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 "X86MachineFunctionInfo.h"
19 #include "X86RegisterInfo.h"
20 #include "X86Subtarget.h"
21 #include "llvm/CodeGen/CallingConvLower.h"
22 #include "llvm/CodeGen/FastISel.h"
23 #include "llvm/CodeGen/SelectionDAG.h"
24 #include "llvm/Target/TargetLowering.h"
25 #include "llvm/Target/TargetOptions.h"
29 // X86 Specific DAG Nodes
31 // Start the numbering where the builtin ops leave off.
32 FIRST_NUMBER = ISD::BUILTIN_OP_END,
34 /// BSF - Bit scan forward.
35 /// BSR - Bit scan reverse.
39 /// SHLD, SHRD - Double shift instructions. These correspond to
40 /// X86::SHLDxx and X86::SHRDxx instructions.
44 /// FAND - Bitwise logical AND of floating point values. This corresponds
45 /// to X86::ANDPS or X86::ANDPD.
48 /// FOR - Bitwise logical OR of floating point values. This corresponds
49 /// to X86::ORPS or X86::ORPD.
52 /// FXOR - Bitwise logical XOR of floating point values. This corresponds
53 /// to X86::XORPS or X86::XORPD.
56 /// FANDN - Bitwise logical ANDNOT of floating point values. This
57 /// corresponds to X86::ANDNPS or X86::ANDNPD.
60 /// FSRL - Bitwise logical right shift of floating point values. These
61 /// corresponds to X86::PSRLDQ.
64 /// CALL - These operations represent an abstract X86 call
65 /// instruction, which includes a bunch of information. In particular the
66 /// operands of these node are:
68 /// #0 - The incoming token chain
70 /// #2 - The number of arg bytes the caller pushes on the stack.
71 /// #3 - The number of arg bytes the callee pops off the stack.
72 /// #4 - The value to pass in AL/AX/EAX (optional)
73 /// #5 - The value to pass in DL/DX/EDX (optional)
75 /// The result values of these nodes are:
77 /// #0 - The outgoing token chain
78 /// #1 - The first register result value (optional)
79 /// #2 - The second register result value (optional)
83 /// RDTSC_DAG - This operation implements the lowering for
87 /// X86 compare and logical compare instructions.
90 /// X86 bit-test instructions.
93 /// X86 SetCC. Operand 0 is condition code, and operand 1 is the EFLAGS
94 /// operand, usually produced by a CMP instruction.
97 // Same as SETCC except it's materialized with a sbb and the value is all
98 // one's or all zero's.
99 SETCC_CARRY, // R = carry_bit ? ~0 : 0
101 /// X86 FP SETCC, implemented with CMP{cc}SS/CMP{cc}SD.
102 /// Operands are two FP values to compare; result is a mask of
103 /// 0s or 1s. Generally DTRT for C/C++ with NaNs.
106 /// X86 MOVMSK{pd|ps}, extracts sign bits of two or four FP values,
107 /// result in an integer GPR. Needs masking for scalar result.
110 /// X86 conditional moves. Operand 0 and operand 1 are the two values
111 /// to select from. Operand 2 is the condition code, and operand 3 is the
112 /// flag operand produced by a CMP or TEST instruction. It also writes a
116 /// X86 conditional branches. Operand 0 is the chain operand, operand 1
117 /// is the block to branch if condition is true, operand 2 is the
118 /// condition code, and operand 3 is the flag operand produced by a CMP
119 /// or TEST instruction.
122 /// Return with a flag operand. Operand 0 is the chain operand, operand
123 /// 1 is the number of bytes of stack to pop.
126 /// REP_STOS - Repeat fill, corresponds to X86::REP_STOSx.
129 /// REP_MOVS - Repeat move, corresponds to X86::REP_MOVSx.
132 /// GlobalBaseReg - On Darwin, this node represents the result of the popl
133 /// at function entry, used for PIC code.
136 /// Wrapper - A wrapper node for TargetConstantPool,
137 /// TargetExternalSymbol, and TargetGlobalAddress.
140 /// WrapperRIP - Special wrapper used under X86-64 PIC mode for RIP
141 /// relative displacements.
144 /// MOVDQ2Q - Copies a 64-bit value from the low word of an XMM vector
145 /// to an MMX vector. If you think this is too close to the previous
146 /// mnemonic, so do I; blame Intel.
149 /// MMX_MOVD2W - Copies a 32-bit value from the low word of a MMX
153 /// PEXTRB - Extract an 8-bit value from a vector and zero extend it to
154 /// i32, corresponds to X86::PEXTRB.
157 /// PEXTRW - Extract a 16-bit value from a vector and zero extend it to
158 /// i32, corresponds to X86::PEXTRW.
