocated to have a byte per basic
3126 // block, and represent the predecessor and successor members of the clique currently being computed.
3127 // *** Access to these will need to be locked in a parallel compiler.
3128 ExpandArray<BYTE> impSpillCliquePredMembers;
3129 ExpandArray<BYTE> impSpillCliqueSuccMembers;
3137 // Abstract class for receiving a callback while walking a spill clique
3138 class SpillCliqueWalker
3141 virtual void Visit(SpillCliqueDir predOrSucc, BasicBlock* blk) = 0;
3144 // This class is used for setting the bbStkTempsIn and bbStkTempsOut on the blocks within a spill clique
3145 class SetSpillTempsBase : public SpillCliqueWalker
3150 SetSpillTempsBase(unsigned baseTmp) : m_baseTmp(baseTmp)
3153 virtual void Visit(SpillCliqueDir predOrSucc, BasicBlock* blk);
3156 // This class is used for implementing impReimportSpillClique part on each block within the spill clique
3157 class ReimportSpillClique : public SpillCliqueWalker
3162 ReimportSpillClique(Compiler* pComp) : m_pComp(pComp)
3165 virtual void Visit(SpillCliqueDir predOrSucc, BasicBlock* blk);
3168 // This is the heart of the algorithm for walking spill cliques. It invokes callback->Visit for each
3169 // predecessor or successor within the spill clique
3170 void impWalkSpillCliqueFromPred(BasicBlock* pred, SpillCliqueWalker* callback);
3172 // For a BasicBlock that has already been imported, the EntryState has an array of GenTrees for the
3173 // incoming locals. This walks that list an resets the types of the GenTrees to match the types of
3174 // the VarDscs. They get out of sync when we have int/native int issues (see impReimportSpillClique).
3175 void impRetypeEntryStateTemps(BasicBlock* blk);
3177 BYTE impSpillCliqueGetMember(SpillCliqueDir predOrSucc, BasicBlock* blk);
3178 void impSpillCliqueSetMember(SpillCliqueDir predOrSucc, BasicBlock* blk, BYTE val);
3180 void impPushVar(GenTree* op, typeInfo tiRetVal);
3181 void impLoadVar(unsigned lclNum, IL_OFFSET offset, typeInfo tiRetVal);
3182 void impLoadVar(unsigned lclNum, IL_OFFSET offset)
3184 impLoadVar(lclNum, offset, lvaTable[lclNum].lvVerTypeInfo);
3186 void impLoadArg(unsigned ilArgNum, IL_OFFSET offset);
3187 void impLoadLoc(unsigned ilLclNum, IL_OFFSET offset);
3188 bool impReturnInstruction(BasicBlock* block, int prefixFlags, OPCODE& opcode);
3191 void impMarkLclDstNotPromotable(unsigned tmpNum, GenTreePtr op, CORINFO_CLASS_HANDLE hClass);
3194 // A free list of linked list nodes used to represent to-do stacks of basic blocks.
3195 struct BlockListNode
3198 BlockListNode* m_next;
3199 BlockListNode(BasicBlock* blk, BlockListNode* next = nullptr) : m_blk(blk), m_next(next)
3202 void* operator new(size_t sz, Compiler* comp);
3204 BlockListNode* impBlockListNodeFreeList;
3206 BlockListNode* AllocBlockListNode();
3207 void FreeBlockListNode(BlockListNode* node);
3209 bool impIsValueType(typeInfo* pTypeInfo);
3210 var_types mangleVarArgsType(var_types type);
3213 regNumber getCallArgIntRegister(regNumber floatReg);
3214 regNumber getCallArgFloatRegister(regNumber intReg);
3215 #endif // FEATURE_VARARG
3218 static unsigned jitTotalMethodCompiled;
3222 static LONG jitNestingLevel;
3225 static BOOL impIsAddressInLocal(GenTreePtr tree, GenTreePtr* lclVarTreeOut);
3227 void impMakeDiscretionaryInlineObservations(InlineInfo* pInlineInfo, InlineResult* inlineResult);
3229 // STATIC inlining decision based on the IL code.
3230 void impCanInlineIL(CORINFO_METHOD_HANDLE fncHandle,
3231 CORINFO_METHOD_INFO* methInfo,
3233 InlineResult* inlineResult);
3235 void impCheckCanInline(GenTreePtr call,
3236 CORINFO_METHOD_HANDLE fncHandle,
3238 CORINFO_CONTEXT_HANDLE exactContextHnd,
3239 InlineCandidateInfo** ppInlineCandidateInfo,
3240 InlineResult* inlineResult);
3242 void impInlineRecordArgInfo(InlineInfo* pInlineInfo,
3243 GenTreePtr curArgVal,
3245 InlineResult* inlineResult);
3247 void impInlineInitVars(InlineInfo* pInlineInfo);
3249 unsigned impInlineFetchLocal(unsigned lclNum DEBUGARG(const char* reason));
3251 GenTreePtr impInlineFetchArg(unsigned lclNum, InlArgInfo* inlArgInfo, InlLclVarInfo* lclTypeInfo);
3253 BOOL impInlineIsThis(GenTreePtr tree, InlArgInfo* inlArgInfo);
3255 BOOL impInlineIsGuaranteedThisDerefBeforeAnySideEffects(GenTreePtr additionalTreesToBeEvaluatedBefore,
3256 GenTreePtr variableBeingDereferenced,
3257 InlArgInfo* inlArgInfo);
3259 void impMarkInlineCandidate(GenTreePtr call, CORINFO_CONTEXT_HANDLE exactContextHnd, CORINFO_CALL_INFO* callInfo);
3261 bool impTailCallRetTypeCompatible(var_types callerRetType,
3262 CORINFO_CLASS_HANDLE callerRetTypeClass,
3263 var_types calleeRetType,
3264 CORINFO_CLASS_HANDLE calleeRetTypeClass);
3266 bool impIsTailCallILPattern(bool tailPrefixed,
3268 const BYTE* codeAddrOfNextOpcode,
3269 const BYTE* codeEnd,
3271 bool* IsCallPopRet = nullptr);
3273 bool impIsImplicitTailCallCandidate(
3274 OPCODE curOpcode, const BYTE* codeAddrOfNextOpcode, const BYTE* codeEnd, int prefixFlags, bool isRecursive);
3277 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
3278 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
3282 XX Info about the basic-blocks, their contents and the flow analysis XX
3284 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
3285 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
3289 BasicBlock* fgFirstBB; // Beginning of the basic block list
3290 BasicBlock* fgLastBB; // End of the basic block list
3291 BasicBlock* fgFirstColdBlock; // First block to be placed in the cold section
3292 #if FEATURE_EH_FUNCLETS
3293 BasicBlock* fgFirstFuncletBB; // First block of outlined funclets (to allow block insertion before the funclets)
3295 BasicBlock* fgFirstBBScratch; // Block inserted for initialization stuff. Is nullptr if no such block has been
3297 BasicBlockList* fgReturnBlocks; // list of BBJ_RETURN blocks
3298 unsigned fgEdgeCount; // # of control flow edges between the BBs
3299 unsigned fgBBcount; // # of BBs in the method
3301 unsigned fgBBcountAtCodegen; // # of BBs in the method at the start of codegen
3303 unsigned fgBBNumMax; // The max bbNum that has been assigned to basic blocks
3304 unsigned fgDomBBcount; // # of BBs for which we have dominator and reachability information
3305 BasicBlock** fgBBInvPostOrder; // The flow graph stored in an array sorted in topological order, needed to compute
3306 // dominance. Indexed by block number. Size: fgBBNumMax + 1.
3308 // After the dominance tree is computed, we cache a DFS preorder number and DFS postorder number to compute
3309 // dominance queries in O(1). fgDomTreePreOrder and fgDomTreePostOrder are arrays giving the block's preorder and
3310 // postorder number, respectively. The arrays are indexed by basic block number. (Note that blocks are numbered
3311 // starting from one. Thus, we always waste element zero. This makes debugging easier and makes the code less likely
3312 // to suffer from bugs stemming from forgetting to add or subtract one from the block number to form an array
3313 // index). The arrays are of size fgBBNumMax + 1.
3314 unsigned* fgDomTreePreOrder;
3315 unsigned* fgDomTreePostOrder;
3317 bool fgBBVarSetsInited;
3319 // Allocate array like T* a = new T[fgBBNumMax + 1];
3320 // Using helper so we don't keep forgetting +1.
3321 template <typename T>
3322 T* fgAllocateTypeForEachBlk(CompMemKind cmk = CMK_Unknown)
3324 return (T*)compGetMem((fgBBNumMax + 1) * sizeof(T), cmk);
3327 // BlockSets are relative to a specific set of BasicBlock numbers. If that changes
3328 // (if the blocks are renumbered), this changes. BlockSets from different epochs
3329 // cannot be meaningfully combined. Note that new blocks can be created with higher
3330 // block numbers without changing the basic block epoch. These blocks *cannot*
3331 // participate in a block set until the blocks are all renumbered, causing the epoch
3332 // to change. This is useful if continuing to use previous block sets is valuable.
3333 // If the epoch is zero, then it is uninitialized, and block sets can't be used.
3334 unsigned fgCurBBEpoch;
3336 unsigned GetCurBasicBlockEpoch()
3338 return fgCurBBEpoch;
3341 // The number of basic blocks in the current epoch. When the blocks are renumbered,
3342 // this is fgBBcount. As blocks are added, fgBBcount increases, fgCurBBEpochSize remains
3343 // the same, until a new BasicBlock epoch is created, such as when the blocks are all renumbered.
3344 unsigned fgCurBBEpochSize;
3346 // The number of "size_t" elements required to hold a bitset large enough for fgCurBBEpochSize
3347 // bits. This is precomputed to avoid doing math every time BasicBlockBitSetTraits::GetArrSize() is called.
3348 unsigned fgBBSetCountInSizeTUnits;
3350 void NewBasicBlockEpoch()
3352 INDEBUG(unsigned oldEpochArrSize = fgBBSetCountInSizeTUnits);
3354 // We have a new epoch. Compute and cache the size needed for new BlockSets.
3356 fgCurBBEpochSize = fgBBNumMax + 1;
3357 fgBBSetCountInSizeTUnits =
3358 unsigned(roundUp(fgCurBBEpochSize, sizeof(size_t) * 8)) / unsigned(sizeof(size_t) * 8);
3361 // All BlockSet objects are now invalid!
3362 fgReachabilitySetsValid = false; // the bbReach sets are now invalid!
3363 fgEnterBlksSetValid = false; // the fgEnterBlks set is now invalid!
3367 unsigned epochArrSize = BasicBlockBitSetTraits::GetArrSize(this, sizeof(size_t));
3368 printf("\nNew BlockSet epoch %d, # of blocks (including unused BB00): %u, bitset array size: %u (%s)",
3369 fgCurBBEpoch, fgCurBBEpochSize, epochArrSize, (epochArrSize <= 1) ? "short" : "long");
3370 if ((fgCurBBEpoch != 1) && ((oldEpochArrSize <= 1) != (epochArrSize <= 1)))
3372 // If we're not just establishing the first epoch, and the epoch array size has changed such that we're
3373 // going to change our bitset representation from short (just a size_t bitset) to long (a pointer to an
3374 // array of size_t bitsets), then print that out.
3375 printf("; NOTE: BlockSet size was previously %s!", (oldEpochArrSize <= 1) ? "short" : "long");
3382 void EnsureBasicBlockEpoch()
3384 if (fgCurBBEpochSize != fgBBNumMax + 1)
3386 NewBasicBlockEpoch();
3390 BasicBlock* fgNewBasicBlock(BBjumpKinds jumpKind);
3391 void fgEnsureFirstBBisScratch();
3392 bool fgFirstBBisScratch();
3393 bool fgBBisScratch(BasicBlock* block);
3395 void fgExtendEHRegionBefore(BasicBlock* block);
3396 void fgExtendEHRegionAfter(BasicBlock* block);
3398 BasicBlock* fgNewBBbefore(BBjumpKinds jumpKind, BasicBlock* block, bool extendRegion);
3400 BasicBlock* fgNewBBafter(BBjumpKinds jumpKind, BasicBlock* block, bool extendRegion);
3402 BasicBlock* fgNewBBinRegion(BBjumpKinds jumpKind,
3405 BasicBlock* nearBlk,
3406 bool putInFilter = false,
3407 bool runRarely = false,
3408 bool insertAtEnd = false);
3410 BasicBlock* fgNewBBinRegion(BBjumpKinds jumpKind,
3412 bool runRarely = false,
3413 bool insertAtEnd = false);
3415 BasicBlock* fgNewBBinRegion(BBjumpKinds jumpKind);
3417 BasicBlock* fgNewBBinRegionWorker(BBjumpKinds jumpKind,
3418 BasicBlock* afterBlk,
3419 unsigned xcptnIndex,
3420 bool putInTryRegion);
3422 void fgInsertBBbefore(BasicBlock* insertBeforeBlk, BasicBlock* newBlk);
3423 void fgInsertBBafter(BasicBlock* insertAfterBlk, BasicBlock* newBlk);
3424 void fgUnlinkBlock(BasicBlock* block);
3426 #if OPT_BOOL_OPS // Used to detect multiple logical "not" assignments.
3427 bool fgMultipleNots;
3430 bool fgModified; // True if the flow graph has been modified recently
3431 bool fgComputePredsDone; // Have we computed the bbPreds list
3432 bool fgCheapPredsValid; // Is the bbCheapPreds list valid?
3433 bool fgDomsComputed; // Have we computed the dominator sets?
3435 bool fgHasSwitch; // any BBJ_SWITCH jumps?
3436 bool fgHasPostfix; // any postfix ++/-- found?
3437 unsigned fgIncrCount; // number of increment nodes found
3439 BlockSet fgEnterBlks; // Set of blocks which have a special transfer of control; the "entry" blocks plus EH handler
3443 bool fgReachabilitySetsValid; // Are the bbReach sets valid?
3444 bool fgEnterBlksSetValid; // Is the fgEnterBlks set valid?
3447 bool fgRemoveRestOfBlock; // true if we know that we will throw
3448 bool fgStmtRemoved; // true if we remove statements -> need new DFA
3450 // There are two modes for ordering of the trees.
3451 // - In FGOrderTree, the dominant ordering is the tree order, and the nodes contained in
3452 // each tree and sub-tree are contiguous, and can be traversed (in gtNext/gtPrev order)
3453 // by traversing the tree according to the order of the operands.
3454 // - In FGOrderLinear, the dominant ordering is the linear order.
3461 FlowGraphOrder fgOrder;
3463 // The following are boolean flags that keep track of the state of internal data structures
3465 bool fgStmtListThreaded;
3466 bool fgCanRelocateEHRegions; // true if we are allowed to relocate the EH regions
3467 bool fgEdgeWeightsComputed; // true after we have called fgComputeEdgeWeights
3468 bool fgHaveValidEdgeWeights; // true if we were successful in computing all of the edge weights
3469 bool fgSlopUsedInEdgeWeights; // true if their was some slop used when computing the edge weights
3470 bool fgRangeUsedInEdgeWeights; // true if some of the edgeWeight are expressed in Min..Max form
3471 bool fgNeedsUpdateFlowGraph; // true if we need to run fgUpdateFlowGraph
3472 BasicBlock::weight_t fgCalledWeight; // count of the number of times this method was called
3473 // This is derived from the profile data
3474 // or is BB_UNITY_WEIGHT when we don't have profile data
3476 #if FEATURE_EH_FUNCLETS
3477 bool fgFuncletsCreated; // true if the funclet creation phase has been run
3478 #endif // FEATURE_EH_FUNCLETS
3480 bool fgGlobalMorph; // indicates if we are during the global morphing phase
3481 // since fgMorphTree can be called from several places
3482 bool fgExpandInline; // indicates that we are creating tree for the inliner
3484 bool impBoxTempInUse; // the temp below is valid and available
3485 unsigned impBoxTemp; // a temporary that is used for boxing
3488 bool jitFallbackCompile; // Are we doing a fallback compile? That is, have we executed a NO_WAY assert,
3489 // and we are trying to compile again in a "safer", minopts mode?
3493 unsigned impInlinedCodeSize;
3496 //-------------------------------------------------------------------------
3504 void fgRemoveEmptyFinally();
3506 void fgCloneFinally();
3508 GenTreePtr fgGetCritSectOfStaticMethod();
3510 #if !defined(_TARGET_X86_)
3512 void fgAddSyncMethodEnterExit();
3514 GenTree* fgCreateMonitorTree(unsigned lvaMonitorBool, unsigned lvaThisVar, BasicBlock* block, bool enter);
3516 void fgConvertSyncReturnToLeave(BasicBlock* block);
3518 #endif // !_TARGET_X86_
3520 void fgAddReversePInvokeEnterExit();
3522 bool fgMoreThanOneReturnBlock();
3524 // The number of separate return points in the method.
3525 unsigned fgReturnCount;
3527 void fgAddInternal();
3529 bool fgFoldConditional(BasicBlock* block);
3531 void fgMorphStmts(BasicBlock* block, bool* mult, bool* lnot, bool* loadw);
3532 void fgMorphBlocks();
3534 bool fgMorphBlockStmt(BasicBlock* block, GenTreeStmt* stmt DEBUGARG(const char* msg));
3536 void fgCheckArgCnt();
3537 void fgSetOptions();
3540 static fgWalkPreFn fgAssertNoQmark;
3541 void fgPreExpandQmarkChecks(GenTreePtr expr);
3542 void fgPostExpandQmarkChecks();
3543 static void fgCheckQmarkAllowedForm(GenTreePtr tree);
3546 IL_OFFSET fgFindBlockILOffset(BasicBlock* block);
3548 BasicBlock* fgSplitBlockAtBeginning(BasicBlock* curr);
3549 BasicBlock* fgSplitBlockAtEnd(BasicBlock* curr);
3550 BasicBlock* fgSplitBlockAfterStatement(BasicBlock* curr, GenTree* stmt);
3551 BasicBlock* fgSplitBlockAfterNode(BasicBlock* curr, GenTree* node); // for LIR
3552 BasicBlock* fgSplitEdge(BasicBlock* curr, BasicBlock* succ);
3554 GenTreeStmt* fgNewStmtFromTree(GenTreePtr tree, BasicBlock* block, IL_OFFSETX offs);
3555 GenTreeStmt* fgNewStmtFromTree(GenTreePtr tree);
3556 GenTreeStmt* fgNewStmtFromTree(GenTreePtr tree, BasicBlock* block);
3557 GenTreeStmt* fgNewStmtFromTree(GenTreePtr tree, IL_OFFSETX offs);
3559 GenTreePtr fgGetTopLevelQmark(GenTreePtr expr, GenTreePtr* ppDst = nullptr);
3560 void fgExpandQmarkForCastInstOf(BasicBlock* block, GenTreePtr stmt);
3561 void fgExpandQmarkStmt(BasicBlock* block, GenTreePtr expr);
3562 void fgExpandQmarkNodes();
3566 // Do "simple lowering." This functionality is (conceptually) part of "general"
3567 // lowering that is distributed between fgMorph and the lowering phase of LSRA.
3568 void fgSimpleLowering();
3570 bool fgShouldCreateAssignOp(GenTreePtr tree, bool* bReverse);
3572 GenTreePtr fgInitThisClass();
3574 GenTreePtr fgGetStaticsCCtorHelper(CORINFO_CLASS_HANDLE cls, CorInfoHelpFunc helper);
3576 GenTreePtr fgGetSharedCCtor(CORINFO_CLASS_HANDLE cls);
3578 void fgLocalVarLiveness();
3580 void fgLocalVarLivenessInit();
3582 #ifdef LEGACY_BACKEND
3583 GenTreePtr fgLegacyPerStatementLocalVarLiveness(GenTreePtr startNode, GenTreePtr relopNode);
3585 void fgPerNodeLocalVarLiveness(GenTree* node);
3587 void fgPerBlockLocalVarLiveness();
3589 VARSET_VALRET_TP fgGetHandlerLiveVars(BasicBlock* block);
3591 void fgLiveVarAnalysis(bool updateInternalOnly = false);
3593 // This is used in the liveness computation, as a temporary. When we use the
3594 // arbitrary-length VarSet representation, it is better not to allocate a new one
3596 VARSET_TP fgMarkIntfUnionVS;
3598 bool fgMarkIntf(VARSET_VALARG_TP varSet);
3600 bool fgMarkIntf(VARSET_VALARG_TP varSet1, VARSET_VALARG_TP varSet2);
3602 void fgUpdateRefCntForClone(BasicBlock* addedToBlock, GenTreePtr clonedTree);
3604 void fgUpdateRefCntForExtract(GenTreePtr wholeTree, GenTreePtr keptTree);
3606 void fgComputeLifeCall(VARSET_TP& life, GenTreeCall* call);
3608 bool fgComputeLifeLocal(VARSET_TP& life, VARSET_TP& keepAliveVars, GenTree* lclVarNode, GenTree* node);
3610 VARSET_VALRET_TP fgComputeLife(VARSET_VALARG_TP life,
3611 GenTreePtr startNode,
3613 VARSET_VALARG_TP volatileVars,
3614 bool* pStmtInfoDirty DEBUGARG(bool* treeModf));
3616 VARSET_VALRET_TP fgComputeLifeLIR(VARSET_VALARG_TP life, BasicBlock* block, VARSET_VALARG_TP volatileVars);
3618 bool fgRemoveDeadStore(GenTree** pTree,
3622 bool* pStmtInfoDirty DEBUGARG(bool* treeModf));
3624 bool fgTryRemoveDeadLIRStore(LIR::Range& blockRange, GenTree* node, GenTree** next);
3626 // For updating liveset during traversal AFTER fgComputeLife has completed
3627 VARSET_VALRET_TP fgGetVarBits(GenTreePtr tree);
3628 VARSET_VALRET_TP fgUpdateLiveSet(VARSET_VALARG_TP liveSet, GenTreePtr tree);
3630 // Returns the set of live variables after endTree,
3631 // assuming that liveSet is the set of live variables BEFORE tree.
3632 // Requires that fgComputeLife has completed, and that tree is in the same
3633 // statement as endTree, and that it comes before endTree in execution order
3635 VARSET_VALRET_TP fgUpdateLiveSet(VARSET_VALARG_TP liveSet, GenTreePtr tree, GenTreePtr endTree)
3637 VARSET_TP VARSET_INIT(this, newLiveSet, liveSet);
3638 while (tree != nullptr && tree != endTree->gtNext)
3640 VarSetOps::AssignNoCopy(this, newLiveSet, fgUpdateLiveSet(newLiveSet, tree));
3641 tree = tree->gtNext;
3643 assert(tree == endTree->gtNext);
3647 void fgInterBlockLocalVarLiveness();
3649 // The presence of "x op= y" operations presents some difficulties for SSA: this is both a use of some SSA name of
3650 // "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
3651 // whether to treat that as the use or def. It chooses the "use", and thus the old SSA name. This map allows us
3652 // to record/recover the "def" SSA number, given the lcl var node for "x" in such a tree.
3653 typedef SimplerHashTable<GenTreePtr, PtrKeyFuncs<GenTree>, unsigned, JitSimplerHashBehavior> NodeToUnsignedMap;
3654 NodeToUnsignedMap* m_opAsgnVarDefSsaNums;
3655 NodeToUnsignedMap* GetOpAsgnVarDefSsaNums()
3657 if (m_opAsgnVarDefSsaNums == nullptr)
3659 m_opAsgnVarDefSsaNums = new (getAllocator()) NodeToUnsignedMap(getAllocator());
3661 return m_opAsgnVarDefSsaNums;
3664 // Requires value numbering phase to have completed. Returns the value number ("gtVN") of the
3665 // "tree," EXCEPT in the case of GTF_VAR_USEASG, because the tree node's gtVN member is the
3666 // "use" VN. Performs a lookup into the map of (use asg tree -> def VN.) to return the "def's"
3668 inline ValueNum GetUseAsgDefVNOrTreeVN(GenTreePtr tree);
3670 // Requires that "lcl" has the GTF_VAR_DEF flag set. Returns the SSA number of "lcl".
3671 // Except: assumes that lcl is a def, and if it is
3672 // a def appearing in "lcl op= rhs" (GTF_VAR_USEASG), looks up and returns the SSA number for the "def",
3673 // rather than the "use" SSA number recorded in the tree "lcl".
3674 inline unsigned GetSsaNumForLocalVarDef(GenTreePtr lcl);
3676 // Some assignments assign to a local "indirectly": they are part of a comma expression that takes the address
3677 // of the local (or a field thereof), assigns this address to a temp, and uses an indirection of this temp as
3678 // the LHS of the assignment. This actually arises in exactly one situation. At the source level we assign one
3679 // struct local to another: "s1 = s2". This becomes a copyblk. If "s2" is promoted into field variables "s2f0",
3680 // ..."s2fn", then the copyblk will morph to a comma expression that takes the address of "s1" and does field-wise
3682 // (byref addrS1 = &s1,
3683 // *(addrS1 * offsetof(f0)) = s2f0,
3685 // *(addrS1 * offsetof(fn)) = s2fn)
3687 // It would be a shame, given the simple form at the source level, to be unable to track the values in the
3688 // fields of "s1" after this. But "s1" does not appear in the assignments that modify it. How, then, to
3689 // give it SSA names and value numbers?
3691 // The solution is to use the side table described below to annotate each of the field-wise assignments at the
3692 // end with an instance of the structure below, whose fields are described in the declaration.
3693 struct IndirectAssignmentAnnotation
3695 unsigned m_lclNum; // The local num that is being indirectly assigned.
3696 FieldSeqNode* m_fieldSeq; // If the LHS of the struct assignment is itself a struct field dereference,
3697 // as in "s0.g = s2", then "m_lclNum" would be "s0", and "m_fieldSeq" would
3698 // be the singleton field sequence "g". The individual assignments would
3699 // further append the fields of "s.g" to that.
3700 bool m_isEntire; // True iff this assignment writes all of m_lclNum. (This can occur if the
3701 // structure has a single field).
3702 unsigned m_defSsaNum; // The new SSA number of "m_lclNum" after the assignment.
3703 unsigned m_useSsaNum; // Only valid if "m_isEntire" is false; if so, the SSA number of "m_lclNum" before the
3706 IndirectAssignmentAnnotation(unsigned lclNum,
3707 FieldSeqNode* fldSeq,
3709 unsigned defSsaNum = SsaConfig::RESERVED_SSA_NUM,
3710 unsigned useSsaNum = SsaConfig::RESERVED_SSA_NUM)
3711 : m_lclNum(lclNum), m_fieldSeq(fldSeq), m_isEntire(isEntire), m_defSsaNum(defSsaNum), m_useSsaNum(useSsaNum)
3715 typedef SimplerHashTable<GenTreePtr, PtrKeyFuncs<GenTree>, IndirectAssignmentAnnotation*, JitSimplerHashBehavior>
3716 NodeToIndirAssignMap;
3717 NodeToIndirAssignMap* m_indirAssignMap;
3718 NodeToIndirAssignMap* GetIndirAssignMap()
3720 if (m_indirAssignMap == nullptr)
3722 // Create a CompAllocator that labels sub-structure with CMK_IndirAssignMap, and use that for allocation.
3723 IAllocator* ialloc = new (this, CMK_IndirAssignMap) CompAllocator(this, CMK_IndirAssignMap);
3724 m_indirAssignMap = new (ialloc) NodeToIndirAssignMap(ialloc);
3726 return m_indirAssignMap;
3729 // Performs SSA conversion.
3732 // Reset any data structures to the state expected by "fgSsaBuild", so it can be run again.
3733 void fgResetForSsa();
3735 unsigned fgSsaPassesCompleted; // Number of times fgSsaBuild has been run.
3737 // Returns "true" iff lcl "lclNum" should be excluded from SSA.
3738 inline bool fgExcludeFromSsa(unsigned lclNum);
3740 // The value numbers for this compilation.
3741 ValueNumStore* vnStore;
3744 ValueNumStore* GetValueNumStore()
3749 // Do value numbering (assign a value number to each
3751 void fgValueNumber();
3753 // Updates "fgCurHeap" via the assignment H[elemTypeEq][arrVN][inx][fldSeq] = rhsVN.
3754 // Assumes that "elemTypeEq" is the (equivalence class rep) of the array element type.
3755 // The 'indType' is the indirection type of the lhs of the assignment and will typically
3756 // match the element type of the array or fldSeq. When this type doesn't match
3757 // or if the fldSeq is 'NotAField' we invalidate the array contents H[elemTypeEq][arrVN]
3759 void fgValueNumberArrIndexAssign(CORINFO_CLASS_HANDLE elemTypeEq,
3762 FieldSeqNode* fldSeq,
3766 // Requires that "tree" is a GT_IND marked as an array index, and that its address argument
3767 // has been parsed to yield the other input arguments. If evaluation of the address
3768 // can raise exceptions, those should be captured in the exception set "excVN."
3769 // Assumes that "elemTypeEq" is the (equivalence class rep) of the array element type.
3770 // Marks "tree" with the VN for H[elemTypeEq][arrVN][inx][fldSeq] (for the liberal VN; a new unique
3771 // VN for the conservative VN.) Also marks the tree's argument as the address of an array element.
3772 // The type tree->TypeGet() will typically match the element type of the array or fldSeq.
3773 // When this type doesn't match or if the fldSeq is 'NotAField' we return a new unique VN
3775 ValueNum fgValueNumberArrIndexVal(GenTreePtr tree,
3776 CORINFO_CLASS_HANDLE elemTypeEq,
3780 FieldSeqNode* fldSeq);
3782 // Requires "funcApp" to be a VNF_PtrToArrElem, and "addrXvn" to represent the exception set thrown
3783 // by evaluating the array index expression "tree". Returns the value number resulting from
3784 // dereferencing the array in the current heap state. If "tree" is non-null, it must be the
3785 // "GT_IND" that does the dereference, and it is given the returned value number.
3786 ValueNum fgValueNumberArrIndexVal(GenTreePtr tree, struct VNFuncApp* funcApp, ValueNum addrXvn);
3788 unsigned fgVNPassesCompleted; // Number of times fgValueNumber has been run.
3790 // Utility functions for fgValueNumber.
3792 // Perform value-numbering for the trees in "blk".
3793 void fgValueNumberBlock(BasicBlock* blk);
3795 // Requires that "entryBlock" is the entry block of loop "loopNum", and that "loopNum" is the
3796 // innermost loop of which "entryBlock" is the entry. Returns the value number that should be
3797 // assumed for the heap at the start "entryBlk".
3798 ValueNum fgHeapVNForLoopSideEffects(BasicBlock* entryBlock, unsigned loopNum);
3800 // Called when an operation (performed by "tree", described by "msg") may cause the global Heap to be mutated.
3801 void fgMutateHeap(GenTreePtr tree DEBUGARG(const char* msg));
3803 // Tree caused an update in the current heap VN. If "tree" has an associated heap SSA #, record that
3804 // value in that SSA #.
3805 void fgValueNumberRecordHeapSsa(GenTreePtr tree);
3807 // The input 'tree' is a leaf node that is a constant
3808 // Assign the proper value number to the tree
3809 void fgValueNumberTreeConst(GenTreePtr tree);
3811 // Assumes that all inputs to "tree" have had value numbers assigned; assigns a VN to tree.
3812 // (With some exceptions: the VN of the lhs of an assignment is assigned as part of the
3814 // If "evalAsgLhsInd" is true, evaluate a GT_IND node, even if it's labeled as the LHS of
3816 void fgValueNumberTree(GenTreePtr tree, bool evalAsgLhsInd = false);
3818 // Does value-numbering for a block assignment.
3819 void fgValueNumberBlockAssignment(GenTreePtr tree, bool evalAsgLhsInd);
3821 // Does value-numbering for a cast tree.
3822 void fgValueNumberCastTree(GenTreePtr tree);
3824 // Does value-numbering for an intrinsic tree.
3825 void fgValueNumberIntrinsic(GenTreePtr tree);
3827 // Does value-numbering for a call. We interpret some helper calls.
3828 void fgValueNumberCall(GenTreeCall* call);
3830 // The VN of some nodes in "args" may have changed -- reassign VNs to the arg list nodes.
3831 void fgUpdateArgListVNs(GenTreeArgList* args);
3833 // Does value-numbering for a helper "call" that has a VN function symbol "vnf".
