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/CodeGen/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.
217 VGETMANT, VGETMANT_RND, VGETMANTS, VGETMANTS_RND,
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.
258 // AVX-512 reciprocal approximations with a little more precision.
259 RSQRT14, RSQRT14S, RCP14, RCP14S,
261 // Thread Local Storage.
264 // Thread Local Storage. A call to get the start address
265 // of the TLS block for the current module.
268 // Thread Local Storage. When calling to an OS provided
269 // thunk at the address from an earlier relocation.
272 // Exception Handling helpers.
275 // SjLj exception handling setjmp.
278 // SjLj exception handling longjmp.
281 // SjLj exception handling dispatch.
282 EH_SJLJ_SETUP_DISPATCH,
284 /// Tail call return. See X86TargetLowering::LowerCall for
285 /// the list of operands.
288 // Vector move to low scalar and zero higher vector elements.
291 // Vector integer zero-extend.
293 // Vector integer signed-extend.
296 // Vector integer truncate.
298 // Vector integer truncate with unsigned/signed saturation.
302 VFPEXT, VFPEXT_RND, VFPEXTS_RND,
305 VFPROUND, VFPROUND_RND, VFPROUNDS_RND,
307 // Convert a vector to mask, set bits base on MSB.
310 // 128-bit vector logical left / right shift
313 // Vector shift elements
316 // Vector variable shift right arithmetic.
317 // Unlike ISD::SRA, in case shift count greater then element size
318 // use sign bit to fill destination data element.
321 // Vector shift elements by immediate
324 // Shifts of mask registers.
327 // Bit rotate by immediate
330 // Vector packed double/float comparison.
333 // Vector integer comparisons.
335 // Vector integer comparisons, the result is in a mask vector.
338 // v8i16 Horizontal minimum and position.
343 /// Vector comparison generating mask bits for fp and
344 /// integer signed and unsigned data types.
347 // Vector comparison with rounding mode for FP values
350 // Arithmetic operations with FLAGS results.
351 ADD, SUB, ADC, SBB, SMUL,
352 INC, DEC, OR, XOR, AND,
354 // LOW, HI, FLAGS = umul LHS, RHS.
357 // 8-bit SMUL/UMUL - AX, FLAGS = smul8/umul8 AL, RHS.
360 // 8-bit divrem that zero-extend the high result (AH).
364 // X86-specific multiply by immediate.
367 // Vector sign bit extraction.
370 // Vector bitwise comparisons.
373 // Vector packed fp sign bitwise comparisons.
376 // Vector "test" in AVX-512, the result is in a mask vector.
380 // OR/AND test for masks.
384 // Several flavors of instructions with vector shuffle behaviors.
385 // Saturated signed/unnsigned packing.
388 // Intra-lane alignr.
390 // AVX512 inter-lane alignr.
396 // VBMI2 Concat & Shift.
401 //Shuffle Packed Values at 128-bit granularity.
419 // Variable Permute (VPERM).
420 // Res = VPERMV MaskV, V0
423 // 3-op Variable Permute (VPERMT2).
424 // Res = VPERMV3 V0, MaskV, V1
427 // 3-op Variable Permute overwriting the index (VPERMI2).
428 // Res = VPERMIV3 V0, MaskV, V1
431 // Bitwise ternary logic.
433 // Fix Up Special Packed Float32/64 values.
436 // Range Restriction Calculation For Packed Pairs of Float32/64 values.
437 VRANGE, VRANGE_RND, VRANGES, VRANGES_RND,
438 // Reduce - Perform Reduction Transformation on scalar\packed FP.
439 VREDUCE, VREDUCE_RND, VREDUCES, VREDUCES_RND,
440 // RndScale - Round FP Values To Include A Given Number Of Fraction Bits.
441 // Also used by the legacy (V)ROUND intrinsics where we mask out the
442 // scaling part of the immediate.
443 VRNDSCALE, VRNDSCALE_RND, VRNDSCALES, VRNDSCALES_RND,
444 // Tests Types Of a FP Values for packed types.
446 // Tests Types Of a FP Values for scalar types.
449 // Broadcast scalar to vector.
451 // Broadcast mask to vector.
453 // Broadcast subvector to vector.
456 /// SSE4A Extraction and Insertion.
459 // XOP arithmetic/logical shifts.
461 // XOP signed/unsigned integer comparisons.
463 // XOP packed permute bytes.
465 // XOP two source permutation.
468 // Vector multiply packed unsigned doubleword integers.
470 // Vector multiply packed signed doubleword integers.
472 // Vector Multiply Packed UnsignedIntegers with Round and Scale.
475 // Multiply and Add Packed Integers.
476 VPMADDUBSW, VPMADDWD,
478 // AVX512IFMA multiply and add.
479 // NOTE: These are different than the instruction and perform
481 VPMADD52L, VPMADD52H,
490 // We use the target independent ISD::FMA for the non-inverted case.
497 // FMA with rounding mode.
505 // FMA4 specific scalar intrinsics bits that zero the non-scalar bits.
506 FMADD4S, FNMADD4S, FMSUB4S, FNMSUB4S,
508 // Scalar intrinsic FMA.
514 // Scalar intrinsic FMA with rounding mode.
515 // Two versions, passthru bits on op1 or op3.
516 FMADDS1_RND, FMADDS3_RND,
517 FNMADDS1_RND, FNMADDS3_RND,
518 FMSUBS1_RND, FMSUBS3_RND,
519 FNMSUBS1_RND, FNMSUBS3_RND,
521 // Compress and expand.
528 // Convert Unsigned/Integer to Floating-Point Value with rounding mode.
529 SINT_TO_FP_RND, UINT_TO_FP_RND,
530 SCALAR_SINT_TO_FP_RND, SCALAR_UINT_TO_FP_RND,
532 // Vector float/double to signed/unsigned integer.
533 CVTP2SI, CVTP2UI, CVTP2SI_RND, CVTP2UI_RND,
534 // Scalar float/double to signed/unsigned integer.
535 CVTS2SI_RND, CVTS2UI_RND,
537 // Vector float/double to signed/unsigned integer with truncation.
538 CVTTP2SI, CVTTP2UI, CVTTP2SI_RND, CVTTP2UI_RND,
539 // Scalar float/double to signed/unsigned integer with truncation.
540 CVTTS2SI_RND, CVTTS2UI_RND,
542 // Vector signed/unsigned integer to float/double.
