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 X86ISELLOWERING_H
16 #define X86ISELLOWERING_H
18 #include "X86Subtarget.h"
19 #include "llvm/CodeGen/CallingConvLower.h"
20 #include "llvm/CodeGen/SelectionDAG.h"
21 #include "llvm/Target/TargetLowering.h"
22 #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 compare and logical compare instructions.
89 /// X86 bit-test instructions.
92 /// X86 SetCC. Operand 0 is condition code, and operand 1 is the EFLAGS
93 /// operand, usually produced by a CMP instruction.
99 // Same as SETCC except it's materialized with a sbb and the value is all
100 // one's or all zero's.
101 SETCC_CARRY, // R = carry_bit ? ~0 : 0
103 /// X86 FP SETCC, implemented with CMP{cc}SS/CMP{cc}SD.
104 /// Operands are two FP values to compare; result is a mask of
105 /// 0s or 1s. Generally DTRT for C/C++ with NaNs.
108 /// X86 MOVMSK{pd|ps}, extracts sign bits of two or four FP values,
109 /// result in an integer GPR. Needs masking for scalar result.
112 /// X86 conditional moves. Operand 0 and operand 1 are the two values
113 /// to select from. Operand 2 is the condition code, and operand 3 is the
114 /// flag operand produced by a CMP or TEST instruction. It also writes a
118 /// X86 conditional branches. Operand 0 is the chain operand, operand 1
119 /// is the block to branch if condition is true, operand 2 is the
120 /// condition code, and operand 3 is the flag operand produced by a CMP
121 /// or TEST instruction.
124 /// Return with a flag operand. Operand 0 is the chain operand, operand
125 /// 1 is the number of bytes of stack to pop.
128 /// REP_STOS - Repeat fill, corresponds to X86::REP_STOSx.
131 /// REP_MOVS - Repeat move, corresponds to X86::REP_MOVSx.
134 /// GlobalBaseReg - On Darwin, this node represents the result of the popl
135 /// at function entry, used for PIC code.
138 /// Wrapper - A wrapper node for TargetConstantPool,
139 /// TargetExternalSymbol, and TargetGlobalAddress.
142 /// WrapperRIP - Special wrapper used under X86-64 PIC mode for RIP
143 /// relative displacements.
146 /// MOVDQ2Q - Copies a 64-bit value from the low word of an XMM vector
147 /// to an MMX vector. If you think this is too close to the previous
148 /// mnemonic, so do I; blame Intel.
151 /// MMX_MOVD2W - Copies a 32-bit value from the low word of a MMX
155 /// PEXTRB - Extract an 8-bit value from a vector and zero extend it to
156 /// i32, corresponds to X86::PEXTRB.
159 /// PEXTRW - Extract a 16-bit value from a vector and zero extend it to
160 /// i32, corresponds to X86::PEXTRW.
163 /// INSERTPS - Insert any element of a 4 x float vector into any element
164 /// of a destination 4 x floatvector.
167 /// PINSRB - Insert the lower 8-bits of a 32-bit value to a vector,
168 /// corresponds to X86::PINSRB.
171 /// PINSRW - Insert the lower 16-bits of a 32-bit value to a vector,
172 /// corresponds to X86::PINSRW.
175 /// PSHUFB - Shuffle 16 8-bit values within a vector.
178 /// ANDNP - Bitwise Logical AND NOT of Packed FP values.
181 /// PSIGN - Copy integer sign.
184 /// BLENDV - Blend where the selector is a register.
187 /// BLENDI - Blend where the selector is an immediate.
190 // SUBUS - Integer sub with unsigned saturation.
193 /// HADD - Integer horizontal add.
196 /// HSUB - Integer horizontal sub.
199 /// FHADD - Floating point horizontal add.
202 /// FHSUB - Floating point horizontal sub.
205 /// UMAX, UMIN - Unsigned integer max and min.
208 /// SMAX, SMIN - Signed integer max and min.
211 /// FMAX, FMIN - Floating point max and min.
215 /// FMAXC, FMINC - Commutative FMIN and FMAX.
218 /// FRSQRT, FRCP - Floating point reciprocal-sqrt and reciprocal
219 /// approximation. Note that these typically require refinement
220 /// in order to obtain suitable precision.
223 // TLSADDR - Thread Local Storage.
226 // TLSBASEADDR - Thread Local Storage. A call to get the start address
227 // of the TLS block for the current module.
230 // TLSCALL - Thread Local Storage. When calling to an OS provided
231 // thunk at the address from an earlier relocation.
234 // EH_RETURN - Exception Handling helpers.
