1 //===-- X86ISelLowering.h - X86 DAG Lowering Interface ----------*- C++ -*-===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file defines the interfaces that X86 uses to lower LLVM code into a
13 //===----------------------------------------------------------------------===//
15 #ifndef X86ISELLOWERING_H
16 #define X86ISELLOWERING_H
18 #include "X86MachineFunctionInfo.h"
19 #include "X86RegisterInfo.h"
20 #include "X86Subtarget.h"
21 #include "llvm/CodeGen/CallingConvLower.h"
22 #include "llvm/CodeGen/FastISel.h"
23 #include "llvm/CodeGen/SelectionDAG.h"
24 #include "llvm/Target/TargetLowering.h"
25 #include "llvm/Target/TargetOptions.h"
29 // X86 Specific DAG Nodes
31 // Start the numbering where the builtin ops leave off.
32 FIRST_NUMBER = ISD::BUILTIN_OP_END,
34 /// BSF - Bit scan forward.
35 /// BSR - Bit scan reverse.
39 /// SHLD, SHRD - Double shift instructions. These correspond to
40 /// X86::SHLDxx and X86::SHRDxx instructions.
44 /// FAND - Bitwise logical AND of floating point values. This corresponds
45 /// to X86::ANDPS or X86::ANDPD.
48 /// FOR - Bitwise logical OR of floating point values. This corresponds
49 /// to X86::ORPS or X86::ORPD.
52 /// FXOR - Bitwise logical XOR of floating point values. This corresponds
53 /// to X86::XORPS or X86::XORPD.
56 /// FANDN - Bitwise logical ANDNOT of floating point values. This
57 /// corresponds to X86::ANDNPS or X86::ANDNPD.
60 /// FSRL - Bitwise logical right shift of floating point values. These
61 /// corresponds to X86::PSRLDQ.
64 /// CALL - These operations represent an abstract X86 call
65 /// instruction, which includes a bunch of information. In particular the
66 /// operands of these node are:
68 /// #0 - The incoming token chain
70 /// #2 - The number of arg bytes the caller pushes on the stack.
71 /// #3 - The number of arg bytes the callee pops off the stack.
72 /// #4 - The value to pass in AL/AX/EAX (optional)
73 /// #5 - The value to pass in DL/DX/EDX (optional)
75 /// The result values of these nodes are:
77 /// #0 - The outgoing token chain
78 /// #1 - The first register result value (optional)
79 /// #2 - The second register result value (optional)
83 /// RDTSC_DAG - This operation implements the lowering for
87 /// X86 compare and logical compare instructions.
90 /// X86 bit-test instructions.
93 /// X86 SetCC. Operand 0 is condition code, and operand 1 is the EFLAGS
94 /// operand, usually produced by a CMP instruction.
97 // Same as SETCC except it's materialized with a sbb and the value is all
98 // one's or all zero's.
99 SETCC_CARRY, // R = carry_bit ? ~0 : 0
101 /// X86 FP SETCC, implemented with CMP{cc}SS/CMP{cc}SD.
102 /// Operands are two FP values to compare; result is a mask of
103 /// 0s or 1s. Generally DTRT for C/C++ with NaNs.
106 /// X86 MOVMSK{pd|ps}, extracts sign bits of two or four FP values,
107 /// result in an integer GPR. Needs masking for scalar result.
110 /// X86 conditional moves. Operand 0 and operand 1 are the two values
111 /// to select from. Operand 2 is the condition code, and operand 3 is the
112 /// flag operand produced by a CMP or TEST instruction. It also writes a
116 /// X86 conditional branches. Operand 0 is the chain operand, operand 1
117 /// is the block to branch if condition is true, operand 2 is the
118 /// condition code, and operand 3 is the flag operand produced by a CMP
119 /// or TEST instruction.
122 /// Return with a flag operand. Operand 0 is the chain operand, operand
123 /// 1 is the number of bytes of stack to pop.
126 /// REP_STOS - Repeat fill, corresponds to X86::REP_STOSx.
129 /// REP_MOVS - Repeat move, corresponds to X86::REP_MOVSx.
132 /// GlobalBaseReg - On Darwin, this node represents the result of the popl
133 /// at function entry, used for PIC code.
136 /// Wrapper - A wrapper node for TargetConstantPool,
137 /// TargetExternalSymbol, and TargetGlobalAddress.
140 /// WrapperRIP - Special wrapper used under X86-64 PIC mode for RIP
141 /// relative displacements.
144 /// MOVDQ2Q - Copies a 64-bit value from the low word of an XMM vector
145 /// to an MMX vector. If you think this is too close to the previous
146 /// mnemonic, so do I; blame Intel.
149 /// MMX_MOVD2W - Copies a 32-bit value from the low word of a MMX
153 /// PEXTRB - Extract an 8-bit value from a vector and zero extend it to
154 /// i32, corresponds to X86::PEXTRB.
