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.
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.
327 // PMULUDQ - Vector multiply packed unsigned doubleword integers
338 // VASTART_SAVE_XMM_REGS - Save xmm argument registers to the stack,
339 // according to %al. An operator is needed so that this can be expanded
340 // with control flow.
341 VASTART_SAVE_XMM_REGS,
343 // WIN_ALLOCA - Windows's _chkstk call to do stack probing.
346 // SEG_ALLOCA - For allocating variable amounts of stack space when using
347 // segmented stacks. Check if the current stacklet has enough space, and
348 // falls back to heap allocation if not.
351 // WIN_FTOL - Windows's _ftol2 runtime routine to do fptoui.
360 // FNSTSW16r - Store FP status word into i16 register.
363 // SAHF - Store contents of %ah into %eflags.
366 // RDRAND - Get a random integer and indicate whether it is valid in CF.
369 // RDSEED - Get a NIST SP800-90B & C compliant random integer and
370 // indicate whether it is valid in CF.
377 // XTEST - Test if in transactional execution.
380 // ATOMADD64_DAG, ATOMSUB64_DAG, ATOMOR64_DAG, ATOMAND64_DAG,
381 // ATOMXOR64_DAG, ATOMNAND64_DAG, ATOMSWAP64_DAG -
382 // Atomic 64-bit binary operations.
383 ATOMADD64_DAG = ISD::FIRST_TARGET_MEMORY_OPCODE,
395 // LCMPXCHG_DAG, LCMPXCHG8_DAG, LCMPXCHG16_DAG - Compare and swap.
400 // VZEXT_LOAD - Load, scalar_to_vector, and zero extend.
403 // FNSTCW16m - Store FP control world into i16 memory.
406 /// FP_TO_INT*_IN_MEM - This instruction implements FP_TO_SINT with the
407 /// integer destination in memory and a FP reg source. This corresponds
408 /// to the X86::FIST*m instructions and the rounding mode change stuff. It
409 /// has two inputs (token chain and address) and two outputs (int value
410 /// and token chain).
415 /// FILD, FILD_FLAG - This instruction implements SINT_TO_FP with the
416 /// integer source in memory and FP reg result. This corresponds to the
417 /// X86::FILD*m instructions. It has three inputs (token chain, address,
418 /// and source type) and two outputs (FP value and token chain). FILD_FLAG
419 /// also produces a flag).
423 /// FLD - This instruction implements an extending load to FP stack slots.
424 /// This corresponds to the X86::FLD32m / X86::FLD64m. It takes a chain
425 /// operand, ptr to load from, and a ValueType node indicating the type
429 /// FST - This instruction implements a truncating store to FP stack
430 /// slots. This corresponds to the X86::FST32m / X86::FST64m. It takes a
431 /// chain operand, value to store, address, and a ValueType to store it
435 /// VAARG_64 - This instruction grabs the address of the next argument
436 /// from a va_list. (reads and modifies the va_list in memory)
439 // WARNING: Do not add anything in the end unless you want the node to
440 // have memop! In fact, starting from ATOMADD64_DAG all opcodes will be
441 // thought as target memory ops!
445 /// Define some predicates that are used for node matching.
447 /// isVEXTRACT128Index - Return true if the specified
448 /// EXTRACT_SUBVECTOR operand specifies a vector extract that is
449 /// suitable for input to VEXTRACTF128, VEXTRACTI128 instructions.
450 bool isVEXTRACT128Index(SDNode *N);
452 /// isVINSERT128Index - Return true if the specified
453 /// INSERT_SUBVECTOR operand specifies a subvector insert that is
454 /// suitable for input to VINSERTF128, VINSERTI128 instructions.
455 bool isVINSERT128Index(SDNode *N);
457 /// isVEXTRACT256Index - Return true if the specified
458 /// EXTRACT_SUBVECTOR operand specifies a vector extract that is
459 /// suitable for input to VEXTRACTF64X4, VEXTRACTI64X4 instructions.