237 // EH_SJLJ_SETJMP - SjLj exception handling setjmp.
240 // EH_SJLJ_LONGJMP - SjLj exception handling longjmp.
243 /// TC_RETURN - Tail call return. See X86TargetLowering::LowerCall for
244 /// the list of operands.
247 // VZEXT_MOVL - Vector move to low scalar and zero higher vector elements.
250 // VZEXT - Vector integer zero-extend.
253 // VSEXT - Vector integer signed-extend.
256 // VTRUNC - Vector integer truncate.
259 // VTRUNC - Vector integer truncate with mask.
262 // VFPEXT - Vector FP extend.
265 // VFPROUND - Vector FP round.
268 // VSHL, VSRL - 128-bit vector logical left / right shift
271 // VSHL, VSRL, VSRA - Vector shift elements
274 // VSHLI, VSRLI, VSRAI - Vector shift elements by immediate
277 // CMPP - Vector packed double/float comparison.
280 // PCMP* - Vector integer comparisons.
282 // PCMP*M - Vector integer comparisons, the result is in a mask vector.
285 /// CMPM, CMPMU - Vector comparison generating mask bits for fp and
286 /// integer signed and unsigned data types.
290 // ADD, SUB, SMUL, etc. - Arithmetic operations with FLAGS results.
291 ADD, SUB, ADC, SBB, SMUL,
292 INC, DEC, OR, XOR, AND,
294 BZHI, // BZHI - Zero high bits
295 BEXTR, // BEXTR - Bit field extract
297 UMUL, // LOW, HI, FLAGS = umul LHS, RHS
299 // MUL_IMM - X86 specific multiply by immediate.
302 // PTEST - Vector bitwise comparisons.
305 // TESTP - Vector packed fp sign bitwise comparisons.
308 // TESTM, TESTNM - Vector "test" in AVX-512, the result is in a mask vector.
312 // OR/AND test for masks
315 // Several flavors of instructions with vector shuffle behaviors.
342 // Insert/Extract vector element
346 // PMULUDQ - Vector multiply packed unsigned doubleword integers
357 // VASTART_SAVE_XMM_REGS - Save xmm argument registers to the stack,
358 // according to %al. An operator is needed so that this can be expanded
359 // with control flow.
360 VASTART_SAVE_XMM_REGS,
362 // WIN_ALLOCA - Windows's _chkstk call to do stack probing.
365 // SEG_ALLOCA - For allocating variable amounts of stack space when using
366 // segmented stacks. Check if the current stacklet has enough space, and
367 // falls back to heap allocation if not.
370 // WIN_FTOL - Windows's _ftol2 runtime routine to do fptoui.
379 // FNSTSW16r - Store FP status word into i16 register.
382 // SAHF - Store contents of %ah into %eflags.
385 // RDRAND - Get a random integer and indicate whether it is valid in CF.
388 // RDSEED - Get a NIST SP800-90B & C compliant random integer and
389 // indicate whether it is valid in CF.
396 // XTEST - Test if in transactional execution.
399 // ATOMADD64_DAG, ATOMSUB64_DAG, ATOMOR64_DAG, ATOMAND64_DAG,
400 // ATOMXOR64_DAG, ATOMNAND64_DAG, ATOMSWAP64_DAG -
401 // Atomic 64-bit binary operations.
402 ATOMADD64_DAG = ISD::FIRST_TARGET_MEMORY_OPCODE,
414 // LCMPXCHG_DAG, LCMPXCHG8_DAG, LCMPXCHG16_DAG - Compare and swap.
419 // VZEXT_LOAD - Load, scalar_to_vector, and zero extend.
422 // FNSTCW16m - Store FP control world into i16 memory.
425 /// FP_TO_INT*_IN_MEM - This instruction implements FP_TO_SINT with the
426 /// integer destination in memory and a FP reg source. This corresponds
427 /// to the X86::FIST*m instructions and the rounding mode change stuff. It
428 /// has two inputs (token chain and address) and two outputs (int value
429 /// and token chain).
434 /// FILD, FILD_FLAG - This instruction implements SINT_TO_FP with the
435 /// integer source in memory and FP reg result. This corresponds to the
436 /// X86::FILD*m instructions. It has three inputs (token chain, address,
437 /// and source type) and two outputs (FP value and token chain). FILD_FLAG
438 /// also produces a flag).
442 /// FLD - This instruction implements an extending load to FP stack slots.