3834 void fgValueNumberHelperCallFunc(GenTreeCall* call, VNFunc vnf, ValueNumPair vnpExc);
3836 // Requires "helpCall" to be a helper call. Assigns it a value number;
3837 // we understand the semantics of some of the calls. Returns "true" if
3838 // the call may modify the heap (we assume arbitrary memory side effects if so).
3839 bool fgValueNumberHelperCall(GenTreeCall* helpCall);
3841 // Requires "helpFunc" to be pure. Returns the corresponding VNFunc.
3842 VNFunc fgValueNumberHelperMethVNFunc(CorInfoHelpFunc helpFunc);
3844 // This is the current value number for the "Heap" implicit variable while
3845 // doing value numbering. This is the value number under the "liberal" interpretation
3846 // of heap values; the "conservative" interpretation needs no VN, since every access of
3847 // the heap yields an unknown value.
3848 ValueNum fgCurHeapVN;
3850 // Return a "pseudo"-class handle for an array element type. If "elemType" is TYP_STRUCT,
3851 // requires "elemStructType" to be non-null (and to have a low-order zero). Otherwise, low order bit
3852 // is 1, and the rest is an encoding of "elemTyp".
3853 static CORINFO_CLASS_HANDLE EncodeElemType(var_types elemTyp, CORINFO_CLASS_HANDLE elemStructType)
3855 if (elemStructType != nullptr)
3857 assert(varTypeIsStruct(elemTyp) || elemTyp == TYP_REF || elemTyp == TYP_BYREF ||
3858 varTypeIsIntegral(elemTyp));
3859 assert((size_t(elemStructType) & 0x1) == 0x0); // Make sure the encoding below is valid.
3860 return elemStructType;
3864 elemTyp = varTypeUnsignedToSigned(elemTyp);
3865 return CORINFO_CLASS_HANDLE(size_t(elemTyp) << 1 | 0x1);
3868 // If "clsHnd" is the result of an "EncodePrim" call, returns true and sets "*pPrimType" to the
3869 // var_types it represents. Otherwise, returns TYP_STRUCT (on the assumption that "clsHnd" is
3870 // the struct type of the element).
3871 static var_types DecodeElemType(CORINFO_CLASS_HANDLE clsHnd)
3873 size_t clsHndVal = size_t(clsHnd);
3874 if (clsHndVal & 0x1)
3876 return var_types(clsHndVal >> 1);
3884 // Convert a BYTE which represents the VM's CorInfoGCtype to the JIT's var_types
3885 var_types getJitGCType(BYTE gcType);
3887 enum structPassingKind
3889 SPK_Unknown, // Invalid value, never returned
3890 SPK_PrimitiveType, // The struct is passed/returned using a primitive type.
3891 SPK_ByValue, // The struct is passed/returned by value (using the ABI rules)
3892 // for ARM64 and UNIX_X64 in multiple registers. (when all of the
3893 // parameters registers are used, then the stack will be used)
3894 // for X86 passed on the stack, for ARM32 passed in registers
3895 // or the stack or split between registers and the stack.
3896 SPK_ByValueAsHfa, // The struct is passed/returned as an HFA in multiple registers.
3898 }; // The struct is passed/returned by reference to a copy/buffer.
3900 // Get the "primitive" type that is is used when we are given a struct of size 'structSize'.
3901 // For pointer sized structs the 'clsHnd' is used to determine if the struct contains GC ref.
3902 // A "primitive" type is one of the scalar types: byte, short, int, long, ref, float, double
3903 // If we can't or shouldn't use a "primitive" type then TYP_UNKNOWN is returned.
3905 var_types getPrimitiveTypeForStruct(unsigned structSize, CORINFO_CLASS_HANDLE clsHnd);
3907 // Get the type that is used to pass values of the given struct type.
3908 // If you have already retrieved the struct size then pass it as the optional third argument
3910 var_types getArgTypeForStruct(CORINFO_CLASS_HANDLE clsHnd,
3911 structPassingKind* wbPassStruct,
3912 unsigned structSize = 0);
3914 // Get the type that is used to return values of the given struct type.
3915 // If you have already retrieved the struct size then pass it as the optional third argument
3917 var_types getReturnTypeForStruct(CORINFO_CLASS_HANDLE clsHnd,
3918 structPassingKind* wbPassStruct = nullptr,
3919 unsigned structSize = 0);
3922 // Print a representation of "vnp" or "vn" on standard output.
3923 // If "level" is non-zero, we also print out a partial expansion of the value.
3924 void vnpPrint(ValueNumPair vnp, unsigned level);
3925 void vnPrint(ValueNum vn, unsigned level);
3928 // Dominator computation member functions
3929 // Not exposed outside Compiler
3931 bool fgDominate(BasicBlock* b1, BasicBlock* b2); // Return true if b1 dominates b2
3933 bool fgReachable(BasicBlock* b1, BasicBlock* b2); // Returns true if block b1 can reach block b2
3935 void fgComputeDoms(); // Computes the immediate dominators for each basic block in the
3936 // flow graph. We first assume the fields bbIDom on each
3937 // basic block are invalid. This computation is needed later
3938 // by fgBuildDomTree to build the dominance tree structure.
3939 // Based on: A Simple, Fast Dominance Algorithm
3940 // by Keith D. Cooper, Timothy J. Harvey, and Ken Kennedy
3942 BlockSet_ValRet_T fgGetDominatorSet(BasicBlock* block); // Returns a set of blocks that dominate the given block.
3943 // Note: this is relatively slow compared to calling fgDominate(),
3944 // especially if dealing with a single block versus block check.
3946 void fgComputeReachabilitySets(); // Compute bbReach sets. (Also sets BBF_GC_SAFE_POINT flag on blocks.)
3948 void fgComputeEnterBlocksSet(); // Compute the set of entry blocks, 'fgEnterBlks'.
3950 bool fgRemoveUnreachableBlocks(); // Remove blocks determined to be unreachable by the bbReach sets.
3952 void fgComputeReachability(); // Perform flow graph node reachability analysis.
3954 BasicBlock* fgIntersectDom(BasicBlock* a, BasicBlock* b); // Intersect two immediate dominator sets.
3956 void fgDfsInvPostOrder(); // In order to compute dominance using fgIntersectDom, the flow graph nodes must be
3957 // processed in topological sort, this function takes care of that.
3959 void fgDfsInvPostOrderHelper(BasicBlock* block, BlockSet& visited, unsigned* count);
3961 BlockSet_ValRet_T fgDomFindStartNodes(); // Computes which basic blocks don't have incoming edges in the flow graph.
3962 // Returns this as a set.
3964 BlockSet_ValRet_T fgDomTreeEntryNodes(BasicBlockList** domTree); // Computes which nodes in the dominance forest are
3965 // root nodes. Returns this as a set.
3968 void fgDispDomTree(BasicBlockList** domTree); // Helper that prints out the Dominator Tree in debug builds.
3971 void fgBuildDomTree(); // Once we compute all the immediate dominator sets for each node in the flow graph
3972 // (performed by fgComputeDoms), this procedure builds the dominance tree represented
3975 // In order to speed up the queries of the form 'Does A dominates B', we can perform a DFS preorder and postorder
3976 // traversal of the dominance tree and the dominance query will become A dominates B iif preOrder(A) <= preOrder(B)
3977 // && postOrder(A) >= postOrder(B) making the computation O(1).
3978 void fgTraverseDomTree(unsigned bbNum, BasicBlockList** domTree, unsigned* preNum, unsigned* postNum);
3980 // When the flow graph changes, we need to update the block numbers, predecessor lists, reachability sets, and
3982 void fgUpdateChangedFlowGraph();
3985 // Compute the predecessors of the blocks in the control flow graph.
3986 void fgComputePreds();
3988 // Remove all predecessor information.
3989 void fgRemovePreds();
3991 // Compute the cheap flow graph predecessors lists. This is used in some early phases
3992 // before the full predecessors lists are computed.
3993 void fgComputeCheapPreds();
3996 void fgAddCheapPred(BasicBlock* block, BasicBlock* blockPred);
3998 void fgRemoveCheapPred(BasicBlock* block, BasicBlock* blockPred);
4008 // Initialize the per-block variable sets (used for liveness analysis).
4009 void fgInitBlockVarSets();
4011 // true if we've gone through and created GC Poll calls.
4012 bool fgGCPollsCreated;
4013 void fgMarkGCPollBlocks();
4014 void fgCreateGCPolls();
4015 bool fgCreateGCPoll(GCPollType pollType, BasicBlock* block);
4017 // Requires that "block" is a block that returns from
4018 // a finally. Returns the number of successors (jump targets of
4019 // of blocks in the covered "try" that did a "LEAVE".)
4020 unsigned fgNSuccsOfFinallyRet(BasicBlock* block);
4022 // Requires that "block" is a block that returns (in the sense of BBJ_EHFINALLYRET) from
4023 // a finally. Returns its "i"th successor (jump targets of
4024 // of blocks in the covered "try" that did a "LEAVE".)
4025 // Requires that "i" < fgNSuccsOfFinallyRet(block).
4026 BasicBlock* fgSuccOfFinallyRet(BasicBlock* block, unsigned i);
4029 // Factor out common portions of the impls of the methods above.
4030 void fgSuccOfFinallyRetWork(BasicBlock* block, unsigned i, BasicBlock** bres, unsigned* nres);
4033 // For many purposes, it is desirable to be able to enumerate the *distinct* targets of a switch statement,
4034 // skipping duplicate targets. (E.g., in flow analyses that are only interested in the set of possible targets.)
4035 // SwitchUniqueSuccSet contains the non-duplicated switch targets.
4036 // (Code that modifies the jump table of a switch has an obligation to call Compiler::UpdateSwitchTableTarget,
4037 // which in turn will call the "UpdateTarget" method of this type if a SwitchUniqueSuccSet has already
4038 // been computed for the switch block. If a switch block is deleted or is transformed into a non-switch,
4039 // we leave the entry associated with the block, but it will no longer be accessed.)
4040 struct SwitchUniqueSuccSet
4042 unsigned numDistinctSuccs; // Number of distinct targets of the switch.
4043 BasicBlock** nonDuplicates; // Array of "numDistinctSuccs", containing all the distinct switch target
4046 // The switch block "switchBlk" just had an entry with value "from" modified to the value "to".
4047 // Update "this" as necessary: if "from" is no longer an element of the jump table of "switchBlk",
4048 // remove it from "this", and ensure that "to" is a member. Use "alloc" to do any required allocation.
4049 void UpdateTarget(IAllocator* alloc, BasicBlock* switchBlk, BasicBlock* from, BasicBlock* to);
4052 typedef SimplerHashTable<BasicBlock*, PtrKeyFuncs<BasicBlock>, SwitchUniqueSuccSet, JitSimplerHashBehavior>
4053 BlockToSwitchDescMap;
4056 // Maps BasicBlock*'s that end in switch statements to SwitchUniqueSuccSets that allow
4057 // iteration over only the distinct successors.
4058 BlockToSwitchDescMap* m_switchDescMap;
4061 BlockToSwitchDescMap* GetSwitchDescMap()
4063 if (m_switchDescMap == nullptr)
4065 m_switchDescMap = new (getAllocator()) BlockToSwitchDescMap(getAllocator());
4067 return m_switchDescMap;
4070 // Invalidate the map of unique switch block successors. For example, since the hash key of the map
4071 // depends on block numbers, we must invalidate the map when the blocks are renumbered, to ensure that
4072 // we don't accidentally look up and return the wrong switch data.
4073 void InvalidateUniqueSwitchSuccMap()
4075 m_switchDescMap = nullptr;
4078 // Requires "switchBlock" to be a block that ends in a switch. Returns
4079 // the corresponding SwitchUniqueSuccSet.
4080 SwitchUniqueSuccSet GetDescriptorForSwitch(BasicBlock* switchBlk);
4082 // The switch block "switchBlk" just had an entry with value "from" modified to the value "to".
4083 // Update "this" as necessary: if "from" is no longer an element of the jump table of "switchBlk",
4084 // remove it from "this", and ensure that "to" is a member.
4085 void UpdateSwitchTableTarget(BasicBlock* switchBlk, BasicBlock* from, BasicBlock* to);
4087 // Remove the "SwitchUniqueSuccSet" of "switchBlk" in the BlockToSwitchDescMap.
4088 void fgInvalidateSwitchDescMapEntry(BasicBlock* switchBlk);
4090 BasicBlock* fgFirstBlockOfHandler(BasicBlock* block);
4092 flowList* fgGetPredForBlock(BasicBlock* block, BasicBlock* blockPred);
4094 flowList* fgGetPredForBlock(BasicBlock* block, BasicBlock* blockPred, flowList*** ptrToPred);
4096 flowList* fgSpliceOutPred(BasicBlock* block, BasicBlock* blockPred);
4098 flowList* fgRemoveRefPred(BasicBlock* block, BasicBlock* blockPred);
4100 flowList* fgRemoveAllRefPreds(BasicBlock* block, BasicBlock* blockPred);
4102 flowList* fgRemoveAllRefPreds(BasicBlock* block, flowList** ptrToPred);
4104 void fgRemoveBlockAsPred(BasicBlock* block);
4106 void fgChangeSwitchBlock(BasicBlock* oldSwitchBlock, BasicBlock* newSwitchBlock);
4108 void fgReplaceSwitchJumpTarget(BasicBlock* blockSwitch, BasicBlock* newTarget, BasicBlock* oldTarget);
4110 void fgReplaceJumpTarget(BasicBlock* block, BasicBlock* newTarget, BasicBlock* oldTarget);
4112 void fgReplacePred(BasicBlock* block, BasicBlock* oldPred, BasicBlock* newPred);
4114 flowList* fgAddRefPred(BasicBlock* block,
4115 BasicBlock* blockPred,
4116 flowList* oldEdge = nullptr,
4117 bool initializingPreds = false); // Only set to 'true' when we are computing preds in
4120 void fgFindBasicBlocks();
4122 bool fgIsBetterFallThrough(BasicBlock* bCur, BasicBlock* bAlt);
4124 bool fgCheckEHCanInsertAfterBlock(BasicBlock* blk, unsigned regionIndex, bool putInTryRegion);
4126 BasicBlock* fgFindInsertPoint(unsigned regionIndex,
4127 bool putInTryRegion,
4128 BasicBlock* startBlk,
4130 BasicBlock* nearBlk,
4131 BasicBlock* jumpBlk,
4134 unsigned fgGetNestingLevel(BasicBlock* block, unsigned* pFinallyNesting = nullptr);
4136 void fgRemoveEmptyBlocks();
4138 void fgRemoveStmt(BasicBlock* block, GenTreePtr stmt, bool updateRefCnt = true);
4140 bool fgCheckRemoveStmt(BasicBlock* block, GenTreePtr stmt);
4142 void fgCreateLoopPreHeader(unsigned lnum);
4144 void fgUnreachableBlock(BasicBlock* block);
4146 void fgRemoveConditionalJump(BasicBlock* block);
4148 BasicBlock* fgLastBBInMainFunction();
4150 BasicBlock* fgEndBBAfterMainFunction();
4152 void fgUnlinkRange(BasicBlock* bBeg, BasicBlock* bEnd);
4154 void fgRemoveBlock(BasicBlock* block, bool unreachable);
4156 bool fgCanCompactBlocks(BasicBlock* block, BasicBlock* bNext);
4158 void fgCompactBlocks(BasicBlock* block, BasicBlock* bNext);
4160 void fgUpdateLoopsAfterCompacting(BasicBlock* block, BasicBlock* bNext);
4162 BasicBlock* fgConnectFallThrough(BasicBlock* bSrc, BasicBlock* bDst);
4164 bool fgRenumberBlocks();
4166 bool fgExpandRarelyRunBlocks();
4168 bool fgEhAllowsMoveBlock(BasicBlock* bBefore, BasicBlock* bAfter);
4170 void fgMoveBlocksAfter(BasicBlock* bStart, BasicBlock* bEnd, BasicBlock* insertAfterBlk);
4172 enum FG_RELOCATE_TYPE
4174 FG_RELOCATE_TRY, // relocate the 'try' region
4175 FG_RELOCATE_HANDLER // relocate the handler region (including the filter if necessary)
4177 BasicBlock* fgRelocateEHRange(unsigned regionIndex, FG_RELOCATE_TYPE relocateType);
4179 #if FEATURE_EH_FUNCLETS
4180 #if defined(_TARGET_ARM_)
4181 void fgClearFinallyTargetBit(BasicBlock* block);
4182 #endif // defined(_TARGET_ARM_)
4183 bool fgIsIntraHandlerPred(BasicBlock* predBlock, BasicBlock* block);
4184 bool fgAnyIntraHandlerPreds(BasicBlock* block);
4185 void fgInsertFuncletPrologBlock(BasicBlock* block);
4186 void fgCreateFuncletPrologBlocks();
4187 void fgCreateFunclets();
4188 #else // !FEATURE_EH_FUNCLETS
4189 bool fgRelocateEHRegions();
4190 #endif // !FEATURE_EH_FUNCLETS
4192 bool fgOptimizeUncondBranchToSimpleCond(BasicBlock* block, BasicBlock* target);
4194 bool fgBlockEndFavorsTailDuplication(BasicBlock* block);
4196 bool fgBlockIsGoodTailDuplicationCandidate(BasicBlock* block);
4198 bool fgOptimizeFallthroughTailDup(BasicBlock* block, BasicBlock* target);
4200 bool fgOptimizeEmptyBlock(BasicBlock* block);
4202 bool fgOptimizeBranchToEmptyUnconditional(BasicBlock* block, BasicBlock* bDest);
4204 bool fgOptimizeBranch(BasicBlock* bJump);
4206 bool fgOptimizeSwitchBranches(BasicBlock* block);
4208 bool fgOptimizeBranchToNext(BasicBlock* block, BasicBlock* bNext, BasicBlock* bPrev);
4210 bool fgOptimizeSwitchJumps();
4212 void fgPrintEdgeWeights();
4214 void fgComputeEdgeWeights();
4216 void fgReorderBlocks();
4218 void fgDetermineFirstColdBlock();
4220 bool fgIsForwardBranch(BasicBlock* bJump, BasicBlock* bSrc = nullptr);
4222 bool fgUpdateFlowGraph(bool doTailDup = false);
4224 void fgFindOperOrder();
4226 // method that returns if you should split here
4227 typedef bool(fgSplitPredicate)(GenTree* tree, GenTree* parent, fgWalkData* data);
4229 void fgSetBlockOrder();
4231 void fgRemoveReturnBlock(BasicBlock* block);
4233 /* Helper code that has been factored out */
4234 inline void fgConvertBBToThrowBB(BasicBlock* block);
4236 bool fgCastNeeded(GenTreePtr tree, var_types toType);
4237 GenTreePtr fgDoNormalizeOnStore(GenTreePtr tree);
4238 GenTreePtr fgMakeTmpArgNode(
4239 unsigned tmpVarNum FEATURE_UNIX_AMD64_STRUCT_PASSING_ONLY_ARG(const bool passedInRegisters));
4241 // The following check for loops that don't execute calls
4242 bool fgLoopCallMarked;
4244 void fgLoopCallTest(BasicBlock* srcBB, BasicBlock* dstBB);
4245 void fgLoopCallMark();
4247 void fgMarkLoopHead(BasicBlock* block);
4249 unsigned fgGetCodeEstimate(BasicBlock* block);
4252 const char* fgProcessEscapes(const char* nameIn, escapeMapping_t* map);
4253 FILE* fgOpenFlowGraphFile(bool* wbDontClose, Phases phase, LPCWSTR type);
4254 bool fgDumpFlowGraph(Phases phase);
4256 #endif // DUMP_FLOWGRAPHS
4261 void fgDispBBLiveness(BasicBlock* block);
4262 void fgDispBBLiveness();
4263 void fgTableDispBasicBlock(BasicBlock* block, int ibcColWidth = 0);
4264 void fgDispBasicBlocks(BasicBlock* firstBlock, BasicBlock* lastBlock, bool dumpTrees);
4265 void fgDispBasicBlocks(bool dumpTrees = false);
4266 void fgDumpStmtTree(GenTreePtr stmt, unsigned blkNum);
4267 void fgDumpBlock(BasicBlock* block);
4268 void fgDumpTrees(BasicBlock* firstBlock, BasicBlock* lastBlock);
4270 static fgWalkPreFn fgStress64RsltMulCB;
4271 void fgStress64RsltMul();
4272 void fgDebugCheckUpdate();
4273 void fgDebugCheckBBlist(bool checkBBNum = false, bool checkBBRefs = true);
4274 void fgDebugCheckBlockLinks();
4275 void fgDebugCheckLinks(bool morphTrees = false);
4276 void fgDebugCheckNodeLinks(BasicBlock* block, GenTreePtr stmt);
4277 void fgDebugCheckFlags(GenTreePtr tree);
4278 void fgDebugCheckFlagsHelper(GenTreePtr tree, unsigned treeFlags, unsigned chkFlags);
4279 void fgDebugCheckTryFinallyExits();
4282 #ifdef LEGACY_BACKEND
4283 static void fgOrderBlockOps(GenTreePtr tree,
4287 GenTreePtr* opsPtr, // OUT
4288 regMaskTP* regsPtr); // OUT
4289 #endif // LEGACY_BACKEND
4291 static GenTreePtr fgGetFirstNode(GenTreePtr tree);
4292 static bool fgTreeIsInStmt(GenTree* tree, GenTreeStmt* stmt);
4294 inline bool fgIsInlining()
4296 return fgExpandInline;
4299 void fgTraverseRPO();
4301 //--------------------- Walking the trees in the IR -----------------------
4306 fgWalkPreFn* wtprVisitorFn;
4307 fgWalkPostFn* wtpoVisitorFn;
4308 void* pCallbackData; // user-provided data
4309 bool wtprLclsOnly; // whether to only visit lclvar nodes
4310 GenTreePtr parent; // parent of current node, provided to callback
4311 GenTreeStack* parentStack; // stack of parent nodes, if asked for
4313 bool printModified; // callback can use this
4317 template <bool computeStack>
4318 static fgWalkResult fgWalkTreePreRec(GenTreePtr* pTree, fgWalkData* fgWalkPre);
4320 // general purpose tree-walker that is capable of doing pre- and post- order
4321 // callbacks at the same time
4322 template <bool doPreOrder, bool doPostOrder>
4323 static fgWalkResult fgWalkTreeRec(GenTreePtr* pTree, fgWalkData* fgWalkPre);
4325 fgWalkResult fgWalkTreePre(GenTreePtr* pTree,
4326 fgWalkPreFn* visitor,
4327 void* pCallBackData = nullptr,
4328 bool lclVarsOnly = false,
4329 bool computeStack = false);
4331 fgWalkResult fgWalkTree(GenTreePtr* pTree,
4332 fgWalkPreFn* preVisitor,
4333 fgWalkPostFn* postVisitor,
4334 void* pCallBackData = nullptr);
4336 void fgWalkAllTreesPre(fgWalkPreFn* visitor, void* pCallBackData);
4340 template <bool computeStack>
4341 static fgWalkResult fgWalkTreePostRec(GenTreePtr* pTree, fgWalkData* fgWalkPre);
4343 fgWalkResult fgWalkTreePost(GenTreePtr* pTree,
4344 fgWalkPostFn* visitor,
4345 void* pCallBackData = nullptr,
4346 bool computeStack = false);
4348 // An fgWalkPreFn that looks for expressions that have inline throws in
4349 // minopts mode. Basically it looks for tress with gtOverflowEx() or
4350 // GTF_IND_RNGCHK. It returns WALK_ABORT if one is found. It
4351 // returns WALK_SKIP_SUBTREES if GTF_EXCEPT is not set (assumes flags
4352 // properly propagated to parent trees). It returns WALK_CONTINUE
4354 static fgWalkResult fgChkThrowCB(GenTreePtr* pTree, Compiler::fgWalkData* data);
4355 static fgWalkResult fgChkLocAllocCB(GenTreePtr* pTree, Compiler::fgWalkData* data);
4356 static fgWalkResult fgChkQmarkCB(GenTreePtr* pTree, Compiler::fgWalkData* data);
4358 /**************************************************************************
4360 *************************************************************************/
4363 friend class SsaBuilder;
4364 friend struct ValueNumberState;
4366 //--------------------- Detect the basic blocks ---------------------------
4368 BasicBlock** fgBBs; // Table of pointers to the BBs
4370 void fgInitBBLookup();
4371 BasicBlock* fgLookupBB(unsigned addr);
4373 void fgMarkJumpTarget(BYTE* jumpTarget, IL_OFFSET offs);
4375 void fgFindJumpTargets(const BYTE* codeAddr, IL_OFFSET codeSize, BYTE* jumpTarget);
4377 void fgMarkBackwardJump(BasicBlock* startBlock, BasicBlock* endBlock);
4379 void fgLinkBasicBlocks();
4381 unsigned fgMakeBasicBlocks(const BYTE* codeAddr, IL_OFFSET codeSize, BYTE* jumpTarget);
4383 void fgCheckBasicBlockControlFlow();
4385 void fgControlFlowPermitted(BasicBlock* blkSrc,
4386 BasicBlock* blkDest,
4387 BOOL IsLeave = false /* is the src a leave block */);
4389 bool fgFlowToFirstBlockOfInnerTry(BasicBlock* blkSrc, BasicBlock* blkDest, bool sibling);
4391 void fgObserveInlineConstants(OPCODE opcode, const FgStack& stack, bool isInlining);
4393 void fgAdjustForAddressExposedOrWrittenThis();
4395 bool fgProfileData_ILSizeMismatch;
4396 ICorJitInfo::ProfileBuffer* fgProfileBuffer;
4397 ULONG fgProfileBufferCount;
4398 ULONG fgNumProfileRuns;
4400 unsigned fgStressBBProf()
4403 unsigned result = JitConfig.JitStressBBProf();
4406 if (compStressCompile(STRESS_BB_PROFILE, 15))
4417 bool fgHaveProfileData();
4418 bool fgGetProfileWeightForBasicBlock(IL_OFFSET offset, unsigned* weight);
4420 bool fgIsUsingProfileWeights()
4422 return (fgHaveProfileData() || fgStressBBProf());
4424 void fgInstrumentMethod();
4426 //-------- Insert a statement at the start or end of a basic block --------
4430 static bool fgBlockContainsStatementBounded(BasicBlock* block, GenTree* stmt, bool answerOnBoundExceeded = true);
4434 GenTreeStmt* fgInsertStmtAtEnd(BasicBlock* block, GenTreePtr node);
4436 public: // Used by linear scan register allocation
4437 GenTreeStmt* fgInsertStmtNearEnd(BasicBlock* block, GenTreePtr node);
4440 GenTreePtr fgInsertStmtAtBeg(BasicBlock* block, GenTreePtr stmt);
4441 GenTreePtr fgInsertStmtAfter(BasicBlock* block, GenTreePtr insertionPoint, GenTreePtr stmt);
4443 public: // Used by linear scan register allocation
4444 GenTreePtr fgInsertStmtBefore(BasicBlock* block, GenTreePtr insertionPoint, GenTreePtr stmt);
4447 GenTreePtr fgInsertStmtListAfter(BasicBlock* block, GenTreePtr stmtAfter, GenTreePtr stmtList);
4449 GenTreePtr fgMorphSplitTree(GenTree** splitPoint, GenTree* stmt, BasicBlock* blk);
4451 // Create a new temporary variable to hold the result of *ppTree,
4452 // and transform the graph accordingly.
4453 GenTree* fgInsertCommaFormTemp(GenTree** ppTree, CORINFO_CLASS_HANDLE structType = nullptr);
4454 GenTree* fgMakeMultiUse(GenTree** ppTree);
4457 // Recognize a bitwise rotation pattern and convert into a GT_ROL or a GT_ROR node.
4458 GenTreePtr fgRecognizeAndMorphBitwiseRotation(GenTreePtr tree);
4459 bool fgOperIsBitwiseRotationRoot(genTreeOps oper);
4461 //-------- Determine the order in which the trees will be evaluated -------
4463 unsigned fgTreeSeqNum;
4464 GenTree* fgTreeSeqLst;
4465 GenTree* fgTreeSeqBeg;
4467 GenTree* fgSetTreeSeq(GenTree* tree, GenTree* prev = nullptr, bool isLIR = false);
4468 void fgSetTreeSeqHelper(GenTree* tree, bool isLIR);
4469 void fgSetTreeSeqFinish(GenTreePtr tree, bool isLIR);
4470 void fgSetStmtSeq(GenTree* tree);
4471 void fgSetBlockOrder(BasicBlock* block);
4473 //------------------------- Morphing --------------------------------------
4475 unsigned fgPtrArgCntCur;
4476 unsigned fgPtrArgCntMax;
4477 hashBv* fgOutgoingArgTemps;
4478 hashBv* fgCurrentlyInUseArgTemps;
4480 bool compCanEncodePtrArgCntMax();
4482 void fgSetRngChkTarget(GenTreePtr tree, bool delay = true);
4485 void fgMoveOpsLeft(GenTreePtr tree);
4488 bool fgIsCommaThrow(GenTreePtr tree, bool forFolding = false);
4490 bool fgIsThrow(GenTreePtr tree);
4492 bool fgInDifferentRegions(BasicBlock* blk1, BasicBlock* blk2);
4493 bool fgIsBlockCold(BasicBlock* block);
4495 GenTreePtr fgMorphCastIntoHelper(GenTreePtr tree, int helper, GenTreePtr oper);
4497 GenTreePtr fgMorphIntoHelperCall(GenTreePtr tree, int helper, GenTreeArgList* args);
4499 GenTreePtr fgMorphStackArgForVarArgs(unsigned lclNum, var_types varType, unsigned lclOffs);
4501 bool fgMorphRelopToQmark(GenTreePtr tree);
4503 // A "MorphAddrContext" carries information from the surrounding context. If we are evaluating a byref address,
4504 // it is useful to know whether the address will be immediately dereferenced, or whether the address value will
4505 // be used, perhaps by passing it as an argument to a called method. This affects how null checking is done:
4506 // for sufficiently small offsets, we can rely on OS page protection to implicitly null-check addresses that we
4507 // know will be dereferenced. To know that reliance on implicit null checking is sound, we must further know that
4508 // all offsets between the top-level indirection and the bottom are constant, and that their sum is sufficiently
4509 // small; hence the other fields of MorphAddrContext.
4510 enum MorphAddrContextKind
4515 struct MorphAddrContext
4517 MorphAddrContextKind m_kind;
4518 bool m_allConstantOffsets; // Valid only for "m_kind == MACK_Ind". True iff all offsets between
4519 // top-level indirection and here have been constants.
4520 size_t m_totalOffset; // Valid only for "m_kind == MACK_Ind", and if "m_allConstantOffsets" is true.
4521 // In that case, is the sum of those constant offsets.
4523 MorphAddrContext(MorphAddrContextKind kind) : m_kind(kind), m_allConstantOffsets(true), m_totalOffset(0)
4528 // A MACK_CopyBlock context is immutable, so we can just make one of these and share it.
4529 static MorphAddrContext s_CopyBlockMAC;
4532 GenTreePtr fgCopySIMDNode(GenTreeSIMD* simdNode);
4533 GenTreePtr getSIMDStructFromField(GenTreePtr tree,
4534 var_types* baseTypeOut,
4536 unsigned* simdSizeOut,
4537 bool ignoreUsedInSIMDIntrinsic = false);
4538 GenTreePtr fgMorphFieldAssignToSIMDIntrinsicSet(GenTreePtr tree);
4539 GenTreePtr fgMorphFieldToSIMDIntrinsicGet(GenTreePtr tree);
4540 bool fgMorphCombineSIMDFieldAssignments(BasicBlock* block, GenTreePtr stmt);
4541 void impMarkContiguousSIMDFieldAssignments(GenTreePtr stmt);
4543 // fgPreviousCandidateSIMDFieldAsgStmt is only used for tracking previous simd field assignment
4544 // in function: Complier::impMarkContiguousSIMDFieldAssignments.