545 // Save xmm argument registers to the stack, according to %al. An operator
546 // is needed so that this can be expanded with control flow.
547 VASTART_SAVE_XMM_REGS,
549 // Windows's _chkstk call to do stack probing.
552 // For allocating variable amounts of stack space when using
553 // segmented stacks. Check if the current stacklet has enough space, and
554 // falls back to heap allocation if not.
561 // Store FP status word into i16 register.
564 // Store contents of %ah into %eflags.
567 // Get a random integer and indicate whether it is valid in CF.
570 // Get a NIST SP800-90B & C compliant random integer and
571 // indicate whether it is valid in CF.
574 // SSE42 string comparisons.
578 // Test if in transactional execution.
582 RSQRT28, RSQRT28S, RCP28, RCP28S, EXP2,
584 // Conversions between float and half-float.
585 CVTPS2PH, CVTPH2PS, CVTPH2PS_RND,
587 // Galois Field Arithmetic Instructions
588 GF2P8AFFINEINVQB, GF2P8AFFINEQB, GF2P8MULB,
590 // LWP insert record.
594 LCMPXCHG_DAG = ISD::FIRST_TARGET_MEMORY_OPCODE,
597 LCMPXCHG8_SAVE_EBX_DAG,
598 LCMPXCHG16_SAVE_RBX_DAG,
600 /// LOCK-prefixed arithmetic read-modify-write instructions.
601 /// EFLAGS, OUTCHAIN = LADD(INCHAIN, PTR, RHS)
602 LADD, LSUB, LOR, LXOR, LAND, LINC, LDEC,
604 // Load, scalar_to_vector, and zero extend.
607 // Store FP control world into i16 memory.
610 /// This instruction implements FP_TO_SINT with the
611 /// integer destination in memory and a FP reg source. This corresponds
612 /// to the X86::FIST*m instructions and the rounding mode change stuff. It
613 /// has two inputs (token chain and address) and two outputs (int value
614 /// and token chain).
619 /// This instruction implements SINT_TO_FP with the
620 /// integer source in memory and FP reg result. This corresponds to the
621 /// X86::FILD*m instructions. It has three inputs (token chain, address,
622 /// and source type) and two outputs (FP value and token chain). FILD_FLAG
623 /// also produces a flag).
627 /// This instruction implements an extending load to FP stack slots.
628 /// This corresponds to the X86::FLD32m / X86::FLD64m. It takes a chain
629 /// operand, ptr to load from, and a ValueType node indicating the type
633 /// This instruction implements a truncating store to FP stack
634 /// slots. This corresponds to the X86::FST32m / X86::FST64m. It takes a
635 /// chain operand, value to store, address, and a ValueType to store it
639 /// This instruction grabs the address of the next argument
640 /// from a va_list. (reads and modifies the va_list in memory)
643 // Vector truncating store with unsigned/signed saturation
644 VTRUNCSTOREUS, VTRUNCSTORES,
645 // Vector truncating masked store with unsigned/signed saturation
646 VMTRUNCSTOREUS, VMTRUNCSTORES,
648 // X86 specific gather and scatter
651 // WARNING: Do not add anything in the end unless you want the node to
652 // have memop! In fact, starting from FIRST_TARGET_MEMORY_OPCODE all
653 // opcodes will be thought as target memory ops!
655 } // end namespace X86ISD
657 /// Define some predicates that are used for node matching.
659 /// Returns true if Elt is a constant zero or floating point constant +0.0.
660 bool isZeroNode(SDValue Elt);
662 /// Returns true of the given offset can be
663 /// fit into displacement field of the instruction.
664 bool isOffsetSuitableForCodeModel(int64_t Offset, CodeModel::Model M,
665 bool hasSymbolicDisplacement = true);
667 /// Determines whether the callee is required to pop its
668 /// own arguments. Callee pop is necessary to support tail calls.
669 bool isCalleePop(CallingConv::ID CallingConv,
670 bool is64Bit, bool IsVarArg, bool GuaranteeTCO);
672 } // end namespace X86
674 //===--------------------------------------------------------------------===//
675 // X86 Implementation of the TargetLowering interface
676 class X86TargetLowering final : public TargetLowering {
678 explicit X86TargetLowering(const X86TargetMachine &TM,
679 const X86Subtarget &STI);
681 unsigned getJumpTableEncoding() const override;
682 bool useSoftFloat() const override;
684 void markLibCallAttributes(MachineFunction *MF, unsigned CC,
685 ArgListTy &Args) const override;
687 MVT getScalarShiftAmountTy(const DataLayout &, EVT VT) const override {
692 LowerCustomJumpTableEntry(const MachineJumpTableInfo *MJTI,
693 const MachineBasicBlock *MBB, unsigned uid,
694 MCContext &Ctx) const override;
696 /// Returns relocation base for the given PIC jumptable.
697 SDValue getPICJumpTableRelocBase(SDValue Table,
698 SelectionDAG &DAG) const override;
700 getPICJumpTableRelocBaseExpr(const MachineFunction *MF,
701 unsigned JTI, MCContext &Ctx) const override;
703 /// Return the desired alignment for ByVal aggregate
704 /// function arguments in the caller parameter area. For X86, aggregates
705 /// that contains are placed at 16-byte boundaries while the rest are at
706 /// 4-byte boundaries.
707 unsigned getByValTypeAlignment(Type *Ty,
708 const DataLayout &DL) const override;
710 /// Returns the target specific optimal type for load
711 /// and store operations as a result of memset, memcpy, and memmove
712 /// lowering. If DstAlign is zero that means it's safe to destination
713 /// alignment can satisfy any constraint. Similarly if SrcAlign is zero it
714 /// means there isn't a need to check it against alignment requirement,
715 /// probably because the source does not need to be loaded. If 'IsMemset' is
716 /// true, that means it's expanding a memset. If 'ZeroMemset' is true, that
717 /// means it's a memset of zero. 'MemcpyStrSrc' indicates whether the memcpy
718 /// source is constant so it does not need to be loaded.
719 /// It returns EVT::Other if the type should be determined using generic
720 /// target-independent logic.
