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
398 // Compress and expand
402 // Save xmm argument registers to the stack, according to %al. An operator
403 // is needed so that this can be expanded with control flow.
404 VASTART_SAVE_XMM_REGS,
406 // Windows's _chkstk call to do stack probing.
409 // For allocating variable amounts of stack space when using
410 // segmented stacks. Check if the current stacklet has enough space, and
411 // falls back to heap allocation if not.
414 // Windows's _ftol2 runtime routine to do fptoui.
423 // Store FP status word into i16 register.
426 // Store contents of %ah into %eflags.
429 // Get a random integer and indicate whether it is valid in CF.
432 // Get a NIST SP800-90B & C compliant random integer and
433 // indicate whether it is valid in CF.
439 // Test if in transactional execution.
443 RSQRT28, RCP28, EXP2,
446 LCMPXCHG_DAG = ISD::FIRST_TARGET_MEMORY_OPCODE,
450 // Load, scalar_to_vector, and zero extend.
453 // Store FP control world into i16 memory.
456 /// This instruction implements FP_TO_SINT with the
457 /// integer destination in memory and a FP reg source. This corresponds
458 /// to the X86::FIST*m instructions and the rounding mode change stuff. It
459 /// has two inputs (token chain and address) and two outputs (int value
460 /// and token chain).
465 /// This instruction implements SINT_TO_FP with the
466 /// integer source in memory and FP reg result. This corresponds to the
467 /// X86::FILD*m instructions. It has three inputs (token chain, address,
468 /// and source type) and two outputs (FP value and token chain). FILD_FLAG
469 /// also produces a flag).
473 /// This instruction implements an extending load to FP stack slots.
474 /// This corresponds to the X86::FLD32m / X86::FLD64m. It takes a chain
475 /// operand, ptr to load from, and a ValueType node indicating the type
479 /// This instruction implements a truncating store to FP stack
480 /// slots. This corresponds to the X86::FST32m / X86::FST64m. It takes a
481 /// chain operand, value to store, address, and a ValueType to store it
485 /// This instruction grabs the address of the next argument
486 /// from a va_list. (reads and modifies the va_list in memory)
489 // WARNING: Do not add anything in the end unless you want the node to
490 // have memop! In fact, starting from ATOMADD64_DAG all opcodes will be
491 // thought as target memory ops!
495 /// Define some predicates that are used for node matching.
497 /// Return true if the specified
498 /// EXTRACT_SUBVECTOR operand specifies a vector extract that is
499 /// suitable for input to VEXTRACTF128, VEXTRACTI128 instructions.
500 bool isVEXTRACT128Index(SDNode *N);
502 /// Return true if the specified
503 /// INSERT_SUBVECTOR operand specifies a subvector insert that is
504 /// suitable for input to VINSERTF128, VINSERTI128 instructions.
505 bool isVINSERT128Index(SDNode *N);
507 /// Return true if the specified
508 /// EXTRACT_SUBVECTOR operand specifies a vector extract that is
509 /// suitable for input to VEXTRACTF64X4, VEXTRACTI64X4 instructions.
510 bool isVEXTRACT256Index(SDNode *N);
512 /// Return true if the specified
513 /// INSERT_SUBVECTOR operand specifies a subvector insert that is
514 /// suitable for input to VINSERTF64X4, VINSERTI64X4 instructions.
515 bool isVINSERT256Index(SDNode *N);
517 /// Return the appropriate
518 /// immediate to extract the specified EXTRACT_SUBVECTOR index
519 /// with VEXTRACTF128, VEXTRACTI128 instructions.
520 unsigned getExtractVEXTRACT128Immediate(SDNode *N);
522 /// Return the appropriate
523 /// immediate to insert at the specified INSERT_SUBVECTOR index
524 /// with VINSERTF128, VINSERT128 instructions.
525 unsigned getInsertVINSERT128Immediate(SDNode *N);
527 /// Return the appropriate
528 /// immediate to extract the specified EXTRACT_SUBVECTOR index
529 /// with VEXTRACTF64X4, VEXTRACTI64x4 instructions.
530 unsigned getExtractVEXTRACT256Immediate(SDNode *N);
532 /// Return the appropriate
533 /// immediate to insert at the specified INSERT_SUBVECTOR index
534 /// with VINSERTF64x4, VINSERTI64x4 instructions.
