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 LLVM_LIB_TARGET_X86_X86ISELLOWERING_H
16 #define LLVM_LIB_TARGET_X86_X86ISELLOWERING_H
18 #include "llvm/CodeGen/CallingConvLower.h"
19 #include "llvm/CodeGen/SelectionDAG.h"
20 #include "llvm/Target/TargetLowering.h"
21 #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 Read Performance Monitoring Counters.
92 /// X86 compare and logical compare instructions.
95 /// X86 bit-test instructions.
98 /// X86 SetCC. Operand 0 is condition code, and operand 1 is the EFLAGS
99 /// operand, usually produced by a CMP instruction.
105 // Same as SETCC except it's materialized with a sbb and the value is all
106 // one's or all zero's.
107 SETCC_CARRY, // R = carry_bit ? ~0 : 0
109 /// X86 FP SETCC, implemented with CMP{cc}SS/CMP{cc}SD.
110 /// Operands are two FP values to compare; result is a mask of
111 /// 0s or 1s. Generally DTRT for C/C++ with NaNs.
114 /// X86 MOVMSK{pd|ps}, extracts sign bits of two or four FP values,
115 /// result in an integer GPR. Needs masking for scalar result.
118 /// X86 conditional moves. Operand 0 and operand 1 are the two values
119 /// to select from. Operand 2 is the condition code, and operand 3 is the
120 /// flag operand produced by a CMP or TEST instruction. It also writes a
124 /// X86 conditional branches. Operand 0 is the chain operand, operand 1
125 /// is the block to branch if condition is true, operand 2 is the
126 /// condition code, and operand 3 is the flag operand produced by a CMP
127 /// or TEST instruction.
130 /// Return with a flag operand. Operand 0 is the chain operand, operand
131 /// 1 is the number of bytes of stack to pop.
134 /// REP_STOS - Repeat fill, corresponds to X86::REP_STOSx.
137 /// REP_MOVS - Repeat move, corresponds to X86::REP_MOVSx.
140 /// GlobalBaseReg - On Darwin, this node represents the result of the popl
141 /// at function entry, used for PIC code.
144 /// Wrapper - A wrapper node for TargetConstantPool,
145 /// TargetExternalSymbol, and TargetGlobalAddress.
148 /// WrapperRIP - Special wrapper used under X86-64 PIC mode for RIP
149 /// relative displacements.
152 /// MOVDQ2Q - Copies a 64-bit value from the low word of an XMM vector
153 /// to an MMX vector. If you think this is too close to the previous
154 /// mnemonic, so do I; blame Intel.
157 /// MMX_MOVD2W - Copies a 32-bit value from the low word of a MMX
161 /// PEXTRB - Extract an 8-bit value from a vector and zero extend it to
162 /// i32, corresponds to X86::PEXTRB.
165 /// PEXTRW - Extract a 16-bit value from a vector and zero extend it to
166 /// i32, corresponds to X86::PEXTRW.
169 /// INSERTPS - Insert any element of a 4 x float vector into any element
170 /// of a destination 4 x floatvector.
173 /// PINSRB - Insert the lower 8-bits of a 32-bit value to a vector,
174 /// corresponds to X86::PINSRB.
177 /// PINSRW - Insert the lower 16-bits of a 32-bit value to a vector,
178 /// corresponds to X86::PINSRW.
181 /// PSHUFB - Shuffle 16 8-bit values within a vector.
184 /// ANDNP - Bitwise Logical AND NOT of Packed FP values.
187 /// PSIGN - Copy integer sign.
190 /// BLENDV - Blend where the selector is a register.
193 /// BLENDI - Blend where the selector is an immediate.
196 /// ADDSUB - Combined add and sub on an FP vector.
199 // SUBUS - Integer sub with unsigned saturation.
202 /// HADD - Integer horizontal add.
205 /// HSUB - Integer horizontal sub.
208 /// FHADD - Floating point horizontal add.
211 /// FHSUB - Floating point horizontal sub.
214 /// UMAX, UMIN - Unsigned integer max and min.
217 /// SMAX, SMIN - Signed integer max and min.
220 /// FMAX, FMIN - Floating point max and min.
224 /// FMAXC, FMINC - Commutative FMIN and FMAX.
227 /// FRSQRT, FRCP - Floating point reciprocal-sqrt and reciprocal
228 /// approximation. Note that these typically require refinement
229 /// in order to obtain suitable precision.
232 // TLSADDR - Thread Local Storage.
235 // TLSBASEADDR - Thread Local Storage. A call to get the start address
236 // of the TLS block for the current module.
239 // TLSCALL - Thread Local Storage. When calling to an OS provided
240 // thunk at the address from an earlier relocation.
243 // EH_RETURN - Exception Handling helpers.
