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"
26 #include "llvm/Target/TargetTransformImpl.h"
30 // X86 Specific DAG Nodes
32 // Start the numbering where the builtin ops leave off.
33 FIRST_NUMBER = ISD::BUILTIN_OP_END,
35 /// BSF - Bit scan forward.
36 /// BSR - Bit scan reverse.
40 /// SHLD, SHRD - Double shift instructions. These correspond to
41 /// X86::SHLDxx and X86::SHRDxx instructions.
45 /// FAND - Bitwise logical AND of floating point values. This corresponds
46 /// to X86::ANDPS or X86::ANDPD.
49 /// FOR - Bitwise logical OR of floating point values. This corresponds
50 /// to X86::ORPS or X86::ORPD.
53 /// FXOR - Bitwise logical XOR of floating point values. This corresponds
54 /// to X86::XORPS or X86::XORPD.
57 /// FSRL - Bitwise logical right shift of floating point values. These
58 /// corresponds to X86::PSRLDQ.
61 /// CALL - These operations represent an abstract X86 call
62 /// instruction, which includes a bunch of information. In particular the
63 /// operands of these node are:
65 /// #0 - The incoming token chain
67 /// #2 - The number of arg bytes the caller pushes on the stack.
68 /// #3 - The number of arg bytes the callee pops off the stack.
69 /// #4 - The value to pass in AL/AX/EAX (optional)
70 /// #5 - The value to pass in DL/DX/EDX (optional)
72 /// The result values of these nodes are:
74 /// #0 - The outgoing token chain
75 /// #1 - The first register result value (optional)
76 /// #2 - The second register result value (optional)
80 /// RDTSC_DAG - This operation implements the lowering for
84 /// X86 compare and logical compare instructions.
87 /// X86 bit-test instructions.
90 /// X86 SetCC. Operand 0 is condition code, and operand 1 is the EFLAGS
91 /// operand, usually produced by a CMP instruction.
94 // Same as SETCC except it's materialized with a sbb and the value is all
95 // one's or all zero's.
96 SETCC_CARRY, // R = carry_bit ? ~0 : 0
98 /// X86 FP SETCC, implemented with CMP{cc}SS/CMP{cc}SD.
99 /// Operands are two FP values to compare; result is a mask of
100 /// 0s or 1s. Generally DTRT for C/C++ with NaNs.
103 /// X86 MOVMSK{pd|ps}, extracts sign bits of two or four FP values,
104 /// result in an integer GPR. Needs masking for scalar result.
107 /// X86 conditional moves. Operand 0 and operand 1 are the two values
108 /// to select from. Operand 2 is the condition code, and operand 3 is the
109 /// flag operand produced by a CMP or TEST instruction. It also writes a
113 /// X86 conditional branches. Operand 0 is the chain operand, operand 1
114 /// is the block to branch if condition is true, operand 2 is the
115 /// condition code, and operand 3 is the flag operand produced by a CMP
116 /// or TEST instruction.
119 /// Return with a flag operand. Operand 0 is the chain operand, operand
120 /// 1 is the number of bytes of stack to pop.
123 /// REP_STOS - Repeat fill, corresponds to X86::REP_STOSx.
126 /// REP_MOVS - Repeat move, corresponds to X86::REP_MOVSx.
129 /// GlobalBaseReg - On Darwin, this node represents the result of the popl
130 /// at function entry, used for PIC code.
133 /// Wrapper - A wrapper node for TargetConstantPool,
134 /// TargetExternalSymbol, and TargetGlobalAddress.
137 /// WrapperRIP - Special wrapper used under X86-64 PIC mode for RIP
138 /// relative displacements.
141 /// MOVDQ2Q - Copies a 64-bit value from the low word of an XMM vector
142 /// to an MMX vector. If you think this is too close to the previous
143 /// mnemonic, so do I; blame Intel.
146 /// MMX_MOVD2W - Copies a 32-bit value from the low word of a MMX
150 /// PEXTRB - Extract an 8-bit value from a vector and zero extend it to
151 /// i32, corresponds to X86::PEXTRB.
154 /// PEXTRW - Extract a 16-bit value from a vector and zero extend it to
155 /// i32, corresponds to X86::PEXTRW.
158 /// INSERTPS - Insert any element of a 4 x float vector into any element
159 /// of a destination 4 x floatvector.
162 /// PINSRB - Insert the lower 8-bits of a 32-bit value to a vector,
163 /// corresponds to X86::PINSRB.
166 /// PINSRW - Insert the lower 16-bits of a 32-bit value to a vector,
167 /// corresponds to X86::PINSRW.
170 /// PSHUFB - Shuffle 16 8-bit values within a vector.
