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 /// FAND - 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.
278 // ADD, SUB, SMUL, etc. - Arithmetic operations with FLAGS results.
279 ADD, SUB, ADC, SBB, SMUL,
280 INC, DEC, OR, XOR, AND,
282 BLSI, // BLSI - Extract lowest set isolated bit
283 BLSMSK, // BLSMSK - Get mask up to lowest set bit
284 BLSR, // BLSR - Reset lowest set bit
286 UMUL, // LOW, HI, FLAGS = umul LHS, RHS
288 // MUL_IMM - X86 specific multiply by immediate.
291 // PTEST - Vector bitwise comparisons
294 // TESTP - Vector packed fp sign bitwise comparisons
297 // OR/AND test for masks
301 // Several flavors of instructions with vector shuffle behaviors.
325 // PMULUDQ - Vector multiply packed unsigned doubleword integers
336 // VASTART_SAVE_XMM_REGS - Save xmm argument registers to the stack,
337 // according to %al. An operator is needed so that this can be expanded
338 // with control flow.
339 VASTART_SAVE_XMM_REGS,
341 // WIN_ALLOCA - Windows's _chkstk call to do stack probing.
344 // SEG_ALLOCA - For allocating variable amounts of stack space when using
345 // segmented stacks. Check if the current stacklet has enough space, and
346 // falls back to heap allocation if not.
349 // WIN_FTOL - Windows's _ftol2 runtime routine to do fptoui.
358 // FNSTSW16r - Store FP status word into i16 register.
361 // SAHF - Store contents of %ah into %eflags.
364 // RDRAND - Get a random integer and indicate whether it is valid in CF.
367 // RDSEED - Get a NIST SP800-90B & C compliant random integer and
368 // indicate whether it is valid in CF.
375 // XTEST - Test if in transactional execution.
378 // ATOMADD64_DAG, ATOMSUB64_DAG, ATOMOR64_DAG, ATOMAND64_DAG,
379 // ATOMXOR64_DAG, ATOMNAND64_DAG, ATOMSWAP64_DAG -
380 // Atomic 64-bit binary operations.
381 ATOMADD64_DAG = ISD::FIRST_TARGET_MEMORY_OPCODE,
393 // LCMPXCHG_DAG, LCMPXCHG8_DAG, LCMPXCHG16_DAG - Compare and swap.
398 // VZEXT_LOAD - Load, scalar_to_vector, and zero extend.
401 // FNSTCW16m - Store FP control world into i16 memory.
404 /// FP_TO_INT*_IN_MEM - This instruction implements FP_TO_SINT with the
405 /// integer destination in memory and a FP reg source. This corresponds
406 /// to the X86::FIST*m instructions and the rounding mode change stuff. It
407 /// has two inputs (token chain and address) and two outputs (int value
408 /// and token chain).
413 /// FILD, FILD_FLAG - This instruction implements SINT_TO_FP with the
414 /// integer source in memory and FP reg result. This corresponds to the
415 /// X86::FILD*m instructions. It has three inputs (token chain, address,
416 /// and source type) and two outputs (FP value and token chain). FILD_FLAG
417 /// also produces a flag).
421 /// FLD - This instruction implements an extending load to FP stack slots.
422 /// This corresponds to the X86::FLD32m / X86::FLD64m. It takes a chain
423 /// operand, ptr to load from, and a ValueType node indicating the type
427 /// FST - This instruction implements a truncating store to FP stack
428 /// slots. This corresponds to the X86::FST32m / X86::FST64m. It takes a
429 /// chain operand, value to store, address, and a ValueType to store it
433 /// VAARG_64 - This instruction grabs the address of the next argument
434 /// from a va_list. (reads and modifies the va_list in memory)
437 // WARNING: Do not add anything in the end unless you want the node to
438 // have memop! In fact, starting from ATOMADD64_DAG all opcodes will be
439 // thought as target memory ops!
443 /// Define some predicates that are used for node matching.