157 /// PEXTRW - Extract a 16-bit value from a vector and zero extend it to
158 /// i32, corresponds to X86::PEXTRW.
161 /// INSERTPS - Insert any element of a 4 x float vector into any element
162 /// of a destination 4 x floatvector.
165 /// PINSRB - Insert the lower 8-bits of a 32-bit value to a vector,
166 /// corresponds to X86::PINSRB.
169 /// PINSRW - Insert the lower 16-bits of a 32-bit value to a vector,
170 /// corresponds to X86::PINSRW.
173 /// PSHUFB - Shuffle 16 8-bit values within a vector.
176 /// ANDNP - Bitwise Logical AND NOT of Packed FP values.
179 /// PSIGN - Copy integer sign.
182 /// BLENDV - Blend where the selector is a register.
185 /// BLENDI - Blend where the selector is an immediate.
188 // SUBUS - Integer sub with unsigned saturation.
191 /// HADD - Integer horizontal add.
194 /// HSUB - Integer horizontal sub.
197 /// FHADD - Floating point horizontal add.
200 /// FHSUB - Floating point horizontal sub.
203 /// UMAX, UMIN - Unsigned integer max and min.
206 /// SMAX, SMIN - Signed integer max and min.
209 /// FMAX, FMIN - Floating point max and min.
213 /// FMAXC, FMINC - Commutative FMIN and FMAX.
216 /// FRSQRT, FRCP - Floating point reciprocal-sqrt and reciprocal
217 /// approximation. Note that these typically require refinement
218 /// in order to obtain suitable precision.
221 // TLSADDR - Thread Local Storage.
224 // TLSBASEADDR - Thread Local Storage. A call to get the start address
225 // of the TLS block for the current module.
228 // TLSCALL - Thread Local Storage. When calling to an OS provided
229 // thunk at the address from an earlier relocation.
232 // EH_RETURN - Exception Handling helpers.
235 // EH_SJLJ_SETJMP - SjLj exception handling setjmp.
238 // EH_SJLJ_LONGJMP - SjLj exception handling longjmp.
241 /// TC_RETURN - Tail call return. See X86TargetLowering::LowerCall for
242 /// the list of operands.
245 // VZEXT_MOVL - Vector move low and zero extend.
248 // VSEXT_MOVL - Vector move low and sign extend.
251 // VZEXT - Vector integer zero-extend.
254 // VSEXT - Vector integer signed-extend.
257 // VTRUNC - Vector integer truncate.
260 // VTRUNC - Vector integer truncate with mask.
263 // VFPEXT - Vector FP extend.
266 // VFPROUND - Vector FP round.
269 // VSHL, VSRL - 128-bit vector logical left / right shift
272 // VSHL, VSRL, VSRA - Vector shift elements
275 // VSHLI, VSRLI, VSRAI - Vector shift elements by immediate
278 // CMPP - Vector packed double/float comparison.
281 // PCMP* - Vector integer comparisons.
283 // PCMP*M - Vector integer comparisons, the result is in a mask vector.
286 /// CMPM, CMPMU - Vector comparison generating mask bits for fp and
287 /// integer signed and unsigned data types.
291 // ADD, SUB, SMUL, etc. - Arithmetic operations with FLAGS results.
292 ADD, SUB, ADC, SBB, SMUL,
293 INC, DEC, OR, XOR, AND,
295 BLSI, // BLSI - Extract lowest set isolated bit
296 BLSMSK, // BLSMSK - Get mask up to lowest set bit
297 BLSR, // BLSR - Reset lowest set bit
299 UMUL, // LOW, HI, FLAGS = umul LHS, RHS
301 // MUL_IMM - X86 specific multiply by immediate.
304 // PTEST - Vector bitwise comparisons.
307 // TESTP - Vector packed fp sign bitwise comparisons.
310 // TESTM - Vector "test" in AVX-512, the result is in a mask vector.
313 // OR/AND test for masks
317 // Several flavors of instructions with vector shuffle behaviors.
344 // PMULUDQ - Vector multiply packed unsigned doubleword integers
355 // VASTART_SAVE_XMM_REGS - Save xmm argument registers to the stack,
356 // according to %al. An operator is needed so that this can be expanded
357 // with control flow.
358 VASTART_SAVE_XMM_REGS,
360 // WIN_ALLOCA - Windows's _chkstk call to do stack probing.
363 // SEG_ALLOCA - For allocating variable amounts of stack space when using
364 // segmented stacks. Check if the current stacklet has enough space, and
365 // falls back to heap allocation if not.
368 // WIN_FTOL - Windows's _ftol2 runtime routine to do fptoui.
377 // FNSTSW16r - Store FP status word into i16 register.