173 /// ANDNP - Bitwise Logical AND NOT of Packed FP values.
176 /// PSIGN - Copy integer sign.
179 /// BLENDV - Blend where the selector is a register.
182 /// BLENDI - Blend where the selector is an immediate.
185 /// HADD - Integer horizontal add.
188 /// HSUB - Integer horizontal sub.
191 /// FHADD - Floating point horizontal add.
194 /// FHSUB - Floating point horizontal sub.
197 /// FMAX, FMIN - Floating point max and min.
201 /// FMAXC, FMINC - Commutative FMIN and FMAX.
204 /// FRSQRT, FRCP - Floating point reciprocal-sqrt and reciprocal
205 /// approximation. Note that these typically require refinement
206 /// in order to obtain suitable precision.
209 // TLSADDR - Thread Local Storage.
212 // TLSBASEADDR - Thread Local Storage. A call to get the start address
213 // of the TLS block for the current module.
216 // TLSCALL - Thread Local Storage. When calling to an OS provided
217 // thunk at the address from an earlier relocation.
220 // EH_RETURN - Exception Handling helpers.
223 // EH_SJLJ_SETJMP - SjLj exception handling setjmp.
226 // EH_SJLJ_LONGJMP - SjLj exception handling longjmp.
229 /// TC_RETURN - Tail call return.
231 /// operand #1 callee (register or absolute)
232 /// operand #2 stack adjustment
233 /// operand #3 optional in flag
236 // VZEXT_MOVL - Vector move low and zero extend.
239 // VSEXT_MOVL - Vector move low and sign extend.
242 // VZEXT - Vector integer zero-extend.
245 // VSEXT - Vector integer signed-extend.
248 // VFPEXT - Vector FP extend.
251 // VFPROUND - Vector FP round.
254 // VSHL, VSRL - 128-bit vector logical left / right shift
257 // VSHL, VSRL, VSRA - Vector shift elements
260 // VSHLI, VSRLI, VSRAI - Vector shift elements by immediate
263 // CMPP - Vector packed double/float comparison.
266 // PCMP* - Vector integer comparisons.
269 // ADD, SUB, SMUL, etc. - Arithmetic operations with FLAGS results.
270 ADD, SUB, ADC, SBB, SMUL,
271 INC, DEC, OR, XOR, AND,
273 ANDN, // ANDN - Bitwise AND NOT with FLAGS results.
275 BLSI, // BLSI - Extract lowest set isolated bit
276 BLSMSK, // BLSMSK - Get mask up to lowest set bit
277 BLSR, // BLSR - Reset lowest set bit
279 UMUL, // LOW, HI, FLAGS = umul LHS, RHS
281 // MUL_IMM - X86 specific multiply by immediate.
284 // PTEST - Vector bitwise comparisons
287 // TESTP - Vector packed fp sign bitwise comparisons
290 // Several flavors of instructions with vector shuffle behaviors.
314 // PMULUDQ - Vector multiply packed unsigned doubleword integers
325 // VASTART_SAVE_XMM_REGS - Save xmm argument registers to the stack,
326 // according to %al. An operator is needed so that this can be expanded
327 // with control flow.
328 VASTART_SAVE_XMM_REGS,
330 // WIN_ALLOCA - Windows's _chkstk call to do stack probing.
333 // SEG_ALLOCA - For allocating variable amounts of stack space when using
334 // segmented stacks. Check if the current stacklet has enough space, and
335 // falls back to heap allocation if not.
338 // WIN_FTOL - Windows's _ftol2 runtime routine to do fptoui.
347 // FNSTSW16r - Store FP status word into i16 register.
350 // SAHF - Store contents of %ah into %eflags.
353 // RDRAND - Get a random integer and indicate whether it is valid in CF.
360 // ATOMADD64_DAG, ATOMSUB64_DAG, ATOMOR64_DAG, ATOMAND64_DAG,
361 // ATOMXOR64_DAG, ATOMNAND64_DAG, ATOMSWAP64_DAG -
362 // Atomic 64-bit binary operations.
363 ATOMADD64_DAG = ISD::FIRST_TARGET_MEMORY_OPCODE,
375 // LCMPXCHG_DAG, LCMPXCHG8_DAG, LCMPXCHG16_DAG - Compare and swap.
380 // VZEXT_LOAD - Load, scalar_to_vector, and zero extend.
383 // FNSTCW16m - Store FP control world into i16 memory.
386 /// FP_TO_INT*_IN_MEM - This instruction implements FP_TO_SINT with the
387 /// integer destination in memory and a FP reg source. This corresponds
388 /// to the X86::FIST*m instructions and the rounding mode change stuff. It
389 /// has two inputs (token chain and address) and two outputs (int value
390 /// and token chain).
