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"
23 class X86TargetMachine;
26 // X86 Specific DAG Nodes
27 enum NodeType : unsigned {
28 // Start the numbering where the builtin ops leave off.
29 FIRST_NUMBER = ISD::BUILTIN_OP_END,
36 /// Double shift instructions. These correspond to
37 /// X86::SHLDxx and X86::SHRDxx instructions.
41 /// Bitwise logical AND of floating point values. This corresponds
42 /// to X86::ANDPS or X86::ANDPD.
45 /// Bitwise logical OR of floating point values. This corresponds
46 /// to X86::ORPS or X86::ORPD.
49 /// Bitwise logical XOR of floating point values. This corresponds
50 /// to X86::XORPS or X86::XORPD.
53 /// Bitwise logical ANDNOT of floating point values. This
54 /// corresponds to X86::ANDNPS or X86::ANDNPD.
57 /// These operations represent an abstract X86 call
58 /// instruction, which includes a bunch of information. In particular the
59 /// operands of these node are:
61 /// #0 - The incoming token chain
63 /// #2 - The number of arg bytes the caller pushes on the stack.
64 /// #3 - The number of arg bytes the callee pops off the stack.
65 /// #4 - The value to pass in AL/AX/EAX (optional)
66 /// #5 - The value to pass in DL/DX/EDX (optional)
68 /// The result values of these nodes are:
70 /// #0 - The outgoing token chain
71 /// #1 - The first register result value (optional)
72 /// #2 - The second register result value (optional)
76 /// Same as call except it adds the NoTrack prefix.
79 /// X86 compare and logical compare instructions.
82 /// X86 bit-test instructions.
85 /// X86 SetCC. Operand 0 is condition code, and operand 1 is the EFLAGS
86 /// operand, usually produced by a CMP instruction.
92 // Same as SETCC except it's materialized with a sbb and the value is all
93 // one's or all zero's.
94 SETCC_CARRY, // R = carry_bit ? ~0 : 0
96 /// X86 FP SETCC, implemented with CMP{cc}SS/CMP{cc}SD.
97 /// Operands are two FP values to compare; result is a mask of
98 /// 0s or 1s. Generally DTRT for C/C++ with NaNs.
101 /// X86 FP SETCC, similar to above, but with output as an i1 mask and
102 /// and a version with SAE.
103 FSETCCM, FSETCCM_SAE,
105 /// X86 conditional moves. Operand 0 and operand 1 are the two values
106 /// to select from. Operand 2 is the condition code, and operand 3 is the
107 /// flag operand produced by a CMP or TEST instruction.
110 /// X86 conditional branches. Operand 0 is the chain operand, operand 1
111 /// is the block to branch if condition is true, operand 2 is the
112 /// condition code, and operand 3 is the flag operand produced by a CMP
113 /// or TEST instruction.
116 /// BRIND node with NoTrack prefix. Operand 0 is the chain operand and
117 /// operand 1 is the target address.
120 /// Return with a flag operand. Operand 0 is the chain operand, operand
121 /// 1 is the number of bytes of stack to pop.
124 /// Return from interrupt. Operand 0 is the number of bytes to pop.
127 /// Repeat fill, corresponds to X86::REP_STOSx.
130 /// Repeat move, corresponds to X86::REP_MOVSx.
133 /// On Darwin, this node represents the result of the popl
134 /// at function entry, used for PIC code.
137 /// A wrapper node for TargetConstantPool, TargetJumpTable,
138 /// TargetExternalSymbol, TargetGlobalAddress, TargetGlobalTLSAddress,
139 /// MCSymbol and TargetBlockAddress.
142 /// Special wrapper used under X86-64 PIC mode for RIP
143 /// relative displacements.
146 /// Copies a 64-bit value from an MMX vector to the low word
147 /// of an XMM vector, with the high word zero filled.
150 /// Copies a 64-bit value from the low word of an XMM vector
151 /// to an MMX vector.
154 /// Copies a 32-bit value from the low word of a MMX
158 /// Copies a GPR into the low 32-bit word of a MMX vector
159 /// and zero out the high word.
162 /// Extract an 8-bit value from a vector and zero extend it to
163 /// i32, corresponds to X86::PEXTRB.
166 /// Extract a 16-bit value from a vector and zero extend it to
167 /// i32, corresponds to X86::PEXTRW.
170 /// Insert any element of a 4 x float vector into any element
171 /// of a destination 4 x floatvector.
174 /// Insert the lower 8-bits of a 32-bit value to a vector,
175 /// corresponds to X86::PINSRB.
178 /// Insert the lower 16-bits of a 32-bit value to a vector,
179 /// corresponds to X86::PINSRW.
182 /// Shuffle 16 8-bit values within a vector.
185 /// Compute Sum of Absolute Differences.
187 /// Compute Double Block Packed Sum-Absolute-Differences
190 /// Bitwise Logical AND NOT of Packed FP values.
193 /// Blend where the selector is an immediate.
196 /// Dynamic (non-constant condition) vector blend where only the sign bits
197 /// of the condition elements are used. This is used to enforce that the
198 /// condition mask is not valid for generic VSELECT optimizations. This
199 /// is also used to implement the intrinsics.
200 /// Operands are in VSELECT order: MASK, TRUE, FALSE
203 /// Combined add and sub on an FP vector.
206 // FP vector ops with rounding mode.
207 FADD_RND, FADDS, FADDS_RND,
208 FSUB_RND, FSUBS, FSUBS_RND,
209 FMUL_RND, FMULS, FMULS_RND,
210 FDIV_RND, FDIVS, FDIVS_RND,
213 FSQRT_RND, FSQRTS, FSQRTS_RND,
215 // FP vector get exponent.
216 FGETEXP, FGETEXP_SAE, FGETEXPS, FGETEXPS_SAE,
217 // Extract Normalized Mantissas.
218 VGETMANT, VGETMANT_SAE, VGETMANTS, VGETMANTS_SAE,
221 SCALEFS, SCALEFS_RND,
223 // Unsigned Integer average.
226 /// Integer horizontal add/sub.
230 /// Floating point horizontal add/sub.
234 // Detect Conflicts Within a Vector
237 /// Floating point max and min.
240 /// Commutative FMIN and FMAX.
243 /// Scalar intrinsic floating point max and min.
246 /// Floating point reciprocal-sqrt and reciprocal approximation.
247 /// Note that these typically require refinement
248 /// in order to obtain suitable precision.
251 // AVX-512 reciprocal approximations with a little more precision.
252 RSQRT14, RSQRT14S, RCP14, RCP14S,
254 // Thread Local Storage.
257 // Thread Local Storage. A call to get the start address
258 // of the TLS block for the current module.
261 // Thread Local Storage. When calling to an OS provided
262 // thunk at the address from an earlier relocation.
265 // Exception Handling helpers.
268 // SjLj exception handling setjmp.
271 // SjLj exception handling longjmp.
274 // SjLj exception handling dispatch.
275 EH_SJLJ_SETUP_DISPATCH,
277 /// Tail call return. See X86TargetLowering::LowerCall for
278 /// the list of operands.
281 // Vector move to low scalar and zero higher vector elements.
284 // Vector integer truncate.
286 // Vector integer truncate with unsigned/signed saturation.
289 // Masked version of the above. Used when less than a 128-bit result is
290 // produced since the mask only applies to the lower elements and can't
291 // be represented by a select.
292 // SRC, PASSTHRU, MASK
293 VMTRUNC, VMTRUNCUS, VMTRUNCS,
296 VFPEXT, VFPEXT_SAE, VFPEXTS, VFPEXTS_SAE,
299 VFPROUND, VFPROUND_RND, VFPROUNDS, VFPROUNDS_RND,
301 // Masked version of above. Used for v2f64->v4f32.
302 // SRC, PASSTHRU, MASK
305 // 128-bit vector logical left / right shift
308 // Vector shift elements
311 // Vector variable shift
314 // Vector shift elements by immediate
317 // Shifts of mask registers.
320 // Bit rotate by immediate
323 // Vector packed double/float comparison.
326 // Vector integer comparisons.
329 // v8i16 Horizontal minimum and position.
334 /// Vector comparison generating mask bits for fp and
335 /// integer signed and unsigned data types.