161 /// INSERTPS - Insert any element of a 4 x float vector into any element
162 /// of a destination 4 x floatvector.
165 /// PINSRB - Insert the lower 8-bits of a 32-bit value to a vector,
166 /// corresponds to X86::PINSRB.
169 /// PINSRW - Insert the lower 16-bits of a 32-bit value to a vector,
170 /// corresponds to X86::PINSRW.
173 /// PSHUFB - Shuffle 16 8-bit values within a vector.
176 /// ANDNP - Bitwise Logical AND NOT of Packed FP values.
179 /// PSIGN - Copy integer sign.
182 /// BLENDV - Blend where the selector is a register.
185 /// BLENDI - Blend where the selector is an immediate.
188 // SUBUS - Integer sub with unsigned saturation.
191 /// HADD - Integer horizontal add.
194 /// HSUB - Integer horizontal sub.
197 /// FHADD - Floating point horizontal add.
200 /// FHSUB - Floating point horizontal sub.
203 /// UMAX, UMIN - Unsigned integer max and min.
206 /// SMAX, SMIN - Signed integer max and min.
209 /// FMAX, FMIN - Floating point max and min.
213 /// FMAXC, FMINC - Commutative FMIN and FMAX.
216 /// FRSQRT, FRCP - Floating point reciprocal-sqrt and reciprocal
217 /// approximation. Note that these typically require refinement
218 /// in order to obtain suitable precision.
221 // TLSADDR - Thread Local Storage.
224 // TLSBASEADDR - Thread Local Storage. A call to get the start address
225 // of the TLS block for the current module.
228 // TLSCALL - Thread Local Storage. When calling to an OS provided
229 // thunk at the address from an earlier relocation.
232 // EH_RETURN - Exception Handling helpers.
235 // EH_SJLJ_SETJMP - SjLj exception handling setjmp.
238 // EH_SJLJ_LONGJMP - SjLj exception handling longjmp.
241 /// TC_RETURN - Tail call return. See X86TargetLowering::LowerCall for
242 /// the list of operands.
245 // VZEXT_MOVL - Vector move low and zero extend.
248 // VSEXT_MOVL - Vector move low and sign extend.
251 // VZEXT - Vector integer zero-extend.
254 // VSEXT - Vector integer signed-extend.
257 // VFPEXT - Vector FP extend.
260 // VFPROUND - Vector FP round.
263 // VSHL, VSRL - 128-bit vector logical left / right shift
266 // VSHL, VSRL, VSRA - Vector shift elements
269 // VSHLI, VSRLI, VSRAI - Vector shift elements by immediate
272 // CMPP - Vector packed double/float comparison.
275 // PCMP* - Vector integer comparisons.
277 // PCMP*M - Vector integer comparisons, the result is in a mask vector.
280 /// CMPM, CMPMU - Vector comparison generating mask bits for fp and
281 /// integer signed and unsigned data types.
285 // ADD, SUB, SMUL, etc. - Arithmetic operations with FLAGS results.
286 ADD, SUB, ADC, SBB, SMUL,
287 INC, DEC, OR, XOR, AND,
289 BLSI, // BLSI - Extract lowest set isolated bit
290 BLSMSK, // BLSMSK - Get mask up to lowest set bit
291 BLSR, // BLSR - Reset lowest set bit
293 UMUL, // LOW, HI, FLAGS = umul LHS, RHS
295 // MUL_IMM - X86 specific multiply by immediate.
298 // PTEST - Vector bitwise comparisons.
301 // TESTP - Vector packed fp sign bitwise comparisons.
304 // TESTM - Vector "test" in AVX-512, the result is in a mask vector.
307 // OR/AND test for masks
311 // Several flavors of instructions with vector shuffle behaviors.
338 // PMULUDQ - Vector multiply packed unsigned doubleword integers
349 // VASTART_SAVE_XMM_REGS - Save xmm argument registers to the stack,
350 // according to %al. An operator is needed so that this can be expanded
351 // with control flow.
352 VASTART_SAVE_XMM_REGS,
354 // WIN_ALLOCA - Windows's _chkstk call to do stack probing.
357 // SEG_ALLOCA - For allocating variable amounts of stack space when using
358 // segmented stacks. Check if the current stacklet has enough space, and
359 // falls back to heap allocation if not.
362 // WIN_FTOL - Windows's _ftol2 runtime routine to do fptoui.
371 // FNSTSW16r - Store FP status word into i16 register.
374 // SAHF - Store contents of %ah into %eflags.
377 // RDRAND - Get a random integer and indicate whether it is valid in CF.
380 // RDSEED - Get a NIST SP800-90B & C compliant random integer and
381 // indicate whether it is valid in CF.
388 // XTEST - Test if in transactional execution.