445 /// isVEXTRACT128Index - Return true if the specified
446 /// EXTRACT_SUBVECTOR operand specifies a vector extract that is
447 /// suitable for input to VEXTRACTF128, VEXTRACTI128 instructions.
448 bool isVEXTRACT128Index(SDNode *N);
450 /// isVINSERT128Index - Return true if the specified
451 /// INSERT_SUBVECTOR operand specifies a subvector insert that is
452 /// suitable for input to VINSERTF128, VINSERTI128 instructions.
453 bool isVINSERT128Index(SDNode *N);
455 /// isVEXTRACT256Index - Return true if the specified
456 /// EXTRACT_SUBVECTOR operand specifies a vector extract that is
457 /// suitable for input to VEXTRACTF64X4, VEXTRACTI64X4 instructions.
458 bool isVEXTRACT256Index(SDNode *N);
460 /// isVINSERT256Index - Return true if the specified
461 /// INSERT_SUBVECTOR operand specifies a subvector insert that is
462 /// suitable for input to VINSERTF64X4, VINSERTI64X4 instructions.
463 bool isVINSERT256Index(SDNode *N);
465 /// getExtractVEXTRACT128Immediate - Return the appropriate
466 /// immediate to extract the specified EXTRACT_SUBVECTOR index
467 /// with VEXTRACTF128, VEXTRACTI128 instructions.
468 unsigned getExtractVEXTRACT128Immediate(SDNode *N);
470 /// getInsertVINSERT128Immediate - Return the appropriate
471 /// immediate to insert at the specified INSERT_SUBVECTOR index
472 /// with VINSERTF128, VINSERT128 instructions.
473 unsigned getInsertVINSERT128Immediate(SDNode *N);
475 /// getExtractVEXTRACT256Immediate - Return the appropriate
476 /// immediate to extract the specified EXTRACT_SUBVECTOR index
477 /// with VEXTRACTF64X4, VEXTRACTI64x4 instructions.
478 unsigned getExtractVEXTRACT256Immediate(SDNode *N);
480 /// getInsertVINSERT256Immediate - Return the appropriate
481 /// immediate to insert at the specified INSERT_SUBVECTOR index
482 /// with VINSERTF64x4, VINSERTI64x4 instructions.
483 unsigned getInsertVINSERT256Immediate(SDNode *N);
485 /// isZeroNode - Returns true if Elt is a constant zero or a floating point
487 bool isZeroNode(SDValue Elt);
489 /// isOffsetSuitableForCodeModel - Returns true of the given offset can be
490 /// fit into displacement field of the instruction.
491 bool isOffsetSuitableForCodeModel(int64_t Offset, CodeModel::Model M,
492 bool hasSymbolicDisplacement = true);
495 /// isCalleePop - Determines whether the callee is required to pop its
496 /// own arguments. Callee pop is necessary to support tail calls.
497 bool isCalleePop(CallingConv::ID CallingConv,
498 bool is64Bit, bool IsVarArg, bool TailCallOpt);
501 //===--------------------------------------------------------------------===//
502 // X86TargetLowering - X86 Implementation of the TargetLowering interface
503 class X86TargetLowering : public TargetLowering {
505 explicit X86TargetLowering(X86TargetMachine &TM);
507 virtual unsigned getJumpTableEncoding() const;
509 virtual MVT getScalarShiftAmountTy(EVT LHSTy) const { return MVT::i8; }
511 virtual const MCExpr *
512 LowerCustomJumpTableEntry(const MachineJumpTableInfo *MJTI,
513 const MachineBasicBlock *MBB, unsigned uid,
514 MCContext &Ctx) const;
516 /// getPICJumpTableRelocaBase - Returns relocation base for the given PIC
518 virtual SDValue getPICJumpTableRelocBase(SDValue Table,
519 SelectionDAG &DAG) const;
520 virtual const MCExpr *
521 getPICJumpTableRelocBaseExpr(const MachineFunction *MF,
522 unsigned JTI, MCContext &Ctx) const;
524 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
525 /// function arguments in the caller parameter area. For X86, aggregates
526 /// that contains are placed at 16-byte boundaries while the rest are at
527 /// 4-byte boundaries.
