1 //===-- X86ISelLowering.h - X86 DAG Lowering Interface ----------*- C++ -*-===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file defines the interfaces that X86 uses to lower LLVM code into a
13 //===----------------------------------------------------------------------===//
15 #ifndef X86ISELLOWERING_H
16 #define X86ISELLOWERING_H
18 #include "X86Subtarget.h"
19 #include "X86RegisterInfo.h"
20 #include "X86MachineFunctionInfo.h"
21 #include "llvm/Target/TargetLowering.h"
22 #include "llvm/Target/TargetOptions.h"
23 #include "llvm/CodeGen/FastISel.h"
24 #include "llvm/CodeGen/SelectionDAG.h"
25 #include "llvm/CodeGen/CallingConvLower.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 /// FSRL - Bitwise logical right shift of floating point values. These
57 /// corresponds to X86::PSRLDQ.
60 /// CALL - These operations represent an abstract X86 call
61 /// instruction, which includes a bunch of information. In particular the
62 /// operands of these node are:
64 /// #0 - The incoming token chain
66 /// #2 - The number of arg bytes the caller pushes on the stack.
67 /// #3 - The number of arg bytes the callee pops off the stack.
68 /// #4 - The value to pass in AL/AX/EAX (optional)
69 /// #5 - The value to pass in DL/DX/EDX (optional)
71 /// The result values of these nodes are:
73 /// #0 - The outgoing token chain
74 /// #1 - The first register result value (optional)
75 /// #2 - The second register result value (optional)
79 /// RDTSC_DAG - This operation implements the lowering for
83 /// X86 compare and logical compare instructions.
86 /// X86 bit-test instructions.
89 /// X86 SetCC. Operand 0 is condition code, and operand 1 is the EFLAGS
90 /// operand, usually produced by a CMP instruction.
93 // Same as SETCC except it's materialized with a sbb and the value is all
94 // one's or all zero's.
95 SETCC_CARRY, // R = carry_bit ? ~0 : 0
97 /// X86 FP SETCC, implemented with CMP{cc}SS/CMP{cc}SD.
98 /// Operands are two FP values to compare; result is a mask of
99 /// 0s or 1s. Generally DTRT for C/C++ with NaNs.
102 /// X86 MOVMSK{pd|ps}, extracts sign bits of two or four FP values,
103 /// result in an integer GPR. Needs masking for scalar result.
106 /// X86 conditional moves. Operand 0 and operand 1 are the two values
107 /// to select from. Operand 2 is the condition code, and operand 3 is the
108 /// flag operand produced by a CMP or TEST instruction. It also writes a
112 /// X86 conditional branches. Operand 0 is the chain operand, operand 1
113 /// is the block to branch if condition is true, operand 2 is the
114 /// condition code, and operand 3 is the flag operand produced by a CMP
115 /// or TEST instruction.
118 /// Return with a flag operand. Operand 0 is the chain operand, operand
119 /// 1 is the number of bytes of stack to pop.
122 /// REP_STOS - Repeat fill, corresponds to X86::REP_STOSx.
125 /// REP_MOVS - Repeat move, corresponds to X86::REP_MOVSx.
128 /// GlobalBaseReg - On Darwin, this node represents the result of the popl
129 /// at function entry, used for PIC code.
132 /// Wrapper - A wrapper node for TargetConstantPool,
133 /// TargetExternalSymbol, and TargetGlobalAddress.
136 /// WrapperRIP - Special wrapper used under X86-64 PIC mode for RIP
137 /// relative displacements.
140 /// MOVQ2DQ - Copies a 64-bit value from an MMX vector to the low word
141 /// of an XMM vector, with the high word zero filled.
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 /// PEXTRB - Extract an 8-bit value from a vector and zero extend it to
150 /// i32, corresponds to X86::PEXTRB.
153 /// PEXTRW - Extract a 16-bit value from a vector and zero extend it to
154 /// i32, corresponds to X86::PEXTRW.
157 /// INSERTPS - Insert any element of a 4 x float vector into any element
158 /// of a destination 4 x floatvector.
161 /// PINSRB - Insert the lower 8-bits of a 32-bit value to a vector,
162 /// corresponds to X86::PINSRB.
165 /// PINSRW - Insert the lower 16-bits of a 32-bit value to a vector,
166 /// corresponds to X86::PINSRW.
169 /// PSHUFB - Shuffle 16 8-bit values within a vector.
172 /// ANDNP - Bitwise Logical AND NOT of Packed FP values.
175 /// PSIGN - Copy integer sign.
178 /// BLEND family of opcodes
181 /// HADD - Integer horizontal add.
184 /// HSUB - Integer horizontal sub.
187 /// FHADD - Floating point horizontal add.
190 /// FHSUB - Floating point horizontal sub.
193 /// FMAX, FMIN - Floating point max and min.
