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 /*****************************************************************************/
29 #include "jithashtable.h"
38 #include "cycletimer.h"
40 #include "arraystack.h"
42 #include "jitexpandarray.h"
43 #include "tinyarray.h"
46 #include "jittelemetry.h"
47 #include "namedintrinsiclist.h"
52 #include "codegeninterface.h"
54 #include "jitgcinfo.h"
56 #if DUMP_GC_TABLES && defined(JIT32_GCENCODER)
62 #include "hwintrinsic.h"
65 // This is only used locally in the JIT to indicate that
66 // a verification block should be inserted
67 #define SEH_VERIFICATION_EXCEPTION 0xe0564552 // VER
69 /*****************************************************************************
70 * Forward declarations
73 struct InfoHdr; // defined in GCInfo.h
74 struct escapeMapping_t; // defined in flowgraph.cpp
75 class emitter; // defined in emit.h
76 struct ShadowParamVarInfo; // defined in GSChecks.cpp
77 struct InitVarDscInfo; // defined in register_arg_convention.h
78 class FgStack; // defined in flowgraph.cpp
80 class CSE_DataFlow; // defined in OptCSE.cpp
86 class Lowering; // defined in lower.h
88 // The following are defined in this file, Compiler.h
92 /*****************************************************************************
98 /*****************************************************************************/
101 // Declare global operator new overloads that use the compiler's arena allocator
104 // I wanted to make the second argument optional, with default = CMK_Unknown, but that
105 // caused these to be ambiguous with the global placement new operators.
106 void* __cdecl operator new(size_t n, Compiler* context, CompMemKind cmk);
107 void* __cdecl operator new[](size_t n, Compiler* context, CompMemKind cmk);
108 void* __cdecl operator new(size_t n, void* p, const jitstd::placement_t& syntax_difference);
110 // Requires the definitions of "operator new" so including "LoopCloning.h" after the definitions.
111 #include "loopcloning.h"
113 /*****************************************************************************/
115 /* This is included here and not earlier as it needs the definition of "CSE"
116 * which is defined in the section above */
118 /*****************************************************************************/
120 unsigned genLog2(unsigned value);
121 unsigned genLog2(unsigned __int64 value);
123 var_types genActualType(var_types type);
124 var_types genUnsignedType(var_types type);
125 var_types genSignedType(var_types type);
127 unsigned ReinterpretHexAsDecimal(unsigned);
129 /*****************************************************************************/
131 const unsigned FLG_CCTOR = (CORINFO_FLG_CONSTRUCTOR | CORINFO_FLG_STATIC);
134 const int BAD_STK_OFFS = 0xBAADF00D; // for LclVarDsc::lvStkOffs
137 // The following holds the Local var info (scope information)
138 typedef const char* VarName; // Actual ASCII string
141 IL_OFFSET vsdLifeBeg; // instr offset of beg of life
142 IL_OFFSET vsdLifeEnd; // instr offset of end of life
143 unsigned vsdVarNum; // (remapped) LclVarDsc number
146 VarName vsdName; // name of the var
149 unsigned vsdLVnum; // 'which' in eeGetLVinfo().
150 // Also, it is the index of this entry in the info.compVarScopes array,
151 // which is useful since the array is also accessed via the
152 // compEnterScopeList and compExitScopeList sorted arrays.
155 // This is the location of a SSA definition.
161 DefLoc() : m_blk(nullptr), m_tree(nullptr)
165 DefLoc(BasicBlock* block, GenTree* tree) : m_blk(block), m_tree(tree)
170 // This class stores information associated with a LclVar SSA definition.
178 LclSsaVarDsc(BasicBlock* block, GenTree* tree) : m_defLoc(block, tree)
182 ValueNumPair m_vnPair;
186 // This class stores information associated with a memory SSA definition.
190 ValueNumPair m_vnPair;
193 //------------------------------------------------------------------------
194 // SsaDefArray: A resizable array of SSA definitions.
196 // Unlike an ordinary resizable array implementation, this allows only element
197 // addition (by calling AllocSsaNum) and has special handling for RESERVED_SSA_NUM
198 // (basically it's a 1-based array). The array doesn't impose any particular
199 // requirements on the elements it stores and AllocSsaNum forwards its arguments
200 // to the array element constructor, this way the array supports both LclSsaVarDsc
201 // and SsaMemDef elements.
203 template <typename T>
207 unsigned m_arraySize;
210 static_assert_no_msg(SsaConfig::RESERVED_SSA_NUM == 0);
211 static_assert_no_msg(SsaConfig::FIRST_SSA_NUM == 1);
213 // Get the minimum valid SSA number.
214 unsigned GetMinSsaNum() const
216 return SsaConfig::FIRST_SSA_NUM;
219 // Increase (double) the size of the array.
220 void GrowArray(CompAllocator alloc)
222 unsigned oldSize = m_arraySize;
223 unsigned newSize = max(2, oldSize * 2);
225 T* newArray = alloc.allocate<T>(newSize);
227 for (unsigned i = 0; i < oldSize; i++)
229 newArray[i] = m_array[i];
233 m_arraySize = newSize;
237 // Construct an empty SsaDefArray.
238 SsaDefArray() : m_array(nullptr), m_arraySize(0), m_count(0)
242 // Reset the array (used only if the SSA form is reconstructed).
248 // Allocate a new SSA number (starting with SsaConfig::FIRST_SSA_NUM).
249 template <class... Args>
250 unsigned AllocSsaNum(CompAllocator alloc, Args&&... args)
252 if (m_count == m_arraySize)
257 unsigned ssaNum = GetMinSsaNum() + m_count;
258 m_array[m_count++] = T(jitstd::forward<Args>(args)...);
260 // Ensure that the first SSA number we allocate is SsaConfig::FIRST_SSA_NUM
261 assert((ssaNum == SsaConfig::FIRST_SSA_NUM) || (m_count > 1));
266 // Get the number of SSA definitions in the array.
267 unsigned GetCount() const
272 // Get a pointer to the SSA definition at the specified index.
273 T* GetSsaDefByIndex(unsigned index)
275 assert(index < m_count);
276 return &m_array[index];
279 // Check if the specified SSA number is valid.
280 bool IsValidSsaNum(unsigned ssaNum) const
282 return (GetMinSsaNum() <= ssaNum) && (ssaNum < (GetMinSsaNum() + m_count));
285 // Get a pointer to the SSA definition associated with the specified SSA number.
286 T* GetSsaDef(unsigned ssaNum)
288 assert(ssaNum != SsaConfig::RESERVED_SSA_NUM);
289 return GetSsaDefByIndex(ssaNum - GetMinSsaNum());
295 RCS_INVALID, // not valid to get/set ref counts
296 RCS_EARLY, // early counts for struct promotion and struct passing
297 RCS_NORMAL, // normal ref counts (from lvaMarkRefs onward)
303 // The constructor. Most things can just be zero'ed.
306 // note this only packs because var_types is a typedef of unsigned char
307 var_types lvType : 5; // TYP_INT/LONG/FLOAT/DOUBLE/REF
309 unsigned char lvIsParam : 1; // is this a parameter?
310 unsigned char lvIsRegArg : 1; // is this a register argument?
311 unsigned char lvFramePointerBased : 1; // 0 = off of REG_SPBASE (e.g., ESP), 1 = off of REG_FPBASE (e.g., EBP)
313 unsigned char lvStructGcCount : 3; // if struct, how many GC pointer (stop counting at 7). The only use of values >1
314 // is to help determine whether to use block init in the prolog.
315 unsigned char lvOnFrame : 1; // (part of) the variable lives on the frame
316 unsigned char lvRegister : 1; // assigned to live in a register? For RyuJIT backend, this is only set if the
317 // variable is in the same register for the entire function.
318 unsigned char lvTracked : 1; // is this a tracked variable?
319 bool lvTrackedNonStruct()
321 return lvTracked && lvType != TYP_STRUCT;
323 unsigned char lvPinned : 1; // is this a pinned variable?
325 unsigned char lvMustInit : 1; // must be initialized
326 unsigned char lvAddrExposed : 1; // The address of this variable is "exposed" -- passed as an argument, stored in a
327 // global location, etc.
328 // We cannot reason reliably about the value of the variable.
329 unsigned char lvDoNotEnregister : 1; // Do not enregister this variable.
330 unsigned char lvFieldAccessed : 1; // The var is a struct local, and a field of the variable is accessed. Affects
333 unsigned char lvInSsa : 1; // The variable is in SSA form (set by SsaBuilder)
336 // These further document the reasons for setting "lvDoNotEnregister". (Note that "lvAddrExposed" is one of the
338 // also, lvType == TYP_STRUCT prevents enregistration. At least one of the reasons should be true.
339 unsigned char lvVMNeedsStackAddr : 1; // The VM may have access to a stack-relative address of the variable, and
340 // read/write its value.
341 unsigned char lvLiveInOutOfHndlr : 1; // The variable was live in or out of an exception handler, and this required
342 // the variable to be
343 // in the stack (at least at those boundaries.)
344 unsigned char lvLclFieldExpr : 1; // The variable is not a struct, but was accessed like one (e.g., reading a
345 // particular byte from an int).
346 unsigned char lvLclBlockOpAddr : 1; // The variable was written to via a block operation that took its address.
347 unsigned char lvLiveAcrossUCall : 1; // The variable is live across an unmanaged call.
349 unsigned char lvIsCSE : 1; // Indicates if this LclVar is a CSE variable.
350 unsigned char lvHasLdAddrOp : 1; // has ldloca or ldarga opcode on this local.
351 unsigned char lvStackByref : 1; // This is a compiler temporary of TYP_BYREF that is known to point into our local
354 unsigned char lvHasILStoreOp : 1; // there is at least one STLOC or STARG on this local
355 unsigned char lvHasMultipleILStoreOp : 1; // there is more than one STLOC on this local
357 unsigned char lvIsTemp : 1; // Short-lifetime compiler temp (if lvIsParam is false), or implicit byref parameter
358 // (if lvIsParam is true)
360 unsigned char lvIsBoolean : 1; // set if variable is boolean
362 unsigned char lvSingleDef : 1; // variable has a single def
363 // before lvaMarkLocalVars: identifies ref type locals that can get type updates
364 // after lvaMarkLocalVars: identifies locals that are suitable for optAddCopies
367 unsigned char lvDisqualify : 1; // variable is no longer OK for add copy optimization
368 unsigned char lvVolatileHint : 1; // hint for AssertionProp
371 #ifndef _TARGET_64BIT_
372 unsigned char lvStructDoubleAlign : 1; // Must we double align this struct?
373 #endif // !_TARGET_64BIT_
374 #ifdef _TARGET_64BIT_
375 unsigned char lvQuirkToLong : 1; // Quirk to allocate this LclVar as a 64-bit long
378 unsigned char lvKeepType : 1; // Don't change the type of this variable
379 unsigned char lvNoLclFldStress : 1; // Can't apply local field stress on this one
381 unsigned char lvIsPtr : 1; // Might this be used in an address computation? (used by buffer overflow security
383 unsigned char lvIsUnsafeBuffer : 1; // Does this contain an unsafe buffer requiring buffer overflow security checks?
384 unsigned char lvPromoted : 1; // True when this local is a promoted struct, a normed struct, or a "split" long on a
385 // 32-bit target. For implicit byref parameters, this gets hijacked between
386 // fgRetypeImplicitByRefArgs and fgMarkDemotedImplicitByRefArgs to indicate whether
387 // references to the arg are being rewritten as references to a promoted shadow local.
388 unsigned char lvIsStructField : 1; // Is this local var a field of a promoted struct local?
389 unsigned char lvOverlappingFields : 1; // True when we have a struct with possibly overlapping fields
390 unsigned char lvContainsHoles : 1; // True when we have a promoted struct that contains holes
391 unsigned char lvCustomLayout : 1; // True when this struct has "CustomLayout"
393 unsigned char lvIsMultiRegArg : 1; // true if this is a multireg LclVar struct used in an argument context
394 unsigned char lvIsMultiRegRet : 1; // true if this is a multireg LclVar struct assigned from a multireg call
397 unsigned char _lvIsHfa : 1; // Is this a struct variable who's class handle is an HFA type
398 unsigned char _lvIsHfaRegArg : 1; // Is this a HFA argument variable? // TODO-CLEANUP: Remove this and replace
399 // with (lvIsRegArg && lvIsHfa())
400 unsigned char _lvHfaTypeIsFloat : 1; // Is the HFA type float or double?
401 #endif // FEATURE_HFA
404 // TODO-Cleanup: See the note on lvSize() - this flag is only in use by asserts that are checking for struct
405 // types, and is needed because of cases where TYP_STRUCT is bashed to an integral type.
406 // Consider cleaning this up so this workaround is not required.
407 unsigned char lvUnusedStruct : 1; // All references to this promoted struct are through its field locals.
408 // I.e. there is no longer any reference to the struct directly.
409 // In this case we can simply remove this struct local.
412 unsigned char lvLRACandidate : 1; // Tracked for linear scan register allocation purposes
415 // Note that both SIMD vector args and locals are marked as lvSIMDType = true, but the
416 // type of an arg node is TYP_BYREF and a local node is TYP_SIMD*.
417 unsigned char lvSIMDType : 1; // This is a SIMD struct
418 unsigned char lvUsedInSIMDIntrinsic : 1; // This tells lclvar is used for simd intrinsic
419 var_types lvBaseType : 5; // Note: this only packs because var_types is a typedef of unsigned char
420 #endif // FEATURE_SIMD
421 unsigned char lvRegStruct : 1; // This is a reg-sized non-field-addressed struct.
423 unsigned char lvClassIsExact : 1; // lvClassHandle is the exact type
426 unsigned char lvClassInfoUpdated : 1; // true if this var has updated class handle or exactness
429 unsigned char lvImplicitlyReferenced : 1; // true if there are non-IR references to this local (prolog, epilog, gc,
433 unsigned lvFieldLclStart; // The index of the local var representing the first field in the promoted struct
434 // local. For implicit byref parameters, this gets hijacked between
435 // fgRetypeImplicitByRefArgs and fgMarkDemotedImplicitByRefArgs to point to the
436 // struct local created to model the parameter's struct promotion, if any.
437 unsigned lvParentLcl; // The index of the local var representing the parent (i.e. the promoted struct local).
438 // Valid on promoted struct local fields.
441 unsigned char lvFieldCnt; // Number of fields in the promoted VarDsc.
442 unsigned char lvFldOffset;
443 unsigned char lvFldOrdinal;
445 #if FEATURE_MULTIREG_ARGS
446 regNumber lvRegNumForSlot(unsigned slotNum)
452 else if (slotNum == 1)
454 return lvOtherArgReg;
458 assert(false && "Invalid slotNum!");
463 #endif // FEATURE_MULTIREG_ARGS
481 bool lvIsHfaRegArg() const
484 return _lvIsHfaRegArg;
490 void lvSetIsHfaRegArg(bool value = true)
493 _lvIsHfaRegArg = value;
497 bool lvHfaTypeIsFloat() const
500 return _lvHfaTypeIsFloat;
506 void lvSetHfaTypeIsFloat(bool value)
509 _lvHfaTypeIsFloat = value;
513 // on Arm64 - Returns 1-4 indicating the number of register slots used by the HFA
514 // on Arm32 - Returns the total number of single FP register slots used by the HFA, max is 8
516 unsigned lvHfaSlots() const
519 assert(varTypeIsStruct(lvType));
521 return lvExactSize / sizeof(float);
522 #else // _TARGET_ARM64_
523 if (lvHfaTypeIsFloat())
525 return lvExactSize / sizeof(float);
529 return lvExactSize / sizeof(double);
531 #endif // _TARGET_ARM64_
534 // lvIsMultiRegArgOrRet()
535 // returns true if this is a multireg LclVar struct used in an argument context
536 // or if this is a multireg LclVar struct assigned from a multireg call
537 bool lvIsMultiRegArgOrRet()
539 return lvIsMultiRegArg || lvIsMultiRegRet;
543 regNumberSmall _lvRegNum; // Used to store the register this variable is in (or, the low register of a
544 // register pair). It is set during codegen any time the
545 // variable is enregistered (lvRegister is only set
546 // to non-zero if the variable gets the same register assignment for its entire
548 #if !defined(_TARGET_64BIT_)
549 regNumberSmall _lvOtherReg; // Used for "upper half" of long var.
550 #endif // !defined(_TARGET_64BIT_)
552 regNumberSmall _lvArgReg; // The register in which this argument is passed.
554 #if FEATURE_MULTIREG_ARGS
555 regNumberSmall _lvOtherArgReg; // Used for the second part of the struct passed in a register.
556 // Note this is defined but not used by ARM32
557 #endif // FEATURE_MULTIREG_ARGS
559 regNumberSmall _lvArgInitReg; // the register into which the argument is moved at entry
562 // The register number is stored in a small format (8 bits), but the getters return and the setters take
563 // a full-size (unsigned) format, to localize the casts here.
565 /////////////////////
567 __declspec(property(get = GetRegNum, put = SetRegNum)) regNumber lvRegNum;
569 regNumber GetRegNum() const
571 return (regNumber)_lvRegNum;
574 void SetRegNum(regNumber reg)
576 _lvRegNum = (regNumberSmall)reg;
577 assert(_lvRegNum == reg);
580 /////////////////////
582 #if defined(_TARGET_64BIT_)
583 __declspec(property(get = GetOtherReg, put = SetOtherReg)) regNumber lvOtherReg;
585 regNumber GetOtherReg() const
587 assert(!"shouldn't get here"); // can't use "unreached();" because it's NORETURN, which causes C4072
588 // "unreachable code" warnings
592 void SetOtherReg(regNumber reg)
594 assert(!"shouldn't get here"); // can't use "unreached();" because it's NORETURN, which causes C4072
595 // "unreachable code" warnings
597 #else // !_TARGET_64BIT_
598 __declspec(property(get = GetOtherReg, put = SetOtherReg)) regNumber lvOtherReg;
600 regNumber GetOtherReg() const
602 return (regNumber)_lvOtherReg;
605 void SetOtherReg(regNumber reg)
607 _lvOtherReg = (regNumberSmall)reg;
608 assert(_lvOtherReg == reg);
610 #endif // !_TARGET_64BIT_
612 /////////////////////
614 __declspec(property(get = GetArgReg, put = SetArgReg)) regNumber lvArgReg;
616 regNumber GetArgReg() const
618 return (regNumber)_lvArgReg;
621 void SetArgReg(regNumber reg)
623 _lvArgReg = (regNumberSmall)reg;
624 assert(_lvArgReg == reg);
627 #if FEATURE_MULTIREG_ARGS
628 __declspec(property(get = GetOtherArgReg, put = SetOtherArgReg)) regNumber lvOtherArgReg;
630 regNumber GetOtherArgReg() const
632 return (regNumber)_lvOtherArgReg;
635 void SetOtherArgReg(regNumber reg)
637 _lvOtherArgReg = (regNumberSmall)reg;
638 assert(_lvOtherArgReg == reg);
640 #endif // FEATURE_MULTIREG_ARGS
643 // Is this is a SIMD struct?
644 bool lvIsSIMDType() const
649 // Is this is a SIMD struct which is used for SIMD intrinsic?
650 bool lvIsUsedInSIMDIntrinsic() const
652 return lvUsedInSIMDIntrinsic;
655 // If feature_simd not enabled, return false
656 bool lvIsSIMDType() const
660 bool lvIsUsedInSIMDIntrinsic() const
666 /////////////////////
668 __declspec(property(get = GetArgInitReg, put = SetArgInitReg)) regNumber lvArgInitReg;
670 regNumber GetArgInitReg() const
672 return (regNumber)_lvArgInitReg;
675 void SetArgInitReg(regNumber reg)
677 _lvArgInitReg = (regNumberSmall)reg;
678 assert(_lvArgInitReg == reg);
681 /////////////////////
683 bool lvIsRegCandidate() const
685 return lvLRACandidate != 0;
688 bool lvIsInReg() const
690 return lvIsRegCandidate() && (lvRegNum != REG_STK);
693 regMaskTP lvRegMask() const
695 regMaskTP regMask = RBM_NONE;
696 if (varTypeIsFloating(TypeGet()))
698 if (lvRegNum != REG_STK)
700 regMask = genRegMaskFloat(lvRegNum, TypeGet());
705 if (lvRegNum != REG_STK)
707 regMask = genRegMask(lvRegNum);
713 unsigned short lvVarIndex; // variable tracking index
716 unsigned short m_lvRefCnt; // unweighted (real) reference count. For implicit by reference
717 // parameters, this gets hijacked from fgMarkImplicitByRefArgs
718 // through fgMarkDemotedImplicitByRefArgs, to provide a static
719 // appearance count (computed during address-exposed analysis)
720 // that fgMakeOutgoingStructArgCopy consults during global morph
721 // to determine if eliding its copy is legal.
723 BasicBlock::weight_t m_lvRefCntWtd; // weighted reference count
726 unsigned short lvRefCnt(RefCountState state = RCS_NORMAL) const;
727 void incLvRefCnt(unsigned short delta, RefCountState state = RCS_NORMAL);
728 void setLvRefCnt(unsigned short newValue, RefCountState state = RCS_NORMAL);
730 BasicBlock::weight_t lvRefCntWtd(RefCountState state = RCS_NORMAL) const;
731 void incLvRefCntWtd(BasicBlock::weight_t delta, RefCountState state = RCS_NORMAL);
732 void setLvRefCntWtd(BasicBlock::weight_t newValue, RefCountState state = RCS_NORMAL);
734 int lvStkOffs; // stack offset of home
735 unsigned lvExactSize; // (exact) size of the type in bytes
737 // Is this a promoted struct?
738 // This method returns true only for structs (including SIMD structs), not for
739 // locals that are split on a 32-bit target.
740 // It is only necessary to use this:
741 // 1) if only structs are wanted, and
742 // 2) if Lowering has already been done.
743 // Otherwise lvPromoted is valid.
744 bool lvPromotedStruct()
746 #if !defined(_TARGET_64BIT_)
747 return (lvPromoted && !varTypeIsLong(lvType));
748 #else // defined(_TARGET_64BIT_)
750 #endif // defined(_TARGET_64BIT_)
753 unsigned lvSize() const // Size needed for storage representation. Only used for structs or TYP_BLK.
755 // TODO-Review: Sometimes we get called on ARM with HFA struct variables that have been promoted,
756 // where the struct itself is no longer used because all access is via its member fields.
757 // When that happens, the struct is marked as unused and its type has been changed to
758 // TYP_INT (to keep the GC tracking code from looking at it).
759 // See Compiler::raAssignVars() for details. For example:
760 // N002 ( 4, 3) [00EA067C] ------------- return struct $346
761 // N001 ( 3, 2) [00EA0628] ------------- lclVar struct(U) V03 loc2
762 // float V03.f1 (offs=0x00) -> V12 tmp7
763 // f8 (last use) (last use) $345
764 // Here, the "struct(U)" shows that the "V03 loc2" variable is unused. Not shown is that V03
765 // is now TYP_INT in the local variable table. It's not really unused, because it's in the tree.
767 assert(varTypeIsStruct(lvType) || (lvType == TYP_BLK) || (lvPromoted && lvUnusedStruct));
769 #if defined(FEATURE_SIMD) && !defined(_TARGET_64BIT_)
770 // For 32-bit architectures, we make local variable SIMD12 types 16 bytes instead of just 12. We can't do
771 // this for arguments, which must be passed according the defined ABI. We don't want to do this for
772 // dependently promoted struct fields, but we don't know that here. See lvaMapSimd12ToSimd16().
773 // (Note that for 64-bits, we are already rounding up to 16.)
774 if ((lvType == TYP_SIMD12) && !lvIsParam)
776 assert(lvExactSize == 12);
779 #endif // defined(FEATURE_SIMD) && !defined(_TARGET_64BIT_)
781 return roundUp(lvExactSize, TARGET_POINTER_SIZE);
784 size_t lvArgStackSize() const;
786 unsigned lvSlotNum; // original slot # (if remapped)
788 typeInfo lvVerTypeInfo; // type info needed for verification
790 CORINFO_CLASS_HANDLE lvClassHnd; // class handle for the local, or null if not known
792 CORINFO_FIELD_HANDLE lvFieldHnd; // field handle for promoted struct fields
794 BYTE* lvGcLayout; // GC layout info for structs
797 BlockSet lvRefBlks; // Set of blocks that contain refs
798 GenTree* lvDefStmt; // Pointer to the statement with the single definition
799 void lvaDisqualifyVar(); // Call to disqualify a local variable from use in optAddCopies
801 var_types TypeGet() const
803 return (var_types)lvType;
805 bool lvStackAligned() const
807 assert(lvIsStructField);
808 return ((lvFldOffset % TARGET_POINTER_SIZE) == 0);
810 bool lvNormalizeOnLoad() const
812 return varTypeIsSmall(TypeGet()) &&
813 // lvIsStructField is treated the same as the aliased local, see fgDoNormalizeOnStore.
814 (lvIsParam || lvAddrExposed || lvIsStructField);
817 bool lvNormalizeOnStore()
819 return varTypeIsSmall(TypeGet()) &&
820 // lvIsStructField is treated the same as the aliased local, see fgDoNormalizeOnStore.
821 !(lvIsParam || lvAddrExposed || lvIsStructField);
824 void incRefCnts(BasicBlock::weight_t weight,
826 RefCountState state = RCS_NORMAL,
827 bool propagate = true);
828 bool IsFloatRegType() const
830 return isFloatRegType(lvType) || lvIsHfaRegArg();
832 var_types GetHfaType() const
834 return lvIsHfa() ? (lvHfaTypeIsFloat() ? TYP_FLOAT : TYP_DOUBLE) : TYP_UNDEF;
836 void SetHfaType(var_types type)
838 assert(varTypeIsFloating(type));
839 lvSetHfaTypeIsFloat(type == TYP_FLOAT);
842 var_types lvaArgType();
844 SsaDefArray<LclSsaVarDsc> lvPerSsaData;
846 // Returns the address of the per-Ssa data for the given ssaNum (which is required
847 // not to be the SsaConfig::RESERVED_SSA_NUM, which indicates that the variable is
848 // not an SSA variable).
849 LclSsaVarDsc* GetPerSsaData(unsigned ssaNum)
851 return lvPerSsaData.GetSsaDef(ssaNum);
856 const char* lvReason;
858 void PrintVarReg() const
860 printf("%s", getRegName(lvRegNum));
864 }; // class LclVarDsc
867 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
868 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
872 XX The temporary lclVars allocated by the compiler for code generation XX
874 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
875 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
878 /*****************************************************************************
880 * The following keeps track of temporaries allocated in the stack frame
881 * during code-generation (after register allocation). These spill-temps are
882 * only used if we run out of registers while evaluating a tree.
884 * These are different from the more common temps allocated by lvaGrabTemp().
895 static const int BAD_TEMP_OFFSET = 0xDDDDDDDD; // used as a sentinel "bad value" for tdOffs in DEBUG
903 TempDsc(int _tdNum, unsigned _tdSize, var_types _tdType) : tdNum(_tdNum), tdSize((BYTE)_tdSize), tdType(_tdType)
907 0); // temps must have a negative number (so they have a different number from all local variables)
908 tdOffs = BAD_TEMP_OFFSET;
912 IMPL_LIMITATION("too many spill temps");
917 bool tdLegalOffset() const
919 return tdOffs != BAD_TEMP_OFFSET;
923 int tdTempOffs() const
925 assert(tdLegalOffset());
928 void tdSetTempOffs(int offs)
931 assert(tdLegalOffset());
933 void tdAdjustTempOffs(int offs)
936 assert(tdLegalOffset());
939 int tdTempNum() const
944 unsigned tdTempSize() const
948 var_types tdTempType() const
954 // interface to hide linearscan implementation from rest of compiler
955 class LinearScanInterface
958 virtual void doLinearScan() = 0;
959 virtual void recordVarLocationsAtStartOfBB(BasicBlock* bb) = 0;
960 virtual bool willEnregisterLocalVars() const = 0;
963 LinearScanInterface* getLinearScanAllocator(Compiler* comp);
965 // Information about arrays: their element type and size, and the offset of the first element.
966 // We label GT_IND's that are array indices with GTF_IND_ARR_INDEX, and, for such nodes,
967 // associate an array info via the map retrieved by GetArrayInfoMap(). This information is used,
968 // for example, in value numbering of array index expressions.
971 var_types m_elemType;
972 CORINFO_CLASS_HANDLE m_elemStructType;
974 unsigned m_elemOffset;
976 ArrayInfo() : m_elemType(TYP_UNDEF), m_elemStructType(nullptr), m_elemSize(0), m_elemOffset(0)
980 ArrayInfo(var_types elemType, unsigned elemSize, unsigned elemOffset, CORINFO_CLASS_HANDLE elemStructType)
981 : m_elemType(elemType), m_elemStructType(elemStructType), m_elemSize(elemSize), m_elemOffset(elemOffset)
986 // This enumeration names the phases into which we divide compilation. The phases should completely
987 // partition a compilation.
990 #define CompPhaseNameMacro(enum_nm, string_nm, short_nm, hasChildren, parent, measureIR) enum_nm,
991 #include "compphases.h"
995 extern const char* PhaseNames[];
996 extern const char* PhaseEnums[];
997 extern const LPCWSTR PhaseShortNames[];
999 // The following enum provides a simple 1:1 mapping to CLR API's
1000 enum API_ICorJitInfo_Names
1002 #define DEF_CLR_API(name) API_##name,
1003 #include "ICorJitInfo_API_names.h"
1007 //---------------------------------------------------------------
1008 // Compilation time.
1011 // A "CompTimeInfo" is a structure for tracking the compilation time of one or more methods.
1012 // We divide a compilation into a sequence of contiguous phases, and track the total (per-thread) cycles
1013 // of the compilation, as well as the cycles for each phase. We also track the number of bytecodes.
1014 // If there is a failure in reading a timer at any point, the "CompTimeInfo" becomes invalid, as indicated
1015 // by "m_timerFailure" being true.
1016 // If FEATURE_JIT_METHOD_PERF is not set, we define a minimal form of this, enough to let other code compile.
1019 #ifdef FEATURE_JIT_METHOD_PERF
1020 // The string names of the phases.
1021 static const char* PhaseNames[];
1023 static bool PhaseHasChildren[];
1024 static int PhaseParent[];
1025 static bool PhaseReportsIRSize[];
1027 unsigned m_byteCodeBytes;
1028 unsigned __int64 m_totalCycles;
1029 unsigned __int64 m_invokesByPhase[PHASE_NUMBER_OF];
1030 unsigned __int64 m_cyclesByPhase[PHASE_NUMBER_OF];
1031 #if MEASURE_CLRAPI_CALLS
1032 unsigned __int64 m_CLRinvokesByPhase[PHASE_NUMBER_OF];
1033 unsigned __int64 m_CLRcyclesByPhase[PHASE_NUMBER_OF];
1036 unsigned m_nodeCountAfterPhase[PHASE_NUMBER_OF];
1038 // For better documentation, we call EndPhase on
1039 // non-leaf phases. We should also call EndPhase on the
1040 // last leaf subphase; obviously, the elapsed cycles between the EndPhase
1041 // for the last leaf subphase and the EndPhase for an ancestor should be very small.
1042 // We add all such "redundant end phase" intervals to this variable below; we print
1043 // it out in a report, so we can verify that it is, indeed, very small. If it ever
1044 // isn't, this means that we're doing something significant between the end of the last
1045 // declared subphase and the end of its parent.
1046 unsigned __int64 m_parentPhaseEndSlop;
1047 bool m_timerFailure;
1049 #if MEASURE_CLRAPI_CALLS
1050 // The following measures the time spent inside each individual CLR API call.
1051 unsigned m_allClrAPIcalls;
1052 unsigned m_perClrAPIcalls[API_ICorJitInfo_Names::API_COUNT];
1053 unsigned __int64 m_allClrAPIcycles;
1054 unsigned __int64 m_perClrAPIcycles[API_ICorJitInfo_Names::API_COUNT];
1055 unsigned __int32 m_maxClrAPIcycles[API_ICorJitInfo_Names::API_COUNT];
1056 #endif // MEASURE_CLRAPI_CALLS
1058 CompTimeInfo(unsigned byteCodeBytes);
1062 #ifdef FEATURE_JIT_METHOD_PERF
1064 #if MEASURE_CLRAPI_CALLS
1065 struct WrapICorJitInfo;
1068 // This class summarizes the JIT time information over the course of a run: the number of methods compiled,
1069 // and the total and maximum timings. (These are instances of the "CompTimeInfo" type described above).
1070 // The operation of adding a single method's timing to the summary may be performed concurrently by several
1071 // threads, so it is protected by a lock.
1072 // This class is intended to be used as a singleton type, with only a single instance.
1073 class CompTimeSummaryInfo
1075 // This lock protects the fields of all CompTimeSummaryInfo(s) (of which we expect there to be one).
1076 static CritSecObject s_compTimeSummaryLock;
1080 CompTimeInfo m_total;
1081 CompTimeInfo m_maximum;
1083 int m_numFilteredMethods;
1084 CompTimeInfo m_filtered;
1086 // This can use what ever data you want to determine if the value to be added
1087 // belongs in the filtered section (it's always included in the unfiltered section)
1088 bool IncludedInFilteredData(CompTimeInfo& info);
1091 // This is the unique CompTimeSummaryInfo object for this instance of the runtime.
1092 static CompTimeSummaryInfo s_compTimeSummary;
1094 CompTimeSummaryInfo()
1095 : m_numMethods(0), m_totMethods(0), m_total(0), m_maximum(0), m_numFilteredMethods(0), m_filtered(0)
1099 // Assumes that "info" is a completed CompTimeInfo for a compilation; adds it to the summary.
1100 // This is thread safe.
1101 void AddInfo(CompTimeInfo& info, bool includePhases);
1103 // Print the summary information to "f".
1104 // This is not thread-safe; assumed to be called by only one thread.
1105 void Print(FILE* f);
1108 // A JitTimer encapsulates a CompTimeInfo for a single compilation. It also tracks the start of compilation,
1109 // and when the current phase started. This is intended to be part of a Compilation object. This is
1110 // disabled (FEATURE_JIT_METHOD_PERF not defined) when FEATURE_CORECLR is set, or on non-windows platforms.
1114 unsigned __int64 m_start; // Start of the compilation.
1115 unsigned __int64 m_curPhaseStart; // Start of the current phase.
1116 #if MEASURE_CLRAPI_CALLS
1117 unsigned __int64 m_CLRcallStart; // Start of the current CLR API call (if any).
1118 unsigned __int64 m_CLRcallInvokes; // CLR API invokes under current outer so far
1119 unsigned __int64 m_CLRcallCycles; // CLR API cycles under current outer so far.
1120 int m_CLRcallAPInum; // The enum/index of the current CLR API call (or -1).
1121 static double s_cyclesPerSec; // Cached for speedier measurements
1124 Phases m_lastPhase; // The last phase that was completed (or (Phases)-1 to start).
1126 CompTimeInfo m_info; // The CompTimeInfo for this compilation.
1128 static CritSecObject s_csvLock; // Lock to protect the time log file.
1129 void PrintCsvMethodStats(Compiler* comp);
1132 void* operator new(size_t);
1133 void* operator new[](size_t);
1134 void operator delete(void*);
1135 void operator delete[](void*);
1138 // Initialized the timer instance
1139 JitTimer(unsigned byteCodeSize);
1141 static JitTimer* Create(Compiler* comp, unsigned byteCodeSize)
1143 return ::new (comp, CMK_Unknown) JitTimer(byteCodeSize);
1146 static void PrintCsvHeader();
1148 // Ends the current phase (argument is for a redundant check).
1149 void EndPhase(Compiler* compiler, Phases phase);
1151 #if MEASURE_CLRAPI_CALLS
1152 // Start and end a timed CLR API call.
1153 void CLRApiCallEnter(unsigned apix);
1154 void CLRApiCallLeave(unsigned apix);
1155 #endif // MEASURE_CLRAPI_CALLS
1157 // Completes the timing of the current method, which is assumed to have "byteCodeBytes" bytes of bytecode,
1158 // and adds it to "sum".
1159 void Terminate(Compiler* comp, CompTimeSummaryInfo& sum, bool includePhases);
1161 // Attempts to query the cycle counter of the current thread. If successful, returns "true" and sets
1162 // *cycles to the cycle counter value. Otherwise, returns false and sets the "m_timerFailure" flag of
1163 // "m_info" to true.
1164 bool GetThreadCycles(unsigned __int64* cycles)
1166 bool res = CycleTimer::GetThreadCyclesS(cycles);
1169 m_info.m_timerFailure = true;
1174 #endif // FEATURE_JIT_METHOD_PERF
1176 //------------------- Function/Funclet info -------------------------------
1177 enum FuncKind : BYTE
1179 FUNC_ROOT, // The main/root function (always id==0)
1180 FUNC_HANDLER, // a funclet associated with an EH handler (finally, fault, catch, filter handler)
1181 FUNC_FILTER, // a funclet associated with an EH filter
1190 BYTE funFlags; // Currently unused, just here for padding
1191 unsigned short funEHIndex; // index, into the ebd table, of innermost EH clause corresponding to this
1192 // funclet. It is only valid if funKind field indicates this is a
1193 // EH-related funclet: FUNC_HANDLER or FUNC_FILTER
1195 #if defined(_TARGET_AMD64_)
1197 // TODO-AMD64-Throughput: make the AMD64 info more like the ARM info to avoid having this large static array.
1198 emitLocation* startLoc;
1199 emitLocation* endLoc;
1200 emitLocation* coldStartLoc; // locations for the cold section, if there is one.
1201 emitLocation* coldEndLoc;
1202 UNWIND_INFO unwindHeader;
1203 // Maximum of 255 UNWIND_CODE 'nodes' and then the unwind header. If there are an odd
1204 // number of codes, the VM or Zapper will 4-byte align the whole thing.
1205 BYTE unwindCodes[offsetof(UNWIND_INFO, UnwindCode) + (0xFF * sizeof(UNWIND_CODE))];
1206 unsigned unwindCodeSlot;
1208 #elif defined(_TARGET_X86_)
1210 #if defined(_TARGET_UNIX_)
1211 emitLocation* startLoc;
1212 emitLocation* endLoc;
1213 emitLocation* coldStartLoc; // locations for the cold section, if there is one.
1214 emitLocation* coldEndLoc;
1215 #endif // _TARGET_UNIX_
1217 #elif defined(_TARGET_ARMARCH_)
1219 UnwindInfo uwi; // Unwind information for this function/funclet's hot section
1220 UnwindInfo* uwiCold; // Unwind information for this function/funclet's cold section
1221 // Note: we only have a pointer here instead of the actual object,
1222 // to save memory in the JIT case (compared to the NGEN case),
1223 // where we don't have any cold section.
1224 // Note 2: we currently don't support hot/cold splitting in functions
1225 // with EH, so uwiCold will be NULL for all funclets.
1227 #if defined(_TARGET_UNIX_)
1228 emitLocation* startLoc;
1229 emitLocation* endLoc;
1230 emitLocation* coldStartLoc; // locations for the cold section, if there is one.
1231 emitLocation* coldEndLoc;
1232 #endif // _TARGET_UNIX_
1234 #endif // _TARGET_ARMARCH_
1236 #if defined(_TARGET_UNIX_)
1237 jitstd::vector<CFI_CODE>* cfiCodes;
1238 #endif // _TARGET_UNIX_
1240 // Eventually we may want to move rsModifiedRegsMask, lvaOutgoingArgSize, and anything else
1241 // that isn't shared between the main function body and funclets.
1244 struct fgArgTabEntry
1246 GenTree* node; // Initially points at the Op1 field of 'parent', but if the argument is replaced with an GT_ASG or
1247 // placeholder it will point at the actual argument in the gtCallLateArgs list.
1248 GenTree* parent; // Points at the GT_LIST node in the gtCallArgs for this argument
1250 unsigned argNum; // The original argument number, also specifies the required argument evaluation order from the IL
1253 regNumberSmall regNums[MAX_ARG_REG_COUNT]; // The registers to use when passing this argument, set to REG_STK for
1254 // arguments passed on the stack
1256 unsigned numRegs; // Count of number of registers that this argument uses.
1257 // Note that on ARM, if we have a double hfa, this reflects the number
1258 // of DOUBLE registers.
1260 // A slot is a pointer sized region in the OutArg area.
1261 unsigned slotNum; // When an argument is passed in the OutArg area this is the slot number in the OutArg area
1262 unsigned numSlots; // Count of number of slots that this argument uses
1264 unsigned alignment; // 1 or 2 (slots/registers)
1266 unsigned _lateArgInx; // index into gtCallLateArgs list; UINT_MAX if this is not a late arg.
1268 unsigned tmpNum; // the LclVar number if we had to force evaluation of this arg
1270 var_types argType; // The type used to pass this argument. This is generally the original argument type, but when a
1271 // struct is passed as a scalar type, this is that type.
1272 // Note that if a struct is passed by reference, this will still be the struct type.
1274 bool needTmp : 1; // True when we force this argument's evaluation into a temp LclVar
1275 bool needPlace : 1; // True when we must replace this argument with a placeholder node
1276 bool isTmp : 1; // True when we setup a temp LclVar for this argument due to size issues with the struct
1277 bool processed : 1; // True when we have decided the evaluation order for this argument in the gtCallLateArgs
1278 bool isBackFilled : 1; // True when the argument fills a register slot skipped due to alignment requirements of
1279 // previous arguments.
1280 bool isNonStandard : 1; // True if it is an arg that is passed in a reg other than a standard arg reg, or is forced
1281 // to be on the stack despite its arg list position.
1282 bool isStruct : 1; // True if this is a struct arg
1283 bool _isVararg : 1; // True if the argument is in a vararg context.
1284 bool passedByRef : 1; // True iff the argument is passed by reference.
1285 #ifdef FEATURE_ARG_SPLIT
1286 bool _isSplit : 1; // True when this argument is split between the registers and OutArg area
1287 #endif // FEATURE_ARG_SPLIT
1289 bool _isHfaArg : 1; // True when the argument is an HFA type.
1290 bool _isDoubleHfa : 1; // True when the argument is an HFA, with an element type of DOUBLE.
1295 bool isLate = (_lateArgInx != UINT_MAX);
1299 __declspec(property(get = getLateArgInx, put = setLateArgInx)) unsigned lateArgInx;
1300 unsigned getLateArgInx()
1302 assert(isLateArg());
1305 void setLateArgInx(unsigned inx)
1309 __declspec(property(get = getRegNum)) regNumber regNum;
1310 regNumber getRegNum()
1312 return (regNumber)regNums[0];
1314 __declspec(property(get = getOtherRegNum)) regNumber otherRegNum;
1315 regNumber getOtherRegNum()
1317 return (regNumber)regNums[1];
1320 #if defined(UNIX_AMD64_ABI)
1321 SYSTEMV_AMD64_CORINFO_STRUCT_REG_PASSING_DESCRIPTOR structDesc;
1324 void setRegNum(unsigned int i, regNumber regNum)
1326 assert(i < MAX_ARG_REG_COUNT);
1327 regNums[i] = (regNumberSmall)regNum;
1329 regNumber getRegNum(unsigned int i)
1331 assert(i < MAX_ARG_REG_COUNT);
1332 return (regNumber)regNums[i];
1335 __declspec(property(get = getIsSplit, put = setIsSplit)) bool isSplit;
1338 #ifdef FEATURE_ARG_SPLIT
1340 #else // FEATURE_ARG_SPLIT
1344 void setIsSplit(bool value)
1346 #ifdef FEATURE_ARG_SPLIT
1351 __declspec(property(get = getIsVararg, put = setIsVararg)) bool isVararg;
1354 #ifdef FEATURE_VARARG
1360 void setIsVararg(bool value)
1362 #ifdef FEATURE_VARARG
1364 #endif // FEATURE_VARARG
1367 __declspec(property(get = getIsHfaArg)) bool isHfaArg;
1377 __declspec(property(get = getIsHfaRegArg)) bool isHfaRegArg;
1378 bool getIsHfaRegArg()
1381 return _isHfaArg && isPassedInRegisters();
1387 __declspec(property(get = getHfaType)) var_types hfaType;
1388 var_types getHfaType()
1391 return _isHfaArg ? (_isDoubleHfa ? TYP_DOUBLE : TYP_FLOAT) : TYP_UNDEF;
1397 void setHfaType(var_types type, unsigned hfaSlots)
1400 if (type != TYP_UNDEF)
1402 // We must already have set the passing mode.
1403 assert(numRegs != 0 || numSlots != 0);
1404 // We originally set numRegs according to the size of the struct, but if the size of the
1405 // hfaType is not the same as the pointer size, we need to correct it.
1406 // Note that hfaSlots is the number of registers we will use. For ARM, that is twice
1407 // the number of "double registers".
1408 unsigned numHfaRegs = hfaSlots;
1409 if (isPassedInRegisters())
1412 if (type == TYP_DOUBLE)
1414 // Must be an even number of registers.
1415 assert((numRegs & 1) == 0);
1416 numHfaRegs = hfaSlots / 2;
1418 #endif // _TARGET_ARM_
1421 // This should already be set correctly.
1422 assert(numRegs == numHfaRegs);
1423 assert(_isDoubleHfa == (type == TYP_DOUBLE));
1427 numRegs = numHfaRegs;
1430 _isDoubleHfa = (type == TYP_DOUBLE);
1433 #endif // FEATURE_HFA
1437 void SetIsBackFilled(bool backFilled)
1439 isBackFilled = backFilled;
1442 bool IsBackFilled() const
1444 return isBackFilled;
1446 #else // !_TARGET_ARM_
1447 void SetIsBackFilled(bool backFilled)
1451 bool IsBackFilled() const
1455 #endif // !_TARGET_ARM_
1457 bool isPassedInRegisters()
1459 return !isSplit && (numRegs != 0);
1462 bool isPassedInFloatRegisters()
1467 return isValidFloatArgReg(regNum);
1471 bool isSingleRegOrSlot()
1473 return !isSplit && ((numRegs == 1) || (numSlots == 1));
1476 // Returns the number of "slots" used, where for this purpose a
1477 // register counts as a slot.
1478 unsigned getSlotCount()
1482 assert(isPassedInRegisters());
1483 assert(numRegs == 1);
1485 else if (regNum == REG_STK)
1487 assert(!isPassedInRegisters());
1488 assert(numRegs == 0);
1492 assert(numRegs > 0);
1494 return numSlots + numRegs;
1497 // Returns the size as a multiple of pointer-size.
1498 // For targets without HFAs, this is the same as getSlotCount().
1501 unsigned size = getSlotCount();
1504 // We counted the number of regs, but if they are DOUBLE hfa regs we have to double the size.
1505 if (isHfaRegArg && (hfaType == TYP_DOUBLE))
1510 #elif defined(_TARGET_ARM64_)
1511 // We counted the number of regs, but if they are FLOAT hfa regs we have to halve the size.
1512 if (isHfaRegArg && (hfaType == TYP_FLOAT))
1514 // Round up in case of odd HFA count.
1515 size = (size + 1) >> 1;
1517 #endif // _TARGET_ARM64_
1522 // Set the register numbers for a multireg argument.
1523 // There's nothing to do on x64/Ux because the structDesc has already been used to set the
1524 // register numbers.
1525 void SetMultiRegNums()
1527 #if FEATURE_MULTIREG_ARGS && !defined(UNIX_AMD64_ABI)
1533 regNumber argReg = getRegNum(0);
1535 unsigned int regSize = (hfaType == TYP_DOUBLE) ? 2 : 1;
1537 unsigned int regSize = 1;
1539 for (unsigned int regIndex = 1; regIndex < numRegs; regIndex++)
1541 argReg = (regNumber)(argReg + regSize);
1542 setRegNum(regIndex, argReg);
1544 #endif // FEATURE_MULTIREG_ARGS && !defined(UNIX_AMD64_ABI)
1547 // Check that the value of 'isStruct' is consistent.
1548 // A struct arg must be one of the following:
1549 // - A node of struct type,
1550 // - A GT_FIELD_LIST, or
1551 // - A node of a scalar type, passed in a single register or slot
1552 // (or two slots in the case of a struct pass on the stack as TYP_DOUBLE).
1554 void checkIsStruct()
1558 if (!varTypeIsStruct(node) && !node->OperIs(GT_FIELD_LIST))
1560 // This is the case where we are passing a struct as a primitive type.
1561 // On most targets, this is always a single register or slot.
1562 // However, on ARM this could be two slots if it is TYP_DOUBLE.
1563 bool isPassedAsPrimitiveType = ((numRegs == 1) || ((numRegs == 0) && (numSlots == 1)));
1565 if (!isPassedAsPrimitiveType)
1567 if (node->TypeGet() == TYP_DOUBLE && numRegs == 0 && (numSlots == 2))
1569 isPassedAsPrimitiveType = true;
1572 #endif // _TARGET_ARM_
1573 assert(isPassedAsPrimitiveType);
1578 assert(!varTypeIsStruct(node));
1587 //-------------------------------------------------------------------------
1589 // The class fgArgInfo is used to handle the arguments
1590 // when morphing a GT_CALL node.
1595 Compiler* compiler; // Back pointer to the compiler instance so that we can allocate memory
1596 GenTreeCall* callTree; // Back pointer to the GT_CALL node for this fgArgInfo
1597 unsigned argCount; // Updatable arg count value
1598 unsigned nextSlotNum; // Updatable slot count value
1599 unsigned stkLevel; // Stack depth when we make this call (for x86)
1601 #if defined(UNIX_X86_ABI)
1602 bool alignmentDone; // Updateable flag, set to 'true' after we've done any required alignment.
1603 unsigned stkSizeBytes; // Size of stack used by this call, in bytes. Calculated during fgMorphArgs().
1604 unsigned padStkAlign; // Stack alignment in bytes required before arguments are pushed for this call.