535 unsigned getInsertVINSERT256Immediate(SDNode *N);
537 /// Returns true if Elt is a constant zero or floating point constant +0.0.
538 bool isZeroNode(SDValue Elt);
540 /// Returns true of the given offset can be
541 /// fit into displacement field of the instruction.
542 bool isOffsetSuitableForCodeModel(int64_t Offset, CodeModel::Model M,
543 bool hasSymbolicDisplacement = true);
546 /// Determines whether the callee is required to pop its
547 /// own arguments. Callee pop is necessary to support tail calls.
548 bool isCalleePop(CallingConv::ID CallingConv,
549 bool is64Bit, bool IsVarArg, bool TailCallOpt);
551 /// AVX512 static rounding constants. These need to match the values in
553 enum STATIC_ROUNDING {
562 //===--------------------------------------------------------------------===//
563 // X86 Implementation of the TargetLowering interface
564 class X86TargetLowering final : public TargetLowering {
566 explicit X86TargetLowering(const X86TargetMachine &TM,
567 const X86Subtarget &STI);
569 unsigned getJumpTableEncoding() const override;
571 MVT getScalarShiftAmountTy(EVT LHSTy) const override { return MVT::i8; }
574 LowerCustomJumpTableEntry(const MachineJumpTableInfo *MJTI,
575 const MachineBasicBlock *MBB, unsigned uid,
576 MCContext &Ctx) const override;
578 /// Returns relocation base for the given PIC jumptable.
579 SDValue getPICJumpTableRelocBase(SDValue Table,
580 SelectionDAG &DAG) const override;
582 getPICJumpTableRelocBaseExpr(const MachineFunction *MF,
583 unsigned JTI, MCContext &Ctx) const override;
585 /// Return the desired alignment for ByVal aggregate
586 /// function arguments in the caller parameter area. For X86, aggregates
587 /// that contains are placed at 16-byte boundaries while the rest are at
588 /// 4-byte boundaries.
589 unsigned getByValTypeAlignment(Type *Ty) const override;
591 /// Returns the target specific optimal type for load
592 /// and store operations as a result of memset, memcpy, and memmove
593 /// lowering. If DstAlign is zero that means it's safe to destination
594 /// alignment can satisfy any constraint. Similarly if SrcAlign is zero it
595 /// means there isn't a need to check it against alignment requirement,
596 /// probably because the source does not need to be loaded. If 'IsMemset' is
597 /// true, that means it's expanding a memset. If 'ZeroMemset' is true, that
598 /// means it's a memset of zero. 'MemcpyStrSrc' indicates whether the memcpy
599 /// source is constant so it does not need to be loaded.
600 /// It returns EVT::Other if the type should be determined using generic
601 /// target-independent logic.
602 EVT getOptimalMemOpType(uint64_t Size, unsigned DstAlign, unsigned SrcAlign,
603 bool IsMemset, bool ZeroMemset, bool MemcpyStrSrc,
604 MachineFunction &MF) const override;
606 /// Returns true if it's safe to use load / store of the
607 /// specified type to expand memcpy / memset inline. This is mostly true
608 /// for all types except for some special cases. For example, on X86
609 /// targets without SSE2 f64 load / store are done with fldl / fstpl which
610 /// also does type conversion. Note the specified type doesn't have to be
611 /// legal as the hook is used before type legalization.
612 bool isSafeMemOpType(MVT VT) const override;
614 /// Returns true if the target allows
615 /// unaligned memory accesses. of the specified type. Returns whether it
616 /// is "fast" by reference in the second argument.
617 bool allowsMisalignedMemoryAccesses(EVT VT, unsigned AS, unsigned Align,
618 bool *Fast) const override;
620 /// Provide custom lowering hooks for some operations.
622 SDValue LowerOperation(SDValue Op, SelectionDAG &DAG) const override;
624 /// Replace the results of node with an illegal result
625 /// type with new values built out of custom code.
627 void ReplaceNodeResults(SDNode *N, SmallVectorImpl<SDValue>&Results,
628 SelectionDAG &DAG) const override;
631 SDValue PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const override;
633 /// Return true if the target has native support for
634 /// the specified value type and it is 'desirable' to use the type for the
635 /// given node type. e.g. On x86 i16 is legal, but undesirable since i16
636 /// instruction encodings are longer and some i16 instructions are slow.
