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
29 enum NodeType : unsigned {
30 // Start the numbering where the builtin ops leave off.
31 FIRST_NUMBER = ISD::BUILTIN_OP_END,
38 /// Double shift instructions. These correspond to
39 /// X86::SHLDxx and X86::SHRDxx instructions.
43 /// Bitwise logical AND of floating point values. This corresponds
44 /// to X86::ANDPS or X86::ANDPD.
47 /// Bitwise logical OR of floating point values. This corresponds
48 /// to X86::ORPS or X86::ORPD.
51 /// Bitwise logical XOR of floating point values. This corresponds
52 /// to X86::XORPS or X86::XORPD.
55 /// Bitwise logical ANDNOT of floating point values. This
56 /// corresponds to X86::ANDNPS or X86::ANDNPD.
59 /// These operations represent an abstract X86 call
60 /// instruction, which includes a bunch of information. In particular the
61 /// operands of these node are:
63 /// #0 - The incoming token chain
65 /// #2 - The number of arg bytes the caller pushes on the stack.
66 /// #3 - The number of arg bytes the callee pops off the stack.
67 /// #4 - The value to pass in AL/AX/EAX (optional)
68 /// #5 - The value to pass in DL/DX/EDX (optional)
70 /// The result values of these nodes are:
72 /// #0 - The outgoing token chain
73 /// #1 - The first register result value (optional)
74 /// #2 - The second register result value (optional)
78 /// This operation implements the lowering for readcyclecounter.
81 /// X86 Read Time-Stamp Counter and Processor ID.
84 /// X86 Read Performance Monitoring Counters.
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.
100 // Same as SETCC except it's materialized with a sbb and the value is all
101 // one's or all zero's.
102 SETCC_CARRY, // R = carry_bit ? ~0 : 0
104 /// X86 FP SETCC, implemented with CMP{cc}SS/CMP{cc}SD.
105 /// Operands are two FP values to compare; result is a mask of
106 /// 0s or 1s. Generally DTRT for C/C++ with NaNs.
109 /// X86 FP SETCC, similar to above, but with output as an i1 mask and
110 /// with optional rounding mode.
111 FSETCCM, FSETCCM_RND,
113 /// X86 conditional moves. Operand 0 and operand 1 are the two values
114 /// to select from. Operand 2 is the condition code, and operand 3 is the
115 /// flag operand produced by a CMP or TEST instruction. It also writes a
119 /// X86 conditional branches. Operand 0 is the chain operand, operand 1
120 /// is the block to branch if condition is true, operand 2 is the
121 /// condition code, and operand 3 is the flag operand produced by a CMP
122 /// or TEST instruction.
125 /// Return with a flag operand. Operand 0 is the chain operand, operand
126 /// 1 is the number of bytes of stack to pop.
129 /// Return from interrupt. Operand 0 is the number of bytes to pop.
132 /// Repeat fill, corresponds to X86::REP_STOSx.
135 /// Repeat move, corresponds to X86::REP_MOVSx.
138 /// On Darwin, this node represents the result of the popl
139 /// at function entry, used for PIC code.
142 /// A wrapper node for TargetConstantPool, TargetJumpTable,
143 /// TargetExternalSymbol, TargetGlobalAddress, TargetGlobalTLSAddress,
144 /// MCSymbol and TargetBlockAddress.
147 /// Special wrapper used under X86-64 PIC mode for RIP
148 /// relative displacements.
151 /// Copies a 64-bit value from the low word of an XMM vector
152 /// to an MMX vector.
155 /// Copies a 32-bit value from the low word of a MMX
159 /// Copies a GPR into the low 32-bit word of a MMX vector
160 /// and zero out the high word.
163 /// Extract an 8-bit value from a vector and zero extend it to
164 /// i32, corresponds to X86::PEXTRB.
167 /// Extract a 16-bit value from a vector and zero extend it to
168 /// i32, corresponds to X86::PEXTRW.
171 /// Insert any element of a 4 x float vector into any element
172 /// of a destination 4 x floatvector.
175 /// Insert the lower 8-bits of a 32-bit value to a vector,
176 /// corresponds to X86::PINSRB.
179 /// Insert the lower 16-bits of a 32-bit value to a vector,
180 /// corresponds to X86::PINSRW.
183 /// Shuffle 16 8-bit values within a vector.
186 /// Compute Sum of Absolute Differences.
188 /// Compute Double Block Packed Sum-Absolute-Differences
191 /// Bitwise Logical AND NOT of Packed FP values.
194 /// Blend where the selector is an immediate.
197 /// Dynamic (non-constant condition) vector blend where only the sign bits
198 /// of the condition elements are used. This is used to enforce that the
199 /// condition mask is not valid for generic VSELECT optimizations.
202 /// Combined add and sub on an FP vector.
205 // FP vector ops with rounding mode.
212 FSQRT_RND, FSQRTS_RND,
214 // FP vector get exponent.
215 FGETEXP_RND, FGETEXPS_RND,
216 // Extract Normalized Mantissas.
222 // Integer add/sub with unsigned saturation.
226 // Integer add/sub with signed saturation.
230 // Unsigned Integer average.
233 /// Integer horizontal add/sub.
237 /// Floating point horizontal add/sub.
241 // Detect Conflicts Within a Vector
244 /// Floating point max and min.
247 /// Commutative FMIN and FMAX.
250 /// Scalar intrinsic floating point max and min.
253 /// Floating point reciprocal-sqrt and reciprocal approximation.
254 /// Note that these typically require refinement
255 /// in order to obtain suitable precision.
259 // Thread Local Storage.
262 // Thread Local Storage. A call to get the start address
263 // of the TLS block for the current module.
266 // Thread Local Storage. When calling to an OS provided
267 // thunk at the address from an earlier relocation.
270 // Exception Handling helpers.
273 // SjLj exception handling setjmp.
276 // SjLj exception handling longjmp.
279 // SjLj exception handling dispatch.
280 EH_SJLJ_SETUP_DISPATCH,
282 /// Tail call return. See X86TargetLowering::LowerCall for
283 /// the list of operands.
286 // Vector move to low scalar and zero higher vector elements.
289 // Vector integer zero-extend.
291 // Vector integer signed-extend.
294 // Vector integer truncate.
296 // Vector integer truncate with unsigned/signed saturation.
300 VFPEXT, VFPEXT_RND, VFPEXTS_RND,
303 VFPROUND, VFPROUND_RND, VFPROUNDS_RND,
305 // Convert a vector to mask, set bits base on MSB.