380 // SAHF - Store contents of %ah into %eflags.
383 // RDRAND - Get a random integer and indicate whether it is valid in CF.
386 // RDSEED - Get a NIST SP800-90B & C compliant random integer and
387 // indicate whether it is valid in CF.
394 // XTEST - Test if in transactional execution.
397 // ATOMADD64_DAG, ATOMSUB64_DAG, ATOMOR64_DAG, ATOMAND64_DAG,
398 // ATOMXOR64_DAG, ATOMNAND64_DAG, ATOMSWAP64_DAG -
399 // Atomic 64-bit binary operations.
400 ATOMADD64_DAG = ISD::FIRST_TARGET_MEMORY_OPCODE,
412 // LCMPXCHG_DAG, LCMPXCHG8_DAG, LCMPXCHG16_DAG - Compare and swap.
417 // VZEXT_LOAD - Load, scalar_to_vector, and zero extend.
420 // FNSTCW16m - Store FP control world into i16 memory.
423 /// FP_TO_INT*_IN_MEM - This instruction implements FP_TO_SINT with the
424 /// integer destination in memory and a FP reg source. This corresponds
425 /// to the X86::FIST*m instructions and the rounding mode change stuff. It
426 /// has two inputs (token chain and address) and two outputs (int value
427 /// and token chain).
432 /// FILD, FILD_FLAG - This instruction implements SINT_TO_FP with the
433 /// integer source in memory and FP reg result. This corresponds to the
434 /// X86::FILD*m instructions. It has three inputs (token chain, address,
435 /// and source type) and two outputs (FP value and token chain). FILD_FLAG
436 /// also produces a flag).
440 /// FLD - This instruction implements an extending load to FP stack slots.
441 /// This corresponds to the X86::FLD32m / X86::FLD64m. It takes a chain
442 /// operand, ptr to load from, and a ValueType node indicating the type
446 /// FST - This instruction implements a truncating store to FP stack
447 /// slots. This corresponds to the X86::FST32m / X86::FST64m. It takes a
448 /// chain operand, value to store, address, and a ValueType to store it
452 /// VAARG_64 - This instruction grabs the address of the next argument
453 /// from a va_list. (reads and modifies the va_list in memory)
456 // WARNING: Do not add anything in the end unless you want the node to
457 // have memop! In fact, starting from ATOMADD64_DAG all opcodes will be
458 // thought as target memory ops!
462 /// Define some predicates that are used for node matching.
464 /// isVEXTRACT128Index - Return true if the specified
465 /// EXTRACT_SUBVECTOR operand specifies a vector extract that is
466 /// suitable for input to VEXTRACTF128, VEXTRACTI128 instructions.
467 bool isVEXTRACT128Index(SDNode *N);
469 /// isVINSERT128Index - Return true if the specified
470 /// INSERT_SUBVECTOR operand specifies a subvector insert that is
471 /// suitable for input to VINSERTF128, VINSERTI128 instructions.
472 bool isVINSERT128Index(SDNode *N);
474 /// isVEXTRACT256Index - Return true if the specified
475 /// EXTRACT_SUBVECTOR operand specifies a vector extract that is
476 /// suitable for input to VEXTRACTF64X4, VEXTRACTI64X4 instructions.
477 bool isVEXTRACT256Index(SDNode *N);
479 /// isVINSERT256Index - Return true if the specified
480 /// INSERT_SUBVECTOR operand specifies a subvector insert that is
481 /// suitable for input to VINSERTF64X4, VINSERTI64X4 instructions.
482 bool isVINSERT256Index(SDNode *N);
484 /// getExtractVEXTRACT128Immediate - Return the appropriate
485 /// immediate to extract the specified EXTRACT_SUBVECTOR index
486 /// with VEXTRACTF128, VEXTRACTI128 instructions.
487 unsigned getExtractVEXTRACT128Immediate(SDNode *N);
489 /// getInsertVINSERT128Immediate - Return the appropriate
490 /// immediate to insert at the specified INSERT_SUBVECTOR index
491 /// with VINSERTF128, VINSERT128 instructions.
492 unsigned getInsertVINSERT128Immediate(SDNode *N);
494 /// getExtractVEXTRACT256Immediate - Return the appropriate
495 /// immediate to extract the specified EXTRACT_SUBVECTOR index
496 /// with VEXTRACTF64X4, VEXTRACTI64x4 instructions.
497 unsigned getExtractVEXTRACT256Immediate(SDNode *N);
499 /// getInsertVINSERT256Immediate - Return the appropriate
500 /// immediate to insert at the specified INSERT_SUBVECTOR index
501 /// with VINSERTF64x4, VINSERTI64x4 instructions.
