1 //===-- X86ISelLowering.h - X86 DAG Lowering Interface ----------*- C++ -*-===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file defines the interfaces that X86 uses to lower LLVM code into a
13 //===----------------------------------------------------------------------===//
15 #ifndef LLVM_LIB_TARGET_X86_X86ISELLOWERING_H
16 #define LLVM_LIB_TARGET_X86_X86ISELLOWERING_H
18 #include "llvm/CodeGen/CallingConvLower.h"
19 #include "llvm/CodeGen/SelectionDAG.h"
20 #include "llvm/Target/TargetLowering.h"
21 #include "llvm/Target/TargetOptions.h"
25 class X86TargetMachine;
28 // X86 Specific DAG Nodes
29 enum NodeType : unsigned {
30 // Start the numbering where the builtin ops leave off.
31 FIRST_NUMBER = ISD::BUILTIN_OP_END,
38 /// Double shift instructions. These correspond to
39 /// X86::SHLDxx and X86::SHRDxx instructions.
43 /// Bitwise logical AND of floating point values. This corresponds
44 /// to X86::ANDPS or X86::ANDPD.
47 /// Bitwise logical OR of floating point values. This corresponds
48 /// to X86::ORPS or X86::ORPD.
51 /// Bitwise logical XOR of floating point values. This corresponds
52 /// to X86::XORPS or X86::XORPD.
55 /// Bitwise logical ANDNOT of floating point values. This
56 /// corresponds to X86::ANDNPS or X86::ANDNPD.
59 /// These operations represent an abstract X86 call
60 /// instruction, which includes a bunch of information. In particular the
61 /// operands of these node are:
63 /// #0 - The incoming token chain
65 /// #2 - The number of arg bytes the caller pushes on the stack.
66 /// #3 - The number of arg bytes the callee pops off the stack.
67 /// #4 - The value to pass in AL/AX/EAX (optional)
68 /// #5 - The value to pass in DL/DX/EDX (optional)
70 /// The result values of these nodes are:
72 /// #0 - The outgoing token chain
73 /// #1 - The first register result value (optional)
74 /// #2 - The second register result value (optional)
78 /// This operation implements the lowering for readcyclecounter
81 /// X86 Read Time-Stamp Counter and Processor ID.
84 /// X86 Read Performance Monitoring Counters.
87 /// X86 compare and logical compare instructions.
90 /// X86 bit-test instructions.
93 /// X86 SetCC. Operand 0 is condition code, and operand 1 is the EFLAGS
94 /// operand, usually produced by a CMP instruction.
100 // Same as SETCC except it's materialized with a sbb and the value is all
101 // one's or all zero's.
102 SETCC_CARRY, // R = carry_bit ? ~0 : 0
104 /// X86 FP SETCC, implemented with CMP{cc}SS/CMP{cc}SD.
105 /// Operands are two FP values to compare; result is a mask of
106 /// 0s or 1s. Generally DTRT for C/C++ with NaNs.
109 /// X86 MOVMSK{pd|ps}, extracts sign bits of two or four FP values,
110 /// result in an integer GPR. Needs masking for scalar result.
113 /// X86 conditional moves. Operand 0 and operand 1 are the two values
114 /// to select from. Operand 2 is the condition code, and operand 3 is the
115 /// flag operand produced by a CMP or TEST instruction. It also writes a
119 /// X86 conditional branches. Operand 0 is the chain operand, operand 1
120 /// is the block to branch if condition is true, operand 2 is the
121 /// condition code, and operand 3 is the flag operand produced by a CMP
122 /// or TEST instruction.
125 /// Return with a flag operand. Operand 0 is the chain operand, operand
126 /// 1 is the number of bytes of stack to pop.
129 /// Repeat fill, corresponds to X86::REP_STOSx.
132 /// Repeat move, corresponds to X86::REP_MOVSx.
135 /// On Darwin, this node represents the result of the popl
136 /// at function entry, used for PIC code.
139 /// A wrapper node for TargetConstantPool,
140 /// TargetExternalSymbol, and TargetGlobalAddress.
143 /// Special wrapper used under X86-64 PIC mode for RIP
144 /// relative displacements.
147 /// Copies a 64-bit value from the low word of an XMM vector
148 /// to an MMX vector. If you think this is too close to the previous
149 /// mnemonic, so do I; blame Intel.
152 /// Copies a 32-bit value from the low word of a MMX
156 /// Copies a GPR into the low 32-bit word of a MMX vector
157 /// and zero out the high word.
160 /// Extract an 8-bit value from a vector and zero extend it to
161 /// i32, corresponds to X86::PEXTRB.
164 /// Extract a 16-bit value from a vector and zero extend it to
165 /// i32, corresponds to X86::PEXTRW.
168 /// Insert any element of a 4 x float vector into any element
169 /// of a destination 4 x floatvector.
172 /// Insert the lower 8-bits of a 32-bit value to a vector,
173 /// corresponds to X86::PINSRB.
176 /// Insert the lower 16-bits of a 32-bit value to a vector,
177 /// corresponds to X86::PINSRW.
