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 /// Bitwise logical right shift of floating point values. This
60 /// corresponds to X86::PSRLDQ.
63 /// These operations represent an abstract X86 call
64 /// instruction, which includes a bunch of information. In particular the
65 /// operands of these node are:
67 /// #0 - The incoming token chain
69 /// #2 - The number of arg bytes the caller pushes on the stack.
70 /// #3 - The number of arg bytes the callee pops off the stack.
71 /// #4 - The value to pass in AL/AX/EAX (optional)
72 /// #5 - The value to pass in DL/DX/EDX (optional)
74 /// The result values of these nodes are:
76 /// #0 - The outgoing token chain
77 /// #1 - The first register result value (optional)
78 /// #2 - The second register result value (optional)
82 /// This operation implements the lowering for readcyclecounter
85 /// X86 Read Time-Stamp Counter and Processor ID.
88 /// X86 Read Performance Monitoring Counters.
91 /// X86 compare and logical compare instructions.
94 /// X86 bit-test instructions.
97 /// X86 SetCC. Operand 0 is condition code, and operand 1 is the EFLAGS
98 /// operand, usually produced by a CMP instruction.
104 // Same as SETCC except it's materialized with a sbb and the value is all
105 // one's or all zero's.
106 SETCC_CARRY, // R = carry_bit ? ~0 : 0
108 /// X86 FP SETCC, implemented with CMP{cc}SS/CMP{cc}SD.
109 /// Operands are two FP values to compare; result is a mask of
110 /// 0s or 1s. Generally DTRT for C/C++ with NaNs.
113 /// X86 MOVMSK{pd|ps}, extracts sign bits of two or four FP values,
114 /// result in an integer GPR. Needs masking for scalar result.
117 /// X86 conditional moves. Operand 0 and operand 1 are the two values
118 /// to select from. Operand 2 is the condition code, and operand 3 is the
119 /// flag operand produced by a CMP or TEST instruction. It also writes a
123 /// X86 conditional branches. Operand 0 is the chain operand, operand 1
124 /// is the block to branch if condition is true, operand 2 is the
125 /// condition code, and operand 3 is the flag operand produced by a CMP
126 /// or TEST instruction.
129 /// Return with a flag operand. Operand 0 is the chain operand, operand
130 /// 1 is the number of bytes of stack to pop.
133 /// Repeat fill, corresponds to X86::REP_STOSx.
136 /// Repeat move, corresponds to X86::REP_MOVSx.
139 /// On Darwin, this node represents the result of the popl
140 /// at function entry, used for PIC code.
143 /// A wrapper node for TargetConstantPool,
144 /// TargetExternalSymbol, and TargetGlobalAddress.
147 /// Special wrapper used under X86-64 PIC mode for RIP
148 /// relative displacements.
151 /// Copies a 64-bit value from the low word of an XMM vector
152 /// to an MMX vector. If you think this is too close to the previous
153 /// mnemonic, so do I; blame Intel.
156 /// Copies a 32-bit value from the low word of a MMX
160 /// Copies a GPR into the low 32-bit word of a MMX vector
161 /// and zero out the high word.
164 /// Extract an 8-bit value from a vector and zero extend it to
165 /// i32, corresponds to X86::PEXTRB.
168 /// Extract a 16-bit value from a vector and zero extend it to
169 /// i32, corresponds to X86::PEXTRW.
172 /// Insert any element of a 4 x float vector into any element
173 /// of a destination 4 x floatvector.
176 /// Insert the lower 8-bits of a 32-bit value to a vector,
177 /// corresponds to X86::PINSRB.
180 /// Insert the lower 16-bits of a 32-bit value to a vector,
181 /// corresponds to X86::PINSRW.
184 /// Shuffle 16 8-bit values within a vector.
187 /// Compute Sum of Absolute Differences.
190 /// Bitwise Logical AND NOT of Packed FP values.
193 /// Copy integer sign.
196 /// Blend where the selector is an immediate.
199 /// Blend where the condition has been shrunk.
200 /// This is used to emphasize that the condition mask is
201 /// no more valid for generic VSELECT optimizations.
204 /// Combined add and sub on an FP vector.
206 // FP vector ops with rounding mode.
214 // Integer add/sub with unsigned saturation.
217 // Integer add/sub with signed saturation.
221 /// Integer horizontal add.
224 /// Integer horizontal sub.
227 /// Floating point horizontal add.
230 /// Floating point horizontal sub.
233 /// Unsigned integer max and min.
236 /// Signed integer max and min.
239 /// Floating point max and min.
242 /// Commutative FMIN and FMAX.
245 /// Floating point reciprocal-sqrt and reciprocal approximation.
246 /// Note that these typically require refinement
247 /// in order to obtain suitable precision.
250 // Thread Local Storage.
253 // Thread Local Storage. A call to get the start address
254 // of the TLS block for the current module.