443 /// This corresponds to the X86::FLD32m / X86::FLD64m. It takes a chain
444 /// operand, ptr to load from, and a ValueType node indicating the type
448 /// FST - This instruction implements a truncating store to FP stack
449 /// slots. This corresponds to the X86::FST32m / X86::FST64m. It takes a
450 /// chain operand, value to store, address, and a ValueType to store it
454 /// VAARG_64 - This instruction grabs the address of the next argument
455 /// from a va_list. (reads and modifies the va_list in memory)
458 // WARNING: Do not add anything in the end unless you want the node to
459 // have memop! In fact, starting from ATOMADD64_DAG all opcodes will be
460 // thought as target memory ops!
464 /// Define some predicates that are used for node matching.
466 /// isVEXTRACT128Index - Return true if the specified
467 /// EXTRACT_SUBVECTOR operand specifies a vector extract that is
468 /// suitable for input to VEXTRACTF128, VEXTRACTI128 instructions.
469 bool isVEXTRACT128Index(SDNode *N);
471 /// isVINSERT128Index - Return true if the specified
472 /// INSERT_SUBVECTOR operand specifies a subvector insert that is
473 /// suitable for input to VINSERTF128, VINSERTI128 instructions.
474 bool isVINSERT128Index(SDNode *N);
476 /// isVEXTRACT256Index - Return true if the specified
477 /// EXTRACT_SUBVECTOR operand specifies a vector extract that is
478 /// suitable for input to VEXTRACTF64X4, VEXTRACTI64X4 instructions.
479 bool isVEXTRACT256Index(SDNode *N);
481 /// isVINSERT256Index - Return true if the specified
482 /// INSERT_SUBVECTOR operand specifies a subvector insert that is
483 /// suitable for input to VINSERTF64X4, VINSERTI64X4 instructions.
484 bool isVINSERT256Index(SDNode *N);
486 /// getExtractVEXTRACT128Immediate - Return the appropriate
487 /// immediate to extract the specified EXTRACT_SUBVECTOR index
488 /// with VEXTRACTF128, VEXTRACTI128 instructions.
489 unsigned getExtractVEXTRACT128Immediate(SDNode *N);
491 /// getInsertVINSERT128Immediate - Return the appropriate
492 /// immediate to insert at the specified INSERT_SUBVECTOR index
493 /// with VINSERTF128, VINSERT128 instructions.
494 unsigned getInsertVINSERT128Immediate(SDNode *N);
496 /// getExtractVEXTRACT256Immediate - Return the appropriate
497 /// immediate to extract the specified EXTRACT_SUBVECTOR index
498 /// with VEXTRACTF64X4, VEXTRACTI64x4 instructions.
499 unsigned getExtractVEXTRACT256Immediate(SDNode *N);
501 /// getInsertVINSERT256Immediate - Return the appropriate
502 /// immediate to insert at the specified INSERT_SUBVECTOR index
503 /// with VINSERTF64x4, VINSERTI64x4 instructions.
504 unsigned getInsertVINSERT256Immediate(SDNode *N);
506 /// isZeroNode - Returns true if Elt is a constant zero or a floating point
508 bool isZeroNode(SDValue Elt);
510 /// isOffsetSuitableForCodeModel - Returns true of the given offset can be
511 /// fit into displacement field of the instruction.
512 bool isOffsetSuitableForCodeModel(int64_t Offset, CodeModel::Model M,
513 bool hasSymbolicDisplacement = true);
516 /// isCalleePop - Determines whether the callee is required to pop its
517 /// own arguments. Callee pop is necessary to support tail calls.
518 bool isCalleePop(CallingConv::ID CallingConv,
519 bool is64Bit, bool IsVarArg, bool TailCallOpt);
522 //===--------------------------------------------------------------------===//
523 // X86TargetLowering - X86 Implementation of the TargetLowering interface
524 class X86TargetLowering : public TargetLowering {
526 explicit X86TargetLowering(X86TargetMachine &TM);
528 unsigned getJumpTableEncoding() const override;
530 MVT getScalarShiftAmountTy(EVT LHSTy) const override { return MVT::i8; }
533 LowerCustomJumpTableEntry(const MachineJumpTableInfo *MJTI,
534 const MachineBasicBlock *MBB, unsigned uid,
535 MCContext &Ctx) const override;
537 /// getPICJumpTableRelocaBase - Returns relocation base for the given PIC
539 SDValue getPICJumpTableRelocBase(SDValue Table,
540 SelectionDAG &DAG) const override;
542 getPICJumpTableRelocBaseExpr(const MachineFunction *MF,
543 unsigned JTI, MCContext &Ctx) const override;
545 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
546 /// function arguments in the caller parameter area. For X86, aggregates
547 /// that contains are placed at 16-byte boundaries while the rest are at
548 /// 4-byte boundaries.