4545 GenTreePtr fgPreviousCandidateSIMDFieldAsgStmt;
4547 #endif // FEATURE_SIMD
4548 GenTreePtr fgMorphArrayIndex(GenTreePtr tree);
4549 GenTreePtr fgMorphCast(GenTreePtr tree);
4550 GenTreePtr fgUnwrapProxy(GenTreePtr objRef);
4551 GenTreeCall* fgMorphArgs(GenTreeCall* call);
4553 void fgMakeOutgoingStructArgCopy(GenTreeCall* call,
4556 CORINFO_CLASS_HANDLE copyBlkClass FEATURE_UNIX_AMD64_STRUCT_PASSING_ONLY_ARG(
4557 const SYSTEMV_AMD64_CORINFO_STRUCT_REG_PASSING_DESCRIPTOR* structDescPtr));
4559 void fgFixupStructReturn(GenTreePtr call);
4560 GenTreePtr fgMorphLocalVar(GenTreePtr tree);
4561 bool fgAddrCouldBeNull(GenTreePtr addr);
4562 GenTreePtr fgMorphField(GenTreePtr tree, MorphAddrContext* mac);
4563 bool fgCanFastTailCall(GenTreeCall* call);
4564 void fgMorphTailCall(GenTreeCall* call);
4565 void fgMorphRecursiveFastTailCallIntoLoop(BasicBlock* block, GenTreeCall* recursiveTailCall);
4566 GenTreePtr fgAssignRecursiveCallArgToCallerParam(GenTreePtr arg,
4567 fgArgTabEntryPtr argTabEntry,
4569 IL_OFFSETX callILOffset,
4570 GenTreePtr tmpAssignmentInsertionPoint,
4571 GenTreePtr paramAssignmentInsertionPoint);
4572 static int fgEstimateCallStackSize(GenTreeCall* call);
4573 GenTreePtr fgMorphCall(GenTreeCall* call);
4574 void fgMorphCallInline(GenTreeCall* call, InlineResult* result);
4575 void fgMorphCallInlineHelper(GenTreeCall* call, InlineResult* result);
4577 void fgNoteNonInlineCandidate(GenTreeStmt* stmt, GenTreeCall* call);
4578 static fgWalkPreFn fgFindNonInlineCandidate;
4580 GenTreePtr fgOptimizeDelegateConstructor(GenTreePtr call, CORINFO_CONTEXT_HANDLE* ExactContextHnd);
4581 GenTreePtr fgMorphLeaf(GenTreePtr tree);
4582 void fgAssignSetVarDef(GenTreePtr tree);
4583 GenTreePtr fgMorphOneAsgBlockOp(GenTreePtr tree);
4584 GenTreePtr fgMorphInitBlock(GenTreePtr tree);
4585 GenTreePtr fgMorphBlkToInd(GenTreeBlk* tree, var_types type);
4586 GenTreePtr fgMorphGetStructAddr(GenTreePtr* pTree, CORINFO_CLASS_HANDLE clsHnd, bool isRValue = false);
4587 GenTreePtr fgMorphBlkNode(GenTreePtr tree, bool isDest);
4588 GenTreePtr fgMorphBlockOperand(GenTreePtr tree, var_types asgType, unsigned blockWidth, bool isDest);
4589 void fgMorphUnsafeBlk(GenTreeObj* obj);
4590 GenTreePtr fgMorphCopyBlock(GenTreePtr tree);
4591 GenTreePtr fgMorphForRegisterFP(GenTreePtr tree);
4592 GenTreePtr fgMorphSmpOp(GenTreePtr tree, MorphAddrContext* mac = nullptr);
4593 GenTreePtr fgMorphSmpOpPre(GenTreePtr tree);
4594 GenTreePtr fgMorphModToSubMulDiv(GenTreeOp* tree);
4595 GenTreePtr fgMorphSmpOpOptional(GenTreeOp* tree);
4596 GenTreePtr fgMorphRecognizeBoxNullable(GenTree* compare);
4598 GenTreePtr fgMorphToEmulatedFP(GenTreePtr tree);
4599 GenTreePtr fgMorphConst(GenTreePtr tree);
4602 GenTreePtr fgMorphTree(GenTreePtr tree, MorphAddrContext* mac = nullptr);
4605 #if LOCAL_ASSERTION_PROP
4606 void fgKillDependentAssertionsSingle(unsigned lclNum DEBUGARG(GenTreePtr tree));
4607 void fgKillDependentAssertions(unsigned lclNum DEBUGARG(GenTreePtr tree));
4609 void fgMorphTreeDone(GenTreePtr tree, GenTreePtr oldTree = nullptr DEBUGARG(int morphNum = 0));
4611 GenTreeStmt* fgMorphStmt;
4613 unsigned fgGetBigOffsetMorphingTemp(var_types type); // We cache one temp per type to be
4614 // used when morphing big offset.
4616 //----------------------- Liveness analysis -------------------------------
4618 VARSET_TP fgCurUseSet; // vars used by block (before an assignment)
4619 VARSET_TP fgCurDefSet; // vars assigned by block (before a use)
4621 bool fgCurHeapUse; // True iff the current basic block uses the heap before defining it.
4622 bool fgCurHeapDef; // True iff the current basic block defines the heap.
4623 bool fgCurHeapHavoc; // True if the current basic block is known to set the heap to a "havoc" value.
4625 void fgMarkUseDef(GenTreeLclVarCommon* tree);
4627 void fgBeginScopeLife(VARSET_TP* inScope, VarScopeDsc* var);
4628 void fgEndScopeLife(VARSET_TP* inScope, VarScopeDsc* var);
4630 void fgMarkInScope(BasicBlock* block, VARSET_VALARG_TP inScope);
4631 void fgUnmarkInScope(BasicBlock* block, VARSET_VALARG_TP unmarkScope);
4633 void fgExtendDbgScopes();
4634 void fgExtendDbgLifetimes();
4637 void fgDispDebugScopes();
4640 //-------------------------------------------------------------------------
4642 // The following keeps track of any code we've added for things like array
4643 // range checking or explicit calls to enable GC, and so on.
4648 AddCodeDsc* acdNext;
4649 BasicBlock* acdDstBlk; // block to which we jump
4651 SpecialCodeKind acdKind; // what kind of a special block is this?
4652 unsigned short acdStkLvl;
4656 static unsigned acdHelper(SpecialCodeKind codeKind);
4658 AddCodeDsc* fgAddCodeList;
4660 bool fgRngChkThrowAdded;
4661 AddCodeDsc* fgExcptnTargetCache[SCK_COUNT];
4663 BasicBlock* fgRngChkTarget(BasicBlock* block, unsigned stkDepth, SpecialCodeKind kind);
4665 BasicBlock* fgAddCodeRef(BasicBlock* srcBlk, unsigned refData, SpecialCodeKind kind, unsigned stkDepth = 0);
4668 AddCodeDsc* fgFindExcptnTarget(SpecialCodeKind kind, unsigned refData);
4671 bool fgIsCodeAdded();
4673 bool fgIsThrowHlpBlk(BasicBlock* block);
4674 unsigned fgThrowHlpBlkStkLevel(BasicBlock* block);
4676 unsigned fgBigOffsetMorphingTemps[TYP_COUNT];
4678 unsigned fgCheckInlineDepthAndRecursion(InlineInfo* inlineInfo);
4679 void fgInvokeInlineeCompiler(GenTreeCall* call, InlineResult* result);
4680 void fgInsertInlineeBlocks(InlineInfo* pInlineInfo);
4681 GenTreePtr fgInlinePrependStatements(InlineInfo* inlineInfo);
4682 void fgInlineAppendStatements(InlineInfo* inlineInfo, BasicBlock* block, GenTreePtr stmt);
4684 #if FEATURE_MULTIREG_RET
4685 GenTreePtr fgGetStructAsStructPtr(GenTreePtr tree);
4686 GenTreePtr fgAssignStructInlineeToVar(GenTreePtr child, CORINFO_CLASS_HANDLE retClsHnd);
4687 void fgAttachStructInlineeToAsg(GenTreePtr tree, GenTreePtr child, CORINFO_CLASS_HANDLE retClsHnd);
4688 #endif // FEATURE_MULTIREG_RET
4690 static fgWalkPreFn fgUpdateInlineReturnExpressionPlaceHolder;
4693 static fgWalkPreFn fgDebugCheckInlineCandidates;
4696 void fgPromoteStructs();
4697 fgWalkResult fgMorphStructField(GenTreePtr tree, fgWalkData* fgWalkPre);
4698 fgWalkResult fgMorphLocalField(GenTreePtr tree, fgWalkData* fgWalkPre);
4699 void fgMarkImplicitByRefArgs();
4700 bool fgMorphImplicitByRefArgs(GenTree** pTree, fgWalkData* fgWalkPre);
4701 static fgWalkPreFn fgMarkAddrTakenLocalsPreCB;
4702 static fgWalkPostFn fgMarkAddrTakenLocalsPostCB;
4703 void fgMarkAddressExposedLocals();
4704 bool fgNodesMayInterfere(GenTree* store, GenTree* load);
4706 // Returns true if the type of tree is of size at least "width", or if "tree" is not a
4708 bool fgFitsInOrNotLoc(GenTreePtr tree, unsigned width);
4710 // The given local variable, required to be a struct variable, is being assigned via
4711 // a "lclField", to make it masquerade as an integral type in the ABI. Make sure that
4712 // the variable is not enregistered, and is therefore not promoted independently.
4713 void fgLclFldAssign(unsigned lclNum);
4715 static fgWalkPreFn gtHasLocalsWithAddrOpCB;
4716 bool gtCanOptimizeTypeEquality(GenTreePtr tree);
4717 bool gtIsTypeHandleToRuntimeTypeHelper(GenTreePtr tree);
4718 bool gtIsActiveCSE_Candidate(GenTreePtr tree);
4721 bool fgPrintInlinedMethods;
4724 bool fgIsBigOffset(size_t offset);
4726 // The following are used when morphing special cases of integer div/mod operations and also by codegen
4727 bool fgIsSignedDivOptimizable(GenTreePtr divisor);
4728 bool fgIsUnsignedDivOptimizable(GenTreePtr divisor);
4729 bool fgIsSignedModOptimizable(GenTreePtr divisor);
4730 bool fgIsUnsignedModOptimizable(GenTreePtr divisor);
4733 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
4734 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
4738 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
4739 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
4746 LclVarDsc* optIsTrackedLocal(GenTreePtr tree);
4749 void optRemoveRangeCheck(
4750 GenTreePtr tree, GenTreePtr stmt, bool updateCSEcounts, unsigned sideEffFlags = 0, bool forceRemove = false);
4751 bool optIsRangeCheckRemovable(GenTreePtr tree);
4754 static fgWalkPreFn optValidRangeCheckIndex;
4755 static fgWalkPreFn optRemoveTreeVisitor; // Helper passed to Compiler::fgWalkAllTreesPre() to decrement the LclVar
4758 void optRemoveTree(GenTreePtr deadTree, GenTreePtr keepList);
4760 /**************************************************************************
4762 *************************************************************************/
4765 // Do hoisting for all loops.
4766 void optHoistLoopCode();
4768 // To represent sets of VN's that have already been hoisted in outer loops.
4769 typedef SimplerHashTable<ValueNum, SmallPrimitiveKeyFuncs<ValueNum>, bool, JitSimplerHashBehavior> VNToBoolMap;
4770 typedef VNToBoolMap VNSet;
4772 struct LoopHoistContext
4775 // The set of variables hoisted in the current loop (or nullptr if there are none).
4776 VNSet* m_pHoistedInCurLoop;
4779 // Value numbers of expressions that have been hoisted in parent loops in the loop nest.
4780 VNSet m_hoistedInParentLoops;
4781 // Value numbers of expressions that have been hoisted in the current (or most recent) loop in the nest.
4782 // Previous decisions on loop-invariance of value numbers in the current loop.
4783 VNToBoolMap m_curLoopVnInvariantCache;
4785 VNSet* GetHoistedInCurLoop(Compiler* comp)
4787 if (m_pHoistedInCurLoop == nullptr)
4789 m_pHoistedInCurLoop = new (comp->getAllocatorLoopHoist()) VNSet(comp->getAllocatorLoopHoist());
4791 return m_pHoistedInCurLoop;
4794 VNSet* ExtractHoistedInCurLoop()
4796 VNSet* res = m_pHoistedInCurLoop;
4797 m_pHoistedInCurLoop = nullptr;
4801 LoopHoistContext(Compiler* comp)
4802 : m_pHoistedInCurLoop(nullptr)
4803 , m_hoistedInParentLoops(comp->getAllocatorLoopHoist())
4804 , m_curLoopVnInvariantCache(comp->getAllocatorLoopHoist())
4809 // Do hoisting for loop "lnum" (an index into the optLoopTable), and all loops nested within it.
4810 // Tracks the expressions that have been hoisted by containing loops by temporary recording their
4811 // value numbers in "m_hoistedInParentLoops". This set is not modified by the call.
4812 void optHoistLoopNest(unsigned lnum, LoopHoistContext* hoistCtxt);
4814 // Do hoisting for a particular loop ("lnum" is an index into the optLoopTable.)
4815 // Assumes that expressions have been hoisted in containing loops if their value numbers are in
4816 // "m_hoistedInParentLoops".
4818 void optHoistThisLoop(unsigned lnum, LoopHoistContext* hoistCtxt);
4820 // Hoist all expressions in "blk" that are invariant in loop "lnum" (an index into the optLoopTable)
4821 // outside of that loop. Exempt expressions whose value number is in "m_hoistedInParentLoops"; add VN's of hoisted
4822 // expressions to "hoistInLoop".
4823 void optHoistLoopExprsForBlock(BasicBlock* blk, unsigned lnum, LoopHoistContext* hoistCtxt);
4825 // Return true if the tree looks profitable to hoist out of loop 'lnum'.
4826 bool optIsProfitableToHoistableTree(GenTreePtr tree, unsigned lnum);
4828 // Hoist all proper sub-expressions of "tree" (which occurs in "stmt", which occurs in "blk")
4829 // that are invariant in loop "lnum" (an index into the optLoopTable)
4830 // outside of that loop. Exempt expressions whose value number is in "hoistedInParents"; add VN's of hoisted
4831 // expressions to "hoistInLoop".
4832 // Returns "true" iff "tree" is loop-invariant (wrt "lnum").
4833 // Assumes that the value of "*firstBlockAndBeforeSideEffect" indicates that we're in the first block, and before
4834 // any possible globally visible side effects. Assume is called in evaluation order, and updates this.
4835 bool optHoistLoopExprsForTree(GenTreePtr tree,
4837 LoopHoistContext* hoistCtxt,
4838 bool* firstBlockAndBeforeSideEffect,
4841 // Performs the hoisting 'tree' into the PreHeader for loop 'lnum'
4842 void optHoistCandidate(GenTreePtr tree, unsigned lnum, LoopHoistContext* hoistCtxt);
4844 // Returns true iff the ValueNum "vn" represents a value that is loop-invariant in "lnum".
4845 // Constants and init values are always loop invariant.
4846 // VNPhi's connect VN's to the SSA definition, so we can know if the SSA def occurs in the loop.
4847 bool optVNIsLoopInvariant(ValueNum vn, unsigned lnum, VNToBoolMap* recordedVNs);
4849 // Returns "true" iff "tree" is valid at the head of loop "lnum", in the context of the hoist substitution
4850 // "subst". If "tree" is a local SSA var, it is valid if its SSA definition occurs outside of the loop, or
4851 // if it is in the domain of "subst" (meaning that it's definition has been previously hoisted, with a "standin"
4852 // local.) If tree is a constant, it is valid. Otherwise, if it is an operator, it is valid iff its children are.
4853 bool optTreeIsValidAtLoopHead(GenTreePtr tree, unsigned lnum);
4855 // If "blk" is the entry block of a natural loop, returns true and sets "*pLnum" to the index of the loop
4856 // in the loop table.
4857 bool optBlockIsLoopEntry(BasicBlock* blk, unsigned* pLnum);
4859 // Records the set of "side effects" of all loops: fields (object instance and static)
4860 // written to, and SZ-array element type equivalence classes updated.
4861 void optComputeLoopSideEffects();
4864 // Requires "lnum" to be the index of an outermost loop in the loop table. Traverses the body of that loop,
4865 // including all nested loops, and records the set of "side effects" of the loop: fields (object instance and
4866 // static) written to, and SZ-array element type equivalence classes updated.
4867 void optComputeLoopNestSideEffects(unsigned lnum);
4869 // Add the side effects of "blk" (which is required to be within a loop) to all loops of which it is a part.
4870 void optComputeLoopSideEffectsOfBlock(BasicBlock* blk);
4872 // Hoist the expression "expr" out of loop "lnum".
4873 void optPerformHoistExpr(GenTreePtr expr, unsigned lnum);
4876 void optOptimizeBools();
4879 GenTree* optIsBoolCond(GenTree* condBranch, GenTree** compPtr, bool* boolPtr);
4881 void optOptimizeBoolsGcStress(BasicBlock* condBlock);
4884 void optOptimizeLayout(); // Optimize the BasicBlock layout of the method
4886 void optOptimizeLoops(); // for "while-do" loops duplicates simple loop conditions and transforms
4887 // the loop into a "do-while" loop
4888 // Also finds all natural loops and records them in the loop table
4890 // Optionally clone loops in the loop table.
4891 void optCloneLoops();
4893 // Clone loop "loopInd" in the loop table.
4894 void optCloneLoop(unsigned loopInd, LoopCloneContext* context);
4896 // Ensure that loop "loopInd" has a unique head block. (If the existing entry has
4897 // non-loop predecessors other than the head entry, create a new, empty block that goes (only) to the entry,
4898 // and redirects the preds of the entry to this new block.) Sets the weight of the newly created block to
4900 void optEnsureUniqueHead(unsigned loopInd, unsigned ambientWeight);
4902 void optUnrollLoops(); // Unrolls loops (needs to have cost info)
4905 // This enumeration describes what is killed by a call.
4909 CALLINT_NONE, // no interference (most helpers)
4910 CALLINT_REF_INDIRS, // kills GC ref indirections (SETFIELD OBJ)
4911 CALLINT_SCL_INDIRS, // kills non GC ref indirections (SETFIELD non-OBJ)
4912 CALLINT_ALL_INDIRS, // kills both GC ref and non GC ref indirections (SETFIELD STRUCT)
4913 CALLINT_ALL, // kills everything (normal method call)
4917 // A "LoopDsc" describes a ("natural") loop. We (currently) require the body of a loop to be a contiguous (in
4918 // bbNext order) sequence of basic blocks. (At times, we may require the blocks in a loop to be "properly numbered"
4919 // in bbNext order; we use comparisons on the bbNum to decide order.)
4920 // The blocks that define the body are
4921 // first <= top <= entry <= bottom .
4922 // The "head" of the loop is a block outside the loop that has "entry" as a successor. We only support loops with a
4923 // single 'head' block. The meanings of these blocks are given in the definitions below. Also see the picture at
4924 // Compiler::optFindNaturalLoops().
4927 BasicBlock* lpHead; // HEAD of the loop (not part of the looping of the loop) -- has ENTRY as a successor.
4928 BasicBlock* lpFirst; // FIRST block (in bbNext order) reachable within this loop. (May be part of a nested
4929 // loop, but not the outer loop.)
4930 BasicBlock* lpTop; // loop TOP (the back edge from lpBottom reaches here) (in most cases FIRST and TOP are the
4932 BasicBlock* lpEntry; // the ENTRY in the loop (in most cases TOP or BOTTOM)
4933 BasicBlock* lpBottom; // loop BOTTOM (from here we have a back edge to the TOP)
4934 BasicBlock* lpExit; // if a single exit loop this is the EXIT (in most cases BOTTOM)
4936 callInterf lpAsgCall; // "callInterf" for calls in the loop
4937 ALLVARSET_TP lpAsgVars; // set of vars assigned within the loop (all vars, not just tracked)
4938 varRefKinds lpAsgInds : 8; // set of inds modified within the loop
4940 unsigned short lpFlags; // Mask of the LPFLG_* constants
4942 unsigned char lpExitCnt; // number of exits from the loop
4944 unsigned char lpParent; // The index of the most-nested loop that completely contains this one,
4945 // or else BasicBlock::NOT_IN_LOOP if no such loop exists.
4946 unsigned char lpChild; // The index of a nested loop, or else BasicBlock::NOT_IN_LOOP if no child exists.
4947 // (Actually, an "immediately" nested loop --
4948 // no other child of this loop is a parent of lpChild.)
4949 unsigned char lpSibling; // The index of another loop that is an immediate child of lpParent,
4950 // or else BasicBlock::NOT_IN_LOOP. One can enumerate all the children of a loop
4951 // by following "lpChild" then "lpSibling" links.
4953 #define LPFLG_DO_WHILE 0x0001 // it's a do-while loop (i.e ENTRY is at the TOP)
4954 #define LPFLG_ONE_EXIT 0x0002 // the loop has only one exit
4956 #define LPFLG_ITER 0x0004 // for (i = icon or lclVar; test_condition(); i++)
4957 #define LPFLG_HOISTABLE 0x0008 // the loop is in a form that is suitable for hoisting expressions
4958 #define LPFLG_CONST 0x0010 // for (i=icon;i<icon;i++){ ... } - constant loop
4960 #define LPFLG_VAR_INIT 0x0020 // iterator is initialized with a local var (var # found in lpVarInit)
4961 #define LPFLG_CONST_INIT 0x0040 // iterator is initialized with a constant (found in lpConstInit)
4963 #define LPFLG_VAR_LIMIT 0x0100 // iterator is compared with a local var (var # found in lpVarLimit)
4964 #define LPFLG_CONST_LIMIT 0x0200 // iterator is compared with a constant (found in lpConstLimit)
4965 #define LPFLG_ARRLEN_LIMIT 0x0400 // iterator is compared with a.len or a[i].len (found in lpArrLenLimit)
4966 #define LPFLG_SIMD_LIMIT 0x0080 // iterator is compared with Vector<T>.Count (found in lpConstLimit)
4968 #define LPFLG_HAS_PREHEAD 0x0800 // lpHead is known to be a preHead for this loop
4969 #define LPFLG_REMOVED 0x1000 // has been removed from the loop table (unrolled or optimized away)
4970 #define LPFLG_DONT_UNROLL 0x2000 // do not unroll this loop
4972 #define LPFLG_ASGVARS_YES 0x4000 // "lpAsgVars" has been computed
4973 #define LPFLG_ASGVARS_INC 0x8000 // "lpAsgVars" is incomplete -- vars beyond those representable in an AllVarSet
4974 // type are assigned to.
4976 bool lpLoopHasHeapHavoc; // The loop contains an operation that we assume has arbitrary heap side effects.
4977 // If this is set, the fields below may not be accurate (since they become irrelevant.)
4978 bool lpContainsCall; // True if executing the loop body *may* execute a call
4980 VARSET_TP lpVarInOut; // The set of variables that are IN or OUT during the execution of this loop
4981 VARSET_TP lpVarUseDef; // The set of variables that are USE or DEF during the execution of this loop
4983 int lpHoistedExprCount; // The register count for the non-FP expressions from inside this loop that have been
4985 int lpLoopVarCount; // The register count for the non-FP LclVars that are read/written inside this loop
4986 int lpVarInOutCount; // The register count for the non-FP LclVars that are alive inside or accross this loop
4988 int lpHoistedFPExprCount; // The register count for the FP expressions from inside this loop that have been
4990 int lpLoopVarFPCount; // The register count for the FP LclVars that are read/written inside this loop
4991 int lpVarInOutFPCount; // The register count for the FP LclVars that are alive inside or accross this loop
4993 typedef SimplerHashTable<CORINFO_FIELD_HANDLE,
4994 PtrKeyFuncs<struct CORINFO_FIELD_STRUCT_>,
4996 JitSimplerHashBehavior>
4998 FieldHandleSet* lpFieldsModified; // This has entries (mappings to "true") for all static field and object
4999 // instance fields modified
5002 typedef SimplerHashTable<CORINFO_CLASS_HANDLE,
5003 PtrKeyFuncs<struct CORINFO_CLASS_STRUCT_>,
5005 JitSimplerHashBehavior>
5007 ClassHandleSet* lpArrayElemTypesModified; // Bits set indicate the set of sz array element types such that
5008 // arrays of that type are modified
5011 // Adds the variable liveness information for 'blk' to 'this' LoopDsc
5012 void AddVariableLiveness(Compiler* comp, BasicBlock* blk);
5014 inline void AddModifiedField(Compiler* comp, CORINFO_FIELD_HANDLE fldHnd);
5015 // This doesn't *always* take a class handle -- it can also take primitive types, encoded as class handles
5016 // (shifted left, with a low-order bit set to distinguish.)
5017 // Use the {Encode/Decode}ElemType methods to construct/destruct these.
5018 inline void AddModifiedElemType(Compiler* comp, CORINFO_CLASS_HANDLE structHnd);
5020 /* The following values are set only for iterator loops, i.e. has the flag LPFLG_ITER set */
5022 GenTreePtr lpIterTree; // The "i <op>= const" tree
5023 unsigned lpIterVar(); // iterator variable #
5024 int lpIterConst(); // the constant with which the iterator is incremented
5025 genTreeOps lpIterOper(); // the type of the operation on the iterator (ASG_ADD, ASG_SUB, etc.)
5026 void VERIFY_lpIterTree();
5028 var_types lpIterOperType(); // For overflow instructions
5031 int lpConstInit; // initial constant value of iterator : Valid if LPFLG_CONST_INIT
5032 unsigned lpVarInit; // initial local var number to which we initialize the iterator : Valid if
5036 /* The following is for LPFLG_ITER loops only (i.e. the loop condition is "i RELOP const or var" */
5038 GenTreePtr lpTestTree; // pointer to the node containing the loop test
5039 genTreeOps lpTestOper(); // the type of the comparison between the iterator and the limit (GT_LE, GT_GE, etc.)
5040 void VERIFY_lpTestTree();
5042 bool lpIsReversed(); // true if the iterator node is the second operand in the loop condition
5043 GenTreePtr lpIterator(); // the iterator node in the loop test
5044 GenTreePtr lpLimit(); // the limit node in the loop test
5046 int lpConstLimit(); // limit constant value of iterator - loop condition is "i RELOP const" : Valid if
5047 // LPFLG_CONST_LIMIT
5048 unsigned lpVarLimit(); // the lclVar # in the loop condition ( "i RELOP lclVar" ) : Valid if
5050 bool lpArrLenLimit(Compiler* comp, ArrIndex* index); // The array length in the loop condition ( "i RELOP
5051 // arr.len" or "i RELOP arr[i][j].len" ) : Valid if
5052 // LPFLG_ARRLEN_LIMIT
5054 // Returns "true" iff "*this" contains the blk.
5055 bool lpContains(BasicBlock* blk)
5057 return lpFirst->bbNum <= blk->bbNum && blk->bbNum <= lpBottom->bbNum;
5059 // Returns "true" iff "*this" (properly) contains the range [first, bottom] (allowing firsts
5060 // to be equal, but requiring bottoms to be different.)
5061 bool lpContains(BasicBlock* first, BasicBlock* bottom)
5063 return lpFirst->bbNum <= first->bbNum && bottom->bbNum < lpBottom->bbNum;
5066 // Returns "true" iff "*this" (properly) contains "lp2" (allowing firsts to be equal, but requiring
5067 // bottoms to be different.)
5068 bool lpContains(const LoopDsc& lp2)
5070 return lpContains(lp2.lpFirst, lp2.lpBottom);
5073 // Returns "true" iff "*this" is (properly) contained by the range [first, bottom]
5074 // (allowing firsts to be equal, but requiring bottoms to be different.)
5075 bool lpContainedBy(BasicBlock* first, BasicBlock* bottom)
5077 return first->bbNum <= lpFirst->bbNum && lpBottom->bbNum < bottom->bbNum;
5080 // Returns "true" iff "*this" is (properly) contained by "lp2"
5081 // (allowing firsts to be equal, but requiring bottoms to be different.)
5082 bool lpContainedBy(const LoopDsc& lp2)
5084 return lpContains(lp2.lpFirst, lp2.lpBottom);
5087 // Returns "true" iff "*this" is disjoint from the range [top, bottom].
5088 bool lpDisjoint(BasicBlock* first, BasicBlock* bottom)
5090 return bottom->bbNum < lpFirst->bbNum || lpBottom->bbNum < first->bbNum;
5092 // Returns "true" iff "*this" is disjoint from "lp2".
5093 bool lpDisjoint(const LoopDsc& lp2)
5095 return lpDisjoint(lp2.lpFirst, lp2.lpBottom);
5097 // Returns "true" iff the loop is well-formed (see code for defn).
5100 return lpFirst->bbNum <= lpTop->bbNum && lpTop->bbNum <= lpEntry->bbNum &&
5101 lpEntry->bbNum <= lpBottom->bbNum &&
5102 (lpHead->bbNum < lpTop->bbNum || lpHead->bbNum > lpBottom->bbNum);
5107 bool fgMightHaveLoop(); // returns true if there are any backedges
5108 bool fgHasLoops; // True if this method has any loops, set in fgComputeReachability
5111 LoopDsc optLoopTable[MAX_LOOP_NUM]; // loop descriptor table
5112 unsigned char optLoopCount; // number of tracked loops
5115 unsigned optCallCount; // number of calls made in the method
5116 unsigned optIndirectCallCount; // number of virtual, interface and indirect calls made in the method
5117 unsigned optNativeCallCount; // number of Pinvoke/Native calls made in the method
5118 unsigned optLoopsCloned; // number of loops cloned in the current method.
5121 unsigned optFindLoopNumberFromBeginBlock(BasicBlock* begBlk);
5122 void optPrintLoopInfo(unsigned loopNum,
5124 BasicBlock* lpFirst,
5126 BasicBlock* lpEntry,
5127 BasicBlock* lpBottom,
5128 unsigned char lpExitCnt,
5130 unsigned parentLoop = BasicBlock::NOT_IN_LOOP);
5131 void optPrintLoopInfo(unsigned lnum);
5132 void optPrintLoopRecording(unsigned lnum);
5134 void optCheckPreds();
5137 void optSetBlockWeights();
5139 void optMarkLoopBlocks(BasicBlock* begBlk, BasicBlock* endBlk, bool excludeEndBlk);
5141 void optUnmarkLoopBlocks(BasicBlock* begBlk, BasicBlock* endBlk);
5143 void optUpdateLoopsBeforeRemoveBlock(BasicBlock* block, bool skipUnmarkLoop = false);
5145 bool optIsLoopTestEvalIntoTemp(GenTreePtr test, GenTreePtr* newTest);
5146 unsigned optIsLoopIncrTree(GenTreePtr incr);
5147 bool optCheckIterInLoopTest(unsigned loopInd, GenTreePtr test, BasicBlock* from, BasicBlock* to, unsigned iterVar);
5148 bool optComputeIterInfo(GenTreePtr incr, BasicBlock* from, BasicBlock* to, unsigned* pIterVar);
5149 bool optPopulateInitInfo(unsigned loopInd, GenTreePtr init, unsigned iterVar);
5150 bool optExtractInitTestIncr(BasicBlock* head,
5155 GenTreePtr* ppIncr);
5157 void optRecordLoop(BasicBlock* head,
5163 unsigned char exitCnt);
5165 void optFindNaturalLoops();
5167 // Ensures that all the loops in the loop nest rooted at "loopInd" (an index into the loop table) are 'canonical' --
5168 // each loop has a unique "top." Returns "true" iff the flowgraph has been modified.
5169 bool optCanonicalizeLoopNest(unsigned char loopInd);
5171 // Ensures that the loop "loopInd" (an index into the loop table) is 'canonical' -- it has a unique "top,"
5172 // unshared with any other loop. Returns "true" iff the flowgraph has been modified
5173 bool optCanonicalizeLoop(unsigned char loopInd);
5175 // Requires "l1" to be a valid loop table index, and not "BasicBlock::NOT_IN_LOOP". Requires "l2" to be
5176 // a valid loop table index, or else "BasicBlock::NOT_IN_LOOP". Returns true
5177 // iff "l2" is not NOT_IN_LOOP, and "l1" contains "l2".
5178 bool optLoopContains(unsigned l1, unsigned l2);
5180 // Requires "loopInd" to be a valid index into the loop table.
5181 // Updates the loop table by changing loop "loopInd", whose head is required
5182 // to be "from", to be "to". Also performs this transformation for any
5183 // loop nested in "loopInd" that shares the same head as "loopInd".
5184 void optUpdateLoopHead(unsigned loopInd, BasicBlock* from, BasicBlock* to);
5186 // Updates the successors of "blk": if "blk2" is a successor of "blk", and there is a mapping for "blk2->blk3" in
5187 // "redirectMap", change "blk" so that "blk3" is this successor. Note that the predecessor lists are not updated.