246 // EH_SJLJ_SETJMP - SjLj exception handling setjmp.
249 // EH_SJLJ_LONGJMP - SjLj exception handling longjmp.
252 /// TC_RETURN - Tail call return. See X86TargetLowering::LowerCall for
253 /// the list of operands.
256 // VZEXT_MOVL - Vector move to low scalar and zero higher vector elements.
259 // VZEXT - Vector integer zero-extend.
262 // VSEXT - Vector integer signed-extend.
265 // VTRUNC - Vector integer truncate.
268 // VTRUNC - Vector integer truncate with mask.
271 // VFPEXT - Vector FP extend.
274 // VFPROUND - Vector FP round.
277 // VSHL, VSRL - 128-bit vector logical left / right shift
280 // VSHL, VSRL, VSRA - Vector shift elements
283 // VSHLI, VSRLI, VSRAI - Vector shift elements by immediate
286 // CMPP - Vector packed double/float comparison.
289 // PCMP* - Vector integer comparisons.
291 // PCMP*M - Vector integer comparisons, the result is in a mask vector.
294 /// CMPM, CMPMU - Vector comparison generating mask bits for fp and
295 /// integer signed and unsigned data types.
299 // ADD, SUB, SMUL, etc. - Arithmetic operations with FLAGS results.
300 ADD, SUB, ADC, SBB, SMUL,
301 INC, DEC, OR, XOR, AND,
303 BEXTR, // BEXTR - Bit field extract
305 UMUL, // LOW, HI, FLAGS = umul LHS, RHS
307 // MUL_IMM - X86 specific multiply by immediate.
310 // PTEST - Vector bitwise comparisons.
313 // TESTP - Vector packed fp sign bitwise comparisons.
316 // TESTM, TESTNM - Vector "test" in AVX-512, the result is in a mask vector.
320 // OR/AND test for masks
323 // Several flavors of instructions with vector shuffle behaviors.
328 // AVX512 inter-lane alignr
356 // Insert/Extract vector element
360 // PMULUDQ - Vector multiply packed unsigned doubleword integers
362 // PMULUDQ - Vector multiply packed signed doubleword integers
373 // VASTART_SAVE_XMM_REGS - Save xmm argument registers to the stack,
374 // according to %al. An operator is needed so that this can be expanded
375 // with control flow.
376 VASTART_SAVE_XMM_REGS,
378 // WIN_ALLOCA - Windows's _chkstk call to do stack probing.
381 // SEG_ALLOCA - For allocating variable amounts of stack space when using
382 // segmented stacks. Check if the current stacklet has enough space, and
383 // falls back to heap allocation if not.
386 // WIN_FTOL - Windows's _ftol2 runtime routine to do fptoui.
395 // FNSTSW16r - Store FP status word into i16 register.
398 // SAHF - Store contents of %ah into %eflags.
401 // RDRAND - Get a random integer and indicate whether it is valid in CF.
404 // RDSEED - Get a NIST SP800-90B & C compliant random integer and
405 // indicate whether it is valid in CF.
412 // XTEST - Test if in transactional execution.
415 // LCMPXCHG_DAG, LCMPXCHG8_DAG, LCMPXCHG16_DAG - Compare and swap.
416 LCMPXCHG_DAG = ISD::FIRST_TARGET_MEMORY_OPCODE,
420 // VZEXT_LOAD - Load, scalar_to_vector, and zero extend.
423 // FNSTCW16m - Store FP control world into i16 memory.
426 /// FP_TO_INT*_IN_MEM - This instruction implements FP_TO_SINT with the
427 /// integer destination in memory and a FP reg source. This corresponds
428 /// to the X86::FIST*m instructions and the rounding mode change stuff. It
429 /// has two inputs (token chain and address) and two outputs (int value
430 /// and token chain).
435 /// FILD, FILD_FLAG - This instruction implements SINT_TO_FP with the
436 /// integer source in memory and FP reg result. This corresponds to the
437 /// X86::FILD*m instructions. It has three inputs (token chain, address,
438 /// and source type) and two outputs (FP value and token chain). FILD_FLAG
439 /// also produces a flag).
443 /// FLD - This instruction implements an extending load to FP stack slots.
444 /// This corresponds to the X86::FLD32m / X86::FLD64m. It takes a chain
445 /// operand, ptr to load from, and a ValueType node indicating the type
449 /// FST - This instruction implements a truncating store to FP stack
450 /// slots. This corresponds to the X86::FST32m / X86::FST64m. It takes a
451 /// chain operand, value to store, address, and a ValueType to store it
455 /// VAARG_64 - This instruction grabs the address of the next argument
456 /// from a va_list. (reads and modifies the va_list in memory)
459 // WARNING: Do not add anything in the end unless you want the node to
460 // have memop! In fact, starting from ATOMADD64_DAG all opcodes will be
461 // thought as target memory ops!