528 virtual unsigned getByValTypeAlignment(Type *Ty) const;
530 /// getOptimalMemOpType - Returns the target specific optimal type for load
531 /// and store operations as a result of memset, memcpy, and memmove
532 /// lowering. If DstAlign is zero that means it's safe to destination
533 /// alignment can satisfy any constraint. Similarly if SrcAlign is zero it
534 /// means there isn't a need to check it against alignment requirement,
535 /// probably because the source does not need to be loaded. If 'IsMemset' is
536 /// true, that means it's expanding a memset. If 'ZeroMemset' is true, that
537 /// means it's a memset of zero. 'MemcpyStrSrc' indicates whether the memcpy
538 /// source is constant so it does not need to be loaded.
539 /// It returns EVT::Other if the type should be determined using generic
540 /// target-independent logic.
542 getOptimalMemOpType(uint64_t Size, unsigned DstAlign, unsigned SrcAlign,
543 bool IsMemset, bool ZeroMemset, bool MemcpyStrSrc,
544 MachineFunction &MF) const;
546 /// isSafeMemOpType - Returns true if it's safe to use load / store of the
547 /// specified type to expand memcpy / memset inline. This is mostly true
548 /// for all types except for some special cases. For example, on X86
549 /// targets without SSE2 f64 load / store are done with fldl / fstpl which
550 /// also does type conversion. Note the specified type doesn't have to be
551 /// legal as the hook is used before type legalization.
552 virtual bool isSafeMemOpType(MVT VT) const;
554 /// allowsUnalignedMemoryAccesses - Returns true if the target allows
555 /// unaligned memory accesses. of the specified type. Returns whether it
556 /// is "fast" by reference in the second argument.
557 virtual bool allowsUnalignedMemoryAccesses(EVT VT, bool *Fast) const;
559 /// LowerOperation - Provide custom lowering hooks for some operations.
561 virtual SDValue LowerOperation(SDValue Op, SelectionDAG &DAG) const;
563 /// ReplaceNodeResults - Replace the results of node with an illegal result
564 /// type with new values built out of custom code.
566 virtual void ReplaceNodeResults(SDNode *N, SmallVectorImpl<SDValue>&Results,
567 SelectionDAG &DAG) const;
570 virtual SDValue PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const;
572 /// isTypeDesirableForOp - Return true if the target has native support for
573 /// the specified value type and it is 'desirable' to use the type for the
574 /// given node type. e.g. On x86 i16 is legal, but undesirable since i16
575 /// instruction encodings are longer and some i16 instructions are slow.
576 virtual bool isTypeDesirableForOp(unsigned Opc, EVT VT) const;
578 /// isTypeDesirable - Return true if the target has native support for the
579 /// specified value type and it is 'desirable' to use the type. e.g. On x86
580 /// i16 is legal, but undesirable since i16 instruction encodings are longer
581 /// and some i16 instructions are slow.
582 virtual bool IsDesirableToPromoteOp(SDValue Op, EVT &PVT) const;
584 virtual MachineBasicBlock *
585 EmitInstrWithCustomInserter(MachineInstr *MI,
586 MachineBasicBlock *MBB) const;
589 /// getTargetNodeName - This method returns the name of a target specific
591 virtual const char *getTargetNodeName(unsigned Opcode) const;
593 /// getSetCCResultType - Return the value type to use for ISD::SETCC.
594 virtual EVT getSetCCResultType(LLVMContext &Context, EVT VT) const;
596 /// computeMaskedBitsForTargetNode - Determine which of the bits specified
597 /// in Mask are known to be either zero or one and return them in the
598 /// KnownZero/KnownOne bitsets.