391 // ATOMADD64_DAG, ATOMSUB64_DAG, ATOMOR64_DAG, ATOMAND64_DAG,
392 // ATOMXOR64_DAG, ATOMNAND64_DAG, ATOMSWAP64_DAG -
393 // Atomic 64-bit binary operations.
394 ATOMADD64_DAG = ISD::FIRST_TARGET_MEMORY_OPCODE,
406 // LCMPXCHG_DAG, LCMPXCHG8_DAG, LCMPXCHG16_DAG - Compare and swap.
411 // VZEXT_LOAD - Load, scalar_to_vector, and zero extend.
414 // FNSTCW16m - Store FP control world into i16 memory.
417 /// FP_TO_INT*_IN_MEM - This instruction implements FP_TO_SINT with the
418 /// integer destination in memory and a FP reg source. This corresponds
419 /// to the X86::FIST*m instructions and the rounding mode change stuff. It
420 /// has two inputs (token chain and address) and two outputs (int value
421 /// and token chain).
426 /// FILD, FILD_FLAG - This instruction implements SINT_TO_FP with the
427 /// integer source in memory and FP reg result. This corresponds to the
428 /// X86::FILD*m instructions. It has three inputs (token chain, address,
429 /// and source type) and two outputs (FP value and token chain). FILD_FLAG
430 /// also produces a flag).
434 /// FLD - This instruction implements an extending load to FP stack slots.
435 /// This corresponds to the X86::FLD32m / X86::FLD64m. It takes a chain
436 /// operand, ptr to load from, and a ValueType node indicating the type
440 /// FST - This instruction implements a truncating store to FP stack
441 /// slots. This corresponds to the X86::FST32m / X86::FST64m. It takes a
442 /// chain operand, value to store, address, and a ValueType to store it
446 /// VAARG_64 - This instruction grabs the address of the next argument
447 /// from a va_list. (reads and modifies the va_list in memory)
450 // WARNING: Do not add anything in the end unless you want the node to
451 // have memop! In fact, starting from ATOMADD64_DAG all opcodes will be
452 // thought as target memory ops!
456 /// Define some predicates that are used for node matching.
458 /// isVEXTRACT128Index - Return true if the specified
459 /// EXTRACT_SUBVECTOR operand specifies a vector extract that is
460 /// suitable for input to VEXTRACTF128, VEXTRACTI128 instructions.
461 bool isVEXTRACT128Index(SDNode *N);
463 /// isVINSERT128Index - Return true if the specified
464 /// INSERT_SUBVECTOR operand specifies a subvector insert that is
465 /// suitable for input to VINSERTF128, VINSERTI128 instructions.
466 bool isVINSERT128Index(SDNode *N);
468 /// isVEXTRACT256Index - Return true if the specified
469 /// EXTRACT_SUBVECTOR operand specifies a vector extract that is
470 /// suitable for input to VEXTRACTF64X4, VEXTRACTI64X4 instructions.
471 bool isVEXTRACT256Index(SDNode *N);
473 /// isVINSERT256Index - Return true if the specified
474 /// INSERT_SUBVECTOR operand specifies a subvector insert that is
475 /// suitable for input to VINSERTF64X4, VINSERTI64X4 instructions.
476 bool isVINSERT256Index(SDNode *N);
478 /// getExtractVEXTRACT128Immediate - Return the appropriate
479 /// immediate to extract the specified EXTRACT_SUBVECTOR index
480 /// with VEXTRACTF128, VEXTRACTI128 instructions.
481 unsigned getExtractVEXTRACT128Immediate(SDNode *N);
483 /// getInsertVINSERT128Immediate - Return the appropriate
484 /// immediate to insert at the specified INSERT_SUBVECTOR index
485 /// with VINSERTF128, VINSERT128 instructions.
486 unsigned getInsertVINSERT128Immediate(SDNode *N);
488 /// getExtractVEXTRACT256Immediate - Return the appropriate
489 /// immediate to extract the specified EXTRACT_SUBVECTOR index
490 /// with VEXTRACTF64X4, VEXTRACTI64x4 instructions.
491 unsigned getExtractVEXTRACT256Immediate(SDNode *N);
493 /// getInsertVINSERT256Immediate - Return the appropriate
494 /// immediate to insert at the specified INSERT_SUBVECTOR index
495 /// with VINSERTF64x4, VINSERTI64x4 instructions.
496 unsigned getInsertVINSERT256Immediate(SDNode *N);
498 /// isZeroNode - Returns true if Elt is a constant zero or a floating point
500 bool isZeroNode(SDValue Elt);
502 /// isOffsetSuitableForCodeModel - Returns true of the given offset can be
503 /// fit into displacement field of the instruction.