637 bool isTypeDesirableForOp(unsigned Opc, EVT VT) const override;
639 /// Return true if the target has native support for the
640 /// specified value type and it is 'desirable' to use the type. e.g. On x86
641 /// i16 is legal, but undesirable since i16 instruction encodings are longer
642 /// and some i16 instructions are slow.
643 bool IsDesirableToPromoteOp(SDValue Op, EVT &PVT) const override;
646 EmitInstrWithCustomInserter(MachineInstr *MI,
647 MachineBasicBlock *MBB) const override;
650 /// This method returns the name of a target specific DAG node.
651 const char *getTargetNodeName(unsigned Opcode) const override;
653 bool isCheapToSpeculateCttz() const override;
655 bool isCheapToSpeculateCtlz() const override;
657 /// Return the value type to use for ISD::SETCC.
658 EVT getSetCCResultType(LLVMContext &Context, EVT VT) const override;
660 /// Determine which of the bits specified in Mask are known to be either
661 /// zero or one and return them in the KnownZero/KnownOne bitsets.
662 void computeKnownBitsForTargetNode(const SDValue Op,
665 const SelectionDAG &DAG,
666 unsigned Depth = 0) const override;
668 /// Determine the number of bits in the operation that are sign bits.
669 unsigned ComputeNumSignBitsForTargetNode(SDValue Op,
670 const SelectionDAG &DAG,
671 unsigned Depth) const override;
673 bool isGAPlusOffset(SDNode *N, const GlobalValue* &GA,
674 int64_t &Offset) const override;
676 SDValue getReturnAddressFrameIndex(SelectionDAG &DAG) const;
678 bool ExpandInlineAsm(CallInst *CI) const override;
681 getConstraintType(const std::string &Constraint) const override;
683 /// Examine constraint string and operand type and determine a weight value.
684 /// The operand object must already have been set up with the operand type.
686 getSingleConstraintMatchWeight(AsmOperandInfo &info,
687 const char *constraint) const override;
689 const char *LowerXConstraint(EVT ConstraintVT) const override;
691 /// Lower the specified operand into the Ops vector. If it is invalid, don't
692 /// add anything to Ops. If hasMemory is true it means one of the asm
693 /// constraint of the inline asm instruction being processed is 'm'.
694 void LowerAsmOperandForConstraint(SDValue Op,
695 std::string &Constraint,
696 std::vector<SDValue> &Ops,
697 SelectionDAG &DAG) const override;
699 /// Given a physical register constraint
700 /// (e.g. {edx}), return the register number and the register class for the
701 /// register. This should only be used for C_Register constraints. On
702 /// error, this returns a register number of 0.
703 std::pair<unsigned, const TargetRegisterClass*>
704 getRegForInlineAsmConstraint(const std::string &Constraint,
705 MVT VT) const override;
707 /// Return true if the addressing mode represented
708 /// by AM is legal for this target, for a load/store of the specified type.
709 bool isLegalAddressingMode(const AddrMode &AM, Type *Ty) const override;
711 /// Return true if the specified immediate is legal
712 /// icmp immediate, that is the target has icmp instructions which can
713 /// compare a register against the immediate without having to materialize
714 /// the immediate into a register.
715 bool isLegalICmpImmediate(int64_t Imm) const override;
717 /// Return true if the specified immediate is legal
718 /// add immediate, that is the target has add instructions which can
719 /// add a register and the immediate without having to materialize
720 /// the immediate into a register.
721 bool isLegalAddImmediate(int64_t Imm) const override;
723 /// \brief Return the cost of the scaling factor used in the addressing
724 /// mode represented by AM for this target, for a load/store
725 /// of the specified type.
726 /// If the AM is supported, the return value must be >= 0.
727 /// If the AM is not supported, it returns a negative value.
728 int getScalingFactorCost(const AddrMode &AM, Type *Ty) const override;
730 bool isVectorShiftByScalarCheap(Type *Ty) const override;
732 /// Return true if it's free to truncate a value of
733 /// type Ty1 to type Ty2. e.g. On x86 it's free to truncate a i32 value in
734 /// register EAX to i16 by referencing its sub-register AX.
