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
30 // Start the numbering where the builtin ops leave off.
31 FIRST_NUMBER = ISD::BUILTIN_OP_END,
33 /// BSF - Bit scan forward.
34 /// BSR - Bit scan reverse.
38 /// SHLD, SHRD - Double shift instructions. These correspond to
39 /// X86::SHLDxx and X86::SHRDxx instructions.
43 /// FAND - Bitwise logical AND of floating point values. This corresponds
44 /// to X86::ANDPS or X86::ANDPD.
47 /// FOR - Bitwise logical OR of floating point values. This corresponds
48 /// to X86::ORPS or X86::ORPD.
51 /// FXOR - Bitwise logical XOR of floating point values. This corresponds
52 /// to X86::XORPS or X86::XORPD.
55 /// FANDN - Bitwise logical ANDNOT of floating point values. This
56 /// corresponds to X86::ANDNPS or X86::ANDNPD.
59 /// FSRL - Bitwise logical right shift of floating point values. These
60 /// corresponds to X86::PSRLDQ.
63 /// CALL - 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 /// RDTSC_DAG - This operation implements the lowering for
86 /// X86 Read Time-Stamp Counter and Processor ID.
89 /// X86 Read Performance Monitoring Counters.
92 /// X86 compare and logical compare instructions.
95 /// X86 bit-test instructions.
98 /// X86 SetCC. Operand 0 is condition code, and operand 1 is the EFLAGS
99 /// operand, usually produced by a CMP instruction.
105 // Same as SETCC except it's materialized with a sbb and the value is all
106 // one's or all zero's.
107 SETCC_CARRY, // R = carry_bit ? ~0 : 0
109 /// X86 FP SETCC, implemented with CMP{cc}SS/CMP{cc}SD.
110 /// Operands are two FP values to compare; result is a mask of
111 /// 0s or 1s. Generally DTRT for C/C++ with NaNs.
114 /// X86 MOVMSK{pd|ps}, extracts sign bits of two or four FP values,
115 /// result in an integer GPR. Needs masking for scalar result.
118 /// X86 conditional moves. Operand 0 and operand 1 are the two values
119 /// to select from. Operand 2 is the condition code, and operand 3 is the
120 /// flag operand produced by a CMP or TEST instruction. It also writes a
124 /// X86 conditional branches. Operand 0 is the chain operand, operand 1
125 /// is the block to branch if condition is true, operand 2 is the
126 /// condition code, and operand 3 is the flag operand produced by a CMP
127 /// or TEST instruction.
130 /// Return with a flag operand. Operand 0 is the chain operand, operand
131 /// 1 is the number of bytes of stack to pop.
134 /// REP_STOS - Repeat fill, corresponds to X86::REP_STOSx.
137 /// REP_MOVS - Repeat move, corresponds to X86::REP_MOVSx.
140 /// GlobalBaseReg - On Darwin, this node represents the result of the popl
141 /// at function entry, used for PIC code.
144 /// Wrapper - A wrapper node for TargetConstantPool,
145 /// TargetExternalSymbol, and TargetGlobalAddress.
148 /// WrapperRIP - Special wrapper used under X86-64 PIC mode for RIP
149 /// relative displacements.
152 /// MOVDQ2Q - Copies a 64-bit value from the low word of an XMM vector
153 /// to an MMX vector. If you think this is too close to the previous
154 /// mnemonic, so do I; blame Intel.
157 /// MMX_MOVD2W - Copies a 32-bit value from the low word of a MMX
161 /// MMX_MOVW2D - Copies a GPR into the low 32-bit word of a MMX vector
162 /// and zero out the high word.
165 /// PEXTRB - Extract an 8-bit value from a vector and zero extend it to
166 /// i32, corresponds to X86::PEXTRB.
169 /// PEXTRW - Extract a 16-bit value from a vector and zero extend it to
170 /// i32, corresponds to X86::PEXTRW.
173 /// INSERTPS - Insert any element of a 4 x float vector into any element
174 /// of a destination 4 x floatvector.
177 /// PINSRB - Insert the lower 8-bits of a 32-bit value to a vector,
178 /// corresponds to X86::PINSRB.
181 /// PINSRW - Insert the lower 16-bits of a 32-bit value to a vector,
182 /// corresponds to X86::PINSRW.
185 /// PSHUFB - Shuffle 16 8-bit values within a vector.
188 /// ANDNP - Bitwise Logical AND NOT of Packed FP values.
191 /// PSIGN - Copy integer sign.
194 /// BLENDI - Blend where the selector is an immediate.
197 /// SHRUNKBLEND - 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 /// ADDSUB - Combined add and sub on an FP vector.
204 // FADD, FSUB, FMUL, FDIV, FMIN, FMAX - FP vector ops with rounding mode.
210 // SUBUS - Integer sub with unsigned saturation.