197 /// FRSQRT, FRCP - Floating point reciprocal-sqrt and reciprocal
198 /// approximation. Note that these typically require refinement
199 /// in order to obtain suitable precision.
202 // TLSADDR - Thread Local Storage.
205 // TLSCALL - Thread Local Storage. When calling to an OS provided
206 // thunk at the address from an earlier relocation.
209 // EH_RETURN - Exception Handling helpers.
212 /// TC_RETURN - Tail call return.
214 /// operand #1 callee (register or absolute)
215 /// operand #2 stack adjustment
216 /// operand #3 optional in flag
219 // VZEXT_MOVL - Vector move low and zero extend.
222 // VZEXT_MOVL - Vector move low and sign extend.
225 // VSHL, VSRL - 128-bit vector logical left / right shift
228 // VSHL, VSRL, VSRA - Vector shift elements
231 // VSHLI, VSRLI, VSRAI - Vector shift elements by immediate
234 // CMPP - Vector packed double/float comparison.
237 // PCMP* - Vector integer comparisons.
240 // VPCOM, VPCOMU - XOP Vector integer comparisons.
243 // ADD, SUB, SMUL, etc. - Arithmetic operations with FLAGS results.
244 ADD, SUB, ADC, SBB, SMUL,
245 INC, DEC, OR, XOR, AND,
247 ANDN, // ANDN - Bitwise AND NOT with FLAGS results.
249 BLSI, // BLSI - Extract lowest set isolated bit
250 BLSMSK, // BLSMSK - Get mask up to lowest set bit
251 BLSR, // BLSR - Reset lowest set bit
253 UMUL, // LOW, HI, FLAGS = umul LHS, RHS
255 // MUL_IMM - X86 specific multiply by immediate.
258 // PTEST - Vector bitwise comparisons
261 // TESTP - Vector packed fp sign bitwise comparisons
264 // Several flavors of instructions with vector shuffle behaviors.
286 // VASTART_SAVE_XMM_REGS - Save xmm argument registers to the stack,
287 // according to %al. An operator is needed so that this can be expanded
288 // with control flow.
289 VASTART_SAVE_XMM_REGS,
291 // WIN_ALLOCA - Windows's _chkstk call to do stack probing.
294 // SEG_ALLOCA - For allocating variable amounts of stack space when using
295 // segmented stacks. Check if the current stacklet has enough space, and
296 // falls back to heap allocation if not.
305 // ATOMADD64_DAG, ATOMSUB64_DAG, ATOMOR64_DAG, ATOMAND64_DAG,
306 // ATOMXOR64_DAG, ATOMNAND64_DAG, ATOMSWAP64_DAG -
307 // Atomic 64-bit binary operations.
308 ATOMADD64_DAG = ISD::FIRST_TARGET_MEMORY_OPCODE,
316 // LCMPXCHG_DAG, LCMPXCHG8_DAG, LCMPXCHG16_DAG - Compare and swap.
321 // VZEXT_LOAD - Load, scalar_to_vector, and zero extend.
324 // FNSTCW16m - Store FP control world into i16 memory.
327 /// FP_TO_INT*_IN_MEM - This instruction implements FP_TO_SINT with the
328 /// integer destination in memory and a FP reg source. This corresponds
329 /// to the X86::FIST*m instructions and the rounding mode change stuff. It
330 /// has two inputs (token chain and address) and two outputs (int value
331 /// and token chain).
336 /// FILD, FILD_FLAG - This instruction implements SINT_TO_FP with the
337 /// integer source in memory and FP reg result. This corresponds to the
338 /// X86::FILD*m instructions. It has three inputs (token chain, address,
339 /// and source type) and two outputs (FP value and token chain). FILD_FLAG
340 /// also produces a flag).
344 /// FLD - This instruction implements an extending load to FP stack slots.
345 /// This corresponds to the X86::FLD32m / X86::FLD64m. It takes a chain
346 /// operand, ptr to load from, and a ValueType node indicating the type
350 /// FST - This instruction implements a truncating store to FP stack
351 /// slots. This corresponds to the X86::FST32m / X86::FST64m. It takes a
352 /// chain operand, value to store, address, and a ValueType to store it
356 /// VAARG_64 - This instruction grabs the address of the next argument
357 /// from a va_list. (reads and modifies the va_list in memory)
360 // WARNING: Do not add anything in the end unless you want the node to
361 // have memop! In fact, starting from ATOMADD64_DAG all opcodes will be
362 // thought as target memory ops!
366 /// Define some predicates that are used for node matching.
368 /// isPSHUFDMask - Return true if the specified VECTOR_SHUFFLE operand
369 /// specifies a shuffle of elements that is suitable for input to PSHUFD.
370 bool isPSHUFDMask(ShuffleVectorSDNode *N);
372 /// isPSHUFHWMask - Return true if the specified VECTOR_SHUFFLE operand
373 /// specifies a shuffle of elements that is suitable for input to PSHUFD.
