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.
186 /// Bitwise Logical AND NOT of Packed FP values.
189 /// Copy integer sign.
192 /// Blend where the selector is an immediate.
195 /// Blend where the condition has been shrunk.
196 /// This is used to emphasize that the condition mask is
197 /// no more valid for generic VSELECT optimizations.
200 /// Combined add and sub on an FP vector.
203 // FP vector ops with rounding mode.
212 // FP vector get exponent
215 // Integer add/sub with unsigned saturation.
218 // Integer add/sub with signed saturation.
222 /// Integer horizontal add.
225 /// Integer horizontal sub.
228 /// Floating point horizontal add.
231 /// Floating point horizontal sub.
234 /// Unsigned integer max and min.
237 /// Signed integer max and min.
240 /// Floating point max and min.
243 /// Commutative FMIN and FMAX.
246 /// Floating point reciprocal-sqrt and reciprocal approximation.
247 /// Note that these typically require refinement
248 /// in order to obtain suitable precision.
251 // Thread Local Storage.
254 // Thread Local Storage. A call to get the start address
255 // of the TLS block for the current module.
258 // Thread Local Storage. When calling to an OS provided
259 // thunk at the address from an earlier relocation.
262 // Exception Handling helpers.
265 // SjLj exception handling setjmp.
268 // SjLj exception handling longjmp.
271 /// Tail call return. See X86TargetLowering::LowerCall for
272 /// the list of operands.
275 // Vector move to low scalar and zero higher vector elements.
278 // Vector integer zero-extend.
281 // Vector integer signed-extend.
284 // Vector integer truncate.
287 // Vector integer truncate with mask.
296 // 128-bit vector logical left / right shift
299 // Vector shift elements
302 // Vector shift elements by immediate
305 // Vector packed double/float comparison.
308 // Vector integer comparisons.
310 // Vector integer comparisons, the result is in a mask vector.
313 /// Vector comparison generating mask bits for fp and
314 /// integer signed and unsigned data types.
317 // Vector comparison with rounding mode for FP values
320 // Arithmetic operations with FLAGS results.
321 ADD, SUB, ADC, SBB, SMUL,
322 INC, DEC, OR, XOR, AND,
324 BEXTR, // Bit field extract
326 UMUL, // LOW, HI, FLAGS = umul LHS, RHS
328 // 8-bit SMUL/UMUL - AX, FLAGS = smul8/umul8 AL, RHS
331 // 8-bit divrem that zero-extend the high result (AH).
335 // X86-specific multiply by immediate.
338 // Vector bitwise comparisons.
341 // Vector packed fp sign bitwise comparisons.
344 // Vector "test" in AVX-512, the result is in a mask vector.
348 // OR/AND test for masks
351 // Several flavors of instructions with vector shuffle behaviors.
356 // AVX512 inter-lane alignr
362 //Shuffle Packed Values at 128-bit granularity
383 //Fix Up Special Packed Float32/64 values
385 //Range Restriction Calculation For Packed Pairs of Float32/64 values
387 // Broadcast scalar to vector
389 // Broadcast subvector to vector
391 // Insert/Extract vector element
395 // Vector multiply packed unsigned doubleword integers
397 // Vector multiply packed signed doubleword integers
407 // FMA with rounding mode
416 // Compress and expand
420 //Convert Unsigned/Integer to Scalar Floating-Point Value
424 // Save xmm argument registers to the stack, according to %al. An operator
425 // is needed so that this can be expanded with control flow.
426 VASTART_SAVE_XMM_REGS,
428 // Windows's _chkstk call to do stack probing.
431 // For allocating variable amounts of stack space when using
432 // segmented stacks. Check if the current stacklet has enough space, and
433 // falls back to heap allocation if not.
436 // Windows's _ftol2 runtime routine to do fptoui.
445 // Store FP status word into i16 register.
448 // Store contents of %ah into %eflags.
451 // Get a random integer and indicate whether it is valid in CF.
454 // Get a NIST SP800-90B & C compliant random integer and
455 // indicate whether it is valid in CF.
461 // Test if in transactional execution.
465 RSQRT28, RCP28, EXP2,
468 LCMPXCHG_DAG = ISD::FIRST_TARGET_MEMORY_OPCODE,
472 // Load, scalar_to_vector, and zero extend.
475 // Store FP control world into i16 memory.