5188 void optRedirectBlock(BasicBlock* blk, BlockToBlockMap* redirectMap);
5190 // Marks the containsCall information to "lnum" and any parent loops.
5191 void AddContainsCallAllContainingLoops(unsigned lnum);
5192 // Adds the variable liveness information from 'blk' to "lnum" and any parent loops.
5193 void AddVariableLivenessAllContainingLoops(unsigned lnum, BasicBlock* blk);
5194 // Adds "fldHnd" to the set of modified fields of "lnum" and any parent loops.
5195 void AddModifiedFieldAllContainingLoops(unsigned lnum, CORINFO_FIELD_HANDLE fldHnd);
5196 // Adds "elemType" to the set of modified array element types of "lnum" and any parent loops.
5197 void AddModifiedElemTypeAllContainingLoops(unsigned lnum, CORINFO_CLASS_HANDLE elemType);
5199 // Requires that "from" and "to" have the same "bbJumpKind" (perhaps because "to" is a clone
5200 // of "from".) Copies the jump destination from "from" to "to".
5201 void optCopyBlkDest(BasicBlock* from, BasicBlock* to);
5203 // The depth of the loop described by "lnum" (an index into the loop table.) (0 == top level)
5204 unsigned optLoopDepth(unsigned lnum)
5206 unsigned par = optLoopTable[lnum].lpParent;
5207 if (par == BasicBlock::NOT_IN_LOOP)
5213 return 1 + optLoopDepth(par);
5217 void fgOptWhileLoop(BasicBlock* block);
5219 bool optComputeLoopRep(int constInit,
5222 genTreeOps iterOper,
5224 genTreeOps testOper,
5227 unsigned* iterCount);
5228 #if FEATURE_STACK_FP_X87
5231 VARSET_TP optAllFloatVars; // mask of all tracked FP variables
5232 VARSET_TP optAllFPregVars; // mask of all enregistered FP variables
5233 VARSET_TP optAllNonFPvars; // mask of all tracked non-FP variables
5234 #endif // FEATURE_STACK_FP_X87
5237 static fgWalkPreFn optIsVarAssgCB;
5240 bool optIsVarAssigned(BasicBlock* beg, BasicBlock* end, GenTreePtr skip, unsigned var);
5242 bool optIsVarAssgLoop(unsigned lnum, unsigned var);
5244 int optIsSetAssgLoop(unsigned lnum, ALLVARSET_VALARG_TP vars, varRefKinds inds = VR_NONE);
5246 bool optNarrowTree(GenTreePtr tree, var_types srct, var_types dstt, ValueNumPair vnpNarrow, bool doit);
5248 /**************************************************************************
5249 * Optimization conditions
5250 *************************************************************************/
5252 bool optFastCodeOrBlendedLoop(BasicBlock::weight_t bbWeight);
5253 bool optPentium4(void);
5254 bool optAvoidIncDec(BasicBlock::weight_t bbWeight);
5255 bool optAvoidIntMult(void);
5260 // The following is the upper limit on how many expressions we'll keep track
5261 // of for the CSE analysis.
5263 static const unsigned MAX_CSE_CNT = EXPSET_SZ;
5265 static const int MIN_CSE_COST = 2;
5267 // Keeps tracked cse indices
5268 BitVecTraits* cseTraits;
5272 /* Generic list of nodes - used by the CSE logic */
5280 typedef struct treeLst* treeLstPtr;
5284 treeStmtLst* tslNext;
5285 GenTreePtr tslTree; // tree node
5286 GenTreePtr tslStmt; // statement containing the tree
5287 BasicBlock* tslBlock; // block containing the statement
5290 typedef struct treeStmtLst* treeStmtLstPtr;
5292 // The following logic keeps track of expressions via a simple hash table.
5296 CSEdsc* csdNextInBucket; // used by the hash table
5298 unsigned csdHashValue; // the orginal hashkey
5300 unsigned csdIndex; // 1..optCSECandidateCount
5301 char csdLiveAcrossCall; // 0 or 1
5303 unsigned short csdDefCount; // definition count
5304 unsigned short csdUseCount; // use count (excluding the implicit uses at defs)
5306 unsigned csdDefWtCnt; // weighted def count
5307 unsigned csdUseWtCnt; // weighted use count (excluding the implicit uses at defs)
5309 GenTreePtr csdTree; // treenode containing the 1st occurance
5310 GenTreePtr csdStmt; // stmt containing the 1st occurance
5311 BasicBlock* csdBlock; // block containing the 1st occurance
5313 treeStmtLstPtr csdTreeList; // list of matching tree nodes: head
5314 treeStmtLstPtr csdTreeLast; // list of matching tree nodes: tail
5317 static const size_t s_optCSEhashSize;
5318 CSEdsc** optCSEhash;
5323 CSEdsc* optCSEfindDsc(unsigned index);
5324 void optUnmarkCSE(GenTreePtr tree);
5326 // user defined callback data for the tree walk function optCSE_MaskHelper()
5327 struct optCSE_MaskData
5329 EXPSET_TP CSE_defMask;
5330 EXPSET_TP CSE_useMask;
5333 // Treewalk helper for optCSE_DefMask and optCSE_UseMask
5334 static fgWalkPreFn optCSE_MaskHelper;
5336 // This function walks all the node for an given tree
5337 // and return the mask of CSE definitions and uses for the tree
5339 void optCSE_GetMaskData(GenTreePtr tree, optCSE_MaskData* pMaskData);
5341 // Given a binary tree node return true if it is safe to swap the order of evaluation for op1 and op2.
5342 bool optCSE_canSwap(GenTree* firstNode, GenTree* secondNode);
5343 bool optCSE_canSwap(GenTree* tree);
5345 static fgWalkPostFn optPropagateNonCSE;
5346 static fgWalkPreFn optHasNonCSEChild;
5348 static fgWalkPreFn optUnmarkCSEs;
5350 static int __cdecl optCSEcostCmpEx(const void* op1, const void* op2);
5351 static int __cdecl optCSEcostCmpSz(const void* op1, const void* op2);
5353 void optCleanupCSEs();
5356 void optEnsureClearCSEInfo();
5359 #endif // FEATURE_ANYCSE
5361 #if FEATURE_VALNUM_CSE
5362 /**************************************************************************
5363 * Value Number based CSEs
5364 *************************************************************************/
5367 void optOptimizeValnumCSEs();
5370 void optValnumCSE_Init();
5371 unsigned optValnumCSE_Index(GenTreePtr tree, GenTreePtr stmt);
5372 unsigned optValnumCSE_Locate();
5373 void optValnumCSE_InitDataFlow();
5374 void optValnumCSE_DataFlow();
5375 void optValnumCSE_Availablity();
5376 void optValnumCSE_Heuristic();
5377 void optValnumCSE_UnmarkCSEs(GenTreePtr deadTree, GenTreePtr keepList);
5379 #endif // FEATURE_VALNUM_CSE
5382 bool optDoCSE; // True when we have found a duplicate CSE tree
5383 bool optValnumCSE_phase; // True when we are executing the optValnumCSE_phase
5384 unsigned optCSECandidateTotal; // Grand total of CSE candidates for both Lexical and ValNum
5385 unsigned optCSECandidateCount; // Count of CSE's candidates, reset for Lexical and ValNum CSE's
5386 unsigned optCSEstart; // The first local variable number that is a CSE
5387 unsigned optCSEcount; // The total count of CSE's introduced.
5388 unsigned optCSEweight; // The weight of the current block when we are
5389 // scanning for CSE expressions
5391 bool optIsCSEcandidate(GenTreePtr tree);
5393 // lclNumIsTrueCSE returns true if the LclVar was introduced by the CSE phase of the compiler
5395 bool lclNumIsTrueCSE(unsigned lclNum) const
5397 return ((optCSEcount > 0) && (lclNum >= optCSEstart) && (lclNum < optCSEstart + optCSEcount));
5400 // lclNumIsCSE returns true if the LclVar should be treated like a CSE with regards to constant prop.
5402 bool lclNumIsCSE(unsigned lclNum) const
5404 return lvaTable[lclNum].lvIsCSE;
5408 bool optConfigDisableCSE();
5409 bool optConfigDisableCSE2();
5411 void optOptimizeCSEs();
5413 #endif // FEATURE_ANYCSE
5421 unsigned ivaVar; // Variable we are interested in, or -1
5422 ALLVARSET_TP ivaMaskVal; // Set of variables assigned to. This is a set of all vars, not tracked vars.
5423 bool ivaMaskIncomplete; // Variables not representable in ivaMaskVal were assigned to.
5424 varRefKinds ivaMaskInd; // What kind of indirect assignments are there?
5425 callInterf ivaMaskCall; // What kind of calls are there?
5428 static callInterf optCallInterf(GenTreePtr call);
5431 // VN based copy propagation.
5432 typedef ArrayStack<GenTreePtr> GenTreePtrStack;
5433 typedef SimplerHashTable<unsigned, SmallPrimitiveKeyFuncs<unsigned>, GenTreePtrStack*, JitSimplerHashBehavior>
5434 LclNumToGenTreePtrStack;
5436 // Kill set to track variables with intervening definitions.
5437 VARSET_TP optCopyPropKillSet;
5439 // Copy propagation functions.
5440 void optCopyProp(BasicBlock* block, GenTreePtr stmt, GenTreePtr tree, LclNumToGenTreePtrStack* curSsaName);
5441 void optBlockCopyPropPopStacks(BasicBlock* block, LclNumToGenTreePtrStack* curSsaName);
5442 void optBlockCopyProp(BasicBlock* block, LclNumToGenTreePtrStack* curSsaName);
5443 bool optIsSsaLocal(GenTreePtr tree);
5444 int optCopyProp_LclVarScore(LclVarDsc* lclVarDsc, LclVarDsc* copyVarDsc, bool preferOp2);
5445 void optVnCopyProp();
5447 /**************************************************************************
5448 * Early value propagation
5449 *************************************************************************/
5455 SSAName(unsigned lvNum, unsigned ssaNum) : m_lvNum(lvNum), m_ssaNum(ssaNum)
5459 static unsigned GetHashCode(SSAName ssaNm)
5461 return (ssaNm.m_lvNum << 16) | (ssaNm.m_ssaNum);
5464 static bool Equals(SSAName ssaNm1, SSAName ssaNm2)
5466 return (ssaNm1.m_lvNum == ssaNm2.m_lvNum) && (ssaNm1.m_ssaNum == ssaNm2.m_ssaNum);
5470 #define OMF_HAS_NEWARRAY 0x00000001 // Method contains 'new' of an array
5471 #define OMF_HAS_NEWOBJ 0x00000002 // Method contains 'new' of an object type.
5472 #define OMF_HAS_ARRAYREF 0x00000004 // Method contains array element loads or stores.
5473 #define OMF_HAS_VTABLEREF 0x00000008 // Method contains method table reference.
5474 #define OMF_HAS_NULLCHECK 0x00000010 // Method contains null check.
5476 unsigned optMethodFlags;
5478 // Recursion bound controls how far we can go backwards tracking for a SSA value.
5479 // No throughput diff was found with backward walk bound between 3-8.
5480 static const int optEarlyPropRecurBound = 5;
5482 enum class optPropKind
5490 bool gtIsVtableRef(GenTreePtr tree);
5491 GenTreePtr getArrayLengthFromAllocation(GenTreePtr tree);
5492 GenTreePtr getObjectHandleNodeFromAllocation(GenTreePtr tree);
5493 GenTreePtr optPropGetValueRec(unsigned lclNum, unsigned ssaNum, optPropKind valueKind, int walkDepth);
5494 GenTreePtr optPropGetValue(unsigned lclNum, unsigned ssaNum, optPropKind valueKind);
5495 bool optEarlyPropRewriteTree(GenTreePtr tree);
5496 bool optDoEarlyPropForBlock(BasicBlock* block);
5497 bool optDoEarlyPropForFunc();
5498 void optEarlyProp();
5499 void optFoldNullCheck(GenTreePtr tree);
5500 bool optCanMoveNullCheckPastTree(GenTreePtr tree, bool isInsideTry);
5503 /**************************************************************************
5504 * Value/Assertion propagation
5505 *************************************************************************/
5507 // Data structures for assertion prop
5508 BitVecTraits* apTraits;
5512 enum optAssertionKind
5527 O1K_ARRLEN_OPER_BND,
5528 O1K_ARRLEN_LOOP_BND,
5529 O1K_CONSTANT_LOOP_BND,
5550 optAssertionKind assertionKind;
5553 unsigned lclNum; // assigned to or property of this local var number
5561 struct AssertionDscOp1
5563 optOp1Kind kind; // a normal LclVar, or Exact-type or Subtype
5570 struct AssertionDscOp2
5572 optOp2Kind kind; // a const or copy assignment
5576 ssize_t iconVal; // integer
5577 unsigned iconFlags; // gtFlags
5579 struct Range // integer subrange
5593 bool IsArrLenArithBound()
5595 return ((assertionKind == OAK_EQUAL || assertionKind == OAK_NOT_EQUAL) && op1.kind == O1K_ARRLEN_OPER_BND);
5597 bool IsArrLenBound()
5599 return ((assertionKind == OAK_EQUAL || assertionKind == OAK_NOT_EQUAL) && op1.kind == O1K_ARRLEN_LOOP_BND);
5601 bool IsConstantBound()
5603 return ((assertionKind == OAK_EQUAL || assertionKind == OAK_NOT_EQUAL) &&
5604 op1.kind == O1K_CONSTANT_LOOP_BND);
5606 bool IsBoundsCheckNoThrow()
5608 return ((assertionKind == OAK_NO_THROW) && (op1.kind == O1K_ARR_BND));
5611 bool IsCopyAssertion()
5613 return ((assertionKind == OAK_EQUAL) && (op1.kind == O1K_LCLVAR) && (op2.kind == O2K_LCLVAR_COPY));
5616 static bool SameKind(AssertionDsc* a1, AssertionDsc* a2)
5618 return a1->assertionKind == a2->assertionKind && a1->op1.kind == a2->op1.kind &&
5619 a1->op2.kind == a2->op2.kind;
5622 static bool ComplementaryKind(optAssertionKind kind, optAssertionKind kind2)
5624 if (kind == OAK_EQUAL)
5626 return kind2 == OAK_NOT_EQUAL;
5628 else if (kind == OAK_NOT_EQUAL)
5630 return kind2 == OAK_EQUAL;
5635 static ssize_t GetLowerBoundForIntegralType(var_types type)
5655 static ssize_t GetUpperBoundForIntegralType(var_types type)
5679 bool HasSameOp1(AssertionDsc* that, bool vnBased)
5681 return (op1.kind == that->op1.kind) &&
5682 ((vnBased && (op1.vn == that->op1.vn)) || (!vnBased && (op1.lcl.lclNum == that->op1.lcl.lclNum)));
5685 bool HasSameOp2(AssertionDsc* that, bool vnBased)
5687 if (op2.kind != that->op2.kind)
5693 case O2K_IND_CNS_INT:
5695 return ((op2.u1.iconVal == that->op2.u1.iconVal) && (op2.u1.iconFlags == that->op2.u1.iconFlags));
5697 case O2K_CONST_LONG:
5698 return (op2.lconVal == that->op2.lconVal);
5700 case O2K_CONST_DOUBLE:
5701 // exact match because of positive and negative zero.
5702 return (memcmp(&op2.dconVal, &that->op2.dconVal, sizeof(double)) == 0);
5704 case O2K_LCLVAR_COPY:
5706 return (op2.lcl.lclNum == that->op2.lcl.lclNum) &&
5707 (!vnBased || op2.lcl.ssaNum == that->op2.lcl.ssaNum);
5710 return ((op2.u2.loBound == that->op2.u2.loBound) && (op2.u2.hiBound == that->op2.u2.hiBound));
5713 // we will return false
5717 assert(!"Unexpected value for op2.kind in AssertionDsc.");
5723 bool Complementary(AssertionDsc* that, bool vnBased)
5725 return ComplementaryKind(assertionKind, that->assertionKind) && HasSameOp1(that, vnBased) &&
5726 HasSameOp2(that, vnBased);
5729 bool Equals(AssertionDsc* that, bool vnBased)
5731 return (assertionKind == that->assertionKind) && HasSameOp1(that, vnBased) && HasSameOp2(that, vnBased);
5735 typedef unsigned short AssertionIndex;
5738 static fgWalkPreFn optAddCopiesCallback;
5739 static fgWalkPreFn optVNAssertionPropCurStmtVisitor;
5740 unsigned optAddCopyLclNum;
5741 GenTreePtr optAddCopyAsgnNode;
5743 bool optLocalAssertionProp; // indicates that we are performing local assertion prop
5744 bool optAssertionPropagated; // set to true if we modified the trees
5745 bool optAssertionPropagatedCurrentStmt;
5747 GenTreePtr optAssertionPropCurrentTree;
5749 AssertionIndex* optComplementaryAssertionMap;
5750 ExpandArray<ASSERT_TP>* optAssertionDep; // table that holds dependent assertions (assertions
5751 // using the value of a local var) for each local var
5752 AssertionDsc* optAssertionTabPrivate; // table that holds info about value assignments
5753 AssertionIndex optAssertionCount; // total number of assertions in the assertion table
5754 AssertionIndex optMaxAssertionCount;
5757 void optVnNonNullPropCurStmt(BasicBlock* block, GenTreePtr stmt, GenTreePtr tree);
5758 fgWalkResult optVNConstantPropCurStmt(BasicBlock* block, GenTreePtr stmt, GenTreePtr tree);
5759 GenTreePtr optVNConstantPropOnRelOp(GenTreePtr tree);
5760 GenTreePtr optVNConstantPropOnJTrue(BasicBlock* block, GenTreePtr stmt, GenTreePtr test);
5761 GenTreePtr optVNConstantPropOnTree(BasicBlock* block, GenTreePtr stmt, GenTreePtr tree);
5762 GenTreePtr optPrepareTreeForReplacement(GenTreePtr extractTree, GenTreePtr replaceTree);
5764 AssertionIndex GetAssertionCount()
5766 return optAssertionCount;
5768 ASSERT_TP* bbJtrueAssertionOut;
5769 typedef SimplerHashTable<ValueNum, SmallPrimitiveKeyFuncs<ValueNum>, ASSERT_TP, JitSimplerHashBehavior>
5770 ValueNumToAssertsMap;
5771 ValueNumToAssertsMap* optValueNumToAsserts;
5773 static const AssertionIndex NO_ASSERTION_INDEX = 0;
5775 // Assertion prop helpers.
5776 ASSERT_TP& GetAssertionDep(unsigned lclNum);
5777 AssertionDsc* optGetAssertion(AssertionIndex assertIndex);
5778 void optAssertionInit(bool isLocalProp);
5779 void optAssertionTraitsInit(AssertionIndex assertionCount);
5780 #if LOCAL_ASSERTION_PROP
5781 void optAssertionReset(AssertionIndex limit);
5782 void optAssertionRemove(AssertionIndex index);
5785 // Assertion prop data flow functions.
5786 void optAssertionPropMain();
5787 GenTreePtr optVNAssertionPropCurStmt(BasicBlock* block, GenTreePtr stmt);
5788 bool optIsTreeKnownIntValue(bool vnBased, GenTreePtr tree, ssize_t* pConstant, unsigned* pIconFlags);
5789 ASSERT_TP* optInitAssertionDataflowFlags();
5790 ASSERT_TP* optComputeAssertionGen();
5792 // Assertion Gen functions.
5793 void optAssertionGen(GenTreePtr tree);
5794 AssertionIndex optAssertionGenPhiDefn(GenTreePtr tree);
5795 AssertionIndex optCreateJTrueBoundsAssertion(GenTreePtr tree);
5796 AssertionIndex optAssertionGenJtrue(GenTreePtr tree);
5797 AssertionIndex optCreateJtrueAssertions(GenTreePtr op1, GenTreePtr op2, Compiler::optAssertionKind assertionKind);
5798 AssertionIndex optFindComplementary(AssertionIndex assertionIndex);
5799 void optMapComplementary(AssertionIndex assertionIndex, AssertionIndex index);
5801 // Assertion creation functions.
5802 AssertionIndex optCreateAssertion(GenTreePtr op1, GenTreePtr op2, optAssertionKind assertionKind);
5803 AssertionIndex optCreateAssertion(GenTreePtr op1,
5805 optAssertionKind assertionKind,
5806 AssertionDsc* assertion);
5807 void optCreateComplementaryAssertion(AssertionIndex assertionIndex, GenTreePtr op1, GenTreePtr op2);
5809 bool optAssertionVnInvolvesNan(AssertionDsc* assertion);
5810 AssertionIndex optAddAssertion(AssertionDsc* assertion);
5811 void optAddVnAssertionMapping(ValueNum vn, AssertionIndex index);
5813 void optPrintVnAssertionMapping();
5815 ASSERT_TP optGetVnMappedAssertions(ValueNum vn);
5817 // Used for respective assertion propagations.
5818 AssertionIndex optAssertionIsSubrange(GenTreePtr tree, var_types toType, ASSERT_VALARG_TP assertions);
5819 AssertionIndex optAssertionIsSubtype(GenTreePtr tree, GenTreePtr methodTableArg, ASSERT_VALARG_TP assertions);
5820 AssertionIndex optAssertionIsNonNullInternal(GenTreePtr op, ASSERT_VALARG_TP assertions);
5821 bool optAssertionIsNonNull(GenTreePtr op,
5822 ASSERT_VALARG_TP assertions DEBUGARG(bool* pVnBased) DEBUGARG(AssertionIndex* pIndex));
5824 // Used for Relop propagation.
5825 AssertionIndex optGlobalAssertionIsEqualOrNotEqual(ASSERT_VALARG_TP assertions, GenTreePtr op1, GenTreePtr op2);
5826 AssertionIndex optLocalAssertionIsEqualOrNotEqual(
5827 optOp1Kind op1Kind, unsigned lclNum, optOp2Kind op2Kind, ssize_t cnsVal, ASSERT_VALARG_TP assertions);
5829 // Assertion prop for lcl var functions.
5830 bool optAssertionProp_LclVarTypeCheck(GenTreePtr tree, LclVarDsc* lclVarDsc, LclVarDsc* copyVarDsc);
5831 GenTreePtr optCopyAssertionProp(AssertionDsc* curAssertion,
5833 GenTreePtr stmt DEBUGARG(AssertionIndex index));
5834 GenTreePtr optConstantAssertionProp(AssertionDsc* curAssertion,
5835 const GenTreePtr tree,
5836 const GenTreePtr stmt DEBUGARG(AssertionIndex index));
5837 GenTreePtr optVnConstantAssertionProp(const GenTreePtr tree, const GenTreePtr stmt);
5839 // Assertion propagation functions.
5840 GenTreePtr optAssertionProp(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
5841 GenTreePtr optAssertionProp_LclVar(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
5842 GenTreePtr optAssertionProp_Ind(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
5843 GenTreePtr optAssertionProp_Cast(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
5844 GenTreePtr optAssertionProp_Call(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
5845 GenTreePtr optAssertionProp_RelOp(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
5846 GenTreePtr optAssertionProp_Comma(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
5847 GenTreePtr optAssertionProp_BndsChk(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
5848 GenTreePtr optAssertionPropGlobal_RelOp(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
5849 GenTreePtr optAssertionPropLocal_RelOp(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
5850 GenTreePtr optAssertionProp_Update(const GenTreePtr newTree, const GenTreePtr tree, const GenTreePtr stmt);
5851 GenTreePtr optNonNullAssertionProp_Call(ASSERT_VALARG_TP assertions, const GenTreePtr tree, const GenTreePtr stmt);
5853 // Implied assertion functions.
5854 void optImpliedAssertions(AssertionIndex assertionIndex, ASSERT_TP& activeAssertions);
5855 void optImpliedByTypeOfAssertions(ASSERT_TP& activeAssertions);
5856 void optImpliedByCopyAssertion(AssertionDsc* copyAssertion, AssertionDsc* depAssertion, ASSERT_TP& result);
5857 void optImpliedByConstAssertion(AssertionDsc* curAssertion, ASSERT_TP& result);
5859 ASSERT_VALRET_TP optNewFullAssertSet();
5860 ASSERT_VALRET_TP optNewEmptyAssertSet();
5863 void optPrintAssertion(AssertionDsc* newAssertion, AssertionIndex assertionIndex = 0);
5864 void optDebugCheckAssertion(AssertionDsc* assertion);
5865 void optDebugCheckAssertions(AssertionIndex AssertionIndex);
5867 void optAddCopies();
5868 #endif // ASSERTION_PROP
5870 /**************************************************************************
5872 *************************************************************************/
5875 struct LoopCloneVisitorInfo
5877 LoopCloneContext* context;
5880 LoopCloneVisitorInfo(LoopCloneContext* context, unsigned loopNum, GenTreePtr stmt)
5881 : context(context), loopNum(loopNum), stmt(nullptr)
5886 bool optIsStackLocalInvariant(unsigned loopNum, unsigned lclNum);
5887 bool optExtractArrIndex(GenTreePtr tree, ArrIndex* result, unsigned lhsNum);
5888 bool optReconstructArrIndex(GenTreePtr tree, ArrIndex* result, unsigned lhsNum);
5889 bool optIdentifyLoopOptInfo(unsigned loopNum, LoopCloneContext* context);
5890 static fgWalkPreFn optCanOptimizeByLoopCloningVisitor;
5891 fgWalkResult optCanOptimizeByLoopCloning(GenTreePtr tree, LoopCloneVisitorInfo* info);
5892 void optObtainLoopCloningOpts(LoopCloneContext* context);
5893 bool optIsLoopClonable(unsigned loopInd);
5895 bool optCanCloneLoops();
5898 void optDebugLogLoopCloning(BasicBlock* block, GenTreePtr insertBefore);
5900 void optPerformStaticOptimizations(unsigned loopNum, LoopCloneContext* context DEBUGARG(bool fastPath));
5901 bool optComputeDerefConditions(unsigned loopNum, LoopCloneContext* context);
5902 bool optDeriveLoopCloningConditions(unsigned loopNum, LoopCloneContext* context);
5903 BasicBlock* optInsertLoopChoiceConditions(LoopCloneContext* context,
5907 void optInsertLoopCloningStress(BasicBlock* head);
5909 #if COUNT_RANGECHECKS
5910 static unsigned optRangeChkRmv;
5911 static unsigned optRangeChkAll;
5920 #define MAX_ARRAYS 4 // a magic max number of arrays tracked for bounds check elimination
5925 RngChkDsc* rcdNextInBucket; // used by the hash table
5927 unsigned short rcdHashValue; // to make matching faster
5928 unsigned short rcdIndex; // 0..optRngChkCount-1
5930 GenTreePtr rcdTree; // the array index tree
5933 unsigned optRngChkCount;
5934 static const size_t optRngChkHashSize;
5936 ssize_t optGetArrayRefScaleAndIndex(GenTreePtr mul, GenTreePtr* pIndex DEBUGARG(bool bRngChk));
5937 GenTreePtr optFindLocalInit(BasicBlock* block, GenTreePtr local, VARSET_TP* pKilledInOut, bool* isKilledAfterInit);
5940 bool optIsNoMore(GenTreePtr op1, GenTreePtr op2, int add1 = 0, int add2 = 0);
5943 bool optReachWithoutCall(BasicBlock* srcBB, BasicBlock* dstBB);
5946 bool optLoopsMarked;
5949 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
5950 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
5954 XX Does the register allocation and puts the remaining lclVars on the stack XX
5956 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
5957 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
5961 #ifndef LEGACY_BACKEND
5966 #else // LEGACY_BACKEND
5971 #endif // LEGACY_BACKEND
5973 #ifdef LEGACY_BACKEND
5975 void raAssignVars(); // register allocation
5976 #endif // LEGACY_BACKEND
5978 VARSET_TP raRegVarsMask; // Set of all enregistered variables (not including FEATURE_STACK_FP_X87 enregistered
5980 regNumber raUpdateRegStateForArg(RegState* regState, LclVarDsc* argDsc);
5982 void raMarkStkVars();
5985 // Some things are used by both LSRA and regpredict allocators.
5987 FrameType rpFrameType;
5988 bool rpMustCreateEBPCalled; // Set to true after we have called rpMustCreateEBPFrame once
5990 #ifdef LEGACY_BACKEND
5991 regMaskTP rpMaskPInvokeEpilogIntf; // pinvoke epilog trashes esi/edi holding stack args needed to setup tail call's
5993 #endif // LEGACY_BACKEND
5995 bool rpMustCreateEBPFrame(INDEBUG(const char** wbReason));
5997 #if FEATURE_FP_REGALLOC
5998 enum enumConfigRegisterFP
6000 CONFIG_REGISTER_FP_NONE = 0x0,
6001 CONFIG_REGISTER_FP_CALLEE_TRASH = 0x1,
6002 CONFIG_REGISTER_FP_CALLEE_SAVED = 0x2,
6003 CONFIG_REGISTER_FP_FULL = 0x3,
6005 enumConfigRegisterFP raConfigRegisterFP();
6006 #endif // FEATURE_FP_REGALLOC
6009 regMaskTP raConfigRestrictMaskFP();
6012 #ifndef LEGACY_BACKEND
6013 LinearScanInterface* m_pLinearScan; // Linear Scan allocator
6014 #else // LEGACY_BACKEND
6015 unsigned raAvoidArgRegMask; // Mask of incoming argument registers that we may need to avoid
6016 VARSET_TP raLclRegIntf[REG_COUNT]; // variable to register interference graph
6017 bool raNewBlocks; // True is we added killing blocks for FPU registers
6018 unsigned rpPasses; // Number of passes made by the register predicter
6019 unsigned rpPassesMax; // Maximum number of passes made by the register predicter
6020 unsigned rpPassesPessimize; // Number of passes non-pessimizing made by the register predicter
6021 unsigned rpStkPredict; // Weighted count of variables were predicted STK (lower means register allocation is better)
6022 unsigned rpPredictSpillCnt; // Predicted number of integer spill tmps for the current tree
6023 regMaskTP rpPredictAssignMask; // Mask of registers to consider in rpPredictAssignRegVars()
6024 VARSET_TP rpLastUseVars; // Set of last use variables in rpPredictTreeRegUse
6025 VARSET_TP rpUseInPlace; // Set of variables that we used in place
6026 int rpAsgVarNum; // VarNum for the target of GT_ASG node
6027 bool rpPredictAssignAgain; // Must rerun the rpPredictAssignRegVars()
6028 bool rpAddedVarIntf; // Set to true if we need to add a new var intf
6029 bool rpLostEnreg; // Set to true if we lost an enregister var that had lvDependReg set
6030 bool rpReverseEBPenreg; // Decided to reverse the enregistration of EBP
6032 bool rpRegAllocDone; // Set to true after we have completed register allocation
6034 regMaskTP rpPredictMap[PREDICT_COUNT]; // Holds the regMaskTP for each of the enum values
6036 void raSetupArgMasks(RegState* r);
6038 const regNumber* raGetRegVarOrder(var_types regType, unsigned* wbVarOrderSize);
6040 void raDumpVarIntf(); // Dump the variable to variable interference graph
6041 void raDumpRegIntf(); // Dump the variable to register interference graph
6043 void raAdjustVarIntf();
6045 regMaskTP rpPredictRegMask(rpPredictReg predictReg, var_types type);
6047 bool rpRecordRegIntf(regMaskTP regMask, VARSET_VALARG_TP life DEBUGARG(const char* msg));
6049 bool rpRecordVarIntf(unsigned varNum, VARSET_VALARG_TP intfVar DEBUGARG(const char* msg));
6050 regMaskTP rpPredictRegPick(var_types type, rpPredictReg predictReg, regMaskTP lockedRegs);
6052 regMaskTP rpPredictGrabReg(var_types type, rpPredictReg predictReg, regMaskTP lockedRegs);
6054 static fgWalkPreFn rpMarkRegIntf;
6056 regMaskTP rpPredictAddressMode(
6057 GenTreePtr tree, var_types type, regMaskTP lockedRegs, regMaskTP rsvdRegs, GenTreePtr lenCSE);
6059 void rpPredictRefAssign(unsigned lclNum);
6061 regMaskTP rpPredictBlkAsgRegUse(GenTreePtr tree, rpPredictReg predictReg, regMaskTP lockedRegs, regMaskTP rsvdRegs);
6063 regMaskTP rpPredictTreeRegUse(GenTreePtr tree, rpPredictReg predictReg, regMaskTP lockedRegs, regMaskTP rsvdRegs);
6065 regMaskTP rpPredictAssignRegVars(regMaskTP regAvail);
6067 void rpPredictRegUse(); // Entry point
6069 unsigned raPredictTreeRegUse(GenTreePtr tree);
6070 unsigned raPredictListRegUse(GenTreePtr list);
6072 void raSetRegVarOrder(var_types regType,
6073 regNumber* customVarOrder,
6074 unsigned* customVarOrderSize,
6076 regMaskTP avoidReg);
6078 // We use (unsigned)-1 as an uninitialized sentinel for rpStkPredict and
6079 // also as the maximum value of lvRefCntWtd. Don't allow overflow, and
6080 // saturate at UINT_MAX - 1, to avoid using the sentinel.