735 bool isTruncateFree(Type *Ty1, Type *Ty2) const override;
736 bool isTruncateFree(EVT VT1, EVT VT2) const override;
738 bool allowTruncateForTailCall(Type *Ty1, Type *Ty2) const override;
740 /// Return true if any actual instruction that defines a
741 /// value of type Ty1 implicit zero-extends the value to Ty2 in the result
742 /// register. This does not necessarily include registers defined in
743 /// unknown ways, such as incoming arguments, or copies from unknown
744 /// virtual registers. Also, if isTruncateFree(Ty2, Ty1) is true, this
745 /// does not necessarily apply to truncate instructions. e.g. on x86-64,
746 /// all instructions that define 32-bit values implicit zero-extend the
747 /// result out to 64 bits.
748 bool isZExtFree(Type *Ty1, Type *Ty2) const override;
749 bool isZExtFree(EVT VT1, EVT VT2) const override;
750 bool isZExtFree(SDValue Val, EVT VT2) const override;
752 /// Return true if folding a vector load into ExtVal (a sign, zero, or any
753 /// extend node) is profitable.
754 bool isVectorLoadExtDesirable(SDValue) const override;
756 /// Return true if an FMA operation is faster than a pair of fmul and fadd
757 /// instructions. fmuladd intrinsics will be expanded to FMAs when this
758 /// method returns true, otherwise fmuladd is expanded to fmul + fadd.
759 bool isFMAFasterThanFMulAndFAdd(EVT VT) const override;
761 /// Return true if it's profitable to narrow
762 /// operations of type VT1 to VT2. e.g. on x86, it's profitable to narrow
763 /// from i32 to i8 but not from i32 to i16.
764 bool isNarrowingProfitable(EVT VT1, EVT VT2) const override;
766 /// Returns true if the target can instruction select the
767 /// specified FP immediate natively. If false, the legalizer will
768 /// materialize the FP immediate as a load from a constant pool.
769 bool isFPImmLegal(const APFloat &Imm, EVT VT) const override;
771 /// Targets can use this to indicate that they only support *some*
772 /// VECTOR_SHUFFLE operations, those with specific masks. By default, if a
773 /// target supports the VECTOR_SHUFFLE node, all mask values are assumed to
775 bool isShuffleMaskLegal(const SmallVectorImpl<int> &Mask,
776 EVT VT) const override;
778 /// Similar to isShuffleMaskLegal. This is used by Targets can use this to
779 /// indicate if there is a suitable VECTOR_SHUFFLE that can be used to
780 /// replace a VAND with a constant pool entry.
781 bool isVectorClearMaskLegal(const SmallVectorImpl<int> &Mask,
782 EVT VT) const override;
784 /// If true, then instruction selection should
785 /// seek to shrink the FP constant of the specified type to a smaller type
786 /// in order to save space and / or reduce runtime.
787 bool ShouldShrinkFPConstant(EVT VT) const override {
788 // Don't shrink FP constpool if SSE2 is available since cvtss2sd is more
789 // expensive than a straight movsd. On the other hand, it's important to
790 // shrink long double fp constant since fldt is very slow.
791 return !X86ScalarSSEf64 || VT == MVT::f80;
794 /// Return true if we believe it is correct and profitable to reduce the
795 /// load node to a smaller type.
796 bool shouldReduceLoadWidth(SDNode *Load, ISD::LoadExtType ExtTy,
797 EVT NewVT) const override;
799 /// Return true if the specified scalar FP type is computed in an SSE
800 /// register, not on the X87 floating point stack.
801 bool isScalarFPTypeInSSEReg(EVT VT) const {
802 return (VT == MVT::f64 && X86ScalarSSEf64) || // f64 is when SSE2
803 (VT == MVT::f32 && X86ScalarSSEf32); // f32 is when SSE1
806 /// Return true if the target uses the MSVC _ftol2 routine for fptoui.
807 bool isTargetFTOL() const;
809 /// Return true if the MSVC _ftol2 routine should be used for fptoui to the
811 bool isIntegerTypeFTOL(EVT VT) const {
812 return isTargetFTOL() && VT == MVT::i64;
815 /// \brief Returns true if it is beneficial to convert a load of a constant
816 /// to just the constant itself.
817 bool shouldConvertConstantLoadToIntImm(const APInt &Imm,
818 Type *Ty) const override;
820 /// Return true if EXTRACT_SUBVECTOR is cheap for this result type
822 bool isExtractSubvectorCheap(EVT ResVT, unsigned Index) const override;
824 /// Intel processors have a unified instruction and data cache
825 const char * getClearCacheBuiltinName() const override {
826 return nullptr; // nothing to do, move along.