337 // Vector comparison with SAE for FP values
340 // Arithmetic operations with FLAGS results.
341 ADD, SUB, ADC, SBB, SMUL, UMUL,
344 // Bit field extract.
347 // Zero High Bits Starting with Specified Bit Position.
350 // X86-specific multiply by immediate.
353 // Vector sign bit extraction.
356 // Vector bitwise comparisons.
359 // Vector packed fp sign bitwise comparisons.
362 // OR/AND test for masks.
369 // Several flavors of instructions with vector shuffle behaviors.
370 // Saturated signed/unnsigned packing.
373 // Intra-lane alignr.
375 // AVX512 inter-lane alignr.
381 // VBMI2 Concat & Shift.
386 //Shuffle Packed Values at 128-bit granularity.
402 // Variable Permute (VPERM).
403 // Res = VPERMV MaskV, V0
406 // 3-op Variable Permute (VPERMT2).
407 // Res = VPERMV3 V0, MaskV, V1
410 // Bitwise ternary logic.
412 // Fix Up Special Packed Float32/64 values.
413 VFIXUPIMM, VFIXUPIMM_SAE,
414 VFIXUPIMMS, VFIXUPIMMS_SAE,
415 // Range Restriction Calculation For Packed Pairs of Float32/64 values.
416 VRANGE, VRANGE_SAE, VRANGES, VRANGES_SAE,
417 // Reduce - Perform Reduction Transformation on scalar\packed FP.
418 VREDUCE, VREDUCE_SAE, VREDUCES, VREDUCES_SAE,
419 // RndScale - Round FP Values To Include A Given Number Of Fraction Bits.
420 // Also used by the legacy (V)ROUND intrinsics where we mask out the
421 // scaling part of the immediate.
422 VRNDSCALE, VRNDSCALE_SAE, VRNDSCALES, VRNDSCALES_SAE,
423 // Tests Types Of a FP Values for packed types.
425 // Tests Types Of a FP Values for scalar types.
428 // Broadcast (splat) scalar or element 0 of a vector. If the operand is
429 // a vector, this node may change the vector length as part of the splat.
431 // Broadcast mask to vector.
433 // Broadcast subvector to vector.
436 /// SSE4A Extraction and Insertion.
439 // XOP arithmetic/logical shifts.
441 // XOP signed/unsigned integer comparisons.
443 // XOP packed permute bytes.
445 // XOP two source permutation.
448 // Vector multiply packed unsigned doubleword integers.
450 // Vector multiply packed signed doubleword integers.
452 // Vector Multiply Packed UnsignedIntegers with Round and Scale.
455 // Multiply and Add Packed Integers.
456 VPMADDUBSW, VPMADDWD,
458 // AVX512IFMA multiply and add.
459 // NOTE: These are different than the instruction and perform
461 VPMADD52L, VPMADD52H,
470 // We use the target independent ISD::FMA for the non-inverted case.
477 // FMA with rounding mode.
485 // Compress and expand.
492 // Convert Unsigned/Integer to Floating-Point Value with rounding mode.
493 SINT_TO_FP_RND, UINT_TO_FP_RND,
494 SCALAR_SINT_TO_FP, SCALAR_UINT_TO_FP,
495 SCALAR_SINT_TO_FP_RND, SCALAR_UINT_TO_FP_RND,
497 // Vector float/double to signed/unsigned integer.
498 CVTP2SI, CVTP2UI, CVTP2SI_RND, CVTP2UI_RND,
499 // Scalar float/double to signed/unsigned integer.
500 CVTS2SI, CVTS2UI, CVTS2SI_RND, CVTS2UI_RND,
502 // Vector float/double to signed/unsigned integer with truncation.
503 CVTTP2SI, CVTTP2UI, CVTTP2SI_SAE, CVTTP2UI_SAE,
504 // Scalar float/double to signed/unsigned integer with truncation.
505 CVTTS2SI, CVTTS2UI, CVTTS2SI_SAE, CVTTS2UI_SAE,
507 // Vector signed/unsigned integer to float/double.
510 // Masked versions of above. Used for v2f64->v4f32.
511 // SRC, PASSTHRU, MASK
512 MCVTP2SI, MCVTP2UI, MCVTTP2SI, MCVTTP2UI,
515 // Vector float to bfloat16.
516 // Convert TWO packed single data to one packed BF16 data
518 // Convert packed single data to packed BF16 data
520 // Masked version of above.
521 // SRC, PASSTHRU, MASK
524 // Dot product of BF16 pairs to accumulated into
525 // packed single precision.
528 // Save xmm argument registers to the stack, according to %al. An operator
529 // is needed so that this can be expanded with control flow.
530 VASTART_SAVE_XMM_REGS,
532 // Windows's _chkstk call to do stack probing.
535 // For allocating variable amounts of stack space when using
536 // segmented stacks. Check if the current stacklet has enough space, and
537 // falls back to heap allocation if not.
544 // Store FP status word into i16 register.
547 // Store contents of %ah into %eflags.
550 // Get a random integer and indicate whether it is valid in CF.
553 // Get a NIST SP800-90B & C compliant random integer and
554 // indicate whether it is valid in CF.
558 // RDPKRU - Operand 0 is chain. Operand 1 is value for ECX.
559 // WRPKRU - Operand 0 is chain. Operand 1 is value for EDX. Operand 2 is
563 // SSE42 string comparisons.
564 // These nodes produce 3 results, index, mask, and flags. X86ISelDAGToDAG
565 // will emit one or two instructions based on which results are used. If
566 // flags and index/mask this allows us to use a single instruction since
567 // we won't have to pick and opcode for flags. Instead we can rely on the
568 // DAG to CSE everything and decide at isel.
572 // Test if in transactional execution.
576 RSQRT28, RSQRT28_SAE, RSQRT28S, RSQRT28S_SAE,
577 RCP28, RCP28_SAE, RCP28S, RCP28S_SAE, EXP2, EXP2_SAE,
579 // Conversions between float and half-float.
580 CVTPS2PH, CVTPH2PS, CVTPH2PS_SAE,
582 // Masked version of above.
583 // SRC, RND, PASSTHRU, MASK
586 // Galois Field Arithmetic Instructions
587 GF2P8AFFINEINVQB, GF2P8AFFINEQB, GF2P8MULB,
589 // LWP insert record.
595 // Enqueue Stores Instructions
598 // For avx512-vp2intersect
601 /// X86 strict FP compare instructions.
602 STRICT_FCMP = ISD::FIRST_TARGET_STRICTFP_OPCODE,
605 // Vector packed double/float comparison.
608 /// Vector comparison generating mask bits for fp and
609 /// integer signed and unsigned data types.
612 // Vector float/double to signed/unsigned integer with truncation.
613 STRICT_CVTTP2SI, STRICT_CVTTP2UI,
621 // RndScale - Round FP Values To Include A Given Number Of Fraction Bits.
622 // Also used by the legacy (V)ROUND intrinsics where we mask out the
623 // scaling part of the immediate.
626 // Vector signed/unsigned integer to float/double.
627 STRICT_CVTSI2P, STRICT_CVTUI2P,
630 LCMPXCHG_DAG = ISD::FIRST_TARGET_MEMORY_OPCODE,
633 LCMPXCHG8_SAVE_EBX_DAG,
634 LCMPXCHG16_SAVE_RBX_DAG,
636 /// LOCK-prefixed arithmetic read-modify-write instructions.
637 /// EFLAGS, OUTCHAIN = LADD(INCHAIN, PTR, RHS)
638 LADD, LSUB, LOR, LXOR, LAND,
640 // Load, scalar_to_vector, and zero extend.
643 // extract_vector_elt, store.
646 // scalar broadcast from memory
649 // Store FP control world into i16 memory.
652 /// This instruction implements FP_TO_SINT with the
653 /// integer destination in memory and a FP reg source. This corresponds
654 /// to the X86::FIST*m instructions and the rounding mode change stuff. It
655 /// has two inputs (token chain and address) and two outputs (int value
656 /// and token chain). Memory VT specifies the type to store to.