460 bool isVEXTRACT256Index(SDNode *N);
462 /// isVINSERT256Index - Return true if the specified
463 /// INSERT_SUBVECTOR operand specifies a subvector insert that is
464 /// suitable for input to VINSERTF64X4, VINSERTI64X4 instructions.
465 bool isVINSERT256Index(SDNode *N);
467 /// getExtractVEXTRACT128Immediate - Return the appropriate
468 /// immediate to extract the specified EXTRACT_SUBVECTOR index
469 /// with VEXTRACTF128, VEXTRACTI128 instructions.
470 unsigned getExtractVEXTRACT128Immediate(SDNode *N);
472 /// getInsertVINSERT128Immediate - Return the appropriate
473 /// immediate to insert at the specified INSERT_SUBVECTOR index
474 /// with VINSERTF128, VINSERT128 instructions.
475 unsigned getInsertVINSERT128Immediate(SDNode *N);
477 /// getExtractVEXTRACT256Immediate - Return the appropriate
478 /// immediate to extract the specified EXTRACT_SUBVECTOR index
479 /// with VEXTRACTF64X4, VEXTRACTI64x4 instructions.
480 unsigned getExtractVEXTRACT256Immediate(SDNode *N);
482 /// getInsertVINSERT256Immediate - Return the appropriate
483 /// immediate to insert at the specified INSERT_SUBVECTOR index
484 /// with VINSERTF64x4, VINSERTI64x4 instructions.
485 unsigned getInsertVINSERT256Immediate(SDNode *N);
487 /// isZeroNode - Returns true if Elt is a constant zero or a floating point
489 bool isZeroNode(SDValue Elt);
491 /// isOffsetSuitableForCodeModel - Returns true of the given offset can be
492 /// fit into displacement field of the instruction.
493 bool isOffsetSuitableForCodeModel(int64_t Offset, CodeModel::Model M,
494 bool hasSymbolicDisplacement = true);
497 /// isCalleePop - Determines whether the callee is required to pop its
498 /// own arguments. Callee pop is necessary to support tail calls.
499 bool isCalleePop(CallingConv::ID CallingConv,
500 bool is64Bit, bool IsVarArg, bool TailCallOpt);
503 //===--------------------------------------------------------------------===//
504 // X86TargetLowering - X86 Implementation of the TargetLowering interface
505 class X86TargetLowering : public TargetLowering {
507 explicit X86TargetLowering(X86TargetMachine &TM);
509 virtual unsigned getJumpTableEncoding() const;
511 virtual MVT getScalarShiftAmountTy(EVT LHSTy) const { return MVT::i8; }
513 virtual const MCExpr *
514 LowerCustomJumpTableEntry(const MachineJumpTableInfo *MJTI,
515 const MachineBasicBlock *MBB, unsigned uid,
516 MCContext &Ctx) const;
518 /// getPICJumpTableRelocaBase - Returns relocation base for the given PIC
520 virtual SDValue getPICJumpTableRelocBase(SDValue Table,
521 SelectionDAG &DAG) const;
522 virtual const MCExpr *
523 getPICJumpTableRelocBaseExpr(const MachineFunction *MF,
524 unsigned JTI, MCContext &Ctx) const;
526 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
527 /// function arguments in the caller parameter area. For X86, aggregates
528 /// that contains are placed at 16-byte boundaries while the rest are at
529 /// 4-byte boundaries.
530 virtual unsigned getByValTypeAlignment(Type *Ty) const;
532 /// getOptimalMemOpType - Returns the target specific optimal type for load
533 /// and store operations as a result of memset, memcpy, and memmove
534 /// lowering. If DstAlign is zero that means it's safe to destination
535 /// alignment can satisfy any constraint. Similarly if SrcAlign is zero it
536 /// means there isn't a need to check it against alignment requirement,
537 /// probably because the source does not need to be loaded. If 'IsMemset' is
538 /// true, that means it's expanding a memset. If 'ZeroMemset' is true, that
539 /// means it's a memset of zero. 'MemcpyStrSrc' indicates whether the memcpy
540 /// source is constant so it does not need to be loaded.