549 unsigned getByValTypeAlignment(Type *Ty) const override;
551 /// getOptimalMemOpType - Returns the target specific optimal type for load
552 /// and store operations as a result of memset, memcpy, and memmove
553 /// lowering. If DstAlign is zero that means it's safe to destination
554 /// alignment can satisfy any constraint. Similarly if SrcAlign is zero it
555 /// means there isn't a need to check it against alignment requirement,
556 /// probably because the source does not need to be loaded. If 'IsMemset' is
557 /// true, that means it's expanding a memset. If 'ZeroMemset' is true, that
558 /// means it's a memset of zero. 'MemcpyStrSrc' indicates whether the memcpy
559 /// source is constant so it does not need to be loaded.
560 /// It returns EVT::Other if the type should be determined using generic
561 /// target-independent logic.
562 EVT getOptimalMemOpType(uint64_t Size, unsigned DstAlign, unsigned SrcAlign,
563 bool IsMemset, bool ZeroMemset, bool MemcpyStrSrc,
564 MachineFunction &MF) const override;
566 /// isSafeMemOpType - Returns true if it's safe to use load / store of the
567 /// specified type to expand memcpy / memset inline. This is mostly true
568 /// for all types except for some special cases. For example, on X86
569 /// targets without SSE2 f64 load / store are done with fldl / fstpl which
570 /// also does type conversion. Note the specified type doesn't have to be
571 /// legal as the hook is used before type legalization.
572 bool isSafeMemOpType(MVT VT) const override;
574 /// allowsUnalignedMemoryAccesses - Returns true if the target allows
575 /// unaligned memory accesses. of the specified type. Returns whether it
576 /// is "fast" by reference in the second argument.
577 bool allowsUnalignedMemoryAccesses(EVT VT, unsigned AS,
578 bool *Fast) const override;
580 /// LowerOperation - Provide custom lowering hooks for some operations.
582 SDValue LowerOperation(SDValue Op, SelectionDAG &DAG) const override;
584 /// ReplaceNodeResults - Replace the results of node with an illegal result
585 /// type with new values built out of custom code.
587 void ReplaceNodeResults(SDNode *N, SmallVectorImpl<SDValue>&Results,
588 SelectionDAG &DAG) const override;
591 SDValue PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const override;
593 /// isTypeDesirableForOp - Return true if the target has native support for
594 /// the specified value type and it is 'desirable' to use the type for the
595 /// given node type. e.g. On x86 i16 is legal, but undesirable since i16
596 /// instruction encodings are longer and some i16 instructions are slow.
597 bool isTypeDesirableForOp(unsigned Opc, EVT VT) const override;
599 /// isTypeDesirable - Return true if the target has native support for the
600 /// specified value type and it is 'desirable' to use the type. e.g. On x86
601 /// i16 is legal, but undesirable since i16 instruction encodings are longer
602 /// and some i16 instructions are slow.
603 bool IsDesirableToPromoteOp(SDValue Op, EVT &PVT) const override;
606 EmitInstrWithCustomInserter(MachineInstr *MI,
607 MachineBasicBlock *MBB) const override;
610 /// getTargetNodeName - This method returns the name of a target specific
612 const char *getTargetNodeName(unsigned Opcode) const override;
614 /// getSetCCResultType - Return the value type to use for ISD::SETCC.
615 EVT getSetCCResultType(LLVMContext &Context, EVT VT) const override;
617 /// computeMaskedBitsForTargetNode - Determine which of the bits specified
618 /// in Mask are known to be either zero or one and return them in the
619 /// KnownZero/KnownOne bitsets.
620 void computeMaskedBitsForTargetNode(const SDValue Op,
623 const SelectionDAG &DAG,
624 unsigned Depth = 0) const override;
626 // ComputeNumSignBitsForTargetNode - Determine the number of bits in the
627 // operation that are sign bits.
628 unsigned ComputeNumSignBitsForTargetNode(SDValue Op,
629 unsigned Depth) const override;
631 bool isGAPlusOffset(SDNode *N, const GlobalValue* &GA,
632 int64_t &Offset) const override;
634 SDValue getReturnAddressFrameIndex(SelectionDAG &DAG) const;
636 bool ExpandInlineAsm(CallInst *CI) const override;
639 getConstraintType(const std::string &Constraint) const override;
641 /// Examine constraint string and operand type and determine a weight value.
642 /// The operand object must already have been set up with the operand type.