502 unsigned getInsertVINSERT256Immediate(SDNode *N);
504 /// isZeroNode - Returns true if Elt is a constant zero or a floating point
506 bool isZeroNode(SDValue Elt);
508 /// isOffsetSuitableForCodeModel - Returns true of the given offset can be
509 /// fit into displacement field of the instruction.
510 bool isOffsetSuitableForCodeModel(int64_t Offset, CodeModel::Model M,
511 bool hasSymbolicDisplacement = true);
514 /// isCalleePop - Determines whether the callee is required to pop its
515 /// own arguments. Callee pop is necessary to support tail calls.
516 bool isCalleePop(CallingConv::ID CallingConv,
517 bool is64Bit, bool IsVarArg, bool TailCallOpt);
520 //===--------------------------------------------------------------------===//
521 // X86TargetLowering - X86 Implementation of the TargetLowering interface
522 class X86TargetLowering : public TargetLowering {
524 explicit X86TargetLowering(X86TargetMachine &TM);
526 virtual unsigned getJumpTableEncoding() const;
528 virtual MVT getScalarShiftAmountTy(EVT LHSTy) const { return MVT::i8; }
530 virtual const MCExpr *
531 LowerCustomJumpTableEntry(const MachineJumpTableInfo *MJTI,
532 const MachineBasicBlock *MBB, unsigned uid,
533 MCContext &Ctx) const;
535 /// getPICJumpTableRelocaBase - Returns relocation base for the given PIC
537 virtual SDValue getPICJumpTableRelocBase(SDValue Table,
538 SelectionDAG &DAG) const;
539 virtual const MCExpr *
540 getPICJumpTableRelocBaseExpr(const MachineFunction *MF,
541 unsigned JTI, MCContext &Ctx) const;
543 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
544 /// function arguments in the caller parameter area. For X86, aggregates
545 /// that contains are placed at 16-byte boundaries while the rest are at
546 /// 4-byte boundaries.
547 virtual unsigned getByValTypeAlignment(Type *Ty) const;
549 /// getOptimalMemOpType - Returns the target specific optimal type for load
550 /// and store operations as a result of memset, memcpy, and memmove
551 /// lowering. If DstAlign is zero that means it's safe to destination
552 /// alignment can satisfy any constraint. Similarly if SrcAlign is zero it
553 /// means there isn't a need to check it against alignment requirement,
554 /// probably because the source does not need to be loaded. If 'IsMemset' is
555 /// true, that means it's expanding a memset. If 'ZeroMemset' is true, that
556 /// means it's a memset of zero. 'MemcpyStrSrc' indicates whether the memcpy
557 /// source is constant so it does not need to be loaded.
558 /// It returns EVT::Other if the type should be determined using generic
559 /// target-independent logic.
561 getOptimalMemOpType(uint64_t Size, unsigned DstAlign, unsigned SrcAlign,
562 bool IsMemset, bool ZeroMemset, bool MemcpyStrSrc,
563 MachineFunction &MF) const;
565 /// isSafeMemOpType - Returns true if it's safe to use load / store of the
566 /// specified type to expand memcpy / memset inline. This is mostly true
567 /// for all types except for some special cases. For example, on X86
568 /// targets without SSE2 f64 load / store are done with fldl / fstpl which
569 /// also does type conversion. Note the specified type doesn't have to be
570 /// legal as the hook is used before type legalization.
571 virtual bool isSafeMemOpType(MVT VT) const;
573 /// allowsUnalignedMemoryAccesses - Returns true if the target allows
574 /// unaligned memory accesses. of the specified type. Returns whether it
575 /// is "fast" by reference in the second argument.
576 virtual bool allowsUnalignedMemoryAccesses(EVT VT, bool *Fast) const;
578 /// LowerOperation - Provide custom lowering hooks for some operations.
580 virtual SDValue LowerOperation(SDValue Op, SelectionDAG &DAG) const;
582 /// ReplaceNodeResults - Replace the results of node with an illegal result
583 /// type with new values built out of custom code.
585 virtual void ReplaceNodeResults(SDNode *N, SmallVectorImpl<SDValue>&Results,
586 SelectionDAG &DAG) const;
589 virtual SDValue PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const;
591 /// isTypeDesirableForOp - Return true if the target has native support for
592 /// the specified value type and it is 'desirable' to use the type for the
593 /// given node type. e.g. On x86 i16 is legal, but undesirable since i16
594 /// instruction encodings are longer and some i16 instructions are slow.
595 virtual bool isTypeDesirableForOp(unsigned Opc, EVT VT) const;
597 /// isTypeDesirable - Return true if the target has native support for the
598 /// specified value type and it is 'desirable' to use the type. e.g. On x86
599 /// i16 is legal, but undesirable since i16 instruction encodings are longer
600 /// and some i16 instructions are slow.
