1 //===-- X86ISelLowering.h - X86 DAG Lowering Interface ----------*- C++ -*-===//
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
7 //===----------------------------------------------------------------------===//
9 // This file defines the interfaces that X86 uses to lower LLVM code into a
12 //===----------------------------------------------------------------------===//
14 #ifndef LLVM_LIB_TARGET_X86_X86ISELLOWERING_H
15 #define LLVM_LIB_TARGET_X86_X86ISELLOWERING_H
17 #include "llvm/CodeGen/CallingConvLower.h"
18 #include "llvm/CodeGen/SelectionDAG.h"
19 #include "llvm/CodeGen/TargetLowering.h"
20 #include "llvm/Target/TargetOptions.h"
24 class X86TargetMachine;
27 // X86 Specific DAG Nodes
28 enum NodeType : unsigned {
29 // Start the numbering where the builtin ops leave off.
30 FIRST_NUMBER = ISD::BUILTIN_OP_END,
37 /// Double shift instructions. These correspond to
38 /// X86::SHLDxx and X86::SHRDxx instructions.
42 /// Bitwise logical AND of floating point values. This corresponds
43 /// to X86::ANDPS or X86::ANDPD.
46 /// Bitwise logical OR of floating point values. This corresponds
47 /// to X86::ORPS or X86::ORPD.
50 /// Bitwise logical XOR of floating point values. This corresponds
51 /// to X86::XORPS or X86::XORPD.
54 /// Bitwise logical ANDNOT of floating point values. This
55 /// corresponds to X86::ANDNPS or X86::ANDNPD.
58 /// These operations represent an abstract X86 call
59 /// instruction, which includes a bunch of information. In particular the
60 /// operands of these node are:
62 /// #0 - The incoming token chain
64 /// #2 - The number of arg bytes the caller pushes on the stack.
65 /// #3 - The number of arg bytes the callee pops off the stack.
66 /// #4 - The value to pass in AL/AX/EAX (optional)
67 /// #5 - The value to pass in DL/DX/EDX (optional)
69 /// The result values of these nodes are:
71 /// #0 - The outgoing token chain
72 /// #1 - The first register result value (optional)
73 /// #2 - The second register result value (optional)
77 /// Same as call except it adds the NoTrack prefix.
80 /// This operation implements the lowering for readcyclecounter.
83 /// X86 Read Time-Stamp Counter and Processor ID.
86 /// X86 Read Performance Monitoring Counters.
89 /// X86 compare and logical compare instructions.
92 /// X86 bit-test instructions.
95 /// X86 SetCC. Operand 0 is condition code, and operand 1 is the EFLAGS
96 /// operand, usually produced by a CMP instruction.
102 // Same as SETCC except it's materialized with a sbb and the value is all
103 // one's or all zero's.
104 SETCC_CARRY, // R = carry_bit ? ~0 : 0
106 /// X86 FP SETCC, implemented with CMP{cc}SS/CMP{cc}SD.
107 /// Operands are two FP values to compare; result is a mask of
108 /// 0s or 1s. Generally DTRT for C/C++ with NaNs.
111 /// X86 FP SETCC, similar to above, but with output as an i1 mask and
112 /// with optional rounding mode.
113 FSETCCM, FSETCCM_RND,
115 /// X86 conditional moves. Operand 0 and operand 1 are the two values
116 /// to select from. Operand 2 is the condition code, and operand 3 is the
117 /// flag operand produced by a CMP or TEST instruction. It also writes a
121 /// X86 conditional branches. Operand 0 is the chain operand, operand 1
122 /// is the block to branch if condition is true, operand 2 is the
123 /// condition code, and operand 3 is the flag operand produced by a CMP
124 /// or TEST instruction.
127 /// BRIND node with NoTrack prefix. Operand 0 is the chain operand and
128 /// operand 1 is the target address.
131 /// Return with a flag operand. Operand 0 is the chain operand, operand
132 /// 1 is the number of bytes of stack to pop.
135 /// Return from interrupt. Operand 0 is the number of bytes to pop.
138 /// Repeat fill, corresponds to X86::REP_STOSx.
141 /// Repeat move, corresponds to X86::REP_MOVSx.
144 /// On Darwin, this node represents the result of the popl
145 /// at function entry, used for PIC code.
148 /// A wrapper node for TargetConstantPool, TargetJumpTable,
149 /// TargetExternalSymbol, TargetGlobalAddress, TargetGlobalTLSAddress,
150 /// MCSymbol and TargetBlockAddress.
153 /// Special wrapper used under X86-64 PIC mode for RIP
154 /// relative displacements.
157 /// Copies a 64-bit value from the low word of an XMM vector
158 /// to an MMX vector.
161 /// Copies a 32-bit value from the low word of a MMX
165 /// Copies a GPR into the low 32-bit word of a MMX vector
166 /// and zero out the high word.
169 /// Extract an 8-bit value from a vector and zero extend it to
170 /// i32, corresponds to X86::PEXTRB.
173 /// Extract a 16-bit value from a vector and zero extend it to
174 /// i32, corresponds to X86::PEXTRW.
177 /// Insert any element of a 4 x float vector into any element
178 /// of a destination 4 x floatvector.
181 /// Insert the lower 8-bits of a 32-bit value to a vector,
182 /// corresponds to X86::PINSRB.
185 /// Insert the lower 16-bits of a 32-bit value to a vector,
186 /// corresponds to X86::PINSRW.
189 /// Shuffle 16 8-bit values within a vector.
192 /// Compute Sum of Absolute Differences.
194 /// Compute Double Block Packed Sum-Absolute-Differences
197 /// Bitwise Logical AND NOT of Packed FP values.
200 /// Blend where the selector is an immediate.
203 /// Dynamic (non-constant condition) vector blend where only the sign bits
204 /// of the condition elements are used. This is used to enforce that the
205 /// condition mask is not valid for generic VSELECT optimizations. This
206 /// is also used to implement the intrinsics.
207 /// Operands are in VSELECT order: MASK, TRUE, FALSE
210 /// Combined add and sub on an FP vector.
213 // FP vector ops with rounding mode.
220 FSQRT_RND, FSQRTS_RND,
222 // FP vector get exponent.
223 FGETEXP_RND, FGETEXPS_RND,
224 // Extract Normalized Mantissas.
225 VGETMANT, VGETMANT_RND, VGETMANTS, VGETMANTS_RND,
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 truncate.
293 // Vector integer truncate with unsigned/signed saturation.
296 // Masked version of the above. Used when less than a 128-bit result is
297 // produced since the mask only applies to the lower elements and can't
298 // be represented by a select.
299 // SRC, PASSTHRU, MASK
300 VMTRUNC, VMTRUNCUS, VMTRUNCS,
303 VFPEXT, VFPEXT_RND, VFPEXTS_RND,
306 VFPROUND, VFPROUND_RND, VFPROUNDS_RND,
308 // Masked version of above. Used for v2f64->v4f32.
309 // SRC, PASSTHRU, MASK
312 // 128-bit vector logical left / right shift
315 // Vector shift elements
318 // Vector variable shift
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.
336 // v8i16 Horizontal minimum and position.
341 /// Vector comparison generating mask bits for fp and
342 /// integer signed and unsigned data types.
344 // Vector comparison with rounding mode for FP values
347 // Arithmetic operations with FLAGS results.
348 ADD, SUB, ADC, SBB, SMUL, UMUL,
351 // Bit field extract.