721 EVT getOptimalMemOpType(uint64_t Size, unsigned DstAlign, unsigned SrcAlign,
722 bool IsMemset, bool ZeroMemset, bool MemcpyStrSrc,
723 MachineFunction &MF) const override;
725 /// Returns true if it's safe to use load / store of the
726 /// specified type to expand memcpy / memset inline. This is mostly true
727 /// for all types except for some special cases. For example, on X86
728 /// targets without SSE2 f64 load / store are done with fldl / fstpl which
729 /// also does type conversion. Note the specified type doesn't have to be
730 /// legal as the hook is used before type legalization.
731 bool isSafeMemOpType(MVT VT) const override;
733 /// Returns true if the target allows unaligned memory accesses of the
734 /// specified type. Returns whether it is "fast" in the last argument.
735 bool allowsMisalignedMemoryAccesses(EVT VT, unsigned AS, unsigned Align,
736 bool *Fast) const override;
738 /// Provide custom lowering hooks for some operations.
740 SDValue LowerOperation(SDValue Op, SelectionDAG &DAG) const override;
742 /// Places new result values for the node in Results (their number
743 /// and types must exactly match those of the original return values of
744 /// the node), or leaves Results empty, which indicates that the node is not
745 /// to be custom lowered after all.
746 void LowerOperationWrapper(SDNode *N,
747 SmallVectorImpl<SDValue> &Results,
748 SelectionDAG &DAG) const override;
750 /// Replace the results of node with an illegal result
751 /// type with new values built out of custom code.
753 void ReplaceNodeResults(SDNode *N, SmallVectorImpl<SDValue>&Results,
754 SelectionDAG &DAG) const override;
756 SDValue PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const override;
758 // Return true if it is profitable to combine a BUILD_VECTOR with a
759 // stride-pattern to a shuffle and a truncate.
760 // Example of such a combine:
761 // v4i32 build_vector((extract_elt V, 1),
762 // (extract_elt V, 3),
763 // (extract_elt V, 5),
764 // (extract_elt V, 7))
766 // v4i32 truncate (bitcast (shuffle<1,u,3,u,4,u,5,u,6,u,7,u> V, u) to
768 bool isDesirableToCombineBuildVectorToShuffleTruncate(
769 ArrayRef<int> ShuffleMask, EVT SrcVT, EVT TruncVT) const override;
771 /// Return true if the target has native support for
772 /// the specified value type and it is 'desirable' to use the type for the
773 /// given node type. e.g. On x86 i16 is legal, but undesirable since i16
774 /// instruction encodings are longer and some i16 instructions are slow.
775 bool isTypeDesirableForOp(unsigned Opc, EVT VT) const override;
777 /// Return true if the target has native support for the
778 /// specified value type and it is 'desirable' to use the type. e.g. On x86
779 /// i16 is legal, but undesirable since i16 instruction encodings are longer
780 /// and some i16 instructions are slow.
781 bool IsDesirableToPromoteOp(SDValue Op, EVT &PVT) const override;
784 EmitInstrWithCustomInserter(MachineInstr &MI,
785 MachineBasicBlock *MBB) const override;
787 /// This method returns the name of a target specific DAG node.
788 const char *getTargetNodeName(unsigned Opcode) const override;
790 bool mergeStoresAfterLegalization() const override { return true; }
792 bool canMergeStoresTo(unsigned AddressSpace, EVT MemVT,
793 const SelectionDAG &DAG) const override;
795 bool isCheapToSpeculateCttz() const override;
797 bool isCheapToSpeculateCtlz() const override;
799 bool isCtlzFast() const override;
801 bool hasBitPreservingFPLogic(EVT VT) const override {
802 return VT == MVT::f32 || VT == MVT::f64 || VT.isVector();
805 bool isMultiStoresCheaperThanBitsMerge(EVT LTy, EVT HTy) const override {
806 // If the pair to store is a mixture of float and int values, we will
807 // save two bitwise instructions and one float-to-int instruction and
808 // increase one store instruction. There is potentially a more
809 // significant benefit because it avoids the float->int domain switch
810 // for input value. So It is more likely a win.
811 if ((LTy.isFloatingPoint() && HTy.isInteger()) ||
812 (LTy.isInteger() && HTy.isFloatingPoint()))
814 // If the pair only contains int values, we will save two bitwise
815 // instructions and increase one store instruction (costing one more
816 // store buffer). Since the benefit is more blurred so we leave
817 // such pair out until we get testcase to prove it is a win.
821 bool isMaskAndCmp0FoldingBeneficial(const Instruction &AndI) const override;
823 bool hasAndNotCompare(SDValue Y) const override;
825 bool convertSetCCLogicToBitwiseLogic(EVT VT) const override {
826 return VT.isScalarInteger();
829 /// Vector-sized comparisons are fast using PCMPEQ + PMOVMSK or PTEST.
830 MVT hasFastEqualityCompare(unsigned NumBits) const override;
832 /// Return the value type to use for ISD::SETCC.
833 EVT getSetCCResultType(const DataLayout &DL, LLVMContext &Context,
834 EVT VT) const override;
836 /// Determine which of the bits specified in Mask are known to be either
837 /// zero or one and return them in the KnownZero/KnownOne bitsets.
838 void computeKnownBitsForTargetNode(const SDValue Op,
840 const APInt &DemandedElts,
841 const SelectionDAG &DAG,
842 unsigned Depth = 0) const override;
844 /// Determine the number of bits in the operation that are sign bits.
845 unsigned ComputeNumSignBitsForTargetNode(SDValue Op,
846 const APInt &DemandedElts,
847 const SelectionDAG &DAG,
848 unsigned Depth) const override;
850 SDValue unwrapAddress(SDValue N) const override;
852 bool isGAPlusOffset(SDNode *N, const GlobalValue* &GA,
853 int64_t &Offset) const override;
855 SDValue getReturnAddressFrameIndex(SelectionDAG &DAG) const;
857 bool ExpandInlineAsm(CallInst *CI) const override;
859 ConstraintType getConstraintType(StringRef Constraint) const override;
861 /// Examine constraint string and operand type and determine a weight value.
862 /// The operand object must already have been set up with the operand type.
864 getSingleConstraintMatchWeight(AsmOperandInfo &info,
865 const char *constraint) const override;
867 const char *LowerXConstraint(EVT ConstraintVT) const override;
869 /// Lower the specified operand into the Ops vector. If it is invalid, don't
870 /// add anything to Ops. If hasMemory is true it means one of the asm
871 /// constraint of the inline asm instruction being processed is 'm'.