601 virtual bool IsDesirableToPromoteOp(SDValue Op, EVT &PVT) const;
603 virtual MachineBasicBlock *
604 EmitInstrWithCustomInserter(MachineInstr *MI,
605 MachineBasicBlock *MBB) const;
608 /// getTargetNodeName - This method returns the name of a target specific
610 virtual const char *getTargetNodeName(unsigned Opcode) const;
612 /// getSetCCResultType - Return the value type to use for ISD::SETCC.
613 virtual EVT getSetCCResultType(LLVMContext &Context, EVT VT) const;
615 /// computeMaskedBitsForTargetNode - Determine which of the bits specified
616 /// in Mask are known to be either zero or one and return them in the
617 /// KnownZero/KnownOne bitsets.
618 virtual void computeMaskedBitsForTargetNode(const SDValue Op,
621 const SelectionDAG &DAG,
622 unsigned Depth = 0) const;
624 // ComputeNumSignBitsForTargetNode - Determine the number of bits in the
625 // operation that are sign bits.
626 virtual unsigned ComputeNumSignBitsForTargetNode(SDValue Op,
627 unsigned Depth) const;
630 isGAPlusOffset(SDNode *N, const GlobalValue* &GA, int64_t &Offset) const;
632 SDValue getReturnAddressFrameIndex(SelectionDAG &DAG) const;
634 virtual bool ExpandInlineAsm(CallInst *CI) const;
636 ConstraintType getConstraintType(const std::string &Constraint) const;
638 /// Examine constraint string and operand type and determine a weight value.
639 /// The operand object must already have been set up with the operand type.
640 virtual ConstraintWeight getSingleConstraintMatchWeight(
641 AsmOperandInfo &info, const char *constraint) const;
643 virtual const char *LowerXConstraint(EVT ConstraintVT) const;
645 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
646 /// vector. If it is invalid, don't add anything to Ops. If hasMemory is
647 /// true it means one of the asm constraint of the inline asm instruction
648 /// being processed is 'm'.
649 virtual void LowerAsmOperandForConstraint(SDValue Op,
650 std::string &Constraint,
651 std::vector<SDValue> &Ops,
652 SelectionDAG &DAG) const;
654 /// getRegForInlineAsmConstraint - Given a physical register constraint
655 /// (e.g. {edx}), return the register number and the register class for the
656 /// register. This should only be used for C_Register constraints. On
657 /// error, this returns a register number of 0.
658 std::pair<unsigned, const TargetRegisterClass*>
659 getRegForInlineAsmConstraint(const std::string &Constraint,
662 /// isLegalAddressingMode - Return true if the addressing mode represented
663 /// by AM is legal for this target, for a load/store of the specified type.
664 virtual bool isLegalAddressingMode(const AddrMode &AM, Type *Ty)const;
666 /// isLegalICmpImmediate - Return true if the specified immediate is legal
667 /// icmp immediate, that is the target has icmp instructions which can
668 /// compare a register against the immediate without having to materialize
669 /// the immediate into a register.
670 virtual bool isLegalICmpImmediate(int64_t Imm) const;
672 /// isLegalAddImmediate - Return true if the specified immediate is legal
673 /// add immediate, that is the target has add instructions which can
674 /// add a register and the immediate without having to materialize
675 /// the immediate into a register.
676 virtual bool isLegalAddImmediate(int64_t Imm) const;
678 /// isTruncateFree - Return true if it's free to truncate a value of
679 /// type Ty1 to type Ty2. e.g. On x86 it's free to truncate a i32 value in
680 /// register EAX to i16 by referencing its sub-register AX.
681 virtual bool isTruncateFree(Type *Ty1, Type *Ty2) const;
682 virtual bool isTruncateFree(EVT VT1, EVT VT2) const;
684 virtual bool allowTruncateForTailCall(Type *Ty1, Type *Ty2) const;
686 /// isZExtFree - Return true if any actual instruction that defines a
687 /// value of type Ty1 implicit zero-extends the value to Ty2 in the result
688 /// register. This does not necessarily include registers defined in
689 /// unknown ways, such as incoming arguments, or copies from unknown
690 /// virtual registers. Also, if isTruncateFree(Ty2, Ty1) is true, this
691 /// does not necessarily apply to truncate instructions. e.g. on x86-64,
692 /// all instructions that define 32-bit values implicit zero-extend the
693 /// result out to 64 bits.
694 virtual bool isZExtFree(Type *Ty1, Type *Ty2) const;
695 virtual bool isZExtFree(EVT VT1, EVT VT2) const;
696 virtual bool isZExtFree(SDValue Val, EVT VT2) const;
698 /// isFMAFasterThanFMulAndFAdd - Return true if an FMA operation is faster
699 /// than a pair of fmul and fadd instructions. fmuladd intrinsics will be
700 /// expanded to FMAs when this method returns true, otherwise fmuladd is
701 /// expanded to fmul + fadd.