829 unsigned getRegisterByName(const char* RegName, EVT VT) const override;
831 /// This method returns a target specific FastISel object,
832 /// or null if the target does not support "fast" ISel.
833 FastISel *createFastISel(FunctionLoweringInfo &funcInfo,
834 const TargetLibraryInfo *libInfo) const override;
836 /// Return true if the target stores stack protector cookies at a fixed
837 /// offset in some non-standard address space, and populates the address
838 /// space and offset as appropriate.
839 bool getStackCookieLocation(unsigned &AddressSpace,
840 unsigned &Offset) const override;
842 SDValue BuildFILD(SDValue Op, EVT SrcVT, SDValue Chain, SDValue StackSlot,
843 SelectionDAG &DAG) const;
845 bool isNoopAddrSpaceCast(unsigned SrcAS, unsigned DestAS) const override;
847 bool useLoadStackGuardNode() const override;
848 /// \brief Customize the preferred legalization strategy for certain types.
849 LegalizeTypeAction getPreferredVectorAction(EVT VT) const override;
852 std::pair<const TargetRegisterClass*, uint8_t>
853 findRepresentativeClass(MVT VT) const override;
856 /// Keep a pointer to the X86Subtarget around so that we can
857 /// make the right decision when generating code for different targets.
858 const X86Subtarget *Subtarget;
859 const DataLayout *TD;
861 /// Select between SSE or x87 floating point ops.
862 /// When SSE is available, use it for f32 operations.
863 /// When SSE2 is available, use it for f64 operations.
864 bool X86ScalarSSEf32;
865 bool X86ScalarSSEf64;
867 /// A list of legal FP immediates.
868 std::vector<APFloat> LegalFPImmediates;
870 /// Indicate that this x86 target can instruction
871 /// select the specified FP immediate natively.
872 void addLegalFPImmediate(const APFloat& Imm) {
873 LegalFPImmediates.push_back(Imm);
876 SDValue LowerCallResult(SDValue Chain, SDValue InFlag,
877 CallingConv::ID CallConv, bool isVarArg,
878 const SmallVectorImpl<ISD::InputArg> &Ins,
879 SDLoc dl, SelectionDAG &DAG,
880 SmallVectorImpl<SDValue> &InVals) const;
881 SDValue LowerMemArgument(SDValue Chain,
882 CallingConv::ID CallConv,
883 const SmallVectorImpl<ISD::InputArg> &ArgInfo,
884 SDLoc dl, SelectionDAG &DAG,
885 const CCValAssign &VA, MachineFrameInfo *MFI,
887 SDValue LowerMemOpCallTo(SDValue Chain, SDValue StackPtr, SDValue Arg,
888 SDLoc dl, SelectionDAG &DAG,
889 const CCValAssign &VA,
890 ISD::ArgFlagsTy Flags) const;
892 // Call lowering helpers.
894 /// Check whether the call is eligible for tail call optimization. Targets
895 /// that want to do tail call optimization should implement this function.
896 bool IsEligibleForTailCallOptimization(SDValue Callee,
897 CallingConv::ID CalleeCC,
899 bool isCalleeStructRet,
900 bool isCallerStructRet,
902 const SmallVectorImpl<ISD::OutputArg> &Outs,
903 const SmallVectorImpl<SDValue> &OutVals,
904 const SmallVectorImpl<ISD::InputArg> &Ins,
905 SelectionDAG& DAG) const;
906 bool IsCalleePop(bool isVarArg, CallingConv::ID CallConv) const;
907 SDValue EmitTailCallLoadRetAddr(SelectionDAG &DAG, SDValue &OutRetAddr,
908 SDValue Chain, bool IsTailCall, bool Is64Bit,
909 int FPDiff, SDLoc dl) const;
911 unsigned GetAlignedArgumentStackSize(unsigned StackSize,
912 SelectionDAG &DAG) const;
914 std::pair<SDValue,SDValue> FP_TO_INTHelper(SDValue Op, SelectionDAG &DAG,
916 bool isReplace) const;
918 SDValue LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) const;
919 SDValue LowerBUILD_VECTORvXi1(SDValue Op, SelectionDAG &DAG) const;
920 SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) const;
921 SDValue LowerVSELECT(SDValue Op, SelectionDAG &DAG) const;
922 SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const;
923 SDValue ExtractBitFromMaskVector(SDValue Op, SelectionDAG &DAG) const;
924 SDValue InsertBitToMaskVector(SDValue Op, SelectionDAG &DAG) const;
926 SDValue LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const;
