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 "X86Subtarget.h"
19 #include "llvm/CodeGen/CallingConvLower.h"
20 #include "llvm/CodeGen/SelectionDAG.h"
21 #include "llvm/Target/TargetLowering.h"
22 #include "llvm/Target/TargetOptions.h"
25 class X86TargetMachine;
28 // X86 Specific DAG Nodes
30 // Start the numbering where the builtin ops leave off.
31 FIRST_NUMBER = ISD::BUILTIN_OP_END,
33 /// BSF - Bit scan forward.
34 /// BSR - Bit scan reverse.
38 /// SHLD, SHRD - Double shift instructions. These correspond to
39 /// X86::SHLDxx and X86::SHRDxx instructions.
43 /// FAND - Bitwise logical AND of floating point values. This corresponds
44 /// to X86::ANDPS or X86::ANDPD.
47 /// FOR - Bitwise logical OR of floating point values. This corresponds
48 /// to X86::ORPS or X86::ORPD.
51 /// FXOR - Bitwise logical XOR of floating point values. This corresponds
52 /// to X86::XORPS or X86::XORPD.
55 /// FANDN - Bitwise logical ANDNOT of floating point values. This
56 /// corresponds to X86::ANDNPS or X86::ANDNPD.
59 /// FSRL - Bitwise logical right shift of floating point values. These
60 /// corresponds to X86::PSRLDQ.
63 /// CALL - These operations represent an abstract X86 call
64 /// instruction, which includes a bunch of information. In particular the
65 /// operands of these node are:
67 /// #0 - The incoming token chain
69 /// #2 - The number of arg bytes the caller pushes on the stack.
70 /// #3 - The number of arg bytes the callee pops off the stack.
71 /// #4 - The value to pass in AL/AX/EAX (optional)
72 /// #5 - The value to pass in DL/DX/EDX (optional)
74 /// The result values of these nodes are:
76 /// #0 - The outgoing token chain
77 /// #1 - The first register result value (optional)
78 /// #2 - The second register result value (optional)
82 /// RDTSC_DAG - This operation implements the lowering for
86 /// X86 Read Time-Stamp Counter and Processor ID.
89 /// X86 compare and logical compare instructions.
92 /// X86 bit-test instructions.
95 /// X86 SetCC. Operand 0 is condition code, and operand 1 is the EFLAGS
96 /// operand, usually produced by a CMP instruction.
102 // Same as SETCC except it's materialized with a sbb and the value is all
103 // one's or all zero's.
104 SETCC_CARRY, // R = carry_bit ? ~0 : 0
106 /// X86 FP SETCC, implemented with CMP{cc}SS/CMP{cc}SD.
107 /// Operands are two FP values to compare; result is a mask of
108 /// 0s or 1s. Generally DTRT for C/C++ with NaNs.
111 /// X86 MOVMSK{pd|ps}, extracts sign bits of two or four FP values,
112 /// result in an integer GPR. Needs masking for scalar result.
115 /// X86 conditional moves. Operand 0 and operand 1 are the two values
116 /// to select from. Operand 2 is the condition code, and operand 3 is the
117 /// flag operand produced by a CMP or TEST instruction. It also writes a
121 /// X86 conditional branches. Operand 0 is the chain operand, operand 1
122 /// is the block to branch if condition is true, operand 2 is the
123 /// condition code, and operand 3 is the flag operand produced by a CMP
124 /// or TEST instruction.
127 /// Return with a flag operand. Operand 0 is the chain operand, operand
128 /// 1 is the number of bytes of stack to pop.
131 /// REP_STOS - Repeat fill, corresponds to X86::REP_STOSx.
134 /// REP_MOVS - Repeat move, corresponds to X86::REP_MOVSx.
137 /// GlobalBaseReg - On Darwin, this node represents the result of the popl
138 /// at function entry, used for PIC code.
141 /// Wrapper - A wrapper node for TargetConstantPool,
142 /// TargetExternalSymbol, and TargetGlobalAddress.
145 /// WrapperRIP - Special wrapper used under X86-64 PIC mode for RIP
146 /// relative displacements.
149 /// MOVDQ2Q - Copies a 64-bit value from the low word of an XMM vector
150 /// to an MMX vector. If you think this is too close to the previous
151 /// mnemonic, so do I; blame Intel.
154 /// MMX_MOVD2W - Copies a 32-bit value from the low word of a MMX
158 /// PEXTRB - Extract an 8-bit value from a vector and zero extend it to
159 /// i32, corresponds to X86::PEXTRB.
162 /// PEXTRW - Extract a 16-bit value from a vector and zero extend it to
163 /// i32, corresponds to X86::PEXTRW.
166 /// INSERTPS - Insert any element of a 4 x float vector into any element
167 /// of a destination 4 x floatvector.