659 /// This instruction implements SINT_TO_FP with the
660 /// integer source in memory and FP reg result. This corresponds to the
661 /// X86::FILD*m instructions. It has two inputs (token chain and address)
662 /// and two outputs (FP value and token chain). FILD_FLAG also produces a
663 /// flag). The integer source type is specified by the memory VT.
667 /// This instruction implements a fp->int store from FP stack
668 /// slots. This corresponds to the fist instruction. It takes a
669 /// chain operand, value to store, address, and glue. The memory VT
670 /// specifies the type to store as.
673 /// This instruction implements an extending load to FP stack slots.
674 /// This corresponds to the X86::FLD32m / X86::FLD64m. It takes a chain
675 /// operand, and ptr to load from. The memory VT specifies the type to
679 /// This instruction implements a truncating store from FP stack
680 /// slots. This corresponds to the X86::FST32m / X86::FST64m. It takes a
681 /// chain operand, value to store, address, and glue. The memory VT
682 /// specifies the type to store as.
685 /// This instruction grabs the address of the next argument
686 /// from a va_list. (reads and modifies the va_list in memory)
689 // Vector truncating store with unsigned/signed saturation
690 VTRUNCSTOREUS, VTRUNCSTORES,
691 // Vector truncating masked store with unsigned/signed saturation
692 VMTRUNCSTOREUS, VMTRUNCSTORES,
694 // X86 specific gather and scatter
697 // WARNING: Do not add anything in the end unless you want the node to
698 // have memop! In fact, starting from FIRST_TARGET_MEMORY_OPCODE all
699 // opcodes will be thought as target memory ops!
701 } // end namespace X86ISD
703 /// Define some predicates that are used for node matching.
705 /// Returns true if Elt is a constant zero or floating point constant +0.0.
706 bool isZeroNode(SDValue Elt);
708 /// Returns true of the given offset can be
709 /// fit into displacement field of the instruction.
710 bool isOffsetSuitableForCodeModel(int64_t Offset, CodeModel::Model M,
711 bool hasSymbolicDisplacement = true);
713 /// Determines whether the callee is required to pop its
714 /// own arguments. Callee pop is necessary to support tail calls.
715 bool isCalleePop(CallingConv::ID CallingConv,
716 bool is64Bit, bool IsVarArg, bool GuaranteeTCO);
718 /// If Op is a constant whose elements are all the same constant or
719 /// undefined, return true and return the constant value in \p SplatVal.
720 bool isConstantSplat(SDValue Op, APInt &SplatVal);
721 } // end namespace X86
723 //===--------------------------------------------------------------------===//
724 // X86 Implementation of the TargetLowering interface
725 class X86TargetLowering final : public TargetLowering {
727 explicit X86TargetLowering(const X86TargetMachine &TM,
728 const X86Subtarget &STI);
730 unsigned getJumpTableEncoding() const override;
731 bool useSoftFloat() const override;
733 void markLibCallAttributes(MachineFunction *MF, unsigned CC,
734 ArgListTy &Args) const override;
736 MVT getScalarShiftAmountTy(const DataLayout &, EVT VT) const override {
741 LowerCustomJumpTableEntry(const MachineJumpTableInfo *MJTI,
742 const MachineBasicBlock *MBB, unsigned uid,
743 MCContext &Ctx) const override;
745 /// Returns relocation base for the given PIC jumptable.
746 SDValue getPICJumpTableRelocBase(SDValue Table,
747 SelectionDAG &DAG) const override;
749 getPICJumpTableRelocBaseExpr(const MachineFunction *MF,
750 unsigned JTI, MCContext &Ctx) const override;
752 /// Return the desired alignment for ByVal aggregate
753 /// function arguments in the caller parameter area. For X86, aggregates
754 /// that contains are placed at 16-byte boundaries while the rest are at
755 /// 4-byte boundaries.
756 unsigned getByValTypeAlignment(Type *Ty,
757 const DataLayout &DL) const override;
759 /// Returns the target specific optimal type for load
760 /// and store operations as a result of memset, memcpy, and memmove
761 /// lowering. If DstAlign is zero that means it's safe to destination
762 /// alignment can satisfy any constraint. Similarly if SrcAlign is zero it
763 /// means there isn't a need to check it against alignment requirement,
764 /// probably because the source does not need to be loaded. If 'IsMemset' is
765 /// true, that means it's expanding a memset. If 'ZeroMemset' is true, that
766 /// means it's a memset of zero. 'MemcpyStrSrc' indicates whether the memcpy
767 /// source is constant so it does not need to be loaded.
768 /// It returns EVT::Other if the type should be determined using generic
769 /// target-independent logic.
770 EVT getOptimalMemOpType(uint64_t Size, unsigned DstAlign, unsigned SrcAlign,
771 bool IsMemset, bool ZeroMemset, bool MemcpyStrSrc,
772 const AttributeList &FuncAttributes) const override;
774 /// Returns true if it's safe to use load / store of the
775 /// specified type to expand memcpy / memset inline. This is mostly true
776 /// for all types except for some special cases. For example, on X86
777 /// targets without SSE2 f64 load / store are done with fldl / fstpl which
778 /// also does type conversion. Note the specified type doesn't have to be
779 /// legal as the hook is used before type legalization.
780 bool isSafeMemOpType(MVT VT) const override;
782 /// Returns true if the target allows unaligned memory accesses of the
783 /// specified type. Returns whether it is "fast" in the last argument.
784 bool allowsMisalignedMemoryAccesses(EVT VT, unsigned AS, unsigned Align,
785 MachineMemOperand::Flags Flags,
786 bool *Fast) const override;
788 /// Provide custom lowering hooks for some operations.
790 SDValue LowerOperation(SDValue Op, SelectionDAG &DAG) const override;
792 /// Places new result values for the node in Results (their number
793 /// and types must exactly match those of the original return values of
794 /// the node), or leaves Results empty, which indicates that the node is not
795 /// to be custom lowered after all.
796 void LowerOperationWrapper(SDNode *N,
797 SmallVectorImpl<SDValue> &Results,
798 SelectionDAG &DAG) const override;
800 /// Replace the results of node with an illegal result
801 /// type with new values built out of custom code.
803 void ReplaceNodeResults(SDNode *N, SmallVectorImpl<SDValue>&Results,
804 SelectionDAG &DAG) const override;
806 SDValue PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const override;
808 // Return true if it is profitable to combine a BUILD_VECTOR with a
809 // stride-pattern to a shuffle and a truncate.
810 // Example of such a combine:
811 // v4i32 build_vector((extract_elt V, 1),
812 // (extract_elt V, 3),
813 // (extract_elt V, 5),
814 // (extract_elt V, 7))
816 // v4i32 truncate (bitcast (shuffle<1,u,3,u,4,u,5,u,6,u,7,u> V, u) to
818 bool isDesirableToCombineBuildVectorToShuffleTruncate(
819 ArrayRef<int> ShuffleMask, EVT SrcVT, EVT TruncVT) const override;
821 /// Return true if the target has native support for
822 /// the specified value type and it is 'desirable' to use the type for the
823 /// given node type. e.g. On x86 i16 is legal, but undesirable since i16
824 /// instruction encodings are longer and some i16 instructions are slow.
825 bool isTypeDesirableForOp(unsigned Opc, EVT VT) const override;
827 /// Return true if the target has native support for the
828 /// specified value type and it is 'desirable' to use the type. e.g. On x86
829 /// i16 is legal, but undesirable since i16 instruction encodings are longer
830 /// and some i16 instructions are slow.
831 bool IsDesirableToPromoteOp(SDValue Op, EVT &PVT) const override;
833 /// Return 1 if we can compute the negated form of the specified expression
834 /// for the same cost as the expression itself, or 2 if we can compute the
835 /// negated form more cheaply than the expression itself. Else return 0.
836 char isNegatibleForFree(SDValue Op, SelectionDAG &DAG, bool LegalOperations,
837 bool ForCodeSize, unsigned Depth) const override;
839 /// If isNegatibleForFree returns true, return the newly negated expression.
840 SDValue getNegatedExpression(SDValue Op, SelectionDAG &DAG,
841 bool LegalOperations, bool ForCodeSize,
842 unsigned Depth) const override;
845 EmitInstrWithCustomInserter(MachineInstr &MI,
846 MachineBasicBlock *MBB) const override;
848 /// This method returns the name of a target specific DAG node.