180 /// Shuffle 16 8-bit values within a vector.
183 /// Compute Sum of Absolute Differences.
185 /// Compute Double Block Packed Sum-Absolute-Differences
188 /// Bitwise Logical AND NOT of Packed FP values.
191 /// Copy integer sign.
194 /// Blend where the selector is an immediate.
197 /// Blend where the condition has been shrunk.
198 /// This is used to emphasize that the condition mask is
199 /// no more valid for generic VSELECT optimizations.
202 /// Combined add and sub on an FP vector.
205 // FP vector ops with rounding mode.
214 // FP vector get exponent
216 // Extract Normalized Mantissas
220 // Integer add/sub with unsigned saturation.
223 // Integer add/sub with signed saturation.
226 // Unsigned Integer average
228 /// Integer horizontal add.
231 /// Integer horizontal sub.
234 /// Floating point horizontal add.
237 /// Floating point horizontal sub.
240 // Integer absolute value
243 /// Floating point max and min.
246 /// Commutative FMIN and FMAX.
249 /// Floating point reciprocal-sqrt and reciprocal approximation.
250 /// Note that these typically require refinement
251 /// in order to obtain suitable precision.
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.
270 // SjLj exception handling setjmp.
273 // SjLj exception handling longjmp.
276 /// Tail call return. See X86TargetLowering::LowerCall for
277 /// the list of operands.
280 // Vector move to low scalar and zero higher vector elements.
283 // Vector integer zero-extend.
286 // Vector integer signed-extend.
289 // Vector integer truncate.
291 // Vector integer truncate with unsigned/signed saturation.
300 // Vector signed/unsigned integer to double.
303 // 128-bit vector logical left / right shift
306 // Vector shift elements
309 // Vector shift elements by immediate
312 // Vector packed double/float comparison.
315 // Vector integer comparisons.
317 // Vector integer comparisons, the result is in a mask vector.
320 /// Vector comparison generating mask bits for fp and
321 /// integer signed and unsigned data types.
324 // Vector comparison with rounding mode for FP values
327 // Arithmetic operations with FLAGS results.
328 ADD, SUB, ADC, SBB, SMUL,
329 INC, DEC, OR, XOR, AND,
331 BEXTR, // Bit field extract
333 UMUL, // LOW, HI, FLAGS = umul LHS, RHS
335 // 8-bit SMUL/UMUL - AX, FLAGS = smul8/umul8 AL, RHS
338 // 8-bit divrem that zero-extend the high result (AH).
342 // X86-specific multiply by immediate.
345 // Vector bitwise comparisons.
348 // Vector packed fp sign bitwise comparisons.
351 // Vector "test" in AVX-512, the result is in a mask vector.
355 // OR/AND test for masks
359 // Several flavors of instructions with vector shuffle behaviors.
364 // AVX512 inter-lane alignr
370 //Shuffle Packed Values at 128-bit granularity
391 //Fix Up Special Packed Float32/64 values
393 //Range Restriction Calculation For Packed Pairs of Float32/64 values
395 // Reduce - Perform Reduction Transformation on scalar\packed FP
397 // RndScale - Round FP Values To Include A Given Number Of Fraction Bits
399 // Broadcast scalar to vector
401 // Broadcast subvector to vector
403 // Insert/Extract vector element
407 /// SSE4A Extraction and Insertion.
410 // Vector multiply packed unsigned doubleword integers
412 // Vector multiply packed signed doubleword integers
414 // Vector Multiply Packed UnsignedIntegers with Round and Scale
416 // Multiply and Add Packed Integers
417 VPMADDUBSW, VPMADDWD,
425 // FMA with rounding mode
433 // Compress and expand
437 //Convert Unsigned/Integer to Scalar Floating-Point Value
442 // Vector float/double to signed/unsigned integer.
443 FP_TO_SINT_RND, FP_TO_UINT_RND,
444 // Save xmm argument registers to the stack, according to %al. An operator
445 // is needed so that this can be expanded with control flow.
446 VASTART_SAVE_XMM_REGS,
448 // Windows's _chkstk call to do stack probing.
451 // For allocating variable amounts of stack space when using
452 // segmented stacks. Check if the current stacklet has enough space, and
453 // falls back to heap allocation if not.
462 // Store FP status word into i16 register.
465 // Store contents of %ah into %eflags.
468 // Get a random integer and indicate whether it is valid in CF.
471 // Get a NIST SP800-90B & C compliant random integer and
472 // indicate whether it is valid in CF.
478 // Test if in transactional execution.
482 RSQRT28, RCP28, EXP2,
485 LCMPXCHG_DAG = ISD::FIRST_TARGET_MEMORY_OPCODE,
489 // Load, scalar_to_vector, and zero extend.
492 // Store FP control world into i16 memory.
495 /// This instruction implements FP_TO_SINT with the
496 /// integer destination in memory and a FP reg source. This corresponds
497 /// to the X86::FIST*m instructions and the rounding mode change stuff. It
498 /// has two inputs (token chain and address) and two outputs (int value
499 /// and token chain).
