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. If you think this is too close to the previous
153 /// mnemonic, so do I; blame Intel.
156 /// Copies a 32-bit value from the low word of a MMX
160 /// Copies a GPR into the low 32-bit word of a MMX vector
161 /// and zero out the high word.
164 /// Extract an 8-bit value from a vector and zero extend it to
165 /// i32, corresponds to X86::PEXTRB.
168 /// Extract a 16-bit value from a vector and zero extend it to
169 /// i32, corresponds to X86::PEXTRW.
172 /// Insert any element of a 4 x float vector into any element
173 /// of a destination 4 x floatvector.
176 /// Insert the lower 8-bits of a 32-bit value to a vector,
177 /// corresponds to X86::PINSRB.
180 /// Insert the lower 16-bits of a 32-bit value to a vector,
181 /// corresponds to X86::PINSRW.
184 /// Shuffle 16 8-bit values within a vector.
187 /// Compute Sum of Absolute Differences.
189 /// Compute Double Block Packed Sum-Absolute-Differences
192 /// Bitwise Logical AND NOT of Packed FP values.
195 /// Blend where the selector is an immediate.
198 /// Blend where the condition has been shrunk.
199 /// This is used to emphasize that the condition mask is
200 /// no more valid for generic VSELECT optimizations.
203 /// Combined add and sub on an FP vector.
206 // FP vector ops with rounding mode.
213 FSQRT_RND, FSQRTS_RND,
215 // FP vector get exponent.
216 FGETEXP_RND, FGETEXPS_RND,
217 // Extract Normalized Mantissas.
223 // Integer add/sub with unsigned saturation.
227 // Integer add/sub with signed saturation.
231 // Unsigned Integer average.
234 /// Integer horizontal add/sub.
238 /// Floating point horizontal add/sub.
242 // Integer absolute value
245 // Detect Conflicts Within a Vector
248 /// Floating point max and min.
251 /// Commutative FMIN and FMAX.
254 /// Floating point reciprocal-sqrt and reciprocal approximation.
255 /// Note that these typically require refinement
256 /// in order to obtain suitable precision.
260 // Thread Local Storage.
263 // Thread Local Storage. A call to get the start address
264 // of the TLS block for the current module.
267 // Thread Local Storage. When calling to an OS provided
268 // thunk at the address from an earlier relocation.
271 // Exception Handling helpers.
274 // SjLj exception handling setjmp.
277 // SjLj exception handling longjmp.
280 // SjLj exception handling dispatch.
281 EH_SJLJ_SETUP_DISPATCH,
283 /// Tail call return. See X86TargetLowering::LowerCall for
284 /// the list of operands.
287 // Vector move to low scalar and zero higher vector elements.
290 // Vector integer zero-extend.
292 // Vector integer signed-extend.
295 // Vector integer truncate.
297 // Vector integer truncate with unsigned/signed saturation.
301 VFPEXT, VFPEXT_RND, VFPEXTS_RND,
304 VFPROUND, VFPROUND_RND, VFPROUNDS_RND,
306 // Convert a vector to mask, set bits base on MSB.
309 // 128-bit vector logical left / right shift
312 // Vector shift elements
315 // Vector variable shift right arithmetic.
316 // Unlike ISD::SRA, in case shift count greater then element size
317 // use sign bit to fill destination data element.
320 // Vector shift elements by immediate
323 // Bit rotate by immediate
326 // Vector packed double/float comparison.
329 // Vector integer comparisons.
331 // Vector integer comparisons, the result is in a mask vector.
336 /// Vector comparison generating mask bits for fp and
337 /// integer signed and unsigned data types.
340 // Vector comparison with rounding mode for FP values
343 // Arithmetic operations with FLAGS results.
344 ADD, SUB, ADC, SBB, SMUL,
345 INC, DEC, OR, XOR, AND,
347 // Bit field extract.
350 // LOW, HI, FLAGS = umul LHS, RHS.
353 // 8-bit SMUL/UMUL - AX, FLAGS = smul8/umul8 AL, RHS.
356 // 8-bit divrem that zero-extend the high result (AH).
360 // X86-specific multiply by immediate.
363 // Vector sign bit extraction.
366 // Vector bitwise comparisons.
369 // Vector packed fp sign bitwise comparisons.
372 // Vector "test" in AVX-512, the result is in a mask vector.
376 // OR/AND test for masks.
380 // Several flavors of instructions with vector shuffle behaviors.
381 // Saturated signed/unnsigned packing.
384 // Intra-lane alignr.
386 // AVX512 inter-lane alignr.
392 //Shuffle Packed Values at 128-bit granularity.
411 // Variable Permute (VPERM).
412 // Res = VPERMV MaskV, V0
415 // 3-op Variable Permute (VPERMT2).
416 // Res = VPERMV3 V0, MaskV, V1
419 // 3-op Variable Permute overwriting the index (VPERMI2).
420 // Res = VPERMIV3 V0, MaskV, V1
423 // Bitwise ternary logic.
425 // Fix Up Special Packed Float32/64 values.
428 // Range Restriction Calculation For Packed Pairs of Float32/64 values.
430 // Reduce - Perform Reduction Transformation on scalar\packed FP.
432 // RndScale - Round FP Values To Include A Given Number Of Fraction Bits.
433 VRNDSCALE, VRNDSCALES,
434 // Tests Types Of a FP Values for packed types.
436 // Tests Types Of a FP Values for scalar types.
439 // Broadcast scalar to vector.
441 // Broadcast mask to vector.
443 // Broadcast subvector to vector.
446 // Insert/Extract vector element.
450 /// SSE4A Extraction and Insertion.
453 // XOP variable/immediate rotations.
455 // XOP arithmetic/logical shifts.
457 // XOP signed/unsigned integer comparisons.
459 // XOP packed permute bytes.
461 // XOP two source permutation.
464 // Vector multiply packed unsigned doubleword integers.
466 // Vector multiply packed signed doubleword integers.