644 getSingleConstraintMatchWeight(AsmOperandInfo &info,
645 const char *constraint) const override;
647 const char *LowerXConstraint(EVT ConstraintVT) const override;
649 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
650 /// vector. If it is invalid, don't add anything to Ops. If hasMemory is
651 /// true it means one of the asm constraint of the inline asm instruction
652 /// being processed is 'm'.
653 void LowerAsmOperandForConstraint(SDValue Op,
654 std::string &Constraint,
655 std::vector<SDValue> &Ops,
656 SelectionDAG &DAG) const override;
658 /// getRegForInlineAsmConstraint - Given a physical register constraint
659 /// (e.g. {edx}), return the register number and the register class for the
660 /// register. This should only be used for C_Register constraints. On
661 /// error, this returns a register number of 0.
662 std::pair<unsigned, const TargetRegisterClass*>
663 getRegForInlineAsmConstraint(const std::string &Constraint,
664 MVT VT) const override;
666 /// isLegalAddressingMode - Return true if the addressing mode represented
667 /// by AM is legal for this target, for a load/store of the specified type.
668 bool isLegalAddressingMode(const AddrMode &AM, Type *Ty) const override;
670 /// isLegalICmpImmediate - Return true if the specified immediate is legal
671 /// icmp immediate, that is the target has icmp instructions which can
672 /// compare a register against the immediate without having to materialize
673 /// the immediate into a register.
674 bool isLegalICmpImmediate(int64_t Imm) const override;
676 /// isLegalAddImmediate - Return true if the specified immediate is legal
677 /// add immediate, that is the target has add instructions which can
678 /// add a register and the immediate without having to materialize
679 /// the immediate into a register.
680 bool isLegalAddImmediate(int64_t Imm) const override;
683 bool isVectorShiftByScalarCheap(Type *Ty) const override;
685 /// isTruncateFree - Return true if it's free to truncate a value of
686 /// type Ty1 to type Ty2. e.g. On x86 it's free to truncate a i32 value in
687 /// register EAX to i16 by referencing its sub-register AX.
688 bool isTruncateFree(Type *Ty1, Type *Ty2) const override;
689 bool isTruncateFree(EVT VT1, EVT VT2) const override;
691 bool allowTruncateForTailCall(Type *Ty1, Type *Ty2) const override;
693 /// isZExtFree - Return true if any actual instruction that defines a
694 /// value of type Ty1 implicit zero-extends the value to Ty2 in the result
695 /// register. This does not necessarily include registers defined in
696 /// unknown ways, such as incoming arguments, or copies from unknown
697 /// virtual registers. Also, if isTruncateFree(Ty2, Ty1) is true, this
698 /// does not necessarily apply to truncate instructions. e.g. on x86-64,
699 /// all instructions that define 32-bit values implicit zero-extend the
700 /// result out to 64 bits.
701 bool isZExtFree(Type *Ty1, Type *Ty2) const override;
702 bool isZExtFree(EVT VT1, EVT VT2) const override;
703 bool isZExtFree(SDValue Val, EVT VT2) const override;
705 /// isFMAFasterThanFMulAndFAdd - Return true if an FMA operation is faster
706 /// than a pair of fmul and fadd instructions. fmuladd intrinsics will be
707 /// expanded to FMAs when this method returns true, otherwise fmuladd is
708 /// expanded to fmul + fadd.
709 bool isFMAFasterThanFMulAndFAdd(EVT VT) const override;
711 /// isNarrowingProfitable - Return true if it's profitable to narrow
712 /// operations of type VT1 to VT2. e.g. on x86, it's profitable to narrow
713 /// from i32 to i8 but not from i32 to i16.
714 bool isNarrowingProfitable(EVT VT1, EVT VT2) const override;
716 /// isFPImmLegal - Returns true if the target can instruction select the
717 /// specified FP immediate natively. If false, the legalizer will
718 /// materialize the FP immediate as a load from a constant pool.
719 bool isFPImmLegal(const APFloat &Imm, EVT VT) const override;
721 /// isShuffleMaskLegal - Targets can use this to indicate that they only
722 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
723 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask
724 /// values are assumed to be legal.
725 bool isShuffleMaskLegal(const SmallVectorImpl<int> &Mask,
726 EVT VT) const override;
728 /// isVectorClearMaskLegal - Similar to isShuffleMaskLegal. This is
729 /// used by Targets can use this to indicate if there is a suitable
730 /// VECTOR_SHUFFLE that can be used to replace a VAND with a constant
732 bool isVectorClearMaskLegal(const SmallVectorImpl<int> &Mask,
733 EVT VT) const override;
735 /// ShouldShrinkFPConstant - If true, then instruction selection should
736 /// seek to shrink the FP constant of the specified type to a smaller type
737 /// in order to save space and / or reduce runtime.