849 const char *getTargetNodeName(unsigned Opcode) const override;
851 /// Do not merge vector stores after legalization because that may conflict
852 /// with x86-specific store splitting optimizations.
853 bool mergeStoresAfterLegalization(EVT MemVT) const override {
854 return !MemVT.isVector();
857 bool canMergeStoresTo(unsigned AddressSpace, EVT MemVT,
858 const SelectionDAG &DAG) const override;
860 bool isCheapToSpeculateCttz() const override;
862 bool isCheapToSpeculateCtlz() const override;
864 bool isCtlzFast() const override;
866 bool hasBitPreservingFPLogic(EVT VT) const override {
867 return VT == MVT::f32 || VT == MVT::f64 || VT.isVector();
870 bool isMultiStoresCheaperThanBitsMerge(EVT LTy, EVT HTy) const override {
871 // If the pair to store is a mixture of float and int values, we will
872 // save two bitwise instructions and one float-to-int instruction and
873 // increase one store instruction. There is potentially a more
874 // significant benefit because it avoids the float->int domain switch
875 // for input value. So It is more likely a win.
876 if ((LTy.isFloatingPoint() && HTy.isInteger()) ||
877 (LTy.isInteger() && HTy.isFloatingPoint()))
879 // If the pair only contains int values, we will save two bitwise
880 // instructions and increase one store instruction (costing one more
881 // store buffer). Since the benefit is more blurred so we leave
882 // such pair out until we get testcase to prove it is a win.
886 bool isMaskAndCmp0FoldingBeneficial(const Instruction &AndI) const override;
888 bool hasAndNotCompare(SDValue Y) const override;
890 bool hasAndNot(SDValue Y) const override;
892 bool hasBitTest(SDValue X, SDValue Y) const override;
894 bool shouldProduceAndByConstByHoistingConstFromShiftsLHSOfAnd(
895 SDValue X, ConstantSDNode *XC, ConstantSDNode *CC, SDValue Y,
896 unsigned OldShiftOpcode, unsigned NewShiftOpcode,
897 SelectionDAG &DAG) const override;
899 bool shouldFoldConstantShiftPairToMask(const SDNode *N,
900 CombineLevel Level) const override;
902 bool shouldFoldMaskToVariableShiftPair(SDValue Y) const override;
905 shouldTransformSignedTruncationCheck(EVT XVT,
906 unsigned KeptBits) const override {
907 // For vectors, we don't have a preference..
911 auto VTIsOk = [](EVT VT) -> bool {
912 return VT == MVT::i8 || VT == MVT::i16 || VT == MVT::i32 ||
916 // We are ok with KeptBitsVT being byte/word/dword, what MOVS supports.
917 // XVT will be larger than KeptBitsVT.
918 MVT KeptBitsVT = MVT::getIntegerVT(KeptBits);
919 return VTIsOk(XVT) && VTIsOk(KeptBitsVT);
922 bool shouldExpandShift(SelectionDAG &DAG, SDNode *N) const override;
924 bool shouldSplatInsEltVarIndex(EVT VT) const override;
926 bool convertSetCCLogicToBitwiseLogic(EVT VT) const override {
927 return VT.isScalarInteger();
930 /// Vector-sized comparisons are fast using PCMPEQ + PMOVMSK or PTEST.
931 MVT hasFastEqualityCompare(unsigned NumBits) const override;
933 /// Return the value type to use for ISD::SETCC.
934 EVT getSetCCResultType(const DataLayout &DL, LLVMContext &Context,
935 EVT VT) const override;
937 bool targetShrinkDemandedConstant(SDValue Op, const APInt &Demanded,
938 TargetLoweringOpt &TLO) const override;
940 /// Determine which of the bits specified in Mask are known to be either
941 /// zero or one and return them in the KnownZero/KnownOne bitsets.
942 void computeKnownBitsForTargetNode(const SDValue Op,
944 const APInt &DemandedElts,
945 const SelectionDAG &DAG,
946 unsigned Depth = 0) const override;
948 /// Determine the number of bits in the operation that are sign bits.
949 unsigned ComputeNumSignBitsForTargetNode(SDValue Op,
950 const APInt &DemandedElts,
951 const SelectionDAG &DAG,
952 unsigned Depth) const override;
954 bool SimplifyDemandedVectorEltsForTargetNode(SDValue Op,
955 const APInt &DemandedElts,
958 TargetLoweringOpt &TLO,
959 unsigned Depth) const override;
961 bool SimplifyDemandedBitsForTargetNode(SDValue Op,
962 const APInt &DemandedBits,
963 const APInt &DemandedElts,
965 TargetLoweringOpt &TLO,
966 unsigned Depth) const override;
968 SDValue SimplifyMultipleUseDemandedBitsForTargetNode(
969 SDValue Op, const APInt &DemandedBits, const APInt &DemandedElts,
970 SelectionDAG &DAG, unsigned Depth) const override;
972 const Constant *getTargetConstantFromLoad(LoadSDNode *LD) const override;
974 SDValue unwrapAddress(SDValue N) const override;
976 SDValue getReturnAddressFrameIndex(SelectionDAG &DAG) const;
978 bool ExpandInlineAsm(CallInst *CI) const override;
980 ConstraintType getConstraintType(StringRef Constraint) const override;
982 /// Examine constraint string and operand type and determine a weight value.
983 /// The operand object must already have been set up with the operand type.
985 getSingleConstraintMatchWeight(AsmOperandInfo &info,
986 const char *constraint) const override;
988 const char *LowerXConstraint(EVT ConstraintVT) const override;
990 /// Lower the specified operand into the Ops vector. If it is invalid, don't
991 /// add anything to Ops. If hasMemory is true it means one of the asm
992 /// constraint of the inline asm instruction being processed is 'm'.
993 void LowerAsmOperandForConstraint(SDValue Op,
994 std::string &Constraint,
995 std::vector<SDValue> &Ops,
996 SelectionDAG &DAG) const override;
999 getInlineAsmMemConstraint(StringRef ConstraintCode) const override {
1000 if (ConstraintCode == "o")
1001 return InlineAsm::Constraint_o;
1002 else if (ConstraintCode == "v")
1003 return InlineAsm::Constraint_v;
1004 else if (ConstraintCode == "X")
1005 return InlineAsm::Constraint_X;
1006 return TargetLowering::getInlineAsmMemConstraint(ConstraintCode);
1009 /// Handle Lowering flag assembly outputs.
1010 SDValue LowerAsmOutputForConstraint(SDValue &Chain, SDValue &Flag, SDLoc DL,
1011 const AsmOperandInfo &Constraint,
1012 SelectionDAG &DAG) const override;
1014 /// Given a physical register constraint
1015 /// (e.g. {edx}), return the register number and the register class for the
1016 /// register. This should only be used for C_Register constraints. On
1017 /// error, this returns a register number of 0.
1018 std::pair<unsigned, const TargetRegisterClass *>
1019 getRegForInlineAsmConstraint(const TargetRegisterInfo *TRI,
1020 StringRef Constraint, MVT VT) const override;
1022 /// Return true if the addressing mode represented
1023 /// by AM is legal for this target, for a load/store of the specified type.
1024 bool isLegalAddressingMode(const DataLayout &DL, const AddrMode &AM,
1025 Type *Ty, unsigned AS,
1026 Instruction *I = nullptr) const override;
1028 /// Return true if the specified immediate is legal
1029 /// icmp immediate, that is the target has icmp instructions which can
1030 /// compare a register against the immediate without having to materialize
1031 /// the immediate into a register.
1032 bool isLegalICmpImmediate(int64_t Imm) const override;
1034 /// Return true if the specified immediate is legal
1035 /// add immediate, that is the target has add instructions which can
1036 /// add a register and the immediate without having to materialize
1037 /// the immediate into a register.
1038 bool isLegalAddImmediate(int64_t Imm) const override;
1040 bool isLegalStoreImmediate(int64_t Imm) const override;
1042 /// Return the cost of the scaling factor used in the addressing
1043 /// mode represented by AM for this target, for a load/store
1044 /// of the specified type.
1045 /// If the AM is supported, the return value must be >= 0.