354 // Zero High Bits Starting with Specified Bit Position.
357 // X86-specific multiply by immediate.
360 // Vector sign bit extraction.
363 // Vector bitwise comparisons.
366 // Vector packed fp sign bitwise comparisons.
369 // OR/AND test for masks.
376 // Several flavors of instructions with vector shuffle behaviors.
377 // Saturated signed/unnsigned packing.
380 // Intra-lane alignr.
382 // AVX512 inter-lane alignr.
388 // VBMI2 Concat & Shift.
393 //Shuffle Packed Values at 128-bit granularity.
409 // Variable Permute (VPERM).
410 // Res = VPERMV MaskV, V0
413 // 3-op Variable Permute (VPERMT2).
414 // Res = VPERMV3 V0, MaskV, V1
417 // Bitwise ternary logic.
419 // Fix Up Special Packed Float32/64 values.
422 // Range Restriction Calculation For Packed Pairs of Float32/64 values.
423 VRANGE, VRANGE_RND, VRANGES, VRANGES_RND,
424 // Reduce - Perform Reduction Transformation on scalar\packed FP.
425 VREDUCE, VREDUCE_RND, VREDUCES, VREDUCES_RND,
426 // RndScale - Round FP Values To Include A Given Number Of Fraction Bits.
427 // Also used by the legacy (V)ROUND intrinsics where we mask out the
428 // scaling part of the immediate.
429 VRNDSCALE, VRNDSCALE_RND, VRNDSCALES, VRNDSCALES_RND,
430 // Tests Types Of a FP Values for packed types.
432 // Tests Types Of a FP Values for scalar types.
435 // Broadcast scalar to vector.
437 // Broadcast mask to vector.
439 // Broadcast subvector to vector.
442 /// SSE4A Extraction and Insertion.
445 // XOP arithmetic/logical shifts.
447 // XOP signed/unsigned integer comparisons.
449 // XOP packed permute bytes.
451 // XOP two source permutation.
454 // Vector multiply packed unsigned doubleword integers.
456 // Vector multiply packed signed doubleword integers.
458 // Vector Multiply Packed UnsignedIntegers with Round and Scale.
461 // Multiply and Add Packed Integers.
462 VPMADDUBSW, VPMADDWD,
464 // AVX512IFMA multiply and add.
465 // NOTE: These are different than the instruction and perform
467 VPMADD52L, VPMADD52H,
476 // We use the target independent ISD::FMA for the non-inverted case.
483 // FMA with rounding mode.
491 // Compress and expand.
498 // Convert Unsigned/Integer to Floating-Point Value with rounding mode.
499 SINT_TO_FP_RND, UINT_TO_FP_RND,
500 SCALAR_SINT_TO_FP_RND, SCALAR_UINT_TO_FP_RND,
502 // Vector float/double to signed/unsigned integer.
503 CVTP2SI, CVTP2UI, CVTP2SI_RND, CVTP2UI_RND,
504 // Scalar float/double to signed/unsigned integer.
505 CVTS2SI, CVTS2UI, CVTS2SI_RND, CVTS2UI_RND,
507 // Vector float/double to signed/unsigned integer with truncation.
508 CVTTP2SI, CVTTP2UI, CVTTP2SI_RND, CVTTP2UI_RND,
509 // Scalar float/double to signed/unsigned integer with truncation.
510 CVTTS2SI, CVTTS2UI, CVTTS2SI_RND, CVTTS2UI_RND,
512 // Vector signed/unsigned integer to float/double.
515 // Masked versions of above. Used for v2f64->v4f32.
516 // SRC, PASSTHRU, MASK
517 MCVTP2SI, MCVTP2UI, MCVTTP2SI, MCVTTP2UI,
520 // Save xmm argument registers to the stack, according to %al. An operator
521 // is needed so that this can be expanded with control flow.
522 VASTART_SAVE_XMM_REGS,
524 // Windows's _chkstk call to do stack probing.
527 // For allocating variable amounts of stack space when using
528 // segmented stacks. Check if the current stacklet has enough space, and
529 // falls back to heap allocation if not.
536 // Store FP status word into i16 register.
539 // Store contents of %ah into %eflags.
542 // Get a random integer and indicate whether it is valid in CF.
545 // Get a NIST SP800-90B & C compliant random integer and
546 // indicate whether it is valid in CF.
549 // SSE42 string comparisons.
550 // These nodes produce 3 results, index, mask, and flags. X86ISelDAGToDAG
551 // will emit one or two instructions based on which results are used. If
552 // flags and index/mask this allows us to use a single instruction since
553 // we won't have to pick and opcode for flags. Instead we can rely on the
554 // DAG to CSE everything and decide at isel.
558 // Test if in transactional execution.
562 RSQRT28, RSQRT28S, RCP28, RCP28S, EXP2,
564 // Conversions between float and half-float.
565 CVTPS2PH, CVTPH2PS, CVTPH2PS_RND,
567 // Masked version of above.
568 // SRC, RND, PASSTHRU, MASK
571 // Galois Field Arithmetic Instructions
572 GF2P8AFFINEINVQB, GF2P8AFFINEQB, GF2P8MULB,
574 // LWP insert record.
581 LCMPXCHG_DAG = ISD::FIRST_TARGET_MEMORY_OPCODE,
584 LCMPXCHG8_SAVE_EBX_DAG,
585 LCMPXCHG16_SAVE_RBX_DAG,
587 /// LOCK-prefixed arithmetic read-modify-write instructions.
588 /// EFLAGS, OUTCHAIN = LADD(INCHAIN, PTR, RHS)
589 LADD, LSUB, LOR, LXOR, LAND,
591 // Load, scalar_to_vector, and zero extend.
594 // Store FP control world into i16 memory.
597 /// This instruction implements FP_TO_SINT with the
598 /// integer destination in memory and a FP reg source. This corresponds
599 /// to the X86::FIST*m instructions and the rounding mode change stuff. It
600 /// has two inputs (token chain and address) and two outputs (int value
601 /// and token chain). Memory VT specifies the type to store to.
604 /// This instruction implements SINT_TO_FP with the
605 /// integer source in memory and FP reg result. This corresponds to the
606 /// X86::FILD*m instructions. It has two inputs (token chain and address)
607 /// and two outputs (FP value and token chain). FILD_FLAG also produces a
608 /// flag). The integer source type is specified by the memory VT.
612 /// This instruction implements an extending load to FP stack slots.
613 /// This corresponds to the X86::FLD32m / X86::FLD64m. It takes a chain
614 /// operand, and ptr to load from. The memory VT specifies the type to
618 /// This instruction implements a truncating store to FP stack
619 /// slots. This corresponds to the X86::FST32m / X86::FST64m. It takes a
620 /// chain operand, value to store, and address. The memory VT specifies
621 /// the type to store as.
624 /// This instruction grabs the address of the next argument
625 /// from a va_list. (reads and modifies the va_list in memory)
628 // Vector truncating store with unsigned/signed saturation
629 VTRUNCSTOREUS, VTRUNCSTORES,
630 // Vector truncating masked store with unsigned/signed saturation
631 VMTRUNCSTOREUS, VMTRUNCSTORES,
633 // X86 specific gather and scatter
636 // WARNING: Do not add anything in the end unless you want the node to
637 // have memop! In fact, starting from FIRST_TARGET_MEMORY_OPCODE all
638 // opcodes will be thought as target memory ops!