738 bool ShouldShrinkFPConstant(EVT VT) const override {
739 // Don't shrink FP constpool if SSE2 is available since cvtss2sd is more
740 // expensive than a straight movsd. On the other hand, it's important to
741 // shrink long double fp constant since fldt is very slow.
742 return !X86ScalarSSEf64 || VT == MVT::f80;
745 const X86Subtarget* getSubtarget() const {
749 /// isScalarFPTypeInSSEReg - Return true if the specified scalar FP type is
750 /// computed in an SSE register, not on the X87 floating point stack.
751 bool isScalarFPTypeInSSEReg(EVT VT) const {
752 return (VT == MVT::f64 && X86ScalarSSEf64) || // f64 is when SSE2
753 (VT == MVT::f32 && X86ScalarSSEf32); // f32 is when SSE1
756 /// isTargetFTOL - Return true if the target uses the MSVC _ftol2 routine
758 bool isTargetFTOL() const {
759 return Subtarget->isTargetWindows() && !Subtarget->is64Bit();
762 /// isIntegerTypeFTOL - Return true if the MSVC _ftol2 routine should be
763 /// used for fptoui to the given type.
764 bool isIntegerTypeFTOL(EVT VT) const {
765 return isTargetFTOL() && VT == MVT::i64;
768 /// \brief Returns true if it is beneficial to convert a load of a constant
769 /// to just the constant itself.
770 bool shouldConvertConstantLoadToIntImm(const APInt &Imm,
771 Type *Ty) const override;
773 /// Intel processors have a unified instruction and data cache
774 const char * getClearCacheBuiltinName() const {
775 return 0; // nothing to do, move along.
778 /// createFastISel - This method returns a target specific FastISel object,
779 /// or null if the target does not support "fast" ISel.
780 FastISel *createFastISel(FunctionLoweringInfo &funcInfo,
781 const TargetLibraryInfo *libInfo) const override;
783 /// getStackCookieLocation - Return true if the target stores stack
784 /// protector cookies at a fixed offset in some non-standard address
785 /// space, and populates the address space and offset as
787 bool getStackCookieLocation(unsigned &AddressSpace,
788 unsigned &Offset) const override;
790 SDValue BuildFILD(SDValue Op, EVT SrcVT, SDValue Chain, SDValue StackSlot,
791 SelectionDAG &DAG) const;
793 bool isNoopAddrSpaceCast(unsigned SrcAS, unsigned DestAS) const override;
795 /// \brief Reset the operation actions based on target options.
796 void resetOperationActions() override;
799 std::pair<const TargetRegisterClass*, uint8_t>
800 findRepresentativeClass(MVT VT) const override;
803 /// Subtarget - Keep a pointer to the X86Subtarget around so that we can
804 /// make the right decision when generating code for different targets.
805 const X86Subtarget *Subtarget;
806 const DataLayout *TD;
808 /// Used to store the TargetOptions so that we don't waste time resetting
809 /// the operation actions unless we have to.
812 /// X86ScalarSSEf32, X86ScalarSSEf64 - Select between SSE or x87
813 /// floating point ops.
814 /// When SSE is available, use it for f32 operations.
815 /// When SSE2 is available, use it for f64 operations.
816 bool X86ScalarSSEf32;
817 bool X86ScalarSSEf64;
819 /// LegalFPImmediates - A list of legal fp immediates.
820 std::vector<APFloat> LegalFPImmediates;
822 /// addLegalFPImmediate - Indicate that this x86 target can instruction
823 /// select the specified FP immediate natively.
824 void addLegalFPImmediate(const APFloat& Imm) {
825 LegalFPImmediates.push_back(Imm);
828 SDValue LowerCallResult(SDValue Chain, SDValue InFlag,
829 CallingConv::ID CallConv, bool isVarArg,
830 const SmallVectorImpl<ISD::InputArg> &Ins,
831 SDLoc dl, SelectionDAG &DAG,
832 SmallVectorImpl<SDValue> &InVals) const;
833 SDValue LowerMemArgument(SDValue Chain,
834 CallingConv::ID CallConv,
835 const SmallVectorImpl<ISD::InputArg> &ArgInfo,
836 SDLoc dl, SelectionDAG &DAG,
837 const CCValAssign &VA, MachineFrameInfo *MFI,
839 SDValue LowerMemOpCallTo(SDValue Chain, SDValue StackPtr, SDValue Arg,
840 SDLoc dl, SelectionDAG &DAG,
841 const CCValAssign &VA,
842 ISD::ArgFlagsTy Flags) const;
844 // Call lowering helpers.