478 /// This instruction implements FP_TO_SINT with the
479 /// integer destination in memory and a FP reg source. This corresponds
480 /// to the X86::FIST*m instructions and the rounding mode change stuff. It
481 /// has two inputs (token chain and address) and two outputs (int value
482 /// and token chain).
487 /// This instruction implements SINT_TO_FP with the
488 /// integer source in memory and FP reg result. This corresponds to the
489 /// X86::FILD*m instructions. It has three inputs (token chain, address,
490 /// and source type) and two outputs (FP value and token chain). FILD_FLAG
491 /// also produces a flag).
495 /// This instruction implements an extending load to FP stack slots.
496 /// This corresponds to the X86::FLD32m / X86::FLD64m. It takes a chain
497 /// operand, ptr to load from, and a ValueType node indicating the type
501 /// This instruction implements a truncating store to FP stack
502 /// slots. This corresponds to the X86::FST32m / X86::FST64m. It takes a
503 /// chain operand, value to store, address, and a ValueType to store it
507 /// This instruction grabs the address of the next argument
508 /// from a va_list. (reads and modifies the va_list in memory)
511 // WARNING: Do not add anything in the end unless you want the node to
512 // have memop! In fact, starting from ATOMADD64_DAG all opcodes will be
513 // thought as target memory ops!
517 /// Define some predicates that are used for node matching.
519 /// Return true if the specified
520 /// EXTRACT_SUBVECTOR operand specifies a vector extract that is
521 /// suitable for input to VEXTRACTF128, VEXTRACTI128 instructions.
522 bool isVEXTRACT128Index(SDNode *N);
524 /// Return true if the specified
525 /// INSERT_SUBVECTOR operand specifies a subvector insert that is
526 /// suitable for input to VINSERTF128, VINSERTI128 instructions.
527 bool isVINSERT128Index(SDNode *N);
529 /// Return true if the specified
530 /// EXTRACT_SUBVECTOR operand specifies a vector extract that is
531 /// suitable for input to VEXTRACTF64X4, VEXTRACTI64X4 instructions.
532 bool isVEXTRACT256Index(SDNode *N);
534 /// Return true if the specified
535 /// INSERT_SUBVECTOR operand specifies a subvector insert that is
536 /// suitable for input to VINSERTF64X4, VINSERTI64X4 instructions.
537 bool isVINSERT256Index(SDNode *N);
539 /// Return the appropriate
540 /// immediate to extract the specified EXTRACT_SUBVECTOR index
541 /// with VEXTRACTF128, VEXTRACTI128 instructions.
542 unsigned getExtractVEXTRACT128Immediate(SDNode *N);
544 /// Return the appropriate
545 /// immediate to insert at the specified INSERT_SUBVECTOR index
546 /// with VINSERTF128, VINSERT128 instructions.
547 unsigned getInsertVINSERT128Immediate(SDNode *N);
549 /// Return the appropriate
550 /// immediate to extract the specified EXTRACT_SUBVECTOR index
551 /// with VEXTRACTF64X4, VEXTRACTI64x4 instructions.
552 unsigned getExtractVEXTRACT256Immediate(SDNode *N);
554 /// Return the appropriate
555 /// immediate to insert at the specified INSERT_SUBVECTOR index
556 /// with VINSERTF64x4, VINSERTI64x4 instructions.
557 unsigned getInsertVINSERT256Immediate(SDNode *N);
559 /// Returns true if Elt is a constant zero or floating point constant +0.0.
560 bool isZeroNode(SDValue Elt);
562 /// Returns true of the given offset can be
563 /// fit into displacement field of the instruction.
564 bool isOffsetSuitableForCodeModel(int64_t Offset, CodeModel::Model M,
565 bool hasSymbolicDisplacement = true);
568 /// Determines whether the callee is required to pop its
569 /// own arguments. Callee pop is necessary to support tail calls.
570 bool isCalleePop(CallingConv::ID CallingConv,
571 bool is64Bit, bool IsVarArg, bool TailCallOpt);
573 /// AVX512 static rounding constants. These need to match the values in
575 enum STATIC_ROUNDING {
584 //===--------------------------------------------------------------------===//
585 // X86 Implementation of the TargetLowering interface
586 class X86TargetLowering final : public TargetLowering {
588 explicit X86TargetLowering(const X86TargetMachine &TM,
589 const X86Subtarget &STI);
591 unsigned getJumpTableEncoding() const override;
592 bool useSoftFloat() const override;
594 MVT getScalarShiftAmountTy(EVT LHSTy) const override { return MVT::i8; }
597 LowerCustomJumpTableEntry(const MachineJumpTableInfo *MJTI,
598 const MachineBasicBlock *MBB, unsigned uid,
599 MCContext &Ctx) const override;
601 /// Returns relocation base for the given PIC jumptable.