504 bool isOffsetSuitableForCodeModel(int64_t Offset, CodeModel::Model M,
505 bool hasSymbolicDisplacement = true);
508 /// isCalleePop - Determines whether the callee is required to pop its
509 /// own arguments. Callee pop is necessary to support tail calls.
510 bool isCalleePop(CallingConv::ID CallingConv,
511 bool is64Bit, bool IsVarArg, bool TailCallOpt);
514 //===--------------------------------------------------------------------===//
515 // X86TargetLowering - X86 Implementation of the TargetLowering interface
516 class X86TargetLowering : public TargetLowering {
518 explicit X86TargetLowering(X86TargetMachine &TM);
520 virtual unsigned getJumpTableEncoding() const;
522 virtual MVT getScalarShiftAmountTy(EVT LHSTy) const { return MVT::i8; }
524 virtual const MCExpr *
525 LowerCustomJumpTableEntry(const MachineJumpTableInfo *MJTI,
526 const MachineBasicBlock *MBB, unsigned uid,
527 MCContext &Ctx) const;
529 /// getPICJumpTableRelocaBase - Returns relocation base for the given PIC
531 virtual SDValue getPICJumpTableRelocBase(SDValue Table,
532 SelectionDAG &DAG) const;
533 virtual const MCExpr *
534 getPICJumpTableRelocBaseExpr(const MachineFunction *MF,
535 unsigned JTI, MCContext &Ctx) const;
537 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
538 /// function arguments in the caller parameter area. For X86, aggregates
539 /// that contains are placed at 16-byte boundaries while the rest are at
540 /// 4-byte boundaries.
541 virtual unsigned getByValTypeAlignment(Type *Ty) const;
543 /// getOptimalMemOpType - Returns the target specific optimal type for load
544 /// and store operations as a result of memset, memcpy, and memmove
545 /// lowering. If DstAlign is zero that means it's safe to destination
546 /// alignment can satisfy any constraint. Similarly if SrcAlign is zero it
547 /// means there isn't a need to check it against alignment requirement,
548 /// probably because the source does not need to be loaded. If 'IsMemset' is
549 /// true, that means it's expanding a memset. If 'ZeroMemset' is true, that
550 /// means it's a memset of zero. 'MemcpyStrSrc' indicates whether the memcpy
551 /// source is constant so it does not need to be loaded.
552 /// It returns EVT::Other if the type should be determined using generic
553 /// target-independent logic.
555 getOptimalMemOpType(uint64_t Size, unsigned DstAlign, unsigned SrcAlign,
556 bool IsMemset, bool ZeroMemset, bool MemcpyStrSrc,
557 MachineFunction &MF) const;
559 /// isSafeMemOpType - Returns true if it's safe to use load / store of the
560 /// specified type to expand memcpy / memset inline. This is mostly true
561 /// for all types except for some special cases. For example, on X86
562 /// targets without SSE2 f64 load / store are done with fldl / fstpl which
563 /// also does type conversion. Note the specified type doesn't have to be
564 /// legal as the hook is used before type legalization.
565 virtual bool isSafeMemOpType(MVT VT) const;
567 /// allowsUnalignedMemoryAccesses - Returns true if the target allows
568 /// unaligned memory accesses. of the specified type. Returns whether it
569 /// is "fast" by reference in the second argument.
570 virtual bool allowsUnalignedMemoryAccesses(EVT VT, bool *Fast) const;
572 /// LowerOperation - Provide custom lowering hooks for some operations.
574 virtual SDValue LowerOperation(SDValue Op, SelectionDAG &DAG) const;
576 /// ReplaceNodeResults - Replace the results of node with an illegal result
577 /// type with new values built out of custom code.
579 virtual void ReplaceNodeResults(SDNode *N, SmallVectorImpl<SDValue>&Results,
580 SelectionDAG &DAG) const;
583 virtual SDValue PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const;
585 /// isTypeDesirableForOp - Return true if the target has native support for
586 /// the specified value type and it is 'desirable' to use the type for the
587 /// given node type. e.g. On x86 i16 is legal, but undesirable since i16
588 /// instruction encodings are longer and some i16 instructions are slow.
589 virtual bool isTypeDesirableForOp(unsigned Opc, EVT VT) const;
591 /// isTypeDesirable - Return true if the target has native support for the
592 /// specified value type and it is 'desirable' to use the type. e.g. On x86
593 /// i16 is legal, but undesirable since i16 instruction encodings are longer
594 /// and some i16 instructions are slow.
595 virtual bool IsDesirableToPromoteOp(SDValue Op, EVT &PVT) const;
597 virtual MachineBasicBlock *
598 EmitInstrWithCustomInserter(MachineInstr *MI,
599 MachineBasicBlock *MBB) const;
602 /// getTargetNodeName - This method returns the name of a target specific
604 virtual const char *getTargetNodeName(unsigned Opcode) const;
606 /// getSetCCResultType - Return the value type to use for ISD::SETCC.