1046 /// If the AM is not supported, it returns a negative value.
1047 int getScalingFactorCost(const DataLayout &DL, const AddrMode &AM, Type *Ty,
1048 unsigned AS) const override;
1050 bool isVectorShiftByScalarCheap(Type *Ty) const override;
1052 /// Add x86-specific opcodes to the default list.
1053 bool isBinOp(unsigned Opcode) const override;
1055 /// Returns true if the opcode is a commutative binary operation.
1056 bool isCommutativeBinOp(unsigned Opcode) const override;
1058 /// Return true if it's free to truncate a value of
1059 /// type Ty1 to type Ty2. e.g. On x86 it's free to truncate a i32 value in
1060 /// register EAX to i16 by referencing its sub-register AX.
1061 bool isTruncateFree(Type *Ty1, Type *Ty2) const override;
1062 bool isTruncateFree(EVT VT1, EVT VT2) const override;
1064 bool allowTruncateForTailCall(Type *Ty1, Type *Ty2) const override;
1066 /// Return true if any actual instruction that defines a
1067 /// value of type Ty1 implicit zero-extends the value to Ty2 in the result
1068 /// register. This does not necessarily include registers defined in
1069 /// unknown ways, such as incoming arguments, or copies from unknown
1070 /// virtual registers. Also, if isTruncateFree(Ty2, Ty1) is true, this
1071 /// does not necessarily apply to truncate instructions. e.g. on x86-64,
1072 /// all instructions that define 32-bit values implicit zero-extend the
1073 /// result out to 64 bits.
1074 bool isZExtFree(Type *Ty1, Type *Ty2) const override;
1075 bool isZExtFree(EVT VT1, EVT VT2) const override;
1076 bool isZExtFree(SDValue Val, EVT VT2) const override;
1078 /// Return true if folding a vector load into ExtVal (a sign, zero, or any
1079 /// extend node) is profitable.
1080 bool isVectorLoadExtDesirable(SDValue) const override;
1082 /// Return true if an FMA operation is faster than a pair of fmul and fadd
1083 /// instructions. fmuladd intrinsics will be expanded to FMAs when this
1084 /// method returns true, otherwise fmuladd is expanded to fmul + fadd.
1085 bool isFMAFasterThanFMulAndFAdd(const MachineFunction &MF,
1086 EVT VT) const override;
1088 /// Return true if it's profitable to narrow
1089 /// operations of type VT1 to VT2. e.g. on x86, it's profitable to narrow
1090 /// from i32 to i8 but not from i32 to i16.
1091 bool isNarrowingProfitable(EVT VT1, EVT VT2) const override;
1093 /// Given an intrinsic, checks if on the target the intrinsic will need to map
1094 /// to a MemIntrinsicNode (touches memory). If this is the case, it returns
1095 /// true and stores the intrinsic information into the IntrinsicInfo that was
1096 /// passed to the function.
1097 bool getTgtMemIntrinsic(IntrinsicInfo &Info, const CallInst &I,
1098 MachineFunction &MF,
1099 unsigned Intrinsic) const override;
1101 /// Returns true if the target can instruction select the
1102 /// specified FP immediate natively. If false, the legalizer will
1103 /// materialize the FP immediate as a load from a constant pool.
1104 bool isFPImmLegal(const APFloat &Imm, EVT VT,
1105 bool ForCodeSize) const override;
1107 /// Targets can use this to indicate that they only support *some*
1108 /// VECTOR_SHUFFLE operations, those with specific masks. By default, if a
1109 /// target supports the VECTOR_SHUFFLE node, all mask values are assumed to
1111 bool isShuffleMaskLegal(ArrayRef<int> Mask, EVT VT) const override;
1113 /// Similar to isShuffleMaskLegal. Targets can use this to indicate if there
1114 /// is a suitable VECTOR_SHUFFLE that can be used to replace a VAND with a
1115 /// constant pool entry.
1116 bool isVectorClearMaskLegal(ArrayRef<int> Mask, EVT VT) const override;
1118 /// Returns true if lowering to a jump table is allowed.
1119 bool areJTsAllowed(const Function *Fn) const override;
1121 /// If true, then instruction selection should
1122 /// seek to shrink the FP constant of the specified type to a smaller type
1123 /// in order to save space and / or reduce runtime.
1124 bool ShouldShrinkFPConstant(EVT VT) const override {
1125 // Don't shrink FP constpool if SSE2 is available since cvtss2sd is more
1126 // expensive than a straight movsd. On the other hand, it's important to
1127 // shrink long double fp constant since fldt is very slow.
1128 return !X86ScalarSSEf64 || VT == MVT::f80;
1131 /// Return true if we believe it is correct and profitable to reduce the
1132 /// load node to a smaller type.
1133 bool shouldReduceLoadWidth(SDNode *Load, ISD::LoadExtType ExtTy,
1134 EVT NewVT) const override;
1136 /// Return true if the specified scalar FP type is computed in an SSE
1137 /// register, not on the X87 floating point stack.
1138 bool isScalarFPTypeInSSEReg(EVT VT) const {
1139 return (VT == MVT::f64 && X86ScalarSSEf64) || // f64 is when SSE2
1140 (VT == MVT::f32 && X86ScalarSSEf32); // f32 is when SSE1
1143 /// Returns true if it is beneficial to convert a load of a constant
1144 /// to just the constant itself.
1145 bool shouldConvertConstantLoadToIntImm(const APInt &Imm,
1146 Type *Ty) const override;
1148 bool reduceSelectOfFPConstantLoads(EVT CmpOpVT) const override;
1150 bool convertSelectOfConstantsToMath(EVT VT) const override;
1152 bool decomposeMulByConstant(LLVMContext &Context, EVT VT,
1153 SDValue C) const override;
1155 /// Return true if EXTRACT_SUBVECTOR is cheap for this result type
1156 /// with this index.
1157 bool isExtractSubvectorCheap(EVT ResVT, EVT SrcVT,
1158 unsigned Index) const override;
1160 /// Scalar ops always have equal or better analysis/performance/power than
1161 /// the vector equivalent, so this always makes sense if the scalar op is
1163 bool shouldScalarizeBinop(SDValue) const override;
1165 /// Extract of a scalar FP value from index 0 of a vector is free.
1166 bool isExtractVecEltCheap(EVT VT, unsigned Index) const override {
1167 EVT EltVT = VT.getScalarType();
1168 return (EltVT == MVT::f32 || EltVT == MVT::f64) && Index == 0;
1171 /// Overflow nodes should get combined/lowered to optimal instructions
1172 /// (they should allow eliminating explicit compares by getting flags from
1174 bool shouldFormOverflowOp(unsigned Opcode, EVT VT) const override;
1176 bool storeOfVectorConstantIsCheap(EVT MemVT, unsigned NumElem,
1177 unsigned AddrSpace) const override {
1178 // If we can replace more than 2 scalar stores, there will be a reduction
1179 // in instructions even after we add a vector constant load.
1183 bool isLoadBitCastBeneficial(EVT LoadVT, EVT BitcastVT,
1184 const SelectionDAG &DAG,
1185 const MachineMemOperand &MMO) const override;
1187 /// Intel processors have a unified instruction and data cache
1188 const char * getClearCacheBuiltinName() const override {
1189 return nullptr; // nothing to do, move along.
1192 Register getRegisterByName(const char* RegName, LLT VT,
1193 const MachineFunction &MF) const override;
1195 /// If a physical register, this returns the register that receives the
1196 /// exception address on entry to an EH pad.
1198 getExceptionPointerRegister(const Constant *PersonalityFn) const override;
1200 /// If a physical register, this returns the register that receives the
1201 /// exception typeid on entry to a landing pad.
1203 getExceptionSelectorRegister(const Constant *PersonalityFn) const override;
1205 virtual bool needsFixedCatchObjects() const override;
1207 /// This method returns a target specific FastISel object,
1208 /// or null if the target does not support "fast" ISel.
1209 FastISel *createFastISel(FunctionLoweringInfo &funcInfo,
1210 const TargetLibraryInfo *libInfo) const override;
1212 /// If the target has a standard location for the stack protector cookie,
1213 /// returns the address of that location. Otherwise, returns nullptr.