599 virtual void computeMaskedBitsForTargetNode(const SDValue Op,
602 const SelectionDAG &DAG,
603 unsigned Depth = 0) const;
605 // ComputeNumSignBitsForTargetNode - Determine the number of bits in the
606 // operation that are sign bits.
607 virtual unsigned ComputeNumSignBitsForTargetNode(SDValue Op,
608 unsigned Depth) const;
611 isGAPlusOffset(SDNode *N, const GlobalValue* &GA, int64_t &Offset) const;
613 SDValue getReturnAddressFrameIndex(SelectionDAG &DAG) const;
615 virtual bool ExpandInlineAsm(CallInst *CI) const;
617 ConstraintType getConstraintType(const std::string &Constraint) const;
619 /// Examine constraint string and operand type and determine a weight value.
620 /// The operand object must already have been set up with the operand type.
621 virtual ConstraintWeight getSingleConstraintMatchWeight(
622 AsmOperandInfo &info, const char *constraint) const;
624 virtual const char *LowerXConstraint(EVT ConstraintVT) const;
626 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
627 /// vector. If it is invalid, don't add anything to Ops. If hasMemory is
628 /// true it means one of the asm constraint of the inline asm instruction
629 /// being processed is 'm'.
630 virtual void LowerAsmOperandForConstraint(SDValue Op,
631 std::string &Constraint,
632 std::vector<SDValue> &Ops,
633 SelectionDAG &DAG) const;
635 /// getRegForInlineAsmConstraint - Given a physical register constraint
636 /// (e.g. {edx}), return the register number and the register class for the
637 /// register. This should only be used for C_Register constraints. On
638 /// error, this returns a register number of 0.
639 std::pair<unsigned, const TargetRegisterClass*>
640 getRegForInlineAsmConstraint(const std::string &Constraint,
643 /// isLegalAddressingMode - Return true if the addressing mode represented
644 /// by AM is legal for this target, for a load/store of the specified type.
645 virtual bool isLegalAddressingMode(const AddrMode &AM, Type *Ty)const;
647 /// isLegalICmpImmediate - Return true if the specified immediate is legal
648 /// icmp immediate, that is the target has icmp instructions which can
649 /// compare a register against the immediate without having to materialize
650 /// the immediate into a register.
651 virtual bool isLegalICmpImmediate(int64_t Imm) const;
653 /// isLegalAddImmediate - Return true if the specified immediate is legal
654 /// add immediate, that is the target has add instructions which can
655 /// add a register and the immediate without having to materialize
656 /// the immediate into a register.
657 virtual bool isLegalAddImmediate(int64_t Imm) const;
659 /// isTruncateFree - Return true if it's free to truncate a value of
660 /// type Ty1 to type Ty2. e.g. On x86 it's free to truncate a i32 value in
661 /// register EAX to i16 by referencing its sub-register AX.
662 virtual bool isTruncateFree(Type *Ty1, Type *Ty2) const;
663 virtual bool isTruncateFree(EVT VT1, EVT VT2) const;
665 /// isZExtFree - Return true if any actual instruction that defines a
666 /// value of type Ty1 implicit zero-extends the value to Ty2 in the result
667 /// register. This does not necessarily include registers defined in
668 /// unknown ways, such as incoming arguments, or copies from unknown
669 /// virtual registers. Also, if isTruncateFree(Ty2, Ty1) is true, this
670 /// does not necessarily apply to truncate instructions. e.g. on x86-64,
671 /// all instructions that define 32-bit values implicit zero-extend the
672 /// result out to 64 bits.
673 virtual bool isZExtFree(Type *Ty1, Type *Ty2) const;
674 virtual bool isZExtFree(EVT VT1, EVT VT2) const;
675 virtual bool isZExtFree(SDValue Val, EVT VT2) const;
677 /// isFMAFasterThanFMulAndFAdd - Return true if an FMA operation is faster
678 /// than a pair of fmul and fadd instructions. fmuladd intrinsics will be
679 /// expanded to FMAs when this method returns true, otherwise fmuladd is
680 /// expanded to fmul + fadd.