872 void LowerAsmOperandForConstraint(SDValue Op,
873 std::string &Constraint,
874 std::vector<SDValue> &Ops,
875 SelectionDAG &DAG) const override;
878 getInlineAsmMemConstraint(StringRef ConstraintCode) const override {
879 if (ConstraintCode == "i")
880 return InlineAsm::Constraint_i;
881 else if (ConstraintCode == "o")
882 return InlineAsm::Constraint_o;
883 else if (ConstraintCode == "v")
884 return InlineAsm::Constraint_v;
885 else if (ConstraintCode == "X")
886 return InlineAsm::Constraint_X;
887 return TargetLowering::getInlineAsmMemConstraint(ConstraintCode);
890 /// Given a physical register constraint
891 /// (e.g. {edx}), return the register number and the register class for the
892 /// register. This should only be used for C_Register constraints. On
893 /// error, this returns a register number of 0.
894 std::pair<unsigned, const TargetRegisterClass *>
895 getRegForInlineAsmConstraint(const TargetRegisterInfo *TRI,
896 StringRef Constraint, MVT VT) const override;
898 /// Return true if the addressing mode represented
899 /// by AM is legal for this target, for a load/store of the specified type.
900 bool isLegalAddressingMode(const DataLayout &DL, const AddrMode &AM,
901 Type *Ty, unsigned AS,
902 Instruction *I = nullptr) const override;
904 /// Return true if the specified immediate is legal
905 /// icmp immediate, that is the target has icmp instructions which can
906 /// compare a register against the immediate without having to materialize
907 /// the immediate into a register.
908 bool isLegalICmpImmediate(int64_t Imm) const override;
910 /// Return true if the specified immediate is legal
911 /// add immediate, that is the target has add instructions which can
912 /// add a register and the immediate without having to materialize
913 /// the immediate into a register.
914 bool isLegalAddImmediate(int64_t Imm) const override;
916 /// \brief Return the cost of the scaling factor used in the addressing
917 /// mode represented by AM for this target, for a load/store
918 /// of the specified type.
919 /// If the AM is supported, the return value must be >= 0.
920 /// If the AM is not supported, it returns a negative value.
921 int getScalingFactorCost(const DataLayout &DL, const AddrMode &AM, Type *Ty,
922 unsigned AS) const override;
924 bool isVectorShiftByScalarCheap(Type *Ty) const override;
926 /// Return true if it's free to truncate a value of
927 /// type Ty1 to type Ty2. e.g. On x86 it's free to truncate a i32 value in
928 /// register EAX to i16 by referencing its sub-register AX.
929 bool isTruncateFree(Type *Ty1, Type *Ty2) const override;
930 bool isTruncateFree(EVT VT1, EVT VT2) const override;
932 bool allowTruncateForTailCall(Type *Ty1, Type *Ty2) const override;
934 /// Return true if any actual instruction that defines a
935 /// value of type Ty1 implicit zero-extends the value to Ty2 in the result
936 /// register. This does not necessarily include registers defined in
937 /// unknown ways, such as incoming arguments, or copies from unknown
938 /// virtual registers. Also, if isTruncateFree(Ty2, Ty1) is true, this
939 /// does not necessarily apply to truncate instructions. e.g. on x86-64,
940 /// all instructions that define 32-bit values implicit zero-extend the
941 /// result out to 64 bits.
942 bool isZExtFree(Type *Ty1, Type *Ty2) const override;
943 bool isZExtFree(EVT VT1, EVT VT2) const override;
944 bool isZExtFree(SDValue Val, EVT VT2) const override;
946 /// Return true if folding a vector load into ExtVal (a sign, zero, or any
947 /// extend node) is profitable.
948 bool isVectorLoadExtDesirable(SDValue) const override;
950 /// Return true if an FMA operation is faster than a pair of fmul and fadd
951 /// instructions. fmuladd intrinsics will be expanded to FMAs when this
952 /// method returns true, otherwise fmuladd is expanded to fmul + fadd.
953 bool isFMAFasterThanFMulAndFAdd(EVT VT) const override;
955 /// Return true if it's profitable to narrow
956 /// operations of type VT1 to VT2. e.g. on x86, it's profitable to narrow
957 /// from i32 to i8 but not from i32 to i16.
958 bool isNarrowingProfitable(EVT VT1, EVT VT2) const override;
960 /// Given an intrinsic, checks if on the target the intrinsic will need to map
961 /// to a MemIntrinsicNode (touches memory). If this is the case, it returns
962 /// true and stores the intrinsic information into the IntrinsicInfo that was
963 /// passed to the function.
964 bool getTgtMemIntrinsic(IntrinsicInfo &Info, const CallInst &I,
966 unsigned Intrinsic) const override;
968 /// Returns true if the target can instruction select the
969 /// specified FP immediate natively. If false, the legalizer will
970 /// materialize the FP immediate as a load from a constant pool.
971 bool isFPImmLegal(const APFloat &Imm, EVT VT) const override;
973 /// Targets can use this to indicate that they only support *some*
974 /// VECTOR_SHUFFLE operations, those with specific masks. By default, if a
975 /// target supports the VECTOR_SHUFFLE node, all mask values are assumed to
977 bool isShuffleMaskLegal(ArrayRef<int> Mask, EVT VT) const override;
979 /// Similar to isShuffleMaskLegal. This is used by Targets can use this to
980 /// indicate if there is a suitable VECTOR_SHUFFLE that can be used to
981 /// replace a VAND with a constant pool entry.
982 bool isVectorClearMaskLegal(const SmallVectorImpl<int> &Mask,
983 EVT VT) const override;
985 /// Returns true if lowering to a jump table is allowed.
986 bool areJTsAllowed(const Function *Fn) const override;
988 /// If true, then instruction selection should
989 /// seek to shrink the FP constant of the specified type to a smaller type
990 /// in order to save space and / or reduce runtime.
991 bool ShouldShrinkFPConstant(EVT VT) const override {
992 // Don't shrink FP constpool if SSE2 is available since cvtss2sd is more
993 // expensive than a straight movsd. On the other hand, it's important to
994 // shrink long double fp constant since fldt is very slow.
995 return !X86ScalarSSEf64 || VT == MVT::f80;
998 /// Return true if we believe it is correct and profitable to reduce the
999 /// load node to a smaller type.