257 // Thread Local Storage. When calling to an OS provided
258 // thunk at the address from an earlier relocation.
261 // Exception Handling helpers.
264 // SjLj exception handling setjmp.
267 // SjLj exception handling longjmp.
270 /// Tail call return. See X86TargetLowering::LowerCall for
271 /// the list of operands.
274 // Vector move to low scalar and zero higher vector elements.
277 // Vector integer zero-extend.
280 // Vector integer signed-extend.
283 // Vector integer truncate.
286 // Vector integer truncate with mask.
295 // 128-bit vector logical left / right shift
298 // Vector shift elements
301 // Vector shift elements by immediate
304 // Vector packed double/float comparison.
307 // Vector integer comparisons.
309 // Vector integer comparisons, the result is in a mask vector.
312 /// Vector comparison generating mask bits for fp and
313 /// integer signed and unsigned data types.
316 // Vector comparison with rounding mode for FP values
319 // Arithmetic operations with FLAGS results.
320 ADD, SUB, ADC, SBB, SMUL,
321 INC, DEC, OR, XOR, AND,
323 BEXTR, // Bit field extract
325 UMUL, // LOW, HI, FLAGS = umul LHS, RHS
327 // 8-bit SMUL/UMUL - AX, FLAGS = smul8/umul8 AL, RHS
330 // 8-bit divrem that zero-extend the high result (AH).
334 // X86-specific multiply by immediate.
337 // Vector bitwise comparisons.
340 // Vector packed fp sign bitwise comparisons.
343 // Vector "test" in AVX-512, the result is in a mask vector.
347 // OR/AND test for masks
350 // Several flavors of instructions with vector shuffle behaviors.
355 // AVX512 inter-lane alignr
380 //Fix Up Special Packed Float32/64 values
382 //Range Restriction Calculation For Packed Pairs of Float32/64 values
384 // Broadcast scalar to vector
386 // Broadcast subvector to vector
388 // Insert/Extract vector element
392 // Vector multiply packed unsigned doubleword integers
394 // Vector multiply packed signed doubleword integers
404 // FMA with rounding mode
413 // Compress and expand
417 // Save xmm argument registers to the stack, according to %al. An operator
418 // is needed so that this can be expanded with control flow.
419 VASTART_SAVE_XMM_REGS,
421 // Windows's _chkstk call to do stack probing.
424 // For allocating variable amounts of stack space when using
425 // segmented stacks. Check if the current stacklet has enough space, and
426 // falls back to heap allocation if not.
429 // Windows's _ftol2 runtime routine to do fptoui.
438 // Store FP status word into i16 register.
441 // Store contents of %ah into %eflags.
444 // Get a random integer and indicate whether it is valid in CF.
447 // Get a NIST SP800-90B & C compliant random integer and
448 // indicate whether it is valid in CF.
454 // Test if in transactional execution.
458 RSQRT28, RCP28, EXP2,
461 LCMPXCHG_DAG = ISD::FIRST_TARGET_MEMORY_OPCODE,
465 // Load, scalar_to_vector, and zero extend.
468 // Store FP control world into i16 memory.
471 /// This instruction implements FP_TO_SINT with the
472 /// integer destination in memory and a FP reg source. This corresponds
473 /// to the X86::FIST*m instructions and the rounding mode change stuff. It
474 /// has two inputs (token chain and address) and two outputs (int value
475 /// and token chain).
480 /// This instruction implements SINT_TO_FP with the
481 /// integer source in memory and FP reg result. This corresponds to the
482 /// X86::FILD*m instructions. It has three inputs (token chain, address,
483 /// and source type) and two outputs (FP value and token chain). FILD_FLAG
484 /// also produces a flag).
488 /// This instruction implements an extending load to FP stack slots.
489 /// This corresponds to the X86::FLD32m / X86::FLD64m. It takes a chain
490 /// operand, ptr to load from, and a ValueType node indicating the type
494 /// This instruction implements a truncating store to FP stack
495 /// slots. This corresponds to the X86::FST32m / X86::FST64m. It takes a
496 /// chain operand, value to store, address, and a ValueType to store it
500 /// This instruction grabs the address of the next argument
501 /// from a va_list. (reads and modifies the va_list in memory)
504 // WARNING: Do not add anything in the end unless you want the node to
505 // have memop! In fact, starting from ATOMADD64_DAG all opcodes will be
506 // thought as target memory ops!
510 /// Define some predicates that are used for node matching.
512 /// Return true if the specified
513 /// EXTRACT_SUBVECTOR operand specifies a vector extract that is
514 /// suitable for input to VEXTRACTF128, VEXTRACTI128 instructions.
515 bool isVEXTRACT128Index(SDNode *N);
517 /// Return true if the specified
518 /// INSERT_SUBVECTOR operand specifies a subvector insert that is
519 /// suitable for input to VINSERTF128, VINSERTI128 instructions.