1605 // Computed dynamically during codegen, based on stkSizeBytes and the current
1606 // stack level (genStackLevel) when the first stack adjustment is made for
1610 #if FEATURE_FIXED_OUT_ARGS
1611 unsigned outArgSize; // Size of the out arg area for the call, will be at least MIN_ARG_AREA_FOR_CALL
1614 unsigned argTableSize; // size of argTable array (equal to the argCount when done with fgMorphArgs)
1615 bool hasRegArgs; // true if we have one or more register arguments
1616 bool hasStackArgs; // true if we have one or more stack arguments
1617 bool argsComplete; // marker for state
1618 bool argsSorted; // marker for state
1619 fgArgTabEntry** argTable; // variable sized array of per argument descrption: (i.e. argTable[argTableSize])
1622 void AddArg(fgArgTabEntry* curArgTabEntry);
1625 fgArgInfo(Compiler* comp, GenTreeCall* call, unsigned argCount);
1626 fgArgInfo(GenTreeCall* newCall, GenTreeCall* oldCall);
1628 fgArgTabEntry* AddRegArg(unsigned argNum,
1635 bool isVararg = false);
1637 #ifdef UNIX_AMD64_ABI
1638 fgArgTabEntry* AddRegArg(unsigned argNum,
1644 const bool isStruct,
1645 const bool isVararg,
1646 const regNumber otherRegNum,
1647 const SYSTEMV_AMD64_CORINFO_STRUCT_REG_PASSING_DESCRIPTOR* const structDescPtr = nullptr);
1648 #endif // UNIX_AMD64_ABI
1650 fgArgTabEntry* AddStkArg(unsigned argNum,
1656 bool isVararg = false);
1658 void RemorphReset();
1659 void UpdateRegArg(fgArgTabEntry* argEntry, GenTree* node, bool reMorphing);
1660 void UpdateStkArg(fgArgTabEntry* argEntry, GenTree* node, bool reMorphing);
1662 void SplitArg(unsigned argNum, unsigned numRegs, unsigned numSlots);
1664 void EvalToTmp(fgArgTabEntry* curArgTabEntry, unsigned tmpNum, GenTree* newNode);
1666 void ArgsComplete();
1670 void EvalArgsToTemps();
1676 fgArgTabEntry** ArgTable()
1680 unsigned GetNextSlotNum()
1690 return hasStackArgs;
1692 bool AreArgsComplete() const
1694 return argsComplete;
1696 #if FEATURE_FIXED_OUT_ARGS
1697 unsigned GetOutArgSize() const
1701 void SetOutArgSize(unsigned newVal)
1703 outArgSize = newVal;
1705 #endif // FEATURE_FIXED_OUT_ARGS
1707 #if defined(UNIX_X86_ABI)
1708 void ComputeStackAlignment(unsigned curStackLevelInBytes)
1710 padStkAlign = AlignmentPad(curStackLevelInBytes, STACK_ALIGN);
1713 unsigned GetStkAlign()
1718 void SetStkSizeBytes(unsigned newStkSizeBytes)
1720 stkSizeBytes = newStkSizeBytes;
1723 unsigned GetStkSizeBytes() const
1725 return stkSizeBytes;
1728 bool IsStkAlignmentDone() const
1730 return alignmentDone;
1733 void SetStkAlignmentDone()
1735 alignmentDone = true;
1737 #endif // defined(UNIX_X86_ABI)
1739 // Get the fgArgTabEntry for the arg at position argNum.
1740 fgArgTabEntry* GetArgEntry(unsigned argNum, bool reMorphing = true)
1742 fgArgTabEntry* curArgTabEntry = nullptr;
1746 // The arg table has not yet been sorted.
1747 curArgTabEntry = argTable[argNum];
1748 assert(curArgTabEntry->argNum == argNum);
1749 return curArgTabEntry;
1752 for (unsigned i = 0; i < argCount; i++)
1754 curArgTabEntry = argTable[i];
1755 if (curArgTabEntry->argNum == argNum)
1757 return curArgTabEntry;
1760 noway_assert(!"GetArgEntry: argNum not found");
1764 // Get the node for the arg at position argIndex.
1765 // Caller must ensure that this index is a valid arg index.
1766 GenTree* GetArgNode(unsigned argIndex)
1768 return GetArgEntry(argIndex)->node;
1771 void Dump(Compiler* compiler);
1775 // XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1776 // We have the ability to mark source expressions with "Test Labels."
1777 // These drive assertions within the JIT, or internal JIT testing. For example, we could label expressions
1778 // that should be CSE defs, and other expressions that should uses of those defs, with a shared label.
1780 enum TestLabel // This must be kept identical to System.Runtime.CompilerServices.JitTestLabel.TestLabel.
1783 TL_VN, // Defines a "VN equivalence class". (For full VN, including exceptions thrown).
1784 TL_VNNorm, // Like above, but uses the non-exceptional value of the expression.
1785 TL_CSE_Def, // This must be identified in the JIT as a CSE def
1786 TL_CSE_Use, // This must be identified in the JIT as a CSE use
1787 TL_LoopHoist, // Expression must (or must not) be hoisted out of the loop.
1790 struct TestLabelAndNum
1795 TestLabelAndNum() : m_tl(TestLabel(0)), m_num(0)
1800 typedef JitHashTable<GenTree*, JitPtrKeyFuncs<GenTree>, TestLabelAndNum> NodeToTestDataMap;
1802 // XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1806 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1807 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1809 XX The big guy. The sections are currently organized as : XX
1811 XX o GenTree and BasicBlock XX
1823 XX o PrologScopeInfo XX
1824 XX o CodeGenerator XX
1829 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1830 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1833 struct HWIntrinsicInfo;
1837 friend class emitter;
1838 friend class UnwindInfo;
1839 friend class UnwindFragmentInfo;
1840 friend class UnwindEpilogInfo;
1841 friend class JitTimer;
1842 friend class LinearScan;
1843 friend class fgArgInfo;
1844 friend class Rationalizer;
1846 friend class Lowering;
1847 friend class CSE_DataFlow;
1848 friend class CSE_Heuristic;
1849 friend class CodeGenInterface;
1850 friend class CodeGen;
1851 friend class LclVarDsc;
1852 friend class TempDsc;
1854 friend class ObjectAllocator;
1855 friend class LocalAddressVisitor;
1856 friend struct GenTree;
1858 #ifdef FEATURE_HW_INTRINSICS
1859 friend struct HWIntrinsicInfo;
1860 #endif // FEATURE_HW_INTRINSICS
1862 #ifndef _TARGET_64BIT_
1863 friend class DecomposeLongs;
1864 #endif // !_TARGET_64BIT_
1867 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1868 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1870 XX Misc structs definitions XX
1872 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1873 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
1877 hashBvGlobalData hbvGlobalData; // Used by the hashBv bitvector package.
1896 bool dumpIRDataflow;
1897 bool dumpIRBlockHeaders;
1899 LPCWSTR dumpIRPhase;
1900 LPCWSTR dumpIRFormat;
1902 bool shouldUseVerboseTrees();
1903 bool asciiTrees; // If true, dump trees using only ASCII characters
1904 bool shouldDumpASCIITrees();
1905 bool verboseSsa; // If true, produce especially verbose dump output in SSA construction.
1906 bool shouldUseVerboseSsa();
1907 bool treesBeforeAfterMorph; // If true, print trees before/after morphing (paired by an intra-compilation id:
1908 int morphNum; // This counts the the trees that have been morphed, allowing us to label each uniquely.
1910 const char* VarNameToStr(VarName name)
1915 DWORD expensiveDebugCheckLevel;
1918 #if FEATURE_MULTIREG_RET
1919 GenTree* impAssignMultiRegTypeToVar(GenTree* op, CORINFO_CLASS_HANDLE hClass);
1920 #endif // FEATURE_MULTIREG_RET
1922 GenTree* impAssignSmallStructTypeToVar(GenTree* op, CORINFO_CLASS_HANDLE hClass);
1925 bool isSingleFloat32Struct(CORINFO_CLASS_HANDLE hClass);
1926 #endif // ARM_SOFTFP
1928 //-------------------------------------------------------------------------
1929 // Functions to handle homogeneous floating-point aggregates (HFAs) in ARM.
1930 // HFAs are one to four element structs where each element is the same
1931 // type, either all float or all double. They are treated specially
1932 // in the ARM Procedure Call Standard, specifically, they are passed in
1933 // floating-point registers instead of the general purpose registers.
1936 bool IsHfa(CORINFO_CLASS_HANDLE hClass);
1937 bool IsHfa(GenTree* tree);
1939 var_types GetHfaType(GenTree* tree);
1940 unsigned GetHfaCount(GenTree* tree);
1942 var_types GetHfaType(CORINFO_CLASS_HANDLE hClass);
1943 unsigned GetHfaCount(CORINFO_CLASS_HANDLE hClass);
1945 bool IsMultiRegReturnedType(CORINFO_CLASS_HANDLE hClass);
1947 //-------------------------------------------------------------------------
1948 // The following is used for validating format of EH table
1952 typedef struct EHNodeDsc* pEHNodeDsc;
1954 EHNodeDsc* ehnTree; // root of the tree comprising the EHnodes.
1955 EHNodeDsc* ehnNext; // root of the tree comprising the EHnodes.
1968 EHBlockType ehnBlockType; // kind of EH block
1969 IL_OFFSET ehnStartOffset; // IL offset of start of the EH block
1970 IL_OFFSET ehnEndOffset; // IL offset past end of the EH block. (TODO: looks like verInsertEhNode() sets this to
1971 // the last IL offset, not "one past the last one", i.e., the range Start to End is
1973 pEHNodeDsc ehnNext; // next (non-nested) block in sequential order
1974 pEHNodeDsc ehnChild; // leftmost nested block
1976 pEHNodeDsc ehnTryNode; // for filters and handlers, the corresponding try node
1977 pEHNodeDsc ehnHandlerNode; // for a try node, the corresponding handler node
1979 pEHNodeDsc ehnFilterNode; // if this is a try node and has a filter, otherwise 0
1980 pEHNodeDsc ehnEquivalent; // if blockType=tryNode, start offset and end offset is same,
1982 inline void ehnSetTryNodeType()
1984 ehnBlockType = TryNode;
1986 inline void ehnSetFilterNodeType()
1988 ehnBlockType = FilterNode;
1990 inline void ehnSetHandlerNodeType()
1992 ehnBlockType = HandlerNode;
1994 inline void ehnSetFinallyNodeType()
1996 ehnBlockType = FinallyNode;
1998 inline void ehnSetFaultNodeType()
2000 ehnBlockType = FaultNode;
2003 inline BOOL ehnIsTryBlock()
2005 return ehnBlockType == TryNode;
2007 inline BOOL ehnIsFilterBlock()
2009 return ehnBlockType == FilterNode;
2011 inline BOOL ehnIsHandlerBlock()
2013 return ehnBlockType == HandlerNode;
2015 inline BOOL ehnIsFinallyBlock()
2017 return ehnBlockType == FinallyNode;
2019 inline BOOL ehnIsFaultBlock()
2021 return ehnBlockType == FaultNode;
2024 // returns true if there is any overlap between the two nodes
2025 static inline BOOL ehnIsOverlap(pEHNodeDsc node1, pEHNodeDsc node2)
2027 if (node1->ehnStartOffset < node2->ehnStartOffset)
2029 return (node1->ehnEndOffset >= node2->ehnStartOffset);
2033 return (node1->ehnStartOffset <= node2->ehnEndOffset);
2037 // fails with BADCODE if inner is not completely nested inside outer
2038 static inline BOOL ehnIsNested(pEHNodeDsc inner, pEHNodeDsc outer)
2040 return ((inner->ehnStartOffset >= outer->ehnStartOffset) && (inner->ehnEndOffset <= outer->ehnEndOffset));
2044 //-------------------------------------------------------------------------
2045 // Exception handling functions
2048 #if !FEATURE_EH_FUNCLETS
2050 bool ehNeedsShadowSPslots()
2052 return (info.compXcptnsCount || opts.compDbgEnC);
2055 // 0 for methods with no EH
2056 // 1 for methods with non-nested EH, or where only the try blocks are nested
2057 // 2 for a method with a catch within a catch
2059 unsigned ehMaxHndNestingCount;
2061 #endif // !FEATURE_EH_FUNCLETS
2063 static bool jitIsBetween(unsigned value, unsigned start, unsigned end);
2064 static bool jitIsBetweenInclusive(unsigned value, unsigned start, unsigned end);
2066 bool bbInCatchHandlerILRange(BasicBlock* blk);
2067 bool bbInFilterILRange(BasicBlock* blk);
2068 bool bbInTryRegions(unsigned regionIndex, BasicBlock* blk);
2069 bool bbInExnFlowRegions(unsigned regionIndex, BasicBlock* blk);
2070 bool bbInHandlerRegions(unsigned regionIndex, BasicBlock* blk);
2071 bool bbInCatchHandlerRegions(BasicBlock* tryBlk, BasicBlock* hndBlk);
2072 unsigned short bbFindInnermostCommonTryRegion(BasicBlock* bbOne, BasicBlock* bbTwo);
2074 unsigned short bbFindInnermostTryRegionContainingHandlerRegion(unsigned handlerIndex);
2075 unsigned short bbFindInnermostHandlerRegionContainingTryRegion(unsigned tryIndex);
2077 // Returns true if "block" is the start of a try region.
2078 bool bbIsTryBeg(BasicBlock* block);
2080 // Returns true if "block" is the start of a handler or filter region.
2081 bool bbIsHandlerBeg(BasicBlock* block);
2083 // Returns true iff "block" is where control flows if an exception is raised in the
2084 // try region, and sets "*regionIndex" to the index of the try for the handler.
2085 // Differs from "IsHandlerBeg" in the case of filters, where this is true for the first
2086 // block of the filter, but not for the filter's handler.
2087 bool bbIsExFlowBlock(BasicBlock* block, unsigned* regionIndex);
2089 bool ehHasCallableHandlers();
2091 // Return the EH descriptor for the given region index.
2092 EHblkDsc* ehGetDsc(unsigned regionIndex);
2094 // Return the EH index given a region descriptor.
2095 unsigned ehGetIndex(EHblkDsc* ehDsc);
2097 // Return the EH descriptor index of the enclosing try, for the given region index.
2098 unsigned ehGetEnclosingTryIndex(unsigned regionIndex);
2100 // Return the EH descriptor index of the enclosing handler, for the given region index.
2101 unsigned ehGetEnclosingHndIndex(unsigned regionIndex);
2103 // Return the EH descriptor for the most nested 'try' region this BasicBlock is a member of (or nullptr if this
2104 // block is not in a 'try' region).
2105 EHblkDsc* ehGetBlockTryDsc(BasicBlock* block);
2107 // Return the EH descriptor for the most nested filter or handler region this BasicBlock is a member of (or nullptr
2108 // if this block is not in a filter or handler region).
2109 EHblkDsc* ehGetBlockHndDsc(BasicBlock* block);
2111 // Return the EH descriptor for the most nested region that may handle exceptions raised in this BasicBlock (or
2112 // nullptr if this block's exceptions propagate to caller).
2113 EHblkDsc* ehGetBlockExnFlowDsc(BasicBlock* block);
2115 EHblkDsc* ehIsBlockTryLast(BasicBlock* block);
2116 EHblkDsc* ehIsBlockHndLast(BasicBlock* block);
2117 bool ehIsBlockEHLast(BasicBlock* block);
2119 bool ehBlockHasExnFlowDsc(BasicBlock* block);
2121 // Return the region index of the most nested EH region this block is in.
2122 unsigned ehGetMostNestedRegionIndex(BasicBlock* block, bool* inTryRegion);
2124 // Find the true enclosing try index, ignoring 'mutual protect' try. Uses IL ranges to check.
2125 unsigned ehTrueEnclosingTryIndexIL(unsigned regionIndex);
2127 // Return the index of the most nested enclosing region for a particular EH region. Returns NO_ENCLOSING_INDEX
2128 // if there is no enclosing region. If the returned index is not NO_ENCLOSING_INDEX, then '*inTryRegion'
2129 // is set to 'true' if the enclosing region is a 'try', or 'false' if the enclosing region is a handler.
2130 // (It can never be a filter.)
2131 unsigned ehGetEnclosingRegionIndex(unsigned regionIndex, bool* inTryRegion);
2133 // A block has been deleted. Update the EH table appropriately.
2134 void ehUpdateForDeletedBlock(BasicBlock* block);
2136 // Determine whether a block can be deleted while preserving the EH normalization rules.
2137 bool ehCanDeleteEmptyBlock(BasicBlock* block);
2139 // Update the 'last' pointers in the EH table to reflect new or deleted blocks in an EH region.
2140 void ehUpdateLastBlocks(BasicBlock* oldLast, BasicBlock* newLast);
2142 // For a finally handler, find the region index that the BBJ_CALLFINALLY lives in that calls the handler,
2143 // or NO_ENCLOSING_INDEX if the BBJ_CALLFINALLY lives in the main function body. Normally, the index
2144 // is the same index as the handler (and the BBJ_CALLFINALLY lives in the 'try' region), but for AMD64 the
2145 // BBJ_CALLFINALLY lives in the enclosing try or handler region, whichever is more nested, or the main function
2146 // body. If the returned index is not NO_ENCLOSING_INDEX, then '*inTryRegion' is set to 'true' if the
2147 // BBJ_CALLFINALLY lives in the returned index's 'try' region, or 'false' if lives in the handler region. (It never
2148 // lives in a filter.)
2149 unsigned ehGetCallFinallyRegionIndex(unsigned finallyIndex, bool* inTryRegion);
2151 // Find the range of basic blocks in which all BBJ_CALLFINALLY will be found that target the 'finallyIndex' region's
2152 // handler. Set begBlk to the first block, and endBlk to the block after the last block of the range
2153 // (nullptr if the last block is the last block in the program).
2154 // Precondition: 'finallyIndex' is the EH region of a try/finally clause.
2155 void ehGetCallFinallyBlockRange(unsigned finallyIndex, BasicBlock** begBlk, BasicBlock** endBlk);
2158 // Given a BBJ_CALLFINALLY block and the EH region index of the finally it is calling, return
2159 // 'true' if the BBJ_CALLFINALLY is in the correct EH region.
2160 bool ehCallFinallyInCorrectRegion(BasicBlock* blockCallFinally, unsigned finallyIndex);
2163 #if FEATURE_EH_FUNCLETS
2164 // Do we need a PSPSym in the main function? For codegen purposes, we only need one
2165 // if there is a filter that protects a region with a nested EH clause (such as a
2166 // try/catch nested in the 'try' body of a try/filter/filter-handler). See
2167 // genFuncletProlog() for more details. However, the VM seems to use it for more
2168 // purposes, maybe including debugging. Until we are sure otherwise, always create
2169 // a PSPSym for functions with any EH.
2170 bool ehNeedsPSPSym() const
2174 #else // _TARGET_X86_
2175 return compHndBBtabCount > 0;
2176 #endif // _TARGET_X86_
2179 bool ehAnyFunclets(); // Are there any funclets in this function?
2180 unsigned ehFuncletCount(); // Return the count of funclets in the function
2182 unsigned bbThrowIndex(BasicBlock* blk); // Get the index to use as the cache key for sharing throw blocks
2183 #else // !FEATURE_EH_FUNCLETS
2184 bool ehAnyFunclets()
2188 unsigned ehFuncletCount()
2193 unsigned bbThrowIndex(BasicBlock* blk)
2195 return blk->bbTryIndex;
2196 } // Get the index to use as the cache key for sharing throw blocks
2197 #endif // !FEATURE_EH_FUNCLETS
2199 // Returns a flowList representing the "EH predecessors" of "blk". These are the normal predecessors of
2200 // "blk", plus one special case: if "blk" is the first block of a handler, considers the predecessor(s) of the first
2201 // first block of the corresponding try region to be "EH predecessors". (If there is a single such predecessor,
2202 // for example, we want to consider that the immediate dominator of the catch clause start block, so it's
2203 // convenient to also consider it a predecessor.)
2204 flowList* BlockPredsWithEH(BasicBlock* blk);
2206 // This table is useful for memoization of the method above.
2207 typedef JitHashTable<BasicBlock*, JitPtrKeyFuncs<BasicBlock>, flowList*> BlockToFlowListMap;
2208 BlockToFlowListMap* m_blockToEHPreds;
2209 BlockToFlowListMap* GetBlockToEHPreds()
2211 if (m_blockToEHPreds == nullptr)
2213 m_blockToEHPreds = new (getAllocator()) BlockToFlowListMap(getAllocator());
2215 return m_blockToEHPreds;
2218 void* ehEmitCookie(BasicBlock* block);
2219 UNATIVE_OFFSET ehCodeOffset(BasicBlock* block);
2221 EHblkDsc* ehInitHndRange(BasicBlock* src, IL_OFFSET* hndBeg, IL_OFFSET* hndEnd, bool* inFilter);
2223 EHblkDsc* ehInitTryRange(BasicBlock* src, IL_OFFSET* tryBeg, IL_OFFSET* tryEnd);
2225 EHblkDsc* ehInitHndBlockRange(BasicBlock* blk, BasicBlock** hndBeg, BasicBlock** hndLast, bool* inFilter);
2227 EHblkDsc* ehInitTryBlockRange(BasicBlock* blk, BasicBlock** tryBeg, BasicBlock** tryLast);
2229 void fgSetTryEnd(EHblkDsc* handlerTab, BasicBlock* newTryLast);
2231 void fgSetHndEnd(EHblkDsc* handlerTab, BasicBlock* newHndLast);
2233 void fgSkipRmvdBlocks(EHblkDsc* handlerTab);
2235 void fgAllocEHTable();
2237 void fgRemoveEHTableEntry(unsigned XTnum);
2239 #if FEATURE_EH_FUNCLETS
2241 EHblkDsc* fgAddEHTableEntry(unsigned XTnum);
2243 #endif // FEATURE_EH_FUNCLETS
2247 #endif // !FEATURE_EH
2249 void fgSortEHTable();
2251 // Causes the EH table to obey some well-formedness conditions, by inserting
2252 // empty BB's when necessary:
2253 // * No block is both the first block of a handler and the first block of a try.
2254 // * No block is the first block of multiple 'try' regions.
2255 // * No block is the last block of multiple EH regions.
2256 void fgNormalizeEH();
2257 bool fgNormalizeEHCase1();
2258 bool fgNormalizeEHCase2();
2259 bool fgNormalizeEHCase3();
2262 void dispIncomingEHClause(unsigned num, const CORINFO_EH_CLAUSE& clause);
2263 void dispOutgoingEHClause(unsigned num, const CORINFO_EH_CLAUSE& clause);
2264 void fgVerifyHandlerTab();
2265 void fgDispHandlerTab();
2268 bool fgNeedToSortEHTable;
2270 void verInitEHTree(unsigned numEHClauses);
2271 void verInsertEhNode(CORINFO_EH_CLAUSE* clause, EHblkDsc* handlerTab);
2272 void verInsertEhNodeInTree(EHNodeDsc** ppRoot, EHNodeDsc* node);
2273 void verInsertEhNodeParent(EHNodeDsc** ppRoot, EHNodeDsc* node);
2274 void verCheckNestingLevel(EHNodeDsc* initRoot);
2277 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
2278 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
2280 XX GenTree and BasicBlock XX
2282 XX Functions to allocate and display the GenTrees and BasicBlocks XX
2284 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
2285 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
2288 // Functions to create nodes
2289 GenTreeStmt* gtNewStmt(GenTree* expr = nullptr, IL_OFFSETX offset = BAD_IL_OFFSET);
2292 GenTree* gtNewOperNode(genTreeOps oper, var_types type, GenTree* op1, bool doSimplifications = TRUE);
2294 // For binary opers.
2295 GenTree* gtNewOperNode(genTreeOps oper, var_types type, GenTree* op1, GenTree* op2);
2297 GenTree* gtNewQmarkNode(var_types type, GenTree* cond, GenTree* colon);
2299 GenTree* gtNewLargeOperNode(genTreeOps oper,
2300 var_types type = TYP_I_IMPL,
2301 GenTree* op1 = nullptr,
2302 GenTree* op2 = nullptr);
2304 GenTreeIntCon* gtNewIconNode(ssize_t value, var_types type = TYP_INT);
2306 GenTree* gtNewPhysRegNode(regNumber reg, var_types type);
2308 GenTree* gtNewJmpTableNode();
2310 GenTree* gtNewIndOfIconHandleNode(var_types indType, size_t value, unsigned iconFlags, bool isInvariant);
2312 GenTree* gtNewIconHandleNode(size_t value, unsigned flags, FieldSeqNode* fields = nullptr);
2314 unsigned gtTokenToIconFlags(unsigned token);
2316 GenTree* gtNewIconEmbHndNode(void* value, void* pValue, unsigned flags, void* compileTimeHandle);
2318 GenTree* gtNewIconEmbScpHndNode(CORINFO_MODULE_HANDLE scpHnd);
2319 GenTree* gtNewIconEmbClsHndNode(CORINFO_CLASS_HANDLE clsHnd);
2320 GenTree* gtNewIconEmbMethHndNode(CORINFO_METHOD_HANDLE methHnd);
2321 GenTree* gtNewIconEmbFldHndNode(CORINFO_FIELD_HANDLE fldHnd);
2323 GenTree* gtNewStringLiteralNode(InfoAccessType iat, void* pValue);
2325 GenTree* gtNewLconNode(__int64 value);
2327 GenTree* gtNewDconNode(double value);
2329 GenTree* gtNewSconNode(int CPX, CORINFO_MODULE_HANDLE scpHandle);
2331 GenTree* gtNewZeroConNode(var_types type);
2333 GenTree* gtNewOneConNode(var_types type);
2336 GenTree* gtNewSIMDVectorZero(var_types simdType, var_types baseType, unsigned size);
2337 GenTree* gtNewSIMDVectorOne(var_types simdType, var_types baseType, unsigned size);
2340 GenTree* gtNewBlkOpNode(GenTree* dst, GenTree* srcOrFillVal, unsigned size, bool isVolatile, bool isCopyBlock);
2342 GenTree* gtNewPutArgReg(var_types type, GenTree* arg, regNumber argReg);
2344 GenTree* gtNewBitCastNode(var_types type, GenTree* arg);
2347 void gtBlockOpInit(GenTree* result, GenTree* dst, GenTree* srcOrFillVal, bool isVolatile);
2350 GenTree* gtNewObjNode(CORINFO_CLASS_HANDLE structHnd, GenTree* addr);
2351 void gtSetObjGcInfo(GenTreeObj* objNode);
2352 GenTree* gtNewStructVal(CORINFO_CLASS_HANDLE structHnd, GenTree* addr);
2353 GenTree* gtNewBlockVal(GenTree* addr, unsigned size);
2355 GenTree* gtNewCpObjNode(GenTree* dst, GenTree* src, CORINFO_CLASS_HANDLE structHnd, bool isVolatile);
2357 GenTreeArgList* gtNewListNode(GenTree* op1, GenTreeArgList* op2);
2359 GenTreeCall* gtNewCallNode(gtCallTypes callType,
2360 CORINFO_METHOD_HANDLE handle,
2362 GenTreeArgList* args,
2363 IL_OFFSETX ilOffset = BAD_IL_OFFSET);
2365 GenTreeCall* gtNewIndCallNode(GenTree* addr,
2367 GenTreeArgList* args,
2368 IL_OFFSETX ilOffset = BAD_IL_OFFSET);
2370 GenTreeCall* gtNewHelperCallNode(unsigned helper, var_types type, GenTreeArgList* args = nullptr);
2372 GenTree* gtNewLclvNode(unsigned lnum, var_types type, IL_OFFSETX ILoffs = BAD_IL_OFFSET);
2375 GenTreeSIMD* gtNewSIMDNode(
2376 var_types type, GenTree* op1, SIMDIntrinsicID simdIntrinsicID, var_types baseType, unsigned size);
2377 GenTreeSIMD* gtNewSIMDNode(
2378 var_types type, GenTree* op1, GenTree* op2, SIMDIntrinsicID simdIntrinsicID, var_types baseType, unsigned size);
2379 void SetOpLclRelatedToSIMDIntrinsic(GenTree* op);
2382 #ifdef FEATURE_HW_INTRINSICS
2383 GenTreeHWIntrinsic* gtNewSimdHWIntrinsicNode(var_types type,
2384 NamedIntrinsic hwIntrinsicID,
2387 GenTreeHWIntrinsic* gtNewSimdHWIntrinsicNode(
2388 var_types type, GenTree* op1, NamedIntrinsic hwIntrinsicID, var_types baseType, unsigned size);
2389 GenTreeHWIntrinsic* gtNewSimdHWIntrinsicNode(
2390 var_types type, GenTree* op1, GenTree* op2, NamedIntrinsic hwIntrinsicID, var_types baseType, unsigned size);
2391 GenTreeHWIntrinsic* gtNewSimdHWIntrinsicNode(var_types type,
2395 NamedIntrinsic hwIntrinsicID,
2398 GenTreeHWIntrinsic* gtNewSimdHWIntrinsicNode(var_types type,
2403 NamedIntrinsic hwIntrinsicID,
2406 GenTreeHWIntrinsic* gtNewScalarHWIntrinsicNode(var_types type, GenTree* op1, NamedIntrinsic hwIntrinsicID);
2407 GenTreeHWIntrinsic* gtNewScalarHWIntrinsicNode(var_types type,
2410 NamedIntrinsic hwIntrinsicID);
2411 GenTreeHWIntrinsic* gtNewScalarHWIntrinsicNode(
2412 var_types type, GenTree* op1, GenTree* op2, GenTree* op3, NamedIntrinsic hwIntrinsicID);
2413 GenTree* gtNewMustThrowException(unsigned helper, var_types type, CORINFO_CLASS_HANDLE clsHnd);
2414 CORINFO_CLASS_HANDLE gtGetStructHandleForHWSIMD(var_types simdType, var_types simdBaseType);
2415 #endif // FEATURE_HW_INTRINSICS
2417 GenTree* gtNewLclLNode(unsigned lnum, var_types type, IL_OFFSETX ILoffs = BAD_IL_OFFSET);
2418 GenTreeLclFld* gtNewLclFldNode(unsigned lnum, var_types type, unsigned offset);
2419 GenTree* gtNewInlineCandidateReturnExpr(GenTree* inlineCandidate, var_types type);
2421 GenTree* gtNewCodeRef(BasicBlock* block);
2423 GenTree* gtNewFieldRef(var_types typ, CORINFO_FIELD_HANDLE fldHnd, GenTree* obj = nullptr, DWORD offset = 0);
2425 GenTree* gtNewIndexRef(var_types typ, GenTree* arrayOp, GenTree* indexOp);
2427 GenTreeArrLen* gtNewArrLen(var_types typ, GenTree* arrayOp, int lenOffset);
2429 GenTree* gtNewIndir(var_types typ, GenTree* addr);
2431 GenTreeArgList* gtNewArgList(GenTree* op);
2432 GenTreeArgList* gtNewArgList(GenTree* op1, GenTree* op2);
2433 GenTreeArgList* gtNewArgList(GenTree* op1, GenTree* op2, GenTree* op3);
2434 GenTreeArgList* gtNewArgList(GenTree* op1, GenTree* op2, GenTree* op3, GenTree* op4);
2436 static fgArgTabEntry* gtArgEntryByArgNum(GenTreeCall* call, unsigned argNum);
2437 static fgArgTabEntry* gtArgEntryByNode(GenTreeCall* call, GenTree* node);
2438 fgArgTabEntry* gtArgEntryByLateArgIndex(GenTreeCall* call, unsigned lateArgInx);
2439 static GenTree* gtArgNodeByLateArgInx(GenTreeCall* call, unsigned lateArgInx);
2440 bool gtArgIsThisPtr(fgArgTabEntry* argEntry);
2442 GenTree* gtNewAssignNode(GenTree* dst, GenTree* src);
2444 GenTree* gtNewTempAssign(unsigned tmp,
2446 GenTree** pAfterStmt = nullptr,
2447 IL_OFFSETX ilOffset = BAD_IL_OFFSET,
2448 BasicBlock* block = nullptr);
2450 GenTree* gtNewRefCOMfield(GenTree* objPtr,
2451 CORINFO_RESOLVED_TOKEN* pResolvedToken,
2452 CORINFO_ACCESS_FLAGS access,
2453 CORINFO_FIELD_INFO* pFieldInfo,
2455 CORINFO_CLASS_HANDLE structType,
2458 GenTree* gtNewNothingNode();
2460 GenTree* gtNewArgPlaceHolderNode(var_types type, CORINFO_CLASS_HANDLE clsHnd);
2462 GenTree* gtUnusedValNode(GenTree* expr);
2464 GenTreeCast* gtNewCastNode(var_types typ, GenTree* op1, bool fromUnsigned, var_types castType);
2466 GenTreeCast* gtNewCastNodeL(var_types typ, GenTree* op1, bool fromUnsigned, var_types castType);
2468 GenTreeAllocObj* gtNewAllocObjNode(
2469 unsigned int helper, bool helperHasSideEffects, CORINFO_CLASS_HANDLE clsHnd, var_types type, GenTree* op1);
2471 GenTreeAllocObj* gtNewAllocObjNode(CORINFO_RESOLVED_TOKEN* pResolvedToken, BOOL useParent);
2473 GenTree* gtNewRuntimeLookup(CORINFO_GENERIC_HANDLE hnd, CorInfoGenericHandleType hndTyp, GenTree* lookupTree);
2475 //------------------------------------------------------------------------
2476 // Other GenTree functions
2478 GenTree* gtClone(GenTree* tree, bool complexOK = false);
2480 // If `tree` is a lclVar with lclNum `varNum`, return an IntCns with value `varVal`; otherwise,
2481 // create a copy of `tree`, adding specified flags, replacing uses of lclVar `deepVarNum` with
2482 // IntCnses with value `deepVarVal`.
2483 GenTree* gtCloneExpr(
2484 GenTree* tree, unsigned addFlags, unsigned varNum, int varVal, unsigned deepVarNum, int deepVarVal);
2486 // Create a copy of `tree`, optionally adding specifed flags, and optionally mapping uses of local
2487 // `varNum` to int constants with value `varVal`.
2488 GenTree* gtCloneExpr(GenTree* tree, unsigned addFlags = 0, unsigned varNum = BAD_VAR_NUM, int varVal = 0)
2490 return gtCloneExpr(tree, addFlags, varNum, varVal, varNum, varVal);
2493 // Internal helper for cloning a call
2494 GenTreeCall* gtCloneExprCallHelper(GenTreeCall* call,
2495 unsigned addFlags = 0,
2496 unsigned deepVarNum = BAD_VAR_NUM,
2497 int deepVarVal = 0);
2499 // Create copy of an inline or guarded devirtualization candidate tree.
2500 GenTreeCall* gtCloneCandidateCall(GenTreeCall* call);
2502 GenTree* gtReplaceTree(GenTree* stmt, GenTree* tree, GenTree* replacementTree);
2504 void gtUpdateSideEffects(GenTree* stmt, GenTree* tree);
2506 void gtUpdateTreeAncestorsSideEffects(GenTree* tree);
2508 void gtUpdateStmtSideEffects(GenTree* stmt);
2510 void gtUpdateNodeSideEffects(GenTree* tree);
2512 void gtUpdateNodeOperSideEffects(GenTree* tree);
2514 // Returns "true" iff the complexity (not formally defined, but first interpretation
2515 // is #of nodes in subtree) of "tree" is greater than "limit".
2516 // (This is somewhat redundant with the "gtCostEx/gtCostSz" fields, but can be used
2517 // before they have been set.)
2518 bool gtComplexityExceeds(GenTree** tree, unsigned limit);
2520 bool gtCompareTree(GenTree* op1, GenTree* op2);
2522 GenTree* gtReverseCond(GenTree* tree);
2524 bool gtHasRef(GenTree* tree, ssize_t lclNum, bool defOnly);
2526 bool gtHasLocalsWithAddrOp(GenTree* tree);
2528 unsigned gtSetListOrder(GenTree* list, bool regs, bool isListCallArgs);
2530 void gtWalkOp(GenTree** op1, GenTree** op2, GenTree* base, bool constOnly);
2533 unsigned gtHashValue(GenTree* tree);
2535 GenTree* gtWalkOpEffectiveVal(GenTree* op);
2538 void gtPrepareCost(GenTree* tree);
2539 bool gtIsLikelyRegVar(GenTree* tree);
2541 // Returns true iff the secondNode can be swapped with firstNode.
2542 bool gtCanSwapOrder(GenTree* firstNode, GenTree* secondNode);
2544 unsigned gtSetEvalOrder(GenTree* tree);
2546 void gtSetStmtInfo(GenTree* stmt);
2548 // Returns "true" iff "node" has any of the side effects in "flags".
2549 bool gtNodeHasSideEffects(GenTree* node, unsigned flags);
2551 // Returns "true" iff "tree" or its (transitive) children have any of the side effects in "flags".
2552 bool gtTreeHasSideEffects(GenTree* tree, unsigned flags);
2554 // Appends 'expr' in front of 'list'
2555 // 'list' will typically start off as 'nullptr'
2556 // when 'list' is non-null a GT_COMMA node is used to insert 'expr'
2557 GenTree* gtBuildCommaList(GenTree* list, GenTree* expr);
2559 void gtExtractSideEffList(GenTree* expr,
2561 unsigned flags = GTF_SIDE_EFFECT,
2562 bool ignoreRoot = false);
2564 GenTree* gtGetThisArg(GenTreeCall* call);
2566 // Static fields of struct types (and sometimes the types that those are reduced to) are represented by having the
2567 // static field contain an object pointer to the boxed struct. This simplifies the GC implementation...but
2568 // complicates the JIT somewhat. This predicate returns "true" iff a node with type "fieldNodeType", representing
2569 // the given "fldHnd", is such an object pointer.
2570 bool gtIsStaticFieldPtrToBoxedStruct(var_types fieldNodeType, CORINFO_FIELD_HANDLE fldHnd);
2572 // Return true if call is a recursive call; return false otherwise.
2573 // Note when inlining, this looks for calls back to the root method.
2574 bool gtIsRecursiveCall(GenTreeCall* call)
2576 return gtIsRecursiveCall(call->gtCallMethHnd);
2579 bool gtIsRecursiveCall(CORINFO_METHOD_HANDLE callMethodHandle)
2581 return (callMethodHandle == impInlineRoot()->info.compMethodHnd);
2584 //-------------------------------------------------------------------------
2586 GenTree* gtFoldExpr(GenTree* tree);
2589 // TODO-Amd64-Unix: Remove this when the clang optimizer is fixed and/or the method implementation is
2590 // refactored in a simpler code. This is a workaround for a bug in the clang-3.5 optimizer. The issue is that in
2591 // release build the optimizer is mistyping (or just wrongly decides to use 32 bit operation for a corner case
2592 // of MIN_LONG) the args of the (ltemp / lval2) to int (it does a 32 bit div operation instead of 64 bit) - see
2593 // the implementation of the method in gentree.cpp. For the case of lval1 and lval2 equal to MIN_LONG
2594 // (0x8000000000000000) this results in raising a SIGFPE. The method implementation is rather complex. Disable
2595 // optimizations for now.
2596 __attribute__((optnone))
2598 gtFoldExprConst(GenTree* tree);
2599 GenTree* gtFoldExprSpecial(GenTree* tree);
2600 GenTree* gtFoldExprCompare(GenTree* tree);
2601 GenTree* gtCreateHandleCompare(genTreeOps oper,
2604 CorInfoInlineTypeCheck typeCheckInliningResult);
2605 GenTree* gtFoldExprCall(GenTreeCall* call);
2606 GenTree* gtFoldTypeCompare(GenTree* tree);
2607 GenTree* gtFoldTypeEqualityCall(CorInfoIntrinsics methodID, GenTree* op1, GenTree* op2);
2609 // Options to control behavior of gtTryRemoveBoxUpstreamEffects
2610 enum BoxRemovalOptions
2612 BR_REMOVE_AND_NARROW, // remove effects, minimize remaining work, return possibly narrowed source tree
2613 BR_REMOVE_AND_NARROW_WANT_TYPE_HANDLE, // remove effects and minimize remaining work, return type handle tree
2614 BR_REMOVE_BUT_NOT_NARROW, // remove effects, return original source tree
2615 BR_DONT_REMOVE, // check if removal is possible, return copy source tree
2616 BR_DONT_REMOVE_WANT_TYPE_HANDLE, // check if removal is possible, return type handle tree
2617 BR_MAKE_LOCAL_COPY // revise box to copy to temp local and return local's address
2620 GenTree* gtTryRemoveBoxUpstreamEffects(GenTree* tree, BoxRemovalOptions options = BR_REMOVE_AND_NARROW);
2621 GenTree* gtOptimizeEnumHasFlag(GenTree* thisOp, GenTree* flagOp);
2623 //-------------------------------------------------------------------------
2624 // Get the handle, if any.
2625 CORINFO_CLASS_HANDLE gtGetStructHandleIfPresent(GenTree* tree);
2626 // Get the handle, and assert if not found.
2627 CORINFO_CLASS_HANDLE gtGetStructHandle(GenTree* tree);
2628 // Get the handle for a ref type.
2629 CORINFO_CLASS_HANDLE gtGetClassHandle(GenTree* tree, bool* pIsExact, bool* pIsNonNull);
2630 // Get the class handle for an helper call
2631 CORINFO_CLASS_HANDLE gtGetHelperCallClassHandle(GenTreeCall* call, bool* pIsExact, bool* pIsNonNull);
2632 // Get the element handle for an array of ref type.
2633 CORINFO_CLASS_HANDLE gtGetArrayElementClassHandle(GenTree* array);
2634 // Get a class handle from a helper call argument
2635 CORINFO_CLASS_HANDLE gtGetHelperArgClassHandle(GenTree* array,
2636 unsigned* runtimeLookupCount = nullptr,
2637 GenTree** handleTree = nullptr);
2638 // Get the class handle for a field
2639 CORINFO_CLASS_HANDLE gtGetFieldClassHandle(CORINFO_FIELD_HANDLE fieldHnd, bool* pIsExact, bool* pIsNonNull);
2640 // Check if this tree is a gc static base helper call
2641 bool gtIsStaticGCBaseHelperCall(GenTree* tree);
2643 //-------------------------------------------------------------------------
2644 // Functions to display the trees
2647 void gtDispNode(GenTree* tree, IndentStack* indentStack, __in_z const char* msg, bool isLIR);
2649 void gtDispVN(GenTree* tree);
2650 void gtDispConst(GenTree* tree);
2651 void gtDispLeaf(GenTree* tree, IndentStack* indentStack);
2652 void gtDispNodeName(GenTree* tree);
2653 void gtDispRegVal(GenTree* tree);
2665 void gtDispChild(GenTree* child,
2666 IndentStack* indentStack,
2668 __in_opt const char* msg = nullptr,
2669 bool topOnly = false);
2670 void gtDispTree(GenTree* tree,
2671 IndentStack* indentStack = nullptr,
2672 __in_opt const char* msg = nullptr,
2673 bool topOnly = false,
2674 bool isLIR = false);
2675 void gtGetLclVarNameInfo(unsigned lclNum, const char** ilKindOut, const char** ilNameOut, unsigned* ilNumOut);
2676 int gtGetLclVarName(unsigned lclNum, char* buf, unsigned buf_remaining);
2677 char* gtGetLclVarName(unsigned lclNum);
2678 void gtDispLclVar(unsigned varNum, bool padForBiggestDisp = true);
2679 void gtDispTreeList(GenTree* tree, IndentStack* indentStack = nullptr);
2680 void gtGetArgMsg(GenTreeCall* call, GenTree* arg, unsigned argNum, int listCount, char* bufp, unsigned bufLength);
2681 void gtGetLateArgMsg(GenTreeCall* call, GenTree* arg, int argNum, int listCount, char* bufp, unsigned bufLength);
2682 void gtDispArgList(GenTreeCall* call, IndentStack* indentStack);
2683 void gtDispFieldSeq(FieldSeqNode* pfsn);
2685 void gtDispRange(LIR::ReadOnlyRange const& range);
2687 void gtDispTreeRange(LIR::Range& containingRange, GenTree* tree);
2689 void gtDispLIRNode(GenTree* node, const char* prefixMsg = nullptr);
2701 typedef fgWalkResult(fgWalkPreFn)(GenTree** pTree, fgWalkData* data);
2702 typedef fgWalkResult(fgWalkPostFn)(GenTree** pTree, fgWalkData* data);
2705 static fgWalkPreFn gtAssertColonCond;
2707 static fgWalkPreFn gtMarkColonCond;
2708 static fgWalkPreFn gtClearColonCond;
2710 GenTree** gtFindLink(GenTree* stmt, GenTree* node);
2711 bool gtHasCatchArg(GenTree* tree);
2713 typedef ArrayStack<GenTree*> GenTreeStack;
2715 static bool gtHasCallOnStack(GenTreeStack* parentStack);
2717 //=========================================================================
2718 // BasicBlock functions
2720 // This is a debug flag we will use to assert when creating block during codegen
2721 // as this interferes with procedure splitting. If you know what you're doing, set
2722 // it to true before creating the block. (DEBUG only)
2723 bool fgSafeBasicBlockCreation;
2726 BasicBlock* bbNewBasicBlock(BBjumpKinds jumpKind);
2729 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
2730 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
2734 XX The variables to be used by the code generator. XX
2736 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
2737 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
2741 // For both PROMOTION_TYPE_NONE and PROMOTION_TYPE_DEPENDENT the struct will
2742 // be placed in the stack frame and it's fields must be laid out sequentially.
2744 // For PROMOTION_TYPE_INDEPENDENT each of the struct's fields is replaced by
2745 // a local variable that can be enregistered or placed in the stack frame.
2746 // The fields do not need to be laid out sequentially
2748 enum lvaPromotionType
2750 PROMOTION_TYPE_NONE, // The struct local is not promoted
2751 PROMOTION_TYPE_INDEPENDENT, // The struct local is promoted,
2752 // and its field locals are independent of its parent struct local.
2753 PROMOTION_TYPE_DEPENDENT // The struct local is promoted,
2754 // but its field locals depend on its parent struct local.
2757 static int __cdecl RefCntCmp(const void* op1, const void* op2);
2758 static int __cdecl WtdRefCntCmp(const void* op1, const void* op2);
2760 /*****************************************************************************/
2762 enum FrameLayoutState
2765 INITIAL_FRAME_LAYOUT,
2766 PRE_REGALLOC_FRAME_LAYOUT,
2767 REGALLOC_FRAME_LAYOUT,
2768 TENTATIVE_FRAME_LAYOUT,
2773 RefCountState lvaRefCountState; // Current local ref count state
2775 bool lvaLocalVarRefCounted() const
2777 return lvaRefCountState == RCS_NORMAL;
2780 bool lvaTrackedFixed; // true: We cannot add new 'tracked' variable
2781 unsigned lvaCount; // total number of locals
2783 unsigned lvaRefCount; // total number of references to locals
2784 LclVarDsc* lvaTable; // variable descriptor table
2785 unsigned lvaTableCnt; // lvaTable size (>= lvaCount)
2787 LclVarDsc** lvaRefSorted; // table sorted by refcount
2789 unsigned short lvaTrackedCount; // actual # of locals being tracked
2790 unsigned lvaTrackedCountInSizeTUnits; // min # of size_t's sufficient to hold a bit for all the locals being tracked
2793 VARSET_TP lvaTrackedVars; // set of tracked variables
2795 #ifndef _TARGET_64BIT_
2796 VARSET_TP lvaLongVars; // set of long (64-bit) variables
2798 VARSET_TP lvaFloatVars; // set of floating-point (32-bit and 64-bit) variables
2800 unsigned lvaCurEpoch; // VarSets are relative to a specific set of tracked var indices.
2801 // It that changes, this changes. VarSets from different epochs
2802 // cannot be meaningfully combined.
2804 unsigned GetCurLVEpoch()
2809 // reverse map of tracked number to var number
2810 unsigned* lvaTrackedToVarNum;
2814 // # of procs compiled a with double-aligned stack
2815 static unsigned s_lvaDoubleAlignedProcsCount;
2819 // Getters and setters for address-exposed and do-not-enregister local var properties.
2820 bool lvaVarAddrExposed(unsigned varNum);
2821 void lvaSetVarAddrExposed(unsigned varNum);
2822 bool lvaVarDoNotEnregister(unsigned varNum);
2824 // Reasons why we can't enregister. Some of these correspond to debug properties of local vars.
2825 enum DoNotEnregisterReason
2830 DNER_VMNeedsStackAddr,
2831 DNER_LiveInOutOfHandler,
2832 DNER_LiveAcrossUnmanagedCall,
2833 DNER_BlockOp, // Is read or written via a block operation that explicitly takes the address.
2834 DNER_IsStructArg, // Is a struct passed as an argument in a way that requires a stack location.
2835 DNER_DepField, // It is a field of a dependently promoted struct
2836 DNER_NoRegVars, // opts.compFlags & CLFLG_REGVAR is not set
2837 DNER_MinOptsGC, // It is a GC Ref and we are compiling MinOpts
2838 #if !defined(_TARGET_64BIT_)
2839 DNER_LongParamField, // It is a decomposed field of a long parameter.
2841 #ifdef JIT32_GCENCODER
2846 void lvaSetVarDoNotEnregister(unsigned varNum DEBUGARG(DoNotEnregisterReason reason));
2848 unsigned lvaVarargsHandleArg;
2850 unsigned lvaVarargsBaseOfStkArgs; // Pointer (computed based on incoming varargs handle) to the start of the stack
2852 #endif // _TARGET_X86_
2854 unsigned lvaInlinedPInvokeFrameVar; // variable representing the InlinedCallFrame
2855 unsigned lvaReversePInvokeFrameVar; // variable representing the reverse PInvoke frame
2856 #if FEATURE_FIXED_OUT_ARGS
2857 unsigned lvaPInvokeFrameRegSaveVar; // variable representing the RegSave for PInvoke inlining.
2859 unsigned lvaMonAcquired; // boolean variable introduced into in synchronized methods
2860 // that tracks whether the lock has been taken
2862 unsigned lvaArg0Var; // The lclNum of arg0. Normally this will be info.compThisArg.
2863 // However, if there is a "ldarga 0" or "starg 0" in the IL,
2864 // we will redirect all "ldarg(a) 0" and "starg 0" to this temp.
2866 unsigned lvaInlineeReturnSpillTemp; // The temp to spill the non-VOID return expression
2867 // in case there are multiple BBJ_RETURN blocks in the inlinee
2868 // or if the inlinee has GC ref locals.
2870 #if FEATURE_FIXED_OUT_ARGS
2871 unsigned lvaOutgoingArgSpaceVar; // dummy TYP_LCLBLK var for fixed outgoing argument space
2872 PhasedVar<unsigned> lvaOutgoingArgSpaceSize; // size of fixed outgoing argument space
2873 #endif // FEATURE_FIXED_OUT_ARGS
2876 // On architectures whose ABIs allow structs to be passed in registers, struct promotion will sometimes
2877 // require us to "rematerialize" a struct from it's separate constituent field variables. Packing several sub-word
2878 // field variables into an argument register is a hard problem. It's easier to reserve a word of memory into which
2879 // such field can be copied, after which the assembled memory word can be read into the register. We will allocate
2880 // this variable to be this scratch word whenever struct promotion occurs.