607 virtual EVT getSetCCResultType(LLVMContext &Context, EVT VT) const;
609 /// computeMaskedBitsForTargetNode - Determine which of the bits specified
610 /// in Mask are known to be either zero or one and return them in the
611 /// KnownZero/KnownOne bitsets.
612 virtual void computeMaskedBitsForTargetNode(const SDValue Op,
615 const SelectionDAG &DAG,
616 unsigned Depth = 0) const;
618 // ComputeNumSignBitsForTargetNode - Determine the number of bits in the
619 // operation that are sign bits.
620 virtual unsigned ComputeNumSignBitsForTargetNode(SDValue Op,
621 unsigned Depth) const;
624 isGAPlusOffset(SDNode *N, const GlobalValue* &GA, int64_t &Offset) const;
626 SDValue getReturnAddressFrameIndex(SelectionDAG &DAG) const;
628 virtual bool ExpandInlineAsm(CallInst *CI) const;
630 ConstraintType getConstraintType(const std::string &Constraint) const;
632 /// Examine constraint string and operand type and determine a weight value.
633 /// The operand object must already have been set up with the operand type.
634 virtual ConstraintWeight getSingleConstraintMatchWeight(
635 AsmOperandInfo &info, const char *constraint) const;
637 virtual const char *LowerXConstraint(EVT ConstraintVT) const;
639 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
640 /// vector. If it is invalid, don't add anything to Ops. If hasMemory is
641 /// true it means one of the asm constraint of the inline asm instruction
642 /// being processed is 'm'.
643 virtual void LowerAsmOperandForConstraint(SDValue Op,
644 std::string &Constraint,
645 std::vector<SDValue> &Ops,
646 SelectionDAG &DAG) const;
648 /// getRegForInlineAsmConstraint - Given a physical register constraint
649 /// (e.g. {edx}), return the register number and the register class for the
650 /// register. This should only be used for C_Register constraints. On
651 /// error, this returns a register number of 0.
652 std::pair<unsigned, const TargetRegisterClass*>
653 getRegForInlineAsmConstraint(const std::string &Constraint,
656 /// isLegalAddressingMode - Return true if the addressing mode represented
657 /// by AM is legal for this target, for a load/store of the specified type.
658 virtual bool isLegalAddressingMode(const AddrMode &AM, Type *Ty)const;
660 /// isLegalICmpImmediate - Return true if the specified immediate is legal
661 /// icmp immediate, that is the target has icmp instructions which can
662 /// compare a register against the immediate without having to materialize
663 /// the immediate into a register.
664 virtual bool isLegalICmpImmediate(int64_t Imm) const;
666 /// isLegalAddImmediate - Return true if the specified immediate is legal
667 /// add immediate, that is the target has add instructions which can
668 /// add a register and the immediate without having to materialize
669 /// the immediate into a register.
670 virtual bool isLegalAddImmediate(int64_t Imm) const;
672 /// isTruncateFree - Return true if it's free to truncate a value of
673 /// type Ty1 to type Ty2. e.g. On x86 it's free to truncate a i32 value in
674 /// register EAX to i16 by referencing its sub-register AX.
675 virtual bool isTruncateFree(Type *Ty1, Type *Ty2) const;
676 virtual bool isTruncateFree(EVT VT1, EVT VT2) const;
678 virtual bool allowTruncateForTailCall(Type *Ty1, Type *Ty2) const;
680 /// isZExtFree - Return true if any actual instruction that defines a
681 /// value of type Ty1 implicit zero-extends the value to Ty2 in the result
682 /// register. This does not necessarily include registers defined in
683 /// unknown ways, such as incoming arguments, or copies from unknown
684 /// virtual registers. Also, if isTruncateFree(Ty2, Ty1) is true, this
685 /// does not necessarily apply to truncate instructions. e.g. on x86-64,
686 /// all instructions that define 32-bit values implicit zero-extend the
687 /// result out to 64 bits.
688 virtual bool isZExtFree(Type *Ty1, Type *Ty2) const;
689 virtual bool isZExtFree(EVT VT1, EVT VT2) const;
690 virtual bool isZExtFree(SDValue Val, EVT VT2) const;
692 /// isFMAFasterThanFMulAndFAdd - Return true if an FMA operation is faster
693 /// than a pair of fmul and fadd instructions. fmuladd intrinsics will be
694 /// expanded to FMAs when this method returns true, otherwise fmuladd is
695 /// expanded to fmul + fadd.