520 bool isVINSERT128Index(SDNode *N);
522 /// Return true if the specified
523 /// EXTRACT_SUBVECTOR operand specifies a vector extract that is
524 /// suitable for input to VEXTRACTF64X4, VEXTRACTI64X4 instructions.
525 bool isVEXTRACT256Index(SDNode *N);
527 /// Return true if the specified
528 /// INSERT_SUBVECTOR operand specifies a subvector insert that is
529 /// suitable for input to VINSERTF64X4, VINSERTI64X4 instructions.
530 bool isVINSERT256Index(SDNode *N);
532 /// Return the appropriate
533 /// immediate to extract the specified EXTRACT_SUBVECTOR index
534 /// with VEXTRACTF128, VEXTRACTI128 instructions.
535 unsigned getExtractVEXTRACT128Immediate(SDNode *N);
537 /// Return the appropriate
538 /// immediate to insert at the specified INSERT_SUBVECTOR index
539 /// with VINSERTF128, VINSERT128 instructions.
540 unsigned getInsertVINSERT128Immediate(SDNode *N);
542 /// Return the appropriate
543 /// immediate to extract the specified EXTRACT_SUBVECTOR index
544 /// with VEXTRACTF64X4, VEXTRACTI64x4 instructions.
545 unsigned getExtractVEXTRACT256Immediate(SDNode *N);
547 /// Return the appropriate
548 /// immediate to insert at the specified INSERT_SUBVECTOR index
549 /// with VINSERTF64x4, VINSERTI64x4 instructions.
550 unsigned getInsertVINSERT256Immediate(SDNode *N);
552 /// Returns true if Elt is a constant zero or floating point constant +0.0.
553 bool isZeroNode(SDValue Elt);
555 /// Returns true of the given offset can be
556 /// fit into displacement field of the instruction.
557 bool isOffsetSuitableForCodeModel(int64_t Offset, CodeModel::Model M,
558 bool hasSymbolicDisplacement = true);
561 /// Determines whether the callee is required to pop its
562 /// own arguments. Callee pop is necessary to support tail calls.
563 bool isCalleePop(CallingConv::ID CallingConv,
564 bool is64Bit, bool IsVarArg, bool TailCallOpt);
566 /// AVX512 static rounding constants. These need to match the values in
568 enum STATIC_ROUNDING {
577 //===--------------------------------------------------------------------===//
578 // X86 Implementation of the TargetLowering interface
579 class X86TargetLowering final : public TargetLowering {
581 explicit X86TargetLowering(const X86TargetMachine &TM,
582 const X86Subtarget &STI);
584 unsigned getJumpTableEncoding() const override;
585 bool useSoftFloat() const override;
587 MVT getScalarShiftAmountTy(EVT LHSTy) const override { return MVT::i8; }
590 LowerCustomJumpTableEntry(const MachineJumpTableInfo *MJTI,
591 const MachineBasicBlock *MBB, unsigned uid,
592 MCContext &Ctx) const override;
594 /// Returns relocation base for the given PIC jumptable.
595 SDValue getPICJumpTableRelocBase(SDValue Table,
596 SelectionDAG &DAG) const override;
598 getPICJumpTableRelocBaseExpr(const MachineFunction *MF,
599 unsigned JTI, MCContext &Ctx) const override;
601 /// Return the desired alignment for ByVal aggregate
602 /// function arguments in the caller parameter area. For X86, aggregates
603 /// that contains are placed at 16-byte boundaries while the rest are at
604 /// 4-byte boundaries.
605 unsigned getByValTypeAlignment(Type *Ty) const override;
607 /// Returns the target specific optimal type for load
608 /// and store operations as a result of memset, memcpy, and memmove
609 /// lowering. If DstAlign is zero that means it's safe to destination
610 /// alignment can satisfy any constraint. Similarly if SrcAlign is zero it
611 /// means there isn't a need to check it against alignment requirement,
612 /// probably because the source does not need to be loaded. If 'IsMemset' is
613 /// true, that means it's expanding a memset. If 'ZeroMemset' is true, that
614 /// means it's a memset of zero. 'MemcpyStrSrc' indicates whether the memcpy
615 /// source is constant so it does not need to be loaded.
616 /// It returns EVT::Other if the type should be determined using generic
617 /// target-independent logic.
618 EVT getOptimalMemOpType(uint64_t Size, unsigned DstAlign, unsigned SrcAlign,
619 bool IsMemset, bool ZeroMemset, bool MemcpyStrSrc,
620 MachineFunction &MF) const override;
622 /// Returns true if it's safe to use load / store of the
623 /// specified type to expand memcpy / memset inline. This is mostly true
624 /// for all types except for some special cases. For example, on X86
625 /// targets without SSE2 f64 load / store are done with fldl / fstpl which
626 /// also does type conversion. Note the specified type doesn't have to be
627 /// legal as the hook is used before type legalization.