308 // 128-bit vector logical left / right shift
311 // Vector shift elements
314 // Vector variable shift right arithmetic.
315 // Unlike ISD::SRA, in case shift count greater then element size
316 // use sign bit to fill destination data element.
319 // Vector shift elements by immediate
322 // Shifts of mask registers.
325 // Bit rotate by immediate
328 // Vector packed double/float comparison.
331 // Vector integer comparisons.
333 // Vector integer comparisons, the result is in a mask vector.
338 /// Vector comparison generating mask bits for fp and
339 /// integer signed and unsigned data types.
342 // Vector comparison with rounding mode for FP values
345 // Arithmetic operations with FLAGS results.
346 ADD, SUB, ADC, SBB, SMUL,
347 INC, DEC, OR, XOR, AND,
349 // Bit field extract.
352 // LOW, HI, FLAGS = umul LHS, RHS.
355 // 8-bit SMUL/UMUL - AX, FLAGS = smul8/umul8 AL, RHS.
358 // 8-bit divrem that zero-extend the high result (AH).
362 // X86-specific multiply by immediate.
365 // Vector sign bit extraction.
368 // Vector bitwise comparisons.
371 // Vector packed fp sign bitwise comparisons.
374 // Vector "test" in AVX-512, the result is in a mask vector.
378 // OR/AND test for masks.
382 // Several flavors of instructions with vector shuffle behaviors.
383 // Saturated signed/unnsigned packing.
386 // Intra-lane alignr.
388 // AVX512 inter-lane alignr.
394 //Shuffle Packed Values at 128-bit granularity.
413 // Variable Permute (VPERM).
414 // Res = VPERMV MaskV, V0
417 // 3-op Variable Permute (VPERMT2).
418 // Res = VPERMV3 V0, MaskV, V1
421 // 3-op Variable Permute overwriting the index (VPERMI2).
422 // Res = VPERMIV3 V0, MaskV, V1
425 // Bitwise ternary logic.
427 // Fix Up Special Packed Float32/64 values.
430 // Range Restriction Calculation For Packed Pairs of Float32/64 values.
432 // Reduce - Perform Reduction Transformation on scalar\packed FP.
434 // RndScale - Round FP Values To Include A Given Number Of Fraction Bits.
435 VRNDSCALE, VRNDSCALES,
436 // Tests Types Of a FP Values for packed types.
438 // Tests Types Of a FP Values for scalar types.
441 // Broadcast scalar to vector.
443 // Broadcast mask to vector.
445 // Broadcast subvector to vector.
448 // Extract vector element.
451 /// SSE4A Extraction and Insertion.
454 // XOP variable/immediate rotations.
456 // XOP arithmetic/logical shifts.
458 // XOP signed/unsigned integer comparisons.
460 // XOP packed permute bytes.
462 // XOP two source permutation.
465 // Vector multiply packed unsigned doubleword integers.
467 // Vector multiply packed signed doubleword integers.
469 // Vector Multiply Packed UnsignedIntegers with Round and Scale.
472 // Multiply and Add Packed Integers.
473 VPMADDUBSW, VPMADDWD,
474 VPMADD52L, VPMADD52H,
484 // FMA with rounding mode.
492 // Scalar intrinsic FMA with rounding mode.
493 // Two versions, passthru bits on op1 or op3.
494 FMADDS1_RND, FMADDS3_RND,
495 FNMADDS1_RND, FNMADDS3_RND,
496 FMSUBS1_RND, FMSUBS3_RND,
497 FNMSUBS1_RND, FNMSUBS3_RND,
499 // Compress and expand.
503 // Convert Unsigned/Integer to Floating-Point Value with rounding mode.
504 SINT_TO_FP_RND, UINT_TO_FP_RND,
505 SCALAR_SINT_TO_FP_RND, SCALAR_UINT_TO_FP_RND,
507 // Vector float/double to signed/unsigned integer.
508 CVTP2SI, CVTP2UI, CVTP2SI_RND, CVTP2UI_RND,
509 // Scalar float/double to signed/unsigned integer.
510 CVTS2SI_RND, CVTS2UI_RND,
512 // Vector float/double to signed/unsigned integer with truncation.
513 CVTTP2SI, CVTTP2UI, CVTTP2SI_RND, CVTTP2UI_RND,
514 // Scalar float/double to signed/unsigned integer with truncation.
515 CVTTS2SI_RND, CVTTS2UI_RND,
517 // Vector signed/unsigned integer to float/double.
520 // Save xmm argument registers to the stack, according to %al. An operator
521 // is needed so that this can be expanded with control flow.
522 VASTART_SAVE_XMM_REGS,
524 // Windows's _chkstk call to do stack probing.
527 // For allocating variable amounts of stack space when using
528 // segmented stacks. Check if the current stacklet has enough space, and
529 // falls back to heap allocation if not.
536 // Store FP status word into i16 register.
539 // Store contents of %ah into %eflags.
542 // Get a random integer and indicate whether it is valid in CF.
545 // Get a NIST SP800-90B & C compliant random integer and
546 // indicate whether it is valid in CF.
549 // SSE42 string comparisons.
553 // Test if in transactional execution.
557 RSQRT28, RSQRT28S, RCP28, RCP28S, EXP2,
559 // Conversions between float and half-float.
563 LCMPXCHG_DAG = ISD::FIRST_TARGET_MEMORY_OPCODE,
566 LCMPXCHG8_SAVE_EBX_DAG,
567 LCMPXCHG16_SAVE_RBX_DAG,
569 /// LOCK-prefixed arithmetic read-modify-write instructions.
570 /// EFLAGS, OUTCHAIN = LADD(INCHAIN, PTR, RHS)
571 LADD, LSUB, LOR, LXOR, LAND,
573 // Load, scalar_to_vector, and zero extend.
576 // Store FP control world into i16 memory.
579 /// This instruction implements FP_TO_SINT with the
580 /// integer destination in memory and a FP reg source. This corresponds
581 /// to the X86::FIST*m instructions and the rounding mode change stuff. It
582 /// has two inputs (token chain and address) and two outputs (int value
583 /// and token chain).
588 /// This instruction implements SINT_TO_FP with the
589 /// integer source in memory and FP reg result. This corresponds to the
590 /// X86::FILD*m instructions. It has three inputs (token chain, address,
591 /// and source type) and two outputs (FP value and token chain). FILD_FLAG
592 /// also produces a flag).
596 /// This instruction implements an extending load to FP stack slots.