927 SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) const;
928 SDValue LowerBlockAddress(SDValue Op, SelectionDAG &DAG) const;
929 SDValue LowerGlobalAddress(const GlobalValue *GV, SDLoc dl,
930 int64_t Offset, SelectionDAG &DAG) const;
931 SDValue LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) const;
932 SDValue LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const;
933 SDValue LowerExternalSymbol(SDValue Op, SelectionDAG &DAG) const;
934 SDValue LowerSINT_TO_FP(SDValue Op, SelectionDAG &DAG) const;
935 SDValue LowerUINT_TO_FP(SDValue Op, SelectionDAG &DAG) const;
936 SDValue LowerUINT_TO_FP_i64(SDValue Op, SelectionDAG &DAG) const;
937 SDValue LowerUINT_TO_FP_i32(SDValue Op, SelectionDAG &DAG) const;
938 SDValue lowerUINT_TO_FP_vec(SDValue Op, SelectionDAG &DAG) const;
939 SDValue LowerTRUNCATE(SDValue Op, SelectionDAG &DAG) const;
940 SDValue LowerFP_TO_SINT(SDValue Op, SelectionDAG &DAG) const;
941 SDValue LowerFP_TO_UINT(SDValue Op, SelectionDAG &DAG) const;
942 SDValue LowerToBT(SDValue And, ISD::CondCode CC,
943 SDLoc dl, SelectionDAG &DAG) const;
944 SDValue LowerSETCC(SDValue Op, SelectionDAG &DAG) const;
945 SDValue LowerSELECT(SDValue Op, SelectionDAG &DAG) const;
946 SDValue LowerBRCOND(SDValue Op, SelectionDAG &DAG) const;
947 SDValue LowerMEMSET(SDValue Op, SelectionDAG &DAG) const;
948 SDValue LowerJumpTable(SDValue Op, SelectionDAG &DAG) const;
949 SDValue LowerDYNAMIC_STACKALLOC(SDValue Op, SelectionDAG &DAG) const;
950 SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG) const;
951 SDValue LowerVAARG(SDValue Op, SelectionDAG &DAG) const;
952 SDValue LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const;
953 SDValue LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const;
954 SDValue LowerFRAME_TO_ARGS_OFFSET(SDValue Op, SelectionDAG &DAG) const;
955 SDValue LowerEH_RETURN(SDValue Op, SelectionDAG &DAG) const;
956 SDValue lowerEH_SJLJ_SETJMP(SDValue Op, SelectionDAG &DAG) const;
957 SDValue lowerEH_SJLJ_LONGJMP(SDValue Op, SelectionDAG &DAG) const;
958 SDValue LowerINIT_TRAMPOLINE(SDValue Op, SelectionDAG &DAG) const;
959 SDValue LowerFLT_ROUNDS_(SDValue Op, SelectionDAG &DAG) const;
960 SDValue LowerSIGN_EXTEND_INREG(SDValue Op, SelectionDAG &DAG) const;
961 SDValue LowerWin64_i128OP(SDValue Op, SelectionDAG &DAG) const;
964 LowerFormalArguments(SDValue Chain,
965 CallingConv::ID CallConv, bool isVarArg,
966 const SmallVectorImpl<ISD::InputArg> &Ins,
967 SDLoc dl, SelectionDAG &DAG,
968 SmallVectorImpl<SDValue> &InVals) const override;
969 SDValue LowerCall(CallLoweringInfo &CLI,
970 SmallVectorImpl<SDValue> &InVals) const override;
972 SDValue LowerReturn(SDValue Chain,
973 CallingConv::ID CallConv, bool isVarArg,
974 const SmallVectorImpl<ISD::OutputArg> &Outs,
975 const SmallVectorImpl<SDValue> &OutVals,
976 SDLoc dl, SelectionDAG &DAG) const override;
978 bool isUsedByReturnOnly(SDNode *N, SDValue &Chain) const override;
980 bool mayBeEmittedAsTailCall(CallInst *CI) const override;
982 EVT getTypeForExtArgOrReturn(LLVMContext &Context, EVT VT,
983 ISD::NodeType ExtendKind) const override;
985 bool CanLowerReturn(CallingConv::ID CallConv, MachineFunction &MF,
987 const SmallVectorImpl<ISD::OutputArg> &Outs,
988 LLVMContext &Context) const override;
990 const MCPhysReg *getScratchRegisters(CallingConv::ID CC) const override;
992 bool shouldExpandAtomicLoadInIR(LoadInst *SI) const override;
993 bool shouldExpandAtomicStoreInIR(StoreInst *SI) const override;
994 bool shouldExpandAtomicRMWInIR(AtomicRMWInst *AI) const override;
997 lowerIdempotentRMWIntoFencedLoad(AtomicRMWInst *AI) const override;
999 bool needsCmpXchgNb(const Type *MemType) const;
1001 /// Utility function to emit atomic-load-arith operations (and, or, xor,
1002 /// nand, max, min, umax, umin). It takes the corresponding instruction to
1003 /// expand, the associated machine basic block, and the associated X86
1004 /// opcodes for reg/reg.