702 virtual bool isFMAFasterThanFMulAndFAdd(EVT VT) const;
704 /// isNarrowingProfitable - Return true if it's profitable to narrow
705 /// operations of type VT1 to VT2. e.g. on x86, it's profitable to narrow
706 /// from i32 to i8 but not from i32 to i16.
707 virtual bool isNarrowingProfitable(EVT VT1, EVT VT2) const;
709 /// isFPImmLegal - Returns true if the target can instruction select the
710 /// specified FP immediate natively. If false, the legalizer will
711 /// materialize the FP immediate as a load from a constant pool.
712 virtual bool isFPImmLegal(const APFloat &Imm, EVT VT) const;
714 /// isShuffleMaskLegal - Targets can use this to indicate that they only
715 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
716 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask
717 /// values are assumed to be legal.
718 virtual bool isShuffleMaskLegal(const SmallVectorImpl<int> &Mask,
721 /// isVectorClearMaskLegal - Similar to isShuffleMaskLegal. This is
722 /// used by Targets can use this to indicate if there is a suitable
723 /// VECTOR_SHUFFLE that can be used to replace a VAND with a constant
725 virtual bool isVectorClearMaskLegal(const SmallVectorImpl<int> &Mask,
728 /// ShouldShrinkFPConstant - If true, then instruction selection should
729 /// seek to shrink the FP constant of the specified type to a smaller type
730 /// in order to save space and / or reduce runtime.
731 virtual bool ShouldShrinkFPConstant(EVT VT) const {
732 // Don't shrink FP constpool if SSE2 is available since cvtss2sd is more
733 // expensive than a straight movsd. On the other hand, it's important to
734 // shrink long double fp constant since fldt is very slow.
735 return !X86ScalarSSEf64 || VT == MVT::f80;
738 const X86Subtarget* getSubtarget() const {
742 /// isScalarFPTypeInSSEReg - Return true if the specified scalar FP type is
743 /// computed in an SSE register, not on the X87 floating point stack.
744 bool isScalarFPTypeInSSEReg(EVT VT) const {
745 return (VT == MVT::f64 && X86ScalarSSEf64) || // f64 is when SSE2
746 (VT == MVT::f32 && X86ScalarSSEf32); // f32 is when SSE1
749 /// isTargetFTOL - Return true if the target uses the MSVC _ftol2 routine
751 bool isTargetFTOL() const {
752 return Subtarget->isTargetWindows() && !Subtarget->is64Bit();
755 /// isIntegerTypeFTOL - Return true if the MSVC _ftol2 routine should be
756 /// used for fptoui to the given type.
757 bool isIntegerTypeFTOL(EVT VT) const {
758 return isTargetFTOL() && VT == MVT::i64;
761 /// createFastISel - This method returns a target specific FastISel object,
762 /// or null if the target does not support "fast" ISel.
763 virtual FastISel *createFastISel(FunctionLoweringInfo &funcInfo,
764 const TargetLibraryInfo *libInfo) const;
766 /// getStackCookieLocation - Return true if the target stores stack
767 /// protector cookies at a fixed offset in some non-standard address
768 /// space, and populates the address space and offset as
770 virtual bool getStackCookieLocation(unsigned &AddressSpace, unsigned &Offset) const;
772 SDValue BuildFILD(SDValue Op, EVT SrcVT, SDValue Chain, SDValue StackSlot,
773 SelectionDAG &DAG) const;
775 /// \brief Reset the operation actions based on target options.
776 virtual void resetOperationActions();
779 std::pair<const TargetRegisterClass*, uint8_t>
780 findRepresentativeClass(MVT VT) const;
783 /// Subtarget - Keep a pointer to the X86Subtarget around so that we can
784 /// make the right decision when generating code for different targets.
785 const X86Subtarget *Subtarget;
786 const DataLayout *TD;
788 /// Used to store the TargetOptions so that we don't waste time resetting
789 /// the operation actions unless we have to.
792 /// X86ScalarSSEf32, X86ScalarSSEf64 - Select between SSE or x87
793 /// floating point ops.
794 /// When SSE is available, use it for f32 operations.
795 /// When SSE2 is available, use it for f64 operations.
796 bool X86ScalarSSEf32;
797 bool X86ScalarSSEf64;
799 /// LegalFPImmediates - A list of legal fp immediates.
800 std::vector<APFloat> LegalFPImmediates;
802 /// addLegalFPImmediate - Indicate that this x86 target can instruction
803 /// select the specified FP immediate natively.
