1 //===-- X86InstrInfo.cpp - X86 Instruction Information --------------------===//
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
7 //===----------------------------------------------------------------------===//
9 // This file contains the X86 implementation of the TargetInstrInfo class.
11 //===----------------------------------------------------------------------===//
13 #include "X86InstrInfo.h"
15 #include "X86InstrBuilder.h"
16 #include "X86InstrFoldTables.h"
17 #include "X86MachineFunctionInfo.h"
18 #include "X86Subtarget.h"
19 #include "X86TargetMachine.h"
20 #include "llvm/ADT/STLExtras.h"
21 #include "llvm/ADT/Sequence.h"
22 #include "llvm/CodeGen/LivePhysRegs.h"
23 #include "llvm/CodeGen/LiveVariables.h"
24 #include "llvm/CodeGen/MachineConstantPool.h"
25 #include "llvm/CodeGen/MachineDominators.h"
26 #include "llvm/CodeGen/MachineFrameInfo.h"
27 #include "llvm/CodeGen/MachineInstrBuilder.h"
28 #include "llvm/CodeGen/MachineModuleInfo.h"
29 #include "llvm/CodeGen/MachineRegisterInfo.h"
30 #include "llvm/CodeGen/StackMaps.h"
31 #include "llvm/IR/DerivedTypes.h"
32 #include "llvm/IR/Function.h"
33 #include "llvm/IR/LLVMContext.h"
34 #include "llvm/MC/MCAsmInfo.h"
35 #include "llvm/MC/MCExpr.h"
36 #include "llvm/MC/MCInst.h"
37 #include "llvm/Support/CommandLine.h"
38 #include "llvm/Support/Debug.h"
39 #include "llvm/Support/ErrorHandling.h"
40 #include "llvm/Support/raw_ostream.h"
41 #include "llvm/Target/TargetOptions.h"
45 #define DEBUG_TYPE "x86-instr-info"
47 #define GET_INSTRINFO_CTOR_DTOR
48 #include "X86GenInstrInfo.inc"
51 NoFusing("disable-spill-fusing",
52 cl::desc("Disable fusing of spill code into instructions"),
55 PrintFailedFusing("print-failed-fuse-candidates",
56 cl::desc("Print instructions that the allocator wants to"
57 " fuse, but the X86 backend currently can't"),
60 ReMatPICStubLoad("remat-pic-stub-load",
61 cl::desc("Re-materialize load from stub in PIC mode"),
62 cl::init(false), cl::Hidden);
63 static cl::opt<unsigned>
64 PartialRegUpdateClearance("partial-reg-update-clearance",
65 cl::desc("Clearance between two register writes "
66 "for inserting XOR to avoid partial "
68 cl::init(64), cl::Hidden);
69 static cl::opt<unsigned>
70 UndefRegClearance("undef-reg-clearance",
71 cl::desc("How many idle instructions we would like before "
72 "certain undef register reads"),
73 cl::init(128), cl::Hidden);
76 // Pin the vtable to this file.
77 void X86InstrInfo::anchor() {}
79 X86InstrInfo::X86InstrInfo(X86Subtarget &STI)
80 : X86GenInstrInfo((STI.isTarget64BitLP64() ? X86::ADJCALLSTACKDOWN64
81 : X86::ADJCALLSTACKDOWN32),
82 (STI.isTarget64BitLP64() ? X86::ADJCALLSTACKUP64
83 : X86::ADJCALLSTACKUP32),
85 (STI.is64Bit() ? X86::RETQ : X86::RETL)),
86 Subtarget(STI), RI(STI.getTargetTriple()) {
90 X86InstrInfo::isCoalescableExtInstr(const MachineInstr &MI,
91 unsigned &SrcReg, unsigned &DstReg,
92 unsigned &SubIdx) const {
93 switch (MI.getOpcode()) {
100 if (!Subtarget.is64Bit())
101 // It's not always legal to reference the low 8-bit of the larger
102 // register in 32-bit mode.
105 case X86::MOVSX32rr16:
106 case X86::MOVZX32rr16:
107 case X86::MOVSX64rr16:
108 case X86::MOVSX64rr32: {
109 if (MI.getOperand(0).getSubReg() || MI.getOperand(1).getSubReg())
112 SrcReg = MI.getOperand(1).getReg();
113 DstReg = MI.getOperand(0).getReg();
114 switch (MI.getOpcode()) {
115 default: llvm_unreachable("Unreachable!");
116 case X86::MOVSX16rr8:
117 case X86::MOVZX16rr8:
118 case X86::MOVSX32rr8:
119 case X86::MOVZX32rr8:
120 case X86::MOVSX64rr8:
121 SubIdx = X86::sub_8bit;
123 case X86::MOVSX32rr16:
124 case X86::MOVZX32rr16:
125 case X86::MOVSX64rr16:
126 SubIdx = X86::sub_16bit;
128 case X86::MOVSX64rr32:
129 SubIdx = X86::sub_32bit;
138 int X86InstrInfo::getSPAdjust(const MachineInstr &MI) const {
139 const MachineFunction *MF = MI.getParent()->getParent();
140 const TargetFrameLowering *TFI = MF->getSubtarget().getFrameLowering();
142 if (isFrameInstr(MI)) {
143 unsigned StackAlign = TFI->getStackAlignment();
144 int SPAdj = alignTo(getFrameSize(MI), StackAlign);
145 SPAdj -= getFrameAdjustment(MI);
146 if (!isFrameSetup(MI))
151 // To know whether a call adjusts the stack, we need information
152 // that is bound to the following ADJCALLSTACKUP pseudo.
153 // Look for the next ADJCALLSTACKUP that follows the call.
155 const MachineBasicBlock *MBB = MI.getParent();
156 auto I = ++MachineBasicBlock::const_iterator(MI);
157 for (auto E = MBB->end(); I != E; ++I) {
158 if (I->getOpcode() == getCallFrameDestroyOpcode() ||
163 // If we could not find a frame destroy opcode, then it has already
164 // been simplified, so we don't care.
165 if (I->getOpcode() != getCallFrameDestroyOpcode())
168 return -(I->getOperand(1).getImm());
171 // Currently handle only PUSHes we can reasonably expect to see
173 switch (MI.getOpcode()) {
191 /// Return true and the FrameIndex if the specified
192 /// operand and follow operands form a reference to the stack frame.
193 bool X86InstrInfo::isFrameOperand(const MachineInstr &MI, unsigned int Op,
194 int &FrameIndex) const {
195 if (MI.getOperand(Op + X86::AddrBaseReg).isFI() &&
196 MI.getOperand(Op + X86::AddrScaleAmt).isImm() &&
197 MI.getOperand(Op + X86::AddrIndexReg).isReg() &&
198 MI.getOperand(Op + X86::AddrDisp).isImm() &&
199 MI.getOperand(Op + X86::AddrScaleAmt).getImm() == 1 &&
200 MI.getOperand(Op + X86::AddrIndexReg).getReg() == 0 &&
201 MI.getOperand(Op + X86::AddrDisp).getImm() == 0) {
202 FrameIndex = MI.getOperand(Op + X86::AddrBaseReg).getIndex();
208 static bool isFrameLoadOpcode(int Opcode, unsigned &MemBytes) {
232 case X86::MMX_MOVD64rm:
233 case X86::MMX_MOVQ64rm:
249 case X86::VMOVAPSZ128rm:
250 case X86::VMOVUPSZ128rm:
251 case X86::VMOVAPSZ128rm_NOVLX:
252 case X86::VMOVUPSZ128rm_NOVLX:
253 case X86::VMOVAPDZ128rm:
254 case X86::VMOVUPDZ128rm:
255 case X86::VMOVDQU8Z128rm:
256 case X86::VMOVDQU16Z128rm:
257 case X86::VMOVDQA32Z128rm:
258 case X86::VMOVDQU32Z128rm:
259 case X86::VMOVDQA64Z128rm:
260 case X86::VMOVDQU64Z128rm:
263 case X86::VMOVAPSYrm:
264 case X86::VMOVUPSYrm:
265 case X86::VMOVAPDYrm:
266 case X86::VMOVUPDYrm:
267 case X86::VMOVDQAYrm:
268 case X86::VMOVDQUYrm:
269 case X86::VMOVAPSZ256rm:
270 case X86::VMOVUPSZ256rm:
271 case X86::VMOVAPSZ256rm_NOVLX:
272 case X86::VMOVUPSZ256rm_NOVLX:
273 case X86::VMOVAPDZ256rm:
274 case X86::VMOVUPDZ256rm:
275 case X86::VMOVDQU8Z256rm:
276 case X86::VMOVDQU16Z256rm:
277 case X86::VMOVDQA32Z256rm:
278 case X86::VMOVDQU32Z256rm:
279 case X86::VMOVDQA64Z256rm:
280 case X86::VMOVDQU64Z256rm:
283 case X86::VMOVAPSZrm:
284 case X86::VMOVUPSZrm:
285 case X86::VMOVAPDZrm:
286 case X86::VMOVUPDZrm:
287 case X86::VMOVDQU8Zrm:
288 case X86::VMOVDQU16Zrm:
289 case X86::VMOVDQA32Zrm:
290 case X86::VMOVDQU32Zrm:
291 case X86::VMOVDQA64Zrm:
292 case X86::VMOVDQU64Zrm:
298 static bool isFrameStoreOpcode(int Opcode, unsigned &MemBytes) {
322 case X86::MMX_MOVD64mr:
323 case X86::MMX_MOVQ64mr:
324 case X86::MMX_MOVNTQmr:
340 case X86::VMOVUPSZ128mr:
341 case X86::VMOVAPSZ128mr:
342 case X86::VMOVUPSZ128mr_NOVLX:
343 case X86::VMOVAPSZ128mr_NOVLX:
344 case X86::VMOVUPDZ128mr:
345 case X86::VMOVAPDZ128mr:
346 case X86::VMOVDQA32Z128mr:
347 case X86::VMOVDQU32Z128mr:
348 case X86::VMOVDQA64Z128mr:
349 case X86::VMOVDQU64Z128mr:
350 case X86::VMOVDQU8Z128mr:
351 case X86::VMOVDQU16Z128mr:
354 case X86::VMOVUPSYmr:
355 case X86::VMOVAPSYmr:
356 case X86::VMOVUPDYmr:
357 case X86::VMOVAPDYmr:
358 case X86::VMOVDQUYmr:
359 case X86::VMOVDQAYmr:
360 case X86::VMOVUPSZ256mr:
361 case X86::VMOVAPSZ256mr:
362 case X86::VMOVUPSZ256mr_NOVLX:
363 case X86::VMOVAPSZ256mr_NOVLX:
364 case X86::VMOVUPDZ256mr:
365 case X86::VMOVAPDZ256mr:
366 case X86::VMOVDQU8Z256mr:
367 case X86::VMOVDQU16Z256mr:
368 case X86::VMOVDQA32Z256mr:
369 case X86::VMOVDQU32Z256mr:
370 case X86::VMOVDQA64Z256mr:
371 case X86::VMOVDQU64Z256mr:
374 case X86::VMOVUPSZmr:
375 case X86::VMOVAPSZmr:
376 case X86::VMOVUPDZmr:
377 case X86::VMOVAPDZmr:
378 case X86::VMOVDQU8Zmr:
379 case X86::VMOVDQU16Zmr:
380 case X86::VMOVDQA32Zmr:
381 case X86::VMOVDQU32Zmr:
382 case X86::VMOVDQA64Zmr:
383 case X86::VMOVDQU64Zmr:
390 unsigned X86InstrInfo::isLoadFromStackSlot(const MachineInstr &MI,
391 int &FrameIndex) const {
393 return X86InstrInfo::isLoadFromStackSlot(MI, FrameIndex, Dummy);
396 unsigned X86InstrInfo::isLoadFromStackSlot(const MachineInstr &MI,
398 unsigned &MemBytes) const {
399 if (isFrameLoadOpcode(MI.getOpcode(), MemBytes))
400 if (MI.getOperand(0).getSubReg() == 0 && isFrameOperand(MI, 1, FrameIndex))
401 return MI.getOperand(0).getReg();
405 unsigned X86InstrInfo::isLoadFromStackSlotPostFE(const MachineInstr &MI,
406 int &FrameIndex) const {
408 if (isFrameLoadOpcode(MI.getOpcode(), Dummy)) {
410 if ((Reg = isLoadFromStackSlot(MI, FrameIndex)))
412 // Check for post-frame index elimination operations
413 SmallVector<const MachineMemOperand *, 1> Accesses;
414 if (hasLoadFromStackSlot(MI, Accesses)) {
416 cast<FixedStackPseudoSourceValue>(Accesses.front()->getPseudoValue())
424 unsigned X86InstrInfo::isStoreToStackSlot(const MachineInstr &MI,
425 int &FrameIndex) const {
427 return X86InstrInfo::isStoreToStackSlot(MI, FrameIndex, Dummy);
430 unsigned X86InstrInfo::isStoreToStackSlot(const MachineInstr &MI,
432 unsigned &MemBytes) const {
433 if (isFrameStoreOpcode(MI.getOpcode(), MemBytes))
434 if (MI.getOperand(X86::AddrNumOperands).getSubReg() == 0 &&
435 isFrameOperand(MI, 0, FrameIndex))
436 return MI.getOperand(X86::AddrNumOperands).getReg();
440 unsigned X86InstrInfo::isStoreToStackSlotPostFE(const MachineInstr &MI,
441 int &FrameIndex) const {
443 if (isFrameStoreOpcode(MI.getOpcode(), Dummy)) {
445 if ((Reg = isStoreToStackSlot(MI, FrameIndex)))
447 // Check for post-frame index elimination operations
448 SmallVector<const MachineMemOperand *, 1> Accesses;
449 if (hasStoreToStackSlot(MI, Accesses)) {
451 cast<FixedStackPseudoSourceValue>(Accesses.front()->getPseudoValue())
459 /// Return true if register is PIC base; i.e.g defined by X86::MOVPC32r.
460 static bool regIsPICBase(unsigned BaseReg, const MachineRegisterInfo &MRI) {
461 // Don't waste compile time scanning use-def chains of physregs.
462 if (!TargetRegisterInfo::isVirtualRegister(BaseReg))
464 bool isPICBase = false;
465 for (MachineRegisterInfo::def_instr_iterator I = MRI.def_instr_begin(BaseReg),
466 E = MRI.def_instr_end(); I != E; ++I) {
467 MachineInstr *DefMI = &*I;
468 if (DefMI->getOpcode() != X86::MOVPC32r)
470 assert(!isPICBase && "More than one PIC base?");
476 bool X86InstrInfo::isReallyTriviallyReMaterializable(const MachineInstr &MI,
477 AliasAnalysis *AA) const {
478 switch (MI.getOpcode()) {
481 case X86::MOV8rm_NOREX:
501 case X86::VMOVAPSYrm:
502 case X86::VMOVUPSYrm:
503 case X86::VMOVAPDYrm:
504 case X86::VMOVUPDYrm:
505 case X86::VMOVDQAYrm:
506 case X86::VMOVDQUYrm:
507 case X86::MMX_MOVD64rm:
508 case X86::MMX_MOVQ64rm:
512 case X86::VMOVAPDZ128rm:
513 case X86::VMOVAPDZ256rm:
514 case X86::VMOVAPDZrm:
515 case X86::VMOVAPSZ128rm:
516 case X86::VMOVAPSZ256rm:
517 case X86::VMOVAPSZ128rm_NOVLX:
518 case X86::VMOVAPSZ256rm_NOVLX:
519 case X86::VMOVAPSZrm:
520 case X86::VMOVDQA32Z128rm:
521 case X86::VMOVDQA32Z256rm:
522 case X86::VMOVDQA32Zrm:
523 case X86::VMOVDQA64Z128rm:
524 case X86::VMOVDQA64Z256rm:
525 case X86::VMOVDQA64Zrm:
526 case X86::VMOVDQU16Z128rm:
527 case X86::VMOVDQU16Z256rm:
528 case X86::VMOVDQU16Zrm:
529 case X86::VMOVDQU32Z128rm:
530 case X86::VMOVDQU32Z256rm:
531 case X86::VMOVDQU32Zrm:
532 case X86::VMOVDQU64Z128rm:
533 case X86::VMOVDQU64Z256rm:
534 case X86::VMOVDQU64Zrm:
535 case X86::VMOVDQU8Z128rm:
536 case X86::VMOVDQU8Z256rm:
537 case X86::VMOVDQU8Zrm:
538 case X86::VMOVUPDZ128rm:
539 case X86::VMOVUPDZ256rm:
540 case X86::VMOVUPDZrm:
541 case X86::VMOVUPSZ128rm:
542 case X86::VMOVUPSZ256rm:
543 case X86::VMOVUPSZ128rm_NOVLX:
544 case X86::VMOVUPSZ256rm_NOVLX:
545 case X86::VMOVUPSZrm: {
546 // Loads from constant pools are trivially rematerializable.
547 if (MI.getOperand(1 + X86::AddrBaseReg).isReg() &&
548 MI.getOperand(1 + X86::AddrScaleAmt).isImm() &&
549 MI.getOperand(1 + X86::AddrIndexReg).isReg() &&
550 MI.getOperand(1 + X86::AddrIndexReg).getReg() == 0 &&
551 MI.isDereferenceableInvariantLoad(AA)) {
552 unsigned BaseReg = MI.getOperand(1 + X86::AddrBaseReg).getReg();
553 if (BaseReg == 0 || BaseReg == X86::RIP)
555 // Allow re-materialization of PIC load.
556 if (!ReMatPICStubLoad && MI.getOperand(1 + X86::AddrDisp).isGlobal())
558 const MachineFunction &MF = *MI.getParent()->getParent();
559 const MachineRegisterInfo &MRI = MF.getRegInfo();
560 return regIsPICBase(BaseReg, MRI);
567 if (MI.getOperand(1 + X86::AddrScaleAmt).isImm() &&
568 MI.getOperand(1 + X86::AddrIndexReg).isReg() &&
569 MI.getOperand(1 + X86::AddrIndexReg).getReg() == 0 &&
570 !MI.getOperand(1 + X86::AddrDisp).isReg()) {
571 // lea fi#, lea GV, etc. are all rematerializable.
572 if (!MI.getOperand(1 + X86::AddrBaseReg).isReg())
574 unsigned BaseReg = MI.getOperand(1 + X86::AddrBaseReg).getReg();
577 // Allow re-materialization of lea PICBase + x.
578 const MachineFunction &MF = *MI.getParent()->getParent();
579 const MachineRegisterInfo &MRI = MF.getRegInfo();
580 return regIsPICBase(BaseReg, MRI);
586 // All other instructions marked M_REMATERIALIZABLE are always trivially
591 void X86InstrInfo::reMaterialize(MachineBasicBlock &MBB,
592 MachineBasicBlock::iterator I,
593 unsigned DestReg, unsigned SubIdx,
594 const MachineInstr &Orig,
595 const TargetRegisterInfo &TRI) const {
596 bool ClobbersEFLAGS = Orig.modifiesRegister(X86::EFLAGS, &TRI);
597 if (ClobbersEFLAGS && !isSafeToClobberEFLAGS(MBB, I)) {
598 // The instruction clobbers EFLAGS. Re-materialize as MOV32ri to avoid side
601 switch (Orig.getOpcode()) {
602 case X86::MOV32r0: Value = 0; break;
603 case X86::MOV32r1: Value = 1; break;
604 case X86::MOV32r_1: Value = -1; break;
606 llvm_unreachable("Unexpected instruction!");
609 const DebugLoc &DL = Orig.getDebugLoc();
610 BuildMI(MBB, I, DL, get(X86::MOV32ri))
611 .add(Orig.getOperand(0))
614 MachineInstr *MI = MBB.getParent()->CloneMachineInstr(&Orig);
618 MachineInstr &NewMI = *std::prev(I);
619 NewMI.substituteRegister(Orig.getOperand(0).getReg(), DestReg, SubIdx, TRI);
622 /// True if MI has a condition code def, e.g. EFLAGS, that is not marked dead.
623 bool X86InstrInfo::hasLiveCondCodeDef(MachineInstr &MI) const {
624 for (unsigned i = 0, e = MI.getNumOperands(); i != e; ++i) {
625 MachineOperand &MO = MI.getOperand(i);
626 if (MO.isReg() && MO.isDef() &&
627 MO.getReg() == X86::EFLAGS && !MO.isDead()) {
634 /// Check whether the shift count for a machine operand is non-zero.
635 inline static unsigned getTruncatedShiftCount(const MachineInstr &MI,
636 unsigned ShiftAmtOperandIdx) {
637 // The shift count is six bits with the REX.W prefix and five bits without.
638 unsigned ShiftCountMask = (MI.getDesc().TSFlags & X86II::REX_W) ? 63 : 31;
639 unsigned Imm = MI.getOperand(ShiftAmtOperandIdx).getImm();
640 return Imm & ShiftCountMask;
643 /// Check whether the given shift count is appropriate
644 /// can be represented by a LEA instruction.
645 inline static bool isTruncatedShiftCountForLEA(unsigned ShAmt) {
646 // Left shift instructions can be transformed into load-effective-address
647 // instructions if we can encode them appropriately.
648 // A LEA instruction utilizes a SIB byte to encode its scale factor.
649 // The SIB.scale field is two bits wide which means that we can encode any
650 // shift amount less than 4.
651 return ShAmt < 4 && ShAmt > 0;
654 bool X86InstrInfo::classifyLEAReg(MachineInstr &MI, const MachineOperand &Src,
655 unsigned Opc, bool AllowSP, unsigned &NewSrc,
656 bool &isKill, MachineOperand &ImplicitOp,
657 LiveVariables *LV) const {
658 MachineFunction &MF = *MI.getParent()->getParent();
659 const TargetRegisterClass *RC;
661 RC = Opc != X86::LEA32r ? &X86::GR64RegClass : &X86::GR32RegClass;
663 RC = Opc != X86::LEA32r ?
664 &X86::GR64_NOSPRegClass : &X86::GR32_NOSPRegClass;
666 unsigned SrcReg = Src.getReg();
668 // For both LEA64 and LEA32 the register already has essentially the right
669 // type (32-bit or 64-bit) we may just need to forbid SP.
670 if (Opc != X86::LEA64_32r) {
672 isKill = Src.isKill();
673 assert(!Src.isUndef() && "Undef op doesn't need optimization");
675 if (TargetRegisterInfo::isVirtualRegister(NewSrc) &&
676 !MF.getRegInfo().constrainRegClass(NewSrc, RC))
682 // This is for an LEA64_32r and incoming registers are 32-bit. One way or
683 // another we need to add 64-bit registers to the final MI.
684 if (TargetRegisterInfo::isPhysicalRegister(SrcReg)) {
686 ImplicitOp.setImplicit();
688 NewSrc = getX86SubSuperRegister(Src.getReg(), 64);
689 isKill = Src.isKill();
690 assert(!Src.isUndef() && "Undef op doesn't need optimization");
692 // Virtual register of the wrong class, we have to create a temporary 64-bit
693 // vreg to feed into the LEA.
694 NewSrc = MF.getRegInfo().createVirtualRegister(RC);
696 BuildMI(*MI.getParent(), MI, MI.getDebugLoc(), get(TargetOpcode::COPY))
697 .addReg(NewSrc, RegState::Define | RegState::Undef, X86::sub_32bit)
700 // Which is obviously going to be dead after we're done with it.
704 LV->replaceKillInstruction(SrcReg, MI, *Copy);
707 // We've set all the parameters without issue.
711 MachineInstr *X86InstrInfo::convertToThreeAddressWithLEA(
712 unsigned MIOpc, MachineFunction::iterator &MFI, MachineInstr &MI,
713 LiveVariables *LV) const {
714 // We handle 8-bit adds and various 16-bit opcodes in the switch below.
715 bool Is16BitOp = !(MIOpc == X86::ADD8rr || MIOpc == X86::ADD8ri);
716 MachineRegisterInfo &RegInfo = MFI->getParent()->getRegInfo();
717 assert((!Is16BitOp || RegInfo.getTargetRegisterInfo()->getRegSizeInBits(
718 *RegInfo.getRegClass(MI.getOperand(0).getReg())) == 16) &&
719 "Unexpected type for LEA transform");
721 // TODO: For a 32-bit target, we need to adjust the LEA variables with
722 // something like this:
723 // Opcode = X86::LEA32r;
724 // InRegLEA = RegInfo.createVirtualRegister(&X86::GR32_NOSPRegClass);
726 // Is8BitOp ? RegInfo.createVirtualRegister(&X86::GR32ABCD_RegClass)
727 // : RegInfo.createVirtualRegister(&X86::GR32RegClass);
728 if (!Subtarget.is64Bit())
731 unsigned Opcode = X86::LEA64_32r;
732 unsigned InRegLEA = RegInfo.createVirtualRegister(&X86::GR64_NOSPRegClass);
733 unsigned OutRegLEA = RegInfo.createVirtualRegister(&X86::GR32RegClass);
735 // Build and insert into an implicit UNDEF value. This is OK because
736 // we will be shifting and then extracting the lower 8/16-bits.
737 // This has the potential to cause partial register stall. e.g.
738 // movw (%rbp,%rcx,2), %dx
739 // leal -65(%rdx), %esi
740 // But testing has shown this *does* help performance in 64-bit mode (at
741 // least on modern x86 machines).
742 MachineBasicBlock::iterator MBBI = MI.getIterator();
743 unsigned Dest = MI.getOperand(0).getReg();
744 unsigned Src = MI.getOperand(1).getReg();
745 bool IsDead = MI.getOperand(0).isDead();
746 bool IsKill = MI.getOperand(1).isKill();
747 unsigned SubReg = Is16BitOp ? X86::sub_16bit : X86::sub_8bit;
748 assert(!MI.getOperand(1).isUndef() && "Undef op doesn't need optimization");
749 BuildMI(*MFI, MBBI, MI.getDebugLoc(), get(X86::IMPLICIT_DEF), InRegLEA);
750 MachineInstr *InsMI =
751 BuildMI(*MFI, MBBI, MI.getDebugLoc(), get(TargetOpcode::COPY))
752 .addReg(InRegLEA, RegState::Define, SubReg)
753 .addReg(Src, getKillRegState(IsKill));
755 MachineInstrBuilder MIB =
756 BuildMI(*MFI, MBBI, MI.getDebugLoc(), get(Opcode), OutRegLEA);
758 default: llvm_unreachable("Unreachable!");
760 unsigned ShAmt = MI.getOperand(2).getImm();
761 MIB.addReg(0).addImm(1ULL << ShAmt)
762 .addReg(InRegLEA, RegState::Kill).addImm(0).addReg(0);
766 addRegOffset(MIB, InRegLEA, true, 1);
769 addRegOffset(MIB, InRegLEA, true, -1);
774 case X86::ADD16ri_DB:
775 case X86::ADD16ri8_DB:
776 addRegOffset(MIB, InRegLEA, true, MI.getOperand(2).getImm());
780 case X86::ADD16rr_DB: {
781 unsigned Src2 = MI.getOperand(2).getReg();
782 bool IsKill2 = MI.getOperand(2).isKill();
783 assert(!MI.getOperand(2).isUndef() && "Undef op doesn't need optimization");
784 unsigned InRegLEA2 = 0;
785 MachineInstr *InsMI2 = nullptr;
787 // ADD8rr/ADD16rr killed %reg1028, %reg1028
788 // just a single insert_subreg.
789 addRegReg(MIB, InRegLEA, true, InRegLEA, false);
791 if (Subtarget.is64Bit())
792 InRegLEA2 = RegInfo.createVirtualRegister(&X86::GR64_NOSPRegClass);
794 InRegLEA2 = RegInfo.createVirtualRegister(&X86::GR32_NOSPRegClass);
795 // Build and insert into an implicit UNDEF value. This is OK because
796 // we will be shifting and then extracting the lower 8/16-bits.
797 BuildMI(*MFI, &*MIB, MI.getDebugLoc(), get(X86::IMPLICIT_DEF), InRegLEA2);
798 InsMI2 = BuildMI(*MFI, &*MIB, MI.getDebugLoc(), get(TargetOpcode::COPY))
799 .addReg(InRegLEA2, RegState::Define, SubReg)
800 .addReg(Src2, getKillRegState(IsKill2));
801 addRegReg(MIB, InRegLEA, true, InRegLEA2, true);
803 if (LV && IsKill2 && InsMI2)
804 LV->replaceKillInstruction(Src2, MI, *InsMI2);
809 MachineInstr *NewMI = MIB;
810 MachineInstr *ExtMI =
811 BuildMI(*MFI, MBBI, MI.getDebugLoc(), get(TargetOpcode::COPY))
812 .addReg(Dest, RegState::Define | getDeadRegState(IsDead))
813 .addReg(OutRegLEA, RegState::Kill, SubReg);
816 // Update live variables.
817 LV->getVarInfo(InRegLEA).Kills.push_back(NewMI);
818 LV->getVarInfo(OutRegLEA).Kills.push_back(ExtMI);
820 LV->replaceKillInstruction(Src, MI, *InsMI);
822 LV->replaceKillInstruction(Dest, MI, *ExtMI);
828 /// This method must be implemented by targets that
829 /// set the M_CONVERTIBLE_TO_3_ADDR flag. When this flag is set, the target
830 /// may be able to convert a two-address instruction into a true
831 /// three-address instruction on demand. This allows the X86 target (for
832 /// example) to convert ADD and SHL instructions into LEA instructions if they
833 /// would require register copies due to two-addressness.
835 /// This method returns a null pointer if the transformation cannot be
836 /// performed, otherwise it returns the new instruction.
839 X86InstrInfo::convertToThreeAddress(MachineFunction::iterator &MFI,
840 MachineInstr &MI, LiveVariables *LV) const {
841 // The following opcodes also sets the condition code register(s). Only
842 // convert them to equivalent lea if the condition code register def's
844 if (hasLiveCondCodeDef(MI))
847 MachineFunction &MF = *MI.getParent()->getParent();
848 // All instructions input are two-addr instructions. Get the known operands.
849 const MachineOperand &Dest = MI.getOperand(0);
850 const MachineOperand &Src = MI.getOperand(1);
852 // Ideally, operations with undef should be folded before we get here, but we
853 // can't guarantee it. Bail out because optimizing undefs is a waste of time.
854 // Without this, we have to forward undef state to new register operands to
855 // avoid machine verifier errors.
858 if (MI.getNumOperands() > 2)
859 if (MI.getOperand(2).isReg() && MI.getOperand(2).isUndef())
862 MachineInstr *NewMI = nullptr;
863 bool Is64Bit = Subtarget.is64Bit();
865 unsigned MIOpc = MI.getOpcode();
867 default: return nullptr;
869 assert(MI.getNumOperands() >= 3 && "Unknown shift instruction!");
870 unsigned ShAmt = getTruncatedShiftCount(MI, 2);
871 if (!isTruncatedShiftCountForLEA(ShAmt)) return nullptr;
873 // LEA can't handle RSP.
874 if (TargetRegisterInfo::isVirtualRegister(Src.getReg()) &&
875 !MF.getRegInfo().constrainRegClass(Src.getReg(),
876 &X86::GR64_NOSPRegClass))
879 NewMI = BuildMI(MF, MI.getDebugLoc(), get(X86::LEA64r))
882 .addImm(1ULL << ShAmt)
889 assert(MI.getNumOperands() >= 3 && "Unknown shift instruction!");
890 unsigned ShAmt = getTruncatedShiftCount(MI, 2);
891 if (!isTruncatedShiftCountForLEA(ShAmt)) return nullptr;
893 unsigned Opc = Is64Bit ? X86::LEA64_32r : X86::LEA32r;
895 // LEA can't handle ESP.