504 /// This instruction implements SINT_TO_FP with the
505 /// integer source in memory and FP reg result. This corresponds to the
506 /// X86::FILD*m instructions. It has three inputs (token chain, address,
507 /// and source type) and two outputs (FP value and token chain). FILD_FLAG
508 /// also produces a flag).
512 /// This instruction implements an extending load to FP stack slots.
513 /// This corresponds to the X86::FLD32m / X86::FLD64m. It takes a chain
514 /// operand, ptr to load from, and a ValueType node indicating the type
518 /// This instruction implements a truncating store to FP stack
519 /// slots. This corresponds to the X86::FST32m / X86::FST64m. It takes a
520 /// chain operand, value to store, address, and a ValueType to store it
524 /// This instruction grabs the address of the next argument
525 /// from a va_list. (reads and modifies the va_list in memory)
528 // WARNING: Do not add anything in the end unless you want the node to
529 // have memop! In fact, starting from ATOMADD64_DAG all opcodes will be
530 // thought as target memory ops!
534 /// Define some predicates that are used for node matching.
536 /// Return true if the specified
537 /// EXTRACT_SUBVECTOR operand specifies a vector extract that is
538 /// suitable for input to VEXTRACTF128, VEXTRACTI128 instructions.
539 bool isVEXTRACT128Index(SDNode *N);
541 /// Return true if the specified
542 /// INSERT_SUBVECTOR operand specifies a subvector insert that is
543 /// suitable for input to VINSERTF128, VINSERTI128 instructions.
544 bool isVINSERT128Index(SDNode *N);
546 /// Return true if the specified
547 /// EXTRACT_SUBVECTOR operand specifies a vector extract that is
548 /// suitable for input to VEXTRACTF64X4, VEXTRACTI64X4 instructions.
549 bool isVEXTRACT256Index(SDNode *N);
551 /// Return true if the specified
552 /// INSERT_SUBVECTOR operand specifies a subvector insert that is
553 /// suitable for input to VINSERTF64X4, VINSERTI64X4 instructions.
554 bool isVINSERT256Index(SDNode *N);
556 /// Return the appropriate
557 /// immediate to extract the specified EXTRACT_SUBVECTOR index
558 /// with VEXTRACTF128, VEXTRACTI128 instructions.
559 unsigned getExtractVEXTRACT128Immediate(SDNode *N);
561 /// Return the appropriate
562 /// immediate to insert at the specified INSERT_SUBVECTOR index
563 /// with VINSERTF128, VINSERT128 instructions.
564 unsigned getInsertVINSERT128Immediate(SDNode *N);
566 /// Return the appropriate
567 /// immediate to extract the specified EXTRACT_SUBVECTOR index
568 /// with VEXTRACTF64X4, VEXTRACTI64x4 instructions.
569 unsigned getExtractVEXTRACT256Immediate(SDNode *N);
571 /// Return the appropriate
572 /// immediate to insert at the specified INSERT_SUBVECTOR index
573 /// with VINSERTF64x4, VINSERTI64x4 instructions.
574 unsigned getInsertVINSERT256Immediate(SDNode *N);
576 /// Returns true if Elt is a constant zero or floating point constant +0.0.
577 bool isZeroNode(SDValue Elt);
579 /// Returns true of the given offset can be
580 /// fit into displacement field of the instruction.
581 bool isOffsetSuitableForCodeModel(int64_t Offset, CodeModel::Model M,
582 bool hasSymbolicDisplacement = true);
585 /// Determines whether the callee is required to pop its
586 /// own arguments. Callee pop is necessary to support tail calls.
587 bool isCalleePop(CallingConv::ID CallingConv,
588 bool is64Bit, bool IsVarArg, bool TailCallOpt);
590 /// AVX512 static rounding constants. These need to match the values in
592 enum STATIC_ROUNDING {
601 //===--------------------------------------------------------------------===//
602 // X86 Implementation of the TargetLowering interface
603 class X86TargetLowering final : public TargetLowering {
605 explicit X86TargetLowering(const X86TargetMachine &TM,
606 const X86Subtarget &STI);
608 unsigned getJumpTableEncoding() const override;
609 bool useSoftFloat() const override;
611 MVT getScalarShiftAmountTy(const DataLayout &, EVT) const override {
616 LowerCustomJumpTableEntry(const MachineJumpTableInfo *MJTI,
617 const MachineBasicBlock *MBB, unsigned uid,
618 MCContext &Ctx) const override;
620 /// Returns relocation base for the given PIC jumptable.
621 SDValue getPICJumpTableRelocBase(SDValue Table,
622 SelectionDAG &DAG) const override;
624 getPICJumpTableRelocBaseExpr(const MachineFunction *MF,
625 unsigned JTI, MCContext &Ctx) const override;
627 /// Return the desired alignment for ByVal aggregate
628 /// function arguments in the caller parameter area. For X86, aggregates
629 /// that contains are placed at 16-byte boundaries while the rest are at
630 /// 4-byte boundaries.