468 // Vector Multiply Packed UnsignedIntegers with Round and Scale.
471 // Multiply and Add Packed Integers.
472 VPMADDUBSW, VPMADDWD,
473 VPMADD52L, VPMADD52H,
483 // FMA with rounding mode.
491 // Scalar intrinsic FMA with rounding mode.
492 // Two versions, passthru bits on op1 or op3.
493 FMADDS1_RND, FMADDS3_RND,
494 FNMADDS1_RND, FNMADDS3_RND,
495 FMSUBS1_RND, FMSUBS3_RND,
496 FNMSUBS1_RND, FNMSUBS3_RND,
498 // Compress and expand.
502 // Convert Unsigned/Integer to Floating-Point Value with rounding mode.
503 SINT_TO_FP_RND, UINT_TO_FP_RND,
504 SCALAR_SINT_TO_FP_RND, SCALAR_UINT_TO_FP_RND,
506 // Vector float/double to signed/unsigned integer.
507 CVTP2SI, CVTP2UI, CVTP2SI_RND, CVTP2UI_RND,
508 // Scalar float/double to signed/unsigned integer.
509 CVTS2SI_RND, CVTS2UI_RND,
511 // Vector float/double to signed/unsigned integer with truncation.
512 CVTTP2SI, CVTTP2UI, CVTTP2SI_RND, CVTTP2UI_RND,
513 // Scalar float/double to signed/unsigned integer with truncation.
514 CVTTS2SI_RND, CVTTS2UI_RND,
516 // Vector signed/unsigned integer to float/double.
519 // Save xmm argument registers to the stack, according to %al. An operator
520 // is needed so that this can be expanded with control flow.
521 VASTART_SAVE_XMM_REGS,
523 // Windows's _chkstk call to do stack probing.
526 // For allocating variable amounts of stack space when using
527 // segmented stacks. Check if the current stacklet has enough space, and
528 // falls back to heap allocation if not.
535 // Store FP status word into i16 register.
538 // Store contents of %ah into %eflags.
541 // Get a random integer and indicate whether it is valid in CF.
544 // Get a NIST SP800-90B & C compliant random integer and
545 // indicate whether it is valid in CF.
548 // SSE42 string comparisons.
552 // Test if in transactional execution.
556 RSQRT28, RSQRT28S, RCP28, RCP28S, EXP2,
558 // Conversions between float and half-float.
562 LCMPXCHG_DAG = ISD::FIRST_TARGET_MEMORY_OPCODE,
565 LCMPXCHG8_SAVE_EBX_DAG,
566 LCMPXCHG16_SAVE_RBX_DAG,
568 /// LOCK-prefixed arithmetic read-modify-write instructions.
569 /// EFLAGS, OUTCHAIN = LADD(INCHAIN, PTR, RHS)
570 LADD, LSUB, LOR, LXOR, LAND,
572 // Load, scalar_to_vector, and zero extend.
575 // Store FP control world into i16 memory.
578 /// This instruction implements FP_TO_SINT with the
579 /// integer destination in memory and a FP reg source. This corresponds
580 /// to the X86::FIST*m instructions and the rounding mode change stuff. It
581 /// has two inputs (token chain and address) and two outputs (int value
582 /// and token chain).
587 /// This instruction implements SINT_TO_FP with the
588 /// integer source in memory and FP reg result. This corresponds to the
589 /// X86::FILD*m instructions. It has three inputs (token chain, address,
590 /// and source type) and two outputs (FP value and token chain). FILD_FLAG
591 /// also produces a flag).
595 /// This instruction implements an extending load to FP stack slots.
596 /// This corresponds to the X86::FLD32m / X86::FLD64m. It takes a chain
597 /// operand, ptr to load from, and a ValueType node indicating the type
601 /// This instruction implements a truncating store to FP stack
602 /// slots. This corresponds to the X86::FST32m / X86::FST64m. It takes a
603 /// chain operand, value to store, address, and a ValueType to store it
607 /// This instruction grabs the address of the next argument
608 /// from a va_list. (reads and modifies the va_list in memory)
611 // Vector truncating store with unsigned/signed saturation
612 VTRUNCSTOREUS, VTRUNCSTORES,
613 // Vector truncating masked store with unsigned/signed saturation
614 VMTRUNCSTOREUS, VMTRUNCSTORES
616 // WARNING: Do not add anything in the end unless you want the node to
617 // have memop! In fact, starting from FIRST_TARGET_MEMORY_OPCODE all
618 // opcodes will be thought as target memory ops!
620 } // end namespace X86ISD
622 /// Define some predicates that are used for node matching.
624 /// Return true if the specified
625 /// EXTRACT_SUBVECTOR operand specifies a vector extract that is
626 /// suitable for input to VEXTRACTF128, VEXTRACTI128 instructions.
627 bool isVEXTRACT128Index(SDNode *N);
629 /// Return true if the specified
630 /// INSERT_SUBVECTOR operand specifies a subvector insert that is
631 /// suitable for input to VINSERTF128, VINSERTI128 instructions.
632 bool isVINSERT128Index(SDNode *N);
634 /// Return true if the specified
635 /// EXTRACT_SUBVECTOR operand specifies a vector extract that is
636 /// suitable for input to VEXTRACTF64X4, VEXTRACTI64X4 instructions.
637 bool isVEXTRACT256Index(SDNode *N);
639 /// Return true if the specified
640 /// INSERT_SUBVECTOR operand specifies a subvector insert that is
641 /// suitable for input to VINSERTF64X4, VINSERTI64X4 instructions.
642 bool isVINSERT256Index(SDNode *N);
644 /// Return the appropriate
645 /// immediate to extract the specified EXTRACT_SUBVECTOR index
646 /// with VEXTRACTF128, VEXTRACTI128 instructions.
647 unsigned getExtractVEXTRACT128Immediate(SDNode *N);
649 /// Return the appropriate
650 /// immediate to insert at the specified INSERT_SUBVECTOR index
651 /// with VINSERTF128, VINSERT128 instructions.