597 /// This corresponds to the X86::FLD32m / X86::FLD64m. It takes a chain
598 /// operand, ptr to load from, and a ValueType node indicating the type
602 /// This instruction implements a truncating store to FP stack
603 /// slots. This corresponds to the X86::FST32m / X86::FST64m. It takes a
604 /// chain operand, value to store, address, and a ValueType to store it
608 /// This instruction grabs the address of the next argument
609 /// from a va_list. (reads and modifies the va_list in memory)
612 // Vector truncating store with unsigned/signed saturation
613 VTRUNCSTOREUS, VTRUNCSTORES,
614 // Vector truncating masked store with unsigned/signed saturation
615 VMTRUNCSTOREUS, VMTRUNCSTORES
617 // WARNING: Do not add anything in the end unless you want the node to
618 // have memop! In fact, starting from FIRST_TARGET_MEMORY_OPCODE all
619 // opcodes will be thought as target memory ops!
621 } // end namespace X86ISD
623 /// Define some predicates that are used for node matching.
625 /// Return true if the specified
626 /// EXTRACT_SUBVECTOR operand specifies a vector extract that is
627 /// suitable for input to VEXTRACTF128, VEXTRACTI128 instructions.
628 bool isVEXTRACT128Index(SDNode *N);
630 /// Return true if the specified
631 /// INSERT_SUBVECTOR operand specifies a subvector insert that is
632 /// suitable for input to VINSERTF128, VINSERTI128 instructions.
633 bool isVINSERT128Index(SDNode *N);
635 /// Return true if the specified
636 /// EXTRACT_SUBVECTOR operand specifies a vector extract that is
637 /// suitable for input to VEXTRACTF64X4, VEXTRACTI64X4 instructions.
638 bool isVEXTRACT256Index(SDNode *N);
640 /// Return true if the specified
641 /// INSERT_SUBVECTOR operand specifies a subvector insert that is
642 /// suitable for input to VINSERTF64X4, VINSERTI64X4 instructions.
643 bool isVINSERT256Index(SDNode *N);
645 /// Return the appropriate
646 /// immediate to extract the specified EXTRACT_SUBVECTOR index
647 /// with VEXTRACTF128, VEXTRACTI128 instructions.
648 unsigned getExtractVEXTRACT128Immediate(SDNode *N);
650 /// Return the appropriate
651 /// immediate to insert at the specified INSERT_SUBVECTOR index
652 /// with VINSERTF128, VINSERT128 instructions.
653 unsigned getInsertVINSERT128Immediate(SDNode *N);
655 /// Return the appropriate
656 /// immediate to extract the specified EXTRACT_SUBVECTOR index
657 /// with VEXTRACTF64X4, VEXTRACTI64x4 instructions.
658 unsigned getExtractVEXTRACT256Immediate(SDNode *N);
660 /// Return the appropriate
661 /// immediate to insert at the specified INSERT_SUBVECTOR index
662 /// with VINSERTF64x4, VINSERTI64x4 instructions.
663 unsigned getInsertVINSERT256Immediate(SDNode *N);
665 /// Returns true if Elt is a constant zero or floating point constant +0.0.
666 bool isZeroNode(SDValue Elt);
668 /// Returns true of the given offset can be
669 /// fit into displacement field of the instruction.
670 bool isOffsetSuitableForCodeModel(int64_t Offset, CodeModel::Model M,
671 bool hasSymbolicDisplacement = true);
673 /// Determines whether the callee is required to pop its
674 /// own arguments. Callee pop is necessary to support tail calls.
675 bool isCalleePop(CallingConv::ID CallingConv,
676 bool is64Bit, bool IsVarArg, bool GuaranteeTCO);
678 } // end namespace X86
680 //===--------------------------------------------------------------------===//
681 // X86 Implementation of the TargetLowering interface
682 class X86TargetLowering final : public TargetLowering {
684 explicit X86TargetLowering(const X86TargetMachine &TM,
685 const X86Subtarget &STI);
687 unsigned getJumpTableEncoding() const override;
688 bool useSoftFloat() const override;
690 void markLibCallAttributes(MachineFunction *MF, unsigned CC,
691 ArgListTy &Args) const override;
693 MVT getScalarShiftAmountTy(const DataLayout &, EVT) const override {
698 LowerCustomJumpTableEntry(const MachineJumpTableInfo *MJTI,
699 const MachineBasicBlock *MBB, unsigned uid,
700 MCContext &Ctx) const override;
702 /// Returns relocation base for the given PIC jumptable.
703 SDValue getPICJumpTableRelocBase(SDValue Table,
704 SelectionDAG &DAG) const override;
706 getPICJumpTableRelocBaseExpr(const MachineFunction *MF,
707 unsigned JTI, MCContext &Ctx) const override;
709 /// Return the desired alignment for ByVal aggregate
710 /// function arguments in the caller parameter area. For X86, aggregates
711 /// that contains are placed at 16-byte boundaries while the rest are at
712 /// 4-byte boundaries.
713 unsigned getByValTypeAlignment(Type *Ty,
714 const DataLayout &DL) const override;
716 /// Returns the target specific optimal type for load
717 /// and store operations as a result of memset, memcpy, and memmove
718 /// lowering. If DstAlign is zero that means it's safe to destination
719 /// alignment can satisfy any constraint. Similarly if SrcAlign is zero it
720 /// means there isn't a need to check it against alignment requirement,
721 /// probably because the source does not need to be loaded. If 'IsMemset' is
722 /// true, that means it's expanding a memset. If 'ZeroMemset' is true, that
723 /// means it's a memset of zero. 'MemcpyStrSrc' indicates whether the memcpy
724 /// source is constant so it does not need to be loaded.
725 /// It returns EVT::Other if the type should be determined using generic
726 /// target-independent logic.
727 EVT getOptimalMemOpType(uint64_t Size, unsigned DstAlign, unsigned SrcAlign,
728 bool IsMemset, bool ZeroMemset, bool MemcpyStrSrc,
729 MachineFunction &MF) const override;
731 /// Returns true if it's safe to use load / store of the
732 /// specified type to expand memcpy / memset inline. This is mostly true
733 /// for all types except for some special cases. For example, on X86
734 /// targets without SSE2 f64 load / store are done with fldl / fstpl which
735 /// also does type conversion. Note the specified type doesn't have to be
736 /// legal as the hook is used before type legalization.