681 virtual bool isFMAFasterThanFMulAndFAdd(EVT VT) const;
683 /// isNarrowingProfitable - Return true if it's profitable to narrow
684 /// operations of type VT1 to VT2. e.g. on x86, it's profitable to narrow
685 /// from i32 to i8 but not from i32 to i16.
686 virtual bool isNarrowingProfitable(EVT VT1, EVT VT2) const;
688 /// isFPImmLegal - Returns true if the target can instruction select the
689 /// specified FP immediate natively. If false, the legalizer will
690 /// materialize the FP immediate as a load from a constant pool.
691 virtual bool isFPImmLegal(const APFloat &Imm, EVT VT) const;
693 /// isShuffleMaskLegal - Targets can use this to indicate that they only
694 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
695 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask
696 /// values are assumed to be legal.
697 virtual bool isShuffleMaskLegal(const SmallVectorImpl<int> &Mask,
700 /// isVectorClearMaskLegal - Similar to isShuffleMaskLegal. This is
701 /// used by Targets can use this to indicate if there is a suitable
702 /// VECTOR_SHUFFLE that can be used to replace a VAND with a constant
704 virtual bool isVectorClearMaskLegal(const SmallVectorImpl<int> &Mask,
707 /// ShouldShrinkFPConstant - If true, then instruction selection should
708 /// seek to shrink the FP constant of the specified type to a smaller type
709 /// in order to save space and / or reduce runtime.
710 virtual bool ShouldShrinkFPConstant(EVT VT) const {
711 // Don't shrink FP constpool if SSE2 is available since cvtss2sd is more
712 // expensive than a straight movsd. On the other hand, it's important to
713 // shrink long double fp constant since fldt is very slow.
714 return !X86ScalarSSEf64 || VT == MVT::f80;
717 const X86Subtarget* getSubtarget() const {
721 /// isScalarFPTypeInSSEReg - Return true if the specified scalar FP type is
722 /// computed in an SSE register, not on the X87 floating point stack.
723 bool isScalarFPTypeInSSEReg(EVT VT) const {
724 return (VT == MVT::f64 && X86ScalarSSEf64) || // f64 is when SSE2
725 (VT == MVT::f32 && X86ScalarSSEf32); // f32 is when SSE1
728 /// isTargetFTOL - Return true if the target uses the MSVC _ftol2 routine
730 bool isTargetFTOL() const {
731 return Subtarget->isTargetWindows() && !Subtarget->is64Bit();
734 /// isIntegerTypeFTOL - Return true if the MSVC _ftol2 routine should be
735 /// used for fptoui to the given type.
736 bool isIntegerTypeFTOL(EVT VT) const {
737 return isTargetFTOL() && VT == MVT::i64;
740 /// createFastISel - This method returns a target specific FastISel object,
741 /// or null if the target does not support "fast" ISel.
742 virtual FastISel *createFastISel(FunctionLoweringInfo &funcInfo,
743 const TargetLibraryInfo *libInfo) const;
745 /// getStackCookieLocation - Return true if the target stores stack
746 /// protector cookies at a fixed offset in some non-standard address
747 /// space, and populates the address space and offset as
749 virtual bool getStackCookieLocation(unsigned &AddressSpace, unsigned &Offset) const;
751 SDValue BuildFILD(SDValue Op, EVT SrcVT, SDValue Chain, SDValue StackSlot,
752 SelectionDAG &DAG) const;
754 /// \brief Reset the operation actions based on target options.
755 virtual void resetOperationActions();
758 std::pair<const TargetRegisterClass*, uint8_t>
759 findRepresentativeClass(MVT VT) const;
762 /// Subtarget - Keep a pointer to the X86Subtarget around so that we can
763 /// make the right decision when generating code for different targets.
764 const X86Subtarget *Subtarget;
765 const DataLayout *TD;
767 /// Used to store the TargetOptions so that we don't waste time resetting
768 /// the operation actions unless we have to.