652 unsigned getInsertVINSERT128Immediate(SDNode *N);
654 /// Return the appropriate
655 /// immediate to extract the specified EXTRACT_SUBVECTOR index
656 /// with VEXTRACTF64X4, VEXTRACTI64x4 instructions.
657 unsigned getExtractVEXTRACT256Immediate(SDNode *N);
659 /// Return the appropriate
660 /// immediate to insert at the specified INSERT_SUBVECTOR index
661 /// with VINSERTF64x4, VINSERTI64x4 instructions.
662 unsigned getInsertVINSERT256Immediate(SDNode *N);
664 /// Returns true if Elt is a constant zero or floating point constant +0.0.
665 bool isZeroNode(SDValue Elt);
667 /// Returns true of the given offset can be
668 /// fit into displacement field of the instruction.
669 bool isOffsetSuitableForCodeModel(int64_t Offset, CodeModel::Model M,
670 bool hasSymbolicDisplacement = true);
672 /// Determines whether the callee is required to pop its
673 /// own arguments. Callee pop is necessary to support tail calls.
674 bool isCalleePop(CallingConv::ID CallingConv,
675 bool is64Bit, bool IsVarArg, bool GuaranteeTCO);
677 } // end namespace X86
679 //===--------------------------------------------------------------------===//
680 // X86 Implementation of the TargetLowering interface
681 class X86TargetLowering final : public TargetLowering {
683 explicit X86TargetLowering(const X86TargetMachine &TM,
684 const X86Subtarget &STI);
686 unsigned getJumpTableEncoding() const override;
687 bool useSoftFloat() const override;
689 MVT getScalarShiftAmountTy(const DataLayout &, EVT) const override {
694 LowerCustomJumpTableEntry(const MachineJumpTableInfo *MJTI,
695 const MachineBasicBlock *MBB, unsigned uid,
696 MCContext &Ctx) const override;
698 /// Returns relocation base for the given PIC jumptable.
699 SDValue getPICJumpTableRelocBase(SDValue Table,
700 SelectionDAG &DAG) const override;
702 getPICJumpTableRelocBaseExpr(const MachineFunction *MF,
703 unsigned JTI, MCContext &Ctx) const override;
705 /// Return the desired alignment for ByVal aggregate
706 /// function arguments in the caller parameter area. For X86, aggregates
707 /// that contains are placed at 16-byte boundaries while the rest are at
708 /// 4-byte boundaries.
709 unsigned getByValTypeAlignment(Type *Ty,
710 const DataLayout &DL) const override;
712 /// Returns the target specific optimal type for load
713 /// and store operations as a result of memset, memcpy, and memmove
714 /// lowering. If DstAlign is zero that means it's safe to destination
715 /// alignment can satisfy any constraint. Similarly if SrcAlign is zero it
716 /// means there isn't a need to check it against alignment requirement,
717 /// probably because the source does not need to be loaded. If 'IsMemset' is
718 /// true, that means it's expanding a memset. If 'ZeroMemset' is true, that
719 /// means it's a memset of zero. 'MemcpyStrSrc' indicates whether the memcpy
720 /// source is constant so it does not need to be loaded.
721 /// It returns EVT::Other if the type should be determined using generic
722 /// target-independent logic.
723 EVT getOptimalMemOpType(uint64_t Size, unsigned DstAlign, unsigned SrcAlign,
724 bool IsMemset, bool ZeroMemset, bool MemcpyStrSrc,
725 MachineFunction &MF) const override;
727 /// Returns true if it's safe to use load / store of the
728 /// specified type to expand memcpy / memset inline. This is mostly true
729 /// for all types except for some special cases. For example, on X86
730 /// targets without SSE2 f64 load / store are done with fldl / fstpl which
731 /// also does type conversion. Note the specified type doesn't have to be
732 /// legal as the hook is used before type legalization.
733 bool isSafeMemOpType(MVT VT) const override;
735 /// Returns true if the target allows unaligned memory accesses of the
736 /// specified type. Returns whether it is "fast" in the last argument.
737 bool allowsMisalignedMemoryAccesses(EVT VT, unsigned AS, unsigned Align,
738 bool *Fast) const override;
740 /// Provide custom lowering hooks for some operations.
742 SDValue LowerOperation(SDValue Op, SelectionDAG &DAG) const override;
744 /// Places new result values for the node in Results (their number
745 /// and types must exactly match those of the original return values of
746 /// the node), or leaves Results empty, which indicates that the node is not
747 /// to be custom lowered after all.
748 void LowerOperationWrapper(SDNode *N,
749 SmallVectorImpl<SDValue> &Results,
750 SelectionDAG &DAG) const override;
752 /// Replace the results of node with an illegal result
753 /// type with new values built out of custom code.
755 void ReplaceNodeResults(SDNode *N, SmallVectorImpl<SDValue>&Results,
756 SelectionDAG &DAG) const override;
758 SDValue PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const override;
760 /// Return true if the target has native support for
761 /// the specified value type and it is 'desirable' to use the type for the
762 /// given node type. e.g. On x86 i16 is legal, but undesirable since i16
763 /// instruction encodings are longer and some i16 instructions are slow.
764 bool isTypeDesirableForOp(unsigned Opc, EVT VT) const override;
766 /// Return true if the target has native support for the
767 /// specified value type and it is 'desirable' to use the type. e.g. On x86
768 /// i16 is legal, but undesirable since i16 instruction encodings are longer
769 /// and some i16 instructions are slow.
770 bool IsDesirableToPromoteOp(SDValue Op, EVT &PVT) const override;
772 /// Return true if the MachineFunction contains a COPY which would imply
773 /// HasOpaqueSPAdjustment.
774 bool hasCopyImplyingStackAdjustment(MachineFunction *MF) const override;
777 EmitInstrWithCustomInserter(MachineInstr &MI,
778 MachineBasicBlock *MBB) const override;
780 /// This method returns the name of a target specific DAG node.