2881 unsigned lvaPromotedStructAssemblyScratchVar;
2882 #endif // _TARGET_ARM_
2884 #if defined(DEBUG) && defined(_TARGET_XARCH_)
2886 unsigned lvaReturnSpCheck; // Stores SP to confirm it is not corrupted on return.
2888 #endif // defined(DEBUG) && defined(_TARGET_XARCH_)
2890 #if defined(DEBUG) && defined(_TARGET_X86_)
2892 unsigned lvaCallSpCheck; // Stores SP to confirm it is not corrupted after every call.
2894 #endif // defined(DEBUG) && defined(_TARGET_X86_)
2896 unsigned lvaGenericsContextUseCount;
2898 bool lvaKeepAliveAndReportThis(); // Synchronized instance method of a reference type, or
2899 // CORINFO_GENERICS_CTXT_FROM_THIS?
2900 bool lvaReportParamTypeArg(); // Exceptions and CORINFO_GENERICS_CTXT_FROM_PARAMTYPEARG?
2902 //-------------------------------------------------------------------------
2903 // All these frame offsets are inter-related and must be kept in sync
2905 #if !FEATURE_EH_FUNCLETS
2906 // This is used for the callable handlers
2907 unsigned lvaShadowSPslotsVar; // TYP_BLK variable for all the shadow SP slots
2908 #endif // FEATURE_EH_FUNCLETS
2910 int lvaCachedGenericContextArgOffs;
2911 int lvaCachedGenericContextArgOffset(); // For CORINFO_CALLCONV_PARAMTYPE and if generic context is passed as
2914 #ifdef JIT32_GCENCODER
2916 unsigned lvaLocAllocSPvar; // variable which stores the value of ESP after the the last alloca/localloc
2918 #endif // JIT32_GCENCODER
2920 unsigned lvaNewObjArrayArgs; // variable with arguments for new MD array helper
2922 // TODO-Review: Prior to reg predict we reserve 24 bytes for Spill temps.
2923 // after the reg predict we will use a computed maxTmpSize
2924 // which is based upon the number of spill temps predicted by reg predict
2925 // All this is necessary because if we under-estimate the size of the spill
2926 // temps we could fail when encoding instructions that reference stack offsets for ARM.
2928 // Pre codegen max spill temp size.
2929 static const unsigned MAX_SPILL_TEMP_SIZE = 24;
2931 //-------------------------------------------------------------------------
2933 unsigned lvaGetMaxSpillTempSize();
2935 bool lvaIsPreSpilled(unsigned lclNum, regMaskTP preSpillMask);
2936 #endif // _TARGET_ARM_
2937 void lvaAssignFrameOffsets(FrameLayoutState curState);
2938 void lvaFixVirtualFrameOffsets();
2939 void lvaUpdateArgsWithInitialReg();
2940 void lvaAssignVirtualFrameOffsetsToArgs();
2941 #ifdef UNIX_AMD64_ABI
2942 int lvaAssignVirtualFrameOffsetToArg(unsigned lclNum, unsigned argSize, int argOffs, int* callerArgOffset);
2943 #else // !UNIX_AMD64_ABI
2944 int lvaAssignVirtualFrameOffsetToArg(unsigned lclNum, unsigned argSize, int argOffs);
2945 #endif // !UNIX_AMD64_ABI
2946 void lvaAssignVirtualFrameOffsetsToLocals();
2947 int lvaAllocLocalAndSetVirtualOffset(unsigned lclNum, unsigned size, int stkOffs);
2948 #ifdef _TARGET_AMD64_
2949 // Returns true if compCalleeRegsPushed (including RBP if used as frame pointer) is even.
2950 bool lvaIsCalleeSavedIntRegCountEven();
2952 void lvaAlignFrame();
2953 void lvaAssignFrameOffsetsToPromotedStructs();
2954 int lvaAllocateTemps(int stkOffs, bool mustDoubleAlign);
2957 void lvaDumpRegLocation(unsigned lclNum);
2958 void lvaDumpFrameLocation(unsigned lclNum);
2959 void lvaDumpEntry(unsigned lclNum, FrameLayoutState curState, size_t refCntWtdWidth = 6);
2960 void lvaTableDump(FrameLayoutState curState = NO_FRAME_LAYOUT); // NO_FRAME_LAYOUT means use the current frame
2961 // layout state defined by lvaDoneFrameLayout
2964 // Limit frames size to 1GB. The maximum is 2GB in theory - make it intentionally smaller
2965 // to avoid bugs from borderline cases.
2966 #define MAX_FrameSize 0x3FFFFFFF
2967 void lvaIncrementFrameSize(unsigned size);
2969 unsigned lvaFrameSize(FrameLayoutState curState);
2971 // Returns the caller-SP-relative offset for the SP/FP relative offset determined by FP based.
2972 int lvaToCallerSPRelativeOffset(int offs, bool isFpBased);
2974 // Returns the caller-SP-relative offset for the local variable "varNum."
2975 int lvaGetCallerSPRelativeOffset(unsigned varNum);
2977 // Returns the SP-relative offset for the local variable "varNum". Illegal to ask this for functions with localloc.
2978 int lvaGetSPRelativeOffset(unsigned varNum);
2980 int lvaToInitialSPRelativeOffset(unsigned offset, bool isFpBased);
2981 int lvaGetInitialSPRelativeOffset(unsigned varNum);
2983 //------------------------ For splitting types ----------------------------
2985 void lvaInitTypeRef();
2987 void lvaInitArgs(InitVarDscInfo* varDscInfo);
2988 void lvaInitThisPtr(InitVarDscInfo* varDscInfo);
2989 void lvaInitRetBuffArg(InitVarDscInfo* varDscInfo);
2990 void lvaInitUserArgs(InitVarDscInfo* varDscInfo);
2991 void lvaInitGenericsCtxt(InitVarDscInfo* varDscInfo);
2992 void lvaInitVarArgsHandle(InitVarDscInfo* varDscInfo);
2994 void lvaInitVarDsc(LclVarDsc* varDsc,
2996 CorInfoType corInfoType,
2997 CORINFO_CLASS_HANDLE typeHnd,
2998 CORINFO_ARG_LIST_HANDLE varList,
2999 CORINFO_SIG_INFO* varSig);
3001 static unsigned lvaTypeRefMask(var_types type);
3003 var_types lvaGetActualType(unsigned lclNum);
3004 var_types lvaGetRealType(unsigned lclNum);
3006 //-------------------------------------------------------------------------
3010 LclVarDsc* lvaGetDesc(unsigned lclNum)
3012 assert(lclNum < lvaCount);
3013 return &lvaTable[lclNum];
3016 LclVarDsc* lvaGetDesc(GenTreeLclVarCommon* lclVar)
3018 assert(lclVar->GetLclNum() < lvaCount);
3019 return &lvaTable[lclVar->GetLclNum()];
3022 unsigned lvaLclSize(unsigned varNum);
3023 unsigned lvaLclExactSize(unsigned varNum);
3025 bool lvaHaveManyLocals() const;
3027 unsigned lvaGrabTemp(bool shortLifetime DEBUGARG(const char* reason));
3028 unsigned lvaGrabTemps(unsigned cnt DEBUGARG(const char* reason));
3029 unsigned lvaGrabTempWithImplicitUse(bool shortLifetime DEBUGARG(const char* reason));
3032 void lvaSortByRefCount();
3033 void lvaDumpRefCounts();
3035 void lvaMarkLocalVars(); // Local variable ref-counting
3036 void lvaComputeRefCounts(bool isRecompute, bool setSlotNumbers);
3037 void lvaMarkLocalVars(BasicBlock* block, bool isRecompute);
3039 void lvaAllocOutgoingArgSpaceVar(); // Set up lvaOutgoingArgSpaceVar
3041 VARSET_VALRET_TP lvaStmtLclMask(GenTree* stmt);
3044 struct lvaStressLclFldArgs
3046 Compiler* m_pCompiler;
3050 static fgWalkPreFn lvaStressLclFldCB;
3051 void lvaStressLclFld();
3053 void lvaDispVarSet(VARSET_VALARG_TP set, VARSET_VALARG_TP allVars);
3054 void lvaDispVarSet(VARSET_VALARG_TP set);
3059 int lvaFrameAddress(int varNum, bool mustBeFPBased, regNumber* pBaseReg, int addrModeOffset, bool isFloatUsage);
3061 int lvaFrameAddress(int varNum, bool* pFPbased);
3064 bool lvaIsParameter(unsigned varNum);
3065 bool lvaIsRegArgument(unsigned varNum);
3066 BOOL lvaIsOriginalThisArg(unsigned varNum); // Is this varNum the original this argument?
3067 BOOL lvaIsOriginalThisReadOnly(); // return TRUE if there is no place in the code
3068 // that writes to arg0
3070 // Struct parameters that are passed by reference are marked as both lvIsParam and lvIsTemp
3071 // (this is an overload of lvIsTemp because there are no temp parameters).
3072 // For x64 this is 3, 5, 6, 7, >8 byte structs that are passed by reference.
3073 // For ARM64, this is structs larger than 16 bytes that are passed by reference.
3074 bool lvaIsImplicitByRefLocal(unsigned varNum)
3076 #if defined(_TARGET_AMD64_) || defined(_TARGET_ARM64_)
3077 LclVarDsc* varDsc = &(lvaTable[varNum]);
3078 if (varDsc->lvIsParam && varDsc->lvIsTemp)
3080 assert(varTypeIsStruct(varDsc) || (varDsc->lvType == TYP_BYREF));
3083 #endif // defined(_TARGET_AMD64_) || defined(_TARGET_ARM64_)
3087 // Returns true if this local var is a multireg struct
3088 bool lvaIsMultiregStruct(LclVarDsc* varDsc, bool isVararg);
3090 // If the local is a TYP_STRUCT, get/set a class handle describing it
3091 CORINFO_CLASS_HANDLE lvaGetStruct(unsigned varNum);
3092 void lvaSetStruct(unsigned varNum, CORINFO_CLASS_HANDLE typeHnd, bool unsafeValueClsCheck, bool setTypeInfo = true);
3093 void lvaSetStructUsedAsVarArg(unsigned varNum);
3095 // If the local is TYP_REF, set or update the associated class information.
3096 void lvaSetClass(unsigned varNum, CORINFO_CLASS_HANDLE clsHnd, bool isExact = false);
3097 void lvaSetClass(unsigned varNum, GenTree* tree, CORINFO_CLASS_HANDLE stackHandle = nullptr);
3098 void lvaUpdateClass(unsigned varNum, CORINFO_CLASS_HANDLE clsHnd, bool isExact = false);
3099 void lvaUpdateClass(unsigned varNum, GenTree* tree, CORINFO_CLASS_HANDLE stackHandle = nullptr);
3101 #define MAX_NumOfFieldsInPromotableStruct 4 // Maximum number of fields in promotable struct
3103 // Info about struct type fields.
3104 struct lvaStructFieldInfo
3106 CORINFO_FIELD_HANDLE fldHnd;
3107 unsigned char fldOffset;
3108 unsigned char fldOrdinal;
3111 CORINFO_CLASS_HANDLE fldTypeHnd;
3113 lvaStructFieldInfo()
3114 : fldHnd(nullptr), fldOffset(0), fldOrdinal(0), fldType(TYP_UNDEF), fldSize(0), fldTypeHnd(nullptr)
3119 // Info about a struct type, instances of which may be candidates for promotion.
3120 struct lvaStructPromotionInfo
3122 CORINFO_CLASS_HANDLE typeHnd;
3127 unsigned char fieldCnt;
3128 lvaStructFieldInfo fields[MAX_NumOfFieldsInPromotableStruct];
3130 lvaStructPromotionInfo(CORINFO_CLASS_HANDLE typeHnd = nullptr)
3133 , containsHoles(false)
3134 , customLayout(false)
3135 , fieldsSorted(false)
3141 static int __cdecl lvaFieldOffsetCmp(const void* field1, const void* field2);
3143 // This class is responsible for checking validity and profitability of struct promotion.
3144 // If it is both legal and profitable, then TryPromoteStructVar promotes the struct and initializes
3145 // nessesary information for fgMorphStructField to use.
3146 class StructPromotionHelper
3149 StructPromotionHelper(Compiler* compiler);
3151 bool CanPromoteStructType(CORINFO_CLASS_HANDLE typeHnd);
3152 bool TryPromoteStructVar(unsigned lclNum);
3155 void CheckRetypedAsScalar(CORINFO_FIELD_HANDLE fieldHnd, var_types requestedType);
3159 bool GetRequiresScratchVar();
3160 #endif // _TARGET_ARM_
3163 bool CanPromoteStructVar(unsigned lclNum);
3164 bool ShouldPromoteStructVar(unsigned lclNum);
3165 void PromoteStructVar(unsigned lclNum);
3166 void SortStructFields();
3168 lvaStructFieldInfo GetFieldInfo(CORINFO_FIELD_HANDLE fieldHnd, BYTE ordinal);
3169 bool TryPromoteStructField(lvaStructFieldInfo& outerFieldInfo);
3173 lvaStructPromotionInfo structPromotionInfo;
3176 bool requiresScratchVar;
3177 #endif // _TARGET_ARM_
3180 typedef JitHashTable<CORINFO_FIELD_HANDLE, JitPtrKeyFuncs<CORINFO_FIELD_STRUCT_>, var_types>
3181 RetypedAsScalarFieldsMap;
3182 RetypedAsScalarFieldsMap retypedFieldsMap;
3186 StructPromotionHelper* structPromotionHelper;
3188 #if !defined(_TARGET_64BIT_)
3189 void lvaPromoteLongVars();
3190 #endif // !defined(_TARGET_64BIT_)
3191 unsigned lvaGetFieldLocal(const LclVarDsc* varDsc, unsigned int fldOffset);
3192 lvaPromotionType lvaGetPromotionType(const LclVarDsc* varDsc);
3193 lvaPromotionType lvaGetPromotionType(unsigned varNum);
3194 lvaPromotionType lvaGetParentPromotionType(const LclVarDsc* varDsc);
3195 lvaPromotionType lvaGetParentPromotionType(unsigned varNum);
3196 bool lvaIsFieldOfDependentlyPromotedStruct(const LclVarDsc* varDsc);
3197 bool lvaIsGCTracked(const LclVarDsc* varDsc);
3199 #if defined(FEATURE_SIMD)
3200 bool lvaMapSimd12ToSimd16(const LclVarDsc* varDsc)
3202 assert(varDsc->lvType == TYP_SIMD12);
3203 assert(varDsc->lvExactSize == 12);
3205 #if defined(_TARGET_64BIT_)
3206 assert(varDsc->lvSize() == 16);
3207 #endif // defined(_TARGET_64BIT_)
3209 // We make local variable SIMD12 types 16 bytes instead of just 12. lvSize()
3210 // already does this calculation. However, we also need to prevent mapping types if the var is a
3211 // dependently promoted struct field, which must remain its exact size within its parent struct.
3212 // However, we don't know this until late, so we may have already pretended the field is bigger
3214 if ((varDsc->lvSize() == 16) && !lvaIsFieldOfDependentlyPromotedStruct(varDsc))
3223 #endif // defined(FEATURE_SIMD)
3225 BYTE* lvaGetGcLayout(unsigned varNum);
3226 bool lvaTypeIsGC(unsigned varNum);
3227 unsigned lvaGSSecurityCookie; // LclVar number
3228 bool lvaTempsHaveLargerOffsetThanVars();
3230 // Returns "true" iff local variable "lclNum" is in SSA form.
3231 bool lvaInSsa(unsigned lclNum)
3233 assert(lclNum < lvaCount);
3234 return lvaTable[lclNum].lvInSsa;
3237 unsigned lvaSecurityObject; // variable representing the security object on the stack
3238 unsigned lvaStubArgumentVar; // variable representing the secret stub argument coming in EAX
3240 #if FEATURE_EH_FUNCLETS
3241 unsigned lvaPSPSym; // variable representing the PSPSym
3244 InlineInfo* impInlineInfo;
3245 InlineStrategy* m_inlineStrategy;
3247 // The Compiler* that is the root of the inlining tree of which "this" is a member.
3248 Compiler* impInlineRoot();
3250 #if defined(DEBUG) || defined(INLINE_DATA)
3251 unsigned __int64 getInlineCycleCount()
3253 return m_compCycles;
3255 #endif // defined(DEBUG) || defined(INLINE_DATA)
3257 bool fgNoStructPromotion; // Set to TRUE to turn off struct promotion for this method.
3258 bool fgNoStructParamPromotion; // Set to TRUE to turn off struct promotion for parameters this method.
3260 //=========================================================================
3262 //=========================================================================
3265 //---------------- Local variable ref-counting ----------------------------
3267 void lvaMarkLclRefs(GenTree* tree, BasicBlock* block, GenTreeStmt* stmt, bool isRecompute);
3268 bool IsDominatedByExceptionalEntry(BasicBlock* block);
3269 void SetVolatileHint(LclVarDsc* varDsc);
3271 // Keeps the mapping from SSA #'s to VN's for the implicit memory variables.
3272 SsaDefArray<SsaMemDef> lvMemoryPerSsaData;
3275 // Returns the address of the per-Ssa data for memory at the given ssaNum (which is required
3276 // not to be the SsaConfig::RESERVED_SSA_NUM, which indicates that the variable is
3277 // not an SSA variable).
3278 SsaMemDef* GetMemoryPerSsaData(unsigned ssaNum)
3280 return lvMemoryPerSsaData.GetSsaDef(ssaNum);
3284 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
3285 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
3289 XX Imports the given method and converts it to semantic trees XX
3291 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
3292 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
3298 void impImport(BasicBlock* method);
3300 CORINFO_CLASS_HANDLE impGetRefAnyClass();
3301 CORINFO_CLASS_HANDLE impGetRuntimeArgumentHandle();
3302 CORINFO_CLASS_HANDLE impGetTypeHandleClass();
3303 CORINFO_CLASS_HANDLE impGetStringClass();
3304 CORINFO_CLASS_HANDLE impGetObjectClass();
3306 // Returns underlying type of handles returned by ldtoken instruction
3307 inline var_types GetRuntimeHandleUnderlyingType()
3309 // RuntimeTypeHandle is backed by raw pointer on CoreRT and by object reference on other runtimes
3310 return IsTargetAbi(CORINFO_CORERT_ABI) ? TYP_I_IMPL : TYP_REF;
3313 void impDevirtualizeCall(GenTreeCall* call,
3314 CORINFO_METHOD_HANDLE* method,
3315 unsigned* methodFlags,
3316 CORINFO_CONTEXT_HANDLE* contextHandle,
3317 CORINFO_CONTEXT_HANDLE* exactContextHandle,
3318 bool isLateDevirtualization);
3320 //=========================================================================
3322 //=========================================================================
3325 //-------------------- Stack manipulation ---------------------------------
3327 unsigned impStkSize; // Size of the full stack
3329 #define SMALL_STACK_SIZE 16 // number of elements in impSmallStack
3331 struct SavedStack // used to save/restore stack contents.
3333 unsigned ssDepth; // number of values on stack
3334 StackEntry* ssTrees; // saved tree values
3337 bool impIsPrimitive(CorInfoType type);
3338 bool impILConsumesAddr(const BYTE* codeAddr, CORINFO_METHOD_HANDLE fncHandle, CORINFO_MODULE_HANDLE scpHandle);
3340 void impResolveToken(const BYTE* addr, CORINFO_RESOLVED_TOKEN* pResolvedToken, CorInfoTokenKind kind);
3342 void impPushOnStack(GenTree* tree, typeInfo ti);
3343 void impPushNullObjRefOnStack();
3344 StackEntry impPopStack();
3345 StackEntry& impStackTop(unsigned n = 0);
3346 unsigned impStackHeight();
3348 void impSaveStackState(SavedStack* savePtr, bool copy);
3349 void impRestoreStackState(SavedStack* savePtr);
3351 GenTree* impImportLdvirtftn(GenTree* thisPtr, CORINFO_RESOLVED_TOKEN* pResolvedToken, CORINFO_CALL_INFO* pCallInfo);
3353 void impImportAndPushBox(CORINFO_RESOLVED_TOKEN* pResolvedToken);
3355 void impImportNewObjArray(CORINFO_RESOLVED_TOKEN* pResolvedToken, CORINFO_CALL_INFO* pCallInfo);
3357 bool impCanPInvokeInline();
3358 bool impCanPInvokeInlineCallSite(BasicBlock* block);
3359 void impCheckForPInvokeCall(
3360 GenTreeCall* call, CORINFO_METHOD_HANDLE methHnd, CORINFO_SIG_INFO* sig, unsigned mflags, BasicBlock* block);
3361 GenTreeCall* impImportIndirectCall(CORINFO_SIG_INFO* sig, IL_OFFSETX ilOffset = BAD_IL_OFFSET);
3362 void impPopArgsForUnmanagedCall(GenTree* call, CORINFO_SIG_INFO* sig);
3364 void impInsertHelperCall(CORINFO_HELPER_DESC* helperCall);
3365 void impHandleAccessAllowed(CorInfoIsAccessAllowedResult result, CORINFO_HELPER_DESC* helperCall);
3366 void impHandleAccessAllowedInternal(CorInfoIsAccessAllowedResult result, CORINFO_HELPER_DESC* helperCall);
3368 var_types impImportCall(OPCODE opcode,
3369 CORINFO_RESOLVED_TOKEN* pResolvedToken,
3370 CORINFO_RESOLVED_TOKEN* pConstrainedResolvedToken, // Is this a "constrained." call on a
3372 GenTree* newobjThis,
3374 CORINFO_CALL_INFO* callInfo,
3375 IL_OFFSET rawILOffset);
3377 CORINFO_CLASS_HANDLE impGetSpecialIntrinsicExactReturnType(CORINFO_METHOD_HANDLE specialIntrinsicHandle);
3379 bool impMethodInfo_hasRetBuffArg(CORINFO_METHOD_INFO* methInfo);
3381 GenTree* impFixupCallStructReturn(GenTreeCall* call, CORINFO_CLASS_HANDLE retClsHnd);
3383 GenTree* impFixupStructReturnType(GenTree* op, CORINFO_CLASS_HANDLE retClsHnd);
3386 var_types impImportJitTestLabelMark(int numArgs);
3389 GenTree* impInitClass(CORINFO_RESOLVED_TOKEN* pResolvedToken);
3391 GenTree* impImportStaticReadOnlyField(void* fldAddr, var_types lclTyp);
3393 GenTree* impImportStaticFieldAccess(CORINFO_RESOLVED_TOKEN* pResolvedToken,
3394 CORINFO_ACCESS_FLAGS access,
3395 CORINFO_FIELD_INFO* pFieldInfo,
3398 static void impBashVarAddrsToI(GenTree* tree1, GenTree* tree2 = nullptr);
3400 GenTree* impImplicitIorI4Cast(GenTree* tree, var_types dstTyp);
3402 GenTree* impImplicitR4orR8Cast(GenTree* tree, var_types dstTyp);
3404 void impImportLeave(BasicBlock* block);
3405 void impResetLeaveBlock(BasicBlock* block, unsigned jmpAddr);
3406 GenTree* impIntrinsic(GenTree* newobjThis,
3407 CORINFO_CLASS_HANDLE clsHnd,
3408 CORINFO_METHOD_HANDLE method,
3409 CORINFO_SIG_INFO* sig,
3410 unsigned methodFlags,
3414 CORINFO_RESOLVED_TOKEN* pContstrainedResolvedToken,
3415 CORINFO_THIS_TRANSFORM constraintCallThisTransform,
3416 CorInfoIntrinsics* pIntrinsicID,
3417 bool* isSpecialIntrinsic = nullptr);
3418 GenTree* impMathIntrinsic(CORINFO_METHOD_HANDLE method,
3419 CORINFO_SIG_INFO* sig,
3421 CorInfoIntrinsics intrinsicID,
3423 NamedIntrinsic lookupNamedIntrinsic(CORINFO_METHOD_HANDLE method);
3425 #ifdef FEATURE_HW_INTRINSICS
3426 GenTree* impBaseIntrinsic(NamedIntrinsic intrinsic,
3427 CORINFO_CLASS_HANDLE clsHnd,
3428 CORINFO_METHOD_HANDLE method,
3429 CORINFO_SIG_INFO* sig);
3430 GenTree* impHWIntrinsic(NamedIntrinsic intrinsic,
3431 CORINFO_METHOD_HANDLE method,
3432 CORINFO_SIG_INFO* sig,
3434 GenTree* impUnsupportedHWIntrinsic(unsigned helper,
3435 CORINFO_METHOD_HANDLE method,
3436 CORINFO_SIG_INFO* sig,
3440 bool compSupportsHWIntrinsic(InstructionSet isa);
3442 #ifdef _TARGET_XARCH_
3443 GenTree* impSSEIntrinsic(NamedIntrinsic intrinsic,
3444 CORINFO_METHOD_HANDLE method,
3445 CORINFO_SIG_INFO* sig,
3447 GenTree* impSSE2Intrinsic(NamedIntrinsic intrinsic,
3448 CORINFO_METHOD_HANDLE method,
3449 CORINFO_SIG_INFO* sig,
3451 GenTree* impSSE42Intrinsic(NamedIntrinsic intrinsic,
3452 CORINFO_METHOD_HANDLE method,
3453 CORINFO_SIG_INFO* sig,
3455 GenTree* impAvxOrAvx2Intrinsic(NamedIntrinsic intrinsic,
3456 CORINFO_METHOD_HANDLE method,
3457 CORINFO_SIG_INFO* sig,
3459 GenTree* impAESIntrinsic(NamedIntrinsic intrinsic,
3460 CORINFO_METHOD_HANDLE method,
3461 CORINFO_SIG_INFO* sig,
3463 GenTree* impBMI1OrBMI2Intrinsic(NamedIntrinsic intrinsic,
3464 CORINFO_METHOD_HANDLE method,
3465 CORINFO_SIG_INFO* sig,
3467 GenTree* impFMAIntrinsic(NamedIntrinsic intrinsic,
3468 CORINFO_METHOD_HANDLE method,
3469 CORINFO_SIG_INFO* sig,
3471 GenTree* impLZCNTIntrinsic(NamedIntrinsic intrinsic,
3472 CORINFO_METHOD_HANDLE method,
3473 CORINFO_SIG_INFO* sig,
3475 GenTree* impPCLMULQDQIntrinsic(NamedIntrinsic intrinsic,
3476 CORINFO_METHOD_HANDLE method,
3477 CORINFO_SIG_INFO* sig,
3479 GenTree* impPOPCNTIntrinsic(NamedIntrinsic intrinsic,
3480 CORINFO_METHOD_HANDLE method,
3481 CORINFO_SIG_INFO* sig,
3485 GenTree* getArgForHWIntrinsic(var_types argType, CORINFO_CLASS_HANDLE argClass);
3486 GenTree* impNonConstFallback(NamedIntrinsic intrinsic, var_types simdType, var_types baseType);
3487 GenTree* addRangeCheckIfNeeded(NamedIntrinsic intrinsic, GenTree* lastOp, bool mustExpand);
3488 #endif // _TARGET_XARCH_
3489 #ifdef _TARGET_ARM64_
3490 InstructionSet lookupHWIntrinsicISA(const char* className);
3491 NamedIntrinsic lookupHWIntrinsic(const char* className, const char* methodName);
3492 bool impCheckImmediate(GenTree* immediateOp, unsigned int max);
3493 #endif // _TARGET_ARM64_
3494 #endif // FEATURE_HW_INTRINSICS
3495 GenTree* impArrayAccessIntrinsic(CORINFO_CLASS_HANDLE clsHnd,
3496 CORINFO_SIG_INFO* sig,
3499 CorInfoIntrinsics intrinsicID);
3500 GenTree* impInitializeArrayIntrinsic(CORINFO_SIG_INFO* sig);
3502 GenTree* impMethodPointer(CORINFO_RESOLVED_TOKEN* pResolvedToken, CORINFO_CALL_INFO* pCallInfo);
3504 GenTree* impTransformThis(GenTree* thisPtr,
3505 CORINFO_RESOLVED_TOKEN* pConstrainedResolvedToken,
3506 CORINFO_THIS_TRANSFORM transform);
3508 //----------------- Manipulating the trees and stmts ----------------------
3510 GenTree* impTreeList; // Trees for the BB being imported
3511 GenTree* impTreeLast; // The last tree for the current BB
3516 CHECK_SPILL_ALL = -1,
3517 CHECK_SPILL_NONE = -2
3520 void impBeginTreeList();
3521 void impEndTreeList(BasicBlock* block, GenTree* firstStmt, GenTree* lastStmt);
3522 void impEndTreeList(BasicBlock* block);
3523 void impAppendStmtCheck(GenTree* stmt, unsigned chkLevel);
3524 void impAppendStmt(GenTree* stmt, unsigned chkLevel);
3525 void impInsertStmtBefore(GenTree* stmt, GenTree* stmtBefore);
3526 GenTree* impAppendTree(GenTree* tree, unsigned chkLevel, IL_OFFSETX offset);
3527 void impInsertTreeBefore(GenTree* tree, IL_OFFSETX offset, GenTree* stmtBefore);
3528 void impAssignTempGen(unsigned tmp,
3531 GenTree** pAfterStmt = nullptr,
3532 IL_OFFSETX ilOffset = BAD_IL_OFFSET,
3533 BasicBlock* block = nullptr);
3534 void impAssignTempGen(unsigned tmpNum,
3536 CORINFO_CLASS_HANDLE structHnd,
3538 GenTree** pAfterStmt = nullptr,
3539 IL_OFFSETX ilOffset = BAD_IL_OFFSET,
3540 BasicBlock* block = nullptr);
3541 GenTree* impCloneExpr(GenTree* tree,
3543 CORINFO_CLASS_HANDLE structHnd,
3545 GenTree** pAfterStmt DEBUGARG(const char* reason));
3546 GenTree* impAssignStruct(GenTree* dest,
3548 CORINFO_CLASS_HANDLE structHnd,
3550 GenTree** pAfterStmt = nullptr,
3551 IL_OFFSETX ilOffset = BAD_IL_OFFSET,
3552 BasicBlock* block = nullptr);
3553 GenTree* impAssignStructPtr(GenTree* dest,
3555 CORINFO_CLASS_HANDLE structHnd,
3557 GenTree** pAfterStmt = nullptr,
3558 IL_OFFSETX ilOffset = BAD_IL_OFFSET,
3559 BasicBlock* block = nullptr);
3561 GenTree* impGetStructAddr(GenTree* structVal, CORINFO_CLASS_HANDLE structHnd, unsigned curLevel, bool willDeref);
3563 var_types impNormStructType(CORINFO_CLASS_HANDLE structHnd,
3564 BYTE* gcLayout = nullptr,
3565 unsigned* numGCVars = nullptr,
3566 var_types* simdBaseType = nullptr);
3568 GenTree* impNormStructVal(GenTree* structVal,
3569 CORINFO_CLASS_HANDLE structHnd,
3571 bool forceNormalization = false);
3573 GenTree* impTokenToHandle(CORINFO_RESOLVED_TOKEN* pResolvedToken,
3574 BOOL* pRuntimeLookup = nullptr,
3575 BOOL mustRestoreHandle = FALSE,
3576 BOOL importParent = FALSE);
3578 GenTree* impParentClassTokenToHandle(CORINFO_RESOLVED_TOKEN* pResolvedToken,
3579 BOOL* pRuntimeLookup = nullptr,
3580 BOOL mustRestoreHandle = FALSE)
3582 return impTokenToHandle(pResolvedToken, pRuntimeLookup, mustRestoreHandle, TRUE);
3585 GenTree* impLookupToTree(CORINFO_RESOLVED_TOKEN* pResolvedToken,
3586 CORINFO_LOOKUP* pLookup,
3588 void* compileTimeHandle);
3590 GenTree* getRuntimeContextTree(CORINFO_RUNTIME_LOOKUP_KIND kind);
3592 GenTree* impRuntimeLookupToTree(CORINFO_RESOLVED_TOKEN* pResolvedToken,
3593 CORINFO_LOOKUP* pLookup,
3594 void* compileTimeHandle);
3596 GenTree* impReadyToRunLookupToTree(CORINFO_CONST_LOOKUP* pLookup, unsigned flags, void* compileTimeHandle);
3598 GenTreeCall* impReadyToRunHelperToTree(CORINFO_RESOLVED_TOKEN* pResolvedToken,
3599 CorInfoHelpFunc helper,
3601 GenTreeArgList* arg = nullptr,
3602 CORINFO_LOOKUP_KIND* pGenericLookupKind = nullptr);
3604 GenTree* impCastClassOrIsInstToTree(GenTree* op1,
3606 CORINFO_RESOLVED_TOKEN* pResolvedToken,
3609 GenTree* impOptimizeCastClassOrIsInst(GenTree* op1, CORINFO_RESOLVED_TOKEN* pResolvedToken, bool isCastClass);
3611 bool VarTypeIsMultiByteAndCanEnreg(
3612 var_types type, CORINFO_CLASS_HANDLE typeClass, unsigned* typeSize, bool forReturn, bool isVarArg);
3614 bool IsIntrinsicImplementedByUserCall(CorInfoIntrinsics intrinsicId);
3615 bool IsTargetIntrinsic(CorInfoIntrinsics intrinsicId);
3616 bool IsMathIntrinsic(CorInfoIntrinsics intrinsicId);
3617 bool IsMathIntrinsic(GenTree* tree);
3620 //----------------- Importing the method ----------------------------------
3622 CORINFO_CONTEXT_HANDLE impTokenLookupContextHandle; // The context used for looking up tokens.
3625 unsigned impCurOpcOffs;
3626 const char* impCurOpcName;
3627 bool impNestedStackSpill;
3629 // For displaying instrs with generated native code (-n:B)
3630 GenTree* impLastILoffsStmt; // oldest stmt added for which we did not gtStmtLastILoffs
3631 void impNoteLastILoffs();
3634 /* IL offset of the stmt currently being imported. It gets set to
3635 BAD_IL_OFFSET after it has been set in the appended trees. Then it gets
3636 updated at IL offsets for which we have to report mapping info.
3637 It also includes flag bits, so use jitGetILoffs()
3638 to get the actual IL offset value.
3641 IL_OFFSETX impCurStmtOffs;
3642 void impCurStmtOffsSet(IL_OFFSET offs);
3644 void impNoteBranchOffs();
3646 unsigned impInitBlockLineInfo();
3648 GenTree* impCheckForNullPointer(GenTree* obj);
3649 bool impIsThis(GenTree* obj);
3650 bool impIsLDFTN_TOKEN(const BYTE* delegateCreateStart, const BYTE* newobjCodeAddr);
3651 bool impIsDUP_LDVIRTFTN_TOKEN(const BYTE* delegateCreateStart, const BYTE* newobjCodeAddr);
3652 bool impIsAnySTLOC(OPCODE opcode)
3654 return ((opcode == CEE_STLOC) || (opcode == CEE_STLOC_S) ||
3655 ((opcode >= CEE_STLOC_0) && (opcode <= CEE_STLOC_3)));
3658 GenTreeArgList* impPopList(unsigned count, CORINFO_SIG_INFO* sig, GenTreeArgList* prefixTree = nullptr);
3660 GenTreeArgList* impPopRevList(unsigned count, CORINFO_SIG_INFO* sig, unsigned skipReverseCount = 0);
3663 * Get current IL offset with stack-empty info incoporated
3665 IL_OFFSETX impCurILOffset(IL_OFFSET offs, bool callInstruction = false);
3667 //---------------- Spilling the importer stack ----------------------------
3669 // The maximum number of bytes of IL processed without clean stack state.
3670 // It allows to limit the maximum tree size and depth.
3671 static const unsigned MAX_TREE_SIZE = 200;
3672 bool impCanSpillNow(OPCODE prevOpcode);
3678 SavedStack pdSavedStack;
3679 ThisInitState pdThisPtrInit;
3682 PendingDsc* impPendingList; // list of BBs currently waiting to be imported.
3683 PendingDsc* impPendingFree; // Freed up dscs that can be reused
3685 // We keep a byte-per-block map (dynamically extended) in the top-level Compiler object of a compilation.
3686 JitExpandArray<BYTE> impPendingBlockMembers;
3688 // Return the byte for "b" (allocating/extending impPendingBlockMembers if necessary.)
3689 // Operates on the map in the top-level ancestor.
3690 BYTE impGetPendingBlockMember(BasicBlock* blk)
3692 return impInlineRoot()->impPendingBlockMembers.Get(blk->bbInd());
3695 // Set the byte for "b" to "val" (allocating/extending impPendingBlockMembers if necessary.)
3696 // Operates on the map in the top-level ancestor.
3697 void impSetPendingBlockMember(BasicBlock* blk, BYTE val)
3699 impInlineRoot()->impPendingBlockMembers.Set(blk->bbInd(), val);
3702 bool impCanReimport;
3704 bool impSpillStackEntry(unsigned level,
3708 bool bAssertOnRecursion,
3713 void impSpillStackEnsure(bool spillLeaves = false);
3714 void impEvalSideEffects();
3715 void impSpillSpecialSideEff();
3716 void impSpillSideEffects(bool spillGlobEffects, unsigned chkLevel DEBUGARG(const char* reason));
3717 void impSpillValueClasses();
3718 void impSpillEvalStack();
3719 static fgWalkPreFn impFindValueClasses;
3720 void impSpillLclRefs(ssize_t lclNum);
3722 BasicBlock* impPushCatchArgOnStack(BasicBlock* hndBlk, CORINFO_CLASS_HANDLE clsHnd, bool isSingleBlockFilter);
3724 void impImportBlockCode(BasicBlock* block);
3726 void impReimportMarkBlock(BasicBlock* block);
3727 void impReimportMarkSuccessors(BasicBlock* block);
3729 void impVerifyEHBlock(BasicBlock* block, bool isTryStart);
3731 void impImportBlockPending(BasicBlock* block);
3733 // Similar to impImportBlockPending, but assumes that block has already been imported once and is being
3734 // reimported for some reason. It specifically does *not* look at verCurrentState to set the EntryState
3735 // for the block, but instead, just re-uses the block's existing EntryState.
3736 void impReimportBlockPending(BasicBlock* block);
3738 var_types impGetByRefResultType(genTreeOps oper, bool fUnsigned, GenTree** pOp1, GenTree** pOp2);
3740 void impImportBlock(BasicBlock* block);
3742 // Assumes that "block" is a basic block that completes with a non-empty stack. We will assign the values
3743 // on the stack to local variables (the "spill temp" variables). The successor blocks will assume that
3744 // its incoming stack contents are in those locals. This requires "block" and its successors to agree on
3745 // the variables that will be used -- and for all the predecessors of those successors, and the
3746 // successors of those predecessors, etc. Call such a set of blocks closed under alternating
3747 // successor/predecessor edges a "spill clique." A block is a "predecessor" or "successor" member of the
3748 // clique (or, conceivably, both). Each block has a specified sequence of incoming and outgoing spill
3749 // temps. If "block" already has its outgoing spill temps assigned (they are always a contiguous series
3750 // of local variable numbers, so we represent them with the base local variable number), returns that.
3751 // Otherwise, picks a set of spill temps, and propagates this choice to all blocks in the spill clique of
3752 // which "block" is a member (asserting, in debug mode, that no block in this clique had its spill temps
3753 // chosen already. More precisely, that the incoming or outgoing spill temps are not chosen, depending
3754 // on which kind of member of the clique the block is).
3755 unsigned impGetSpillTmpBase(BasicBlock* block);
3757 // Assumes that "block" is a basic block that completes with a non-empty stack. We have previously
3758 // assigned the values on the stack to local variables (the "spill temp" variables). The successor blocks
3759 // will assume that its incoming stack contents are in those locals. This requires "block" and its
3760 // successors to agree on the variables and their types that will be used. The CLI spec allows implicit
3761 // conversions between 'int' and 'native int' or 'float' and 'double' stack types. So one predecessor can
3762 // push an int and another can push a native int. For 64-bit we have chosen to implement this by typing
3763 // the "spill temp" as native int, and then importing (or re-importing as needed) so that all the
3764 // predecessors in the "spill clique" push a native int (sign-extending if needed), and all the
3765 // successors receive a native int. Similarly float and double are unified to double.
3766 // This routine is called after a type-mismatch is detected, and it will walk the spill clique to mark
3767 // blocks for re-importation as appropriate (both successors, so they get the right incoming type, and
3768 // predecessors, so they insert an upcast if needed).
3769 void impReimportSpillClique(BasicBlock* block);
3771 // When we compute a "spill clique" (see above) these byte-maps are allocated to have a byte per basic
3772 // block, and represent the predecessor and successor members of the clique currently being computed.
3773 // *** Access to these will need to be locked in a parallel compiler.
3774 JitExpandArray<BYTE> impSpillCliquePredMembers;
3775 JitExpandArray<BYTE> impSpillCliqueSuccMembers;
3783 // Abstract class for receiving a callback while walking a spill clique
3784 class SpillCliqueWalker
3787 virtual void Visit(SpillCliqueDir predOrSucc, BasicBlock* blk) = 0;
3790 // This class is used for setting the bbStkTempsIn and bbStkTempsOut on the blocks within a spill clique
3791 class SetSpillTempsBase : public SpillCliqueWalker
3796 SetSpillTempsBase(unsigned baseTmp) : m_baseTmp(baseTmp)
3799 virtual void Visit(SpillCliqueDir predOrSucc, BasicBlock* blk);
3802 // This class is used for implementing impReimportSpillClique part on each block within the spill clique
3803 class ReimportSpillClique : public SpillCliqueWalker
3808 ReimportSpillClique(Compiler* pComp) : m_pComp(pComp)
3811 virtual void Visit(SpillCliqueDir predOrSucc, BasicBlock* blk);
3814 // This is the heart of the algorithm for walking spill cliques. It invokes callback->Visit for each
3815 // predecessor or successor within the spill clique
3816 void impWalkSpillCliqueFromPred(BasicBlock* pred, SpillCliqueWalker* callback);
3818 // For a BasicBlock that has already been imported, the EntryState has an array of GenTrees for the
3819 // incoming locals. This walks that list an resets the types of the GenTrees to match the types of
3820 // the VarDscs. They get out of sync when we have int/native int issues (see impReimportSpillClique).
3821 void impRetypeEntryStateTemps(BasicBlock* blk);
3823 BYTE impSpillCliqueGetMember(SpillCliqueDir predOrSucc, BasicBlock* blk);
3824 void impSpillCliqueSetMember(SpillCliqueDir predOrSucc, BasicBlock* blk, BYTE val);
3826 void impPushVar(GenTree* op, typeInfo tiRetVal);
3827 void impLoadVar(unsigned lclNum, IL_OFFSET offset, typeInfo tiRetVal);
3828 void impLoadVar(unsigned lclNum, IL_OFFSET offset)
3830 impLoadVar(lclNum, offset, lvaTable[lclNum].lvVerTypeInfo);
3832 void impLoadArg(unsigned ilArgNum, IL_OFFSET offset);
3833 void impLoadLoc(unsigned ilLclNum, IL_OFFSET offset);
3834 bool impReturnInstruction(BasicBlock* block, int prefixFlags, OPCODE& opcode);
3837 void impMarkLclDstNotPromotable(unsigned tmpNum, GenTree* op, CORINFO_CLASS_HANDLE hClass);
3840 // A free list of linked list nodes used to represent to-do stacks of basic blocks.
3841 struct BlockListNode
3844 BlockListNode* m_next;
3845 BlockListNode(BasicBlock* blk, BlockListNode* next = nullptr) : m_blk(blk), m_next(next)
3848 void* operator new(size_t sz, Compiler* comp);
3850 BlockListNode* impBlockListNodeFreeList;
3852 void FreeBlockListNode(BlockListNode* node);
3854 bool impIsValueType(typeInfo* pTypeInfo);
3855 var_types mangleVarArgsType(var_types type);
3858 regNumber getCallArgIntRegister(regNumber floatReg);
3859 regNumber getCallArgFloatRegister(regNumber intReg);
3860 #endif // FEATURE_VARARG
3863 static unsigned jitTotalMethodCompiled;
3867 static LONG jitNestingLevel;
3870 static BOOL impIsAddressInLocal(GenTree* tree, GenTree** lclVarTreeOut);
3872 void impMakeDiscretionaryInlineObservations(InlineInfo* pInlineInfo, InlineResult* inlineResult);
3874 // STATIC inlining decision based on the IL code.
3875 void impCanInlineIL(CORINFO_METHOD_HANDLE fncHandle,
3876 CORINFO_METHOD_INFO* methInfo,
3878 InlineResult* inlineResult);
3880 void impCheckCanInline(GenTreeCall* call,
3881 CORINFO_METHOD_HANDLE fncHandle,
3883 CORINFO_CONTEXT_HANDLE exactContextHnd,
3884 InlineCandidateInfo** ppInlineCandidateInfo,
3885 InlineResult* inlineResult);
3887 void impInlineRecordArgInfo(InlineInfo* pInlineInfo,
3890 InlineResult* inlineResult);
3892 void impInlineInitVars(InlineInfo* pInlineInfo);
3894 unsigned impInlineFetchLocal(unsigned lclNum DEBUGARG(const char* reason));
3896 GenTree* impInlineFetchArg(unsigned lclNum, InlArgInfo* inlArgInfo, InlLclVarInfo* lclTypeInfo);
3898 BOOL impInlineIsThis(GenTree* tree, InlArgInfo* inlArgInfo);
3900 BOOL impInlineIsGuaranteedThisDerefBeforeAnySideEffects(GenTree* additionalTreesToBeEvaluatedBefore,
3901 GenTree* variableBeingDereferenced,
3902 InlArgInfo* inlArgInfo);
3904 void impMarkInlineCandidate(GenTree* call,
3905 CORINFO_CONTEXT_HANDLE exactContextHnd,
3906 bool exactContextNeedsRuntimeLookup,
3907 CORINFO_CALL_INFO* callInfo);
3909 void impMarkInlineCandidateHelper(GenTreeCall* call,
3910 CORINFO_CONTEXT_HANDLE exactContextHnd,
3911 bool exactContextNeedsRuntimeLookup,
3912 CORINFO_CALL_INFO* callInfo);
3914 bool impTailCallRetTypeCompatible(var_types callerRetType,
3915 CORINFO_CLASS_HANDLE callerRetTypeClass,
3916 var_types calleeRetType,
3917 CORINFO_CLASS_HANDLE calleeRetTypeClass);
3919 bool impIsTailCallILPattern(bool tailPrefixed,
3921 const BYTE* codeAddrOfNextOpcode,
3922 const BYTE* codeEnd,
3924 bool* IsCallPopRet = nullptr);
3926 bool impIsImplicitTailCallCandidate(
3927 OPCODE curOpcode, const BYTE* codeAddrOfNextOpcode, const BYTE* codeEnd, int prefixFlags, bool isRecursive);
3929 CORINFO_RESOLVED_TOKEN* impAllocateToken(CORINFO_RESOLVED_TOKEN token);
3932 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
3933 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
3937 XX Info about the basic-blocks, their contents and the flow analysis XX
3939 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
3940 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
3944 BasicBlock* fgFirstBB; // Beginning of the basic block list
3945 BasicBlock* fgLastBB; // End of the basic block list
3946 BasicBlock* fgFirstColdBlock; // First block to be placed in the cold section
3947 #if FEATURE_EH_FUNCLETS
3948 BasicBlock* fgFirstFuncletBB; // First block of outlined funclets (to allow block insertion before the funclets)
3950 BasicBlock* fgFirstBBScratch; // Block inserted for initialization stuff. Is nullptr if no such block has been
3952 BasicBlockList* fgReturnBlocks; // list of BBJ_RETURN blocks
3953 unsigned fgEdgeCount; // # of control flow edges between the BBs
3954 unsigned fgBBcount; // # of BBs in the method
3956 unsigned fgBBcountAtCodegen; // # of BBs in the method at the start of codegen
3958 unsigned fgBBNumMax; // The max bbNum that has been assigned to basic blocks
3959 unsigned fgDomBBcount; // # of BBs for which we have dominator and reachability information
3960 BasicBlock** fgBBInvPostOrder; // The flow graph stored in an array sorted in topological order, needed to compute
3961 // dominance. Indexed by block number. Size: fgBBNumMax + 1.
3963 // After the dominance tree is computed, we cache a DFS preorder number and DFS postorder number to compute
3964 // dominance queries in O(1). fgDomTreePreOrder and fgDomTreePostOrder are arrays giving the block's preorder and
3965 // postorder number, respectively. The arrays are indexed by basic block number. (Note that blocks are numbered
3966 // starting from one. Thus, we always waste element zero. This makes debugging easier and makes the code less likely
3967 // to suffer from bugs stemming from forgetting to add or subtract one from the block number to form an array
3968 // index). The arrays are of size fgBBNumMax + 1.
3969 unsigned* fgDomTreePreOrder;
3970 unsigned* fgDomTreePostOrder;
3972 bool fgBBVarSetsInited;
3974 // Allocate array like T* a = new T[fgBBNumMax + 1];
3975 // Using helper so we don't keep forgetting +1.
3976 template <typename T>
3977 T* fgAllocateTypeForEachBlk(CompMemKind cmk = CMK_Unknown)
3979 return getAllocator(cmk).allocate<T>(fgBBNumMax + 1);
3982 // BlockSets are relative to a specific set of BasicBlock numbers. If that changes
3983 // (if the blocks are renumbered), this changes. BlockSets from different epochs
3984 // cannot be meaningfully combined. Note that new blocks can be created with higher
3985 // block numbers without changing the basic block epoch. These blocks *cannot*
3986 // participate in a block set until the blocks are all renumbered, causing the epoch
3987 // to change. This is useful if continuing to use previous block sets is valuable.
3988 // If the epoch is zero, then it is uninitialized, and block sets can't be used.
3989 unsigned fgCurBBEpoch;
3991 unsigned GetCurBasicBlockEpoch()
3993 return fgCurBBEpoch;
3996 // The number of basic blocks in the current epoch. When the blocks are renumbered,
3997 // this is fgBBcount. As blocks are added, fgBBcount increases, fgCurBBEpochSize remains
3998 // the same, until a new BasicBlock epoch is created, such as when the blocks are all renumbered.
3999 unsigned fgCurBBEpochSize;
4001 // The number of "size_t" elements required to hold a bitset large enough for fgCurBBEpochSize
4002 // bits. This is precomputed to avoid doing math every time BasicBlockBitSetTraits::GetArrSize() is called.
4003 unsigned fgBBSetCountInSizeTUnits;
4005 void NewBasicBlockEpoch()
4007 INDEBUG(unsigned oldEpochArrSize = fgBBSetCountInSizeTUnits);
4009 // We have a new epoch. Compute and cache the size needed for new BlockSets.