628 bool isSafeMemOpType(MVT VT) const override;
630 /// Returns true if the target allows
631 /// unaligned memory accesses. of the specified type. Returns whether it
632 /// is "fast" by reference in the second argument.
633 bool allowsMisalignedMemoryAccesses(EVT VT, unsigned AS, unsigned Align,
634 bool *Fast) const override;
636 /// Provide custom lowering hooks for some operations.
638 SDValue LowerOperation(SDValue Op, SelectionDAG &DAG) const override;
640 /// Replace the results of node with an illegal result
641 /// type with new values built out of custom code.
643 void ReplaceNodeResults(SDNode *N, SmallVectorImpl<SDValue>&Results,
644 SelectionDAG &DAG) const override;
647 SDValue PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const override;
649 /// Return true if the target has native support for
650 /// the specified value type and it is 'desirable' to use the type for the
651 /// given node type. e.g. On x86 i16 is legal, but undesirable since i16
652 /// instruction encodings are longer and some i16 instructions are slow.
653 bool isTypeDesirableForOp(unsigned Opc, EVT VT) const override;
655 /// Return true if the target has native support for the
656 /// specified value type and it is 'desirable' to use the type. e.g. On x86
657 /// i16 is legal, but undesirable since i16 instruction encodings are longer
658 /// and some i16 instructions are slow.
659 bool IsDesirableToPromoteOp(SDValue Op, EVT &PVT) const override;
662 EmitInstrWithCustomInserter(MachineInstr *MI,
663 MachineBasicBlock *MBB) const override;
666 /// This method returns the name of a target specific DAG node.
667 const char *getTargetNodeName(unsigned Opcode) const override;
669 bool isCheapToSpeculateCttz() const override;
671 bool isCheapToSpeculateCtlz() const override;
673 /// Return the value type to use for ISD::SETCC.
674 EVT getSetCCResultType(LLVMContext &Context, EVT VT) const override;
676 /// Determine which of the bits specified in Mask are known to be either
677 /// zero or one and return them in the KnownZero/KnownOne bitsets.
678 void computeKnownBitsForTargetNode(const SDValue Op,
681 const SelectionDAG &DAG,
682 unsigned Depth = 0) const override;
684 /// Determine the number of bits in the operation that are sign bits.
685 unsigned ComputeNumSignBitsForTargetNode(SDValue Op,
686 const SelectionDAG &DAG,
687 unsigned Depth) const override;
689 bool isGAPlusOffset(SDNode *N, const GlobalValue* &GA,
690 int64_t &Offset) const override;
692 SDValue getReturnAddressFrameIndex(SelectionDAG &DAG) const;
694 bool ExpandInlineAsm(CallInst *CI) const override;
697 getConstraintType(const std::string &Constraint) const override;
699 /// Examine constraint string and operand type and determine a weight value.
700 /// The operand object must already have been set up with the operand type.
702 getSingleConstraintMatchWeight(AsmOperandInfo &info,
703 const char *constraint) const override;
705 const char *LowerXConstraint(EVT ConstraintVT) const override;
707 /// Lower the specified operand into the Ops vector. If it is invalid, don't
708 /// add anything to Ops. If hasMemory is true it means one of the asm
709 /// constraint of the inline asm instruction being processed is 'm'.
710 void LowerAsmOperandForConstraint(SDValue Op,
711 std::string &Constraint,
712 std::vector<SDValue> &Ops,
713 SelectionDAG &DAG) const override;
715 unsigned getInlineAsmMemConstraint(
716 const std::string &ConstraintCode) const override {
717 if (ConstraintCode == "i")
718 return InlineAsm::Constraint_i;
719 else if (ConstraintCode == "o")
720 return InlineAsm::Constraint_o;
721 else if (ConstraintCode == "v")
722 return InlineAsm::Constraint_v;
723 else if (ConstraintCode == "X")
724 return InlineAsm::Constraint_X;
725 return TargetLowering::getInlineAsmMemConstraint(ConstraintCode);
728 /// Given a physical register constraint
729 /// (e.g. {edx}), return the register number and the register class for the
730 /// register. This should only be used for C_Register constraints. On
731 /// error, this returns a register number of 0.
732 std::pair<unsigned, const TargetRegisterClass *>
733 getRegForInlineAsmConstraint(const TargetRegisterInfo *TRI,
734 const std::string &Constraint,
735 MVT VT) const override;
737 /// Return true if the addressing mode represented
738 /// by AM is legal for this target, for a load/store of the specified type.
739 bool isLegalAddressingMode(const AddrMode &AM, Type *Ty,
740 unsigned AS) const override;
742 /// Return true if the specified immediate is legal
743 /// icmp immediate, that is the target has icmp instructions which can
744 /// compare a register against the immediate without having to materialize
745 /// the immediate into a register.