541 /// It returns EVT::Other if the type should be determined using generic
542 /// target-independent logic.
544 getOptimalMemOpType(uint64_t Size, unsigned DstAlign, unsigned SrcAlign,
545 bool IsMemset, bool ZeroMemset, bool MemcpyStrSrc,
546 MachineFunction &MF) const;
548 /// isSafeMemOpType - Returns true if it's safe to use load / store of the
549 /// specified type to expand memcpy / memset inline. This is mostly true
550 /// for all types except for some special cases. For example, on X86
551 /// targets without SSE2 f64 load / store are done with fldl / fstpl which
552 /// also does type conversion. Note the specified type doesn't have to be
553 /// legal as the hook is used before type legalization.
554 virtual bool isSafeMemOpType(MVT VT) const;
556 /// allowsUnalignedMemoryAccesses - Returns true if the target allows
557 /// unaligned memory accesses. of the specified type. Returns whether it
558 /// is "fast" by reference in the second argument.
559 virtual bool allowsUnalignedMemoryAccesses(EVT VT, bool *Fast) const;
561 /// LowerOperation - Provide custom lowering hooks for some operations.
563 virtual SDValue LowerOperation(SDValue Op, SelectionDAG &DAG) const;
565 /// ReplaceNodeResults - Replace the results of node with an illegal result
566 /// type with new values built out of custom code.
568 virtual void ReplaceNodeResults(SDNode *N, SmallVectorImpl<SDValue>&Results,
569 SelectionDAG &DAG) const;
572 virtual SDValue PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const;
574 /// isTypeDesirableForOp - Return true if the target has native support for
575 /// the specified value type and it is 'desirable' to use the type for the
576 /// given node type. e.g. On x86 i16 is legal, but undesirable since i16
577 /// instruction encodings are longer and some i16 instructions are slow.
578 virtual bool isTypeDesirableForOp(unsigned Opc, EVT VT) const;
580 /// isTypeDesirable - Return true if the target has native support for the
581 /// specified value type and it is 'desirable' to use the type. e.g. On x86
582 /// i16 is legal, but undesirable since i16 instruction encodings are longer
583 /// and some i16 instructions are slow.
584 virtual bool IsDesirableToPromoteOp(SDValue Op, EVT &PVT) const;
586 virtual MachineBasicBlock *
587 EmitInstrWithCustomInserter(MachineInstr *MI,
588 MachineBasicBlock *MBB) const;
591 /// getTargetNodeName - This method returns the name of a target specific
593 virtual const char *getTargetNodeName(unsigned Opcode) const;
595 /// getSetCCResultType - Return the value type to use for ISD::SETCC.
596 virtual EVT getSetCCResultType(LLVMContext &Context, EVT VT) const;
598 /// computeMaskedBitsForTargetNode - Determine which of the bits specified
599 /// in Mask are known to be either zero or one and return them in the
600 /// KnownZero/KnownOne bitsets.
601 virtual void computeMaskedBitsForTargetNode(const SDValue Op,
604 const SelectionDAG &DAG,
605 unsigned Depth = 0) const;
607 // ComputeNumSignBitsForTargetNode - Determine the number of bits in the
608 // operation that are sign bits.
609 virtual unsigned ComputeNumSignBitsForTargetNode(SDValue Op,
610 unsigned Depth) const;
613 isGAPlusOffset(SDNode *N, const GlobalValue* &GA, int64_t &Offset) const;
615 SDValue getReturnAddressFrameIndex(SelectionDAG &DAG) const;
617 virtual bool ExpandInlineAsm(CallInst *CI) const;
619 ConstraintType getConstraintType(const std::string &Constraint) const;
621 /// Examine constraint string and operand type and determine a weight value.