846 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
847 /// for tail call optimization. Targets which want to do tail call
848 /// optimization should implement this function.
849 bool IsEligibleForTailCallOptimization(SDValue Callee,
850 CallingConv::ID CalleeCC,
852 bool isCalleeStructRet,
853 bool isCallerStructRet,
855 const SmallVectorImpl<ISD::OutputArg> &Outs,
856 const SmallVectorImpl<SDValue> &OutVals,
857 const SmallVectorImpl<ISD::InputArg> &Ins,
858 SelectionDAG& DAG) const;
859 bool IsCalleePop(bool isVarArg, CallingConv::ID CallConv) const;
860 SDValue EmitTailCallLoadRetAddr(SelectionDAG &DAG, SDValue &OutRetAddr,
861 SDValue Chain, bool IsTailCall, bool Is64Bit,
862 int FPDiff, SDLoc dl) const;
864 unsigned GetAlignedArgumentStackSize(unsigned StackSize,
865 SelectionDAG &DAG) const;
867 std::pair<SDValue,SDValue> FP_TO_INTHelper(SDValue Op, SelectionDAG &DAG,
869 bool isReplace) const;
871 SDValue LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) const;
872 SDValue LowerBUILD_VECTORvXi1(SDValue Op, SelectionDAG &DAG) const;
873 SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) const;
874 SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const;
875 SDValue ExtractBitFromMaskVector(SDValue Op, SelectionDAG &DAG) const;
876 SDValue LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const;
877 SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) const;
878 SDValue LowerBlockAddress(SDValue Op, SelectionDAG &DAG) const;
879 SDValue LowerGlobalAddress(const GlobalValue *GV, SDLoc dl,
880 int64_t Offset, SelectionDAG &DAG) const;
881 SDValue LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) const;
882 SDValue LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const;
883 SDValue LowerExternalSymbol(SDValue Op, SelectionDAG &DAG) const;
884 SDValue LowerSINT_TO_FP(SDValue Op, SelectionDAG &DAG) const;
885 SDValue LowerUINT_TO_FP(SDValue Op, SelectionDAG &DAG) const;
886 SDValue LowerUINT_TO_FP_i64(SDValue Op, SelectionDAG &DAG) const;
887 SDValue LowerUINT_TO_FP_i32(SDValue Op, SelectionDAG &DAG) const;
888 SDValue lowerUINT_TO_FP_vec(SDValue Op, SelectionDAG &DAG) const;
889 SDValue LowerTRUNCATE(SDValue Op, SelectionDAG &DAG) const;
890 SDValue LowerFP_TO_SINT(SDValue Op, SelectionDAG &DAG) const;
891 SDValue LowerFP_TO_UINT(SDValue Op, SelectionDAG &DAG) const;
892 SDValue LowerToBT(SDValue And, ISD::CondCode CC,
893 SDLoc dl, SelectionDAG &DAG) const;
894 SDValue LowerSETCC(SDValue Op, SelectionDAG &DAG) const;
895 SDValue LowerSELECT(SDValue Op, SelectionDAG &DAG) const;
896 SDValue LowerBRCOND(SDValue Op, SelectionDAG &DAG) const;
897 SDValue LowerMEMSET(SDValue Op, SelectionDAG &DAG) const;
898 SDValue LowerJumpTable(SDValue Op, SelectionDAG &DAG) const;
899 SDValue LowerDYNAMIC_STACKALLOC(SDValue Op, SelectionDAG &DAG) const;
900 SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG) const;
901 SDValue LowerVAARG(SDValue Op, SelectionDAG &DAG) const;
902 SDValue LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const;
903 SDValue LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const;
904 SDValue LowerFRAME_TO_ARGS_OFFSET(SDValue Op, SelectionDAG &DAG) const;
905 SDValue LowerEH_RETURN(SDValue Op, SelectionDAG &DAG) const;
906 SDValue lowerEH_SJLJ_SETJMP(SDValue Op, SelectionDAG &DAG) const;
907 SDValue lowerEH_SJLJ_LONGJMP(SDValue Op, SelectionDAG &DAG) const;
908 SDValue LowerINIT_TRAMPOLINE(SDValue Op, SelectionDAG &DAG) const;
909 SDValue LowerFLT_ROUNDS_(SDValue Op, SelectionDAG &DAG) const;