170 /// PINSRB - Insert the lower 8-bits of a 32-bit value to a vector,
171 /// corresponds to X86::PINSRB.
174 /// PINSRW - Insert the lower 16-bits of a 32-bit value to a vector,
175 /// corresponds to X86::PINSRW.
178 /// PSHUFB - Shuffle 16 8-bit values within a vector.
181 /// ANDNP - Bitwise Logical AND NOT of Packed FP values.
184 /// PSIGN - Copy integer sign.
187 /// BLENDV - Blend where the selector is a register.
190 /// BLENDI - Blend where the selector is an immediate.
193 // SUBUS - Integer sub with unsigned saturation.
196 /// HADD - Integer horizontal add.
199 /// HSUB - Integer horizontal sub.
202 /// FHADD - Floating point horizontal add.
205 /// FHSUB - Floating point horizontal sub.
208 /// UMAX, UMIN - Unsigned integer max and min.
211 /// SMAX, SMIN - Signed integer max and min.
214 /// FMAX, FMIN - Floating point max and min.
218 /// FMAXC, FMINC - Commutative FMIN and FMAX.
221 /// FRSQRT, FRCP - Floating point reciprocal-sqrt and reciprocal
222 /// approximation. Note that these typically require refinement
223 /// in order to obtain suitable precision.
226 // TLSADDR - Thread Local Storage.
229 // TLSBASEADDR - Thread Local Storage. A call to get the start address
230 // of the TLS block for the current module.
233 // TLSCALL - Thread Local Storage. When calling to an OS provided
234 // thunk at the address from an earlier relocation.
237 // EH_RETURN - Exception Handling helpers.
240 // EH_SJLJ_SETJMP - SjLj exception handling setjmp.
243 // EH_SJLJ_LONGJMP - SjLj exception handling longjmp.
246 /// TC_RETURN - Tail call return. See X86TargetLowering::LowerCall for
247 /// the list of operands.
250 // VZEXT_MOVL - Vector move to low scalar and zero higher vector elements.
253 // VZEXT - Vector integer zero-extend.
256 // VSEXT - Vector integer signed-extend.
259 // VTRUNC - Vector integer truncate.
262 // VTRUNC - Vector integer truncate with mask.
265 // VFPEXT - Vector FP extend.
268 // VFPROUND - Vector FP round.
271 // VSHL, VSRL - 128-bit vector logical left / right shift
274 // VSHL, VSRL, VSRA - Vector shift elements
277 // VSHLI, VSRLI, VSRAI - Vector shift elements by immediate
280 // CMPP - Vector packed double/float comparison.
283 // PCMP* - Vector integer comparisons.
285 // PCMP*M - Vector integer comparisons, the result is in a mask vector.
288 /// CMPM, CMPMU - Vector comparison generating mask bits for fp and
289 /// integer signed and unsigned data types.
293 // ADD, SUB, SMUL, etc. - Arithmetic operations with FLAGS results.
294 ADD, SUB, ADC, SBB, SMUL,
295 INC, DEC, OR, XOR, AND,
297 BEXTR, // BEXTR - Bit field extract
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, TESTNM - Vector "test" in AVX-512, the result is in a mask vector.
314 // OR/AND test for masks
317 // Several flavors of instructions with vector shuffle behaviors.
344 // Insert/Extract vector element
348 // PMULUDQ - Vector multiply packed unsigned doubleword integers
359 // VASTART_SAVE_XMM_REGS - Save xmm argument registers to the stack,
360 // according to %al. An operator is needed so that this can be expanded
361 // with control flow.
362 VASTART_SAVE_XMM_REGS,
364 // WIN_ALLOCA - Windows's _chkstk call to do stack probing.
367 // SEG_ALLOCA - For allocating variable amounts of stack space when using
368 // segmented stacks. Check if the current stacklet has enough space, and
369 // falls back to heap allocation if not.
372 // WIN_FTOL - Windows's _ftol2 runtime routine to do fptoui.
381 // FNSTSW16r - Store FP status word into i16 register.
384 // SAHF - Store contents of %ah into %eflags.
387 // RDRAND - Get a random integer and indicate whether it is valid in CF.
390 // RDSEED - Get a NIST SP800-90B & C compliant random integer and
391 // indicate whether it is valid in CF.
398 // XTEST - Test if in transactional execution.
401 // ATOMADD64_DAG, ATOMSUB64_DAG, ATOMOR64_DAG, ATOMAND64_DAG,
402 // ATOMXOR64_DAG, ATOMNAND64_DAG, ATOMSWAP64_DAG -
403 // Atomic 64-bit binary operations.
404 ATOMADD64_DAG = ISD::FIRST_TARGET_MEMORY_OPCODE,
416 // LCMPXCHG_DAG, LCMPXCHG8_DAG, LCMPXCHG16_DAG - Compare and swap.