4011 fgCurBBEpochSize = fgBBNumMax + 1;
4012 fgBBSetCountInSizeTUnits =
4013 roundUp(fgCurBBEpochSize, (unsigned)(sizeof(size_t) * 8)) / unsigned(sizeof(size_t) * 8);
4016 // All BlockSet objects are now invalid!
4017 fgReachabilitySetsValid = false; // the bbReach sets are now invalid!
4018 fgEnterBlksSetValid = false; // the fgEnterBlks set is now invalid!
4022 unsigned epochArrSize = BasicBlockBitSetTraits::GetArrSize(this, sizeof(size_t));
4023 printf("\nNew BlockSet epoch %d, # of blocks (including unused BB00): %u, bitset array size: %u (%s)",
4024 fgCurBBEpoch, fgCurBBEpochSize, epochArrSize, (epochArrSize <= 1) ? "short" : "long");
4025 if ((fgCurBBEpoch != 1) && ((oldEpochArrSize <= 1) != (epochArrSize <= 1)))
4027 // If we're not just establishing the first epoch, and the epoch array size has changed such that we're
4028 // going to change our bitset representation from short (just a size_t bitset) to long (a pointer to an
4029 // array of size_t bitsets), then print that out.
4030 printf("; NOTE: BlockSet size was previously %s!", (oldEpochArrSize <= 1) ? "short" : "long");
4037 void EnsureBasicBlockEpoch()
4039 if (fgCurBBEpochSize != fgBBNumMax + 1)
4041 NewBasicBlockEpoch();
4045 BasicBlock* fgNewBasicBlock(BBjumpKinds jumpKind);
4046 void fgEnsureFirstBBisScratch();
4047 bool fgFirstBBisScratch();
4048 bool fgBBisScratch(BasicBlock* block);
4050 void fgExtendEHRegionBefore(BasicBlock* block);
4051 void fgExtendEHRegionAfter(BasicBlock* block);
4053 BasicBlock* fgNewBBbefore(BBjumpKinds jumpKind, BasicBlock* block, bool extendRegion);
4055 BasicBlock* fgNewBBafter(BBjumpKinds jumpKind, BasicBlock* block, bool extendRegion);
4057 BasicBlock* fgNewBBinRegion(BBjumpKinds jumpKind,
4060 BasicBlock* nearBlk,
4061 bool putInFilter = false,
4062 bool runRarely = false,
4063 bool insertAtEnd = false);
4065 BasicBlock* fgNewBBinRegion(BBjumpKinds jumpKind,
4067 bool runRarely = false,
4068 bool insertAtEnd = false);
4070 BasicBlock* fgNewBBinRegion(BBjumpKinds jumpKind);
4072 BasicBlock* fgNewBBinRegionWorker(BBjumpKinds jumpKind,
4073 BasicBlock* afterBlk,
4074 unsigned xcptnIndex,
4075 bool putInTryRegion);
4077 void fgInsertBBbefore(BasicBlock* insertBeforeBlk, BasicBlock* newBlk);
4078 void fgInsertBBafter(BasicBlock* insertAfterBlk, BasicBlock* newBlk);
4079 void fgUnlinkBlock(BasicBlock* block);
4081 unsigned fgMeasureIR();
4083 bool fgModified; // True if the flow graph has been modified recently
4084 bool fgComputePredsDone; // Have we computed the bbPreds list
4085 bool fgCheapPredsValid; // Is the bbCheapPreds list valid?
4086 bool fgDomsComputed; // Have we computed the dominator sets?
4087 bool fgOptimizedFinally; // Did we optimize any try-finallys?
4089 bool fgHasSwitch; // any BBJ_SWITCH jumps?
4091 BlockSet fgEnterBlks; // Set of blocks which have a special transfer of control; the "entry" blocks plus EH handler
4095 bool fgReachabilitySetsValid; // Are the bbReach sets valid?
4096 bool fgEnterBlksSetValid; // Is the fgEnterBlks set valid?
4099 bool fgRemoveRestOfBlock; // true if we know that we will throw
4100 bool fgStmtRemoved; // true if we remove statements -> need new DFA
4102 // There are two modes for ordering of the trees.
4103 // - In FGOrderTree, the dominant ordering is the tree order, and the nodes contained in
4104 // each tree and sub-tree are contiguous, and can be traversed (in gtNext/gtPrev order)
4105 // by traversing the tree according to the order of the operands.
4106 // - In FGOrderLinear, the dominant ordering is the linear order.
4113 FlowGraphOrder fgOrder;
4115 // The following are boolean flags that keep track of the state of internal data structures
4117 bool fgStmtListThreaded; // true if the node list is now threaded
4118 bool fgCanRelocateEHRegions; // true if we are allowed to relocate the EH regions
4119 bool fgEdgeWeightsComputed; // true after we have called fgComputeEdgeWeights
4120 bool fgHaveValidEdgeWeights; // true if we were successful in computing all of the edge weights
4121 bool fgSlopUsedInEdgeWeights; // true if their was some slop used when computing the edge weights
4122 bool fgRangeUsedInEdgeWeights; // true if some of the edgeWeight are expressed in Min..Max form
4123 bool fgNeedsUpdateFlowGraph; // true if we need to run fgUpdateFlowGraph
4124 BasicBlock::weight_t fgCalledCount; // count of the number of times this method was called
4125 // This is derived from the profile data
4126 // or is BB_UNITY_WEIGHT when we don't have profile data
4128 #if FEATURE_EH_FUNCLETS
4129 bool fgFuncletsCreated; // true if the funclet creation phase has been run
4130 #endif // FEATURE_EH_FUNCLETS
4132 bool fgGlobalMorph; // indicates if we are during the global morphing phase
4133 // since fgMorphTree can be called from several places
4135 bool impBoxTempInUse; // the temp below is valid and available
4136 unsigned impBoxTemp; // a temporary that is used for boxing
4139 bool jitFallbackCompile; // Are we doing a fallback compile? That is, have we executed a NO_WAY assert,
4140 // and we are trying to compile again in a "safer", minopts mode?
4144 unsigned impInlinedCodeSize;
4147 //-------------------------------------------------------------------------
4153 void fgTransformIndirectCalls();
4157 void fgRemoveEmptyTry();
4159 void fgRemoveEmptyFinally();
4161 void fgMergeFinallyChains();
4163 void fgCloneFinally();
4165 void fgCleanupContinuation(BasicBlock* continuation);
4167 void fgUpdateFinallyTargetFlags();
4169 void fgClearAllFinallyTargetBits();
4171 void fgAddFinallyTargetFlags();
4173 #if FEATURE_EH_FUNCLETS && defined(_TARGET_ARM_)
4174 // Sometimes we need to defer updating the BBF_FINALLY_TARGET bit. fgNeedToAddFinallyTargetBits signals
4175 // when this is necessary.
4176 bool fgNeedToAddFinallyTargetBits;
4177 #endif // FEATURE_EH_FUNCLETS && defined(_TARGET_ARM_)
4179 bool fgRetargetBranchesToCanonicalCallFinally(BasicBlock* block,
4180 BasicBlock* handler,
4181 BlockToBlockMap& continuationMap);
4183 GenTree* fgGetCritSectOfStaticMethod();
4185 #if FEATURE_EH_FUNCLETS
4187 void fgAddSyncMethodEnterExit();
4189 GenTree* fgCreateMonitorTree(unsigned lvaMonitorBool, unsigned lvaThisVar, BasicBlock* block, bool enter);
4191 void fgConvertSyncReturnToLeave(BasicBlock* block);
4193 #endif // FEATURE_EH_FUNCLETS
4195 void fgAddReversePInvokeEnterExit();
4197 bool fgMoreThanOneReturnBlock();
4199 // The number of separate return points in the method.
4200 unsigned fgReturnCount;
4202 void fgAddInternal();
4204 bool fgFoldConditional(BasicBlock* block);
4206 void fgMorphStmts(BasicBlock* block, bool* lnot, bool* loadw);
4207 void fgMorphBlocks();
4209 bool fgMorphBlockStmt(BasicBlock* block, GenTreeStmt* stmt DEBUGARG(const char* msg));
4211 void fgSetOptions();
4214 static fgWalkPreFn fgAssertNoQmark;
4215 void fgPreExpandQmarkChecks(GenTree* expr);
4216 void fgPostExpandQmarkChecks();
4217 static void fgCheckQmarkAllowedForm(GenTree* tree);
4220 IL_OFFSET fgFindBlockILOffset(BasicBlock* block);
4222 BasicBlock* fgSplitBlockAtBeginning(BasicBlock* curr);
4223 BasicBlock* fgSplitBlockAtEnd(BasicBlock* curr);
4224 BasicBlock* fgSplitBlockAfterStatement(BasicBlock* curr, GenTree* stmt);
4225 BasicBlock* fgSplitBlockAfterNode(BasicBlock* curr, GenTree* node); // for LIR
4226 BasicBlock* fgSplitEdge(BasicBlock* curr, BasicBlock* succ);
4228 GenTreeStmt* fgNewStmtFromTree(GenTree* tree, BasicBlock* block, IL_OFFSETX offs);
4229 GenTreeStmt* fgNewStmtFromTree(GenTree* tree);
4230 GenTreeStmt* fgNewStmtFromTree(GenTree* tree, BasicBlock* block);
4231 GenTreeStmt* fgNewStmtFromTree(GenTree* tree, IL_OFFSETX offs);
4233 GenTree* fgGetTopLevelQmark(GenTree* expr, GenTree** ppDst = nullptr);
4234 void fgExpandQmarkForCastInstOf(BasicBlock* block, GenTree* stmt);
4235 void fgExpandQmarkStmt(BasicBlock* block, GenTree* expr);
4236 void fgExpandQmarkNodes();
4240 // Do "simple lowering." This functionality is (conceptually) part of "general"
4241 // lowering that is distributed between fgMorph and the lowering phase of LSRA.
4242 void fgSimpleLowering();
4244 GenTree* fgInitThisClass();
4246 GenTreeCall* fgGetStaticsCCtorHelper(CORINFO_CLASS_HANDLE cls, CorInfoHelpFunc helper);
4248 GenTreeCall* fgGetSharedCCtor(CORINFO_CLASS_HANDLE cls);
4250 inline bool backendRequiresLocalVarLifetimes()
4252 return !opts.MinOpts() || m_pLinearScan->willEnregisterLocalVars();
4255 void fgLocalVarLiveness();
4257 void fgLocalVarLivenessInit();
4259 void fgPerNodeLocalVarLiveness(GenTree* node);
4260 void fgPerBlockLocalVarLiveness();
4262 VARSET_VALRET_TP fgGetHandlerLiveVars(BasicBlock* block);
4264 void fgLiveVarAnalysis(bool updateInternalOnly = false);
4266 void fgComputeLifeCall(VARSET_TP& life, GenTreeCall* call);
4268 void fgComputeLifeTrackedLocalUse(VARSET_TP& life, LclVarDsc& varDsc, GenTreeLclVarCommon* node);
4269 bool fgComputeLifeTrackedLocalDef(VARSET_TP& life,
4270 VARSET_VALARG_TP keepAliveVars,
4272 GenTreeLclVarCommon* node);
4273 void fgComputeLifeUntrackedLocal(VARSET_TP& life,
4274 VARSET_VALARG_TP keepAliveVars,
4276 GenTreeLclVarCommon* lclVarNode);
4277 bool fgComputeLifeLocal(VARSET_TP& life, VARSET_VALARG_TP keepAliveVars, GenTree* lclVarNode);
4279 void fgComputeLife(VARSET_TP& life,
4282 VARSET_VALARG_TP volatileVars,
4283 bool* pStmtInfoDirty DEBUGARG(bool* treeModf));
4285 void fgComputeLifeLIR(VARSET_TP& life, BasicBlock* block, VARSET_VALARG_TP volatileVars);
4287 bool fgRemoveDeadStore(GenTree** pTree,
4289 VARSET_VALARG_TP life,
4291 bool* pStmtInfoDirty DEBUGARG(bool* treeModf));
4293 // For updating liveset during traversal AFTER fgComputeLife has completed
4294 VARSET_VALRET_TP fgGetVarBits(GenTree* tree);
4295 VARSET_VALRET_TP fgUpdateLiveSet(VARSET_VALARG_TP liveSet, GenTree* tree);
4297 // Returns the set of live variables after endTree,
4298 // assuming that liveSet is the set of live variables BEFORE tree.
4299 // Requires that fgComputeLife has completed, and that tree is in the same
4300 // statement as endTree, and that it comes before endTree in execution order
4302 VARSET_VALRET_TP fgUpdateLiveSet(VARSET_VALARG_TP liveSet, GenTree* tree, GenTree* endTree)
4304 VARSET_TP newLiveSet(VarSetOps::MakeCopy(this, liveSet));
4305 while (tree != nullptr && tree != endTree->gtNext)
4307 VarSetOps::AssignNoCopy(this, newLiveSet, fgUpdateLiveSet(newLiveSet, tree));
4308 tree = tree->gtNext;
4310 assert(tree == endTree->gtNext);
4314 void fgInterBlockLocalVarLiveness();
4316 // The presence of a partial definition presents some difficulties for SSA: this is both a use of some SSA name
4317 // of "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
4318 // whether to treat that as the use or def. It chooses the "use", and thus the old SSA name. This map allows us
4319 // to record/recover the "def" SSA number, given the lcl var node for "x" in such a tree.
4320 typedef JitHashTable<GenTree*, JitPtrKeyFuncs<GenTree>, unsigned> NodeToUnsignedMap;
4321 NodeToUnsignedMap* m_opAsgnVarDefSsaNums;
4322 NodeToUnsignedMap* GetOpAsgnVarDefSsaNums()
4324 if (m_opAsgnVarDefSsaNums == nullptr)
4326 m_opAsgnVarDefSsaNums = new (getAllocator()) NodeToUnsignedMap(getAllocator());
4328 return m_opAsgnVarDefSsaNums;
4331 // Requires value numbering phase to have completed. Returns the value number ("gtVN") of the
4332 // "tree," EXCEPT in the case of GTF_VAR_USEASG, because the tree node's gtVN member is the
4333 // "use" VN. Performs a lookup into the map of (use asg tree -> def VN.) to return the "def's"
4335 inline ValueNum GetUseAsgDefVNOrTreeVN(GenTree* tree);
4337 // Requires that "lcl" has the GTF_VAR_DEF flag set. Returns the SSA number of "lcl".
4338 // Except: assumes that lcl is a def, and if it is
4339 // a partial def (GTF_VAR_USEASG), looks up and returns the SSA number for the "def",
4340 // rather than the "use" SSA number recorded in the tree "lcl".
4341 inline unsigned GetSsaNumForLocalVarDef(GenTree* lcl);
4343 // Performs SSA conversion.
4346 // Reset any data structures to the state expected by "fgSsaBuild", so it can be run again.
4347 void fgResetForSsa();
4349 unsigned fgSsaPassesCompleted; // Number of times fgSsaBuild has been run.
4351 // Returns "true" if a struct temp of the given type requires needs zero init in this block
4352 inline bool fgStructTempNeedsExplicitZeroInit(LclVarDsc* varDsc, BasicBlock* block);
4354 // The value numbers for this compilation.
4355 ValueNumStore* vnStore;
4358 ValueNumStore* GetValueNumStore()
4363 // Do value numbering (assign a value number to each
4365 void fgValueNumber();
4367 // Computes new GcHeap VN via the assignment H[elemTypeEq][arrVN][inx][fldSeq] = rhsVN.
4368 // Assumes that "elemTypeEq" is the (equivalence class rep) of the array element type.
4369 // The 'indType' is the indirection type of the lhs of the assignment and will typically
4370 // match the element type of the array or fldSeq. When this type doesn't match
4371 // or if the fldSeq is 'NotAField' we invalidate the array contents H[elemTypeEq][arrVN]
4373 ValueNum fgValueNumberArrIndexAssign(CORINFO_CLASS_HANDLE elemTypeEq,
4376 FieldSeqNode* fldSeq,
4380 // Requires that "tree" is a GT_IND marked as an array index, and that its address argument
4381 // has been parsed to yield the other input arguments. If evaluation of the address
4382 // can raise exceptions, those should be captured in the exception set "excVN."
4383 // Assumes that "elemTypeEq" is the (equivalence class rep) of the array element type.
4384 // Marks "tree" with the VN for H[elemTypeEq][arrVN][inx][fldSeq] (for the liberal VN; a new unique
4385 // VN for the conservative VN.) Also marks the tree's argument as the address of an array element.
4386 // The type tree->TypeGet() will typically match the element type of the array or fldSeq.
4387 // When this type doesn't match or if the fldSeq is 'NotAField' we return a new unique VN
4389 ValueNum fgValueNumberArrIndexVal(GenTree* tree,
4390 CORINFO_CLASS_HANDLE elemTypeEq,
4394 FieldSeqNode* fldSeq);
4396 // Requires "funcApp" to be a VNF_PtrToArrElem, and "addrXvn" to represent the exception set thrown
4397 // by evaluating the array index expression "tree". Returns the value number resulting from
4398 // dereferencing the array in the current GcHeap state. If "tree" is non-null, it must be the
4399 // "GT_IND" that does the dereference, and it is given the returned value number.
4400 ValueNum fgValueNumberArrIndexVal(GenTree* tree, struct VNFuncApp* funcApp, ValueNum addrXvn);
4402 // Compute the value number for a byref-exposed load of the given type via the given pointerVN.
4403 ValueNum fgValueNumberByrefExposedLoad(var_types type, ValueNum pointerVN);
4405 unsigned fgVNPassesCompleted; // Number of times fgValueNumber has been run.
4407 // Utility functions for fgValueNumber.
4409 // Perform value-numbering for the trees in "blk".
4410 void fgValueNumberBlock(BasicBlock* blk);
4412 // Requires that "entryBlock" is the entry block of loop "loopNum", and that "loopNum" is the
4413 // innermost loop of which "entryBlock" is the entry. Returns the value number that should be
4414 // assumed for the memoryKind at the start "entryBlk".
4415 ValueNum fgMemoryVNForLoopSideEffects(MemoryKind memoryKind, BasicBlock* entryBlock, unsigned loopNum);
4417 // Called when an operation (performed by "tree", described by "msg") may cause the GcHeap to be mutated.
4418 // As GcHeap is a subset of ByrefExposed, this will also annotate the ByrefExposed mutation.
4419 void fgMutateGcHeap(GenTree* tree DEBUGARG(const char* msg));
4421 // Called when an operation (performed by "tree", described by "msg") may cause an address-exposed local to be
4423 void fgMutateAddressExposedLocal(GenTree* tree DEBUGARG(const char* msg));
4425 // For a GC heap store at curTree, record the new curMemoryVN's and update curTree's MemorySsaMap.
4426 // As GcHeap is a subset of ByrefExposed, this will also record the ByrefExposed store.
4427 void recordGcHeapStore(GenTree* curTree, ValueNum gcHeapVN DEBUGARG(const char* msg));
4429 // For a store to an address-exposed local at curTree, record the new curMemoryVN and update curTree's MemorySsaMap.
4430 void recordAddressExposedLocalStore(GenTree* curTree, ValueNum memoryVN DEBUGARG(const char* msg));
4432 // Tree caused an update in the current memory VN. If "tree" has an associated heap SSA #, record that
4433 // value in that SSA #.
4434 void fgValueNumberRecordMemorySsa(MemoryKind memoryKind, GenTree* tree);
4436 // The input 'tree' is a leaf node that is a constant
4437 // Assign the proper value number to the tree
4438 void fgValueNumberTreeConst(GenTree* tree);
4440 // Assumes that all inputs to "tree" have had value numbers assigned; assigns a VN to tree.
4441 // (With some exceptions: the VN of the lhs of an assignment is assigned as part of the
4443 void fgValueNumberTree(GenTree* tree);
4445 // Does value-numbering for a block assignment.
4446 void fgValueNumberBlockAssignment(GenTree* tree);
4448 // Does value-numbering for a cast tree.
4449 void fgValueNumberCastTree(GenTree* tree);
4451 // Does value-numbering for an intrinsic tree.
4452 void fgValueNumberIntrinsic(GenTree* tree);
4454 // Does value-numbering for a call. We interpret some helper calls.
4455 void fgValueNumberCall(GenTreeCall* call);
4457 // The VN of some nodes in "args" may have changed -- reassign VNs to the arg list nodes.
4458 void fgUpdateArgListVNs(GenTreeArgList* args);
4460 // Does value-numbering for a helper "call" that has a VN function symbol "vnf".
4461 void fgValueNumberHelperCallFunc(GenTreeCall* call, VNFunc vnf, ValueNumPair vnpExc);
4463 // Requires "helpCall" to be a helper call. Assigns it a value number;
4464 // we understand the semantics of some of the calls. Returns "true" if
4465 // the call may modify the heap (we assume arbitrary memory side effects if so).
4466 bool fgValueNumberHelperCall(GenTreeCall* helpCall);
4468 // Requires that "helpFunc" is one of the pure Jit Helper methods.
4469 // Returns the corresponding VNFunc to use for value numbering
4470 VNFunc fgValueNumberJitHelperMethodVNFunc(CorInfoHelpFunc helpFunc);
4472 // Adds the exception set for the current tree node which has a memory indirection operation
4473 void fgValueNumberAddExceptionSetForIndirection(GenTree* tree, GenTree* baseAddr);
4475 // Adds the exception sets for the current tree node which is performing a division or modulus operation
4476 void fgValueNumberAddExceptionSetForDivision(GenTree* tree);
4478 // Adds the exception set for the current tree node which is performing a overflow checking operation
4479 void fgValueNumberAddExceptionSetForOverflow(GenTree* tree);
4481 // Adds the exception set for the current tree node which is performing a ckfinite operation
4482 void fgValueNumberAddExceptionSetForCkFinite(GenTree* tree);
4484 // Adds the exception sets for the current tree node
4485 void fgValueNumberAddExceptionSet(GenTree* tree);
4487 // These are the current value number for the memory implicit variables while
4488 // doing value numbering. These are the value numbers under the "liberal" interpretation
4489 // of memory values; the "conservative" interpretation needs no VN, since every access of
4490 // memory yields an unknown value.
4491 ValueNum fgCurMemoryVN[MemoryKindCount];
4493 // Return a "pseudo"-class handle for an array element type. If "elemType" is TYP_STRUCT,
4494 // requires "elemStructType" to be non-null (and to have a low-order zero). Otherwise, low order bit
4495 // is 1, and the rest is an encoding of "elemTyp".
4496 static CORINFO_CLASS_HANDLE EncodeElemType(var_types elemTyp, CORINFO_CLASS_HANDLE elemStructType)
4498 if (elemStructType != nullptr)
4500 assert(varTypeIsStruct(elemTyp) || elemTyp == TYP_REF || elemTyp == TYP_BYREF ||
4501 varTypeIsIntegral(elemTyp));
4502 assert((size_t(elemStructType) & 0x1) == 0x0); // Make sure the encoding below is valid.
4503 return elemStructType;
4507 elemTyp = varTypeUnsignedToSigned(elemTyp);
4508 return CORINFO_CLASS_HANDLE(size_t(elemTyp) << 1 | 0x1);
4511 // If "clsHnd" is the result of an "EncodePrim" call, returns true and sets "*pPrimType" to the
4512 // var_types it represents. Otherwise, returns TYP_STRUCT (on the assumption that "clsHnd" is
4513 // the struct type of the element).
4514 static var_types DecodeElemType(CORINFO_CLASS_HANDLE clsHnd)
4516 size_t clsHndVal = size_t(clsHnd);
4517 if (clsHndVal & 0x1)
4519 return var_types(clsHndVal >> 1);
4527 // Convert a BYTE which represents the VM's CorInfoGCtype to the JIT's var_types
4528 var_types getJitGCType(BYTE gcType);
4530 enum structPassingKind
4532 SPK_Unknown, // Invalid value, never returned
4533 SPK_PrimitiveType, // The struct is passed/returned using a primitive type.
4534 SPK_EnclosingType, // Like SPK_Primitive type, but used for return types that
4535 // require a primitive type temp that is larger than the struct size.
4536 // Currently used for structs of size 3, 5, 6, or 7 bytes.
4537 SPK_ByValue, // The struct is passed/returned by value (using the ABI rules)
4538 // for ARM64 and UNIX_X64 in multiple registers. (when all of the
4539 // parameters registers are used, then the stack will be used)
4540 // for X86 passed on the stack, for ARM32 passed in registers
4541 // or the stack or split between registers and the stack.
4542 SPK_ByValueAsHfa, // The struct is passed/returned as an HFA in multiple registers.
4544 }; // The struct is passed/returned by reference to a copy/buffer.
4546 // Get the "primitive" type that is is used when we are given a struct of size 'structSize'.
4547 // For pointer sized structs the 'clsHnd' is used to determine if the struct contains GC ref.
4548 // A "primitive" type is one of the scalar types: byte, short, int, long, ref, float, double
4549 // If we can't or shouldn't use a "primitive" type then TYP_UNKNOWN is returned.
4551 // isVarArg is passed for use on Windows Arm64 to change the decision returned regarding
4554 var_types getPrimitiveTypeForStruct(unsigned structSize, CORINFO_CLASS_HANDLE clsHnd, bool isVarArg);
4556 // Get the type that is used to pass values of the given struct type.
4557 // isVarArg is passed for use on Windows Arm64 to change the decision returned regarding
4560 var_types getArgTypeForStruct(CORINFO_CLASS_HANDLE clsHnd,
4561 structPassingKind* wbPassStruct,
4563 unsigned structSize);
4565 // Get the type that is used to return values of the given struct type.
4566 // If the size is unknown, pass 0 and it will be determined from 'clsHnd'.
4567 var_types getReturnTypeForStruct(CORINFO_CLASS_HANDLE clsHnd,
4568 structPassingKind* wbPassStruct = nullptr,
4569 unsigned structSize = 0);
4572 // Print a representation of "vnp" or "vn" on standard output.
4573 // If "level" is non-zero, we also print out a partial expansion of the value.
4574 void vnpPrint(ValueNumPair vnp, unsigned level);
4575 void vnPrint(ValueNum vn, unsigned level);
4578 bool fgDominate(BasicBlock* b1, BasicBlock* b2); // Return true if b1 dominates b2
4580 // Dominator computation member functions
4581 // Not exposed outside Compiler
4583 bool fgReachable(BasicBlock* b1, BasicBlock* b2); // Returns true if block b1 can reach block b2
4585 void fgComputeDoms(); // Computes the immediate dominators for each basic block in the
4586 // flow graph. We first assume the fields bbIDom on each
4587 // basic block are invalid. This computation is needed later
4588 // by fgBuildDomTree to build the dominance tree structure.
4589 // Based on: A Simple, Fast Dominance Algorithm
4590 // by Keith D. Cooper, Timothy J. Harvey, and Ken Kennedy
4592 void fgCompDominatedByExceptionalEntryBlocks();
4594 BlockSet_ValRet_T fgGetDominatorSet(BasicBlock* block); // Returns a set of blocks that dominate the given block.
4595 // Note: this is relatively slow compared to calling fgDominate(),
4596 // especially if dealing with a single block versus block check.
4598 void fgComputeReachabilitySets(); // Compute bbReach sets. (Also sets BBF_GC_SAFE_POINT flag on blocks.)
4600 void fgComputeEnterBlocksSet(); // Compute the set of entry blocks, 'fgEnterBlks'.
4602 bool fgRemoveUnreachableBlocks(); // Remove blocks determined to be unreachable by the bbReach sets.
4604 void fgComputeReachability(); // Perform flow graph node reachability analysis.
4606 BasicBlock* fgIntersectDom(BasicBlock* a, BasicBlock* b); // Intersect two immediate dominator sets.
4608 void fgDfsInvPostOrder(); // In order to compute dominance using fgIntersectDom, the flow graph nodes must be
4609 // processed in topological sort, this function takes care of that.
4611 void fgDfsInvPostOrderHelper(BasicBlock* block, BlockSet& visited, unsigned* count);
4613 BlockSet_ValRet_T fgDomFindStartNodes(); // Computes which basic blocks don't have incoming edges in the flow graph.
4614 // Returns this as a set.
4616 BlockSet_ValRet_T fgDomTreeEntryNodes(BasicBlockList** domTree); // Computes which nodes in the dominance forest are
4617 // root nodes. Returns this as a set.
4620 void fgDispDomTree(BasicBlockList** domTree); // Helper that prints out the Dominator Tree in debug builds.
4623 void fgBuildDomTree(); // Once we compute all the immediate dominator sets for each node in the flow graph
4624 // (performed by fgComputeDoms), this procedure builds the dominance tree represented
4627 // In order to speed up the queries of the form 'Does A dominates B', we can perform a DFS preorder and postorder
4628 // traversal of the dominance tree and the dominance query will become A dominates B iif preOrder(A) <= preOrder(B)
4629 // && postOrder(A) >= postOrder(B) making the computation O(1).
4630 void fgTraverseDomTree(unsigned bbNum, BasicBlockList** domTree, unsigned* preNum, unsigned* postNum);
4632 // When the flow graph changes, we need to update the block numbers, predecessor lists, reachability sets, and
4634 void fgUpdateChangedFlowGraph();
4637 // Compute the predecessors of the blocks in the control flow graph.
4638 void fgComputePreds();
4640 // Remove all predecessor information.
4641 void fgRemovePreds();
4643 // Compute the cheap flow graph predecessors lists. This is used in some early phases
4644 // before the full predecessors lists are computed.
4645 void fgComputeCheapPreds();
4648 void fgAddCheapPred(BasicBlock* block, BasicBlock* blockPred);
4650 void fgRemoveCheapPred(BasicBlock* block, BasicBlock* blockPred);
4660 // Initialize the per-block variable sets (used for liveness analysis).
4661 void fgInitBlockVarSets();
4663 // true if we've gone through and created GC Poll calls.
4664 bool fgGCPollsCreated;
4665 void fgMarkGCPollBlocks();
4666 void fgCreateGCPolls();
4667 bool fgCreateGCPoll(GCPollType pollType, BasicBlock* block);
4669 // Requires that "block" is a block that returns from
4670 // a finally. Returns the number of successors (jump targets of
4671 // of blocks in the covered "try" that did a "LEAVE".)
4672 unsigned fgNSuccsOfFinallyRet(BasicBlock* block);
4674 // Requires that "block" is a block that returns (in the sense of BBJ_EHFINALLYRET) from
4675 // a finally. Returns its "i"th successor (jump targets of
4676 // of blocks in the covered "try" that did a "LEAVE".)
4677 // Requires that "i" < fgNSuccsOfFinallyRet(block).
4678 BasicBlock* fgSuccOfFinallyRet(BasicBlock* block, unsigned i);
4681 // Factor out common portions of the impls of the methods above.
4682 void fgSuccOfFinallyRetWork(BasicBlock* block, unsigned i, BasicBlock** bres, unsigned* nres);
4685 // For many purposes, it is desirable to be able to enumerate the *distinct* targets of a switch statement,
4686 // skipping duplicate targets. (E.g., in flow analyses that are only interested in the set of possible targets.)
4687 // SwitchUniqueSuccSet contains the non-duplicated switch targets.
4688 // (Code that modifies the jump table of a switch has an obligation to call Compiler::UpdateSwitchTableTarget,
4689 // which in turn will call the "UpdateTarget" method of this type if a SwitchUniqueSuccSet has already
4690 // been computed for the switch block. If a switch block is deleted or is transformed into a non-switch,
4691 // we leave the entry associated with the block, but it will no longer be accessed.)
4692 struct SwitchUniqueSuccSet
4694 unsigned numDistinctSuccs; // Number of distinct targets of the switch.
4695 BasicBlock** nonDuplicates; // Array of "numDistinctSuccs", containing all the distinct switch target
4698 // The switch block "switchBlk" just had an entry with value "from" modified to the value "to".
4699 // Update "this" as necessary: if "from" is no longer an element of the jump table of "switchBlk",
4700 // remove it from "this", and ensure that "to" is a member. Use "alloc" to do any required allocation.
4701 void UpdateTarget(CompAllocator alloc, BasicBlock* switchBlk, BasicBlock* from, BasicBlock* to);
4704 typedef JitHashTable<BasicBlock*, JitPtrKeyFuncs<BasicBlock>, SwitchUniqueSuccSet> BlockToSwitchDescMap;
4707 // Maps BasicBlock*'s that end in switch statements to SwitchUniqueSuccSets that allow
4708 // iteration over only the distinct successors.
4709 BlockToSwitchDescMap* m_switchDescMap;
4712 BlockToSwitchDescMap* GetSwitchDescMap(bool createIfNull = true)
4714 if ((m_switchDescMap == nullptr) && createIfNull)
4716 m_switchDescMap = new (getAllocator()) BlockToSwitchDescMap(getAllocator());
4718 return m_switchDescMap;
4721 // Invalidate the map of unique switch block successors. For example, since the hash key of the map
4722 // depends on block numbers, we must invalidate the map when the blocks are renumbered, to ensure that
4723 // we don't accidentally look up and return the wrong switch data.
4724 void InvalidateUniqueSwitchSuccMap()
4726 m_switchDescMap = nullptr;
4729 // Requires "switchBlock" to be a block that ends in a switch. Returns
4730 // the corresponding SwitchUniqueSuccSet.
4731 SwitchUniqueSuccSet GetDescriptorForSwitch(BasicBlock* switchBlk);
4733 // The switch block "switchBlk" just had an entry with value "from" modified to the value "to".
4734 // Update "this" as necessary: if "from" is no longer an element of the jump table of "switchBlk",
4735 // remove it from "this", and ensure that "to" is a member.
4736 void UpdateSwitchTableTarget(BasicBlock* switchBlk, BasicBlock* from, BasicBlock* to);
4738 // Remove the "SwitchUniqueSuccSet" of "switchBlk" in the BlockToSwitchDescMap.
4739 void fgInvalidateSwitchDescMapEntry(BasicBlock* switchBlk);
4741 BasicBlock* fgFirstBlockOfHandler(BasicBlock* block);
4743 flowList* fgGetPredForBlock(BasicBlock* block, BasicBlock* blockPred);
4745 flowList* fgGetPredForBlock(BasicBlock* block, BasicBlock* blockPred, flowList*** ptrToPred);
4747 flowList* fgSpliceOutPred(BasicBlock* block, BasicBlock* blockPred);
4749 flowList* fgRemoveRefPred(BasicBlock* block, BasicBlock* blockPred);
4751 flowList* fgRemoveAllRefPreds(BasicBlock* block, BasicBlock* blockPred);
4753 flowList* fgRemoveAllRefPreds(BasicBlock* block, flowList** ptrToPred);
4755 void fgRemoveBlockAsPred(BasicBlock* block);
4757 void fgChangeSwitchBlock(BasicBlock* oldSwitchBlock, BasicBlock* newSwitchBlock);
4759 void fgReplaceSwitchJumpTarget(BasicBlock* blockSwitch, BasicBlock* newTarget, BasicBlock* oldTarget);
4761 void fgReplaceJumpTarget(BasicBlock* block, BasicBlock* newTarget, BasicBlock* oldTarget);
4763 void fgReplacePred(BasicBlock* block, BasicBlock* oldPred, BasicBlock* newPred);
4765 flowList* fgAddRefPred(BasicBlock* block,
4766 BasicBlock* blockPred,
4767 flowList* oldEdge = nullptr,
4768 bool initializingPreds = false); // Only set to 'true' when we are computing preds in
4771 void fgFindBasicBlocks();
4773 bool fgIsBetterFallThrough(BasicBlock* bCur, BasicBlock* bAlt);
4775 bool fgCheckEHCanInsertAfterBlock(BasicBlock* blk, unsigned regionIndex, bool putInTryRegion);
4777 BasicBlock* fgFindInsertPoint(unsigned regionIndex,
4778 bool putInTryRegion,
4779 BasicBlock* startBlk,
4781 BasicBlock* nearBlk,
4782 BasicBlock* jumpBlk,
4785 unsigned fgGetNestingLevel(BasicBlock* block, unsigned* pFinallyNesting = nullptr);
4787 void fgRemoveEmptyBlocks();
4789 void fgRemoveStmt(BasicBlock* block, GenTree* stmt);
4791 bool fgCheckRemoveStmt(BasicBlock* block, GenTree* stmt);
4793 void fgCreateLoopPreHeader(unsigned lnum);
4795 void fgUnreachableBlock(BasicBlock* block);
4797 void fgRemoveConditionalJump(BasicBlock* block);
4799 BasicBlock* fgLastBBInMainFunction();
4801 BasicBlock* fgEndBBAfterMainFunction();
4803 void fgUnlinkRange(BasicBlock* bBeg, BasicBlock* bEnd);
4805 void fgRemoveBlock(BasicBlock* block, bool unreachable);
4807 bool fgCanCompactBlocks(BasicBlock* block, BasicBlock* bNext);
4809 void fgCompactBlocks(BasicBlock* block, BasicBlock* bNext);
4811 void fgUpdateLoopsAfterCompacting(BasicBlock* block, BasicBlock* bNext);
4813 BasicBlock* fgConnectFallThrough(BasicBlock* bSrc, BasicBlock* bDst);
4815 bool fgRenumberBlocks();
4817 bool fgExpandRarelyRunBlocks();
4819 bool fgEhAllowsMoveBlock(BasicBlock* bBefore, BasicBlock* bAfter);
4821 void fgMoveBlocksAfter(BasicBlock* bStart, BasicBlock* bEnd, BasicBlock* insertAfterBlk);
4823 enum FG_RELOCATE_TYPE
4825 FG_RELOCATE_TRY, // relocate the 'try' region
4826 FG_RELOCATE_HANDLER // relocate the handler region (including the filter if necessary)
4828 BasicBlock* fgRelocateEHRange(unsigned regionIndex, FG_RELOCATE_TYPE relocateType);
4830 #if FEATURE_EH_FUNCLETS
4831 #if defined(_TARGET_ARM_)
4832 void fgClearFinallyTargetBit(BasicBlock* block);
4833 #endif // defined(_TARGET_ARM_)
4834 bool fgIsIntraHandlerPred(BasicBlock* predBlock, BasicBlock* block);
4835 bool fgAnyIntraHandlerPreds(BasicBlock* block);
4836 void fgInsertFuncletPrologBlock(BasicBlock* block);
4837 void fgCreateFuncletPrologBlocks();
4838 void fgCreateFunclets();
4839 #else // !FEATURE_EH_FUNCLETS
4840 bool fgRelocateEHRegions();
4841 #endif // !FEATURE_EH_FUNCLETS
4843 bool fgOptimizeUncondBranchToSimpleCond(BasicBlock* block, BasicBlock* target);
4845 bool fgBlockEndFavorsTailDuplication(BasicBlock* block);
4847 bool fgBlockIsGoodTailDuplicationCandidate(BasicBlock* block);
4849 bool fgOptimizeEmptyBlock(BasicBlock* block);
4851 bool fgOptimizeBranchToEmptyUnconditional(BasicBlock* block, BasicBlock* bDest);
4853 bool fgOptimizeBranch(BasicBlock* bJump);
4855 bool fgOptimizeSwitchBranches(BasicBlock* block);
4857 bool fgOptimizeBranchToNext(BasicBlock* block, BasicBlock* bNext, BasicBlock* bPrev);
4859 bool fgOptimizeSwitchJumps();
4861 void fgPrintEdgeWeights();
4863 void fgComputeBlockAndEdgeWeights();
4864 BasicBlock::weight_t fgComputeMissingBlockWeights();
4865 void fgComputeCalledCount(BasicBlock::weight_t returnWeight);
4866 void fgComputeEdgeWeights();
4868 void fgReorderBlocks();
4870 void fgDetermineFirstColdBlock();
4872 bool fgIsForwardBranch(BasicBlock* bJump, BasicBlock* bSrc = nullptr);
4874 bool fgUpdateFlowGraph(bool doTailDup = false);
4876 void fgFindOperOrder();
4878 // method that returns if you should split here
4879 typedef bool(fgSplitPredicate)(GenTree* tree, GenTree* parent, fgWalkData* data);
4881 void fgSetBlockOrder();
4883 void fgRemoveReturnBlock(BasicBlock* block);
4885 /* Helper code that has been factored out */
4886 inline void fgConvertBBToThrowBB(BasicBlock* block);
4888 bool fgCastNeeded(GenTree* tree, var_types toType);
4889 GenTree* fgDoNormalizeOnStore(GenTree* tree);
4890 GenTree* fgMakeTmpArgNode(fgArgTabEntry* curArgTabEntry);
4892 // The following check for loops that don't execute calls
4893 bool fgLoopCallMarked;
4895 void fgLoopCallTest(BasicBlock* srcBB, BasicBlock* dstBB);
4896 void fgLoopCallMark();
4898 void fgMarkLoopHead(BasicBlock* block);
4900 unsigned fgGetCodeEstimate(BasicBlock* block);
4903 const char* fgProcessEscapes(const char* nameIn, escapeMapping_t* map);
4904 FILE* fgOpenFlowGraphFile(bool* wbDontClose, Phases phase, LPCWSTR type);
4905 bool fgDumpFlowGraph(Phases phase);
4907 #endif // DUMP_FLOWGRAPHS
4912 void fgDispBBLiveness(BasicBlock* block);
4913 void fgDispBBLiveness();
4914 void fgTableDispBasicBlock(BasicBlock* block, int ibcColWidth = 0);
4915 void fgDispBasicBlocks(BasicBlock* firstBlock, BasicBlock* lastBlock, bool dumpTrees);
4916 void fgDispBasicBlocks(bool dumpTrees = false);
4917 void fgDumpStmtTree(GenTree* stmt, unsigned bbNum);
4918 void fgDumpBlock(BasicBlock* block);
4919 void fgDumpTrees(BasicBlock* firstBlock, BasicBlock* lastBlock);
4921 static fgWalkPreFn fgStress64RsltMulCB;
4922 void fgStress64RsltMul();
4923 void fgDebugCheckUpdate();
4924 void fgDebugCheckBBlist(bool checkBBNum = false, bool checkBBRefs = true);
4925 void fgDebugCheckBlockLinks();
4926 void fgDebugCheckLinks(bool morphTrees = false);
4927 void fgDebugCheckStmtsList(BasicBlock* block, bool morphTrees);
4928 void fgDebugCheckNodeLinks(BasicBlock* block, GenTree* stmt);
4929 void fgDebugCheckNodesUniqueness();
4931 void fgDebugCheckFlags(GenTree* tree);
4932 void fgDebugCheckFlagsHelper(GenTree* tree, unsigned treeFlags, unsigned chkFlags);
4933 void fgDebugCheckTryFinallyExits();
4936 static GenTree* fgGetFirstNode(GenTree* tree);
4938 //--------------------- Walking the trees in the IR -----------------------
4943 fgWalkPreFn* wtprVisitorFn;
4944 fgWalkPostFn* wtpoVisitorFn;
4945 void* pCallbackData; // user-provided data
4946 bool wtprLclsOnly; // whether to only visit lclvar nodes
4947 GenTree* parent; // parent of current node, provided to callback
4948 GenTreeStack* parentStack; // stack of parent nodes, if asked for
4950 bool printModified; // callback can use this
4954 fgWalkResult fgWalkTreePre(GenTree** pTree,
4955 fgWalkPreFn* visitor,
4956 void* pCallBackData = nullptr,
4957 bool lclVarsOnly = false,
4958 bool computeStack = false);
4960 fgWalkResult fgWalkTree(GenTree** pTree,
4961 fgWalkPreFn* preVisitor,
4962 fgWalkPostFn* postVisitor,
4963 void* pCallBackData = nullptr);
4965 void fgWalkAllTreesPre(fgWalkPreFn* visitor, void* pCallBackData);
4969 fgWalkResult fgWalkTreePost(GenTree** pTree,
4970 fgWalkPostFn* visitor,
4971 void* pCallBackData = nullptr,
4972 bool computeStack = false);
4974 // An fgWalkPreFn that looks for expressions that have inline throws in
4975 // minopts mode. Basically it looks for tress with gtOverflowEx() or
4976 // GTF_IND_RNGCHK. It returns WALK_ABORT if one is found. It
4977 // returns WALK_SKIP_SUBTREES if GTF_EXCEPT is not set (assumes flags
4978 // properly propagated to parent trees). It returns WALK_CONTINUE
4980 static fgWalkResult fgChkThrowCB(GenTree** pTree, Compiler::fgWalkData* data);
4981 static fgWalkResult fgChkLocAllocCB(GenTree** pTree, Compiler::fgWalkData* data);
4982 static fgWalkResult fgChkQmarkCB(GenTree** pTree, Compiler::fgWalkData* data);
4984 /**************************************************************************
4986 *************************************************************************/
4989 friend class SsaBuilder;
4990 friend struct ValueNumberState;
4992 //--------------------- Detect the basic blocks ---------------------------
4994 BasicBlock** fgBBs; // Table of pointers to the BBs
4996 void fgInitBBLookup();
4997 BasicBlock* fgLookupBB(unsigned addr);
4999 void fgFindJumpTargets(const BYTE* codeAddr, IL_OFFSET codeSize, FixedBitVect* jumpTarget);
5001 void fgMarkBackwardJump(BasicBlock* startBlock, BasicBlock* endBlock);
5003 void fgLinkBasicBlocks();
5005 unsigned fgMakeBasicBlocks(const BYTE* codeAddr, IL_OFFSET codeSize, FixedBitVect* jumpTarget);
5007 void fgCheckBasicBlockControlFlow();
5009 void fgControlFlowPermitted(BasicBlock* blkSrc,
5010 BasicBlock* blkDest,
5011 BOOL IsLeave = false /* is the src a leave block */);
5013 bool fgFlowToFirstBlockOfInnerTry(BasicBlock* blkSrc, BasicBlock* blkDest, bool sibling);
5015 void fgObserveInlineConstants(OPCODE opcode, const FgStack& stack, bool isInlining);
5017 void fgAdjustForAddressExposedOrWrittenThis();
5019 bool fgProfileData_ILSizeMismatch;
5020 ICorJitInfo::ProfileBuffer* fgProfileBuffer;
5021 ULONG fgProfileBufferCount;
5022 ULONG fgNumProfileRuns;
5024 unsigned fgStressBBProf()
5027 unsigned result = JitConfig.JitStressBBProf();
5030 if (compStressCompile(STRESS_BB_PROFILE, 15))
5041 bool fgHaveProfileData();
5042 bool fgGetProfileWeightForBasicBlock(IL_OFFSET offset, unsigned* weight);
5043 void fgInstrumentMethod();
5046 // fgIsUsingProfileWeights - returns true if we have real profile data for this method
5047 // or if we have some fake profile data for the stress mode
5048 bool fgIsUsingProfileWeights()
5050 return (fgHaveProfileData() || fgStressBBProf());
5053 // fgProfileRunsCount - returns total number of scenario runs for the profile data
5054 // or BB_UNITY_WEIGHT when we aren't using profile data.
5055 unsigned fgProfileRunsCount()
5057 return fgIsUsingProfileWeights() ? fgNumProfileRuns : BB_UNITY_WEIGHT;
5060 //-------- Insert a statement at the start or end of a basic block --------
5064 static bool fgBlockContainsStatementBounded(BasicBlock* block, GenTree* stmt, bool answerOnBoundExceeded = true);
5068 GenTreeStmt* fgInsertStmtAtEnd(BasicBlock* block, GenTree* node);
5070 public: // Used by linear scan register allocation
5071 GenTreeStmt* fgInsertStmtNearEnd(BasicBlock* block, GenTree* node);
5074 GenTree* fgInsertStmtAtBeg(BasicBlock* block, GenTree* stmt);
5075 GenTree* fgInsertStmtAfter(BasicBlock* block, GenTree* insertionPoint, GenTree* stmt);
5077 public: // Used by linear scan register allocation
5078 GenTree* fgInsertStmtBefore(BasicBlock* block, GenTree* insertionPoint, GenTree* stmt);
5081 GenTree* fgInsertStmtListAfter(BasicBlock* block, GenTree* stmtAfter, GenTree* stmtList);
5083 // Create a new temporary variable to hold the result of *ppTree,
5084 // and transform the graph accordingly.
5085 GenTree* fgInsertCommaFormTemp(GenTree** ppTree, CORINFO_CLASS_HANDLE structType = nullptr);
5086 GenTree* fgMakeMultiUse(GenTree** ppTree);
5089 // Recognize a bitwise rotation pattern and convert into a GT_ROL or a GT_ROR node.
5090 GenTree* fgRecognizeAndMorphBitwiseRotation(GenTree* tree);
5091 bool fgOperIsBitwiseRotationRoot(genTreeOps oper);
5093 //-------- Determine the order in which the trees will be evaluated -------
5095 unsigned fgTreeSeqNum;
5096 GenTree* fgTreeSeqLst;
5097 GenTree* fgTreeSeqBeg;
5099 GenTree* fgSetTreeSeq(GenTree* tree, GenTree* prev = nullptr, bool isLIR = false);
5100 void fgSetTreeSeqHelper(GenTree* tree, bool isLIR);
5101 void fgSetTreeSeqFinish(GenTree* tree, bool isLIR);
5102 void fgSetStmtSeq(GenTree* tree);
5103 void fgSetBlockOrder(BasicBlock* block);
5105 //------------------------- Morphing --------------------------------------
5107 unsigned fgPtrArgCntMax;
5110 //------------------------------------------------------------------------
5111 // fgGetPtrArgCntMax: Return the maximum number of pointer-sized stack arguments that calls inside this method
5112 // can push on the stack. This value is calculated during morph.
5115 // Returns fgPtrArgCntMax, that is a private field.
5117 unsigned fgGetPtrArgCntMax() const
5119 return fgPtrArgCntMax;
5122 //------------------------------------------------------------------------
5123 // fgSetPtrArgCntMax: Set the maximum number of pointer-sized stack arguments that calls inside this method
5124 // can push on the stack. This function is used during StackLevelSetter to fix incorrect morph calculations.
5126 void fgSetPtrArgCntMax(unsigned argCntMax)
5128 fgPtrArgCntMax = argCntMax;
5131 bool compCanEncodePtrArgCntMax();
5134 hashBv* fgOutgoingArgTemps;
5135 hashBv* fgCurrentlyInUseArgTemps;
5137 void fgSetRngChkTarget(GenTree* tree, bool delay = true);
5139 BasicBlock* fgSetRngChkTargetInner(SpecialCodeKind kind, bool delay);
5142 void fgMoveOpsLeft(GenTree* tree);
5145 bool fgIsCommaThrow(GenTree* tree, bool forFolding = false);
5147 bool fgIsThrow(GenTree* tree);
5149 bool fgInDifferentRegions(BasicBlock* blk1, BasicBlock* blk2);
5150 bool fgIsBlockCold(BasicBlock* block);
5152 GenTree* fgMorphCastIntoHelper(GenTree* tree, int helper, GenTree* oper);
5154 GenTree* fgMorphIntoHelperCall(GenTree* tree, int helper, GenTreeArgList* args, bool morphArgs = true);
5156 GenTree* fgMorphStackArgForVarArgs(unsigned lclNum, var_types varType, unsigned lclOffs);
5158 // A "MorphAddrContext" carries information from the surrounding context. If we are evaluating a byref address,
5159 // it is useful to know whether the address will be immediately dereferenced, or whether the address value will
5160 // be used, perhaps by passing it as an argument to a called method. This affects how null checking is done:
5161 // for sufficiently small offsets, we can rely on OS page protection to implicitly null-check addresses that we
5162 // know will be dereferenced. To know that reliance on implicit null checking is sound, we must further know that
5163 // all offsets between the top-level indirection and the bottom are constant, and that their sum is sufficiently
5164 // small; hence the other fields of MorphAddrContext.