696 virtual bool isFMAFasterThanFMulAndFAdd(EVT VT) const;
698 /// isNarrowingProfitable - Return true if it's profitable to narrow
699 /// operations of type VT1 to VT2. e.g. on x86, it's profitable to narrow
700 /// from i32 to i8 but not from i32 to i16.
701 virtual bool isNarrowingProfitable(EVT VT1, EVT VT2) const;
703 /// isFPImmLegal - Returns true if the target can instruction select the
704 /// specified FP immediate natively. If false, the legalizer will
705 /// materialize the FP immediate as a load from a constant pool.
706 virtual bool isFPImmLegal(const APFloat &Imm, EVT VT) const;
708 /// isShuffleMaskLegal - Targets can use this to indicate that they only
709 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
710 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask
711 /// values are assumed to be legal.
712 virtual bool isShuffleMaskLegal(const SmallVectorImpl<int> &Mask,
715 /// isVectorClearMaskLegal - Similar to isShuffleMaskLegal. This is
716 /// used by Targets can use this to indicate if there is a suitable
717 /// VECTOR_SHUFFLE that can be used to replace a VAND with a constant
719 virtual bool isVectorClearMaskLegal(const SmallVectorImpl<int> &Mask,
722 /// ShouldShrinkFPConstant - If true, then instruction selection should
723 /// seek to shrink the FP constant of the specified type to a smaller type
724 /// in order to save space and / or reduce runtime.
725 virtual bool ShouldShrinkFPConstant(EVT VT) const {
726 // Don't shrink FP constpool if SSE2 is available since cvtss2sd is more
727 // expensive than a straight movsd. On the other hand, it's important to
728 // shrink long double fp constant since fldt is very slow.
729 return !X86ScalarSSEf64 || VT == MVT::f80;
732 const X86Subtarget* getSubtarget() const {
736 /// isScalarFPTypeInSSEReg - Return true if the specified scalar FP type is
737 /// computed in an SSE register, not on the X87 floating point stack.
738 bool isScalarFPTypeInSSEReg(EVT VT) const {
739 return (VT == MVT::f64 && X86ScalarSSEf64) || // f64 is when SSE2
740 (VT == MVT::f32 && X86ScalarSSEf32); // f32 is when SSE1
743 /// isTargetFTOL - Return true if the target uses the MSVC _ftol2 routine
745 bool isTargetFTOL() const {
746 return Subtarget->isTargetWindows() && !Subtarget->is64Bit();
749 /// isIntegerTypeFTOL - Return true if the MSVC _ftol2 routine should be
750 /// used for fptoui to the given type.
751 bool isIntegerTypeFTOL(EVT VT) const {
752 return isTargetFTOL() && VT == MVT::i64;
755 /// createFastISel - This method returns a target specific FastISel object,
756 /// or null if the target does not support "fast" ISel.
757 virtual FastISel *createFastISel(FunctionLoweringInfo &funcInfo,
758 const TargetLibraryInfo *libInfo) const;
760 /// getStackCookieLocation - Return true if the target stores stack
761 /// protector cookies at a fixed offset in some non-standard address
762 /// space, and populates the address space and offset as
764 virtual bool getStackCookieLocation(unsigned &AddressSpace, unsigned &Offset) const;
766 SDValue BuildFILD(SDValue Op, EVT SrcVT, SDValue Chain, SDValue StackSlot,
767 SelectionDAG &DAG) const;
769 /// \brief Reset the operation actions based on target options.
770 virtual void resetOperationActions();
773 std::pair<const TargetRegisterClass*, uint8_t>
774 findRepresentativeClass(MVT VT) const;
777 /// Subtarget - Keep a pointer to the X86Subtarget around so that we can
778 /// make the right decision when generating code for different targets.
779 const X86Subtarget *Subtarget;
780 const DataLayout *TD;
782 /// Used to store the TargetOptions so that we don't waste time resetting
783 /// the operation actions unless we have to.
786 /// X86ScalarSSEf32, X86ScalarSSEf64 - Select between SSE or x87
787 /// floating point ops.
788 /// When SSE is available, use it for f32 operations.
789 /// When SSE2 is available, use it for f64 operations.
790 bool X86ScalarSSEf32;
791 bool X86ScalarSSEf64;
793 /// LegalFPImmediates - A list of legal fp immediates.
794 std::vector<APFloat> LegalFPImmediates;
796 /// addLegalFPImmediate - Indicate that this x86 target can instruction
797 /// select the specified FP immediate natively.