465 /// Define some predicates that are used for node matching.
467 /// isVEXTRACT128Index - Return true if the specified
468 /// EXTRACT_SUBVECTOR operand specifies a vector extract that is
469 /// suitable for input to VEXTRACTF128, VEXTRACTI128 instructions.
470 bool isVEXTRACT128Index(SDNode *N);
472 /// isVINSERT128Index - Return true if the specified
473 /// INSERT_SUBVECTOR operand specifies a subvector insert that is
474 /// suitable for input to VINSERTF128, VINSERTI128 instructions.
475 bool isVINSERT128Index(SDNode *N);
477 /// isVEXTRACT256Index - Return true if the specified
478 /// EXTRACT_SUBVECTOR operand specifies a vector extract that is
479 /// suitable for input to VEXTRACTF64X4, VEXTRACTI64X4 instructions.
480 bool isVEXTRACT256Index(SDNode *N);
482 /// isVINSERT256Index - Return true if the specified
483 /// INSERT_SUBVECTOR operand specifies a subvector insert that is
484 /// suitable for input to VINSERTF64X4, VINSERTI64X4 instructions.
485 bool isVINSERT256Index(SDNode *N);
487 /// getExtractVEXTRACT128Immediate - Return the appropriate
488 /// immediate to extract the specified EXTRACT_SUBVECTOR index
489 /// with VEXTRACTF128, VEXTRACTI128 instructions.
490 unsigned getExtractVEXTRACT128Immediate(SDNode *N);
492 /// getInsertVINSERT128Immediate - Return the appropriate
493 /// immediate to insert at the specified INSERT_SUBVECTOR index
494 /// with VINSERTF128, VINSERT128 instructions.
495 unsigned getInsertVINSERT128Immediate(SDNode *N);
497 /// getExtractVEXTRACT256Immediate - Return the appropriate
498 /// immediate to extract the specified EXTRACT_SUBVECTOR index
499 /// with VEXTRACTF64X4, VEXTRACTI64x4 instructions.
500 unsigned getExtractVEXTRACT256Immediate(SDNode *N);
502 /// getInsertVINSERT256Immediate - Return the appropriate
503 /// immediate to insert at the specified INSERT_SUBVECTOR index
504 /// with VINSERTF64x4, VINSERTI64x4 instructions.
505 unsigned getInsertVINSERT256Immediate(SDNode *N);
507 /// isZeroNode - Returns true if Elt is a constant zero or a floating point
509 bool isZeroNode(SDValue Elt);
511 /// isOffsetSuitableForCodeModel - Returns true of the given offset can be
512 /// fit into displacement field of the instruction.
513 bool isOffsetSuitableForCodeModel(int64_t Offset, CodeModel::Model M,
514 bool hasSymbolicDisplacement = true);
517 /// isCalleePop - Determines whether the callee is required to pop its
518 /// own arguments. Callee pop is necessary to support tail calls.
519 bool isCalleePop(CallingConv::ID CallingConv,
520 bool is64Bit, bool IsVarArg, bool TailCallOpt);
522 /// AVX512 static rounding constants. These need to match the values in
524 enum STATIC_ROUNDING {
533 //===--------------------------------------------------------------------===//
534 // X86TargetLowering - X86 Implementation of the TargetLowering interface
535 class X86TargetLowering final : public TargetLowering {
537 explicit X86TargetLowering(X86TargetMachine &TM);
539 unsigned getJumpTableEncoding() const override;
541 MVT getScalarShiftAmountTy(EVT LHSTy) const override { return MVT::i8; }
544 LowerCustomJumpTableEntry(const MachineJumpTableInfo *MJTI,
545 const MachineBasicBlock *MBB, unsigned uid,
546 MCContext &Ctx) const override;
548 /// getPICJumpTableRelocaBase - Returns relocation base for the given PIC
550 SDValue getPICJumpTableRelocBase(SDValue Table,
551 SelectionDAG &DAG) const override;
553 getPICJumpTableRelocBaseExpr(const MachineFunction *MF,
554 unsigned JTI, MCContext &Ctx) const override;
556 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
557 /// function arguments in the caller parameter area. For X86, aggregates
558 /// that contains are placed at 16-byte boundaries while the rest are at
559 /// 4-byte boundaries.