374 bool isPSHUFHWMask(ShuffleVectorSDNode *N);
376 /// isPSHUFLWMask - Return true if the specified VECTOR_SHUFFLE operand
377 /// specifies a shuffle of elements that is suitable for input to PSHUFD.
378 bool isPSHUFLWMask(ShuffleVectorSDNode *N);
380 /// isSHUFPMask - Return true if the specified VECTOR_SHUFFLE operand
381 /// specifies a shuffle of elements that is suitable for input to SHUFP*.
382 bool isSHUFPMask(ShuffleVectorSDNode *N, bool HasAVX);
384 /// isMOVHLPSMask - Return true if the specified VECTOR_SHUFFLE operand
385 /// specifies a shuffle of elements that is suitable for input to MOVHLPS.
386 bool isMOVHLPSMask(ShuffleVectorSDNode *N);
388 /// isMOVHLPS_v_undef_Mask - Special case of isMOVHLPSMask for canonical form
389 /// of vector_shuffle v, v, <2, 3, 2, 3>, i.e. vector_shuffle v, undef,
391 bool isMOVHLPS_v_undef_Mask(ShuffleVectorSDNode *N);
393 /// isMOVLPMask - Return true if the specified VECTOR_SHUFFLE operand
394 /// specifies a shuffle of elements that is suitable for MOVLP{S|D}.
395 bool isMOVLPMask(ShuffleVectorSDNode *N);
397 /// isMOVHPMask - Return true if the specified VECTOR_SHUFFLE operand
398 /// specifies a shuffle of elements that is suitable for MOVHP{S|D}.
399 /// as well as MOVLHPS.
400 bool isMOVLHPSMask(ShuffleVectorSDNode *N);
402 /// isUNPCKLMask - Return true if the specified VECTOR_SHUFFLE operand
403 /// specifies a shuffle of elements that is suitable for input to UNPCKL.
404 bool isUNPCKLMask(ShuffleVectorSDNode *N, bool HasAVX2,
405 bool V2IsSplat = false);
407 /// isUNPCKHMask - Return true if the specified VECTOR_SHUFFLE operand
408 /// specifies a shuffle of elements that is suitable for input to UNPCKH.
409 bool isUNPCKHMask(ShuffleVectorSDNode *N, bool HasAVX2,
410 bool V2IsSplat = false);
412 /// isUNPCKL_v_undef_Mask - Special case of isUNPCKLMask for canonical form
413 /// of vector_shuffle v, v, <0, 4, 1, 5>, i.e. vector_shuffle v, undef,
415 bool isUNPCKL_v_undef_Mask(ShuffleVectorSDNode *N, bool HasAVX2);
417 /// isUNPCKH_v_undef_Mask - Special case of isUNPCKHMask for canonical form
418 /// of vector_shuffle v, v, <2, 6, 3, 7>, i.e. vector_shuffle v, undef,
420 bool isUNPCKH_v_undef_Mask(ShuffleVectorSDNode *N, bool HasAVX2);
422 /// isMOVLMask - Return true if the specified VECTOR_SHUFFLE operand
423 /// specifies a shuffle of elements that is suitable for input to MOVSS,
424 /// MOVSD, and MOVD, i.e. setting the lowest element.
425 bool isMOVLMask(ShuffleVectorSDNode *N);
427 /// isMOVSHDUPMask - Return true if the specified VECTOR_SHUFFLE operand
428 /// specifies a shuffle of elements that is suitable for input to MOVSHDUP.
429 bool isMOVSHDUPMask(ShuffleVectorSDNode *N, const X86Subtarget *Subtarget);
431 /// isMOVSLDUPMask - Return true if the specified VECTOR_SHUFFLE operand
432 /// specifies a shuffle of elements that is suitable for input to MOVSLDUP.
433 bool isMOVSLDUPMask(ShuffleVectorSDNode *N, const X86Subtarget *Subtarget);
435 /// isMOVDDUPMask - Return true if the specified VECTOR_SHUFFLE operand
436 /// specifies a shuffle of elements that is suitable for input to MOVDDUP.
437 bool isMOVDDUPMask(ShuffleVectorSDNode *N);
439 /// isVEXTRACTF128Index - Return true if the specified
440 /// EXTRACT_SUBVECTOR operand specifies a vector extract that is
441 /// suitable for input to VEXTRACTF128.
442 bool isVEXTRACTF128Index(SDNode *N);
444 /// isVINSERTF128Index - Return true if the specified
445 /// INSERT_SUBVECTOR operand specifies a subvector insert that is
446 /// suitable for input to VINSERTF128.
447 bool isVINSERTF128Index(SDNode *N);
449 /// getShuffleSHUFImmediate - Return the appropriate immediate to shuffle
450 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUF* and SHUFP*
451 /// instructions. Handles 128-bit and 256-bit.