798 void addLegalFPImmediate(const APFloat& Imm) {
799 LegalFPImmediates.push_back(Imm);
802 SDValue LowerCallResult(SDValue Chain, SDValue InFlag,
803 CallingConv::ID CallConv, bool isVarArg,
804 const SmallVectorImpl<ISD::InputArg> &Ins,
805 SDLoc dl, SelectionDAG &DAG,
806 SmallVectorImpl<SDValue> &InVals) const;
807 SDValue LowerMemArgument(SDValue Chain,
808 CallingConv::ID CallConv,
809 const SmallVectorImpl<ISD::InputArg> &ArgInfo,
810 SDLoc dl, SelectionDAG &DAG,
811 const CCValAssign &VA, MachineFrameInfo *MFI,
813 SDValue LowerMemOpCallTo(SDValue Chain, SDValue StackPtr, SDValue Arg,
814 SDLoc dl, SelectionDAG &DAG,
815 const CCValAssign &VA,
816 ISD::ArgFlagsTy Flags) const;
818 // Call lowering helpers.
820 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
821 /// for tail call optimization. Targets which want to do tail call
822 /// optimization should implement this function.
823 bool IsEligibleForTailCallOptimization(SDValue Callee,
824 CallingConv::ID CalleeCC,
826 bool isCalleeStructRet,
827 bool isCallerStructRet,
829 const SmallVectorImpl<ISD::OutputArg> &Outs,
830 const SmallVectorImpl<SDValue> &OutVals,
831 const SmallVectorImpl<ISD::InputArg> &Ins,
832 SelectionDAG& DAG) const;
833 bool IsCalleePop(bool isVarArg, CallingConv::ID CallConv) const;
834 SDValue EmitTailCallLoadRetAddr(SelectionDAG &DAG, SDValue &OutRetAddr,
835 SDValue Chain, bool IsTailCall, bool Is64Bit,
836 int FPDiff, SDLoc dl) const;
838 unsigned GetAlignedArgumentStackSize(unsigned StackSize,
839 SelectionDAG &DAG) const;
841 std::pair<SDValue,SDValue> FP_TO_INTHelper(SDValue Op, SelectionDAG &DAG,
843 bool isReplace) const;
845 SDValue LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) const;
846 SDValue LowerBUILD_VECTORvXi1(SDValue Op, SelectionDAG &DAG) const;
847 SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) const;
848 SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const;
849 SDValue LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const;
850 SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) const;
851 SDValue LowerBlockAddress(SDValue Op, SelectionDAG &DAG) const;
852 SDValue LowerGlobalAddress(const GlobalValue *GV, SDLoc dl,
853 int64_t Offset, SelectionDAG &DAG) const;
854 SDValue LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) const;
855 SDValue LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const;
856 SDValue LowerExternalSymbol(SDValue Op, SelectionDAG &DAG) const;
857 SDValue LowerShiftParts(SDValue Op, SelectionDAG &DAG) const;
858 SDValue LowerSINT_TO_FP(SDValue Op, SelectionDAG &DAG) const;
859 SDValue LowerUINT_TO_FP(SDValue Op, SelectionDAG &DAG) const;
860 SDValue LowerUINT_TO_FP_i64(SDValue Op, SelectionDAG &DAG) const;
861 SDValue LowerUINT_TO_FP_i32(SDValue Op, SelectionDAG &DAG) const;
862 SDValue lowerUINT_TO_FP_vec(SDValue Op, SelectionDAG &DAG) const;
863 SDValue LowerTRUNCATE(SDValue Op, SelectionDAG &DAG) const;
864 SDValue LowerFP_TO_SINT(SDValue Op, SelectionDAG &DAG) const;
865 SDValue LowerFP_TO_UINT(SDValue Op, SelectionDAG &DAG) const;
866 SDValue LowerFABS(SDValue Op, SelectionDAG &DAG) const;
867 SDValue LowerFNEG(SDValue Op, SelectionDAG &DAG) const;
868 SDValue LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const;
869 SDValue LowerToBT(SDValue And, ISD::CondCode CC,
870 SDLoc dl, SelectionDAG &DAG) const;
871 SDValue LowerSETCC(SDValue Op, SelectionDAG &DAG) const;
872 SDValue LowerSELECT(SDValue Op, SelectionDAG &DAG) const;
873 SDValue LowerBRCOND(SDValue Op, SelectionDAG &DAG) const;
874 SDValue LowerMEMSET(SDValue Op, SelectionDAG &DAG) const;
875 SDValue LowerJumpTable(SDValue Op, SelectionDAG &DAG) const;
876 SDValue LowerDYNAMIC_STACKALLOC(SDValue Op, SelectionDAG &DAG) const;
877 SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG) const;