746 bool isLegalICmpImmediate(int64_t Imm) const override;
748 /// Return true if the specified immediate is legal
749 /// add immediate, that is the target has add instructions which can
750 /// add a register and the immediate without having to materialize
751 /// the immediate into a register.
752 bool isLegalAddImmediate(int64_t Imm) const override;
754 /// \brief Return the cost of the scaling factor used in the addressing
755 /// mode represented by AM for this target, for a load/store
756 /// of the specified type.
757 /// If the AM is supported, the return value must be >= 0.
758 /// If the AM is not supported, it returns a negative value.
759 int getScalingFactorCost(const AddrMode &AM, Type *Ty,
760 unsigned AS) const override;
762 bool isVectorShiftByScalarCheap(Type *Ty) const override;
764 /// Return true if it's free to truncate a value of
765 /// type Ty1 to type Ty2. e.g. On x86 it's free to truncate a i32 value in
766 /// register EAX to i16 by referencing its sub-register AX.
767 bool isTruncateFree(Type *Ty1, Type *Ty2) const override;
768 bool isTruncateFree(EVT VT1, EVT VT2) const override;
770 bool allowTruncateForTailCall(Type *Ty1, Type *Ty2) const override;
772 /// Return true if any actual instruction that defines a
773 /// value of type Ty1 implicit zero-extends the value to Ty2 in the result
774 /// register. This does not necessarily include registers defined in
775 /// unknown ways, such as incoming arguments, or copies from unknown
776 /// virtual registers. Also, if isTruncateFree(Ty2, Ty1) is true, this
777 /// does not necessarily apply to truncate instructions. e.g. on x86-64,
778 /// all instructions that define 32-bit values implicit zero-extend the
779 /// result out to 64 bits.
780 bool isZExtFree(Type *Ty1, Type *Ty2) const override;
781 bool isZExtFree(EVT VT1, EVT VT2) const override;
782 bool isZExtFree(SDValue Val, EVT VT2) const override;
784 /// Return true if folding a vector load into ExtVal (a sign, zero, or any
785 /// extend node) is profitable.
786 bool isVectorLoadExtDesirable(SDValue) const override;
788 /// Return true if an FMA operation is faster than a pair of fmul and fadd
789 /// instructions. fmuladd intrinsics will be expanded to FMAs when this
790 /// method returns true, otherwise fmuladd is expanded to fmul + fadd.
791 bool isFMAFasterThanFMulAndFAdd(EVT VT) const override;
793 /// Return true if it's profitable to narrow
794 /// operations of type VT1 to VT2. e.g. on x86, it's profitable to narrow
795 /// from i32 to i8 but not from i32 to i16.
796 bool isNarrowingProfitable(EVT VT1, EVT VT2) const override;
798 /// Returns true if the target can instruction select the
799 /// specified FP immediate natively. If false, the legalizer will
800 /// materialize the FP immediate as a load from a constant pool.
801 bool isFPImmLegal(const APFloat &Imm, EVT VT) const override;
803 /// Targets can use this to indicate that they only support *some*
804 /// VECTOR_SHUFFLE operations, those with specific masks. By default, if a
805 /// target supports the VECTOR_SHUFFLE node, all mask values are assumed to
807 bool isShuffleMaskLegal(const SmallVectorImpl<int> &Mask,
808 EVT VT) const override;
810 /// Similar to isShuffleMaskLegal. This is used by Targets can use this to
811 /// indicate if there is a suitable VECTOR_SHUFFLE that can be used to
812 /// replace a VAND with a constant pool entry.
813 bool isVectorClearMaskLegal(const SmallVectorImpl<int> &Mask,
814 EVT VT) const override;
816 /// If true, then instruction selection should
817 /// seek to shrink the FP constant of the specified type to a smaller type
818 /// in order to save space and / or reduce runtime.
819 bool ShouldShrinkFPConstant(EVT VT) const override {
820 // Don't shrink FP constpool if SSE2 is available since cvtss2sd is more
821 // expensive than a straight movsd. On the other hand, it's important to
822 // shrink long double fp constant since fldt is very slow.
823 return !X86ScalarSSEf64 || VT == MVT::f80;
826 /// Return true if we believe it is correct and profitable to reduce the
827 /// load node to a smaller type.
828 bool shouldReduceLoadWidth(SDNode *Load, ISD::LoadExtType ExtTy,
829 EVT NewVT) const override;
831 /// Return true if the specified scalar FP type is computed in an SSE
832 /// register, not on the X87 floating point stack.