602 SDValue getPICJumpTableRelocBase(SDValue Table,
603 SelectionDAG &DAG) const override;
605 getPICJumpTableRelocBaseExpr(const MachineFunction *MF,
606 unsigned JTI, MCContext &Ctx) const override;
608 /// Return the desired alignment for ByVal aggregate
609 /// function arguments in the caller parameter area. For X86, aggregates
610 /// that contains are placed at 16-byte boundaries while the rest are at
611 /// 4-byte boundaries.
612 unsigned getByValTypeAlignment(Type *Ty) const override;
614 /// Returns the target specific optimal type for load
615 /// and store operations as a result of memset, memcpy, and memmove
616 /// lowering. If DstAlign is zero that means it's safe to destination
617 /// alignment can satisfy any constraint. Similarly if SrcAlign is zero it
618 /// means there isn't a need to check it against alignment requirement,
619 /// probably because the source does not need to be loaded. If 'IsMemset' is
620 /// true, that means it's expanding a memset. If 'ZeroMemset' is true, that
621 /// means it's a memset of zero. 'MemcpyStrSrc' indicates whether the memcpy
622 /// source is constant so it does not need to be loaded.
623 /// It returns EVT::Other if the type should be determined using generic
624 /// target-independent logic.
625 EVT getOptimalMemOpType(uint64_t Size, unsigned DstAlign, unsigned SrcAlign,
626 bool IsMemset, bool ZeroMemset, bool MemcpyStrSrc,
627 MachineFunction &MF) const override;
629 /// Returns true if it's safe to use load / store of the
630 /// specified type to expand memcpy / memset inline. This is mostly true
631 /// for all types except for some special cases. For example, on X86
632 /// targets without SSE2 f64 load / store are done with fldl / fstpl which
633 /// also does type conversion. Note the specified type doesn't have to be
634 /// legal as the hook is used before type legalization.
635 bool isSafeMemOpType(MVT VT) const override;
637 /// Returns true if the target allows
638 /// unaligned memory accesses. of the specified type. Returns whether it
639 /// is "fast" by reference in the second argument.
640 bool allowsMisalignedMemoryAccesses(EVT VT, unsigned AS, unsigned Align,
641 bool *Fast) const override;
643 /// Provide custom lowering hooks for some operations.
645 SDValue LowerOperation(SDValue Op, SelectionDAG &DAG) const override;
647 /// Replace the results of node with an illegal result
648 /// type with new values built out of custom code.
650 void ReplaceNodeResults(SDNode *N, SmallVectorImpl<SDValue>&Results,
651 SelectionDAG &DAG) const override;
654 SDValue PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const override;
656 /// Return true if the target has native support for
657 /// the specified value type and it is 'desirable' to use the type for the
658 /// given node type. e.g. On x86 i16 is legal, but undesirable since i16
659 /// instruction encodings are longer and some i16 instructions are slow.
660 bool isTypeDesirableForOp(unsigned Opc, EVT VT) const override;
662 /// Return true if the target has native support for the
663 /// specified value type and it is 'desirable' to use the type. e.g. On x86
664 /// i16 is legal, but undesirable since i16 instruction encodings are longer
665 /// and some i16 instructions are slow.
666 bool IsDesirableToPromoteOp(SDValue Op, EVT &PVT) const override;
669 EmitInstrWithCustomInserter(MachineInstr *MI,
670 MachineBasicBlock *MBB) const override;
673 /// This method returns the name of a target specific DAG node.
674 const char *getTargetNodeName(unsigned Opcode) const override;
676 bool isCheapToSpeculateCttz() const override;
678 bool isCheapToSpeculateCtlz() const override;
680 /// Return the value type to use for ISD::SETCC.
681 EVT getSetCCResultType(LLVMContext &Context, EVT VT) const override;
683 /// Determine which of the bits specified in Mask are known to be either
684 /// zero or one and return them in the KnownZero/KnownOne bitsets.