452 unsigned getShuffleSHUFImmediate(ShuffleVectorSDNode *N);
454 /// getShufflePSHUFHWImmediate - Return the appropriate immediate to shuffle
455 /// the specified VECTOR_SHUFFLE mask with PSHUFHW instruction.
456 unsigned getShufflePSHUFHWImmediate(SDNode *N);
458 /// getShufflePSHUFLWImmediate - Return the appropriate immediate to shuffle
459 /// the specified VECTOR_SHUFFLE mask with PSHUFLW instruction.
460 unsigned getShufflePSHUFLWImmediate(SDNode *N);
462 /// getExtractVEXTRACTF128Immediate - Return the appropriate
463 /// immediate to extract the specified EXTRACT_SUBVECTOR index
464 /// with VEXTRACTF128 instructions.
465 unsigned getExtractVEXTRACTF128Immediate(SDNode *N);
467 /// getInsertVINSERTF128Immediate - Return the appropriate
468 /// immediate to insert at the specified INSERT_SUBVECTOR index
469 /// with VINSERTF128 instructions.
470 unsigned getInsertVINSERTF128Immediate(SDNode *N);
472 /// isZeroNode - Returns true if Elt is a constant zero or a floating point
474 bool isZeroNode(SDValue Elt);
476 /// isOffsetSuitableForCodeModel - Returns true of the given offset can be
477 /// fit into displacement field of the instruction.
478 bool isOffsetSuitableForCodeModel(int64_t Offset, CodeModel::Model M,
479 bool hasSymbolicDisplacement = true);
482 /// isCalleePop - Determines whether the callee is required to pop its
483 /// own arguments. Callee pop is necessary to support tail calls.
484 bool isCalleePop(CallingConv::ID CallingConv,
485 bool is64Bit, bool IsVarArg, bool TailCallOpt);
488 //===--------------------------------------------------------------------===//
489 // X86TargetLowering - X86 Implementation of the TargetLowering interface
490 class X86TargetLowering : public TargetLowering {
492 explicit X86TargetLowering(X86TargetMachine &TM);
494 virtual unsigned getJumpTableEncoding() const;
496 virtual MVT getShiftAmountTy(EVT LHSTy) const { return MVT::i8; }
498 virtual const MCExpr *
499 LowerCustomJumpTableEntry(const MachineJumpTableInfo *MJTI,
500 const MachineBasicBlock *MBB, unsigned uid,
501 MCContext &Ctx) const;
503 /// getPICJumpTableRelocaBase - Returns relocation base for the given PIC
505 virtual SDValue getPICJumpTableRelocBase(SDValue Table,
506 SelectionDAG &DAG) const;
507 virtual const MCExpr *
508 getPICJumpTableRelocBaseExpr(const MachineFunction *MF,
509 unsigned JTI, MCContext &Ctx) const;
511 /// getStackPtrReg - Return the stack pointer register we are using: either
513 unsigned getStackPtrReg() const { return X86StackPtr; }
515 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
516 /// function arguments in the caller parameter area. For X86, aggregates
517 /// that contains are placed at 16-byte boundaries while the rest are at
518 /// 4-byte boundaries.
519 virtual unsigned getByValTypeAlignment(Type *Ty) const;
521 /// getOptimalMemOpType - Returns the target specific optimal type for load
522 /// and store operations as a result of memset, memcpy, and memmove
523 /// lowering. If DstAlign is zero that means it's safe to destination
524 /// alignment can satisfy any constraint. Similarly if SrcAlign is zero it
525 /// means there isn't a need to check it against alignment requirement,
526 /// probably because the source does not need to be loaded. If
527 /// 'IsZeroVal' is true, that means it's safe to return a
528 /// non-scalar-integer type, e.g. empty string source, constant, or loaded
529 /// from memory. 'MemcpyStrSrc' indicates whether the memcpy source is
530 /// constant so it does not need to be loaded.
531 /// It returns EVT::Other if the type should be determined using generic
532 /// target-independent logic.
534 getOptimalMemOpType(uint64_t Size, unsigned DstAlign, unsigned SrcAlign,
535 bool IsZeroVal, bool MemcpyStrSrc,
536 MachineFunction &MF) const;
538 /// allowsUnalignedMemoryAccesses - Returns true if the target allows
539 /// unaligned memory accesses. of the specified type.
540 virtual bool allowsUnalignedMemoryAccesses(EVT VT) const {
544 /// LowerOperation - Provide custom lowering hooks for some operations.
546 virtual SDValue LowerOperation(SDValue Op, SelectionDAG &DAG) const;
548 /// ReplaceNodeResults - Replace the results of node with an illegal result
549 /// type with new values built out of custom code.