213 /// HADD - Integer horizontal add.
216 /// HSUB - Integer horizontal sub.
219 /// FHADD - Floating point horizontal add.
222 /// FHSUB - Floating point horizontal sub.
225 /// UMAX, UMIN - Unsigned integer max and min.
228 /// SMAX, SMIN - Signed integer max and min.
231 /// FMAX, FMIN - Floating point max and min.
235 /// FMAXC, FMINC - Commutative FMIN and FMAX.
238 /// FRSQRT, FRCP - Floating point reciprocal-sqrt and reciprocal
239 /// approximation. Note that these typically require refinement
240 /// in order to obtain suitable precision.
243 // TLSADDR - Thread Local Storage.
246 // TLSBASEADDR - Thread Local Storage. A call to get the start address
247 // of the TLS block for the current module.
250 // TLSCALL - Thread Local Storage. When calling to an OS provided
251 // thunk at the address from an earlier relocation.
254 // EH_RETURN - Exception Handling helpers.
257 // EH_SJLJ_SETJMP - SjLj exception handling setjmp.
260 // EH_SJLJ_LONGJMP - SjLj exception handling longjmp.
263 /// TC_RETURN - Tail call return. See X86TargetLowering::LowerCall for
264 /// the list of operands.
267 // VZEXT_MOVL - Vector move to low scalar and zero higher vector elements.
270 // VZEXT - Vector integer zero-extend.
273 // VSEXT - Vector integer signed-extend.
276 // VTRUNC - Vector integer truncate.
279 // VTRUNC - Vector integer truncate with mask.
282 // VFPEXT - Vector FP extend.
285 // VFPROUND - Vector FP round.
288 // VSHL, VSRL - 128-bit vector logical left / right shift
291 // VSHL, VSRL, VSRA - Vector shift elements
294 // VSHLI, VSRLI, VSRAI - Vector shift elements by immediate
297 // CMPP - Vector packed double/float comparison.
300 // PCMP* - Vector integer comparisons.
302 // PCMP*M - Vector integer comparisons, the result is in a mask vector.
305 /// CMPM, CMPMU - Vector comparison generating mask bits for fp and
306 /// integer signed and unsigned data types.
310 // ADD, SUB, SMUL, etc. - Arithmetic operations with FLAGS results.
311 ADD, SUB, ADC, SBB, SMUL,
312 INC, DEC, OR, XOR, AND,
314 BEXTR, // BEXTR - Bit field extract
316 UMUL, // LOW, HI, FLAGS = umul LHS, RHS
318 // 8-bit SMUL/UMUL - AX, FLAGS = smul8/umul8 AL, RHS
321 // 8-bit divrem that zero-extend the high result (AH).
325 // MUL_IMM - X86 specific multiply by immediate.
328 // PTEST - Vector bitwise comparisons.
331 // TESTP - Vector packed fp sign bitwise comparisons.
334 // TESTM, TESTNM - Vector "test" in AVX-512, the result is in a mask vector.
338 // OR/AND test for masks
341 // Several flavors of instructions with vector shuffle behaviors.
346 // AVX512 inter-lane alignr
374 // Insert/Extract vector element
378 // Vector multiply packed unsigned doubleword integers
380 // Vector multiply packed signed doubleword integers
390 // FMA with rounding mode
399 // Compress and expand
403 // Save xmm argument registers to the stack, according to %al. An operator
404 // is needed so that this can be expanded with control flow.
405 VASTART_SAVE_XMM_REGS,
407 // Windows's _chkstk call to do stack probing.
410 // For allocating variable amounts of stack space when using
411 // segmented stacks. Check if the current stacklet has enough space, and
412 // falls back to heap allocation if not.
415 // Windows's _ftol2 runtime routine to do fptoui.
424 // Store FP status word into i16 register.
427 // Store contents of %ah into %eflags.
430 // Get a random integer and indicate whether it is valid in CF.
433 // Get a NIST SP800-90B & C compliant random integer and
434 // indicate whether it is valid in CF.
440 // Test if in transactional execution.
444 RSQRT28, RCP28, EXP2,
447 LCMPXCHG_DAG = ISD::FIRST_TARGET_MEMORY_OPCODE,
451 // Load, scalar_to_vector, and zero extend.
454 // Store FP control world into i16 memory.
457 /// This instruction implements FP_TO_SINT with the
458 /// integer destination in memory and a FP reg source. This corresponds
459 /// to the X86::FIST*m instructions and the rounding mode change stuff. It
460 /// has two inputs (token chain and address) and two outputs (int value
461 /// and token chain).
466 /// This instruction implements SINT_TO_FP with the
467 /// integer source in memory and FP reg result. This corresponds to the
468 /// X86::FILD*m instructions. It has three inputs (token chain, address,
469 /// and source type) and two outputs (FP value and token chain). FILD_FLAG
470 /// also produces a flag).