898 MachineOperand ImplicitOp = MachineOperand::CreateReg(0, false);
899 if (!classifyLEAReg(MI, Src, Opc, /*AllowSP=*/ false,
900 SrcReg, isKill, ImplicitOp, LV))
903 MachineInstrBuilder MIB =
904 BuildMI(MF, MI.getDebugLoc(), get(Opc))
907 .addImm(1ULL << ShAmt)
908 .addReg(SrcReg, getKillRegState(isKill))
911 if (ImplicitOp.getReg() != 0)
918 assert(MI.getNumOperands() >= 3 && "Unknown shift instruction!");
919 unsigned ShAmt = getTruncatedShiftCount(MI, 2);
920 if (!isTruncatedShiftCountForLEA(ShAmt))
922 return convertToThreeAddressWithLEA(MIOpc, MFI, MI, LV);
926 assert(MI.getNumOperands() >= 2 && "Unknown inc instruction!");
927 unsigned Opc = MIOpc == X86::INC64r ? X86::LEA64r :
928 (Is64Bit ? X86::LEA64_32r : X86::LEA32r);
931 MachineOperand ImplicitOp = MachineOperand::CreateReg(0, false);
932 if (!classifyLEAReg(MI, Src, Opc, /*AllowSP=*/ false, SrcReg, isKill,
936 MachineInstrBuilder MIB =
937 BuildMI(MF, MI.getDebugLoc(), get(Opc))
939 .addReg(SrcReg, getKillRegState(isKill));
940 if (ImplicitOp.getReg() != 0)
943 NewMI = addOffset(MIB, 1);
947 return convertToThreeAddressWithLEA(MIOpc, MFI, MI, LV);
950 assert(MI.getNumOperands() >= 2 && "Unknown dec instruction!");
951 unsigned Opc = MIOpc == X86::DEC64r ? X86::LEA64r
952 : (Is64Bit ? X86::LEA64_32r : X86::LEA32r);
956 MachineOperand ImplicitOp = MachineOperand::CreateReg(0, false);
957 if (!classifyLEAReg(MI, Src, Opc, /*AllowSP=*/ false, SrcReg, isKill,
961 MachineInstrBuilder MIB = BuildMI(MF, MI.getDebugLoc(), get(Opc))
963 .addReg(SrcReg, getKillRegState(isKill));
964 if (ImplicitOp.getReg() != 0)
967 NewMI = addOffset(MIB, -1);
972 return convertToThreeAddressWithLEA(MIOpc, MFI, MI, LV);
974 case X86::ADD64rr_DB:
976 case X86::ADD32rr_DB: {
977 assert(MI.getNumOperands() >= 3 && "Unknown add instruction!");
979 if (MIOpc == X86::ADD64rr || MIOpc == X86::ADD64rr_DB)
982 Opc = Is64Bit ? X86::LEA64_32r : X86::LEA32r;
986 MachineOperand ImplicitOp = MachineOperand::CreateReg(0, false);
987 if (!classifyLEAReg(MI, Src, Opc, /*AllowSP=*/ true,
988 SrcReg, isKill, ImplicitOp, LV))
991 const MachineOperand &Src2 = MI.getOperand(2);
994 MachineOperand ImplicitOp2 = MachineOperand::CreateReg(0, false);
995 if (!classifyLEAReg(MI, Src2, Opc, /*AllowSP=*/ false,
996 SrcReg2, isKill2, ImplicitOp2, LV))
999 MachineInstrBuilder MIB = BuildMI(MF, MI.getDebugLoc(), get(Opc)).add(Dest);
1000 if (ImplicitOp.getReg() != 0)
1001 MIB.add(ImplicitOp);
1002 if (ImplicitOp2.getReg() != 0)
1003 MIB.add(ImplicitOp2);
1005 NewMI = addRegReg(MIB, SrcReg, isKill, SrcReg2, isKill2);
1006 if (LV && Src2.isKill())
1007 LV->replaceKillInstruction(SrcReg2, MI, *NewMI);
1012 case X86::ADD16rr_DB:
1013 return convertToThreeAddressWithLEA(MIOpc, MFI, MI, LV);
1014 case X86::ADD64ri32:
1016 case X86::ADD64ri32_DB:
1017 case X86::ADD64ri8_DB:
1018 assert(MI.getNumOperands() >= 3 && "Unknown add instruction!");
1020 BuildMI(MF, MI.getDebugLoc(), get(X86::LEA64r)).add(Dest).add(Src),
1025 case X86::ADD32ri_DB:
1026 case X86::ADD32ri8_DB: {
1027 assert(MI.getNumOperands() >= 3 && "Unknown add instruction!");
1028 unsigned Opc = Is64Bit ? X86::LEA64_32r : X86::LEA32r;
1032 MachineOperand ImplicitOp = MachineOperand::CreateReg(0, false);
1033 if (!classifyLEAReg(MI, Src, Opc, /*AllowSP=*/ true,
1034 SrcReg, isKill, ImplicitOp, LV))
1037 MachineInstrBuilder MIB = BuildMI(MF, MI.getDebugLoc(), get(Opc))
1039 .addReg(SrcReg, getKillRegState(isKill));
1040 if (ImplicitOp.getReg() != 0)
1041 MIB.add(ImplicitOp);
1043 NewMI = addOffset(MIB, MI.getOperand(2));
1049 case X86::ADD16ri_DB:
1050 case X86::ADD16ri8_DB:
1051 return convertToThreeAddressWithLEA(MIOpc, MFI, MI, LV);
1052 case X86::VMOVDQU8Z128rmk:
1053 case X86::VMOVDQU8Z256rmk:
1054 case X86::VMOVDQU8Zrmk:
1055 case X86::VMOVDQU16Z128rmk:
1056 case X86::VMOVDQU16Z256rmk:
1057 case X86::VMOVDQU16Zrmk:
1058 case X86::VMOVDQU32Z128rmk: case X86::VMOVDQA32Z128rmk:
1059 case X86::VMOVDQU32Z256rmk: case X86::VMOVDQA32Z256rmk:
1060 case X86::VMOVDQU32Zrmk: case X86::VMOVDQA32Zrmk:
1061 case X86::VMOVDQU64Z128rmk: case X86::VMOVDQA64Z128rmk:
1062 case X86::VMOVDQU64Z256rmk: case X86::VMOVDQA64Z256rmk:
1063 case X86::VMOVDQU64Zrmk: case X86::VMOVDQA64Zrmk:
1064 case X86::VMOVUPDZ128rmk: case X86::VMOVAPDZ128rmk:
1065 case X86::VMOVUPDZ256rmk: case X86::VMOVAPDZ256rmk:
1066 case X86::VMOVUPDZrmk: case X86::VMOVAPDZrmk:
1067 case X86::VMOVUPSZ128rmk: case X86::VMOVAPSZ128rmk:
1068 case X86::VMOVUPSZ256rmk: case X86::VMOVAPSZ256rmk:
1069 case X86::VMOVUPSZrmk: case X86::VMOVAPSZrmk: {
1072 default: llvm_unreachable("Unreachable!");
1073 case X86::VMOVDQU8Z128rmk: Opc = X86::VPBLENDMBZ128rmk; break;
1074 case X86::VMOVDQU8Z256rmk: Opc = X86::VPBLENDMBZ256rmk; break;
1075 case X86::VMOVDQU8Zrmk: Opc = X86::VPBLENDMBZrmk; break;
1076 case X86::VMOVDQU16Z128rmk: Opc = X86::VPBLENDMWZ128rmk; break;
1077 case X86::VMOVDQU16Z256rmk: Opc = X86::VPBLENDMWZ256rmk; break;
1078 case X86::VMOVDQU16Zrmk: Opc = X86::VPBLENDMWZrmk; break;
1079 case X86::VMOVDQU32Z128rmk: Opc = X86::VPBLENDMDZ128rmk; break;
1080 case X86::VMOVDQU32Z256rmk: Opc = X86::VPBLENDMDZ256rmk; break;
1081 case X86::VMOVDQU32Zrmk: Opc = X86::VPBLENDMDZrmk; break;
1082 case X86::VMOVDQU64Z128rmk: Opc = X86::VPBLENDMQZ128rmk; break;
1083 case X86::VMOVDQU64Z256rmk: Opc = X86::VPBLENDMQZ256rmk; break;
1084 case X86::VMOVDQU64Zrmk: Opc = X86::VPBLENDMQZrmk; break;
1085 case X86::VMOVUPDZ128rmk: Opc = X86::VBLENDMPDZ128rmk; break;
1086 case X86::VMOVUPDZ256rmk: Opc = X86::VBLENDMPDZ256rmk; break;
1087 case X86::VMOVUPDZrmk: Opc = X86::VBLENDMPDZrmk; break;
1088 case X86::VMOVUPSZ128rmk: Opc = X86::VBLENDMPSZ128rmk; break;
1089 case X86::VMOVUPSZ256rmk: Opc = X86::VBLENDMPSZ256rmk; break;
1090 case X86::VMOVUPSZrmk: Opc = X86::VBLENDMPSZrmk; break;
1091 case X86::VMOVDQA32Z128rmk: Opc = X86::VPBLENDMDZ128rmk; break;
1092 case X86::VMOVDQA32Z256rmk: Opc = X86::VPBLENDMDZ256rmk; break;
1093 case X86::VMOVDQA32Zrmk: Opc = X86::VPBLENDMDZrmk; break;
1094 case X86::VMOVDQA64Z128rmk: Opc = X86::VPBLENDMQZ128rmk; break;
1095 case X86::VMOVDQA64Z256rmk: Opc = X86::VPBLENDMQZ256rmk; break;
1096 case X86::VMOVDQA64Zrmk: Opc = X86::VPBLENDMQZrmk; break;
1097 case X86::VMOVAPDZ128rmk: Opc = X86::VBLENDMPDZ128rmk; break;
1098 case X86::VMOVAPDZ256rmk: Opc = X86::VBLENDMPDZ256rmk; break;
1099 case X86::VMOVAPDZrmk: Opc = X86::VBLENDMPDZrmk; break;
1100 case X86::VMOVAPSZ128rmk: Opc = X86::VBLENDMPSZ128rmk; break;
1101 case X86::VMOVAPSZ256rmk: Opc = X86::VBLENDMPSZ256rmk; break;
1102 case X86::VMOVAPSZrmk: Opc = X86::VBLENDMPSZrmk; break;
1105 NewMI = BuildMI(MF, MI.getDebugLoc(), get(Opc))
1107 .add(MI.getOperand(2))
1109 .add(MI.getOperand(3))
1110 .add(MI.getOperand(4))
1111 .add(MI.getOperand(5))
1112 .add(MI.getOperand(6))
1113 .add(MI.getOperand(7));
1116 case X86::VMOVDQU8Z128rrk:
1117 case X86::VMOVDQU8Z256rrk:
1118 case X86::VMOVDQU8Zrrk:
1119 case X86::VMOVDQU16Z128rrk:
1120 case X86::VMOVDQU16Z256rrk:
1121 case X86::VMOVDQU16Zrrk:
1122 case X86::VMOVDQU32Z128rrk: case X86::VMOVDQA32Z128rrk:
1123 case X86::VMOVDQU32Z256rrk: case X86::VMOVDQA32Z256rrk:
1124 case X86::VMOVDQU32Zrrk: case X86::VMOVDQA32Zrrk:
1125 case X86::VMOVDQU64Z128rrk: case X86::VMOVDQA64Z128rrk:
1126 case X86::VMOVDQU64Z256rrk: case X86::VMOVDQA64Z256rrk:
1127 case X86::VMOVDQU64Zrrk: case X86::VMOVDQA64Zrrk:
1128 case X86::VMOVUPDZ128rrk: case X86::VMOVAPDZ128rrk:
1129 case X86::VMOVUPDZ256rrk: case X86::VMOVAPDZ256rrk:
1130 case X86::VMOVUPDZrrk: case X86::VMOVAPDZrrk:
1131 case X86::VMOVUPSZ128rrk: case X86::VMOVAPSZ128rrk:
1132 case X86::VMOVUPSZ256rrk: case X86::VMOVAPSZ256rrk:
1133 case X86::VMOVUPSZrrk: case X86::VMOVAPSZrrk: {
1136 default: llvm_unreachable("Unreachable!");
1137 case X86::VMOVDQU8Z128rrk: Opc = X86::VPBLENDMBZ128rrk; break;
1138 case X86::VMOVDQU8Z256rrk: Opc = X86::VPBLENDMBZ256rrk; break;
1139 case X86::VMOVDQU8Zrrk: Opc = X86::VPBLENDMBZrrk; break;
1140 case X86::VMOVDQU16Z128rrk: Opc = X86::VPBLENDMWZ128rrk; break;
1141 case X86::VMOVDQU16Z256rrk: Opc = X86::VPBLENDMWZ256rrk; break;
1142 case X86::VMOVDQU16Zrrk: Opc = X86::VPBLENDMWZrrk; break;
1143 case X86::VMOVDQU32Z128rrk: Opc = X86::VPBLENDMDZ128rrk; break;
1144 case X86::VMOVDQU32Z256rrk: Opc = X86::VPBLENDMDZ256rrk; break;
1145 case X86::VMOVDQU32Zrrk: Opc = X86::VPBLENDMDZrrk; break;
1146 case X86::VMOVDQU64Z128rrk: Opc = X86::VPBLENDMQZ128rrk; break;
1147 case X86::VMOVDQU64Z256rrk: Opc = X86::VPBLENDMQZ256rrk; break;
1148 case X86::VMOVDQU64Zrrk: Opc = X86::VPBLENDMQZrrk; break;
1149 case X86::VMOVUPDZ128rrk: Opc = X86::VBLENDMPDZ128rrk; break;
1150 case X86::VMOVUPDZ256rrk: Opc = X86::VBLENDMPDZ256rrk; break;
1151 case X86::VMOVUPDZrrk: Opc = X86::VBLENDMPDZrrk; break;
1152 case X86::VMOVUPSZ128rrk: Opc = X86::VBLENDMPSZ128rrk; break;
1153 case X86::VMOVUPSZ256rrk: Opc = X86::VBLENDMPSZ256rrk; break;
1154 case X86::VMOVUPSZrrk: Opc = X86::VBLENDMPSZrrk; break;
1155 case X86::VMOVDQA32Z128rrk: Opc = X86::VPBLENDMDZ128rrk; break;
1156 case X86::VMOVDQA32Z256rrk: Opc = X86::VPBLENDMDZ256rrk; break;
1157 case X86::VMOVDQA32Zrrk: Opc = X86::VPBLENDMDZrrk; break;
1158 case X86::VMOVDQA64Z128rrk: Opc = X86::VPBLENDMQZ128rrk; break;
1159 case X86::VMOVDQA64Z256rrk: Opc = X86::VPBLENDMQZ256rrk; break;
1160 case X86::VMOVDQA64Zrrk: Opc = X86::VPBLENDMQZrrk; break;
1161 case X86::VMOVAPDZ128rrk: Opc = X86::VBLENDMPDZ128rrk; break;
1162 case X86::VMOVAPDZ256rrk: Opc = X86::VBLENDMPDZ256rrk; break;
1163 case X86::VMOVAPDZrrk: Opc = X86::VBLENDMPDZrrk; break;
1164 case X86::VMOVAPSZ128rrk: Opc = X86::VBLENDMPSZ128rrk; break;
1165 case X86::VMOVAPSZ256rrk: Opc = X86::VBLENDMPSZ256rrk; break;
1166 case X86::VMOVAPSZrrk: Opc = X86::VBLENDMPSZrrk; break;
1169 NewMI = BuildMI(MF, MI.getDebugLoc(), get(Opc))
1171 .add(MI.getOperand(2))
1173 .add(MI.getOperand(3));
1178 if (!NewMI) return nullptr;
1180 if (LV) { // Update live variables
1182 LV->replaceKillInstruction(Src.getReg(), MI, *NewMI);
1184 LV->replaceKillInstruction(Dest.getReg(), MI, *NewMI);
1187 MFI->insert(MI.getIterator(), NewMI); // Insert the new inst
1191 /// This determines which of three possible cases of a three source commute
1192 /// the source indexes correspond to taking into account any mask operands.
1193 /// All prevents commuting a passthru operand. Returns -1 if the commute isn't
1195 /// Case 0 - Possible to commute the first and second operands.
1196 /// Case 1 - Possible to commute the first and third operands.
1197 /// Case 2 - Possible to commute the second and third operands.
1198 static unsigned getThreeSrcCommuteCase(uint64_t TSFlags, unsigned SrcOpIdx1,
1199 unsigned SrcOpIdx2) {
1200 // Put the lowest index to SrcOpIdx1 to simplify the checks below.
1201 if (SrcOpIdx1 > SrcOpIdx2)
1202 std::swap(SrcOpIdx1, SrcOpIdx2);
1204 unsigned Op1 = 1, Op2 = 2, Op3 = 3;
1205 if (X86II::isKMasked(TSFlags)) {
1210 if (SrcOpIdx1 == Op1 && SrcOpIdx2 == Op2)
1212 if (SrcOpIdx1 == Op1 && SrcOpIdx2 == Op3)
1214 if (SrcOpIdx1 == Op2 && SrcOpIdx2 == Op3)
1216 llvm_unreachable("Unknown three src commute case.");
1219 unsigned X86InstrInfo::getFMA3OpcodeToCommuteOperands(
1220 const MachineInstr &MI, unsigned SrcOpIdx1, unsigned SrcOpIdx2,
1221 const X86InstrFMA3Group &FMA3Group) const {
1223 unsigned Opc = MI.getOpcode();
1225 // TODO: Commuting the 1st operand of FMA*_Int requires some additional
1226 // analysis. The commute optimization is legal only if all users of FMA*_Int
1227 // use only the lowest element of the FMA*_Int instruction. Such analysis are
1228 // not implemented yet. So, just return 0 in that case.
1229 // When such analysis are available this place will be the right place for
1231 assert(!(FMA3Group.isIntrinsic() && (SrcOpIdx1 == 1 || SrcOpIdx2 == 1)) &&
1232 "Intrinsic instructions can't commute operand 1");
1234 // Determine which case this commute is or if it can't be done.
1235 unsigned Case = getThreeSrcCommuteCase(MI.getDesc().TSFlags, SrcOpIdx1,
1237 assert(Case < 3 && "Unexpected case number!");
1239 // Define the FMA forms mapping array that helps to map input FMA form
1240 // to output FMA form to preserve the operation semantics after
1241 // commuting the operands.
1242 const unsigned Form132Index = 0;
1243 const unsigned Form213Index = 1;
1244 const unsigned Form231Index = 2;
1245 static const unsigned FormMapping[][3] = {
1246 // 0: SrcOpIdx1 == 1 && SrcOpIdx2 == 2;
1247 // FMA132 A, C, b; ==> FMA231 C, A, b;
1248 // FMA213 B, A, c; ==> FMA213 A, B, c;
1249 // FMA231 C, A, b; ==> FMA132 A, C, b;
1250 { Form231Index, Form213Index, Form132Index },
1251 // 1: SrcOpIdx1 == 1 && SrcOpIdx2 == 3;
1252 // FMA132 A, c, B; ==> FMA132 B, c, A;
1253 // FMA213 B, a, C; ==> FMA231 C, a, B;
1254 // FMA231 C, a, B; ==> FMA213 B, a, C;
1255 { Form132Index, Form231Index, Form213Index },
1256 // 2: SrcOpIdx1 == 2 && SrcOpIdx2 == 3;
1257 // FMA132 a, C, B; ==> FMA213 a, B, C;
1258 // FMA213 b, A, C; ==> FMA132 b, C, A;
1259 // FMA231 c, A, B; ==> FMA231 c, B, A;
1260 { Form213Index, Form132Index, Form231Index }
1263 unsigned FMAForms[3];
1264 FMAForms[0] = FMA3Group.get132Opcode();
1265 FMAForms[1] = FMA3Group.get213Opcode();
1266 FMAForms[2] = FMA3Group.get231Opcode();
1268 for (FormIndex = 0; FormIndex < 3; FormIndex++)
1269 if (Opc == FMAForms[FormIndex])
1272 // Everything is ready, just adjust the FMA opcode and return it.
1273 FormIndex = FormMapping[Case][FormIndex];
1274 return FMAForms[FormIndex];
1277 static void commuteVPTERNLOG(MachineInstr &MI, unsigned SrcOpIdx1,
1278 unsigned SrcOpIdx2) {
1279 // Determine which case this commute is or if it can't be done.
1280 unsigned Case = getThreeSrcCommuteCase(MI.getDesc().TSFlags, SrcOpIdx1,
1282 assert(Case < 3 && "Unexpected case value!");
1284 // For each case we need to swap two pairs of bits in the final immediate.
1285 static const uint8_t SwapMasks[3][4] = {
1286 { 0x04, 0x10, 0x08, 0x20 }, // Swap bits 2/4 and 3/5.
1287 { 0x02, 0x10, 0x08, 0x40 }, // Swap bits 1/4 and 3/6.
1288 { 0x02, 0x04, 0x20, 0x40 }, // Swap bits 1/2 and 5/6.
1291 uint8_t Imm = MI.getOperand(MI.getNumOperands()-1).getImm();
1292 // Clear out the bits we are swapping.
1293 uint8_t NewImm = Imm & ~(SwapMasks[Case][0] | SwapMasks[Case][1] |
1294 SwapMasks[Case][2] | SwapMasks[Case][3]);
1295 // If the immediate had a bit of the pair set, then set the opposite bit.
1296 if (Imm & SwapMasks[Case][0]) NewImm |= SwapMasks[Case][1];
1297 if (Imm & SwapMasks[Case][1]) NewImm |= SwapMasks[Case][0];
1298 if (Imm & SwapMasks[Case][2]) NewImm |= SwapMasks[Case][3];
1299 if (Imm & SwapMasks[Case][3]) NewImm |= SwapMasks[Case][2];
1300 MI.getOperand(MI.getNumOperands()-1).setImm(NewImm);
1303 // Returns true if this is a VPERMI2 or VPERMT2 instruction that can be
1305 static bool isCommutableVPERMV3Instruction(unsigned Opcode) {
1306 #define VPERM_CASES(Suffix) \
1307 case X86::VPERMI2##Suffix##128rr: case X86::VPERMT2##Suffix##128rr: \
1308 case X86::VPERMI2##Suffix##256rr: case X86::VPERMT2##Suffix##256rr: \
1309 case X86::VPERMI2##Suffix##rr: case X86::VPERMT2##Suffix##rr: \
1310 case X86::VPERMI2##Suffix##128rm: case X86::VPERMT2##Suffix##128rm: \
1311 case X86::VPERMI2##Suffix##256rm: case X86::VPERMT2##Suffix##256rm: \
1312 case X86::VPERMI2##Suffix##rm: case X86::VPERMT2##Suffix##rm: \
1313 case X86::VPERMI2##Suffix##128rrkz: case X86::VPERMT2##Suffix##128rrkz: \
1314 case X86::VPERMI2##Suffix##256rrkz: case X86::VPERMT2##Suffix##256rrkz: \
1315 case X86::VPERMI2##Suffix##rrkz: case X86::VPERMT2##Suffix##rrkz: \
1316 case X86::VPERMI2##Suffix##128rmkz: case X86::VPERMT2##Suffix##128rmkz: \
1317 case X86::VPERMI2##Suffix##256rmkz: case X86::VPERMT2##Suffix##256rmkz: \
1318 case X86::VPERMI2##Suffix##rmkz: case X86::VPERMT2##Suffix##rmkz:
1320 #define VPERM_CASES_BROADCAST(Suffix) \
1321 VPERM_CASES(Suffix) \
1322 case X86::VPERMI2##Suffix##128rmb: case X86::VPERMT2##Suffix##128rmb: \
1323 case X86::VPERMI2##Suffix##256rmb: case X86::VPERMT2##Suffix##256rmb: \
1324 case X86::VPERMI2##Suffix##rmb: case X86::VPERMT2##Suffix##rmb: \
1325 case X86::VPERMI2##Suffix##128rmbkz: case X86::VPERMT2##Suffix##128rmbkz: \
1326 case X86::VPERMI2##Suffix##256rmbkz: case X86::VPERMT2##Suffix##256rmbkz: \
1327 case X86::VPERMI2##Suffix##rmbkz: case X86::VPERMT2##Suffix##rmbkz:
1330 default: return false;
1332 VPERM_CASES_BROADCAST(D)
1333 VPERM_CASES_BROADCAST(PD)
1334 VPERM_CASES_BROADCAST(PS)
1335 VPERM_CASES_BROADCAST(Q)
1339 #undef VPERM_CASES_BROADCAST
1343 // Returns commuted opcode for VPERMI2 and VPERMT2 instructions by switching
1344 // from the I opcode to the T opcode and vice versa.
1345 static unsigned getCommutedVPERMV3Opcode(unsigned Opcode) {
1346 #define VPERM_CASES(Orig, New) \
1347 case X86::Orig##128rr: return X86::New##128rr; \
1348 case X86::Orig##128rrkz: return X86::New##128rrkz; \
1349 case X86::Orig##128rm: return X86::New##128rm; \
1350 case X86::Orig##128rmkz: return X86::New##128rmkz; \
1351 case X86::Orig##256rr: return X86::New##256rr; \
1352 case X86::Orig##256rrkz: return X86::New##256rrkz; \
1353 case X86::Orig##256rm: return X86::New##256rm; \
1354 case X86::Orig##256rmkz: return X86::New##256rmkz; \
1355 case X86::Orig##rr: return X86::New##rr; \
1356 case X86::Orig##rrkz: return X86::New##rrkz; \
1357 case X86::Orig##rm: return X86::New##rm; \
1358 case X86::Orig##rmkz: return X86::New##rmkz;
1360 #define VPERM_CASES_BROADCAST(Orig, New) \
1361 VPERM_CASES(Orig, New) \
1362 case X86::Orig##128rmb: return X86::New##128rmb; \
1363 case X86::Orig##128rmbkz: return X86::New##128rmbkz; \
1364 case X86::Orig##256rmb: return X86::New##256rmb; \
1365 case X86::Orig##256rmbkz: return X86::New##256rmbkz; \
1366 case X86::Orig##rmb: return X86::New##rmb; \
1367 case X86::Orig##rmbkz: return X86::New##rmbkz;
1370 VPERM_CASES(VPERMI2B, VPERMT2B)
1371 VPERM_CASES_BROADCAST(VPERMI2D, VPERMT2D)
1372 VPERM_CASES_BROADCAST(VPERMI2PD, VPERMT2PD)
1373 VPERM_CASES_BROADCAST(VPERMI2PS, VPERMT2PS)
1374 VPERM_CASES_BROADCAST(VPERMI2Q, VPERMT2Q)
1375 VPERM_CASES(VPERMI2W, VPERMT2W)
1376 VPERM_CASES(VPERMT2B, VPERMI2B)
1377 VPERM_CASES_BROADCAST(VPERMT2D, VPERMI2D)
1378 VPERM_CASES_BROADCAST(VPERMT2PD, VPERMI2PD)
1379 VPERM_CASES_BROADCAST(VPERMT2PS, VPERMI2PS)
1380 VPERM_CASES_BROADCAST(VPERMT2Q, VPERMI2Q)
1381 VPERM_CASES(VPERMT2W, VPERMI2W)
1384 llvm_unreachable("Unreachable!");
1385 #undef VPERM_CASES_BROADCAST
1389 MachineInstr *X86InstrInfo::commuteInstructionImpl(MachineInstr &MI, bool NewMI,
1391 unsigned OpIdx2) const {
1392 auto cloneIfNew = [NewMI](MachineInstr &MI) -> MachineInstr & {
1394 return *MI.getParent()->getParent()->CloneMachineInstr(&MI);
1398 switch (MI.getOpcode()) {
1399 case X86::SHRD16rri8: // A = SHRD16rri8 B, C, I -> A = SHLD16rri8 C, B, (16-I)
1400 case X86::SHLD16rri8: // A = SHLD16rri8 B, C, I -> A = SHRD16rri8 C, B, (16-I)
1401 case X86::SHRD32rri8: // A = SHRD32rri8 B, C, I -> A = SHLD32rri8 C, B, (32-I)
1402 case X86::SHLD32rri8: // A = SHLD32rri8 B, C, I -> A = SHRD32rri8 C, B, (32-I)
1403 case X86::SHRD64rri8: // A = SHRD64rri8 B, C, I -> A = SHLD64rri8 C, B, (64-I)
1404 case X86::SHLD64rri8:{// A = SHLD64rri8 B, C, I -> A = SHRD64rri8 C, B, (64-I)
1407 switch (MI.getOpcode()) {
1408 default: llvm_unreachable("Unreachable!");
1409 case X86::SHRD16rri8: Size = 16; Opc = X86::SHLD16rri8; break;
1410 case X86::SHLD16rri8: Size = 16; Opc = X86::SHRD16rri8; break;
1411 case X86::SHRD32rri8: Size = 32; Opc = X86::SHLD32rri8; break;
1412 case X86::SHLD32rri8: Size = 32; Opc = X86::SHRD32rri8; break;
1413 case X86::SHRD64rri8: Size = 64; Opc = X86::SHLD64rri8; break;
1414 case X86::SHLD64rri8: Size = 64; Opc = X86::SHRD64rri8; break;
1416 unsigned Amt = MI.getOperand(3).getImm();
1417 auto &WorkingMI = cloneIfNew(MI);
1418 WorkingMI.setDesc(get(Opc));
1419 WorkingMI.getOperand(3).setImm(Size - Amt);
1420 return TargetInstrInfo::commuteInstructionImpl(WorkingMI, /*NewMI=*/false,
1424 case X86::PFSUBRrr: {
1425 // PFSUB x, y: x = x - y
1426 // PFSUBR x, y: x = y - x
1428 (X86::PFSUBRrr == MI.getOpcode() ? X86::PFSUBrr : X86::PFSUBRrr);
1429 auto &WorkingMI = cloneIfNew(MI);
1430 WorkingMI.setDesc(get(Opc));
1431 return TargetInstrInfo::commuteInstructionImpl(WorkingMI, /*NewMI=*/false,
1434 case X86::BLENDPDrri:
1435 case X86::BLENDPSrri:
1436 case X86::VBLENDPDrri:
1437 case X86::VBLENDPSrri:
1438 // If we're optimizing for size, try to use MOVSD/MOVSS.
1439 if (MI.getParent()->getParent()->getFunction().optForSize()) {
1441 switch (MI.getOpcode()) {
1442 default: llvm_unreachable("Unreachable!");
1443 case X86::BLENDPDrri: Opc = X86::MOVSDrr; Mask = 0x03; break;
1444 case X86::BLENDPSrri: Opc = X86::MOVSSrr; Mask = 0x0F; break;
1445 case X86::VBLENDPDrri: Opc = X86::VMOVSDrr; Mask = 0x03; break;
1446 case X86::VBLENDPSrri: Opc = X86::VMOVSSrr; Mask = 0x0F; break;
1448 if ((MI.getOperand(3).getImm() ^ Mask) == 1) {
1449 auto &WorkingMI = cloneIfNew(MI);
1450 WorkingMI.setDesc(get(Opc));
1451 WorkingMI.RemoveOperand(3);
1452 return TargetInstrInfo::commuteInstructionImpl(WorkingMI,
1458 case X86::PBLENDWrri:
1459 case X86::VBLENDPDYrri:
1460 case X86::VBLENDPSYrri:
1461 case X86::VPBLENDDrri:
1462 case X86::VPBLENDWrri:
1463 case X86::VPBLENDDYrri:
1464 case X86::VPBLENDWYrri:{
1466 switch (MI.getOpcode()) {
1467 default: llvm_unreachable("Unreachable!");
1468 case X86::BLENDPDrri: Mask = (int8_t)0x03; break;
1469 case X86::BLENDPSrri: Mask = (int8_t)0x0F; break;
1470 case X86::PBLENDWrri: Mask = (int8_t)0xFF; break;
1471 case X86::VBLENDPDrri: Mask = (int8_t)0x03; break;
1472 case X86::VBLENDPSrri: Mask = (int8_t)0x0F; break;
1473 case X86::VBLENDPDYrri: Mask = (int8_t)0x0F; break;
1474 case X86::VBLENDPSYrri: Mask = (int8_t)0xFF; break;
1475 case X86::VPBLENDDrri: Mask = (int8_t)0x0F; break;
1476 case X86::VPBLENDWrri: Mask = (int8_t)0xFF; break;
1477 case X86::VPBLENDDYrri: Mask = (int8_t)0xFF; break;
1478 case X86::VPBLENDWYrri: Mask = (int8_t)0xFF; break;
1480 // Only the least significant bits of Imm are used.
1481 // Using int8_t to ensure it will be sign extended to the int64_t that
1482 // setImm takes in order to match isel behavior.
1483 int8_t Imm = MI.getOperand(3).getImm() & Mask;
1484 auto &WorkingMI = cloneIfNew(MI);
1485 WorkingMI.getOperand(3).setImm(Mask ^ Imm);
1486 return TargetInstrInfo::commuteInstructionImpl(WorkingMI, /*NewMI=*/false,
1489 case X86::INSERTPSrr:
1490 case X86::VINSERTPSrr:
1491 case X86::VINSERTPSZrr: {
1492 unsigned Imm = MI.getOperand(MI.getNumOperands() - 1).getImm();
1493 unsigned ZMask = Imm & 15;
1494 unsigned DstIdx = (Imm >> 4) & 3;
1495 unsigned SrcIdx = (Imm >> 6) & 3;
1497 // We can commute insertps if we zero 2 of the elements, the insertion is
1498 // "inline" and we don't override the insertion with a zero.
1499 if (DstIdx == SrcIdx && (ZMask & (1 << DstIdx)) == 0 &&
1500 countPopulation(ZMask) == 2) {
1501 unsigned AltIdx = findFirstSet((ZMask | (1 << DstIdx)) ^ 15);
1502 assert(AltIdx < 4 && "Illegal insertion index");
1503 unsigned AltImm = (AltIdx << 6) | (AltIdx << 4) | ZMask;
1504 auto &WorkingMI = cloneIfNew(MI);
1505 WorkingMI.getOperand(MI.getNumOperands() - 1).setImm(AltImm);
1506 return TargetInstrInfo::commuteInstructionImpl(WorkingMI, /*NewMI=*/false,
1514 case X86::VMOVSSrr:{
1515 // On SSE41 or later we can commute a MOVSS/MOVSD to a BLENDPS/BLENDPD.