395 /// FILD, FILD_FLAG - This instruction implements SINT_TO_FP with the
396 /// integer source in memory and FP reg result. This corresponds to the
397 /// X86::FILD*m instructions. It has three inputs (token chain, address,
398 /// and source type) and two outputs (FP value and token chain). FILD_FLAG
399 /// also produces a flag).
403 /// FLD - This instruction implements an extending load to FP stack slots.
404 /// This corresponds to the X86::FLD32m / X86::FLD64m. It takes a chain
405 /// operand, ptr to load from, and a ValueType node indicating the type
409 /// FST - This instruction implements a truncating store to FP stack
410 /// slots. This corresponds to the X86::FST32m / X86::FST64m. It takes a
411 /// chain operand, value to store, address, and a ValueType to store it
415 /// VAARG_64 - This instruction grabs the address of the next argument
416 /// from a va_list. (reads and modifies the va_list in memory)
419 // WARNING: Do not add anything in the end unless you want the node to
420 // have memop! In fact, starting from ATOMADD64_DAG all opcodes will be
421 // thought as target memory ops!
425 /// Define some predicates that are used for node matching.
427 /// isVEXTRACTF128Index - Return true if the specified
428 /// EXTRACT_SUBVECTOR operand specifies a vector extract that is
429 /// suitable for input to VEXTRACTF128.
430 bool isVEXTRACTF128Index(SDNode *N);
432 /// isVINSERTF128Index - Return true if the specified
433 /// INSERT_SUBVECTOR operand specifies a subvector insert that is
434 /// suitable for input to VINSERTF128.
435 bool isVINSERTF128Index(SDNode *N);
437 /// getExtractVEXTRACTF128Immediate - Return the appropriate
438 /// immediate to extract the specified EXTRACT_SUBVECTOR index
439 /// with VEXTRACTF128 instructions.
440 unsigned getExtractVEXTRACTF128Immediate(SDNode *N);
442 /// getInsertVINSERTF128Immediate - Return the appropriate
443 /// immediate to insert at the specified INSERT_SUBVECTOR index
444 /// with VINSERTF128 instructions.
445 unsigned getInsertVINSERTF128Immediate(SDNode *N);
447 /// isZeroNode - Returns true if Elt is a constant zero or a floating point
449 bool isZeroNode(SDValue Elt);
451 /// isOffsetSuitableForCodeModel - Returns true of the given offset can be
452 /// fit into displacement field of the instruction.
453 bool isOffsetSuitableForCodeModel(int64_t Offset, CodeModel::Model M,
454 bool hasSymbolicDisplacement = true);
457 /// isCalleePop - Determines whether the callee is required to pop its
458 /// own arguments. Callee pop is necessary to support tail calls.
459 bool isCalleePop(CallingConv::ID CallingConv,
460 bool is64Bit, bool IsVarArg, bool TailCallOpt);
463 //===--------------------------------------------------------------------===//
464 // X86TargetLowering - X86 Implementation of the TargetLowering interface
465 class X86TargetLowering : public TargetLowering {
467 explicit X86TargetLowering(X86TargetMachine &TM);
469 virtual unsigned getJumpTableEncoding() const;
471 virtual MVT getShiftAmountTy(EVT LHSTy) const { return MVT::i8; }
473 virtual const MCExpr *
474 LowerCustomJumpTableEntry(const MachineJumpTableInfo *MJTI,
475 const MachineBasicBlock *MBB, unsigned uid,
476 MCContext &Ctx) const;
478 /// getPICJumpTableRelocaBase - Returns relocation base for the given PIC
480 virtual SDValue getPICJumpTableRelocBase(SDValue Table,
481 SelectionDAG &DAG) const;
482 virtual const MCExpr *
483 getPICJumpTableRelocBaseExpr(const MachineFunction *MF,
484 unsigned JTI, MCContext &Ctx) const;
486 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
487 /// function arguments in the caller parameter area. For X86, aggregates
488 /// that contains are placed at 16-byte boundaries while the rest are at
489 /// 4-byte boundaries.
490 virtual unsigned getByValTypeAlignment(Type *Ty) const;
492 /// getOptimalMemOpType - Returns the target specific optimal type for load
493 /// and store operations as a result of memset, memcpy, and memmove
494 /// lowering. If DstAlign is zero that means it's safe to destination
495 /// alignment can satisfy any constraint. Similarly if SrcAlign is zero it
496 /// means there isn't a need to check it against alignment requirement,
497 /// probably because the source does not need to be loaded. If
498 /// 'IsZeroVal' is true, that means it's safe to return a
499 /// non-scalar-integer type, e.g. empty string source, constant, or loaded
500 /// from memory. 'MemcpyStrSrc' indicates whether the memcpy source is
501 /// constant so it does not need to be loaded.