737 bool isSafeMemOpType(MVT VT) const override;
739 /// Returns true if the target allows unaligned memory accesses of the
740 /// specified type. Returns whether it is "fast" in the last argument.
741 bool allowsMisalignedMemoryAccesses(EVT VT, unsigned AS, unsigned Align,
742 bool *Fast) const override;
744 /// Provide custom lowering hooks for some operations.
746 SDValue LowerOperation(SDValue Op, SelectionDAG &DAG) const override;
748 /// Places new result values for the node in Results (their number
749 /// and types must exactly match those of the original return values of
750 /// the node), or leaves Results empty, which indicates that the node is not
751 /// to be custom lowered after all.
752 void LowerOperationWrapper(SDNode *N,
753 SmallVectorImpl<SDValue> &Results,
754 SelectionDAG &DAG) const override;
756 /// Replace the results of node with an illegal result
757 /// type with new values built out of custom code.
759 void ReplaceNodeResults(SDNode *N, SmallVectorImpl<SDValue>&Results,
760 SelectionDAG &DAG) const override;
762 SDValue PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const override;
764 /// Return true if the target has native support for
765 /// the specified value type and it is 'desirable' to use the type for the
766 /// given node type. e.g. On x86 i16 is legal, but undesirable since i16
767 /// instruction encodings are longer and some i16 instructions are slow.
768 bool isTypeDesirableForOp(unsigned Opc, EVT VT) const override;
770 /// Return true if the target has native support for the
771 /// specified value type and it is 'desirable' to use the type. e.g. On x86
772 /// i16 is legal, but undesirable since i16 instruction encodings are longer
773 /// and some i16 instructions are slow.
774 bool IsDesirableToPromoteOp(SDValue Op, EVT &PVT) const override;
776 /// Return true if the MachineFunction contains a COPY which would imply
777 /// HasOpaqueSPAdjustment.
778 bool hasCopyImplyingStackAdjustment(MachineFunction *MF) const override;
781 EmitInstrWithCustomInserter(MachineInstr &MI,
782 MachineBasicBlock *MBB) const override;
784 /// This method returns the name of a target specific DAG node.
785 const char *getTargetNodeName(unsigned Opcode) const override;
787 bool isCheapToSpeculateCttz() const override;
789 bool isCheapToSpeculateCtlz() const override;
791 bool isCtlzFast() const override;
793 bool hasBitPreservingFPLogic(EVT VT) const override {
794 return VT == MVT::f32 || VT == MVT::f64 || VT.isVector();
797 bool isMultiStoresCheaperThanBitsMerge(EVT LTy, EVT HTy) const override {
798 // If the pair to store is a mixture of float and int values, we will
799 // save two bitwise instructions and one float-to-int instruction and
800 // increase one store instruction. There is potentially a more
801 // significant benefit because it avoids the float->int domain switch
802 // for input value. So It is more likely a win.
803 if ((LTy.isFloatingPoint() && HTy.isInteger()) ||
804 (LTy.isInteger() && HTy.isFloatingPoint()))
806 // If the pair only contains int values, we will save two bitwise
807 // instructions and increase one store instruction (costing one more
808 // store buffer). Since the benefit is more blurred so we leave
809 // such pair out until we get testcase to prove it is a win.
813 bool isMaskAndCmp0FoldingBeneficial(const Instruction &AndI) const override;
815 bool hasAndNotCompare(SDValue Y) const override;
817 bool convertSetCCLogicToBitwiseLogic(EVT VT) const override {
818 return VT.isScalarInteger();
821 /// Vector-sized comparisons are fast using PCMPEQ + PMOVMSK or PTEST.
822 MVT hasFastEqualityCompare(unsigned NumBits) const override;
824 /// Return the value type to use for ISD::SETCC.
825 EVT getSetCCResultType(const DataLayout &DL, LLVMContext &Context,
826 EVT VT) const override;
828 /// Determine which of the bits specified in Mask are known to be either
829 /// zero or one and return them in the KnownZero/KnownOne bitsets.
830 void computeKnownBitsForTargetNode(const SDValue Op,
833 const APInt &DemandedElts,
834 const SelectionDAG &DAG,
835 unsigned Depth = 0) const override;
837 /// Determine the number of bits in the operation that are sign bits.
838 unsigned ComputeNumSignBitsForTargetNode(SDValue Op,
839 const APInt &DemandedElts,
840 const SelectionDAG &DAG,
841 unsigned Depth) const override;
843 bool isGAPlusOffset(SDNode *N, const GlobalValue* &GA,
844 int64_t &Offset) const override;
846 SDValue getReturnAddressFrameIndex(SelectionDAG &DAG) const;
848 bool ExpandInlineAsm(CallInst *CI) const override;
850 ConstraintType getConstraintType(StringRef Constraint) const override;
852 /// Examine constraint string and operand type and determine a weight value.
853 /// The operand object must already have been set up with the operand type.
855 getSingleConstraintMatchWeight(AsmOperandInfo &info,
856 const char *constraint) const override;
858 const char *LowerXConstraint(EVT ConstraintVT) const override;
860 /// Lower the specified operand into the Ops vector. If it is invalid, don't
861 /// add anything to Ops. If hasMemory is true it means one of the asm
862 /// constraint of the inline asm instruction being processed is 'm'.
863 void LowerAsmOperandForConstraint(SDValue Op,
864 std::string &Constraint,
865 std::vector<SDValue> &Ops,
866 SelectionDAG &DAG) const override;
869 getInlineAsmMemConstraint(StringRef ConstraintCode) const override {
870 if (ConstraintCode == "i")
871 return InlineAsm::Constraint_i;
872 else if (ConstraintCode == "o")
873 return InlineAsm::Constraint_o;
874 else if (ConstraintCode == "v")
875 return InlineAsm::Constraint_v;
876 else if (ConstraintCode == "X")
877 return InlineAsm::Constraint_X;
878 return TargetLowering::getInlineAsmMemConstraint(ConstraintCode);
881 /// Given a physical register constraint
882 /// (e.g. {edx}), return the register number and the register class for the
883 /// register. This should only be used for C_Register constraints. On
884 /// error, this returns a register number of 0.
885 std::pair<unsigned, const TargetRegisterClass *>
886 getRegForInlineAsmConstraint(const TargetRegisterInfo *TRI,
887 StringRef Constraint, MVT VT) const override;
889 /// Return true if the addressing mode represented
890 /// by AM is legal for this target, for a load/store of the specified type.