631 unsigned getByValTypeAlignment(Type *Ty,
632 const DataLayout &DL) const override;
634 /// Returns the target specific optimal type for load
635 /// and store operations as a result of memset, memcpy, and memmove
636 /// lowering. If DstAlign is zero that means it's safe to destination
637 /// alignment can satisfy any constraint. Similarly if SrcAlign is zero it
638 /// means there isn't a need to check it against alignment requirement,
639 /// probably because the source does not need to be loaded. If 'IsMemset' is
640 /// true, that means it's expanding a memset. If 'ZeroMemset' is true, that
641 /// means it's a memset of zero. 'MemcpyStrSrc' indicates whether the memcpy
642 /// source is constant so it does not need to be loaded.
643 /// It returns EVT::Other if the type should be determined using generic
644 /// target-independent logic.
645 EVT getOptimalMemOpType(uint64_t Size, unsigned DstAlign, unsigned SrcAlign,
646 bool IsMemset, bool ZeroMemset, bool MemcpyStrSrc,
647 MachineFunction &MF) const override;
649 /// Returns true if it's safe to use load / store of the
650 /// specified type to expand memcpy / memset inline. This is mostly true
651 /// for all types except for some special cases. For example, on X86
652 /// targets without SSE2 f64 load / store are done with fldl / fstpl which
653 /// also does type conversion. Note the specified type doesn't have to be
654 /// legal as the hook is used before type legalization.
655 bool isSafeMemOpType(MVT VT) const override;
657 /// Returns true if the target allows unaligned memory accesses of the
658 /// specified type. Returns whether it is "fast" in the last argument.
659 bool allowsMisalignedMemoryAccesses(EVT VT, unsigned AS, unsigned Align,
660 bool *Fast) const override;
662 /// Provide custom lowering hooks for some operations.
664 SDValue LowerOperation(SDValue Op, SelectionDAG &DAG) const override;
666 /// Replace the results of node with an illegal result
667 /// type with new values built out of custom code.
669 void ReplaceNodeResults(SDNode *N, SmallVectorImpl<SDValue>&Results,
670 SelectionDAG &DAG) const override;
673 SDValue PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const override;
675 /// Return true if the target has native support for
676 /// the specified value type and it is 'desirable' to use the type for the
677 /// given node type. e.g. On x86 i16 is legal, but undesirable since i16
678 /// instruction encodings are longer and some i16 instructions are slow.
679 bool isTypeDesirableForOp(unsigned Opc, EVT VT) const override;
681 /// Return true if the target has native support for the
682 /// specified value type and it is 'desirable' to use the type. e.g. On x86
683 /// i16 is legal, but undesirable since i16 instruction encodings are longer
684 /// and some i16 instructions are slow.
685 bool IsDesirableToPromoteOp(SDValue Op, EVT &PVT) const override;
688 EmitInstrWithCustomInserter(MachineInstr *MI,
689 MachineBasicBlock *MBB) const override;
692 /// This method returns the name of a target specific DAG node.
693 const char *getTargetNodeName(unsigned Opcode) const override;
695 bool isCheapToSpeculateCttz() const override;
697 bool isCheapToSpeculateCtlz() const override;
699 /// Return the value type to use for ISD::SETCC.
700 EVT getSetCCResultType(const DataLayout &DL, LLVMContext &Context,
701 EVT VT) const override;
703 /// Determine which of the bits specified in Mask are known to be either
704 /// zero or one and return them in the KnownZero/KnownOne bitsets.
705 void computeKnownBitsForTargetNode(const SDValue Op,
708 const SelectionDAG &DAG,
709 unsigned Depth = 0) const override;
711 /// Determine the number of bits in the operation that are sign bits.
712 unsigned ComputeNumSignBitsForTargetNode(SDValue Op,
713 const SelectionDAG &DAG,
714 unsigned Depth) const override;
716 bool isGAPlusOffset(SDNode *N, const GlobalValue* &GA,
717 int64_t &Offset) const override;
719 SDValue getReturnAddressFrameIndex(SelectionDAG &DAG) const;
721 bool ExpandInlineAsm(CallInst *CI) const override;
723 ConstraintType getConstraintType(StringRef Constraint) const override;
725 /// Examine constraint string and operand type and determine a weight value.
726 /// The operand object must already have been set up with the operand type.
728 getSingleConstraintMatchWeight(AsmOperandInfo &info,
729 const char *constraint) const override;
731 const char *LowerXConstraint(EVT ConstraintVT) const override;
733 /// Lower the specified operand into the Ops vector. If it is invalid, don't
734 /// add anything to Ops. If hasMemory is true it means one of the asm
735 /// constraint of the inline asm instruction being processed is 'm'.
736 void LowerAsmOperandForConstraint(SDValue Op,
737 std::string &Constraint,
738 std::vector<SDValue> &Ops,
739 SelectionDAG &DAG) const override;
742 getInlineAsmMemConstraint(StringRef ConstraintCode) const override {
743 if (ConstraintCode == "i")
744 return InlineAsm::Constraint_i;
745 else if (ConstraintCode == "o")
746 return InlineAsm::Constraint_o;
747 else if (ConstraintCode == "v")
748 return InlineAsm::Constraint_v;
749 else if (ConstraintCode == "X")
750 return InlineAsm::Constraint_X;
751 return TargetLowering::getInlineAsmMemConstraint(ConstraintCode);
754 /// Given a physical register constraint
755 /// (e.g. {edx}), return the register number and the register class for the
756 /// register. This should only be used for C_Register constraints. On
757 /// error, this returns a register number of 0.