1516 assert(Subtarget.hasSSE41() && "Commuting MOVSD/MOVSS requires SSE41!");
1519 switch (MI.getOpcode()) {
1520 default: llvm_unreachable("Unreachable!");
1521 case X86::MOVSDrr: Opc = X86::BLENDPDrri; Mask = 0x02; break;
1522 case X86::MOVSSrr: Opc = X86::BLENDPSrri; Mask = 0x0E; break;
1523 case X86::VMOVSDrr: Opc = X86::VBLENDPDrri; Mask = 0x02; break;
1524 case X86::VMOVSSrr: Opc = X86::VBLENDPSrri; Mask = 0x0E; break;
1527 auto &WorkingMI = cloneIfNew(MI);
1528 WorkingMI.setDesc(get(Opc));
1529 WorkingMI.addOperand(MachineOperand::CreateImm(Mask));
1530 return TargetInstrInfo::commuteInstructionImpl(WorkingMI, /*NewMI=*/false,
1533 case X86::PCLMULQDQrr:
1534 case X86::VPCLMULQDQrr:
1535 case X86::VPCLMULQDQYrr:
1536 case X86::VPCLMULQDQZrr:
1537 case X86::VPCLMULQDQZ128rr:
1538 case X86::VPCLMULQDQZ256rr: {
1539 // SRC1 64bits = Imm[0] ? SRC1[127:64] : SRC1[63:0]
1540 // SRC2 64bits = Imm[4] ? SRC2[127:64] : SRC2[63:0]
1541 unsigned Imm = MI.getOperand(3).getImm();
1542 unsigned Src1Hi = Imm & 0x01;
1543 unsigned Src2Hi = Imm & 0x10;
1544 auto &WorkingMI = cloneIfNew(MI);
1545 WorkingMI.getOperand(3).setImm((Src1Hi << 4) | (Src2Hi >> 4));
1546 return TargetInstrInfo::commuteInstructionImpl(WorkingMI, /*NewMI=*/false,
1549 case X86::VPCMPBZ128rri: case X86::VPCMPUBZ128rri:
1550 case X86::VPCMPBZ256rri: case X86::VPCMPUBZ256rri:
1551 case X86::VPCMPBZrri: case X86::VPCMPUBZrri:
1552 case X86::VPCMPDZ128rri: case X86::VPCMPUDZ128rri:
1553 case X86::VPCMPDZ256rri: case X86::VPCMPUDZ256rri:
1554 case X86::VPCMPDZrri: case X86::VPCMPUDZrri:
1555 case X86::VPCMPQZ128rri: case X86::VPCMPUQZ128rri:
1556 case X86::VPCMPQZ256rri: case X86::VPCMPUQZ256rri:
1557 case X86::VPCMPQZrri: case X86::VPCMPUQZrri:
1558 case X86::VPCMPWZ128rri: case X86::VPCMPUWZ128rri:
1559 case X86::VPCMPWZ256rri: case X86::VPCMPUWZ256rri:
1560 case X86::VPCMPWZrri: case X86::VPCMPUWZrri:
1561 case X86::VPCMPBZ128rrik: case X86::VPCMPUBZ128rrik:
1562 case X86::VPCMPBZ256rrik: case X86::VPCMPUBZ256rrik:
1563 case X86::VPCMPBZrrik: case X86::VPCMPUBZrrik:
1564 case X86::VPCMPDZ128rrik: case X86::VPCMPUDZ128rrik:
1565 case X86::VPCMPDZ256rrik: case X86::VPCMPUDZ256rrik:
1566 case X86::VPCMPDZrrik: case X86::VPCMPUDZrrik:
1567 case X86::VPCMPQZ128rrik: case X86::VPCMPUQZ128rrik:
1568 case X86::VPCMPQZ256rrik: case X86::VPCMPUQZ256rrik:
1569 case X86::VPCMPQZrrik: case X86::VPCMPUQZrrik:
1570 case X86::VPCMPWZ128rrik: case X86::VPCMPUWZ128rrik:
1571 case X86::VPCMPWZ256rrik: case X86::VPCMPUWZ256rrik:
1572 case X86::VPCMPWZrrik: case X86::VPCMPUWZrrik: {
1573 // Flip comparison mode immediate (if necessary).
1574 unsigned Imm = MI.getOperand(MI.getNumOperands() - 1).getImm() & 0x7;
1575 Imm = X86::getSwappedVPCMPImm(Imm);
1576 auto &WorkingMI = cloneIfNew(MI);
1577 WorkingMI.getOperand(MI.getNumOperands() - 1).setImm(Imm);
1578 return TargetInstrInfo::commuteInstructionImpl(WorkingMI, /*NewMI=*/false,
1581 case X86::VPCOMBri: case X86::VPCOMUBri:
1582 case X86::VPCOMDri: case X86::VPCOMUDri:
1583 case X86::VPCOMQri: case X86::VPCOMUQri:
1584 case X86::VPCOMWri: case X86::VPCOMUWri: {
1585 // Flip comparison mode immediate (if necessary).
1586 unsigned Imm = MI.getOperand(3).getImm() & 0x7;
1587 Imm = X86::getSwappedVPCOMImm(Imm);
1588 auto &WorkingMI = cloneIfNew(MI);
1589 WorkingMI.getOperand(3).setImm(Imm);
1590 return TargetInstrInfo::commuteInstructionImpl(WorkingMI, /*NewMI=*/false,
1593 case X86::VPERM2F128rr:
1594 case X86::VPERM2I128rr: {
1595 // Flip permute source immediate.
1596 // Imm & 0x02: lo = if set, select Op1.lo/hi else Op0.lo/hi.
1597 // Imm & 0x20: hi = if set, select Op1.lo/hi else Op0.lo/hi.
1598 int8_t Imm = MI.getOperand(3).getImm() & 0xFF;
1599 auto &WorkingMI = cloneIfNew(MI);
1600 WorkingMI.getOperand(3).setImm(Imm ^ 0x22);
1601 return TargetInstrInfo::commuteInstructionImpl(WorkingMI, /*NewMI=*/false,
1604 case X86::MOVHLPSrr:
1605 case X86::UNPCKHPDrr:
1606 case X86::VMOVHLPSrr:
1607 case X86::VUNPCKHPDrr:
1608 case X86::VMOVHLPSZrr:
1609 case X86::VUNPCKHPDZ128rr: {
1610 assert(Subtarget.hasSSE2() && "Commuting MOVHLP/UNPCKHPD requires SSE2!");
1612 unsigned Opc = MI.getOpcode();
1614 default: llvm_unreachable("Unreachable!");
1615 case X86::MOVHLPSrr: Opc = X86::UNPCKHPDrr; break;
1616 case X86::UNPCKHPDrr: Opc = X86::MOVHLPSrr; break;
1617 case X86::VMOVHLPSrr: Opc = X86::VUNPCKHPDrr; break;
1618 case X86::VUNPCKHPDrr: Opc = X86::VMOVHLPSrr; break;
1619 case X86::VMOVHLPSZrr: Opc = X86::VUNPCKHPDZ128rr; break;
1620 case X86::VUNPCKHPDZ128rr: Opc = X86::VMOVHLPSZrr; break;
1622 auto &WorkingMI = cloneIfNew(MI);
1623 WorkingMI.setDesc(get(Opc));
1624 return TargetInstrInfo::commuteInstructionImpl(WorkingMI, /*NewMI=*/false,
1627 case X86::CMOVB16rr: case X86::CMOVB32rr: case X86::CMOVB64rr:
1628 case X86::CMOVAE16rr: case X86::CMOVAE32rr: case X86::CMOVAE64rr:
1629 case X86::CMOVE16rr: case X86::CMOVE32rr: case X86::CMOVE64rr:
1630 case X86::CMOVNE16rr: case X86::CMOVNE32rr: case X86::CMOVNE64rr:
1631 case X86::CMOVBE16rr: case X86::CMOVBE32rr: case X86::CMOVBE64rr:
1632 case X86::CMOVA16rr: case X86::CMOVA32rr: case X86::CMOVA64rr:
1633 case X86::CMOVL16rr: case X86::CMOVL32rr: case X86::CMOVL64rr:
1634 case X86::CMOVGE16rr: case X86::CMOVGE32rr: case X86::CMOVGE64rr:
1635 case X86::CMOVLE16rr: case X86::CMOVLE32rr: case X86::CMOVLE64rr:
1636 case X86::CMOVG16rr: case X86::CMOVG32rr: case X86::CMOVG64rr:
1637 case X86::CMOVS16rr: case X86::CMOVS32rr: case X86::CMOVS64rr:
1638 case X86::CMOVNS16rr: case X86::CMOVNS32rr: case X86::CMOVNS64rr:
1639 case X86::CMOVP16rr: case X86::CMOVP32rr: case X86::CMOVP64rr:
1640 case X86::CMOVNP16rr: case X86::CMOVNP32rr: case X86::CMOVNP64rr:
1641 case X86::CMOVO16rr: case X86::CMOVO32rr: case X86::CMOVO64rr:
1642 case X86::CMOVNO16rr: case X86::CMOVNO32rr: case X86::CMOVNO64rr: {
1644 switch (MI.getOpcode()) {
1645 default: llvm_unreachable("Unreachable!");
1646 case X86::CMOVB16rr: Opc = X86::CMOVAE16rr; break;
1647 case X86::CMOVB32rr: Opc = X86::CMOVAE32rr; break;
1648 case X86::CMOVB64rr: Opc = X86::CMOVAE64rr; break;
1649 case X86::CMOVAE16rr: Opc = X86::CMOVB16rr; break;
1650 case X86::CMOVAE32rr: Opc = X86::CMOVB32rr; break;
1651 case X86::CMOVAE64rr: Opc = X86::CMOVB64rr; break;
1652 case X86::CMOVE16rr: Opc = X86::CMOVNE16rr; break;
1653 case X86::CMOVE32rr: Opc = X86::CMOVNE32rr; break;
1654 case X86::CMOVE64rr: Opc = X86::CMOVNE64rr; break;
1655 case X86::CMOVNE16rr: Opc = X86::CMOVE16rr; break;
1656 case X86::CMOVNE32rr: Opc = X86::CMOVE32rr; break;
1657 case X86::CMOVNE64rr: Opc = X86::CMOVE64rr; break;
1658 case X86::CMOVBE16rr: Opc = X86::CMOVA16rr; break;
1659 case X86::CMOVBE32rr: Opc = X86::CMOVA32rr; break;
1660 case X86::CMOVBE64rr: Opc = X86::CMOVA64rr; break;
1661 case X86::CMOVA16rr: Opc = X86::CMOVBE16rr; break;
1662 case X86::CMOVA32rr: Opc = X86::CMOVBE32rr; break;
1663 case X86::CMOVA64rr: Opc = X86::CMOVBE64rr; break;
1664 case X86::CMOVL16rr: Opc = X86::CMOVGE16rr; break;
1665 case X86::CMOVL32rr: Opc = X86::CMOVGE32rr; break;
1666 case X86::CMOVL64rr: Opc = X86::CMOVGE64rr; break;
1667 case X86::CMOVGE16rr: Opc = X86::CMOVL16rr; break;
1668 case X86::CMOVGE32rr: Opc = X86::CMOVL32rr; break;
1669 case X86::CMOVGE64rr: Opc = X86::CMOVL64rr; break;
1670 case X86::CMOVLE16rr: Opc = X86::CMOVG16rr; break;
1671 case X86::CMOVLE32rr: Opc = X86::CMOVG32rr; break;
1672 case X86::CMOVLE64rr: Opc = X86::CMOVG64rr; break;
1673 case X86::CMOVG16rr: Opc = X86::CMOVLE16rr; break;
1674 case X86::CMOVG32rr: Opc = X86::CMOVLE32rr; break;
1675 case X86::CMOVG64rr: Opc = X86::CMOVLE64rr; break;
1676 case X86::CMOVS16rr: Opc = X86::CMOVNS16rr; break;
1677 case X86::CMOVS32rr: Opc = X86::CMOVNS32rr; break;
1678 case X86::CMOVS64rr: Opc = X86::CMOVNS64rr; break;
1679 case X86::CMOVNS16rr: Opc = X86::CMOVS16rr; break;
1680 case X86::CMOVNS32rr: Opc = X86::CMOVS32rr; break;
1681 case X86::CMOVNS64rr: Opc = X86::CMOVS64rr; break;
1682 case X86::CMOVP16rr: Opc = X86::CMOVNP16rr; break;
1683 case X86::CMOVP32rr: Opc = X86::CMOVNP32rr; break;
1684 case X86::CMOVP64rr: Opc = X86::CMOVNP64rr; break;
1685 case X86::CMOVNP16rr: Opc = X86::CMOVP16rr; break;
1686 case X86::CMOVNP32rr: Opc = X86::CMOVP32rr; break;
1687 case X86::CMOVNP64rr: Opc = X86::CMOVP64rr; break;
1688 case X86::CMOVO16rr: Opc = X86::CMOVNO16rr; break;
1689 case X86::CMOVO32rr: Opc = X86::CMOVNO32rr; break;
1690 case X86::CMOVO64rr: Opc = X86::CMOVNO64rr; break;
1691 case X86::CMOVNO16rr: Opc = X86::CMOVO16rr; break;
1692 case X86::CMOVNO32rr: Opc = X86::CMOVO32rr; break;
1693 case X86::CMOVNO64rr: Opc = X86::CMOVO64rr; break;
1695 auto &WorkingMI = cloneIfNew(MI);
1696 WorkingMI.setDesc(get(Opc));
1697 return TargetInstrInfo::commuteInstructionImpl(WorkingMI, /*NewMI=*/false,
1700 case X86::VPTERNLOGDZrri: case X86::VPTERNLOGDZrmi:
1701 case X86::VPTERNLOGDZ128rri: case X86::VPTERNLOGDZ128rmi:
1702 case X86::VPTERNLOGDZ256rri: case X86::VPTERNLOGDZ256rmi:
1703 case X86::VPTERNLOGQZrri: case X86::VPTERNLOGQZrmi:
1704 case X86::VPTERNLOGQZ128rri: case X86::VPTERNLOGQZ128rmi:
1705 case X86::VPTERNLOGQZ256rri: case X86::VPTERNLOGQZ256rmi:
1706 case X86::VPTERNLOGDZrrik:
1707 case X86::VPTERNLOGDZ128rrik:
1708 case X86::VPTERNLOGDZ256rrik:
1709 case X86::VPTERNLOGQZrrik:
1710 case X86::VPTERNLOGQZ128rrik:
1711 case X86::VPTERNLOGQZ256rrik:
1712 case X86::VPTERNLOGDZrrikz: case X86::VPTERNLOGDZrmikz:
1713 case X86::VPTERNLOGDZ128rrikz: case X86::VPTERNLOGDZ128rmikz:
1714 case X86::VPTERNLOGDZ256rrikz: case X86::VPTERNLOGDZ256rmikz:
1715 case X86::VPTERNLOGQZrrikz: case X86::VPTERNLOGQZrmikz:
1716 case X86::VPTERNLOGQZ128rrikz: case X86::VPTERNLOGQZ128rmikz:
1717 case X86::VPTERNLOGQZ256rrikz: case X86::VPTERNLOGQZ256rmikz:
1718 case X86::VPTERNLOGDZ128rmbi:
1719 case X86::VPTERNLOGDZ256rmbi:
1720 case X86::VPTERNLOGDZrmbi:
1721 case X86::VPTERNLOGQZ128rmbi:
1722 case X86::VPTERNLOGQZ256rmbi:
1723 case X86::VPTERNLOGQZrmbi:
1724 case X86::VPTERNLOGDZ128rmbikz:
1725 case X86::VPTERNLOGDZ256rmbikz:
1726 case X86::VPTERNLOGDZrmbikz:
1727 case X86::VPTERNLOGQZ128rmbikz:
1728 case X86::VPTERNLOGQZ256rmbikz:
1729 case X86::VPTERNLOGQZrmbikz: {
1730 auto &WorkingMI = cloneIfNew(MI);
1731 commuteVPTERNLOG(WorkingMI, OpIdx1, OpIdx2);
1732 return TargetInstrInfo::commuteInstructionImpl(WorkingMI, /*NewMI=*/false,
1736 if (isCommutableVPERMV3Instruction(MI.getOpcode())) {
1737 unsigned Opc = getCommutedVPERMV3Opcode(MI.getOpcode());
1738 auto &WorkingMI = cloneIfNew(MI);
1739 WorkingMI.setDesc(get(Opc));
1740 return TargetInstrInfo::commuteInstructionImpl(WorkingMI, /*NewMI=*/false,
1744 const X86InstrFMA3Group *FMA3Group = getFMA3Group(MI.getOpcode(),
1745 MI.getDesc().TSFlags);
1748 getFMA3OpcodeToCommuteOperands(MI, OpIdx1, OpIdx2, *FMA3Group);
1749 auto &WorkingMI = cloneIfNew(MI);
1750 WorkingMI.setDesc(get(Opc));
1751 return TargetInstrInfo::commuteInstructionImpl(WorkingMI, /*NewMI=*/false,
1755 return TargetInstrInfo::commuteInstructionImpl(MI, NewMI, OpIdx1, OpIdx2);
1761 X86InstrInfo::findThreeSrcCommutedOpIndices(const MachineInstr &MI,
1762 unsigned &SrcOpIdx1,
1763 unsigned &SrcOpIdx2,
1764 bool IsIntrinsic) const {
1765 uint64_t TSFlags = MI.getDesc().TSFlags;
1767 unsigned FirstCommutableVecOp = 1;
1768 unsigned LastCommutableVecOp = 3;
1769 unsigned KMaskOp = -1U;
1770 if (X86II::isKMasked(TSFlags)) {
1771 // For k-zero-masked operations it is Ok to commute the first vector
1773 // For regular k-masked operations a conservative choice is done as the
1774 // elements of the first vector operand, for which the corresponding bit
1775 // in the k-mask operand is set to 0, are copied to the result of the
1777 // TODO/FIXME: The commute still may be legal if it is known that the
1778 // k-mask operand is set to either all ones or all zeroes.
1779 // It is also Ok to commute the 1st operand if all users of MI use only
1780 // the elements enabled by the k-mask operand. For example,
1781 // v4 = VFMADD213PSZrk v1, k, v2, v3; // v1[i] = k[i] ? v2[i]*v1[i]+v3[i]
1783 // VMOVAPSZmrk <mem_addr>, k, v4; // this is the ONLY user of v4 ->
1784 // // Ok, to commute v1 in FMADD213PSZrk.
1786 // The k-mask operand has index = 2 for masked and zero-masked operations.
1789 // The operand with index = 1 is used as a source for those elements for
1790 // which the corresponding bit in the k-mask is set to 0.
1791 if (X86II::isKMergeMasked(TSFlags))
1792 FirstCommutableVecOp = 3;
1794 LastCommutableVecOp++;
1795 } else if (IsIntrinsic) {
1796 // Commuting the first operand of an intrinsic instruction isn't possible
1797 // unless we can prove that only the lowest element of the result is used.
1798 FirstCommutableVecOp = 2;
1801 if (isMem(MI, LastCommutableVecOp))
1802 LastCommutableVecOp--;
1804 // Only the first RegOpsNum operands are commutable.
1805 // Also, the value 'CommuteAnyOperandIndex' is valid here as it means
1806 // that the operand is not specified/fixed.
1807 if (SrcOpIdx1 != CommuteAnyOperandIndex &&
1808 (SrcOpIdx1 < FirstCommutableVecOp || SrcOpIdx1 > LastCommutableVecOp ||
1809 SrcOpIdx1 == KMaskOp))
1811 if (SrcOpIdx2 != CommuteAnyOperandIndex &&
1812 (SrcOpIdx2 < FirstCommutableVecOp || SrcOpIdx2 > LastCommutableVecOp ||
1813 SrcOpIdx2 == KMaskOp))
1816 // Look for two different register operands assumed to be commutable
1817 // regardless of the FMA opcode. The FMA opcode is adjusted later.
1818 if (SrcOpIdx1 == CommuteAnyOperandIndex ||
1819 SrcOpIdx2 == CommuteAnyOperandIndex) {
1820 unsigned CommutableOpIdx1 = SrcOpIdx1;
1821 unsigned CommutableOpIdx2 = SrcOpIdx2;
1823 // At least one of operands to be commuted is not specified and
1824 // this method is free to choose appropriate commutable operands.
1825 if (SrcOpIdx1 == SrcOpIdx2)
1826 // Both of operands are not fixed. By default set one of commutable
1827 // operands to the last register operand of the instruction.
1828 CommutableOpIdx2 = LastCommutableVecOp;
1829 else if (SrcOpIdx2 == CommuteAnyOperandIndex)
1830 // Only one of operands is not fixed.
1831 CommutableOpIdx2 = SrcOpIdx1;
1833 // CommutableOpIdx2 is well defined now. Let's choose another commutable
1834 // operand and assign its index to CommutableOpIdx1.
1835 unsigned Op2Reg = MI.getOperand(CommutableOpIdx2).getReg();
1836 for (CommutableOpIdx1 = LastCommutableVecOp;
1837 CommutableOpIdx1 >= FirstCommutableVecOp; CommutableOpIdx1--) {
1838 // Just ignore and skip the k-mask operand.
1839 if (CommutableOpIdx1 == KMaskOp)
1842 // The commuted operands must have different registers.
1843 // Otherwise, the commute transformation does not change anything and
1845 if (Op2Reg != MI.getOperand(CommutableOpIdx1).getReg())
1849 // No appropriate commutable operands were found.
1850 if (CommutableOpIdx1 < FirstCommutableVecOp)
1853 // Assign the found pair of commutable indices to SrcOpIdx1 and SrcOpidx2
1854 // to return those values.
1855 if (!fixCommutedOpIndices(SrcOpIdx1, SrcOpIdx2,
1856 CommutableOpIdx1, CommutableOpIdx2))
1863 bool X86InstrInfo::findCommutedOpIndices(MachineInstr &MI, unsigned &SrcOpIdx1,
1864 unsigned &SrcOpIdx2) const {
1865 const MCInstrDesc &Desc = MI.getDesc();
1866 if (!Desc.isCommutable())
1869 switch (MI.getOpcode()) {
1876 case X86::VCMPPDrri:
1877 case X86::VCMPPSrri:
1878 case X86::VCMPPDYrri:
1879 case X86::VCMPPSYrri:
1880 case X86::VCMPSDZrr:
1881 case X86::VCMPSSZrr:
1882 case X86::VCMPPDZrri:
1883 case X86::VCMPPSZrri:
1884 case X86::VCMPPDZ128rri:
1885 case X86::VCMPPSZ128rri:
1886 case X86::VCMPPDZ256rri:
1887 case X86::VCMPPSZ256rri: {
1888 // Float comparison can be safely commuted for
1889 // Ordered/Unordered/Equal/NotEqual tests
1890 unsigned Imm = MI.getOperand(3).getImm() & 0x7;
1893 case 0x03: // UNORDERED
1894 case 0x04: // NOT EQUAL
1895 case 0x07: // ORDERED
1896 // The indices of the commutable operands are 1 and 2.
1897 // Assign them to the returned operand indices here.
1898 return fixCommutedOpIndices(SrcOpIdx1, SrcOpIdx2, 1, 2);
1906 if (Subtarget.hasSSE41())
1907 return TargetInstrInfo::findCommutedOpIndices(MI, SrcOpIdx1, SrcOpIdx2);
1909 case X86::MOVHLPSrr:
1910 case X86::UNPCKHPDrr:
1911 case X86::VMOVHLPSrr:
1912 case X86::VUNPCKHPDrr:
1913 case X86::VMOVHLPSZrr:
1914 case X86::VUNPCKHPDZ128rr:
1915 if (Subtarget.hasSSE2())
1916 return TargetInstrInfo::findCommutedOpIndices(MI, SrcOpIdx1, SrcOpIdx2);
1918 case X86::VPTERNLOGDZrri: case X86::VPTERNLOGDZrmi:
1919 case X86::VPTERNLOGDZ128rri: case X86::VPTERNLOGDZ128rmi:
1920 case X86::VPTERNLOGDZ256rri: case X86::VPTERNLOGDZ256rmi:
1921 case X86::VPTERNLOGQZrri: case X86::VPTERNLOGQZrmi:
1922 case X86::VPTERNLOGQZ128rri: case X86::VPTERNLOGQZ128rmi:
1923 case X86::VPTERNLOGQZ256rri: case X86::VPTERNLOGQZ256rmi:
1924 case X86::VPTERNLOGDZrrik:
1925 case X86::VPTERNLOGDZ128rrik:
1926 case X86::VPTERNLOGDZ256rrik:
1927 case X86::VPTERNLOGQZrrik:
1928 case X86::VPTERNLOGQZ128rrik:
1929 case X86::VPTERNLOGQZ256rrik:
1930 case X86::VPTERNLOGDZrrikz: case X86::VPTERNLOGDZrmikz:
1931 case X86::VPTERNLOGDZ128rrikz: case X86::VPTERNLOGDZ128rmikz:
1932 case X86::VPTERNLOGDZ256rrikz: case X86::VPTERNLOGDZ256rmikz:
1933 case X86::VPTERNLOGQZrrikz: case X86::VPTERNLOGQZrmikz:
1934 case X86::VPTERNLOGQZ128rrikz: case X86::VPTERNLOGQZ128rmikz:
1935 case X86::VPTERNLOGQZ256rrikz: case X86::VPTERNLOGQZ256rmikz:
1936 case X86::VPTERNLOGDZ128rmbi:
1937 case X86::VPTERNLOGDZ256rmbi:
1938 case X86::VPTERNLOGDZrmbi:
1939 case X86::VPTERNLOGQZ128rmbi:
1940 case X86::VPTERNLOGQZ256rmbi:
1941 case X86::VPTERNLOGQZrmbi:
1942 case X86::VPTERNLOGDZ128rmbikz:
1943 case X86::VPTERNLOGDZ256rmbikz:
1944 case X86::VPTERNLOGDZrmbikz:
1945 case X86::VPTERNLOGQZ128rmbikz:
1946 case X86::VPTERNLOGQZ256rmbikz:
1947 case X86::VPTERNLOGQZrmbikz:
1948 return findThreeSrcCommutedOpIndices(MI, SrcOpIdx1, SrcOpIdx2);
1949 case X86::VPMADD52HUQZ128r:
1950 case X86::VPMADD52HUQZ128rk:
1951 case X86::VPMADD52HUQZ128rkz:
1952 case X86::VPMADD52HUQZ256r:
1953 case X86::VPMADD52HUQZ256rk:
1954 case X86::VPMADD52HUQZ256rkz:
1955 case X86::VPMADD52HUQZr:
1956 case X86::VPMADD52HUQZrk:
1957 case X86::VPMADD52HUQZrkz:
1958 case X86::VPMADD52LUQZ128r:
1959 case X86::VPMADD52LUQZ128rk:
1960 case X86::VPMADD52LUQZ128rkz:
1961 case X86::VPMADD52LUQZ256r:
1962 case X86::VPMADD52LUQZ256rk:
1963 case X86::VPMADD52LUQZ256rkz:
1964 case X86::VPMADD52LUQZr:
1965 case X86::VPMADD52LUQZrk:
1966 case X86::VPMADD52LUQZrkz: {
1967 unsigned CommutableOpIdx1 = 2;
1968 unsigned CommutableOpIdx2 = 3;
1969 if (X86II::isKMasked(Desc.TSFlags)) {
1970 // Skip the mask register.
1974 if (!fixCommutedOpIndices(SrcOpIdx1, SrcOpIdx2,
1975 CommutableOpIdx1, CommutableOpIdx2))
1977 if (!MI.getOperand(SrcOpIdx1).isReg() ||
1978 !MI.getOperand(SrcOpIdx2).isReg())
1985 const X86InstrFMA3Group *FMA3Group = getFMA3Group(MI.getOpcode(),
1986 MI.getDesc().TSFlags);
1988 return findThreeSrcCommutedOpIndices(MI, SrcOpIdx1, SrcOpIdx2,
1989 FMA3Group->isIntrinsic());
1991 // Handled masked instructions since we need to skip over the mask input
1992 // and the preserved input.
1993 if (X86II::isKMasked(Desc.TSFlags)) {
1994 // First assume that the first input is the mask operand and skip past it.
1995 unsigned CommutableOpIdx1 = Desc.getNumDefs() + 1;
1996 unsigned CommutableOpIdx2 = Desc.getNumDefs() + 2;
1997 // Check if the first input is tied. If there isn't one then we only
1998 // need to skip the mask operand which we did above.
1999 if ((MI.getDesc().getOperandConstraint(Desc.getNumDefs(),
2000 MCOI::TIED_TO) != -1)) {
2001 // If this is zero masking instruction with a tied operand, we need to
2002 // move the first index back to the first input since this must
2003 // be a 3 input instruction and we want the first two non-mask inputs.
2004 // Otherwise this is a 2 input instruction with a preserved input and
2005 // mask, so we need to move the indices to skip one more input.
2006 if (X86II::isKMergeMasked(Desc.TSFlags)) {
2014 if (!fixCommutedOpIndices(SrcOpIdx1, SrcOpIdx2,
2015 CommutableOpIdx1, CommutableOpIdx2))
2018 if (!MI.getOperand(SrcOpIdx1).isReg() ||
2019 !MI.getOperand(SrcOpIdx2).isReg())
2025 return TargetInstrInfo::findCommutedOpIndices(MI, SrcOpIdx1, SrcOpIdx2);
2030 X86::CondCode X86::getCondFromBranchOpc(unsigned BrOpc) {
2032 default: return X86::COND_INVALID;
2033 case X86::JE_1: return X86::COND_E;
2034 case X86::JNE_1: return X86::COND_NE;
2035 case X86::JL_1: return X86::COND_L;
2036 case X86::JLE_1: return X86::COND_LE;
2037 case X86::JG_1: return X86::COND_G;
2038 case X86::JGE_1: return X86::COND_GE;
2039 case X86::JB_1: return X86::COND_B;
2040 case X86::JBE_1: return X86::COND_BE;
2041 case X86::JA_1: return X86::COND_A;
2042 case X86::JAE_1: return X86::COND_AE;
2043 case X86::JS_1: return X86::COND_S;
2044 case X86::JNS_1: return X86::COND_NS;
2045 case X86::JP_1: return X86::COND_P;
2046 case X86::JNP_1: return X86::COND_NP;
2047 case X86::JO_1: return X86::COND_O;
2048 case X86::JNO_1: return X86::COND_NO;
2052 /// Return condition code of a SET opcode.
2053 X86::CondCode X86::getCondFromSETOpc(unsigned Opc) {
2055 default: return X86::COND_INVALID;
2056 case X86::SETAr: case X86::SETAm: return X86::COND_A;
2057 case X86::SETAEr: case X86::SETAEm: return X86::COND_AE;
2058 case X86::SETBr: case X86::SETBm: return X86::COND_B;
2059 case X86::SETBEr: case X86::SETBEm: return X86::COND_BE;
2060 case X86::SETEr: case X86::SETEm: return X86::COND_E;
2061 case X86::SETGr: case X86::SETGm: return X86::COND_G;
2062 case X86::SETGEr: case X86::SETGEm: return X86::COND_GE;
2063 case X86::SETLr: case X86::SETLm: return X86::COND_L;
2064 case X86::SETLEr: case X86::SETLEm: return X86::COND_LE;
2065 case X86::SETNEr: case X86::SETNEm: return X86::COND_NE;
2066 case X86::SETNOr: case X86::SETNOm: return X86::COND_NO;
2067 case X86::SETNPr: case X86::SETNPm: return X86::COND_NP;
2068 case X86::SETNSr: case X86::SETNSm: return X86::COND_NS;
2069 case X86::SETOr: case X86::SETOm: return X86::COND_O;
2070 case X86::SETPr: case X86::SETPm: return X86::COND_P;
2071 case X86::SETSr: case X86::SETSm: return X86::COND_S;
2075 /// Return condition code of a CMov opcode.