891 bool isLegalAddressingMode(const DataLayout &DL, const AddrMode &AM,
892 Type *Ty, unsigned AS) const override;
894 /// Return true if the specified immediate is legal
895 /// icmp immediate, that is the target has icmp instructions which can
896 /// compare a register against the immediate without having to materialize
897 /// the immediate into a register.
898 bool isLegalICmpImmediate(int64_t Imm) const override;
900 /// Return true if the specified immediate is legal
901 /// add immediate, that is the target has add instructions which can
902 /// add a register and the immediate without having to materialize
903 /// the immediate into a register.
904 bool isLegalAddImmediate(int64_t Imm) const override;
906 /// \brief Return the cost of the scaling factor used in the addressing
907 /// mode represented by AM for this target, for a load/store
908 /// of the specified type.
909 /// If the AM is supported, the return value must be >= 0.
910 /// If the AM is not supported, it returns a negative value.
911 int getScalingFactorCost(const DataLayout &DL, const AddrMode &AM, Type *Ty,
912 unsigned AS) const override;
914 bool isVectorShiftByScalarCheap(Type *Ty) const override;
916 /// Return true if it's free to truncate a value of
917 /// type Ty1 to type Ty2. e.g. On x86 it's free to truncate a i32 value in
918 /// register EAX to i16 by referencing its sub-register AX.
919 bool isTruncateFree(Type *Ty1, Type *Ty2) const override;
920 bool isTruncateFree(EVT VT1, EVT VT2) const override;
922 bool allowTruncateForTailCall(Type *Ty1, Type *Ty2) const override;
924 /// Return true if any actual instruction that defines a
925 /// value of type Ty1 implicit zero-extends the value to Ty2 in the result
926 /// register. This does not necessarily include registers defined in
927 /// unknown ways, such as incoming arguments, or copies from unknown
928 /// virtual registers. Also, if isTruncateFree(Ty2, Ty1) is true, this
929 /// does not necessarily apply to truncate instructions. e.g. on x86-64,
930 /// all instructions that define 32-bit values implicit zero-extend the
931 /// result out to 64 bits.
932 bool isZExtFree(Type *Ty1, Type *Ty2) const override;
933 bool isZExtFree(EVT VT1, EVT VT2) const override;
934 bool isZExtFree(SDValue Val, EVT VT2) const override;
936 /// Return true if folding a vector load into ExtVal (a sign, zero, or any
937 /// extend node) is profitable.
938 bool isVectorLoadExtDesirable(SDValue) const override;
940 /// Return true if an FMA operation is faster than a pair of fmul and fadd
941 /// instructions. fmuladd intrinsics will be expanded to FMAs when this
942 /// method returns true, otherwise fmuladd is expanded to fmul + fadd.
943 bool isFMAFasterThanFMulAndFAdd(EVT VT) const override;
945 /// Return true if it's profitable to narrow
946 /// operations of type VT1 to VT2. e.g. on x86, it's profitable to narrow
947 /// from i32 to i8 but not from i32 to i16.
948 bool isNarrowingProfitable(EVT VT1, EVT VT2) const override;
950 /// Given an intrinsic, checks if on the target the intrinsic will need to map
951 /// to a MemIntrinsicNode (touches memory). If this is the case, it returns
952 /// true and stores the intrinsic information into the IntrinsicInfo that was
953 /// passed to the function.
954 bool getTgtMemIntrinsic(IntrinsicInfo &Info, const CallInst &I,
955 unsigned Intrinsic) const override;
957 /// Returns true if the target can instruction select the
958 /// specified FP immediate natively. If false, the legalizer will
959 /// materialize the FP immediate as a load from a constant pool.
960 bool isFPImmLegal(const APFloat &Imm, EVT VT) const override;
962 /// Targets can use this to indicate that they only support *some*
963 /// VECTOR_SHUFFLE operations, those with specific masks. By default, if a
964 /// target supports the VECTOR_SHUFFLE node, all mask values are assumed to
966 bool isShuffleMaskLegal(const SmallVectorImpl<int> &Mask,
967 EVT VT) const override;
969 /// Similar to isShuffleMaskLegal. This is used by Targets can use this to
970 /// indicate if there is a suitable VECTOR_SHUFFLE that can be used to
971 /// replace a VAND with a constant pool entry.
972 bool isVectorClearMaskLegal(const SmallVectorImpl<int> &Mask,
973 EVT VT) const override;
975 /// If true, then instruction selection should
976 /// seek to shrink the FP constant of the specified type to a smaller type
977 /// in order to save space and / or reduce runtime.
978 bool ShouldShrinkFPConstant(EVT VT) const override {
979 // Don't shrink FP constpool if SSE2 is available since cvtss2sd is more
980 // expensive than a straight movsd. On the other hand, it's important to
981 // shrink long double fp constant since fldt is very slow.
982 return !X86ScalarSSEf64 || VT == MVT::f80;
985 /// Return true if we believe it is correct and profitable to reduce the
986 /// load node to a smaller type.
987 bool shouldReduceLoadWidth(SDNode *Load, ISD::LoadExtType ExtTy,
988 EVT NewVT) const override;
990 /// Return true if the specified scalar FP type is computed in an SSE
991 /// register, not on the X87 floating point stack.
992 bool isScalarFPTypeInSSEReg(EVT VT) const {
993 return (VT == MVT::f64 && X86ScalarSSEf64) || // f64 is when SSE2
994 (VT == MVT::f32 && X86ScalarSSEf32); // f32 is when SSE1
997 /// \brief Returns true if it is beneficial to convert a load of a constant
998 /// to just the constant itself.
999 bool shouldConvertConstantLoadToIntImm(const APInt &Imm,
1000 Type *Ty) const override;
1002 bool convertSelectOfConstantsToMath() const override {
1006 /// Return true if EXTRACT_SUBVECTOR is cheap for this result type
1007 /// with this index.
1008 bool isExtractSubvectorCheap(EVT ResVT, unsigned Index) const override;
1010 /// Intel processors have a unified instruction and data cache
1011 const char * getClearCacheBuiltinName() const override {
1012 return nullptr; // nothing to do, move along.
1015 unsigned getRegisterByName(const char* RegName, EVT VT,
1016 SelectionDAG &DAG) const override;
1018 /// If a physical register, this returns the register that receives the
1019 /// exception address on entry to an EH pad.
1021 getExceptionPointerRegister(const Constant *PersonalityFn) const override;
1023 /// If a physical register, this returns the register that receives the
1024 /// exception typeid on entry to a landing pad.