551 virtual void ReplaceNodeResults(SDNode *N, SmallVectorImpl<SDValue>&Results,
552 SelectionDAG &DAG) const;
555 virtual SDValue PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const;
557 /// isTypeDesirableForOp - Return true if the target has native support for
558 /// the specified value type and it is 'desirable' to use the type for the
559 /// given node type. e.g. On x86 i16 is legal, but undesirable since i16
560 /// instruction encodings are longer and some i16 instructions are slow.
561 virtual bool isTypeDesirableForOp(unsigned Opc, EVT VT) const;
563 /// isTypeDesirable - Return true if the target has native support for the
564 /// specified value type and it is 'desirable' to use the type. e.g. On x86
565 /// i16 is legal, but undesirable since i16 instruction encodings are longer
566 /// and some i16 instructions are slow.
567 virtual bool IsDesirableToPromoteOp(SDValue Op, EVT &PVT) const;
569 virtual MachineBasicBlock *
570 EmitInstrWithCustomInserter(MachineInstr *MI,
571 MachineBasicBlock *MBB) const;
574 /// getTargetNodeName - This method returns the name of a target specific
576 virtual const char *getTargetNodeName(unsigned Opcode) const;
578 /// getSetCCResultType - Return the value type to use for ISD::SETCC.
579 virtual EVT getSetCCResultType(EVT VT) const;
581 /// computeMaskedBitsForTargetNode - Determine which of the bits specified
582 /// in Mask are known to be either zero or one and return them in the
583 /// KnownZero/KnownOne bitsets.
584 virtual void computeMaskedBitsForTargetNode(const SDValue Op,
588 const SelectionDAG &DAG,
589 unsigned Depth = 0) const;
591 // ComputeNumSignBitsForTargetNode - Determine the number of bits in the
592 // operation that are sign bits.
593 virtual unsigned ComputeNumSignBitsForTargetNode(SDValue Op,
594 unsigned Depth) const;
597 isGAPlusOffset(SDNode *N, const GlobalValue* &GA, int64_t &Offset) const;
599 SDValue getReturnAddressFrameIndex(SelectionDAG &DAG) const;
601 virtual bool ExpandInlineAsm(CallInst *CI) const;
603 ConstraintType getConstraintType(const std::string &Constraint) const;
605 /// Examine constraint string and operand type and determine a weight value.
606 /// The operand object must already have been set up with the operand type.
607 virtual ConstraintWeight getSingleConstraintMatchWeight(
608 AsmOperandInfo &info, const char *constraint) const;
610 virtual const char *LowerXConstraint(EVT ConstraintVT) const;
612 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
613 /// vector. If it is invalid, don't add anything to Ops. If hasMemory is
614 /// true it means one of the asm constraint of the inline asm instruction
615 /// being processed is 'm'.
616 virtual void LowerAsmOperandForConstraint(SDValue Op,
617 std::string &Constraint,
618 std::vector<SDValue> &Ops,
619 SelectionDAG &DAG) const;
621 /// getRegForInlineAsmConstraint - Given a physical register constraint
622 /// (e.g. {edx}), return the register number and the register class for the
623 /// register. This should only be used for C_Register constraints. On
624 /// error, this returns a register number of 0.
625 std::pair<unsigned, const TargetRegisterClass*>
626 getRegForInlineAsmConstraint(const std::string &Constraint,
629 /// isLegalAddressingMode - Return true if the addressing mode represented
630 /// by AM is legal for this target, for a load/store of the specified type.
631 virtual bool isLegalAddressingMode(const AddrMode &AM, Type *Ty)const;
633 /// isTruncateFree - Return true if it's free to truncate a value of
634 /// type Ty1 to type Ty2. e.g. On x86 it's free to truncate a i32 value in
635 /// register EAX to i16 by referencing its sub-register AX.
636 virtual bool isTruncateFree(Type *Ty1, Type *Ty2) const;
637 virtual bool isTruncateFree(EVT VT1, EVT VT2) const;
639 /// isZExtFree - Return true if any actual instruction that defines a
640 /// value of type Ty1 implicit zero-extends the value to Ty2 in the result
641 /// register. This does not necessarily include registers defined in
642 /// unknown ways, such as incoming arguments, or copies from unknown
643 /// virtual registers. Also, if isTruncateFree(Ty2, Ty1) is true, this
644 /// does not necessarily apply to truncate instructions. e.g. on x86-64,
645 /// all instructions that define 32-bit values implicit zero-extend the
646 /// result out to 64 bits.
647 virtual bool isZExtFree(Type *Ty1, Type *Ty2) const;
648 virtual bool isZExtFree(EVT VT1, EVT VT2) const;
650 /// isNarrowingProfitable - Return true if it's profitable to narrow
651 /// operations of type VT1 to VT2. e.g. on x86, it's profitable to narrow
652 /// from i32 to i8 but not from i32 to i16.