878 SDValue LowerVAARG(SDValue Op, SelectionDAG &DAG) const;
879 SDValue LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const;
880 SDValue LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const;
881 SDValue LowerFRAME_TO_ARGS_OFFSET(SDValue Op, SelectionDAG &DAG) const;
882 SDValue LowerEH_RETURN(SDValue Op, SelectionDAG &DAG) const;
883 SDValue lowerEH_SJLJ_SETJMP(SDValue Op, SelectionDAG &DAG) const;
884 SDValue lowerEH_SJLJ_LONGJMP(SDValue Op, SelectionDAG &DAG) const;
885 SDValue LowerINIT_TRAMPOLINE(SDValue Op, SelectionDAG &DAG) const;
886 SDValue LowerFLT_ROUNDS_(SDValue Op, SelectionDAG &DAG) const;
887 SDValue LowerSIGN_EXTEND_INREG(SDValue Op, SelectionDAG &DAG) const;
890 LowerFormalArguments(SDValue Chain,
891 CallingConv::ID CallConv, bool isVarArg,
892 const SmallVectorImpl<ISD::InputArg> &Ins,
893 SDLoc dl, SelectionDAG &DAG,
894 SmallVectorImpl<SDValue> &InVals) const;
896 LowerCall(CallLoweringInfo &CLI,
897 SmallVectorImpl<SDValue> &InVals) const;
900 LowerReturn(SDValue Chain,
901 CallingConv::ID CallConv, bool isVarArg,
902 const SmallVectorImpl<ISD::OutputArg> &Outs,
903 const SmallVectorImpl<SDValue> &OutVals,
904 SDLoc dl, SelectionDAG &DAG) const;
906 virtual bool isUsedByReturnOnly(SDNode *N, SDValue &Chain) const;
908 virtual bool mayBeEmittedAsTailCall(CallInst *CI) const;
911 getTypeForExtArgOrReturn(MVT VT, ISD::NodeType ExtendKind) const;
914 CanLowerReturn(CallingConv::ID CallConv, MachineFunction &MF,
916 const SmallVectorImpl<ISD::OutputArg> &Outs,
917 LLVMContext &Context) const;
919 /// Utility function to emit atomic-load-arith operations (and, or, xor,
920 /// nand, max, min, umax, umin). It takes the corresponding instruction to
921 /// expand, the associated machine basic block, and the associated X86
922 /// opcodes for reg/reg.
923 MachineBasicBlock *EmitAtomicLoadArith(MachineInstr *MI,
924 MachineBasicBlock *MBB) const;
926 /// Utility function to emit atomic-load-arith operations (and, or, xor,
927 /// nand, add, sub, swap) for 64-bit operands on 32-bit target.
928 MachineBasicBlock *EmitAtomicLoadArith6432(MachineInstr *MI,
929 MachineBasicBlock *MBB) const;
931 // Utility function to emit the low-level va_arg code for X86-64.
932 MachineBasicBlock *EmitVAARG64WithCustomInserter(
934 MachineBasicBlock *MBB) const;
936 /// Utility function to emit the xmm reg save portion of va_start.
937 MachineBasicBlock *EmitVAStartSaveXMMRegsWithCustomInserter(
938 MachineInstr *BInstr,
939 MachineBasicBlock *BB) const;
941 MachineBasicBlock *EmitLoweredSelect(MachineInstr *I,
942 MachineBasicBlock *BB) const;
944 MachineBasicBlock *EmitLoweredWinAlloca(MachineInstr *MI,
945 MachineBasicBlock *BB) const;
947 MachineBasicBlock *EmitLoweredSegAlloca(MachineInstr *MI,
948 MachineBasicBlock *BB,
951 MachineBasicBlock *EmitLoweredTLSCall(MachineInstr *MI,
952 MachineBasicBlock *BB) const;
954 MachineBasicBlock *emitLoweredTLSAddr(MachineInstr *MI,
955 MachineBasicBlock *BB) const;
957 MachineBasicBlock *emitEHSjLjSetJmp(MachineInstr *MI,
958 MachineBasicBlock *MBB) const;
960 MachineBasicBlock *emitEHSjLjLongJmp(MachineInstr *MI,
961 MachineBasicBlock *MBB) const;
963 /// Emit nodes that will be selected as "test Op0,Op0", or something
964 /// equivalent, for use with the given x86 condition code.
965 SDValue EmitTest(SDValue Op0, unsigned X86CC, SelectionDAG &DAG) const;
967 /// Emit nodes that will be selected as "cmp Op0,Op1", or something
968 /// equivalent, for use with the given x86 condition code.
969 SDValue EmitCmp(SDValue Op0, SDValue Op1, unsigned X86CC,
970 SelectionDAG &DAG) const;
972 /// Convert a comparison if required by the subtarget.
973 SDValue ConvertCmpIfNecessary(SDValue Cmp, SelectionDAG &DAG) const;
977 FastISel *createFastISel(FunctionLoweringInfo &funcInfo,
978 const TargetLibraryInfo *libInfo);
982 #endif // X86ISELLOWERING_H