2076 X86::CondCode X86::getCondFromCMovOpc(unsigned Opc) {
2078 default: return X86::COND_INVALID;
2079 case X86::CMOVA16rm: case X86::CMOVA16rr: case X86::CMOVA32rm:
2080 case X86::CMOVA32rr: case X86::CMOVA64rm: case X86::CMOVA64rr:
2082 case X86::CMOVAE16rm: case X86::CMOVAE16rr: case X86::CMOVAE32rm:
2083 case X86::CMOVAE32rr: case X86::CMOVAE64rm: case X86::CMOVAE64rr:
2084 return X86::COND_AE;
2085 case X86::CMOVB16rm: case X86::CMOVB16rr: case X86::CMOVB32rm:
2086 case X86::CMOVB32rr: case X86::CMOVB64rm: case X86::CMOVB64rr:
2088 case X86::CMOVBE16rm: case X86::CMOVBE16rr: case X86::CMOVBE32rm:
2089 case X86::CMOVBE32rr: case X86::CMOVBE64rm: case X86::CMOVBE64rr:
2090 return X86::COND_BE;
2091 case X86::CMOVE16rm: case X86::CMOVE16rr: case X86::CMOVE32rm:
2092 case X86::CMOVE32rr: case X86::CMOVE64rm: case X86::CMOVE64rr:
2094 case X86::CMOVG16rm: case X86::CMOVG16rr: case X86::CMOVG32rm:
2095 case X86::CMOVG32rr: case X86::CMOVG64rm: case X86::CMOVG64rr:
2097 case X86::CMOVGE16rm: case X86::CMOVGE16rr: case X86::CMOVGE32rm:
2098 case X86::CMOVGE32rr: case X86::CMOVGE64rm: case X86::CMOVGE64rr:
2099 return X86::COND_GE;
2100 case X86::CMOVL16rm: case X86::CMOVL16rr: case X86::CMOVL32rm:
2101 case X86::CMOVL32rr: case X86::CMOVL64rm: case X86::CMOVL64rr:
2103 case X86::CMOVLE16rm: case X86::CMOVLE16rr: case X86::CMOVLE32rm:
2104 case X86::CMOVLE32rr: case X86::CMOVLE64rm: case X86::CMOVLE64rr:
2105 return X86::COND_LE;
2106 case X86::CMOVNE16rm: case X86::CMOVNE16rr: case X86::CMOVNE32rm:
2107 case X86::CMOVNE32rr: case X86::CMOVNE64rm: case X86::CMOVNE64rr:
2108 return X86::COND_NE;
2109 case X86::CMOVNO16rm: case X86::CMOVNO16rr: case X86::CMOVNO32rm:
2110 case X86::CMOVNO32rr: case X86::CMOVNO64rm: case X86::CMOVNO64rr:
2111 return X86::COND_NO;
2112 case X86::CMOVNP16rm: case X86::CMOVNP16rr: case X86::CMOVNP32rm:
2113 case X86::CMOVNP32rr: case X86::CMOVNP64rm: case X86::CMOVNP64rr:
2114 return X86::COND_NP;
2115 case X86::CMOVNS16rm: case X86::CMOVNS16rr: case X86::CMOVNS32rm:
2116 case X86::CMOVNS32rr: case X86::CMOVNS64rm: case X86::CMOVNS64rr:
2117 return X86::COND_NS;
2118 case X86::CMOVO16rm: case X86::CMOVO16rr: case X86::CMOVO32rm:
2119 case X86::CMOVO32rr: case X86::CMOVO64rm: case X86::CMOVO64rr:
2121 case X86::CMOVP16rm: case X86::CMOVP16rr: case X86::CMOVP32rm:
2122 case X86::CMOVP32rr: case X86::CMOVP64rm: case X86::CMOVP64rr:
2124 case X86::CMOVS16rm: case X86::CMOVS16rr: case X86::CMOVS32rm:
2125 case X86::CMOVS32rr: case X86::CMOVS64rm: case X86::CMOVS64rr:
2130 unsigned X86::GetCondBranchFromCond(X86::CondCode CC) {
2132 default: llvm_unreachable("Illegal condition code!");
2133 case X86::COND_E: return X86::JE_1;
2134 case X86::COND_NE: return X86::JNE_1;
2135 case X86::COND_L: return X86::JL_1;
2136 case X86::COND_LE: return X86::JLE_1;
2137 case X86::COND_G: return X86::JG_1;
2138 case X86::COND_GE: return X86::JGE_1;
2139 case X86::COND_B: return X86::JB_1;
2140 case X86::COND_BE: return X86::JBE_1;
2141 case X86::COND_A: return X86::JA_1;
2142 case X86::COND_AE: return X86::JAE_1;
2143 case X86::COND_S: return X86::JS_1;
2144 case X86::COND_NS: return X86::JNS_1;
2145 case X86::COND_P: return X86::JP_1;
2146 case X86::COND_NP: return X86::JNP_1;
2147 case X86::COND_O: return X86::JO_1;
2148 case X86::COND_NO: return X86::JNO_1;
2152 /// Return the inverse of the specified condition,
2153 /// e.g. turning COND_E to COND_NE.
2154 X86::CondCode X86::GetOppositeBranchCondition(X86::CondCode CC) {
2156 default: llvm_unreachable("Illegal condition code!");
2157 case X86::COND_E: return X86::COND_NE;
2158 case X86::COND_NE: return X86::COND_E;
2159 case X86::COND_L: return X86::COND_GE;
2160 case X86::COND_LE: return X86::COND_G;
2161 case X86::COND_G: return X86::COND_LE;
2162 case X86::COND_GE: return X86::COND_L;
2163 case X86::COND_B: return X86::COND_AE;
2164 case X86::COND_BE: return X86::COND_A;
2165 case X86::COND_A: return X86::COND_BE;
2166 case X86::COND_AE: return X86::COND_B;
2167 case X86::COND_S: return X86::COND_NS;
2168 case X86::COND_NS: return X86::COND_S;
2169 case X86::COND_P: return X86::COND_NP;
2170 case X86::COND_NP: return X86::COND_P;
2171 case X86::COND_O: return X86::COND_NO;
2172 case X86::COND_NO: return X86::COND_O;
2173 case X86::COND_NE_OR_P: return X86::COND_E_AND_NP;
2174 case X86::COND_E_AND_NP: return X86::COND_NE_OR_P;
2178 /// Assuming the flags are set by MI(a,b), return the condition code if we
2179 /// modify the instructions such that flags are set by MI(b,a).
2180 static X86::CondCode getSwappedCondition(X86::CondCode CC) {
2182 default: return X86::COND_INVALID;
2183 case X86::COND_E: return X86::COND_E;
2184 case X86::COND_NE: return X86::COND_NE;
2185 case X86::COND_L: return X86::COND_G;
2186 case X86::COND_LE: return X86::COND_GE;
2187 case X86::COND_G: return X86::COND_L;
2188 case X86::COND_GE: return X86::COND_LE;
2189 case X86::COND_B: return X86::COND_A;
2190 case X86::COND_BE: return X86::COND_AE;
2191 case X86::COND_A: return X86::COND_B;
2192 case X86::COND_AE: return X86::COND_BE;
2196 std::pair<X86::CondCode, bool>
2197 X86::getX86ConditionCode(CmpInst::Predicate Predicate) {
2198 X86::CondCode CC = X86::COND_INVALID;
2199 bool NeedSwap = false;
2200 switch (Predicate) {
2202 // Floating-point Predicates
2203 case CmpInst::FCMP_UEQ: CC = X86::COND_E; break;
2204 case CmpInst::FCMP_OLT: NeedSwap = true; LLVM_FALLTHROUGH;
2205 case CmpInst::FCMP_OGT: CC = X86::COND_A; break;
2206 case CmpInst::FCMP_OLE: NeedSwap = true; LLVM_FALLTHROUGH;
2207 case CmpInst::FCMP_OGE: CC = X86::COND_AE; break;
2208 case CmpInst::FCMP_UGT: NeedSwap = true; LLVM_FALLTHROUGH;
2209 case CmpInst::FCMP_ULT: CC = X86::COND_B; break;
2210 case CmpInst::FCMP_UGE: NeedSwap = true; LLVM_FALLTHROUGH;
2211 case CmpInst::FCMP_ULE: CC = X86::COND_BE; break;
2212 case CmpInst::FCMP_ONE: CC = X86::COND_NE; break;
2213 case CmpInst::FCMP_UNO: CC = X86::COND_P; break;
2214 case CmpInst::FCMP_ORD: CC = X86::COND_NP; break;
2215 case CmpInst::FCMP_OEQ: LLVM_FALLTHROUGH;
2216 case CmpInst::FCMP_UNE: CC = X86::COND_INVALID; break;
2218 // Integer Predicates
2219 case CmpInst::ICMP_EQ: CC = X86::COND_E; break;
2220 case CmpInst::ICMP_NE: CC = X86::COND_NE; break;
2221 case CmpInst::ICMP_UGT: CC = X86::COND_A; break;
2222 case CmpInst::ICMP_UGE: CC = X86::COND_AE; break;
2223 case CmpInst::ICMP_ULT: CC = X86::COND_B; break;
2224 case CmpInst::ICMP_ULE: CC = X86::COND_BE; break;
2225 case CmpInst::ICMP_SGT: CC = X86::COND_G; break;
2226 case CmpInst::ICMP_SGE: CC = X86::COND_GE; break;
2227 case CmpInst::ICMP_SLT: CC = X86::COND_L; break;
2228 case CmpInst::ICMP_SLE: CC = X86::COND_LE; break;
2231 return std::make_pair(CC, NeedSwap);
2234 /// Return a set opcode for the given condition and
2235 /// whether it has memory operand.
2236 unsigned X86::getSETFromCond(CondCode CC, bool HasMemoryOperand) {
2237 static const uint16_t Opc[16][2] = {
2238 { X86::SETAr, X86::SETAm },
2239 { X86::SETAEr, X86::SETAEm },
2240 { X86::SETBr, X86::SETBm },
2241 { X86::SETBEr, X86::SETBEm },
2242 { X86::SETEr, X86::SETEm },
2243 { X86::SETGr, X86::SETGm },
2244 { X86::SETGEr, X86::SETGEm },
2245 { X86::SETLr, X86::SETLm },
2246 { X86::SETLEr, X86::SETLEm },
2247 { X86::SETNEr, X86::SETNEm },
2248 { X86::SETNOr, X86::SETNOm },
2249 { X86::SETNPr, X86::SETNPm },
2250 { X86::SETNSr, X86::SETNSm },
2251 { X86::SETOr, X86::SETOm },
2252 { X86::SETPr, X86::SETPm },
2253 { X86::SETSr, X86::SETSm }
2256 assert(CC <= LAST_VALID_COND && "Can only handle standard cond codes");
2257 return Opc[CC][HasMemoryOperand ? 1 : 0];
2260 /// Return a cmov opcode for the given condition,
2261 /// register size in bytes, and operand type.
2262 unsigned X86::getCMovFromCond(CondCode CC, unsigned RegBytes,
2263 bool HasMemoryOperand) {
2264 static const uint16_t Opc[32][3] = {
2265 { X86::CMOVA16rr, X86::CMOVA32rr, X86::CMOVA64rr },
2266 { X86::CMOVAE16rr, X86::CMOVAE32rr, X86::CMOVAE64rr },
2267 { X86::CMOVB16rr, X86::CMOVB32rr, X86::CMOVB64rr },
2268 { X86::CMOVBE16rr, X86::CMOVBE32rr, X86::CMOVBE64rr },
2269 { X86::CMOVE16rr, X86::CMOVE32rr, X86::CMOVE64rr },
2270 { X86::CMOVG16rr, X86::CMOVG32rr, X86::CMOVG64rr },
2271 { X86::CMOVGE16rr, X86::CMOVGE32rr, X86::CMOVGE64rr },
2272 { X86::CMOVL16rr, X86::CMOVL32rr, X86::CMOVL64rr },
2273 { X86::CMOVLE16rr, X86::CMOVLE32rr, X86::CMOVLE64rr },
2274 { X86::CMOVNE16rr, X86::CMOVNE32rr, X86::CMOVNE64rr },
2275 { X86::CMOVNO16rr, X86::CMOVNO32rr, X86::CMOVNO64rr },
2276 { X86::CMOVNP16rr, X86::CMOVNP32rr, X86::CMOVNP64rr },
2277 { X86::CMOVNS16rr, X86::CMOVNS32rr, X86::CMOVNS64rr },
2278 { X86::CMOVO16rr, X86::CMOVO32rr, X86::CMOVO64rr },
2279 { X86::CMOVP16rr, X86::CMOVP32rr, X86::CMOVP64rr },
2280 { X86::CMOVS16rr, X86::CMOVS32rr, X86::CMOVS64rr },
2281 { X86::CMOVA16rm, X86::CMOVA32rm, X86::CMOVA64rm },
2282 { X86::CMOVAE16rm, X86::CMOVAE32rm, X86::CMOVAE64rm },
2283 { X86::CMOVB16rm, X86::CMOVB32rm, X86::CMOVB64rm },
2284 { X86::CMOVBE16rm, X86::CMOVBE32rm, X86::CMOVBE64rm },
2285 { X86::CMOVE16rm, X86::CMOVE32rm, X86::CMOVE64rm },
2286 { X86::CMOVG16rm, X86::CMOVG32rm, X86::CMOVG64rm },
2287 { X86::CMOVGE16rm, X86::CMOVGE32rm, X86::CMOVGE64rm },
2288 { X86::CMOVL16rm, X86::CMOVL32rm, X86::CMOVL64rm },
2289 { X86::CMOVLE16rm, X86::CMOVLE32rm, X86::CMOVLE64rm },
2290 { X86::CMOVNE16rm, X86::CMOVNE32rm, X86::CMOVNE64rm },
2291 { X86::CMOVNO16rm, X86::CMOVNO32rm, X86::CMOVNO64rm },
2292 { X86::CMOVNP16rm, X86::CMOVNP32rm, X86::CMOVNP64rm },
2293 { X86::CMOVNS16rm, X86::CMOVNS32rm, X86::CMOVNS64rm },
2294 { X86::CMOVO16rm, X86::CMOVO32rm, X86::CMOVO64rm },
2295 { X86::CMOVP16rm, X86::CMOVP32rm, X86::CMOVP64rm },
2296 { X86::CMOVS16rm, X86::CMOVS32rm, X86::CMOVS64rm }
2299 assert(CC < 16 && "Can only handle standard cond codes");
2300 unsigned Idx = HasMemoryOperand ? 16+CC : CC;
2302 default: llvm_unreachable("Illegal register size!");
2303 case 2: return Opc[Idx][0];
2304 case 4: return Opc[Idx][1];
2305 case 8: return Opc[Idx][2];
2309 /// Get the VPCMP immediate for the given condition.
2310 unsigned X86::getVPCMPImmForCond(ISD::CondCode CC) {
2312 default: llvm_unreachable("Unexpected SETCC condition");
2313 case ISD::SETNE: return 4;
2314 case ISD::SETEQ: return 0;
2316 case ISD::SETLT: return 1;
2318 case ISD::SETGT: return 6;
2320 case ISD::SETGE: return 5;
2322 case ISD::SETLE: return 2;
2326 /// Get the VPCMP immediate if the opcodes are swapped.
2327 unsigned X86::getSwappedVPCMPImm(unsigned Imm) {
2329 default: llvm_unreachable("Unreachable!");
2330 case 0x01: Imm = 0x06; break; // LT -> NLE
2331 case 0x02: Imm = 0x05; break; // LE -> NLT
2332 case 0x05: Imm = 0x02; break; // NLT -> LE
2333 case 0x06: Imm = 0x01; break; // NLE -> LT
2344 /// Get the VPCOM immediate if the opcodes are swapped.
2345 unsigned X86::getSwappedVPCOMImm(unsigned Imm) {
2347 default: llvm_unreachable("Unreachable!");
2348 case 0x00: Imm = 0x02; break; // LT -> GT
2349 case 0x01: Imm = 0x03; break; // LE -> GE
2350 case 0x02: Imm = 0x00; break; // GT -> LT
2351 case 0x03: Imm = 0x01; break; // GE -> LE
2362 bool X86InstrInfo::isUnpredicatedTerminator(const MachineInstr &MI) const {
2363 if (!MI.isTerminator()) return false;
2365 // Conditional branch is a special case.
2366 if (MI.isBranch() && !MI.isBarrier())
2368 if (!MI.isPredicable())
2370 return !isPredicated(MI);
2373 bool X86InstrInfo::isUnconditionalTailCall(const MachineInstr &MI) const {
2374 switch (MI.getOpcode()) {
2375 case X86::TCRETURNdi:
2376 case X86::TCRETURNri:
2377 case X86::TCRETURNmi:
2378 case X86::TCRETURNdi64:
2379 case X86::TCRETURNri64:
2380 case X86::TCRETURNmi64:
2387 bool X86InstrInfo::canMakeTailCallConditional(
2388 SmallVectorImpl<MachineOperand> &BranchCond,
2389 const MachineInstr &TailCall) const {
2390 if (TailCall.getOpcode() != X86::TCRETURNdi &&
2391 TailCall.getOpcode() != X86::TCRETURNdi64) {
2392 // Only direct calls can be done with a conditional branch.
2396 const MachineFunction *MF = TailCall.getParent()->getParent();
2397 if (Subtarget.isTargetWin64() && MF->hasWinCFI()) {
2398 // Conditional tail calls confuse the Win64 unwinder.
2402 assert(BranchCond.size() == 1);
2403 if (BranchCond[0].getImm() > X86::LAST_VALID_COND) {
2404 // Can't make a conditional tail call with this condition.
2408 const X86MachineFunctionInfo *X86FI = MF->getInfo<X86MachineFunctionInfo>();
2409 if (X86FI->getTCReturnAddrDelta() != 0 ||
2410 TailCall.getOperand(1).getImm() != 0) {
2411 // A conditional tail call cannot do any stack adjustment.
2418 void X86InstrInfo::replaceBranchWithTailCall(
2419 MachineBasicBlock &MBB, SmallVectorImpl<MachineOperand> &BranchCond,
2420 const MachineInstr &TailCall) const {
2421 assert(canMakeTailCallConditional(BranchCond, TailCall));
2423 MachineBasicBlock::iterator I = MBB.end();
2424 while (I != MBB.begin()) {
2426 if (I->isDebugInstr())
2429 assert(0 && "Can't find the branch to replace!");
2431 X86::CondCode CC = X86::getCondFromBranchOpc(I->getOpcode());
2432 assert(BranchCond.size() == 1);
2433 if (CC != BranchCond[0].getImm())
2439 unsigned Opc = TailCall.getOpcode() == X86::TCRETURNdi ? X86::TCRETURNdicc
2440 : X86::TCRETURNdi64cc;
2442 auto MIB = BuildMI(MBB, I, MBB.findDebugLoc(I), get(Opc));
2443 MIB->addOperand(TailCall.getOperand(0)); // Destination.
2444 MIB.addImm(0); // Stack offset (not used).
2445 MIB->addOperand(BranchCond[0]); // Condition.
2446 MIB.copyImplicitOps(TailCall); // Regmask and (imp-used) parameters.
2448 // Add implicit uses and defs of all live regs potentially clobbered by the
2449 // call. This way they still appear live across the call.
2450 LivePhysRegs LiveRegs(getRegisterInfo());
2451 LiveRegs.addLiveOuts(MBB);
2452 SmallVector<std::pair<MCPhysReg, const MachineOperand *>, 8> Clobbers;
2453 LiveRegs.stepForward(*MIB, Clobbers);
2454 for (const auto &C : Clobbers) {
2455 MIB.addReg(C.first, RegState::Implicit);
2456 MIB.addReg(C.first, RegState::Implicit | RegState::Define);
2459 I->eraseFromParent();
2462 // Given a MBB and its TBB, find the FBB which was a fallthrough MBB (it may
2463 // not be a fallthrough MBB now due to layout changes). Return nullptr if the
2464 // fallthrough MBB cannot be identified.
2465 static MachineBasicBlock *getFallThroughMBB(MachineBasicBlock *MBB,
2466 MachineBasicBlock *TBB) {
2467 // Look for non-EHPad successors other than TBB. If we find exactly one, it
2468 // is the fallthrough MBB. If we find zero, then TBB is both the target MBB
2469 // and fallthrough MBB. If we find more than one, we cannot identify the
2470 // fallthrough MBB and should return nullptr.
2471 MachineBasicBlock *FallthroughBB = nullptr;
2472 for (auto SI = MBB->succ_begin(), SE = MBB->succ_end(); SI != SE; ++SI) {
2473 if ((*SI)->isEHPad() || (*SI == TBB && FallthroughBB))
2475 // Return a nullptr if we found more than one fallthrough successor.
2476 if (FallthroughBB && FallthroughBB != TBB)
2478 FallthroughBB = *SI;
2480 return FallthroughBB;
2483 bool X86InstrInfo::AnalyzeBranchImpl(
2484 MachineBasicBlock &MBB, MachineBasicBlock *&TBB, MachineBasicBlock *&FBB,
2485 SmallVectorImpl<MachineOperand> &Cond,
2486 SmallVectorImpl<MachineInstr *> &CondBranches, bool AllowModify) const {
2488 // Start from the bottom of the block and work up, examining the
2489 // terminator instructions.
2490 MachineBasicBlock::iterator I = MBB.end();
2491 MachineBasicBlock::iterator UnCondBrIter = MBB.end();
2492 while (I != MBB.begin()) {
2494 if (I->isDebugInstr())
2497 // Working from the bottom, when we see a non-terminator instruction, we're
2499 if (!isUnpredicatedTerminator(*I))
2502 // A terminator that isn't a branch can't easily be handled by this
2507 // Handle unconditional branches.
2508 if (I->getOpcode() == X86::JMP_1) {
2512 TBB = I->getOperand(0).getMBB();
2516 // If the block has any instructions after a JMP, delete them.
2517 while (std::next(I) != MBB.end())
2518 std::next(I)->eraseFromParent();
2523 // Delete the JMP if it's equivalent to a fall-through.
2524 if (MBB.isLayoutSuccessor(I->getOperand(0).getMBB())) {
2526 I->eraseFromParent();
2528 UnCondBrIter = MBB.end();
2532 // TBB is used to indicate the unconditional destination.
2533 TBB = I->getOperand(0).getMBB();
2537 // Handle conditional branches.
2538 X86::CondCode BranchCode = X86::getCondFromBranchOpc(I->getOpcode());
2539 if (BranchCode == X86::COND_INVALID)
2540 return true; // Can't handle indirect branch.
2542 // In practice we should never have an undef eflags operand, if we do
2543 // abort here as we are not prepared to preserve the flag.
2544 if (I->getOperand(1).isUndef())
2547 // Working from the bottom, handle the first conditional branch.
2549 MachineBasicBlock *TargetBB = I->getOperand(0).getMBB();
2550 if (AllowModify && UnCondBrIter != MBB.end() &&
2551 MBB.isLayoutSuccessor(TargetBB)) {
2552 // If we can modify the code and it ends in something like:
2560 // Then we can change this to:
2567 // Which is a bit more efficient.
2568 // We conditionally jump to the fall-through block.
2569 BranchCode = GetOppositeBranchCondition(BranchCode);
2570 unsigned JNCC = GetCondBranchFromCond(BranchCode);
2571 MachineBasicBlock::iterator OldInst = I;
2573 BuildMI(MBB, UnCondBrIter, MBB.findDebugLoc(I), get(JNCC))
2574 .addMBB(UnCondBrIter->getOperand(0).getMBB());
2575 BuildMI(MBB, UnCondBrIter, MBB.findDebugLoc(I), get(X86::JMP_1))
2578 OldInst->eraseFromParent();
2579 UnCondBrIter->eraseFromParent();
2581 // Restart the analysis.
2582 UnCondBrIter = MBB.end();
2588 TBB = I->getOperand(0).getMBB();
2589 Cond.push_back(MachineOperand::CreateImm(BranchCode));
2590 CondBranches.push_back(&*I);
2594 // Handle subsequent conditional branches. Only handle the case where all
2595 // conditional branches branch to the same destination and their condition
2596 // opcodes fit one of the special multi-branch idioms.
2597 assert(Cond.size() == 1);
2600 // If the conditions are the same, we can leave them alone.
2601 X86::CondCode OldBranchCode = (X86::CondCode)Cond[0].getImm();
2602 auto NewTBB = I->getOperand(0).getMBB();
2603 if (OldBranchCode == BranchCode && TBB == NewTBB)
2606 // If they differ, see if they fit one of the known patterns. Theoretically,
2607 // we could handle more patterns here, but we shouldn't expect to see them
2608 // if instruction selection has done a reasonable job.
2609 if (TBB == NewTBB &&
2610 ((OldBranchCode == X86::COND_P && BranchCode == X86::COND_NE) ||
2611 (OldBranchCode == X86::COND_NE && BranchCode == X86::COND_P))) {
2612 BranchCode = X86::COND_NE_OR_P;
2613 } else if ((OldBranchCode == X86::COND_NP && BranchCode == X86::COND_NE) ||
2614 (OldBranchCode == X86::COND_E && BranchCode == X86::COND_P)) {
2615 if (NewTBB != (FBB ? FBB : getFallThroughMBB(&MBB, TBB)))
2618 // X86::COND_E_AND_NP usually has two different branch destinations.
2626 // Here this condition branches to B2 only if NP && E. It has another
2635 // Similarly it branches to B2 only if E && NP. That is why this condition
2636 // is named with COND_E_AND_NP.
2637 BranchCode = X86::COND_E_AND_NP;
2641 // Update the MachineOperand.
2642 Cond[0].setImm(BranchCode);
2643 CondBranches.push_back(&*I);
2649 bool X86InstrInfo::analyzeBranch(MachineBasicBlock &MBB,
2650 MachineBasicBlock *&TBB,
2651 MachineBasicBlock *&FBB,
2652 SmallVectorImpl<MachineOperand> &Cond,
2653 bool AllowModify) const {
2654 SmallVector<MachineInstr *, 4> CondBranches;
2655 return AnalyzeBranchImpl(MBB, TBB, FBB, Cond, CondBranches, AllowModify);
2658 bool X86InstrInfo::analyzeBranchPredicate(MachineBasicBlock &MBB,
2659 MachineBranchPredicate &MBP,
2660 bool AllowModify) const {
2661 using namespace std::placeholders;
2663 SmallVector<MachineOperand, 4> Cond;
2664 SmallVector<MachineInstr *, 4> CondBranches;
2665 if (AnalyzeBranchImpl(MBB, MBP.TrueDest, MBP.FalseDest, Cond, CondBranches,
2669 if (Cond.size() != 1)
2672 assert(MBP.TrueDest && "expected!");
2675 MBP.FalseDest = MBB.getNextNode();
2677 const TargetRegisterInfo *TRI = &getRegisterInfo();
2679 MachineInstr *ConditionDef = nullptr;
2680 bool SingleUseCondition = true;
2682 for (auto I = std::next(MBB.rbegin()), E = MBB.rend(); I != E; ++I) {
2683 if (I->modifiesRegister(X86::EFLAGS, TRI)) {
2688 if (I->readsRegister(X86::EFLAGS, TRI))
2689 SingleUseCondition = false;
2695 if (SingleUseCondition) {
2696 for (auto *Succ : MBB.successors())
2697 if (Succ->isLiveIn(X86::EFLAGS))
2698 SingleUseCondition = false;
2701 MBP.ConditionDef = ConditionDef;
2702 MBP.SingleUseCondition = SingleUseCondition;
2704 // Currently we only recognize the simple pattern:
2709 const unsigned TestOpcode =
2710 Subtarget.is64Bit() ? X86::TEST64rr : X86::TEST32rr;
2712 if (ConditionDef->getOpcode() == TestOpcode &&
2713 ConditionDef->getNumOperands() == 3 &&
2714 ConditionDef->getOperand(0).isIdenticalTo(ConditionDef->getOperand(1)) &&
2715 (Cond[0].getImm() == X86::COND_NE || Cond[0].getImm() == X86::COND_E)) {
2716 MBP.LHS = ConditionDef->getOperand(0);
2717 MBP.RHS = MachineOperand::CreateImm(0);
2718 MBP.Predicate = Cond[0].getImm() == X86::COND_NE
2719 ? MachineBranchPredicate::PRED_NE
2720 : MachineBranchPredicate::PRED_EQ;
2727 unsigned X86InstrInfo::removeBranch(MachineBasicBlock &MBB,
2728 int *BytesRemoved) const {
2729 assert(!BytesRemoved && "code size not handled");
2731 MachineBasicBlock::iterator I = MBB.end();
2734 while (I != MBB.begin()) {
2736 if (I->isDebugInstr())
2738 if (I->getOpcode() != X86::JMP_1 &&
2739 X86::getCondFromBranchOpc(I->getOpcode()) == X86::COND_INVALID)
2741 // Remove the branch.
2742 I->eraseFromParent();
2750 unsigned X86InstrInfo::insertBranch(MachineBasicBlock &MBB,
2751 MachineBasicBlock *TBB,
2752 MachineBasicBlock *FBB,
2753 ArrayRef<MachineOperand> Cond,
2755 int *BytesAdded) const {
2756 // Shouldn't be a fall through.
2757 assert(TBB && "insertBranch must not be told to insert a fallthrough");
2758 assert((Cond.size() == 1 || Cond.size() == 0) &&
2759 "X86 branch conditions have one component!");
2760 assert(!BytesAdded && "code size not handled");
2763 // Unconditional branch?
2764 assert(!FBB && "Unconditional branch with multiple successors!");
2765 BuildMI(&MBB, DL, get(X86::JMP_1)).addMBB(TBB);
2769 // If FBB is null, it is implied to be a fall-through block.
2770 bool FallThru = FBB == nullptr;
2772 // Conditional branch.
2774 X86::CondCode CC = (X86::CondCode)Cond[0].getImm();
2776 case X86::COND_NE_OR_P:
2777 // Synthesize NE_OR_P with two branches.
2778 BuildMI(&MBB, DL, get(X86::JNE_1)).addMBB(TBB);
2780 BuildMI(&MBB, DL, get(X86::JP_1)).addMBB(TBB);
2783 case X86::COND_E_AND_NP:
2784 // Use the next block of MBB as FBB if it is null.
2785 if (FBB == nullptr) {
2786 FBB = getFallThroughMBB(&MBB, TBB);
2787 assert(FBB && "MBB cannot be the last block in function when the false "
2788 "body is a fall-through.");
2790 // Synthesize COND_E_AND_NP with two branches.
2791 BuildMI(&MBB, DL, get(X86::JNE_1)).addMBB(FBB);
2793 BuildMI(&MBB, DL, get(X86::JNP_1)).addMBB(TBB);
2797 unsigned Opc = GetCondBranchFromCond(CC);
2798 BuildMI(&MBB, DL, get(Opc)).addMBB(TBB);
2803 // Two-way Conditional branch. Insert the second branch.
2804 BuildMI(&MBB, DL, get(X86::JMP_1)).addMBB(FBB);
2811 canInsertSelect(const MachineBasicBlock &MBB,
2812 ArrayRef<MachineOperand> Cond,
2813 unsigned TrueReg, unsigned FalseReg,
2814 int &CondCycles, int &TrueCycles, int &FalseCycles) const {
2815 // Not all subtargets have cmov instructions.
2816 if (!Subtarget.hasCMov())
2818 if (Cond.size() != 1)
2820 // We cannot do the composite conditions, at least not in SSA form.
2821 if ((X86::CondCode)Cond[0].getImm() > X86::COND_S)
2824 // Check register classes.
2825 const MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo();
2826 const TargetRegisterClass *RC =
2827 RI.getCommonSubClass(MRI.getRegClass(TrueReg), MRI.getRegClass(FalseReg));
2831 // We have cmov instructions for 16, 32, and 64 bit general purpose registers.
2832 if (X86::GR16RegClass.hasSubClassEq(RC) ||
2833 X86::GR32RegClass.hasSubClassEq(RC) ||
2834 X86::GR64RegClass.hasSubClassEq(RC)) {
2835 // This latency applies to Pentium M, Merom, Wolfdale, Nehalem, and Sandy
2836 // Bridge. Probably Ivy Bridge as well.
2843 // Can't do vectors.
2847 void X86InstrInfo::insertSelect(MachineBasicBlock &MBB,
2848 MachineBasicBlock::iterator I,
2849 const DebugLoc &DL, unsigned DstReg,
2850 ArrayRef<MachineOperand> Cond, unsigned TrueReg,
2851 unsigned FalseReg) const {
2852 MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo();
2853 const TargetRegisterInfo &TRI = *MRI.getTargetRegisterInfo();
2854 const TargetRegisterClass &RC = *MRI.getRegClass(DstReg);
2855 assert(Cond.size() == 1 && "Invalid Cond array");
2856 unsigned Opc = getCMovFromCond((X86::CondCode)Cond[0].getImm(),
2857 TRI.getRegSizeInBits(RC) / 8,
2858 false /*HasMemoryOperand*/);
2859 BuildMI(MBB, I, DL, get(Opc), DstReg).addReg(FalseReg).addReg(TrueReg);
2862 /// Test if the given register is a physical h register.
2863 static bool isHReg(unsigned Reg) {
2864 return X86::GR8_ABCD_HRegClass.contains(Reg);
2867 // Try and copy between VR128/VR64 and GR64 registers.