804 void addLegalFPImmediate(const APFloat& Imm) {
805 LegalFPImmediates.push_back(Imm);
808 SDValue LowerCallResult(SDValue Chain, SDValue InFlag,
809 CallingConv::ID CallConv, bool isVarArg,
810 const SmallVectorImpl<ISD::InputArg> &Ins,
811 SDLoc dl, SelectionDAG &DAG,
812 SmallVectorImpl<SDValue> &InVals) const;
813 SDValue LowerMemArgument(SDValue Chain,
814 CallingConv::ID CallConv,
815 const SmallVectorImpl<ISD::InputArg> &ArgInfo,
816 SDLoc dl, SelectionDAG &DAG,
817 const CCValAssign &VA, MachineFrameInfo *MFI,
819 SDValue LowerMemOpCallTo(SDValue Chain, SDValue StackPtr, SDValue Arg,
820 SDLoc dl, SelectionDAG &DAG,
821 const CCValAssign &VA,
822 ISD::ArgFlagsTy Flags) const;
824 // Call lowering helpers.
826 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
827 /// for tail call optimization. Targets which want to do tail call
828 /// optimization should implement this function.
829 bool IsEligibleForTailCallOptimization(SDValue Callee,
830 CallingConv::ID CalleeCC,
832 bool isCalleeStructRet,
833 bool isCallerStructRet,
835 const SmallVectorImpl<ISD::OutputArg> &Outs,
836 const SmallVectorImpl<SDValue> &OutVals,
837 const SmallVectorImpl<ISD::InputArg> &Ins,
838 SelectionDAG& DAG) const;
839 bool IsCalleePop(bool isVarArg, CallingConv::ID CallConv) const;
840 SDValue EmitTailCallLoadRetAddr(SelectionDAG &DAG, SDValue &OutRetAddr,
841 SDValue Chain, bool IsTailCall, bool Is64Bit,
842 int FPDiff, SDLoc dl) const;
844 unsigned GetAlignedArgumentStackSize(unsigned StackSize,
845 SelectionDAG &DAG) const;
847 std::pair<SDValue,SDValue> FP_TO_INTHelper(SDValue Op, SelectionDAG &DAG,
849 bool isReplace) const;
851 SDValue LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) const;
852 SDValue LowerBUILD_VECTORvXi1(SDValue Op, SelectionDAG &DAG) const;
853 SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) const;
854 SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const;
855 SDValue LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const;
856 SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) const;
857 SDValue LowerBlockAddress(SDValue Op, SelectionDAG &DAG) const;
858 SDValue LowerGlobalAddress(const GlobalValue *GV, SDLoc dl,
859 int64_t Offset, SelectionDAG &DAG) const;
860 SDValue LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) const;
861 SDValue LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const;
862 SDValue LowerExternalSymbol(SDValue Op, SelectionDAG &DAG) const;
863 SDValue LowerShiftParts(SDValue Op, SelectionDAG &DAG) const;
864 SDValue LowerSINT_TO_FP(SDValue Op, SelectionDAG &DAG) const;
865 SDValue LowerUINT_TO_FP(SDValue Op, SelectionDAG &DAG) const;
866 SDValue LowerUINT_TO_FP_i64(SDValue Op, SelectionDAG &DAG) const;
867 SDValue LowerUINT_TO_FP_i32(SDValue Op, SelectionDAG &DAG) const;
868 SDValue lowerUINT_TO_FP_vec(SDValue Op, SelectionDAG &DAG) const;
869 SDValue LowerTRUNCATE(SDValue Op, SelectionDAG &DAG) const;
870 SDValue LowerFP_TO_SINT(SDValue Op, SelectionDAG &DAG) const;
871 SDValue LowerFP_TO_UINT(SDValue Op, SelectionDAG &DAG) const;
872 SDValue LowerFABS(SDValue Op, SelectionDAG &DAG) const;
873 SDValue LowerFNEG(SDValue Op, SelectionDAG &DAG) const;
874 SDValue LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const;
875 SDValue LowerToBT(SDValue And, ISD::CondCode CC,
876 SDLoc dl, SelectionDAG &DAG) const;
877 SDValue LowerSETCC(SDValue Op, SelectionDAG &DAG) const;
878 SDValue LowerSELECT(SDValue Op, SelectionDAG &DAG) const;
879 SDValue LowerBRCOND(SDValue Op, SelectionDAG &DAG) const;
880 SDValue LowerMEMSET(SDValue Op, SelectionDAG &DAG) const;
881 SDValue LowerJumpTable(SDValue Op, SelectionDAG &DAG) const;
882 SDValue LowerDYNAMIC_STACKALLOC(SDValue Op, SelectionDAG &DAG) const;
883 SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG) const;