781 const char *getTargetNodeName(unsigned Opcode) const override;
783 bool isCheapToSpeculateCttz() const override;
785 bool isCheapToSpeculateCtlz() const override;
787 bool isCtlzFast() const override;
789 bool hasBitPreservingFPLogic(EVT VT) const override {
790 return VT == MVT::f32 || VT == MVT::f64 || VT.isVector();
793 bool isMultiStoresCheaperThanBitsMerge(EVT LTy, EVT HTy) const override {
794 // If the pair to store is a mixture of float and int values, we will
795 // save two bitwise instructions and one float-to-int instruction and
796 // increase one store instruction. There is potentially a more
797 // significant benefit because it avoids the float->int domain switch
798 // for input value. So It is more likely a win.
799 if ((LTy.isFloatingPoint() && HTy.isInteger()) ||
800 (LTy.isInteger() && HTy.isFloatingPoint()))
802 // If the pair only contains int values, we will save two bitwise
803 // instructions and increase one store instruction (costing one more
804 // store buffer). Since the benefit is more blurred so we leave
805 // such pair out until we get testcase to prove it is a win.
809 bool hasAndNotCompare(SDValue Y) const override;
811 /// Return the value type to use for ISD::SETCC.
812 EVT getSetCCResultType(const DataLayout &DL, LLVMContext &Context,
813 EVT VT) const override;
815 /// Determine which of the bits specified in Mask are known to be either
816 /// zero or one and return them in the KnownZero/KnownOne bitsets.
817 void computeKnownBitsForTargetNode(const SDValue Op,
820 const SelectionDAG &DAG,
821 unsigned Depth = 0) const override;
823 /// Determine the number of bits in the operation that are sign bits.
824 unsigned ComputeNumSignBitsForTargetNode(SDValue Op,
825 const SelectionDAG &DAG,
826 unsigned Depth) const override;
828 bool isGAPlusOffset(SDNode *N, const GlobalValue* &GA,
829 int64_t &Offset) const override;
831 SDValue getReturnAddressFrameIndex(SelectionDAG &DAG) const;
833 bool ExpandInlineAsm(CallInst *CI) const override;
835 ConstraintType getConstraintType(StringRef Constraint) const override;
837 /// Examine constraint string and operand type and determine a weight value.
838 /// The operand object must already have been set up with the operand type.
840 getSingleConstraintMatchWeight(AsmOperandInfo &info,
841 const char *constraint) const override;
843 const char *LowerXConstraint(EVT ConstraintVT) const override;
845 /// Lower the specified operand into the Ops vector. If it is invalid, don't
846 /// add anything to Ops. If hasMemory is true it means one of the asm
847 /// constraint of the inline asm instruction being processed is 'm'.
848 void LowerAsmOperandForConstraint(SDValue Op,
849 std::string &Constraint,
850 std::vector<SDValue> &Ops,
851 SelectionDAG &DAG) const override;
854 getInlineAsmMemConstraint(StringRef ConstraintCode) const override {
855 if (ConstraintCode == "i")
856 return InlineAsm::Constraint_i;
857 else if (ConstraintCode == "o")
858 return InlineAsm::Constraint_o;
859 else if (ConstraintCode == "v")
860 return InlineAsm::Constraint_v;
861 else if (ConstraintCode == "X")
862 return InlineAsm::Constraint_X;
863 return TargetLowering::getInlineAsmMemConstraint(ConstraintCode);
866 /// Given a physical register constraint
867 /// (e.g. {edx}), return the register number and the register class for the
868 /// register. This should only be used for C_Register constraints. On
869 /// error, this returns a register number of 0.
870 std::pair<unsigned, const TargetRegisterClass *>
871 getRegForInlineAsmConstraint(const TargetRegisterInfo *TRI,
872 StringRef Constraint, MVT VT) const override;
874 /// Return true if the addressing mode represented
875 /// by AM is legal for this target, for a load/store of the specified type.
876 bool isLegalAddressingMode(const DataLayout &DL, const AddrMode &AM,
877 Type *Ty, unsigned AS) const override;
879 /// Return true if the specified immediate is legal
880 /// icmp immediate, that is the target has icmp instructions which can
881 /// compare a register against the immediate without having to materialize
882 /// the immediate into a register.
883 bool isLegalICmpImmediate(int64_t Imm) const override;
885 /// Return true if the specified immediate is legal
886 /// add immediate, that is the target has add instructions which can
887 /// add a register and the immediate without having to materialize
888 /// the immediate into a register.
889 bool isLegalAddImmediate(int64_t Imm) const override;
891 /// \brief Return the cost of the scaling factor used in the addressing
892 /// mode represented by AM for this target, for a load/store
893 /// of the specified type.
894 /// If the AM is supported, the return value must be >= 0.
895 /// If the AM is not supported, it returns a negative value.
896 int getScalingFactorCost(const DataLayout &DL, const AddrMode &AM, Type *Ty,
897 unsigned AS) const override;
899 bool isVectorShiftByScalarCheap(Type *Ty) const override;
901 /// Return true if it's free to truncate a value of
902 /// type Ty1 to type Ty2. e.g. On x86 it's free to truncate a i32 value in
903 /// register EAX to i16 by referencing its sub-register AX.
904 bool isTruncateFree(Type *Ty1, Type *Ty2) const override;
905 bool isTruncateFree(EVT VT1, EVT VT2) const override;
907 bool allowTruncateForTailCall(Type *Ty1, Type *Ty2) const override;
909 /// Return true if any actual instruction that defines a
910 /// value of type Ty1 implicit zero-extends the value to Ty2 in the result
911 /// register. This does not necessarily include registers defined in
912 /// unknown ways, such as incoming arguments, or copies from unknown
913 /// virtual registers. Also, if isTruncateFree(Ty2, Ty1) is true, this
914 /// does not necessarily apply to truncate instructions. e.g. on x86-64,
915 /// all instructions that define 32-bit values implicit zero-extend the
916 /// result out to 64 bits.