502 /// It returns EVT::Other if the type should be determined using generic
503 /// target-independent logic.
505 getOptimalMemOpType(uint64_t Size, unsigned DstAlign, unsigned SrcAlign,
506 bool IsZeroVal, bool MemcpyStrSrc,
507 MachineFunction &MF) const;
509 /// allowsUnalignedMemoryAccesses - Returns true if the target allows
510 /// unaligned memory accesses. of the specified type.
511 virtual bool allowsUnalignedMemoryAccesses(EVT VT) const {
515 /// LowerOperation - Provide custom lowering hooks for some operations.
517 virtual SDValue LowerOperation(SDValue Op, SelectionDAG &DAG) const;
519 /// ReplaceNodeResults - Replace the results of node with an illegal result
520 /// type with new values built out of custom code.
522 virtual void ReplaceNodeResults(SDNode *N, SmallVectorImpl<SDValue>&Results,
523 SelectionDAG &DAG) const;
526 virtual SDValue PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const;
528 /// isTypeDesirableForOp - Return true if the target has native support for
529 /// the specified value type and it is 'desirable' to use the type for the
530 /// given node type. e.g. On x86 i16 is legal, but undesirable since i16
531 /// instruction encodings are longer and some i16 instructions are slow.
532 virtual bool isTypeDesirableForOp(unsigned Opc, EVT VT) const;
534 /// isTypeDesirable - Return true if the target has native support for the
535 /// specified value type and it is 'desirable' to use the type. e.g. On x86
536 /// i16 is legal, but undesirable since i16 instruction encodings are longer
537 /// and some i16 instructions are slow.
538 virtual bool IsDesirableToPromoteOp(SDValue Op, EVT &PVT) const;
540 virtual MachineBasicBlock *
541 EmitInstrWithCustomInserter(MachineInstr *MI,
542 MachineBasicBlock *MBB) const;
545 /// getTargetNodeName - This method returns the name of a target specific
547 virtual const char *getTargetNodeName(unsigned Opcode) const;
549 /// getSetCCResultType - Return the value type to use for ISD::SETCC.
550 virtual EVT getSetCCResultType(EVT VT) const;
552 /// computeMaskedBitsForTargetNode - Determine which of the bits specified
553 /// in Mask are known to be either zero or one and return them in the
554 /// KnownZero/KnownOne bitsets.
555 virtual void computeMaskedBitsForTargetNode(const SDValue Op,
558 const SelectionDAG &DAG,
559 unsigned Depth = 0) const;
561 // ComputeNumSignBitsForTargetNode - Determine the number of bits in the
562 // operation that are sign bits.
563 virtual unsigned ComputeNumSignBitsForTargetNode(SDValue Op,
564 unsigned Depth) const;
567 isGAPlusOffset(SDNode *N, const GlobalValue* &GA, int64_t &Offset) const;
569 SDValue getReturnAddressFrameIndex(SelectionDAG &DAG) const;
571 virtual bool ExpandInlineAsm(CallInst *CI) const;
573 ConstraintType getConstraintType(const std::string &Constraint) const;
575 /// Examine constraint string and operand type and determine a weight value.
576 /// The operand object must already have been set up with the operand type.
577 virtual ConstraintWeight getSingleConstraintMatchWeight(
578 AsmOperandInfo &info, const char *constraint) const;
580 virtual const char *LowerXConstraint(EVT ConstraintVT) const;
582 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
583 /// vector. If it is invalid, don't add anything to Ops. If hasMemory is
584 /// true it means one of the asm constraint of the inline asm instruction
585 /// being processed is 'm'.
586 virtual void LowerAsmOperandForConstraint(SDValue Op,
587 std::string &Constraint,
588 std::vector<SDValue> &Ops,
589 SelectionDAG &DAG) const;
591 /// getRegForInlineAsmConstraint - Given a physical register constraint
592 /// (e.g. {edx}), return the register number and the register class for the
593 /// register. This should only be used for C_Register constraints. On
594 /// error, this returns a register number of 0.
595 std::pair<unsigned, const TargetRegisterClass*>
596 getRegForInlineAsmConstraint(const std::string &Constraint,
599 /// isLegalAddressingMode - Return true if the addressing mode represented
600 /// by AM is legal for this target, for a load/store of the specified type.