758 std::pair<unsigned, const TargetRegisterClass *>
759 getRegForInlineAsmConstraint(const TargetRegisterInfo *TRI,
760 StringRef Constraint, MVT VT) const override;
762 /// Return true if the addressing mode represented
763 /// by AM is legal for this target, for a load/store of the specified type.
764 bool isLegalAddressingMode(const DataLayout &DL, const AddrMode &AM,
765 Type *Ty, unsigned AS) const override;
767 /// Return true if the specified immediate is legal
768 /// icmp immediate, that is the target has icmp instructions which can
769 /// compare a register against the immediate without having to materialize
770 /// the immediate into a register.
771 bool isLegalICmpImmediate(int64_t Imm) const override;
773 /// Return true if the specified immediate is legal
774 /// add immediate, that is the target has add instructions which can
775 /// add a register and the immediate without having to materialize
776 /// the immediate into a register.
777 bool isLegalAddImmediate(int64_t Imm) const override;
779 /// \brief Return the cost of the scaling factor used in the addressing
780 /// mode represented by AM for this target, for a load/store
781 /// of the specified type.
782 /// If the AM is supported, the return value must be >= 0.
783 /// If the AM is not supported, it returns a negative value.
784 int getScalingFactorCost(const DataLayout &DL, const AddrMode &AM, Type *Ty,
785 unsigned AS) const override;
787 bool isVectorShiftByScalarCheap(Type *Ty) const override;
789 /// Return true if it's free to truncate a value of
790 /// type Ty1 to type Ty2. e.g. On x86 it's free to truncate a i32 value in
791 /// register EAX to i16 by referencing its sub-register AX.
792 bool isTruncateFree(Type *Ty1, Type *Ty2) const override;
793 bool isTruncateFree(EVT VT1, EVT VT2) const override;
795 bool allowTruncateForTailCall(Type *Ty1, Type *Ty2) const override;
797 /// Return true if any actual instruction that defines a
798 /// value of type Ty1 implicit zero-extends the value to Ty2 in the result
799 /// register. This does not necessarily include registers defined in
800 /// unknown ways, such as incoming arguments, or copies from unknown
801 /// virtual registers. Also, if isTruncateFree(Ty2, Ty1) is true, this
802 /// does not necessarily apply to truncate instructions. e.g. on x86-64,
803 /// all instructions that define 32-bit values implicit zero-extend the
804 /// result out to 64 bits.
805 bool isZExtFree(Type *Ty1, Type *Ty2) const override;
806 bool isZExtFree(EVT VT1, EVT VT2) const override;
807 bool isZExtFree(SDValue Val, EVT VT2) const override;
809 /// Return true if folding a vector load into ExtVal (a sign, zero, or any
810 /// extend node) is profitable.
811 bool isVectorLoadExtDesirable(SDValue) const override;
813 /// Return true if an FMA operation is faster than a pair of fmul and fadd
814 /// instructions. fmuladd intrinsics will be expanded to FMAs when this
815 /// method returns true, otherwise fmuladd is expanded to fmul + fadd.
816 bool isFMAFasterThanFMulAndFAdd(EVT VT) const override;
818 /// Return true if it's profitable to narrow
819 /// operations of type VT1 to VT2. e.g. on x86, it's profitable to narrow
820 /// from i32 to i8 but not from i32 to i16.
821 bool isNarrowingProfitable(EVT VT1, EVT VT2) const override;
823 /// Returns true if the target can instruction select the
824 /// specified FP immediate natively. If false, the legalizer will
825 /// materialize the FP immediate as a load from a constant pool.
826 bool isFPImmLegal(const APFloat &Imm, EVT VT) const override;
828 /// Targets can use this to indicate that they only support *some*
829 /// VECTOR_SHUFFLE operations, those with specific masks. By default, if a
830 /// target supports the VECTOR_SHUFFLE node, all mask values are assumed to
832 bool isShuffleMaskLegal(const SmallVectorImpl<int> &Mask,
833 EVT VT) const override;
835 /// Similar to isShuffleMaskLegal. This is used by Targets can use this to
836 /// indicate if there is a suitable VECTOR_SHUFFLE that can be used to
837 /// replace a VAND with a constant pool entry.
838 bool isVectorClearMaskLegal(const SmallVectorImpl<int> &Mask,
839 EVT VT) const override;
841 /// If true, then instruction selection should
842 /// seek to shrink the FP constant of the specified type to a smaller type
843 /// in order to save space and / or reduce runtime.
844 bool ShouldShrinkFPConstant(EVT VT) const override {
845 // Don't shrink FP constpool if SSE2 is available since cvtss2sd is more
846 // expensive than a straight movsd. On the other hand, it's important to
847 // shrink long double fp constant since fldt is very slow.