771 /// X86ScalarSSEf32, X86ScalarSSEf64 - Select between SSE or x87
772 /// floating point ops.
773 /// When SSE is available, use it for f32 operations.
774 /// When SSE2 is available, use it for f64 operations.
775 bool X86ScalarSSEf32;
776 bool X86ScalarSSEf64;
778 /// LegalFPImmediates - A list of legal fp immediates.
779 std::vector<APFloat> LegalFPImmediates;
781 /// addLegalFPImmediate - Indicate that this x86 target can instruction
782 /// select the specified FP immediate natively.
783 void addLegalFPImmediate(const APFloat& Imm) {
784 LegalFPImmediates.push_back(Imm);
787 SDValue LowerCallResult(SDValue Chain, SDValue InFlag,
788 CallingConv::ID CallConv, bool isVarArg,
789 const SmallVectorImpl<ISD::InputArg> &Ins,
790 SDLoc dl, SelectionDAG &DAG,
791 SmallVectorImpl<SDValue> &InVals) const;
792 SDValue LowerMemArgument(SDValue Chain,
793 CallingConv::ID CallConv,
794 const SmallVectorImpl<ISD::InputArg> &ArgInfo,
795 SDLoc dl, SelectionDAG &DAG,
796 const CCValAssign &VA, MachineFrameInfo *MFI,
798 SDValue LowerMemOpCallTo(SDValue Chain, SDValue StackPtr, SDValue Arg,
799 SDLoc dl, SelectionDAG &DAG,
800 const CCValAssign &VA,
801 ISD::ArgFlagsTy Flags) const;
803 // Call lowering helpers.
805 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
806 /// for tail call optimization. Targets which want to do tail call
807 /// optimization should implement this function.
808 bool IsEligibleForTailCallOptimization(SDValue Callee,
809 CallingConv::ID CalleeCC,
811 bool isCalleeStructRet,
812 bool isCallerStructRet,
814 const SmallVectorImpl<ISD::OutputArg> &Outs,
815 const SmallVectorImpl<SDValue> &OutVals,
816 const SmallVectorImpl<ISD::InputArg> &Ins,
817 SelectionDAG& DAG) const;
818 bool IsCalleePop(bool isVarArg, CallingConv::ID CallConv) const;
819 SDValue EmitTailCallLoadRetAddr(SelectionDAG &DAG, SDValue &OutRetAddr,
820 SDValue Chain, bool IsTailCall, bool Is64Bit,
821 int FPDiff, SDLoc dl) const;
823 unsigned GetAlignedArgumentStackSize(unsigned StackSize,
824 SelectionDAG &DAG) const;
826 std::pair<SDValue,SDValue> FP_TO_INTHelper(SDValue Op, SelectionDAG &DAG,
828 bool isReplace) const;
830 SDValue LowerAsSplatVectorLoad(SDValue SrcOp, EVT VT, SDLoc dl,
831 SelectionDAG &DAG) const;
832 SDValue LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) const;
833 SDValue LowerBUILD_VECTORvXi1(SDValue Op, SelectionDAG &DAG) const;
834 SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) const;
835 SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const;
836 SDValue LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const;
837 SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) const;
838 SDValue LowerBlockAddress(SDValue Op, SelectionDAG &DAG) const;
839 SDValue LowerGlobalAddress(const GlobalValue *GV, SDLoc dl,
840 int64_t Offset, SelectionDAG &DAG) const;
841 SDValue LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) const;
842 SDValue LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const;
843 SDValue LowerExternalSymbol(SDValue Op, SelectionDAG &DAG) const;