917 bool isZExtFree(Type *Ty1, Type *Ty2) const override;
918 bool isZExtFree(EVT VT1, EVT VT2) const override;
919 bool isZExtFree(SDValue Val, EVT VT2) const override;
921 /// Return true if folding a vector load into ExtVal (a sign, zero, or any
922 /// extend node) is profitable.
923 bool isVectorLoadExtDesirable(SDValue) const override;
925 /// Return true if an FMA operation is faster than a pair of fmul and fadd
926 /// instructions. fmuladd intrinsics will be expanded to FMAs when this
927 /// method returns true, otherwise fmuladd is expanded to fmul + fadd.
928 bool isFMAFasterThanFMulAndFAdd(EVT VT) const override;
930 /// Return true if it's profitable to narrow
931 /// operations of type VT1 to VT2. e.g. on x86, it's profitable to narrow
932 /// from i32 to i8 but not from i32 to i16.
933 bool isNarrowingProfitable(EVT VT1, EVT VT2) const override;
935 /// Given an intrinsic, checks if on the target the intrinsic will need to map
936 /// to a MemIntrinsicNode (touches memory). If this is the case, it returns
937 /// true and stores the intrinsic information into the IntrinsicInfo that was
938 /// passed to the function.
939 bool getTgtMemIntrinsic(IntrinsicInfo &Info, const CallInst &I,
940 unsigned Intrinsic) const override;
942 /// Returns true if the target can instruction select the
943 /// specified FP immediate natively. If false, the legalizer will
944 /// materialize the FP immediate as a load from a constant pool.
945 bool isFPImmLegal(const APFloat &Imm, EVT VT) const override;
947 /// Targets can use this to indicate that they only support *some*
948 /// VECTOR_SHUFFLE operations, those with specific masks. By default, if a
949 /// target supports the VECTOR_SHUFFLE node, all mask values are assumed to
951 bool isShuffleMaskLegal(const SmallVectorImpl<int> &Mask,
952 EVT VT) const override;
954 /// Similar to isShuffleMaskLegal. This is used by Targets can use this to
955 /// indicate if there is a suitable VECTOR_SHUFFLE that can be used to
956 /// replace a VAND with a constant pool entry.
957 bool isVectorClearMaskLegal(const SmallVectorImpl<int> &Mask,
958 EVT VT) const override;
960 /// If true, then instruction selection should
961 /// seek to shrink the FP constant of the specified type to a smaller type
962 /// in order to save space and / or reduce runtime.
963 bool ShouldShrinkFPConstant(EVT VT) const override {
964 // Don't shrink FP constpool if SSE2 is available since cvtss2sd is more
965 // expensive than a straight movsd. On the other hand, it's important to
966 // shrink long double fp constant since fldt is very slow.
967 return !X86ScalarSSEf64 || VT == MVT::f80;
970 /// Return true if we believe it is correct and profitable to reduce the
971 /// load node to a smaller type.
972 bool shouldReduceLoadWidth(SDNode *Load, ISD::LoadExtType ExtTy,
973 EVT NewVT) const override;
975 /// Return true if the specified scalar FP type is computed in an SSE
976 /// register, not on the X87 floating point stack.
977 bool isScalarFPTypeInSSEReg(EVT VT) const {
978 return (VT == MVT::f64 && X86ScalarSSEf64) || // f64 is when SSE2
979 (VT == MVT::f32 && X86ScalarSSEf32); // f32 is when SSE1
982 /// \brief Returns true if it is beneficial to convert a load of a constant
983 /// to just the constant itself.
984 bool shouldConvertConstantLoadToIntImm(const APInt &Imm,
985 Type *Ty) const override;
987 /// Return true if EXTRACT_SUBVECTOR is cheap for this result type
989 bool isExtractSubvectorCheap(EVT ResVT, unsigned Index) const override;
991 /// Intel processors have a unified instruction and data cache
992 const char * getClearCacheBuiltinName() const override {
993 return nullptr; // nothing to do, move along.
996 unsigned getRegisterByName(const char* RegName, EVT VT,
997 SelectionDAG &DAG) const override;
999 /// If a physical register, this returns the register that receives the
1000 /// exception address on entry to an EH pad.
1002 getExceptionPointerRegister(const Constant *PersonalityFn) const override;
1004 /// If a physical register, this returns the register that receives the
1005 /// exception typeid on entry to a landing pad.
1007 getExceptionSelectorRegister(const Constant *PersonalityFn) const override;
1009 virtual bool needsFixedCatchObjects() const override;
1011 /// This method returns a target specific FastISel object,
1012 /// or null if the target does not support "fast" ISel.
1013 FastISel *createFastISel(FunctionLoweringInfo &funcInfo,
1014 const TargetLibraryInfo *libInfo) const override;
1016 /// If the target has a standard location for the stack protector cookie,
1017 /// returns the address of that location. Otherwise, returns nullptr.
1018 Value *getIRStackGuard(IRBuilder<> &IRB) const override;
1020 bool useLoadStackGuardNode() const override;
1021 void insertSSPDeclarations(Module &M) const override;
1022 Value *getSDagStackGuard(const Module &M) const override;
1023 Value *getSSPStackGuardCheck(const Module &M) const override;
1025 /// Return true if the target stores SafeStack pointer at a fixed offset in
1026 /// some non-standard address space, and populates the address space and
1027 /// offset as appropriate.
1028 Value *getSafeStackPointerLocation(IRBuilder<> &IRB) const override;
1030 SDValue BuildFILD(SDValue Op, EVT SrcVT, SDValue Chain, SDValue StackSlot,
1031 SelectionDAG &DAG) const;
1033 bool isNoopAddrSpaceCast(unsigned SrcAS, unsigned DestAS) const override;
1035 /// \brief Customize the preferred legalization strategy for certain types.
1036 LegalizeTypeAction getPreferredVectorAction(EVT VT) const override;
1038 bool isIntDivCheap(EVT VT, AttributeSet Attr) const override;
1040 bool supportSwiftError() const override;
1042 unsigned getMaxSupportedInterleaveFactor() const override { return 4; }
1044 /// \brief Lower interleaved load(s) into target specific
1045 /// instructions/intrinsics.