474 /// This instruction implements an extending load to FP stack slots.
475 /// This corresponds to the X86::FLD32m / X86::FLD64m. It takes a chain
476 /// operand, ptr to load from, and a ValueType node indicating the type
480 /// This instruction implements a truncating store to FP stack
481 /// slots. This corresponds to the X86::FST32m / X86::FST64m. It takes a
482 /// chain operand, value to store, address, and a ValueType to store it
486 /// This instruction grabs the address of the next argument
487 /// from a va_list. (reads and modifies the va_list in memory)
490 // WARNING: Do not add anything in the end unless you want the node to
491 // have memop! In fact, starting from ATOMADD64_DAG all opcodes will be
492 // thought as target memory ops!
496 /// Define some predicates that are used for node matching.
498 /// Return true if the specified
499 /// EXTRACT_SUBVECTOR operand specifies a vector extract that is
500 /// suitable for input to VEXTRACTF128, VEXTRACTI128 instructions.
501 bool isVEXTRACT128Index(SDNode *N);
503 /// Return true if the specified
504 /// INSERT_SUBVECTOR operand specifies a subvector insert that is
505 /// suitable for input to VINSERTF128, VINSERTI128 instructions.
506 bool isVINSERT128Index(SDNode *N);
508 /// Return true if the specified
509 /// EXTRACT_SUBVECTOR operand specifies a vector extract that is
510 /// suitable for input to VEXTRACTF64X4, VEXTRACTI64X4 instructions.
511 bool isVEXTRACT256Index(SDNode *N);
513 /// Return true if the specified
514 /// INSERT_SUBVECTOR operand specifies a subvector insert that is
515 /// suitable for input to VINSERTF64X4, VINSERTI64X4 instructions.
516 bool isVINSERT256Index(SDNode *N);
518 /// Return the appropriate
519 /// immediate to extract the specified EXTRACT_SUBVECTOR index
520 /// with VEXTRACTF128, VEXTRACTI128 instructions.
521 unsigned getExtractVEXTRACT128Immediate(SDNode *N);
523 /// Return the appropriate
524 /// immediate to insert at the specified INSERT_SUBVECTOR index
525 /// with VINSERTF128, VINSERT128 instructions.
526 unsigned getInsertVINSERT128Immediate(SDNode *N);
528 /// Return the appropriate
529 /// immediate to extract the specified EXTRACT_SUBVECTOR index
530 /// with VEXTRACTF64X4, VEXTRACTI64x4 instructions.
531 unsigned getExtractVEXTRACT256Immediate(SDNode *N);
533 /// Return the appropriate
534 /// immediate to insert at the specified INSERT_SUBVECTOR index
535 /// with VINSERTF64x4, VINSERTI64x4 instructions.
536 unsigned getInsertVINSERT256Immediate(SDNode *N);
538 /// Returns true if Elt is a constant zero or floating point constant +0.0.
539 bool isZeroNode(SDValue Elt);
541 /// Returns true of the given offset can be
542 /// fit into displacement field of the instruction.
543 bool isOffsetSuitableForCodeModel(int64_t Offset, CodeModel::Model M,
544 bool hasSymbolicDisplacement = true);
547 /// Determines whether the callee is required to pop its
548 /// own arguments. Callee pop is necessary to support tail calls.
549 bool isCalleePop(CallingConv::ID CallingConv,
550 bool is64Bit, bool IsVarArg, bool TailCallOpt);
552 /// AVX512 static rounding constants. These need to match the values in
554 enum STATIC_ROUNDING {
563 //===--------------------------------------------------------------------===//
564 // X86 Implementation of the TargetLowering interface
565 class X86TargetLowering final : public TargetLowering {
567 explicit X86TargetLowering(const X86TargetMachine &TM,
568 const X86Subtarget &STI);
570 unsigned getJumpTableEncoding() const override;
572 MVT getScalarShiftAmountTy(EVT LHSTy) const override { return MVT::i8; }
575 LowerCustomJumpTableEntry(const MachineJumpTableInfo *MJTI,
576 const MachineBasicBlock *MBB, unsigned uid,
577 MCContext &Ctx) const override;
579 /// Returns relocation base for the given PIC jumptable.
580 SDValue getPICJumpTableRelocBase(SDValue Table,
581 SelectionDAG &DAG) const override;
583 getPICJumpTableRelocBaseExpr(const MachineFunction *MF,
584 unsigned JTI, MCContext &Ctx) const override;
586 /// Return the desired alignment for ByVal aggregate
587 /// function arguments in the caller parameter area. For X86, aggregates
588 /// that contains are placed at 16-byte boundaries while the rest are at
589 /// 4-byte boundaries.