1214 Value *getIRStackGuard(IRBuilder<> &IRB) const override;
1216 bool useLoadStackGuardNode() const override;
1217 bool useStackGuardXorFP() const override;
1218 void insertSSPDeclarations(Module &M) const override;
1219 Value *getSDagStackGuard(const Module &M) const override;
1220 Function *getSSPStackGuardCheck(const Module &M) const override;
1221 SDValue emitStackGuardXorFP(SelectionDAG &DAG, SDValue Val,
1222 const SDLoc &DL) const override;
1225 /// Return true if the target stores SafeStack pointer at a fixed offset in
1226 /// some non-standard address space, and populates the address space and
1227 /// offset as appropriate.
1228 Value *getSafeStackPointerLocation(IRBuilder<> &IRB) const override;
1230 std::pair<SDValue, SDValue> BuildFILD(SDValue Op, EVT SrcVT, SDValue Chain,
1232 SelectionDAG &DAG) const;
1234 bool isNoopAddrSpaceCast(unsigned SrcAS, unsigned DestAS) const override;
1236 /// Customize the preferred legalization strategy for certain types.
1237 LegalizeTypeAction getPreferredVectorAction(MVT VT) const override;
1239 MVT getRegisterTypeForCallingConv(LLVMContext &Context, CallingConv::ID CC,
1240 EVT VT) const override;
1242 unsigned getNumRegistersForCallingConv(LLVMContext &Context,
1244 EVT VT) const override;
1246 unsigned getVectorTypeBreakdownForCallingConv(
1247 LLVMContext &Context, CallingConv::ID CC, EVT VT, EVT &IntermediateVT,
1248 unsigned &NumIntermediates, MVT &RegisterVT) const override;
1250 bool isIntDivCheap(EVT VT, AttributeList Attr) const override;
1252 bool supportSwiftError() const override;
1254 StringRef getStackProbeSymbolName(MachineFunction &MF) const override;
1256 unsigned getStackProbeSize(MachineFunction &MF) const;
1258 bool hasVectorBlend() const override { return true; }
1260 unsigned getMaxSupportedInterleaveFactor() const override { return 4; }
1262 /// Lower interleaved load(s) into target specific
1263 /// instructions/intrinsics.
1264 bool lowerInterleavedLoad(LoadInst *LI,
1265 ArrayRef<ShuffleVectorInst *> Shuffles,
1266 ArrayRef<unsigned> Indices,
1267 unsigned Factor) const override;
1269 /// Lower interleaved store(s) into target specific
1270 /// instructions/intrinsics.
1271 bool lowerInterleavedStore(StoreInst *SI, ShuffleVectorInst *SVI,
1272 unsigned Factor) const override;
1274 SDValue expandIndirectJTBranch(const SDLoc& dl, SDValue Value,
1275 SDValue Addr, SelectionDAG &DAG)
1279 std::pair<const TargetRegisterClass *, uint8_t>
1280 findRepresentativeClass(const TargetRegisterInfo *TRI,
1281 MVT VT) const override;
1284 /// Keep a reference to the X86Subtarget around so that we can
1285 /// make the right decision when generating code for different targets.
1286 const X86Subtarget &Subtarget;
1288 /// Select between SSE or x87 floating point ops.
1289 /// When SSE is available, use it for f32 operations.
1290 /// When SSE2 is available, use it for f64 operations.
1291 bool X86ScalarSSEf32;
1292 bool X86ScalarSSEf64;
1294 /// A list of legal FP immediates.
1295 std::vector<APFloat> LegalFPImmediates;
1297 /// Indicate that this x86 target can instruction
1298 /// select the specified FP immediate natively.
1299 void addLegalFPImmediate(const APFloat& Imm) {
1300 LegalFPImmediates.push_back(Imm);
1303 SDValue LowerCallResult(SDValue Chain, SDValue InFlag,
1304 CallingConv::ID CallConv, bool isVarArg,
1305 const SmallVectorImpl<ISD::InputArg> &Ins,
1306 const SDLoc &dl, SelectionDAG &DAG,
1307 SmallVectorImpl<SDValue> &InVals,
1308 uint32_t *RegMask) const;
1309 SDValue LowerMemArgument(SDValue Chain, CallingConv::ID CallConv,
1310 const SmallVectorImpl<ISD::InputArg> &ArgInfo,
1311 const SDLoc &dl, SelectionDAG &DAG,
1312 const CCValAssign &VA, MachineFrameInfo &MFI,
1314 SDValue LowerMemOpCallTo(SDValue Chain, SDValue StackPtr, SDValue Arg,
1315 const SDLoc &dl, SelectionDAG &DAG,
1316 const CCValAssign &VA,
1317 ISD::ArgFlagsTy Flags) const;
1319 // Call lowering helpers.
1321 /// Check whether the call is eligible for tail call optimization. Targets
1322 /// that want to do tail call optimization should implement this function.
1323 bool IsEligibleForTailCallOptimization(SDValue Callee,
1324 CallingConv::ID CalleeCC,
1326 bool isCalleeStructRet,
1327 bool isCallerStructRet,
1329 const SmallVectorImpl<ISD::OutputArg> &Outs,
1330 const SmallVectorImpl<SDValue> &OutVals,
1331 const SmallVectorImpl<ISD::InputArg> &Ins,
1332 SelectionDAG& DAG) const;
1333 SDValue EmitTailCallLoadRetAddr(SelectionDAG &DAG, SDValue &OutRetAddr,
1334 SDValue Chain, bool IsTailCall,
1335 bool Is64Bit, int FPDiff,
1336 const SDLoc &dl) const;
1338 unsigned GetAlignedArgumentStackSize(unsigned StackSize,
1339 SelectionDAG &DAG) const;
1341 unsigned getAddressSpace(void) const;
1343 SDValue FP_TO_INTHelper(SDValue Op, SelectionDAG &DAG, bool isSigned,
1344 SDValue &Chain) const;
1346 SDValue LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) const;
1347 SDValue LowerVSELECT(SDValue Op, SelectionDAG &DAG) const;
1348 SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const;
1349 SDValue LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const;
1351 unsigned getGlobalWrapperKind(const GlobalValue *GV = nullptr,
1352 const unsigned char OpFlags = 0) const;
1353 SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) const;
1354 SDValue LowerBlockAddress(SDValue Op, SelectionDAG &DAG) const;
1355 SDValue LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) const;
1356 SDValue LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const;
1357 SDValue LowerExternalSymbol(SDValue Op, SelectionDAG &DAG) const;
1359 /// Creates target global address or external symbol nodes for calls or
1361 SDValue LowerGlobalOrExternal(SDValue Op, SelectionDAG &DAG,
1362 bool ForCall) const;
1364 SDValue LowerSINT_TO_FP(SDValue Op, SelectionDAG &DAG) const;
1365 SDValue LowerUINT_TO_FP(SDValue Op, SelectionDAG &DAG) const;
1366 SDValue LowerTRUNCATE(SDValue Op, SelectionDAG &DAG) const;
1367 SDValue LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG) const;
1368 SDValue LowerSETCC(SDValue Op, SelectionDAG &DAG) const;
1369 SDValue LowerSTRICT_FSETCC(SDValue Op, SelectionDAG &DAG) const;
1370 SDValue LowerSETCCCARRY(SDValue Op, SelectionDAG &DAG) const;
1371 SDValue LowerSELECT(SDValue Op, SelectionDAG &DAG) const;
1372 SDValue LowerBRCOND(SDValue Op, SelectionDAG &DAG) const;
1373 SDValue LowerJumpTable(SDValue Op, SelectionDAG &DAG) const;
1374 SDValue LowerDYNAMIC_STACKALLOC(SDValue Op, SelectionDAG &DAG) const;
1375 SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG) const;
1376 SDValue LowerVAARG(SDValue Op, SelectionDAG &DAG) const;