2868 static unsigned CopyToFromAsymmetricReg(unsigned DestReg, unsigned SrcReg,
2869 const X86Subtarget &Subtarget) {
2870 bool HasAVX = Subtarget.hasAVX();
2871 bool HasAVX512 = Subtarget.hasAVX512();
2873 // SrcReg(MaskReg) -> DestReg(GR64)
2874 // SrcReg(MaskReg) -> DestReg(GR32)
2876 // All KMASK RegClasses hold the same k registers, can be tested against anyone.
2877 if (X86::VK16RegClass.contains(SrcReg)) {
2878 if (X86::GR64RegClass.contains(DestReg)) {
2879 assert(Subtarget.hasBWI());
2880 return X86::KMOVQrk;
2882 if (X86::GR32RegClass.contains(DestReg))
2883 return Subtarget.hasBWI() ? X86::KMOVDrk : X86::KMOVWrk;
2886 // SrcReg(GR64) -> DestReg(MaskReg)
2887 // SrcReg(GR32) -> DestReg(MaskReg)
2889 // All KMASK RegClasses hold the same k registers, can be tested against anyone.
2890 if (X86::VK16RegClass.contains(DestReg)) {
2891 if (X86::GR64RegClass.contains(SrcReg)) {
2892 assert(Subtarget.hasBWI());
2893 return X86::KMOVQkr;
2895 if (X86::GR32RegClass.contains(SrcReg))
2896 return Subtarget.hasBWI() ? X86::KMOVDkr : X86::KMOVWkr;
2900 // SrcReg(VR128) -> DestReg(GR64)
2901 // SrcReg(VR64) -> DestReg(GR64)
2902 // SrcReg(GR64) -> DestReg(VR128)
2903 // SrcReg(GR64) -> DestReg(VR64)
2905 if (X86::GR64RegClass.contains(DestReg)) {
2906 if (X86::VR128XRegClass.contains(SrcReg))
2907 // Copy from a VR128 register to a GR64 register.
2908 return HasAVX512 ? X86::VMOVPQIto64Zrr :
2909 HasAVX ? X86::VMOVPQIto64rr :
2911 if (X86::VR64RegClass.contains(SrcReg))
2912 // Copy from a VR64 register to a GR64 register.
2913 return X86::MMX_MOVD64from64rr;
2914 } else if (X86::GR64RegClass.contains(SrcReg)) {
2915 // Copy from a GR64 register to a VR128 register.
2916 if (X86::VR128XRegClass.contains(DestReg))
2917 return HasAVX512 ? X86::VMOV64toPQIZrr :
2918 HasAVX ? X86::VMOV64toPQIrr :
2920 // Copy from a GR64 register to a VR64 register.
2921 if (X86::VR64RegClass.contains(DestReg))
2922 return X86::MMX_MOVD64to64rr;
2925 // SrcReg(FR32) -> DestReg(GR32)
2926 // SrcReg(GR32) -> DestReg(FR32)
2928 if (X86::GR32RegClass.contains(DestReg) &&
2929 X86::FR32XRegClass.contains(SrcReg))
2930 // Copy from a FR32 register to a GR32 register.
2931 return HasAVX512 ? X86::VMOVSS2DIZrr :
2932 HasAVX ? X86::VMOVSS2DIrr :
2935 if (X86::FR32XRegClass.contains(DestReg) &&
2936 X86::GR32RegClass.contains(SrcReg))
2937 // Copy from a GR32 register to a FR32 register.
2938 return HasAVX512 ? X86::VMOVDI2SSZrr :
2939 HasAVX ? X86::VMOVDI2SSrr :
2944 void X86InstrInfo::copyPhysReg(MachineBasicBlock &MBB,
2945 MachineBasicBlock::iterator MI,
2946 const DebugLoc &DL, unsigned DestReg,
2947 unsigned SrcReg, bool KillSrc) const {
2948 // First deal with the normal symmetric copies.
2949 bool HasAVX = Subtarget.hasAVX();
2950 bool HasVLX = Subtarget.hasVLX();
2952 if (X86::GR64RegClass.contains(DestReg, SrcReg))
2954 else if (X86::GR32RegClass.contains(DestReg, SrcReg))
2956 else if (X86::GR16RegClass.contains(DestReg, SrcReg))
2958 else if (X86::GR8RegClass.contains(DestReg, SrcReg)) {
2959 // Copying to or from a physical H register on x86-64 requires a NOREX
2960 // move. Otherwise use a normal move.
2961 if ((isHReg(DestReg) || isHReg(SrcReg)) &&
2962 Subtarget.is64Bit()) {
2963 Opc = X86::MOV8rr_NOREX;
2964 // Both operands must be encodable without an REX prefix.
2965 assert(X86::GR8_NOREXRegClass.contains(SrcReg, DestReg) &&
2966 "8-bit H register can not be copied outside GR8_NOREX");
2970 else if (X86::VR64RegClass.contains(DestReg, SrcReg))
2971 Opc = X86::MMX_MOVQ64rr;
2972 else if (X86::VR128XRegClass.contains(DestReg, SrcReg)) {
2974 Opc = X86::VMOVAPSZ128rr;
2975 else if (X86::VR128RegClass.contains(DestReg, SrcReg))
2976 Opc = HasAVX ? X86::VMOVAPSrr : X86::MOVAPSrr;
2978 // If this an extended register and we don't have VLX we need to use a
2980 Opc = X86::VMOVAPSZrr;
2981 const TargetRegisterInfo *TRI = &getRegisterInfo();
2982 DestReg = TRI->getMatchingSuperReg(DestReg, X86::sub_xmm,
2983 &X86::VR512RegClass);
2984 SrcReg = TRI->getMatchingSuperReg(SrcReg, X86::sub_xmm,
2985 &X86::VR512RegClass);
2987 } else if (X86::VR256XRegClass.contains(DestReg, SrcReg)) {
2989 Opc = X86::VMOVAPSZ256rr;
2990 else if (X86::VR256RegClass.contains(DestReg, SrcReg))
2991 Opc = X86::VMOVAPSYrr;
2993 // If this an extended register and we don't have VLX we need to use a
2995 Opc = X86::VMOVAPSZrr;
2996 const TargetRegisterInfo *TRI = &getRegisterInfo();
2997 DestReg = TRI->getMatchingSuperReg(DestReg, X86::sub_ymm,
2998 &X86::VR512RegClass);
2999 SrcReg = TRI->getMatchingSuperReg(SrcReg, X86::sub_ymm,
3000 &X86::VR512RegClass);
3002 } else if (X86::VR512RegClass.contains(DestReg, SrcReg))
3003 Opc = X86::VMOVAPSZrr;
3004 // All KMASK RegClasses hold the same k registers, can be tested against anyone.
3005 else if (X86::VK16RegClass.contains(DestReg, SrcReg))
3006 Opc = Subtarget.hasBWI() ? X86::KMOVQkk : X86::KMOVWkk;
3008 Opc = CopyToFromAsymmetricReg(DestReg, SrcReg, Subtarget);
3011 BuildMI(MBB, MI, DL, get(Opc), DestReg)
3012 .addReg(SrcReg, getKillRegState(KillSrc));
3016 if (SrcReg == X86::EFLAGS || DestReg == X86::EFLAGS) {
3017 // FIXME: We use a fatal error here because historically LLVM has tried
3018 // lower some of these physreg copies and we want to ensure we get
3019 // reasonable bug reports if someone encounters a case no other testing
3020 // found. This path should be removed after the LLVM 7 release.
3021 report_fatal_error("Unable to copy EFLAGS physical register!");
3024 LLVM_DEBUG(dbgs() << "Cannot copy " << RI.getName(SrcReg) << " to "
3025 << RI.getName(DestReg) << '\n');
3026 report_fatal_error("Cannot emit physreg copy instruction");
3029 bool X86InstrInfo::isCopyInstrImpl(const MachineInstr &MI,
3030 const MachineOperand *&Src,
3031 const MachineOperand *&Dest) const {
3032 if (MI.isMoveReg()) {
3033 Dest = &MI.getOperand(0);
3034 Src = &MI.getOperand(1);
3040 static unsigned getLoadStoreRegOpcode(unsigned Reg,
3041 const TargetRegisterClass *RC,
3042 bool isStackAligned,
3043 const X86Subtarget &STI,
3045 bool HasAVX = STI.hasAVX();
3046 bool HasAVX512 = STI.hasAVX512();
3047 bool HasVLX = STI.hasVLX();
3049 switch (STI.getRegisterInfo()->getSpillSize(*RC)) {
3051 llvm_unreachable("Unknown spill size");
3053 assert(X86::GR8RegClass.hasSubClassEq(RC) && "Unknown 1-byte regclass");
3055 // Copying to or from a physical H register on x86-64 requires a NOREX
3056 // move. Otherwise use a normal move.
3057 if (isHReg(Reg) || X86::GR8_ABCD_HRegClass.hasSubClassEq(RC))
3058 return load ? X86::MOV8rm_NOREX : X86::MOV8mr_NOREX;
3059 return load ? X86::MOV8rm : X86::MOV8mr;
3061 if (X86::VK16RegClass.hasSubClassEq(RC))
3062 return load ? X86::KMOVWkm : X86::KMOVWmk;
3063 assert(X86::GR16RegClass.hasSubClassEq(RC) && "Unknown 2-byte regclass");
3064 return load ? X86::MOV16rm : X86::MOV16mr;
3066 if (X86::GR32RegClass.hasSubClassEq(RC))
3067 return load ? X86::MOV32rm : X86::MOV32mr;
3068 if (X86::FR32XRegClass.hasSubClassEq(RC))
3070 (HasAVX512 ? X86::VMOVSSZrm : HasAVX ? X86::VMOVSSrm : X86::MOVSSrm) :
3071 (HasAVX512 ? X86::VMOVSSZmr : HasAVX ? X86::VMOVSSmr : X86::MOVSSmr);
3072 if (X86::RFP32RegClass.hasSubClassEq(RC))
3073 return load ? X86::LD_Fp32m : X86::ST_Fp32m;
3074 if (X86::VK32RegClass.hasSubClassEq(RC)) {
3075 assert(STI.hasBWI() && "KMOVD requires BWI");
3076 return load ? X86::KMOVDkm : X86::KMOVDmk;
3078 llvm_unreachable("Unknown 4-byte regclass");
3080 if (X86::GR64RegClass.hasSubClassEq(RC))
3081 return load ? X86::MOV64rm : X86::MOV64mr;
3082 if (X86::FR64XRegClass.hasSubClassEq(RC))
3084 (HasAVX512 ? X86::VMOVSDZrm : HasAVX ? X86::VMOVSDrm : X86::MOVSDrm) :
3085 (HasAVX512 ? X86::VMOVSDZmr : HasAVX ? X86::VMOVSDmr : X86::MOVSDmr);
3086 if (X86::VR64RegClass.hasSubClassEq(RC))
3087 return load ? X86::MMX_MOVQ64rm : X86::MMX_MOVQ64mr;
3088 if (X86::RFP64RegClass.hasSubClassEq(RC))
3089 return load ? X86::LD_Fp64m : X86::ST_Fp64m;
3090 if (X86::VK64RegClass.hasSubClassEq(RC)) {
3091 assert(STI.hasBWI() && "KMOVQ requires BWI");
3092 return load ? X86::KMOVQkm : X86::KMOVQmk;
3094 llvm_unreachable("Unknown 8-byte regclass");
3096 assert(X86::RFP80RegClass.hasSubClassEq(RC) && "Unknown 10-byte regclass");
3097 return load ? X86::LD_Fp80m : X86::ST_FpP80m;
3099 if (X86::VR128XRegClass.hasSubClassEq(RC)) {
3100 // If stack is realigned we can use aligned stores.
3103 (HasVLX ? X86::VMOVAPSZ128rm :
3104 HasAVX512 ? X86::VMOVAPSZ128rm_NOVLX :
3105 HasAVX ? X86::VMOVAPSrm :
3107 (HasVLX ? X86::VMOVAPSZ128mr :
3108 HasAVX512 ? X86::VMOVAPSZ128mr_NOVLX :
3109 HasAVX ? X86::VMOVAPSmr :
3113 (HasVLX ? X86::VMOVUPSZ128rm :
3114 HasAVX512 ? X86::VMOVUPSZ128rm_NOVLX :
3115 HasAVX ? X86::VMOVUPSrm :
3117 (HasVLX ? X86::VMOVUPSZ128mr :
3118 HasAVX512 ? X86::VMOVUPSZ128mr_NOVLX :
3119 HasAVX ? X86::VMOVUPSmr :
3122 if (X86::BNDRRegClass.hasSubClassEq(RC)) {
3124 return load ? X86::BNDMOV64rm : X86::BNDMOV64mr;
3126 return load ? X86::BNDMOV32rm : X86::BNDMOV32mr;
3128 llvm_unreachable("Unknown 16-byte regclass");
3131 assert(X86::VR256XRegClass.hasSubClassEq(RC) && "Unknown 32-byte regclass");
3132 // If stack is realigned we can use aligned stores.
3135 (HasVLX ? X86::VMOVAPSZ256rm :
3136 HasAVX512 ? X86::VMOVAPSZ256rm_NOVLX :
3138 (HasVLX ? X86::VMOVAPSZ256mr :
3139 HasAVX512 ? X86::VMOVAPSZ256mr_NOVLX :
3143 (HasVLX ? X86::VMOVUPSZ256rm :
3144 HasAVX512 ? X86::VMOVUPSZ256rm_NOVLX :
3146 (HasVLX ? X86::VMOVUPSZ256mr :
3147 HasAVX512 ? X86::VMOVUPSZ256mr_NOVLX :
3150 assert(X86::VR512RegClass.hasSubClassEq(RC) && "Unknown 64-byte regclass");
3151 assert(STI.hasAVX512() && "Using 512-bit register requires AVX512");
3153 return load ? X86::VMOVAPSZrm : X86::VMOVAPSZmr;
3155 return load ? X86::VMOVUPSZrm : X86::VMOVUPSZmr;
3159 bool X86InstrInfo::getMemOperandWithOffset(
3160 MachineInstr &MemOp, MachineOperand *&BaseOp, int64_t &Offset,
3161 const TargetRegisterInfo *TRI) const {
3162 const MCInstrDesc &Desc = MemOp.getDesc();
3163 int MemRefBegin = X86II::getMemoryOperandNo(Desc.TSFlags);
3164 if (MemRefBegin < 0)
3167 MemRefBegin += X86II::getOperandBias(Desc);
3169 BaseOp = &MemOp.getOperand(MemRefBegin + X86::AddrBaseReg);
3170 if (!BaseOp->isReg()) // Can be an MO_FrameIndex
3173 if (MemOp.getOperand(MemRefBegin + X86::AddrScaleAmt).getImm() != 1)
3176 if (MemOp.getOperand(MemRefBegin + X86::AddrIndexReg).getReg() !=
3180 const MachineOperand &DispMO = MemOp.getOperand(MemRefBegin + X86::AddrDisp);
3182 // Displacement can be symbolic
3183 if (!DispMO.isImm())
3186 Offset = DispMO.getImm();
3188 assert(BaseOp->isReg() && "getMemOperandWithOffset only supports base "
3189 "operands of type register.");
3193 static unsigned getStoreRegOpcode(unsigned SrcReg,
3194 const TargetRegisterClass *RC,
3195 bool isStackAligned,
3196 const X86Subtarget &STI) {
3197 return getLoadStoreRegOpcode(SrcReg, RC, isStackAligned, STI, false);
3201 static unsigned getLoadRegOpcode(unsigned DestReg,
3202 const TargetRegisterClass *RC,
3203 bool isStackAligned,
3204 const X86Subtarget &STI) {
3205 return getLoadStoreRegOpcode(DestReg, RC, isStackAligned, STI, true);
3208 void X86InstrInfo::storeRegToStackSlot(MachineBasicBlock &MBB,
3209 MachineBasicBlock::iterator MI,
3210 unsigned SrcReg, bool isKill, int FrameIdx,
3211 const TargetRegisterClass *RC,
3212 const TargetRegisterInfo *TRI) const {
3213 const MachineFunction &MF = *MBB.getParent();
3214 assert(MF.getFrameInfo().getObjectSize(FrameIdx) >= TRI->getSpillSize(*RC) &&
3215 "Stack slot too small for store");
3216 unsigned Alignment = std::max<uint32_t>(TRI->getSpillSize(*RC), 16);
3218 (Subtarget.getFrameLowering()->getStackAlignment() >= Alignment) ||
3219 RI.canRealignStack(MF);
3220 unsigned Opc = getStoreRegOpcode(SrcReg, RC, isAligned, Subtarget);
3221 addFrameReference(BuildMI(MBB, MI, DebugLoc(), get(Opc)), FrameIdx)
3222 .addReg(SrcReg, getKillRegState(isKill));
3225 void X86InstrInfo::storeRegToAddr(
3226 MachineFunction &MF, unsigned SrcReg, bool isKill,
3227 SmallVectorImpl<MachineOperand> &Addr, const TargetRegisterClass *RC,
3228 ArrayRef<MachineMemOperand *> MMOs,
3229 SmallVectorImpl<MachineInstr *> &NewMIs) const {
3230 const TargetRegisterInfo &TRI = *MF.getSubtarget().getRegisterInfo();
3231 unsigned Alignment = std::max<uint32_t>(TRI.getSpillSize(*RC), 16);
3232 bool isAligned = !MMOs.empty() && MMOs.front()->getAlignment() >= Alignment;
3233 unsigned Opc = getStoreRegOpcode(SrcReg, RC, isAligned, Subtarget);
3235 MachineInstrBuilder MIB = BuildMI(MF, DL, get(Opc));
3236 for (unsigned i = 0, e = Addr.size(); i != e; ++i)
3238 MIB.addReg(SrcReg, getKillRegState(isKill));
3239 MIB.setMemRefs(MMOs);
3240 NewMIs.push_back(MIB);
3244 void X86InstrInfo::loadRegFromStackSlot(MachineBasicBlock &MBB,
3245 MachineBasicBlock::iterator MI,
3246 unsigned DestReg, int FrameIdx,
3247 const TargetRegisterClass *RC,
3248 const TargetRegisterInfo *TRI) const {
3249 const MachineFunction &MF = *MBB.getParent();
3250 unsigned Alignment = std::max<uint32_t>(TRI->getSpillSize(*RC), 16);
3252 (Subtarget.getFrameLowering()->getStackAlignment() >= Alignment) ||
3253 RI.canRealignStack(MF);
3254 unsigned Opc = getLoadRegOpcode(DestReg, RC, isAligned, Subtarget);
3255 addFrameReference(BuildMI(MBB, MI, DebugLoc(), get(Opc), DestReg), FrameIdx);
3258 void X86InstrInfo::loadRegFromAddr(
3259 MachineFunction &MF, unsigned DestReg,
3260 SmallVectorImpl<MachineOperand> &Addr, const TargetRegisterClass *RC,
3261 ArrayRef<MachineMemOperand *> MMOs,
3262 SmallVectorImpl<MachineInstr *> &NewMIs) const {
3263 const TargetRegisterInfo &TRI = *MF.getSubtarget().getRegisterInfo();
3264 unsigned Alignment = std::max<uint32_t>(TRI.getSpillSize(*RC), 16);
3265 bool isAligned = !MMOs.empty() && MMOs.front()->getAlignment() >= Alignment;
3266 unsigned Opc = getLoadRegOpcode(DestReg, RC, isAligned, Subtarget);
3268 MachineInstrBuilder MIB = BuildMI(MF, DL, get(Opc), DestReg);
3269 for (unsigned i = 0, e = Addr.size(); i != e; ++i)
3271 MIB.setMemRefs(MMOs);
3272 NewMIs.push_back(MIB);
3275 bool X86InstrInfo::analyzeCompare(const MachineInstr &MI, unsigned &SrcReg,
3276 unsigned &SrcReg2, int &CmpMask,
3277 int &CmpValue) const {
3278 switch (MI.getOpcode()) {
3280 case X86::CMP64ri32:
3287 SrcReg = MI.getOperand(0).getReg();
3289 if (MI.getOperand(1).isImm()) {
3291 CmpValue = MI.getOperand(1).getImm();
3293 CmpMask = CmpValue = 0;
3296 // A SUB can be used to perform comparison.
3301 SrcReg = MI.getOperand(1).getReg();
3310 SrcReg = MI.getOperand(1).getReg();
3311 SrcReg2 = MI.getOperand(2).getReg();
3315 case X86::SUB64ri32:
3322 SrcReg = MI.getOperand(1).getReg();
3324 if (MI.getOperand(2).isImm()) {
3326 CmpValue = MI.getOperand(2).getImm();
3328 CmpMask = CmpValue = 0;
3335 SrcReg = MI.getOperand(0).getReg();
3336 SrcReg2 = MI.getOperand(1).getReg();
3344 SrcReg = MI.getOperand(0).getReg();
3345 if (MI.getOperand(1).getReg() != SrcReg)
3347 // Compare against zero.
3356 /// Check whether the first instruction, whose only
3357 /// purpose is to update flags, can be made redundant.
3358 /// CMPrr can be made redundant by SUBrr if the operands are the same.
3359 /// This function can be extended later on.
3360 /// SrcReg, SrcRegs: register operands for FlagI.
3361 /// ImmValue: immediate for FlagI if it takes an immediate.
3362 inline static bool isRedundantFlagInstr(const MachineInstr &FlagI,
3363 unsigned SrcReg, unsigned SrcReg2,
3364 int ImmMask, int ImmValue,
3365 const MachineInstr &OI) {
3366 if (((FlagI.getOpcode() == X86::CMP64rr && OI.getOpcode() == X86::SUB64rr) ||
3367 (FlagI.getOpcode() == X86::CMP32rr && OI.getOpcode() == X86::SUB32rr) ||
3368 (FlagI.getOpcode() == X86::CMP16rr && OI.getOpcode() == X86::SUB16rr) ||
3369 (FlagI.getOpcode() == X86::CMP8rr && OI.getOpcode() == X86::SUB8rr)) &&
3370 ((OI.getOperand(1).getReg() == SrcReg &&
3371 OI.getOperand(2).getReg() == SrcReg2) ||
3372 (OI.getOperand(1).getReg() == SrcReg2 &&
3373 OI.getOperand(2).getReg() == SrcReg)))
3377 ((FlagI.getOpcode() == X86::CMP64ri32 &&
3378 OI.getOpcode() == X86::SUB64ri32) ||
3379 (FlagI.getOpcode() == X86::CMP64ri8 &&
3380 OI.getOpcode() == X86::SUB64ri8) ||
3381 (FlagI.getOpcode() == X86::CMP32ri && OI.getOpcode() == X86::SUB32ri) ||
3382 (FlagI.getOpcode() == X86::CMP32ri8 &&
3383 OI.getOpcode() == X86::SUB32ri8) ||
3384 (FlagI.getOpcode() == X86::CMP16ri && OI.getOpcode() == X86::SUB16ri) ||
3385 (FlagI.getOpcode() == X86::CMP16ri8 &&
3386 OI.getOpcode() == X86::SUB16ri8) ||
3387 (FlagI.getOpcode() == X86::CMP8ri && OI.getOpcode() == X86::SUB8ri)) &&
3388 OI.getOperand(1).getReg() == SrcReg &&
3389 OI.getOperand(2).getImm() == ImmValue)
3394 /// Check whether the definition can be converted
3395 /// to remove a comparison against zero.
3396 inline static bool isDefConvertible(const MachineInstr &MI, bool &NoSignFlag) {
3399 switch (MI.getOpcode()) {
3400 default: return false;
3402 // The shift instructions only modify ZF if their shift count is non-zero.
3403 // N.B.: The processor truncates the shift count depending on the encoding.
3404 case X86::SAR8ri: case X86::SAR16ri: case X86::SAR32ri:case X86::SAR64ri:
3405 case X86::SHR8ri: case X86::SHR16ri: case X86::SHR32ri:case X86::SHR64ri:
3406 return getTruncatedShiftCount(MI, 2) != 0;
3408 // Some left shift instructions can be turned into LEA instructions but only
3409 // if their flags aren't used. Avoid transforming such instructions.
3410 case X86::SHL8ri: case X86::SHL16ri: case X86::SHL32ri:case X86::SHL64ri:{
3411 unsigned ShAmt = getTruncatedShiftCount(MI, 2);
3412 if (isTruncatedShiftCountForLEA(ShAmt)) return false;
3416 case X86::SHRD16rri8:case X86::SHRD32rri8:case X86::SHRD64rri8:
3417 case X86::SHLD16rri8:case X86::SHLD32rri8:case X86::SHLD64rri8:
3418 return getTruncatedShiftCount(MI, 3) != 0;
3420 case X86::SUB64ri32: case X86::SUB64ri8: case X86::SUB32ri:
3421 case X86::SUB32ri8: case X86::SUB16ri: case X86::SUB16ri8:
3422 case X86::SUB8ri: case X86::SUB64rr: case X86::SUB32rr:
3423 case X86::SUB16rr: case X86::SUB8rr: case X86::SUB64rm:
3424 case X86::SUB32rm: case X86::SUB16rm: case X86::SUB8rm:
3425 case X86::DEC64r: case X86::DEC32r: case X86::DEC16r: case X86::DEC8r:
3426 case X86::ADD64ri32: case X86::ADD64ri8: case X86::ADD32ri:
3427 case X86::ADD32ri8: case X86::ADD16ri: case X86::ADD16ri8:
3428 case X86::ADD8ri: case X86::ADD64rr: case X86::ADD32rr:
3429 case X86::ADD16rr: case X86::ADD8rr: case X86::ADD64rm:
3430 case X86::ADD32rm: case X86::ADD16rm: case X86::ADD8rm:
3431 case X86::INC64r: case X86::INC32r: case X86::INC16r: case X86::INC8r:
3432 case X86::AND64ri32: case X86::AND64ri8: case X86::AND32ri:
3433 case X86::AND32ri8: case X86::AND16ri: case X86::AND16ri8:
3434 case X86::AND8ri: case X86::AND64rr: case X86::AND32rr:
3435 case X86::AND16rr: case X86::AND8rr: case X86::AND64rm:
3436 case X86::AND32rm: case X86::AND16rm: case X86::AND8rm:
3437 case X86::XOR64ri32: case X86::XOR64ri8: case X86::XOR32ri:
3438 case X86::XOR32ri8: case X86::XOR16ri: case X86::XOR16ri8:
3439 case X86::XOR8ri: case X86::XOR64rr: case X86::XOR32rr:
3440 case X86::XOR16rr: case X86::XOR8rr: case X86::XOR64rm:
3441 case X86::XOR32rm: case X86::XOR16rm: case X86::XOR8rm:
3442 case X86::OR64ri32: case X86::OR64ri8: case X86::OR32ri:
3443 case X86::OR32ri8: case X86::OR16ri: case X86::OR16ri8:
3444 case X86::OR8ri: case X86::OR64rr: case X86::OR32rr:
3445 case X86::OR16rr: case X86::OR8rr: case X86::OR64rm:
3446 case X86::OR32rm: case X86::OR16rm: case X86::OR8rm:
3447 case X86::ADC64ri32: case X86::ADC64ri8: case X86::ADC32ri:
3448 case X86::ADC32ri8: case X86::ADC16ri: case X86::ADC16ri8:
3449 case X86::ADC8ri: case X86::ADC64rr: case X86::ADC32rr:
3450 case X86::ADC16rr: case X86::ADC8rr: case X86::ADC64rm:
3451 case X86::ADC32rm: case X86::ADC16rm: case X86::ADC8rm:
3452 case X86::SBB64ri32: case X86::SBB64ri8: case X86::SBB32ri:
3453 case X86::SBB32ri8: case X86::SBB16ri: case X86::SBB16ri8:
3454 case X86::SBB8ri: case X86::SBB64rr: case X86::SBB32rr:
3455 case X86::SBB16rr: case X86::SBB8rr: case X86::SBB64rm:
3456 case X86::SBB32rm: case X86::SBB16rm: case X86::SBB8rm:
3457 case X86::NEG8r: case X86::NEG16r: case X86::NEG32r: case X86::NEG64r:
3458 case X86::SAR8r1: case X86::SAR16r1: case X86::SAR32r1:case X86::SAR64r1:
3459 case X86::SHR8r1: case X86::SHR16r1: case X86::SHR32r1:case X86::SHR64r1:
3460 case X86::SHL8r1: case X86::SHL16r1: case X86::SHL32r1:case X86::SHL64r1:
3461 case X86::ANDN32rr: case X86::ANDN32rm:
3462 case X86::ANDN64rr: case X86::ANDN64rm:
3463 case X86::BLSI32rr: case X86::BLSI32rm:
3464 case X86::BLSI64rr: case X86::BLSI64rm:
3465 case X86::BLSMSK32rr:case X86::BLSMSK32rm:
3466 case X86::BLSMSK64rr:case X86::BLSMSK64rm:
3467 case X86::BLSR32rr: case X86::BLSR32rm:
3468 case X86::BLSR64rr: case X86::BLSR64rm:
3469 case X86::BZHI32rr: case X86::BZHI32rm:
3470 case X86::BZHI64rr: case X86::BZHI64rm:
3471 case X86::LZCNT16rr: case X86::LZCNT16rm:
3472 case X86::LZCNT32rr: case X86::LZCNT32rm:
3473 case X86::LZCNT64rr: case X86::LZCNT64rm:
3474 case X86::POPCNT16rr:case X86::POPCNT16rm:
3475 case X86::POPCNT32rr:case X86::POPCNT32rm:
3476 case X86::POPCNT64rr:case X86::POPCNT64rm:
3477 case X86::TZCNT16rr: case X86::TZCNT16rm:
3478 case X86::TZCNT32rr: case X86::TZCNT32rm:
3479 case X86::TZCNT64rr: case X86::TZCNT64rm:
3480 case X86::BLCFILL32rr: case X86::BLCFILL32rm:
3481 case X86::BLCFILL64rr: case X86::BLCFILL64rm:
3482 case X86::BLCI32rr: case X86::BLCI32rm:
3483 case X86::BLCI64rr: case X86::BLCI64rm:
3484 case X86::BLCIC32rr: case X86::BLCIC32rm:
3485 case X86::BLCIC64rr: case X86::BLCIC64rm:
3486 case X86::BLCMSK32rr: case X86::BLCMSK32rm:
3487 case X86::BLCMSK64rr: case X86::BLCMSK64rm:
3488 case X86::BLCS32rr: case X86::BLCS32rm:
3489 case X86::BLCS64rr: case X86::BLCS64rm:
3490 case X86::BLSFILL32rr: case X86::BLSFILL32rm:
3491 case X86::BLSFILL64rr: case X86::BLSFILL64rm:
3492 case X86::BLSIC32rr: case X86::BLSIC32rm:
3493 case X86::BLSIC64rr: case X86::BLSIC64rm:
3494 case X86::T1MSKC32rr: case X86::T1MSKC32rm:
3495 case X86::T1MSKC64rr: case X86::T1MSKC64rm:
3496 case X86::TZMSK32rr: case X86::TZMSK32rm:
3497 case X86::TZMSK64rr: case X86::TZMSK64rm:
3499 case X86::BEXTR32rr: case X86::BEXTR64rr:
3500 case X86::BEXTR32rm: case X86::BEXTR64rm:
3501 case X86::BEXTRI32ri: case X86::BEXTRI32mi:
3502 case X86::BEXTRI64ri: case X86::BEXTRI64mi:
3503 // BEXTR doesn't update the sign flag so we can't use it.
3509 /// Check whether the use can be converted to remove a comparison against zero.
3510 static X86::CondCode isUseDefConvertible(const MachineInstr &MI) {
3511 switch (MI.getOpcode()) {
3512 default: return X86::COND_INVALID;
3513 case X86::LZCNT16rr: case X86::LZCNT16rm:
3514 case X86::LZCNT32rr: case X86::LZCNT32rm:
3515 case X86::LZCNT64rr: case X86::LZCNT64rm:
3517 case X86::POPCNT16rr:case X86::POPCNT16rm:
3518 case X86::POPCNT32rr:case X86::POPCNT32rm:
3519 case X86::POPCNT64rr:case X86::POPCNT64rm:
3521 case X86::TZCNT16rr: case X86::TZCNT16rm:
3522 case X86::TZCNT32rr: case X86::TZCNT32rm:
3523 case X86::TZCNT64rr: case X86::TZCNT64rm:
3525 case X86::BSF16rr: case X86::BSF16rm:
3526 case X86::BSF32rr: case X86::BSF32rm:
3527 case X86::BSF64rr: case X86::BSF64rm:
3528 case X86::BSR16rr: case X86::BSR16rm:
3529 case X86::BSR32rr: case X86::BSR32rm:
3530 case X86::BSR64rr: case X86::BSR64rm:
3535 /// Check if there exists an earlier instruction that
3536 /// operates on the same source operands and sets flags in the same way as
3537 /// Compare; remove Compare if possible.