590 unsigned getByValTypeAlignment(Type *Ty) const override;
592 /// Returns the target specific optimal type for load
593 /// and store operations as a result of memset, memcpy, and memmove
594 /// lowering. If DstAlign is zero that means it's safe to destination
595 /// alignment can satisfy any constraint. Similarly if SrcAlign is zero it
596 /// means there isn't a need to check it against alignment requirement,
597 /// probably because the source does not need to be loaded. If 'IsMemset' is
598 /// true, that means it's expanding a memset. If 'ZeroMemset' is true, that
599 /// means it's a memset of zero. 'MemcpyStrSrc' indicates whether the memcpy
600 /// source is constant so it does not need to be loaded.
601 /// It returns EVT::Other if the type should be determined using generic
602 /// target-independent logic.
603 EVT getOptimalMemOpType(uint64_t Size, unsigned DstAlign, unsigned SrcAlign,
604 bool IsMemset, bool ZeroMemset, bool MemcpyStrSrc,
605 MachineFunction &MF) const override;
607 /// Returns true if it's safe to use load / store of the
608 /// specified type to expand memcpy / memset inline. This is mostly true
609 /// for all types except for some special cases. For example, on X86
610 /// targets without SSE2 f64 load / store are done with fldl / fstpl which
611 /// also does type conversion. Note the specified type doesn't have to be
612 /// legal as the hook is used before type legalization.
613 bool isSafeMemOpType(MVT VT) const override;
615 /// Returns true if the target allows
616 /// unaligned memory accesses. of the specified type. Returns whether it
617 /// is "fast" by reference in the second argument.
618 bool allowsMisalignedMemoryAccesses(EVT VT, unsigned AS, unsigned Align,
619 bool *Fast) const override;
621 /// Provide custom lowering hooks for some operations.
623 SDValue LowerOperation(SDValue Op, SelectionDAG &DAG) const override;
625 /// Replace the results of node with an illegal result
626 /// type with new values built out of custom code.
628 void ReplaceNodeResults(SDNode *N, SmallVectorImpl<SDValue>&Results,
629 SelectionDAG &DAG) const override;
632 SDValue PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const override;
634 /// Return true if the target has native support for
635 /// the specified value type and it is 'desirable' to use the type for the
636 /// given node type. e.g. On x86 i16 is legal, but undesirable since i16
637 /// instruction encodings are longer and some i16 instructions are slow.
638 bool isTypeDesirableForOp(unsigned Opc, EVT VT) const override;
640 /// Return true if the target has native support for the
641 /// specified value type and it is 'desirable' to use the type. e.g. On x86
642 /// i16 is legal, but undesirable since i16 instruction encodings are longer
643 /// and some i16 instructions are slow.
644 bool IsDesirableToPromoteOp(SDValue Op, EVT &PVT) const override;
647 EmitInstrWithCustomInserter(MachineInstr *MI,
648 MachineBasicBlock *MBB) const override;
651 /// This method returns the name of a target specific DAG node.
652 const char *getTargetNodeName(unsigned Opcode) const override;
654 bool isCheapToSpeculateCttz() const override;
656 bool isCheapToSpeculateCtlz() const override;
658 /// Return the value type to use for ISD::SETCC.
659 EVT getSetCCResultType(LLVMContext &Context, EVT VT) const override;
661 /// Determine which of the bits specified in Mask are known to be either
662 /// zero or one and return them in the KnownZero/KnownOne bitsets.
663 void computeKnownBitsForTargetNode(const SDValue Op,
666 const SelectionDAG &DAG,
667 unsigned Depth = 0) const override;
669 /// Determine the number of bits in the operation that are sign bits.
670 unsigned ComputeNumSignBitsForTargetNode(SDValue Op,
671 const SelectionDAG &DAG,
672 unsigned Depth) const override;
674 bool isGAPlusOffset(SDNode *N, const GlobalValue* &GA,
675 int64_t &Offset) const override;
677 SDValue getReturnAddressFrameIndex(SelectionDAG &DAG) const;
679 bool ExpandInlineAsm(CallInst *CI) const override;
682 getConstraintType(const std::string &Constraint) const override;
684 /// Examine constraint string and operand type and determine a weight value.
685 /// The operand object must already have been set up with the operand type.
687 getSingleConstraintMatchWeight(AsmOperandInfo &info,
688 const char *constraint) const override;
690 const char *LowerXConstraint(EVT ConstraintVT) const override;
692 /// Lower the specified operand into the Ops vector. If it is invalid, don't
693 /// add anything to Ops. If hasMemory is true it means one of the asm
694 /// constraint of the inline asm instruction being processed is 'm'.
695 void LowerAsmOperandForConstraint(SDValue Op,
696 std::string &Constraint,
697 std::vector<SDValue> &Ops,
698 SelectionDAG &DAG) const override;
700 /// Given a physical register constraint
701 /// (e.g. {edx}), return the register number and the register class for the
702 /// register. This should only be used for C_Register constraints. On
703 /// error, this returns a register number of 0.