1046 bool lowerInterleavedLoad(LoadInst *LI,
1047 ArrayRef<ShuffleVectorInst *> Shuffles,
1048 ArrayRef<unsigned> Indices,
1049 unsigned Factor) const override;
1051 std::pair<const TargetRegisterClass *, uint8_t>
1052 findRepresentativeClass(const TargetRegisterInfo *TRI,
1053 MVT VT) const override;
1056 /// Keep a reference to the X86Subtarget around so that we can
1057 /// make the right decision when generating code for different targets.
1058 const X86Subtarget &Subtarget;
1060 /// Select between SSE or x87 floating point ops.
1061 /// When SSE is available, use it for f32 operations.
1062 /// When SSE2 is available, use it for f64 operations.
1063 bool X86ScalarSSEf32;
1064 bool X86ScalarSSEf64;
1066 /// A list of legal FP immediates.
1067 std::vector<APFloat> LegalFPImmediates;
1069 /// Indicate that this x86 target can instruction
1070 /// select the specified FP immediate natively.
1071 void addLegalFPImmediate(const APFloat& Imm) {
1072 LegalFPImmediates.push_back(Imm);
1075 SDValue LowerCallResult(SDValue Chain, SDValue InFlag,
1076 CallingConv::ID CallConv, bool isVarArg,
1077 const SmallVectorImpl<ISD::InputArg> &Ins,
1078 const SDLoc &dl, SelectionDAG &DAG,
1079 SmallVectorImpl<SDValue> &InVals) const;
1080 SDValue LowerMemArgument(SDValue Chain, CallingConv::ID CallConv,
1081 const SmallVectorImpl<ISD::InputArg> &ArgInfo,
1082 const SDLoc &dl, SelectionDAG &DAG,
1083 const CCValAssign &VA, MachineFrameInfo &MFI,
1085 SDValue LowerMemOpCallTo(SDValue Chain, SDValue StackPtr, SDValue Arg,
1086 const SDLoc &dl, SelectionDAG &DAG,
1087 const CCValAssign &VA,
1088 ISD::ArgFlagsTy Flags) const;
1090 // Call lowering helpers.
1092 /// Check whether the call is eligible for tail call optimization. Targets
1093 /// that want to do tail call optimization should implement this function.
1094 bool IsEligibleForTailCallOptimization(SDValue Callee,
1095 CallingConv::ID CalleeCC,
1097 bool isCalleeStructRet,
1098 bool isCallerStructRet,
1100 const SmallVectorImpl<ISD::OutputArg> &Outs,
1101 const SmallVectorImpl<SDValue> &OutVals,
1102 const SmallVectorImpl<ISD::InputArg> &Ins,
1103 SelectionDAG& DAG) const;
1104 SDValue EmitTailCallLoadRetAddr(SelectionDAG &DAG, SDValue &OutRetAddr,
1105 SDValue Chain, bool IsTailCall,
1106 bool Is64Bit, int FPDiff,
1107 const SDLoc &dl) const;
1109 unsigned GetAlignedArgumentStackSize(unsigned StackSize,
1110 SelectionDAG &DAG) const;
1112 unsigned getAddressSpace(void) const;
1114 std::pair<SDValue,SDValue> FP_TO_INTHelper(SDValue Op, SelectionDAG &DAG,
1116 bool isReplace) const;
1118 SDValue LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) const;
1119 SDValue LowerBUILD_VECTORvXi1(SDValue Op, SelectionDAG &DAG) const;
1120 SDValue LowerVSELECT(SDValue Op, SelectionDAG &DAG) const;
1121 SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const;
1122 SDValue ExtractBitFromMaskVector(SDValue Op, SelectionDAG &DAG) const;
1123 SDValue InsertBitToMaskVector(SDValue Op, SelectionDAG &DAG) const;
1124 SDValue LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const;
1126 unsigned getGlobalWrapperKind(const GlobalValue *GV = nullptr) const;
1127 SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) const;
1128 SDValue LowerBlockAddress(SDValue Op, SelectionDAG &DAG) const;
1129 SDValue LowerGlobalAddress(const GlobalValue *GV, const SDLoc &dl,
1130 int64_t Offset, SelectionDAG &DAG) const;
1131 SDValue LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) const;
1132 SDValue LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const;
1133 SDValue LowerExternalSymbol(SDValue Op, SelectionDAG &DAG) const;
1135 SDValue LowerSINT_TO_FP(SDValue Op, SelectionDAG &DAG) const;
1136 SDValue LowerUINT_TO_FP(SDValue Op, SelectionDAG &DAG) const;
1137 SDValue LowerUINT_TO_FP_i64(SDValue Op, SelectionDAG &DAG) const;
1138 SDValue LowerUINT_TO_FP_i32(SDValue Op, SelectionDAG &DAG) const;
1139 SDValue lowerUINT_TO_FP_vec(SDValue Op, SelectionDAG &DAG) const;
1140 SDValue LowerTRUNCATE(SDValue Op, SelectionDAG &DAG) const;
1141 SDValue LowerFP_TO_INT(SDValue Op, const X86Subtarget &Subtarget,
1142 SelectionDAG &DAG) const;
1143 SDValue LowerToBT(SDValue And, ISD::CondCode CC, const SDLoc &dl,
1144 SelectionDAG &DAG) const;
1145 SDValue LowerSETCC(SDValue Op, SelectionDAG &DAG) const;
1146 SDValue LowerSETCCE(SDValue Op, SelectionDAG &DAG) const;
1147 SDValue LowerSELECT(SDValue Op, SelectionDAG &DAG) const;
1148 SDValue LowerBRCOND(SDValue Op, SelectionDAG &DAG) const;
1149 SDValue LowerJumpTable(SDValue Op, SelectionDAG &DAG) const;
1150 SDValue LowerDYNAMIC_STACKALLOC(SDValue Op, SelectionDAG &DAG) const;