1377 SDValue LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const;
1378 SDValue LowerADDROFRETURNADDR(SDValue Op, SelectionDAG &DAG) const;
1379 SDValue LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const;
1380 SDValue LowerFRAME_TO_ARGS_OFFSET(SDValue Op, SelectionDAG &DAG) const;
1381 SDValue LowerEH_RETURN(SDValue Op, SelectionDAG &DAG) const;
1382 SDValue lowerEH_SJLJ_SETJMP(SDValue Op, SelectionDAG &DAG) const;
1383 SDValue lowerEH_SJLJ_LONGJMP(SDValue Op, SelectionDAG &DAG) const;
1384 SDValue lowerEH_SJLJ_SETUP_DISPATCH(SDValue Op, SelectionDAG &DAG) const;
1385 SDValue LowerINIT_TRAMPOLINE(SDValue Op, SelectionDAG &DAG) const;
1386 SDValue LowerFLT_ROUNDS_(SDValue Op, SelectionDAG &DAG) const;
1387 SDValue LowerWin64_i128OP(SDValue Op, SelectionDAG &DAG) const;
1388 SDValue LowerGC_TRANSITION(SDValue Op, SelectionDAG &DAG) const;
1389 SDValue LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) const;
1390 SDValue lowerFaddFsub(SDValue Op, SelectionDAG &DAG) const;
1391 SDValue LowerFP_EXTEND(SDValue Op, SelectionDAG &DAG) const;
1392 SDValue LowerFP_ROUND(SDValue Op, SelectionDAG &DAG) const;
1394 SDValue LowerF128Call(SDValue Op, SelectionDAG &DAG,
1395 RTLIB::Libcall Call) const;
1398 LowerFormalArguments(SDValue Chain, CallingConv::ID CallConv, bool isVarArg,
1399 const SmallVectorImpl<ISD::InputArg> &Ins,
1400 const SDLoc &dl, SelectionDAG &DAG,
1401 SmallVectorImpl<SDValue> &InVals) const override;
1402 SDValue LowerCall(CallLoweringInfo &CLI,
1403 SmallVectorImpl<SDValue> &InVals) const override;
1405 SDValue LowerReturn(SDValue Chain, CallingConv::ID CallConv, bool isVarArg,
1406 const SmallVectorImpl<ISD::OutputArg> &Outs,
1407 const SmallVectorImpl<SDValue> &OutVals,
1408 const SDLoc &dl, SelectionDAG &DAG) const override;
1410 bool supportSplitCSR(MachineFunction *MF) const override {
1411 return MF->getFunction().getCallingConv() == CallingConv::CXX_FAST_TLS &&
1412 MF->getFunction().hasFnAttribute(Attribute::NoUnwind);
1414 void initializeSplitCSR(MachineBasicBlock *Entry) const override;
1415 void insertCopiesSplitCSR(
1416 MachineBasicBlock *Entry,
1417 const SmallVectorImpl<MachineBasicBlock *> &Exits) const override;
1419 bool isUsedByReturnOnly(SDNode *N, SDValue &Chain) const override;
1421 bool mayBeEmittedAsTailCall(const CallInst *CI) const override;
1423 EVT getTypeForExtReturn(LLVMContext &Context, EVT VT,
1424 ISD::NodeType ExtendKind) const override;
1426 bool CanLowerReturn(CallingConv::ID CallConv, MachineFunction &MF,
1428 const SmallVectorImpl<ISD::OutputArg> &Outs,
1429 LLVMContext &Context) const override;
1431 const MCPhysReg *getScratchRegisters(CallingConv::ID CC) const override;
1433 TargetLoweringBase::AtomicExpansionKind
1434 shouldExpandAtomicLoadInIR(LoadInst *SI) const override;
1435 bool shouldExpandAtomicStoreInIR(StoreInst *SI) const override;
1436 TargetLoweringBase::AtomicExpansionKind
1437 shouldExpandAtomicRMWInIR(AtomicRMWInst *AI) const override;
1440 lowerIdempotentRMWIntoFencedLoad(AtomicRMWInst *AI) const override;
1442 bool lowerAtomicStoreAsStoreSDNode(const StoreInst &SI) const override;
1443 bool lowerAtomicLoadAsLoadSDNode(const LoadInst &LI) const override;
1445 bool needsCmpXchgNb(Type *MemType) const;
1447 void SetupEntryBlockForSjLj(MachineInstr &MI, MachineBasicBlock *MBB,
1448 MachineBasicBlock *DispatchBB, int FI) const;
1450 // Utility function to emit the low-level va_arg code for X86-64.
1452 EmitVAARG64WithCustomInserter(MachineInstr &MI,
1453 MachineBasicBlock *MBB) const;
1455 /// Utility function to emit the xmm reg save portion of va_start.
1457 EmitVAStartSaveXMMRegsWithCustomInserter(MachineInstr &BInstr,
1458 MachineBasicBlock *BB) const;
1460 MachineBasicBlock *EmitLoweredCascadedSelect(MachineInstr &MI1,
1462 MachineBasicBlock *BB) const;
1464 MachineBasicBlock *EmitLoweredSelect(MachineInstr &I,
1465 MachineBasicBlock *BB) const;
1467 MachineBasicBlock *EmitLoweredAtomicFP(MachineInstr &I,
1468 MachineBasicBlock *BB) const;
1470 MachineBasicBlock *EmitLoweredCatchRet(MachineInstr &MI,
1471 MachineBasicBlock *BB) const;
1473 MachineBasicBlock *EmitLoweredCatchPad(MachineInstr &MI,
1474 MachineBasicBlock *BB) const;
1476 MachineBasicBlock *EmitLoweredSegAlloca(MachineInstr &MI,
1477 MachineBasicBlock *BB) const;
1479 MachineBasicBlock *EmitLoweredTLSAddr(MachineInstr &MI,
1480 MachineBasicBlock *BB) const;
1482 MachineBasicBlock *EmitLoweredTLSCall(MachineInstr &MI,
1483 MachineBasicBlock *BB) const;
1485 MachineBasicBlock *EmitLoweredIndirectThunk(MachineInstr &MI,
1486 MachineBasicBlock *BB) const;
1488 MachineBasicBlock *emitEHSjLjSetJmp(MachineInstr &MI,
1489 MachineBasicBlock *MBB) const;
1491 void emitSetJmpShadowStackFix(MachineInstr &MI,
1492 MachineBasicBlock *MBB) const;
1494 MachineBasicBlock *emitEHSjLjLongJmp(MachineInstr &MI,
1495 MachineBasicBlock *MBB) const;
1497 MachineBasicBlock *emitLongJmpShadowStackFix(MachineInstr &MI,
1498 MachineBasicBlock *MBB) const;
1500 MachineBasicBlock *emitFMA3Instr(MachineInstr &MI,
1501 MachineBasicBlock *MBB) const;
1503 MachineBasicBlock *EmitSjLjDispatchBlock(MachineInstr &MI,
1504 MachineBasicBlock *MBB) const;
1506 /// Convert a comparison if required by the subtarget.
1507 SDValue ConvertCmpIfNecessary(SDValue Cmp, SelectionDAG &DAG) const;
1509 /// Emit flags for the given setcc condition and operands. Also returns the
1510 /// corresponding X86 condition code constant in X86CC.
1511 SDValue emitFlagsForSetcc(SDValue Op0, SDValue Op1, ISD::CondCode CC,
1512 const SDLoc &dl, SelectionDAG &DAG,
1513 SDValue &X86CC, SDValue &Chain,
1514 bool IsSignaling) const;
1516 /// Check if replacement of SQRT with RSQRT should be disabled.
1517 bool isFsqrtCheap(SDValue Operand, SelectionDAG &DAG) const override;
1519 /// Use rsqrt* to speed up sqrt calculations.
1520 SDValue getSqrtEstimate(SDValue Operand, SelectionDAG &DAG, int Enabled,
1521 int &RefinementSteps, bool &UseOneConstNR,
1522 bool Reciprocal) const override;
1524 /// Use rcp* to speed up fdiv calculations.
1525 SDValue getRecipEstimate(SDValue Operand, SelectionDAG &DAG, int Enabled,
1526 int &RefinementSteps) const override;
1528 /// Reassociate floating point divisions into multiply by reciprocal.
1529 unsigned combineRepeatedFPDivisors() const override;
1531 SDValue BuildSDIVPow2(SDNode *N, const APInt &Divisor, SelectionDAG &DAG,
1532 SmallVectorImpl<SDNode *> &Created) const override;
1536 FastISel *createFastISel(FunctionLoweringInfo &funcInfo,
1537 const TargetLibraryInfo *libInfo);
1538 } // end namespace X86
1540 // Base class for all X86 non-masked store operations.