704 std::pair<unsigned, const TargetRegisterClass*>
705 getRegForInlineAsmConstraint(const std::string &Constraint,
706 MVT VT) const override;
708 /// Return true if the addressing mode represented
709 /// by AM is legal for this target, for a load/store of the specified type.
710 bool isLegalAddressingMode(const AddrMode &AM, Type *Ty) const override;
712 /// Return true if the specified immediate is legal
713 /// icmp immediate, that is the target has icmp instructions which can
714 /// compare a register against the immediate without having to materialize
715 /// the immediate into a register.
716 bool isLegalICmpImmediate(int64_t Imm) const override;
718 /// Return true if the specified immediate is legal
719 /// add immediate, that is the target has add instructions which can
720 /// add a register and the immediate without having to materialize
721 /// the immediate into a register.
722 bool isLegalAddImmediate(int64_t Imm) const override;
724 /// \brief Return the cost of the scaling factor used in the addressing
725 /// mode represented by AM for this target, for a load/store
726 /// of the specified type.
727 /// If the AM is supported, the return value must be >= 0.
728 /// If the AM is not supported, it returns a negative value.
729 int getScalingFactorCost(const AddrMode &AM, Type *Ty) const override;
731 bool isVectorShiftByScalarCheap(Type *Ty) const override;
733 /// Return true if it's free to truncate a value of
734 /// type Ty1 to type Ty2. e.g. On x86 it's free to truncate a i32 value in
735 /// register EAX to i16 by referencing its sub-register AX.
736 bool isTruncateFree(Type *Ty1, Type *Ty2) const override;
737 bool isTruncateFree(EVT VT1, EVT VT2) const override;
739 bool allowTruncateForTailCall(Type *Ty1, Type *Ty2) const override;
741 /// Return true if any actual instruction that defines a
742 /// value of type Ty1 implicit zero-extends the value to Ty2 in the result
743 /// register. This does not necessarily include registers defined in
744 /// unknown ways, such as incoming arguments, or copies from unknown
745 /// virtual registers. Also, if isTruncateFree(Ty2, Ty1) is true, this
746 /// does not necessarily apply to truncate instructions. e.g. on x86-64,
747 /// all instructions that define 32-bit values implicit zero-extend the
748 /// result out to 64 bits.
749 bool isZExtFree(Type *Ty1, Type *Ty2) const override;
750 bool isZExtFree(EVT VT1, EVT VT2) const override;
751 bool isZExtFree(SDValue Val, EVT VT2) const override;
753 /// Return true if folding a vector load into ExtVal (a sign, zero, or any
754 /// extend node) is profitable.
755 bool isVectorLoadExtDesirable(SDValue) const override;
757 /// Return true if an FMA operation is faster than a pair of fmul and fadd
758 /// instructions. fmuladd intrinsics will be expanded to FMAs when this
759 /// method returns true, otherwise fmuladd is expanded to fmul + fadd.
760 bool isFMAFasterThanFMulAndFAdd(EVT VT) const override;
762 /// Return true if it's profitable to narrow
763 /// operations of type VT1 to VT2. e.g. on x86, it's profitable to narrow
764 /// from i32 to i8 but not from i32 to i16.
765 bool isNarrowingProfitable(EVT VT1, EVT VT2) const override;
767 /// Returns true if the target can instruction select the
768 /// specified FP immediate natively. If false, the legalizer will
769 /// materialize the FP immediate as a load from a constant pool.
770 bool isFPImmLegal(const APFloat &Imm, EVT VT) const override;
772 /// Targets can use this to indicate that they only support *some*
773 /// VECTOR_SHUFFLE operations, those with specific masks. By default, if a
774 /// target supports the VECTOR_SHUFFLE node, all mask values are assumed to
776 bool isShuffleMaskLegal(const SmallVectorImpl<int> &Mask,
777 EVT VT) const override;
779 /// Similar to isShuffleMaskLegal. This is used by Targets can use this to
780 /// indicate if there is a suitable VECTOR_SHUFFLE that can be used to
781 /// replace a VAND with a constant pool entry.
782 bool isVectorClearMaskLegal(const SmallVectorImpl<int> &Mask,
783 EVT VT) const override;
785 /// If true, then instruction selection should
786 /// seek to shrink the FP constant of the specified type to a smaller type
787 /// in order to save space and / or reduce runtime.
788 bool ShouldShrinkFPConstant(EVT VT) const override {
789 // Don't shrink FP constpool if SSE2 is available since cvtss2sd is more
790 // expensive than a straight movsd. On the other hand, it's important to
791 // shrink long double fp constant since fldt is very slow.
792 return !X86ScalarSSEf64 || VT == MVT::f80;
795 /// Return true if we believe it is correct and profitable to reduce the
796 /// load node to a smaller type.