601 virtual bool isLegalAddressingMode(const AddrMode &AM, Type *Ty)const;
603 /// isLegalICmpImmediate - Return true if the specified immediate is legal
604 /// icmp immediate, that is the target has icmp instructions which can
605 /// compare a register against the immediate without having to materialize
606 /// the immediate into a register.
607 virtual bool isLegalICmpImmediate(int64_t Imm) const;
609 /// isLegalAddImmediate - Return true if the specified immediate is legal
610 /// add immediate, that is the target has add instructions which can
611 /// add a register and the immediate without having to materialize
612 /// the immediate into a register.
613 virtual bool isLegalAddImmediate(int64_t Imm) const;
615 /// isTruncateFree - Return true if it's free to truncate a value of
616 /// type Ty1 to type Ty2. e.g. On x86 it's free to truncate a i32 value in
617 /// register EAX to i16 by referencing its sub-register AX.
618 virtual bool isTruncateFree(Type *Ty1, Type *Ty2) const;
619 virtual bool isTruncateFree(EVT VT1, EVT VT2) const;
621 /// isZExtFree - Return true if any actual instruction that defines a
622 /// value of type Ty1 implicit zero-extends the value to Ty2 in the result
623 /// register. This does not necessarily include registers defined in
624 /// unknown ways, such as incoming arguments, or copies from unknown
625 /// virtual registers. Also, if isTruncateFree(Ty2, Ty1) is true, this
626 /// does not necessarily apply to truncate instructions. e.g. on x86-64,
627 /// all instructions that define 32-bit values implicit zero-extend the
628 /// result out to 64 bits.
629 virtual bool isZExtFree(Type *Ty1, Type *Ty2) const;
630 virtual bool isZExtFree(EVT VT1, EVT VT2) const;
631 virtual bool isZExtFree(SDValue Val, EVT VT2) const;
633 /// isFMAFasterThanMulAndAdd - Return true if an FMA operation is faster than
634 /// a pair of mul and add instructions. fmuladd intrinsics will be expanded to
635 /// FMAs when this method returns true (and FMAs are legal), otherwise fmuladd
636 /// is expanded to mul + add.
637 virtual bool isFMAFasterThanMulAndAdd(EVT) const { return true; }
639 /// isNarrowingProfitable - Return true if it's profitable to narrow
640 /// operations of type VT1 to VT2. e.g. on x86, it's profitable to narrow
641 /// from i32 to i8 but not from i32 to i16.
642 virtual bool isNarrowingProfitable(EVT VT1, EVT VT2) const;
644 /// isFPImmLegal - Returns true if the target can instruction select the
645 /// specified FP immediate natively. If false, the legalizer will
646 /// materialize the FP immediate as a load from a constant pool.
647 virtual bool isFPImmLegal(const APFloat &Imm, EVT VT) const;
649 /// isShuffleMaskLegal - Targets can use this to indicate that they only
650 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
651 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask
652 /// values are assumed to be legal.
653 virtual bool isShuffleMaskLegal(const SmallVectorImpl<int> &Mask,
656 /// isVectorClearMaskLegal - Similar to isShuffleMaskLegal. This is
657 /// used by Targets can use this to indicate if there is a suitable
658 /// VECTOR_SHUFFLE that can be used to replace a VAND with a constant
660 virtual bool isVectorClearMaskLegal(const SmallVectorImpl<int> &Mask,
663 /// ShouldShrinkFPConstant - If true, then instruction selection should
664 /// seek to shrink the FP constant of the specified type to a smaller type
665 /// in order to save space and / or reduce runtime.
666 virtual bool ShouldShrinkFPConstant(EVT VT) const {
667 // Don't shrink FP constpool if SSE2 is available since cvtss2sd is more
668 // expensive than a straight movsd. On the other hand, it's important to
669 // shrink long double fp constant since fldt is very slow.
670 return !X86ScalarSSEf64 || VT == MVT::f80;
673 const X86Subtarget* getSubtarget() const {
677 /// isScalarFPTypeInSSEReg - Return true if the specified scalar FP type is
678 /// computed in an SSE register, not on the X87 floating point stack.
679 bool isScalarFPTypeInSSEReg(EVT VT) const {
680 return (VT == MVT::f64 && X86ScalarSSEf64) || // f64 is when SSE2
681 (VT == MVT::f32 && X86ScalarSSEf32); // f32 is when SSE1
684 /// isTargetFTOL - Return true if the target uses the MSVC _ftol2 routine
686 bool isTargetFTOL() const {
687 return Subtarget->isTargetWindows() && !Subtarget->is64Bit();
690 /// isIntegerTypeFTOL - Return true if the MSVC _ftol2 routine should be
691 /// used for fptoui to the given type.