6081 void raAddToStkPredict(unsigned val)
6083 unsigned newStkPredict = rpStkPredict + val;
6084 if ((newStkPredict < rpStkPredict) || (newStkPredict == UINT_MAX))
6085 rpStkPredict = UINT_MAX - 1;
6087 rpStkPredict = newStkPredict;
6091 #if !FEATURE_FP_REGALLOC
6092 void raDispFPlifeInfo();
6096 regMaskTP genReturnRegForTree(GenTreePtr tree);
6097 #endif // LEGACY_BACKEND
6099 /* raIsVarargsStackArg is called by raMaskStkVars and by
6100 lvaSortByRefCount. It identifies the special case
6101 where a varargs function has a parameter passed on the
6102 stack, other than the special varargs handle. Such parameters
6103 require special treatment, because they cannot be tracked
6104 by the GC (their offsets in the stack are not known
6108 bool raIsVarargsStackArg(unsigned lclNum)
6112 LclVarDsc* varDsc = &lvaTable[lclNum];
6114 assert(varDsc->lvIsParam);
6116 return (info.compIsVarArgs && !varDsc->lvIsRegArg && (lclNum != lvaVarargsHandleArg));
6118 #else // _TARGET_X86_
6122 #endif // _TARGET_X86_
6125 #ifdef LEGACY_BACKEND
6126 // Records the current prediction, if it's better than any previous recorded prediction.
6127 void rpRecordPrediction();
6128 // Applies the best recorded prediction, if one exists and is better than the current prediction.
6129 void rpUseRecordedPredictionIfBetter();
6131 // Data members used in the methods above.
6132 unsigned rpBestRecordedStkPredict;
6133 struct VarRegPrediction
6135 bool m_isEnregistered;
6136 regNumberSmall m_regNum;
6137 regNumberSmall m_otherReg;
6139 VarRegPrediction* rpBestRecordedPrediction;
6140 #endif // LEGACY_BACKEND
6143 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6144 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6148 XX Get to the class and method info from the Execution Engine given XX
6149 XX tokens for the class and method XX
6151 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6152 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6156 /* These are the different addressing modes used to access a local var.
6157 * The JIT has to report the location of the locals back to the EE
6158 * for debugging purposes.
6164 VLT_REG_BYREF, // this type is currently only used for value types on X64
6167 VLT_STK_BYREF, // this type is currently only used for value types on X64
6181 siVarLocType vlType;
6184 // VLT_REG/VLT_REG_FP -- Any pointer-sized enregistered value (TYP_INT, TYP_REF, etc)
6186 // VLT_REG_BYREF -- the specified register contains the address of the variable
6194 // VLT_STK -- Any 32 bit value which is on the stack
6195 // eg. [ESP+0x20], or [EBP-0x28]
6196 // VLT_STK_BYREF -- the specified stack location contains the address of the variable
6197 // eg. mov EAX, [ESP+0x20]; [EAX]
6201 regNumber vlsBaseReg;
6202 NATIVE_OFFSET vlsOffset;
6205 // VLT_REG_REG -- TYP_LONG/TYP_DOUBLE with both DWords enregistered
6214 // VLT_REG_STK -- Partly enregistered TYP_LONG/TYP_DOUBLE
6215 // eg { LowerDWord=EAX UpperDWord=[ESP+0x8] }
6223 regNumber vlrssBaseReg;
6224 NATIVE_OFFSET vlrssOffset;
6228 // VLT_STK_REG -- Partly enregistered TYP_LONG/TYP_DOUBLE
6229 // eg { LowerDWord=[ESP+0x8] UpperDWord=EAX }
6235 regNumber vlsrsBaseReg;
6236 NATIVE_OFFSET vlsrsOffset;
6242 // VLT_STK2 -- Any 64 bit value which is on the stack, in 2 successsive DWords
6243 // eg 2 DWords at [ESP+0x10]
6247 regNumber vls2BaseReg;
6248 NATIVE_OFFSET vls2Offset;
6251 // VLT_FPSTK -- enregisterd TYP_DOUBLE (on the FP stack)
6252 // eg. ST(3). Actually it is ST("FPstkHeight - vpFpStk")
6259 // VLT_FIXED_VA -- fixed argument of a varargs function.
6260 // The argument location depends on the size of the variable
6261 // arguments (...). Inspecting the VARARGS_HANDLE indicates the
6262 // location of the first arg. This argument can then be accessed
6263 // relative to the position of the first arg
6267 unsigned vlfvOffset;
6274 void* rpValue; // pointer to the in-process
6275 // location of the value.
6281 bool vlIsInReg(regNumber reg);
6282 bool vlIsOnStk(regNumber reg, signed offset);
6285 /*************************************************************************/
6290 void eeGetCallInfo(CORINFO_RESOLVED_TOKEN* pResolvedToken,
6291 CORINFO_RESOLVED_TOKEN* pConstrainedToken,
6292 CORINFO_CALLINFO_FLAGS flags,
6293 CORINFO_CALL_INFO* pResult);
6294 inline CORINFO_CALLINFO_FLAGS addVerifyFlag(CORINFO_CALLINFO_FLAGS flags);
6296 void eeGetFieldInfo(CORINFO_RESOLVED_TOKEN* pResolvedToken,
6297 CORINFO_ACCESS_FLAGS flags,
6298 CORINFO_FIELD_INFO* pResult);
6302 BOOL eeIsValueClass(CORINFO_CLASS_HANDLE clsHnd);
6304 #if defined(DEBUG) || defined(FEATURE_JIT_METHOD_PERF) || defined(FEATURE_SIMD) || defined(TRACK_LSRA_STATS)
6306 bool IsSuperPMIException(unsigned code)
6308 // Copied from NDP\clr\src\ToolBox\SuperPMI\SuperPMI-Shared\ErrorHandling.h
6310 const unsigned EXCEPTIONCODE_DebugBreakorAV = 0xe0421000;
6311 const unsigned EXCEPTIONCODE_MC = 0xe0422000;
6312 const unsigned EXCEPTIONCODE_LWM = 0xe0423000;
6313 const unsigned EXCEPTIONCODE_SASM = 0xe0424000;
6314 const unsigned EXCEPTIONCODE_SSYM = 0xe0425000;
6315 const unsigned EXCEPTIONCODE_CALLUTILS = 0xe0426000;
6316 const unsigned EXCEPTIONCODE_TYPEUTILS = 0xe0427000;
6317 const unsigned EXCEPTIONCODE_ASSERT = 0xe0440000;
6321 case EXCEPTIONCODE_DebugBreakorAV:
6322 case EXCEPTIONCODE_MC:
6323 case EXCEPTIONCODE_LWM:
6324 case EXCEPTIONCODE_SASM:
6325 case EXCEPTIONCODE_SSYM:
6326 case EXCEPTIONCODE_CALLUTILS:
6327 case EXCEPTIONCODE_TYPEUTILS:
6328 case EXCEPTIONCODE_ASSERT:
6335 const char* eeGetMethodName(CORINFO_METHOD_HANDLE hnd, const char** className);
6336 const char* eeGetMethodFullName(CORINFO_METHOD_HANDLE hnd);
6338 bool eeIsNativeMethod(CORINFO_METHOD_HANDLE method);
6339 CORINFO_METHOD_HANDLE eeGetMethodHandleForNative(CORINFO_METHOD_HANDLE method);
6342 var_types eeGetArgType(CORINFO_ARG_LIST_HANDLE list, CORINFO_SIG_INFO* sig);
6343 var_types eeGetArgType(CORINFO_ARG_LIST_HANDLE list, CORINFO_SIG_INFO* sig, bool* isPinned);
6344 unsigned eeGetArgSize(CORINFO_ARG_LIST_HANDLE list, CORINFO_SIG_INFO* sig);
6346 // VOM info, method sigs
6348 void eeGetSig(unsigned sigTok,
6349 CORINFO_MODULE_HANDLE scope,
6350 CORINFO_CONTEXT_HANDLE context,
6351 CORINFO_SIG_INFO* retSig);
6353 void eeGetCallSiteSig(unsigned sigTok,
6354 CORINFO_MODULE_HANDLE scope,
6355 CORINFO_CONTEXT_HANDLE context,
6356 CORINFO_SIG_INFO* retSig);
6358 void eeGetMethodSig(CORINFO_METHOD_HANDLE methHnd, CORINFO_SIG_INFO* retSig, CORINFO_CLASS_HANDLE owner = nullptr);
6360 // Method entry-points, instrs
6362 void* eeGetFieldAddress(CORINFO_FIELD_HANDLE handle, void*** ppIndir);
6364 CORINFO_METHOD_HANDLE eeMarkNativeTarget(CORINFO_METHOD_HANDLE method);
6366 CORINFO_EE_INFO eeInfo;
6367 bool eeInfoInitialized;
6369 CORINFO_EE_INFO* eeGetEEInfo();
6371 // Gets the offset of a SDArray's first element
6372 unsigned eeGetArrayDataOffset(var_types type);
6373 // Gets the offset of a MDArray's first element
6374 unsigned eeGetMDArrayDataOffset(var_types type, unsigned rank);
6376 GenTreePtr eeGetPInvokeCookie(CORINFO_SIG_INFO* szMetaSig);
6378 // Returns the page size for the target machine as reported by the EE.
6379 inline size_t eeGetPageSize()
6381 #if COR_JIT_EE_VERSION > 460
6382 return eeGetEEInfo()->osPageSize;
6383 #else // COR_JIT_EE_VERSION <= 460
6384 return CORINFO_PAGE_SIZE;
6385 #endif // COR_JIT_EE_VERSION > 460
6388 // Returns the frame size at which we will generate a loop to probe the stack.
6389 inline size_t getVeryLargeFrameSize()
6392 // The looping probe code is 40 bytes, whereas the straight-line probing for
6393 // the (0x2000..0x3000) case is 44, so use looping for anything 0x2000 bytes
6394 // or greater, to generate smaller code.
6395 return 2 * eeGetPageSize();
6397 return 3 * eeGetPageSize();
6401 inline bool IsTargetAbi(CORINFO_RUNTIME_ABI abi)
6403 #if COR_JIT_EE_VERSION > 460
6404 return eeGetEEInfo()->targetAbi == abi;
6406 return CORINFO_DESKTOP_ABI == abi;
6410 inline bool generateCFIUnwindCodes()
6412 #ifdef UNIX_AMD64_ABI
6413 return IsTargetAbi(CORINFO_CORERT_ABI);
6421 unsigned eeGetEHcount(CORINFO_METHOD_HANDLE handle);
6423 // Debugging support - Line number info
6425 void eeGetStmtOffsets();
6427 unsigned eeBoundariesCount;
6429 struct boundariesDsc
6431 UNATIVE_OFFSET nativeIP;
6433 unsigned sourceReason;
6434 } * eeBoundaries; // Boundaries to report to EE
6435 void eeSetLIcount(unsigned count);
6436 void eeSetLIinfo(unsigned which, UNATIVE_OFFSET offs, unsigned srcIP, bool stkEmpty, bool callInstruction);
6440 static void eeDispILOffs(IL_OFFSET offs);
6441 static void eeDispLineInfo(const boundariesDsc* line);
6442 void eeDispLineInfos();
6445 // Debugging support - Local var info
6449 unsigned eeVarsCount;
6451 struct VarResultInfo
6453 UNATIVE_OFFSET startOffset;
6454 UNATIVE_OFFSET endOffset;
6458 void eeSetLVcount(unsigned count);
6459 void eeSetLVinfo(unsigned which,
6460 UNATIVE_OFFSET startOffs,
6461 UNATIVE_OFFSET length,
6466 const siVarLoc& loc);
6470 void eeDispVar(ICorDebugInfo::NativeVarInfo* var);
6471 void eeDispVars(CORINFO_METHOD_HANDLE ftn, ULONG32 cVars, ICorDebugInfo::NativeVarInfo* vars);
6474 // ICorJitInfo wrappers
6476 void eeReserveUnwindInfo(BOOL isFunclet, BOOL isColdCode, ULONG unwindSize);
6478 void eeAllocUnwindInfo(BYTE* pHotCode,
6484 CorJitFuncKind funcKind);
6486 void eeSetEHcount(unsigned cEH);
6488 void eeSetEHinfo(unsigned EHnumber, const CORINFO_EH_CLAUSE* clause);
6490 WORD eeGetRelocTypeHint(void* target);
6492 // ICorStaticInfo wrapper functions
6494 bool eeTryResolveToken(CORINFO_RESOLVED_TOKEN* resolvedToken);
6496 #if defined(FEATURE_UNIX_AMD64_STRUCT_PASSING)
6498 static void dumpSystemVClassificationType(SystemVClassificationType ct);
6501 void eeGetSystemVAmd64PassStructInRegisterDescriptor(
6502 /*IN*/ CORINFO_CLASS_HANDLE structHnd,
6503 /*OUT*/ SYSTEMV_AMD64_CORINFO_STRUCT_REG_PASSING_DESCRIPTOR* structPassInRegDescPtr);
6504 #endif // FEATURE_UNIX_AMD64_STRUCT_PASSING
6506 template <typename ParamType>
6507 bool eeRunWithErrorTrap(void (*function)(ParamType*), ParamType* param)
6509 return eeRunWithErrorTrapImp(reinterpret_cast<void (*)(void*)>(function), reinterpret_cast<void*>(param));
6512 bool eeRunWithErrorTrapImp(void (*function)(void*), void* param);
6514 // Utility functions
6516 const char* eeGetFieldName(CORINFO_FIELD_HANDLE fieldHnd, const char** classNamePtr = nullptr);
6519 const wchar_t* eeGetCPString(size_t stringHandle);
6522 const char* eeGetClassName(CORINFO_CLASS_HANDLE clsHnd);
6524 static CORINFO_METHOD_HANDLE eeFindHelper(unsigned helper);
6525 static CorInfoHelpFunc eeGetHelperNum(CORINFO_METHOD_HANDLE method);
6527 static fgWalkPreFn CountSharedStaticHelper;
6528 static bool IsSharedStaticHelper(GenTreePtr tree);
6529 static bool IsTreeAlwaysHoistable(GenTreePtr tree);
6531 static CORINFO_FIELD_HANDLE eeFindJitDataOffs(unsigned jitDataOffs);
6532 // returns true/false if 'field' is a Jit Data offset
6533 static bool eeIsJitDataOffs(CORINFO_FIELD_HANDLE field);
6534 // returns a number < 0 if 'field' is not a Jit Data offset, otherwise the data offset (limited to 2GB)
6535 static int eeGetJitDataOffs(CORINFO_FIELD_HANDLE field);
6537 /*****************************************************************************/
6542 enum TEMP_USAGE_TYPE
6548 static var_types tmpNormalizeType(var_types type);
6549 TempDsc* tmpGetTemp(var_types type); // get temp for the given type
6550 void tmpRlsTemp(TempDsc* temp);
6551 TempDsc* tmpFindNum(int temp, TEMP_USAGE_TYPE usageType = TEMP_USAGE_FREE) const;
6554 TempDsc* tmpListBeg(TEMP_USAGE_TYPE usageType = TEMP_USAGE_FREE) const;
6555 TempDsc* tmpListNxt(TempDsc* curTemp, TEMP_USAGE_TYPE usageType = TEMP_USAGE_FREE) const;
6559 bool tmpAllFree() const;
6562 #ifndef LEGACY_BACKEND
6563 void tmpPreAllocateTemps(var_types type, unsigned count);
6564 #endif // !LEGACY_BACKEND
6567 #ifdef LEGACY_BACKEND
6568 unsigned tmpIntSpillMax; // number of int-sized spill temps
6569 unsigned tmpDoubleSpillMax; // number of double-sized spill temps
6570 #endif // LEGACY_BACKEND
6572 unsigned tmpCount; // Number of temps
6573 unsigned tmpSize; // Size of all the temps
6576 // Used by RegSet::rsSpillChk()
6577 unsigned tmpGetCount; // Temps which haven't been released yet
6580 static unsigned tmpSlot(unsigned size); // which slot in tmpFree[] or tmpUsed[] to use
6582 TempDsc* tmpFree[TEMP_MAX_SIZE / sizeof(int)];
6583 TempDsc* tmpUsed[TEMP_MAX_SIZE / sizeof(int)];
6586 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6587 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6591 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6592 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6596 CodeGenInterface* codeGen;
6598 // The following holds information about instr offsets in terms of generated code.
6602 IPmappingDsc* ipmdNext; // next line# record
6603 IL_OFFSETX ipmdILoffsx; // the instr offset
6604 emitLocation ipmdNativeLoc; // the emitter location of the native code corresponding to the IL offset
6605 bool ipmdIsLabel; // Can this code be a branch label?
6608 // Record the instr offset mapping to the generated code
6610 IPmappingDsc* genIPmappingList;
6611 IPmappingDsc* genIPmappingLast;
6613 // Managed RetVal - A side hash table meant to record the mapping from a
6614 // GT_CALL node to its IL offset. This info is used to emit sequence points
6615 // that can be used by debugger to determine the native offset at which the
6616 // managed RetVal will be available.
6618 // In fact we can store IL offset in a GT_CALL node. This was ruled out in
6619 // favor of a side table for two reasons: 1) We need IL offset for only those
6620 // GT_CALL nodes (created during importation) that correspond to an IL call and
6621 // whose return type is other than TYP_VOID. 2) GT_CALL node is a frequently used
6622 // structure and IL offset is needed only when generating debuggable code. Therefore
6623 // it is desirable to avoid memory size penalty in retail scenarios.
6624 typedef SimplerHashTable<GenTreePtr, PtrKeyFuncs<GenTree>, IL_OFFSETX, JitSimplerHashBehavior>
6625 CallSiteILOffsetTable;
6626 CallSiteILOffsetTable* genCallSite2ILOffsetMap;
6628 unsigned genReturnLocal; // Local number for the return value when applicable.
6629 BasicBlock* genReturnBB; // jumped to when not optimizing for speed.
6631 // The following properties are part of CodeGenContext. Getters are provided here for
6632 // convenience and backward compatibility, but the properties can only be set by invoking
6633 // the setter on CodeGenContext directly.
6635 __declspec(property(get = getEmitter)) emitter* genEmitter;
6636 emitter* getEmitter()
6638 return codeGen->getEmitter();
6641 const bool isFramePointerUsed()
6643 return codeGen->isFramePointerUsed();
6646 __declspec(property(get = getInterruptible, put = setInterruptible)) bool genInterruptible;
6647 bool getInterruptible()
6649 return codeGen->genInterruptible;
6651 void setInterruptible(bool value)
6653 codeGen->setInterruptible(value);
6657 const bool genDoubleAlign()
6659 return codeGen->doDoubleAlign();
6661 DWORD getCanDoubleAlign();
6662 bool shouldDoubleAlign(unsigned refCntStk,
6664 unsigned refCntWtdReg,
6665 unsigned refCntStkParam,
6666 unsigned refCntWtdStkDbl);
6667 #endif // DOUBLE_ALIGN
6669 __declspec(property(get = getFullPtrRegMap, put = setFullPtrRegMap)) bool genFullPtrRegMap;
6670 bool getFullPtrRegMap()
6672 return codeGen->genFullPtrRegMap;
6674 void setFullPtrRegMap(bool value)
6676 codeGen->setFullPtrRegMap(value);
6679 // Things that MAY belong either in CodeGen or CodeGenContext
6681 #if FEATURE_EH_FUNCLETS
6682 FuncInfoDsc* compFuncInfos;
6683 unsigned short compCurrFuncIdx;
6684 unsigned short compFuncInfoCount;
6686 unsigned short compFuncCount()
6688 assert(fgFuncletsCreated);
6689 return compFuncInfoCount;
6692 #else // !FEATURE_EH_FUNCLETS
6694 // This is a no-op when there are no funclets!
6695 void genUpdateCurrentFunclet(BasicBlock* block)
6700 FuncInfoDsc compFuncInfoRoot;
6702 static const unsigned compCurrFuncIdx = 0;
6704 unsigned short compFuncCount()
6709 #endif // !FEATURE_EH_FUNCLETS
6711 FuncInfoDsc* funCurrentFunc();
6712 void funSetCurrentFunc(unsigned funcIdx);
6713 FuncInfoDsc* funGetFunc(unsigned funcIdx);
6714 unsigned int funGetFuncIdx(BasicBlock* block);
6718 VARSET_TP compCurLife; // current live variables
6719 GenTreePtr compCurLifeTree; // node after which compCurLife has been computed
6721 template <bool ForCodeGen>
6722 void compChangeLife(VARSET_VALARG_TP newLife DEBUGARG(GenTreePtr tree));
6724 void genChangeLife(VARSET_VALARG_TP newLife DEBUGARG(GenTreePtr tree))
6726 compChangeLife</*ForCodeGen*/ true>(newLife DEBUGARG(tree));
6729 template <bool ForCodeGen>
6730 void compUpdateLife(GenTreePtr tree);
6732 // Updates "compCurLife" to its state after evaluate of "true". If "pLastUseVars" is
6733 // non-null, sets "*pLastUseVars" to the set of tracked variables for which "tree" was a last
6734 // use. (Can be more than one var in the case of dependently promoted struct vars.)
6735 template <bool ForCodeGen>
6736 void compUpdateLifeVar(GenTreePtr tree, VARSET_TP* pLastUseVars = nullptr);
6738 template <bool ForCodeGen>
6739 inline void compUpdateLife(VARSET_VALARG_TP newLife);
6741 // Gets a register mask that represent the kill set for a helper call since
6742 // not all JIT Helper calls follow the standard ABI on the target architecture.
6743 regMaskTP compHelperCallKillSet(CorInfoHelpFunc helper);
6745 // Gets a register mask that represent the kill set for a NoGC helper call.
6746 regMaskTP compNoGCHelperCallKillSet(CorInfoHelpFunc helper);
6749 // Requires that "varDsc" be a promoted struct local variable being passed as an argument, beginning at
6750 // "firstArgRegNum", which is assumed to have already been aligned to the register alignment restriction of the
6751 // struct type. Adds bits to "*pArgSkippedRegMask" for any argument registers *not* used in passing "varDsc" --
6752 // i.e., internal "holes" caused by internal alignment constraints. For example, if the struct contained an int and
6753 // a double, and we at R0 (on ARM), then R1 would be skipped, and the bit for R1 would be added to the mask.
6754 void fgAddSkippedRegsInPromotedStructArg(LclVarDsc* varDsc, unsigned firstArgRegNum, regMaskTP* pArgSkippedRegMask);
6755 #endif // _TARGET_ARM_
6757 // If "tree" is a indirection (GT_IND, or GT_OBJ) whose arg is an ADDR, whose arg is a LCL_VAR, return that LCL_VAR
6759 static GenTreePtr fgIsIndirOfAddrOfLocal(GenTreePtr tree);
6761 // This is indexed by GT_OBJ nodes that are address of promoted struct variables, which
6762 // have been annotated with the GTF_VAR_DEATH flag. If such a node is *not* mapped in this
6763 // table, one may assume that all the (tracked) field vars die at this point. Otherwise,
6764 // the node maps to a pointer to a VARSET_TP, containing set bits for each of the tracked field
6765 // vars of the promoted struct local that go dead at the given node (the set bits are the bits
6766 // for the tracked var indices of the field vars, as in a live var set).
6767 NodeToVarsetPtrMap* m_promotedStructDeathVars;
6769 NodeToVarsetPtrMap* GetPromotedStructDeathVars()
6771 if (m_promotedStructDeathVars == nullptr)
6773 m_promotedStructDeathVars = new (getAllocator()) NodeToVarsetPtrMap(getAllocator());
6775 return m_promotedStructDeathVars;
6779 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6780 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6784 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6785 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6788 #if !defined(__GNUC__)
6789 #pragma region Unwind information
6794 // Infrastructure functions: start/stop/reserve/emit.
6797 void unwindBegProlog();
6798 void unwindEndProlog();
6799 void unwindBegEpilog();
6800 void unwindEndEpilog();
6801 void unwindReserve();
6802 void unwindEmit(void* pHotCode, void* pColdCode);
6805 // Specific unwind information functions: called by code generation to indicate a particular
6806 // prolog or epilog unwindable instruction has been generated.
6809 void unwindPush(regNumber reg);
6810 void unwindAllocStack(unsigned size);
6811 void unwindSetFrameReg(regNumber reg, unsigned offset);
6812 void unwindSaveReg(regNumber reg, unsigned offset);
6814 #if defined(_TARGET_ARM_)
6815 void unwindPushMaskInt(regMaskTP mask);
6816 void unwindPushMaskFloat(regMaskTP mask);
6817 void unwindPopMaskInt(regMaskTP mask);
6818 void unwindPopMaskFloat(regMaskTP mask);
6819 void unwindBranch16(); // The epilog terminates with a 16-bit branch (e.g., "bx lr")
6820 void unwindNop(unsigned codeSizeInBytes); // Generate unwind NOP code. 'codeSizeInBytes' is 2 or 4 bytes. Only
6821 // called via unwindPadding().
6822 void unwindPadding(); // Generate a sequence of unwind NOP codes representing instructions between the last
6823 // instruction and the current location.
6824 #endif // _TARGET_ARM_
6826 #if defined(_TARGET_ARM64_)
6828 void unwindPadding(); // Generate a sequence of unwind NOP codes representing instructions between the last
6829 // instruction and the current location.
6830 void unwindSaveReg(regNumber reg, int offset); // str reg, [sp, #offset]
6831 void unwindSaveRegPreindexed(regNumber reg, int offset); // str reg, [sp, #offset]!
6832 void unwindSaveRegPair(regNumber reg1, regNumber reg2, int offset); // stp reg1, reg2, [sp, #offset]
6833 void unwindSaveRegPairPreindexed(regNumber reg1, regNumber reg2, int offset); // stp reg1, reg2, [sp, #offset]!
6834 void unwindSaveNext(); // unwind code: save_next
6835 void unwindReturn(regNumber reg); // ret lr
6836 #endif // defined(_TARGET_ARM64_)
6839 // Private "helper" functions for the unwind implementation.
6843 #if FEATURE_EH_FUNCLETS
6844 void unwindGetFuncLocations(FuncInfoDsc* func,
6845 bool getHotSectionData,
6846 /* OUT */ emitLocation** ppStartLoc,
6847 /* OUT */ emitLocation** ppEndLoc);
6848 #endif // FEATURE_EH_FUNCLETS
6850 void unwindReserveFunc(FuncInfoDsc* func);
6851 void unwindEmitFunc(FuncInfoDsc* func, void* pHotCode, void* pColdCode);
6853 #if defined(_TARGET_AMD64_)
6855 void unwindReserveFuncHelper(FuncInfoDsc* func, bool isHotCode);
6856 void unwindEmitFuncHelper(FuncInfoDsc* func, void* pHotCode, void* pColdCode, bool isHotCode);
6857 UNATIVE_OFFSET unwindGetCurrentOffset(FuncInfoDsc* func);
6859 void unwindBegPrologWindows();
6860 void unwindPushWindows(regNumber reg);
6861 void unwindAllocStackWindows(unsigned size);
6862 void unwindSetFrameRegWindows(regNumber reg, unsigned offset);
6863 void unwindSaveRegWindows(regNumber reg, unsigned offset);
6865 #ifdef UNIX_AMD64_ABI
6866 void unwindBegPrologCFI();
6867 void unwindPushCFI(regNumber reg);
6868 void unwindAllocStackCFI(unsigned size);
6869 void unwindSetFrameRegCFI(regNumber reg, unsigned offset);
6870 void unwindSaveRegCFI(regNumber reg, unsigned offset);
6871 int mapRegNumToDwarfReg(regNumber reg);
6872 void createCfiCode(FuncInfoDsc* func, UCHAR codeOffset, UCHAR opcode, USHORT dwarfReg, INT offset = 0);
6873 #endif // UNIX_AMD64_ABI
6874 #elif defined(_TARGET_ARM_)
6876 void unwindPushPopMaskInt(regMaskTP mask, bool useOpsize16);
6877 void unwindPushPopMaskFloat(regMaskTP mask);
6878 void unwindSplit(FuncInfoDsc* func);
6880 #endif // _TARGET_ARM_
6882 #if !defined(__GNUC__)
6883 #pragma endregion // Note: region is NOT under !defined(__GNUC__)
6887 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6888 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6892 XX Info about SIMD types, methods and the SIMD assembly (i.e. the assembly XX
6893 XX that contains the distinguished, well-known SIMD type definitions). XX
6895 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6896 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6899 // Get highest available instruction set for floating point codegen
6900 InstructionSet getFloatingPointInstructionSet()
6902 #if defined(_TARGET_XARCH_) && !defined(LEGACY_BACKEND)
6905 return InstructionSet_AVX;
6910 return InstructionSet_SSE3_4;
6914 assert(canUseSSE2());
6915 return InstructionSet_SSE2;
6917 assert(!"getFPInstructionSet() is not implemented for target arch");
6919 return InstructionSet_NONE;
6923 // Get highest available instruction set for SIMD codegen
6924 InstructionSet getSIMDInstructionSet()
6926 #if defined(_TARGET_XARCH_) && !defined(LEGACY_BACKEND)
6927 return getFloatingPointInstructionSet();
6929 assert(!"Available instruction set(s) for SIMD codegen is not defined for target arch");
6931 return InstructionSet_NONE;
6937 // Should we support SIMD intrinsics?
6940 // This is a temp lclVar allocated on the stack as TYP_SIMD. It is used to implement intrinsics
6941 // that require indexed access to the individual fields of the vector, which is not well supported
6942 // by the hardware. It is allocated when/if such situations are encountered during Lowering.
6943 unsigned lvaSIMDInitTempVarNum;
6946 CORINFO_CLASS_HANDLE SIMDFloatHandle;
6947 CORINFO_CLASS_HANDLE SIMDDoubleHandle;
6948 CORINFO_CLASS_HANDLE SIMDIntHandle;
6949 CORINFO_CLASS_HANDLE SIMDUShortHandle;
6950 CORINFO_CLASS_HANDLE SIMDUByteHandle;
6951 CORINFO_CLASS_HANDLE SIMDShortHandle;
6952 CORINFO_CLASS_HANDLE SIMDByteHandle;
6953 CORINFO_CLASS_HANDLE SIMDLongHandle;
6954 CORINFO_CLASS_HANDLE SIMDUIntHandle;
6955 CORINFO_CLASS_HANDLE SIMDULongHandle;
6956 CORINFO_CLASS_HANDLE SIMDVector2Handle;
6957 CORINFO_CLASS_HANDLE SIMDVector3Handle;
6958 CORINFO_CLASS_HANDLE SIMDVector4Handle;
6959 CORINFO_CLASS_HANDLE SIMDVectorHandle;
6961 // Get the handle for a SIMD type.
6962 CORINFO_CLASS_HANDLE gtGetStructHandleForSIMD(var_types simdType, var_types simdBaseType)
6964 if (simdBaseType == TYP_FLOAT)
6969 return SIMDVector2Handle;
6971 return SIMDVector3Handle;
6973 if ((getSIMDVectorType() == TYP_SIMD32) || (SIMDVector4Handle != NO_CLASS_HANDLE))
6975 return SIMDVector4Handle;
6984 assert(simdType == getSIMDVectorType());
6985 switch (simdBaseType)
6988 return SIMDFloatHandle;
6990 return SIMDDoubleHandle;
6992 return SIMDIntHandle;
6994 return SIMDUShortHandle;
6996 return SIMDUShortHandle;
6998 return SIMDUByteHandle;
7000 return SIMDShortHandle;
7002 return SIMDByteHandle;
7004 return SIMDLongHandle;
7006 return SIMDUIntHandle;
7008 return SIMDULongHandle;
7010 assert(!"Didn't find a class handle for simdType");
7012 return NO_CLASS_HANDLE;
7016 CORINFO_METHOD_HANDLE SIMDVectorFloat_set_Item;
7017 CORINFO_METHOD_HANDLE SIMDVectorFloat_get_Length;
7018 CORINFO_METHOD_HANDLE SIMDVectorFloat_op_Addition;
7020 // Returns true if the tree corresponds to a TYP_SIMD lcl var.