797 bool shouldReduceLoadWidth(SDNode *Load, ISD::LoadExtType ExtTy,
798 EVT NewVT) const override;
800 /// Return true if the specified scalar FP type is computed in an SSE
801 /// register, not on the X87 floating point stack.
802 bool isScalarFPTypeInSSEReg(EVT VT) const {
803 return (VT == MVT::f64 && X86ScalarSSEf64) || // f64 is when SSE2
804 (VT == MVT::f32 && X86ScalarSSEf32); // f32 is when SSE1
807 /// Return true if the target uses the MSVC _ftol2 routine for fptoui.
808 bool isTargetFTOL() const;
810 /// Return true if the MSVC _ftol2 routine should be used for fptoui to the
812 bool isIntegerTypeFTOL(EVT VT) const {
813 return isTargetFTOL() && VT == MVT::i64;
816 /// \brief Returns true if it is beneficial to convert a load of a constant
817 /// to just the constant itself.
818 bool shouldConvertConstantLoadToIntImm(const APInt &Imm,
819 Type *Ty) const override;
821 /// Return true if EXTRACT_SUBVECTOR is cheap for this result type
823 bool isExtractSubvectorCheap(EVT ResVT, unsigned Index) const override;
825 /// Intel processors have a unified instruction and data cache
826 const char * getClearCacheBuiltinName() const override {
827 return nullptr; // nothing to do, move along.
830 unsigned getRegisterByName(const char* RegName, EVT VT) const override;
832 /// This method returns a target specific FastISel object,
833 /// or null if the target does not support "fast" ISel.
834 FastISel *createFastISel(FunctionLoweringInfo &funcInfo,
835 const TargetLibraryInfo *libInfo) const override;
837 /// Return true if the target stores stack protector cookies at a fixed
838 /// offset in some non-standard address space, and populates the address
839 /// space and offset as appropriate.
840 bool getStackCookieLocation(unsigned &AddressSpace,
841 unsigned &Offset) const override;
843 SDValue BuildFILD(SDValue Op, EVT SrcVT, SDValue Chain, SDValue StackSlot,
844 SelectionDAG &DAG) const;
846 bool isNoopAddrSpaceCast(unsigned SrcAS, unsigned DestAS) const override;
848 bool useLoadStackGuardNode() const override;
849 /// \brief Customize the preferred legalization strategy for certain types.
850 LegalizeTypeAction getPreferredVectorAction(EVT VT) const override;
853 std::pair<const TargetRegisterClass*, uint8_t>
854 findRepresentativeClass(MVT VT) const override;
857 /// Keep a pointer to the X86Subtarget around so that we can
858 /// make the right decision when generating code for different targets.
859 const X86Subtarget *Subtarget;
860 const DataLayout *TD;
862 /// Select between SSE or x87 floating point ops.
863 /// When SSE is available, use it for f32 operations.
864 /// When SSE2 is available, use it for f64 operations.
865 bool X86ScalarSSEf32;
866 bool X86ScalarSSEf64;
868 /// A list of legal FP immediates.
869 std::vector<APFloat> LegalFPImmediates;
871 /// Indicate that this x86 target can instruction
872 /// select the specified FP immediate natively.
873 void addLegalFPImmediate(const APFloat& Imm) {
874 LegalFPImmediates.push_back(Imm);
877 SDValue LowerCallResult(SDValue Chain, SDValue InFlag,
878 CallingConv::ID CallConv, bool isVarArg,
879 const SmallVectorImpl<ISD::InputArg> &Ins,
880 SDLoc dl, SelectionDAG &DAG,
881 SmallVectorImpl<SDValue> &InVals) const;
882 SDValue LowerMemArgument(SDValue Chain,
883 CallingConv::ID CallConv,
884 const SmallVectorImpl<ISD::InputArg> &ArgInfo,
885 SDLoc dl, SelectionDAG &DAG,
886 const CCValAssign &VA, MachineFrameInfo *MFI,
888 SDValue LowerMemOpCallTo(SDValue Chain, SDValue StackPtr, SDValue Arg,
889 SDLoc dl, SelectionDAG &DAG,
890 const CCValAssign &VA,
891 ISD::ArgFlagsTy Flags) const;
893 // Call lowering helpers.
895 /// Check whether the call is eligible for tail call optimization. Targets
896 /// that want to do tail call optimization should implement this function.