622 /// The operand object must already have been set up with the operand type.
623 virtual ConstraintWeight getSingleConstraintMatchWeight(
624 AsmOperandInfo &info, const char *constraint) const;
626 virtual const char *LowerXConstraint(EVT ConstraintVT) const;
628 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
629 /// vector. If it is invalid, don't add anything to Ops. If hasMemory is
630 /// true it means one of the asm constraint of the inline asm instruction
631 /// being processed is 'm'.
632 virtual void LowerAsmOperandForConstraint(SDValue Op,
633 std::string &Constraint,
634 std::vector<SDValue> &Ops,
635 SelectionDAG &DAG) const;
637 /// getRegForInlineAsmConstraint - Given a physical register constraint
638 /// (e.g. {edx}), return the register number and the register class for the
639 /// register. This should only be used for C_Register constraints. On
640 /// error, this returns a register number of 0.
641 std::pair<unsigned, const TargetRegisterClass*>
642 getRegForInlineAsmConstraint(const std::string &Constraint,
645 /// isLegalAddressingMode - Return true if the addressing mode represented
646 /// by AM is legal for this target, for a load/store of the specified type.
647 virtual bool isLegalAddressingMode(const AddrMode &AM, Type *Ty)const;
649 /// isLegalICmpImmediate - Return true if the specified immediate is legal
650 /// icmp immediate, that is the target has icmp instructions which can
651 /// compare a register against the immediate without having to materialize
652 /// the immediate into a register.
653 virtual bool isLegalICmpImmediate(int64_t Imm) const;
655 /// isLegalAddImmediate - Return true if the specified immediate is legal
656 /// add immediate, that is the target has add instructions which can
657 /// add a register and the immediate without having to materialize
658 /// the immediate into a register.
659 virtual bool isLegalAddImmediate(int64_t Imm) const;
661 /// isTruncateFree - Return true if it's free to truncate a value of
662 /// type Ty1 to type Ty2. e.g. On x86 it's free to truncate a i32 value in
663 /// register EAX to i16 by referencing its sub-register AX.
664 virtual bool isTruncateFree(Type *Ty1, Type *Ty2) const;
665 virtual bool isTruncateFree(EVT VT1, EVT VT2) const;
667 virtual bool allowTruncateForTailCall(Type *Ty1, Type *Ty2) const;
669 /// isZExtFree - Return true if any actual instruction that defines a
670 /// value of type Ty1 implicit zero-extends the value to Ty2 in the result
671 /// register. This does not necessarily include registers defined in
672 /// unknown ways, such as incoming arguments, or copies from unknown
673 /// virtual registers. Also, if isTruncateFree(Ty2, Ty1) is true, this
674 /// does not necessarily apply to truncate instructions. e.g. on x86-64,
675 /// all instructions that define 32-bit values implicit zero-extend the
676 /// result out to 64 bits.
677 virtual bool isZExtFree(Type *Ty1, Type *Ty2) const;
678 virtual bool isZExtFree(EVT VT1, EVT VT2) const;
679 virtual bool isZExtFree(SDValue Val, EVT VT2) const;
681 /// isFMAFasterThanFMulAndFAdd - Return true if an FMA operation is faster
682 /// than a pair of fmul and fadd instructions. fmuladd intrinsics will be
683 /// expanded to FMAs when this method returns true, otherwise fmuladd is
684 /// expanded to fmul + fadd.
685 virtual bool isFMAFasterThanFMulAndFAdd(EVT VT) const;
687 /// isNarrowingProfitable - Return true if it's profitable to narrow
688 /// operations of type VT1 to VT2. e.g. on x86, it's profitable to narrow
689 /// from i32 to i8 but not from i32 to i16.