560 unsigned getByValTypeAlignment(Type *Ty) const override;
562 /// getOptimalMemOpType - Returns the target specific optimal type for load
563 /// and store operations as a result of memset, memcpy, and memmove
564 /// lowering. If DstAlign is zero that means it's safe to destination
565 /// alignment can satisfy any constraint. Similarly if SrcAlign is zero it
566 /// means there isn't a need to check it against alignment requirement,
567 /// probably because the source does not need to be loaded. If 'IsMemset' is
568 /// true, that means it's expanding a memset. If 'ZeroMemset' is true, that
569 /// means it's a memset of zero. 'MemcpyStrSrc' indicates whether the memcpy
570 /// source is constant so it does not need to be loaded.
571 /// It returns EVT::Other if the type should be determined using generic
572 /// target-independent logic.
573 EVT getOptimalMemOpType(uint64_t Size, unsigned DstAlign, unsigned SrcAlign,
574 bool IsMemset, bool ZeroMemset, bool MemcpyStrSrc,
575 MachineFunction &MF) const override;
577 /// isSafeMemOpType - Returns true if it's safe to use load / store of the
578 /// specified type to expand memcpy / memset inline. This is mostly true
579 /// for all types except for some special cases. For example, on X86
580 /// targets without SSE2 f64 load / store are done with fldl / fstpl which
581 /// also does type conversion. Note the specified type doesn't have to be
582 /// legal as the hook is used before type legalization.
583 bool isSafeMemOpType(MVT VT) const override;
585 /// allowsMisalignedMemoryAccesses - Returns true if the target allows
586 /// unaligned memory accesses. of the specified type. Returns whether it
587 /// is "fast" by reference in the second argument.
588 bool allowsMisalignedMemoryAccesses(EVT VT, unsigned AS, unsigned Align,
589 bool *Fast) const override;
591 /// LowerOperation - Provide custom lowering hooks for some operations.
593 SDValue LowerOperation(SDValue Op, SelectionDAG &DAG) const override;
595 /// ReplaceNodeResults - Replace the results of node with an illegal result
596 /// type with new values built out of custom code.
598 void ReplaceNodeResults(SDNode *N, SmallVectorImpl<SDValue>&Results,
599 SelectionDAG &DAG) const override;
602 SDValue PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const override;
604 /// isTypeDesirableForOp - Return true if the target has native support for
605 /// the specified value type and it is 'desirable' to use the type for the
606 /// given node type. e.g. On x86 i16 is legal, but undesirable since i16
607 /// instruction encodings are longer and some i16 instructions are slow.
608 bool isTypeDesirableForOp(unsigned Opc, EVT VT) const override;
610 /// isTypeDesirable - Return true if the target has native support for the
611 /// specified value type and it is 'desirable' to use the type. e.g. On x86
612 /// i16 is legal, but undesirable since i16 instruction encodings are longer
613 /// and some i16 instructions are slow.
614 bool IsDesirableToPromoteOp(SDValue Op, EVT &PVT) const override;
617 EmitInstrWithCustomInserter(MachineInstr *MI,
618 MachineBasicBlock *MBB) const override;
621 /// getTargetNodeName - This method returns the name of a target specific
623 const char *getTargetNodeName(unsigned Opcode) const override;
625 /// getSetCCResultType - Return the value type to use for ISD::SETCC.
626 EVT getSetCCResultType(LLVMContext &Context, EVT VT) const override;
628 /// computeKnownBitsForTargetNode - Determine which of the bits specified
629 /// in Mask are known to be either zero or one and return them in the
630 /// KnownZero/KnownOne bitsets.
631 void computeKnownBitsForTargetNode(const SDValue Op,
634 const SelectionDAG &DAG,
635 unsigned Depth = 0) const override;
637 // ComputeNumSignBitsForTargetNode - Determine the number of bits in the
638 // operation that are sign bits.
639 unsigned ComputeNumSignBitsForTargetNode(SDValue Op,
640 const SelectionDAG &DAG,
641 unsigned Depth) const override;
643 bool isGAPlusOffset(SDNode *N, const GlobalValue* &GA,
644 int64_t &Offset) const override;
646 SDValue getReturnAddressFrameIndex(SelectionDAG &DAG) const;
648 bool ExpandInlineAsm(CallInst *CI) const override;
651 getConstraintType(const std::string &Constraint) const override;
653 /// Examine constraint string and operand type and determine a weight value.
654 /// The operand object must already have been set up with the operand type.
656 getSingleConstraintMatchWeight(AsmOperandInfo &info,
657 const char *constraint) const override;
659 const char *LowerXConstraint(EVT ConstraintVT) const override;
661 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
662 /// vector. If it is invalid, don't add anything to Ops. If hasMemory is
663 /// true it means one of the asm constraint of the inline asm instruction
664 /// being processed is 'm'.
665 void LowerAsmOperandForConstraint(SDValue Op,
666 std::string &Constraint,
667 std::vector<SDValue> &Ops,
668 SelectionDAG &DAG) const override;
670 /// getRegForInlineAsmConstraint - Given a physical register constraint
671 /// (e.g. {edx}), return the register number and the register class for the
672 /// register. This should only be used for C_Register constraints. On
673 /// error, this returns a register number of 0.