1026 getExceptionSelectorRegister(const Constant *PersonalityFn) const override;
1028 virtual bool needsFixedCatchObjects() const override;
1030 /// This method returns a target specific FastISel object,
1031 /// or null if the target does not support "fast" ISel.
1032 FastISel *createFastISel(FunctionLoweringInfo &funcInfo,
1033 const TargetLibraryInfo *libInfo) const override;
1035 /// If the target has a standard location for the stack protector cookie,
1036 /// returns the address of that location. Otherwise, returns nullptr.
1037 Value *getIRStackGuard(IRBuilder<> &IRB) const override;
1039 bool useLoadStackGuardNode() const override;
1040 void insertSSPDeclarations(Module &M) const override;
1041 Value *getSDagStackGuard(const Module &M) const override;
1042 Value *getSSPStackGuardCheck(const Module &M) const override;
1044 /// Return true if the target stores SafeStack pointer at a fixed offset in
1045 /// some non-standard address space, and populates the address space and
1046 /// offset as appropriate.
1047 Value *getSafeStackPointerLocation(IRBuilder<> &IRB) const override;
1049 SDValue BuildFILD(SDValue Op, EVT SrcVT, SDValue Chain, SDValue StackSlot,
1050 SelectionDAG &DAG) const;
1052 bool isNoopAddrSpaceCast(unsigned SrcAS, unsigned DestAS) const override;
1054 /// \brief Customize the preferred legalization strategy for certain types.
1055 LegalizeTypeAction getPreferredVectorAction(EVT VT) const override;
1057 bool isIntDivCheap(EVT VT, AttributeList Attr) const override;
1059 bool supportSwiftError() const override;
1061 unsigned getMaxSupportedInterleaveFactor() const override { return 4; }
1063 /// \brief Lower interleaved load(s) into target specific
1064 /// instructions/intrinsics.
1065 bool lowerInterleavedLoad(LoadInst *LI,
1066 ArrayRef<ShuffleVectorInst *> Shuffles,
1067 ArrayRef<unsigned> Indices,
1068 unsigned Factor) const override;
1070 std::pair<const TargetRegisterClass *, uint8_t>
1071 findRepresentativeClass(const TargetRegisterInfo *TRI,
1072 MVT VT) const override;
1075 /// Keep a reference to the X86Subtarget around so that we can
1076 /// make the right decision when generating code for different targets.
1077 const X86Subtarget &Subtarget;
1079 /// Select between SSE or x87 floating point ops.
1080 /// When SSE is available, use it for f32 operations.
1081 /// When SSE2 is available, use it for f64 operations.
1082 bool X86ScalarSSEf32;
1083 bool X86ScalarSSEf64;
1085 /// A list of legal FP immediates.
1086 std::vector<APFloat> LegalFPImmediates;
1088 /// Indicate that this x86 target can instruction
1089 /// select the specified FP immediate natively.
1090 void addLegalFPImmediate(const APFloat& Imm) {
1091 LegalFPImmediates.push_back(Imm);
1094 SDValue LowerCallResult(SDValue Chain, SDValue InFlag,
1095 CallingConv::ID CallConv, bool isVarArg,
1096 const SmallVectorImpl<ISD::InputArg> &Ins,
1097 const SDLoc &dl, SelectionDAG &DAG,
1098 SmallVectorImpl<SDValue> &InVals,
1099 uint32_t *RegMask) const;
1100 SDValue LowerMemArgument(SDValue Chain, CallingConv::ID CallConv,
1101 const SmallVectorImpl<ISD::InputArg> &ArgInfo,
1102 const SDLoc &dl, SelectionDAG &DAG,
1103 const CCValAssign &VA, MachineFrameInfo &MFI,
1105 SDValue LowerMemOpCallTo(SDValue Chain, SDValue StackPtr, SDValue Arg,
1106 const SDLoc &dl, SelectionDAG &DAG,
1107 const CCValAssign &VA,
1108 ISD::ArgFlagsTy Flags) const;
1110 // Call lowering helpers.
1112 /// Check whether the call is eligible for tail call optimization. Targets
1113 /// that want to do tail call optimization should implement this function.
1114 bool IsEligibleForTailCallOptimization(SDValue Callee,
1115 CallingConv::ID CalleeCC,
1117 bool isCalleeStructRet,
1118 bool isCallerStructRet,
1120 const SmallVectorImpl<ISD::OutputArg> &Outs,
1121 const SmallVectorImpl<SDValue> &OutVals,
1122 const SmallVectorImpl<ISD::InputArg> &Ins,
1123 SelectionDAG& DAG) const;
1124 SDValue EmitTailCallLoadRetAddr(SelectionDAG &DAG, SDValue &OutRetAddr,
1125 SDValue Chain, bool IsTailCall,
1126 bool Is64Bit, int FPDiff,
1127 const SDLoc &dl) const;
1129 unsigned GetAlignedArgumentStackSize(unsigned StackSize,
1130 SelectionDAG &DAG) const;
1132 unsigned getAddressSpace(void) const;
1134 std::pair<SDValue,SDValue> FP_TO_INTHelper(SDValue Op, SelectionDAG &DAG,
1136 bool isReplace) const;
1138 SDValue LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) const;
1139 SDValue LowerBUILD_VECTORvXi1(SDValue Op, SelectionDAG &DAG) const;
1140 SDValue LowerVSELECT(SDValue Op, SelectionDAG &DAG) const;
1141 SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const;
1142 SDValue ExtractBitFromMaskVector(SDValue Op, SelectionDAG &DAG) const;
1143 SDValue InsertBitToMaskVector(SDValue Op, SelectionDAG &DAG) const;
1144 SDValue LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const;
1146 unsigned getGlobalWrapperKind(const GlobalValue *GV = nullptr) const;
1147 SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) const;
1148 SDValue LowerBlockAddress(SDValue Op, SelectionDAG &DAG) const;
1149 SDValue LowerGlobalAddress(const GlobalValue *GV, const SDLoc &dl,
1150 int64_t Offset, SelectionDAG &DAG) const;
1151 SDValue LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) const;
1152 SDValue LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const;
1153 SDValue LowerExternalSymbol(SDValue Op, SelectionDAG &DAG) const;
1155 SDValue LowerSINT_TO_FP(SDValue Op, SelectionDAG &DAG) const;
1156 SDValue LowerUINT_TO_FP(SDValue Op, SelectionDAG &DAG) const;