653 virtual bool isNarrowingProfitable(EVT VT1, EVT VT2) const;
655 /// isFPImmLegal - Returns true if the target can instruction select the
656 /// specified FP immediate natively. If false, the legalizer will
657 /// materialize the FP immediate as a load from a constant pool.
658 virtual bool isFPImmLegal(const APFloat &Imm, EVT VT) const;
660 /// isShuffleMaskLegal - Targets can use this to indicate that they only
661 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
662 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask
663 /// values are assumed to be legal.
664 virtual bool isShuffleMaskLegal(const SmallVectorImpl<int> &Mask,
667 /// isVectorClearMaskLegal - Similar to isShuffleMaskLegal. This is
668 /// used by Targets can use this to indicate if there is a suitable
669 /// VECTOR_SHUFFLE that can be used to replace a VAND with a constant
671 virtual bool isVectorClearMaskLegal(const SmallVectorImpl<int> &Mask,
674 /// ShouldShrinkFPConstant - If true, then instruction selection should
675 /// seek to shrink the FP constant of the specified type to a smaller type
676 /// in order to save space and / or reduce runtime.
677 virtual bool ShouldShrinkFPConstant(EVT VT) const {
678 // Don't shrink FP constpool if SSE2 is available since cvtss2sd is more
679 // expensive than a straight movsd. On the other hand, it's important to
680 // shrink long double fp constant since fldt is very slow.
681 return !X86ScalarSSEf64 || VT == MVT::f80;
684 const X86Subtarget* getSubtarget() const {
688 /// isScalarFPTypeInSSEReg - Return true if the specified scalar FP type is
689 /// computed in an SSE register, not on the X87 floating point stack.
690 bool isScalarFPTypeInSSEReg(EVT VT) const {
691 return (VT == MVT::f64 && X86ScalarSSEf64) || // f64 is when SSE2
692 (VT == MVT::f32 && X86ScalarSSEf32); // f32 is when SSE1
695 /// createFastISel - This method returns a target specific FastISel object,
696 /// or null if the target does not support "fast" ISel.
697 virtual FastISel *createFastISel(FunctionLoweringInfo &funcInfo) const;
699 /// getStackCookieLocation - Return true if the target stores stack
700 /// protector cookies at a fixed offset in some non-standard address
701 /// space, and populates the address space and offset as
703 virtual bool getStackCookieLocation(unsigned &AddressSpace, unsigned &Offset) const;
705 SDValue BuildFILD(SDValue Op, EVT SrcVT, SDValue Chain, SDValue StackSlot,
706 SelectionDAG &DAG) const;
709 std::pair<const TargetRegisterClass*, uint8_t>
710 findRepresentativeClass(EVT VT) const;
713 /// Subtarget - Keep a pointer to the X86Subtarget around so that we can
714 /// make the right decision when generating code for different targets.
715 const X86Subtarget *Subtarget;
716 const X86RegisterInfo *RegInfo;
717 const TargetData *TD;
719 /// X86StackPtr - X86 physical register used as stack ptr.
720 unsigned X86StackPtr;
722 /// X86ScalarSSEf32, X86ScalarSSEf64 - Select between SSE or x87
723 /// floating point ops.
724 /// When SSE is available, use it for f32 operations.
725 /// When SSE2 is available, use it for f64 operations.
726 bool X86ScalarSSEf32;
727 bool X86ScalarSSEf64;
729 /// LegalFPImmediates - A list of legal fp immediates.
730 std::vector<APFloat> LegalFPImmediates;
732 /// addLegalFPImmediate - Indicate that this x86 target can instruction
733 /// select the specified FP immediate natively.
734 void addLegalFPImmediate(const APFloat& Imm) {
735 LegalFPImmediates.push_back(Imm);
738 SDValue LowerCallResult(SDValue Chain, SDValue InFlag,
739 CallingConv::ID CallConv, bool isVarArg,
740 const SmallVectorImpl<ISD::InputArg> &Ins,
741 DebugLoc dl, SelectionDAG &DAG,
742 SmallVectorImpl<SDValue> &InVals) const;
743 SDValue LowerMemArgument(SDValue Chain,
744 CallingConv::ID CallConv,
745 const SmallVectorImpl<ISD::InputArg> &ArgInfo,
746 DebugLoc dl, SelectionDAG &DAG,
747 const CCValAssign &VA, MachineFrameInfo *MFI,
749 SDValue LowerMemOpCallTo(SDValue Chain, SDValue StackPtr, SDValue Arg,
750 DebugLoc dl, SelectionDAG &DAG,
751 const CCValAssign &VA,
752 ISD::ArgFlagsTy Flags) const;
754 // Call lowering helpers.
756 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
757 /// for tail call optimization. Targets which want to do tail call
758 /// optimization should implement this function.