1000 bool shouldReduceLoadWidth(SDNode *Load, ISD::LoadExtType ExtTy,
1001 EVT NewVT) const override;
1003 /// Return true if the specified scalar FP type is computed in an SSE
1004 /// register, not on the X87 floating point stack.
1005 bool isScalarFPTypeInSSEReg(EVT VT) const {
1006 return (VT == MVT::f64 && X86ScalarSSEf64) || // f64 is when SSE2
1007 (VT == MVT::f32 && X86ScalarSSEf32); // f32 is when SSE1
1010 /// \brief Returns true if it is beneficial to convert a load of a constant
1011 /// to just the constant itself.
1012 bool shouldConvertConstantLoadToIntImm(const APInt &Imm,
1013 Type *Ty) const override;
1015 bool convertSelectOfConstantsToMath(EVT VT) const override;
1017 /// Return true if EXTRACT_SUBVECTOR is cheap for this result type
1018 /// with this index.
1019 bool isExtractSubvectorCheap(EVT ResVT, EVT SrcVT,
1020 unsigned Index) const override;
1022 bool storeOfVectorConstantIsCheap(EVT MemVT, unsigned NumElem,
1023 unsigned AddrSpace) const override {
1024 // If we can replace more than 2 scalar stores, there will be a reduction
1025 // in instructions even after we add a vector constant load.
1029 bool isLoadBitCastBeneficial(EVT LoadVT, EVT BitcastVT) const override;
1031 /// Intel processors have a unified instruction and data cache
1032 const char * getClearCacheBuiltinName() const override {
1033 return nullptr; // nothing to do, move along.
1036 unsigned getRegisterByName(const char* RegName, EVT VT,
1037 SelectionDAG &DAG) const override;
1039 /// If a physical register, this returns the register that receives the
1040 /// exception address on entry to an EH pad.
1042 getExceptionPointerRegister(const Constant *PersonalityFn) const override;
1044 /// If a physical register, this returns the register that receives the
1045 /// exception typeid on entry to a landing pad.
1047 getExceptionSelectorRegister(const Constant *PersonalityFn) const override;
1049 virtual bool needsFixedCatchObjects() const override;
1051 /// This method returns a target specific FastISel object,
1052 /// or null if the target does not support "fast" ISel.
1053 FastISel *createFastISel(FunctionLoweringInfo &funcInfo,
1054 const TargetLibraryInfo *libInfo) const override;
1056 /// If the target has a standard location for the stack protector cookie,
1057 /// returns the address of that location. Otherwise, returns nullptr.
1058 Value *getIRStackGuard(IRBuilder<> &IRB) const override;
1060 bool useLoadStackGuardNode() const override;
1061 bool useStackGuardXorFP() const override;
1062 void insertSSPDeclarations(Module &M) const override;
1063 Value *getSDagStackGuard(const Module &M) const override;
1064 Value *getSSPStackGuardCheck(const Module &M) const override;
1065 SDValue emitStackGuardXorFP(SelectionDAG &DAG, SDValue Val,
1066 const SDLoc &DL) const override;
1069 /// Return true if the target stores SafeStack pointer at a fixed offset in
1070 /// some non-standard address space, and populates the address space and
1071 /// offset as appropriate.
1072 Value *getSafeStackPointerLocation(IRBuilder<> &IRB) const override;
1074 SDValue BuildFILD(SDValue Op, EVT SrcVT, SDValue Chain, SDValue StackSlot,
1075 SelectionDAG &DAG) const;
1077 bool isNoopAddrSpaceCast(unsigned SrcAS, unsigned DestAS) const override;
1079 /// \brief Customize the preferred legalization strategy for certain types.
1080 LegalizeTypeAction getPreferredVectorAction(EVT VT) const override;
1082 bool isIntDivCheap(EVT VT, AttributeList Attr) const override;
1084 bool supportSwiftError() const override;
1086 StringRef getStackProbeSymbolName(MachineFunction &MF) const override;
1088 unsigned getMaxSupportedInterleaveFactor() const override { return 4; }
1090 /// \brief Lower interleaved load(s) into target specific
1091 /// instructions/intrinsics.
1092 bool lowerInterleavedLoad(LoadInst *LI,
1093 ArrayRef<ShuffleVectorInst *> Shuffles,
1094 ArrayRef<unsigned> Indices,
1095 unsigned Factor) const override;
1097 /// \brief Lower interleaved store(s) into target specific
1098 /// instructions/intrinsics.
1099 bool lowerInterleavedStore(StoreInst *SI, ShuffleVectorInst *SVI,
1100 unsigned Factor) const override;
1103 void finalizeLowering(MachineFunction &MF) const override;
1106 std::pair<const TargetRegisterClass *, uint8_t>
1107 findRepresentativeClass(const TargetRegisterInfo *TRI,
1108 MVT VT) const override;
1111 /// Keep a reference to the X86Subtarget around so that we can
1112 /// make the right decision when generating code for different targets.
1113 const X86Subtarget &Subtarget;
1115 /// Select between SSE or x87 floating point ops.
1116 /// When SSE is available, use it for f32 operations.
1117 /// When SSE2 is available, use it for f64 operations.
1118 bool X86ScalarSSEf32;
1119 bool X86ScalarSSEf64;
1121 /// A list of legal FP immediates.
1122 std::vector<APFloat> LegalFPImmediates;
1124 /// Indicate that this x86 target can instruction
1125 /// select the specified FP immediate natively.
1126 void addLegalFPImmediate(const APFloat& Imm) {
1127 LegalFPImmediates.push_back(Imm);
1130 SDValue LowerCallResult(SDValue Chain, SDValue InFlag,
1131 CallingConv::ID CallConv, bool isVarArg,
1132 const SmallVectorImpl<ISD::InputArg> &Ins,
1133 const SDLoc &dl, SelectionDAG &DAG,
1134 SmallVectorImpl<SDValue> &InVals,
1135 uint32_t *RegMask) const;
1136 SDValue LowerMemArgument(SDValue Chain, CallingConv::ID CallConv,
1137 const SmallVectorImpl<ISD::InputArg> &ArgInfo,
1138 const SDLoc &dl, SelectionDAG &DAG,
1139 const CCValAssign &VA, MachineFrameInfo &MFI,
1141 SDValue LowerMemOpCallTo(SDValue Chain, SDValue StackPtr, SDValue Arg,
1142 const SDLoc &dl, SelectionDAG &DAG,
1143 const CCValAssign &VA,
1144 ISD::ArgFlagsTy Flags) const;
1146 // Call lowering helpers.