1151 SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG) const;
1152 SDValue LowerVAARG(SDValue Op, SelectionDAG &DAG) const;
1153 SDValue LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const;
1154 SDValue LowerADDROFRETURNADDR(SDValue Op, SelectionDAG &DAG) const;
1155 SDValue LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const;
1156 SDValue LowerFRAME_TO_ARGS_OFFSET(SDValue Op, SelectionDAG &DAG) const;
1157 SDValue LowerEH_RETURN(SDValue Op, SelectionDAG &DAG) const;
1158 SDValue lowerEH_SJLJ_SETJMP(SDValue Op, SelectionDAG &DAG) const;
1159 SDValue lowerEH_SJLJ_LONGJMP(SDValue Op, SelectionDAG &DAG) const;
1160 SDValue lowerEH_SJLJ_SETUP_DISPATCH(SDValue Op, SelectionDAG &DAG) const;
1161 SDValue LowerINIT_TRAMPOLINE(SDValue Op, SelectionDAG &DAG) const;
1162 SDValue LowerFLT_ROUNDS_(SDValue Op, SelectionDAG &DAG) const;
1163 SDValue LowerWin64_i128OP(SDValue Op, SelectionDAG &DAG) const;
1164 SDValue LowerGC_TRANSITION_START(SDValue Op, SelectionDAG &DAG) const;
1165 SDValue LowerGC_TRANSITION_END(SDValue Op, SelectionDAG &DAG) const;
1168 LowerFormalArguments(SDValue Chain, CallingConv::ID CallConv, bool isVarArg,
1169 const SmallVectorImpl<ISD::InputArg> &Ins,
1170 const SDLoc &dl, SelectionDAG &DAG,
1171 SmallVectorImpl<SDValue> &InVals) const override;
1172 SDValue LowerCall(CallLoweringInfo &CLI,
1173 SmallVectorImpl<SDValue> &InVals) const override;
1175 SDValue LowerReturn(SDValue Chain, CallingConv::ID CallConv, bool isVarArg,
1176 const SmallVectorImpl<ISD::OutputArg> &Outs,
1177 const SmallVectorImpl<SDValue> &OutVals,
1178 const SDLoc &dl, SelectionDAG &DAG) const override;
1180 bool supportSplitCSR(MachineFunction *MF) const override {
1181 return MF->getFunction()->getCallingConv() == CallingConv::CXX_FAST_TLS &&
1182 MF->getFunction()->hasFnAttribute(Attribute::NoUnwind);
1184 void initializeSplitCSR(MachineBasicBlock *Entry) const override;
1185 void insertCopiesSplitCSR(
1186 MachineBasicBlock *Entry,
1187 const SmallVectorImpl<MachineBasicBlock *> &Exits) const override;
1189 bool isUsedByReturnOnly(SDNode *N, SDValue &Chain) const override;
1191 bool mayBeEmittedAsTailCall(CallInst *CI) const override;
1193 EVT getTypeForExtReturn(LLVMContext &Context, EVT VT,
1194 ISD::NodeType ExtendKind) const override;
1196 bool CanLowerReturn(CallingConv::ID CallConv, MachineFunction &MF,
1198 const SmallVectorImpl<ISD::OutputArg> &Outs,
1199 LLVMContext &Context) const override;
1201 const MCPhysReg *getScratchRegisters(CallingConv::ID CC) const override;
1203 TargetLoweringBase::AtomicExpansionKind
1204 shouldExpandAtomicLoadInIR(LoadInst *SI) const override;
1205 bool shouldExpandAtomicStoreInIR(StoreInst *SI) const override;
1206 TargetLoweringBase::AtomicExpansionKind
1207 shouldExpandAtomicRMWInIR(AtomicRMWInst *AI) const override;
1210 lowerIdempotentRMWIntoFencedLoad(AtomicRMWInst *AI) const override;
1212 bool needsCmpXchgNb(Type *MemType) const;
1214 void SetupEntryBlockForSjLj(MachineInstr &MI, MachineBasicBlock *MBB,
1215 MachineBasicBlock *DispatchBB, int FI) const;
1217 // Utility function to emit the low-level va_arg code for X86-64.
1219 EmitVAARG64WithCustomInserter(MachineInstr &MI,
1220 MachineBasicBlock *MBB) const;
1222 /// Utility function to emit the xmm reg save portion of va_start.
1224 EmitVAStartSaveXMMRegsWithCustomInserter(MachineInstr &BInstr,
1225 MachineBasicBlock *BB) const;
1227 MachineBasicBlock *EmitLoweredSelect(MachineInstr &I,
1228 MachineBasicBlock *BB) const;
1230 MachineBasicBlock *EmitLoweredAtomicFP(MachineInstr &I,
1231 MachineBasicBlock *BB) const;
1233 MachineBasicBlock *EmitLoweredCatchRet(MachineInstr &MI,
1234 MachineBasicBlock *BB) const;
1236 MachineBasicBlock *EmitLoweredCatchPad(MachineInstr &MI,
1237 MachineBasicBlock *BB) const;
1239 MachineBasicBlock *EmitLoweredSegAlloca(MachineInstr &MI,
1240 MachineBasicBlock *BB) const;
1242 MachineBasicBlock *EmitLoweredTLSAddr(MachineInstr &MI,
1243 MachineBasicBlock *BB) const;
1245 MachineBasicBlock *EmitLoweredTLSCall(MachineInstr &MI,
1246 MachineBasicBlock *BB) const;
1248 MachineBasicBlock *emitEHSjLjSetJmp(MachineInstr &MI,
1249 MachineBasicBlock *MBB) const;
1251 MachineBasicBlock *emitEHSjLjLongJmp(MachineInstr &MI,
1252 MachineBasicBlock *MBB) const;
1254 MachineBasicBlock *emitFMA3Instr(MachineInstr &MI,
1255 MachineBasicBlock *MBB) const;
1257 MachineBasicBlock *EmitSjLjDispatchBlock(MachineInstr &MI,
1258 MachineBasicBlock *MBB) const;
1260 /// Emit nodes that will be selected as "test Op0,Op0", or something
1261 /// equivalent, for use with the given x86 condition code.