640 } // end namespace X86ISD
642 /// Define some predicates that are used for node matching.
644 /// Returns true if Elt is a constant zero or floating point constant +0.0.
645 bool isZeroNode(SDValue Elt);
647 /// Returns true of the given offset can be
648 /// fit into displacement field of the instruction.
649 bool isOffsetSuitableForCodeModel(int64_t Offset, CodeModel::Model M,
650 bool hasSymbolicDisplacement = true);
652 /// Determines whether the callee is required to pop its
653 /// own arguments. Callee pop is necessary to support tail calls.
654 bool isCalleePop(CallingConv::ID CallingConv,
655 bool is64Bit, bool IsVarArg, bool GuaranteeTCO);
657 } // end namespace X86
659 //===--------------------------------------------------------------------===//
660 // X86 Implementation of the TargetLowering interface
661 class X86TargetLowering final : public TargetLowering {
663 explicit X86TargetLowering(const X86TargetMachine &TM,
664 const X86Subtarget &STI);
666 unsigned getJumpTableEncoding() const override;
667 bool useSoftFloat() const override;
669 void markLibCallAttributes(MachineFunction *MF, unsigned CC,
670 ArgListTy &Args) const override;
672 MVT getScalarShiftAmountTy(const DataLayout &, EVT VT) const override {
677 LowerCustomJumpTableEntry(const MachineJumpTableInfo *MJTI,
678 const MachineBasicBlock *MBB, unsigned uid,
679 MCContext &Ctx) const override;
681 /// Returns relocation base for the given PIC jumptable.
682 SDValue getPICJumpTableRelocBase(SDValue Table,
683 SelectionDAG &DAG) const override;
685 getPICJumpTableRelocBaseExpr(const MachineFunction *MF,
686 unsigned JTI, MCContext &Ctx) const override;
688 /// Return the desired alignment for ByVal aggregate
689 /// function arguments in the caller parameter area. For X86, aggregates
690 /// that contains are placed at 16-byte boundaries while the rest are at
691 /// 4-byte boundaries.
692 unsigned getByValTypeAlignment(Type *Ty,
693 const DataLayout &DL) const override;
695 /// Returns the target specific optimal type for load
696 /// and store operations as a result of memset, memcpy, and memmove
697 /// lowering. If DstAlign is zero that means it's safe to destination
698 /// alignment can satisfy any constraint. Similarly if SrcAlign is zero it
699 /// means there isn't a need to check it against alignment requirement,
700 /// probably because the source does not need to be loaded. If 'IsMemset' is
701 /// true, that means it's expanding a memset. If 'ZeroMemset' is true, that
702 /// means it's a memset of zero. 'MemcpyStrSrc' indicates whether the memcpy
703 /// source is constant so it does not need to be loaded.
704 /// It returns EVT::Other if the type should be determined using generic
705 /// target-independent logic.
706 EVT getOptimalMemOpType(uint64_t Size, unsigned DstAlign, unsigned SrcAlign,
707 bool IsMemset, bool ZeroMemset, bool MemcpyStrSrc,
708 MachineFunction &MF) const override;
710 /// Returns true if it's safe to use load / store of the
711 /// specified type to expand memcpy / memset inline. This is mostly true
712 /// for all types except for some special cases. For example, on X86
713 /// targets without SSE2 f64 load / store are done with fldl / fstpl which
714 /// also does type conversion. Note the specified type doesn't have to be
715 /// legal as the hook is used before type legalization.
716 bool isSafeMemOpType(MVT VT) const override;
718 /// Returns true if the target allows unaligned memory accesses of the
719 /// specified type. Returns whether it is "fast" in the last argument.
720 bool allowsMisalignedMemoryAccesses(EVT VT, unsigned AS, unsigned Align,
721 bool *Fast) const override;
723 /// Provide custom lowering hooks for some operations.
725 SDValue LowerOperation(SDValue Op, SelectionDAG &DAG) const override;
727 /// Places new result values for the node in Results (their number
728 /// and types must exactly match those of the original return values of
729 /// the node), or leaves Results empty, which indicates that the node is not
730 /// to be custom lowered after all.
731 void LowerOperationWrapper(SDNode *N,
732 SmallVectorImpl<SDValue> &Results,
733 SelectionDAG &DAG) const override;
735 /// Replace the results of node with an illegal result
736 /// type with new values built out of custom code.
738 void ReplaceNodeResults(SDNode *N, SmallVectorImpl<SDValue>&Results,
739 SelectionDAG &DAG) const override;
741 SDValue PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const override;
743 // Return true if it is profitable to combine a BUILD_VECTOR with a
744 // stride-pattern to a shuffle and a truncate.
745 // Example of such a combine:
746 // v4i32 build_vector((extract_elt V, 1),
747 // (extract_elt V, 3),
748 // (extract_elt V, 5),
749 // (extract_elt V, 7))
751 // v4i32 truncate (bitcast (shuffle<1,u,3,u,4,u,5,u,6,u,7,u> V, u) to
753 bool isDesirableToCombineBuildVectorToShuffleTruncate(
754 ArrayRef<int> ShuffleMask, EVT SrcVT, EVT TruncVT) const override;
756 /// Return true if the target has native support for
757 /// the specified value type and it is 'desirable' to use the type for the
758 /// given node type. e.g. On x86 i16 is legal, but undesirable since i16
759 /// instruction encodings are longer and some i16 instructions are slow.
760 bool isTypeDesirableForOp(unsigned Opc, EVT VT) const override;
762 /// Return true if the target has native support for the
763 /// specified value type and it is 'desirable' to use the type. e.g. On x86
764 /// i16 is legal, but undesirable since i16 instruction encodings are longer
765 /// and some i16 instructions are slow.
766 bool IsDesirableToPromoteOp(SDValue Op, EVT &PVT) const override;
769 EmitInstrWithCustomInserter(MachineInstr &MI,
770 MachineBasicBlock *MBB) const override;
772 /// This method returns the name of a target specific DAG node.
773 const char *getTargetNodeName(unsigned Opcode) const override;
775 bool mergeStoresAfterLegalization() const override { return true; }
777 bool canMergeStoresTo(unsigned AddressSpace, EVT MemVT,
778 const SelectionDAG &DAG) const override;
780 bool isCheapToSpeculateCttz() const override;
782 bool isCheapToSpeculateCtlz() const override;
784 bool isCtlzFast() const override;
786 bool hasBitPreservingFPLogic(EVT VT) const override {
787 return VT == MVT::f32 || VT == MVT::f64 || VT.isVector();
790 bool isMultiStoresCheaperThanBitsMerge(EVT LTy, EVT HTy) const override {
791 // If the pair to store is a mixture of float and int values, we will
792 // save two bitwise instructions and one float-to-int instruction and
793 // increase one store instruction. There is potentially a more
794 // significant benefit because it avoids the float->int domain switch
795 // for input value. So It is more likely a win.
796 if ((LTy.isFloatingPoint() && HTy.isInteger()) ||
797 (LTy.isInteger() && HTy.isFloatingPoint()))
799 // If the pair only contains int values, we will save two bitwise
800 // instructions and increase one store instruction (costing one more
801 // store buffer). Since the benefit is more blurred so we leave
802 // such pair out until we get testcase to prove it is a win.