7021 // Note that both SIMD vector args and locals are mared as lvSIMDType = true, but
7022 // type of an arg node is TYP_BYREF and a local node is TYP_SIMD or TYP_STRUCT.
7023 bool isSIMDTypeLocal(GenTree* tree)
7025 return tree->OperIsLocal() && lvaTable[tree->AsLclVarCommon()->gtLclNum].lvSIMDType;
7028 // Returns true if the type of the tree is a byref of TYP_SIMD
7029 bool isAddrOfSIMDType(GenTree* tree)
7031 if (tree->TypeGet() == TYP_BYREF || tree->TypeGet() == TYP_I_IMPL)
7033 switch (tree->OperGet())
7036 return varTypeIsSIMD(tree->gtGetOp1());
7038 case GT_LCL_VAR_ADDR:
7039 return lvaTable[tree->AsLclVarCommon()->gtLclNum].lvSIMDType;
7042 return isSIMDTypeLocal(tree);
7049 static bool isRelOpSIMDIntrinsic(SIMDIntrinsicID intrinsicId)
7051 return (intrinsicId == SIMDIntrinsicEqual || intrinsicId == SIMDIntrinsicLessThan ||
7052 intrinsicId == SIMDIntrinsicLessThanOrEqual || intrinsicId == SIMDIntrinsicGreaterThan ||
7053 intrinsicId == SIMDIntrinsicGreaterThanOrEqual);
7056 // Returns base type of a TYP_SIMD local.
7057 // Returns TYP_UNKNOWN if the local is not TYP_SIMD.
7058 var_types getBaseTypeOfSIMDLocal(GenTree* tree)
7060 if (isSIMDTypeLocal(tree))
7062 return lvaTable[tree->AsLclVarCommon()->gtLclNum].lvBaseType;
7068 bool isSIMDClass(CORINFO_CLASS_HANDLE clsHnd)
7070 return info.compCompHnd->isInSIMDModule(clsHnd);
7073 bool isSIMDClass(typeInfo* pTypeInfo)
7075 return pTypeInfo->IsStruct() && isSIMDClass(pTypeInfo->GetClassHandleForValueClass());
7078 // Get the base (element) type and size in bytes for a SIMD type. Returns TYP_UNKNOWN
7079 // if it is not a SIMD type or is an unsupported base type.
7080 var_types getBaseTypeAndSizeOfSIMDType(CORINFO_CLASS_HANDLE typeHnd, unsigned* sizeBytes = nullptr);
7082 var_types getBaseTypeOfSIMDType(CORINFO_CLASS_HANDLE typeHnd)
7084 return getBaseTypeAndSizeOfSIMDType(typeHnd, nullptr);
7087 // Get SIMD Intrinsic info given the method handle.
7088 // Also sets typeHnd, argCount, baseType and sizeBytes out params.
7089 const SIMDIntrinsicInfo* getSIMDIntrinsicInfo(CORINFO_CLASS_HANDLE* typeHnd,
7090 CORINFO_METHOD_HANDLE methodHnd,
7091 CORINFO_SIG_INFO* sig,
7094 var_types* baseType,
7095 unsigned* sizeBytes);
7097 // Pops and returns GenTree node from importers type stack.
7098 // Normalizes TYP_STRUCT value in case of GT_CALL, GT_RET_EXPR and arg nodes.
7099 GenTreePtr impSIMDPopStack(var_types type, bool expectAddr = false);
7101 // Create a GT_SIMD tree for a Get property of SIMD vector with a fixed index.
7102 GenTreeSIMD* impSIMDGetFixed(var_types simdType, var_types baseType, unsigned simdSize, int index);
7104 // Creates a GT_SIMD tree for Select operation
7105 GenTreePtr impSIMDSelect(CORINFO_CLASS_HANDLE typeHnd,
7107 unsigned simdVectorSize,
7112 // Creates a GT_SIMD tree for Min/Max operation
7113 GenTreePtr impSIMDMinMax(SIMDIntrinsicID intrinsicId,
7114 CORINFO_CLASS_HANDLE typeHnd,
7116 unsigned simdVectorSize,
7120 // Transforms operands and returns the SIMD intrinsic to be applied on
7121 // transformed operands to obtain given relop result.
7122 SIMDIntrinsicID impSIMDRelOp(SIMDIntrinsicID relOpIntrinsicId,
7123 CORINFO_CLASS_HANDLE typeHnd,
7124 unsigned simdVectorSize,
7125 var_types* baseType,
7129 // Creates a GT_SIMD tree for Abs intrinsic.
7130 GenTreePtr impSIMDAbs(CORINFO_CLASS_HANDLE typeHnd, var_types baseType, unsigned simdVectorSize, GenTree* op1);
7132 #if defined(_TARGET_XARCH_) && !defined(LEGACY_BACKEND)
7133 // Transforms operands and returns the SIMD intrinsic to be applied on
7134 // transformed operands to obtain == comparison result.
7135 SIMDIntrinsicID impSIMDLongRelOpEqual(CORINFO_CLASS_HANDLE typeHnd,
7136 unsigned simdVectorSize,
7140 // Transforms operands and returns the SIMD intrinsic to be applied on
7141 // transformed operands to obtain > comparison result.
7142 SIMDIntrinsicID impSIMDLongRelOpGreaterThan(CORINFO_CLASS_HANDLE typeHnd,
7143 unsigned simdVectorSize,
7147 // Transforms operands and returns the SIMD intrinsic to be applied on
7148 // transformed operands to obtain >= comparison result.
7149 SIMDIntrinsicID impSIMDLongRelOpGreaterThanOrEqual(CORINFO_CLASS_HANDLE typeHnd,
7150 unsigned simdVectorSize,
7154 // Transforms operands and returns the SIMD intrinsic to be applied on
7155 // transformed operands to obtain >= comparison result in case of int32
7156 // and small int base type vectors.
7157 SIMDIntrinsicID impSIMDIntegralRelOpGreaterThanOrEqual(
7158 CORINFO_CLASS_HANDLE typeHnd, unsigned simdVectorSize, var_types baseType, GenTree** op1, GenTree** op2);
7159 #endif // defined(_TARGET_XARCH_) && !defined(LEGACY_BACKEND)
7161 void setLclRelatedToSIMDIntrinsic(GenTreePtr tree);
7162 bool areFieldsContiguous(GenTreePtr op1, GenTreePtr op2);
7163 bool areArrayElementsContiguous(GenTreePtr op1, GenTreePtr op2);
7164 bool areArgumentsContiguous(GenTreePtr op1, GenTreePtr op2);
7165 GenTreePtr createAddressNodeForSIMDInit(GenTreePtr tree, unsigned simdSize);
7167 // check methodHnd to see if it is a SIMD method that is expanded as an intrinsic in the JIT.
7168 GenTreePtr impSIMDIntrinsic(OPCODE opcode,
7169 GenTreePtr newobjThis,
7170 CORINFO_CLASS_HANDLE clsHnd,
7171 CORINFO_METHOD_HANDLE method,
7172 CORINFO_SIG_INFO* sig,
7175 GenTreePtr getOp1ForConstructor(OPCODE opcode, GenTreePtr newobjThis, CORINFO_CLASS_HANDLE clsHnd);
7177 // Whether SIMD vector occupies part of SIMD register.
7178 // SSE2: vector2f/3f are considered sub register SIMD types.
7179 // AVX: vector2f, 3f and 4f are all considered sub register SIMD types.
7180 bool isSubRegisterSIMDType(CORINFO_CLASS_HANDLE typeHnd)
7182 unsigned sizeBytes = 0;
7183 var_types baseType = getBaseTypeAndSizeOfSIMDType(typeHnd, &sizeBytes);
7184 return (baseType == TYP_FLOAT) && (sizeBytes < getSIMDVectorRegisterByteLength());
7187 bool isSubRegisterSIMDType(GenTreeSIMD* simdNode)
7189 return (simdNode->gtSIMDSize < getSIMDVectorRegisterByteLength());
7192 // Get the type for the hardware SIMD vector.
7193 // This is the maximum SIMD type supported for this target.
7194 var_types getSIMDVectorType()
7196 #if defined(_TARGET_XARCH_) && !defined(LEGACY_BACKEND)
7203 assert(canUseSSE2());
7207 assert(!"getSIMDVectorType() unimplemented on target arch");
7212 // Get the size of the SIMD type in bytes
7213 int getSIMDTypeSizeInBytes(CORINFO_CLASS_HANDLE typeHnd)
7215 unsigned sizeBytes = 0;
7216 (void)getBaseTypeAndSizeOfSIMDType(typeHnd, &sizeBytes);
7220 // Get the the number of elements of basetype of SIMD vector given by its size and baseType
7221 static int getSIMDVectorLength(unsigned simdSize, var_types baseType);
7223 // Get the the number of elements of basetype of SIMD vector given by its type handle
7224 int getSIMDVectorLength(CORINFO_CLASS_HANDLE typeHnd);
7226 // Get preferred alignment of SIMD type.
7227 int getSIMDTypeAlignment(var_types simdType);
7229 // Get the number of bytes in a SIMD Vector for the current compilation.
7230 unsigned getSIMDVectorRegisterByteLength()
7232 #if defined(_TARGET_XARCH_) && !defined(LEGACY_BACKEND)
7235 return YMM_REGSIZE_BYTES;
7239 assert(canUseSSE2());
7240 return XMM_REGSIZE_BYTES;
7243 assert(!"getSIMDVectorRegisterByteLength() unimplemented on target arch");
7248 // The minimum and maximum possible number of bytes in a SIMD vector.
7249 unsigned int maxSIMDStructBytes()
7251 return getSIMDVectorRegisterByteLength();
7253 unsigned int minSIMDStructBytes()
7255 return emitTypeSize(TYP_SIMD8);
7258 #ifdef FEATURE_AVX_SUPPORT
7259 // (maxPossibleSIMDStructBytes is for use in a context that requires a compile-time constant.)
7260 static const unsigned maxPossibleSIMDStructBytes = 32;
7261 #else // !FEATURE_AVX_SUPPORT
7262 static const unsigned maxPossibleSIMDStructBytes = 16;
7263 #endif // !FEATURE_AVX_SUPPORT
7265 // Returns the codegen type for a given SIMD size.
7266 var_types getSIMDTypeForSize(unsigned size)
7268 var_types simdType = TYP_UNDEF;
7271 simdType = TYP_SIMD8;
7273 else if (size == 12)
7275 simdType = TYP_SIMD12;
7277 else if (size == 16)
7279 simdType = TYP_SIMD16;
7281 #ifdef FEATURE_AVX_SUPPORT
7282 else if (size == 32)
7284 simdType = TYP_SIMD32;
7286 #endif // FEATURE_AVX_SUPPORT
7289 noway_assert(!"Unexpected size for SIMD type");
7294 unsigned getSIMDInitTempVarNum()
7296 if (lvaSIMDInitTempVarNum == BAD_VAR_NUM)
7298 lvaSIMDInitTempVarNum = lvaGrabTempWithImplicitUse(false DEBUGARG("SIMDInitTempVar"));
7299 lvaTable[lvaSIMDInitTempVarNum].lvType = getSIMDVectorType();
7301 return lvaSIMDInitTempVarNum;
7304 #endif // FEATURE_SIMD
7307 //------------------------------------------------------------------------
7308 // largestEnregisterableStruct: The size in bytes of the largest struct that can be enregistered.
7310 // Notes: It is not guaranteed that the struct of this size or smaller WILL be a
7311 // candidate for enregistration.
7313 unsigned largestEnregisterableStructSize()
7316 unsigned vectorRegSize = getSIMDVectorRegisterByteLength();
7317 if (vectorRegSize > TARGET_POINTER_SIZE)
7319 return vectorRegSize;
7322 #endif // FEATURE_SIMD
7324 return TARGET_POINTER_SIZE;
7329 // These routines need not be enclosed under FEATURE_SIMD since lvIsSIMDType()
7330 // is defined for both FEATURE_SIMD and !FEATURE_SIMD apropriately. The use
7331 // of this routines also avoids the need of #ifdef FEATURE_SIMD specific code.
7333 // Is this var is of type simd struct?
7334 bool lclVarIsSIMDType(unsigned varNum)
7336 LclVarDsc* varDsc = lvaTable + varNum;
7337 return varDsc->lvIsSIMDType();
7340 // Is this Local node a SIMD local?
7341 bool lclVarIsSIMDType(GenTreeLclVarCommon* lclVarTree)
7343 return lclVarIsSIMDType(lclVarTree->gtLclNum);
7346 // Returns true if the TYP_SIMD locals on stack are aligned at their
7347 // preferred byte boundary specified by getSIMDTypeAlignment().
7349 // As per the Intel manual, the preferred alignment for AVX vectors is 32-bytes. On Amd64,
7350 // RSP/EBP is aligned at 16-bytes, therefore to align SIMD types at 32-bytes we need even
7351 // RSP/EBP to be 32-byte aligned. It is not clear whether additional stack space used in
7352 // aligning stack is worth the benefit and for now will use 16-byte alignment for AVX
7353 // 256-bit vectors with unaligned load/stores to/from memory. On x86, the stack frame
7354 // is aligned to 4 bytes. We need to extend existing support for double (8-byte) alignment
7355 // to 16 or 32 byte alignment for frames with local SIMD vars, if that is determined to be
7358 bool isSIMDTypeLocalAligned(unsigned varNum)
7360 #if defined(FEATURE_SIMD) && ALIGN_SIMD_TYPES
7361 if (lclVarIsSIMDType(varNum) && lvaTable[varNum].lvType != TYP_BYREF)
7364 int off = lvaFrameAddress(varNum, &ebpBased);
7365 // TODO-Cleanup: Can't this use the lvExactSize on the varDsc?
7366 int alignment = getSIMDTypeAlignment(lvaTable[varNum].lvType);
7367 bool isAligned = (alignment <= STACK_ALIGN) && ((off % alignment) == 0);
7370 #endif // FEATURE_SIMD
7375 // Whether SSE2 is available
7376 bool canUseSSE2() const
7378 #ifdef _TARGET_XARCH_
7379 return opts.compCanUseSSE2;
7385 // Whether SSE3, SSE3, SSE4.1 and SSE4.2 is available
7386 bool CanUseSSE3_4() const
7388 #ifdef _TARGET_XARCH_
7389 return opts.compCanUseSSE3_4;
7395 bool canUseAVX() const
7397 #ifdef FEATURE_AVX_SUPPORT
7398 return opts.compCanUseAVX;
7405 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7406 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7410 XX Generic info about the compilation and the method being compiled. XX
7411 XX It is responsible for driving the other phases. XX
7412 XX It is also responsible for all the memory management. XX
7414 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7415 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7419 Compiler* InlineeCompiler; // The Compiler instance for the inlinee
7421 InlineResult* compInlineResult; // The result of importing the inlinee method.
7423 bool compDoAggressiveInlining; // If true, mark every method as CORINFO_FLG_FORCEINLINE
7424 bool compJmpOpUsed; // Does the method do a JMP
7425 bool compLongUsed; // Does the method use TYP_LONG
7426 bool compFloatingPointUsed; // Does the method use TYP_FLOAT or TYP_DOUBLE
7427 bool compTailCallUsed; // Does the method do a tailcall
7428 bool compLocallocUsed; // Does the method use localloc.
7429 bool compQmarkUsed; // Does the method use GT_QMARK/GT_COLON
7430 bool compQmarkRationalized; // Is it allowed to use a GT_QMARK/GT_COLON node.
7431 bool compUnsafeCastUsed; // Does the method use LDIND/STIND to cast between scalar/refernce types
7433 // NOTE: These values are only reliable after
7434 // the importing is completely finished.
7436 ExpandArrayStack<GenTreePtr>* compQMarks; // The set of QMark nodes created in the current compilation, so
7437 // we can iterate over these efficiently.
7439 #if CPU_USES_BLOCK_MOVE
7440 bool compBlkOpUsed; // Does the method do a COPYBLK or INITBLK
7444 // State information - which phases have completed?
7445 // These are kept together for easy discoverability
7447 bool bRangeAllowStress;
7448 bool compCodeGenDone;
7449 int64_t compNumStatementLinksTraversed; // # of links traversed while doing debug checks
7450 bool fgNormalizeEHDone; // Has the flowgraph EH normalization phase been done?
7451 size_t compSizeEstimate; // The estimated size of the method as per `gtSetEvalOrder`.
7452 size_t compCycleEstimate; // The estimated cycle count of the method as per `gtSetEvalOrder`
7455 bool fgLocalVarLivenessDone; // Note that this one is used outside of debug.
7456 bool fgLocalVarLivenessChanged;
7458 bool compStackProbePrologDone;
7460 #ifndef LEGACY_BACKEND
7462 #endif // !LEGACY_BACKEND
7463 bool compRationalIRForm;
7465 bool compUsesThrowHelper; // There is a call to a THOROW_HELPER for the compiled method.
7467 bool compGeneratingProlog;
7468 bool compGeneratingEpilog;
7469 bool compNeedsGSSecurityCookie; // There is an unsafe buffer (or localloc) on the stack.
7470 // Insert cookie on frame and code to check the cookie, like VC++ -GS.
7471 bool compGSReorderStackLayout; // There is an unsafe buffer on the stack, reorder locals and make local
7472 // copies of susceptible parameters to avoid buffer overrun attacks through locals/params
7473 bool getNeedsGSSecurityCookie() const
7475 return compNeedsGSSecurityCookie;
7477 void setNeedsGSSecurityCookie()
7479 compNeedsGSSecurityCookie = true;
7482 FrameLayoutState lvaDoneFrameLayout; // The highest frame layout state that we've completed. During
7483 // frame layout calculations, this is the level we are currently
7486 //---------------------------- JITing options -----------------------------
7499 JitFlags* jitFlags; // all flags passed from the EE
7500 unsigned compFlags; // method attributes
7502 codeOptimize compCodeOpt; // what type of code optimizations
7506 #ifdef _TARGET_XARCH_
7507 bool compCanUseSSE2; // Allow CodeGen to use "movq XMM" instructions
7508 bool compCanUseSSE3_4; // Allow CodeGen to use SSE3, SSSE3, SSE4.1 and SSE4.2 instructions
7510 #ifdef FEATURE_AVX_SUPPORT
7511 bool compCanUseAVX; // Allow CodeGen to use AVX 256-bit vectors for SIMD operations
7512 #endif // FEATURE_AVX_SUPPORT
7513 #endif // _TARGET_XARCH_
7515 // optimize maximally and/or favor speed over size?
7517 #define DEFAULT_MIN_OPTS_CODE_SIZE 60000
7518 #define DEFAULT_MIN_OPTS_INSTR_COUNT 20000
7519 #define DEFAULT_MIN_OPTS_BB_COUNT 2000
7520 #define DEFAULT_MIN_OPTS_LV_NUM_COUNT 2000
7521 #define DEFAULT_MIN_OPTS_LV_REF_COUNT 8000
7523 // Maximun number of locals before turning off the inlining
7524 #define MAX_LV_NUM_COUNT_FOR_INLINING 512
7527 unsigned instrCount;
7528 unsigned lvRefCount;
7529 bool compMinOptsIsSet;
7531 bool compMinOptsIsUsed;
7533 inline bool MinOpts()
7535 assert(compMinOptsIsSet);
7536 compMinOptsIsUsed = true;
7539 inline bool IsMinOptsSet()
7541 return compMinOptsIsSet;
7544 inline bool MinOpts()
7548 inline bool IsMinOptsSet()
7550 return compMinOptsIsSet;
7553 inline void SetMinOpts(bool val)
7555 assert(!compMinOptsIsUsed);
7556 assert(!compMinOptsIsSet || (compMinOpts == val));
7558 compMinOptsIsSet = true;
7561 // true if the CLFLG_* for an optimization is set.
7562 inline bool OptEnabled(unsigned optFlag)
7564 return !!(compFlags & optFlag);
7567 #ifdef FEATURE_READYTORUN_COMPILER
7568 inline bool IsReadyToRun()
7570 return jitFlags->IsSet(JitFlags::JIT_FLAG_READYTORUN);
7573 inline bool IsReadyToRun()
7579 // true if we should use the PINVOKE_{BEGIN,END} helpers instead of generating
7580 // PInvoke transitions inline (e.g. when targeting CoreRT).
7581 inline bool ShouldUsePInvokeHelpers()
7583 #if COR_JIT_EE_VERSION > 460
7584 return jitFlags->IsSet(JitFlags::JIT_FLAG_USE_PINVOKE_HELPERS);
7590 // true if we should use insert the REVERSE_PINVOKE_{ENTER,EXIT} helpers in the method
7592 inline bool IsReversePInvoke()
7594 #if COR_JIT_EE_VERSION > 460
7595 return jitFlags->IsSet(JitFlags::JIT_FLAG_REVERSE_PINVOKE);
7601 // true if we must generate code compatible with JIT32 quirks
7602 inline bool IsJit32Compat()
7604 #if defined(_TARGET_X86_) && COR_JIT_EE_VERSION > 460
7605 return jitFlags->IsSet(JitFlags::JIT_FLAG_DESKTOP_QUIRKS);
7611 // true if we must generate code compatible with Jit64 quirks
7612 inline bool IsJit64Compat()
7614 #if defined(_TARGET_AMD64_) && COR_JIT_EE_VERSION > 460
7615 return jitFlags->IsSet(JitFlags::JIT_FLAG_DESKTOP_QUIRKS);
7616 #elif defined(_TARGET_AMD64_) && !defined(FEATURE_CORECLR)
7623 bool compScopeInfo; // Generate the LocalVar info ?
7624 bool compDbgCode; // Generate debugger-friendly code?
7625 bool compDbgInfo; // Gather debugging info?
7628 #ifdef PROFILING_SUPPORTED
7629 bool compNoPInvokeInlineCB;
7631 static const bool compNoPInvokeInlineCB;
7635 bool compGcChecks; // Check arguments and return values to ensure they are sane
7636 bool compStackCheckOnRet; // Check ESP on return to ensure it is correct
7637 bool compStackCheckOnCall; // Check ESP after every call to ensure it is correct
7641 bool compNeedSecurityCheck; // This flag really means where or not a security object needs
7642 // to be allocated on the stack.
7643 // It will be set to true in the following cases:
7644 // 1. When the method being compiled has a declarative security
7645 // (i.e. when CORINFO_FLG_NOSECURITYWRAP is reset for the current method).
7646 // This is also the case when we inject a prolog and epilog in the method.
7648 // 2. When the method being compiled has imperative security (i.e. the method
7649 // calls into another method that has CORINFO_FLG_SECURITYCHECK flag set).
7651 // 3. When opts.compDbgEnC is true. (See also Compiler::compCompile).
7653 // When this flag is set, jit will allocate a gc-reference local variable (lvaSecurityObject),
7654 // which gets reported as a GC root to stackwalker.
7655 // (See also ICodeManager::GetAddrOfSecurityObject.)
7662 #if defined(_TARGET_XARCH_) && !defined(LEGACY_BACKEND)
7663 bool compEnablePCRelAddr; // Whether absolute addr be encoded as PC-rel offset by RyuJIT where possible
7667 #ifdef UNIX_AMD64_ABI
7668 // This flag is indicating if there is a need to align the frame.
7669 // On AMD64-Windows, if there are calls, 4 slots for the outgoing ars are allocated, except for
7670 // FastTailCall. This slots makes the frame size non-zero, so alignment logic will be called.
7671 // On AMD64-Unix, there are no such slots. There is a possibility to have calls in the method with frame size of
7672 // 0. The frame alignment logic won't kick in. This flags takes care of the AMD64-Unix case by remembering that
7673 // there are calls and making sure the frame alignment logic is executed.
7674 bool compNeedToAlignFrame;
7675 #endif // UNIX_AMD64_ABI
7677 bool compProcedureSplitting; // Separate cold code from hot code
7679 bool genFPorder; // Preserve FP order (operations are non-commutative)
7680 bool genFPopt; // Can we do frame-pointer-omission optimization?
7681 bool altJit; // True if we are an altjit and are compiling this method
7684 bool optRepeat; // Repeat optimizer phases k times
7685 bool compProcedureSplittingEH; // Separate cold code from hot code for functions with EH
7686 bool dspCode; // Display native code generated
7687 bool dspEHTable; // Display the EH table reported to the VM
7688 bool dspInstrs; // Display the IL instructions intermixed with the native code output
7689 bool dspEmit; // Display emitter output
7690 bool dspLines; // Display source-code lines intermixed with native code output
7691 bool dmpHex; // Display raw bytes in hex of native code output
7692 bool varNames; // Display variables names in native code output
7693 bool disAsm; // Display native code as it is generated
7694 bool disAsmSpilled; // Display native code when any register spilling occurs
7695 bool disDiffable; // Makes the Disassembly code 'diff-able'
7696 bool disAsm2; // Display native code after it is generated using external disassembler
7697 bool dspOrder; // Display names of each of the methods that we ngen/jit
7698 bool dspUnwind; // Display the unwind info output
7699 bool dspDiffable; // Makes the Jit Dump 'diff-able' (currently uses same COMPlus_* flag as disDiffable)
7700 bool compLongAddress; // Force using large pseudo instructions for long address
7701 // (IF_LARGEJMP/IF_LARGEADR/IF_LARGLDC)
7702 bool dspGCtbls; // Display the GC tables
7706 bool doLateDisasm; // Run the late disassembler
7707 #endif // LATE_DISASM
7709 #if DUMP_GC_TABLES && !defined(DEBUG) && defined(JIT32_GCENCODER)
7710 // Only the JIT32_GCENCODER implements GC dumping in non-DEBUG code.
7711 #pragma message("NOTE: this non-debug build has GC ptr table dumping always enabled!")
7712 static const bool dspGCtbls = true;
7715 // We need stack probes to guarantee that we won't trigger a stack overflow
7716 // when calling unmanaged code until they get a chance to set up a frame, because
7717 // the EE will have no idea where it is.
7719 // We will only be doing this currently for hosted environments. Unfortunately
7720 // we need to take care of stubs, so potentially, we will have to do the probes
7721 // for any call. We have a plan for not needing for stubs though
7722 bool compNeedStackProbes;
7724 #ifdef PROFILING_SUPPORTED
7725 // Whether to emit Enter/Leave/TailCall hooks using a dummy stub (DummyProfilerELTStub()).
7726 // This option helps make the JIT behave as if it is running under a profiler.
7727 bool compJitELTHookEnabled;
7728 #endif // PROFILING_SUPPORTED
7730 #if FEATURE_TAILCALL_OPT
7731 // Whether opportunistic or implicit tail call optimization is enabled.
7732 bool compTailCallOpt;
7733 // Whether optimization of transforming a recursive tail call into a loop is enabled.
7734 bool compTailCallLoopOpt;
7738 static const bool compUseSoftFP = true;
7739 #else // !ARM_SOFTFP
7740 static const bool compUseSoftFP = false;
7743 GCPollType compGCPollType;
7747 static bool s_pAltJitExcludeAssembliesListInitialized;
7748 static AssemblyNamesList2* s_pAltJitExcludeAssembliesList;
7753 template <typename T>
7756 return (p == ZERO) ? ZERO : (opts.dspDiffable ? T(0xD1FFAB1E) : p);
7759 template <typename T>
7762 return (o == ZERO) ? ZERO : (opts.dspDiffable ? T(0xD1FFAB1E) : o);
7765 static int dspTreeID(GenTree* tree)
7767 return tree->gtTreeID;
7769 static void printTreeID(GenTree* tree)
7771 if (tree == nullptr)
7777 printf("[%06d]", dspTreeID(tree));
7784 #define STRESS_MODES \
7788 /* "Variations" stress areas which we try to mix up with each other. */ \
7789 /* These should not be exhaustively used as they might */ \
7790 /* hide/trivialize other areas */ \
7792 STRESS_MODE(REGS) STRESS_MODE(DBL_ALN) STRESS_MODE(LCL_FLDS) STRESS_MODE(UNROLL_LOOPS) \
7793 STRESS_MODE(MAKE_CSE) STRESS_MODE(LEGACY_INLINE) STRESS_MODE(CLONE_EXPR) \
7794 STRESS_MODE(USE_FCOMI) STRESS_MODE(USE_CMOV) STRESS_MODE(FOLD) \
7795 STRESS_MODE(BB_PROFILE) STRESS_MODE(OPT_BOOLS_GC) STRESS_MODE(REMORPH_TREES) \
7796 STRESS_MODE(64RSLT_MUL) STRESS_MODE(DO_WHILE_LOOPS) STRESS_MODE(MIN_OPTS) \
7797 STRESS_MODE(REVERSE_FLAG) /* Will set GTF_REVERSE_OPS whenever we can */ \
7798 STRESS_MODE(REVERSE_COMMA) /* Will reverse commas created with gtNewCommaNode */ \
7799 STRESS_MODE(TAILCALL) /* Will make the call as a tailcall whenever legal */ \
7800 STRESS_MODE(CATCH_ARG) /* Will spill catch arg */ \
7801 STRESS_MODE(UNSAFE_BUFFER_CHECKS) \
7802 STRESS_MODE(NULL_OBJECT_CHECK) \
7803 STRESS_MODE(PINVOKE_RESTORE_ESP) \
7804 STRESS_MODE(RANDOM_INLINE) \
7806 STRESS_MODE(GENERIC_VARN) STRESS_MODE(COUNT_VARN) \
7808 /* "Check" stress areas that can be exhaustively used if we */ \
7809 /* dont care about performance at all */ \
7811 STRESS_MODE(FORCE_INLINE) /* Treat every method as AggressiveInlining */ \
7812 STRESS_MODE(CHK_FLOW_UPDATE) \
7813 STRESS_MODE(EMITTER) STRESS_MODE(CHK_REIMPORT) STRESS_MODE(FLATFP) \
7815 STRESS_MODE(GENERIC_CHECK) STRESS_MODE(COUNT) \
7819 #define STRESS_MODE(mode) STRESS_##mode,
7826 static const LPCWSTR s_compStressModeNames[STRESS_COUNT + 1];
7827 BYTE compActiveStressModes[STRESS_COUNT];
7830 #define MAX_STRESS_WEIGHT 100
7832 bool compStressCompile(compStressArea stressArea, unsigned weightPercentage);
7836 bool compInlineStress()
7838 return compStressCompile(STRESS_LEGACY_INLINE, 50);
7841 bool compRandomInlineStress()
7843 return compStressCompile(STRESS_RANDOM_INLINE, 50);
7848 bool compTailCallStress()
7851 return (JitConfig.TailcallStress() != 0 || compStressCompile(STRESS_TAILCALL, 5));
7857 codeOptimize compCodeOpt()
7860 // Switching between size & speed has measurable throughput impact
7861 // (3.5% on NGen mscorlib when measured). It used to be enabled for
7862 // DEBUG, but should generate identical code between CHK & RET builds,
7863 // so that's not acceptable.
7864 // TODO-Throughput: Figure out what to do about size vs. speed & throughput.
7865 // Investigate the cause of the throughput regression.
7867 return opts.compCodeOpt;
7869 return BLENDED_CODE;
7873 //--------------------- Info about the procedure --------------------------
7877 COMP_HANDLE compCompHnd;
7878 CORINFO_MODULE_HANDLE compScopeHnd;
7879 CORINFO_CLASS_HANDLE compClassHnd;
7880 CORINFO_METHOD_HANDLE compMethodHnd;
7881 CORINFO_METHOD_INFO* compMethodInfo;
7883 BOOL hasCircularClassConstraints;
7884 BOOL hasCircularMethodConstraints;
7886 #if defined(DEBUG) || defined(LATE_DISASM)
7887 const char* compMethodName;
7888 const char* compClassName;
7889 const char* compFullName;
7890 #endif // defined(DEBUG) || defined(LATE_DISASM)
7892 #if defined(DEBUG) || defined(INLINE_DATA)
7893 // Method hash is logcally const, but computed
7895 mutable unsigned compMethodHashPrivate;
7896 unsigned compMethodHash() const;
7897 #endif // defined(DEBUG) || defined(INLINE_DATA)
7899 #ifdef PSEUDORANDOM_NOP_INSERTION
7900 // things for pseudorandom nop insertion
7901 unsigned compChecksum;
7905 // The following holds the FLG_xxxx flags for the method we're compiling.