690 virtual bool isNarrowingProfitable(EVT VT1, EVT VT2) const;
692 /// isFPImmLegal - Returns true if the target can instruction select the
693 /// specified FP immediate natively. If false, the legalizer will
694 /// materialize the FP immediate as a load from a constant pool.
695 virtual bool isFPImmLegal(const APFloat &Imm, EVT VT) const;
697 /// isShuffleMaskLegal - Targets can use this to indicate that they only
698 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
699 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask
700 /// values are assumed to be legal.
701 virtual bool isShuffleMaskLegal(const SmallVectorImpl<int> &Mask,
704 /// isVectorClearMaskLegal - Similar to isShuffleMaskLegal. This is
705 /// used by Targets can use this to indicate if there is a suitable
706 /// VECTOR_SHUFFLE that can be used to replace a VAND with a constant
708 virtual bool isVectorClearMaskLegal(const SmallVectorImpl<int> &Mask,
711 /// ShouldShrinkFPConstant - If true, then instruction selection should
712 /// seek to shrink the FP constant of the specified type to a smaller type
713 /// in order to save space and / or reduce runtime.
714 virtual bool ShouldShrinkFPConstant(EVT VT) const {
715 // Don't shrink FP constpool if SSE2 is available since cvtss2sd is more
716 // expensive than a straight movsd. On the other hand, it's important to
717 // shrink long double fp constant since fldt is very slow.
718 return !X86ScalarSSEf64 || VT == MVT::f80;
721 const X86Subtarget* getSubtarget() const {
725 /// isScalarFPTypeInSSEReg - Return true if the specified scalar FP type is
726 /// computed in an SSE register, not on the X87 floating point stack.
727 bool isScalarFPTypeInSSEReg(EVT VT) const {
728 return (VT == MVT::f64 && X86ScalarSSEf64) || // f64 is when SSE2
729 (VT == MVT::f32 && X86ScalarSSEf32); // f32 is when SSE1
732 /// isTargetFTOL - Return true if the target uses the MSVC _ftol2 routine
734 bool isTargetFTOL() const {
735 return Subtarget->isTargetWindows() && !Subtarget->is64Bit();
738 /// isIntegerTypeFTOL - Return true if the MSVC _ftol2 routine should be
739 /// used for fptoui to the given type.
740 bool isIntegerTypeFTOL(EVT VT) const {
741 return isTargetFTOL() && VT == MVT::i64;
744 /// createFastISel - This method returns a target specific FastISel object,
745 /// or null if the target does not support "fast" ISel.
746 virtual FastISel *createFastISel(FunctionLoweringInfo &funcInfo,
747 const TargetLibraryInfo *libInfo) const;
749 /// getStackCookieLocation - Return true if the target stores stack
750 /// protector cookies at a fixed offset in some non-standard address
751 /// space, and populates the address space and offset as
753 virtual bool getStackCookieLocation(unsigned &AddressSpace, unsigned &Offset) const;
755 SDValue BuildFILD(SDValue Op, EVT SrcVT, SDValue Chain, SDValue StackSlot,
756 SelectionDAG &DAG) const;
758 /// \brief Reset the operation actions based on target options.
759 virtual void resetOperationActions();
762 std::pair<const TargetRegisterClass*, uint8_t>
763 findRepresentativeClass(MVT VT) const;
766 /// Subtarget - Keep a pointer to the X86Subtarget around so that we can
767 /// make the right decision when generating code for different targets.
768 const X86Subtarget *Subtarget;
769 const DataLayout *TD;
771 /// Used to store the TargetOptions so that we don't waste time resetting
772 /// the operation actions unless we have to.
775 /// X86ScalarSSEf32, X86ScalarSSEf64 - Select between SSE or x87
776 /// floating point ops.
777 /// When SSE is available, use it for f32 operations.
778 /// When SSE2 is available, use it for f64 operations.
779 bool X86ScalarSSEf32;
780 bool X86ScalarSSEf64;
782 /// LegalFPImmediates - A list of legal fp immediates.