806 bool isMaskAndCmp0FoldingBeneficial(const Instruction &AndI) const override;
808 bool hasAndNotCompare(SDValue Y) const override;
810 bool hasAndNot(SDValue Y) const override;
812 bool preferShiftsToClearExtremeBits(SDValue Y) const override;
815 shouldTransformSignedTruncationCheck(EVT XVT,
816 unsigned KeptBits) const override {
817 // For vectors, we don't have a preference..
821 auto VTIsOk = [](EVT VT) -> bool {
822 return VT == MVT::i8 || VT == MVT::i16 || VT == MVT::i32 ||
826 // We are ok with KeptBitsVT being byte/word/dword, what MOVS supports.
827 // XVT will be larger than KeptBitsVT.
828 MVT KeptBitsVT = MVT::getIntegerVT(KeptBits);
829 return VTIsOk(XVT) && VTIsOk(KeptBitsVT);
832 bool shouldExpandShift(SelectionDAG &DAG, SDNode *N) const override {
833 if (DAG.getMachineFunction().getFunction().optForMinSize())
838 bool shouldSplatInsEltVarIndex(EVT VT) const override;
840 bool convertSetCCLogicToBitwiseLogic(EVT VT) const override {
841 return VT.isScalarInteger();
844 /// Vector-sized comparisons are fast using PCMPEQ + PMOVMSK or PTEST.
845 MVT hasFastEqualityCompare(unsigned NumBits) const override;
847 /// Allow multiple load pairs per block for smaller and faster code.
848 unsigned getMemcmpEqZeroLoadsPerBlock() const override {
852 /// Return the value type to use for ISD::SETCC.
853 EVT getSetCCResultType(const DataLayout &DL, LLVMContext &Context,
854 EVT VT) const override;
856 bool targetShrinkDemandedConstant(SDValue Op, const APInt &Demanded,
857 TargetLoweringOpt &TLO) const override;
859 /// Determine which of the bits specified in Mask are known to be either
860 /// zero or one and return them in the KnownZero/KnownOne bitsets.
861 void computeKnownBitsForTargetNode(const SDValue Op,
863 const APInt &DemandedElts,
864 const SelectionDAG &DAG,
865 unsigned Depth = 0) const override;
867 /// Determine the number of bits in the operation that are sign bits.
868 unsigned ComputeNumSignBitsForTargetNode(SDValue Op,
869 const APInt &DemandedElts,
870 const SelectionDAG &DAG,
871 unsigned Depth) const override;
873 bool SimplifyDemandedVectorEltsForTargetNode(SDValue Op,
874 const APInt &DemandedElts,
877 TargetLoweringOpt &TLO,
878 unsigned Depth) const override;
880 bool SimplifyDemandedBitsForTargetNode(SDValue Op,
881 const APInt &DemandedBits,
882 const APInt &DemandedElts,
884 TargetLoweringOpt &TLO,
885 unsigned Depth) const override;
887 SDValue unwrapAddress(SDValue N) const override;
889 SDValue getReturnAddressFrameIndex(SelectionDAG &DAG) const;
891 bool ExpandInlineAsm(CallInst *CI) const override;
893 ConstraintType getConstraintType(StringRef Constraint) const override;
895 /// Examine constraint string and operand type and determine a weight value.
896 /// The operand object must already have been set up with the operand type.
898 getSingleConstraintMatchWeight(AsmOperandInfo &info,
899 const char *constraint) const override;
901 const char *LowerXConstraint(EVT ConstraintVT) const override;
903 /// Lower the specified operand into the Ops vector. If it is invalid, don't
904 /// add anything to Ops. If hasMemory is true it means one of the asm
905 /// constraint of the inline asm instruction being processed is 'm'.
906 void LowerAsmOperandForConstraint(SDValue Op,
907 std::string &Constraint,
908 std::vector<SDValue> &Ops,
909 SelectionDAG &DAG) const override;
912 getInlineAsmMemConstraint(StringRef ConstraintCode) const override {
913 if (ConstraintCode == "i")
914 return InlineAsm::Constraint_i;
915 else if (ConstraintCode == "o")
916 return InlineAsm::Constraint_o;
917 else if (ConstraintCode == "v")
918 return InlineAsm::Constraint_v;
919 else if (ConstraintCode == "X")
920 return InlineAsm::Constraint_X;
921 return TargetLowering::getInlineAsmMemConstraint(ConstraintCode);
924 /// Handle Lowering flag assembly outputs.
925 SDValue LowerAsmOutputForConstraint(SDValue &Chain, SDValue &Flag, SDLoc DL,
926 const AsmOperandInfo &Constraint,
927 SelectionDAG &DAG) const override;
929 /// Given a physical register constraint
930 /// (e.g. {edx}), return the register number and the register class for the
931 /// register. This should only be used for C_Register constraints. On
932 /// error, this returns a register number of 0.
933 std::pair<unsigned, const TargetRegisterClass *>
934 getRegForInlineAsmConstraint(const TargetRegisterInfo *TRI,
935 StringRef Constraint, MVT VT) const override;
937 /// Return true if the addressing mode represented
938 /// by AM is legal for this target, for a load/store of the specified type.
939 bool isLegalAddressingMode(const DataLayout &DL, const AddrMode &AM,
940 Type *Ty, unsigned AS,
941 Instruction *I = nullptr) const override;
943 /// Return true if the specified immediate is legal
944 /// icmp immediate, that is the target has icmp instructions which can
945 /// compare a register against the immediate without having to materialize
946 /// the immediate into a register.
947 bool isLegalICmpImmediate(int64_t Imm) const override;
949 /// Return true if the specified immediate is legal
950 /// add immediate, that is the target has add instructions which can
951 /// add a register and the immediate without having to materialize
952 /// the immediate into a register.
953 bool isLegalAddImmediate(int64_t Imm) const override;
955 bool isLegalStoreImmediate(int64_t Imm) const override;
957 /// Return the cost of the scaling factor used in the addressing
958 /// mode represented by AM for this target, for a load/store
959 /// of the specified type.
960 /// If the AM is supported, the return value must be >= 0.
961 /// If the AM is not supported, it returns a negative value.
962 int getScalingFactorCost(const DataLayout &DL, const AddrMode &AM, Type *Ty,
963 unsigned AS) const override;
965 bool isVectorShiftByScalarCheap(Type *Ty) const override;
967 /// Return true if it's free to truncate a value of
968 /// type Ty1 to type Ty2. e.g. On x86 it's free to truncate a i32 value in
969 /// register EAX to i16 by referencing its sub-register AX.
970 bool isTruncateFree(Type *Ty1, Type *Ty2) const override;
971 bool isTruncateFree(EVT VT1, EVT VT2) const override;
973 bool allowTruncateForTailCall(Type *Ty1, Type *Ty2) const override;
975 /// Return true if any actual instruction that defines a
976 /// value of type Ty1 implicit zero-extends the value to Ty2 in the result
977 /// register. This does not necessarily include registers defined in
978 /// unknown ways, such as incoming arguments, or copies from unknown
979 /// virtual registers. Also, if isTruncateFree(Ty2, Ty1) is true, this
980 /// does not necessarily apply to truncate instructions. e.g. on x86-64,
981 /// all instructions that define 32-bit values implicit zero-extend the
982 /// result out to 64 bits.