421 // VZEXT_LOAD - Load, scalar_to_vector, and zero extend.
424 // FNSTCW16m - Store FP control world into i16 memory.
427 /// FP_TO_INT*_IN_MEM - This instruction implements FP_TO_SINT with the
428 /// integer destination in memory and a FP reg source. This corresponds
429 /// to the X86::FIST*m instructions and the rounding mode change stuff. It
430 /// has two inputs (token chain and address) and two outputs (int value
431 /// and token chain).
436 /// FILD, FILD_FLAG - This instruction implements SINT_TO_FP with the
437 /// integer source in memory and FP reg result. This corresponds to the
438 /// X86::FILD*m instructions. It has three inputs (token chain, address,
439 /// and source type) and two outputs (FP value and token chain). FILD_FLAG
440 /// also produces a flag).
444 /// FLD - This instruction implements an extending load to FP stack slots.
445 /// This corresponds to the X86::FLD32m / X86::FLD64m. It takes a chain
446 /// operand, ptr to load from, and a ValueType node indicating the type
450 /// FST - This instruction implements a truncating store to FP stack
451 /// slots. This corresponds to the X86::FST32m / X86::FST64m. It takes a
452 /// chain operand, value to store, address, and a ValueType to store it
456 /// VAARG_64 - This instruction grabs the address of the next argument
457 /// from a va_list. (reads and modifies the va_list in memory)
460 // WARNING: Do not add anything in the end unless you want the node to
461 // have memop! In fact, starting from ATOMADD64_DAG all opcodes will be
462 // thought as target memory ops!
466 /// Define some predicates that are used for node matching.
468 /// isVEXTRACT128Index - Return true if the specified
469 /// EXTRACT_SUBVECTOR operand specifies a vector extract that is
470 /// suitable for input to VEXTRACTF128, VEXTRACTI128 instructions.
471 bool isVEXTRACT128Index(SDNode *N);
473 /// isVINSERT128Index - Return true if the specified
474 /// INSERT_SUBVECTOR operand specifies a subvector insert that is
475 /// suitable for input to VINSERTF128, VINSERTI128 instructions.
476 bool isVINSERT128Index(SDNode *N);
478 /// isVEXTRACT256Index - Return true if the specified
479 /// EXTRACT_SUBVECTOR operand specifies a vector extract that is
480 /// suitable for input to VEXTRACTF64X4, VEXTRACTI64X4 instructions.
481 bool isVEXTRACT256Index(SDNode *N);
483 /// isVINSERT256Index - Return true if the specified
484 /// INSERT_SUBVECTOR operand specifies a subvector insert that is
485 /// suitable for input to VINSERTF64X4, VINSERTI64X4 instructions.
486 bool isVINSERT256Index(SDNode *N);
488 /// getExtractVEXTRACT128Immediate - Return the appropriate
489 /// immediate to extract the specified EXTRACT_SUBVECTOR index
490 /// with VEXTRACTF128, VEXTRACTI128 instructions.
491 unsigned getExtractVEXTRACT128Immediate(SDNode *N);
493 /// getInsertVINSERT128Immediate - Return the appropriate
494 /// immediate to insert at the specified INSERT_SUBVECTOR index
495 /// with VINSERTF128, VINSERT128 instructions.
496 unsigned getInsertVINSERT128Immediate(SDNode *N);
498 /// getExtractVEXTRACT256Immediate - Return the appropriate
499 /// immediate to extract the specified EXTRACT_SUBVECTOR index
500 /// with VEXTRACTF64X4, VEXTRACTI64x4 instructions.
501 unsigned getExtractVEXTRACT256Immediate(SDNode *N);
503 /// getInsertVINSERT256Immediate - Return the appropriate
504 /// immediate to insert at the specified INSERT_SUBVECTOR index
505 /// with VINSERTF64x4, VINSERTI64x4 instructions.
506 unsigned getInsertVINSERT256Immediate(SDNode *N);
508 /// isZeroNode - Returns true if Elt is a constant zero or a floating point
510 bool isZeroNode(SDValue Elt);
512 /// isOffsetSuitableForCodeModel - Returns true of the given offset can be
513 /// fit into displacement field of the instruction.
514 bool isOffsetSuitableForCodeModel(int64_t Offset, CodeModel::Model M,
515 bool hasSymbolicDisplacement = true);
518 /// isCalleePop - Determines whether the callee is required to pop its
519 /// own arguments. Callee pop is necessary to support tail calls.