1157 SDValue LowerUINT_TO_FP_i64(SDValue Op, SelectionDAG &DAG) const;
1158 SDValue LowerUINT_TO_FP_i32(SDValue Op, SelectionDAG &DAG) const;
1159 SDValue lowerUINT_TO_FP_vec(SDValue Op, SelectionDAG &DAG) const;
1160 SDValue LowerTRUNCATE(SDValue Op, SelectionDAG &DAG) const;
1161 SDValue LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG) const;
1162 SDValue LowerToBT(SDValue And, ISD::CondCode CC, const SDLoc &dl,
1163 SelectionDAG &DAG) const;
1164 SDValue LowerSETCC(SDValue Op, SelectionDAG &DAG) const;
1165 SDValue LowerSETCCE(SDValue Op, SelectionDAG &DAG) const;
1166 SDValue LowerSELECT(SDValue Op, SelectionDAG &DAG) const;
1167 SDValue LowerBRCOND(SDValue Op, SelectionDAG &DAG) const;
1168 SDValue LowerJumpTable(SDValue Op, SelectionDAG &DAG) const;
1169 SDValue LowerDYNAMIC_STACKALLOC(SDValue Op, SelectionDAG &DAG) const;
1170 SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG) const;
1171 SDValue LowerVAARG(SDValue Op, SelectionDAG &DAG) const;
1172 SDValue LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const;
1173 SDValue LowerADDROFRETURNADDR(SDValue Op, SelectionDAG &DAG) const;
1174 SDValue LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const;
1175 SDValue LowerFRAME_TO_ARGS_OFFSET(SDValue Op, SelectionDAG &DAG) const;
1176 SDValue LowerEH_RETURN(SDValue Op, SelectionDAG &DAG) const;
1177 SDValue lowerEH_SJLJ_SETJMP(SDValue Op, SelectionDAG &DAG) const;
1178 SDValue lowerEH_SJLJ_LONGJMP(SDValue Op, SelectionDAG &DAG) const;
1179 SDValue lowerEH_SJLJ_SETUP_DISPATCH(SDValue Op, SelectionDAG &DAG) const;
1180 SDValue LowerINIT_TRAMPOLINE(SDValue Op, SelectionDAG &DAG) const;
1181 SDValue LowerFLT_ROUNDS_(SDValue Op, SelectionDAG &DAG) const;
1182 SDValue LowerWin64_i128OP(SDValue Op, SelectionDAG &DAG) const;
1183 SDValue LowerGC_TRANSITION_START(SDValue Op, SelectionDAG &DAG) const;
1184 SDValue LowerGC_TRANSITION_END(SDValue Op, SelectionDAG &DAG) const;
1187 LowerFormalArguments(SDValue Chain, CallingConv::ID CallConv, bool isVarArg,
1188 const SmallVectorImpl<ISD::InputArg> &Ins,
1189 const SDLoc &dl, SelectionDAG &DAG,
1190 SmallVectorImpl<SDValue> &InVals) const override;
1191 SDValue LowerCall(CallLoweringInfo &CLI,
1192 SmallVectorImpl<SDValue> &InVals) const override;
1194 SDValue LowerReturn(SDValue Chain, CallingConv::ID CallConv, bool isVarArg,
1195 const SmallVectorImpl<ISD::OutputArg> &Outs,
1196 const SmallVectorImpl<SDValue> &OutVals,
1197 const SDLoc &dl, SelectionDAG &DAG) const override;
1199 bool supportSplitCSR(MachineFunction *MF) const override {
1200 return MF->getFunction()->getCallingConv() == CallingConv::CXX_FAST_TLS &&
1201 MF->getFunction()->hasFnAttribute(Attribute::NoUnwind);
1203 void initializeSplitCSR(MachineBasicBlock *Entry) const override;
1204 void insertCopiesSplitCSR(
1205 MachineBasicBlock *Entry,
1206 const SmallVectorImpl<MachineBasicBlock *> &Exits) const override;
1208 bool isUsedByReturnOnly(SDNode *N, SDValue &Chain) const override;
1210 bool mayBeEmittedAsTailCall(const CallInst *CI) const override;
1212 EVT getTypeForExtReturn(LLVMContext &Context, EVT VT,
1213 ISD::NodeType ExtendKind) const override;
1215 bool CanLowerReturn(CallingConv::ID CallConv, MachineFunction &MF,
1217 const SmallVectorImpl<ISD::OutputArg> &Outs,
1218 LLVMContext &Context) const override;
1220 const MCPhysReg *getScratchRegisters(CallingConv::ID CC) const override;
1222 TargetLoweringBase::AtomicExpansionKind
1223 shouldExpandAtomicLoadInIR(LoadInst *SI) const override;
1224 bool shouldExpandAtomicStoreInIR(StoreInst *SI) const override;
1225 TargetLoweringBase::AtomicExpansionKind
1226 shouldExpandAtomicRMWInIR(AtomicRMWInst *AI) const override;
1229 lowerIdempotentRMWIntoFencedLoad(AtomicRMWInst *AI) const override;
1231 bool needsCmpXchgNb(Type *MemType) const;
1233 void SetupEntryBlockForSjLj(MachineInstr &MI, MachineBasicBlock *MBB,
1234 MachineBasicBlock *DispatchBB, int FI) const;
1236 // Utility function to emit the low-level va_arg code for X86-64.
1238 EmitVAARG64WithCustomInserter(MachineInstr &MI,
1239 MachineBasicBlock *MBB) const;
1241 /// Utility function to emit the xmm reg save portion of va_start.
1243 EmitVAStartSaveXMMRegsWithCustomInserter(MachineInstr &BInstr,
1244 MachineBasicBlock *BB) const;
1246 MachineBasicBlock *EmitLoweredSelect(MachineInstr &I,
1247 MachineBasicBlock *BB) const;
1249 MachineBasicBlock *EmitLoweredAtomicFP(MachineInstr &I,
1250 MachineBasicBlock *BB) const;
1252 MachineBasicBlock *EmitLoweredCatchRet(MachineInstr &MI,
1253 MachineBasicBlock *BB) const;
1255 MachineBasicBlock *EmitLoweredCatchPad(MachineInstr &MI,
1256 MachineBasicBlock *BB) const;
1258 MachineBasicBlock *EmitLoweredSegAlloca(MachineInstr &MI,
1259 MachineBasicBlock *BB) const;
1261 MachineBasicBlock *EmitLoweredTLSAddr(MachineInstr &MI,
1262 MachineBasicBlock *BB) const;
1264 MachineBasicBlock *EmitLoweredTLSCall(MachineInstr &MI,
1265 MachineBasicBlock *BB) const;
1267 MachineBasicBlock *emitEHSjLjSetJmp(MachineInstr &MI,
1268 MachineBasicBlock *MBB) const;
1270 MachineBasicBlock *emitEHSjLjLongJmp(MachineInstr &MI,
1271 MachineBasicBlock *MBB) const;
1273 MachineBasicBlock *emitFMA3Instr(MachineInstr &MI,
1274 MachineBasicBlock *MBB) const;
1276 MachineBasicBlock *EmitSjLjDispatchBlock(MachineInstr &MI,
1277 MachineBasicBlock *MBB) const;
1279 /// Emit nodes that will be selected as "test Op0,Op0", or something
1280 /// equivalent, for use with the given x86 condition code.