910 SDValue LowerSIGN_EXTEND_INREG(SDValue Op, SelectionDAG &DAG) const;
913 LowerFormalArguments(SDValue Chain,
914 CallingConv::ID CallConv, bool isVarArg,
915 const SmallVectorImpl<ISD::InputArg> &Ins,
916 SDLoc dl, SelectionDAG &DAG,
917 SmallVectorImpl<SDValue> &InVals) const override;
918 SDValue LowerCall(CallLoweringInfo &CLI,
919 SmallVectorImpl<SDValue> &InVals) const override;
921 SDValue LowerReturn(SDValue Chain,
922 CallingConv::ID CallConv, bool isVarArg,
923 const SmallVectorImpl<ISD::OutputArg> &Outs,
924 const SmallVectorImpl<SDValue> &OutVals,
925 SDLoc dl, SelectionDAG &DAG) const override;
927 bool isUsedByReturnOnly(SDNode *N, SDValue &Chain) const override;
929 bool mayBeEmittedAsTailCall(CallInst *CI) const override;
931 MVT getTypeForExtArgOrReturn(MVT VT,
932 ISD::NodeType ExtendKind) const override;
934 bool CanLowerReturn(CallingConv::ID CallConv, MachineFunction &MF,
936 const SmallVectorImpl<ISD::OutputArg> &Outs,
937 LLVMContext &Context) const override;
939 const uint16_t *getScratchRegisters(CallingConv::ID CC) const override;
941 /// Utility function to emit atomic-load-arith operations (and, or, xor,
942 /// nand, max, min, umax, umin). It takes the corresponding instruction to
943 /// expand, the associated machine basic block, and the associated X86
944 /// opcodes for reg/reg.
945 MachineBasicBlock *EmitAtomicLoadArith(MachineInstr *MI,
946 MachineBasicBlock *MBB) const;
948 /// Utility function to emit atomic-load-arith operations (and, or, xor,
949 /// nand, add, sub, swap) for 64-bit operands on 32-bit target.
950 MachineBasicBlock *EmitAtomicLoadArith6432(MachineInstr *MI,
951 MachineBasicBlock *MBB) const;
953 // Utility function to emit the low-level va_arg code for X86-64.
954 MachineBasicBlock *EmitVAARG64WithCustomInserter(
956 MachineBasicBlock *MBB) const;
958 /// Utility function to emit the xmm reg save portion of va_start.
959 MachineBasicBlock *EmitVAStartSaveXMMRegsWithCustomInserter(
960 MachineInstr *BInstr,
961 MachineBasicBlock *BB) const;
963 MachineBasicBlock *EmitLoweredSelect(MachineInstr *I,
964 MachineBasicBlock *BB) const;
966 MachineBasicBlock *EmitLoweredWinAlloca(MachineInstr *MI,
967 MachineBasicBlock *BB) const;
969 MachineBasicBlock *EmitLoweredSegAlloca(MachineInstr *MI,
970 MachineBasicBlock *BB,
973 MachineBasicBlock *EmitLoweredTLSCall(MachineInstr *MI,
974 MachineBasicBlock *BB) const;
976 MachineBasicBlock *emitLoweredTLSAddr(MachineInstr *MI,
977 MachineBasicBlock *BB) const;
979 MachineBasicBlock *emitEHSjLjSetJmp(MachineInstr *MI,
980 MachineBasicBlock *MBB) const;
982 MachineBasicBlock *emitEHSjLjLongJmp(MachineInstr *MI,
983 MachineBasicBlock *MBB) const;
985 MachineBasicBlock *emitFMA3Instr(MachineInstr *MI,
986 MachineBasicBlock *MBB) const;
988 /// Emit nodes that will be selected as "test Op0,Op0", or something
989 /// equivalent, for use with the given x86 condition code.
990 SDValue EmitTest(SDValue Op0, unsigned X86CC, SelectionDAG &DAG) const;
992 /// Emit nodes that will be selected as "cmp Op0,Op1", or something
993 /// equivalent, for use with the given x86 condition code.
994 SDValue EmitCmp(SDValue Op0, SDValue Op1, unsigned X86CC,
995 SelectionDAG &DAG) const;
997 /// Convert a comparison if required by the subtarget.
998 SDValue ConvertCmpIfNecessary(SDValue Cmp, SelectionDAG &DAG) const;
1002 FastISel *createFastISel(FunctionLoweringInfo &funcInfo,
1003 const TargetLibraryInfo *libInfo);
1007 #endif // X86ISELLOWERING_H