5165 enum MorphAddrContextKind
5170 struct MorphAddrContext
5172 MorphAddrContextKind m_kind;
5173 bool m_allConstantOffsets; // Valid only for "m_kind == MACK_Ind". True iff all offsets between
5174 // top-level indirection and here have been constants.
5175 size_t m_totalOffset; // Valid only for "m_kind == MACK_Ind", and if "m_allConstantOffsets" is true.
5176 // In that case, is the sum of those constant offsets.
5178 MorphAddrContext(MorphAddrContextKind kind) : m_kind(kind), m_allConstantOffsets(true), m_totalOffset(0)
5183 // A MACK_CopyBlock context is immutable, so we can just make one of these and share it.
5184 static MorphAddrContext s_CopyBlockMAC;
5187 GenTree* getSIMDStructFromField(GenTree* tree,
5188 var_types* baseTypeOut,
5190 unsigned* simdSizeOut,
5191 bool ignoreUsedInSIMDIntrinsic = false);
5192 GenTree* fgMorphFieldAssignToSIMDIntrinsicSet(GenTree* tree);
5193 GenTree* fgMorphFieldToSIMDIntrinsicGet(GenTree* tree);
5194 bool fgMorphCombineSIMDFieldAssignments(BasicBlock* block, GenTree* stmt);
5195 void impMarkContiguousSIMDFieldAssignments(GenTree* stmt);
5197 // fgPreviousCandidateSIMDFieldAsgStmt is only used for tracking previous simd field assignment
5198 // in function: Complier::impMarkContiguousSIMDFieldAssignments.
5199 GenTree* fgPreviousCandidateSIMDFieldAsgStmt;
5201 #endif // FEATURE_SIMD
5202 GenTree* fgMorphArrayIndex(GenTree* tree);
5203 GenTree* fgMorphCast(GenTree* tree);
5204 GenTree* fgUnwrapProxy(GenTree* objRef);
5205 GenTreeFieldList* fgMorphLclArgToFieldlist(GenTreeLclVarCommon* lcl);
5206 void fgInitArgInfo(GenTreeCall* call);
5207 GenTreeCall* fgMorphArgs(GenTreeCall* call);
5208 GenTreeArgList* fgMorphArgList(GenTreeArgList* args, MorphAddrContext* mac);
5210 void fgMakeOutgoingStructArgCopy(GenTreeCall* call,
5213 CORINFO_CLASS_HANDLE copyBlkClass);
5215 void fgFixupStructReturn(GenTree* call);
5216 GenTree* fgMorphLocalVar(GenTree* tree, bool forceRemorph);
5219 bool fgAddrCouldBeNull(GenTree* addr);
5222 GenTree* fgMorphField(GenTree* tree, MorphAddrContext* mac);
5223 bool fgCanFastTailCall(GenTreeCall* call);
5224 bool fgCheckStmtAfterTailCall();
5225 void fgMorphTailCall(GenTreeCall* call, void* pfnCopyArgs);
5226 GenTree* fgGetStubAddrArg(GenTreeCall* call);
5227 void fgMorphRecursiveFastTailCallIntoLoop(BasicBlock* block, GenTreeCall* recursiveTailCall);
5228 GenTree* fgAssignRecursiveCallArgToCallerParam(GenTree* arg,
5229 fgArgTabEntry* argTabEntry,
5231 IL_OFFSETX callILOffset,
5232 GenTree* tmpAssignmentInsertionPoint,
5233 GenTree* paramAssignmentInsertionPoint);
5234 static int fgEstimateCallStackSize(GenTreeCall* call);
5235 GenTree* fgMorphCall(GenTreeCall* call);
5236 void fgMorphCallInline(GenTreeCall* call, InlineResult* result);
5237 void fgMorphCallInlineHelper(GenTreeCall* call, InlineResult* result);
5239 void fgNoteNonInlineCandidate(GenTreeStmt* stmt, GenTreeCall* call);
5240 static fgWalkPreFn fgFindNonInlineCandidate;
5242 GenTree* fgOptimizeDelegateConstructor(GenTreeCall* call,
5243 CORINFO_CONTEXT_HANDLE* ExactContextHnd,
5244 CORINFO_RESOLVED_TOKEN* ldftnToken);
5245 GenTree* fgMorphLeaf(GenTree* tree);
5246 void fgAssignSetVarDef(GenTree* tree);
5247 GenTree* fgMorphOneAsgBlockOp(GenTree* tree);
5248 GenTree* fgMorphInitBlock(GenTree* tree);
5249 GenTree* fgMorphBlkToInd(GenTreeBlk* tree, var_types type);
5250 GenTree* fgMorphGetStructAddr(GenTree** pTree, CORINFO_CLASS_HANDLE clsHnd, bool isRValue = false);
5251 GenTree* fgMorphBlkNode(GenTree* tree, bool isDest);
5252 GenTree* fgMorphBlockOperand(GenTree* tree, var_types asgType, unsigned blockWidth, bool isDest);
5253 void fgMorphUnsafeBlk(GenTreeObj* obj);
5254 GenTree* fgMorphCopyBlock(GenTree* tree);
5255 GenTree* fgMorphForRegisterFP(GenTree* tree);
5256 GenTree* fgMorphSmpOp(GenTree* tree, MorphAddrContext* mac = nullptr);
5257 GenTree* fgMorphModToSubMulDiv(GenTreeOp* tree);
5258 GenTree* fgMorphSmpOpOptional(GenTreeOp* tree);
5259 GenTree* fgMorphRecognizeBoxNullable(GenTree* compare);
5261 GenTree* fgMorphToEmulatedFP(GenTree* tree);
5262 GenTree* fgMorphConst(GenTree* tree);
5265 GenTree* fgMorphTree(GenTree* tree, MorphAddrContext* mac = nullptr);
5268 #if LOCAL_ASSERTION_PROP
5269 void fgKillDependentAssertionsSingle(unsigned lclNum DEBUGARG(GenTree* tree));
5270 void fgKillDependentAssertions(unsigned lclNum DEBUGARG(GenTree* tree));
5272 void fgMorphTreeDone(GenTree* tree, GenTree* oldTree = nullptr DEBUGARG(int morphNum = 0));
5274 GenTreeStmt* fgMorphStmt;
5276 unsigned fgGetBigOffsetMorphingTemp(var_types type); // We cache one temp per type to be
5277 // used when morphing big offset.
5279 //----------------------- Liveness analysis -------------------------------
5281 VARSET_TP fgCurUseSet; // vars used by block (before an assignment)
5282 VARSET_TP fgCurDefSet; // vars assigned by block (before a use)
5284 MemoryKindSet fgCurMemoryUse; // True iff the current basic block uses memory.
5285 MemoryKindSet fgCurMemoryDef; // True iff the current basic block modifies memory.
5286 MemoryKindSet fgCurMemoryHavoc; // True if the current basic block is known to set memory to a "havoc" value.
5288 bool byrefStatesMatchGcHeapStates; // True iff GcHeap and ByrefExposed memory have all the same def points.
5290 void fgMarkUseDef(GenTreeLclVarCommon* tree);
5292 void fgBeginScopeLife(VARSET_TP* inScope, VarScopeDsc* var);
5293 void fgEndScopeLife(VARSET_TP* inScope, VarScopeDsc* var);
5295 void fgMarkInScope(BasicBlock* block, VARSET_VALARG_TP inScope);
5296 void fgUnmarkInScope(BasicBlock* block, VARSET_VALARG_TP unmarkScope);
5298 void fgExtendDbgScopes();
5299 void fgExtendDbgLifetimes();
5302 void fgDispDebugScopes();
5305 //-------------------------------------------------------------------------
5307 // The following keeps track of any code we've added for things like array
5308 // range checking or explicit calls to enable GC, and so on.
5313 AddCodeDsc* acdNext;
5314 BasicBlock* acdDstBlk; // block to which we jump
5316 SpecialCodeKind acdKind; // what kind of a special block is this?
5317 #if !FEATURE_FIXED_OUT_ARGS
5318 bool acdStkLvlInit; // has acdStkLvl value been already set?
5320 #endif // !FEATURE_FIXED_OUT_ARGS
5324 static unsigned acdHelper(SpecialCodeKind codeKind);
5326 AddCodeDsc* fgAddCodeList;
5328 bool fgRngChkThrowAdded;
5329 AddCodeDsc* fgExcptnTargetCache[SCK_COUNT];
5331 BasicBlock* fgRngChkTarget(BasicBlock* block, SpecialCodeKind kind);
5333 BasicBlock* fgAddCodeRef(BasicBlock* srcBlk, unsigned refData, SpecialCodeKind kind);
5336 AddCodeDsc* fgFindExcptnTarget(SpecialCodeKind kind, unsigned refData);
5338 bool fgUseThrowHelperBlocks();
5340 AddCodeDsc* fgGetAdditionalCodeDescriptors()
5342 return fgAddCodeList;
5346 bool fgIsCodeAdded();
5348 bool fgIsThrowHlpBlk(BasicBlock* block);
5350 #if !FEATURE_FIXED_OUT_ARGS
5351 unsigned fgThrowHlpBlkStkLevel(BasicBlock* block);
5352 #endif // !FEATURE_FIXED_OUT_ARGS
5354 unsigned fgBigOffsetMorphingTemps[TYP_COUNT];
5356 unsigned fgCheckInlineDepthAndRecursion(InlineInfo* inlineInfo);
5357 void fgInvokeInlineeCompiler(GenTreeCall* call, InlineResult* result);
5358 void fgInsertInlineeBlocks(InlineInfo* pInlineInfo);
5359 GenTree* fgInlinePrependStatements(InlineInfo* inlineInfo);
5360 void fgInlineAppendStatements(InlineInfo* inlineInfo, BasicBlock* block, GenTree* stmt);
5362 #if FEATURE_MULTIREG_RET
5363 GenTree* fgGetStructAsStructPtr(GenTree* tree);
5364 GenTree* fgAssignStructInlineeToVar(GenTree* child, CORINFO_CLASS_HANDLE retClsHnd);
5365 void fgAttachStructInlineeToAsg(GenTree* tree, GenTree* child, CORINFO_CLASS_HANDLE retClsHnd);
5366 #endif // FEATURE_MULTIREG_RET
5368 static fgWalkPreFn fgUpdateInlineReturnExpressionPlaceHolder;
5369 static fgWalkPostFn fgLateDevirtualization;
5372 static fgWalkPreFn fgDebugCheckInlineCandidates;
5374 void CheckNoTransformableIndirectCallsRemain();
5375 static fgWalkPreFn fgDebugCheckForTransformableIndirectCalls;
5378 void fgPromoteStructs();
5379 void fgMorphStructField(GenTree* tree, GenTree* parent);
5380 void fgMorphLocalField(GenTree* tree, GenTree* parent);
5382 // Identify which parameters are implicit byrefs, and flag their LclVarDscs.
5383 void fgMarkImplicitByRefArgs();
5385 // Change implicit byrefs' types from struct to pointer, and for any that were
5386 // promoted, create new promoted struct temps.
5387 void fgRetypeImplicitByRefArgs();
5389 // Rewrite appearances of implicit byrefs (manifest the implied additional level of indirection).
5390 bool fgMorphImplicitByRefArgs(GenTree* tree);
5391 GenTree* fgMorphImplicitByRefArgs(GenTree* tree, bool isAddr);
5393 // Clear up annotations for any struct promotion temps created for implicit byrefs.
5394 void fgMarkDemotedImplicitByRefArgs();
5396 void fgMarkAddressExposedLocals();
5398 static fgWalkPreFn fgUpdateSideEffectsPre;
5399 static fgWalkPostFn fgUpdateSideEffectsPost;
5401 // The given local variable, required to be a struct variable, is being assigned via
5402 // a "lclField", to make it masquerade as an integral type in the ABI. Make sure that
5403 // the variable is not enregistered, and is therefore not promoted independently.
5404 void fgLclFldAssign(unsigned lclNum);
5406 static fgWalkPreFn gtHasLocalsWithAddrOpCB;
5408 enum TypeProducerKind
5410 TPK_Unknown = 0, // May not be a RuntimeType
5411 TPK_Handle = 1, // RuntimeType via handle
5412 TPK_GetType = 2, // RuntimeType via Object.get_Type()
5413 TPK_Null = 3, // Tree value is null
5414 TPK_Other = 4 // RuntimeType via other means
5417 TypeProducerKind gtGetTypeProducerKind(GenTree* tree);
5418 bool gtIsTypeHandleToRuntimeTypeHelper(GenTreeCall* call);
5419 bool gtIsTypeHandleToRuntimeTypeHandleHelper(GenTreeCall* call, CorInfoHelpFunc* pHelper = nullptr);
5420 bool gtIsActiveCSE_Candidate(GenTree* tree);
5423 bool fgPrintInlinedMethods;
5426 bool fgIsBigOffset(size_t offset);
5428 bool fgNeedReturnSpillTemp();
5431 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
5432 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
5436 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
5437 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
5443 void optRemoveRangeCheck(GenTree* tree, GenTree* stmt);
5444 bool optIsRangeCheckRemovable(GenTree* tree);
5447 static fgWalkPreFn optValidRangeCheckIndex;
5448 static fgWalkPreFn optRemoveTreeVisitor; // Helper passed to Compiler::fgWalkAllTreesPre() to decrement the LclVar
5451 void optRemoveTree(GenTree* deadTree, GenTree* keepList);
5453 /**************************************************************************
5455 *************************************************************************/
5458 // Do hoisting for all loops.
5459 void optHoistLoopCode();
5461 // To represent sets of VN's that have already been hoisted in outer loops.
5462 typedef JitHashTable<ValueNum, JitSmallPrimitiveKeyFuncs<ValueNum>, bool> VNToBoolMap;
5463 typedef VNToBoolMap VNSet;
5465 struct LoopHoistContext
5468 // The set of variables hoisted in the current loop (or nullptr if there are none).
5469 VNSet* m_pHoistedInCurLoop;
5472 // Value numbers of expressions that have been hoisted in parent loops in the loop nest.
5473 VNSet m_hoistedInParentLoops;
5474 // Value numbers of expressions that have been hoisted in the current (or most recent) loop in the nest.
5475 // Previous decisions on loop-invariance of value numbers in the current loop.
5476 VNToBoolMap m_curLoopVnInvariantCache;
5478 VNSet* GetHoistedInCurLoop(Compiler* comp)
5480 if (m_pHoistedInCurLoop == nullptr)
5482 m_pHoistedInCurLoop = new (comp->getAllocatorLoopHoist()) VNSet(comp->getAllocatorLoopHoist());
5484 return m_pHoistedInCurLoop;
5487 VNSet* ExtractHoistedInCurLoop()
5489 VNSet* res = m_pHoistedInCurLoop;
5490 m_pHoistedInCurLoop = nullptr;
5494 LoopHoistContext(Compiler* comp)
5495 : m_pHoistedInCurLoop(nullptr)
5496 , m_hoistedInParentLoops(comp->getAllocatorLoopHoist())
5497 , m_curLoopVnInvariantCache(comp->getAllocatorLoopHoist())
5502 // Do hoisting for loop "lnum" (an index into the optLoopTable), and all loops nested within it.
5503 // Tracks the expressions that have been hoisted by containing loops by temporary recording their
5504 // value numbers in "m_hoistedInParentLoops". This set is not modified by the call.
5505 void optHoistLoopNest(unsigned lnum, LoopHoistContext* hoistCtxt);
5507 // Do hoisting for a particular loop ("lnum" is an index into the optLoopTable.)
5508 // Assumes that expressions have been hoisted in containing loops if their value numbers are in
5509 // "m_hoistedInParentLoops".
5511 void optHoistThisLoop(unsigned lnum, LoopHoistContext* hoistCtxt);
5513 // Hoist all expressions in "blk" that are invariant in loop "lnum" (an index into the optLoopTable)
5514 // outside of that loop. Exempt expressions whose value number is in "m_hoistedInParentLoops"; add VN's of hoisted
5515 // expressions to "hoistInLoop".
5516 void optHoistLoopExprsForBlock(BasicBlock* blk, unsigned lnum, LoopHoistContext* hoistCtxt);
5518 // Return true if the tree looks profitable to hoist out of loop 'lnum'.
5519 bool optIsProfitableToHoistableTree(GenTree* tree, unsigned lnum);
5521 // Hoist all proper sub-expressions of "tree" (which occurs in "stmt", which occurs in "blk")
5522 // that are invariant in loop "lnum" (an index into the optLoopTable)
5523 // outside of that loop. Exempt expressions whose value number is in "hoistedInParents"; add VN's of hoisted
5524 // expressions to "hoistInLoop".
5525 // Returns "true" iff "tree" is loop-invariant (wrt "lnum").
5526 // Assumes that the value of "*firstBlockAndBeforeSideEffect" indicates that we're in the first block, and before
5527 // any possible globally visible side effects. Assume is called in evaluation order, and updates this.
5528 bool optHoistLoopExprsForTree(GenTree* tree,
5530 LoopHoistContext* hoistCtxt,
5531 bool* firstBlockAndBeforeSideEffect,
5533 bool* pCctorDependent);
5535 // Performs the hoisting 'tree' into the PreHeader for loop 'lnum'
5536 void optHoistCandidate(GenTree* tree, unsigned lnum, LoopHoistContext* hoistCtxt);
5538 // Returns true iff the ValueNum "vn" represents a value that is loop-invariant in "lnum".
5539 // Constants and init values are always loop invariant.
5540 // VNPhi's connect VN's to the SSA definition, so we can know if the SSA def occurs in the loop.
5541 bool optVNIsLoopInvariant(ValueNum vn, unsigned lnum, VNToBoolMap* recordedVNs);
5543 // Returns "true" iff "tree" is valid at the head of loop "lnum", in the context of the hoist substitution
5544 // "subst". If "tree" is a local SSA var, it is valid if its SSA definition occurs outside of the loop, or
5545 // if it is in the domain of "subst" (meaning that it's definition has been previously hoisted, with a "standin"
5546 // local.) If tree is a constant, it is valid. Otherwise, if it is an operator, it is valid iff its children are.
5547 bool optTreeIsValidAtLoopHead(GenTree* tree, unsigned lnum);
5549 // If "blk" is the entry block of a natural loop, returns true and sets "*pLnum" to the index of the loop
5550 // in the loop table.
5551 bool optBlockIsLoopEntry(BasicBlock* blk, unsigned* pLnum);
5553 // Records the set of "side effects" of all loops: fields (object instance and static)
5554 // written to, and SZ-array element type equivalence classes updated.
5555 void optComputeLoopSideEffects();
5558 // Requires "lnum" to be the index of an outermost loop in the loop table. Traverses the body of that loop,
5559 // including all nested loops, and records the set of "side effects" of the loop: fields (object instance and
5560 // static) written to, and SZ-array element type equivalence classes updated.
5561 void optComputeLoopNestSideEffects(unsigned lnum);
5563 // Add the side effects of "blk" (which is required to be within a loop) to all loops of which it is a part.
5564 void optComputeLoopSideEffectsOfBlock(BasicBlock* blk);
5566 // Hoist the expression "expr" out of loop "lnum".
5567 void optPerformHoistExpr(GenTree* expr, unsigned lnum);
5570 void optOptimizeBools();
5573 GenTree* optIsBoolCond(GenTree* condBranch, GenTree** compPtr, bool* boolPtr);
5575 void optOptimizeBoolsGcStress(BasicBlock* condBlock);
5578 void optOptimizeLayout(); // Optimize the BasicBlock layout of the method
5580 void optOptimizeLoops(); // for "while-do" loops duplicates simple loop conditions and transforms
5581 // the loop into a "do-while" loop
5582 // Also finds all natural loops and records them in the loop table
5584 // Optionally clone loops in the loop table.
5585 void optCloneLoops();
5587 // Clone loop "loopInd" in the loop table.
5588 void optCloneLoop(unsigned loopInd, LoopCloneContext* context);
5590 // Ensure that loop "loopInd" has a unique head block. (If the existing entry has
5591 // non-loop predecessors other than the head entry, create a new, empty block that goes (only) to the entry,
5592 // and redirects the preds of the entry to this new block.) Sets the weight of the newly created block to
5594 void optEnsureUniqueHead(unsigned loopInd, unsigned ambientWeight);
5596 void optUnrollLoops(); // Unrolls loops (needs to have cost info)
5599 // This enumeration describes what is killed by a call.
5603 CALLINT_NONE, // no interference (most helpers)
5604 CALLINT_REF_INDIRS, // kills GC ref indirections (SETFIELD OBJ)
5605 CALLINT_SCL_INDIRS, // kills non GC ref indirections (SETFIELD non-OBJ)
5606 CALLINT_ALL_INDIRS, // kills both GC ref and non GC ref indirections (SETFIELD STRUCT)
5607 CALLINT_ALL, // kills everything (normal method call)
5611 // A "LoopDsc" describes a ("natural") loop. We (currently) require the body of a loop to be a contiguous (in
5612 // bbNext order) sequence of basic blocks. (At times, we may require the blocks in a loop to be "properly numbered"
5613 // in bbNext order; we use comparisons on the bbNum to decide order.)
5614 // The blocks that define the body are
5615 // first <= top <= entry <= bottom .
5616 // The "head" of the loop is a block outside the loop that has "entry" as a successor. We only support loops with a
5617 // single 'head' block. The meanings of these blocks are given in the definitions below. Also see the picture at
5618 // Compiler::optFindNaturalLoops().
5621 BasicBlock* lpHead; // HEAD of the loop (not part of the looping of the loop) -- has ENTRY as a successor.
5622 BasicBlock* lpFirst; // FIRST block (in bbNext order) reachable within this loop. (May be part of a nested
5623 // loop, but not the outer loop.)
5624 BasicBlock* lpTop; // loop TOP (the back edge from lpBottom reaches here) (in most cases FIRST and TOP are the
5626 BasicBlock* lpEntry; // the ENTRY in the loop (in most cases TOP or BOTTOM)
5627 BasicBlock* lpBottom; // loop BOTTOM (from here we have a back edge to the TOP)
5628 BasicBlock* lpExit; // if a single exit loop this is the EXIT (in most cases BOTTOM)
5630 callInterf lpAsgCall; // "callInterf" for calls in the loop
5631 ALLVARSET_TP lpAsgVars; // set of vars assigned within the loop (all vars, not just tracked)
5632 varRefKinds lpAsgInds : 8; // set of inds modified within the loop
5634 unsigned short lpFlags; // Mask of the LPFLG_* constants
5636 unsigned char lpExitCnt; // number of exits from the loop
5638 unsigned char lpParent; // The index of the most-nested loop that completely contains this one,
5639 // or else BasicBlock::NOT_IN_LOOP if no such loop exists.
5640 unsigned char lpChild; // The index of a nested loop, or else BasicBlock::NOT_IN_LOOP if no child exists.
5641 // (Actually, an "immediately" nested loop --
5642 // no other child of this loop is a parent of lpChild.)
5643 unsigned char lpSibling; // The index of another loop that is an immediate child of lpParent,
5644 // or else BasicBlock::NOT_IN_LOOP. One can enumerate all the children of a loop
5645 // by following "lpChild" then "lpSibling" links.
5647 #define LPFLG_DO_WHILE 0x0001 // it's a do-while loop (i.e ENTRY is at the TOP)
5648 #define LPFLG_ONE_EXIT 0x0002 // the loop has only one exit
5650 #define LPFLG_ITER 0x0004 // for (i = icon or lclVar; test_condition(); i++)
5651 #define LPFLG_HOISTABLE 0x0008 // the loop is in a form that is suitable for hoisting expressions
5652 #define LPFLG_CONST 0x0010 // for (i=icon;i<icon;i++){ ... } - constant loop
5654 #define LPFLG_VAR_INIT 0x0020 // iterator is initialized with a local var (var # found in lpVarInit)
5655 #define LPFLG_CONST_INIT 0x0040 // iterator is initialized with a constant (found in lpConstInit)
5657 #define LPFLG_VAR_LIMIT 0x0100 // iterator is compared with a local var (var # found in lpVarLimit)
5658 #define LPFLG_CONST_LIMIT 0x0200 // iterator is compared with a constant (found in lpConstLimit)
5659 #define LPFLG_ARRLEN_LIMIT 0x0400 // iterator is compared with a.len or a[i].len (found in lpArrLenLimit)
5660 #define LPFLG_SIMD_LIMIT 0x0080 // iterator is compared with Vector<T>.Count (found in lpConstLimit)
5662 #define LPFLG_HAS_PREHEAD 0x0800 // lpHead is known to be a preHead for this loop
5663 #define LPFLG_REMOVED 0x1000 // has been removed from the loop table (unrolled or optimized away)
5664 #define LPFLG_DONT_UNROLL 0x2000 // do not unroll this loop
5666 #define LPFLG_ASGVARS_YES 0x4000 // "lpAsgVars" has been computed
5667 #define LPFLG_ASGVARS_INC 0x8000 // "lpAsgVars" is incomplete -- vars beyond those representable in an AllVarSet
5668 // type are assigned to.
5670 bool lpLoopHasMemoryHavoc[MemoryKindCount]; // The loop contains an operation that we assume has arbitrary
5671 // memory side effects. If this is set, the fields below
5672 // may not be accurate (since they become irrelevant.)
5673 bool lpContainsCall; // True if executing the loop body *may* execute a call
5675 VARSET_TP lpVarInOut; // The set of variables that are IN or OUT during the execution of this loop
5676 VARSET_TP lpVarUseDef; // The set of variables that are USE or DEF during the execution of this loop
5678 int lpHoistedExprCount; // The register count for the non-FP expressions from inside this loop that have been
5680 int lpLoopVarCount; // The register count for the non-FP LclVars that are read/written inside this loop
5681 int lpVarInOutCount; // The register count for the non-FP LclVars that are alive inside or accross this loop
5683 int lpHoistedFPExprCount; // The register count for the FP expressions from inside this loop that have been
5685 int lpLoopVarFPCount; // The register count for the FP LclVars that are read/written inside this loop
5686 int lpVarInOutFPCount; // The register count for the FP LclVars that are alive inside or accross this loop
5688 typedef JitHashTable<CORINFO_FIELD_HANDLE, JitPtrKeyFuncs<struct CORINFO_FIELD_STRUCT_>, bool> FieldHandleSet;
5689 FieldHandleSet* lpFieldsModified; // This has entries (mappings to "true") for all static field and object
5690 // instance fields modified
5693 typedef JitHashTable<CORINFO_CLASS_HANDLE, JitPtrKeyFuncs<struct CORINFO_CLASS_STRUCT_>, bool> ClassHandleSet;
5694 ClassHandleSet* lpArrayElemTypesModified; // Bits set indicate the set of sz array element types such that
5695 // arrays of that type are modified
5698 // Adds the variable liveness information for 'blk' to 'this' LoopDsc
5699 void AddVariableLiveness(Compiler* comp, BasicBlock* blk);
5701 inline void AddModifiedField(Compiler* comp, CORINFO_FIELD_HANDLE fldHnd);
5702 // This doesn't *always* take a class handle -- it can also take primitive types, encoded as class handles
5703 // (shifted left, with a low-order bit set to distinguish.)
5704 // Use the {Encode/Decode}ElemType methods to construct/destruct these.
5705 inline void AddModifiedElemType(Compiler* comp, CORINFO_CLASS_HANDLE structHnd);
5707 /* The following values are set only for iterator loops, i.e. has the flag LPFLG_ITER set */
5709 GenTree* lpIterTree; // The "i = i <op> const" tree
5710 unsigned lpIterVar(); // iterator variable #
5711 int lpIterConst(); // the constant with which the iterator is incremented
5712 genTreeOps lpIterOper(); // the type of the operation on the iterator (ASG_ADD, ASG_SUB, etc.)
5713 void VERIFY_lpIterTree();
5715 var_types lpIterOperType(); // For overflow instructions
5718 int lpConstInit; // initial constant value of iterator : Valid if LPFLG_CONST_INIT
5719 unsigned lpVarInit; // initial local var number to which we initialize the iterator : Valid if
5723 /* The following is for LPFLG_ITER loops only (i.e. the loop condition is "i RELOP const or var" */
5725 GenTree* lpTestTree; // pointer to the node containing the loop test
5726 genTreeOps lpTestOper(); // the type of the comparison between the iterator and the limit (GT_LE, GT_GE, etc.)
5727 void VERIFY_lpTestTree();
5729 bool lpIsReversed(); // true if the iterator node is the second operand in the loop condition
5730 GenTree* lpIterator(); // the iterator node in the loop test
5731 GenTree* lpLimit(); // the limit node in the loop test
5733 int lpConstLimit(); // limit constant value of iterator - loop condition is "i RELOP const" : Valid if
5734 // LPFLG_CONST_LIMIT
5735 unsigned lpVarLimit(); // the lclVar # in the loop condition ( "i RELOP lclVar" ) : Valid if
5737 bool lpArrLenLimit(Compiler* comp, ArrIndex* index); // The array length in the loop condition ( "i RELOP
5738 // arr.len" or "i RELOP arr[i][j].len" ) : Valid if
5739 // LPFLG_ARRLEN_LIMIT
5741 // Returns "true" iff "*this" contains the blk.
5742 bool lpContains(BasicBlock* blk)
5744 return lpFirst->bbNum <= blk->bbNum && blk->bbNum <= lpBottom->bbNum;
5746 // Returns "true" iff "*this" (properly) contains the range [first, bottom] (allowing firsts
5747 // to be equal, but requiring bottoms to be different.)
5748 bool lpContains(BasicBlock* first, BasicBlock* bottom)
5750 return lpFirst->bbNum <= first->bbNum && bottom->bbNum < lpBottom->bbNum;
5753 // Returns "true" iff "*this" (properly) contains "lp2" (allowing firsts to be equal, but requiring
5754 // bottoms to be different.)
5755 bool lpContains(const LoopDsc& lp2)
5757 return lpContains(lp2.lpFirst, lp2.lpBottom);
5760 // Returns "true" iff "*this" is (properly) contained by the range [first, bottom]
5761 // (allowing firsts to be equal, but requiring bottoms to be different.)
5762 bool lpContainedBy(BasicBlock* first, BasicBlock* bottom)
5764 return first->bbNum <= lpFirst->bbNum && lpBottom->bbNum < bottom->bbNum;
5767 // Returns "true" iff "*this" is (properly) contained by "lp2"
5768 // (allowing firsts to be equal, but requiring bottoms to be different.)
5769 bool lpContainedBy(const LoopDsc& lp2)
5771 return lpContains(lp2.lpFirst, lp2.lpBottom);
5774 // Returns "true" iff "*this" is disjoint from the range [top, bottom].
5775 bool lpDisjoint(BasicBlock* first, BasicBlock* bottom)
5777 return bottom->bbNum < lpFirst->bbNum || lpBottom->bbNum < first->bbNum;
5779 // Returns "true" iff "*this" is disjoint from "lp2".
5780 bool lpDisjoint(const LoopDsc& lp2)
5782 return lpDisjoint(lp2.lpFirst, lp2.lpBottom);
5784 // Returns "true" iff the loop is well-formed (see code for defn).
5787 return lpFirst->bbNum <= lpTop->bbNum && lpTop->bbNum <= lpEntry->bbNum &&
5788 lpEntry->bbNum <= lpBottom->bbNum &&
5789 (lpHead->bbNum < lpTop->bbNum || lpHead->bbNum > lpBottom->bbNum);
5794 bool fgMightHaveLoop(); // returns true if there are any backedges
5795 bool fgHasLoops; // True if this method has any loops, set in fgComputeReachability
5798 LoopDsc* optLoopTable; // loop descriptor table
5799 unsigned char optLoopCount; // number of tracked loops
5801 bool optRecordLoop(BasicBlock* head,
5807 unsigned char exitCnt);
5810 unsigned optCallCount; // number of calls made in the method
5811 unsigned optIndirectCallCount; // number of virtual, interface and indirect calls made in the method
5812 unsigned optNativeCallCount; // number of Pinvoke/Native calls made in the method
5813 unsigned optLoopsCloned; // number of loops cloned in the current method.
5816 unsigned optFindLoopNumberFromBeginBlock(BasicBlock* begBlk);
5817 void optPrintLoopInfo(unsigned loopNum,
5819 BasicBlock* lpFirst,
5821 BasicBlock* lpEntry,
5822 BasicBlock* lpBottom,
5823 unsigned char lpExitCnt,
5825 unsigned parentLoop = BasicBlock::NOT_IN_LOOP);
5826 void optPrintLoopInfo(unsigned lnum);
5827 void optPrintLoopRecording(unsigned lnum);
5829 void optCheckPreds();
5832 void optSetBlockWeights();
5834 void optMarkLoopBlocks(BasicBlock* begBlk, BasicBlock* endBlk, bool excludeEndBlk);
5836 void optUnmarkLoopBlocks(BasicBlock* begBlk, BasicBlock* endBlk);
5838 void optUpdateLoopsBeforeRemoveBlock(BasicBlock* block, bool skipUnmarkLoop = false);
5840 bool optIsLoopTestEvalIntoTemp(GenTree* test, GenTree** newTest);
5841 unsigned optIsLoopIncrTree(GenTree* incr);
5842 bool optCheckIterInLoopTest(unsigned loopInd, GenTree* test, BasicBlock* from, BasicBlock* to, unsigned iterVar);
5843 bool optComputeIterInfo(GenTree* incr, BasicBlock* from, BasicBlock* to, unsigned* pIterVar);
5844 bool optPopulateInitInfo(unsigned loopInd, GenTree* init, unsigned iterVar);
5845 bool optExtractInitTestIncr(
5846 BasicBlock* head, BasicBlock* bottom, BasicBlock* exit, GenTree** ppInit, GenTree** ppTest, GenTree** ppIncr);
5848 void optFindNaturalLoops();
5850 // Ensures that all the loops in the loop nest rooted at "loopInd" (an index into the loop table) are 'canonical' --
5851 // each loop has a unique "top." Returns "true" iff the flowgraph has been modified.
5852 bool optCanonicalizeLoopNest(unsigned char loopInd);
5854 // Ensures that the loop "loopInd" (an index into the loop table) is 'canonical' -- it has a unique "top,"
5855 // unshared with any other loop. Returns "true" iff the flowgraph has been modified
5856 bool optCanonicalizeLoop(unsigned char loopInd);
5858 // Requires "l1" to be a valid loop table index, and not "BasicBlock::NOT_IN_LOOP". Requires "l2" to be
5859 // a valid loop table index, or else "BasicBlock::NOT_IN_LOOP". Returns true
5860 // iff "l2" is not NOT_IN_LOOP, and "l1" contains "l2".
5861 bool optLoopContains(unsigned l1, unsigned l2);
5863 // Requires "loopInd" to be a valid index into the loop table.
5864 // Updates the loop table by changing loop "loopInd", whose head is required
5865 // to be "from", to be "to". Also performs this transformation for any
5866 // loop nested in "loopInd" that shares the same head as "loopInd".
5867 void optUpdateLoopHead(unsigned loopInd, BasicBlock* from, BasicBlock* to);
5869 // Updates the successors of "blk": if "blk2" is a successor of "blk", and there is a mapping for "blk2->blk3" in
5870 // "redirectMap", change "blk" so that "blk3" is this successor. Note that the predecessor lists are not updated.
5871 void optRedirectBlock(BasicBlock* blk, BlockToBlockMap* redirectMap);
5873 // Marks the containsCall information to "lnum" and any parent loops.
5874 void AddContainsCallAllContainingLoops(unsigned lnum);
5875 // Adds the variable liveness information from 'blk' to "lnum" and any parent loops.
5876 void AddVariableLivenessAllContainingLoops(unsigned lnum, BasicBlock* blk);
5877 // Adds "fldHnd" to the set of modified fields of "lnum" and any parent loops.
5878 void AddModifiedFieldAllContainingLoops(unsigned lnum, CORINFO_FIELD_HANDLE fldHnd);
5879 // Adds "elemType" to the set of modified array element types of "lnum" and any parent loops.
5880 void AddModifiedElemTypeAllContainingLoops(unsigned lnum, CORINFO_CLASS_HANDLE elemType);
5882 // Requires that "from" and "to" have the same "bbJumpKind" (perhaps because "to" is a clone
5883 // of "from".) Copies the jump destination from "from" to "to".
5884 void optCopyBlkDest(BasicBlock* from, BasicBlock* to);
5886 // The depth of the loop described by "lnum" (an index into the loop table.) (0 == top level)
5887 unsigned optLoopDepth(unsigned lnum)
5889 unsigned par = optLoopTable[lnum].lpParent;
5890 if (par == BasicBlock::NOT_IN_LOOP)
5896 return 1 + optLoopDepth(par);
5900 void fgOptWhileLoop(BasicBlock* block);
5902 bool optComputeLoopRep(int constInit,
5905 genTreeOps iterOper,
5907 genTreeOps testOper,
5910 unsigned* iterCount);
5913 static fgWalkPreFn optIsVarAssgCB;
5916 bool optIsVarAssigned(BasicBlock* beg, BasicBlock* end, GenTree* skip, unsigned var);
5918 bool optIsVarAssgLoop(unsigned lnum, unsigned var);
5920 int optIsSetAssgLoop(unsigned lnum, ALLVARSET_VALARG_TP vars, varRefKinds inds = VR_NONE);
5922 bool optNarrowTree(GenTree* tree, var_types srct, var_types dstt, ValueNumPair vnpNarrow, bool doit);
5924 /**************************************************************************
5925 * Optimization conditions
5926 *************************************************************************/
5928 bool optFastCodeOrBlendedLoop(BasicBlock::weight_t bbWeight);
5929 bool optPentium4(void);
5930 bool optAvoidIncDec(BasicBlock::weight_t bbWeight);
5931 bool optAvoidIntMult(void);
5936 // The following is the upper limit on how many expressions we'll keep track
5937 // of for the CSE analysis.
5939 static const unsigned MAX_CSE_CNT = EXPSET_SZ;
5941 static const int MIN_CSE_COST = 2;
5943 // Keeps tracked cse indices
5944 BitVecTraits* cseTraits;
5947 /* Generic list of nodes - used by the CSE logic */
5957 treeStmtLst* tslNext;
5958 GenTree* tslTree; // tree node
5959 GenTree* tslStmt; // statement containing the tree
5960 BasicBlock* tslBlock; // block containing the statement
5963 // The following logic keeps track of expressions via a simple hash table.
5967 CSEdsc* csdNextInBucket; // used by the hash table
5969 unsigned csdHashKey; // the orginal hashkey
5971 unsigned csdIndex; // 1..optCSECandidateCount
5972 char csdLiveAcrossCall; // 0 or 1
5974 unsigned short csdDefCount; // definition count
5975 unsigned short csdUseCount; // use count (excluding the implicit uses at defs)
5977 unsigned csdDefWtCnt; // weighted def count
5978 unsigned csdUseWtCnt; // weighted use count (excluding the implicit uses at defs)
5980 GenTree* csdTree; // treenode containing the 1st occurance
5981 GenTree* csdStmt; // stmt containing the 1st occurance
5982 BasicBlock* csdBlock; // block containing the 1st occurance
5984 treeStmtLst* csdTreeList; // list of matching tree nodes: head
5985 treeStmtLst* csdTreeLast; // list of matching tree nodes: tail
5987 ValueNum defExcSetPromise; // The exception set that is now required for all defs of this CSE.
5988 // This will be set to NoVN if we decide to abandon this CSE
5990 ValueNum defExcSetCurrent; // The set of exceptions we currently can use for CSE uses.
5992 ValueNum defConservNormVN; // if all def occurrences share the same conservative normal value
5993 // number, this will reflect it; otherwise, NoVN.
5996 static const size_t s_optCSEhashSize;
5997 CSEdsc** optCSEhash;
6000 typedef JitHashTable<GenTree*, JitPtrKeyFuncs<GenTree>, GenTree*> NodeToNodeMap;
6002 NodeToNodeMap* optCseCheckedBoundMap; // Maps bound nodes to ancestor compares that should be
6003 // re-numbered with the bound to improve range check elimination
6005 // Given a compare, look for a cse candidate checked bound feeding it and add a map entry if found.
6006 void optCseUpdateCheckedBoundMap(GenTree* compare);
6010 CSEdsc* optCSEfindDsc(unsigned index);
6011 bool optUnmarkCSE(GenTree* tree);
6013 // user defined callback data for the tree walk function optCSE_MaskHelper()
6014 struct optCSE_MaskData
6016 EXPSET_TP CSE_defMask;
6017 EXPSET_TP CSE_useMask;
6020 // Treewalk helper for optCSE_DefMask and optCSE_UseMask
6021 static fgWalkPreFn optCSE_MaskHelper;
6023 // This function walks all the node for an given tree
6024 // and return the mask of CSE definitions and uses for the tree
6026 void optCSE_GetMaskData(GenTree* tree, optCSE_MaskData* pMaskData);
6028 // Given a binary tree node return true if it is safe to swap the order of evaluation for op1 and op2.
6029 bool optCSE_canSwap(GenTree* firstNode, GenTree* secondNode);
6030 bool optCSE_canSwap(GenTree* tree);
6032 static int __cdecl optCSEcostCmpEx(const void* op1, const void* op2);
6033 static int __cdecl optCSEcostCmpSz(const void* op1, const void* op2);
6035 void optCleanupCSEs();
6038 void optEnsureClearCSEInfo();
6041 #endif // FEATURE_ANYCSE
6043 #if FEATURE_VALNUM_CSE
6044 /**************************************************************************
6045 * Value Number based CSEs
6046 *************************************************************************/
6049 void optOptimizeValnumCSEs();
6052 void optValnumCSE_Init();
6053 unsigned optValnumCSE_Index(GenTree* tree, GenTree* stmt);
6054 unsigned optValnumCSE_Locate();
6055 void optValnumCSE_InitDataFlow();
6056 void optValnumCSE_DataFlow();
6057 void optValnumCSE_Availablity();
6058 void optValnumCSE_Heuristic();
6060 #endif // FEATURE_VALNUM_CSE
6063 bool optDoCSE; // True when we have found a duplicate CSE tree
6064 bool optValnumCSE_phase; // True when we are executing the optValnumCSE_phase
6065 unsigned optCSECandidateTotal; // Grand total of CSE candidates for both Lexical and ValNum
6066 unsigned optCSECandidateCount; // Count of CSE's candidates, reset for Lexical and ValNum CSE's
6067 unsigned optCSEstart; // The first local variable number that is a CSE
6068 unsigned optCSEcount; // The total count of CSE's introduced.
6069 unsigned optCSEweight; // The weight of the current block when we are
6070 // scanning for CSE expressions
6072 bool optIsCSEcandidate(GenTree* tree);
6074 // lclNumIsTrueCSE returns true if the LclVar was introduced by the CSE phase of the compiler
6076 bool lclNumIsTrueCSE(unsigned lclNum) const
6078 return ((optCSEcount > 0) && (lclNum >= optCSEstart) && (lclNum < optCSEstart + optCSEcount));
6081 // lclNumIsCSE returns true if the LclVar should be treated like a CSE with regards to constant prop.
6083 bool lclNumIsCSE(unsigned lclNum) const
6085 return lvaTable[lclNum].lvIsCSE;
6089 bool optConfigDisableCSE();
6090 bool optConfigDisableCSE2();
6092 void optOptimizeCSEs();
6094 #endif // FEATURE_ANYCSE
6102 unsigned ivaVar; // Variable we are interested in, or -1
6103 ALLVARSET_TP ivaMaskVal; // Set of variables assigned to. This is a set of all vars, not tracked vars.
6104 bool ivaMaskIncomplete; // Variables not representable in ivaMaskVal were assigned to.
6105 varRefKinds ivaMaskInd; // What kind of indirect assignments are there?
6106 callInterf ivaMaskCall; // What kind of calls are there?
6109 static callInterf optCallInterf(GenTreeCall* call);
6112 // VN based copy propagation.
6113 typedef ArrayStack<GenTree*> GenTreePtrStack;
6114 typedef JitHashTable<unsigned, JitSmallPrimitiveKeyFuncs<unsigned>, GenTreePtrStack*> LclNumToGenTreePtrStack;
6116 // Kill set to track variables with intervening definitions.
6117 VARSET_TP optCopyPropKillSet;
6119 // Copy propagation functions.
6120 void optCopyProp(BasicBlock* block, GenTree* stmt, GenTree* tree, LclNumToGenTreePtrStack* curSsaName);
6121 void optBlockCopyPropPopStacks(BasicBlock* block, LclNumToGenTreePtrStack* curSsaName);
6122 void optBlockCopyProp(BasicBlock* block, LclNumToGenTreePtrStack* curSsaName);
6123 bool optIsSsaLocal(GenTree* tree);
6124 int optCopyProp_LclVarScore(LclVarDsc* lclVarDsc, LclVarDsc* copyVarDsc, bool preferOp2);
6125 void optVnCopyProp();
6126 INDEBUG(void optDumpCopyPropStack(LclNumToGenTreePtrStack* curSsaName));
6128 /**************************************************************************
6129 * Early value propagation
6130 *************************************************************************/
6136 SSAName(unsigned lvNum, unsigned ssaNum) : m_lvNum(lvNum), m_ssaNum(ssaNum)
6140 static unsigned GetHashCode(SSAName ssaNm)
6142 return (ssaNm.m_lvNum << 16) | (ssaNm.m_ssaNum);
6145 static bool Equals(SSAName ssaNm1, SSAName ssaNm2)
6147 return (ssaNm1.m_lvNum == ssaNm2.m_lvNum) && (ssaNm1.m_ssaNum == ssaNm2.m_ssaNum);
6151 #define OMF_HAS_NEWARRAY 0x00000001 // Method contains 'new' of an array
6152 #define OMF_HAS_NEWOBJ 0x00000002 // Method contains 'new' of an object type.
6153 #define OMF_HAS_ARRAYREF 0x00000004 // Method contains array element loads or stores.
6154 #define OMF_HAS_VTABLEREF 0x00000008 // Method contains method table reference.
6155 #define OMF_HAS_NULLCHECK 0x00000010 // Method contains null check.
6156 #define OMF_HAS_FATPOINTER 0x00000020 // Method contains call, that needs fat pointer transformation.
6157 #define OMF_HAS_OBJSTACKALLOC 0x00000040 // Method contains an object allocated on the stack.
6158 #define OMF_HAS_GUARDEDDEVIRT 0x00000080 // Method contains guarded devirtualization candidate
6160 bool doesMethodHaveFatPointer()
6162 return (optMethodFlags & OMF_HAS_FATPOINTER) != 0;
6165 void setMethodHasFatPointer()
6167 optMethodFlags |= OMF_HAS_FATPOINTER;
6170 void clearMethodHasFatPointer()
6172 optMethodFlags &= ~OMF_HAS_FATPOINTER;
6175 void addFatPointerCandidate(GenTreeCall* call);
6177 bool doesMethodHaveGuardedDevirtualization()
6179 return (optMethodFlags & OMF_HAS_GUARDEDDEVIRT) != 0;
6182 void setMethodHasGuardedDevirtualization()
6184 optMethodFlags |= OMF_HAS_GUARDEDDEVIRT;
6187 void clearMethodHasGuardedDevirtualization()
6189 optMethodFlags &= ~OMF_HAS_GUARDEDDEVIRT;
6192 void addGuardedDevirtualizationCandidate(GenTreeCall* call,
6193 CORINFO_METHOD_HANDLE methodHandle,
6194 CORINFO_CLASS_HANDLE classHandle,
6195 unsigned methodAttr,
6196 unsigned classAttr);
6198 unsigned optMethodFlags;
6200 // Recursion bound controls how far we can go backwards tracking for a SSA value.
6201 // No throughput diff was found with backward walk bound between 3-8.
6202 static const int optEarlyPropRecurBound = 5;
6204 enum class optPropKind
6212 bool gtIsVtableRef(GenTree* tree);
6213 GenTree* getArrayLengthFromAllocation(GenTree* tree);
6214 GenTree* getObjectHandleNodeFromAllocation(GenTree* tree);
6215 GenTree* optPropGetValueRec(unsigned lclNum, unsigned ssaNum, optPropKind valueKind, int walkDepth);
6216 GenTree* optPropGetValue(unsigned lclNum, unsigned ssaNum, optPropKind valueKind);
6217 GenTree* optEarlyPropRewriteTree(GenTree* tree);
6218 bool optDoEarlyPropForBlock(BasicBlock* block);
6219 bool optDoEarlyPropForFunc();
6220 void optEarlyProp();
6221 void optFoldNullCheck(GenTree* tree);
6222 bool optCanMoveNullCheckPastTree(GenTree* tree, bool isInsideTry);
6225 /**************************************************************************
6226 * Value/Assertion propagation
6227 *************************************************************************/
6229 // Data structures for assertion prop
6230 BitVecTraits* apTraits;
6233 enum optAssertionKind
6250 O1K_CONSTANT_LOOP_BND,
6271 optAssertionKind assertionKind;
6274 unsigned lclNum; // assigned to or property of this local var number
6282 struct AssertionDscOp1
6284 optOp1Kind kind; // a normal LclVar, or Exact-type or Subtype
6291 struct AssertionDscOp2
6293 optOp2Kind kind; // a const or copy assignment
6297 ssize_t iconVal; // integer
6298 unsigned iconFlags; // gtFlags
6300 struct Range // integer subrange
6314 bool IsCheckedBoundArithBound()
6316 return ((assertionKind == OAK_EQUAL || assertionKind == OAK_NOT_EQUAL) && op1.kind == O1K_BOUND_OPER_BND);
6318 bool IsCheckedBoundBound()
6320 return ((assertionKind == OAK_EQUAL || assertionKind == OAK_NOT_EQUAL) && op1.kind == O1K_BOUND_LOOP_BND);
6322 bool IsConstantBound()
6324 return ((assertionKind == OAK_EQUAL || assertionKind == OAK_NOT_EQUAL) &&
6325 op1.kind == O1K_CONSTANT_LOOP_BND);
6327 bool IsBoundsCheckNoThrow()
6329 return ((assertionKind == OAK_NO_THROW) && (op1.kind == O1K_ARR_BND));
6332 bool IsCopyAssertion()
6334 return ((assertionKind == OAK_EQUAL) && (op1.kind == O1K_LCLVAR) && (op2.kind == O2K_LCLVAR_COPY));
6337 static bool SameKind(AssertionDsc* a1, AssertionDsc* a2)
6339 return a1->assertionKind == a2->assertionKind && a1->op1.kind == a2->op1.kind &&
6340 a1->op2.kind == a2->op2.kind;
6343 static bool ComplementaryKind(optAssertionKind kind, optAssertionKind kind2)
6345 if (kind == OAK_EQUAL)
6347 return kind2 == OAK_NOT_EQUAL;
6349 else if (kind == OAK_NOT_EQUAL)
6351 return kind2 == OAK_EQUAL;
6356 static ssize_t GetLowerBoundForIntegralType(var_types type)
6375 static ssize_t GetUpperBoundForIntegralType(var_types type)
6398 bool HasSameOp1(AssertionDsc* that, bool vnBased)
6400 if (op1.kind != that->op1.kind)
6404 else if (op1.kind == O1K_ARR_BND)
6407 return (op1.bnd.vnIdx == that->op1.bnd.vnIdx) && (op1.bnd.vnLen == that->op1.bnd.vnLen);
6411 return ((vnBased && (op1.vn == that->op1.vn)) ||
6412 (!vnBased && (op1.lcl.lclNum == that->op1.lcl.lclNum)));
6416 bool HasSameOp2(AssertionDsc* that, bool vnBased)
6418 if (op2.kind != that->op2.kind)
6424 case O2K_IND_CNS_INT:
6426 return ((op2.u1.iconVal == that->op2.u1.iconVal) && (op2.u1.iconFlags == that->op2.u1.iconFlags));
6428 case O2K_CONST_LONG:
6429 return (op2.lconVal == that->op2.lconVal);
6431 case O2K_CONST_DOUBLE:
6432 // exact match because of positive and negative zero.