833 bool isScalarFPTypeInSSEReg(EVT VT) const {
834 return (VT == MVT::f64 && X86ScalarSSEf64) || // f64 is when SSE2
835 (VT == MVT::f32 && X86ScalarSSEf32); // f32 is when SSE1
838 /// Return true if the target uses the MSVC _ftol2 routine for fptoui.
839 bool isTargetFTOL() const;
841 /// Return true if the MSVC _ftol2 routine should be used for fptoui to the
843 bool isIntegerTypeFTOL(EVT VT) const {
844 return isTargetFTOL() && VT == MVT::i64;
847 /// \brief Returns true if it is beneficial to convert a load of a constant
848 /// to just the constant itself.
849 bool shouldConvertConstantLoadToIntImm(const APInt &Imm,
850 Type *Ty) const override;
852 /// Return true if EXTRACT_SUBVECTOR is cheap for this result type
854 bool isExtractSubvectorCheap(EVT ResVT, unsigned Index) const override;
856 /// Intel processors have a unified instruction and data cache
857 const char * getClearCacheBuiltinName() const override {
858 return nullptr; // nothing to do, move along.
861 unsigned getRegisterByName(const char* RegName, EVT VT) const override;
863 /// This method returns a target specific FastISel object,
864 /// or null if the target does not support "fast" ISel.
865 FastISel *createFastISel(FunctionLoweringInfo &funcInfo,
866 const TargetLibraryInfo *libInfo) const override;
868 /// Return true if the target stores stack protector cookies at a fixed
869 /// offset in some non-standard address space, and populates the address
870 /// space and offset as appropriate.
871 bool getStackCookieLocation(unsigned &AddressSpace,
872 unsigned &Offset) const override;
874 SDValue BuildFILD(SDValue Op, EVT SrcVT, SDValue Chain, SDValue StackSlot,
875 SelectionDAG &DAG) const;
877 bool isNoopAddrSpaceCast(unsigned SrcAS, unsigned DestAS) const override;
879 bool useLoadStackGuardNode() const override;
880 /// \brief Customize the preferred legalization strategy for certain types.
881 LegalizeTypeAction getPreferredVectorAction(EVT VT) const override;
884 std::pair<const TargetRegisterClass *, uint8_t>
885 findRepresentativeClass(const TargetRegisterInfo *TRI,
886 MVT VT) const override;
889 /// Keep a pointer to the X86Subtarget around so that we can
890 /// make the right decision when generating code for different targets.
891 const X86Subtarget *Subtarget;
892 const DataLayout *TD;
894 /// Select between SSE or x87 floating point ops.
895 /// When SSE is available, use it for f32 operations.
896 /// When SSE2 is available, use it for f64 operations.
897 bool X86ScalarSSEf32;
898 bool X86ScalarSSEf64;
900 /// A list of legal FP immediates.
901 std::vector<APFloat> LegalFPImmediates;
903 /// Indicate that this x86 target can instruction
904 /// select the specified FP immediate natively.
905 void addLegalFPImmediate(const APFloat& Imm) {
906 LegalFPImmediates.push_back(Imm);
909 SDValue LowerCallResult(SDValue Chain, SDValue InFlag,
910 CallingConv::ID CallConv, bool isVarArg,
911 const SmallVectorImpl<ISD::InputArg> &Ins,
912 SDLoc dl, SelectionDAG &DAG,
913 SmallVectorImpl<SDValue> &InVals) const;
914 SDValue LowerMemArgument(SDValue Chain,
915 CallingConv::ID CallConv,
916 const SmallVectorImpl<ISD::InputArg> &ArgInfo,
917 SDLoc dl, SelectionDAG &DAG,
918 const CCValAssign &VA, MachineFrameInfo *MFI,
920 SDValue LowerMemOpCallTo(SDValue Chain, SDValue StackPtr, SDValue Arg,
921 SDLoc dl, SelectionDAG &DAG,
922 const CCValAssign &VA,
923 ISD::ArgFlagsTy Flags) const;
925 // Call lowering helpers.
927 /// Check whether the call is eligible for tail call optimization. Targets
928 /// that want to do tail call optimization should implement this function.