844 SDValue LowerShiftParts(SDValue Op, SelectionDAG &DAG) const;
845 SDValue LowerBITCAST(SDValue op, SelectionDAG &DAG) const;
846 SDValue LowerSINT_TO_FP(SDValue Op, SelectionDAG &DAG) const;
847 SDValue LowerUINT_TO_FP(SDValue Op, SelectionDAG &DAG) const;
848 SDValue LowerUINT_TO_FP_i64(SDValue Op, SelectionDAG &DAG) const;
849 SDValue LowerUINT_TO_FP_i32(SDValue Op, SelectionDAG &DAG) const;
850 SDValue lowerUINT_TO_FP_vec(SDValue Op, SelectionDAG &DAG) const;
851 SDValue LowerTRUNCATE(SDValue Op, SelectionDAG &DAG) const;
852 SDValue LowerZERO_EXTEND(SDValue Op, SelectionDAG &DAG) const;
853 SDValue LowerSIGN_EXTEND(SDValue Op, SelectionDAG &DAG) const;
854 SDValue LowerANY_EXTEND(SDValue Op, SelectionDAG &DAG) const;
855 SDValue LowerFP_TO_SINT(SDValue Op, SelectionDAG &DAG) const;
856 SDValue LowerFP_TO_UINT(SDValue Op, SelectionDAG &DAG) const;
857 SDValue LowerFABS(SDValue Op, SelectionDAG &DAG) const;
858 SDValue LowerFNEG(SDValue Op, SelectionDAG &DAG) const;
859 SDValue LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const;
860 SDValue LowerToBT(SDValue And, ISD::CondCode CC,
861 SDLoc dl, SelectionDAG &DAG) const;
862 SDValue LowerSETCC(SDValue Op, SelectionDAG &DAG) const;
863 SDValue LowerSELECT(SDValue Op, SelectionDAG &DAG) const;
864 SDValue LowerBRCOND(SDValue Op, SelectionDAG &DAG) const;
865 SDValue LowerMEMSET(SDValue Op, SelectionDAG &DAG) const;
866 SDValue LowerJumpTable(SDValue Op, SelectionDAG &DAG) const;
867 SDValue LowerDYNAMIC_STACKALLOC(SDValue Op, SelectionDAG &DAG) const;
868 SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG) const;
869 SDValue LowerVAARG(SDValue Op, SelectionDAG &DAG) const;
870 SDValue LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const;
871 SDValue LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const;
872 SDValue LowerFRAME_TO_ARGS_OFFSET(SDValue Op, SelectionDAG &DAG) const;
873 SDValue LowerEH_RETURN(SDValue Op, SelectionDAG &DAG) const;
874 SDValue lowerEH_SJLJ_SETJMP(SDValue Op, SelectionDAG &DAG) const;
875 SDValue lowerEH_SJLJ_LONGJMP(SDValue Op, SelectionDAG &DAG) const;
876 SDValue LowerINIT_TRAMPOLINE(SDValue Op, SelectionDAG &DAG) const;
877 SDValue LowerFLT_ROUNDS_(SDValue Op, SelectionDAG &DAG) const;
878 SDValue LowerShift(SDValue Op, SelectionDAG &DAG) const;
879 SDValue LowerSDIV(SDValue Op, SelectionDAG &DAG) const;
880 SDValue LowerSIGN_EXTEND_INREG(SDValue Op, SelectionDAG &DAG) const;
881 SDValue LowerFSINCOS(SDValue Op, SelectionDAG &DAG) const;
883 // Utility functions to help LowerVECTOR_SHUFFLE & LowerBUILD_VECTOR
884 SDValue LowerVectorBroadcast(SDValue Op, SelectionDAG &DAG) const;
885 SDValue NormalizeVectorShuffle(SDValue Op, SelectionDAG &DAG) const;
886 SDValue buildFromShuffleMostly(SDValue Op, SelectionDAG &DAG) const;
888 SDValue LowerVectorAllZeroTest(SDValue Op, SelectionDAG &DAG) const;
890 SDValue LowerVectorIntExtend(SDValue Op, SelectionDAG &DAG) const;
893 LowerFormalArguments(SDValue Chain,
894 CallingConv::ID CallConv, bool isVarArg,
895 const SmallVectorImpl<ISD::InputArg> &Ins,