520 bool isCalleePop(CallingConv::ID CallingConv,
521 bool is64Bit, bool IsVarArg, bool TailCallOpt);
524 //===--------------------------------------------------------------------===//
525 // X86TargetLowering - X86 Implementation of the TargetLowering interface
526 class X86TargetLowering final : public TargetLowering {
528 explicit X86TargetLowering(X86TargetMachine &TM);
530 unsigned getJumpTableEncoding() const override;
532 MVT getScalarShiftAmountTy(EVT LHSTy) const override { return MVT::i8; }
535 LowerCustomJumpTableEntry(const MachineJumpTableInfo *MJTI,
536 const MachineBasicBlock *MBB, unsigned uid,
537 MCContext &Ctx) const override;
539 /// getPICJumpTableRelocaBase - Returns relocation base for the given PIC
541 SDValue getPICJumpTableRelocBase(SDValue Table,
542 SelectionDAG &DAG) const override;
544 getPICJumpTableRelocBaseExpr(const MachineFunction *MF,
545 unsigned JTI, MCContext &Ctx) const override;
547 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
548 /// function arguments in the caller parameter area. For X86, aggregates
549 /// that contains are placed at 16-byte boundaries while the rest are at
550 /// 4-byte boundaries.
551 unsigned getByValTypeAlignment(Type *Ty) const override;
553 /// getOptimalMemOpType - Returns the target specific optimal type for load
554 /// and store operations as a result of memset, memcpy, and memmove
555 /// lowering. If DstAlign is zero that means it's safe to destination
556 /// alignment can satisfy any constraint. Similarly if SrcAlign is zero it
557 /// means there isn't a need to check it against alignment requirement,
558 /// probably because the source does not need to be loaded. If 'IsMemset' is
559 /// true, that means it's expanding a memset. If 'ZeroMemset' is true, that
560 /// means it's a memset of zero. 'MemcpyStrSrc' indicates whether the memcpy
561 /// source is constant so it does not need to be loaded.
562 /// It returns EVT::Other if the type should be determined using generic
563 /// target-independent logic.
564 EVT getOptimalMemOpType(uint64_t Size, unsigned DstAlign, unsigned SrcAlign,
565 bool IsMemset, bool ZeroMemset, bool MemcpyStrSrc,
566 MachineFunction &MF) const override;
568 /// isSafeMemOpType - Returns true if it's safe to use load / store of the
569 /// specified type to expand memcpy / memset inline. This is mostly true
570 /// for all types except for some special cases. For example, on X86
571 /// targets without SSE2 f64 load / store are done with fldl / fstpl which
572 /// also does type conversion. Note the specified type doesn't have to be
573 /// legal as the hook is used before type legalization.
574 bool isSafeMemOpType(MVT VT) const override;
576 /// allowsUnalignedMemoryAccesses - Returns true if the target allows
577 /// unaligned memory accesses. of the specified type. Returns whether it
578 /// is "fast" by reference in the second argument.
579 bool allowsUnalignedMemoryAccesses(EVT VT, unsigned AS,
580 bool *Fast) const override;
582 /// LowerOperation - Provide custom lowering hooks for some operations.
584 SDValue LowerOperation(SDValue Op, SelectionDAG &DAG) const override;
586 /// ReplaceNodeResults - Replace the results of node with an illegal result
587 /// type with new values built out of custom code.
589 void ReplaceNodeResults(SDNode *N, SmallVectorImpl<SDValue>&Results,
590 SelectionDAG &DAG) const override;
593 SDValue PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const override;
595 /// isTypeDesirableForOp - Return true if the target has native support for
596 /// the specified value type and it is 'desirable' to use the type for the
597 /// given node type. e.g. On x86 i16 is legal, but undesirable since i16
598 /// instruction encodings are longer and some i16 instructions are slow.
599 bool isTypeDesirableForOp(unsigned Opc, EVT VT) const override;
601 /// isTypeDesirable - Return true if the target has native support for the
602 /// specified value type and it is 'desirable' to use the type. e.g. On x86
603 /// i16 is legal, but undesirable since i16 instruction encodings are longer
604 /// and some i16 instructions are slow.
605 bool IsDesirableToPromoteOp(SDValue Op, EVT &PVT) const override;
608 EmitInstrWithCustomInserter(MachineInstr *MI,
609 MachineBasicBlock *MBB) const override;
612 /// getTargetNodeName - This method returns the name of a target specific
614 const char *getTargetNodeName(unsigned Opcode) const override;
616 /// getSetCCResultType - Return the value type to use for ISD::SETCC.
617 EVT getSetCCResultType(LLVMContext &Context, EVT VT) const override;
619 /// computeMaskedBitsForTargetNode - Determine which of the bits specified
620 /// in Mask are known to be either zero or one and return them in the
621 /// KnownZero/KnownOne bitsets.
622 void computeMaskedBitsForTargetNode(const SDValue Op,
625 const SelectionDAG &DAG,
626 unsigned Depth = 0) const override;
628 // ComputeNumSignBitsForTargetNode - Determine the number of bits in the
629 // operation that are sign bits.