674 std::pair<unsigned, const TargetRegisterClass*>
675 getRegForInlineAsmConstraint(const std::string &Constraint,
676 MVT VT) const override;
678 /// isLegalAddressingMode - Return true if the addressing mode represented
679 /// by AM is legal for this target, for a load/store of the specified type.
680 bool isLegalAddressingMode(const AddrMode &AM, Type *Ty) const override;
682 /// isLegalICmpImmediate - Return true if the specified immediate is legal
683 /// icmp immediate, that is the target has icmp instructions which can
684 /// compare a register against the immediate without having to materialize
685 /// the immediate into a register.
686 bool isLegalICmpImmediate(int64_t Imm) const override;
688 /// isLegalAddImmediate - Return true if the specified immediate is legal
689 /// add immediate, that is the target has add instructions which can
690 /// add a register and the immediate without having to materialize
691 /// the immediate into a register.
692 bool isLegalAddImmediate(int64_t Imm) const override;
694 /// \brief Return the cost of the scaling factor used in the addressing
695 /// mode represented by AM for this target, for a load/store
696 /// of the specified type.
697 /// If the AM is supported, the return value must be >= 0.
698 /// If the AM is not supported, it returns a negative value.
699 int getScalingFactorCost(const AddrMode &AM, Type *Ty) const override;
701 bool isVectorShiftByScalarCheap(Type *Ty) const override;
703 /// isTruncateFree - Return true if it's free to truncate a value of
704 /// type Ty1 to type Ty2. e.g. On x86 it's free to truncate a i32 value in
705 /// register EAX to i16 by referencing its sub-register AX.
706 bool isTruncateFree(Type *Ty1, Type *Ty2) const override;
707 bool isTruncateFree(EVT VT1, EVT VT2) const override;
709 bool allowTruncateForTailCall(Type *Ty1, Type *Ty2) const override;
711 /// isZExtFree - Return true if any actual instruction that defines a
712 /// value of type Ty1 implicit zero-extends the value to Ty2 in the result
713 /// register. This does not necessarily include registers defined in
714 /// unknown ways, such as incoming arguments, or copies from unknown
715 /// virtual registers. Also, if isTruncateFree(Ty2, Ty1) is true, this
716 /// does not necessarily apply to truncate instructions. e.g. on x86-64,
717 /// all instructions that define 32-bit values implicit zero-extend the
718 /// result out to 64 bits.
719 bool isZExtFree(Type *Ty1, Type *Ty2) const override;
720 bool isZExtFree(EVT VT1, EVT VT2) const override;
721 bool isZExtFree(SDValue Val, EVT VT2) const override;
723 /// isFMAFasterThanFMulAndFAdd - Return true if an FMA operation is faster
724 /// than a pair of fmul and fadd instructions. fmuladd intrinsics will be
725 /// expanded to FMAs when this method returns true, otherwise fmuladd is
726 /// expanded to fmul + fadd.
727 bool isFMAFasterThanFMulAndFAdd(EVT VT) const override;
729 /// isNarrowingProfitable - Return true if it's profitable to narrow
730 /// operations of type VT1 to VT2. e.g. on x86, it's profitable to narrow
731 /// from i32 to i8 but not from i32 to i16.
732 bool isNarrowingProfitable(EVT VT1, EVT VT2) const override;
734 /// isFPImmLegal - Returns true if the target can instruction select the
735 /// specified FP immediate natively. If false, the legalizer will
736 /// materialize the FP immediate as a load from a constant pool.
737 bool isFPImmLegal(const APFloat &Imm, EVT VT) const override;
739 /// isShuffleMaskLegal - Targets can use this to indicate that they only
740 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
741 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask
742 /// values are assumed to be legal.
743 bool isShuffleMaskLegal(const SmallVectorImpl<int> &Mask,
744 EVT VT) const override;
746 /// isVectorClearMaskLegal - Similar to isShuffleMaskLegal. This is
747 /// used by Targets can use this to indicate if there is a suitable
748 /// VECTOR_SHUFFLE that can be used to replace a VAND with a constant
750 bool isVectorClearMaskLegal(const SmallVectorImpl<int> &Mask,
751 EVT VT) const override;
753 /// ShouldShrinkFPConstant - If true, then instruction selection should
754 /// seek to shrink the FP constant of the specified type to a smaller type
755 /// in order to save space and / or reduce runtime.
756 bool ShouldShrinkFPConstant(EVT VT) const override {
757 // Don't shrink FP constpool if SSE2 is available since cvtss2sd is more
758 // expensive than a straight movsd. On the other hand, it's important to
759 // shrink long double fp constant since fldt is very slow.