6433 return (memcmp(&op2.dconVal, &that->op2.dconVal, sizeof(double)) == 0);
6435 case O2K_LCLVAR_COPY:
6437 return (op2.lcl.lclNum == that->op2.lcl.lclNum) &&
6438 (!vnBased || op2.lcl.ssaNum == that->op2.lcl.ssaNum);
6441 return ((op2.u2.loBound == that->op2.u2.loBound) && (op2.u2.hiBound == that->op2.u2.hiBound));
6444 // we will return false
6448 assert(!"Unexpected value for op2.kind in AssertionDsc.");
6454 bool Complementary(AssertionDsc* that, bool vnBased)
6456 return ComplementaryKind(assertionKind, that->assertionKind) && HasSameOp1(that, vnBased) &&
6457 HasSameOp2(that, vnBased);
6460 bool Equals(AssertionDsc* that, bool vnBased)
6462 if (assertionKind != that->assertionKind)
6466 else if (assertionKind == OAK_NO_THROW)
6468 assert(op2.kind == O2K_INVALID);
6469 return HasSameOp1(that, vnBased);
6473 return HasSameOp1(that, vnBased) && HasSameOp2(that, vnBased);
6479 static fgWalkPreFn optAddCopiesCallback;
6480 static fgWalkPreFn optVNAssertionPropCurStmtVisitor;
6481 unsigned optAddCopyLclNum;
6482 GenTree* optAddCopyAsgnNode;
6484 bool optLocalAssertionProp; // indicates that we are performing local assertion prop
6485 bool optAssertionPropagated; // set to true if we modified the trees
6486 bool optAssertionPropagatedCurrentStmt;
6488 GenTree* optAssertionPropCurrentTree;
6490 AssertionIndex* optComplementaryAssertionMap;
6491 JitExpandArray<ASSERT_TP>* optAssertionDep; // table that holds dependent assertions (assertions
6492 // using the value of a local var) for each local var
6493 AssertionDsc* optAssertionTabPrivate; // table that holds info about value assignments
6494 AssertionIndex optAssertionCount; // total number of assertions in the assertion table
6495 AssertionIndex optMaxAssertionCount;
6498 void optVnNonNullPropCurStmt(BasicBlock* block, GenTree* stmt, GenTree* tree);
6499 fgWalkResult optVNConstantPropCurStmt(BasicBlock* block, GenTree* stmt, GenTree* tree);
6500 GenTree* optVNConstantPropOnJTrue(BasicBlock* block, GenTree* stmt, GenTree* test);
6501 GenTree* optVNConstantPropOnTree(BasicBlock* block, GenTree* stmt, GenTree* tree);
6502 GenTree* optPrepareTreeForReplacement(GenTree* extractTree, GenTree* replaceTree);
6504 AssertionIndex GetAssertionCount()
6506 return optAssertionCount;
6508 ASSERT_TP* bbJtrueAssertionOut;
6509 typedef JitHashTable<ValueNum, JitSmallPrimitiveKeyFuncs<ValueNum>, ASSERT_TP> ValueNumToAssertsMap;
6510 ValueNumToAssertsMap* optValueNumToAsserts;
6512 // Assertion prop helpers.
6513 ASSERT_TP& GetAssertionDep(unsigned lclNum);
6514 AssertionDsc* optGetAssertion(AssertionIndex assertIndex);
6515 void optAssertionInit(bool isLocalProp);
6516 void optAssertionTraitsInit(AssertionIndex assertionCount);
6517 #if LOCAL_ASSERTION_PROP
6518 void optAssertionReset(AssertionIndex limit);
6519 void optAssertionRemove(AssertionIndex index);
6522 // Assertion prop data flow functions.
6523 void optAssertionPropMain();
6524 GenTree* optVNAssertionPropCurStmt(BasicBlock* block, GenTree* stmt);
6525 bool optIsTreeKnownIntValue(bool vnBased, GenTree* tree, ssize_t* pConstant, unsigned* pIconFlags);
6526 ASSERT_TP* optInitAssertionDataflowFlags();
6527 ASSERT_TP* optComputeAssertionGen();
6529 // Assertion Gen functions.
6530 void optAssertionGen(GenTree* tree);
6531 AssertionIndex optAssertionGenPhiDefn(GenTree* tree);
6532 AssertionInfo optCreateJTrueBoundsAssertion(GenTree* tree);
6533 AssertionInfo optAssertionGenJtrue(GenTree* tree);
6534 AssertionIndex optCreateJtrueAssertions(GenTree* op1, GenTree* op2, Compiler::optAssertionKind assertionKind);
6535 AssertionIndex optFindComplementary(AssertionIndex assertionIndex);
6536 void optMapComplementary(AssertionIndex assertionIndex, AssertionIndex index);
6538 // Assertion creation functions.
6539 AssertionIndex optCreateAssertion(GenTree* op1, GenTree* op2, optAssertionKind assertionKind);
6540 AssertionIndex optCreateAssertion(GenTree* op1,
6542 optAssertionKind assertionKind,
6543 AssertionDsc* assertion);
6544 void optCreateComplementaryAssertion(AssertionIndex assertionIndex, GenTree* op1, GenTree* op2);
6546 bool optAssertionVnInvolvesNan(AssertionDsc* assertion);
6547 AssertionIndex optAddAssertion(AssertionDsc* assertion);
6548 void optAddVnAssertionMapping(ValueNum vn, AssertionIndex index);
6550 void optPrintVnAssertionMapping();
6552 ASSERT_TP optGetVnMappedAssertions(ValueNum vn);
6554 // Used for respective assertion propagations.
6555 AssertionIndex optAssertionIsSubrange(GenTree* tree, var_types toType, ASSERT_VALARG_TP assertions);
6556 AssertionIndex optAssertionIsSubtype(GenTree* tree, GenTree* methodTableArg, ASSERT_VALARG_TP assertions);
6557 AssertionIndex optAssertionIsNonNullInternal(GenTree* op, ASSERT_VALARG_TP assertions);
6558 bool optAssertionIsNonNull(GenTree* op,
6559 ASSERT_VALARG_TP assertions DEBUGARG(bool* pVnBased) DEBUGARG(AssertionIndex* pIndex));
6561 // Used for Relop propagation.
6562 AssertionIndex optGlobalAssertionIsEqualOrNotEqual(ASSERT_VALARG_TP assertions, GenTree* op1, GenTree* op2);
6563 AssertionIndex optGlobalAssertionIsEqualOrNotEqualZero(ASSERT_VALARG_TP assertions, GenTree* op1);
6564 AssertionIndex optLocalAssertionIsEqualOrNotEqual(
6565 optOp1Kind op1Kind, unsigned lclNum, optOp2Kind op2Kind, ssize_t cnsVal, ASSERT_VALARG_TP assertions);
6567 // Assertion prop for lcl var functions.
6568 bool optAssertionProp_LclVarTypeCheck(GenTree* tree, LclVarDsc* lclVarDsc, LclVarDsc* copyVarDsc);
6569 GenTree* optCopyAssertionProp(AssertionDsc* curAssertion,
6571 GenTree* stmt DEBUGARG(AssertionIndex index));
6572 GenTree* optConstantAssertionProp(AssertionDsc* curAssertion,
6574 GenTree* stmt DEBUGARG(AssertionIndex index));
6576 // Assertion propagation functions.
6577 GenTree* optAssertionProp(ASSERT_VALARG_TP assertions, GenTree* tree, GenTree* stmt);
6578 GenTree* optAssertionProp_LclVar(ASSERT_VALARG_TP assertions, GenTree* tree, GenTree* stmt);
6579 GenTree* optAssertionProp_Ind(ASSERT_VALARG_TP assertions, GenTree* tree, GenTree* stmt);
6580 GenTree* optAssertionProp_Cast(ASSERT_VALARG_TP assertions, GenTree* tree, GenTree* stmt);
6581 GenTree* optAssertionProp_Call(ASSERT_VALARG_TP assertions, GenTreeCall* call, GenTree* stmt);
6582 GenTree* optAssertionProp_RelOp(ASSERT_VALARG_TP assertions, GenTree* tree, GenTree* stmt);
6583 GenTree* optAssertionProp_Comma(ASSERT_VALARG_TP assertions, GenTree* tree, GenTree* stmt);
6584 GenTree* optAssertionProp_BndsChk(ASSERT_VALARG_TP assertions, GenTree* tree, GenTree* stmt);
6585 GenTree* optAssertionPropGlobal_RelOp(ASSERT_VALARG_TP assertions, GenTree* tree, GenTree* stmt);
6586 GenTree* optAssertionPropLocal_RelOp(ASSERT_VALARG_TP assertions, GenTree* tree, GenTree* stmt);
6587 GenTree* optAssertionProp_Update(GenTree* newTree, GenTree* tree, GenTree* stmt);
6588 GenTree* optNonNullAssertionProp_Call(ASSERT_VALARG_TP assertions, GenTreeCall* call, GenTree* stmt);
6590 // Implied assertion functions.
6591 void optImpliedAssertions(AssertionIndex assertionIndex, ASSERT_TP& activeAssertions);
6592 void optImpliedByTypeOfAssertions(ASSERT_TP& activeAssertions);
6593 void optImpliedByCopyAssertion(AssertionDsc* copyAssertion, AssertionDsc* depAssertion, ASSERT_TP& result);
6594 void optImpliedByConstAssertion(AssertionDsc* curAssertion, ASSERT_TP& result);
6597 void optPrintAssertion(AssertionDsc* newAssertion, AssertionIndex assertionIndex = 0);
6598 void optDebugCheckAssertion(AssertionDsc* assertion);
6599 void optDebugCheckAssertions(AssertionIndex AssertionIndex);
6601 void optAddCopies();
6602 #endif // ASSERTION_PROP
6604 /**************************************************************************
6606 *************************************************************************/
6609 struct LoopCloneVisitorInfo
6611 LoopCloneContext* context;
6614 LoopCloneVisitorInfo(LoopCloneContext* context, unsigned loopNum, GenTree* stmt)
6615 : context(context), loopNum(loopNum), stmt(nullptr)
6620 bool optIsStackLocalInvariant(unsigned loopNum, unsigned lclNum);
6621 bool optExtractArrIndex(GenTree* tree, ArrIndex* result, unsigned lhsNum);
6622 bool optReconstructArrIndex(GenTree* tree, ArrIndex* result, unsigned lhsNum);
6623 bool optIdentifyLoopOptInfo(unsigned loopNum, LoopCloneContext* context);
6624 static fgWalkPreFn optCanOptimizeByLoopCloningVisitor;
6625 fgWalkResult optCanOptimizeByLoopCloning(GenTree* tree, LoopCloneVisitorInfo* info);
6626 void optObtainLoopCloningOpts(LoopCloneContext* context);
6627 bool optIsLoopClonable(unsigned loopInd);
6629 bool optCanCloneLoops();
6632 void optDebugLogLoopCloning(BasicBlock* block, GenTree* insertBefore);
6634 void optPerformStaticOptimizations(unsigned loopNum, LoopCloneContext* context DEBUGARG(bool fastPath));
6635 bool optComputeDerefConditions(unsigned loopNum, LoopCloneContext* context);
6636 bool optDeriveLoopCloningConditions(unsigned loopNum, LoopCloneContext* context);
6637 BasicBlock* optInsertLoopChoiceConditions(LoopCloneContext* context,
6643 ssize_t optGetArrayRefScaleAndIndex(GenTree* mul, GenTree** pIndex DEBUGARG(bool bRngChk));
6645 bool optReachWithoutCall(BasicBlock* srcBB, BasicBlock* dstBB);
6648 bool optLoopsMarked;
6651 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6652 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6656 XX Does the register allocation and puts the remaining lclVars on the stack XX
6658 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6659 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6663 regNumber raUpdateRegStateForArg(RegState* regState, LclVarDsc* argDsc);
6665 void raMarkStkVars();
6668 // Some things are used by both LSRA and regpredict allocators.
6670 FrameType rpFrameType;
6671 bool rpMustCreateEBPCalled; // Set to true after we have called rpMustCreateEBPFrame once
6673 bool rpMustCreateEBPFrame(INDEBUG(const char** wbReason));
6676 Lowering* m_pLowering; // Lowering; needed to Lower IR that's added or modified after Lowering.
6677 LinearScanInterface* m_pLinearScan; // Linear Scan allocator
6679 /* raIsVarargsStackArg is called by raMaskStkVars and by
6680 lvaSortByRefCount. It identifies the special case
6681 where a varargs function has a parameter passed on the
6682 stack, other than the special varargs handle. Such parameters
6683 require special treatment, because they cannot be tracked
6684 by the GC (their offsets in the stack are not known
6688 bool raIsVarargsStackArg(unsigned lclNum)
6692 LclVarDsc* varDsc = &lvaTable[lclNum];
6694 assert(varDsc->lvIsParam);
6696 return (info.compIsVarArgs && !varDsc->lvIsRegArg && (lclNum != lvaVarargsHandleArg));
6698 #else // _TARGET_X86_
6702 #endif // _TARGET_X86_
6706 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6707 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6711 XX Get to the class and method info from the Execution Engine given XX
6712 XX tokens for the class and method XX
6714 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6715 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6719 /* These are the different addressing modes used to access a local var.
6720 * The JIT has to report the location of the locals back to the EE
6721 * for debugging purposes.
6727 VLT_REG_BYREF, // this type is currently only used for value types on X64
6730 VLT_STK_BYREF, // this type is currently only used for value types on X64
6744 siVarLocType vlType;
6747 // VLT_REG/VLT_REG_FP -- Any pointer-sized enregistered value (TYP_INT, TYP_REF, etc)
6749 // VLT_REG_BYREF -- the specified register contains the address of the variable
6757 // VLT_STK -- Any 32 bit value which is on the stack
6758 // eg. [ESP+0x20], or [EBP-0x28]
6759 // VLT_STK_BYREF -- the specified stack location contains the address of the variable
6760 // eg. mov EAX, [ESP+0x20]; [EAX]
6764 regNumber vlsBaseReg;
6765 NATIVE_OFFSET vlsOffset;
6768 // VLT_REG_REG -- TYP_LONG/TYP_DOUBLE with both DWords enregistered
6777 // VLT_REG_STK -- Partly enregistered TYP_LONG/TYP_DOUBLE
6778 // eg { LowerDWord=EAX UpperDWord=[ESP+0x8] }
6786 regNumber vlrssBaseReg;
6787 NATIVE_OFFSET vlrssOffset;
6791 // VLT_STK_REG -- Partly enregistered TYP_LONG/TYP_DOUBLE
6792 // eg { LowerDWord=[ESP+0x8] UpperDWord=EAX }
6798 regNumber vlsrsBaseReg;
6799 NATIVE_OFFSET vlsrsOffset;
6805 // VLT_STK2 -- Any 64 bit value which is on the stack, in 2 successsive DWords
6806 // eg 2 DWords at [ESP+0x10]
6810 regNumber vls2BaseReg;
6811 NATIVE_OFFSET vls2Offset;
6814 // VLT_FPSTK -- enregisterd TYP_DOUBLE (on the FP stack)
6815 // eg. ST(3). Actually it is ST("FPstkHeight - vpFpStk")
6822 // VLT_FIXED_VA -- fixed argument of a varargs function.
6823 // The argument location depends on the size of the variable
6824 // arguments (...). Inspecting the VARARGS_HANDLE indicates the
6825 // location of the first arg. This argument can then be accessed
6826 // relative to the position of the first arg
6830 unsigned vlfvOffset;
6837 void* rpValue; // pointer to the in-process
6838 // location of the value.
6844 bool vlIsInReg(regNumber reg);
6845 bool vlIsOnStk(regNumber reg, signed offset);
6848 /*************************************************************************/
6853 void eeGetCallInfo(CORINFO_RESOLVED_TOKEN* pResolvedToken,
6854 CORINFO_RESOLVED_TOKEN* pConstrainedToken,
6855 CORINFO_CALLINFO_FLAGS flags,
6856 CORINFO_CALL_INFO* pResult);
6857 inline CORINFO_CALLINFO_FLAGS addVerifyFlag(CORINFO_CALLINFO_FLAGS flags);
6859 void eeGetFieldInfo(CORINFO_RESOLVED_TOKEN* pResolvedToken,
6860 CORINFO_ACCESS_FLAGS flags,
6861 CORINFO_FIELD_INFO* pResult);
6865 BOOL eeIsValueClass(CORINFO_CLASS_HANDLE clsHnd);
6867 #if defined(DEBUG) || defined(FEATURE_JIT_METHOD_PERF) || defined(FEATURE_SIMD) || defined(TRACK_LSRA_STATS)
6869 bool IsSuperPMIException(unsigned code)
6871 // Copied from NDP\clr\src\ToolBox\SuperPMI\SuperPMI-Shared\ErrorHandling.h
6873 const unsigned EXCEPTIONCODE_DebugBreakorAV = 0xe0421000;
6874 const unsigned EXCEPTIONCODE_MC = 0xe0422000;
6875 const unsigned EXCEPTIONCODE_LWM = 0xe0423000;
6876 const unsigned EXCEPTIONCODE_SASM = 0xe0424000;
6877 const unsigned EXCEPTIONCODE_SSYM = 0xe0425000;
6878 const unsigned EXCEPTIONCODE_CALLUTILS = 0xe0426000;
6879 const unsigned EXCEPTIONCODE_TYPEUTILS = 0xe0427000;
6880 const unsigned EXCEPTIONCODE_ASSERT = 0xe0440000;
6884 case EXCEPTIONCODE_DebugBreakorAV:
6885 case EXCEPTIONCODE_MC:
6886 case EXCEPTIONCODE_LWM:
6887 case EXCEPTIONCODE_SASM:
6888 case EXCEPTIONCODE_SSYM:
6889 case EXCEPTIONCODE_CALLUTILS:
6890 case EXCEPTIONCODE_TYPEUTILS:
6891 case EXCEPTIONCODE_ASSERT:
6898 const char* eeGetMethodName(CORINFO_METHOD_HANDLE hnd, const char** className);
6899 const char* eeGetMethodFullName(CORINFO_METHOD_HANDLE hnd);
6901 bool eeIsNativeMethod(CORINFO_METHOD_HANDLE method);
6902 CORINFO_METHOD_HANDLE eeGetMethodHandleForNative(CORINFO_METHOD_HANDLE method);
6905 var_types eeGetArgType(CORINFO_ARG_LIST_HANDLE list, CORINFO_SIG_INFO* sig);
6906 var_types eeGetArgType(CORINFO_ARG_LIST_HANDLE list, CORINFO_SIG_INFO* sig, bool* isPinned);
6907 unsigned eeGetArgSize(CORINFO_ARG_LIST_HANDLE list, CORINFO_SIG_INFO* sig);
6909 // VOM info, method sigs
6911 void eeGetSig(unsigned sigTok,
6912 CORINFO_MODULE_HANDLE scope,
6913 CORINFO_CONTEXT_HANDLE context,
6914 CORINFO_SIG_INFO* retSig);
6916 void eeGetCallSiteSig(unsigned sigTok,
6917 CORINFO_MODULE_HANDLE scope,
6918 CORINFO_CONTEXT_HANDLE context,
6919 CORINFO_SIG_INFO* retSig);
6921 void eeGetMethodSig(CORINFO_METHOD_HANDLE methHnd, CORINFO_SIG_INFO* retSig, CORINFO_CLASS_HANDLE owner = nullptr);
6923 // Method entry-points, instrs
6925 CORINFO_METHOD_HANDLE eeMarkNativeTarget(CORINFO_METHOD_HANDLE method);
6927 CORINFO_EE_INFO eeInfo;
6928 bool eeInfoInitialized;
6930 CORINFO_EE_INFO* eeGetEEInfo();
6932 // Gets the offset of a SDArray's first element
6933 unsigned eeGetArrayDataOffset(var_types type);
6934 // Gets the offset of a MDArray's first element
6935 unsigned eeGetMDArrayDataOffset(var_types type, unsigned rank);
6937 GenTree* eeGetPInvokeCookie(CORINFO_SIG_INFO* szMetaSig);
6939 // Returns the page size for the target machine as reported by the EE.
6940 inline target_size_t eeGetPageSize()
6942 return (target_size_t)eeGetEEInfo()->osPageSize;
6945 // Returns the frame size at which we will generate a loop to probe the stack.
6946 inline target_size_t getVeryLargeFrameSize()
6949 // The looping probe code is 40 bytes, whereas the straight-line probing for
6950 // the (0x2000..0x3000) case is 44, so use looping for anything 0x2000 bytes
6951 // or greater, to generate smaller code.
6952 return 2 * eeGetPageSize();
6954 return 3 * eeGetPageSize();
6958 //------------------------------------------------------------------------
6959 // VirtualStubParam: virtual stub dispatch extra parameter (slot address).
6961 // It represents Abi and target specific registers for the parameter.
6963 class VirtualStubParamInfo
6966 VirtualStubParamInfo(bool isCoreRTABI)
6968 #if defined(_TARGET_X86_)
6971 #elif defined(_TARGET_AMD64_)
6982 #elif defined(_TARGET_ARM_)
6993 #elif defined(_TARGET_ARM64_)
6997 #error Unsupported or unset target architecture
7001 regNumber GetReg() const
7006 _regMask_enum GetRegMask() const
7013 _regMask_enum regMask;
7016 VirtualStubParamInfo* virtualStubParamInfo;
7018 inline bool IsTargetAbi(CORINFO_RUNTIME_ABI abi)
7020 return eeGetEEInfo()->targetAbi == abi;
7023 inline bool generateCFIUnwindCodes()
7025 #if defined(_TARGET_UNIX_)
7026 return IsTargetAbi(CORINFO_CORERT_ABI);
7032 // Debugging support - Line number info
7034 void eeGetStmtOffsets();
7036 unsigned eeBoundariesCount;
7038 struct boundariesDsc
7040 UNATIVE_OFFSET nativeIP;
7042 unsigned sourceReason;
7043 } * eeBoundaries; // Boundaries to report to EE
7044 void eeSetLIcount(unsigned count);
7045 void eeSetLIinfo(unsigned which, UNATIVE_OFFSET offs, unsigned srcIP, bool stkEmpty, bool callInstruction);
7049 static void eeDispILOffs(IL_OFFSET offs);
7050 static void eeDispLineInfo(const boundariesDsc* line);
7051 void eeDispLineInfos();
7054 // Debugging support - Local var info
7058 unsigned eeVarsCount;
7060 struct VarResultInfo
7062 UNATIVE_OFFSET startOffset;
7063 UNATIVE_OFFSET endOffset;
7067 void eeSetLVcount(unsigned count);
7068 void eeSetLVinfo(unsigned which,
7069 UNATIVE_OFFSET startOffs,
7070 UNATIVE_OFFSET length,
7075 const siVarLoc& loc);
7079 void eeDispVar(ICorDebugInfo::NativeVarInfo* var);
7080 void eeDispVars(CORINFO_METHOD_HANDLE ftn, ULONG32 cVars, ICorDebugInfo::NativeVarInfo* vars);
7083 // ICorJitInfo wrappers
7085 void eeReserveUnwindInfo(BOOL isFunclet, BOOL isColdCode, ULONG unwindSize);
7087 void eeAllocUnwindInfo(BYTE* pHotCode,
7093 CorJitFuncKind funcKind);
7095 void eeSetEHcount(unsigned cEH);
7097 void eeSetEHinfo(unsigned EHnumber, const CORINFO_EH_CLAUSE* clause);
7099 WORD eeGetRelocTypeHint(void* target);
7101 // ICorStaticInfo wrapper functions
7103 bool eeTryResolveToken(CORINFO_RESOLVED_TOKEN* resolvedToken);
7105 #if defined(UNIX_AMD64_ABI)
7107 static void dumpSystemVClassificationType(SystemVClassificationType ct);
7110 void eeGetSystemVAmd64PassStructInRegisterDescriptor(
7111 /*IN*/ CORINFO_CLASS_HANDLE structHnd,
7112 /*OUT*/ SYSTEMV_AMD64_CORINFO_STRUCT_REG_PASSING_DESCRIPTOR* structPassInRegDescPtr);
7113 #endif // UNIX_AMD64_ABI
7115 template <typename ParamType>
7116 bool eeRunWithErrorTrap(void (*function)(ParamType*), ParamType* param)
7118 return eeRunWithErrorTrapImp(reinterpret_cast<void (*)(void*)>(function), reinterpret_cast<void*>(param));
7121 bool eeRunWithErrorTrapImp(void (*function)(void*), void* param);
7123 // Utility functions
7125 const char* eeGetFieldName(CORINFO_FIELD_HANDLE fieldHnd, const char** classNamePtr = nullptr);
7128 const wchar_t* eeGetCPString(size_t stringHandle);
7131 const char* eeGetClassName(CORINFO_CLASS_HANDLE clsHnd);
7133 static CORINFO_METHOD_HANDLE eeFindHelper(unsigned helper);
7134 static CorInfoHelpFunc eeGetHelperNum(CORINFO_METHOD_HANDLE method);
7136 static fgWalkPreFn CountSharedStaticHelper;
7137 static bool IsSharedStaticHelper(GenTree* tree);
7138 static bool IsTreeAlwaysHoistable(GenTree* tree);
7139 static bool IsGcSafePoint(GenTree* tree);
7141 static CORINFO_FIELD_HANDLE eeFindJitDataOffs(unsigned jitDataOffs);
7142 // returns true/false if 'field' is a Jit Data offset
7143 static bool eeIsJitDataOffs(CORINFO_FIELD_HANDLE field);
7144 // returns a number < 0 if 'field' is not a Jit Data offset, otherwise the data offset (limited to 2GB)
7145 static int eeGetJitDataOffs(CORINFO_FIELD_HANDLE field);
7147 /*****************************************************************************/
7150 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7151 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7155 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7156 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7160 CodeGenInterface* codeGen;
7162 // The following holds information about instr offsets in terms of generated code.
7166 IPmappingDsc* ipmdNext; // next line# record
7167 IL_OFFSETX ipmdILoffsx; // the instr offset
7168 emitLocation ipmdNativeLoc; // the emitter location of the native code corresponding to the IL offset
7169 bool ipmdIsLabel; // Can this code be a branch label?
7172 // Record the instr offset mapping to the generated code
7174 IPmappingDsc* genIPmappingList;
7175 IPmappingDsc* genIPmappingLast;
7177 // Managed RetVal - A side hash table meant to record the mapping from a
7178 // GT_CALL node to its IL offset. This info is used to emit sequence points
7179 // that can be used by debugger to determine the native offset at which the
7180 // managed RetVal will be available.
7182 // In fact we can store IL offset in a GT_CALL node. This was ruled out in
7183 // favor of a side table for two reasons: 1) We need IL offset for only those
7184 // GT_CALL nodes (created during importation) that correspond to an IL call and
7185 // whose return type is other than TYP_VOID. 2) GT_CALL node is a frequently used
7186 // structure and IL offset is needed only when generating debuggable code. Therefore
7187 // it is desirable to avoid memory size penalty in retail scenarios.
7188 typedef JitHashTable<GenTree*, JitPtrKeyFuncs<GenTree>, IL_OFFSETX> CallSiteILOffsetTable;
7189 CallSiteILOffsetTable* genCallSite2ILOffsetMap;
7191 unsigned genReturnLocal; // Local number for the return value when applicable.
7192 BasicBlock* genReturnBB; // jumped to when not optimizing for speed.
7194 // The following properties are part of CodeGenContext. Getters are provided here for
7195 // convenience and backward compatibility, but the properties can only be set by invoking
7196 // the setter on CodeGenContext directly.
7198 __declspec(property(get = getEmitter)) emitter* genEmitter;
7199 emitter* getEmitter()
7201 return codeGen->getEmitter();
7204 bool isFramePointerUsed()
7206 return codeGen->isFramePointerUsed();
7209 __declspec(property(get = getInterruptible, put = setInterruptible)) bool genInterruptible;
7210 bool getInterruptible()
7212 return codeGen->genInterruptible;
7214 void setInterruptible(bool value)
7216 codeGen->setInterruptible(value);
7219 #ifdef _TARGET_ARMARCH_
7220 __declspec(property(get = getHasTailCalls, put = setHasTailCalls)) bool hasTailCalls;
7221 bool getHasTailCalls()
7223 return codeGen->hasTailCalls;
7225 void setHasTailCalls(bool value)
7227 codeGen->setHasTailCalls(value);
7229 #endif // _TARGET_ARMARCH_
7232 const bool genDoubleAlign()
7234 return codeGen->doDoubleAlign();
7236 DWORD getCanDoubleAlign();
7237 bool shouldDoubleAlign(unsigned refCntStk,
7239 unsigned refCntWtdReg,
7240 unsigned refCntStkParam,
7241 unsigned refCntWtdStkDbl);
7242 #endif // DOUBLE_ALIGN
7244 __declspec(property(get = getFullPtrRegMap, put = setFullPtrRegMap)) bool genFullPtrRegMap;
7245 bool getFullPtrRegMap()
7247 return codeGen->genFullPtrRegMap;
7249 void setFullPtrRegMap(bool value)
7251 codeGen->setFullPtrRegMap(value);
7254 // Things that MAY belong either in CodeGen or CodeGenContext
7256 #if FEATURE_EH_FUNCLETS
7257 FuncInfoDsc* compFuncInfos;
7258 unsigned short compCurrFuncIdx;
7259 unsigned short compFuncInfoCount;
7261 unsigned short compFuncCount()
7263 assert(fgFuncletsCreated);
7264 return compFuncInfoCount;
7267 #else // !FEATURE_EH_FUNCLETS
7269 // This is a no-op when there are no funclets!
7270 void genUpdateCurrentFunclet(BasicBlock* block)
7275 FuncInfoDsc compFuncInfoRoot;
7277 static const unsigned compCurrFuncIdx = 0;
7279 unsigned short compFuncCount()
7284 #endif // !FEATURE_EH_FUNCLETS
7286 FuncInfoDsc* funCurrentFunc();
7287 void funSetCurrentFunc(unsigned funcIdx);
7288 FuncInfoDsc* funGetFunc(unsigned funcIdx);
7289 unsigned int funGetFuncIdx(BasicBlock* block);
7293 VARSET_TP compCurLife; // current live variables
7294 GenTree* compCurLifeTree; // node after which compCurLife has been computed
7296 template <bool ForCodeGen>
7297 void compChangeLife(VARSET_VALARG_TP newLife);
7299 void genChangeLife(VARSET_VALARG_TP newLife)
7301 compChangeLife</*ForCodeGen*/ true>(newLife);
7304 template <bool ForCodeGen>
7305 inline void compUpdateLife(VARSET_VALARG_TP newLife);
7307 // Gets a register mask that represent the kill set for a helper call since
7308 // not all JIT Helper calls follow the standard ABI on the target architecture.
7309 regMaskTP compHelperCallKillSet(CorInfoHelpFunc helper);
7311 // Gets a register mask that represent the kill set for a NoGC helper call.
7312 regMaskTP compNoGCHelperCallKillSet(CorInfoHelpFunc helper);
7315 // Requires that "varDsc" be a promoted struct local variable being passed as an argument, beginning at
7316 // "firstArgRegNum", which is assumed to have already been aligned to the register alignment restriction of the
7317 // struct type. Adds bits to "*pArgSkippedRegMask" for any argument registers *not* used in passing "varDsc" --
7318 // i.e., internal "holes" caused by internal alignment constraints. For example, if the struct contained an int and
7319 // a double, and we at R0 (on ARM), then R1 would be skipped, and the bit for R1 would be added to the mask.
7320 void fgAddSkippedRegsInPromotedStructArg(LclVarDsc* varDsc, unsigned firstArgRegNum, regMaskTP* pArgSkippedRegMask);
7321 #endif // _TARGET_ARM_
7323 // 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
7325 static GenTree* fgIsIndirOfAddrOfLocal(GenTree* tree);
7327 // This map is indexed by GT_OBJ nodes that are address of promoted struct variables, which
7328 // have been annotated with the GTF_VAR_DEATH flag. If such a node is *not* mapped in this
7329 // table, one may assume that all the (tracked) field vars die at this GT_OBJ. Otherwise,
7330 // the node maps to a pointer to a VARSET_TP, containing set bits for each of the tracked field
7331 // vars of the promoted struct local that go dead at the given node (the set bits are the bits
7332 // for the tracked var indices of the field vars, as in a live var set).
7334 // The map is allocated on demand so all map operations should use one of the following three
7337 NodeToVarsetPtrMap* m_promotedStructDeathVars;
7339 NodeToVarsetPtrMap* GetPromotedStructDeathVars()
7341 if (m_promotedStructDeathVars == nullptr)
7343 m_promotedStructDeathVars = new (getAllocator()) NodeToVarsetPtrMap(getAllocator());
7345 return m_promotedStructDeathVars;
7348 void ClearPromotedStructDeathVars()
7350 if (m_promotedStructDeathVars != nullptr)
7352 m_promotedStructDeathVars->RemoveAll();
7356 bool LookupPromotedStructDeathVars(GenTree* tree, VARSET_TP** bits)
7359 bool result = false;
7361 if (m_promotedStructDeathVars != nullptr)
7363 result = m_promotedStructDeathVars->Lookup(tree, bits);
7370 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7371 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7375 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7376 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7379 #if !defined(__GNUC__)
7380 #pragma region Unwind information
7385 // Infrastructure functions: start/stop/reserve/emit.
7388 void unwindBegProlog();
7389 void unwindEndProlog();
7390 void unwindBegEpilog();
7391 void unwindEndEpilog();
7392 void unwindReserve();
7393 void unwindEmit(void* pHotCode, void* pColdCode);
7396 // Specific unwind information functions: called by code generation to indicate a particular
7397 // prolog or epilog unwindable instruction has been generated.
7400 void unwindPush(regNumber reg);
7401 void unwindAllocStack(unsigned size);
7402 void unwindSetFrameReg(regNumber reg, unsigned offset);
7403 void unwindSaveReg(regNumber reg, unsigned offset);
7405 #if defined(_TARGET_ARM_)
7406 void unwindPushMaskInt(regMaskTP mask);
7407 void unwindPushMaskFloat(regMaskTP mask);
7408 void unwindPopMaskInt(regMaskTP mask);
7409 void unwindPopMaskFloat(regMaskTP mask);
7410 void unwindBranch16(); // The epilog terminates with a 16-bit branch (e.g., "bx lr")
7411 void unwindNop(unsigned codeSizeInBytes); // Generate unwind NOP code. 'codeSizeInBytes' is 2 or 4 bytes. Only
7412 // called via unwindPadding().
7413 void unwindPadding(); // Generate a sequence of unwind NOP codes representing instructions between the last
7414 // instruction and the current location.
7415 #endif // _TARGET_ARM_
7417 #if defined(_TARGET_ARM64_)
7419 void unwindPadding(); // Generate a sequence of unwind NOP codes representing instructions between the last
7420 // instruction and the current location.
7421 void unwindSaveReg(regNumber reg, int offset); // str reg, [sp, #offset]
7422 void unwindSaveRegPreindexed(regNumber reg, int offset); // str reg, [sp, #offset]!
7423 void unwindSaveRegPair(regNumber reg1, regNumber reg2, int offset); // stp reg1, reg2, [sp, #offset]
7424 void unwindSaveRegPairPreindexed(regNumber reg1, regNumber reg2, int offset); // stp reg1, reg2, [sp, #offset]!
7425 void unwindSaveNext(); // unwind code: save_next
7426 void unwindReturn(regNumber reg); // ret lr
7427 #endif // defined(_TARGET_ARM64_)
7430 // Private "helper" functions for the unwind implementation.
7434 #if FEATURE_EH_FUNCLETS
7435 void unwindGetFuncLocations(FuncInfoDsc* func,
7436 bool getHotSectionData,
7437 /* OUT */ emitLocation** ppStartLoc,
7438 /* OUT */ emitLocation** ppEndLoc);
7439 #endif // FEATURE_EH_FUNCLETS
7441 void unwindReserveFunc(FuncInfoDsc* func);
7442 void unwindEmitFunc(FuncInfoDsc* func, void* pHotCode, void* pColdCode);
7444 #if defined(_TARGET_AMD64_) || (defined(_TARGET_X86_) && FEATURE_EH_FUNCLETS)
7446 void unwindReserveFuncHelper(FuncInfoDsc* func, bool isHotCode);
7447 void unwindEmitFuncHelper(FuncInfoDsc* func, void* pHotCode, void* pColdCode, bool isHotCode);
7449 #endif // _TARGET_AMD64_ || (_TARGET_X86_ && FEATURE_EH_FUNCLETS)
7451 UNATIVE_OFFSET unwindGetCurrentOffset(FuncInfoDsc* func);
7453 #if defined(_TARGET_AMD64_)
7455 void unwindBegPrologWindows();
7456 void unwindPushWindows(regNumber reg);
7457 void unwindAllocStackWindows(unsigned size);
7458 void unwindSetFrameRegWindows(regNumber reg, unsigned offset);
7459 void unwindSaveRegWindows(regNumber reg, unsigned offset);
7461 #ifdef UNIX_AMD64_ABI
7462 void unwindSaveRegCFI(regNumber reg, unsigned offset);
7463 #endif // UNIX_AMD64_ABI
7464 #elif defined(_TARGET_ARM_)
7466 void unwindPushPopMaskInt(regMaskTP mask, bool useOpsize16);
7467 void unwindPushPopMaskFloat(regMaskTP mask);
7469 #endif // _TARGET_ARM_
7471 #if defined(_TARGET_UNIX_)
7472 int mapRegNumToDwarfReg(regNumber reg);
7473 void createCfiCode(FuncInfoDsc* func, UCHAR codeOffset, UCHAR opcode, USHORT dwarfReg, INT offset = 0);
7474 void unwindPushPopCFI(regNumber reg);
7475 void unwindBegPrologCFI();
7476 void unwindPushPopMaskCFI(regMaskTP regMask, bool isFloat);
7477 void unwindAllocStackCFI(unsigned size);
7478 void unwindSetFrameRegCFI(regNumber reg, unsigned offset);
7479 void unwindEmitFuncCFI(FuncInfoDsc* func, void* pHotCode, void* pColdCode);
7481 void DumpCfiInfo(bool isHotCode,
7482 UNATIVE_OFFSET startOffset,
7483 UNATIVE_OFFSET endOffset,
7485 const CFI_CODE* const pCfiCode);
7488 #endif // _TARGET_UNIX_
7490 #if !defined(__GNUC__)
7491 #pragma endregion // Note: region is NOT under !defined(__GNUC__)
7495 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7496 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7500 XX Info about SIMD types, methods and the SIMD assembly (i.e. the assembly XX
7501 XX that contains the distinguished, well-known SIMD type definitions). XX
7503 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7504 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
7507 // Get highest available level for SIMD codegen
7508 SIMDLevel getSIMDSupportLevel()
7510 #if defined(_TARGET_XARCH_)
7511 if (compSupports(InstructionSet_AVX2))
7513 return SIMD_AVX2_Supported;
7516 if (compSupports(InstructionSet_SSE42))
7518 return SIMD_SSE4_Supported;
7522 return SIMD_SSE2_Supported;
7524 assert(!"Available instruction set(s) for SIMD codegen is not defined for target arch");
7526 return SIMD_Not_Supported;
7532 // Should we support SIMD intrinsics?
7535 // Have we identified any SIMD types?
7536 // This is currently used by struct promotion to avoid getting type information for a struct
7537 // field to see if it is a SIMD type, if we haven't seen any SIMD types or operations in
7539 bool _usesSIMDTypes;
7540 bool usesSIMDTypes()
7542 return _usesSIMDTypes;
7544 void setUsesSIMDTypes(bool value)
7546 _usesSIMDTypes = value;
7549 // This is a temp lclVar allocated on the stack as TYP_SIMD. It is used to implement intrinsics
7550 // that require indexed access to the individual fields of the vector, which is not well supported
7551 // by the hardware. It is allocated when/if such situations are encountered during Lowering.
7552 unsigned lvaSIMDInitTempVarNum;
7554 struct SIMDHandlesCache
7557 CORINFO_CLASS_HANDLE SIMDFloatHandle;
7558 CORINFO_CLASS_HANDLE SIMDDoubleHandle;
7559 CORINFO_CLASS_HANDLE SIMDIntHandle;
7560 CORINFO_CLASS_HANDLE SIMDUShortHandle;
7561 CORINFO_CLASS_HANDLE SIMDUByteHandle;
7562 CORINFO_CLASS_HANDLE SIMDShortHandle;
7563 CORINFO_CLASS_HANDLE SIMDByteHandle;
7564 CORINFO_CLASS_HANDLE SIMDLongHandle;
7565 CORINFO_CLASS_HANDLE SIMDUIntHandle;
7566 CORINFO_CLASS_HANDLE SIMDULongHandle;
7567 CORINFO_CLASS_HANDLE SIMDVector2Handle;
7568 CORINFO_CLASS_HANDLE SIMDVector3Handle;
7569 CORINFO_CLASS_HANDLE SIMDVector4Handle;
7570 CORINFO_CLASS_HANDLE SIMDVectorHandle;
7572 #ifdef FEATURE_HW_INTRINSICS
7573 #if defined(_TARGET_ARM64_)
7574 CORINFO_CLASS_HANDLE Vector64FloatHandle;
7575 CORINFO_CLASS_HANDLE Vector64IntHandle;
7576 CORINFO_CLASS_HANDLE Vector64UShortHandle;
7577 CORINFO_CLASS_HANDLE Vector64UByteHandle;
7578 CORINFO_CLASS_HANDLE Vector64ShortHandle;
7579 CORINFO_CLASS_HANDLE Vector64ByteHandle;
7580 CORINFO_CLASS_HANDLE Vector64UIntHandle;
7581 #endif // defined(_TARGET_ARM64_)
7582 CORINFO_CLASS_HANDLE Vector128FloatHandle;
7583 CORINFO_CLASS_HANDLE Vector128DoubleHandle;
7584 CORINFO_CLASS_HANDLE Vector128IntHandle;
7585 CORINFO_CLASS_HANDLE Vector128UShortHandle;
7586 CORINFO_CLASS_HANDLE Vector128UByteHandle;
7587 CORINFO_CLASS_HANDLE Vector128ShortHandle;
7588 CORINFO_CLASS_HANDLE Vector128ByteHandle;
7589 CORINFO_CLASS_HANDLE Vector128LongHandle;
7590 CORINFO_CLASS_HANDLE Vector128UIntHandle;
7591 CORINFO_CLASS_HANDLE Vector128ULongHandle;
7592 #if defined(_TARGET_XARCH_)
7593 CORINFO_CLASS_HANDLE Vector256FloatHandle;
7594 CORINFO_CLASS_HANDLE Vector256DoubleHandle;
7595 CORINFO_CLASS_HANDLE Vector256IntHandle;
7596 CORINFO_CLASS_HANDLE Vector256UShortHandle;
7597 CORINFO_CLASS_HANDLE Vector256UByteHandle;
7598 CORINFO_CLASS_HANDLE Vector256ShortHandle;
7599 CORINFO_CLASS_HANDLE Vector256ByteHandle;
7600 CORINFO_CLASS_HANDLE Vector256LongHandle;
7601 CORINFO_CLASS_HANDLE Vector256UIntHandle;
7602 CORINFO_CLASS_HANDLE Vector256ULongHandle;
7603 #endif // defined(_TARGET_XARCH_)
7604 #endif // FEATURE_HW_INTRINSICS
7608 memset(this, 0, sizeof(*this));
7612 SIMDHandlesCache* m_simdHandleCache;
7614 // Get an appropriate "zero" for the given type and class handle.
7615 GenTree* gtGetSIMDZero(var_types simdType, var_types baseType, CORINFO_CLASS_HANDLE simdHandle);
7617 // Get the handle for a SIMD type.
7618 CORINFO_CLASS_HANDLE gtGetStructHandleForSIMD(var_types simdType, var_types simdBaseType)
7620 if (m_simdHandleCache == nullptr)
7622 // This may happen if the JIT generates SIMD node on its own, without importing them.
7623 // Otherwise getBaseTypeAndSizeOfSIMDType should have created the cache.
7624 return NO_CLASS_HANDLE;
7627 if (simdBaseType == TYP_FLOAT)
7632 return m_simdHandleCache->SIMDVector2Handle;
7634 return m_simdHandleCache->SIMDVector3Handle;
7636 if ((getSIMDVectorType() == TYP_SIMD32) ||
7637 (m_simdHandleCache->SIMDVector4Handle != NO_CLASS_HANDLE))
7639 return m_simdHandleCache->SIMDVector4Handle;
7648 assert(emitTypeSize(simdType) <= maxSIMDStructBytes());
7649 switch (simdBaseType)
7652 return m_simdHandleCache->SIMDFloatHandle;
7654 return m_simdHandleCache->SIMDDoubleHandle;
7656 return m_simdHandleCache->SIMDIntHandle;
7658 return m_simdHandleCache->SIMDUShortHandle;
7660 return m_simdHandleCache->SIMDUByteHandle;
7662 return m_simdHandleCache->SIMDShortHandle;
7664 return m_simdHandleCache->SIMDByteHandle;
7666 return m_simdHandleCache->SIMDLongHandle;
7668 return m_simdHandleCache->SIMDUIntHandle;
7670 return m_simdHandleCache->SIMDULongHandle;
7672 assert(!"Didn't find a class handle for simdType");
7674 return NO_CLASS_HANDLE;
7677 // Returns true if this is a SIMD type that should be considered an opaque
7678 // vector type (i.e. do not analyze or promote its fields).
7679 // Note that all but the fixed vector types are opaque, even though they may
7680 // actually be declared as having fields.
7681 bool isOpaqueSIMDType(CORINFO_CLASS_HANDLE structHandle)
7683 return ((m_simdHandleCache != nullptr) && (structHandle != m_simdHandleCache->SIMDVector2Handle) &&
7684 (structHandle != m_simdHandleCache->SIMDVector3Handle) &&
7685 (structHandle != m_simdHandleCache->SIMDVector4Handle));
7688 // Returns true if the tree corresponds to a TYP_SIMD lcl var.
7689 // Note that both SIMD vector args and locals are mared as lvSIMDType = true, but
7690 // type of an arg node is TYP_BYREF and a local node is TYP_SIMD or TYP_STRUCT.
7691 bool isSIMDTypeLocal(GenTree* tree)
7693 return tree->OperIsLocal() && lvaTable[tree->AsLclVarCommon()->gtLclNum].lvSIMDType;
7696 // Returns true if the lclVar is an opaque SIMD type.
7697 bool isOpaqueSIMDLclVar(LclVarDsc* varDsc)
7699 if (!varDsc->lvSIMDType)
7703 return isOpaqueSIMDType(varDsc->lvVerTypeInfo.GetClassHandle());
7706 // Returns true if the type of the tree is a byref of TYP_SIMD
7707 bool isAddrOfSIMDType(GenTree* tree)
7709 if (tree->TypeGet() == TYP_BYREF || tree->TypeGet() == TYP_I_IMPL)
7711 switch (tree->OperGet())
7714 return varTypeIsSIMD(tree->gtGetOp1());
7716 case GT_LCL_VAR_ADDR:
7717 return lvaTable[tree->AsLclVarCommon()->gtLclNum].lvSIMDType;
7720 return isSIMDTypeLocal(tree);
7727 static bool isRelOpSIMDIntrinsic(SIMDIntrinsicID intrinsicId)
7729 return (intrinsicId == SIMDIntrinsicEqual || intrinsicId == SIMDIntrinsicLessThan ||
7730 intrinsicId == SIMDIntrinsicLessThanOrEqual || intrinsicId == SIMDIntrinsicGreaterThan ||
7731 intrinsicId == SIMDIntrinsicGreaterThanOrEqual);
7734 // Returns base type of a TYP_SIMD local.
7735 // Returns TYP_UNKNOWN if the local is not TYP_SIMD.