983 bool isZExtFree(Type *Ty1, Type *Ty2) const override;
984 bool isZExtFree(EVT VT1, EVT VT2) const override;
985 bool isZExtFree(SDValue Val, EVT VT2) const override;
987 /// Return true if folding a vector load into ExtVal (a sign, zero, or any
988 /// extend node) is profitable.
989 bool isVectorLoadExtDesirable(SDValue) const override;
991 /// Return true if an FMA operation is faster than a pair of fmul and fadd
992 /// instructions. fmuladd intrinsics will be expanded to FMAs when this
993 /// method returns true, otherwise fmuladd is expanded to fmul + fadd.
994 bool isFMAFasterThanFMulAndFAdd(EVT VT) const override;
996 /// Return true if it's profitable to narrow
997 /// operations of type VT1 to VT2. e.g. on x86, it's profitable to narrow
998 /// from i32 to i8 but not from i32 to i16.
999 bool isNarrowingProfitable(EVT VT1, EVT VT2) const override;
1001 /// Given an intrinsic, checks if on the target the intrinsic will need to map
1002 /// to a MemIntrinsicNode (touches memory). If this is the case, it returns
1003 /// true and stores the intrinsic information into the IntrinsicInfo that was
1004 /// passed to the function.
1005 bool getTgtMemIntrinsic(IntrinsicInfo &Info, const CallInst &I,
1006 MachineFunction &MF,
1007 unsigned Intrinsic) const override;
1009 /// Returns true if the target can instruction select the
1010 /// specified FP immediate natively. If false, the legalizer will
1011 /// materialize the FP immediate as a load from a constant pool.
1012 bool isFPImmLegal(const APFloat &Imm, EVT VT) const override;
1014 /// Targets can use this to indicate that they only support *some*
1015 /// VECTOR_SHUFFLE operations, those with specific masks. By default, if a
1016 /// target supports the VECTOR_SHUFFLE node, all mask values are assumed to
1018 bool isShuffleMaskLegal(ArrayRef<int> Mask, EVT VT) const override;
1020 /// Similar to isShuffleMaskLegal. Targets can use this to indicate if there
1021 /// is a suitable VECTOR_SHUFFLE that can be used to replace a VAND with a
1022 /// constant pool entry.
1023 bool isVectorClearMaskLegal(ArrayRef<int> Mask, EVT VT) const override;
1025 /// Returns true if lowering to a jump table is allowed.
1026 bool areJTsAllowed(const Function *Fn) const override;
1028 /// If true, then instruction selection should
1029 /// seek to shrink the FP constant of the specified type to a smaller type
1030 /// in order to save space and / or reduce runtime.
1031 bool ShouldShrinkFPConstant(EVT VT) const override {
1032 // Don't shrink FP constpool if SSE2 is available since cvtss2sd is more
1033 // expensive than a straight movsd. On the other hand, it's important to
1034 // shrink long double fp constant since fldt is very slow.
1035 return !X86ScalarSSEf64 || VT == MVT::f80;
1038 /// Return true if we believe it is correct and profitable to reduce the
1039 /// load node to a smaller type.
1040 bool shouldReduceLoadWidth(SDNode *Load, ISD::LoadExtType ExtTy,
1041 EVT NewVT) const override;
1043 /// Return true if the specified scalar FP type is computed in an SSE
1044 /// register, not on the X87 floating point stack.
1045 bool isScalarFPTypeInSSEReg(EVT VT) const {
1046 return (VT == MVT::f64 && X86ScalarSSEf64) || // f64 is when SSE2
1047 (VT == MVT::f32 && X86ScalarSSEf32); // f32 is when SSE1
1050 /// Returns true if it is beneficial to convert a load of a constant
1051 /// to just the constant itself.
1052 bool shouldConvertConstantLoadToIntImm(const APInt &Imm,
1053 Type *Ty) const override;
1055 bool reduceSelectOfFPConstantLoads(bool IsFPSetCC) const override;
1057 bool convertSelectOfConstantsToMath(EVT VT) const override;
1059 bool decomposeMulByConstant(EVT VT, SDValue C) const override;
1061 bool shouldUseStrictFP_TO_INT(EVT FpVT, EVT IntVT,
1062 bool IsSigned) const override;
1064 /// Return true if EXTRACT_SUBVECTOR is cheap for this result type
1065 /// with this index.
1066 bool isExtractSubvectorCheap(EVT ResVT, EVT SrcVT,
1067 unsigned Index) const override;
1069 /// Scalar ops always have equal or better analysis/performance/power than
1070 /// the vector equivalent, so this always makes sense if the scalar op is
1072 bool shouldScalarizeBinop(SDValue) const override;
1074 bool storeOfVectorConstantIsCheap(EVT MemVT, unsigned NumElem,
1075 unsigned AddrSpace) const override {
1076 // If we can replace more than 2 scalar stores, there will be a reduction
1077 // in instructions even after we add a vector constant load.
1081 bool isLoadBitCastBeneficial(EVT LoadVT, EVT BitcastVT) const override;
1083 /// Intel processors have a unified instruction and data cache
1084 const char * getClearCacheBuiltinName() const override {
1085 return nullptr; // nothing to do, move along.
1088 unsigned getRegisterByName(const char* RegName, EVT VT,
1089 SelectionDAG &DAG) const override;
1091 /// If a physical register, this returns the register that receives the
1092 /// exception address on entry to an EH pad.
1094 getExceptionPointerRegister(const Constant *PersonalityFn) const override;
1096 /// If a physical register, this returns the register that receives the
1097 /// exception typeid on entry to a landing pad.
1099 getExceptionSelectorRegister(const Constant *PersonalityFn) const override;
1101 virtual bool needsFixedCatchObjects() const override;
1103 /// This method returns a target specific FastISel object,
1104 /// or null if the target does not support "fast" ISel.
1105 FastISel *createFastISel(FunctionLoweringInfo &funcInfo,
1106 const TargetLibraryInfo *libInfo) const override;
1108 /// If the target has a standard location for the stack protector cookie,
1109 /// returns the address of that location. Otherwise, returns nullptr.
1110 Value *getIRStackGuard(IRBuilder<> &IRB) const override;
1112 bool useLoadStackGuardNode() const override;
1113 bool useStackGuardXorFP() const override;
1114 void insertSSPDeclarations(Module &M) const override;
1115 Value *getSDagStackGuard(const Module &M) const override;
1116 Function *getSSPStackGuardCheck(const Module &M) const override;
1117 SDValue emitStackGuardXorFP(SelectionDAG &DAG, SDValue Val,
1118 const SDLoc &DL) const override;
1121 /// Return true if the target stores SafeStack pointer at a fixed offset in
1122 /// some non-standard address space, and populates the address space and
1123 /// offset as appropriate.
1124 Value *getSafeStackPointerLocation(IRBuilder<> &IRB) const override;
1126 SDValue BuildFILD(SDValue Op, EVT SrcVT, SDValue Chain, SDValue StackSlot,
1127 SelectionDAG &DAG) const;
1129 bool isNoopAddrSpaceCast(unsigned SrcAS, unsigned DestAS) const override;
1131 /// Customize the preferred legalization strategy for certain types.