760 return !X86ScalarSSEf64 || VT == MVT::f80;
763 const X86Subtarget* getSubtarget() const {
767 /// isScalarFPTypeInSSEReg - Return true if the specified scalar FP type is
768 /// computed in an SSE register, not on the X87 floating point stack.
769 bool isScalarFPTypeInSSEReg(EVT VT) const {
770 return (VT == MVT::f64 && X86ScalarSSEf64) || // f64 is when SSE2
771 (VT == MVT::f32 && X86ScalarSSEf32); // f32 is when SSE1
774 /// isTargetFTOL - Return true if the target uses the MSVC _ftol2 routine
776 bool isTargetFTOL() const;
778 /// isIntegerTypeFTOL - Return true if the MSVC _ftol2 routine should be
779 /// used for fptoui to the given type.
780 bool isIntegerTypeFTOL(EVT VT) const {
781 return isTargetFTOL() && VT == MVT::i64;
784 /// \brief Returns true if it is beneficial to convert a load of a constant
785 /// to just the constant itself.
786 bool shouldConvertConstantLoadToIntImm(const APInt &Imm,
787 Type *Ty) const override;
789 /// Intel processors have a unified instruction and data cache
790 const char * getClearCacheBuiltinName() const override {
791 return nullptr; // nothing to do, move along.
794 unsigned getRegisterByName(const char* RegName, EVT VT) const override;
796 /// createFastISel - This method returns a target specific FastISel object,
797 /// or null if the target does not support "fast" ISel.
798 FastISel *createFastISel(FunctionLoweringInfo &funcInfo,
799 const TargetLibraryInfo *libInfo) const override;
801 /// getStackCookieLocation - Return true if the target stores stack
802 /// protector cookies at a fixed offset in some non-standard address
803 /// space, and populates the address space and offset as
805 bool getStackCookieLocation(unsigned &AddressSpace,
806 unsigned &Offset) const override;
808 SDValue BuildFILD(SDValue Op, EVT SrcVT, SDValue Chain, SDValue StackSlot,
809 SelectionDAG &DAG) const;
811 bool isNoopAddrSpaceCast(unsigned SrcAS, unsigned DestAS) const override;
813 /// \brief Reset the operation actions based on target options.
814 void resetOperationActions() override;
816 bool useLoadStackGuardNode() const override;
817 /// \brief Customize the preferred legalization strategy for certain types.
818 LegalizeTypeAction getPreferredVectorAction(EVT VT) const override;
821 std::pair<const TargetRegisterClass*, uint8_t>
822 findRepresentativeClass(MVT VT) const override;
825 /// Subtarget - Keep a pointer to the X86Subtarget around so that we can
826 /// make the right decision when generating code for different targets.
827 const X86Subtarget *Subtarget;
828 const DataLayout *TD;
830 /// Used to store the TargetOptions so that we don't waste time resetting
831 /// the operation actions unless we have to.
834 /// X86ScalarSSEf32, X86ScalarSSEf64 - Select between SSE or x87
835 /// floating point ops.
836 /// When SSE is available, use it for f32 operations.
837 /// When SSE2 is available, use it for f64 operations.
838 bool X86ScalarSSEf32;
839 bool X86ScalarSSEf64;
841 /// LegalFPImmediates - A list of legal fp immediates.
842 std::vector<APFloat> LegalFPImmediates;
844 /// addLegalFPImmediate - Indicate that this x86 target can instruction
845 /// select the specified FP immediate natively.
846 void addLegalFPImmediate(const APFloat& Imm) {
847 LegalFPImmediates.push_back(Imm);
850 SDValue LowerCallResult(SDValue Chain, SDValue InFlag,
851 CallingConv::ID CallConv, bool isVarArg,
852 const SmallVectorImpl<ISD::InputArg> &Ins,
853 SDLoc dl, SelectionDAG &DAG,
854 SmallVectorImpl<SDValue> &InVals) const;
855 SDValue LowerMemArgument(SDValue Chain,
856 CallingConv::ID CallConv,
857 const SmallVectorImpl<ISD::InputArg> &ArgInfo,
858 SDLoc dl, SelectionDAG &DAG,
859 const CCValAssign &VA, MachineFrameInfo *MFI,
861 SDValue LowerMemOpCallTo(SDValue Chain, SDValue StackPtr, SDValue Arg,
862 SDLoc dl, SelectionDAG &DAG,
863 const CCValAssign &VA,
864 ISD::ArgFlagsTy Flags) const;
866 // Call lowering helpers.
868 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
869 /// for tail call optimization. Targets which want to do tail call
870 /// optimization should implement this function.