630 unsigned ComputeNumSignBitsForTargetNode(SDValue Op,
631 const SelectionDAG &DAG,
632 unsigned Depth) const override;
634 bool isGAPlusOffset(SDNode *N, const GlobalValue* &GA,
635 int64_t &Offset) const override;
637 SDValue getReturnAddressFrameIndex(SelectionDAG &DAG) const;
639 bool ExpandInlineAsm(CallInst *CI) const override;
642 getConstraintType(const std::string &Constraint) const override;
644 /// Examine constraint string and operand type and determine a weight value.
645 /// The operand object must already have been set up with the operand type.
647 getSingleConstraintMatchWeight(AsmOperandInfo &info,
648 const char *constraint) const override;
650 const char *LowerXConstraint(EVT ConstraintVT) const override;
652 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
653 /// vector. If it is invalid, don't add anything to Ops. If hasMemory is
654 /// true it means one of the asm constraint of the inline asm instruction
655 /// being processed is 'm'.
656 void LowerAsmOperandForConstraint(SDValue Op,
657 std::string &Constraint,
658 std::vector<SDValue> &Ops,
659 SelectionDAG &DAG) const override;
661 /// getRegForInlineAsmConstraint - Given a physical register constraint
662 /// (e.g. {edx}), return the register number and the register class for the
663 /// register. This should only be used for C_Register constraints. On
664 /// error, this returns a register number of 0.
665 std::pair<unsigned, const TargetRegisterClass*>
666 getRegForInlineAsmConstraint(const std::string &Constraint,
667 MVT VT) const override;
669 /// isLegalAddressingMode - Return true if the addressing mode represented
670 /// by AM is legal for this target, for a load/store of the specified type.
671 bool isLegalAddressingMode(const AddrMode &AM, Type *Ty) const override;
673 /// isLegalICmpImmediate - Return true if the specified immediate is legal
674 /// icmp immediate, that is the target has icmp instructions which can
675 /// compare a register against the immediate without having to materialize
676 /// the immediate into a register.
677 bool isLegalICmpImmediate(int64_t Imm) const override;
679 /// isLegalAddImmediate - Return true if the specified immediate is legal
680 /// add immediate, that is the target has add instructions which can
681 /// add a register and the immediate without having to materialize
682 /// the immediate into a register.
683 bool isLegalAddImmediate(int64_t Imm) const override;
685 /// \brief Return the cost of the scaling factor used in the addressing
686 /// mode represented by AM for this target, for a load/store
687 /// of the specified type.
688 /// If the AM is supported, the return value must be >= 0.
689 /// If the AM is not supported, it returns a negative value.
690 int getScalingFactorCost(const AddrMode &AM, Type *Ty) const override;
692 bool isVectorShiftByScalarCheap(Type *Ty) const override;
694 /// isTruncateFree - Return true if it's free to truncate a value of
695 /// type Ty1 to type Ty2. e.g. On x86 it's free to truncate a i32 value in
696 /// register EAX to i16 by referencing its sub-register AX.
697 bool isTruncateFree(Type *Ty1, Type *Ty2) const override;
698 bool isTruncateFree(EVT VT1, EVT VT2) const override;
700 bool allowTruncateForTailCall(Type *Ty1, Type *Ty2) const override;
702 /// isZExtFree - Return true if any actual instruction that defines a
703 /// value of type Ty1 implicit zero-extends the value to Ty2 in the result
704 /// register. This does not necessarily include registers defined in
705 /// unknown ways, such as incoming arguments, or copies from unknown
706 /// virtual registers. Also, if isTruncateFree(Ty2, Ty1) is true, this
707 /// does not necessarily apply to truncate instructions. e.g. on x86-64,
708 /// all instructions that define 32-bit values implicit zero-extend the
709 /// result out to 64 bits.
710 bool isZExtFree(Type *Ty1, Type *Ty2) const override;
711 bool isZExtFree(EVT VT1, EVT VT2) const override;
712 bool isZExtFree(SDValue Val, EVT VT2) const override;
714 /// isFMAFasterThanFMulAndFAdd - Return true if an FMA operation is faster
715 /// than a pair of fmul and fadd instructions. fmuladd intrinsics will be
716 /// expanded to FMAs when this method returns true, otherwise fmuladd is
717 /// expanded to fmul + fadd.
718 bool isFMAFasterThanFMulAndFAdd(EVT VT) const override;
720 /// isNarrowingProfitable - Return true if it's profitable to narrow
721 /// operations of type VT1 to VT2. e.g. on x86, it's profitable to narrow
722 /// from i32 to i8 but not from i32 to i16.
723 bool isNarrowingProfitable(EVT VT1, EVT VT2) const override;
725 /// isFPImmLegal - Returns true if the target can instruction select the
726 /// specified FP immediate natively. If false, the legalizer will
727 /// materialize the FP immediate as a load from a constant pool.