3538 bool X86InstrInfo::optimizeCompareInstr(MachineInstr &CmpInstr, unsigned SrcReg,
3539 unsigned SrcReg2, int CmpMask,
3541 const MachineRegisterInfo *MRI) const {
3542 // Check whether we can replace SUB with CMP.
3543 unsigned NewOpcode = 0;
3544 switch (CmpInstr.getOpcode()) {
3546 case X86::SUB64ri32:
3561 if (!MRI->use_nodbg_empty(CmpInstr.getOperand(0).getReg()))
3563 // There is no use of the destination register, we can replace SUB with CMP.
3564 switch (CmpInstr.getOpcode()) {
3565 default: llvm_unreachable("Unreachable!");
3566 case X86::SUB64rm: NewOpcode = X86::CMP64rm; break;
3567 case X86::SUB32rm: NewOpcode = X86::CMP32rm; break;
3568 case X86::SUB16rm: NewOpcode = X86::CMP16rm; break;
3569 case X86::SUB8rm: NewOpcode = X86::CMP8rm; break;
3570 case X86::SUB64rr: NewOpcode = X86::CMP64rr; break;
3571 case X86::SUB32rr: NewOpcode = X86::CMP32rr; break;
3572 case X86::SUB16rr: NewOpcode = X86::CMP16rr; break;
3573 case X86::SUB8rr: NewOpcode = X86::CMP8rr; break;
3574 case X86::SUB64ri32: NewOpcode = X86::CMP64ri32; break;
3575 case X86::SUB64ri8: NewOpcode = X86::CMP64ri8; break;
3576 case X86::SUB32ri: NewOpcode = X86::CMP32ri; break;
3577 case X86::SUB32ri8: NewOpcode = X86::CMP32ri8; break;
3578 case X86::SUB16ri: NewOpcode = X86::CMP16ri; break;
3579 case X86::SUB16ri8: NewOpcode = X86::CMP16ri8; break;
3580 case X86::SUB8ri: NewOpcode = X86::CMP8ri; break;
3582 CmpInstr.setDesc(get(NewOpcode));
3583 CmpInstr.RemoveOperand(0);
3584 // Fall through to optimize Cmp if Cmp is CMPrr or CMPri.
3585 if (NewOpcode == X86::CMP64rm || NewOpcode == X86::CMP32rm ||
3586 NewOpcode == X86::CMP16rm || NewOpcode == X86::CMP8rm)
3591 // Get the unique definition of SrcReg.
3592 MachineInstr *MI = MRI->getUniqueVRegDef(SrcReg);
3593 if (!MI) return false;
3595 // CmpInstr is the first instruction of the BB.
3596 MachineBasicBlock::iterator I = CmpInstr, Def = MI;
3598 // If we are comparing against zero, check whether we can use MI to update
3599 // EFLAGS. If MI is not in the same BB as CmpInstr, do not optimize.
3600 bool IsCmpZero = (CmpMask != 0 && CmpValue == 0);
3601 if (IsCmpZero && MI->getParent() != CmpInstr.getParent())
3604 // If we have a use of the source register between the def and our compare
3605 // instruction we can eliminate the compare iff the use sets EFLAGS in the
3607 bool ShouldUpdateCC = false;
3608 bool NoSignFlag = false;
3609 X86::CondCode NewCC = X86::COND_INVALID;
3610 if (IsCmpZero && !isDefConvertible(*MI, NoSignFlag)) {
3611 // Scan forward from the use until we hit the use we're looking for or the
3612 // compare instruction.
3613 for (MachineBasicBlock::iterator J = MI;; ++J) {
3614 // Do we have a convertible instruction?
3615 NewCC = isUseDefConvertible(*J);
3616 if (NewCC != X86::COND_INVALID && J->getOperand(1).isReg() &&
3617 J->getOperand(1).getReg() == SrcReg) {
3618 assert(J->definesRegister(X86::EFLAGS) && "Must be an EFLAGS def!");
3619 ShouldUpdateCC = true; // Update CC later on.
3620 // This is not a def of SrcReg, but still a def of EFLAGS. Keep going
3621 // with the new def.
3632 // We are searching for an earlier instruction that can make CmpInstr
3633 // redundant and that instruction will be saved in Sub.
3634 MachineInstr *Sub = nullptr;
3635 const TargetRegisterInfo *TRI = &getRegisterInfo();
3637 // We iterate backward, starting from the instruction before CmpInstr and
3638 // stop when reaching the definition of a source register or done with the BB.
3639 // RI points to the instruction before CmpInstr.
3640 // If the definition is in this basic block, RE points to the definition;
3641 // otherwise, RE is the rend of the basic block.
3642 MachineBasicBlock::reverse_iterator
3643 RI = ++I.getReverse(),
3644 RE = CmpInstr.getParent() == MI->getParent()
3645 ? Def.getReverse() /* points to MI */
3646 : CmpInstr.getParent()->rend();
3647 MachineInstr *Movr0Inst = nullptr;
3648 for (; RI != RE; ++RI) {
3649 MachineInstr &Instr = *RI;
3650 // Check whether CmpInstr can be made redundant by the current instruction.
3651 if (!IsCmpZero && isRedundantFlagInstr(CmpInstr, SrcReg, SrcReg2, CmpMask,
3657 if (Instr.modifiesRegister(X86::EFLAGS, TRI) ||
3658 Instr.readsRegister(X86::EFLAGS, TRI)) {
3659 // This instruction modifies or uses EFLAGS.
3661 // MOV32r0 etc. are implemented with xor which clobbers condition code.
3662 // They are safe to move up, if the definition to EFLAGS is dead and
3663 // earlier instructions do not read or write EFLAGS.
3664 if (!Movr0Inst && Instr.getOpcode() == X86::MOV32r0 &&
3665 Instr.registerDefIsDead(X86::EFLAGS, TRI)) {
3670 // We can't remove CmpInstr.
3675 // Return false if no candidates exist.
3676 if (!IsCmpZero && !Sub)
3679 bool IsSwapped = (SrcReg2 != 0 && Sub->getOperand(1).getReg() == SrcReg2 &&
3680 Sub->getOperand(2).getReg() == SrcReg);
3682 // Scan forward from the instruction after CmpInstr for uses of EFLAGS.
3683 // It is safe to remove CmpInstr if EFLAGS is redefined or killed.
3684 // If we are done with the basic block, we need to check whether EFLAGS is
3686 bool IsSafe = false;
3687 SmallVector<std::pair<MachineInstr*, unsigned /*NewOpc*/>, 4> OpsToUpdate;
3688 MachineBasicBlock::iterator E = CmpInstr.getParent()->end();
3689 for (++I; I != E; ++I) {
3690 const MachineInstr &Instr = *I;
3691 bool ModifyEFLAGS = Instr.modifiesRegister(X86::EFLAGS, TRI);
3692 bool UseEFLAGS = Instr.readsRegister(X86::EFLAGS, TRI);
3693 // We should check the usage if this instruction uses and updates EFLAGS.
3694 if (!UseEFLAGS && ModifyEFLAGS) {
3695 // It is safe to remove CmpInstr if EFLAGS is updated again.
3699 if (!UseEFLAGS && !ModifyEFLAGS)
3702 // EFLAGS is used by this instruction.
3703 X86::CondCode OldCC = X86::COND_INVALID;
3704 bool OpcIsSET = false;
3705 if (IsCmpZero || IsSwapped) {
3706 // We decode the condition code from opcode.
3707 if (Instr.isBranch())
3708 OldCC = X86::getCondFromBranchOpc(Instr.getOpcode());
3710 OldCC = X86::getCondFromSETOpc(Instr.getOpcode());
3711 if (OldCC != X86::COND_INVALID)
3714 OldCC = X86::getCondFromCMovOpc(Instr.getOpcode());
3716 if (OldCC == X86::COND_INVALID) return false;
3718 X86::CondCode ReplacementCC = X86::COND_INVALID;
3722 case X86::COND_A: case X86::COND_AE:
3723 case X86::COND_B: case X86::COND_BE:
3724 case X86::COND_G: case X86::COND_GE:
3725 case X86::COND_L: case X86::COND_LE:
3726 case X86::COND_O: case X86::COND_NO:
3727 // CF and OF are used, we can't perform this optimization.
3729 case X86::COND_S: case X86::COND_NS:
3730 // If SF is used, but the instruction doesn't update the SF, then we
3731 // can't do the optimization.
3737 // If we're updating the condition code check if we have to reverse the
3744 ReplacementCC = NewCC;
3747 ReplacementCC = GetOppositeBranchCondition(NewCC);
3750 } else if (IsSwapped) {
3751 // If we have SUB(r1, r2) and CMP(r2, r1), the condition code needs
3752 // to be changed from r2 > r1 to r1 < r2, from r2 < r1 to r1 > r2, etc.
3753 // We swap the condition code and synthesize the new opcode.
3754 ReplacementCC = getSwappedCondition(OldCC);
3755 if (ReplacementCC == X86::COND_INVALID) return false;
3758 if ((ShouldUpdateCC || IsSwapped) && ReplacementCC != OldCC) {
3759 // Synthesize the new opcode.
3760 bool HasMemoryOperand = Instr.hasOneMemOperand();
3762 if (Instr.isBranch())
3763 NewOpc = GetCondBranchFromCond(ReplacementCC);
3765 NewOpc = getSETFromCond(ReplacementCC, HasMemoryOperand);
3767 unsigned DstReg = Instr.getOperand(0).getReg();
3768 const TargetRegisterClass *DstRC = MRI->getRegClass(DstReg);
3769 NewOpc = getCMovFromCond(ReplacementCC, TRI->getRegSizeInBits(*DstRC)/8,
3773 // Push the MachineInstr to OpsToUpdate.
3774 // If it is safe to remove CmpInstr, the condition code of these
3775 // instructions will be modified.
3776 OpsToUpdate.push_back(std::make_pair(&*I, NewOpc));
3778 if (ModifyEFLAGS || Instr.killsRegister(X86::EFLAGS, TRI)) {
3779 // It is safe to remove CmpInstr if EFLAGS is updated again or killed.
3785 // If EFLAGS is not killed nor re-defined, we should check whether it is
3786 // live-out. If it is live-out, do not optimize.
3787 if ((IsCmpZero || IsSwapped) && !IsSafe) {
3788 MachineBasicBlock *MBB = CmpInstr.getParent();
3789 for (MachineBasicBlock *Successor : MBB->successors())
3790 if (Successor->isLiveIn(X86::EFLAGS))
3794 // The instruction to be updated is either Sub or MI.
3795 Sub = IsCmpZero ? MI : Sub;
3796 // Move Movr0Inst to the appropriate place before Sub.
3798 // Look backwards until we find a def that doesn't use the current EFLAGS.
3800 MachineBasicBlock::reverse_iterator InsertI = Def.getReverse(),
3801 InsertE = Sub->getParent()->rend();
3802 for (; InsertI != InsertE; ++InsertI) {
3803 MachineInstr *Instr = &*InsertI;
3804 if (!Instr->readsRegister(X86::EFLAGS, TRI) &&
3805 Instr->modifiesRegister(X86::EFLAGS, TRI)) {
3806 Sub->getParent()->remove(Movr0Inst);
3807 Instr->getParent()->insert(MachineBasicBlock::iterator(Instr),
3812 if (InsertI == InsertE)
3816 // Make sure Sub instruction defines EFLAGS and mark the def live.
3817 unsigned i = 0, e = Sub->getNumOperands();
3818 for (; i != e; ++i) {
3819 MachineOperand &MO = Sub->getOperand(i);
3820 if (MO.isReg() && MO.isDef() && MO.getReg() == X86::EFLAGS) {
3821 MO.setIsDead(false);
3825 assert(i != e && "Unable to locate a def EFLAGS operand");
3827 CmpInstr.eraseFromParent();
3829 // Modify the condition code of instructions in OpsToUpdate.
3830 for (auto &Op : OpsToUpdate)
3831 Op.first->setDesc(get(Op.second));
3835 /// Try to remove the load by folding it to a register
3836 /// operand at the use. We fold the load instructions if load defines a virtual
3837 /// register, the virtual register is used once in the same BB, and the
3838 /// instructions in-between do not load or store, and have no side effects.
3839 MachineInstr *X86InstrInfo::optimizeLoadInstr(MachineInstr &MI,
3840 const MachineRegisterInfo *MRI,
3841 unsigned &FoldAsLoadDefReg,
3842 MachineInstr *&DefMI) const {
3843 // Check whether we can move DefMI here.
3844 DefMI = MRI->getVRegDef(FoldAsLoadDefReg);
3846 bool SawStore = false;
3847 if (!DefMI->isSafeToMove(nullptr, SawStore))
3850 // Collect information about virtual register operands of MI.
3851 SmallVector<unsigned, 1> SrcOperandIds;
3852 for (unsigned i = 0, e = MI.getNumOperands(); i != e; ++i) {
3853 MachineOperand &MO = MI.getOperand(i);
3856 unsigned Reg = MO.getReg();
3857 if (Reg != FoldAsLoadDefReg)
3859 // Do not fold if we have a subreg use or a def.
3860 if (MO.getSubReg() || MO.isDef())
3862 SrcOperandIds.push_back(i);
3864 if (SrcOperandIds.empty())
3867 // Check whether we can fold the def into SrcOperandId.
3868 if (MachineInstr *FoldMI = foldMemoryOperand(MI, SrcOperandIds, *DefMI)) {
3869 FoldAsLoadDefReg = 0;
3876 /// Expand a single-def pseudo instruction to a two-addr
3877 /// instruction with two undef reads of the register being defined.
3878 /// This is used for mapping:
3881 /// %xmm4 = PXORrr undef %xmm4, undef %xmm4
3883 static bool Expand2AddrUndef(MachineInstrBuilder &MIB,
3884 const MCInstrDesc &Desc) {
3885 assert(Desc.getNumOperands() == 3 && "Expected two-addr instruction.");
3886 unsigned Reg = MIB->getOperand(0).getReg();
3889 // MachineInstr::addOperand() will insert explicit operands before any
3890 // implicit operands.
3891 MIB.addReg(Reg, RegState::Undef).addReg(Reg, RegState::Undef);
3892 // But we don't trust that.
3893 assert(MIB->getOperand(1).getReg() == Reg &&
3894 MIB->getOperand(2).getReg() == Reg && "Misplaced operand");
3898 /// Expand a single-def pseudo instruction to a two-addr
3899 /// instruction with two %k0 reads.
3900 /// This is used for mapping:
3903 /// %k4 = KXNORrr %k0, %k0
3904 static bool Expand2AddrKreg(MachineInstrBuilder &MIB,
3905 const MCInstrDesc &Desc, unsigned Reg) {
3906 assert(Desc.getNumOperands() == 3 && "Expected two-addr instruction.");
3908 MIB.addReg(Reg, RegState::Undef).addReg(Reg, RegState::Undef);
3912 static bool expandMOV32r1(MachineInstrBuilder &MIB, const TargetInstrInfo &TII,
3914 MachineBasicBlock &MBB = *MIB->getParent();
3915 DebugLoc DL = MIB->getDebugLoc();
3916 unsigned Reg = MIB->getOperand(0).getReg();
3919 BuildMI(MBB, MIB.getInstr(), DL, TII.get(X86::XOR32rr), Reg)
3920 .addReg(Reg, RegState::Undef)
3921 .addReg(Reg, RegState::Undef);
3923 // Turn the pseudo into an INC or DEC.
3924 MIB->setDesc(TII.get(MinusOne ? X86::DEC32r : X86::INC32r));
3930 static bool ExpandMOVImmSExti8(MachineInstrBuilder &MIB,
3931 const TargetInstrInfo &TII,
3932 const X86Subtarget &Subtarget) {
3933 MachineBasicBlock &MBB = *MIB->getParent();
3934 DebugLoc DL = MIB->getDebugLoc();
3935 int64_t Imm = MIB->getOperand(1).getImm();
3936 assert(Imm != 0 && "Using push/pop for 0 is not efficient.");
3937 MachineBasicBlock::iterator I = MIB.getInstr();
3939 int StackAdjustment;
3941 if (Subtarget.is64Bit()) {
3942 assert(MIB->getOpcode() == X86::MOV64ImmSExti8 ||
3943 MIB->getOpcode() == X86::MOV32ImmSExti8);
3945 // Can't use push/pop lowering if the function might write to the red zone.
3946 X86MachineFunctionInfo *X86FI =
3947 MBB.getParent()->getInfo<X86MachineFunctionInfo>();
3948 if (X86FI->getUsesRedZone()) {
3949 MIB->setDesc(TII.get(MIB->getOpcode() ==
3950 X86::MOV32ImmSExti8 ? X86::MOV32ri : X86::MOV64ri));
3954 // 64-bit mode doesn't have 32-bit push/pop, so use 64-bit operations and
3955 // widen the register if necessary.
3956 StackAdjustment = 8;
3957 BuildMI(MBB, I, DL, TII.get(X86::PUSH64i8)).addImm(Imm);
3958 MIB->setDesc(TII.get(X86::POP64r));
3960 .setReg(getX86SubSuperRegister(MIB->getOperand(0).getReg(), 64));
3962 assert(MIB->getOpcode() == X86::MOV32ImmSExti8);
3963 StackAdjustment = 4;
3964 BuildMI(MBB, I, DL, TII.get(X86::PUSH32i8)).addImm(Imm);
3965 MIB->setDesc(TII.get(X86::POP32r));
3968 // Build CFI if necessary.
3969 MachineFunction &MF = *MBB.getParent();
3970 const X86FrameLowering *TFL = Subtarget.getFrameLowering();
3971 bool IsWin64Prologue = MF.getTarget().getMCAsmInfo()->usesWindowsCFI();
3972 bool NeedsDwarfCFI =
3974 (MF.getMMI().hasDebugInfo() || MF.getFunction().needsUnwindTableEntry());
3975 bool EmitCFI = !TFL->hasFP(MF) && NeedsDwarfCFI;
3977 TFL->BuildCFI(MBB, I, DL,
3978 MCCFIInstruction::createAdjustCfaOffset(nullptr, StackAdjustment));
3979 TFL->BuildCFI(MBB, std::next(I), DL,
3980 MCCFIInstruction::createAdjustCfaOffset(nullptr, -StackAdjustment));
3986 // LoadStackGuard has so far only been implemented for 64-bit MachO. Different
3987 // code sequence is needed for other targets.
3988 static void expandLoadStackGuard(MachineInstrBuilder &MIB,
3989 const TargetInstrInfo &TII) {
3990 MachineBasicBlock &MBB = *MIB->getParent();
3991 DebugLoc DL = MIB->getDebugLoc();
3992 unsigned Reg = MIB->getOperand(0).getReg();
3993 const GlobalValue *GV =
3994 cast<GlobalValue>((*MIB->memoperands_begin())->getValue());
3995 auto Flags = MachineMemOperand::MOLoad |
3996 MachineMemOperand::MODereferenceable |
3997 MachineMemOperand::MOInvariant;
3998 MachineMemOperand *MMO = MBB.getParent()->getMachineMemOperand(
3999 MachinePointerInfo::getGOT(*MBB.getParent()), Flags, 8, 8);
4000 MachineBasicBlock::iterator I = MIB.getInstr();
4002 BuildMI(MBB, I, DL, TII.get(X86::MOV64rm), Reg).addReg(X86::RIP).addImm(1)
4003 .addReg(0).addGlobalAddress(GV, 0, X86II::MO_GOTPCREL).addReg(0)
4004 .addMemOperand(MMO);
4005 MIB->setDebugLoc(DL);
4006 MIB->setDesc(TII.get(X86::MOV64rm));
4007 MIB.addReg(Reg, RegState::Kill).addImm(1).addReg(0).addImm(0).addReg(0);
4010 static bool expandXorFP(MachineInstrBuilder &MIB, const TargetInstrInfo &TII) {
4011 MachineBasicBlock &MBB = *MIB->getParent();
4012 MachineFunction &MF = *MBB.getParent();
4013 const X86Subtarget &Subtarget = MF.getSubtarget<X86Subtarget>();
4014 const X86RegisterInfo *TRI = Subtarget.getRegisterInfo();
4016 MIB->getOpcode() == X86::XOR64_FP ? X86::XOR64rr : X86::XOR32rr;
4017 MIB->setDesc(TII.get(XorOp));
4018 MIB.addReg(TRI->getFrameRegister(MF), RegState::Undef);
4022 // This is used to handle spills for 128/256-bit registers when we have AVX512,
4023 // but not VLX. If it uses an extended register we need to use an instruction
4024 // that loads the lower 128/256-bit, but is available with only AVX512F.
4025 static bool expandNOVLXLoad(MachineInstrBuilder &MIB,
4026 const TargetRegisterInfo *TRI,
4027 const MCInstrDesc &LoadDesc,
4028 const MCInstrDesc &BroadcastDesc,
4030 unsigned DestReg = MIB->getOperand(0).getReg();
4031 // Check if DestReg is XMM16-31 or YMM16-31.
4032 if (TRI->getEncodingValue(DestReg) < 16) {
4033 // We can use a normal VEX encoded load.
4034 MIB->setDesc(LoadDesc);
4036 // Use a 128/256-bit VBROADCAST instruction.
4037 MIB->setDesc(BroadcastDesc);
4038 // Change the destination to a 512-bit register.
4039 DestReg = TRI->getMatchingSuperReg(DestReg, SubIdx, &X86::VR512RegClass);
4040 MIB->getOperand(0).setReg(DestReg);
4045 // This is used to handle spills for 128/256-bit registers when we have AVX512,
4046 // but not VLX. If it uses an extended register we need to use an instruction
4047 // that stores the lower 128/256-bit, but is available with only AVX512F.
4048 static bool expandNOVLXStore(MachineInstrBuilder &MIB,
4049 const TargetRegisterInfo *TRI,
4050 const MCInstrDesc &StoreDesc,
4051 const MCInstrDesc &ExtractDesc,
4053 unsigned SrcReg = MIB->getOperand(X86::AddrNumOperands).getReg();
4054 // Check if DestReg is XMM16-31 or YMM16-31.
4055 if (TRI->getEncodingValue(SrcReg) < 16) {
4056 // We can use a normal VEX encoded store.
4057 MIB->setDesc(StoreDesc);
4059 // Use a VEXTRACTF instruction.
4060 MIB->setDesc(ExtractDesc);
4061 // Change the destination to a 512-bit register.
4062 SrcReg = TRI->getMatchingSuperReg(SrcReg, SubIdx, &X86::VR512RegClass);
4063 MIB->getOperand(X86::AddrNumOperands).setReg(SrcReg);
4064 MIB.addImm(0x0); // Append immediate to extract from the lower bits.
4069 bool X86InstrInfo::expandPostRAPseudo(MachineInstr &MI) const {
4070 bool HasAVX = Subtarget.hasAVX();
4071 MachineInstrBuilder MIB(*MI.getParent()->getParent(), MI);
4072 switch (MI.getOpcode()) {
4074 return Expand2AddrUndef(MIB, get(X86::XOR32rr));
4076 return expandMOV32r1(MIB, *this, /*MinusOne=*/ false);
4078 return expandMOV32r1(MIB, *this, /*MinusOne=*/ true);
4079 case X86::MOV32ImmSExti8:
4080 case X86::MOV64ImmSExti8:
4081 return ExpandMOVImmSExti8(MIB, *this, Subtarget);
4083 return Expand2AddrUndef(MIB, get(X86::SBB8rr));
4084 case X86::SETB_C16r:
4085 return Expand2AddrUndef(MIB, get(X86::SBB16rr));
4086 case X86::SETB_C32r:
4087 return Expand2AddrUndef(MIB, get(X86::SBB32rr));
4088 case X86::SETB_C64r:
4089 return Expand2AddrUndef(MIB, get(X86::SBB64rr));
4091 return Expand2AddrUndef(MIB, get(X86::MMX_PXORirr));
4095 return Expand2AddrUndef(MIB, get(HasAVX ? X86::VXORPSrr : X86::XORPSrr));
4096 case X86::AVX_SET0: {
4097 assert(HasAVX && "AVX not supported");
4098 const TargetRegisterInfo *TRI = &getRegisterInfo();
4099 unsigned SrcReg = MIB->getOperand(0).getReg();
4100 unsigned XReg = TRI->getSubReg(SrcReg, X86::sub_xmm);
4101 MIB->getOperand(0).setReg(XReg);
4102 Expand2AddrUndef(MIB, get(X86::VXORPSrr));
4103 MIB.addReg(SrcReg, RegState::ImplicitDefine);
4106 case X86::AVX512_128_SET0:
4107 case X86::AVX512_FsFLD0SS:
4108 case X86::AVX512_FsFLD0SD: {
4109 bool HasVLX = Subtarget.hasVLX();
4110 unsigned SrcReg = MIB->getOperand(0).getReg();
4111 const TargetRegisterInfo *TRI = &getRegisterInfo();
4112 if (HasVLX || TRI->getEncodingValue(SrcReg) < 16)
4113 return Expand2AddrUndef(MIB,
4114 get(HasVLX ? X86::VPXORDZ128rr : X86::VXORPSrr));
4115 // Extended register without VLX. Use a larger XOR.
4117 TRI->getMatchingSuperReg(SrcReg, X86::sub_xmm, &X86::VR512RegClass);
4118 MIB->getOperand(0).setReg(SrcReg);
4119 return Expand2AddrUndef(MIB, get(X86::VPXORDZrr));
4121 case X86::AVX512_256_SET0:
4122 case X86::AVX512_512_SET0: {
4123 bool HasVLX = Subtarget.hasVLX();
4124 unsigned SrcReg = MIB->getOperand(0).getReg();
4125 const TargetRegisterInfo *TRI = &getRegisterInfo();
4126 if (HasVLX || TRI->getEncodingValue(SrcReg) < 16) {
4127 unsigned XReg = TRI->getSubReg(SrcReg, X86::sub_xmm);
4128 MIB->getOperand(0).setReg(XReg);
4129 Expand2AddrUndef(MIB,
4130 get(HasVLX ? X86::VPXORDZ128rr : X86::VXORPSrr));
4131 MIB.addReg(SrcReg, RegState::ImplicitDefine);
4134 return Expand2AddrUndef(MIB, get(X86::VPXORDZrr));
4136 case X86::V_SETALLONES:
4137 return Expand2AddrUndef(MIB, get(HasAVX ? X86::VPCMPEQDrr : X86::PCMPEQDrr));
4138 case X86::AVX2_SETALLONES:
4139 return Expand2AddrUndef(MIB, get(X86::VPCMPEQDYrr));
4140 case X86::AVX1_SETALLONES: {
4141 unsigned Reg = MIB->getOperand(0).getReg();
4142 // VCMPPSYrri with an immediate 0xf should produce VCMPTRUEPS.
4143 MIB->setDesc(get(X86::VCMPPSYrri));
4144 MIB.addReg(Reg, RegState::Undef).addReg(Reg, RegState::Undef).addImm(0xf);
4147 case X86::AVX512_512_SETALLONES: {
4148 unsigned Reg = MIB->getOperand(0).getReg();
4149 MIB->setDesc(get(X86::VPTERNLOGDZrri));
4150 // VPTERNLOGD needs 3 register inputs and an immediate.
4151 // 0xff will return 1s for any input.
4152 MIB.addReg(Reg, RegState::Undef).addReg(Reg, RegState::Undef)
4153 .addReg(Reg, RegState::Undef).addImm(0xff);
4156 case X86::AVX512_512_SEXT_MASK_32:
4157 case X86::AVX512_512_SEXT_MASK_64: {
4158 unsigned Reg = MIB->getOperand(0).getReg();
4159 unsigned MaskReg = MIB->getOperand(1).getReg();
4160 unsigned MaskState = getRegState(MIB->getOperand(1));
4161 unsigned Opc = (MI.getOpcode() == X86::AVX512_512_SEXT_MASK_64) ?
4162 X86::VPTERNLOGQZrrikz : X86::VPTERNLOGDZrrikz;
4163 MI.RemoveOperand(1);
4164 MIB->setDesc(get(Opc));
4165 // VPTERNLOG needs 3 register inputs and an immediate.
4166 // 0xff will return 1s for any input.
4167 MIB.addReg(Reg, RegState::Undef).addReg(MaskReg, MaskState)
4168 .addReg(Reg, RegState::Undef).addReg(Reg, RegState::Undef).addImm(0xff);
4171 case X86::VMOVAPSZ128rm_NOVLX:
4172 return expandNOVLXLoad(MIB, &getRegisterInfo(), get(X86::VMOVAPSrm),
4173 get(X86::VBROADCASTF32X4rm), X86::sub_xmm);
4174 case X86::VMOVUPSZ128rm_NOVLX:
4175 return expandNOVLXLoad(MIB, &getRegisterInfo(), get(X86::VMOVUPSrm),
4176 get(X86::VBROADCASTF32X4rm), X86::sub_xmm);
4177 case X86::VMOVAPSZ256rm_NOVLX:
4178 return expandNOVLXLoad(MIB, &getRegisterInfo(), get(X86::VMOVAPSYrm),
4179 get(X86::VBROADCASTF64X4rm), X86::sub_ymm);
4180 case X86::VMOVUPSZ256rm_NOVLX:
4181 return expandNOVLXLoad(MIB, &getRegisterInfo(), get(X86::VMOVUPSYrm),
4182 get(X86::VBROADCASTF64X4rm), X86::sub_ymm);
4183 case X86::VMOVAPSZ128mr_NOVLX:
4184 return expandNOVLXStore(MIB, &getRegisterInfo(), get(X86::VMOVAPSmr),
4185 get(X86::VEXTRACTF32x4Zmr), X86::sub_xmm);
4186 case X86::VMOVUPSZ128mr_NOVLX:
4187 return expandNOVLXStore(MIB, &getRegisterInfo(), get(X86::VMOVUPSmr),
4188 get(X86::VEXTRACTF32x4Zmr), X86::sub_xmm);
4189 case X86::VMOVAPSZ256mr_NOVLX:
4190 return expandNOVLXStore(MIB, &getRegisterInfo(), get(X86::VMOVAPSYmr),
4191 get(X86::VEXTRACTF64x4Zmr), X86::sub_ymm);
4192 case X86::VMOVUPSZ256mr_NOVLX:
4193 return expandNOVLXStore(MIB, &getRegisterInfo(), get(X86::VMOVUPSYmr),
4194 get(X86::VEXTRACTF64x4Zmr), X86::sub_ymm);
4195 case X86::MOV32ri64: {
4196 unsigned Reg = MIB->getOperand(0).getReg();
4197 unsigned Reg32 = RI.getSubReg(Reg, X86::sub_32bit);
4198 MI.setDesc(get(X86::MOV32ri));
4199 MIB->getOperand(0).setReg(Reg32);
4200 MIB.addReg(Reg, RegState::ImplicitDefine);
4204 // KNL does not recognize dependency-breaking idioms for mask registers,
4205 // so kxnor %k1, %k1, %k2 has a RAW dependence on %k1.
4206 // Using %k0 as the undef input register is a performance heuristic based
4207 // on the assumption that %k0 is used less frequently than the other mask
4208 // registers, since it is not usable as a write mask.
4209 // FIXME: A more advanced approach would be to choose the best input mask
4210 // register based on context.
4211 case X86::KSET0W: return Expand2AddrKreg(MIB, get(X86::KXORWrr), X86::K0);
4212 case X86::KSET0D: return Expand2AddrKreg(MIB, get(X86::KXORDrr), X86::K0);
4213 case X86::KSET0Q: return Expand2AddrKreg(MIB, get(X86::KXORQrr), X86::K0);
4214 case X86::KSET1W: return Expand2AddrKreg(MIB, get(X86::KXNORWrr), X86::K0);
4215 case X86::KSET1D: return Expand2AddrKreg(MIB, get(X86::KXNORDrr), X86::K0);
4216 case X86::KSET1Q: return Expand2AddrKreg(MIB, get(X86::KXNORQrr), X86::K0);
4217 case TargetOpcode::LOAD_STACK_GUARD:
4218 expandLoadStackGuard(MIB, *this);
4222 return expandXorFP(MIB, *this);
4227 /// Return true for all instructions that only update
4228 /// the first 32 or 64-bits of the destination register and leave the rest
4229 /// unmodified. This can be used to avoid folding loads if the instructions
4230 /// only update part of the destination register, and the non-updated part is
4231 /// not needed. e.g. cvtss2sd, sqrtss. Unfolding the load from these
4232 /// instructions breaks the partial register dependency and it can improve
4233 /// performance. e.g.:
4235 /// movss (%rdi), %xmm0
4236 /// cvtss2sd %xmm0, %xmm0
4239 /// cvtss2sd (%rdi), %xmm0
4241 /// FIXME: This should be turned into a TSFlags.