1005 MachineBasicBlock *EmitAtomicLoadArith(MachineInstr *MI,
1006 MachineBasicBlock *MBB) const;
1008 /// Utility function to emit atomic-load-arith operations (and, or, xor,
1009 /// nand, add, sub, swap) for 64-bit operands on 32-bit target.
1010 MachineBasicBlock *EmitAtomicLoadArith6432(MachineInstr *MI,
1011 MachineBasicBlock *MBB) const;
1013 // Utility function to emit the low-level va_arg code for X86-64.
1014 MachineBasicBlock *EmitVAARG64WithCustomInserter(
1016 MachineBasicBlock *MBB) const;
1018 /// Utility function to emit the xmm reg save portion of va_start.
1019 MachineBasicBlock *EmitVAStartSaveXMMRegsWithCustomInserter(
1020 MachineInstr *BInstr,
1021 MachineBasicBlock *BB) const;
1023 MachineBasicBlock *EmitLoweredSelect(MachineInstr *I,
1024 MachineBasicBlock *BB) const;
1026 MachineBasicBlock *EmitLoweredWinAlloca(MachineInstr *MI,
1027 MachineBasicBlock *BB) const;
1029 MachineBasicBlock *EmitLoweredSegAlloca(MachineInstr *MI,
1030 MachineBasicBlock *BB) const;
1032 MachineBasicBlock *EmitLoweredTLSCall(MachineInstr *MI,
1033 MachineBasicBlock *BB) const;
1035 MachineBasicBlock *emitLoweredTLSAddr(MachineInstr *MI,
1036 MachineBasicBlock *BB) const;
1038 MachineBasicBlock *emitEHSjLjSetJmp(MachineInstr *MI,
1039 MachineBasicBlock *MBB) const;
1041 MachineBasicBlock *emitEHSjLjLongJmp(MachineInstr *MI,
1042 MachineBasicBlock *MBB) const;
1044 MachineBasicBlock *emitFMA3Instr(MachineInstr *MI,
1045 MachineBasicBlock *MBB) const;
1047 /// Emit nodes that will be selected as "test Op0,Op0", or something
1048 /// equivalent, for use with the given x86 condition code.
1049 SDValue EmitTest(SDValue Op0, unsigned X86CC, SDLoc dl,
1050 SelectionDAG &DAG) const;
1052 /// Emit nodes that will be selected as "cmp Op0,Op1", or something
1053 /// equivalent, for use with the given x86 condition code.
1054 SDValue EmitCmp(SDValue Op0, SDValue Op1, unsigned X86CC, SDLoc dl,
1055 SelectionDAG &DAG) const;
1057 /// Convert a comparison if required by the subtarget.
1058 SDValue ConvertCmpIfNecessary(SDValue Cmp, SelectionDAG &DAG) const;
1060 /// Use rsqrt* to speed up sqrt calculations.
1061 SDValue getRsqrtEstimate(SDValue Operand, DAGCombinerInfo &DCI,
1062 unsigned &RefinementSteps,
1063 bool &UseOneConstNR) const override;
1065 /// Use rcp* to speed up fdiv calculations.
1066 SDValue getRecipEstimate(SDValue Operand, DAGCombinerInfo &DCI,
1067 unsigned &RefinementSteps) const override;
1071 FastISel *createFastISel(FunctionLoweringInfo &funcInfo,
1072 const TargetLibraryInfo *libInfo);
1076 #endif // X86ISELLOWERING_H