1132 LegalizeTypeAction getPreferredVectorAction(MVT VT) const override;
1134 MVT getRegisterTypeForCallingConv(LLVMContext &Context, CallingConv::ID CC,
1135 EVT VT) const override;
1137 unsigned getNumRegistersForCallingConv(LLVMContext &Context,
1139 EVT VT) const override;
1141 bool isIntDivCheap(EVT VT, AttributeList Attr) const override;
1143 bool supportSwiftError() const override;
1145 StringRef getStackProbeSymbolName(MachineFunction &MF) const override;
1147 bool hasVectorBlend() const override { return true; }
1149 unsigned getMaxSupportedInterleaveFactor() const override { return 4; }
1151 /// Lower interleaved load(s) into target specific
1152 /// instructions/intrinsics.
1153 bool lowerInterleavedLoad(LoadInst *LI,
1154 ArrayRef<ShuffleVectorInst *> Shuffles,
1155 ArrayRef<unsigned> Indices,
1156 unsigned Factor) const override;
1158 /// Lower interleaved store(s) into target specific
1159 /// instructions/intrinsics.
1160 bool lowerInterleavedStore(StoreInst *SI, ShuffleVectorInst *SVI,
1161 unsigned Factor) const override;
1163 SDValue expandIndirectJTBranch(const SDLoc& dl, SDValue Value,
1164 SDValue Addr, SelectionDAG &DAG)
1168 std::pair<const TargetRegisterClass *, uint8_t>
1169 findRepresentativeClass(const TargetRegisterInfo *TRI,
1170 MVT VT) const override;
1173 /// Keep a reference to the X86Subtarget around so that we can
1174 /// make the right decision when generating code for different targets.
1175 const X86Subtarget &Subtarget;
1177 /// Select between SSE or x87 floating point ops.
1178 /// When SSE is available, use it for f32 operations.
1179 /// When SSE2 is available, use it for f64 operations.
1180 bool X86ScalarSSEf32;
1181 bool X86ScalarSSEf64;
1183 /// A list of legal FP immediates.
1184 std::vector<APFloat> LegalFPImmediates;
1186 /// Indicate that this x86 target can instruction
1187 /// select the specified FP immediate natively.
1188 void addLegalFPImmediate(const APFloat& Imm) {
1189 LegalFPImmediates.push_back(Imm);
1192 SDValue LowerCallResult(SDValue Chain, SDValue InFlag,
1193 CallingConv::ID CallConv, bool isVarArg,
1194 const SmallVectorImpl<ISD::InputArg> &Ins,
1195 const SDLoc &dl, SelectionDAG &DAG,
1196 SmallVectorImpl<SDValue> &InVals,
1197 uint32_t *RegMask) const;
1198 SDValue LowerMemArgument(SDValue Chain, CallingConv::ID CallConv,
1199 const SmallVectorImpl<ISD::InputArg> &ArgInfo,
1200 const SDLoc &dl, SelectionDAG &DAG,
1201 const CCValAssign &VA, MachineFrameInfo &MFI,
1203 SDValue LowerMemOpCallTo(SDValue Chain, SDValue StackPtr, SDValue Arg,
1204 const SDLoc &dl, SelectionDAG &DAG,
1205 const CCValAssign &VA,
1206 ISD::ArgFlagsTy Flags) const;
1208 // Call lowering helpers.
1210 /// Check whether the call is eligible for tail call optimization. Targets
1211 /// that want to do tail call optimization should implement this function.
1212 bool IsEligibleForTailCallOptimization(SDValue Callee,
1213 CallingConv::ID CalleeCC,
1215 bool isCalleeStructRet,
1216 bool isCallerStructRet,
1218 const SmallVectorImpl<ISD::OutputArg> &Outs,
1219 const SmallVectorImpl<SDValue> &OutVals,
1220 const SmallVectorImpl<ISD::InputArg> &Ins,
1221 SelectionDAG& DAG) const;
1222 SDValue EmitTailCallLoadRetAddr(SelectionDAG &DAG, SDValue &OutRetAddr,
1223 SDValue Chain, bool IsTailCall,
1224 bool Is64Bit, int FPDiff,
1225 const SDLoc &dl) const;
1227 unsigned GetAlignedArgumentStackSize(unsigned StackSize,
1228 SelectionDAG &DAG) const;
1230 unsigned getAddressSpace(void) const;
1232 SDValue FP_TO_INTHelper(SDValue Op, SelectionDAG &DAG, bool isSigned) const;
1234 SDValue LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) const;
1235 SDValue LowerVSELECT(SDValue Op, SelectionDAG &DAG) const;
1236 SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const;
1237 SDValue LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const;
1239 unsigned getGlobalWrapperKind(const GlobalValue *GV = nullptr,
1240 const unsigned char OpFlags = 0) const;
1241 SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) const;
1242 SDValue LowerBlockAddress(SDValue Op, SelectionDAG &DAG) const;
1243 SDValue LowerGlobalAddress(const GlobalValue *GV, const SDLoc &dl,
1244 int64_t Offset, SelectionDAG &DAG) const;
1245 SDValue LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) const;
1246 SDValue LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const;
1247 SDValue LowerExternalSymbol(SDValue Op, SelectionDAG &DAG) const;
1249 SDValue LowerSINT_TO_FP(SDValue Op, SelectionDAG &DAG) const;
1250 SDValue LowerUINT_TO_FP(SDValue Op, SelectionDAG &DAG) const;
1251 SDValue LowerTRUNCATE(SDValue Op, SelectionDAG &DAG) const;
1252 SDValue LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG) const;
1253 SDValue LowerSETCC(SDValue Op, SelectionDAG &DAG) const;
1254 SDValue LowerSETCCCARRY(SDValue Op, SelectionDAG &DAG) const;
1255 SDValue LowerSELECT(SDValue Op, SelectionDAG &DAG) const;
1256 SDValue LowerBRCOND(SDValue Op, SelectionDAG &DAG) const;
1257 SDValue LowerJumpTable(SDValue Op, SelectionDAG &DAG) const;
1258 SDValue LowerDYNAMIC_STACKALLOC(SDValue Op, SelectionDAG &DAG) const;
1259 SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG) const;
1260 SDValue LowerVAARG(SDValue Op, SelectionDAG &DAG) const;
1261 SDValue LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const;
1262 SDValue LowerADDROFRETURNADDR(SDValue Op, SelectionDAG &DAG) const;
1263 SDValue LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const;
1264 SDValue LowerFRAME_TO_ARGS_OFFSET(SDValue Op, SelectionDAG &DAG) const;
1265 SDValue LowerEH_RETURN(SDValue Op, SelectionDAG &DAG) const;
1266 SDValue lowerEH_SJLJ_SETJMP(SDValue Op, SelectionDAG &DAG) const;
1267 SDValue lowerEH_SJLJ_LONGJMP(SDValue Op, SelectionDAG &DAG) const;
1268 SDValue lowerEH_SJLJ_SETUP_DISPATCH(SDValue Op, SelectionDAG &DAG) const;
1269 SDValue LowerINIT_TRAMPOLINE(SDValue Op, SelectionDAG &DAG) const;
1270 SDValue LowerFLT_ROUNDS_(SDValue Op, SelectionDAG &DAG) const;
1271 SDValue LowerWin64_i128OP(SDValue Op, SelectionDAG &DAG) const;