871 bool IsEligibleForTailCallOptimization(SDValue Callee,
872 CallingConv::ID CalleeCC,
874 bool isCalleeStructRet,
875 bool isCallerStructRet,
877 const SmallVectorImpl<ISD::OutputArg> &Outs,
878 const SmallVectorImpl<SDValue> &OutVals,
879 const SmallVectorImpl<ISD::InputArg> &Ins,
880 SelectionDAG& DAG) const;
881 bool IsCalleePop(bool isVarArg, CallingConv::ID CallConv) const;
882 SDValue EmitTailCallLoadRetAddr(SelectionDAG &DAG, SDValue &OutRetAddr,
883 SDValue Chain, bool IsTailCall, bool Is64Bit,
884 int FPDiff, SDLoc dl) const;
886 unsigned GetAlignedArgumentStackSize(unsigned StackSize,
887 SelectionDAG &DAG) const;
889 std::pair<SDValue,SDValue> FP_TO_INTHelper(SDValue Op, SelectionDAG &DAG,
891 bool isReplace) const;
893 SDValue LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) const;
894 SDValue LowerBUILD_VECTORvXi1(SDValue Op, SelectionDAG &DAG) const;
895 SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) const;
896 SDValue LowerVSELECT(SDValue Op, SelectionDAG &DAG) const;
897 SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const;
898 SDValue ExtractBitFromMaskVector(SDValue Op, SelectionDAG &DAG) const;
899 SDValue InsertBitToMaskVector(SDValue Op, SelectionDAG &DAG) const;
901 SDValue LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const;
902 SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) const;
903 SDValue LowerBlockAddress(SDValue Op, SelectionDAG &DAG) const;
904 SDValue LowerGlobalAddress(const GlobalValue *GV, SDLoc dl,
905 int64_t Offset, SelectionDAG &DAG) const;
906 SDValue LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) const;
907 SDValue LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const;
908 SDValue LowerExternalSymbol(SDValue Op, SelectionDAG &DAG) const;
909 SDValue LowerSINT_TO_FP(SDValue Op, SelectionDAG &DAG) const;
910 SDValue LowerUINT_TO_FP(SDValue Op, SelectionDAG &DAG) const;
911 SDValue LowerUINT_TO_FP_i64(SDValue Op, SelectionDAG &DAG) const;
912 SDValue LowerUINT_TO_FP_i32(SDValue Op, SelectionDAG &DAG) const;
913 SDValue lowerUINT_TO_FP_vec(SDValue Op, SelectionDAG &DAG) const;
914 SDValue LowerTRUNCATE(SDValue Op, SelectionDAG &DAG) const;
915 SDValue LowerFP_TO_SINT(SDValue Op, SelectionDAG &DAG) const;
916 SDValue LowerFP_TO_UINT(SDValue Op, SelectionDAG &DAG) const;
917 SDValue LowerToBT(SDValue And, ISD::CondCode CC,
918 SDLoc dl, SelectionDAG &DAG) const;
919 SDValue LowerSETCC(SDValue Op, SelectionDAG &DAG) const;
920 SDValue LowerSELECT(SDValue Op, SelectionDAG &DAG) const;
921 SDValue LowerBRCOND(SDValue Op, SelectionDAG &DAG) const;
922 SDValue LowerMEMSET(SDValue Op, SelectionDAG &DAG) const;
923 SDValue LowerJumpTable(SDValue Op, SelectionDAG &DAG) const;
924 SDValue LowerDYNAMIC_STACKALLOC(SDValue Op, SelectionDAG &DAG) const;
925 SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG) const;
926 SDValue LowerVAARG(SDValue Op, SelectionDAG &DAG) const;
927 SDValue LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const;
928 SDValue LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const;
929 SDValue LowerFRAME_TO_ARGS_OFFSET(SDValue Op, SelectionDAG &DAG) const;
930 SDValue LowerEH_RETURN(SDValue Op, SelectionDAG &DAG) const;
931 SDValue lowerEH_SJLJ_SETJMP(SDValue Op, SelectionDAG &DAG) const;
932 SDValue lowerEH_SJLJ_LONGJMP(SDValue Op, SelectionDAG &DAG) const;
933 SDValue LowerINIT_TRAMPOLINE(SDValue Op, SelectionDAG &DAG) const;
934 SDValue LowerFLT_ROUNDS_(SDValue Op, SelectionDAG &DAG) const;
935 SDValue LowerSIGN_EXTEND_INREG(SDValue Op, SelectionDAG &DAG) const;
936 SDValue LowerWin64_i128OP(SDValue Op, SelectionDAG &DAG) const;
939 LowerFormalArguments(SDValue Chain,