728 bool isFPImmLegal(const APFloat &Imm, EVT VT) const override;
730 /// isShuffleMaskLegal - Targets can use this to indicate that they only
731 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
732 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask
733 /// values are assumed to be legal.
734 bool isShuffleMaskLegal(const SmallVectorImpl<int> &Mask,
735 EVT VT) const override;
737 /// isVectorClearMaskLegal - Similar to isShuffleMaskLegal. This is
738 /// used by Targets can use this to indicate if there is a suitable
739 /// VECTOR_SHUFFLE that can be used to replace a VAND with a constant
741 bool isVectorClearMaskLegal(const SmallVectorImpl<int> &Mask,
742 EVT VT) const override;
744 /// ShouldShrinkFPConstant - If true, then instruction selection should
745 /// seek to shrink the FP constant of the specified type to a smaller type
746 /// in order to save space and / or reduce runtime.
747 bool ShouldShrinkFPConstant(EVT VT) const override {
748 // Don't shrink FP constpool if SSE2 is available since cvtss2sd is more
749 // expensive than a straight movsd. On the other hand, it's important to
750 // shrink long double fp constant since fldt is very slow.
751 return !X86ScalarSSEf64 || VT == MVT::f80;
754 const X86Subtarget* getSubtarget() const {
758 /// isScalarFPTypeInSSEReg - Return true if the specified scalar FP type is
759 /// computed in an SSE register, not on the X87 floating point stack.
760 bool isScalarFPTypeInSSEReg(EVT VT) const {
761 return (VT == MVT::f64 && X86ScalarSSEf64) || // f64 is when SSE2
762 (VT == MVT::f32 && X86ScalarSSEf32); // f32 is when SSE1
765 /// isTargetFTOL - Return true if the target uses the MSVC _ftol2 routine
767 bool isTargetFTOL() const {
768 return Subtarget->isTargetKnownWindowsMSVC() && !Subtarget->is64Bit();
771 /// isIntegerTypeFTOL - Return true if the MSVC _ftol2 routine should be
772 /// used for fptoui to the given type.
773 bool isIntegerTypeFTOL(EVT VT) const {
774 return isTargetFTOL() && VT == MVT::i64;
777 /// \brief Returns true if it is beneficial to convert a load of a constant
778 /// to just the constant itself.
779 bool shouldConvertConstantLoadToIntImm(const APInt &Imm,
780 Type *Ty) const override;
782 /// Intel processors have a unified instruction and data cache
783 const char * getClearCacheBuiltinName() const {
784 return 0; // nothing to do, move along.
787 /// createFastISel - This method returns a target specific FastISel object,
788 /// or null if the target does not support "fast" ISel.
789 FastISel *createFastISel(FunctionLoweringInfo &funcInfo,
790 const TargetLibraryInfo *libInfo) const override;
792 /// getStackCookieLocation - Return true if the target stores stack
793 /// protector cookies at a fixed offset in some non-standard address
794 /// space, and populates the address space and offset as
796 bool getStackCookieLocation(unsigned &AddressSpace,
797 unsigned &Offset) const override;
799 SDValue BuildFILD(SDValue Op, EVT SrcVT, SDValue Chain, SDValue StackSlot,
800 SelectionDAG &DAG) const;
802 bool isNoopAddrSpaceCast(unsigned SrcAS, unsigned DestAS) const override;
804 /// \brief Reset the operation actions based on target options.
805 void resetOperationActions() override;
808 std::pair<const TargetRegisterClass*, uint8_t>
809 findRepresentativeClass(MVT VT) const override;
812 /// Subtarget - Keep a pointer to the X86Subtarget around so that we can
813 /// make the right decision when generating code for different targets.
814 const X86Subtarget *Subtarget;
815 const DataLayout *TD;
817 /// Used to store the TargetOptions so that we don't waste time resetting
818 /// the operation actions unless we have to.
821 /// X86ScalarSSEf32, X86ScalarSSEf64 - Select between SSE or x87
822 /// floating point ops.
823 /// When SSE is available, use it for f32 operations.
824 /// When SSE2 is available, use it for f64 operations.
825 bool X86ScalarSSEf32;
826 bool X86ScalarSSEf64;
828 /// LegalFPImmediates - A list of legal fp immediates.
829 std::vector<APFloat> LegalFPImmediates;
831 /// addLegalFPImmediate - Indicate that this x86 target can instruction
832 /// select the specified FP immediate natively.