1541 class X86StoreSDNode : public MemSDNode {
1543 X86StoreSDNode(unsigned Opcode, unsigned Order, const DebugLoc &dl,
1544 SDVTList VTs, EVT MemVT,
1545 MachineMemOperand *MMO)
1546 :MemSDNode(Opcode, Order, dl, VTs, MemVT, MMO) {}
1547 const SDValue &getValue() const { return getOperand(1); }
1548 const SDValue &getBasePtr() const { return getOperand(2); }
1550 static bool classof(const SDNode *N) {
1551 return N->getOpcode() == X86ISD::VTRUNCSTORES ||
1552 N->getOpcode() == X86ISD::VTRUNCSTOREUS;
1556 // Base class for all X86 masked store operations.
1557 // The class has the same order of operands as MaskedStoreSDNode for
1559 class X86MaskedStoreSDNode : public MemSDNode {
1561 X86MaskedStoreSDNode(unsigned Opcode, unsigned Order,
1562 const DebugLoc &dl, SDVTList VTs, EVT MemVT,
1563 MachineMemOperand *MMO)
1564 : MemSDNode(Opcode, Order, dl, VTs, MemVT, MMO) {}
1566 const SDValue &getValue() const { return getOperand(1); }
1567 const SDValue &getBasePtr() const { return getOperand(2); }
1568 const SDValue &getMask() const { return getOperand(3); }
1570 static bool classof(const SDNode *N) {
1571 return N->getOpcode() == X86ISD::VMTRUNCSTORES ||
1572 N->getOpcode() == X86ISD::VMTRUNCSTOREUS;
1576 // X86 Truncating Store with Signed saturation.
1577 class TruncSStoreSDNode : public X86StoreSDNode {
1579 TruncSStoreSDNode(unsigned Order, const DebugLoc &dl,
1580 SDVTList VTs, EVT MemVT, MachineMemOperand *MMO)
1581 : X86StoreSDNode(X86ISD::VTRUNCSTORES, Order, dl, VTs, MemVT, MMO) {}
1583 static bool classof(const SDNode *N) {
1584 return N->getOpcode() == X86ISD::VTRUNCSTORES;
1588 // X86 Truncating Store with Unsigned saturation.
1589 class TruncUSStoreSDNode : public X86StoreSDNode {
1591 TruncUSStoreSDNode(unsigned Order, const DebugLoc &dl,
1592 SDVTList VTs, EVT MemVT, MachineMemOperand *MMO)
1593 : X86StoreSDNode(X86ISD::VTRUNCSTOREUS, Order, dl, VTs, MemVT, MMO) {}
1595 static bool classof(const SDNode *N) {
1596 return N->getOpcode() == X86ISD::VTRUNCSTOREUS;
1600 // X86 Truncating Masked Store with Signed saturation.
1601 class MaskedTruncSStoreSDNode : public X86MaskedStoreSDNode {
1603 MaskedTruncSStoreSDNode(unsigned Order,
1604 const DebugLoc &dl, SDVTList VTs, EVT MemVT,
1605 MachineMemOperand *MMO)
1606 : X86MaskedStoreSDNode(X86ISD::VMTRUNCSTORES, Order, dl, VTs, MemVT, MMO) {}
1608 static bool classof(const SDNode *N) {
1609 return N->getOpcode() == X86ISD::VMTRUNCSTORES;
1613 // X86 Truncating Masked Store with Unsigned saturation.
1614 class MaskedTruncUSStoreSDNode : public X86MaskedStoreSDNode {
1616 MaskedTruncUSStoreSDNode(unsigned Order,
1617 const DebugLoc &dl, SDVTList VTs, EVT MemVT,
1618 MachineMemOperand *MMO)
1619 : X86MaskedStoreSDNode(X86ISD::VMTRUNCSTOREUS, Order, dl, VTs, MemVT, MMO) {}
1621 static bool classof(const SDNode *N) {
1622 return N->getOpcode() == X86ISD::VMTRUNCSTOREUS;
1626 // X86 specific Gather/Scatter nodes.
1627 // The class has the same order of operands as MaskedGatherScatterSDNode for
1629 class X86MaskedGatherScatterSDNode : public MemSDNode {
1631 X86MaskedGatherScatterSDNode(unsigned Opc, unsigned Order,
1632 const DebugLoc &dl, SDVTList VTs, EVT MemVT,
1633 MachineMemOperand *MMO)
1634 : MemSDNode(Opc, Order, dl, VTs, MemVT, MMO) {}
1636 const SDValue &getBasePtr() const { return getOperand(3); }
1637 const SDValue &getIndex() const { return getOperand(4); }
1638 const SDValue &getMask() const { return getOperand(2); }
1639 const SDValue &getScale() const { return getOperand(5); }
1641 static bool classof(const SDNode *N) {
1642 return N->getOpcode() == X86ISD::MGATHER ||
1643 N->getOpcode() == X86ISD::MSCATTER;
1647 class X86MaskedGatherSDNode : public X86MaskedGatherScatterSDNode {
1649 X86MaskedGatherSDNode(unsigned Order, const DebugLoc &dl, SDVTList VTs,
1650 EVT MemVT, MachineMemOperand *MMO)
1651 : X86MaskedGatherScatterSDNode(X86ISD::MGATHER, Order, dl, VTs, MemVT,
1654 const SDValue &getPassThru() const { return getOperand(1); }
1656 static bool classof(const SDNode *N) {
1657 return N->getOpcode() == X86ISD::MGATHER;
1661 class X86MaskedScatterSDNode : public X86MaskedGatherScatterSDNode {
1663 X86MaskedScatterSDNode(unsigned Order, const DebugLoc &dl, SDVTList VTs,
1664 EVT MemVT, MachineMemOperand *MMO)
1665 : X86MaskedGatherScatterSDNode(X86ISD::MSCATTER, Order, dl, VTs, MemVT,
1668 const SDValue &getValue() const { return getOperand(1); }
1670 static bool classof(const SDNode *N) {
1671 return N->getOpcode() == X86ISD::MSCATTER;
1675 /// Generate unpacklo/unpackhi shuffle mask.
1676 template <typename T = int>
1677 void createUnpackShuffleMask(MVT VT, SmallVectorImpl<T> &Mask, bool Lo,
1679 assert(Mask.empty() && "Expected an empty shuffle mask vector");
1680 int NumElts = VT.getVectorNumElements();
1681 int NumEltsInLane = 128 / VT.getScalarSizeInBits();
1682 for (int i = 0; i < NumElts; ++i) {
1683 unsigned LaneStart = (i / NumEltsInLane) * NumEltsInLane;
1684 int Pos = (i % NumEltsInLane) / 2 + LaneStart;
1685 Pos += (Unary ? 0 : NumElts * (i % 2));
1686 Pos += (Lo ? 0 : NumEltsInLane / 2);
1687 Mask.push_back(Pos);
1691 /// Helper function to scale a shuffle or target shuffle mask, replacing each
1692 /// mask index with the scaled sequential indices for an equivalent narrowed
1693 /// mask. This is the reverse process to canWidenShuffleElements, but can
1695 template <typename T>
1696 void scaleShuffleMask(size_t Scale, ArrayRef<T> Mask,
1697 SmallVectorImpl<T> &ScaledMask) {
1698 assert(0 < Scale && "Unexpected scaling factor");
1699 size_t NumElts = Mask.size();
1700 ScaledMask.assign(NumElts * Scale, -1);
1702 for (size_t i = 0; i != NumElts; ++i) {
1705 // Repeat sentinel values in every mask element.
1707 for (size_t s = 0; s != Scale; ++s)
1708 ScaledMask[(Scale * i) + s] = M;
1712 // Scale mask element and increment across each mask element.
1713 for (size_t s = 0; s != Scale; ++s)
1714 ScaledMask[(Scale * i) + s] = (Scale * M) + s;
1717 } // end namespace llvm
1719 #endif // LLVM_LIB_TARGET_X86_X86ISELLOWERING_H