897 bool IsEligibleForTailCallOptimization(SDValue Callee,
898 CallingConv::ID CalleeCC,
900 bool isCalleeStructRet,
901 bool isCallerStructRet,
903 const SmallVectorImpl<ISD::OutputArg> &Outs,
904 const SmallVectorImpl<SDValue> &OutVals,
905 const SmallVectorImpl<ISD::InputArg> &Ins,
906 SelectionDAG& DAG) const;
907 bool IsCalleePop(bool isVarArg, CallingConv::ID CallConv) const;
908 SDValue EmitTailCallLoadRetAddr(SelectionDAG &DAG, SDValue &OutRetAddr,
909 SDValue Chain, bool IsTailCall, bool Is64Bit,
910 int FPDiff, SDLoc dl) const;
912 unsigned GetAlignedArgumentStackSize(unsigned StackSize,
913 SelectionDAG &DAG) const;
915 std::pair<SDValue,SDValue> FP_TO_INTHelper(SDValue Op, SelectionDAG &DAG,
917 bool isReplace) const;
919 SDValue LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) const;
920 SDValue LowerBUILD_VECTORvXi1(SDValue Op, SelectionDAG &DAG) const;
921 SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) const;
922 SDValue LowerVSELECT(SDValue Op, SelectionDAG &DAG) const;
923 SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const;
924 SDValue ExtractBitFromMaskVector(SDValue Op, SelectionDAG &DAG) const;
925 SDValue InsertBitToMaskVector(SDValue Op, SelectionDAG &DAG) const;
927 SDValue LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const;
928 SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) const;
929 SDValue LowerBlockAddress(SDValue Op, SelectionDAG &DAG) const;
930 SDValue LowerGlobalAddress(const GlobalValue *GV, SDLoc dl,
931 int64_t Offset, SelectionDAG &DAG) const;
932 SDValue LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) const;
933 SDValue LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const;
934 SDValue LowerExternalSymbol(SDValue Op, SelectionDAG &DAG) const;
935 SDValue LowerSINT_TO_FP(SDValue Op, SelectionDAG &DAG) const;
936 SDValue LowerUINT_TO_FP(SDValue Op, SelectionDAG &DAG) const;
937 SDValue LowerUINT_TO_FP_i64(SDValue Op, SelectionDAG &DAG) const;
938 SDValue LowerUINT_TO_FP_i32(SDValue Op, SelectionDAG &DAG) const;
939 SDValue lowerUINT_TO_FP_vec(SDValue Op, SelectionDAG &DAG) const;
940 SDValue LowerTRUNCATE(SDValue Op, SelectionDAG &DAG) const;
941 SDValue LowerFP_TO_SINT(SDValue Op, SelectionDAG &DAG) const;
942 SDValue LowerFP_TO_UINT(SDValue Op, SelectionDAG &DAG) const;
943 SDValue LowerToBT(SDValue And, ISD::CondCode CC,
944 SDLoc dl, SelectionDAG &DAG) const;
945 SDValue LowerSETCC(SDValue Op, SelectionDAG &DAG) const;
946 SDValue LowerSELECT(SDValue Op, SelectionDAG &DAG) const;
947 SDValue LowerBRCOND(SDValue Op, SelectionDAG &DAG) const;
948 SDValue LowerMEMSET(SDValue Op, SelectionDAG &DAG) const;
949 SDValue LowerJumpTable(SDValue Op, SelectionDAG &DAG) const;
950 SDValue LowerDYNAMIC_STACKALLOC(SDValue Op, SelectionDAG &DAG) const;
951 SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG) const;
952 SDValue LowerVAARG(SDValue Op, SelectionDAG &DAG) const;
953 SDValue LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const;
954 SDValue LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const;
955 SDValue LowerFRAME_TO_ARGS_OFFSET(SDValue Op, SelectionDAG &DAG) const;
956 SDValue LowerEH_RETURN(SDValue Op, SelectionDAG &DAG) const;
957 SDValue lowerEH_SJLJ_SETJMP(SDValue Op, SelectionDAG &DAG) const;
958 SDValue lowerEH_SJLJ_LONGJMP(SDValue Op, SelectionDAG &DAG) const;
959 SDValue LowerINIT_TRAMPOLINE(SDValue Op, SelectionDAG &DAG) const;
960 SDValue LowerFLT_ROUNDS_(SDValue Op, SelectionDAG &DAG) const;
961 SDValue LowerSIGN_EXTEND_INREG(SDValue Op, SelectionDAG &DAG) const;
962 SDValue LowerWin64_i128OP(SDValue Op, SelectionDAG &DAG) const;
965 LowerFormalArguments(SDValue Chain,
966 CallingConv::ID CallConv, bool isVarArg,
967 const SmallVectorImpl<ISD::InputArg> &Ins,
968 SDLoc dl, SelectionDAG &DAG,
969 SmallVectorImpl<SDValue> &InVals) const override;
970 SDValue LowerCall(CallLoweringInfo &CLI,
971 SmallVectorImpl<SDValue> &InVals) const override;
973 SDValue LowerReturn(SDValue Chain,
974 CallingConv::ID CallConv, bool isVarArg,
975 const SmallVectorImpl<ISD::OutputArg> &Outs,
976 const SmallVectorImpl<SDValue> &OutVals,
977 SDLoc dl, SelectionDAG &DAG) const override;
979 bool isUsedByReturnOnly(SDNode *N, SDValue &Chain) const override;
981 bool mayBeEmittedAsTailCall(CallInst *CI) const override;
983 EVT getTypeForExtArgOrReturn(LLVMContext &Context, EVT VT,
984 ISD::NodeType ExtendKind) const override;
986 bool CanLowerReturn(CallingConv::ID CallConv, MachineFunction &MF,
988 const SmallVectorImpl<ISD::OutputArg> &Outs,
989 LLVMContext &Context) const override;
991 const MCPhysReg *getScratchRegisters(CallingConv::ID CC) const override;
993 bool shouldExpandAtomicLoadInIR(LoadInst *SI) const override;
994 bool shouldExpandAtomicStoreInIR(StoreInst *SI) const override;
995 bool shouldExpandAtomicRMWInIR(AtomicRMWInst *AI) const override;
998 lowerIdempotentRMWIntoFencedLoad(AtomicRMWInst *AI) const override;
1000 bool needsCmpXchgNb(const Type *MemType) const;
1002 /// Utility function to emit atomic-load-arith operations (and, or, xor,
1003 /// nand, max, min, umax, umin). It takes the corresponding instruction to
1004 /// expand, the associated machine basic block, and the associated X86
1005 /// opcodes for reg/reg.