7908 // The following holds the class attributes for the method we're compiling.
7909 unsigned compClassAttr;
7911 const BYTE* compCode;
7912 IL_OFFSET compILCodeSize; // The IL code size
7913 UNATIVE_OFFSET compNativeCodeSize; // The native code size, after instructions are issued. This
7914 // is less than (compTotalHotCodeSize + compTotalColdCodeSize) only if:
7915 // (1) the code is not hot/cold split, and we issued less code than we expected, or
7916 // (2) the code is hot/cold split, and we issued less code than we expected
7917 // in the cold section (the hot section will always be padded out to compTotalHotCodeSize).
7919 bool compIsStatic : 1; // Is the method static (no 'this' pointer)?
7920 bool compIsVarArgs : 1; // Does the method have varargs parameters?
7921 bool compIsContextful : 1; // contextful method
7922 bool compInitMem : 1; // Is the CORINFO_OPT_INIT_LOCALS bit set in the method info options?
7923 bool compUnwrapContextful : 1; // JIT should unwrap proxies when possible
7924 bool compProfilerCallback : 1; // JIT inserted a profiler Enter callback
7925 bool compPublishStubParam : 1; // EAX captured in prolog will be available through an instrinsic
7926 bool compRetBuffDefStack : 1; // The ret buff argument definitely points into the stack.
7928 var_types compRetType; // Return type of the method as declared in IL
7929 var_types compRetNativeType; // Normalized return type as per target arch ABI
7930 unsigned compILargsCount; // Number of arguments (incl. implicit but not hidden)
7931 unsigned compArgsCount; // Number of arguments (incl. implicit and hidden)
7932 unsigned compRetBuffArg; // position of hidden return param var (0, 1) (BAD_VAR_NUM means not present);
7933 int compTypeCtxtArg; // position of hidden param for type context for generic code (CORINFO_CALLCONV_PARAMTYPE)
7934 unsigned compThisArg; // position of implicit this pointer param (not to be confused with lvaArg0Var)
7935 unsigned compILlocalsCount; // Number of vars : args + locals (incl. implicit but not hidden)
7936 unsigned compLocalsCount; // Number of vars : args + locals (incl. implicit and hidden)
7937 unsigned compMaxStack;
7938 UNATIVE_OFFSET compTotalHotCodeSize; // Total number of bytes of Hot Code in the method
7939 UNATIVE_OFFSET compTotalColdCodeSize; // Total number of bytes of Cold Code in the method
7941 unsigned compCallUnmanaged; // count of unmanaged calls
7942 unsigned compLvFrameListRoot; // lclNum for the Frame root
7943 unsigned compXcptnsCount; // Number of exception-handling clauses read in the method's IL.
7944 // You should generally use compHndBBtabCount instead: it is the
7945 // current number of EH clauses (after additions like synchronized
7946 // methods and funclets, and removals like unreachable code deletion).
7948 bool compMatchedVM; // true if the VM is "matched": either the JIT is a cross-compiler
7949 // and the VM expects that, or the JIT is a "self-host" compiler
7950 // (e.g., x86 hosted targeting x86) and the VM expects that.
7952 /* The following holds IL scope information about local variables.
7955 unsigned compVarScopesCount;
7956 VarScopeDsc* compVarScopes;
7958 /* The following holds information about instr offsets for
7959 * which we need to report IP-mappings
7962 IL_OFFSET* compStmtOffsets; // sorted
7963 unsigned compStmtOffsetsCount;
7964 ICorDebugInfo::BoundaryTypes compStmtOffsetsImplicit;
7966 #define CPU_X86 0x0100 // The generic X86 CPU
7967 #define CPU_X86_PENTIUM_4 0x0110
7969 #define CPU_X64 0x0200 // The generic x64 CPU
7970 #define CPU_AMD_X64 0x0210 // AMD x64 CPU
7971 #define CPU_INTEL_X64 0x0240 // Intel x64 CPU
7973 #define CPU_ARM 0x0300 // The generic ARM CPU
7975 unsigned genCPU; // What CPU are we running on
7978 // Returns true if the method being compiled returns a non-void and non-struct value.
7979 // Note that lvaInitTypeRef() normalizes compRetNativeType for struct returns in a
7980 // single register as per target arch ABI (e.g on Amd64 Windows structs of size 1, 2,
7981 // 4 or 8 gets normalized to TYP_BYTE/TYP_SHORT/TYP_INT/TYP_LONG; On Arm HFA structs).
7982 // Methods returning such structs are considered to return non-struct return value and
7983 // this method returns true in that case.
7984 bool compMethodReturnsNativeScalarType()
7986 return (info.compRetType != TYP_VOID) && !varTypeIsStruct(info.compRetNativeType);
7989 // Returns true if the method being compiled returns RetBuf addr as its return value
7990 bool compMethodReturnsRetBufAddr()
7992 // There are cases where implicit RetBuf argument should be explicitly returned in a register.
7993 // In such cases the return type is changed to TYP_BYREF and appropriate IR is generated.
7995 // 1. Profiler Leave calllback expects the address of retbuf as return value for
7996 // methods with hidden RetBuf argument. impReturnInstruction() when profiler
7997 // callbacks are needed creates GT_RETURN(TYP_BYREF, op1 = Addr of RetBuf) for
7998 // methods with hidden RetBufArg.
8000 // 2. As per the System V ABI, the address of RetBuf needs to be returned by
8001 // methods with hidden RetBufArg in RAX. In such case GT_RETURN is of TYP_BYREF,
8002 // returning the address of RetBuf.
8004 // 3. Windows 64-bit native calling convention also requires the address of RetBuff
8005 // to be returned in RAX.
8006 CLANG_FORMAT_COMMENT_ANCHOR;
8008 #ifdef _TARGET_AMD64_
8009 return (info.compRetBuffArg != BAD_VAR_NUM);
8010 #else // !_TARGET_AMD64_
8011 return (compIsProfilerHookNeeded()) && (info.compRetBuffArg != BAD_VAR_NUM);
8012 #endif // !_TARGET_AMD64_
8015 // Returns true if the method returns a value in more than one return register
8016 // TODO-ARM-Bug: Deal with multi-register genReturnLocaled structs?
8017 // TODO-ARM64: Does this apply for ARM64 too?
8018 bool compMethodReturnsMultiRegRetType()
8020 #if FEATURE_MULTIREG_RET
8021 #if defined(_TARGET_X86_)
8022 // On x86 only 64-bit longs are returned in multiple registers
8023 return varTypeIsLong(info.compRetNativeType);
8024 #else // targets: X64-UNIX, ARM64 or ARM32
8025 // On all other targets that support multireg return values:
8026 // Methods returning a struct in multiple registers have a return value of TYP_STRUCT.
8027 // Such method's compRetNativeType is TYP_STRUCT without a hidden RetBufArg
8028 return varTypeIsStruct(info.compRetNativeType) && (info.compRetBuffArg == BAD_VAR_NUM);
8029 #endif // TARGET_XXX
8031 #else // not FEATURE_MULTIREG_RET
8033 // For this architecture there are no multireg returns
8036 #endif // FEATURE_MULTIREG_RET
8039 #if FEATURE_MULTIREG_ARGS
8040 // Given a GenTree node of TYP_STRUCT that represents a pass by value argument
8041 // return the gcPtr layout for the pointers sized fields
8042 void getStructGcPtrsFromOp(GenTreePtr op, BYTE* gcPtrsOut);
8043 #endif // FEATURE_MULTIREG_ARGS
8045 // Returns true if the method being compiled returns a value
8046 bool compMethodHasRetVal()
8048 return compMethodReturnsNativeScalarType() || compMethodReturnsRetBufAddr() ||
8049 compMethodReturnsMultiRegRetType();
8054 void compDispLocalVars();
8058 //-------------------------- Global Compiler Data ------------------------------------
8061 static unsigned s_compMethodsCount; // to produce unique label names
8062 unsigned compGenTreeID;
8065 BasicBlock* compCurBB; // the current basic block in process
8066 GenTreePtr compCurStmt; // the current statement in process
8068 unsigned compCurStmtNum; // to give all statements an increasing StmtNum when printing dumps
8071 // The following is used to create the 'method JIT info' block.
8072 size_t compInfoBlkSize;
8073 BYTE* compInfoBlkAddr;
8075 EHblkDsc* compHndBBtab; // array of EH data
8076 unsigned compHndBBtabCount; // element count of used elements in EH data array
8077 unsigned compHndBBtabAllocCount; // element count of allocated elements in EH data array
8079 #if defined(_TARGET_X86_)
8081 //-------------------------------------------------------------------------
8082 // Tracking of region covered by the monitor in synchronized methods
8083 void* syncStartEmitCookie; // the emitter cookie for first instruction after the call to MON_ENTER
8084 void* syncEndEmitCookie; // the emitter cookie for first instruction after the call to MON_EXIT
8086 #endif // !_TARGET_X86_
8088 Phases previousCompletedPhase; // the most recently completed phase
8090 //-------------------------------------------------------------------------
8091 // The following keeps track of how many bytes of local frame space we've
8092 // grabbed so far in the current function, and how many argument bytes we
8093 // need to pop when we return.
8096 unsigned compLclFrameSize; // secObject+lclBlk+locals+temps
8098 // Count of callee-saved regs we pushed in the prolog.
8099 // Does not include EBP for isFramePointerUsed() and double-aligned frames.
8100 // In case of Amd64 this doesn't include float regs saved on stack.
8101 unsigned compCalleeRegsPushed;
8103 #if defined(_TARGET_XARCH_) && !FEATURE_STACK_FP_X87
8104 // Mask of callee saved float regs on stack.
8105 regMaskTP compCalleeFPRegsSavedMask;
8107 #ifdef _TARGET_AMD64_
8108 // Quirk for VS debug-launch scenario to work:
8109 // Bytes of padding between save-reg area and locals.
8110 #define VSQUIRK_STACK_PAD (2 * REGSIZE_BYTES)
8111 unsigned compVSQuirkStackPaddingNeeded;
8112 bool compQuirkForPPPflag;
8115 unsigned compArgSize; // total size of arguments in bytes (including register args (lvIsRegArg))
8117 unsigned compMapILargNum(unsigned ILargNum); // map accounting for hidden args
8118 unsigned compMapILvarNum(unsigned ILvarNum); // map accounting for hidden args
8119 unsigned compMap2ILvarNum(unsigned varNum); // map accounting for hidden args
8121 //-------------------------------------------------------------------------
8123 static void compStartup(); // One-time initialization
8124 static void compShutdown(); // One-time finalization
8126 void compInit(ArenaAllocator* pAlloc, InlineInfo* inlineInfo);
8129 static void compDisplayStaticSizes(FILE* fout);
8131 //------------ Some utility functions --------------
8133 void* compGetHelperFtn(CorInfoHelpFunc ftnNum, /* IN */
8134 void** ppIndirection); /* OUT */
8136 // Several JIT/EE interface functions return a CorInfoType, and also return a
8137 // class handle as an out parameter if the type is a value class. Returns the
8138 // size of the type these describe.
8139 unsigned compGetTypeSize(CorInfoType cit, CORINFO_CLASS_HANDLE clsHnd);
8142 // Components used by the compiler may write unit test suites, and
8143 // have them run within this method. They will be run only once per process, and only
8144 // in debug. (Perhaps should be under the control of a COMPlus_ flag.)
8145 // These should fail by asserting.
8146 void compDoComponentUnitTestsOnce();
8149 int compCompile(CORINFO_METHOD_HANDLE methodHnd,
8150 CORINFO_MODULE_HANDLE classPtr,
8151 COMP_HANDLE compHnd,
8152 CORINFO_METHOD_INFO* methodInfo,
8153 void** methodCodePtr,
8154 ULONG* methodCodeSize,
8155 JitFlags* compileFlags);
8156 void compCompileFinish();
8157 int compCompileHelper(CORINFO_MODULE_HANDLE classPtr,
8158 COMP_HANDLE compHnd,
8159 CORINFO_METHOD_INFO* methodInfo,
8160 void** methodCodePtr,
8161 ULONG* methodCodeSize,
8162 JitFlags* compileFlags,
8163 CorInfoInstantiationVerification instVerInfo);
8165 ArenaAllocator* compGetAllocator();
8167 #if MEASURE_MEM_ALLOC
8169 static bool s_dspMemStats; // Display per-phase memory statistics for every function
8173 unsigned allocCnt; // # of allocs
8174 UINT64 allocSz; // total size of those alloc.
8175 UINT64 allocSzMax; // Maximum single allocation.
8176 UINT64 allocSzByKind[CMK_Count]; // Classified by "kind".
8177 UINT64 nraTotalSizeAlloc;
8178 UINT64 nraTotalSizeUsed;
8180 static const char* s_CompMemKindNames[]; // Names of the kinds.
8182 MemStats() : allocCnt(0), allocSz(0), allocSzMax(0), nraTotalSizeAlloc(0), nraTotalSizeUsed(0)
8184 for (int i = 0; i < CMK_Count; i++)
8186 allocSzByKind[i] = 0;
8189 MemStats(const MemStats& ms)
8190 : allocCnt(ms.allocCnt)
8191 , allocSz(ms.allocSz)
8192 , allocSzMax(ms.allocSzMax)
8193 , nraTotalSizeAlloc(ms.nraTotalSizeAlloc)
8194 , nraTotalSizeUsed(ms.nraTotalSizeUsed)
8196 for (int i = 0; i < CMK_Count; i++)
8198 allocSzByKind[i] = ms.allocSzByKind[i];
8202 // Until we have ubiquitous constructors.
8205 this->MemStats::MemStats();
8208 void AddAlloc(size_t sz, CompMemKind cmk)
8212 if (sz > allocSzMax)
8216 allocSzByKind[cmk] += sz;
8219 void Print(FILE* f); // Print these stats to f.
8220 void PrintByKind(FILE* f); // Do just the by-kind histogram part.
8222 MemStats genMemStats;
8224 struct AggregateMemStats : public MemStats
8228 AggregateMemStats() : MemStats(), nMethods(0)
8232 void Add(const MemStats& ms)
8235 allocCnt += ms.allocCnt;
8236 allocSz += ms.allocSz;
8237 allocSzMax = max(allocSzMax, ms.allocSzMax);
8238 for (int i = 0; i < CMK_Count; i++)
8240 allocSzByKind[i] += ms.allocSzByKind[i];
8242 nraTotalSizeAlloc += ms.nraTotalSizeAlloc;
8243 nraTotalSizeUsed += ms.nraTotalSizeUsed;
8246 void Print(FILE* f); // Print these stats to jitstdout.
8249 static CritSecObject s_memStatsLock; // This lock protects the data structures below.
8250 static MemStats s_maxCompMemStats; // Stats for the compilation with the largest amount allocated.
8251 static AggregateMemStats s_aggMemStats; // Aggregates statistics for all compilations.
8253 #endif // MEASURE_MEM_ALLOC
8255 #if LOOP_HOIST_STATS
8256 unsigned m_loopsConsidered;
8257 bool m_curLoopHasHoistedExpression;
8258 unsigned m_loopsWithHoistedExpressions;
8259 unsigned m_totalHoistedExpressions;
8261 void AddLoopHoistStats();
8262 void PrintPerMethodLoopHoistStats();
8264 static CritSecObject s_loopHoistStatsLock; // This lock protects the data structures below.
8265 static unsigned s_loopsConsidered;
8266 static unsigned s_loopsWithHoistedExpressions;
8267 static unsigned s_totalHoistedExpressions;
8269 static void PrintAggregateLoopHoistStats(FILE* f);
8270 #endif // LOOP_HOIST_STATS
8272 void* compGetMemArray(size_t numElem, size_t elemSize, CompMemKind cmk = CMK_Unknown);
8273 void* compGetMemArrayA(size_t numElem, size_t elemSize, CompMemKind cmk = CMK_Unknown);
8274 void* compGetMem(size_t sz, CompMemKind cmk = CMK_Unknown);
8275 void* compGetMemA(size_t sz, CompMemKind cmk = CMK_Unknown);
8276 static void* compGetMemCallback(void*, size_t, CompMemKind cmk = CMK_Unknown);
8277 void compFreeMem(void*);
8279 bool compIsForImportOnly();
8280 bool compIsForInlining();
8281 bool compDonotInline();
8284 const char* compLocalVarName(unsigned varNum, unsigned offs);
8285 VarName compVarName(regNumber reg, bool isFloatReg = false);
8286 const char* compRegVarName(regNumber reg, bool displayVar = false, bool isFloatReg = false);
8287 const char* compRegPairName(regPairNo regPair);
8288 const char* compRegNameForSize(regNumber reg, size_t size);
8289 const char* compFPregVarName(unsigned fpReg, bool displayVar = false);
8290 void compDspSrcLinesByNativeIP(UNATIVE_OFFSET curIP);
8291 void compDspSrcLinesByLineNum(unsigned line, bool seek = false);
8294 //-------------------------------------------------------------------------
8296 typedef ListNode<VarScopeDsc*> VarScopeListNode;
8298 struct VarScopeMapInfo
8300 VarScopeListNode* head;
8301 VarScopeListNode* tail;
8302 static VarScopeMapInfo* Create(VarScopeListNode* node, IAllocator* alloc)
8304 VarScopeMapInfo* info = new (alloc) VarScopeMapInfo;
8311 // Max value of scope count for which we would use linear search; for larger values we would use hashtable lookup.
8312 static const unsigned MAX_LINEAR_FIND_LCL_SCOPELIST = 32;
8314 typedef SimplerHashTable<unsigned, SmallPrimitiveKeyFuncs<unsigned>, VarScopeMapInfo*, JitSimplerHashBehavior>
8315 VarNumToScopeDscMap;
8317 // Map to keep variables' scope indexed by varNum containing it's scope dscs at the index.
8318 VarNumToScopeDscMap* compVarScopeMap;
8320 VarScopeDsc* compFindLocalVar(unsigned varNum, unsigned lifeBeg, unsigned lifeEnd);
8322 VarScopeDsc* compFindLocalVar(unsigned varNum, unsigned offs);
8324 VarScopeDsc* compFindLocalVarLinear(unsigned varNum, unsigned offs);
8326 void compInitVarScopeMap();
8328 VarScopeDsc** compEnterScopeList; // List has the offsets where variables
8329 // enter scope, sorted by instr offset
8330 unsigned compNextEnterScope;
8332 VarScopeDsc** compExitScopeList; // List has the offsets where variables
8333 // go out of scope, sorted by instr offset
8334 unsigned compNextExitScope;
8336 void compInitScopeLists();
8338 void compResetScopeLists();
8340 VarScopeDsc* compGetNextEnterScope(unsigned offs, bool scan = false);
8342 VarScopeDsc* compGetNextExitScope(unsigned offs, bool scan = false);
8344 void compProcessScopesUntil(unsigned offset,
8346 void (Compiler::*enterScopeFn)(VARSET_TP* inScope, VarScopeDsc*),
8347 void (Compiler::*exitScopeFn)(VARSET_TP* inScope, VarScopeDsc*));
8350 void compDispScopeLists();
8353 bool compIsProfilerHookNeeded();
8355 //-------------------------------------------------------------------------
8356 /* Statistical Data Gathering */
8358 void compJitStats(); // call this function and enable
8359 // various ifdef's below for statistical data
8362 void compCallArgStats();
8363 static void compDispCallArgStats(FILE* fout);
8366 //-------------------------------------------------------------------------
8373 ArenaAllocator* compAllocator;
8376 // This one presents an implementation of the "IAllocator" abstract class that uses "compAllocator",
8377 // suitable for use by utilcode collection types.
8378 IAllocator* compAsIAllocator;
8380 #if MEASURE_MEM_ALLOC
8381 IAllocator* compAsIAllocatorBitset; // An allocator that uses the CMK_bitset tracker.
8382 IAllocator* compAsIAllocatorGC; // An allocator that uses the CMK_GC tracker.
8383 IAllocator* compAsIAllocatorLoopHoist; // An allocator that uses the CMK_LoopHoist tracker.
8385 IAllocator* compAsIAllocatorDebugOnly; // An allocator that uses the CMK_DebugOnly tracker.
8387 #endif // MEASURE_MEM_ALLOC
8389 void compFunctionTraceStart();
8390 void compFunctionTraceEnd(void* methodCodePtr, ULONG methodCodeSize, bool isNYI);
8393 size_t compMaxUncheckedOffsetForNullObject;
8395 void compInitOptions(JitFlags* compileFlags);
8397 void compSetProcessor();
8398 void compInitDebuggingInfo();
8399 void compSetOptimizationLevel();
8400 #ifdef _TARGET_ARMARCH_
8401 bool compRsvdRegCheck(FrameLayoutState curState);
8403 void compCompile(void** methodCodePtr, ULONG* methodCodeSize, JitFlags* compileFlags);
8405 // Clear annotations produced during optimizations; to be used between iterations when repeating opts.
8406 void ResetOptAnnotations();
8408 // Regenerate loop descriptors; to be used between iterations when repeating opts.
8409 void RecomputeLoopInfo();
8411 #ifdef PROFILING_SUPPORTED
8412 // Data required for generating profiler Enter/Leave/TailCall hooks
8414 bool compProfilerHookNeeded; // Whether profiler Enter/Leave/TailCall hook needs to be generated for the method
8415 void* compProfilerMethHnd; // Profiler handle of the method being compiled. Passed as param to ELT callbacks
8416 bool compProfilerMethHndIndirected; // Whether compProfilerHandle is pointer to the handle or is an actual handle
8419 #ifdef _TARGET_AMD64_
8420 bool compQuirkForPPP(); // Check if this method should be Quirked for the PPP issue
8423 // Assumes called as part of process shutdown; does any compiler-specific work associated with that.
8424 static void ProcessShutdownWork(ICorStaticInfo* statInfo);
8426 IAllocator* getAllocator()
8428 return compAsIAllocator;
8431 #if MEASURE_MEM_ALLOC
8432 IAllocator* getAllocatorBitset()
8434 return compAsIAllocatorBitset;
8436 IAllocator* getAllocatorGC()
8438 return compAsIAllocatorGC;
8440 IAllocator* getAllocatorLoopHoist()
8442 return compAsIAllocatorLoopHoist;
8444 #else // !MEASURE_MEM_ALLOC
8445 IAllocator* getAllocatorBitset()
8447 return compAsIAllocator;
8449 IAllocator* getAllocatorGC()
8451 return compAsIAllocator;
8453 IAllocator* getAllocatorLoopHoist()
8455 return compAsIAllocator;
8457 #endif // !MEASURE_MEM_ALLOC
8460 IAllocator* getAllocatorDebugOnly()
8462 #if MEASURE_MEM_ALLOC
8463 return compAsIAllocatorDebugOnly;
8464 #else // !MEASURE_MEM_ALLOC
8465 return compAsIAllocator;
8466 #endif // !MEASURE_MEM_ALLOC
8471 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8472 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8476 XX Checks for type compatibility and merges types XX
8478 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8479 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8483 // Set to TRUE if verification cannot be skipped for this method
8484 // If we detect unverifiable code, we will lazily check
8485 // canSkipMethodVerification() to see if verification is REALLY needed.
8486 BOOL tiVerificationNeeded;
8488 // It it initially TRUE, and it gets set to FALSE if we run into unverifiable code
8489 // Note that this is valid only if tiVerificationNeeded was ever TRUE.
8490 BOOL tiIsVerifiableCode;
8492 // Set to TRUE if runtime callout is needed for this method
8493 BOOL tiRuntimeCalloutNeeded;
8495 // Set to TRUE if security prolog/epilog callout is needed for this method
8496 // Note: This flag is different than compNeedSecurityCheck.
8497 // compNeedSecurityCheck means whether or not a security object needs
8498 // to be allocated on the stack, which is currently true for EnC as well.
8499 // tiSecurityCalloutNeeded means whether or not security callouts need
8500 // to be inserted in the jitted code.
8501 BOOL tiSecurityCalloutNeeded;
8503 // Returns TRUE if child is equal to or a subtype of parent for merge purposes
8504 // This support is necessary to suport attributes that are not described in
8505 // for example, signatures. For example, the permanent home byref (byref that
8506 // points to the gc heap), isn't a property of method signatures, therefore,
8507 // it is safe to have mismatches here (that tiCompatibleWith will not flag),
8508 // but when deciding if we need to reimport a block, we need to take these
8510 BOOL tiMergeCompatibleWith(const typeInfo& pChild, const typeInfo& pParent, bool normalisedForStack) const;
8512 // Returns TRUE if child is equal to or a subtype of parent.
8513 // normalisedForStack indicates that both types are normalised for the stack
8514 BOOL tiCompatibleWith(const typeInfo& pChild, const typeInfo& pParent, bool normalisedForStack) const;
8516 // Merges pDest and pSrc. Returns FALSE if merge is undefined.
8517 // *pDest is modified to represent the merged type. Sets "*changed" to true
8518 // if this changes "*pDest".
8519 BOOL tiMergeToCommonParent(typeInfo* pDest, const typeInfo* pSrc, bool* changed) const;
8521 // Set pDest from the primitive value type.
8522 // Eg. System.Int32 -> ELEMENT_TYPE_I4
8524 BOOL tiFromPrimitiveValueClass(typeInfo* pDest, const typeInfo* pVC) const;
8527 // <BUGNUM> VSW 471305
8528 // IJW allows assigning REF to BYREF. The following allows us to temporarily
8529 // bypass the assert check in gcMarkRegSetGCref and gcMarkRegSetByref
8530 // We use a "short" as we need to push/pop this scope.
8532 short compRegSetCheckLevel;
8536 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8537 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8539 XX IL verification stuff XX
8542 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8543 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8547 // The following is used to track liveness of local variables, initialization
8548 // of valueclass constructors, and type safe use of IL instructions.
8550 // dynamic state info needed for verification
8551 EntryState verCurrentState;
8553 // this ptr of object type .ctors are considered intited only after
8554 // the base class ctor is called, or an alternate ctor is called.
8555 // An uninited this ptr can be used to access fields, but cannot
8556 // be used to call a member function.
8557 BOOL verTrackObjCtorInitState;
8559 void verInitBBEntryState(BasicBlock* block, EntryState* currentState);
8561 // Requires that "tis" is not TIS_Bottom -- it's a definite init/uninit state.
8562 void verSetThisInit(BasicBlock* block, ThisInitState tis);
8563 void verInitCurrentState();
8564 void verResetCurrentState(BasicBlock* block, EntryState* currentState);
8566 // Merges the current verification state into the entry state of "block", return FALSE if that merge fails,
8567 // TRUE if it succeeds. Further sets "*changed" to true if this changes the entry state of "block".
8568 BOOL verMergeEntryStates(BasicBlock* block, bool* changed);
8570 void verConvertBBToThrowVerificationException(BasicBlock* block DEBUGARG(bool logMsg));
8571 void verHandleVerificationFailure(BasicBlock* block DEBUGARG(bool logMsg));
8572 typeInfo verMakeTypeInfo(CORINFO_CLASS_HANDLE clsHnd,
8573 bool bashStructToRef = false); // converts from jit type representation to typeInfo
8574 typeInfo verMakeTypeInfo(CorInfoType ciType,
8575 CORINFO_CLASS_HANDLE clsHnd); // converts from jit type representation to typeInfo
8576 BOOL verIsSDArray(typeInfo ti);
8577 typeInfo verGetArrayElemType(typeInfo ti);
8579 typeInfo verParseArgSigToTypeInfo(CORINFO_SIG_INFO* sig, CORINFO_ARG_LIST_HANDLE args);
8580 BOOL verNeedsVerification();
8581 BOOL verIsByRefLike(const typeInfo& ti);
8582 BOOL verIsSafeToReturnByRef(const typeInfo& ti);
8584 // generic type variables range over types that satisfy IsBoxable
8585 BOOL verIsBoxable(const typeInfo& ti);
8587 void DECLSPEC_NORETURN verRaiseVerifyException(INDEBUG(const char* reason) DEBUGARG(const char* file)
8588 DEBUGARG(unsigned line));
8589 void verRaiseVerifyExceptionIfNeeded(INDEBUG(const char* reason) DEBUGARG(const char* file)
8590 DEBUGARG(unsigned line));
8591 bool verCheckTailCallConstraint(OPCODE opcode,
8592 CORINFO_RESOLVED_TOKEN* pResolvedToken,
8593 CORINFO_RESOLVED_TOKEN* pConstrainedResolvedToken, // Is this a "constrained." call
8594 // on a type parameter?
8595 bool speculative // If true, won't throw if verificatoin fails. Instead it will
8596 // return false to the caller.
8597 // If false, it will throw.
8599 bool verIsBoxedValueType(typeInfo ti);
8601 void verVerifyCall(OPCODE opcode,
8602 CORINFO_RESOLVED_TOKEN* pResolvedToken,
8603 CORINFO_RESOLVED_TOKEN* pConstrainedResolvedToken,
8605 bool readonlyCall, // is this a "readonly." call?
8606 const BYTE* delegateCreateStart,
8607 const BYTE* codeAddr,
8608 CORINFO_CALL_INFO* callInfo DEBUGARG(const char* methodName));
8610 BOOL verCheckDelegateCreation(const BYTE* delegateCreateStart, const BYTE* codeAddr, mdMemberRef& targetMemberRef);
8612 typeInfo verVerifySTIND(const typeInfo& ptr, const typeInfo& value, const typeInfo& instrType);
8613 typeInfo verVerifyLDIND(const typeInfo& ptr, const typeInfo& instrType);
8614 void verVerifyField(CORINFO_RESOLVED_TOKEN* pResolvedToken,
8615 const CORINFO_FIELD_INFO& fieldInfo,
8616 const typeInfo* tiThis,
8618 BOOL allowPlainStructAsThis = FALSE);
8619 void verVerifyCond(const typeInfo& tiOp1, const typeInfo& tiOp2, unsigned opcode);
8620 void verVerifyThisPtrInitialised();
8621 BOOL verIsCallToInitThisPtr(CORINFO_CLASS_HANDLE context, CORINFO_CLASS_HANDLE target);
8623 // Register allocator
8624 void raInitStackFP();
8625 void raEnregisterVarsPrePassStackFP();
8626 void raSetRegLclBirthDeath(GenTreePtr tree, VARSET_VALARG_TP lastlife, bool fromLDOBJ);
8627 void raEnregisterVarsPostPassStackFP();
8628 void raGenerateFPRefCounts();
8629 void raEnregisterVarsStackFP();
8630 void raUpdateHeightsForVarsStackFP(VARSET_VALARG_TP mask);
8632 regNumber raRegForVarStackFP(unsigned varTrackedIndex);
8633 void raAddPayloadStackFP(VARSET_VALARG_TP mask, unsigned weight);
8635 // returns true if enregistering v1 would save more mem accesses than v2
8636 bool raVarIsGreaterValueStackFP(LclVarDsc* lv1, LclVarDsc* lv2);
8639 void raDumpHeightsStackFP();
8640 void raDumpVariableRegIntfFloat();
8643 #if FEATURE_STACK_FP_X87
8645 // Currently, we use FP transition blocks in only 2 situations:
8647 // -conditional jump on longs where FP stack differs with target: it's not strictly
8648 // necessary, but its low frequency and the code would get complicated if we try to
8649 // inline the FP stack adjustment, as we have a lot of special casing going on to try
8650 // minimize the way we generate the jump code.
8651 // -case statements of switch where the FP stack differs with the one of evaluating the switch () statement
8652 // We do this as we want to codegen switch as a jumptable. Again, this is low frequency.
8654 // However, transition blocks have 2 problems
8656 // - Procedure splitting: current implementation of procedure splitting requires all basic blocks to
8657 // be known at codegen time, as it generates all hot blocks first and cold blocks later. This ties
8658 // us up in codegen and is a solvable problem (we could make procedure splitting generate blocks
8659 // in the right place without preordering them), this causes us to have to generate the transition
8660 // blocks in the cold area if we want procedure splitting.
8663 // - Thread abort exceptions and transition blocks. Transition blocks were designed under the assumption
8664 // that no exceptions can happen inside them. Unfortunately Thread.Abort can happen in any instruction,
8665 // and if we have handlers we will have to try to call them. Fixing this the right way would imply
8666 // having multiple try native code regions for a single try il region. This is doable and shouldnt be
8667 // a big change in the exception.