1272 SDValue LowerGC_TRANSITION_START(SDValue Op, SelectionDAG &DAG) const;
1273 SDValue LowerGC_TRANSITION_END(SDValue Op, SelectionDAG &DAG) const;
1274 SDValue LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) const;
1277 LowerFormalArguments(SDValue Chain, CallingConv::ID CallConv, bool isVarArg,
1278 const SmallVectorImpl<ISD::InputArg> &Ins,
1279 const SDLoc &dl, SelectionDAG &DAG,
1280 SmallVectorImpl<SDValue> &InVals) const override;
1281 SDValue LowerCall(CallLoweringInfo &CLI,
1282 SmallVectorImpl<SDValue> &InVals) const override;
1284 SDValue LowerReturn(SDValue Chain, CallingConv::ID CallConv, bool isVarArg,
1285 const SmallVectorImpl<ISD::OutputArg> &Outs,
1286 const SmallVectorImpl<SDValue> &OutVals,
1287 const SDLoc &dl, SelectionDAG &DAG) const override;
1289 bool supportSplitCSR(MachineFunction *MF) const override {
1290 return MF->getFunction().getCallingConv() == CallingConv::CXX_FAST_TLS &&
1291 MF->getFunction().hasFnAttribute(Attribute::NoUnwind);
1293 void initializeSplitCSR(MachineBasicBlock *Entry) const override;
1294 void insertCopiesSplitCSR(
1295 MachineBasicBlock *Entry,
1296 const SmallVectorImpl<MachineBasicBlock *> &Exits) const override;
1298 bool isUsedByReturnOnly(SDNode *N, SDValue &Chain) const override;
1300 bool mayBeEmittedAsTailCall(const CallInst *CI) const override;
1302 EVT getTypeForExtReturn(LLVMContext &Context, EVT VT,
1303 ISD::NodeType ExtendKind) const override;
1305 bool CanLowerReturn(CallingConv::ID CallConv, MachineFunction &MF,
1307 const SmallVectorImpl<ISD::OutputArg> &Outs,
1308 LLVMContext &Context) const override;
1310 const MCPhysReg *getScratchRegisters(CallingConv::ID CC) const override;
1312 TargetLoweringBase::AtomicExpansionKind
1313 shouldExpandAtomicLoadInIR(LoadInst *SI) const override;
1314 bool shouldExpandAtomicStoreInIR(StoreInst *SI) const override;
1315 TargetLoweringBase::AtomicExpansionKind
1316 shouldExpandAtomicRMWInIR(AtomicRMWInst *AI) const override;
1319 lowerIdempotentRMWIntoFencedLoad(AtomicRMWInst *AI) const override;
1321 bool needsCmpXchgNb(Type *MemType) const;
1323 void SetupEntryBlockForSjLj(MachineInstr &MI, MachineBasicBlock *MBB,
1324 MachineBasicBlock *DispatchBB, int FI) const;
1326 // Utility function to emit the low-level va_arg code for X86-64.
1328 EmitVAARG64WithCustomInserter(MachineInstr &MI,
1329 MachineBasicBlock *MBB) const;
1331 /// Utility function to emit the xmm reg save portion of va_start.
1333 EmitVAStartSaveXMMRegsWithCustomInserter(MachineInstr &BInstr,
1334 MachineBasicBlock *BB) const;
1336 MachineBasicBlock *EmitLoweredCascadedSelect(MachineInstr &MI1,
1338 MachineBasicBlock *BB) const;
1340 MachineBasicBlock *EmitLoweredSelect(MachineInstr &I,
1341 MachineBasicBlock *BB) const;
1343 MachineBasicBlock *EmitLoweredAtomicFP(MachineInstr &I,
1344 MachineBasicBlock *BB) const;
1346 MachineBasicBlock *EmitLoweredCatchRet(MachineInstr &MI,
1347 MachineBasicBlock *BB) const;
1349 MachineBasicBlock *EmitLoweredCatchPad(MachineInstr &MI,
1350 MachineBasicBlock *BB) const;
1352 MachineBasicBlock *EmitLoweredSegAlloca(MachineInstr &MI,
1353 MachineBasicBlock *BB) const;
1355 MachineBasicBlock *EmitLoweredTLSAddr(MachineInstr &MI,
1356 MachineBasicBlock *BB) const;
1358 MachineBasicBlock *EmitLoweredTLSCall(MachineInstr &MI,
1359 MachineBasicBlock *BB) const;
1361 MachineBasicBlock *EmitLoweredRetpoline(MachineInstr &MI,
1362 MachineBasicBlock *BB) const;
1364 MachineBasicBlock *emitEHSjLjSetJmp(MachineInstr &MI,
1365 MachineBasicBlock *MBB) const;
1367 void emitSetJmpShadowStackFix(MachineInstr &MI,
1368 MachineBasicBlock *MBB) const;
1370 MachineBasicBlock *emitEHSjLjLongJmp(MachineInstr &MI,
1371 MachineBasicBlock *MBB) const;
1373 MachineBasicBlock *emitLongJmpShadowStackFix(MachineInstr &MI,
1374 MachineBasicBlock *MBB) const;
1376 MachineBasicBlock *emitFMA3Instr(MachineInstr &MI,
1377 MachineBasicBlock *MBB) const;
1379 MachineBasicBlock *EmitSjLjDispatchBlock(MachineInstr &MI,
1380 MachineBasicBlock *MBB) const;
1382 /// Emit nodes that will be selected as "cmp Op0,Op1", or something
1383 /// equivalent, for use with the given x86 condition code.
1384 SDValue EmitCmp(SDValue Op0, SDValue Op1, unsigned X86CC, const SDLoc &dl,
1385 SelectionDAG &DAG) const;
1387 /// Convert a comparison if required by the subtarget.
1388 SDValue ConvertCmpIfNecessary(SDValue Cmp, SelectionDAG &DAG) const;
1390 /// Emit flags for the given setcc condition and operands. Also returns the
1391 /// corresponding X86 condition code constant in X86CC.
1392 SDValue emitFlagsForSetcc(SDValue Op0, SDValue Op1,
1393 ISD::CondCode CC, const SDLoc &dl,
1395 SDValue &X86CC) const;
1397 /// Check if replacement of SQRT with RSQRT should be disabled.
1398 bool isFsqrtCheap(SDValue Operand, SelectionDAG &DAG) const override;
1400 /// Use rsqrt* to speed up sqrt calculations.
1401 SDValue getSqrtEstimate(SDValue Operand, SelectionDAG &DAG, int Enabled,
1402 int &RefinementSteps, bool &UseOneConstNR,
1403 bool Reciprocal) const override;
1405 /// Use rcp* to speed up fdiv calculations.
1406 SDValue getRecipEstimate(SDValue Operand, SelectionDAG &DAG, int Enabled,
1407 int &RefinementSteps) const override;
1409 /// Reassociate floating point divisions into multiply by reciprocal.
1410 unsigned combineRepeatedFPDivisors() const override;
1414 FastISel *createFastISel(FunctionLoweringInfo &funcInfo,
1415 const TargetLibraryInfo *libInfo);
1416 } // end namespace X86
1418 // Base class for all X86 non-masked store operations.
1419 class X86StoreSDNode : public MemSDNode {
1421 X86StoreSDNode(unsigned Opcode, unsigned Order, const DebugLoc &dl,
1422 SDVTList VTs, EVT MemVT,
1423 MachineMemOperand *MMO)
1424 :MemSDNode(Opcode, Order, dl, VTs, MemVT, MMO) {}
1425 const SDValue &getValue() const { return getOperand(1); }
1426 const SDValue &getBasePtr() const { return getOperand(2); }
1428 static bool classof(const SDNode *N) {
1429 return N->getOpcode() == X86ISD::VTRUNCSTORES ||
1430 N->getOpcode() == X86ISD::VTRUNCSTOREUS;
1434 // Base class for all X86 masked store operations.