685 void computeKnownBitsForTargetNode(const SDValue Op,
688 const SelectionDAG &DAG,
689 unsigned Depth = 0) const override;
691 /// Determine the number of bits in the operation that are sign bits.
692 unsigned ComputeNumSignBitsForTargetNode(SDValue Op,
693 const SelectionDAG &DAG,
694 unsigned Depth) const override;
696 bool isGAPlusOffset(SDNode *N, const GlobalValue* &GA,
697 int64_t &Offset) const override;
699 SDValue getReturnAddressFrameIndex(SelectionDAG &DAG) const;
701 bool ExpandInlineAsm(CallInst *CI) const override;
704 getConstraintType(const std::string &Constraint) const override;
706 /// Examine constraint string and operand type and determine a weight value.
707 /// The operand object must already have been set up with the operand type.
709 getSingleConstraintMatchWeight(AsmOperandInfo &info,
710 const char *constraint) const override;
712 const char *LowerXConstraint(EVT ConstraintVT) const override;
714 /// Lower the specified operand into the Ops vector. If it is invalid, don't
715 /// add anything to Ops. If hasMemory is true it means one of the asm
716 /// constraint of the inline asm instruction being processed is 'm'.
717 void LowerAsmOperandForConstraint(SDValue Op,
718 std::string &Constraint,
719 std::vector<SDValue> &Ops,
720 SelectionDAG &DAG) const override;
722 unsigned getInlineAsmMemConstraint(
723 const std::string &ConstraintCode) const override {
724 if (ConstraintCode == "i")
725 return InlineAsm::Constraint_i;
726 else if (ConstraintCode == "o")
727 return InlineAsm::Constraint_o;
728 else if (ConstraintCode == "v")
729 return InlineAsm::Constraint_v;
730 else if (ConstraintCode == "X")
731 return InlineAsm::Constraint_X;
732 return TargetLowering::getInlineAsmMemConstraint(ConstraintCode);
735 /// Given a physical register constraint
736 /// (e.g. {edx}), return the register number and the register class for the
737 /// register. This should only be used for C_Register constraints. On
738 /// error, this returns a register number of 0.
739 std::pair<unsigned, const TargetRegisterClass *>
740 getRegForInlineAsmConstraint(const TargetRegisterInfo *TRI,
741 const std::string &Constraint,
742 MVT VT) const override;
744 /// Return true if the addressing mode represented
745 /// by AM is legal for this target, for a load/store of the specified type.
746 bool isLegalAddressingMode(const AddrMode &AM, Type *Ty,
747 unsigned AS) const override;
749 /// Return true if the specified immediate is legal
750 /// icmp immediate, that is the target has icmp instructions which can
751 /// compare a register against the immediate without having to materialize
752 /// the immediate into a register.
753 bool isLegalICmpImmediate(int64_t Imm) const override;
755 /// Return true if the specified immediate is legal
756 /// add immediate, that is the target has add instructions which can
757 /// add a register and the immediate without having to materialize
758 /// the immediate into a register.
759 bool isLegalAddImmediate(int64_t Imm) const override;
761 /// \brief Return the cost of the scaling factor used in the addressing
762 /// mode represented by AM for this target, for a load/store
763 /// of the specified type.
764 /// If the AM is supported, the return value must be >= 0.
765 /// If the AM is not supported, it returns a negative value.
766 int getScalingFactorCost(const AddrMode &AM, Type *Ty,
767 unsigned AS) const override;
769 bool isVectorShiftByScalarCheap(Type *Ty) const override;
771 /// Return true if it's free to truncate a value of
772 /// type Ty1 to type Ty2. e.g. On x86 it's free to truncate a i32 value in
773 /// register EAX to i16 by referencing its sub-register AX.
774 bool isTruncateFree(Type *Ty1, Type *Ty2) const override;
775 bool isTruncateFree(EVT VT1, EVT VT2) const override;
777 bool allowTruncateForTailCall(Type *Ty1, Type *Ty2) const override;
779 /// Return true if any actual instruction that defines a
780 /// value of type Ty1 implicit zero-extends the value to Ty2 in the result
781 /// register. This does not necessarily include registers defined in
782 /// unknown ways, such as incoming arguments, or copies from unknown
783 /// virtual registers. Also, if isTruncateFree(Ty2, Ty1) is true, this
784 /// does not necessarily apply to truncate instructions. e.g. on x86-64,
785 /// all instructions that define 32-bit values implicit zero-extend the
786 /// result out to 64 bits.