7736 var_types getBaseTypeOfSIMDLocal(GenTree* tree)
7738 if (isSIMDTypeLocal(tree))
7740 return lvaTable[tree->AsLclVarCommon()->gtLclNum].lvBaseType;
7746 bool isSIMDClass(CORINFO_CLASS_HANDLE clsHnd)
7748 return info.compCompHnd->isInSIMDModule(clsHnd);
7751 bool isIntrinsicType(CORINFO_CLASS_HANDLE clsHnd)
7753 return (info.compCompHnd->getClassAttribs(clsHnd) & CORINFO_FLG_INTRINSIC_TYPE) != 0;
7756 const char* getClassNameFromMetadata(CORINFO_CLASS_HANDLE cls, const char** namespaceName)
7758 return info.compCompHnd->getClassNameFromMetadata(cls, namespaceName);
7761 CORINFO_CLASS_HANDLE getTypeInstantiationArgument(CORINFO_CLASS_HANDLE cls, unsigned index)
7763 return info.compCompHnd->getTypeInstantiationArgument(cls, index);
7766 bool isSIMDClass(typeInfo* pTypeInfo)
7768 return pTypeInfo->IsStruct() && isSIMDClass(pTypeInfo->GetClassHandleForValueClass());
7771 bool isHWSIMDClass(CORINFO_CLASS_HANDLE clsHnd)
7773 #ifdef FEATURE_HW_INTRINSICS
7774 if (isIntrinsicType(clsHnd))
7776 const char* namespaceName = nullptr;
7777 (void)getClassNameFromMetadata(clsHnd, &namespaceName);
7778 return strcmp(namespaceName, "System.Runtime.Intrinsics") == 0;
7780 #endif // FEATURE_HW_INTRINSICS
7784 bool isHWSIMDClass(typeInfo* pTypeInfo)
7786 #ifdef FEATURE_HW_INTRINSICS
7787 return pTypeInfo->IsStruct() && isHWSIMDClass(pTypeInfo->GetClassHandleForValueClass());
7793 bool isSIMDorHWSIMDClass(CORINFO_CLASS_HANDLE clsHnd)
7795 return isSIMDClass(clsHnd) || isHWSIMDClass(clsHnd);
7798 bool isSIMDorHWSIMDClass(typeInfo* pTypeInfo)
7800 return isSIMDClass(pTypeInfo) || isHWSIMDClass(pTypeInfo);
7803 // Get the base (element) type and size in bytes for a SIMD type. Returns TYP_UNKNOWN
7804 // if it is not a SIMD type or is an unsupported base type.
7805 var_types getBaseTypeAndSizeOfSIMDType(CORINFO_CLASS_HANDLE typeHnd, unsigned* sizeBytes = nullptr);
7807 var_types getBaseTypeOfSIMDType(CORINFO_CLASS_HANDLE typeHnd)
7809 return getBaseTypeAndSizeOfSIMDType(typeHnd, nullptr);
7812 // Get SIMD Intrinsic info given the method handle.
7813 // Also sets typeHnd, argCount, baseType and sizeBytes out params.
7814 const SIMDIntrinsicInfo* getSIMDIntrinsicInfo(CORINFO_CLASS_HANDLE* typeHnd,
7815 CORINFO_METHOD_HANDLE methodHnd,
7816 CORINFO_SIG_INFO* sig,
7819 var_types* baseType,
7820 unsigned* sizeBytes);
7822 // Pops and returns GenTree node from importers type stack.
7823 // Normalizes TYP_STRUCT value in case of GT_CALL, GT_RET_EXPR and arg nodes.
7824 GenTree* impSIMDPopStack(var_types type, bool expectAddr = false, CORINFO_CLASS_HANDLE structType = nullptr);
7826 // Create a GT_SIMD tree for a Get property of SIMD vector with a fixed index.
7827 GenTreeSIMD* impSIMDGetFixed(var_types simdType, var_types baseType, unsigned simdSize, int index);
7829 // Creates a GT_SIMD tree for Select operation
7830 GenTree* impSIMDSelect(CORINFO_CLASS_HANDLE typeHnd,
7832 unsigned simdVectorSize,
7837 // Creates a GT_SIMD tree for Min/Max operation
7838 GenTree* impSIMDMinMax(SIMDIntrinsicID intrinsicId,
7839 CORINFO_CLASS_HANDLE typeHnd,
7841 unsigned simdVectorSize,
7845 // Transforms operands and returns the SIMD intrinsic to be applied on
7846 // transformed operands to obtain given relop result.
7847 SIMDIntrinsicID impSIMDRelOp(SIMDIntrinsicID relOpIntrinsicId,
7848 CORINFO_CLASS_HANDLE typeHnd,
7849 unsigned simdVectorSize,
7850 var_types* baseType,
7854 // Creates a GT_SIMD tree for Abs intrinsic.
7855 GenTree* impSIMDAbs(CORINFO_CLASS_HANDLE typeHnd, var_types baseType, unsigned simdVectorSize, GenTree* op1);
7857 #if defined(_TARGET_XARCH_)
7859 // Transforms operands and returns the SIMD intrinsic to be applied on
7860 // transformed operands to obtain == comparison result.
7861 SIMDIntrinsicID impSIMDLongRelOpEqual(CORINFO_CLASS_HANDLE typeHnd,
7862 unsigned simdVectorSize,
7866 // Transforms operands and returns the SIMD intrinsic to be applied on
7867 // transformed operands to obtain > comparison result.
7868 SIMDIntrinsicID impSIMDLongRelOpGreaterThan(CORINFO_CLASS_HANDLE typeHnd,
7869 unsigned simdVectorSize,
7873 // Transforms operands and returns the SIMD intrinsic to be applied on
7874 // transformed operands to obtain >= comparison result.
7875 SIMDIntrinsicID impSIMDLongRelOpGreaterThanOrEqual(CORINFO_CLASS_HANDLE typeHnd,
7876 unsigned simdVectorSize,
7880 // Transforms operands and returns the SIMD intrinsic to be applied on
7881 // transformed operands to obtain >= comparison result in case of int32
7882 // and small int base type vectors.
7883 SIMDIntrinsicID impSIMDIntegralRelOpGreaterThanOrEqual(
7884 CORINFO_CLASS_HANDLE typeHnd, unsigned simdVectorSize, var_types baseType, GenTree** op1, GenTree** op2);
7886 #endif // defined(_TARGET_XARCH_)
7888 void setLclRelatedToSIMDIntrinsic(GenTree* tree);
7889 bool areFieldsContiguous(GenTree* op1, GenTree* op2);
7890 bool areArrayElementsContiguous(GenTree* op1, GenTree* op2);
7891 bool areArgumentsContiguous(GenTree* op1, GenTree* op2);
7892 GenTree* createAddressNodeForSIMDInit(GenTree* tree, unsigned simdSize);
7894 // check methodHnd to see if it is a SIMD method that is expanded as an intrinsic in the JIT.
7895 GenTree* impSIMDIntrinsic(OPCODE opcode,
7896 GenTree* newobjThis,
7897 CORINFO_CLASS_HANDLE clsHnd,
7898 CORINFO_METHOD_HANDLE method,
7899 CORINFO_SIG_INFO* sig,
7900 unsigned methodFlags,
7903 GenTree* getOp1ForConstructor(OPCODE opcode, GenTree* newobjThis, CORINFO_CLASS_HANDLE clsHnd);
7905 // Whether SIMD vector occupies part of SIMD register.
7906 // SSE2: vector2f/3f are considered sub register SIMD types.
7907 // AVX: vector2f, 3f and 4f are all considered sub register SIMD types.
7908 bool isSubRegisterSIMDType(CORINFO_CLASS_HANDLE typeHnd)
7910 unsigned sizeBytes = 0;
7911 var_types baseType = getBaseTypeAndSizeOfSIMDType(typeHnd, &sizeBytes);
7912 return (baseType == TYP_FLOAT) && (sizeBytes < getSIMDVectorRegisterByteLength());
7915 bool isSubRegisterSIMDType(GenTreeSIMD* simdNode)
7917 return (simdNode->gtSIMDSize < getSIMDVectorRegisterByteLength());
7920 // Get the type for the hardware SIMD vector.
7921 // This is the maximum SIMD type supported for this target.
7922 var_types getSIMDVectorType()
7924 #if defined(_TARGET_XARCH_)
7925 if (getSIMDSupportLevel() == SIMD_AVX2_Supported)
7931 assert(getSIMDSupportLevel() >= SIMD_SSE2_Supported);
7934 #elif defined(_TARGET_ARM64_)
7937 assert(!"getSIMDVectorType() unimplemented on target arch");
7942 // Get the size of the SIMD type in bytes
7943 int getSIMDTypeSizeInBytes(CORINFO_CLASS_HANDLE typeHnd)
7945 unsigned sizeBytes = 0;
7946 (void)getBaseTypeAndSizeOfSIMDType(typeHnd, &sizeBytes);
7950 // Get the the number of elements of basetype of SIMD vector given by its size and baseType
7951 static int getSIMDVectorLength(unsigned simdSize, var_types baseType);
7953 // Get the the number of elements of basetype of SIMD vector given by its type handle
7954 int getSIMDVectorLength(CORINFO_CLASS_HANDLE typeHnd);
7956 // Get preferred alignment of SIMD type.
7957 int getSIMDTypeAlignment(var_types simdType);
7959 // Get the number of bytes in a System.Numeric.Vector<T> for the current compilation.
7960 // Note - cannot be used for System.Runtime.Intrinsic
7961 unsigned getSIMDVectorRegisterByteLength()
7963 #if defined(_TARGET_XARCH_)
7964 if (getSIMDSupportLevel() == SIMD_AVX2_Supported)
7966 return YMM_REGSIZE_BYTES;
7970 assert(getSIMDSupportLevel() >= SIMD_SSE2_Supported);
7971 return XMM_REGSIZE_BYTES;
7973 #elif defined(_TARGET_ARM64_)
7974 return FP_REGSIZE_BYTES;
7976 assert(!"getSIMDVectorRegisterByteLength() unimplemented on target arch");
7981 // The minimum and maximum possible number of bytes in a SIMD vector.
7983 // maxSIMDStructBytes
7984 // The minimum SIMD size supported by System.Numeric.Vectors or System.Runtime.Intrinsic
7985 // SSE: 16-byte Vector<T> and Vector128<T>
7986 // AVX: 32-byte Vector256<T> (Vector<T> is 16-byte)
7987 // AVX2: 32-byte Vector<T> and Vector256<T>
7988 unsigned int maxSIMDStructBytes()
7990 #if defined(FEATURE_HW_INTRINSICS) && defined(_TARGET_XARCH_)
7991 if (compSupports(InstructionSet_AVX))
7993 return YMM_REGSIZE_BYTES;
7997 assert(getSIMDSupportLevel() >= SIMD_SSE2_Supported);
7998 return XMM_REGSIZE_BYTES;
8001 return getSIMDVectorRegisterByteLength();
8004 unsigned int minSIMDStructBytes()
8006 return emitTypeSize(TYP_SIMD8);
8009 // Returns the codegen type for a given SIMD size.
8010 var_types getSIMDTypeForSize(unsigned size)
8012 var_types simdType = TYP_UNDEF;
8015 simdType = TYP_SIMD8;
8017 else if (size == 12)
8019 simdType = TYP_SIMD12;
8021 else if (size == 16)
8023 simdType = TYP_SIMD16;
8025 else if (size == 32)
8027 simdType = TYP_SIMD32;
8031 noway_assert(!"Unexpected size for SIMD type");
8036 unsigned getSIMDInitTempVarNum()
8038 if (lvaSIMDInitTempVarNum == BAD_VAR_NUM)
8040 lvaSIMDInitTempVarNum = lvaGrabTempWithImplicitUse(false DEBUGARG("SIMDInitTempVar"));
8041 lvaTable[lvaSIMDInitTempVarNum].lvType = getSIMDVectorType();
8043 return lvaSIMDInitTempVarNum;
8046 #else // !FEATURE_SIMD
8047 bool isOpaqueSIMDLclVar(LclVarDsc* varDsc)
8051 #endif // FEATURE_SIMD
8054 //------------------------------------------------------------------------
8055 // largestEnregisterableStruct: The size in bytes of the largest struct that can be enregistered.
8057 // Notes: It is not guaranteed that the struct of this size or smaller WILL be a
8058 // candidate for enregistration.
8060 unsigned largestEnregisterableStructSize()
8063 unsigned vectorRegSize = getSIMDVectorRegisterByteLength();
8064 if (vectorRegSize > TARGET_POINTER_SIZE)
8066 return vectorRegSize;
8069 #endif // FEATURE_SIMD
8071 return TARGET_POINTER_SIZE;
8076 // These routines need not be enclosed under FEATURE_SIMD since lvIsSIMDType()
8077 // is defined for both FEATURE_SIMD and !FEATURE_SIMD apropriately. The use
8078 // of this routines also avoids the need of #ifdef FEATURE_SIMD specific code.
8080 // Is this var is of type simd struct?
8081 bool lclVarIsSIMDType(unsigned varNum)
8083 LclVarDsc* varDsc = lvaTable + varNum;
8084 return varDsc->lvIsSIMDType();
8087 // Is this Local node a SIMD local?
8088 bool lclVarIsSIMDType(GenTreeLclVarCommon* lclVarTree)
8090 return lclVarIsSIMDType(lclVarTree->gtLclNum);
8093 // Returns true if the TYP_SIMD locals on stack are aligned at their
8094 // preferred byte boundary specified by getSIMDTypeAlignment().
8096 // As per the Intel manual, the preferred alignment for AVX vectors is 32-bytes. On Amd64,
8097 // RSP/EBP is aligned at 16-bytes, therefore to align SIMD types at 32-bytes we need even
8098 // RSP/EBP to be 32-byte aligned. It is not clear whether additional stack space used in
8099 // aligning stack is worth the benefit and for now will use 16-byte alignment for AVX
8100 // 256-bit vectors with unaligned load/stores to/from memory. On x86, the stack frame
8101 // is aligned to 4 bytes. We need to extend existing support for double (8-byte) alignment
8102 // to 16 or 32 byte alignment for frames with local SIMD vars, if that is determined to be
8105 bool isSIMDTypeLocalAligned(unsigned varNum)
8107 #if defined(FEATURE_SIMD) && ALIGN_SIMD_TYPES
8108 if (lclVarIsSIMDType(varNum) && lvaTable[varNum].lvType != TYP_BYREF)
8111 int off = lvaFrameAddress(varNum, &ebpBased);
8112 // TODO-Cleanup: Can't this use the lvExactSize on the varDsc?
8113 int alignment = getSIMDTypeAlignment(lvaTable[varNum].lvType);
8114 bool isAligned = (alignment <= STACK_ALIGN) && ((off % alignment) == 0);
8117 #endif // FEATURE_SIMD
8122 bool compSupports(InstructionSet isa) const
8124 #if defined(_TARGET_XARCH_) || defined(_TARGET_ARM64_)
8125 return (opts.compSupportsISA & (1ULL << isa)) != 0;
8131 bool canUseVexEncoding() const
8133 #ifdef _TARGET_XARCH_
8134 return compSupports(InstructionSet_AVX);
8141 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8142 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8146 XX Generic info about the compilation and the method being compiled. XX
8147 XX It is responsible for driving the other phases. XX
8148 XX It is also responsible for all the memory management. XX
8150 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8151 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
8155 Compiler* InlineeCompiler; // The Compiler instance for the inlinee
8157 InlineResult* compInlineResult; // The result of importing the inlinee method.
8159 bool compDoAggressiveInlining; // If true, mark every method as CORINFO_FLG_FORCEINLINE
8160 bool compJmpOpUsed; // Does the method do a JMP
8161 bool compLongUsed; // Does the method use TYP_LONG
8162 bool compFloatingPointUsed; // Does the method use TYP_FLOAT or TYP_DOUBLE
8163 bool compTailCallUsed; // Does the method do a tailcall
8164 bool compLocallocUsed; // Does the method use localloc.
8165 bool compLocallocOptimized; // Does the method have an optimized localloc
8166 bool compQmarkUsed; // Does the method use GT_QMARK/GT_COLON
8167 bool compQmarkRationalized; // Is it allowed to use a GT_QMARK/GT_COLON node.
8168 bool compUnsafeCastUsed; // Does the method use LDIND/STIND to cast between scalar/refernce types
8170 // NOTE: These values are only reliable after
8171 // the importing is completely finished.
8174 // State information - which phases have completed?
8175 // These are kept together for easy discoverability
8177 bool bRangeAllowStress;
8178 bool compCodeGenDone;
8179 int64_t compNumStatementLinksTraversed; // # of links traversed while doing debug checks
8180 bool fgNormalizeEHDone; // Has the flowgraph EH normalization phase been done?
8181 size_t compSizeEstimate; // The estimated size of the method as per `gtSetEvalOrder`.
8182 size_t compCycleEstimate; // The estimated cycle count of the method as per `gtSetEvalOrder`
8185 bool fgLocalVarLivenessDone; // Note that this one is used outside of debug.
8186 bool fgLocalVarLivenessChanged;
8188 bool compStackProbePrologDone;
8191 bool compRationalIRForm;
8193 bool compUsesThrowHelper; // There is a call to a THOROW_HELPER for the compiled method.
8195 bool compGeneratingProlog;
8196 bool compGeneratingEpilog;
8197 bool compNeedsGSSecurityCookie; // There is an unsafe buffer (or localloc) on the stack.
8198 // Insert cookie on frame and code to check the cookie, like VC++ -GS.
8199 bool compGSReorderStackLayout; // There is an unsafe buffer on the stack, reorder locals and make local
8200 // copies of susceptible parameters to avoid buffer overrun attacks through locals/params
8201 bool getNeedsGSSecurityCookie() const
8203 return compNeedsGSSecurityCookie;
8205 void setNeedsGSSecurityCookie()
8207 compNeedsGSSecurityCookie = true;
8210 FrameLayoutState lvaDoneFrameLayout; // The highest frame layout state that we've completed. During
8211 // frame layout calculations, this is the level we are currently
8214 //---------------------------- JITing options -----------------------------
8227 JitFlags* jitFlags; // all flags passed from the EE
8228 unsigned compFlags; // method attributes
8230 codeOptimize compCodeOpt; // what type of code optimizations
8235 #if defined(_TARGET_XARCH_) || defined(_TARGET_ARM64_)
8236 uint64_t compSupportsISA;
8237 void setSupportedISA(InstructionSet isa)
8239 compSupportsISA |= 1ULL << isa;
8243 // optimize maximally and/or favor speed over size?
8245 #define DEFAULT_MIN_OPTS_CODE_SIZE 60000
8246 #define DEFAULT_MIN_OPTS_INSTR_COUNT 20000
8247 #define DEFAULT_MIN_OPTS_BB_COUNT 2000
8248 #define DEFAULT_MIN_OPTS_LV_NUM_COUNT 2000
8249 #define DEFAULT_MIN_OPTS_LV_REF_COUNT 8000
8251 // Maximun number of locals before turning off the inlining
8252 #define MAX_LV_NUM_COUNT_FOR_INLINING 512
8255 unsigned instrCount;
8256 unsigned lvRefCount;
8257 bool compMinOptsIsSet;
8259 bool compMinOptsIsUsed;
8261 inline bool MinOpts()
8263 assert(compMinOptsIsSet);
8264 compMinOptsIsUsed = true;
8267 inline bool IsMinOptsSet()
8269 return compMinOptsIsSet;
8272 inline bool MinOpts()
8276 inline bool IsMinOptsSet()
8278 return compMinOptsIsSet;
8281 inline void SetMinOpts(bool val)
8283 assert(!compMinOptsIsUsed);
8284 assert(!compMinOptsIsSet || (compMinOpts == val));
8286 compMinOptsIsSet = true;
8289 // true if the CLFLG_* for an optimization is set.
8290 inline bool OptEnabled(unsigned optFlag)
8292 return !!(compFlags & optFlag);
8295 #ifdef FEATURE_READYTORUN_COMPILER
8296 inline bool IsReadyToRun()
8298 return jitFlags->IsSet(JitFlags::JIT_FLAG_READYTORUN);
8301 inline bool IsReadyToRun()
8307 // true if we should use the PINVOKE_{BEGIN,END} helpers instead of generating
8308 // PInvoke transitions inline (e.g. when targeting CoreRT).
8309 inline bool ShouldUsePInvokeHelpers()
8311 return jitFlags->IsSet(JitFlags::JIT_FLAG_USE_PINVOKE_HELPERS);
8314 // true if we should use insert the REVERSE_PINVOKE_{ENTER,EXIT} helpers in the method
8316 inline bool IsReversePInvoke()
8318 return jitFlags->IsSet(JitFlags::JIT_FLAG_REVERSE_PINVOKE);
8321 // true if we must generate code compatible with JIT32 quirks
8322 inline bool IsJit32Compat()
8324 #if defined(_TARGET_X86_)
8325 return jitFlags->IsSet(JitFlags::JIT_FLAG_DESKTOP_QUIRKS);
8331 // true if we must generate code compatible with Jit64 quirks
8332 inline bool IsJit64Compat()
8334 #if defined(_TARGET_AMD64_)
8335 return jitFlags->IsSet(JitFlags::JIT_FLAG_DESKTOP_QUIRKS);
8336 #elif !defined(FEATURE_CORECLR)
8343 bool compScopeInfo; // Generate the LocalVar info ?
8344 bool compDbgCode; // Generate debugger-friendly code?
8345 bool compDbgInfo; // Gather debugging info?
8348 #ifdef PROFILING_SUPPORTED
8349 bool compNoPInvokeInlineCB;
8351 static const bool compNoPInvokeInlineCB;
8355 bool compGcChecks; // Check arguments and return values to ensure they are sane
8358 #if defined(DEBUG) && defined(_TARGET_XARCH_)
8360 bool compStackCheckOnRet; // Check stack pointer on return to ensure it is correct.
8362 #endif // defined(DEBUG) && defined(_TARGET_XARCH_)
8364 #if defined(DEBUG) && defined(_TARGET_X86_)
8366 bool compStackCheckOnCall; // Check stack pointer after call to ensure it is correct. Only for x86.
8368 #endif // defined(DEBUG) && defined(_TARGET_X86_)
8370 bool compNeedSecurityCheck; // This flag really means where or not a security object needs
8371 // to be allocated on the stack.
8372 // It will be set to true in the following cases:
8373 // 1. When the method being compiled has a declarative security
8374 // (i.e. when CORINFO_FLG_NOSECURITYWRAP is reset for the current method).
8375 // This is also the case when we inject a prolog and epilog in the method.
8377 // 2. When the method being compiled has imperative security (i.e. the method
8378 // calls into another method that has CORINFO_FLG_SECURITYCHECK flag set).
8380 // 3. When opts.compDbgEnC is true. (See also Compiler::compCompile).
8382 // When this flag is set, jit will allocate a gc-reference local variable (lvaSecurityObject),
8383 // which gets reported as a GC root to stackwalker.
8384 // (See also ICodeManager::GetAddrOfSecurityObject.)
8386 bool compReloc; // Generate relocs for pointers in code, true for all ngen/prejit codegen
8389 #if defined(_TARGET_XARCH_)
8390 bool compEnablePCRelAddr; // Whether absolute addr be encoded as PC-rel offset by RyuJIT where possible
8394 #ifdef UNIX_AMD64_ABI
8395 // This flag is indicating if there is a need to align the frame.
8396 // On AMD64-Windows, if there are calls, 4 slots for the outgoing ars are allocated, except for
8397 // FastTailCall. This slots makes the frame size non-zero, so alignment logic will be called.
8398 // On AMD64-Unix, there are no such slots. There is a possibility to have calls in the method with frame size of
8399 // 0. The frame alignment logic won't kick in. This flags takes care of the AMD64-Unix case by remembering that
8400 // there are calls and making sure the frame alignment logic is executed.
8401 bool compNeedToAlignFrame;
8402 #endif // UNIX_AMD64_ABI
8404 bool compProcedureSplitting; // Separate cold code from hot code
8406 bool genFPorder; // Preserve FP order (operations are non-commutative)
8407 bool genFPopt; // Can we do frame-pointer-omission optimization?
8408 bool altJit; // True if we are an altjit and are compiling this method
8411 bool optRepeat; // Repeat optimizer phases k times
8415 bool compProcedureSplittingEH; // Separate cold code from hot code for functions with EH
8416 bool dspCode; // Display native code generated
8417 bool dspEHTable; // Display the EH table reported to the VM
8418 bool dspDebugInfo; // Display the Debug info reported to the VM
8419 bool dspInstrs; // Display the IL instructions intermixed with the native code output
8420 bool dspEmit; // Display emitter output
8421 bool dspLines; // Display source-code lines intermixed with native code output
8422 bool dmpHex; // Display raw bytes in hex of native code output
8423 bool varNames; // Display variables names in native code output
8424 bool disAsm; // Display native code as it is generated
8425 bool disAsmSpilled; // Display native code when any register spilling occurs
8426 bool disDiffable; // Makes the Disassembly code 'diff-able'
8427 bool disAsm2; // Display native code after it is generated using external disassembler
8428 bool dspOrder; // Display names of each of the methods that we ngen/jit
8429 bool dspUnwind; // Display the unwind info output
8430 bool dspDiffable; // Makes the Jit Dump 'diff-able' (currently uses same COMPlus_* flag as disDiffable)
8431 bool compLongAddress; // Force using large pseudo instructions for long address
8432 // (IF_LARGEJMP/IF_LARGEADR/IF_LARGLDC)
8433 bool dspGCtbls; // Display the GC tables
8437 bool doLateDisasm; // Run the late disassembler
8438 #endif // LATE_DISASM
8440 #if DUMP_GC_TABLES && !defined(DEBUG) && defined(JIT32_GCENCODER)
8441 // Only the JIT32_GCENCODER implements GC dumping in non-DEBUG code.
8442 #pragma message("NOTE: this non-debug build has GC ptr table dumping always enabled!")
8443 static const bool dspGCtbls = true;
8446 // We need stack probes to guarantee that we won't trigger a stack overflow
8447 // when calling unmanaged code until they get a chance to set up a frame, because
8448 // the EE will have no idea where it is.
8450 // We will only be doing this currently for hosted environments. Unfortunately
8451 // we need to take care of stubs, so potentially, we will have to do the probes
8452 // for any call. We have a plan for not needing for stubs though
8453 bool compNeedStackProbes;
8455 #ifdef PROFILING_SUPPORTED
8456 // Whether to emit Enter/Leave/TailCall hooks using a dummy stub (DummyProfilerELTStub()).
8457 // This option helps make the JIT behave as if it is running under a profiler.
8458 bool compJitELTHookEnabled;
8459 #endif // PROFILING_SUPPORTED
8461 #if FEATURE_TAILCALL_OPT
8462 // Whether opportunistic or implicit tail call optimization is enabled.
8463 bool compTailCallOpt;
8464 // Whether optimization of transforming a recursive tail call into a loop is enabled.
8465 bool compTailCallLoopOpt;
8469 static const bool compUseSoftFP = true;
8470 #else // !ARM_SOFTFP
8471 static const bool compUseSoftFP = false;
8474 GCPollType compGCPollType;
8478 static bool s_pAltJitExcludeAssembliesListInitialized;
8479 static AssemblyNamesList2* s_pAltJitExcludeAssembliesList;
8483 static bool s_pJitDisasmIncludeAssembliesListInitialized;
8484 static AssemblyNamesList2* s_pJitDisasmIncludeAssembliesList;
8488 // silence warning of cast to greater size. It is easier to silence than construct code the compiler is happy with, and
8489 // it is safe in this case
8490 #pragma warning(push)
8491 #pragma warning(disable : 4312)
8493 template <typename T>
8496 return (p == ZERO) ? ZERO : (opts.dspDiffable ? T(0xD1FFAB1E) : p);
8499 template <typename T>
8502 return (o == ZERO) ? ZERO : (opts.dspDiffable ? T(0xD1FFAB1E) : o);
8504 #pragma warning(pop)
8506 static int dspTreeID(GenTree* tree)
8508 return tree->gtTreeID;
8510 static void printTreeID(GenTree* tree)
8512 if (tree == nullptr)
8518 printf("[%06d]", dspTreeID(tree));
8525 #define STRESS_MODES \
8529 /* "Variations" stress areas which we try to mix up with each other. */ \
8530 /* These should not be exhaustively used as they might */ \
8531 /* hide/trivialize other areas */ \
8534 STRESS_MODE(DBL_ALN) \
8535 STRESS_MODE(LCL_FLDS) \
8536 STRESS_MODE(UNROLL_LOOPS) \
8537 STRESS_MODE(MAKE_CSE) \
8538 STRESS_MODE(LEGACY_INLINE) \
8539 STRESS_MODE(CLONE_EXPR) \
8540 STRESS_MODE(USE_FCOMI) \
8541 STRESS_MODE(USE_CMOV) \
8543 STRESS_MODE(BB_PROFILE) \
8544 STRESS_MODE(OPT_BOOLS_GC) \
8545 STRESS_MODE(REMORPH_TREES) \
8546 STRESS_MODE(64RSLT_MUL) \
8547 STRESS_MODE(DO_WHILE_LOOPS) \
8548 STRESS_MODE(MIN_OPTS) \
8549 STRESS_MODE(REVERSE_FLAG) /* Will set GTF_REVERSE_OPS whenever we can */ \
8550 STRESS_MODE(REVERSE_COMMA) /* Will reverse commas created with gtNewCommaNode */ \
8551 STRESS_MODE(TAILCALL) /* Will make the call as a tailcall whenever legal */ \
8552 STRESS_MODE(CATCH_ARG) /* Will spill catch arg */ \
8553 STRESS_MODE(UNSAFE_BUFFER_CHECKS) \
8554 STRESS_MODE(NULL_OBJECT_CHECK) \
8555 STRESS_MODE(PINVOKE_RESTORE_ESP) \
8556 STRESS_MODE(RANDOM_INLINE) \
8557 STRESS_MODE(SWITCH_CMP_BR_EXPANSION) \
8558 STRESS_MODE(GENERIC_VARN) \
8560 /* After COUNT_VARN, stress level 2 does all of these all the time */ \
8562 STRESS_MODE(COUNT_VARN) \
8564 /* "Check" stress areas that can be exhaustively used if we */ \
8565 /* dont care about performance at all */ \
8567 STRESS_MODE(FORCE_INLINE) /* Treat every method as AggressiveInlining */ \
8568 STRESS_MODE(CHK_FLOW_UPDATE) \
8569 STRESS_MODE(EMITTER) \
8570 STRESS_MODE(CHK_REIMPORT) \
8571 STRESS_MODE(FLATFP) \
8572 STRESS_MODE(GENERIC_CHECK) \
8577 #define STRESS_MODE(mode) STRESS_##mode,
8584 static const LPCWSTR s_compStressModeNames[STRESS_COUNT + 1];
8585 BYTE compActiveStressModes[STRESS_COUNT];
8588 #define MAX_STRESS_WEIGHT 100
8590 bool compStressCompile(compStressArea stressArea, unsigned weightPercentage);
8594 bool compInlineStress()
8596 return compStressCompile(STRESS_LEGACY_INLINE, 50);
8599 bool compRandomInlineStress()
8601 return compStressCompile(STRESS_RANDOM_INLINE, 50);
8606 bool compTailCallStress()
8609 return (JitConfig.TailcallStress() != 0 || compStressCompile(STRESS_TAILCALL, 5));
8615 codeOptimize compCodeOpt()
8618 // Switching between size & speed has measurable throughput impact
8619 // (3.5% on NGen mscorlib when measured). It used to be enabled for
8620 // DEBUG, but should generate identical code between CHK & RET builds,
8621 // so that's not acceptable.
8622 // TODO-Throughput: Figure out what to do about size vs. speed & throughput.
8623 // Investigate the cause of the throughput regression.
8625 return opts.compCodeOpt;
8627 return BLENDED_CODE;
8631 //--------------------- Info about the procedure --------------------------
8635 COMP_HANDLE compCompHnd;
8636 CORINFO_MODULE_HANDLE compScopeHnd;
8637 CORINFO_CLASS_HANDLE compClassHnd;
8638 CORINFO_METHOD_HANDLE compMethodHnd;
8639 CORINFO_METHOD_INFO* compMethodInfo;
8641 BOOL hasCircularClassConstraints;
8642 BOOL hasCircularMethodConstraints;
8644 #if defined(DEBUG) || defined(LATE_DISASM)
8645 const char* compMethodName;
8646 const char* compClassName;
8647 const char* compFullName;
8648 #endif // defined(DEBUG) || defined(LATE_DISASM)
8650 #if defined(DEBUG) || defined(INLINE_DATA)
8651 // Method hash is logcally const, but computed
8653 mutable unsigned compMethodHashPrivate;
8654 unsigned compMethodHash() const;
8655 #endif // defined(DEBUG) || defined(INLINE_DATA)
8657 #ifdef PSEUDORANDOM_NOP_INSERTION
8658 // things for pseudorandom nop insertion
8659 unsigned compChecksum;
8663 // The following holds the FLG_xxxx flags for the method we're compiling.
8666 // The following holds the class attributes for the method we're compiling.
8667 unsigned compClassAttr;
8669 const BYTE* compCode;
8670 IL_OFFSET compILCodeSize; // The IL code size
8671 UNATIVE_OFFSET compNativeCodeSize; // The native code size, after instructions are issued. This
8672 // is less than (compTotalHotCodeSize + compTotalColdCodeSize) only if:
8673 // (1) the code is not hot/cold split, and we issued less code than we expected, or
8674 // (2) the code is hot/cold split, and we issued less code than we expected
8675 // in the cold section (the hot section will always be padded out to compTotalHotCodeSize).
8677 bool compIsStatic : 1; // Is the method static (no 'this' pointer)?
8678 bool compIsVarArgs : 1; // Does the method have varargs parameters?
8679 bool compIsContextful : 1; // contextful method
8680 bool compInitMem : 1; // Is the CORINFO_OPT_INIT_LOCALS bit set in the method info options?
8681 bool compUnwrapContextful : 1; // JIT should unwrap proxies when possible
8682 bool compProfilerCallback : 1; // JIT inserted a profiler Enter callback
8683 bool compPublishStubParam : 1; // EAX captured in prolog will be available through an instrinsic
8684 bool compRetBuffDefStack : 1; // The ret buff argument definitely points into the stack.
8686 var_types compRetType; // Return type of the method as declared in IL
8687 var_types compRetNativeType; // Normalized return type as per target arch ABI
8688 unsigned compILargsCount; // Number of arguments (incl. implicit but not hidden)
8689 unsigned compArgsCount; // Number of arguments (incl. implicit and hidden)
8691 #if FEATURE_FASTTAILCALL
8692 size_t compArgStackSize; // Incoming argument stack size in bytes
8693 #endif // FEATURE_FASTTAILCALL
8695 unsigned compRetBuffArg; // position of hidden return param var (0, 1) (BAD_VAR_NUM means not present);
8696 int compTypeCtxtArg; // position of hidden param for type context for generic code (CORINFO_CALLCONV_PARAMTYPE)
8697 unsigned compThisArg; // position of implicit this pointer param (not to be confused with lvaArg0Var)
8698 unsigned compILlocalsCount; // Number of vars : args + locals (incl. implicit but not hidden)
8699 unsigned compLocalsCount; // Number of vars : args + locals (incl. implicit and hidden)
8700 unsigned compMaxStack;
8701 UNATIVE_OFFSET compTotalHotCodeSize; // Total number of bytes of Hot Code in the method
8702 UNATIVE_OFFSET compTotalColdCodeSize; // Total number of bytes of Cold Code in the method
8704 unsigned compCallUnmanaged; // count of unmanaged calls
8705 unsigned compLvFrameListRoot; // lclNum for the Frame root
8706 unsigned compXcptnsCount; // Number of exception-handling clauses read in the method's IL.
8707 // You should generally use compHndBBtabCount instead: it is the
8708 // current number of EH clauses (after additions like synchronized
8709 // methods and funclets, and removals like unreachable code deletion).
8711 bool compMatchedVM; // true if the VM is "matched": either the JIT is a cross-compiler
8712 // and the VM expects that, or the JIT is a "self-host" compiler
8713 // (e.g., x86 hosted targeting x86) and the VM expects that.
8715 /* The following holds IL scope information about local variables.
8718 unsigned compVarScopesCount;
8719 VarScopeDsc* compVarScopes;
8721 /* The following holds information about instr offsets for
8722 * which we need to report IP-mappings
8725 IL_OFFSET* compStmtOffsets; // sorted
8726 unsigned compStmtOffsetsCount;
8727 ICorDebugInfo::BoundaryTypes compStmtOffsetsImplicit;
8729 #define CPU_X86 0x0100 // The generic X86 CPU
8730 #define CPU_X86_PENTIUM_4 0x0110
8732 #define CPU_X64 0x0200 // The generic x64 CPU
8733 #define CPU_AMD_X64 0x0210 // AMD x64 CPU
8734 #define CPU_INTEL_X64 0x0240 // Intel x64 CPU
8736 #define CPU_ARM 0x0300 // The generic ARM CPU
8737 #define CPU_ARM64 0x0400 // The generic ARM64 CPU
8739 unsigned genCPU; // What CPU are we running on
8742 // Returns true if the method being compiled returns a non-void and non-struct value.
8743 // Note that lvaInitTypeRef() normalizes compRetNativeType for struct returns in a
8744 // single register as per target arch ABI (e.g on Amd64 Windows structs of size 1, 2,
8745 // 4 or 8 gets normalized to TYP_BYTE/TYP_SHORT/TYP_INT/TYP_LONG; On Arm HFA structs).
8746 // Methods returning such structs are considered to return non-struct return value and
8747 // this method returns true in that case.
8748 bool compMethodReturnsNativeScalarType()
8750 return (info.compRetType != TYP_VOID) && !varTypeIsStruct(info.compRetNativeType);
8753 // Returns true if the method being compiled returns RetBuf addr as its return value
8754 bool compMethodReturnsRetBufAddr()
8756 // There are cases where implicit RetBuf argument should be explicitly returned in a register.
8757 // In such cases the return type is changed to TYP_BYREF and appropriate IR is generated.
8759 // 1. Profiler Leave calllback expects the address of retbuf as return value for
8760 // methods with hidden RetBuf argument. impReturnInstruction() when profiler
8761 // callbacks are needed creates GT_RETURN(TYP_BYREF, op1 = Addr of RetBuf) for
8762 // methods with hidden RetBufArg.
8764 // 2. As per the System V ABI, the address of RetBuf needs to be returned by
8765 // methods with hidden RetBufArg in RAX. In such case GT_RETURN is of TYP_BYREF,
8766 // returning the address of RetBuf.
8768 // 3. Windows 64-bit native calling convention also requires the address of RetBuff
8769 // to be returned in RAX.
8770 CLANG_FORMAT_COMMENT_ANCHOR;
8772 #ifdef _TARGET_AMD64_
8773 return (info.compRetBuffArg != BAD_VAR_NUM);
8774 #else // !_TARGET_AMD64_
8775 return (compIsProfilerHookNeeded()) && (info.compRetBuffArg != BAD_VAR_NUM);
8776 #endif // !_TARGET_AMD64_
8779 // Returns true if the method returns a value in more than one return register
8780 // TODO-ARM-Bug: Deal with multi-register genReturnLocaled structs?
8781 // TODO-ARM64: Does this apply for ARM64 too?
8782 bool compMethodReturnsMultiRegRetType()
8784 #if FEATURE_MULTIREG_RET
8785 #if defined(_TARGET_X86_)
8786 // On x86 only 64-bit longs are returned in multiple registers
8787 return varTypeIsLong(info.compRetNativeType);
8788 #else // targets: X64-UNIX, ARM64 or ARM32
8789 // On all other targets that support multireg return values:
8790 // Methods returning a struct in multiple registers have a return value of TYP_STRUCT.
8791 // Such method's compRetNativeType is TYP_STRUCT without a hidden RetBufArg
8792 return varTypeIsStruct(info.compRetNativeType) && (info.compRetBuffArg == BAD_VAR_NUM);
8793 #endif // TARGET_XXX
8795 #else // not FEATURE_MULTIREG_RET
8797 // For this architecture there are no multireg returns
8800 #endif // FEATURE_MULTIREG_RET
8803 #if FEATURE_MULTIREG_ARGS
8804 // Given a GenTree node of TYP_STRUCT that represents a pass by value argument
8805 // return the gcPtr layout for the pointers sized fields
8806 void getStructGcPtrsFromOp(GenTree* op, BYTE* gcPtrsOut);
8807 #endif // FEATURE_MULTIREG_ARGS
8809 // Returns true if the method being compiled returns a value
8810 bool compMethodHasRetVal()
8812 return compMethodReturnsNativeScalarType() || compMethodReturnsRetBufAddr() ||
8813 compMethodReturnsMultiRegRetType();
8818 void compDispLocalVars();
8822 //-------------------------- Global Compiler Data ------------------------------------
8825 static unsigned s_compMethodsCount; // to produce unique label names
8826 unsigned compGenTreeID;
8827 unsigned compBasicBlockID;
8830 BasicBlock* compCurBB; // the current basic block in process
8831 GenTree* compCurStmt; // the current statement in process
8833 unsigned compCurStmtNum; // to give all statements an increasing StmtNum when printing dumps
8836 // The following is used to create the 'method JIT info' block.
8837 size_t compInfoBlkSize;
8838 BYTE* compInfoBlkAddr;
8840 EHblkDsc* compHndBBtab; // array of EH data
8841 unsigned compHndBBtabCount; // element count of used elements in EH data array
8842 unsigned compHndBBtabAllocCount; // element count of allocated elements in EH data array
8844 #if defined(_TARGET_X86_)
8846 //-------------------------------------------------------------------------
8847 // Tracking of region covered by the monitor in synchronized methods
8848 void* syncStartEmitCookie; // the emitter cookie for first instruction after the call to MON_ENTER
8849 void* syncEndEmitCookie; // the emitter cookie for first instruction after the call to MON_EXIT
8851 #endif // !_TARGET_X86_
8853 Phases previousCompletedPhase; // the most recently completed phase
8855 //-------------------------------------------------------------------------
8856 // The following keeps track of how many bytes of local frame space we've
8857 // grabbed so far in the current function, and how many argument bytes we
8858 // need to pop when we return.
8861 unsigned compLclFrameSize; // secObject+lclBlk+locals+temps
8863 // Count of callee-saved regs we pushed in the prolog.
8864 // Does not include EBP for isFramePointerUsed() and double-aligned frames.
8865 // In case of Amd64 this doesn't include float regs saved on stack.
8866 unsigned compCalleeRegsPushed;
8868 #if defined(_TARGET_XARCH_)
8869 // Mask of callee saved float regs on stack.
8870 regMaskTP compCalleeFPRegsSavedMask;
8872 #ifdef _TARGET_AMD64_
8873 // Quirk for VS debug-launch scenario to work:
8874 // Bytes of padding between save-reg area and locals.
8875 #define VSQUIRK_STACK_PAD (2 * REGSIZE_BYTES)
8876 unsigned compVSQuirkStackPaddingNeeded;
8877 bool compQuirkForPPPflag;
8880 unsigned compArgSize; // total size of arguments in bytes (including register args (lvIsRegArg))
8882 unsigned compMapILargNum(unsigned ILargNum); // map accounting for hidden args
8883 unsigned compMapILvarNum(unsigned ILvarNum); // map accounting for hidden args
8884 unsigned compMap2ILvarNum(unsigned varNum); // map accounting for hidden args
8886 //-------------------------------------------------------------------------
8888 static void compStartup(); // One-time initialization
8889 static void compShutdown(); // One-time finalization
8891 void compInit(ArenaAllocator* pAlloc, InlineInfo* inlineInfo);
8894 static void compDisplayStaticSizes(FILE* fout);
8896 //------------ Some utility functions --------------
8898 void* compGetHelperFtn(CorInfoHelpFunc ftnNum, /* IN */
8899 void** ppIndirection); /* OUT */
8901 // Several JIT/EE interface functions return a CorInfoType, and also return a
8902 // class handle as an out parameter if the type is a value class. Returns the
8903 // size of the type these describe.
8904 unsigned compGetTypeSize(CorInfoType cit, CORINFO_CLASS_HANDLE clsHnd);
8907 // Components used by the compiler may write unit test suites, and
8908 // have them run within this method. They will be run only once per process, and only
8909 // in debug. (Perhaps should be under the control of a COMPlus_ flag.)
8910 // These should fail by asserting.
8911 void compDoComponentUnitTestsOnce();
8914 int compCompile(CORINFO_METHOD_HANDLE methodHnd,
8915 CORINFO_MODULE_HANDLE classPtr,
8916 COMP_HANDLE compHnd,
8917 CORINFO_METHOD_INFO* methodInfo,
8918 void** methodCodePtr,
8919 ULONG* methodCodeSize,
8920 JitFlags* compileFlags);
8921 void compCompileFinish();
8922 int compCompileHelper(CORINFO_MODULE_HANDLE classPtr,
8923 COMP_HANDLE compHnd,
8924 CORINFO_METHOD_INFO* methodInfo,
8925 void** methodCodePtr,
8926 ULONG* methodCodeSize,
8927 JitFlags* compileFlags,
8928 CorInfoInstantiationVerification instVerInfo);
8930 ArenaAllocator* compGetArenaAllocator();
8932 #if MEASURE_MEM_ALLOC
8933 static bool s_dspMemStats; // Display per-phase memory statistics for every function
8934 #endif // MEASURE_MEM_ALLOC
8936 #if LOOP_HOIST_STATS
8937 unsigned m_loopsConsidered;
8938 bool m_curLoopHasHoistedExpression;
8939 unsigned m_loopsWithHoistedExpressions;
8940 unsigned m_totalHoistedExpressions;
8942 void AddLoopHoistStats();
8943 void PrintPerMethodLoopHoistStats();
8945 static CritSecObject s_loopHoistStatsLock; // This lock protects the data structures below.
8946 static unsigned s_loopsConsidered;
8947 static unsigned s_loopsWithHoistedExpressions;
8948 static unsigned s_totalHoistedExpressions;
8950 static void PrintAggregateLoopHoistStats(FILE* f);
8951 #endif // LOOP_HOIST_STATS
8953 bool compIsForImportOnly();
8954 bool compIsForInlining();
8955 bool compDonotInline();
8958 unsigned char compGetJitDefaultFill(); // Get the default fill char value
8959 // we randomize this value when JitStress is enabled
8961 const char* compLocalVarName(unsigned varNum, unsigned offs);
8962 VarName compVarName(regNumber reg, bool isFloatReg = false);
8963 const char* compRegVarName(regNumber reg, bool displayVar = false, bool isFloatReg = false);
8964 const char* compRegNameForSize(regNumber reg, size_t size);
8965 const char* compFPregVarName(unsigned fpReg, bool displayVar = false);
8966 void compDspSrcLinesByNativeIP(UNATIVE_OFFSET curIP);
8967 void compDspSrcLinesByLineNum(unsigned line, bool seek = false);
8970 //-------------------------------------------------------------------------
8972 struct VarScopeListNode
8975 VarScopeListNode* next;
8976 static VarScopeListNode* Create(VarScopeDsc* value, CompAllocator alloc)
8978 VarScopeListNode* node = new (alloc) VarScopeListNode;
8980 node->next = nullptr;
8985 struct VarScopeMapInfo
8987 VarScopeListNode* head;
8988 VarScopeListNode* tail;
8989 static VarScopeMapInfo* Create(VarScopeListNode* node, CompAllocator alloc)
8991 VarScopeMapInfo* info = new (alloc) VarScopeMapInfo;
8998 // Max value of scope count for which we would use linear search; for larger values we would use hashtable lookup.
8999 static const unsigned MAX_LINEAR_FIND_LCL_SCOPELIST = 32;
9001 typedef JitHashTable<unsigned, JitSmallPrimitiveKeyFuncs<unsigned>, VarScopeMapInfo*> VarNumToScopeDscMap;
9003 // Map to keep variables' scope indexed by varNum containing it's scope dscs at the index.
9004 VarNumToScopeDscMap* compVarScopeMap;
9006 VarScopeDsc* compFindLocalVar(unsigned varNum, unsigned lifeBeg, unsigned lifeEnd);
9008 VarScopeDsc* compFindLocalVar(unsigned varNum, unsigned offs);
9010 VarScopeDsc* compFindLocalVarLinear(unsigned varNum, unsigned offs);
9012 void compInitVarScopeMap();
9014 VarScopeDsc** compEnterScopeList; // List has the offsets where variables
9015 // enter scope, sorted by instr offset
9016 unsigned compNextEnterScope;
9018 VarScopeDsc** compExitScopeList; // List has the offsets where variables
9019 // go out of scope, sorted by instr offset
9020 unsigned compNextExitScope;
9022 void compInitScopeLists();
9024 void compResetScopeLists();
9026 VarScopeDsc* compGetNextEnterScope(unsigned offs, bool scan = false);
9028 VarScopeDsc* compGetNextExitScope(unsigned offs, bool scan = false);
9030 void compProcessScopesUntil(unsigned offset,
9032 void (Compiler::*enterScopeFn)(VARSET_TP* inScope, VarScopeDsc*),
9033 void (Compiler::*exitScopeFn)(VARSET_TP* inScope, VarScopeDsc*));
9036 void compDispScopeLists();
9039 bool compIsProfilerHookNeeded();
9041 //-------------------------------------------------------------------------
9042 /* Statistical Data Gathering */
9044 void compJitStats(); // call this function and enable
9045 // various ifdef's below for statistical data
9048 void compCallArgStats();
9049 static void compDispCallArgStats(FILE* fout);
9052 //-------------------------------------------------------------------------
9059 ArenaAllocator* compArenaAllocator;
9062 void compFunctionTraceStart();
9063 void compFunctionTraceEnd(void* methodCodePtr, ULONG methodCodeSize, bool isNYI);
9066 size_t compMaxUncheckedOffsetForNullObject;
9068 void compInitOptions(JitFlags* compileFlags);
9070 void compSetProcessor();
9071 void compInitDebuggingInfo();
9072 void compSetOptimizationLevel();
9073 #ifdef _TARGET_ARMARCH_
9074 bool compRsvdRegCheck(FrameLayoutState curState);
9076 void compCompile(void** methodCodePtr, ULONG* methodCodeSize, JitFlags* compileFlags);
9078 // Clear annotations produced during optimizations; to be used between iterations when repeating opts.
9079 void ResetOptAnnotations();
9081 // Regenerate loop descriptors; to be used between iterations when repeating opts.
9082 void RecomputeLoopInfo();
9084 #ifdef PROFILING_SUPPORTED
9085 // Data required for generating profiler Enter/Leave/TailCall hooks
9087 bool compProfilerHookNeeded; // Whether profiler Enter/Leave/TailCall hook needs to be generated for the method
9088 void* compProfilerMethHnd; // Profiler handle of the method being compiled. Passed as param to ELT callbacks
9089 bool compProfilerMethHndIndirected; // Whether compProfilerHandle is pointer to the handle or is an actual handle
9092 #ifdef _TARGET_AMD64_
9093 bool compQuirkForPPP(); // Check if this method should be Quirked for the PPP issue
9096 // Assumes called as part of process shutdown; does any compiler-specific work associated with that.