848 return !X86ScalarSSEf64 || VT == MVT::f80;
851 /// Return true if we believe it is correct and profitable to reduce the
852 /// load node to a smaller type.
853 bool shouldReduceLoadWidth(SDNode *Load, ISD::LoadExtType ExtTy,
854 EVT NewVT) const override;
856 /// Return true if the specified scalar FP type is computed in an SSE
857 /// register, not on the X87 floating point stack.
858 bool isScalarFPTypeInSSEReg(EVT VT) const {
859 return (VT == MVT::f64 && X86ScalarSSEf64) || // f64 is when SSE2
860 (VT == MVT::f32 && X86ScalarSSEf32); // f32 is when SSE1
863 /// \brief Returns true if it is beneficial to convert a load of a constant
864 /// to just the constant itself.
865 bool shouldConvertConstantLoadToIntImm(const APInt &Imm,
866 Type *Ty) const override;
868 /// Return true if EXTRACT_SUBVECTOR is cheap for this result type
870 bool isExtractSubvectorCheap(EVT ResVT, unsigned Index) const override;
872 /// Intel processors have a unified instruction and data cache
873 const char * getClearCacheBuiltinName() const override {
874 return nullptr; // nothing to do, move along.
877 unsigned getRegisterByName(const char* RegName, EVT VT,
878 SelectionDAG &DAG) const override;
880 /// This method returns a target specific FastISel object,
881 /// or null if the target does not support "fast" ISel.
882 FastISel *createFastISel(FunctionLoweringInfo &funcInfo,
883 const TargetLibraryInfo *libInfo) const override;
885 /// Return true if the target stores stack protector cookies at a fixed
886 /// offset in some non-standard address space, and populates the address
887 /// space and offset as appropriate.
888 bool getStackCookieLocation(unsigned &AddressSpace,
889 unsigned &Offset) const override;
891 SDValue BuildFILD(SDValue Op, EVT SrcVT, SDValue Chain, SDValue StackSlot,
892 SelectionDAG &DAG) const;
894 bool isNoopAddrSpaceCast(unsigned SrcAS, unsigned DestAS) const override;
896 bool useLoadStackGuardNode() const override;
897 /// \brief Customize the preferred legalization strategy for certain types.
898 LegalizeTypeAction getPreferredVectorAction(EVT VT) const override;
900 bool isIntDivCheap(EVT VT, AttributeSet Attr) const override;
903 std::pair<const TargetRegisterClass *, uint8_t>
904 findRepresentativeClass(const TargetRegisterInfo *TRI,
905 MVT VT) const override;
908 /// Keep a pointer to the X86Subtarget around so that we can
909 /// make the right decision when generating code for different targets.
910 const X86Subtarget *Subtarget;
911 const DataLayout *TD;
913 /// Select between SSE or x87 floating point ops.
914 /// When SSE is available, use it for f32 operations.
915 /// When SSE2 is available, use it for f64 operations.
916 bool X86ScalarSSEf32;
917 bool X86ScalarSSEf64;
919 /// A list of legal FP immediates.
920 std::vector<APFloat> LegalFPImmediates;
922 /// Indicate that this x86 target can instruction
923 /// select the specified FP immediate natively.
924 void addLegalFPImmediate(const APFloat& Imm) {
925 LegalFPImmediates.push_back(Imm);
928 SDValue LowerCallResult(SDValue Chain, SDValue InFlag,
929 CallingConv::ID CallConv, bool isVarArg,
930 const SmallVectorImpl<ISD::InputArg> &Ins,
931 SDLoc dl, SelectionDAG &DAG,
932 SmallVectorImpl<SDValue> &InVals) const;
933 SDValue LowerMemArgument(SDValue Chain,
934 CallingConv::ID CallConv,
935 const SmallVectorImpl<ISD::InputArg> &ArgInfo,
936 SDLoc dl, SelectionDAG &DAG,
937 const CCValAssign &VA, MachineFrameInfo *MFI,
939 SDValue LowerMemOpCallTo(SDValue Chain, SDValue StackPtr, SDValue Arg,
940 SDLoc dl, SelectionDAG &DAG,
941 const CCValAssign &VA,
942 ISD::ArgFlagsTy Flags) const;
944 // Call lowering helpers.
946 /// Check whether the call is eligible for tail call optimization. Targets
947 /// that want to do tail call optimization should implement this function.