1148 /// Check whether the call is eligible for tail call optimization. Targets
1149 /// that want to do tail call optimization should implement this function.
1150 bool IsEligibleForTailCallOptimization(SDValue Callee,
1151 CallingConv::ID CalleeCC,
1153 bool isCalleeStructRet,
1154 bool isCallerStructRet,
1156 const SmallVectorImpl<ISD::OutputArg> &Outs,
1157 const SmallVectorImpl<SDValue> &OutVals,
1158 const SmallVectorImpl<ISD::InputArg> &Ins,
1159 SelectionDAG& DAG) const;
1160 SDValue EmitTailCallLoadRetAddr(SelectionDAG &DAG, SDValue &OutRetAddr,
1161 SDValue Chain, bool IsTailCall,
1162 bool Is64Bit, int FPDiff,
1163 const SDLoc &dl) const;
1165 unsigned GetAlignedArgumentStackSize(unsigned StackSize,
1166 SelectionDAG &DAG) const;
1168 unsigned getAddressSpace(void) const;
1170 std::pair<SDValue,SDValue> FP_TO_INTHelper(SDValue Op, SelectionDAG &DAG,
1172 bool isReplace) const;
1174 SDValue LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) const;
1175 SDValue LowerVSELECT(SDValue Op, SelectionDAG &DAG) const;
1176 SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const;
1177 SDValue LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const;
1179 unsigned getGlobalWrapperKind(const GlobalValue *GV = nullptr) const;
1180 SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) const;
1181 SDValue LowerBlockAddress(SDValue Op, SelectionDAG &DAG) const;
1182 SDValue LowerGlobalAddress(const GlobalValue *GV, const SDLoc &dl,
1183 int64_t Offset, SelectionDAG &DAG) const;
1184 SDValue LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) const;
1185 SDValue LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const;
1186 SDValue LowerExternalSymbol(SDValue Op, SelectionDAG &DAG) const;
1188 SDValue LowerSINT_TO_FP(SDValue Op, SelectionDAG &DAG) const;
1189 SDValue LowerUINT_TO_FP(SDValue Op, SelectionDAG &DAG) const;
1190 SDValue LowerTRUNCATE(SDValue Op, SelectionDAG &DAG) const;
1191 SDValue LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG) const;
1192 SDValue LowerSETCC(SDValue Op, SelectionDAG &DAG) const;
1193 SDValue LowerSETCCCARRY(SDValue Op, SelectionDAG &DAG) const;
1194 SDValue LowerSELECT(SDValue Op, SelectionDAG &DAG) const;
1195 SDValue LowerBRCOND(SDValue Op, SelectionDAG &DAG) const;
1196 SDValue LowerJumpTable(SDValue Op, SelectionDAG &DAG) const;
1197 SDValue LowerDYNAMIC_STACKALLOC(SDValue Op, SelectionDAG &DAG) const;
1198 SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG) const;
1199 SDValue LowerVAARG(SDValue Op, SelectionDAG &DAG) const;
1200 SDValue LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const;
1201 SDValue LowerADDROFRETURNADDR(SDValue Op, SelectionDAG &DAG) const;
1202 SDValue LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const;
1203 SDValue LowerFRAME_TO_ARGS_OFFSET(SDValue Op, SelectionDAG &DAG) const;
1204 SDValue LowerEH_RETURN(SDValue Op, SelectionDAG &DAG) const;
1205 SDValue lowerEH_SJLJ_SETJMP(SDValue Op, SelectionDAG &DAG) const;
1206 SDValue lowerEH_SJLJ_LONGJMP(SDValue Op, SelectionDAG &DAG) const;
1207 SDValue lowerEH_SJLJ_SETUP_DISPATCH(SDValue Op, SelectionDAG &DAG) const;
1208 SDValue LowerINIT_TRAMPOLINE(SDValue Op, SelectionDAG &DAG) const;
1209 SDValue LowerFLT_ROUNDS_(SDValue Op, SelectionDAG &DAG) const;
1210 SDValue LowerWin64_i128OP(SDValue Op, SelectionDAG &DAG) const;
1211 SDValue LowerGC_TRANSITION_START(SDValue Op, SelectionDAG &DAG) const;
1212 SDValue LowerGC_TRANSITION_END(SDValue Op, SelectionDAG &DAG) const;
1213 SDValue LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) const;
1216 LowerFormalArguments(SDValue Chain, CallingConv::ID CallConv, bool isVarArg,
1217 const SmallVectorImpl<ISD::InputArg> &Ins,
1218 const SDLoc &dl, SelectionDAG &DAG,
1219 SmallVectorImpl<SDValue> &InVals) const override;
1220 SDValue LowerCall(CallLoweringInfo &CLI,
1221 SmallVectorImpl<SDValue> &InVals) const override;
1223 SDValue LowerReturn(SDValue Chain, CallingConv::ID CallConv, bool isVarArg,
1224 const SmallVectorImpl<ISD::OutputArg> &Outs,
1225 const SmallVectorImpl<SDValue> &OutVals,
1226 const SDLoc &dl, SelectionDAG &DAG) const override;
1228 bool supportSplitCSR(MachineFunction *MF) const override {
1229 return MF->getFunction().getCallingConv() == CallingConv::CXX_FAST_TLS &&
1230 MF->getFunction().hasFnAttribute(Attribute::NoUnwind);
1232 void initializeSplitCSR(MachineBasicBlock *Entry) const override;
1233 void insertCopiesSplitCSR(
1234 MachineBasicBlock *Entry,
1235 const SmallVectorImpl<MachineBasicBlock *> &Exits) const override;
1237 bool isUsedByReturnOnly(SDNode *N, SDValue &Chain) const override;
1239 bool mayBeEmittedAsTailCall(const CallInst *CI) const override;
1241 EVT getTypeForExtReturn(LLVMContext &Context, EVT VT,
1242 ISD::NodeType ExtendKind) const override;
1244 bool CanLowerReturn(CallingConv::ID CallConv, MachineFunction &MF,
1246 const SmallVectorImpl<ISD::OutputArg> &Outs,
1247 LLVMContext &Context) const override;
1249 const MCPhysReg *getScratchRegisters(CallingConv::ID CC) const override;
1251 TargetLoweringBase::AtomicExpansionKind
1252 shouldExpandAtomicLoadInIR(LoadInst *SI) const override;
1253 bool shouldExpandAtomicStoreInIR(StoreInst *SI) const override;
1254 TargetLoweringBase::AtomicExpansionKind
1255 shouldExpandAtomicRMWInIR(AtomicRMWInst *AI) const override;
1258 lowerIdempotentRMWIntoFencedLoad(AtomicRMWInst *AI) const override;
1260 bool needsCmpXchgNb(Type *MemType) const;
1262 void SetupEntryBlockForSjLj(MachineInstr &MI, MachineBasicBlock *MBB,
1263 MachineBasicBlock *DispatchBB, int FI) const;
1265 // Utility function to emit the low-level va_arg code for X86-64.