929 bool IsEligibleForTailCallOptimization(SDValue Callee,
930 CallingConv::ID CalleeCC,
932 bool isCalleeStructRet,
933 bool isCallerStructRet,
935 const SmallVectorImpl<ISD::OutputArg> &Outs,
936 const SmallVectorImpl<SDValue> &OutVals,
937 const SmallVectorImpl<ISD::InputArg> &Ins,
938 SelectionDAG& DAG) const;
939 bool IsCalleePop(bool isVarArg, CallingConv::ID CallConv) const;
940 SDValue EmitTailCallLoadRetAddr(SelectionDAG &DAG, SDValue &OutRetAddr,
941 SDValue Chain, bool IsTailCall, bool Is64Bit,
942 int FPDiff, SDLoc dl) const;
944 unsigned GetAlignedArgumentStackSize(unsigned StackSize,
945 SelectionDAG &DAG) const;
947 std::pair<SDValue,SDValue> FP_TO_INTHelper(SDValue Op, SelectionDAG &DAG,
949 bool isReplace) const;
951 SDValue LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) const;
952 SDValue LowerBUILD_VECTORvXi1(SDValue Op, SelectionDAG &DAG) const;
953 SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) const;
954 SDValue LowerVSELECT(SDValue Op, SelectionDAG &DAG) const;
955 SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const;
956 SDValue ExtractBitFromMaskVector(SDValue Op, SelectionDAG &DAG) const;
957 SDValue InsertBitToMaskVector(SDValue Op, SelectionDAG &DAG) const;
959 SDValue LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const;
960 SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) const;
961 SDValue LowerBlockAddress(SDValue Op, SelectionDAG &DAG) const;
962 SDValue LowerGlobalAddress(const GlobalValue *GV, SDLoc dl,
963 int64_t Offset, SelectionDAG &DAG) const;
964 SDValue LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) const;
965 SDValue LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const;
966 SDValue LowerExternalSymbol(SDValue Op, SelectionDAG &DAG) const;
967 SDValue LowerSINT_TO_FP(SDValue Op, SelectionDAG &DAG) const;
968 SDValue LowerUINT_TO_FP(SDValue Op, SelectionDAG &DAG) const;
969 SDValue LowerUINT_TO_FP_i64(SDValue Op, SelectionDAG &DAG) const;
970 SDValue LowerUINT_TO_FP_i32(SDValue Op, SelectionDAG &DAG) const;
971 SDValue lowerUINT_TO_FP_vec(SDValue Op, SelectionDAG &DAG) const;
972 SDValue LowerTRUNCATE(SDValue Op, SelectionDAG &DAG) const;
973 SDValue LowerFP_TO_SINT(SDValue Op, SelectionDAG &DAG) const;
974 SDValue LowerFP_TO_UINT(SDValue Op, SelectionDAG &DAG) const;
975 SDValue LowerToBT(SDValue And, ISD::CondCode CC,
976 SDLoc dl, SelectionDAG &DAG) const;
977 SDValue LowerSETCC(SDValue Op, SelectionDAG &DAG) const;
978 SDValue LowerSELECT(SDValue Op, SelectionDAG &DAG) const;
979 SDValue LowerBRCOND(SDValue Op, SelectionDAG &DAG) const;
980 SDValue LowerMEMSET(SDValue Op, SelectionDAG &DAG) const;
981 SDValue LowerJumpTable(SDValue Op, SelectionDAG &DAG) const;
982 SDValue LowerDYNAMIC_STACKALLOC(SDValue Op, SelectionDAG &DAG) const;
983 SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG) const;
984 SDValue LowerVAARG(SDValue Op, SelectionDAG &DAG) const;
985 SDValue LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const;
986 SDValue LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const;
987 SDValue LowerFRAME_TO_ARGS_OFFSET(SDValue Op, SelectionDAG &DAG) const;
988 SDValue LowerEH_RETURN(SDValue Op, SelectionDAG &DAG) const;
989 SDValue lowerEH_SJLJ_SETJMP(SDValue Op, SelectionDAG &DAG) const;
990 SDValue lowerEH_SJLJ_LONGJMP(SDValue Op, SelectionDAG &DAG) const;
991 SDValue LowerINIT_TRAMPOLINE(SDValue Op, SelectionDAG &DAG) const;
992 SDValue LowerFLT_ROUNDS_(SDValue Op, SelectionDAG &DAG) const;
993 SDValue LowerWin64_i128OP(SDValue Op, SelectionDAG &DAG) const;
994 SDValue LowerGC_TRANSITION_START(SDValue Op, SelectionDAG &DAG) const;
995 SDValue LowerGC_TRANSITION_END(SDValue Op, SelectionDAG &DAG) const;
998 LowerFormalArguments(SDValue Chain,
999 CallingConv::ID CallConv, bool isVarArg,
1000 const SmallVectorImpl<ISD::InputArg> &Ins,
1001 SDLoc dl, SelectionDAG &DAG,
1002 SmallVectorImpl<SDValue> &InVals) const override;
1003 SDValue LowerCall(CallLoweringInfo &CLI,
1004 SmallVectorImpl<SDValue> &InVals) const override;
1006 SDValue LowerReturn(SDValue Chain,
1007 CallingConv::ID CallConv, bool isVarArg,
1008 const SmallVectorImpl<ISD::OutputArg> &Outs,
1009 const SmallVectorImpl<SDValue> &OutVals,
1010 SDLoc dl, SelectionDAG &DAG) const override;
1012 bool isUsedByReturnOnly(SDNode *N, SDValue &Chain) const override;
1014 bool mayBeEmittedAsTailCall(CallInst *CI) const override;
1016 EVT getTypeForExtArgOrReturn(LLVMContext &Context, EVT VT,
1017 ISD::NodeType ExtendKind) const override;
1019 bool CanLowerReturn(CallingConv::ID CallConv, MachineFunction &MF,
1021 const SmallVectorImpl<ISD::OutputArg> &Outs,
1022 LLVMContext &Context) const override;
1024 const MCPhysReg *getScratchRegisters(CallingConv::ID CC) const override;
1026 bool shouldExpandAtomicLoadInIR(LoadInst *SI) const override;
1027 bool shouldExpandAtomicStoreInIR(StoreInst *SI) const override;
1028 TargetLoweringBase::AtomicRMWExpansionKind
1029 shouldExpandAtomicRMWInIR(AtomicRMWInst *AI) const override;
1032 lowerIdempotentRMWIntoFencedLoad(AtomicRMWInst *AI) const override;
1034 bool needsCmpXchgNb(const Type *MemType) const;
1036 /// Utility function to emit atomic-load-arith operations (and, or, xor,
1037 /// nand, max, min, umax, umin). It takes the corresponding instruction to
1038 /// expand, the associated machine basic block, and the associated X86
1039 /// opcodes for reg/reg.