783 std::vector<APFloat> LegalFPImmediates;
785 /// addLegalFPImmediate - Indicate that this x86 target can instruction
786 /// select the specified FP immediate natively.
787 void addLegalFPImmediate(const APFloat& Imm) {
788 LegalFPImmediates.push_back(Imm);
791 SDValue LowerCallResult(SDValue Chain, SDValue InFlag,
792 CallingConv::ID CallConv, bool isVarArg,
793 const SmallVectorImpl<ISD::InputArg> &Ins,
794 SDLoc dl, SelectionDAG &DAG,
795 SmallVectorImpl<SDValue> &InVals) const;
796 SDValue LowerMemArgument(SDValue Chain,
797 CallingConv::ID CallConv,
798 const SmallVectorImpl<ISD::InputArg> &ArgInfo,
799 SDLoc dl, SelectionDAG &DAG,
800 const CCValAssign &VA, MachineFrameInfo *MFI,
802 SDValue LowerMemOpCallTo(SDValue Chain, SDValue StackPtr, SDValue Arg,
803 SDLoc dl, SelectionDAG &DAG,
804 const CCValAssign &VA,
805 ISD::ArgFlagsTy Flags) const;
807 // Call lowering helpers.
809 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
810 /// for tail call optimization. Targets which want to do tail call
811 /// optimization should implement this function.
812 bool IsEligibleForTailCallOptimization(SDValue Callee,
813 CallingConv::ID CalleeCC,
815 bool isCalleeStructRet,
816 bool isCallerStructRet,
818 const SmallVectorImpl<ISD::OutputArg> &Outs,
819 const SmallVectorImpl<SDValue> &OutVals,
820 const SmallVectorImpl<ISD::InputArg> &Ins,
821 SelectionDAG& DAG) const;
822 bool IsCalleePop(bool isVarArg, CallingConv::ID CallConv) const;
823 SDValue EmitTailCallLoadRetAddr(SelectionDAG &DAG, SDValue &OutRetAddr,
824 SDValue Chain, bool IsTailCall, bool Is64Bit,
825 int FPDiff, SDLoc dl) const;
827 unsigned GetAlignedArgumentStackSize(unsigned StackSize,
828 SelectionDAG &DAG) const;
830 std::pair<SDValue,SDValue> FP_TO_INTHelper(SDValue Op, SelectionDAG &DAG,
832 bool isReplace) const;
834 SDValue LowerAsSplatVectorLoad(SDValue SrcOp, EVT VT, SDLoc dl,
835 SelectionDAG &DAG) const;
836 SDValue LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) const;
837 SDValue LowerBUILD_VECTORvXi1(SDValue Op, SelectionDAG &DAG) const;
838 SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) const;
839 SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const;
840 SDValue LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const;
841 SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) const;
842 SDValue LowerBlockAddress(SDValue Op, SelectionDAG &DAG) const;
843 SDValue LowerGlobalAddress(const GlobalValue *GV, SDLoc dl,
844 int64_t Offset, SelectionDAG &DAG) const;
845 SDValue LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) const;
846 SDValue LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const;
847 SDValue LowerExternalSymbol(SDValue Op, SelectionDAG &DAG) const;
848 SDValue LowerShiftParts(SDValue Op, SelectionDAG &DAG) const;
849 SDValue LowerBITCAST(SDValue op, SelectionDAG &DAG) const;
850 SDValue LowerSINT_TO_FP(SDValue Op, SelectionDAG &DAG) const;
851 SDValue LowerUINT_TO_FP(SDValue Op, SelectionDAG &DAG) const;