1267 EmitVAARG64WithCustomInserter(MachineInstr &MI,
1268 MachineBasicBlock *MBB) const;
1270 /// Utility function to emit the xmm reg save portion of va_start.
1272 EmitVAStartSaveXMMRegsWithCustomInserter(MachineInstr &BInstr,
1273 MachineBasicBlock *BB) const;
1275 MachineBasicBlock *EmitLoweredCascadedSelect(MachineInstr &MI1,
1277 MachineBasicBlock *BB) const;
1279 MachineBasicBlock *EmitLoweredSelect(MachineInstr &I,
1280 MachineBasicBlock *BB) const;
1282 MachineBasicBlock *EmitLoweredAtomicFP(MachineInstr &I,
1283 MachineBasicBlock *BB) const;
1285 MachineBasicBlock *EmitLoweredCatchRet(MachineInstr &MI,
1286 MachineBasicBlock *BB) const;
1288 MachineBasicBlock *EmitLoweredCatchPad(MachineInstr &MI,
1289 MachineBasicBlock *BB) const;
1291 MachineBasicBlock *EmitLoweredSegAlloca(MachineInstr &MI,
1292 MachineBasicBlock *BB) const;
1294 MachineBasicBlock *EmitLoweredTLSAddr(MachineInstr &MI,
1295 MachineBasicBlock *BB) const;
1297 MachineBasicBlock *EmitLoweredTLSCall(MachineInstr &MI,
1298 MachineBasicBlock *BB) const;
1300 MachineBasicBlock *EmitLoweredRetpoline(MachineInstr &MI,
1301 MachineBasicBlock *BB) const;
1303 MachineBasicBlock *emitEHSjLjSetJmp(MachineInstr &MI,
1304 MachineBasicBlock *MBB) const;
1306 MachineBasicBlock *emitEHSjLjLongJmp(MachineInstr &MI,
1307 MachineBasicBlock *MBB) const;
1309 MachineBasicBlock *emitFMA3Instr(MachineInstr &MI,
1310 MachineBasicBlock *MBB) const;
1312 MachineBasicBlock *EmitSjLjDispatchBlock(MachineInstr &MI,
1313 MachineBasicBlock *MBB) const;
1315 /// Emit nodes that will be selected as "test Op0,Op0", or something
1316 /// equivalent, for use with the given x86 condition code.
1317 SDValue EmitTest(SDValue Op0, unsigned X86CC, const SDLoc &dl,
1318 SelectionDAG &DAG) const;
1320 /// Emit nodes that will be selected as "cmp Op0,Op1", or something
1321 /// equivalent, for use with the given x86 condition code.
1322 SDValue EmitCmp(SDValue Op0, SDValue Op1, unsigned X86CC, const SDLoc &dl,
1323 SelectionDAG &DAG) const;
1325 /// Convert a comparison if required by the subtarget.
1326 SDValue ConvertCmpIfNecessary(SDValue Cmp, SelectionDAG &DAG) const;
1328 /// Check if replacement of SQRT with RSQRT should be disabled.
1329 bool isFsqrtCheap(SDValue Operand, SelectionDAG &DAG) const override;
1331 /// Use rsqrt* to speed up sqrt calculations.
1332 SDValue getSqrtEstimate(SDValue Operand, SelectionDAG &DAG, int Enabled,
1333 int &RefinementSteps, bool &UseOneConstNR,
1334 bool Reciprocal) const override;
1336 /// Use rcp* to speed up fdiv calculations.
1337 SDValue getRecipEstimate(SDValue Operand, SelectionDAG &DAG, int Enabled,
1338 int &RefinementSteps) const override;
1340 /// Reassociate floating point divisions into multiply by reciprocal.
1341 unsigned combineRepeatedFPDivisors() const override;
1345 FastISel *createFastISel(FunctionLoweringInfo &funcInfo,
1346 const TargetLibraryInfo *libInfo);
1347 } // end namespace X86
1349 // Base class for all X86 non-masked store operations.
1350 class X86StoreSDNode : public MemSDNode {
1352 X86StoreSDNode(unsigned Opcode, unsigned Order, const DebugLoc &dl,
1353 SDVTList VTs, EVT MemVT,
1354 MachineMemOperand *MMO)
1355 :MemSDNode(Opcode, Order, dl, VTs, MemVT, MMO) {}
1356 const SDValue &getValue() const { return getOperand(1); }
1357 const SDValue &getBasePtr() const { return getOperand(2); }
1359 static bool classof(const SDNode *N) {
1360 return N->getOpcode() == X86ISD::VTRUNCSTORES ||
1361 N->getOpcode() == X86ISD::VTRUNCSTOREUS;
1365 // Base class for all X86 masked store operations.
1366 // The class has the same order of operands as MaskedStoreSDNode for
1368 class X86MaskedStoreSDNode : public MemSDNode {
1370 X86MaskedStoreSDNode(unsigned Opcode, unsigned Order,
1371 const DebugLoc &dl, SDVTList VTs, EVT MemVT,
1372 MachineMemOperand *MMO)
1373 : MemSDNode(Opcode, Order, dl, VTs, MemVT, MMO) {}
1375 const SDValue &getBasePtr() const { return getOperand(1); }
1376 const SDValue &getMask() const { return getOperand(2); }
1377 const SDValue &getValue() const { return getOperand(3); }
1379 static bool classof(const SDNode *N) {
1380 return N->getOpcode() == X86ISD::VMTRUNCSTORES ||
1381 N->getOpcode() == X86ISD::VMTRUNCSTOREUS;
1385 // X86 Truncating Store with Signed saturation.