4243 static bool hasPartialRegUpdate(unsigned Opcode,
4244 const X86Subtarget &Subtarget,
4245 bool ForLoadFold = false) {
4247 case X86::CVTSI2SSrr:
4248 case X86::CVTSI2SSrm:
4249 case X86::CVTSI642SSrr:
4250 case X86::CVTSI642SSrm:
4251 case X86::CVTSI2SDrr:
4252 case X86::CVTSI2SDrm:
4253 case X86::CVTSI642SDrr:
4254 case X86::CVTSI642SDrm:
4255 // Load folding won't effect the undef register update since the input is
4257 return !ForLoadFold;
4258 case X86::CVTSD2SSrr:
4259 case X86::CVTSD2SSrm:
4260 case X86::CVTSS2SDrr:
4261 case X86::CVTSS2SDrm:
4268 case X86::RCPSSr_Int:
4269 case X86::RCPSSm_Int:
4276 case X86::RSQRTSSr_Int:
4277 case X86::RSQRTSSm_Int:
4280 case X86::SQRTSSr_Int:
4281 case X86::SQRTSSm_Int:
4284 case X86::SQRTSDr_Int:
4285 case X86::SQRTSDm_Int:
4288 case X86::POPCNT32rm:
4289 case X86::POPCNT32rr:
4290 case X86::POPCNT64rm:
4291 case X86::POPCNT64rr:
4292 return Subtarget.hasPOPCNTFalseDeps();
4293 case X86::LZCNT32rm:
4294 case X86::LZCNT32rr:
4295 case X86::LZCNT64rm:
4296 case X86::LZCNT64rr:
4297 case X86::TZCNT32rm:
4298 case X86::TZCNT32rr:
4299 case X86::TZCNT64rm:
4300 case X86::TZCNT64rr:
4301 return Subtarget.hasLZCNTFalseDeps();
4307 /// Inform the BreakFalseDeps pass how many idle
4308 /// instructions we would like before a partial register update.
4309 unsigned X86InstrInfo::getPartialRegUpdateClearance(
4310 const MachineInstr &MI, unsigned OpNum,
4311 const TargetRegisterInfo *TRI) const {
4312 if (OpNum != 0 || !hasPartialRegUpdate(MI.getOpcode(), Subtarget))
4315 // If MI is marked as reading Reg, the partial register update is wanted.
4316 const MachineOperand &MO = MI.getOperand(0);
4317 unsigned Reg = MO.getReg();
4318 if (TargetRegisterInfo::isVirtualRegister(Reg)) {
4319 if (MO.readsReg() || MI.readsVirtualRegister(Reg))
4322 if (MI.readsRegister(Reg, TRI))
4326 // If any instructions in the clearance range are reading Reg, insert a
4327 // dependency breaking instruction, which is inexpensive and is likely to
4328 // be hidden in other instruction's cycles.
4329 return PartialRegUpdateClearance;
4332 // Return true for any instruction the copies the high bits of the first source
4333 // operand into the unused high bits of the destination operand.
4334 static bool hasUndefRegUpdate(unsigned Opcode, bool ForLoadFold = false) {
4336 case X86::VCVTSI2SSrr:
4337 case X86::VCVTSI2SSrm:
4338 case X86::VCVTSI2SSrr_Int:
4339 case X86::VCVTSI2SSrm_Int:
4340 case X86::VCVTSI642SSrr:
4341 case X86::VCVTSI642SSrm:
4342 case X86::VCVTSI642SSrr_Int:
4343 case X86::VCVTSI642SSrm_Int:
4344 case X86::VCVTSI2SDrr:
4345 case X86::VCVTSI2SDrm:
4346 case X86::VCVTSI2SDrr_Int:
4347 case X86::VCVTSI2SDrm_Int:
4348 case X86::VCVTSI642SDrr:
4349 case X86::VCVTSI642SDrm:
4350 case X86::VCVTSI642SDrr_Int:
4351 case X86::VCVTSI642SDrm_Int:
4353 case X86::VCVTSI2SSZrr:
4354 case X86::VCVTSI2SSZrm:
4355 case X86::VCVTSI2SSZrr_Int:
4356 case X86::VCVTSI2SSZrrb_Int:
4357 case X86::VCVTSI2SSZrm_Int:
4358 case X86::VCVTSI642SSZrr:
4359 case X86::VCVTSI642SSZrm:
4360 case X86::VCVTSI642SSZrr_Int:
4361 case X86::VCVTSI642SSZrrb_Int:
4362 case X86::VCVTSI642SSZrm_Int:
4363 case X86::VCVTSI2SDZrr:
4364 case X86::VCVTSI2SDZrm:
4365 case X86::VCVTSI2SDZrr_Int:
4366 case X86::VCVTSI2SDZrrb_Int:
4367 case X86::VCVTSI2SDZrm_Int:
4368 case X86::VCVTSI642SDZrr:
4369 case X86::VCVTSI642SDZrm:
4370 case X86::VCVTSI642SDZrr_Int:
4371 case X86::VCVTSI642SDZrrb_Int:
4372 case X86::VCVTSI642SDZrm_Int:
4373 case X86::VCVTUSI2SSZrr:
4374 case X86::VCVTUSI2SSZrm:
4375 case X86::VCVTUSI2SSZrr_Int:
4376 case X86::VCVTUSI2SSZrrb_Int:
4377 case X86::VCVTUSI2SSZrm_Int:
4378 case X86::VCVTUSI642SSZrr:
4379 case X86::VCVTUSI642SSZrm:
4380 case X86::VCVTUSI642SSZrr_Int:
4381 case X86::VCVTUSI642SSZrrb_Int:
4382 case X86::VCVTUSI642SSZrm_Int:
4383 case X86::VCVTUSI2SDZrr:
4384 case X86::VCVTUSI2SDZrm:
4385 case X86::VCVTUSI2SDZrr_Int:
4386 case X86::VCVTUSI2SDZrm_Int:
4387 case X86::VCVTUSI642SDZrr:
4388 case X86::VCVTUSI642SDZrm:
4389 case X86::VCVTUSI642SDZrr_Int:
4390 case X86::VCVTUSI642SDZrrb_Int:
4391 case X86::VCVTUSI642SDZrm_Int:
4392 // Load folding won't effect the undef register update since the input is
4394 return !ForLoadFold;
4395 case X86::VCVTSD2SSrr:
4396 case X86::VCVTSD2SSrm:
4397 case X86::VCVTSD2SSrr_Int:
4398 case X86::VCVTSD2SSrm_Int:
4399 case X86::VCVTSS2SDrr:
4400 case X86::VCVTSS2SDrm:
4401 case X86::VCVTSS2SDrr_Int:
4402 case X86::VCVTSS2SDrm_Int:
4404 case X86::VRCPSSr_Int:
4406 case X86::VRCPSSm_Int:
4407 case X86::VROUNDSDr:
4408 case X86::VROUNDSDm:
4409 case X86::VROUNDSDr_Int:
4410 case X86::VROUNDSDm_Int:
4411 case X86::VROUNDSSr:
4412 case X86::VROUNDSSm:
4413 case X86::VROUNDSSr_Int:
4414 case X86::VROUNDSSm_Int:
4415 case X86::VRSQRTSSr:
4416 case X86::VRSQRTSSr_Int:
4417 case X86::VRSQRTSSm:
4418 case X86::VRSQRTSSm_Int:
4420 case X86::VSQRTSSr_Int:
4422 case X86::VSQRTSSm_Int:
4424 case X86::VSQRTSDr_Int:
4426 case X86::VSQRTSDm_Int:
4428 case X86::VCVTSD2SSZrr:
4429 case X86::VCVTSD2SSZrr_Int:
4430 case X86::VCVTSD2SSZrrb_Int:
4431 case X86::VCVTSD2SSZrm:
4432 case X86::VCVTSD2SSZrm_Int:
4433 case X86::VCVTSS2SDZrr:
4434 case X86::VCVTSS2SDZrr_Int:
4435 case X86::VCVTSS2SDZrrb_Int:
4436 case X86::VCVTSS2SDZrm:
4437 case X86::VCVTSS2SDZrm_Int:
4438 case X86::VGETEXPSDZr:
4439 case X86::VGETEXPSDZrb:
4440 case X86::VGETEXPSDZm:
4441 case X86::VGETEXPSSZr:
4442 case X86::VGETEXPSSZrb:
4443 case X86::VGETEXPSSZm:
4444 case X86::VGETMANTSDZrri:
4445 case X86::VGETMANTSDZrrib:
4446 case X86::VGETMANTSDZrmi:
4447 case X86::VGETMANTSSZrri:
4448 case X86::VGETMANTSSZrrib:
4449 case X86::VGETMANTSSZrmi:
4450 case X86::VRNDSCALESDZr:
4451 case X86::VRNDSCALESDZr_Int:
4452 case X86::VRNDSCALESDZrb_Int:
4453 case X86::VRNDSCALESDZm:
4454 case X86::VRNDSCALESDZm_Int:
4455 case X86::VRNDSCALESSZr:
4456 case X86::VRNDSCALESSZr_Int:
4457 case X86::VRNDSCALESSZrb_Int:
4458 case X86::VRNDSCALESSZm:
4459 case X86::VRNDSCALESSZm_Int:
4460 case X86::VRCP14SDZrr:
4461 case X86::VRCP14SDZrm:
4462 case X86::VRCP14SSZrr:
4463 case X86::VRCP14SSZrm:
4464 case X86::VRCP28SDZr:
4465 case X86::VRCP28SDZrb:
4466 case X86::VRCP28SDZm:
4467 case X86::VRCP28SSZr:
4468 case X86::VRCP28SSZrb:
4469 case X86::VRCP28SSZm:
4470 case X86::VREDUCESSZrmi:
4471 case X86::VREDUCESSZrri:
4472 case X86::VREDUCESSZrrib:
4473 case X86::VRSQRT14SDZrr:
4474 case X86::VRSQRT14SDZrm:
4475 case X86::VRSQRT14SSZrr:
4476 case X86::VRSQRT14SSZrm:
4477 case X86::VRSQRT28SDZr:
4478 case X86::VRSQRT28SDZrb:
4479 case X86::VRSQRT28SDZm:
4480 case X86::VRSQRT28SSZr:
4481 case X86::VRSQRT28SSZrb:
4482 case X86::VRSQRT28SSZm:
4483 case X86::VSQRTSSZr:
4484 case X86::VSQRTSSZr_Int:
4485 case X86::VSQRTSSZrb_Int:
4486 case X86::VSQRTSSZm:
4487 case X86::VSQRTSSZm_Int:
4488 case X86::VSQRTSDZr:
4489 case X86::VSQRTSDZr_Int:
4490 case X86::VSQRTSDZrb_Int:
4491 case X86::VSQRTSDZm:
4492 case X86::VSQRTSDZm_Int:
4499 /// Inform the BreakFalseDeps pass how many idle instructions we would like
4500 /// before certain undef register reads.
4502 /// This catches the VCVTSI2SD family of instructions:
4504 /// vcvtsi2sdq %rax, undef %xmm0, %xmm14
4506 /// We should to be careful *not* to catch VXOR idioms which are presumably
4507 /// handled specially in the pipeline:
4509 /// vxorps undef %xmm1, undef %xmm1, %xmm1
4511 /// Like getPartialRegUpdateClearance, this makes a strong assumption that the
4512 /// high bits that are passed-through are not live.
4514 X86InstrInfo::getUndefRegClearance(const MachineInstr &MI, unsigned &OpNum,
4515 const TargetRegisterInfo *TRI) const {
4516 if (!hasUndefRegUpdate(MI.getOpcode()))
4519 // Set the OpNum parameter to the first source operand.
4522 const MachineOperand &MO = MI.getOperand(OpNum);
4523 if (MO.isUndef() && TargetRegisterInfo::isPhysicalRegister(MO.getReg())) {
4524 return UndefRegClearance;
4529 void X86InstrInfo::breakPartialRegDependency(
4530 MachineInstr &MI, unsigned OpNum, const TargetRegisterInfo *TRI) const {
4531 unsigned Reg = MI.getOperand(OpNum).getReg();
4532 // If MI kills this register, the false dependence is already broken.
4533 if (MI.killsRegister(Reg, TRI))
4536 if (X86::VR128RegClass.contains(Reg)) {
4537 // These instructions are all floating point domain, so xorps is the best
4539 unsigned Opc = Subtarget.hasAVX() ? X86::VXORPSrr : X86::XORPSrr;
4540 BuildMI(*MI.getParent(), MI, MI.getDebugLoc(), get(Opc), Reg)
4541 .addReg(Reg, RegState::Undef)
4542 .addReg(Reg, RegState::Undef);
4543 MI.addRegisterKilled(Reg, TRI, true);
4544 } else if (X86::VR256RegClass.contains(Reg)) {
4545 // Use vxorps to clear the full ymm register.
4546 // It wants to read and write the xmm sub-register.
4547 unsigned XReg = TRI->getSubReg(Reg, X86::sub_xmm);
4548 BuildMI(*MI.getParent(), MI, MI.getDebugLoc(), get(X86::VXORPSrr), XReg)
4549 .addReg(XReg, RegState::Undef)
4550 .addReg(XReg, RegState::Undef)
4551 .addReg(Reg, RegState::ImplicitDefine);
4552 MI.addRegisterKilled(Reg, TRI, true);
4553 } else if (X86::GR64RegClass.contains(Reg)) {
4554 // Using XOR32rr because it has shorter encoding and zeros up the upper bits
4556 unsigned XReg = TRI->getSubReg(Reg, X86::sub_32bit);
4557 BuildMI(*MI.getParent(), MI, MI.getDebugLoc(), get(X86::XOR32rr), XReg)
4558 .addReg(XReg, RegState::Undef)
4559 .addReg(XReg, RegState::Undef)
4560 .addReg(Reg, RegState::ImplicitDefine);
4561 MI.addRegisterKilled(Reg, TRI, true);
4562 } else if (X86::GR32RegClass.contains(Reg)) {
4563 BuildMI(*MI.getParent(), MI, MI.getDebugLoc(), get(X86::XOR32rr), Reg)
4564 .addReg(Reg, RegState::Undef)
4565 .addReg(Reg, RegState::Undef);
4566 MI.addRegisterKilled(Reg, TRI, true);
4570 static void addOperands(MachineInstrBuilder &MIB, ArrayRef<MachineOperand> MOs,
4571 int PtrOffset = 0) {
4572 unsigned NumAddrOps = MOs.size();
4574 if (NumAddrOps < 4) {
4575 // FrameIndex only - add an immediate offset (whether its zero or not).
4576 for (unsigned i = 0; i != NumAddrOps; ++i)
4578 addOffset(MIB, PtrOffset);
4580 // General Memory Addressing - we need to add any offset to an existing
4582 assert(MOs.size() == 5 && "Unexpected memory operand list length");
4583 for (unsigned i = 0; i != NumAddrOps; ++i) {
4584 const MachineOperand &MO = MOs[i];
4585 if (i == 3 && PtrOffset != 0) {
4586 MIB.addDisp(MO, PtrOffset);
4594 static void updateOperandRegConstraints(MachineFunction &MF,
4595 MachineInstr &NewMI,
4596 const TargetInstrInfo &TII) {
4597 MachineRegisterInfo &MRI = MF.getRegInfo();
4598 const TargetRegisterInfo &TRI = *MRI.getTargetRegisterInfo();
4600 for (int Idx : llvm::seq<int>(0, NewMI.getNumOperands())) {
4601 MachineOperand &MO = NewMI.getOperand(Idx);
4602 // We only need to update constraints on virtual register operands.
4605 unsigned Reg = MO.getReg();
4606 if (!TRI.isVirtualRegister(Reg))
4609 auto *NewRC = MRI.constrainRegClass(
4610 Reg, TII.getRegClass(NewMI.getDesc(), Idx, &TRI, MF));
4613 dbgs() << "WARNING: Unable to update register constraint for operand "
4614 << Idx << " of instruction:\n";
4615 NewMI.dump(); dbgs() << "\n");
4620 static MachineInstr *FuseTwoAddrInst(MachineFunction &MF, unsigned Opcode,
4621 ArrayRef<MachineOperand> MOs,
4622 MachineBasicBlock::iterator InsertPt,
4624 const TargetInstrInfo &TII) {
4625 // Create the base instruction with the memory operand as the first part.
4626 // Omit the implicit operands, something BuildMI can't do.
4627 MachineInstr *NewMI =
4628 MF.CreateMachineInstr(TII.get(Opcode), MI.getDebugLoc(), true);
4629 MachineInstrBuilder MIB(MF, NewMI);
4630 addOperands(MIB, MOs);
4632 // Loop over the rest of the ri operands, converting them over.
4633 unsigned NumOps = MI.getDesc().getNumOperands() - 2;
4634 for (unsigned i = 0; i != NumOps; ++i) {
4635 MachineOperand &MO = MI.getOperand(i + 2);
4638 for (unsigned i = NumOps + 2, e = MI.getNumOperands(); i != e; ++i) {
4639 MachineOperand &MO = MI.getOperand(i);
4643 updateOperandRegConstraints(MF, *NewMI, TII);
4645 MachineBasicBlock *MBB = InsertPt->getParent();
4646 MBB->insert(InsertPt, NewMI);
4651 static MachineInstr *FuseInst(MachineFunction &MF, unsigned Opcode,
4652 unsigned OpNo, ArrayRef<MachineOperand> MOs,
4653 MachineBasicBlock::iterator InsertPt,
4654 MachineInstr &MI, const TargetInstrInfo &TII,
4655 int PtrOffset = 0) {
4656 // Omit the implicit operands, something BuildMI can't do.
4657 MachineInstr *NewMI =
4658 MF.CreateMachineInstr(TII.get(Opcode), MI.getDebugLoc(), true);
4659 MachineInstrBuilder MIB(MF, NewMI);
4661 for (unsigned i = 0, e = MI.getNumOperands(); i != e; ++i) {
4662 MachineOperand &MO = MI.getOperand(i);
4664 assert(MO.isReg() && "Expected to fold into reg operand!");
4665 addOperands(MIB, MOs, PtrOffset);
4671 updateOperandRegConstraints(MF, *NewMI, TII);
4673 MachineBasicBlock *MBB = InsertPt->getParent();
4674 MBB->insert(InsertPt, NewMI);
4679 static MachineInstr *MakeM0Inst(const TargetInstrInfo &TII, unsigned Opcode,
4680 ArrayRef<MachineOperand> MOs,
4681 MachineBasicBlock::iterator InsertPt,
4683 MachineInstrBuilder MIB = BuildMI(*InsertPt->getParent(), InsertPt,
4684 MI.getDebugLoc(), TII.get(Opcode));
4685 addOperands(MIB, MOs);
4686 return MIB.addImm(0);
4689 MachineInstr *X86InstrInfo::foldMemoryOperandCustom(
4690 MachineFunction &MF, MachineInstr &MI, unsigned OpNum,
4691 ArrayRef<MachineOperand> MOs, MachineBasicBlock::iterator InsertPt,
4692 unsigned Size, unsigned Align) const {
4693 switch (MI.getOpcode()) {
4694 case X86::INSERTPSrr:
4695 case X86::VINSERTPSrr:
4696 case X86::VINSERTPSZrr:
4697 // Attempt to convert the load of inserted vector into a fold load
4698 // of a single float.
4700 unsigned Imm = MI.getOperand(MI.getNumOperands() - 1).getImm();
4701 unsigned ZMask = Imm & 15;
4702 unsigned DstIdx = (Imm >> 4) & 3;
4703 unsigned SrcIdx = (Imm >> 6) & 3;
4705 const TargetRegisterInfo &TRI = *MF.getSubtarget().getRegisterInfo();
4706 const TargetRegisterClass *RC = getRegClass(MI.getDesc(), OpNum, &RI, MF);
4707 unsigned RCSize = TRI.getRegSizeInBits(*RC) / 8;
4708 if (Size <= RCSize && 4 <= Align) {
4709 int PtrOffset = SrcIdx * 4;
4710 unsigned NewImm = (DstIdx << 4) | ZMask;
4711 unsigned NewOpCode =
4712 (MI.getOpcode() == X86::VINSERTPSZrr) ? X86::VINSERTPSZrm :
4713 (MI.getOpcode() == X86::VINSERTPSrr) ? X86::VINSERTPSrm :
4715 MachineInstr *NewMI =
4716 FuseInst(MF, NewOpCode, OpNum, MOs, InsertPt, MI, *this, PtrOffset);
4717 NewMI->getOperand(NewMI->getNumOperands() - 1).setImm(NewImm);
4722 case X86::MOVHLPSrr:
4723 case X86::VMOVHLPSrr:
4724 case X86::VMOVHLPSZrr:
4725 // Move the upper 64-bits of the second operand to the lower 64-bits.
4726 // To fold the load, adjust the pointer to the upper and use (V)MOVLPS.
4727 // TODO: In most cases AVX doesn't have a 8-byte alignment requirement.
4729 const TargetRegisterInfo &TRI = *MF.getSubtarget().getRegisterInfo();
4730 const TargetRegisterClass *RC = getRegClass(MI.getDesc(), OpNum, &RI, MF);
4731 unsigned RCSize = TRI.getRegSizeInBits(*RC) / 8;
4732 if (Size <= RCSize && 8 <= Align) {
4733 unsigned NewOpCode =
4734 (MI.getOpcode() == X86::VMOVHLPSZrr) ? X86::VMOVLPSZ128rm :
4735 (MI.getOpcode() == X86::VMOVHLPSrr) ? X86::VMOVLPSrm :
4737 MachineInstr *NewMI =
4738 FuseInst(MF, NewOpCode, OpNum, MOs, InsertPt, MI, *this, 8);
4748 static bool shouldPreventUndefRegUpdateMemFold(MachineFunction &MF,
4750 if (!hasUndefRegUpdate(MI.getOpcode(), /*ForLoadFold*/true) ||
4751 !MI.getOperand(1).isReg())
4754 // The are two cases we need to handle depending on where in the pipeline
4755 // the folding attempt is being made.
4756 // -Register has the undef flag set.
4757 // -Register is produced by the IMPLICIT_DEF instruction.
4759 if (MI.getOperand(1).isUndef())
4762 MachineRegisterInfo &RegInfo = MF.getRegInfo();
4763 MachineInstr *VRegDef = RegInfo.getUniqueVRegDef(MI.getOperand(1).getReg());
4764 return VRegDef && VRegDef->isImplicitDef();
4768 MachineInstr *X86InstrInfo::foldMemoryOperandImpl(
4769 MachineFunction &MF, MachineInstr &MI, unsigned OpNum,
4770 ArrayRef<MachineOperand> MOs, MachineBasicBlock::iterator InsertPt,
4771 unsigned Size, unsigned Align, bool AllowCommute) const {
4772 bool isSlowTwoMemOps = Subtarget.slowTwoMemOps();
4773 bool isTwoAddrFold = false;
4775 // For CPUs that favor the register form of a call or push,
4776 // do not fold loads into calls or pushes, unless optimizing for size
4778 if (isSlowTwoMemOps && !MF.getFunction().optForMinSize() &&
4779 (MI.getOpcode() == X86::CALL32r || MI.getOpcode() == X86::CALL64r ||
4780 MI.getOpcode() == X86::PUSH16r || MI.getOpcode() == X86::PUSH32r ||
4781 MI.getOpcode() == X86::PUSH64r))
4784 // Avoid partial and undef register update stalls unless optimizing for size.
4785 if (!MF.getFunction().optForSize() &&
4786 (hasPartialRegUpdate(MI.getOpcode(), Subtarget, /*ForLoadFold*/true) ||
4787 shouldPreventUndefRegUpdateMemFold(MF, MI)))
4790 unsigned NumOps = MI.getDesc().getNumOperands();
4792 NumOps > 1 && MI.getDesc().getOperandConstraint(1, MCOI::TIED_TO) != -1;
4794 // FIXME: AsmPrinter doesn't know how to handle
4795 // X86II::MO_GOT_ABSOLUTE_ADDRESS after folding.
4796 if (MI.getOpcode() == X86::ADD32ri &&
4797 MI.getOperand(2).getTargetFlags() == X86II::MO_GOT_ABSOLUTE_ADDRESS)
4800 // GOTTPOFF relocation loads can only be folded into add instructions.
4801 // FIXME: Need to exclude other relocations that only support specific
4803 if (MOs.size() == X86::AddrNumOperands &&
4804 MOs[X86::AddrDisp].getTargetFlags() == X86II::MO_GOTTPOFF &&
4805 MI.getOpcode() != X86::ADD64rr)
4808 MachineInstr *NewMI = nullptr;
4810 // Attempt to fold any custom cases we have.
4811 if (MachineInstr *CustomMI =
4812 foldMemoryOperandCustom(MF, MI, OpNum, MOs, InsertPt, Size, Align))
4815 const X86MemoryFoldTableEntry *I = nullptr;
4817 // Folding a memory location into the two-address part of a two-address
4818 // instruction is different than folding it other places. It requires
4819 // replacing the *two* registers with the memory location.
4820 if (isTwoAddr && NumOps >= 2 && OpNum < 2 && MI.getOperand(0).isReg() &&
4821 MI.getOperand(1).isReg() &&
4822 MI.getOperand(0).getReg() == MI.getOperand(1).getReg()) {
4823 I = lookupTwoAddrFoldTable(MI.getOpcode());
4824 isTwoAddrFold = true;
4827 if (MI.getOpcode() == X86::MOV32r0) {
4828 NewMI = MakeM0Inst(*this, X86::MOV32mi, MOs, InsertPt, MI);
4834 I = lookupFoldTable(MI.getOpcode(), OpNum);
4838 unsigned Opcode = I->DstOp;
4839 unsigned MinAlign = (I->Flags & TB_ALIGN_MASK) >> TB_ALIGN_SHIFT;
4840 if (Align < MinAlign)
4842 bool NarrowToMOV32rm = false;
4844 const TargetRegisterInfo &TRI = *MF.getSubtarget().getRegisterInfo();
4845 const TargetRegisterClass *RC = getRegClass(MI.getDesc(), OpNum,
4847 unsigned RCSize = TRI.getRegSizeInBits(*RC) / 8;
4848 if (Size < RCSize) {
4849 // Check if it's safe to fold the load. If the size of the object is
4850 // narrower than the load width, then it's not.
4851 if (Opcode != X86::MOV64rm || RCSize != 8 || Size != 4)
4853 // If this is a 64-bit load, but the spill slot is 32, then we can do
4854 // a 32-bit load which is implicitly zero-extended. This likely is
4855 // due to live interval analysis remat'ing a load from stack slot.
4856 if (MI.getOperand(0).getSubReg() || MI.getOperand(1).getSubReg())
4858 Opcode = X86::MOV32rm;
4859 NarrowToMOV32rm = true;
4864 NewMI = FuseTwoAddrInst(MF, Opcode, MOs, InsertPt, MI, *this);
4866 NewMI = FuseInst(MF, Opcode, OpNum, MOs, InsertPt, MI, *this);
4868 if (NarrowToMOV32rm) {
4869 // If this is the special case where we use a MOV32rm to load a 32-bit
4870 // value and zero-extend the top bits. Change the destination register
4872 unsigned DstReg = NewMI->getOperand(0).getReg();
4873 if (TargetRegisterInfo::isPhysicalRegister(DstReg))
4874 NewMI->getOperand(0).setReg(RI.getSubReg(DstReg, X86::sub_32bit));
4876 NewMI->getOperand(0).setSubReg(X86::sub_32bit);
4881 // If the instruction and target operand are commutable, commute the
4882 // instruction and try again.
4884 unsigned CommuteOpIdx1 = OpNum, CommuteOpIdx2 = CommuteAnyOperandIndex;
4885 if (findCommutedOpIndices(MI, CommuteOpIdx1, CommuteOpIdx2)) {
4886 bool HasDef = MI.getDesc().getNumDefs();
4887 unsigned Reg0 = HasDef ? MI.getOperand(0).getReg() : 0;
4888 unsigned Reg1 = MI.getOperand(CommuteOpIdx1).getReg();
4889 unsigned Reg2 = MI.getOperand(CommuteOpIdx2).getReg();
4891 0 == MI.getDesc().getOperandConstraint(CommuteOpIdx1, MCOI::TIED_TO);
4893 0 == MI.getDesc().getOperandConstraint(CommuteOpIdx2, MCOI::TIED_TO);
4895 // If either of the commutable operands are tied to the destination
4896 // then we can not commute + fold.
4897 if ((HasDef && Reg0 == Reg1 && Tied1) ||
4898 (HasDef && Reg0 == Reg2 && Tied2))
4901 MachineInstr *CommutedMI =
4902 commuteInstruction(MI, false, CommuteOpIdx1, CommuteOpIdx2);
4904 // Unable to commute.
4907 if (CommutedMI != &MI) {
4908 // New instruction. We can't fold from this.
4909 CommutedMI->eraseFromParent();
4913 // Attempt to fold with the commuted version of the instruction.
4914 NewMI = foldMemoryOperandImpl(MF, MI, CommuteOpIdx2, MOs, InsertPt,
4915 Size, Align, /*AllowCommute=*/false);
4919 // Folding failed again - undo the commute before returning.
4920 MachineInstr *UncommutedMI =
4921 commuteInstruction(MI, false, CommuteOpIdx1, CommuteOpIdx2);
4922 if (!UncommutedMI) {
4923 // Unable to commute.
4926 if (UncommutedMI != &MI) {
4927 // New instruction. It doesn't need to be kept.
4928 UncommutedMI->eraseFromParent();
4932 // Return here to prevent duplicate fuse failure report.
4938 if (PrintFailedFusing && !MI.isCopy())
4939 dbgs() << "We failed to fuse operand " << OpNum << " in " << MI;
4944 X86InstrInfo::foldMemoryOperandImpl(MachineFunction &MF, MachineInstr &MI,
4945 ArrayRef<unsigned> Ops,
4946 MachineBasicBlock::iterator InsertPt,
4947 int FrameIndex, LiveIntervals *LIS) const {
4948 // Check switch flag
4952 // Avoid partial and undef register update stalls unless optimizing for size.
4953 if (!MF.getFunction().optForSize() &&
4954 (hasPartialRegUpdate(MI.getOpcode(), Subtarget, /*ForLoadFold*/true) ||
4955 shouldPreventUndefRegUpdateMemFold(MF, MI)))
4958 // Don't fold subreg spills, or reloads that use a high subreg.
4959 for (auto Op : Ops) {
4960 MachineOperand &MO = MI.getOperand(Op);
4961 auto SubReg = MO.getSubReg();
4962 if (SubReg && (MO.isDef() || SubReg == X86::sub_8bit_hi))
4966 const MachineFrameInfo &MFI = MF.getFrameInfo();
4967 unsigned Size = MFI.getObjectSize(FrameIndex);
4968 unsigned Alignment = MFI.getObjectAlignment(FrameIndex);
4969 // If the function stack isn't realigned we don't want to fold instructions
4970 // that need increased alignment.
4971 if (!RI.needsStackRealignment(MF))
4973 std::min(Alignment, Subtarget.getFrameLowering()->getStackAlignment());
4974 if (Ops.size() == 2 && Ops[0] == 0 && Ops[1] == 1) {
4975 unsigned NewOpc = 0;
4976 unsigned RCSize = 0;
4977 switch (MI.getOpcode()) {
4978 default: return nullptr;
4979 case X86::TEST8rr: NewOpc = X86::CMP8ri; RCSize = 1; break;
4980 case X86::TEST16rr: NewOpc = X86::CMP16ri8; RCSize = 2; break;
4981 case X86::TEST32rr: NewOpc = X86::CMP32ri8; RCSize = 4; break;
4982 case X86::TEST64rr: NewOpc = X86::CMP64ri8; RCSize = 8; break;
4984 // Check if it's safe to fold the load. If the size of the object is
4985 // narrower than the load width, then it's not.
4988 // Change to CMPXXri r, 0 first.
4989 MI.setDesc(get(NewOpc));
4990 MI.getOperand(1).ChangeToImmediate(0);
4991 } else if (Ops.size() != 1)
4994 return foldMemoryOperandImpl(MF, MI, Ops[0],
4995 MachineOperand::CreateFI(FrameIndex), InsertPt,
4996 Size, Alignment, /*AllowCommute=*/true);
4999 /// Check if \p LoadMI is a partial register load that we can't fold into \p MI
5000 /// because the latter uses contents that wouldn't be defined in the folded
5001 /// version. For instance, this transformation isn't legal:
5002 /// movss (%rdi), %xmm0
5003 /// addps %xmm0, %xmm0
5005 /// addps (%rdi), %xmm0
5007 /// But this one is:
5008 /// movss (%rdi), %xmm0
5009 /// addss %xmm0, %xmm0
5011 /// addss (%rdi), %xmm0
5013 static bool isNonFoldablePartialRegisterLoad(const MachineInstr &LoadMI,
5014 const MachineInstr &UserMI,
5015 const MachineFunction &MF) {
5016 unsigned Opc = LoadMI.getOpcode();
5017 unsigned UserOpc = UserMI.getOpcode();
5018 const TargetRegisterInfo &TRI = *MF.getSubtarget().getRegisterInfo();
5019 const TargetRegisterClass *RC =
5020 MF.getRegInfo().getRegClass(LoadMI.getOperand(0).getReg());
5021 unsigned RegSize = TRI.getRegSizeInBits(*RC);
5023 if ((Opc == X86::MOVSSrm || Opc == X86::VMOVSSrm || Opc == X86::VMOVSSZrm) &&
5025 // These instructions only load 32 bits, we can't fold them if the
5026 // destination register is wider than 32 bits (4 bytes), and its user
5027 // instruction isn't scalar (SS).