759 bool IsEligibleForTailCallOptimization(SDValue Callee,
760 CallingConv::ID CalleeCC,
762 bool isCalleeStructRet,
763 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,
777 bool isSigned) const;
779 SDValue LowerAsSplatVectorLoad(SDValue SrcOp, EVT VT, DebugLoc dl,
780 SelectionDAG &DAG) const;
781 SDValue LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) const;
782 SDValue LowerCONCAT_VECTORS(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 LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) const;
789 SDValue LowerEXTRACT_SUBVECTOR(SDValue Op, SelectionDAG &DAG) const;
790 SDValue LowerINSERT_SUBVECTOR(SDValue Op, SelectionDAG &DAG) const;
791 SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) const;
792 SDValue LowerBlockAddress(SDValue Op, SelectionDAG &DAG) const;
793 SDValue LowerGlobalAddress(const GlobalValue *GV, DebugLoc dl,
794 int64_t Offset, SelectionDAG &DAG) const;
795 SDValue LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) const;
796 SDValue LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const;
797 SDValue LowerExternalSymbol(SDValue Op, SelectionDAG &DAG) const;
798 SDValue LowerShiftParts(SDValue Op, SelectionDAG &DAG) const;
799 SDValue LowerBITCAST(SDValue op, SelectionDAG &DAG) const;
800 SDValue LowerSINT_TO_FP(SDValue Op, SelectionDAG &DAG) const;
801 SDValue LowerUINT_TO_FP(SDValue Op, SelectionDAG &DAG) const;
802 SDValue LowerUINT_TO_FP_i64(SDValue Op, SelectionDAG &DAG) const;
803 SDValue LowerUINT_TO_FP_i32(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 LowerFABS(SDValue Op, SelectionDAG &DAG) const;
807 SDValue LowerFNEG(SDValue Op, SelectionDAG &DAG) const;
808 SDValue LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const;
809 SDValue LowerFGETSIGN(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 LowerVACOPY(SDValue Op, SelectionDAG &DAG) const;
822 SDValue LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) const;
823 SDValue LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const;
824 SDValue LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const;
825 SDValue LowerFRAME_TO_ARGS_OFFSET(SDValue Op, SelectionDAG &DAG) const;
826 SDValue LowerEH_RETURN(SDValue Op, SelectionDAG &DAG) const;
827 SDValue LowerINIT_TRAMPOLINE(SDValue Op, SelectionDAG &DAG) const;
828 SDValue LowerADJUST_TRAMPOLINE(SDValue Op, SelectionDAG &DAG) const;
829 SDValue LowerFLT_ROUNDS_(SDValue Op, SelectionDAG &DAG) const;
830 SDValue LowerCTLZ(SDValue Op, SelectionDAG &DAG) const;
831 SDValue LowerCTLZ_ZERO_UNDEF(SDValue Op, SelectionDAG &DAG) const;
832 SDValue LowerCTTZ(SDValue Op, SelectionDAG &DAG) const;
833 SDValue LowerADD(SDValue Op, SelectionDAG &DAG) const;
834 SDValue LowerSUB(SDValue Op, SelectionDAG &DAG) const;
835 SDValue LowerMUL(SDValue Op, SelectionDAG &DAG) const;
836 SDValue LowerShift(SDValue Op, SelectionDAG &DAG) const;
837 SDValue LowerXALUO(SDValue Op, SelectionDAG &DAG) const;
839 SDValue LowerCMP_SWAP(SDValue Op, SelectionDAG &DAG) const;
840 SDValue LowerLOAD_SUB(SDValue Op, SelectionDAG &DAG) const;
841 SDValue LowerREADCYCLECOUNTER(SDValue Op, SelectionDAG &DAG) const;
842 SDValue LowerMEMBARRIER(SDValue Op, SelectionDAG &DAG) const;
843 SDValue LowerATOMIC_FENCE(SDValue Op, SelectionDAG &DAG) const;
844 SDValue LowerSIGN_EXTEND_INREG(SDValue Op, SelectionDAG &DAG) const;
845 SDValue PerformTruncateCombine(SDNode* N, SelectionDAG &DAG, DAGCombinerInfo &DCI) const;
847 // Utility functions to help LowerVECTOR_SHUFFLE
848 SDValue LowerVECTOR_SHUFFLEv8i16(SDValue Op, SelectionDAG &DAG) const;
851 LowerFormalArguments(SDValue Chain,
852 CallingConv::ID CallConv, bool isVarArg,
853 const SmallVectorImpl<ISD::InputArg> &Ins,
854 DebugLoc dl, SelectionDAG &DAG,
855 SmallVectorImpl<SDValue> &InVals) const;
857 LowerCall(SDValue Chain, SDValue Callee,
858 CallingConv::ID CallConv, bool isVarArg, bool &isTailCall,
859 const SmallVectorImpl<ISD::OutputArg> &Outs,
860 const SmallVectorImpl<SDValue> &OutVals,
861 const SmallVectorImpl<ISD::InputArg> &Ins,
862 DebugLoc dl, SelectionDAG &DAG,
863 SmallVectorImpl<SDValue> &InVals) const;
866 LowerReturn(SDValue Chain,
867 CallingConv::ID CallConv, bool isVarArg,
868 const SmallVectorImpl<ISD::OutputArg> &Outs,
869 const SmallVectorImpl<SDValue> &OutVals,
870 DebugLoc dl, SelectionDAG &DAG) const;
872 virtual bool isUsedByReturnOnly(SDNode *N) const;
874 virtual bool mayBeEmittedAsTailCall(CallInst *CI) const;
877 getTypeForExtArgOrReturn(LLVMContext &Context, EVT VT,
878 ISD::NodeType ExtendKind) const;
881 CanLowerReturn(CallingConv::ID CallConv, MachineFunction &MF,
883 const SmallVectorImpl<ISD::OutputArg> &Outs,
884 LLVMContext &Context) const;
886 void ReplaceATOMIC_BINARY_64(SDNode *N, SmallVectorImpl<SDValue> &Results,
887 SelectionDAG &DAG, unsigned NewOp) const;
889 /// Utility function to emit string processing sse4.2 instructions
890 /// that return in xmm0.