692 bool isIntegerTypeFTOL(EVT VT) const {
693 return isTargetFTOL() && VT == MVT::i64;
696 /// createFastISel - This method returns a target specific FastISel object,
697 /// or null if the target does not support "fast" ISel.
698 virtual FastISel *createFastISel(FunctionLoweringInfo &funcInfo,
699 const TargetLibraryInfo *libInfo) const;
701 /// getStackCookieLocation - Return true if the target stores stack
702 /// protector cookies at a fixed offset in some non-standard address
703 /// space, and populates the address space and offset as
705 virtual bool getStackCookieLocation(unsigned &AddressSpace, unsigned &Offset) const;
707 SDValue BuildFILD(SDValue Op, EVT SrcVT, SDValue Chain, SDValue StackSlot,
708 SelectionDAG &DAG) const;
711 std::pair<const TargetRegisterClass*, uint8_t>
712 findRepresentativeClass(EVT VT) const;
715 /// Subtarget - Keep a pointer to the X86Subtarget around so that we can
716 /// make the right decision when generating code for different targets.
717 const X86Subtarget *Subtarget;
718 const X86RegisterInfo *RegInfo;
719 const DataLayout *TD;
721 /// X86ScalarSSEf32, X86ScalarSSEf64 - Select between SSE or x87
722 /// floating point ops.
723 /// When SSE is available, use it for f32 operations.
724 /// When SSE2 is available, use it for f64 operations.
725 bool X86ScalarSSEf32;
726 bool X86ScalarSSEf64;
728 /// LegalFPImmediates - A list of legal fp immediates.
729 std::vector<APFloat> LegalFPImmediates;
731 /// addLegalFPImmediate - Indicate that this x86 target can instruction
732 /// select the specified FP immediate natively.
733 void addLegalFPImmediate(const APFloat& Imm) {
734 LegalFPImmediates.push_back(Imm);
737 SDValue LowerCallResult(SDValue Chain, SDValue InFlag,
738 CallingConv::ID CallConv, bool isVarArg,
739 const SmallVectorImpl<ISD::InputArg> &Ins,
740 DebugLoc dl, SelectionDAG &DAG,
741 SmallVectorImpl<SDValue> &InVals) const;
742 SDValue LowerMemArgument(SDValue Chain,
743 CallingConv::ID CallConv,
744 const SmallVectorImpl<ISD::InputArg> &ArgInfo,
745 DebugLoc dl, SelectionDAG &DAG,
746 const CCValAssign &VA, MachineFrameInfo *MFI,
748 SDValue LowerMemOpCallTo(SDValue Chain, SDValue StackPtr, SDValue Arg,
749 DebugLoc dl, SelectionDAG &DAG,
750 const CCValAssign &VA,
751 ISD::ArgFlagsTy Flags) const;
753 // Call lowering helpers.
755 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
756 /// for tail call optimization. Targets which want to do tail call
757 /// optimization should implement this function.
758 bool IsEligibleForTailCallOptimization(SDValue Callee,
759 CallingConv::ID CalleeCC,
761 bool isCalleeStructRet,
762 bool isCallerStructRet,
764 const SmallVectorImpl<ISD::OutputArg> &Outs,
765 const SmallVectorImpl<SDValue> &OutVals,
766 const SmallVectorImpl<ISD::InputArg> &Ins,
767 SelectionDAG& DAG) const;
768 bool IsCalleePop(bool isVarArg, CallingConv::ID CallConv) const;
769 SDValue EmitTailCallLoadRetAddr(SelectionDAG &DAG, SDValue &OutRetAddr,
770 SDValue Chain, bool IsTailCall, bool Is64Bit,
771 int FPDiff, DebugLoc dl) const;
773 unsigned GetAlignedArgumentStackSize(unsigned StackSize,
774 SelectionDAG &DAG) const;
776 std::pair<SDValue,SDValue> FP_TO_INTHelper(SDValue Op, SelectionDAG &DAG,
778 bool isReplace) const;
780 SDValue LowerAsSplatVectorLoad(SDValue SrcOp, EVT VT, DebugLoc dl,
781 SelectionDAG &DAG) const;
782 SDValue LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) const;
783 SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) const;
784 SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const;
785 SDValue LowerEXTRACT_VECTOR_ELT_SSE4(SDValue Op, SelectionDAG &DAG) const;
786 SDValue LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const;
787 SDValue LowerINSERT_VECTOR_ELT_SSE4(SDValue Op, SelectionDAG &DAG) const;
788 SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) const;
789 SDValue LowerBlockAddress(SDValue Op, SelectionDAG &DAG) const;
790 SDValue LowerGlobalAddress(const GlobalValue *GV, DebugLoc dl,
791 int64_t Offset, SelectionDAG &DAG) const;
792 SDValue LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) const;