940 CallingConv::ID CallConv, bool isVarArg,
941 const SmallVectorImpl<ISD::InputArg> &Ins,
942 SDLoc dl, SelectionDAG &DAG,
943 SmallVectorImpl<SDValue> &InVals) const override;
944 SDValue LowerCall(CallLoweringInfo &CLI,
945 SmallVectorImpl<SDValue> &InVals) const override;
947 SDValue LowerReturn(SDValue Chain,
948 CallingConv::ID CallConv, bool isVarArg,
949 const SmallVectorImpl<ISD::OutputArg> &Outs,
950 const SmallVectorImpl<SDValue> &OutVals,
951 SDLoc dl, SelectionDAG &DAG) const override;
953 bool isUsedByReturnOnly(SDNode *N, SDValue &Chain) const override;
955 bool mayBeEmittedAsTailCall(CallInst *CI) const override;
957 EVT getTypeForExtArgOrReturn(LLVMContext &Context, EVT VT,
958 ISD::NodeType ExtendKind) const override;
960 bool CanLowerReturn(CallingConv::ID CallConv, MachineFunction &MF,
962 const SmallVectorImpl<ISD::OutputArg> &Outs,
963 LLVMContext &Context) const override;
965 const MCPhysReg *getScratchRegisters(CallingConv::ID CC) const override;
967 bool shouldExpandAtomicLoadInIR(LoadInst *SI) const override;
968 bool shouldExpandAtomicStoreInIR(StoreInst *SI) const override;
969 bool shouldExpandAtomicRMWInIR(AtomicRMWInst *AI) const override;
971 bool needsCmpXchgNb(const Type *MemType) const;
973 /// Utility function to emit atomic-load-arith operations (and, or, xor,
974 /// nand, max, min, umax, umin). It takes the corresponding instruction to
975 /// expand, the associated machine basic block, and the associated X86
976 /// opcodes for reg/reg.
977 MachineBasicBlock *EmitAtomicLoadArith(MachineInstr *MI,
978 MachineBasicBlock *MBB) const;
980 /// Utility function to emit atomic-load-arith operations (and, or, xor,
981 /// nand, add, sub, swap) for 64-bit operands on 32-bit target.
982 MachineBasicBlock *EmitAtomicLoadArith6432(MachineInstr *MI,
983 MachineBasicBlock *MBB) const;
985 // Utility function to emit the low-level va_arg code for X86-64.
986 MachineBasicBlock *EmitVAARG64WithCustomInserter(
988 MachineBasicBlock *MBB) const;
990 /// Utility function to emit the xmm reg save portion of va_start.
991 MachineBasicBlock *EmitVAStartSaveXMMRegsWithCustomInserter(
992 MachineInstr *BInstr,
993 MachineBasicBlock *BB) const;
995 MachineBasicBlock *EmitLoweredSelect(MachineInstr *I,
996 MachineBasicBlock *BB) const;
998 MachineBasicBlock *EmitLoweredWinAlloca(MachineInstr *MI,
999 MachineBasicBlock *BB) const;
1001 MachineBasicBlock *EmitLoweredSegAlloca(MachineInstr *MI,
1002 MachineBasicBlock *BB) const;
1004 MachineBasicBlock *EmitLoweredTLSCall(MachineInstr *MI,
1005 MachineBasicBlock *BB) const;
1007 MachineBasicBlock *emitLoweredTLSAddr(MachineInstr *MI,
1008 MachineBasicBlock *BB) const;
1010 MachineBasicBlock *emitEHSjLjSetJmp(MachineInstr *MI,
1011 MachineBasicBlock *MBB) const;
1013 MachineBasicBlock *emitEHSjLjLongJmp(MachineInstr *MI,
1014 MachineBasicBlock *MBB) const;
1016 MachineBasicBlock *emitFMA3Instr(MachineInstr *MI,
1017 MachineBasicBlock *MBB) const;
1019 /// Emit nodes that will be selected as "test Op0,Op0", or something
1020 /// equivalent, for use with the given x86 condition code.
1021 SDValue EmitTest(SDValue Op0, unsigned X86CC, SDLoc dl,
1022 SelectionDAG &DAG) const;
1024 /// Emit nodes that will be selected as "cmp Op0,Op1", or something
1025 /// equivalent, for use with the given x86 condition code.
1026 SDValue EmitCmp(SDValue Op0, SDValue Op1, unsigned X86CC, SDLoc dl,
1027 SelectionDAG &DAG) const;
1029 /// Convert a comparison if required by the subtarget.
1030 SDValue ConvertCmpIfNecessary(SDValue Cmp, SelectionDAG &DAG) const;
1034 FastISel *createFastISel(FunctionLoweringInfo &funcInfo,
1035 const TargetLibraryInfo *libInfo);
1039 #endif // X86ISELLOWERING_H