787 bool isZExtFree(Type *Ty1, Type *Ty2) const override;
788 bool isZExtFree(EVT VT1, EVT VT2) const override;
789 bool isZExtFree(SDValue Val, EVT VT2) const override;
791 /// Return true if folding a vector load into ExtVal (a sign, zero, or any
792 /// extend node) is profitable.
793 bool isVectorLoadExtDesirable(SDValue) const override;
795 /// Return true if an FMA operation is faster than a pair of fmul and fadd
796 /// instructions. fmuladd intrinsics will be expanded to FMAs when this
797 /// method returns true, otherwise fmuladd is expanded to fmul + fadd.
798 bool isFMAFasterThanFMulAndFAdd(EVT VT) const override;
800 /// Return true if it's profitable to narrow
801 /// operations of type VT1 to VT2. e.g. on x86, it's profitable to narrow
802 /// from i32 to i8 but not from i32 to i16.
803 bool isNarrowingProfitable(EVT VT1, EVT VT2) const override;
805 /// Returns true if the target can instruction select the
806 /// specified FP immediate natively. If false, the legalizer will
807 /// materialize the FP immediate as a load from a constant pool.
808 bool isFPImmLegal(const APFloat &Imm, EVT VT) const override;
810 /// Targets can use this to indicate that they only support *some*
811 /// VECTOR_SHUFFLE operations, those with specific masks. By default, if a
812 /// target supports the VECTOR_SHUFFLE node, all mask values are assumed to
814 bool isShuffleMaskLegal(const SmallVectorImpl<int> &Mask,
815 EVT VT) const override;
817 /// Similar to isShuffleMaskLegal. This is used by Targets can use this to
818 /// indicate if there is a suitable VECTOR_SHUFFLE that can be used to
819 /// replace a VAND with a constant pool entry.
820 bool isVectorClearMaskLegal(const SmallVectorImpl<int> &Mask,
821 EVT VT) const override;
823 /// If true, then instruction selection should
824 /// seek to shrink the FP constant of the specified type to a smaller type
825 /// in order to save space and / or reduce runtime.
826 bool ShouldShrinkFPConstant(EVT VT) const override {
827 // Don't shrink FP constpool if SSE2 is available since cvtss2sd is more
828 // expensive than a straight movsd. On the other hand, it's important to
829 // shrink long double fp constant since fldt is very slow.
830 return !X86ScalarSSEf64 || VT == MVT::f80;
833 /// Return true if we believe it is correct and profitable to reduce the
834 /// load node to a smaller type.
835 bool shouldReduceLoadWidth(SDNode *Load, ISD::LoadExtType ExtTy,
836 EVT NewVT) const override;
838 /// Return true if the specified scalar FP type is computed in an SSE
839 /// register, not on the X87 floating point stack.
840 bool isScalarFPTypeInSSEReg(EVT VT) const {
841 return (VT == MVT::f64 && X86ScalarSSEf64) || // f64 is when SSE2
842 (VT == MVT::f32 && X86ScalarSSEf32); // f32 is when SSE1
845 /// Return true if the target uses the MSVC _ftol2 routine for fptoui.
846 bool isTargetFTOL() const;
848 /// Return true if the MSVC _ftol2 routine should be used for fptoui to the
850 bool isIntegerTypeFTOL(EVT VT) const {
851 return isTargetFTOL() && VT == MVT::i64;
854 /// \brief Returns true if it is beneficial to convert a load of a constant
855 /// to just the constant itself.
856 bool shouldConvertConstantLoadToIntImm(const APInt &Imm,
857 Type *Ty) const override;
859 /// Return true if EXTRACT_SUBVECTOR is cheap for this result type
861 bool isExtractSubvectorCheap(EVT ResVT, unsigned Index) const override;
863 /// Intel processors have a unified instruction and data cache
864 const char * getClearCacheBuiltinName() const override {
865 return nullptr; // nothing to do, move along.
868 unsigned getRegisterByName(const char* RegName, EVT VT) const override;
870 /// This method returns a target specific FastISel object,
871 /// or null if the target does not support "fast" ISel.
872 FastISel *createFastISel(FunctionLoweringInfo &funcInfo,
873 const TargetLibraryInfo *libInfo) const override;
875 /// Return true if the target stores stack protector cookies at a fixed
876 /// offset in some non-standard address space, and populates the address
877 /// space and offset as appropriate.