852 SDValue LowerUINT_TO_FP_i64(SDValue Op, SelectionDAG &DAG) const;
853 SDValue LowerUINT_TO_FP_i32(SDValue Op, SelectionDAG &DAG) const;
854 SDValue lowerUINT_TO_FP_vec(SDValue Op, SelectionDAG &DAG) const;
855 SDValue LowerTRUNCATE(SDValue Op, SelectionDAG &DAG) const;
856 SDValue LowerZERO_EXTEND(SDValue Op, SelectionDAG &DAG) const;
857 SDValue LowerZERO_EXTEND_AVX512(SDValue Op, SelectionDAG &DAG) const;
858 SDValue LowerSIGN_EXTEND(SDValue Op, SelectionDAG &DAG) const;
859 SDValue LowerSIGN_EXTEND_AVX512(SDValue Op, SelectionDAG &DAG) const;
860 SDValue LowerANY_EXTEND(SDValue Op, SelectionDAG &DAG) const;
861 SDValue LowerFP_TO_SINT(SDValue Op, SelectionDAG &DAG) const;
862 SDValue LowerFP_TO_UINT(SDValue Op, SelectionDAG &DAG) const;
863 SDValue LowerFABS(SDValue Op, SelectionDAG &DAG) const;
864 SDValue LowerFNEG(SDValue Op, SelectionDAG &DAG) const;
865 SDValue LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const;
866 SDValue LowerToBT(SDValue And, ISD::CondCode CC,
867 SDLoc dl, SelectionDAG &DAG) const;
868 SDValue LowerSETCC(SDValue Op, SelectionDAG &DAG) const;
869 SDValue LowerSELECT(SDValue Op, SelectionDAG &DAG) const;
870 SDValue LowerBRCOND(SDValue Op, SelectionDAG &DAG) const;
871 SDValue LowerMEMSET(SDValue Op, SelectionDAG &DAG) const;
872 SDValue LowerJumpTable(SDValue Op, SelectionDAG &DAG) const;
873 SDValue LowerDYNAMIC_STACKALLOC(SDValue Op, SelectionDAG &DAG) const;
874 SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG) const;
875 SDValue LowerVAARG(SDValue Op, SelectionDAG &DAG) const;
876 SDValue LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const;
877 SDValue LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const;
878 SDValue LowerFRAME_TO_ARGS_OFFSET(SDValue Op, SelectionDAG &DAG) const;
879 SDValue LowerEH_RETURN(SDValue Op, SelectionDAG &DAG) const;
880 SDValue lowerEH_SJLJ_SETJMP(SDValue Op, SelectionDAG &DAG) const;
881 SDValue lowerEH_SJLJ_LONGJMP(SDValue Op, SelectionDAG &DAG) const;
882 SDValue LowerINIT_TRAMPOLINE(SDValue Op, SelectionDAG &DAG) const;
883 SDValue LowerFLT_ROUNDS_(SDValue Op, SelectionDAG &DAG) const;
884 SDValue LowerShift(SDValue Op, SelectionDAG &DAG) const;
885 SDValue LowerSDIV(SDValue Op, SelectionDAG &DAG) const;
886 SDValue LowerSIGN_EXTEND_INREG(SDValue Op, SelectionDAG &DAG) const;
887 SDValue LowerFSINCOS(SDValue Op, SelectionDAG &DAG) const;
889 // Utility functions to help LowerVECTOR_SHUFFLE & LowerBUILD_VECTOR
890 SDValue LowerVectorBroadcast(SDValue Op, SelectionDAG &DAG) const;
891 SDValue NormalizeVectorShuffle(SDValue Op, SelectionDAG &DAG) const;
892 SDValue buildFromShuffleMostly(SDValue Op, SelectionDAG &DAG) const;
894 SDValue LowerVectorAllZeroTest(SDValue Op, SelectionDAG &DAG) const;
896 SDValue LowerVectorIntExtend(SDValue Op, SelectionDAG &DAG) const;
899 LowerFormalArguments(SDValue Chain,
900 CallingConv::ID CallConv, bool isVarArg,
901 const SmallVectorImpl<ISD::InputArg> &Ins,
902 SDLoc dl, SelectionDAG &DAG,