1262 SDValue EmitTest(SDValue Op0, unsigned X86CC, const SDLoc &dl,
1263 SelectionDAG &DAG) const;
1265 /// Emit nodes that will be selected as "cmp Op0,Op1", or something
1266 /// equivalent, for use with the given x86 condition code.
1267 SDValue EmitCmp(SDValue Op0, SDValue Op1, unsigned X86CC, const SDLoc &dl,
1268 SelectionDAG &DAG) const;
1270 /// Convert a comparison if required by the subtarget.
1271 SDValue ConvertCmpIfNecessary(SDValue Cmp, SelectionDAG &DAG) const;
1273 /// Check if replacement of SQRT with RSQRT should be disabled.
1274 bool isFsqrtCheap(SDValue Operand, SelectionDAG &DAG) const override;
1276 /// Use rsqrt* to speed up sqrt calculations.
1277 SDValue getSqrtEstimate(SDValue Operand, SelectionDAG &DAG, int Enabled,
1278 int &RefinementSteps, bool &UseOneConstNR,
1279 bool Reciprocal) const override;
1281 /// Use rcp* to speed up fdiv calculations.
1282 SDValue getRecipEstimate(SDValue Operand, SelectionDAG &DAG, int Enabled,
1283 int &RefinementSteps) const override;
1285 /// Reassociate floating point divisions into multiply by reciprocal.
1286 unsigned combineRepeatedFPDivisors() const override;
1290 FastISel *createFastISel(FunctionLoweringInfo &funcInfo,
1291 const TargetLibraryInfo *libInfo);
1292 } // end namespace X86
1294 // Base class for all X86 non-masked store operations.
1295 class X86StoreSDNode : public MemSDNode {
1297 X86StoreSDNode(unsigned Opcode, unsigned Order, const DebugLoc &dl,
1298 SDVTList VTs, EVT MemVT,
1299 MachineMemOperand *MMO)
1300 :MemSDNode(Opcode, Order, dl, VTs, MemVT, MMO) {}
1301 const SDValue &getValue() const { return getOperand(1); }
1302 const SDValue &getBasePtr() const { return getOperand(2); }
1304 static bool classof(const SDNode *N) {
1305 return N->getOpcode() == X86ISD::VTRUNCSTORES ||
1306 N->getOpcode() == X86ISD::VTRUNCSTOREUS;
1310 // Base class for all X86 masked store operations.
1311 // The class has the same order of operands as MaskedStoreSDNode for
1313 class X86MaskedStoreSDNode : public MemSDNode {
1315 X86MaskedStoreSDNode(unsigned Opcode, unsigned Order,
1316 const DebugLoc &dl, SDVTList VTs, EVT MemVT,
1317 MachineMemOperand *MMO)
1318 : MemSDNode(Opcode, Order, dl, VTs, MemVT, MMO) {}
1320 const SDValue &getBasePtr() const { return getOperand(1); }
1321 const SDValue &getMask() const { return getOperand(2); }
1322 const SDValue &getValue() const { return getOperand(3); }
1324 static bool classof(const SDNode *N) {
1325 return N->getOpcode() == X86ISD::VMTRUNCSTORES ||
1326 N->getOpcode() == X86ISD::VMTRUNCSTOREUS;
1330 // X86 Truncating Store with Signed saturation.
1331 class TruncSStoreSDNode : public X86StoreSDNode {
1333 TruncSStoreSDNode(unsigned Order, const DebugLoc &dl,
1334 SDVTList VTs, EVT MemVT, MachineMemOperand *MMO)
1335 : X86StoreSDNode(X86ISD::VTRUNCSTORES, Order, dl, VTs, MemVT, MMO) {}
1337 static bool classof(const SDNode *N) {
1338 return N->getOpcode() == X86ISD::VTRUNCSTORES;
1342 // X86 Truncating Store with Unsigned saturation.
1343 class TruncUSStoreSDNode : public X86StoreSDNode {
1345 TruncUSStoreSDNode(unsigned Order, const DebugLoc &dl,
1346 SDVTList VTs, EVT MemVT, MachineMemOperand *MMO)
1347 : X86StoreSDNode(X86ISD::VTRUNCSTOREUS, Order, dl, VTs, MemVT, MMO) {}
1349 static bool classof(const SDNode *N) {
1350 return N->getOpcode() == X86ISD::VTRUNCSTOREUS;
1354 // X86 Truncating Masked Store with Signed saturation.
1355 class MaskedTruncSStoreSDNode : public X86MaskedStoreSDNode {
1357 MaskedTruncSStoreSDNode(unsigned Order,
1358 const DebugLoc &dl, SDVTList VTs, EVT MemVT,
1359 MachineMemOperand *MMO)
1360 : X86MaskedStoreSDNode(X86ISD::VMTRUNCSTORES, Order, dl, VTs, MemVT, MMO) {}
1362 static bool classof(const SDNode *N) {
1363 return N->getOpcode() == X86ISD::VMTRUNCSTORES;
1367 // X86 Truncating Masked Store with Unsigned saturation.
1368 class MaskedTruncUSStoreSDNode : public X86MaskedStoreSDNode {
1370 MaskedTruncUSStoreSDNode(unsigned Order,
1371 const DebugLoc &dl, SDVTList VTs, EVT MemVT,
1372 MachineMemOperand *MMO)
1373 : X86MaskedStoreSDNode(X86ISD::VMTRUNCSTOREUS, Order, dl, VTs, MemVT, MMO) {}
1375 static bool classof(const SDNode *N) {
1376 return N->getOpcode() == X86ISD::VMTRUNCSTOREUS;
1380 } // end namespace llvm
1382 #endif // LLVM_LIB_TARGET_X86_X86ISELLOWERING_H