833 void addLegalFPImmediate(const APFloat& Imm) {
834 LegalFPImmediates.push_back(Imm);
837 SDValue LowerCallResult(SDValue Chain, SDValue InFlag,
838 CallingConv::ID CallConv, bool isVarArg,
839 const SmallVectorImpl<ISD::InputArg> &Ins,
840 SDLoc dl, SelectionDAG &DAG,
841 SmallVectorImpl<SDValue> &InVals) const;
842 SDValue LowerMemArgument(SDValue Chain,
843 CallingConv::ID CallConv,
844 const SmallVectorImpl<ISD::InputArg> &ArgInfo,
845 SDLoc dl, SelectionDAG &DAG,
846 const CCValAssign &VA, MachineFrameInfo *MFI,
848 SDValue LowerMemOpCallTo(SDValue Chain, SDValue StackPtr, SDValue Arg,
849 SDLoc dl, SelectionDAG &DAG,
850 const CCValAssign &VA,
851 ISD::ArgFlagsTy Flags) const;
853 // Call lowering helpers.
855 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
856 /// for tail call optimization. Targets which want to do tail call
857 /// optimization should implement this function.
858 bool IsEligibleForTailCallOptimization(SDValue Callee,
859 CallingConv::ID CalleeCC,
861 bool isCalleeStructRet,
862 bool isCallerStructRet,
864 const SmallVectorImpl<ISD::OutputArg> &Outs,
865 const SmallVectorImpl<SDValue> &OutVals,
866 const SmallVectorImpl<ISD::InputArg> &Ins,
867 SelectionDAG& DAG) const;
868 bool IsCalleePop(bool isVarArg, CallingConv::ID CallConv) const;
869 SDValue EmitTailCallLoadRetAddr(SelectionDAG &DAG, SDValue &OutRetAddr,
870 SDValue Chain, bool IsTailCall, bool Is64Bit,
871 int FPDiff, SDLoc dl) const;
873 unsigned GetAlignedArgumentStackSize(unsigned StackSize,
874 SelectionDAG &DAG) const;
876 std::pair<SDValue,SDValue> FP_TO_INTHelper(SDValue Op, SelectionDAG &DAG,
878 bool isReplace) const;
880 SDValue LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) const;
881 SDValue LowerBUILD_VECTORvXi1(SDValue Op, SelectionDAG &DAG) const;
882 SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) const;
883 SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const;
884 SDValue ExtractBitFromMaskVector(SDValue Op, SelectionDAG &DAG) const;
885 SDValue InsertBitToMaskVector(SDValue Op, SelectionDAG &DAG) const;
887 SDValue LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const;
888 SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) const;
889 SDValue LowerBlockAddress(SDValue Op, SelectionDAG &DAG) const;
890 SDValue LowerGlobalAddress(const GlobalValue *GV, SDLoc dl,
891 int64_t Offset, SelectionDAG &DAG) const;
892 SDValue LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) const;
893 SDValue LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const;
894 SDValue LowerExternalSymbol(SDValue Op, SelectionDAG &DAG) const;
895 SDValue LowerSINT_TO_FP(SDValue Op, SelectionDAG &DAG) const;
896 SDValue LowerUINT_TO_FP(SDValue Op, SelectionDAG &DAG) const;
897 SDValue LowerUINT_TO_FP_i64(SDValue Op, SelectionDAG &DAG) const;
898 SDValue LowerUINT_TO_FP_i32(SDValue Op, SelectionDAG &DAG) const;
899 SDValue lowerUINT_TO_FP_vec(SDValue Op, SelectionDAG &DAG) const;
900 SDValue LowerTRUNCATE(SDValue Op, SelectionDAG &DAG) const;
901 SDValue LowerFP_TO_SINT(SDValue Op, SelectionDAG &DAG) const;
902 SDValue LowerFP_TO_UINT(SDValue Op, SelectionDAG &DAG) const;
903 SDValue LowerToBT(SDValue And, ISD::CondCode CC,
904 SDLoc dl, SelectionDAG &DAG) const;
905 SDValue LowerSETCC(SDValue Op, SelectionDAG &DAG) const;
906 SDValue LowerSELECT(SDValue Op, SelectionDAG &DAG) const;
907 SDValue LowerBRCOND(SDValue Op, SelectionDAG &DAG) const;
908 SDValue LowerMEMSET(SDValue Op, SelectionDAG &DAG) const;
909 SDValue LowerJumpTable(SDValue Op, SelectionDAG &DAG) const;
910 SDValue LowerDYNAMIC_STACKALLOC(SDValue Op, SelectionDAG &DAG) const;
911 SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG) const;
912 SDValue LowerVAARG(SDValue Op, SelectionDAG &DAG) const;
913 SDValue LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const;
914 SDValue LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const;