1281 SDValue EmitTest(SDValue Op0, unsigned X86CC, const SDLoc &dl,
1282 SelectionDAG &DAG) const;
1284 /// Emit nodes that will be selected as "cmp Op0,Op1", or something
1285 /// equivalent, for use with the given x86 condition code.
1286 SDValue EmitCmp(SDValue Op0, SDValue Op1, unsigned X86CC, const SDLoc &dl,
1287 SelectionDAG &DAG) const;
1289 /// Convert a comparison if required by the subtarget.
1290 SDValue ConvertCmpIfNecessary(SDValue Cmp, SelectionDAG &DAG) const;
1292 /// Check if replacement of SQRT with RSQRT should be disabled.
1293 bool isFsqrtCheap(SDValue Operand, SelectionDAG &DAG) const override;
1295 /// Use rsqrt* to speed up sqrt calculations.
1296 SDValue getSqrtEstimate(SDValue Operand, SelectionDAG &DAG, int Enabled,
1297 int &RefinementSteps, bool &UseOneConstNR,
1298 bool Reciprocal) const override;
1300 /// Use rcp* to speed up fdiv calculations.
1301 SDValue getRecipEstimate(SDValue Operand, SelectionDAG &DAG, int Enabled,
1302 int &RefinementSteps) const override;
1304 /// Reassociate floating point divisions into multiply by reciprocal.
1305 unsigned combineRepeatedFPDivisors() const override;
1309 FastISel *createFastISel(FunctionLoweringInfo &funcInfo,
1310 const TargetLibraryInfo *libInfo);
1311 } // end namespace X86
1313 // Base class for all X86 non-masked store operations.
1314 class X86StoreSDNode : public MemSDNode {
1316 X86StoreSDNode(unsigned Opcode, unsigned Order, const DebugLoc &dl,
1317 SDVTList VTs, EVT MemVT,
1318 MachineMemOperand *MMO)
1319 :MemSDNode(Opcode, Order, dl, VTs, MemVT, MMO) {}
1320 const SDValue &getValue() const { return getOperand(1); }
1321 const SDValue &getBasePtr() const { return getOperand(2); }
1323 static bool classof(const SDNode *N) {
1324 return N->getOpcode() == X86ISD::VTRUNCSTORES ||
1325 N->getOpcode() == X86ISD::VTRUNCSTOREUS;
1329 // Base class for all X86 masked store operations.
1330 // The class has the same order of operands as MaskedStoreSDNode for
1332 class X86MaskedStoreSDNode : public MemSDNode {
1334 X86MaskedStoreSDNode(unsigned Opcode, unsigned Order,
1335 const DebugLoc &dl, SDVTList VTs, EVT MemVT,
1336 MachineMemOperand *MMO)
1337 : MemSDNode(Opcode, Order, dl, VTs, MemVT, MMO) {}
1339 const SDValue &getBasePtr() const { return getOperand(1); }
1340 const SDValue &getMask() const { return getOperand(2); }
1341 const SDValue &getValue() const { return getOperand(3); }
1343 static bool classof(const SDNode *N) {
1344 return N->getOpcode() == X86ISD::VMTRUNCSTORES ||
1345 N->getOpcode() == X86ISD::VMTRUNCSTOREUS;
1349 // X86 Truncating Store with Signed saturation.
1350 class TruncSStoreSDNode : public X86StoreSDNode {
1352 TruncSStoreSDNode(unsigned Order, const DebugLoc &dl,
1353 SDVTList VTs, EVT MemVT, MachineMemOperand *MMO)
1354 : X86StoreSDNode(X86ISD::VTRUNCSTORES, Order, dl, VTs, MemVT, MMO) {}
1356 static bool classof(const SDNode *N) {
1357 return N->getOpcode() == X86ISD::VTRUNCSTORES;
1361 // X86 Truncating Store with Unsigned saturation.
1362 class TruncUSStoreSDNode : public X86StoreSDNode {
1364 TruncUSStoreSDNode(unsigned Order, const DebugLoc &dl,
1365 SDVTList VTs, EVT MemVT, MachineMemOperand *MMO)
1366 : X86StoreSDNode(X86ISD::VTRUNCSTOREUS, Order, dl, VTs, MemVT, MMO) {}
1368 static bool classof(const SDNode *N) {
1369 return N->getOpcode() == X86ISD::VTRUNCSTOREUS;
1373 // X86 Truncating Masked Store with Signed saturation.
1374 class MaskedTruncSStoreSDNode : public X86MaskedStoreSDNode {
1376 MaskedTruncSStoreSDNode(unsigned Order,
1377 const DebugLoc &dl, SDVTList VTs, EVT MemVT,
1378 MachineMemOperand *MMO)
1379 : X86MaskedStoreSDNode(X86ISD::VMTRUNCSTORES, Order, dl, VTs, MemVT, MMO) {}
1381 static bool classof(const SDNode *N) {
1382 return N->getOpcode() == X86ISD::VMTRUNCSTORES;
1386 // X86 Truncating Masked Store with Unsigned saturation.
1387 class MaskedTruncUSStoreSDNode : public X86MaskedStoreSDNode {
1389 MaskedTruncUSStoreSDNode(unsigned Order,
1390 const DebugLoc &dl, SDVTList VTs, EVT MemVT,
1391 MachineMemOperand *MMO)
1392 : X86MaskedStoreSDNode(X86ISD::VMTRUNCSTOREUS, Order, dl, VTs, MemVT, MMO) {}
1394 static bool classof(const SDNode *N) {
1395 return N->getOpcode() == X86ISD::VMTRUNCSTOREUS;
1399 } // end namespace llvm
1401 #endif // LLVM_LIB_TARGET_X86_X86ISELLOWERING_H