896 SDLoc dl, SelectionDAG &DAG,
897 SmallVectorImpl<SDValue> &InVals) const;
899 LowerCall(CallLoweringInfo &CLI,
900 SmallVectorImpl<SDValue> &InVals) const;
903 LowerReturn(SDValue Chain,
904 CallingConv::ID CallConv, bool isVarArg,
905 const SmallVectorImpl<ISD::OutputArg> &Outs,
906 const SmallVectorImpl<SDValue> &OutVals,
907 SDLoc dl, SelectionDAG &DAG) const;
909 virtual bool isUsedByReturnOnly(SDNode *N, SDValue &Chain) const;
911 virtual bool mayBeEmittedAsTailCall(CallInst *CI) const;
914 getTypeForExtArgOrReturn(MVT VT, ISD::NodeType ExtendKind) const;
917 CanLowerReturn(CallingConv::ID CallConv, MachineFunction &MF,
919 const SmallVectorImpl<ISD::OutputArg> &Outs,
920 LLVMContext &Context) const;
922 /// Utility function to emit atomic-load-arith operations (and, or, xor,
923 /// nand, max, min, umax, umin). It takes the corresponding instruction to
924 /// expand, the associated machine basic block, and the associated X86
925 /// opcodes for reg/reg.
926 MachineBasicBlock *EmitAtomicLoadArith(MachineInstr *MI,
927 MachineBasicBlock *MBB) const;
929 /// Utility function to emit atomic-load-arith operations (and, or, xor,
930 /// nand, add, sub, swap) for 64-bit operands on 32-bit target.
931 MachineBasicBlock *EmitAtomicLoadArith6432(MachineInstr *MI,
932 MachineBasicBlock *MBB) const;
934 // Utility function to emit the low-level va_arg code for X86-64.
935 MachineBasicBlock *EmitVAARG64WithCustomInserter(
937 MachineBasicBlock *MBB) const;
939 /// Utility function to emit the xmm reg save portion of va_start.
940 MachineBasicBlock *EmitVAStartSaveXMMRegsWithCustomInserter(
941 MachineInstr *BInstr,
942 MachineBasicBlock *BB) const;
944 MachineBasicBlock *EmitLoweredSelect(MachineInstr *I,
945 MachineBasicBlock *BB) const;
947 MachineBasicBlock *EmitLoweredWinAlloca(MachineInstr *MI,
948 MachineBasicBlock *BB) const;
950 MachineBasicBlock *EmitLoweredSegAlloca(MachineInstr *MI,
951 MachineBasicBlock *BB,
954 MachineBasicBlock *EmitLoweredTLSCall(MachineInstr *MI,
955 MachineBasicBlock *BB) const;
957 MachineBasicBlock *emitLoweredTLSAddr(MachineInstr *MI,
958 MachineBasicBlock *BB) const;
960 MachineBasicBlock *emitEHSjLjSetJmp(MachineInstr *MI,
961 MachineBasicBlock *MBB) const;
963 MachineBasicBlock *emitEHSjLjLongJmp(MachineInstr *MI,
964 MachineBasicBlock *MBB) const;
966 /// Emit nodes that will be selected as "test Op0,Op0", or something
967 /// equivalent, for use with the given x86 condition code.
968 SDValue EmitTest(SDValue Op0, unsigned X86CC, SelectionDAG &DAG) const;
970 /// Emit nodes that will be selected as "cmp Op0,Op1", or something
971 /// equivalent, for use with the given x86 condition code.
972 SDValue EmitCmp(SDValue Op0, SDValue Op1, unsigned X86CC,
973 SelectionDAG &DAG) const;
975 /// Convert a comparison if required by the subtarget.
976 SDValue ConvertCmpIfNecessary(SDValue Cmp, SelectionDAG &DAG) const;
980 FastISel *createFastISel(FunctionLoweringInfo &funcInfo,
981 const TargetLibraryInfo *libInfo);
985 #endif // X86ISELLOWERING_H