1435 // The class has the same order of operands as MaskedStoreSDNode for
1437 class X86MaskedStoreSDNode : public MemSDNode {
1439 X86MaskedStoreSDNode(unsigned Opcode, unsigned Order,
1440 const DebugLoc &dl, SDVTList VTs, EVT MemVT,
1441 MachineMemOperand *MMO)
1442 : MemSDNode(Opcode, Order, dl, VTs, MemVT, MMO) {}
1444 const SDValue &getValue() const { return getOperand(1); }
1445 const SDValue &getBasePtr() const { return getOperand(2); }
1446 const SDValue &getMask() const { return getOperand(3); }
1448 static bool classof(const SDNode *N) {
1449 return N->getOpcode() == X86ISD::VMTRUNCSTORES ||
1450 N->getOpcode() == X86ISD::VMTRUNCSTOREUS;
1454 // X86 Truncating Store with Signed saturation.
1455 class TruncSStoreSDNode : public X86StoreSDNode {
1457 TruncSStoreSDNode(unsigned Order, const DebugLoc &dl,
1458 SDVTList VTs, EVT MemVT, MachineMemOperand *MMO)
1459 : X86StoreSDNode(X86ISD::VTRUNCSTORES, Order, dl, VTs, MemVT, MMO) {}
1461 static bool classof(const SDNode *N) {
1462 return N->getOpcode() == X86ISD::VTRUNCSTORES;
1466 // X86 Truncating Store with Unsigned saturation.
1467 class TruncUSStoreSDNode : public X86StoreSDNode {
1469 TruncUSStoreSDNode(unsigned Order, const DebugLoc &dl,
1470 SDVTList VTs, EVT MemVT, MachineMemOperand *MMO)
1471 : X86StoreSDNode(X86ISD::VTRUNCSTOREUS, Order, dl, VTs, MemVT, MMO) {}
1473 static bool classof(const SDNode *N) {
1474 return N->getOpcode() == X86ISD::VTRUNCSTOREUS;
1478 // X86 Truncating Masked Store with Signed saturation.
1479 class MaskedTruncSStoreSDNode : public X86MaskedStoreSDNode {
1481 MaskedTruncSStoreSDNode(unsigned Order,
1482 const DebugLoc &dl, SDVTList VTs, EVT MemVT,
1483 MachineMemOperand *MMO)
1484 : X86MaskedStoreSDNode(X86ISD::VMTRUNCSTORES, Order, dl, VTs, MemVT, MMO) {}
1486 static bool classof(const SDNode *N) {
1487 return N->getOpcode() == X86ISD::VMTRUNCSTORES;
1491 // X86 Truncating Masked Store with Unsigned saturation.
1492 class MaskedTruncUSStoreSDNode : public X86MaskedStoreSDNode {
1494 MaskedTruncUSStoreSDNode(unsigned Order,
1495 const DebugLoc &dl, SDVTList VTs, EVT MemVT,
1496 MachineMemOperand *MMO)
1497 : X86MaskedStoreSDNode(X86ISD::VMTRUNCSTOREUS, Order, dl, VTs, MemVT, MMO) {}
1499 static bool classof(const SDNode *N) {
1500 return N->getOpcode() == X86ISD::VMTRUNCSTOREUS;
1504 // X86 specific Gather/Scatter nodes.
1505 // The class has the same order of operands as MaskedGatherScatterSDNode for
1507 class X86MaskedGatherScatterSDNode : public MemSDNode {
1509 X86MaskedGatherScatterSDNode(unsigned Opc, unsigned Order,
1510 const DebugLoc &dl, SDVTList VTs, EVT MemVT,
1511 MachineMemOperand *MMO)
1512 : MemSDNode(Opc, Order, dl, VTs, MemVT, MMO) {}
1514 const SDValue &getBasePtr() const { return getOperand(3); }
1515 const SDValue &getIndex() const { return getOperand(4); }
1516 const SDValue &getMask() const { return getOperand(2); }
1517 const SDValue &getScale() const { return getOperand(5); }
1519 static bool classof(const SDNode *N) {
1520 return N->getOpcode() == X86ISD::MGATHER ||
1521 N->getOpcode() == X86ISD::MSCATTER;
1525 class X86MaskedGatherSDNode : public X86MaskedGatherScatterSDNode {
1527 X86MaskedGatherSDNode(unsigned Order, const DebugLoc &dl, SDVTList VTs,
1528 EVT MemVT, MachineMemOperand *MMO)
1529 : X86MaskedGatherScatterSDNode(X86ISD::MGATHER, Order, dl, VTs, MemVT,
1532 const SDValue &getPassThru() const { return getOperand(1); }
1534 static bool classof(const SDNode *N) {
1535 return N->getOpcode() == X86ISD::MGATHER;
1539 class X86MaskedScatterSDNode : public X86MaskedGatherScatterSDNode {
1541 X86MaskedScatterSDNode(unsigned Order, const DebugLoc &dl, SDVTList VTs,
1542 EVT MemVT, MachineMemOperand *MMO)
1543 : X86MaskedGatherScatterSDNode(X86ISD::MSCATTER, Order, dl, VTs, MemVT,
1546 const SDValue &getValue() const { return getOperand(1); }
1548 static bool classof(const SDNode *N) {
1549 return N->getOpcode() == X86ISD::MSCATTER;
1553 /// Generate unpacklo/unpackhi shuffle mask.
1554 template <typename T = int>
1555 void createUnpackShuffleMask(MVT VT, SmallVectorImpl<T> &Mask, bool Lo,
1557 assert(Mask.empty() && "Expected an empty shuffle mask vector");
1558 int NumElts = VT.getVectorNumElements();
1559 int NumEltsInLane = 128 / VT.getScalarSizeInBits();
1560 for (int i = 0; i < NumElts; ++i) {
1561 unsigned LaneStart = (i / NumEltsInLane) * NumEltsInLane;
1562 int Pos = (i % NumEltsInLane) / 2 + LaneStart;
1563 Pos += (Unary ? 0 : NumElts * (i % 2));
1564 Pos += (Lo ? 0 : NumEltsInLane / 2);
1565 Mask.push_back(Pos);
1569 /// Helper function to scale a shuffle or target shuffle mask, replacing each
1570 /// mask index with the scaled sequential indices for an equivalent narrowed
1571 /// mask. This is the reverse process to canWidenShuffleElements, but can
1573 template <typename T>
1574 void scaleShuffleMask(int Scale, ArrayRef<T> Mask,
1575 SmallVectorImpl<T> &ScaledMask) {
1576 assert(0 < Scale && "Unexpected scaling factor");
1577 int NumElts = Mask.size();
1578 ScaledMask.assign(static_cast<size_t>(NumElts * Scale), -1);
1580 for (int i = 0; i != NumElts; ++i) {
1583 // Repeat sentinel values in every mask element.
1585 for (int s = 0; s != Scale; ++s)
1586 ScaledMask[(Scale * i) + s] = M;
1590 // Scale mask element and increment across each mask element.
1591 for (int s = 0; s != Scale; ++s)
1592 ScaledMask[(Scale * i) + s] = (Scale * M) + s;
1595 } // end namespace llvm
1597 #endif // LLVM_LIB_TARGET_X86_X86ISELLOWERING_H