884 SDValue LowerVAARG(SDValue Op, SelectionDAG &DAG) const;
885 SDValue LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const;
886 SDValue LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const;
887 SDValue LowerFRAME_TO_ARGS_OFFSET(SDValue Op, SelectionDAG &DAG) const;
888 SDValue LowerEH_RETURN(SDValue Op, SelectionDAG &DAG) const;
889 SDValue lowerEH_SJLJ_SETJMP(SDValue Op, SelectionDAG &DAG) const;
890 SDValue lowerEH_SJLJ_LONGJMP(SDValue Op, SelectionDAG &DAG) const;
891 SDValue LowerINIT_TRAMPOLINE(SDValue Op, SelectionDAG &DAG) const;
892 SDValue LowerFLT_ROUNDS_(SDValue Op, SelectionDAG &DAG) const;
893 SDValue LowerSIGN_EXTEND_INREG(SDValue Op, SelectionDAG &DAG) const;
896 LowerFormalArguments(SDValue Chain,
897 CallingConv::ID CallConv, bool isVarArg,
898 const SmallVectorImpl<ISD::InputArg> &Ins,
899 SDLoc dl, SelectionDAG &DAG,
900 SmallVectorImpl<SDValue> &InVals) const;
902 LowerCall(CallLoweringInfo &CLI,
903 SmallVectorImpl<SDValue> &InVals) const;
906 LowerReturn(SDValue Chain,
907 CallingConv::ID CallConv, bool isVarArg,
908 const SmallVectorImpl<ISD::OutputArg> &Outs,
909 const SmallVectorImpl<SDValue> &OutVals,
910 SDLoc dl, SelectionDAG &DAG) const;
912 virtual bool isUsedByReturnOnly(SDNode *N, SDValue &Chain) const;
914 virtual bool mayBeEmittedAsTailCall(CallInst *CI) const;
917 getTypeForExtArgOrReturn(MVT VT, ISD::NodeType ExtendKind) const;
920 CanLowerReturn(CallingConv::ID CallConv, MachineFunction &MF,
922 const SmallVectorImpl<ISD::OutputArg> &Outs,
923 LLVMContext &Context) const;
925 /// Utility function to emit atomic-load-arith operations (and, or, xor,
926 /// nand, max, min, umax, umin). It takes the corresponding instruction to
927 /// expand, the associated machine basic block, and the associated X86
928 /// opcodes for reg/reg.
929 MachineBasicBlock *EmitAtomicLoadArith(MachineInstr *MI,
930 MachineBasicBlock *MBB) const;
932 /// Utility function to emit atomic-load-arith operations (and, or, xor,
933 /// nand, add, sub, swap) for 64-bit operands on 32-bit target.
934 MachineBasicBlock *EmitAtomicLoadArith6432(MachineInstr *MI,
935 MachineBasicBlock *MBB) const;
937 // Utility function to emit the low-level va_arg code for X86-64.
938 MachineBasicBlock *EmitVAARG64WithCustomInserter(
940 MachineBasicBlock *MBB) const;
942 /// Utility function to emit the xmm reg save portion of va_start.
943 MachineBasicBlock *EmitVAStartSaveXMMRegsWithCustomInserter(
944 MachineInstr *BInstr,
945 MachineBasicBlock *BB) const;
947 MachineBasicBlock *EmitLoweredSelect(MachineInstr *I,
948 MachineBasicBlock *BB) const;
950 MachineBasicBlock *EmitLoweredWinAlloca(MachineInstr *MI,
951 MachineBasicBlock *BB) const;
953 MachineBasicBlock *EmitLoweredSegAlloca(MachineInstr *MI,
954 MachineBasicBlock *BB,
957 MachineBasicBlock *EmitLoweredTLSCall(MachineInstr *MI,
958 MachineBasicBlock *BB) const;
960 MachineBasicBlock *emitLoweredTLSAddr(MachineInstr *MI,
961 MachineBasicBlock *BB) const;
963 MachineBasicBlock *emitEHSjLjSetJmp(MachineInstr *MI,
964 MachineBasicBlock *MBB) const;
966 MachineBasicBlock *emitEHSjLjLongJmp(MachineInstr *MI,
967 MachineBasicBlock *MBB) const;
969 /// Emit nodes that will be selected as "test Op0,Op0", or something
970 /// equivalent, for use with the given x86 condition code.
971 SDValue EmitTest(SDValue Op0, unsigned X86CC, SelectionDAG &DAG) const;
973 /// Emit nodes that will be selected as "cmp Op0,Op1", or something
974 /// equivalent, for use with the given x86 condition code.
975 SDValue EmitCmp(SDValue Op0, SDValue Op1, unsigned X86CC,
976 SelectionDAG &DAG) const;
978 /// Convert a comparison if required by the subtarget.
979 SDValue ConvertCmpIfNecessary(SDValue Cmp, SelectionDAG &DAG) const;
983 FastISel *createFastISel(FunctionLoweringInfo &funcInfo,
984 const TargetLibraryInfo *libInfo);
988 #endif // X86ISELLOWERING_H