903 SmallVectorImpl<SDValue> &InVals) const;
905 LowerCall(CallLoweringInfo &CLI,
906 SmallVectorImpl<SDValue> &InVals) const;
909 LowerReturn(SDValue Chain,
910 CallingConv::ID CallConv, bool isVarArg,
911 const SmallVectorImpl<ISD::OutputArg> &Outs,
912 const SmallVectorImpl<SDValue> &OutVals,
913 SDLoc dl, SelectionDAG &DAG) const;
915 virtual bool isUsedByReturnOnly(SDNode *N, SDValue &Chain) const;
917 virtual bool mayBeEmittedAsTailCall(CallInst *CI) const;
920 getTypeForExtArgOrReturn(MVT VT, ISD::NodeType ExtendKind) const;
923 CanLowerReturn(CallingConv::ID CallConv, MachineFunction &MF,
925 const SmallVectorImpl<ISD::OutputArg> &Outs,
926 LLVMContext &Context) const;
928 /// Utility function to emit atomic-load-arith operations (and, or, xor,
929 /// nand, max, min, umax, umin). It takes the corresponding instruction to
930 /// expand, the associated machine basic block, and the associated X86
931 /// opcodes for reg/reg.
932 MachineBasicBlock *EmitAtomicLoadArith(MachineInstr *MI,
933 MachineBasicBlock *MBB) const;
935 /// Utility function to emit atomic-load-arith operations (and, or, xor,
936 /// nand, add, sub, swap) for 64-bit operands on 32-bit target.
937 MachineBasicBlock *EmitAtomicLoadArith6432(MachineInstr *MI,
938 MachineBasicBlock *MBB) const;
940 // Utility function to emit the low-level va_arg code for X86-64.
941 MachineBasicBlock *EmitVAARG64WithCustomInserter(
943 MachineBasicBlock *MBB) const;
945 /// Utility function to emit the xmm reg save portion of va_start.
946 MachineBasicBlock *EmitVAStartSaveXMMRegsWithCustomInserter(
947 MachineInstr *BInstr,
948 MachineBasicBlock *BB) const;
950 MachineBasicBlock *EmitLoweredSelect(MachineInstr *I,
951 MachineBasicBlock *BB) const;
953 MachineBasicBlock *EmitLoweredWinAlloca(MachineInstr *MI,
954 MachineBasicBlock *BB) const;
956 MachineBasicBlock *EmitLoweredSegAlloca(MachineInstr *MI,
957 MachineBasicBlock *BB,
960 MachineBasicBlock *EmitLoweredTLSCall(MachineInstr *MI,
961 MachineBasicBlock *BB) const;
963 MachineBasicBlock *emitLoweredTLSAddr(MachineInstr *MI,
964 MachineBasicBlock *BB) const;
966 MachineBasicBlock *emitEHSjLjSetJmp(MachineInstr *MI,
967 MachineBasicBlock *MBB) const;
969 MachineBasicBlock *emitEHSjLjLongJmp(MachineInstr *MI,
970 MachineBasicBlock *MBB) const;
972 /// Emit nodes that will be selected as "test Op0,Op0", or something
973 /// equivalent, for use with the given x86 condition code.
974 SDValue EmitTest(SDValue Op0, unsigned X86CC, SelectionDAG &DAG) const;
976 /// Emit nodes that will be selected as "cmp Op0,Op1", or something
977 /// equivalent, for use with the given x86 condition code.
978 SDValue EmitCmp(SDValue Op0, SDValue Op1, unsigned X86CC,
979 SelectionDAG &DAG) const;
981 /// Convert a comparison if required by the subtarget.
982 SDValue ConvertCmpIfNecessary(SDValue Cmp, SelectionDAG &DAG) const;
986 FastISel *createFastISel(FunctionLoweringInfo &funcInfo,
987 const TargetLibraryInfo *libInfo);
991 #endif // X86ISELLOWERING_H