793 SDValue LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const;
794 SDValue LowerExternalSymbol(SDValue Op, SelectionDAG &DAG) const;
795 SDValue LowerShiftParts(SDValue Op, SelectionDAG &DAG) const;
796 SDValue LowerBITCAST(SDValue op, SelectionDAG &DAG) const;
797 SDValue LowerSINT_TO_FP(SDValue Op, SelectionDAG &DAG) const;
798 SDValue LowerUINT_TO_FP(SDValue Op, SelectionDAG &DAG) const;
799 SDValue LowerUINT_TO_FP_i64(SDValue Op, SelectionDAG &DAG) const;
800 SDValue LowerUINT_TO_FP_i32(SDValue Op, SelectionDAG &DAG) const;
801 SDValue lowerUINT_TO_FP_vec(SDValue Op, SelectionDAG &DAG) const;
802 SDValue lowerTRUNCATE(SDValue Op, SelectionDAG &DAG) const;
803 SDValue lowerZERO_EXTEND(SDValue Op, SelectionDAG &DAG) const;
804 SDValue LowerFP_TO_SINT(SDValue Op, SelectionDAG &DAG) const;
805 SDValue LowerFP_TO_UINT(SDValue Op, SelectionDAG &DAG) const;
806 SDValue lowerFP_EXTEND(SDValue Op, SelectionDAG &DAG) const;
807 SDValue LowerFABS(SDValue Op, SelectionDAG &DAG) const;
808 SDValue LowerFNEG(SDValue Op, SelectionDAG &DAG) const;
809 SDValue LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const;
810 SDValue LowerToBT(SDValue And, ISD::CondCode CC,
811 DebugLoc dl, SelectionDAG &DAG) const;
812 SDValue LowerSETCC(SDValue Op, SelectionDAG &DAG) const;
813 SDValue LowerVSETCC(SDValue Op, SelectionDAG &DAG) const;
814 SDValue LowerSELECT(SDValue Op, SelectionDAG &DAG) const;
815 SDValue LowerBRCOND(SDValue Op, SelectionDAG &DAG) const;
816 SDValue LowerMEMSET(SDValue Op, SelectionDAG &DAG) const;
817 SDValue LowerJumpTable(SDValue Op, SelectionDAG &DAG) const;
818 SDValue LowerDYNAMIC_STACKALLOC(SDValue Op, SelectionDAG &DAG) const;
819 SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG) const;
820 SDValue LowerVAARG(SDValue Op, SelectionDAG &DAG) const;
821 SDValue LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const;
822 SDValue LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const;
823 SDValue LowerFRAME_TO_ARGS_OFFSET(SDValue Op, SelectionDAG &DAG) const;
824 SDValue LowerEH_RETURN(SDValue Op, SelectionDAG &DAG) const;
825 SDValue lowerEH_SJLJ_SETJMP(SDValue Op, SelectionDAG &DAG) const;
826 SDValue lowerEH_SJLJ_LONGJMP(SDValue Op, SelectionDAG &DAG) const;
827 SDValue LowerINIT_TRAMPOLINE(SDValue Op, SelectionDAG &DAG) const;
828 SDValue LowerFLT_ROUNDS_(SDValue Op, SelectionDAG &DAG) const;
829 SDValue LowerShift(SDValue Op, SelectionDAG &DAG) const;
831 SDValue LowerSIGN_EXTEND_INREG(SDValue Op, SelectionDAG &DAG) const;
833 // Utility functions to help LowerVECTOR_SHUFFLE & LowerBUILD_VECTOR
834 SDValue LowerVectorBroadcast(SDValue Op, SelectionDAG &DAG) const;
835 SDValue NormalizeVectorShuffle(SDValue Op, SelectionDAG &DAG) const;
836 SDValue buildFromShuffleMostly(SDValue Op, SelectionDAG &DAG) const;
838 SDValue LowerVectorAllZeroTest(SDValue Op, SelectionDAG &DAG) const;
840 SDValue lowerVectorIntExtend(SDValue Op, SelectionDAG &DAG) const;
843 LowerFormalArguments(SDValue Chain,
844 CallingConv::ID CallConv, bool isVarArg,
845 const SmallVectorImpl<ISD::InputArg> &Ins,
846 DebugLoc dl, SelectionDAG &DAG,
847 SmallVectorImpl<SDValue> &InVals) const;
849 LowerCall(CallLoweringInfo &CLI,
850 SmallVectorImpl<SDValue> &InVals) const;
853 LowerReturn(SDValue Chain,
854 CallingConv::ID CallConv, bool isVarArg,
855 const SmallVectorImpl<ISD::OutputArg> &Outs,
856 const SmallVectorImpl<SDValue> &OutVals,
857 DebugLoc dl, SelectionDAG &DAG) const;
859 virtual bool isUsedByReturnOnly(SDNode *N, SDValue &Chain) const;
861 virtual bool mayBeEmittedAsTailCall(CallInst *CI) const;
864 getTypeForExtArgOrReturn(LLVMContext &Context, EVT VT,
865 ISD::NodeType ExtendKind) const;
868 CanLowerReturn(CallingConv::ID CallConv, MachineFunction &MF,
870 const SmallVectorImpl<ISD::OutputArg> &Outs,
871 LLVMContext &Context) const;
873 /// Utility function to emit atomic-load-arith operations (and, or, xor,
874 /// nand, max, min, umax, umin). It takes the corresponding instruction to
875 /// expand, the associated machine basic block, and the associated X86
876 /// opcodes for reg/reg.