878 bool getStackCookieLocation(unsigned &AddressSpace,
879 unsigned &Offset) const override;
881 SDValue BuildFILD(SDValue Op, EVT SrcVT, SDValue Chain, SDValue StackSlot,
882 SelectionDAG &DAG) const;
884 bool isNoopAddrSpaceCast(unsigned SrcAS, unsigned DestAS) const override;
886 bool useLoadStackGuardNode() const override;
887 /// \brief Customize the preferred legalization strategy for certain types.
888 LegalizeTypeAction getPreferredVectorAction(EVT VT) const override;
891 std::pair<const TargetRegisterClass *, uint8_t>
892 findRepresentativeClass(const TargetRegisterInfo *TRI,
893 MVT VT) const override;
896 /// Keep a pointer to the X86Subtarget around so that we can
897 /// make the right decision when generating code for different targets.
898 const X86Subtarget *Subtarget;
899 const DataLayout *TD;
901 /// Select between SSE or x87 floating point ops.
902 /// When SSE is available, use it for f32 operations.
903 /// When SSE2 is available, use it for f64 operations.
904 bool X86ScalarSSEf32;
905 bool X86ScalarSSEf64;
907 /// A list of legal FP immediates.
908 std::vector<APFloat> LegalFPImmediates;
910 /// Indicate that this x86 target can instruction
911 /// select the specified FP immediate natively.
912 void addLegalFPImmediate(const APFloat& Imm) {
913 LegalFPImmediates.push_back(Imm);
916 SDValue LowerCallResult(SDValue Chain, SDValue InFlag,
917 CallingConv::ID CallConv, bool isVarArg,
918 const SmallVectorImpl<ISD::InputArg> &Ins,
919 SDLoc dl, SelectionDAG &DAG,
920 SmallVectorImpl<SDValue> &InVals) const;
921 SDValue LowerMemArgument(SDValue Chain,
922 CallingConv::ID CallConv,
923 const SmallVectorImpl<ISD::InputArg> &ArgInfo,
924 SDLoc dl, SelectionDAG &DAG,
925 const CCValAssign &VA, MachineFrameInfo *MFI,
927 SDValue LowerMemOpCallTo(SDValue Chain, SDValue StackPtr, SDValue Arg,
928 SDLoc dl, SelectionDAG &DAG,
929 const CCValAssign &VA,
930 ISD::ArgFlagsTy Flags) const;
932 // Call lowering helpers.
934 /// Check whether the call is eligible for tail call optimization. Targets
935 /// that want to do tail call optimization should implement this function.
936 bool IsEligibleForTailCallOptimization(SDValue Callee,
937 CallingConv::ID CalleeCC,
939 bool isCalleeStructRet,
940 bool isCallerStructRet,
942 const SmallVectorImpl<ISD::OutputArg> &Outs,
943 const SmallVectorImpl<SDValue> &OutVals,
944 const SmallVectorImpl<ISD::InputArg> &Ins,
945 SelectionDAG& DAG) const;
946 bool IsCalleePop(bool isVarArg, CallingConv::ID CallConv) const;
947 SDValue EmitTailCallLoadRetAddr(SelectionDAG &DAG, SDValue &OutRetAddr,
948 SDValue Chain, bool IsTailCall, bool Is64Bit,
949 int FPDiff, SDLoc dl) const;
951 unsigned GetAlignedArgumentStackSize(unsigned StackSize,
952 SelectionDAG &DAG) const;
954 std::pair<SDValue,SDValue> FP_TO_INTHelper(SDValue Op, SelectionDAG &DAG,
956 bool isReplace) const;
958 SDValue LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) const;
959 SDValue LowerBUILD_VECTORvXi1(SDValue Op, SelectionDAG &DAG) const;
960 SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) const;
961 SDValue LowerVSELECT(SDValue Op, SelectionDAG &DAG) const;
962 SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const;
963 SDValue ExtractBitFromMaskVector(SDValue Op, SelectionDAG &DAG) const;
964 SDValue InsertBitToMaskVector(SDValue Op, SelectionDAG &DAG) const;
966 SDValue LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const;
967 SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) const;
968 SDValue LowerBlockAddress(SDValue Op, SelectionDAG &DAG) const;
969 SDValue LowerGlobalAddress(const GlobalValue *GV, SDLoc dl,
970 int64_t Offset, SelectionDAG &DAG) const;
971 SDValue LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) const;