915 SDValue LowerFRAME_TO_ARGS_OFFSET(SDValue Op, SelectionDAG &DAG) const;
916 SDValue LowerEH_RETURN(SDValue Op, SelectionDAG &DAG) const;
917 SDValue lowerEH_SJLJ_SETJMP(SDValue Op, SelectionDAG &DAG) const;
918 SDValue lowerEH_SJLJ_LONGJMP(SDValue Op, SelectionDAG &DAG) const;
919 SDValue LowerINIT_TRAMPOLINE(SDValue Op, SelectionDAG &DAG) const;
920 SDValue LowerFLT_ROUNDS_(SDValue Op, SelectionDAG &DAG) const;
921 SDValue LowerSIGN_EXTEND_INREG(SDValue Op, SelectionDAG &DAG) const;
924 LowerFormalArguments(SDValue Chain,
925 CallingConv::ID CallConv, bool isVarArg,
926 const SmallVectorImpl<ISD::InputArg> &Ins,
927 SDLoc dl, SelectionDAG &DAG,
928 SmallVectorImpl<SDValue> &InVals) const override;
929 SDValue LowerCall(CallLoweringInfo &CLI,
930 SmallVectorImpl<SDValue> &InVals) const override;
932 SDValue LowerReturn(SDValue Chain,
933 CallingConv::ID CallConv, bool isVarArg,
934 const SmallVectorImpl<ISD::OutputArg> &Outs,
935 const SmallVectorImpl<SDValue> &OutVals,
936 SDLoc dl, SelectionDAG &DAG) const override;
938 bool isUsedByReturnOnly(SDNode *N, SDValue &Chain) const override;
940 bool mayBeEmittedAsTailCall(CallInst *CI) const override;
942 MVT getTypeForExtArgOrReturn(MVT VT,
943 ISD::NodeType ExtendKind) const override;
945 bool CanLowerReturn(CallingConv::ID CallConv, MachineFunction &MF,
947 const SmallVectorImpl<ISD::OutputArg> &Outs,
948 LLVMContext &Context) const override;
950 const MCPhysReg *getScratchRegisters(CallingConv::ID CC) const override;
952 /// Utility function to emit atomic-load-arith operations (and, or, xor,
953 /// nand, max, min, umax, umin). It takes the corresponding instruction to
954 /// expand, the associated machine basic block, and the associated X86
955 /// opcodes for reg/reg.
956 MachineBasicBlock *EmitAtomicLoadArith(MachineInstr *MI,
957 MachineBasicBlock *MBB) const;
959 /// Utility function to emit atomic-load-arith operations (and, or, xor,
960 /// nand, add, sub, swap) for 64-bit operands on 32-bit target.
961 MachineBasicBlock *EmitAtomicLoadArith6432(MachineInstr *MI,
962 MachineBasicBlock *MBB) const;
964 // Utility function to emit the low-level va_arg code for X86-64.
965 MachineBasicBlock *EmitVAARG64WithCustomInserter(
967 MachineBasicBlock *MBB) const;
969 /// Utility function to emit the xmm reg save portion of va_start.
970 MachineBasicBlock *EmitVAStartSaveXMMRegsWithCustomInserter(
971 MachineInstr *BInstr,
972 MachineBasicBlock *BB) const;
974 MachineBasicBlock *EmitLoweredSelect(MachineInstr *I,
975 MachineBasicBlock *BB) const;
977 MachineBasicBlock *EmitLoweredWinAlloca(MachineInstr *MI,
978 MachineBasicBlock *BB) const;
980 MachineBasicBlock *EmitLoweredSegAlloca(MachineInstr *MI,
981 MachineBasicBlock *BB,
984 MachineBasicBlock *EmitLoweredTLSCall(MachineInstr *MI,
985 MachineBasicBlock *BB) const;
987 MachineBasicBlock *emitLoweredTLSAddr(MachineInstr *MI,
988 MachineBasicBlock *BB) const;
990 MachineBasicBlock *emitEHSjLjSetJmp(MachineInstr *MI,
991 MachineBasicBlock *MBB) const;
993 MachineBasicBlock *emitEHSjLjLongJmp(MachineInstr *MI,
994 MachineBasicBlock *MBB) const;
996 MachineBasicBlock *emitFMA3Instr(MachineInstr *MI,
997 MachineBasicBlock *MBB) const;
999 /// Emit nodes that will be selected as "test Op0,Op0", or something
1000 /// equivalent, for use with the given x86 condition code.
1001 SDValue EmitTest(SDValue Op0, unsigned X86CC, SDLoc dl,
1002 SelectionDAG &DAG) const;
1004 /// Emit nodes that will be selected as "cmp Op0,Op1", or something
1005 /// equivalent, for use with the given x86 condition code.
1006 SDValue EmitCmp(SDValue Op0, SDValue Op1, unsigned X86CC, SDLoc dl,
1007 SelectionDAG &DAG) const;
1009 /// Convert a comparison if required by the subtarget.
1010 SDValue ConvertCmpIfNecessary(SDValue Cmp, SelectionDAG &DAG) const;
1014 FastISel *createFastISel(FunctionLoweringInfo &funcInfo,
1015 const TargetLibraryInfo *libInfo);
1019 #endif // X86ISELLOWERING_H