948 bool IsEligibleForTailCallOptimization(SDValue Callee,
949 CallingConv::ID CalleeCC,
951 bool isCalleeStructRet,
952 bool isCallerStructRet,
954 const SmallVectorImpl<ISD::OutputArg> &Outs,
955 const SmallVectorImpl<SDValue> &OutVals,
956 const SmallVectorImpl<ISD::InputArg> &Ins,
957 SelectionDAG& DAG) const;
958 bool IsCalleePop(bool isVarArg, CallingConv::ID CallConv) const;
959 SDValue EmitTailCallLoadRetAddr(SelectionDAG &DAG, SDValue &OutRetAddr,
960 SDValue Chain, bool IsTailCall, bool Is64Bit,
961 int FPDiff, SDLoc dl) const;
963 unsigned GetAlignedArgumentStackSize(unsigned StackSize,
964 SelectionDAG &DAG) const;
966 std::pair<SDValue,SDValue> FP_TO_INTHelper(SDValue Op, SelectionDAG &DAG,
968 bool isReplace) const;
970 SDValue LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) const;
971 SDValue LowerBUILD_VECTORvXi1(SDValue Op, SelectionDAG &DAG) const;
972 SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) const;
973 SDValue LowerVSELECT(SDValue Op, SelectionDAG &DAG) const;
974 SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const;
975 SDValue ExtractBitFromMaskVector(SDValue Op, SelectionDAG &DAG) const;
976 SDValue InsertBitToMaskVector(SDValue Op, SelectionDAG &DAG) const;
978 SDValue LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const;
979 SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) const;
980 SDValue LowerBlockAddress(SDValue Op, SelectionDAG &DAG) const;
981 SDValue LowerGlobalAddress(const GlobalValue *GV, SDLoc dl,
982 int64_t Offset, SelectionDAG &DAG) const;
983 SDValue LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) const;
984 SDValue LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const;
985 SDValue LowerExternalSymbol(SDValue Op, SelectionDAG &DAG) const;
986 SDValue LowerSINT_TO_FP(SDValue Op, SelectionDAG &DAG) const;
987 SDValue LowerUINT_TO_FP(SDValue Op, SelectionDAG &DAG) const;
988 SDValue LowerUINT_TO_FP_i64(SDValue Op, SelectionDAG &DAG) const;
989 SDValue LowerUINT_TO_FP_i32(SDValue Op, SelectionDAG &DAG) const;
990 SDValue lowerUINT_TO_FP_vec(SDValue Op, SelectionDAG &DAG) const;
991 SDValue LowerTRUNCATE(SDValue Op, SelectionDAG &DAG) const;
992 SDValue LowerFP_TO_SINT(SDValue Op, SelectionDAG &DAG) const;
993 SDValue LowerFP_TO_UINT(SDValue Op, SelectionDAG &DAG) const;
994 SDValue LowerToBT(SDValue And, ISD::CondCode CC,
995 SDLoc dl, SelectionDAG &DAG) const;
996 SDValue LowerSETCC(SDValue Op, SelectionDAG &DAG) const;
997 SDValue LowerSELECT(SDValue Op, SelectionDAG &DAG) const;
998 SDValue LowerBRCOND(SDValue Op, SelectionDAG &DAG) const;
999 SDValue LowerMEMSET(SDValue Op, SelectionDAG &DAG) const;
1000 SDValue LowerJumpTable(SDValue Op, SelectionDAG &DAG) const;
1001 SDValue LowerDYNAMIC_STACKALLOC(SDValue Op, SelectionDAG &DAG) const;
1002 SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG) const;
1003 SDValue LowerVAARG(SDValue Op, SelectionDAG &DAG) const;
1004 SDValue LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const;
1005 SDValue LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const;
1006 SDValue LowerFRAME_TO_ARGS_OFFSET(SDValue Op, SelectionDAG &DAG) const;
1007 SDValue LowerEH_RETURN(SDValue Op, SelectionDAG &DAG) const;
1008 SDValue LowerCATCHRET(SDValue Op, SelectionDAG &DAG) const;
1009 SDValue lowerEH_SJLJ_SETJMP(SDValue Op, SelectionDAG &DAG) const;
1010 SDValue lowerEH_SJLJ_LONGJMP(SDValue Op, SelectionDAG &DAG) const;
1011 SDValue LowerINIT_TRAMPOLINE(SDValue Op, SelectionDAG &DAG) const;
1012 SDValue LowerFLT_ROUNDS_(SDValue Op, SelectionDAG &DAG) const;
1013 SDValue LowerWin64_i128OP(SDValue Op, SelectionDAG &DAG) const;
1014 SDValue LowerGC_TRANSITION_START(SDValue Op, SelectionDAG &DAG) const;
1015 SDValue LowerGC_TRANSITION_END(SDValue Op, SelectionDAG &DAG) const;
1018 LowerFormalArguments(SDValue Chain,
1019 CallingConv::ID CallConv, bool isVarArg,
1020 const SmallVectorImpl<ISD::InputArg> &Ins,
1021 SDLoc dl, SelectionDAG &DAG,
1022 SmallVectorImpl<SDValue> &InVals) const override;
1023 SDValue LowerCall(CallLoweringInfo &CLI,
1024 SmallVectorImpl<SDValue> &InVals) const override;
1026 SDValue LowerReturn(SDValue Chain,
1027 CallingConv::ID CallConv, bool isVarArg,
1028 const SmallVectorImpl<ISD::OutputArg> &Outs,
1029 const SmallVectorImpl<SDValue> &OutVals,
1030 SDLoc dl, SelectionDAG &DAG) const override;
1032 bool isUsedByReturnOnly(SDNode *N, SDValue &Chain) const override;
1034 bool mayBeEmittedAsTailCall(CallInst *CI) const override;
1036 EVT getTypeForExtArgOrReturn(LLVMContext &Context, EVT VT,
1037 ISD::NodeType ExtendKind) const override;
1039 bool CanLowerReturn(CallingConv::ID CallConv, MachineFunction &MF,
1041 const SmallVectorImpl<ISD::OutputArg> &Outs,
1042 LLVMContext &Context) const override;
1044 const MCPhysReg *getScratchRegisters(CallingConv::ID CC) const override;
1046 bool shouldExpandAtomicLoadInIR(LoadInst *SI) const override;
1047 bool shouldExpandAtomicStoreInIR(StoreInst *SI) const override;
1048 TargetLoweringBase::AtomicRMWExpansionKind
1049 shouldExpandAtomicRMWInIR(AtomicRMWInst *AI) const override;
1052 lowerIdempotentRMWIntoFencedLoad(AtomicRMWInst *AI) const override;
1054 bool needsCmpXchgNb(Type *MemType) const;
1056 /// Utility function to emit atomic-load-arith operations (and, or, xor,
1057 /// nand, max, min, umax, umin). It takes the corresponding instruction to
1058 /// expand, the associated machine basic block, and the associated X86
1059 /// opcodes for reg/reg.