1040 MachineBasicBlock *EmitAtomicLoadArith(MachineInstr *MI,
1041 MachineBasicBlock *MBB) const;
1043 /// Utility function to emit atomic-load-arith operations (and, or, xor,
1044 /// nand, add, sub, swap) for 64-bit operands on 32-bit target.
1045 MachineBasicBlock *EmitAtomicLoadArith6432(MachineInstr *MI,
1046 MachineBasicBlock *MBB) const;
1048 // Utility function to emit the low-level va_arg code for X86-64.
1049 MachineBasicBlock *EmitVAARG64WithCustomInserter(
1051 MachineBasicBlock *MBB) const;
1053 /// Utility function to emit the xmm reg save portion of va_start.
1054 MachineBasicBlock *EmitVAStartSaveXMMRegsWithCustomInserter(
1055 MachineInstr *BInstr,
1056 MachineBasicBlock *BB) const;
1058 MachineBasicBlock *EmitLoweredSelect(MachineInstr *I,
1059 MachineBasicBlock *BB) const;
1061 MachineBasicBlock *EmitLoweredWinAlloca(MachineInstr *MI,
1062 MachineBasicBlock *BB) const;
1064 MachineBasicBlock *EmitLoweredSegAlloca(MachineInstr *MI,
1065 MachineBasicBlock *BB) const;
1067 MachineBasicBlock *EmitLoweredTLSCall(MachineInstr *MI,
1068 MachineBasicBlock *BB) const;
1070 MachineBasicBlock *emitLoweredTLSAddr(MachineInstr *MI,
1071 MachineBasicBlock *BB) const;
1073 MachineBasicBlock *emitEHSjLjSetJmp(MachineInstr *MI,
1074 MachineBasicBlock *MBB) const;
1076 MachineBasicBlock *emitEHSjLjLongJmp(MachineInstr *MI,
1077 MachineBasicBlock *MBB) const;
1079 MachineBasicBlock *emitFMA3Instr(MachineInstr *MI,
1080 MachineBasicBlock *MBB) const;
1082 /// Emit nodes that will be selected as "test Op0,Op0", or something
1083 /// equivalent, for use with the given x86 condition code.
1084 SDValue EmitTest(SDValue Op0, unsigned X86CC, SDLoc dl,
1085 SelectionDAG &DAG) const;
1087 /// Emit nodes that will be selected as "cmp Op0,Op1", or something
1088 /// equivalent, for use with the given x86 condition code.
1089 SDValue EmitCmp(SDValue Op0, SDValue Op1, unsigned X86CC, SDLoc dl,
1090 SelectionDAG &DAG) const;
1092 /// Convert a comparison if required by the subtarget.
1093 SDValue ConvertCmpIfNecessary(SDValue Cmp, SelectionDAG &DAG) const;
1095 /// Use rsqrt* to speed up sqrt calculations.
1096 SDValue getRsqrtEstimate(SDValue Operand, DAGCombinerInfo &DCI,
1097 unsigned &RefinementSteps,
1098 bool &UseOneConstNR) const override;
1100 /// Use rcp* to speed up fdiv calculations.
1101 SDValue getRecipEstimate(SDValue Operand, DAGCombinerInfo &DCI,
1102 unsigned &RefinementSteps) const override;
1104 /// Reassociate floating point divisions into multiply by reciprocal.
1105 bool combineRepeatedFPDivisors(unsigned NumUsers) const override;
1109 FastISel *createFastISel(FunctionLoweringInfo &funcInfo,
1110 const TargetLibraryInfo *libInfo);
1114 #endif // X86ISELLOWERING_H