891 /// This takes the instruction to expand, the associated machine basic
892 /// block, the number of args, and whether or not the second arg is
893 /// in memory or not.
894 MachineBasicBlock *EmitPCMP(MachineInstr *BInstr, MachineBasicBlock *BB,
895 unsigned argNum, bool inMem) const;
897 /// Utility functions to emit monitor and mwait instructions. These
898 /// need to make sure that the arguments to the intrinsic are in the
899 /// correct registers.
900 MachineBasicBlock *EmitMonitor(MachineInstr *MI,
901 MachineBasicBlock *BB) const;
902 MachineBasicBlock *EmitMwait(MachineInstr *MI, MachineBasicBlock *BB) const;
904 /// Utility function to emit atomic bitwise operations (and, or, xor).
905 /// It takes the bitwise instruction to expand, the associated machine basic
906 /// block, and the associated X86 opcodes for reg/reg and reg/imm.
907 MachineBasicBlock *EmitAtomicBitwiseWithCustomInserter(
908 MachineInstr *BInstr,
909 MachineBasicBlock *BB,
916 TargetRegisterClass *RC,
917 bool invSrc = false) const;
919 MachineBasicBlock *EmitAtomicBit6432WithCustomInserter(
920 MachineInstr *BInstr,
921 MachineBasicBlock *BB,
926 bool invSrc = false) const;
928 /// Utility function to emit atomic min and max. It takes the min/max
929 /// instruction to expand, the associated basic block, and the associated
930 /// cmov opcode for moving the min or max value.
931 MachineBasicBlock *EmitAtomicMinMaxWithCustomInserter(MachineInstr *BInstr,
932 MachineBasicBlock *BB,
933 unsigned cmovOpc) const;
935 // Utility function to emit the low-level va_arg code for X86-64.
936 MachineBasicBlock *EmitVAARG64WithCustomInserter(
938 MachineBasicBlock *MBB) const;
940 /// Utility function to emit the xmm reg save portion of va_start.
941 MachineBasicBlock *EmitVAStartSaveXMMRegsWithCustomInserter(
942 MachineInstr *BInstr,
943 MachineBasicBlock *BB) const;
945 MachineBasicBlock *EmitLoweredSelect(MachineInstr *I,
946 MachineBasicBlock *BB) const;
948 MachineBasicBlock *EmitLoweredWinAlloca(MachineInstr *MI,
949 MachineBasicBlock *BB) const;
951 MachineBasicBlock *EmitLoweredSegAlloca(MachineInstr *MI,
952 MachineBasicBlock *BB,
955 MachineBasicBlock *EmitLoweredTLSCall(MachineInstr *MI,
956 MachineBasicBlock *BB) const;
958 MachineBasicBlock *emitLoweredTLSAddr(MachineInstr *MI,
959 MachineBasicBlock *BB) const;
961 /// Emit nodes that will be selected as "test Op0,Op0", or something
962 /// equivalent, for use with the given x86 condition code.
963 SDValue EmitTest(SDValue Op0, unsigned X86CC, SelectionDAG &DAG) const;
965 /// Emit nodes that will be selected as "cmp Op0,Op1", or something
966 /// equivalent, for use with the given x86 condition code.
967 SDValue EmitCmp(SDValue Op0, SDValue Op1, unsigned X86CC,
968 SelectionDAG &DAG) const;
972 FastISel *createFastISel(FunctionLoweringInfo &funcInfo);
976 #endif // X86ISELLOWERING_H