1060 MachineBasicBlock *EmitAtomicLoadArith(MachineInstr *MI,
1061 MachineBasicBlock *MBB) const;
1063 /// Utility function to emit atomic-load-arith operations (and, or, xor,
1064 /// nand, add, sub, swap) for 64-bit operands on 32-bit target.
1065 MachineBasicBlock *EmitAtomicLoadArith6432(MachineInstr *MI,
1066 MachineBasicBlock *MBB) const;
1068 // Utility function to emit the low-level va_arg code for X86-64.
1069 MachineBasicBlock *EmitVAARG64WithCustomInserter(
1071 MachineBasicBlock *MBB) const;
1073 /// Utility function to emit the xmm reg save portion of va_start.
1074 MachineBasicBlock *EmitVAStartSaveXMMRegsWithCustomInserter(
1075 MachineInstr *BInstr,
1076 MachineBasicBlock *BB) const;
1078 MachineBasicBlock *EmitLoweredSelect(MachineInstr *I,
1079 MachineBasicBlock *BB) const;
1081 MachineBasicBlock *EmitLoweredAtomicFP(MachineInstr *I,
1082 MachineBasicBlock *BB) const;
1084 MachineBasicBlock *EmitLoweredWinAlloca(MachineInstr *MI,
1085 MachineBasicBlock *BB) const;
1087 MachineBasicBlock *EmitLoweredSegAlloca(MachineInstr *MI,
1088 MachineBasicBlock *BB) const;
1090 MachineBasicBlock *EmitLoweredTLSCall(MachineInstr *MI,
1091 MachineBasicBlock *BB) const;
1093 MachineBasicBlock *emitLoweredTLSAddr(MachineInstr *MI,
1094 MachineBasicBlock *BB) const;
1096 MachineBasicBlock *emitEHSjLjSetJmp(MachineInstr *MI,
1097 MachineBasicBlock *MBB) const;
1099 MachineBasicBlock *emitEHSjLjLongJmp(MachineInstr *MI,
1100 MachineBasicBlock *MBB) const;
1102 MachineBasicBlock *emitFMA3Instr(MachineInstr *MI,
1103 MachineBasicBlock *MBB) const;
1105 /// Emit nodes that will be selected as "test Op0,Op0", or something
1106 /// equivalent, for use with the given x86 condition code.
1107 SDValue EmitTest(SDValue Op0, unsigned X86CC, SDLoc dl,
1108 SelectionDAG &DAG) const;
1110 /// Emit nodes that will be selected as "cmp Op0,Op1", or something
1111 /// equivalent, for use with the given x86 condition code.
1112 SDValue EmitCmp(SDValue Op0, SDValue Op1, unsigned X86CC, SDLoc dl,
1113 SelectionDAG &DAG) const;
1115 /// Convert a comparison if required by the subtarget.
1116 SDValue ConvertCmpIfNecessary(SDValue Cmp, SelectionDAG &DAG) const;
1118 /// Use rsqrt* to speed up sqrt calculations.
1119 SDValue getRsqrtEstimate(SDValue Operand, DAGCombinerInfo &DCI,
1120 unsigned &RefinementSteps,
1121 bool &UseOneConstNR) const override;
1123 /// Use rcp* to speed up fdiv calculations.
1124 SDValue getRecipEstimate(SDValue Operand, DAGCombinerInfo &DCI,
1125 unsigned &RefinementSteps) const override;
1127 /// Reassociate floating point divisions into multiply by reciprocal.
1128 unsigned combineRepeatedFPDivisors() const override;
1132 FastISel *createFastISel(FunctionLoweringInfo &funcInfo,
1133 const TargetLibraryInfo *libInfo);
1137 #endif // X86ISELLOWERING_H