1006 MachineBasicBlock *EmitAtomicLoadArith(MachineInstr *MI,
1007 MachineBasicBlock *MBB) const;
1009 /// Utility function to emit atomic-load-arith operations (and, or, xor,
1010 /// nand, add, sub, swap) for 64-bit operands on 32-bit target.
1011 MachineBasicBlock *EmitAtomicLoadArith6432(MachineInstr *MI,
1012 MachineBasicBlock *MBB) const;
1014 // Utility function to emit the low-level va_arg code for X86-64.
1015 MachineBasicBlock *EmitVAARG64WithCustomInserter(
1017 MachineBasicBlock *MBB) const;
1019 /// Utility function to emit the xmm reg save portion of va_start.
1020 MachineBasicBlock *EmitVAStartSaveXMMRegsWithCustomInserter(
1021 MachineInstr *BInstr,
1022 MachineBasicBlock *BB) const;
1024 MachineBasicBlock *EmitLoweredSelect(MachineInstr *I,
1025 MachineBasicBlock *BB) const;
1027 MachineBasicBlock *EmitLoweredWinAlloca(MachineInstr *MI,
1028 MachineBasicBlock *BB) const;
1030 MachineBasicBlock *EmitLoweredSegAlloca(MachineInstr *MI,
1031 MachineBasicBlock *BB) const;
1033 MachineBasicBlock *EmitLoweredTLSCall(MachineInstr *MI,
1034 MachineBasicBlock *BB) const;
1036 MachineBasicBlock *emitLoweredTLSAddr(MachineInstr *MI,
1037 MachineBasicBlock *BB) const;
1039 MachineBasicBlock *emitEHSjLjSetJmp(MachineInstr *MI,
1040 MachineBasicBlock *MBB) const;
1042 MachineBasicBlock *emitEHSjLjLongJmp(MachineInstr *MI,
1043 MachineBasicBlock *MBB) const;
1045 MachineBasicBlock *emitFMA3Instr(MachineInstr *MI,
1046 MachineBasicBlock *MBB) const;
1048 /// Emit nodes that will be selected as "test Op0,Op0", or something
1049 /// equivalent, for use with the given x86 condition code.
1050 SDValue EmitTest(SDValue Op0, unsigned X86CC, SDLoc dl,
1051 SelectionDAG &DAG) const;
1053 /// Emit nodes that will be selected as "cmp Op0,Op1", or something
1054 /// equivalent, for use with the given x86 condition code.
1055 SDValue EmitCmp(SDValue Op0, SDValue Op1, unsigned X86CC, SDLoc dl,
1056 SelectionDAG &DAG) const;
1058 /// Convert a comparison if required by the subtarget.
1059 SDValue ConvertCmpIfNecessary(SDValue Cmp, SelectionDAG &DAG) const;
1061 /// Use rsqrt* to speed up sqrt calculations.
1062 SDValue getRsqrtEstimate(SDValue Operand, DAGCombinerInfo &DCI,
1063 unsigned &RefinementSteps,
1064 bool &UseOneConstNR) const override;
1066 /// Use rcp* to speed up fdiv calculations.
1067 SDValue getRecipEstimate(SDValue Operand, DAGCombinerInfo &DCI,
1068 unsigned &RefinementSteps) const override;
1072 FastISel *createFastISel(FunctionLoweringInfo &funcInfo,
1073 const TargetLibraryInfo *libInfo);
1077 #endif // X86ISELLOWERING_H