972 SDValue LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const;
973 SDValue LowerExternalSymbol(SDValue Op, SelectionDAG &DAG) const;
974 SDValue LowerSINT_TO_FP(SDValue Op, SelectionDAG &DAG) const;
975 SDValue LowerUINT_TO_FP(SDValue Op, SelectionDAG &DAG) const;
976 SDValue LowerUINT_TO_FP_i64(SDValue Op, SelectionDAG &DAG) const;
977 SDValue LowerUINT_TO_FP_i32(SDValue Op, SelectionDAG &DAG) const;
978 SDValue lowerUINT_TO_FP_vec(SDValue Op, SelectionDAG &DAG) const;
979 SDValue LowerTRUNCATE(SDValue Op, SelectionDAG &DAG) const;
980 SDValue LowerFP_TO_SINT(SDValue Op, SelectionDAG &DAG) const;
981 SDValue LowerFP_TO_UINT(SDValue Op, SelectionDAG &DAG) const;
982 SDValue LowerToBT(SDValue And, ISD::CondCode CC,
983 SDLoc dl, SelectionDAG &DAG) const;
984 SDValue LowerSETCC(SDValue Op, SelectionDAG &DAG) const;
985 SDValue LowerSELECT(SDValue Op, SelectionDAG &DAG) const;
986 SDValue LowerBRCOND(SDValue Op, SelectionDAG &DAG) const;
987 SDValue LowerMEMSET(SDValue Op, SelectionDAG &DAG) const;
988 SDValue LowerJumpTable(SDValue Op, SelectionDAG &DAG) const;
989 SDValue LowerDYNAMIC_STACKALLOC(SDValue Op, SelectionDAG &DAG) const;
990 SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG) const;
991 SDValue LowerVAARG(SDValue Op, SelectionDAG &DAG) const;
992 SDValue LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const;
993 SDValue LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const;
994 SDValue LowerFRAME_TO_ARGS_OFFSET(SDValue Op, SelectionDAG &DAG) const;
995 SDValue LowerEH_RETURN(SDValue Op, SelectionDAG &DAG) const;
996 SDValue lowerEH_SJLJ_SETJMP(SDValue Op, SelectionDAG &DAG) const;
997 SDValue lowerEH_SJLJ_LONGJMP(SDValue Op, SelectionDAG &DAG) const;
998 SDValue LowerINIT_TRAMPOLINE(SDValue Op, SelectionDAG &DAG) const;
999 SDValue LowerFLT_ROUNDS_(SDValue Op, SelectionDAG &DAG) const;
1000 SDValue LowerWin64_i128OP(SDValue Op, SelectionDAG &DAG) const;
1001 SDValue LowerGC_TRANSITION_START(SDValue Op, SelectionDAG &DAG) const;
1002 SDValue LowerGC_TRANSITION_END(SDValue Op, SelectionDAG &DAG) const;
1005 LowerFormalArguments(SDValue Chain,
1006 CallingConv::ID CallConv, bool isVarArg,
1007 const SmallVectorImpl<ISD::InputArg> &Ins,
1008 SDLoc dl, SelectionDAG &DAG,
1009 SmallVectorImpl<SDValue> &InVals) const override;
1010 SDValue LowerCall(CallLoweringInfo &CLI,
1011 SmallVectorImpl<SDValue> &InVals) const override;
1013 SDValue LowerReturn(SDValue Chain,
1014 CallingConv::ID CallConv, bool isVarArg,
1015 const SmallVectorImpl<ISD::OutputArg> &Outs,
1016 const SmallVectorImpl<SDValue> &OutVals,
1017 SDLoc dl, SelectionDAG &DAG) const override;
1019 bool isUsedByReturnOnly(SDNode *N, SDValue &Chain) const override;
1021 bool mayBeEmittedAsTailCall(CallInst *CI) const override;
1023 EVT getTypeForExtArgOrReturn(LLVMContext &Context, EVT VT,
1024 ISD::NodeType ExtendKind) const override;
1026 bool CanLowerReturn(CallingConv::ID CallConv, MachineFunction &MF,
1028 const SmallVectorImpl<ISD::OutputArg> &Outs,
1029 LLVMContext &Context) const override;
1031 const MCPhysReg *getScratchRegisters(CallingConv::ID CC) const override;
1033 bool shouldExpandAtomicLoadInIR(LoadInst *SI) const override;
1034 bool shouldExpandAtomicStoreInIR(StoreInst *SI) const override;
1035 TargetLoweringBase::AtomicRMWExpansionKind
1036 shouldExpandAtomicRMWInIR(AtomicRMWInst *AI) const override;
1039 lowerIdempotentRMWIntoFencedLoad(AtomicRMWInst *AI) const override;
1041 bool needsCmpXchgNb(const Type *MemType) const;
1043 /// Utility function to emit atomic-load-arith operations (and, or, xor,
1044 /// nand, max, min, umax, umin). It takes the corresponding instruction to
1045 /// expand, the associated machine basic block, and the associated X86
1046 /// opcodes for reg/reg.