1386 class TruncSStoreSDNode : public X86StoreSDNode {
1388 TruncSStoreSDNode(unsigned Order, const DebugLoc &dl,
1389 SDVTList VTs, EVT MemVT, MachineMemOperand *MMO)
1390 : X86StoreSDNode(X86ISD::VTRUNCSTORES, Order, dl, VTs, MemVT, MMO) {}
1392 static bool classof(const SDNode *N) {
1393 return N->getOpcode() == X86ISD::VTRUNCSTORES;
1397 // X86 Truncating Store with Unsigned saturation.
1398 class TruncUSStoreSDNode : public X86StoreSDNode {
1400 TruncUSStoreSDNode(unsigned Order, const DebugLoc &dl,
1401 SDVTList VTs, EVT MemVT, MachineMemOperand *MMO)
1402 : X86StoreSDNode(X86ISD::VTRUNCSTOREUS, Order, dl, VTs, MemVT, MMO) {}
1404 static bool classof(const SDNode *N) {
1405 return N->getOpcode() == X86ISD::VTRUNCSTOREUS;
1409 // X86 Truncating Masked Store with Signed saturation.
1410 class MaskedTruncSStoreSDNode : public X86MaskedStoreSDNode {
1412 MaskedTruncSStoreSDNode(unsigned Order,
1413 const DebugLoc &dl, SDVTList VTs, EVT MemVT,
1414 MachineMemOperand *MMO)
1415 : X86MaskedStoreSDNode(X86ISD::VMTRUNCSTORES, Order, dl, VTs, MemVT, MMO) {}
1417 static bool classof(const SDNode *N) {
1418 return N->getOpcode() == X86ISD::VMTRUNCSTORES;
1422 // X86 Truncating Masked Store with Unsigned saturation.
1423 class MaskedTruncUSStoreSDNode : public X86MaskedStoreSDNode {
1425 MaskedTruncUSStoreSDNode(unsigned Order,
1426 const DebugLoc &dl, SDVTList VTs, EVT MemVT,
1427 MachineMemOperand *MMO)
1428 : X86MaskedStoreSDNode(X86ISD::VMTRUNCSTOREUS, Order, dl, VTs, MemVT, MMO) {}
1430 static bool classof(const SDNode *N) {
1431 return N->getOpcode() == X86ISD::VMTRUNCSTOREUS;
1435 // X86 specific Gather/Scatter nodes.
1436 // The class has the same order of operands as MaskedGatherScatterSDNode for
1438 class X86MaskedGatherScatterSDNode : public MemSDNode {
1440 X86MaskedGatherScatterSDNode(unsigned Opc, unsigned Order,
1441 const DebugLoc &dl, SDVTList VTs, EVT MemVT,
1442 MachineMemOperand *MMO)
1443 : MemSDNode(Opc, Order, dl, VTs, MemVT, MMO) {}
1445 const SDValue &getBasePtr() const { return getOperand(3); }
1446 const SDValue &getIndex() const { return getOperand(4); }
1447 const SDValue &getMask() const { return getOperand(2); }
1448 const SDValue &getValue() const { return getOperand(1); }
1450 static bool classof(const SDNode *N) {
1451 return N->getOpcode() == X86ISD::MGATHER ||
1452 N->getOpcode() == X86ISD::MSCATTER;
1456 class X86MaskedGatherSDNode : public X86MaskedGatherScatterSDNode {
1458 X86MaskedGatherSDNode(unsigned Order, const DebugLoc &dl, SDVTList VTs,
1459 EVT MemVT, MachineMemOperand *MMO)
1460 : X86MaskedGatherScatterSDNode(X86ISD::MGATHER, Order, dl, VTs, MemVT,
1463 static bool classof(const SDNode *N) {
1464 return N->getOpcode() == X86ISD::MGATHER;
1468 class X86MaskedScatterSDNode : public X86MaskedGatherScatterSDNode {
1470 X86MaskedScatterSDNode(unsigned Order, const DebugLoc &dl, SDVTList VTs,
1471 EVT MemVT, MachineMemOperand *MMO)
1472 : X86MaskedGatherScatterSDNode(X86ISD::MSCATTER, Order, dl, VTs, MemVT,
1475 static bool classof(const SDNode *N) {
1476 return N->getOpcode() == X86ISD::MSCATTER;
1480 /// Generate unpacklo/unpackhi shuffle mask.
1481 template <typename T = int>
1482 void createUnpackShuffleMask(MVT VT, SmallVectorImpl<T> &Mask, bool Lo,
1484 assert(Mask.empty() && "Expected an empty shuffle mask vector");
1485 int NumElts = VT.getVectorNumElements();
1486 int NumEltsInLane = 128 / VT.getScalarSizeInBits();
1487 for (int i = 0; i < NumElts; ++i) {
1488 unsigned LaneStart = (i / NumEltsInLane) * NumEltsInLane;
1489 int Pos = (i % NumEltsInLane) / 2 + LaneStart;
1490 Pos += (Unary ? 0 : NumElts * (i % 2));
1491 Pos += (Lo ? 0 : NumEltsInLane / 2);
1492 Mask.push_back(Pos);
1496 /// Helper function to scale a shuffle or target shuffle mask, replacing each
1497 /// mask index with the scaled sequential indices for an equivalent narrowed
1498 /// mask. This is the reverse process to canWidenShuffleElements, but can
1500 template <typename T>
1501 void scaleShuffleMask(int Scale, ArrayRef<T> Mask,
1502 SmallVectorImpl<T> &ScaledMask) {
1503 assert(0 < Scale && "Unexpected scaling factor");
1504 int NumElts = Mask.size();
1505 ScaledMask.assign(static_cast<size_t>(NumElts * Scale), -1);
1507 for (int i = 0; i != NumElts; ++i) {
1510 // Repeat sentinel values in every mask element.
1512 for (int s = 0; s != Scale; ++s)
1513 ScaledMask[(Scale * i) + s] = M;
1517 // Scale mask element and increment across each mask element.
1518 for (int s = 0; s != Scale; ++s)
1519 ScaledMask[(Scale * i) + s] = (Scale * M) + s;
1522 } // end namespace llvm
1524 #endif // LLVM_LIB_TARGET_X86_X86ISELLOWERING_H