9097 static void ProcessShutdownWork(ICorStaticInfo* statInfo);
9099 CompAllocator getAllocator(CompMemKind cmk = CMK_Generic)
9101 return CompAllocator(compArenaAllocator, cmk);
9104 CompAllocator getAllocatorGC()
9106 return getAllocator(CMK_GC);
9109 CompAllocator getAllocatorLoopHoist()
9111 return getAllocator(CMK_LoopHoist);
9115 CompAllocator getAllocatorDebugOnly()
9117 return getAllocator(CMK_DebugOnly);
9122 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9123 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9127 XX Checks for type compatibility and merges types XX
9129 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9130 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9134 // Set to TRUE if verification cannot be skipped for this method
9135 // If we detect unverifiable code, we will lazily check
9136 // canSkipMethodVerification() to see if verification is REALLY needed.
9137 BOOL tiVerificationNeeded;
9139 // It it initially TRUE, and it gets set to FALSE if we run into unverifiable code
9140 // Note that this is valid only if tiVerificationNeeded was ever TRUE.
9141 BOOL tiIsVerifiableCode;
9143 // Set to TRUE if runtime callout is needed for this method
9144 BOOL tiRuntimeCalloutNeeded;
9146 // Set to TRUE if security prolog/epilog callout is needed for this method
9147 // Note: This flag is different than compNeedSecurityCheck.
9148 // compNeedSecurityCheck means whether or not a security object needs
9149 // to be allocated on the stack, which is currently true for EnC as well.
9150 // tiSecurityCalloutNeeded means whether or not security callouts need
9151 // to be inserted in the jitted code.
9152 BOOL tiSecurityCalloutNeeded;
9154 // Returns TRUE if child is equal to or a subtype of parent for merge purposes
9155 // This support is necessary to suport attributes that are not described in
9156 // for example, signatures. For example, the permanent home byref (byref that
9157 // points to the gc heap), isn't a property of method signatures, therefore,
9158 // it is safe to have mismatches here (that tiCompatibleWith will not flag),
9159 // but when deciding if we need to reimport a block, we need to take these
9161 BOOL tiMergeCompatibleWith(const typeInfo& pChild, const typeInfo& pParent, bool normalisedForStack) const;
9163 // Returns TRUE if child is equal to or a subtype of parent.
9164 // normalisedForStack indicates that both types are normalised for the stack
9165 BOOL tiCompatibleWith(const typeInfo& pChild, const typeInfo& pParent, bool normalisedForStack) const;
9167 // Merges pDest and pSrc. Returns FALSE if merge is undefined.
9168 // *pDest is modified to represent the merged type. Sets "*changed" to true
9169 // if this changes "*pDest".
9170 BOOL tiMergeToCommonParent(typeInfo* pDest, const typeInfo* pSrc, bool* changed) const;
9173 // <BUGNUM> VSW 471305
9174 // IJW allows assigning REF to BYREF. The following allows us to temporarily
9175 // bypass the assert check in gcMarkRegSetGCref and gcMarkRegSetByref
9176 // We use a "short" as we need to push/pop this scope.
9178 short compRegSetCheckLevel;
9182 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9183 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9185 XX IL verification stuff XX
9188 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9189 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9193 // The following is used to track liveness of local variables, initialization
9194 // of valueclass constructors, and type safe use of IL instructions.
9196 // dynamic state info needed for verification
9197 EntryState verCurrentState;
9199 // this ptr of object type .ctors are considered intited only after
9200 // the base class ctor is called, or an alternate ctor is called.
9201 // An uninited this ptr can be used to access fields, but cannot
9202 // be used to call a member function.
9203 BOOL verTrackObjCtorInitState;
9205 void verInitBBEntryState(BasicBlock* block, EntryState* currentState);
9207 // Requires that "tis" is not TIS_Bottom -- it's a definite init/uninit state.
9208 void verSetThisInit(BasicBlock* block, ThisInitState tis);
9209 void verInitCurrentState();
9210 void verResetCurrentState(BasicBlock* block, EntryState* currentState);
9212 // Merges the current verification state into the entry state of "block", return FALSE if that merge fails,
9213 // TRUE if it succeeds. Further sets "*changed" to true if this changes the entry state of "block".
9214 BOOL verMergeEntryStates(BasicBlock* block, bool* changed);
9216 void verConvertBBToThrowVerificationException(BasicBlock* block DEBUGARG(bool logMsg));
9217 void verHandleVerificationFailure(BasicBlock* block DEBUGARG(bool logMsg));
9218 typeInfo verMakeTypeInfo(CORINFO_CLASS_HANDLE clsHnd,
9219 bool bashStructToRef = false); // converts from jit type representation to typeInfo
9220 typeInfo verMakeTypeInfo(CorInfoType ciType,
9221 CORINFO_CLASS_HANDLE clsHnd); // converts from jit type representation to typeInfo
9222 BOOL verIsSDArray(typeInfo ti);
9223 typeInfo verGetArrayElemType(typeInfo ti);
9225 typeInfo verParseArgSigToTypeInfo(CORINFO_SIG_INFO* sig, CORINFO_ARG_LIST_HANDLE args);
9226 BOOL verNeedsVerification();
9227 BOOL verIsByRefLike(const typeInfo& ti);
9228 BOOL verIsSafeToReturnByRef(const typeInfo& ti);
9230 // generic type variables range over types that satisfy IsBoxable
9231 BOOL verIsBoxable(const typeInfo& ti);
9233 void DECLSPEC_NORETURN verRaiseVerifyException(INDEBUG(const char* reason) DEBUGARG(const char* file)
9234 DEBUGARG(unsigned line));
9235 void verRaiseVerifyExceptionIfNeeded(INDEBUG(const char* reason) DEBUGARG(const char* file)
9236 DEBUGARG(unsigned line));
9237 bool verCheckTailCallConstraint(OPCODE opcode,
9238 CORINFO_RESOLVED_TOKEN* pResolvedToken,
9239 CORINFO_RESOLVED_TOKEN* pConstrainedResolvedToken, // Is this a "constrained." call
9240 // on a type parameter?
9241 bool speculative // If true, won't throw if verificatoin fails. Instead it will
9242 // return false to the caller.
9243 // If false, it will throw.
9245 bool verIsBoxedValueType(typeInfo ti);
9247 void verVerifyCall(OPCODE opcode,
9248 CORINFO_RESOLVED_TOKEN* pResolvedToken,
9249 CORINFO_RESOLVED_TOKEN* pConstrainedResolvedToken,
9251 bool readonlyCall, // is this a "readonly." call?
9252 const BYTE* delegateCreateStart,
9253 const BYTE* codeAddr,
9254 CORINFO_CALL_INFO* callInfo DEBUGARG(const char* methodName));
9256 BOOL verCheckDelegateCreation(const BYTE* delegateCreateStart, const BYTE* codeAddr, mdMemberRef& targetMemberRef);
9258 typeInfo verVerifySTIND(const typeInfo& ptr, const typeInfo& value, const typeInfo& instrType);
9259 typeInfo verVerifyLDIND(const typeInfo& ptr, const typeInfo& instrType);
9260 void verVerifyField(CORINFO_RESOLVED_TOKEN* pResolvedToken,
9261 const CORINFO_FIELD_INFO& fieldInfo,
9262 const typeInfo* tiThis,
9264 BOOL allowPlainStructAsThis = FALSE);
9265 void verVerifyCond(const typeInfo& tiOp1, const typeInfo& tiOp2, unsigned opcode);
9266 void verVerifyThisPtrInitialised();
9267 BOOL verIsCallToInitThisPtr(CORINFO_CLASS_HANDLE context, CORINFO_CLASS_HANDLE target);
9271 // One line log function. Default level is 0. Increasing it gives you
9272 // more log information
9274 // levels are currently unused: #define JITDUMP(level,...) ();
9275 void JitLogEE(unsigned level, const char* fmt, ...);
9277 bool compDebugBreak;
9279 bool compJitHaltMethod();
9284 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9285 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9287 XX GS Security checks for unsafe buffers XX
9289 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9290 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
9293 struct ShadowParamVarInfo
9295 FixedBitVect* assignGroup; // the closure set of variables whose values depend on each other
9296 unsigned shadowCopy; // Lcl var num, valid only if not set to NO_SHADOW_COPY
9298 static bool mayNeedShadowCopy(LclVarDsc* varDsc)
9300 #if defined(_TARGET_AMD64_)
9301 // GS cookie logic to create shadow slots, create trees to copy reg args to shadow
9302 // slots and update all trees to refer to shadow slots is done immediately after
9303 // fgMorph(). Lsra could potentially mark a param as DoNotEnregister after JIT determines
9304 // not to shadow a parameter. Also, LSRA could potentially spill a param which is passed
9305 // in register. Therefore, conservatively all params may need a shadow copy. Note that
9306 // GS cookie logic further checks whether the param is a ptr or an unsafe buffer before
9307 // creating a shadow slot even though this routine returns true.
9309 // TODO-AMD64-CQ: Revisit this conservative approach as it could create more shadow slots than
9310 // required. There are two cases under which a reg arg could potentially be used from its
9312 // a) LSRA marks it as DoNotEnregister (see LinearScan::identifyCandidates())
9313 // b) LSRA spills it
9315 // Possible solution to address case (a)
9316 // - The conditions under which LSRA marks a varDsc as DoNotEnregister could be checked
9317 // in this routine. Note that live out of exception handler is something we may not be
9318 // able to do it here since GS cookie logic is invoked ahead of liveness computation.
9319 // Therefore, for methods with exception handling and need GS cookie check we might have
9320 // to take conservative approach.
9322 // Possible solution to address case (b)
9323 // - Whenver a parameter passed in an argument register needs to be spilled by LSRA, we
9324 // create a new spill temp if the method needs GS cookie check.
9325 return varDsc->lvIsParam;
9326 #else // !defined(_TARGET_AMD64_)
9327 return varDsc->lvIsParam && !varDsc->lvIsRegArg;
9334 printf("assignGroup [%p]; shadowCopy: [%d];\n", assignGroup, shadowCopy);
9339 GSCookie* gsGlobalSecurityCookieAddr; // Address of global cookie for unsafe buffer checks
9340 GSCookie gsGlobalSecurityCookieVal; // Value of global cookie if addr is NULL
9341 ShadowParamVarInfo* gsShadowVarInfo; // Table used by shadow param analysis code
9343 void gsGSChecksInitCookie(); // Grabs cookie variable
9344 void gsCopyShadowParams(); // Identify vulnerable params and create dhadow copies
9345 bool gsFindVulnerableParams(); // Shadow param analysis code
9346 void gsParamsToShadows(); // Insert copy code and replave param uses by shadow
9348 static fgWalkPreFn gsMarkPtrsAndAssignGroups; // Shadow param analysis tree-walk
9349 static fgWalkPreFn gsReplaceShadowParams; // Shadow param replacement tree-walk
9351 #define DEFAULT_MAX_INLINE_SIZE 100 // Methods with > DEFAULT_MAX_INLINE_SIZE IL bytes will never be inlined.
9352 // This can be overwritten by setting complus_JITInlineSize env variable.
9354 #define DEFAULT_MAX_INLINE_DEPTH 20 // Methods at more than this level deep will not be inlined
9356 #define DEFAULT_MAX_LOCALLOC_TO_LOCAL_SIZE 32 // fixed locallocs of this size or smaller will convert to local buffers
9359 #ifdef FEATURE_JIT_METHOD_PERF
9360 JitTimer* pCompJitTimer; // Timer data structure (by phases) for current compilation.
9361 static CompTimeSummaryInfo s_compJitTimerSummary; // Summary of the Timer information for the whole run.
9363 static LPCWSTR JitTimeLogCsv(); // Retrieve the file name for CSV from ConfigDWORD.
9364 static LPCWSTR compJitTimeLogFilename; // If a log file for JIT time is desired, filename to write it to.
9366 inline void EndPhase(Phases phase); // Indicate the end of the given phase.
9368 #if MEASURE_CLRAPI_CALLS
9369 // Thin wrappers that call into JitTimer (if present).
9370 inline void CLRApiCallEnter(unsigned apix);
9371 inline void CLRApiCallLeave(unsigned apix);
9374 inline void CLR_API_Enter(API_ICorJitInfo_Names ename);
9375 inline void CLR_API_Leave(API_ICorJitInfo_Names ename);
9380 #if defined(DEBUG) || defined(INLINE_DATA) || defined(FEATURE_CLRSQM)
9381 // These variables are associated with maintaining SQM data about compile time.
9382 unsigned __int64 m_compCyclesAtEndOfInlining; // The thread-virtualized cycle count at the end of the inlining phase
9383 // in the current compilation.
9384 unsigned __int64 m_compCycles; // Net cycle count for current compilation
9385 DWORD m_compTickCountAtEndOfInlining; // The result of GetTickCount() (# ms since some epoch marker) at the end of
9386 // the inlining phase in the current compilation.
9387 #endif // defined(DEBUG) || defined(INLINE_DATA) || defined(FEATURE_CLRSQM)
9389 // Records the SQM-relevant (cycles and tick count). Should be called after inlining is complete.
9390 // (We do this after inlining because this marks the last point at which the JIT is likely to cause
9391 // type-loading and class initialization).
9392 void RecordStateAtEndOfInlining();
9393 // Assumes being called at the end of compilation. Update the SQM state.
9394 void RecordStateAtEndOfCompilation();
9396 #ifdef FEATURE_CLRSQM
9397 // Does anything SQM related necessary at process shutdown time.
9398 static void ProcessShutdownSQMWork(ICorStaticInfo* statInfo);
9399 #endif // FEATURE_CLRSQM
9402 #if FUNC_INFO_LOGGING
9403 static LPCWSTR compJitFuncInfoFilename; // If a log file for per-function information is required, this is the
9404 // filename to write it to.
9405 static FILE* compJitFuncInfoFile; // And this is the actual FILE* to write to.
9406 #endif // FUNC_INFO_LOGGING
9408 Compiler* prevCompiler; // Previous compiler on stack for TLS Compiler* linked list for reentrant compilers.
9410 // Is the compilation in a full trust context?
9411 bool compIsFullTrust();
9414 void RecordNowayAssert(const char* filename, unsigned line, const char* condStr);
9415 #endif // MEASURE_NOWAY
9417 #ifndef FEATURE_TRACELOGGING
9418 // Should we actually fire the noway assert body and the exception handler?
9419 bool compShouldThrowOnNoway();
9420 #else // FEATURE_TRACELOGGING
9421 // Should we actually fire the noway assert body and the exception handler?
9422 bool compShouldThrowOnNoway(const char* filename, unsigned line);
9424 // Telemetry instance to use per method compilation.
9425 JitTelemetry compJitTelemetry;
9427 // Get common parameters that have to be logged with most telemetry data.
9428 void compGetTelemetryDefaults(const char** assemblyName,
9429 const char** scopeName,
9430 const char** methodName,
9431 unsigned* methodHash);
9432 #endif // !FEATURE_TRACELOGGING
9436 NodeToTestDataMap* m_nodeTestData;
9438 static const unsigned FIRST_LOOP_HOIST_CSE_CLASS = 1000;
9439 unsigned m_loopHoistCSEClass; // LoopHoist test annotations turn into CSE requirements; we
9440 // label them with CSE Class #'s starting at FIRST_LOOP_HOIST_CSE_CLASS.
9441 // Current kept in this.
9443 NodeToTestDataMap* GetNodeTestData()
9445 Compiler* compRoot = impInlineRoot();
9446 if (compRoot->m_nodeTestData == nullptr)
9448 compRoot->m_nodeTestData = new (getAllocatorDebugOnly()) NodeToTestDataMap(getAllocatorDebugOnly());
9450 return compRoot->m_nodeTestData;
9453 typedef JitHashTable<GenTree*, JitPtrKeyFuncs<GenTree>, int> NodeToIntMap;
9455 // Returns the set (i.e., the domain of the result map) of nodes that are keys in m_nodeTestData, and
9456 // currently occur in the AST graph.
9457 NodeToIntMap* FindReachableNodesInNodeTestData();
9459 // Node "from" is being eliminated, and being replaced by node "to". If "from" had any associated
9460 // test data, associate that data with "to".
9461 void TransferTestDataToNode(GenTree* from, GenTree* to);
9463 // Requires that "to" is a clone of "from". If any nodes in the "from" tree
9464 // have annotations, attach similar annotations to the corresponding nodes in "to".
9465 void CopyTestDataToCloneTree(GenTree* from, GenTree* to);
9467 // These are the methods that test that the various conditions implied by the
9468 // test attributes are satisfied.
9469 void JitTestCheckSSA(); // SSA builder tests.
9470 void JitTestCheckVN(); // Value numbering tests.
9473 // The "FieldSeqStore", for canonicalizing field sequences. See the definition of FieldSeqStore for
9475 FieldSeqStore* m_fieldSeqStore;
9477 FieldSeqStore* GetFieldSeqStore()
9479 Compiler* compRoot = impInlineRoot();
9480 if (compRoot->m_fieldSeqStore == nullptr)
9482 // Create a CompAllocator that labels sub-structure with CMK_FieldSeqStore, and use that for allocation.
9483 CompAllocator ialloc(getAllocator(CMK_FieldSeqStore));
9484 compRoot->m_fieldSeqStore = new (ialloc) FieldSeqStore(ialloc);
9486 return compRoot->m_fieldSeqStore;
9489 typedef JitHashTable<GenTree*, JitPtrKeyFuncs<GenTree>, FieldSeqNode*> NodeToFieldSeqMap;
9491 // Some nodes of "TYP_BYREF" or "TYP_I_IMPL" actually represent the address of a field within a struct, but since
9492 // the offset of the field is zero, there's no "GT_ADD" node. We normally attach a field sequence to the constant
9493 // that is added, but what do we do when that constant is zero, and is thus not present? We use this mechanism to
9494 // attach the field sequence directly to the address node.
9495 NodeToFieldSeqMap* m_zeroOffsetFieldMap;
9497 NodeToFieldSeqMap* GetZeroOffsetFieldMap()
9499 // Don't need to worry about inlining here
9500 if (m_zeroOffsetFieldMap == nullptr)
9502 // Create a CompAllocator that labels sub-structure with CMK_ZeroOffsetFieldMap, and use that for
9504 CompAllocator ialloc(getAllocator(CMK_ZeroOffsetFieldMap));
9505 m_zeroOffsetFieldMap = new (ialloc) NodeToFieldSeqMap(ialloc);
9507 return m_zeroOffsetFieldMap;
9510 // Requires that "op1" is a node of type "TYP_BYREF" or "TYP_I_IMPL". We are dereferencing this with the fields in
9511 // "fieldSeq", whose offsets are required all to be zero. Ensures that any field sequence annotation currently on
9512 // "op1" or its components is augmented by appending "fieldSeq". In practice, if "op1" is a GT_LCL_FLD, it has
9513 // a field sequence as a member; otherwise, it may be the addition of an a byref and a constant, where the const
9514 // has a field sequence -- in this case "fieldSeq" is appended to that of the constant; otherwise, we
9515 // record the the field sequence using the ZeroOffsetFieldMap described above.
9517 // One exception above is that "op1" is a node of type "TYP_REF" where "op1" is a GT_LCL_VAR.
9518 // This happens when System.Object vtable pointer is a regular field at offset 0 in System.Private.CoreLib in
9519 // CoreRT. Such case is handled same as the default case.
9520 void fgAddFieldSeqForZeroOffset(GenTree* op1, FieldSeqNode* fieldSeq);
9522 typedef JitHashTable<const GenTree*, JitPtrKeyFuncs<GenTree>, ArrayInfo> NodeToArrayInfoMap;
9523 NodeToArrayInfoMap* m_arrayInfoMap;
9525 NodeToArrayInfoMap* GetArrayInfoMap()
9527 Compiler* compRoot = impInlineRoot();
9528 if (compRoot->m_arrayInfoMap == nullptr)
9530 // Create a CompAllocator that labels sub-structure with CMK_ArrayInfoMap, and use that for allocation.
9531 CompAllocator ialloc(getAllocator(CMK_ArrayInfoMap));
9532 compRoot->m_arrayInfoMap = new (ialloc) NodeToArrayInfoMap(ialloc);
9534 return compRoot->m_arrayInfoMap;
9537 //-----------------------------------------------------------------------------------------------------------------
9538 // Compiler::TryGetArrayInfo:
9539 // Given an indirection node, checks to see whether or not that indirection represents an array access, and
9540 // if so returns information about the array.
9543 // indir - The `GT_IND` node.
9544 // arrayInfo (out) - Information about the accessed array if this function returns true. Undefined otherwise.
9547 // True if the `GT_IND` node represents an array access; false otherwise.
9548 inline bool TryGetArrayInfo(GenTreeIndir* indir, ArrayInfo* arrayInfo)
9550 if ((indir->gtFlags & GTF_IND_ARR_INDEX) == 0)
9555 if (indir->gtOp1->OperIs(GT_INDEX_ADDR))
9557 GenTreeIndexAddr* const indexAddr = indir->gtOp1->AsIndexAddr();
9558 *arrayInfo = ArrayInfo(indexAddr->gtElemType, indexAddr->gtElemSize, indexAddr->gtElemOffset,
9559 indexAddr->gtStructElemClass);
9563 bool found = GetArrayInfoMap()->Lookup(indir, arrayInfo);
9568 NodeToUnsignedMap* m_memorySsaMap[MemoryKindCount];
9570 // In some cases, we want to assign intermediate SSA #'s to memory states, and know what nodes create those memory
9571 // states. (We do this for try blocks, where, if the try block doesn't do a call that loses track of the memory
9572 // state, all the possible memory states are possible initial states of the corresponding catch block(s).)
9573 NodeToUnsignedMap* GetMemorySsaMap(MemoryKind memoryKind)
9575 if (memoryKind == GcHeap && byrefStatesMatchGcHeapStates)
9577 // Use the same map for GCHeap and ByrefExposed when their states match.
9578 memoryKind = ByrefExposed;
9581 assert(memoryKind < MemoryKindCount);
9582 Compiler* compRoot = impInlineRoot();
9583 if (compRoot->m_memorySsaMap[memoryKind] == nullptr)
9585 // Create a CompAllocator that labels sub-structure with CMK_ArrayInfoMap, and use that for allocation.
9586 CompAllocator ialloc(getAllocator(CMK_ArrayInfoMap));
9587 compRoot->m_memorySsaMap[memoryKind] = new (ialloc) NodeToUnsignedMap(ialloc);
9589 return compRoot->m_memorySsaMap[memoryKind];
9592 // The Refany type is the only struct type whose structure is implicitly assumed by IL. We need its fields.
9593 CORINFO_CLASS_HANDLE m_refAnyClass;
9594 CORINFO_FIELD_HANDLE GetRefanyDataField()
9596 if (m_refAnyClass == nullptr)
9598 m_refAnyClass = info.compCompHnd->getBuiltinClass(CLASSID_TYPED_BYREF);
9600 return info.compCompHnd->getFieldInClass(m_refAnyClass, 0);
9602 CORINFO_FIELD_HANDLE GetRefanyTypeField()
9604 if (m_refAnyClass == nullptr)
9606 m_refAnyClass = info.compCompHnd->getBuiltinClass(CLASSID_TYPED_BYREF);
9608 return info.compCompHnd->getFieldInClass(m_refAnyClass, 1);
9612 static BitSetSupport::BitSetOpCounter m_varsetOpCounter;
9614 #if ALLVARSET_COUNTOPS
9615 static BitSetSupport::BitSetOpCounter m_allvarsetOpCounter;
9618 static HelperCallProperties s_helperCallProperties;
9620 #ifdef UNIX_AMD64_ABI
9621 static var_types GetTypeFromClassificationAndSizes(SystemVClassificationType classType, int size);
9622 static var_types GetEightByteType(const SYSTEMV_AMD64_CORINFO_STRUCT_REG_PASSING_DESCRIPTOR& structDesc,
9625 static void GetStructTypeOffset(const SYSTEMV_AMD64_CORINFO_STRUCT_REG_PASSING_DESCRIPTOR& structDesc,
9628 unsigned __int8* offset0,
9629 unsigned __int8* offset1);
9631 void GetStructTypeOffset(CORINFO_CLASS_HANDLE typeHnd,
9634 unsigned __int8* offset0,
9635 unsigned __int8* offset1);
9637 #endif // defined(UNIX_AMD64_ABI)
9639 void fgMorphMultiregStructArgs(GenTreeCall* call);
9640 GenTree* fgMorphMultiregStructArg(GenTree* arg, fgArgTabEntry* fgEntryPtr);
9642 bool killGCRefs(GenTree* tree);
9644 }; // end of class Compiler
9646 // LclVarDsc constructor. Uses Compiler, so must come after Compiler definition.
9647 inline LclVarDsc::LclVarDsc()
9648 : // Initialize the ArgRegs to REG_STK.
9649 // The morph will do the right thing to change
9650 // to the right register if passed in register.
9653 #if FEATURE_MULTIREG_ARGS
9654 _lvOtherArgReg(REG_STK)
9656 #endif // FEATURE_MULTIREG_ARGS
9658 lvRefBlks(BlockSetOps::UninitVal())
9660 #endif // ASSERTION_PROP
9665 //---------------------------------------------------------------------------------------------------------------------
9666 // GenTreeVisitor: a flexible tree walker implemented using the curiosly-recurring-template pattern.
9668 // This class implements a configurable walker for IR trees. There are five configuration options (defaults values are
9669 // shown in parentheses):
9671 // - ComputeStack (false): when true, the walker will push each node onto the `m_ancestors` stack. "Ancestors" is a bit
9672 // of a misnomer, as the first entry will always be the current node.
9674 // - DoPreOrder (false): when true, the walker will invoke `TVisitor::PreOrderVisit` with the current node as an
9675 // argument before visiting the node's operands.
9677 // - DoPostOrder (false): when true, the walker will invoke `TVisitor::PostOrderVisit` with the current node as an
9678 // argument after visiting the node's operands.
9680 // - DoLclVarsOnly (false): when true, the walker will only invoke `TVisitor::PreOrderVisit` for lclVar nodes.
9681 // `DoPreOrder` must be true if this option is true.
9683 // - UseExecutionOrder (false): when true, then walker will visit a node's operands in execution order (e.g. if a
9684 // binary operator has the `GTF_REVERSE_OPS` flag set, the second operand will be
9685 // visited before the first).
9687 // At least one of `DoPreOrder` and `DoPostOrder` must be specified.
9689 // A simple pre-order visitor might look something like the following:
9691 // class CountingVisitor final : public GenTreeVisitor<CountingVisitor>
9696 // DoPreOrder = true
9699 // unsigned m_count;
9701 // CountingVisitor(Compiler* compiler)
9702 // : GenTreeVisitor<CountingVisitor>(compiler), m_count(0)
9706 // Compiler::fgWalkResult PreOrderVisit(GenTree* node)
9712 // This visitor would then be used like so:
9714 // CountingVisitor countingVisitor(compiler);
9715 // countingVisitor.WalkTree(root);
9717 template <typename TVisitor>
9718 class GenTreeVisitor
9721 typedef Compiler::fgWalkResult fgWalkResult;
9725 ComputeStack = false,
9727 DoPostOrder = false,
9728 DoLclVarsOnly = false,
9729 UseExecutionOrder = false,
9732 Compiler* m_compiler;
9733 ArrayStack<GenTree*> m_ancestors;
9735 GenTreeVisitor(Compiler* compiler) : m_compiler(compiler), m_ancestors(compiler->getAllocator(CMK_ArrayStack))
9737 assert(compiler != nullptr);
9739 static_assert_no_msg(TVisitor::DoPreOrder || TVisitor::DoPostOrder);
9740 static_assert_no_msg(!TVisitor::DoLclVarsOnly || TVisitor::DoPreOrder);
9743 fgWalkResult PreOrderVisit(GenTree** use, GenTree* user)
9745 return fgWalkResult::WALK_CONTINUE;
9748 fgWalkResult PostOrderVisit(GenTree** use, GenTree* user)
9750 return fgWalkResult::WALK_CONTINUE;
9754 fgWalkResult WalkTree(GenTree** use, GenTree* user)
9756 assert(use != nullptr);
9758 GenTree* node = *use;
9760 if (TVisitor::ComputeStack)
9762 m_ancestors.Push(node);
9765 fgWalkResult result = fgWalkResult::WALK_CONTINUE;
9766 if (TVisitor::DoPreOrder && !TVisitor::DoLclVarsOnly)
9768 result = reinterpret_cast<TVisitor*>(this)->PreOrderVisit(use, user);
9769 if (result == fgWalkResult::WALK_ABORT)
9775 if ((node == nullptr) || (result == fgWalkResult::WALK_SKIP_SUBTREES))
9781 switch (node->OperGet())
9786 case GT_LCL_VAR_ADDR:
9787 case GT_LCL_FLD_ADDR:
9788 if (TVisitor::DoLclVarsOnly)
9790 result = reinterpret_cast<TVisitor*>(this)->PreOrderVisit(use, user);
9791 if (result == fgWalkResult::WALK_ABORT)
9807 case GT_MEMORYBARRIER:
9812 case GT_START_NONGC:
9814 #if !FEATURE_EH_FUNCLETS
9816 #endif // !FEATURE_EH_FUNCLETS
9820 case GT_CLS_VAR_ADDR:
9824 case GT_PINVOKE_PROLOG:
9825 case GT_PINVOKE_EPILOG:
9829 // Lclvar unary operators
9830 case GT_STORE_LCL_VAR:
9831 case GT_STORE_LCL_FLD:
9832 if (TVisitor::DoLclVarsOnly)
9834 result = reinterpret_cast<TVisitor*>(this)->PreOrderVisit(use, user);
9835 if (result == fgWalkResult::WALK_ABORT)
9842 // Standard unary operators
9871 case GT_RUNTIMELOOKUP:
9873 GenTreeUnOp* const unOp = node->AsUnOp();
9874 if (unOp->gtOp1 != nullptr)
9876 result = WalkTree(&unOp->gtOp1, unOp);
9877 if (result == fgWalkResult::WALK_ABORT)
9888 GenTreeCmpXchg* const cmpXchg = node->AsCmpXchg();
9890 result = WalkTree(&cmpXchg->gtOpLocation, cmpXchg);
9891 if (result == fgWalkResult::WALK_ABORT)
9895 result = WalkTree(&cmpXchg->gtOpValue, cmpXchg);
9896 if (result == fgWalkResult::WALK_ABORT)
9900 result = WalkTree(&cmpXchg->gtOpComparand, cmpXchg);
9901 if (result == fgWalkResult::WALK_ABORT)
9908 case GT_ARR_BOUNDS_CHECK:
9911 #endif // FEATURE_SIMD
9912 #ifdef FEATURE_HW_INTRINSICS
9913 case GT_HW_INTRINSIC_CHK:
9914 #endif // FEATURE_HW_INTRINSICS
9916 GenTreeBoundsChk* const boundsChk = node->AsBoundsChk();
9918 result = WalkTree(&boundsChk->gtIndex, boundsChk);
9919 if (result == fgWalkResult::WALK_ABORT)
9923 result = WalkTree(&boundsChk->gtArrLen, boundsChk);
9924 if (result == fgWalkResult::WALK_ABORT)
9933 GenTreeField* const field = node->AsField();
9935 if (field->gtFldObj != nullptr)
9937 result = WalkTree(&field->gtFldObj, field);
9938 if (result == fgWalkResult::WALK_ABORT)
9948 GenTreeArrElem* const arrElem = node->AsArrElem();
9950 result = WalkTree(&arrElem->gtArrObj, arrElem);
9951 if (result == fgWalkResult::WALK_ABORT)
9956 const unsigned rank = arrElem->gtArrRank;
9957 for (unsigned dim = 0; dim < rank; dim++)
9959 result = WalkTree(&arrElem->gtArrInds[dim], arrElem);
9960 if (result == fgWalkResult::WALK_ABORT)
9970 GenTreeArrOffs* const arrOffs = node->AsArrOffs();
9972 result = WalkTree(&arrOffs->gtOffset, arrOffs);
9973 if (result == fgWalkResult::WALK_ABORT)
9977 result = WalkTree(&arrOffs->gtIndex, arrOffs);
9978 if (result == fgWalkResult::WALK_ABORT)
9982 result = WalkTree(&arrOffs->gtArrObj, arrOffs);
9983 if (result == fgWalkResult::WALK_ABORT)
9992 GenTreeDynBlk* const dynBlock = node->AsDynBlk();
9994 GenTree** op1Use = &dynBlock->gtOp1;
9995 GenTree** op2Use = &dynBlock->gtDynamicSize;
9997 if (TVisitor::UseExecutionOrder && dynBlock->gtEvalSizeFirst)
9999 std::swap(op1Use, op2Use);
10002 result = WalkTree(op1Use, dynBlock);
10003 if (result == fgWalkResult::WALK_ABORT)
10007 result = WalkTree(op2Use, dynBlock);
10008 if (result == fgWalkResult::WALK_ABORT)
10015 case GT_STORE_DYN_BLK:
10017 GenTreeDynBlk* const dynBlock = node->AsDynBlk();
10019 GenTree** op1Use = &dynBlock->gtOp1;
10020 GenTree** op2Use = &dynBlock->gtOp2;
10021 GenTree** op3Use = &dynBlock->gtDynamicSize;
10023 if (TVisitor::UseExecutionOrder)
10025 if (dynBlock->IsReverseOp())
10027 std::swap(op1Use, op2Use);
10029 if (dynBlock->gtEvalSizeFirst)
10031 std::swap(op3Use, op2Use);
10032 std::swap(op2Use, op1Use);
10036 result = WalkTree(op1Use, dynBlock);
10037 if (result == fgWalkResult::WALK_ABORT)
10041 result = WalkTree(op2Use, dynBlock);
10042 if (result == fgWalkResult::WALK_ABORT)
10046 result = WalkTree(op3Use, dynBlock);
10047 if (result == fgWalkResult::WALK_ABORT)
10056 GenTreeCall* const call = node->AsCall();
10058 if (call->gtCallObjp != nullptr)
10060 result = WalkTree(&call->gtCallObjp, call);
10061 if (result == fgWalkResult::WALK_ABORT)
10067 for (GenTreeArgList* args = call->gtCallArgs; args != nullptr; args = args->Rest())
10069 result = WalkTree(args->pCurrent(), call);
10070 if (result == fgWalkResult::WALK_ABORT)
10076 for (GenTreeArgList* args = call->gtCallLateArgs; args != nullptr; args = args->Rest())
10078 result = WalkTree(args->pCurrent(), call);
10079 if (result == fgWalkResult::WALK_ABORT)
10085 if (call->gtCallType == CT_INDIRECT)
10087 if (call->gtCallCookie != nullptr)
10089 result = WalkTree(&call->gtCallCookie, call);
10090 if (result == fgWalkResult::WALK_ABORT)
10096 result = WalkTree(&call->gtCallAddr, call);
10097 if (result == fgWalkResult::WALK_ABORT)
10103 if (call->gtControlExpr != nullptr)
10105 result = WalkTree(&call->gtControlExpr, call);
10106 if (result == fgWalkResult::WALK_ABORT)
10118 assert(node->OperIsBinary());
10120 GenTreeOp* const op = node->AsOp();
10122 GenTree** op1Use = &op->gtOp1;
10123 GenTree** op2Use = &op->gtOp2;
10125 if (TVisitor::UseExecutionOrder && node->IsReverseOp())
10127 std::swap(op1Use, op2Use);
10130 if (*op1Use != nullptr)
10132 result = WalkTree(op1Use, op);
10133 if (result == fgWalkResult::WALK_ABORT)
10139 if (*op2Use != nullptr)
10141 result = WalkTree(op2Use, op);
10142 if (result == fgWalkResult::WALK_ABORT)
10152 // Finally, visit the current node
10153 if (TVisitor::DoPostOrder)
10155 result = reinterpret_cast<TVisitor*>(this)->PostOrderVisit(use, user);
10158 if (TVisitor::ComputeStack)
10167 template <bool computeStack, bool doPreOrder, bool doPostOrder, bool doLclVarsOnly, bool useExecutionOrder>
10168 class GenericTreeWalker final
10169 : public GenTreeVisitor<GenericTreeWalker<computeStack, doPreOrder, doPostOrder, doLclVarsOnly, useExecutionOrder>>
10174 ComputeStack = computeStack,
10175 DoPreOrder = doPreOrder,
10176 DoPostOrder = doPostOrder,
10177 DoLclVarsOnly = doLclVarsOnly,
10178 UseExecutionOrder = useExecutionOrder,
10182 Compiler::fgWalkData* m_walkData;
10185 GenericTreeWalker(Compiler::fgWalkData* walkData)
10186 : GenTreeVisitor<GenericTreeWalker<computeStack, doPreOrder, doPostOrder, doLclVarsOnly, useExecutionOrder>>(
10187 walkData->compiler)
10188 , m_walkData(walkData)
10190 assert(walkData != nullptr);
10194 walkData->parentStack = &this->m_ancestors;
10198 Compiler::fgWalkResult PreOrderVisit(GenTree** use, GenTree* user)
10200 m_walkData->parent = user;
10201 return m_walkData->wtprVisitorFn(use, m_walkData);
10204 Compiler::fgWalkResult PostOrderVisit(GenTree** use, GenTree* user)
10206 m_walkData->parent = user;
10207 return m_walkData->wtpoVisitorFn(use, m_walkData);
10212 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
10213 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
10215 XX Miscellaneous Compiler stuff XX
10217 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
10218 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
10221 // Values used to mark the types a stack slot is used for
10223 const unsigned TYPE_REF_INT = 0x01; // slot used as a 32-bit int
10224 const unsigned TYPE_REF_LNG = 0x02; // slot used as a 64-bit long
10225 const unsigned TYPE_REF_FLT = 0x04; // slot used as a 32-bit float
10226 const unsigned TYPE_REF_DBL = 0x08; // slot used as a 64-bit float
10227 const unsigned TYPE_REF_PTR = 0x10; // slot used as a 32-bit pointer
10228 const unsigned TYPE_REF_BYR = 0x20; // slot used as a byref pointer
10229 const unsigned TYPE_REF_STC = 0x40; // slot used as a struct
10230 const unsigned TYPE_REF_TYPEMASK = 0x7F; // bits that represent the type
10232 // const unsigned TYPE_REF_ADDR_TAKEN = 0x80; // slots address was taken
10234 /*****************************************************************************
10236 * Variables to keep track of total code amounts.
10241 extern size_t grossVMsize;
10242 extern size_t grossNCsize;
10243 extern size_t totalNCsize;
10245 extern unsigned genMethodICnt;
10246 extern unsigned genMethodNCnt;
10247 extern size_t gcHeaderISize;
10248 extern size_t gcPtrMapISize;
10249 extern size_t gcHeaderNSize;
10250 extern size_t gcPtrMapNSize;
10252 #endif // DISPLAY_SIZES
10254 /*****************************************************************************
10256 * Variables to keep track of basic block counts (more data on 1 BB methods)
10259 #if COUNT_BASIC_BLOCKS
10260 extern Histogram bbCntTable;
10261 extern Histogram bbOneBBSizeTable;
10264 /*****************************************************************************
10266 * Used by optFindNaturalLoops to gather statistical information such as
10267 * - total number of natural loops
10268 * - number of loops with 1, 2, ... exit conditions
10269 * - number of loops that have an iterator (for like)
10270 * - number of loops that have a constant iterator
10275 extern unsigned totalLoopMethods; // counts the total number of methods that have natural loops
10276 extern unsigned maxLoopsPerMethod; // counts the maximum number of loops a method has
10277 extern unsigned totalLoopOverflows; // # of methods that identified more loops than we can represent
10278 extern unsigned totalLoopCount; // counts the total number of natural loops
10279 extern unsigned totalUnnatLoopCount; // counts the total number of (not-necessarily natural) loops
10280 extern unsigned totalUnnatLoopOverflows; // # of methods that identified more unnatural loops than we can represent
10281 extern unsigned iterLoopCount; // counts the # of loops with an iterator (for like)
10282 extern unsigned simpleTestLoopCount; // counts the # of loops with an iterator and a simple loop condition (iter <
10284 extern unsigned constIterLoopCount; // counts the # of loops with a constant iterator (for like)
10285 extern bool hasMethodLoops; // flag to keep track if we already counted a method as having loops
10286 extern unsigned loopsThisMethod; // counts the number of loops in the current method
10287 extern bool loopOverflowThisMethod; // True if we exceeded the max # of loops in the method.
10288 extern Histogram loopCountTable; // Histogram of loop counts
10289 extern Histogram loopExitCountTable; // Histogram of loop exit counts
10291 #endif // COUNT_LOOPS
10293 /*****************************************************************************
10294 * variables to keep track of how many iterations we go in a dataflow pass
10299 extern unsigned CSEiterCount; // counts the # of iteration for the CSE dataflow
10300 extern unsigned CFiterCount; // counts the # of iteration for the Const Folding dataflow
10302 #endif // DATAFLOW_ITER
10304 #if MEASURE_BLOCK_SIZE
10305 extern size_t genFlowNodeSize;
10306 extern size_t genFlowNodeCnt;
10307 #endif // MEASURE_BLOCK_SIZE
10309 #if MEASURE_NODE_SIZE
10310 struct NodeSizeStats
10314 genTreeNodeCnt = 0;
10315 genTreeNodeSize = 0;
10316 genTreeNodeActualSize = 0;
10319 // Count of tree nodes allocated.
10320 unsigned __int64 genTreeNodeCnt;
10322 // The size we allocate.
10323 unsigned __int64 genTreeNodeSize;
10325 // The actual size of the node. Note that the actual size will likely be smaller
10326 // than the allocated size, but we sometimes use SetOper()/ChangeOper() to change
10327 // a smaller node to a larger one. TODO-Cleanup: add stats on
10328 // SetOper()/ChangeOper() usage to quantify this.
10329 unsigned __int64 genTreeNodeActualSize;
10331 extern NodeSizeStats genNodeSizeStats; // Total node size stats
10332 extern NodeSizeStats genNodeSizeStatsPerFunc; // Per-function node size stats
10333 extern Histogram genTreeNcntHist;
10334 extern Histogram genTreeNsizHist;
10335 #endif // MEASURE_NODE_SIZE
10337 /*****************************************************************************
10338 * Count fatal errors (including noway_asserts).
10342 extern unsigned fatal_badCode;
10343 extern unsigned fatal_noWay;
10344 extern unsigned fatal_NOMEM;
10345 extern unsigned fatal_noWayAssertBody;
10347 extern unsigned fatal_noWayAssertBodyArgs;
10349 extern unsigned fatal_NYI;
10350 #endif // MEASURE_FATAL
10352 /*****************************************************************************
10356 #ifdef _TARGET_XARCH_
10358 const instruction INS_SHIFT_LEFT_LOGICAL = INS_shl;
10359 const instruction INS_SHIFT_RIGHT_LOGICAL = INS_shr;
10360 const instruction INS_SHIFT_RIGHT_ARITHM = INS_sar;
10362 const instruction INS_AND = INS_and;
10363 const instruction INS_OR = INS_or;
10364 const instruction INS_XOR = INS_xor;
10365 const instruction INS_NEG = INS_neg;
10366 const instruction INS_TEST = INS_test;
10367 const instruction INS_MUL = INS_imul;
10368 const instruction INS_SIGNED_DIVIDE = INS_idiv;
10369 const instruction INS_UNSIGNED_DIVIDE = INS_div;
10370 const instruction INS_BREAKPOINT = INS_int3;
10371 const instruction INS_ADDC = INS_adc;
10372 const instruction INS_SUBC = INS_sbb;
10373 const instruction INS_NOT = INS_not;
10375 #endif // _TARGET_XARCH_
10377 #ifdef _TARGET_ARM_
10379 const instruction INS_SHIFT_LEFT_LOGICAL = INS_lsl;
10380 const instruction INS_SHIFT_RIGHT_LOGICAL = INS_lsr;
10381 const instruction INS_SHIFT_RIGHT_ARITHM = INS_asr;
10383 const instruction INS_AND = INS_and;
10384 const instruction INS_OR = INS_orr;
10385 const instruction INS_XOR = INS_eor;
10386 const instruction INS_NEG = INS_rsb;
10387 const instruction INS_TEST = INS_tst;
10388 const instruction INS_MUL = INS_mul;
10389 const instruction INS_MULADD = INS_mla;
10390 const instruction INS_SIGNED_DIVIDE = INS_sdiv;
10391 const instruction INS_UNSIGNED_DIVIDE = INS_udiv;
10392 const instruction INS_BREAKPOINT = INS_bkpt;
10393 const instruction INS_ADDC = INS_adc;
10394 const instruction INS_SUBC = INS_sbc;
10395 const instruction INS_NOT = INS_mvn;
10397 const instruction INS_ABS = INS_vabs;
10398 const instruction INS_SQRT = INS_vsqrt;
10400 #endif // _TARGET_ARM_
10402 #ifdef _TARGET_ARM64_
10404 const instruction INS_MULADD = INS_madd;
10405 const instruction INS_BREAKPOINT = INS_bkpt;
10407 const instruction INS_ABS = INS_fabs;
10408 const instruction INS_SQRT = INS_fsqrt;
10410 #endif // _TARGET_ARM64_
10412 /*****************************************************************************/
10414 extern const BYTE genTypeSizes[];
10415 extern const BYTE genTypeAlignments[];
10416 extern const BYTE genTypeStSzs[];
10417 extern const BYTE genActualTypes[];
10419 /*****************************************************************************/
10421 // VERY_LARGE_FRAME_SIZE_REG_MASK is the set of registers we need to use for
10422 // the probing loop generated for very large stack frames (see `getVeryLargeFrameSize`).
10424 #ifdef _TARGET_ARM_
10425 #define VERY_LARGE_FRAME_SIZE_REG_MASK (RBM_R4 | RBM_R5 | RBM_R6)
10426 #elif defined(_TARGET_ARM64_)
10427 #define VERY_LARGE_FRAME_SIZE_REG_MASK (RBM_R9 | RBM_R10 | RBM_R11)
10430 /*****************************************************************************/
10432 extern BasicBlock dummyBB;
10434 /*****************************************************************************/
10435 /*****************************************************************************/
10437 // foreach_treenode_execution_order: An iterator that iterates through all the tree
10438 // nodes of a statement in execution order.
10439 // __stmt: a GT_STMT type GenTree*
10440 // __node: a GenTree*, already declared, that gets updated with each node in the statement, in execution order
10442 #define foreach_treenode_execution_order(__node, __stmt) \
10443 for ((__node) = (__stmt)->gtStmt.gtStmtList; (__node); (__node) = (__node)->gtNext)
10445 // foreach_block: An iterator over all blocks in the function.
10446 // __compiler: the Compiler* object
10447 // __block : a BasicBlock*, already declared, that gets updated each iteration.
10449 #define foreach_block(__compiler, __block) \
10450 for ((__block) = (__compiler)->fgFirstBB; (__block); (__block) = (__block)->bbNext)
10452 /*****************************************************************************/
10453 /*****************************************************************************/
10457 void dumpConvertedVarSet(Compiler* comp, VARSET_VALARG_TP vars);
10459 /*XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
10460 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
10462 XX Debugging helpers XX
10464 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
10465 XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
10468 /*****************************************************************************/
10469 /* The following functions are intended to be called from the debugger, to dump
10470 * various data structures. The can be used in the debugger Watch or Quick Watch
10471 * windows. They are designed to be short to type and take as few arguments as
10472 * possible. The 'c' versions take a Compiler*, whereas the 'd' versions use the TlsCompiler.
10473 * See the function definition comment for more details.
10476 void cBlock(Compiler* comp, BasicBlock* block);
10477 void cBlocks(Compiler* comp);
10478 void cBlocksV(Compiler* comp);
10479 void cTree(Compiler* comp, GenTree* tree);
10480 void cTrees(Compiler* comp);
10481 void cEH(Compiler* comp);
10482 void cVar(Compiler* comp, unsigned lclNum);
10483 void cVarDsc(Compiler* comp, LclVarDsc* varDsc);
10484 void cVars(Compiler* comp);
10485 void cVarsFinal(Compiler* comp);
10486 void cBlockPreds(Compiler* comp, BasicBlock* block);
10487 void cReach(Compiler* comp);
10488 void cDoms(Compiler* comp);
10489 void cLiveness(Compiler* comp);
10490 void cCVarSet(Compiler* comp, VARSET_VALARG_TP vars);
10492 void cFuncIR(Compiler* comp);
10493 void cBlockIR(Compiler* comp, BasicBlock* block);
10494 void cLoopIR(Compiler* comp, Compiler::LoopDsc* loop);
10495 void cTreeIR(Compiler* comp, GenTree* tree);
10496 int cTreeTypeIR(Compiler* comp, GenTree* tree);
10497 int cTreeKindsIR(Compiler* comp, GenTree* tree);
10498 int cTreeFlagsIR(Compiler* comp, GenTree* tree);
10499 int cOperandIR(Compiler* comp, GenTree* operand);
10500 int cLeafIR(Compiler* comp, GenTree* tree);
10501 int cIndirIR(Compiler* comp, GenTree* tree);
10502 int cListIR(Compiler* comp, GenTree* list);
10503 int cSsaNumIR(Compiler* comp, GenTree* tree);
10504 int cValNumIR(Compiler* comp, GenTree* tree);
10505 int cDependsIR(Compiler* comp, GenTree* comma, bool* first);
10507 void dBlock(BasicBlock* block);
10510 void dTree(GenTree* tree);
10513 void dVar(unsigned lclNum);
10514 void dVarDsc(LclVarDsc* varDsc);
10517 void dBlockPreds(BasicBlock* block);
10521 void dCVarSet(VARSET_VALARG_TP vars);
10523 void dRegMask(regMaskTP mask);
10526 void dBlockIR(BasicBlock* block);
10527 void dTreeIR(GenTree* tree);
10528 void dLoopIR(Compiler::LoopDsc* loop);
10529 void dLoopNumIR(unsigned loopNum);
10530 int dTabStopIR(int curr, int tabstop);
10531 int dTreeTypeIR(GenTree* tree);
10532 int dTreeKindsIR(GenTree* tree);
10533 int dTreeFlagsIR(GenTree* tree);
10534 int dOperandIR(GenTree* operand);
10535 int dLeafIR(GenTree* tree);
10536 int dIndirIR(GenTree* tree);
10537 int dListIR(GenTree* list);
10538 int dSsaNumIR(GenTree* tree);
10539 int dValNumIR(GenTree* tree);
10540 int dDependsIR(GenTree* comma);
10543 GenTree* dFindTree(GenTree* tree, unsigned id);
10544 GenTree* dFindTree(unsigned id);
10545 GenTreeStmt* dFindStmt(unsigned id);
10546 BasicBlock* dFindBlock(unsigned bbNum);
10550 #include "compiler.hpp" // All the shared inline functions
10552 /*****************************************************************************/
10553 #endif //_COMPILER_H_
10554 /*****************************************************************************/