1047 MachineBasicBlock *EmitAtomicLoadArith(MachineInstr *MI,
1048 MachineBasicBlock *MBB) const;
1050 /// Utility function to emit atomic-load-arith operations (and, or, xor,
1051 /// nand, add, sub, swap) for 64-bit operands on 32-bit target.
1052 MachineBasicBlock *EmitAtomicLoadArith6432(MachineInstr *MI,
1053 MachineBasicBlock *MBB) const;
1055 // Utility function to emit the low-level va_arg code for X86-64.
1056 MachineBasicBlock *EmitVAARG64WithCustomInserter(
1058 MachineBasicBlock *MBB) const;
1060 /// Utility function to emit the xmm reg save portion of va_start.
1061 MachineBasicBlock *EmitVAStartSaveXMMRegsWithCustomInserter(
1062 MachineInstr *BInstr,
1063 MachineBasicBlock *BB) const;
1065 MachineBasicBlock *EmitLoweredSelect(MachineInstr *I,
1066 MachineBasicBlock *BB) const;
1068 MachineBasicBlock *EmitLoweredWinAlloca(MachineInstr *MI,
1069 MachineBasicBlock *BB) const;
1071 MachineBasicBlock *EmitLoweredSegAlloca(MachineInstr *MI,
1072 MachineBasicBlock *BB) const;
1074 MachineBasicBlock *EmitLoweredTLSCall(MachineInstr *MI,
1075 MachineBasicBlock *BB) const;
1077 MachineBasicBlock *emitLoweredTLSAddr(MachineInstr *MI,
1078 MachineBasicBlock *BB) const;
1080 MachineBasicBlock *emitEHSjLjSetJmp(MachineInstr *MI,
1081 MachineBasicBlock *MBB) const;
1083 MachineBasicBlock *emitEHSjLjLongJmp(MachineInstr *MI,
1084 MachineBasicBlock *MBB) const;
1086 MachineBasicBlock *emitFMA3Instr(MachineInstr *MI,
1087 MachineBasicBlock *MBB) const;
1089 /// Emit nodes that will be selected as "test Op0,Op0", or something
1090 /// equivalent, for use with the given x86 condition code.
1091 SDValue EmitTest(SDValue Op0, unsigned X86CC, SDLoc dl,
1092 SelectionDAG &DAG) const;
1094 /// Emit nodes that will be selected as "cmp Op0,Op1", or something
1095 /// equivalent, for use with the given x86 condition code.
1096 SDValue EmitCmp(SDValue Op0, SDValue Op1, unsigned X86CC, SDLoc dl,
1097 SelectionDAG &DAG) const;
1099 /// Convert a comparison if required by the subtarget.
1100 SDValue ConvertCmpIfNecessary(SDValue Cmp, SelectionDAG &DAG) const;
1102 /// Use rsqrt* to speed up sqrt calculations.
1103 SDValue getRsqrtEstimate(SDValue Operand, DAGCombinerInfo &DCI,
1104 unsigned &RefinementSteps,
1105 bool &UseOneConstNR) const override;
1107 /// Use rcp* to speed up fdiv calculations.
1108 SDValue getRecipEstimate(SDValue Operand, DAGCombinerInfo &DCI,
1109 unsigned &RefinementSteps) const override;
1111 /// Reassociate floating point divisions into multiply by reciprocal.
1112 bool combineRepeatedFPDivisors(unsigned NumUsers) const override;
1116 FastISel *createFastISel(FunctionLoweringInfo &funcInfo,
1117 const TargetLibraryInfo *libInfo);
1121 #endif // X86ISELLOWERING_H