877 MachineBasicBlock *EmitAtomicLoadArith(MachineInstr *MI,
878 MachineBasicBlock *MBB) const;
880 /// Utility function to emit atomic-load-arith operations (and, or, xor,
881 /// nand, add, sub, swap) for 64-bit operands on 32-bit target.
882 MachineBasicBlock *EmitAtomicLoadArith6432(MachineInstr *MI,
883 MachineBasicBlock *MBB) const;
885 // Utility function to emit the low-level va_arg code for X86-64.
886 MachineBasicBlock *EmitVAARG64WithCustomInserter(
888 MachineBasicBlock *MBB) const;
890 /// Utility function to emit the xmm reg save portion of va_start.
891 MachineBasicBlock *EmitVAStartSaveXMMRegsWithCustomInserter(
892 MachineInstr *BInstr,
893 MachineBasicBlock *BB) const;
895 MachineBasicBlock *EmitLoweredSelect(MachineInstr *I,
896 MachineBasicBlock *BB) const;
898 MachineBasicBlock *EmitLoweredWinAlloca(MachineInstr *MI,
899 MachineBasicBlock *BB) const;
901 MachineBasicBlock *EmitLoweredSegAlloca(MachineInstr *MI,
902 MachineBasicBlock *BB,
905 MachineBasicBlock *EmitLoweredTLSCall(MachineInstr *MI,
906 MachineBasicBlock *BB) const;
908 MachineBasicBlock *emitLoweredTLSAddr(MachineInstr *MI,
909 MachineBasicBlock *BB) const;
911 MachineBasicBlock *emitEHSjLjSetJmp(MachineInstr *MI,
912 MachineBasicBlock *MBB) const;
914 MachineBasicBlock *emitEHSjLjLongJmp(MachineInstr *MI,
915 MachineBasicBlock *MBB) const;
917 /// Emit nodes that will be selected as "test Op0,Op0", or something
918 /// equivalent, for use with the given x86 condition code.
919 SDValue EmitTest(SDValue Op0, unsigned X86CC, SelectionDAG &DAG) const;
921 /// Emit nodes that will be selected as "cmp Op0,Op1", or something
922 /// equivalent, for use with the given x86 condition code.
923 SDValue EmitCmp(SDValue Op0, SDValue Op1, unsigned X86CC,
924 SelectionDAG &DAG) const;
926 /// Convert a comparison if required by the subtarget.
927 SDValue ConvertCmpIfNecessary(SDValue Cmp, SelectionDAG &DAG) const;
931 FastISel *createFastISel(FunctionLoweringInfo &funcInfo,
932 const TargetLibraryInfo *libInfo);
935 class X86ScalarTargetTransformImpl : public ScalarTargetTransformImpl {
937 explicit X86ScalarTargetTransformImpl(const TargetLowering *TL) :
938 ScalarTargetTransformImpl(TL) {};
940 virtual PopcntHwSupport getPopcntHwSupport(unsigned TyWidth) const;
943 class X86VectorTargetTransformInfo : public VectorTargetTransformImpl {
945 explicit X86VectorTargetTransformInfo(const TargetLowering *TL) :
946 VectorTargetTransformImpl(TL) {}
948 virtual unsigned getArithmeticInstrCost(unsigned Opcode, Type *Ty) const;
950 virtual unsigned getVectorInstrCost(unsigned Opcode, Type *Val,
951 unsigned Index) const;
953 unsigned getCmpSelInstrCost(unsigned Opcode, Type *ValTy,
956 virtual unsigned getCastInstrCost(unsigned Opcode, Type *Dst,
961 #endif // X86ISELLOWERING_H