8669 // Given the low frequency of the cases where we have transition blocks, I've decided to dumb down
8670 // optimizations. For these 2 cases:
8672 // - When there is a chance that we will have FP transition blocks, we won't do procedure splitting.
8673 // - When a method has a handler, it won't enregister any FP variables that go thru a conditional long or
8674 // a switch statement.
8676 // If at any point we find we need to optimize this, we should throw work at unblocking the restrictions our
8677 // current procedure splitting and exception code have.
8678 bool compMayHaveTransitionBlocks;
8680 VARSET_TP raMaskDontEnregFloat; // mask for additional restrictions
8682 VARSET_TP raLclRegIntfFloat[REG_FPCOUNT];
8684 unsigned raCntStkStackFP;
8685 unsigned raCntWtdStkDblStackFP;
8686 unsigned raCntStkParamDblStackFP;
8688 // Payload in mem accesses for enregistering a variable (we dont want to mix with refcounts)
8689 // TODO: Do we want to put this in LclVarDsc?
8690 unsigned raPayloadStackFP[lclMAX_TRACKED];
8691 unsigned raHeightsStackFP[lclMAX_TRACKED][FP_VIRTUALREGISTERS + 1];
8693 // Useful for debugging
8694 unsigned raHeightsNonWeightedStackFP[lclMAX_TRACKED][FP_VIRTUALREGISTERS + 1];
8696 #endif // FEATURE_STACK_FP_X87
8699 // One line log function. Default level is 0. Increasing it gives you
8700 // more log information
8702 // levels are currently unused: #define JITDUMP(level,...) ();
8703 void JitLogEE(unsigned level, const char* fmt, ...);
8705 bool compDebugBreak;
8707 bool compJitHaltMethod();
8712 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8713 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8715 XX GS Security checks for unsafe buffers XX
8717 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8718 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8721 struct ShadowParamVarInfo
8723 FixedBitVect* assignGroup; // the closure set of variables whose values depend on each other
8724 unsigned shadowCopy; // Lcl var num, valid only if not set to NO_SHADOW_COPY
8726 static bool mayNeedShadowCopy(LclVarDsc* varDsc)
8728 #if defined(_TARGET_AMD64_) && !defined(LEGACY_BACKEND)
8729 // GS cookie logic to create shadow slots, create trees to copy reg args to shadow
8730 // slots and update all trees to refer to shadow slots is done immediately after
8731 // fgMorph(). Lsra could potentially mark a param as DoNotEnregister after JIT determines
8732 // not to shadow a parameter. Also, LSRA could potentially spill a param which is passed
8733 // in register. Therefore, conservatively all params may need a shadow copy. Note that
8734 // GS cookie logic further checks whether the param is a ptr or an unsafe buffer before
8735 // creating a shadow slot even though this routine returns true.
8737 // TODO-AMD64-CQ: Revisit this conservative approach as it could create more shadow slots than
8738 // required. There are two cases under which a reg arg could potentially be used from its
8740 // a) LSRA marks it as DoNotEnregister (see LinearScan::identifyCandidates())
8741 // b) LSRA spills it
8743 // Possible solution to address case (a)
8744 // - The conditions under which LSRA marks a varDsc as DoNotEnregister could be checked
8745 // in this routine. Note that live out of exception handler is something we may not be
8746 // able to do it here since GS cookie logic is invoked ahead of liveness computation.
8747 // Therefore, for methods with exception handling and need GS cookie check we might have
8748 // to take conservative approach.
8750 // Possible solution to address case (b)
8751 // - Whenver a parameter passed in an argument register needs to be spilled by LSRA, we
8752 // create a new spill temp if the method needs GS cookie check.
8753 return varDsc->lvIsParam;
8754 #else // !(defined(_TARGET_AMD64_) && defined(LEGACY_BACKEND))
8755 return varDsc->lvIsParam && !varDsc->lvIsRegArg;
8762 printf("assignGroup [%p]; shadowCopy: [%d];\n", assignGroup, shadowCopy);
8767 GSCookie* gsGlobalSecurityCookieAddr; // Address of global cookie for unsafe buffer checks
8768 GSCookie gsGlobalSecurityCookieVal; // Value of global cookie if addr is NULL
8769 ShadowParamVarInfo* gsShadowVarInfo; // Table used by shadow param analysis code
8771 void gsGSChecksInitCookie(); // Grabs cookie variable
8772 void gsCopyShadowParams(); // Identify vulnerable params and create dhadow copies
8773 bool gsFindVulnerableParams(); // Shadow param analysis code
8774 void gsParamsToShadows(); // Insert copy code and replave param uses by shadow
8776 static fgWalkPreFn gsMarkPtrsAndAssignGroups; // Shadow param analysis tree-walk
8777 static fgWalkPreFn gsReplaceShadowParams; // Shadow param replacement tree-walk
8779 #define DEFAULT_MAX_INLINE_SIZE 100 // Methods with > DEFAULT_MAX_INLINE_SIZE IL bytes will never be inlined.
8780 // This can be overwritten by setting complus_JITInlineSize env variable.
8782 #define DEFAULT_MAX_INLINE_DEPTH 20 // Methods at more than this level deep will not be inlined
8785 #ifdef FEATURE_JIT_METHOD_PERF
8786 JitTimer* pCompJitTimer; // Timer data structure (by phases) for current compilation.
8787 static CompTimeSummaryInfo s_compJitTimerSummary; // Summary of the Timer information for the whole run.
8789 static LPCWSTR JitTimeLogCsv(); // Retrieve the file name for CSV from ConfigDWORD.
8790 static LPCWSTR compJitTimeLogFilename; // If a log file for JIT time is desired, filename to write it to.
8792 inline void EndPhase(Phases phase); // Indicate the end of the given phase.
8794 #if MEASURE_CLRAPI_CALLS
8795 // Thin wrappers that call into JitTimer (if present).
8796 inline void CLRApiCallEnter(unsigned apix);
8797 inline void CLRApiCallLeave(unsigned apix);
8800 inline void CLR_API_Enter(API_ICorJitInfo_Names ename);
8801 inline void CLR_API_Leave(API_ICorJitInfo_Names ename);
8806 #if defined(DEBUG) || defined(INLINE_DATA) || defined(FEATURE_CLRSQM)
8807 // These variables are associated with maintaining SQM data about compile time.
8808 unsigned __int64 m_compCyclesAtEndOfInlining; // The thread-virtualized cycle count at the end of the inlining phase
8809 // in the current compilation.
8810 unsigned __int64 m_compCycles; // Net cycle count for current compilation
8811 DWORD m_compTickCountAtEndOfInlining; // The result of GetTickCount() (# ms since some epoch marker) at the end of
8812 // the inlining phase in the current compilation.
8813 #endif // defined(DEBUG) || defined(INLINE_DATA) || defined(FEATURE_CLRSQM)
8815 // Records the SQM-relevant (cycles and tick count). Should be called after inlining is complete.
8816 // (We do this after inlining because this marks the last point at which the JIT is likely to cause
8817 // type-loading and class initialization).
8818 void RecordStateAtEndOfInlining();
8819 // Assumes being called at the end of compilation. Update the SQM state.
8820 void RecordStateAtEndOfCompilation();
8822 #ifdef FEATURE_CLRSQM
8823 // Does anything SQM related necessary at process shutdown time.
8824 static void ProcessShutdownSQMWork(ICorStaticInfo* statInfo);
8825 #endif // FEATURE_CLRSQM
8828 #if FUNC_INFO_LOGGING
8829 static LPCWSTR compJitFuncInfoFilename; // If a log file for per-function information is required, this is the
8830 // filename to write it to.
8831 static FILE* compJitFuncInfoFile; // And this is the actual FILE* to write to.
8832 #endif // FUNC_INFO_LOGGING
8834 Compiler* prevCompiler; // Previous compiler on stack for TLS Compiler* linked list for reentrant compilers.
8836 // Is the compilation in a full trust context?
8837 bool compIsFullTrust();
8839 #ifndef FEATURE_TRACELOGGING
8840 // Should we actually fire the noway assert body and the exception handler?
8841 bool compShouldThrowOnNoway();
8842 #else // FEATURE_TRACELOGGING
8843 // Should we actually fire the noway assert body and the exception handler?
8844 bool compShouldThrowOnNoway(const char* filename, unsigned line);
8846 // Telemetry instance to use per method compilation.
8847 JitTelemetry compJitTelemetry;
8849 // Get common parameters that have to be logged with most telemetry data.
8850 void compGetTelemetryDefaults(const char** assemblyName,
8851 const char** scopeName,
8852 const char** methodName,
8853 unsigned* methodHash);
8854 #endif // !FEATURE_TRACELOGGING
8858 NodeToTestDataMap* m_nodeTestData;
8860 static const unsigned FIRST_LOOP_HOIST_CSE_CLASS = 1000;
8861 unsigned m_loopHoistCSEClass; // LoopHoist test annotations turn into CSE requirements; we
8862 // label them with CSE Class #'s starting at FIRST_LOOP_HOIST_CSE_CLASS.
8863 // Current kept in this.
8865 NodeToTestDataMap* GetNodeTestData()
8867 Compiler* compRoot = impInlineRoot();
8868 if (compRoot->m_nodeTestData == nullptr)
8870 compRoot->m_nodeTestData = new (getAllocatorDebugOnly()) NodeToTestDataMap(getAllocatorDebugOnly());
8872 return compRoot->m_nodeTestData;
8875 typedef SimplerHashTable<GenTreePtr, PtrKeyFuncs<GenTree>, int, JitSimplerHashBehavior> NodeToIntMap;
8877 // Returns the set (i.e., the domain of the result map) of nodes that are keys in m_nodeTestData, and
8878 // currently occur in the AST graph.
8879 NodeToIntMap* FindReachableNodesInNodeTestData();
8881 // Node "from" is being eliminated, and being replaced by node "to". If "from" had any associated
8882 // test data, associate that data with "to".
8883 void TransferTestDataToNode(GenTreePtr from, GenTreePtr to);
8885 // Requires that "to" is a clone of "from". If any nodes in the "from" tree
8886 // have annotations, attach similar annotations to the corresponding nodes in "to".
8887 void CopyTestDataToCloneTree(GenTreePtr from, GenTreePtr to);
8889 // These are the methods that test that the various conditions implied by the
8890 // test attributes are satisfied.
8891 void JitTestCheckSSA(); // SSA builder tests.
8892 void JitTestCheckVN(); // Value numbering tests.
8895 // The "FieldSeqStore", for canonicalizing field sequences. See the definition of FieldSeqStore for
8897 FieldSeqStore* m_fieldSeqStore;
8899 FieldSeqStore* GetFieldSeqStore()
8901 Compiler* compRoot = impInlineRoot();
8902 if (compRoot->m_fieldSeqStore == nullptr)
8904 // Create a CompAllocator that labels sub-structure with CMK_FieldSeqStore, and use that for allocation.
8905 IAllocator* ialloc = new (this, CMK_FieldSeqStore) CompAllocator(this, CMK_FieldSeqStore);
8906 compRoot->m_fieldSeqStore = new (ialloc) FieldSeqStore(ialloc);
8908 return compRoot->m_fieldSeqStore;
8911 typedef SimplerHashTable<GenTreePtr, PtrKeyFuncs<GenTree>, FieldSeqNode*, JitSimplerHashBehavior> NodeToFieldSeqMap;
8913 // Some nodes of "TYP_BYREF" or "TYP_I_IMPL" actually represent the address of a field within a struct, but since
8914 // the offset of the field is zero, there's no "GT_ADD" node. We normally attach a field sequence to the constant
8915 // that is added, but what do we do when that constant is zero, and is thus not present? We use this mechanism to
8916 // attach the field sequence directly to the address node.
8917 NodeToFieldSeqMap* m_zeroOffsetFieldMap;
8919 NodeToFieldSeqMap* GetZeroOffsetFieldMap()
8921 // Don't need to worry about inlining here
8922 if (m_zeroOffsetFieldMap == nullptr)
8924 // Create a CompAllocator that labels sub-structure with CMK_ZeroOffsetFieldMap, and use that for
8926 IAllocator* ialloc = new (this, CMK_ZeroOffsetFieldMap) CompAllocator(this, CMK_ZeroOffsetFieldMap);
8927 m_zeroOffsetFieldMap = new (ialloc) NodeToFieldSeqMap(ialloc);
8929 return m_zeroOffsetFieldMap;
8932 // Requires that "op1" is a node of type "TYP_BYREF" or "TYP_I_IMPL". We are dereferencing this with the fields in
8933 // "fieldSeq", whose offsets are required all to be zero. Ensures that any field sequence annotation currently on
8934 // "op1" or its components is augmented by appending "fieldSeq". In practice, if "op1" is a GT_LCL_FLD, it has
8935 // a field sequence as a member; otherwise, it may be the addition of an a byref and a constant, where the const
8936 // has a field sequence -- in this case "fieldSeq" is appended to that of the constant; otherwise, we
8937 // record the the field sequence using the ZeroOffsetFieldMap described above.
8939 // One exception above is that "op1" is a node of type "TYP_REF" where "op1" is a GT_LCL_VAR.
8940 // This happens when System.Object vtable pointer is a regular field at offset 0 in System.Private.CoreLib in
8941 // CoreRT. Such case is handled same as the default case.
8942 void fgAddFieldSeqForZeroOffset(GenTreePtr op1, FieldSeqNode* fieldSeq);
8944 typedef SimplerHashTable<const GenTree*, PtrKeyFuncs<GenTree>, ArrayInfo, JitSimplerHashBehavior>
8946 NodeToArrayInfoMap* m_arrayInfoMap;
8948 NodeToArrayInfoMap* GetArrayInfoMap()
8950 Compiler* compRoot = impInlineRoot();
8951 if (compRoot->m_arrayInfoMap == nullptr)
8953 // Create a CompAllocator that labels sub-structure with CMK_ArrayInfoMap, and use that for allocation.
8954 IAllocator* ialloc = new (this, CMK_ArrayInfoMap) CompAllocator(this, CMK_ArrayInfoMap);
8955 compRoot->m_arrayInfoMap = new (ialloc) NodeToArrayInfoMap(ialloc);
8957 return compRoot->m_arrayInfoMap;
8960 NodeToUnsignedMap* m_heapSsaMap;
8962 // In some cases, we want to assign intermediate SSA #'s to heap states, and know what nodes create those heap
8963 // states. (We do this for try blocks, where, if the try block doesn't do a call that loses track of the heap state,
8964 // all the possible heap states are possible initial states of the corresponding catch block(s).)
8965 NodeToUnsignedMap* GetHeapSsaMap()
8967 Compiler* compRoot = impInlineRoot();
8968 if (compRoot->m_heapSsaMap == nullptr)
8970 // Create a CompAllocator that labels sub-structure with CMK_ArrayInfoMap, and use that for allocation.
8971 IAllocator* ialloc = new (this, CMK_ArrayInfoMap) CompAllocator(this, CMK_ArrayInfoMap);
8972 compRoot->m_heapSsaMap = new (ialloc) NodeToUnsignedMap(ialloc);
8974 return compRoot->m_heapSsaMap;
8977 // The Refany type is the only struct type whose structure is implicitly assumed by IL. We need its fields.
8978 CORINFO_CLASS_HANDLE m_refAnyClass;
8979 CORINFO_FIELD_HANDLE GetRefanyDataField()
8981 if (m_refAnyClass == nullptr)
8983 m_refAnyClass = info.compCompHnd->getBuiltinClass(CLASSID_TYPED_BYREF);
8985 return info.compCompHnd->getFieldInClass(m_refAnyClass, 0);
8987 CORINFO_FIELD_HANDLE GetRefanyTypeField()
8989 if (m_refAnyClass == nullptr)
8991 m_refAnyClass = info.compCompHnd->getBuiltinClass(CLASSID_TYPED_BYREF);
8993 return info.compCompHnd->getFieldInClass(m_refAnyClass, 1);
8997 static BitSetSupport::BitSetOpCounter m_varsetOpCounter;
8999 #if ALLVARSET_COUNTOPS
9000 static BitSetSupport::BitSetOpCounter m_allvarsetOpCounter;
9003 static HelperCallProperties s_helperCallProperties;
9005 #ifdef FEATURE_UNIX_AMD64_STRUCT_PASSING
9006 static var_types GetTypeFromClassificationAndSizes(SystemVClassificationType classType, int size);
9007 static var_types GetEightByteType(const SYSTEMV_AMD64_CORINFO_STRUCT_REG_PASSING_DESCRIPTOR& structDesc,
9009 static void GetStructTypeOffset(const SYSTEMV_AMD64_CORINFO_STRUCT_REG_PASSING_DESCRIPTOR& structDesc,
9012 unsigned __int8* offset0,
9013 unsigned __int8* offset1);
9014 void fgMorphSystemVStructArgs(GenTreeCall* call, bool hasStructArgument);
9015 #endif // defined(FEATURE_UNIX_AMD64_STRUCT_PASSING)
9017 void fgMorphMultiregStructArgs(GenTreeCall* call);
9018 GenTreePtr fgMorphMultiregStructArg(GenTreePtr arg, fgArgTabEntryPtr fgEntryPtr);
9020 }; // end of class Compiler
9022 // Inline methods of CompAllocator.
9023 void* CompAllocator::Alloc(size_t sz)
9025 #if MEASURE_MEM_ALLOC
9026 return m_comp->compGetMem(sz, m_cmk);
9028 return m_comp->compGetMem(sz);
9032 void* CompAllocator::ArrayAlloc(size_t elems, size_t elemSize)
9034 #if MEASURE_MEM_ALLOC
9035 return m_comp->compGetMemArray(elems, elemSize, m_cmk);
9037 return m_comp->compGetMemArray(elems, elemSize);
9041 // LclVarDsc constructor. Uses Compiler, so must come after Compiler definition.
9042 inline LclVarDsc::LclVarDsc(Compiler* comp)
9043 : // Initialize the ArgRegs to REG_STK.
9044 // The morph will do the right thing to change
9045 // to the right register if passed in register.
9048 #if FEATURE_MULTIREG_ARGS
9049 _lvOtherArgReg(REG_STK)
9051 #endif // FEATURE_MULTIREG_ARGS
9053 lvRefBlks(BlockSetOps::UninitVal())
9055 #endif // ASSERTION_PROP
9056 lvPerSsaData(comp->getAllocator())
9061 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9062 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9064 XX Miscellaneous Compiler stuff XX
9066 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9067 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9070 // Values used to mark the types a stack slot is used for
9072 const unsigned TYPE_REF_INT = 0x01; // slot used as a 32-bit int
9073 const unsigned TYPE_REF_LNG = 0x02; // slot used as a 64-bit long
9074 const unsigned TYPE_REF_FLT = 0x04; // slot used as a 32-bit float
9075 const unsigned TYPE_REF_DBL = 0x08; // slot used as a 64-bit float
9076 const unsigned TYPE_REF_PTR = 0x10; // slot used as a 32-bit pointer
9077 const unsigned TYPE_REF_BYR = 0x20; // slot used as a byref pointer
9078 const unsigned TYPE_REF_STC = 0x40; // slot used as a struct
9079 const unsigned TYPE_REF_TYPEMASK = 0x7F; // bits that represent the type
9081 // const unsigned TYPE_REF_ADDR_TAKEN = 0x80; // slots address was taken
9083 /*****************************************************************************
9085 * Variables to keep track of total code amounts.
9090 extern size_t grossVMsize;
9091 extern size_t grossNCsize;
9092 extern size_t totalNCsize;
9094 extern unsigned genMethodICnt;
9095 extern unsigned genMethodNCnt;
9096 extern size_t gcHeaderISize;
9097 extern size_t gcPtrMapISize;
9098 extern size_t gcHeaderNSize;
9099 extern size_t gcPtrMapNSize;
9101 #endif // DISPLAY_SIZES
9103 /*****************************************************************************
9105 * Variables to keep track of basic block counts (more data on 1 BB methods)
9108 #if COUNT_BASIC_BLOCKS
9109 extern Histogram bbCntTable;
9110 extern Histogram bbOneBBSizeTable;
9113 /*****************************************************************************
9115 * Used by optFindNaturalLoops to gather statistical information such as
9116 * - total number of natural loops
9117 * - number of loops with 1, 2, ... exit conditions
9118 * - number of loops that have an iterator (for like)
9119 * - number of loops that have a constant iterator
9124 extern unsigned totalLoopMethods; // counts the total number of methods that have natural loops
9125 extern unsigned maxLoopsPerMethod; // counts the maximum number of loops a method has
9126 extern unsigned totalLoopOverflows; // # of methods that identified more loops than we can represent
9127 extern unsigned totalLoopCount; // counts the total number of natural loops
9128 extern unsigned totalUnnatLoopCount; // counts the total number of (not-necessarily natural) loops
9129 extern unsigned totalUnnatLoopOverflows; // # of methods that identified more unnatural loops than we can represent
9130 extern unsigned iterLoopCount; // counts the # of loops with an iterator (for like)
9131 extern unsigned simpleTestLoopCount; // counts the # of loops with an iterator and a simple loop condition (iter <
9133 extern unsigned constIterLoopCount; // counts the # of loops with a constant iterator (for like)
9134 extern bool hasMethodLoops; // flag to keep track if we already counted a method as having loops
9135 extern unsigned loopsThisMethod; // counts the number of loops in the current method
9136 extern bool loopOverflowThisMethod; // True if we exceeded the max # of loops in the method.
9137 extern Histogram loopCountTable; // Histogram of loop counts
9138 extern Histogram loopExitCountTable; // Histogram of loop exit counts
9140 #endif // COUNT_LOOPS
9142 /*****************************************************************************
9143 * variables to keep track of how many iterations we go in a dataflow pass
9148 extern unsigned CSEiterCount; // counts the # of iteration for the CSE dataflow
9149 extern unsigned CFiterCount; // counts the # of iteration for the Const Folding dataflow
9151 #endif // DATAFLOW_ITER
9153 #if MEASURE_BLOCK_SIZE
9154 extern size_t genFlowNodeSize;
9155 extern size_t genFlowNodeCnt;
9156 #endif // MEASURE_BLOCK_SIZE
9158 #if MEASURE_NODE_SIZE
9159 struct NodeSizeStats
9164 genTreeNodeSize = 0;
9165 genTreeNodeActualSize = 0;
9168 size_t genTreeNodeCnt;
9169 size_t genTreeNodeSize; // The size we allocate
9170 size_t genTreeNodeActualSize; // The actual size of the node. Note that the actual size will likely be smaller
9171 // than the allocated size, but we sometimes use SetOper()/ChangeOper() to change
9172 // a smaller node to a larger one. TODO-Cleanup: add stats on
9173 // SetOper()/ChangeOper() usage to quanitfy this.
9175 extern NodeSizeStats genNodeSizeStats; // Total node size stats
9176 extern NodeSizeStats genNodeSizeStatsPerFunc; // Per-function node size stats
9177 extern Histogram genTreeNcntHist;
9178 extern Histogram genTreeNsizHist;
9179 #endif // MEASURE_NODE_SIZE
9181 /*****************************************************************************
9182 * Count fatal errors (including noway_asserts).
9186 extern unsigned fatal_badCode;
9187 extern unsigned fatal_noWay;
9188 extern unsigned fatal_NOMEM;
9189 extern unsigned fatal_noWayAssertBody;
9191 extern unsigned fatal_noWayAssertBodyArgs;
9193 extern unsigned fatal_NYI;
9194 #endif // MEASURE_FATAL
9196 /*****************************************************************************
9200 #ifdef _TARGET_XARCH_
9202 const instruction INS_SHIFT_LEFT_LOGICAL = INS_shl;
9203 const instruction INS_SHIFT_RIGHT_LOGICAL = INS_shr;
9204 const instruction INS_SHIFT_RIGHT_ARITHM = INS_sar;
9206 const instruction INS_AND = INS_and;
9207 const instruction INS_OR = INS_or;
9208 const instruction INS_XOR = INS_xor;
9209 const instruction INS_NEG = INS_neg;
9210 const instruction INS_TEST = INS_test;
9211 const instruction INS_MUL = INS_imul;
9212 const instruction INS_SIGNED_DIVIDE = INS_idiv;
9213 const instruction INS_UNSIGNED_DIVIDE = INS_div;
9214 const instruction INS_BREAKPOINT = INS_int3;
9215 const instruction INS_ADDC = INS_adc;
9216 const instruction INS_SUBC = INS_sbb;
9217 const instruction INS_NOT = INS_not;
9223 const instruction INS_SHIFT_LEFT_LOGICAL = INS_lsl;
9224 const instruction INS_SHIFT_RIGHT_LOGICAL = INS_lsr;
9225 const instruction INS_SHIFT_RIGHT_ARITHM = INS_asr;
9227 const instruction INS_AND = INS_and;
9228 const instruction INS_OR = INS_orr;
9229 const instruction INS_XOR = INS_eor;
9230 const instruction INS_NEG = INS_rsb;
9231 const instruction INS_TEST = INS_tst;
9232 const instruction INS_MUL = INS_mul;
9233 const instruction INS_SIGNED_DIVIDE = INS_sdiv;
9234 const instruction INS_UNSIGNED_DIVIDE = INS_udiv;
9235 const instruction INS_BREAKPOINT = INS_bkpt;
9236 const instruction INS_ADDC = INS_adc;
9237 const instruction INS_SUBC = INS_sbc;
9238 const instruction INS_NOT = INS_mvn;
9242 #ifdef _TARGET_ARM64_
9244 const instruction INS_SHIFT_LEFT_LOGICAL = INS_lsl;
9245 const instruction INS_SHIFT_RIGHT_LOGICAL = INS_lsr;
9246 const instruction INS_SHIFT_RIGHT_ARITHM = INS_asr;
9248 const instruction INS_AND = INS_and;
9249 const instruction INS_OR = INS_orr;
9250 const instruction INS_XOR = INS_eor;
9251 const instruction INS_NEG = INS_neg;
9252 const instruction INS_TEST = INS_tst;
9253 const instruction INS_MUL = INS_mul;
9254 const instruction INS_SIGNED_DIVIDE = INS_sdiv;
9255 const instruction INS_UNSIGNED_DIVIDE = INS_udiv;
9256 const instruction INS_BREAKPOINT = INS_bkpt;
9257 const instruction INS_ADDC = INS_adc;
9258 const instruction INS_SUBC = INS_sbc;
9259 const instruction INS_NOT = INS_mvn;
9263 /*****************************************************************************/
9265 extern const BYTE genTypeSizes[];
9266 extern const BYTE genTypeAlignments[];
9267 extern const BYTE genTypeStSzs[];
9268 extern const BYTE genActualTypes[];
9270 /*****************************************************************************/
9272 // VERY_LARGE_FRAME_SIZE_REG_MASK is the set of registers we need to use for
9273 // the probing loop generated for very large stack frames (see `getVeryLargeFrameSize`).
9276 #define VERY_LARGE_FRAME_SIZE_REG_MASK (RBM_R4 | RBM_R5 | RBM_R6)
9277 #elif defined(_TARGET_ARM64_)
9278 #define VERY_LARGE_FRAME_SIZE_REG_MASK (RBM_R9 | RBM_R10 | RBM_R11)
9281 /*****************************************************************************/
9283 #define REG_CORRUPT regNumber(REG_NA + 1)
9284 #define RBM_CORRUPT (RBM_ILLEGAL | regMaskTP(1))
9285 #define REG_PAIR_CORRUPT regPairNo(REG_PAIR_NONE + 1)
9287 /*****************************************************************************/
9289 extern BasicBlock dummyBB;
9291 /*****************************************************************************/
9292 /*****************************************************************************/
9294 // foreach_treenode_execution_order: An iterator that iterates through all the tree
9295 // nodes of a statement in execution order.
9296 // __stmt: a GT_STMT type GenTree*
9297 // __node: a GenTree*, already declared, that gets updated with each node in the statement, in execution order
9299 #define foreach_treenode_execution_order(__node, __stmt) \
9300 for ((__node) = (__stmt)->gtStmt.gtStmtList; (__node); (__node) = (__node)->gtNext)
9302 // foreach_block: An iterator over all blocks in the function.
9303 // __compiler: the Compiler* object
9304 // __block : a BasicBlock*, already declared, that gets updated each iteration.
9306 #define foreach_block(__compiler, __block) \
9307 for ((__block) = (__compiler)->fgFirstBB; (__block); (__block) = (__block)->bbNext)
9309 /*****************************************************************************/
9310 /*****************************************************************************/
9314 void dumpConvertedVarSet(Compiler* comp, VARSET_VALARG_TP vars);
9316 /*XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9317 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9319 XX Debugging helpers XX
9321 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9322 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9325 /*****************************************************************************/
9326 /* The following functions are intended to be called from the debugger, to dump
9327 * various data structures. The can be used in the debugger Watch or Quick Watch
9328 * windows. They are designed to be short to type and take as few arguments as
9329 * possible. The 'c' versions take a Compiler*, whereas the 'd' versions use the TlsCompiler.
9330 * See the function definition comment for more details.
9333 void cBlock(Compiler* comp, BasicBlock* block);
9334 void cBlocks(Compiler* comp);
9335 void cBlocksV(Compiler* comp);
9336 void cTree(Compiler* comp, GenTree* tree);
9337 void cTrees(Compiler* comp);
9338 void cEH(Compiler* comp);
9339 void cVar(Compiler* comp, unsigned lclNum);
9340 void cVarDsc(Compiler* comp, LclVarDsc* varDsc);
9341 void cVars(Compiler* comp);
9342 void cVarsFinal(Compiler* comp);
9343 void cBlockPreds(Compiler* comp, BasicBlock* block);
9344 void cReach(Compiler* comp);
9345 void cDoms(Compiler* comp);
9346 void cLiveness(Compiler* comp);
9347 void cCVarSet(Compiler* comp, VARSET_VALARG_TP vars);
9349 void cFuncIR(Compiler* comp);
9350 void cBlockIR(Compiler* comp, BasicBlock* block);
9351 void cLoopIR(Compiler* comp, Compiler::LoopDsc* loop);
9352 void cTreeIR(Compiler* comp, GenTree* tree);
9353 int cTreeTypeIR(Compiler* comp, GenTree* tree);
9354 int cTreeKindsIR(Compiler* comp, GenTree* tree);
9355 int cTreeFlagsIR(Compiler* comp, GenTree* tree);
9356 int cOperandIR(Compiler* comp, GenTree* operand);
9357 int cLeafIR(Compiler* comp, GenTree* tree);
9358 int cIndirIR(Compiler* comp, GenTree* tree);
9359 int cListIR(Compiler* comp, GenTree* list);
9360 int cSsaNumIR(Compiler* comp, GenTree* tree);
9361 int cValNumIR(Compiler* comp, GenTree* tree);
9362 int cDependsIR(Compiler* comp, GenTree* comma, bool* first);
9364 void dBlock(BasicBlock* block);
9367 void dTree(GenTree* tree);
9370 void dVar(unsigned lclNum);
9371 void dVarDsc(LclVarDsc* varDsc);
9374 void dBlockPreds(BasicBlock* block);
9378 void dCVarSet(VARSET_VALARG_TP vars);
9380 void dVarSet(VARSET_VALARG_TP vars);
9381 void dRegMask(regMaskTP mask);
9384 void dBlockIR(BasicBlock* block);
9385 void dTreeIR(GenTree* tree);
9386 void dLoopIR(Compiler::LoopDsc* loop);
9387 void dLoopNumIR(unsigned loopNum);
9388 int dTabStopIR(int curr, int tabstop);
9389 int dTreeTypeIR(GenTree* tree);
9390 int dTreeKindsIR(GenTree* tree);
9391 int dTreeFlagsIR(GenTree* tree);
9392 int dOperandIR(GenTree* operand);
9393 int dLeafIR(GenTree* tree);
9394 int dIndirIR(GenTree* tree);
9395 int dListIR(GenTree* list);
9396 int dSsaNumIR(GenTree* tree);
9397 int dValNumIR(GenTree* tree);
9398 int dDependsIR(GenTree* comma);
9401 GenTree* dFindTree(GenTree* tree, unsigned id);
9402 GenTree* dFindTree(unsigned id);
9403 GenTreeStmt* dFindStmt(unsigned id);
9404 BasicBlock* dFindBlock(unsigned bbNum);
9408 #include "compiler.hpp" // All the shared inline functions
9410 /*****************************************************************************/
9411 #endif //_COMPILER_H_
9412 /*****************************************************************************/