5029 case X86::ADDSSrr_Int: case X86::VADDSSrr_Int: case X86::VADDSSZrr_Int:
5030 case X86::CMPSSrr_Int: case X86::VCMPSSrr_Int: case X86::VCMPSSZrr_Int:
5031 case X86::DIVSSrr_Int: case X86::VDIVSSrr_Int: case X86::VDIVSSZrr_Int:
5032 case X86::MAXSSrr_Int: case X86::VMAXSSrr_Int: case X86::VMAXSSZrr_Int:
5033 case X86::MINSSrr_Int: case X86::VMINSSrr_Int: case X86::VMINSSZrr_Int:
5034 case X86::MULSSrr_Int: case X86::VMULSSrr_Int: case X86::VMULSSZrr_Int:
5035 case X86::SUBSSrr_Int: case X86::VSUBSSrr_Int: case X86::VSUBSSZrr_Int:
5036 case X86::VADDSSZrr_Intk: case X86::VADDSSZrr_Intkz:
5037 case X86::VDIVSSZrr_Intk: case X86::VDIVSSZrr_Intkz:
5038 case X86::VMAXSSZrr_Intk: case X86::VMAXSSZrr_Intkz:
5039 case X86::VMINSSZrr_Intk: case X86::VMINSSZrr_Intkz:
5040 case X86::VMULSSZrr_Intk: case X86::VMULSSZrr_Intkz:
5041 case X86::VSUBSSZrr_Intk: case X86::VSUBSSZrr_Intkz:
5042 case X86::VFMADDSS4rr_Int: case X86::VFNMADDSS4rr_Int:
5043 case X86::VFMSUBSS4rr_Int: case X86::VFNMSUBSS4rr_Int:
5044 case X86::VFMADD132SSr_Int: case X86::VFNMADD132SSr_Int:
5045 case X86::VFMADD213SSr_Int: case X86::VFNMADD213SSr_Int:
5046 case X86::VFMADD231SSr_Int: case X86::VFNMADD231SSr_Int:
5047 case X86::VFMSUB132SSr_Int: case X86::VFNMSUB132SSr_Int:
5048 case X86::VFMSUB213SSr_Int: case X86::VFNMSUB213SSr_Int:
5049 case X86::VFMSUB231SSr_Int: case X86::VFNMSUB231SSr_Int:
5050 case X86::VFMADD132SSZr_Int: case X86::VFNMADD132SSZr_Int:
5051 case X86::VFMADD213SSZr_Int: case X86::VFNMADD213SSZr_Int:
5052 case X86::VFMADD231SSZr_Int: case X86::VFNMADD231SSZr_Int:
5053 case X86::VFMSUB132SSZr_Int: case X86::VFNMSUB132SSZr_Int:
5054 case X86::VFMSUB213SSZr_Int: case X86::VFNMSUB213SSZr_Int:
5055 case X86::VFMSUB231SSZr_Int: case X86::VFNMSUB231SSZr_Int:
5056 case X86::VFMADD132SSZr_Intk: case X86::VFNMADD132SSZr_Intk:
5057 case X86::VFMADD213SSZr_Intk: case X86::VFNMADD213SSZr_Intk:
5058 case X86::VFMADD231SSZr_Intk: case X86::VFNMADD231SSZr_Intk:
5059 case X86::VFMSUB132SSZr_Intk: case X86::VFNMSUB132SSZr_Intk:
5060 case X86::VFMSUB213SSZr_Intk: case X86::VFNMSUB213SSZr_Intk:
5061 case X86::VFMSUB231SSZr_Intk: case X86::VFNMSUB231SSZr_Intk:
5062 case X86::VFMADD132SSZr_Intkz: case X86::VFNMADD132SSZr_Intkz:
5063 case X86::VFMADD213SSZr_Intkz: case X86::VFNMADD213SSZr_Intkz:
5064 case X86::VFMADD231SSZr_Intkz: case X86::VFNMADD231SSZr_Intkz:
5065 case X86::VFMSUB132SSZr_Intkz: case X86::VFNMSUB132SSZr_Intkz:
5066 case X86::VFMSUB213SSZr_Intkz: case X86::VFNMSUB213SSZr_Intkz:
5067 case X86::VFMSUB231SSZr_Intkz: case X86::VFNMSUB231SSZr_Intkz:
5074 if ((Opc == X86::MOVSDrm || Opc == X86::VMOVSDrm || Opc == X86::VMOVSDZrm) &&
5076 // These instructions only load 64 bits, we can't fold them if the
5077 // destination register is wider than 64 bits (8 bytes), and its user
5078 // instruction isn't scalar (SD).
5080 case X86::ADDSDrr_Int: case X86::VADDSDrr_Int: case X86::VADDSDZrr_Int:
5081 case X86::CMPSDrr_Int: case X86::VCMPSDrr_Int: case X86::VCMPSDZrr_Int:
5082 case X86::DIVSDrr_Int: case X86::VDIVSDrr_Int: case X86::VDIVSDZrr_Int:
5083 case X86::MAXSDrr_Int: case X86::VMAXSDrr_Int: case X86::VMAXSDZrr_Int:
5084 case X86::MINSDrr_Int: case X86::VMINSDrr_Int: case X86::VMINSDZrr_Int:
5085 case X86::MULSDrr_Int: case X86::VMULSDrr_Int: case X86::VMULSDZrr_Int:
5086 case X86::SUBSDrr_Int: case X86::VSUBSDrr_Int: case X86::VSUBSDZrr_Int:
5087 case X86::VADDSDZrr_Intk: case X86::VADDSDZrr_Intkz:
5088 case X86::VDIVSDZrr_Intk: case X86::VDIVSDZrr_Intkz:
5089 case X86::VMAXSDZrr_Intk: case X86::VMAXSDZrr_Intkz:
5090 case X86::VMINSDZrr_Intk: case X86::VMINSDZrr_Intkz:
5091 case X86::VMULSDZrr_Intk: case X86::VMULSDZrr_Intkz:
5092 case X86::VSUBSDZrr_Intk: case X86::VSUBSDZrr_Intkz:
5093 case X86::VFMADDSD4rr_Int: case X86::VFNMADDSD4rr_Int:
5094 case X86::VFMSUBSD4rr_Int: case X86::VFNMSUBSD4rr_Int:
5095 case X86::VFMADD132SDr_Int: case X86::VFNMADD132SDr_Int:
5096 case X86::VFMADD213SDr_Int: case X86::VFNMADD213SDr_Int:
5097 case X86::VFMADD231SDr_Int: case X86::VFNMADD231SDr_Int:
5098 case X86::VFMSUB132SDr_Int: case X86::VFNMSUB132SDr_Int:
5099 case X86::VFMSUB213SDr_Int: case X86::VFNMSUB213SDr_Int:
5100 case X86::VFMSUB231SDr_Int: case X86::VFNMSUB231SDr_Int:
5101 case X86::VFMADD132SDZr_Int: case X86::VFNMADD132SDZr_Int:
5102 case X86::VFMADD213SDZr_Int: case X86::VFNMADD213SDZr_Int:
5103 case X86::VFMADD231SDZr_Int: case X86::VFNMADD231SDZr_Int:
5104 case X86::VFMSUB132SDZr_Int: case X86::VFNMSUB132SDZr_Int:
5105 case X86::VFMSUB213SDZr_Int: case X86::VFNMSUB213SDZr_Int:
5106 case X86::VFMSUB231SDZr_Int: case X86::VFNMSUB231SDZr_Int:
5107 case X86::VFMADD132SDZr_Intk: case X86::VFNMADD132SDZr_Intk:
5108 case X86::VFMADD213SDZr_Intk: case X86::VFNMADD213SDZr_Intk:
5109 case X86::VFMADD231SDZr_Intk: case X86::VFNMADD231SDZr_Intk:
5110 case X86::VFMSUB132SDZr_Intk: case X86::VFNMSUB132SDZr_Intk:
5111 case X86::VFMSUB213SDZr_Intk: case X86::VFNMSUB213SDZr_Intk:
5112 case X86::VFMSUB231SDZr_Intk: case X86::VFNMSUB231SDZr_Intk:
5113 case X86::VFMADD132SDZr_Intkz: case X86::VFNMADD132SDZr_Intkz:
5114 case X86::VFMADD213SDZr_Intkz: case X86::VFNMADD213SDZr_Intkz:
5115 case X86::VFMADD231SDZr_Intkz: case X86::VFNMADD231SDZr_Intkz:
5116 case X86::VFMSUB132SDZr_Intkz: case X86::VFNMSUB132SDZr_Intkz:
5117 case X86::VFMSUB213SDZr_Intkz: case X86::VFNMSUB213SDZr_Intkz:
5118 case X86::VFMSUB231SDZr_Intkz: case X86::VFNMSUB231SDZr_Intkz:
5128 MachineInstr *X86InstrInfo::foldMemoryOperandImpl(
5129 MachineFunction &MF, MachineInstr &MI, ArrayRef<unsigned> Ops,
5130 MachineBasicBlock::iterator InsertPt, MachineInstr &LoadMI,
5131 LiveIntervals *LIS) const {
5133 // TODO: Support the case where LoadMI loads a wide register, but MI
5134 // only uses a subreg.
5135 for (auto Op : Ops) {
5136 if (MI.getOperand(Op).getSubReg())
5140 // If loading from a FrameIndex, fold directly from the FrameIndex.
5141 unsigned NumOps = LoadMI.getDesc().getNumOperands();
5143 if (isLoadFromStackSlot(LoadMI, FrameIndex)) {
5144 if (isNonFoldablePartialRegisterLoad(LoadMI, MI, MF))
5146 return foldMemoryOperandImpl(MF, MI, Ops, InsertPt, FrameIndex, LIS);
5149 // Check switch flag
5150 if (NoFusing) return nullptr;
5152 // Avoid partial and undef register update stalls unless optimizing for size.
5153 if (!MF.getFunction().optForSize() &&
5154 (hasPartialRegUpdate(MI.getOpcode(), Subtarget, /*ForLoadFold*/true) ||
5155 shouldPreventUndefRegUpdateMemFold(MF, MI)))
5158 // Determine the alignment of the load.
5159 unsigned Alignment = 0;
5160 if (LoadMI.hasOneMemOperand())
5161 Alignment = (*LoadMI.memoperands_begin())->getAlignment();
5163 switch (LoadMI.getOpcode()) {
5164 case X86::AVX512_512_SET0:
5165 case X86::AVX512_512_SETALLONES:
5168 case X86::AVX2_SETALLONES:
5169 case X86::AVX1_SETALLONES:
5171 case X86::AVX512_256_SET0:
5175 case X86::V_SETALLONES:
5176 case X86::AVX512_128_SET0:
5181 case X86::AVX512_FsFLD0SD:
5185 case X86::AVX512_FsFLD0SS:
5191 if (Ops.size() == 2 && Ops[0] == 0 && Ops[1] == 1) {
5192 unsigned NewOpc = 0;
5193 switch (MI.getOpcode()) {
5194 default: return nullptr;
5195 case X86::TEST8rr: NewOpc = X86::CMP8ri; break;
5196 case X86::TEST16rr: NewOpc = X86::CMP16ri8; break;
5197 case X86::TEST32rr: NewOpc = X86::CMP32ri8; break;
5198 case X86::TEST64rr: NewOpc = X86::CMP64ri8; break;
5200 // Change to CMPXXri r, 0 first.
5201 MI.setDesc(get(NewOpc));
5202 MI.getOperand(1).ChangeToImmediate(0);
5203 } else if (Ops.size() != 1)
5206 // Make sure the subregisters match.
5207 // Otherwise we risk changing the size of the load.
5208 if (LoadMI.getOperand(0).getSubReg() != MI.getOperand(Ops[0]).getSubReg())
5211 SmallVector<MachineOperand,X86::AddrNumOperands> MOs;
5212 switch (LoadMI.getOpcode()) {
5215 case X86::V_SETALLONES:
5216 case X86::AVX2_SETALLONES:
5217 case X86::AVX1_SETALLONES:
5219 case X86::AVX512_128_SET0:
5220 case X86::AVX512_256_SET0:
5221 case X86::AVX512_512_SET0:
5222 case X86::AVX512_512_SETALLONES:
5224 case X86::AVX512_FsFLD0SD:
5226 case X86::AVX512_FsFLD0SS: {
5227 // Folding a V_SET0 or V_SETALLONES as a load, to ease register pressure.
5228 // Create a constant-pool entry and operands to load from it.
5230 // Medium and large mode can't fold loads this way.
5231 if (MF.getTarget().getCodeModel() != CodeModel::Small &&
5232 MF.getTarget().getCodeModel() != CodeModel::Kernel)
5235 // x86-32 PIC requires a PIC base register for constant pools.
5236 unsigned PICBase = 0;
5237 if (MF.getTarget().isPositionIndependent()) {
5238 if (Subtarget.is64Bit())
5241 // FIXME: PICBase = getGlobalBaseReg(&MF);
5242 // This doesn't work for several reasons.
5243 // 1. GlobalBaseReg may have been spilled.
5244 // 2. It may not be live at MI.
5248 // Create a constant-pool entry.
5249 MachineConstantPool &MCP = *MF.getConstantPool();
5251 unsigned Opc = LoadMI.getOpcode();
5252 if (Opc == X86::FsFLD0SS || Opc == X86::AVX512_FsFLD0SS)
5253 Ty = Type::getFloatTy(MF.getFunction().getContext());
5254 else if (Opc == X86::FsFLD0SD || Opc == X86::AVX512_FsFLD0SD)
5255 Ty = Type::getDoubleTy(MF.getFunction().getContext());
5256 else if (Opc == X86::AVX512_512_SET0 || Opc == X86::AVX512_512_SETALLONES)
5257 Ty = VectorType::get(Type::getInt32Ty(MF.getFunction().getContext()),16);
5258 else if (Opc == X86::AVX2_SETALLONES || Opc == X86::AVX_SET0 ||
5259 Opc == X86::AVX512_256_SET0 || Opc == X86::AVX1_SETALLONES)
5260 Ty = VectorType::get(Type::getInt32Ty(MF.getFunction().getContext()), 8);
5261 else if (Opc == X86::MMX_SET0)
5262 Ty = VectorType::get(Type::getInt32Ty(MF.getFunction().getContext()), 2);
5264 Ty = VectorType::get(Type::getInt32Ty(MF.getFunction().getContext()), 4);
5266 bool IsAllOnes = (Opc == X86::V_SETALLONES || Opc == X86::AVX2_SETALLONES ||
5267 Opc == X86::AVX512_512_SETALLONES ||
5268 Opc == X86::AVX1_SETALLONES);
5269 const Constant *C = IsAllOnes ? Constant::getAllOnesValue(Ty) :
5270 Constant::getNullValue(Ty);
5271 unsigned CPI = MCP.getConstantPoolIndex(C, Alignment);
5273 // Create operands to load from the constant pool entry.
5274 MOs.push_back(MachineOperand::CreateReg(PICBase, false));
5275 MOs.push_back(MachineOperand::CreateImm(1));
5276 MOs.push_back(MachineOperand::CreateReg(0, false));
5277 MOs.push_back(MachineOperand::CreateCPI(CPI, 0));
5278 MOs.push_back(MachineOperand::CreateReg(0, false));
5282 if (isNonFoldablePartialRegisterLoad(LoadMI, MI, MF))
5285 // Folding a normal load. Just copy the load's address operands.
5286 MOs.append(LoadMI.operands_begin() + NumOps - X86::AddrNumOperands,
5287 LoadMI.operands_begin() + NumOps);
5291 return foldMemoryOperandImpl(MF, MI, Ops[0], MOs, InsertPt,
5292 /*Size=*/0, Alignment, /*AllowCommute=*/true);
5295 static SmallVector<MachineMemOperand *, 2>
5296 extractLoadMMOs(ArrayRef<MachineMemOperand *> MMOs, MachineFunction &MF) {
5297 SmallVector<MachineMemOperand *, 2> LoadMMOs;
5299 for (MachineMemOperand *MMO : MMOs) {
5303 if (!MMO->isStore()) {
5305 LoadMMOs.push_back(MMO);
5307 // Clone the MMO and unset the store flag.
5308 LoadMMOs.push_back(MF.getMachineMemOperand(
5309 MMO->getPointerInfo(), MMO->getFlags() & ~MachineMemOperand::MOStore,
5310 MMO->getSize(), MMO->getBaseAlignment(), MMO->getAAInfo(), nullptr,
5311 MMO->getSyncScopeID(), MMO->getOrdering(),
5312 MMO->getFailureOrdering()));
5319 static SmallVector<MachineMemOperand *, 2>
5320 extractStoreMMOs(ArrayRef<MachineMemOperand *> MMOs, MachineFunction &MF) {
5321 SmallVector<MachineMemOperand *, 2> StoreMMOs;
5323 for (MachineMemOperand *MMO : MMOs) {
5324 if (!MMO->isStore())
5327 if (!MMO->isLoad()) {
5329 StoreMMOs.push_back(MMO);
5331 // Clone the MMO and unset the load flag.
5332 StoreMMOs.push_back(MF.getMachineMemOperand(
5333 MMO->getPointerInfo(), MMO->getFlags() & ~MachineMemOperand::MOLoad,
5334 MMO->getSize(), MMO->getBaseAlignment(), MMO->getAAInfo(), nullptr,
5335 MMO->getSyncScopeID(), MMO->getOrdering(),
5336 MMO->getFailureOrdering()));
5343 bool X86InstrInfo::unfoldMemoryOperand(
5344 MachineFunction &MF, MachineInstr &MI, unsigned Reg, bool UnfoldLoad,
5345 bool UnfoldStore, SmallVectorImpl<MachineInstr *> &NewMIs) const {
5346 const X86MemoryFoldTableEntry *I = lookupUnfoldTable(MI.getOpcode());
5349 unsigned Opc = I->DstOp;
5350 unsigned Index = I->Flags & TB_INDEX_MASK;
5351 bool FoldedLoad = I->Flags & TB_FOLDED_LOAD;
5352 bool FoldedStore = I->Flags & TB_FOLDED_STORE;
5353 if (UnfoldLoad && !FoldedLoad)
5355 UnfoldLoad &= FoldedLoad;
5356 if (UnfoldStore && !FoldedStore)
5358 UnfoldStore &= FoldedStore;
5360 const MCInstrDesc &MCID = get(Opc);
5361 const TargetRegisterClass *RC = getRegClass(MCID, Index, &RI, MF);
5362 // TODO: Check if 32-byte or greater accesses are slow too?
5363 if (!MI.hasOneMemOperand() && RC == &X86::VR128RegClass &&
5364 Subtarget.isUnalignedMem16Slow())
5365 // Without memoperands, loadRegFromAddr and storeRegToStackSlot will
5366 // conservatively assume the address is unaligned. That's bad for
5369 SmallVector<MachineOperand, X86::AddrNumOperands> AddrOps;
5370 SmallVector<MachineOperand,2> BeforeOps;
5371 SmallVector<MachineOperand,2> AfterOps;
5372 SmallVector<MachineOperand,4> ImpOps;
5373 for (unsigned i = 0, e = MI.getNumOperands(); i != e; ++i) {
5374 MachineOperand &Op = MI.getOperand(i);
5375 if (i >= Index && i < Index + X86::AddrNumOperands)
5376 AddrOps.push_back(Op);
5377 else if (Op.isReg() && Op.isImplicit())
5378 ImpOps.push_back(Op);
5380 BeforeOps.push_back(Op);
5382 AfterOps.push_back(Op);
5385 // Emit the load instruction.
5387 auto MMOs = extractLoadMMOs(MI.memoperands(), MF);
5388 loadRegFromAddr(MF, Reg, AddrOps, RC, MMOs, NewMIs);
5390 // Address operands cannot be marked isKill.
5391 for (unsigned i = 1; i != 1 + X86::AddrNumOperands; ++i) {
5392 MachineOperand &MO = NewMIs[0]->getOperand(i);
5394 MO.setIsKill(false);
5399 // Emit the data processing instruction.
5400 MachineInstr *DataMI = MF.CreateMachineInstr(MCID, MI.getDebugLoc(), true);
5401 MachineInstrBuilder MIB(MF, DataMI);
5404 MIB.addReg(Reg, RegState::Define);
5405 for (MachineOperand &BeforeOp : BeforeOps)
5409 for (MachineOperand &AfterOp : AfterOps)
5411 for (MachineOperand &ImpOp : ImpOps) {
5412 MIB.addReg(ImpOp.getReg(),
5413 getDefRegState(ImpOp.isDef()) |
5414 RegState::Implicit |
5415 getKillRegState(ImpOp.isKill()) |
5416 getDeadRegState(ImpOp.isDead()) |
5417 getUndefRegState(ImpOp.isUndef()));
5419 // Change CMP32ri r, 0 back to TEST32rr r, r, etc.
5420 switch (DataMI->getOpcode()) {
5422 case X86::CMP64ri32:
5429 MachineOperand &MO0 = DataMI->getOperand(0);
5430 MachineOperand &MO1 = DataMI->getOperand(1);
5431 if (MO1.getImm() == 0) {
5433 switch (DataMI->getOpcode()) {
5434 default: llvm_unreachable("Unreachable!");
5436 case X86::CMP64ri32: NewOpc = X86::TEST64rr; break;
5438 case X86::CMP32ri: NewOpc = X86::TEST32rr; break;
5440 case X86::CMP16ri: NewOpc = X86::TEST16rr; break;
5441 case X86::CMP8ri: NewOpc = X86::TEST8rr; break;
5443 DataMI->setDesc(get(NewOpc));
5444 MO1.ChangeToRegister(MO0.getReg(), false);
5448 NewMIs.push_back(DataMI);
5450 // Emit the store instruction.
5452 const TargetRegisterClass *DstRC = getRegClass(MCID, 0, &RI, MF);
5453 auto MMOs = extractStoreMMOs(MI.memoperands(), MF);
5454 storeRegToAddr(MF, Reg, true, AddrOps, DstRC, MMOs, NewMIs);
5461 X86InstrInfo::unfoldMemoryOperand(SelectionDAG &DAG, SDNode *N,
5462 SmallVectorImpl<SDNode*> &NewNodes) const {
5463 if (!N->isMachineOpcode())
5466 const X86MemoryFoldTableEntry *I = lookupUnfoldTable(N->getMachineOpcode());
5469 unsigned Opc = I->DstOp;
5470 unsigned Index = I->Flags & TB_INDEX_MASK;
5471 bool FoldedLoad = I->Flags & TB_FOLDED_LOAD;
5472 bool FoldedStore = I->Flags & TB_FOLDED_STORE;
5473 const MCInstrDesc &MCID = get(Opc);
5474 MachineFunction &MF = DAG.getMachineFunction();
5475 const TargetRegisterInfo &TRI = *MF.getSubtarget().getRegisterInfo();
5476 const TargetRegisterClass *RC = getRegClass(MCID, Index, &RI, MF);
5477 unsigned NumDefs = MCID.NumDefs;
5478 std::vector<SDValue> AddrOps;
5479 std::vector<SDValue> BeforeOps;
5480 std::vector<SDValue> AfterOps;
5482 unsigned NumOps = N->getNumOperands();
5483 for (unsigned i = 0; i != NumOps-1; ++i) {
5484 SDValue Op = N->getOperand(i);
5485 if (i >= Index-NumDefs && i < Index-NumDefs + X86::AddrNumOperands)
5486 AddrOps.push_back(Op);
5487 else if (i < Index-NumDefs)
5488 BeforeOps.push_back(Op);
5489 else if (i > Index-NumDefs)
5490 AfterOps.push_back(Op);
5492 SDValue Chain = N->getOperand(NumOps-1);
5493 AddrOps.push_back(Chain);
5495 // Emit the load instruction.
5496 SDNode *Load = nullptr;
5498 EVT VT = *TRI.legalclasstypes_begin(*RC);
5499 auto MMOs = extractLoadMMOs(cast<MachineSDNode>(N)->memoperands(), MF);
5500 if (MMOs.empty() && RC == &X86::VR128RegClass &&
5501 Subtarget.isUnalignedMem16Slow())
5502 // Do not introduce a slow unaligned load.
5504 // FIXME: If a VR128 can have size 32, we should be checking if a 32-byte
5505 // memory access is slow above.
5506 unsigned Alignment = std::max<uint32_t>(TRI.getSpillSize(*RC), 16);
5507 bool isAligned = !MMOs.empty() && MMOs.front()->getAlignment() >= Alignment;
5508 Load = DAG.getMachineNode(getLoadRegOpcode(0, RC, isAligned, Subtarget), dl,
5509 VT, MVT::Other, AddrOps);
5510 NewNodes.push_back(Load);
5512 // Preserve memory reference information.
5513 DAG.setNodeMemRefs(cast<MachineSDNode>(Load), MMOs);
5516 // Emit the data processing instruction.
5517 std::vector<EVT> VTs;
5518 const TargetRegisterClass *DstRC = nullptr;
5519 if (MCID.getNumDefs() > 0) {
5520 DstRC = getRegClass(MCID, 0, &RI, MF);
5521 VTs.push_back(*TRI.legalclasstypes_begin(*DstRC));
5523 for (unsigned i = 0, e = N->getNumValues(); i != e; ++i) {
5524 EVT VT = N->getValueType(i);
5525 if (VT != MVT::Other && i >= (unsigned)MCID.getNumDefs())
5529 BeforeOps.push_back(SDValue(Load, 0));
5530 BeforeOps.insert(BeforeOps.end(), AfterOps.begin(), AfterOps.end());
5531 // Change CMP32ri r, 0 back to TEST32rr r, r, etc.
5534 case X86::CMP64ri32:
5541 if (isNullConstant(BeforeOps[1])) {
5543 default: llvm_unreachable("Unreachable!");
5545 case X86::CMP64ri32: Opc = X86::TEST64rr; break;
5547 case X86::CMP32ri: Opc = X86::TEST32rr; break;
5549 case X86::CMP16ri: Opc = X86::TEST16rr; break;
5550 case X86::CMP8ri: Opc = X86::TEST8rr; break;
5552 BeforeOps[1] = BeforeOps[0];
5555 SDNode *NewNode= DAG.getMachineNode(Opc, dl, VTs, BeforeOps);
5556 NewNodes.push_back(NewNode);
5558 // Emit the store instruction.
5561 AddrOps.push_back(SDValue(NewNode, 0));
5562 AddrOps.push_back(Chain);
5563 auto MMOs = extractStoreMMOs(cast<MachineSDNode>(N)->memoperands(), MF);
5564 if (MMOs.empty() && RC == &X86::VR128RegClass &&
5565 Subtarget.isUnalignedMem16Slow())
5566 // Do not introduce a slow unaligned store.
5568 // FIXME: If a VR128 can have size 32, we should be checking if a 32-byte
5569 // memory access is slow above.
5570 unsigned Alignment = std::max<uint32_t>(TRI.getSpillSize(*RC), 16);
5571 bool isAligned = !MMOs.empty() && MMOs.front()->getAlignment() >= Alignment;
5573 DAG.getMachineNode(getStoreRegOpcode(0, DstRC, isAligned, Subtarget),
5574 dl, MVT::Other, AddrOps);
5575 NewNodes.push_back(Store);
5577 // Preserve memory reference information.
5578 DAG.setNodeMemRefs(cast<MachineSDNode>(Store), MMOs);
5584 unsigned X86InstrInfo::getOpcodeAfterMemoryUnfold(unsigned Opc,
5585 bool UnfoldLoad, bool UnfoldStore,
5586 unsigned *LoadRegIndex) const {
5587 const X86MemoryFoldTableEntry *I = lookupUnfoldTable(Opc);
5590 bool FoldedLoad = I->Flags & TB_FOLDED_LOAD;
5591 bool FoldedStore = I->Flags & TB_FOLDED_STORE;
5592 if (UnfoldLoad && !FoldedLoad)
5594 if (UnfoldStore && !FoldedStore)
5597 *LoadRegIndex = I->Flags & TB_INDEX_MASK;
5602 X86InstrInfo::areLoadsFromSameBasePtr(SDNode *Load1, SDNode *Load2,
5603 int64_t &Offset1, int64_t &Offset2) const {
5604 if (!Load1->isMachineOpcode() || !Load2->isMachineOpcode())
5606 unsigned Opc1 = Load1->getMachineOpcode();
5607 unsigned Opc2 = Load2->getMachineOpcode();
5609 default: return false;
5619 case X86::MMX_MOVD64rm:
5620 case X86::MMX_MOVQ64rm:
5627 // AVX load instructions
5630 case X86::VMOVAPSrm:
5631 case X86::VMOVUPSrm:
5632 case X86::VMOVAPDrm:
5633 case X86::VMOVUPDrm:
5634 case X86::VMOVDQArm:
5635 case X86::VMOVDQUrm:
5636 case X86::VMOVAPSYrm:
5637 case X86::VMOVUPSYrm:
5638 case X86::VMOVAPDYrm:
5639 case X86::VMOVUPDYrm:
5640 case X86::VMOVDQAYrm:
5641 case X86::VMOVDQUYrm:
5642 // AVX512 load instructions
5643 case X86::VMOVSSZrm:
5644 case X86::VMOVSDZrm:
5645 case X86::VMOVAPSZ128rm:
5646 case X86::VMOVUPSZ128rm:
5647 case X86::VMOVAPSZ128rm_NOVLX:
5648 case X86::VMOVUPSZ128rm_NOVLX:
5649 case X86::VMOVAPDZ128rm:
5650 case X86::VMOVUPDZ128rm:
5651 case X86::VMOVDQU8Z128rm:
5652 case X86::VMOVDQU16Z128rm:
5653 case X86::VMOVDQA32Z128rm:
5654 case X86::VMOVDQU32Z128rm:
5655 case X86::VMOVDQA64Z128rm:
5656 case X86::VMOVDQU64Z128rm:
5657 case X86::VMOVAPSZ256rm:
5658 case X86::VMOVUPSZ256rm:
5659 case X86::VMOVAPSZ256rm_NOVLX:
5660 case X86::VMOVUPSZ256rm_NOVLX:
5661 case X86::VMOVAPDZ256rm:
5662 case X86::VMOVUPDZ256rm:
5663 case X86::VMOVDQU8Z256rm:
5664 case X86::VMOVDQU16Z256rm:
5665 case X86::VMOVDQA32Z256rm:
5666 case X86::VMOVDQU32Z256rm:
5667 case X86::VMOVDQA64Z256rm:
5668 case X86::VMOVDQU64Z256rm:
5669 case X86::VMOVAPSZrm:
5670 case X86::VMOVUPSZrm:
5671 case X86::VMOVAPDZrm:
5672 case X86::VMOVUPDZrm:
5673 case X86::VMOVDQU8Zrm:
5674 case X86::VMOVDQU16Zrm:
5675 case X86::VMOVDQA32Zrm:
5676 case X86::VMOVDQU32Zrm:
5677 case X86::VMOVDQA64Zrm:
5678 case X86::VMOVDQU64Zrm:
5686 default: return false;
5696 case X86::MMX_MOVD64rm:
5697 case X86::MMX_MOVQ64rm:
5704 // AVX load instructions
5707 case X86::VMOVAPSrm:
5708 case X86::VMOVUPSrm:
5709 case X86::VMOVAPDrm:
5710 case X86::VMOVUPDrm:
5711 case X86::VMOVDQArm:
5712 case X86::VMOVDQUrm:
5713 case X86::VMOVAPSYrm:
5714 case X86::VMOVUPSYrm:
5715 case X86::VMOVAPDYrm:
5716 case X86::VMOVUPDYrm:
5717 case X86::VMOVDQAYrm:
5718 case X86::VMOVDQUYrm:
5719 // AVX512 load instructions
5720 case X86::VMOVSSZrm:
5721 case X86::VMOVSDZrm:
5722 case X86::VMOVAPSZ128rm:
5723 case X86::VMOVUPSZ128rm:
5724 case X86::VMOVAPSZ128rm_NOVLX:
5725 case X86::VMOVUPSZ128rm_NOVLX:
5726 case X86::VMOVAPDZ128rm:
5727 case X86::VMOVUPDZ128rm:
5728 case X86::VMOVDQU8Z128rm:
5729 case X86::VMOVDQU16Z128rm:
5730 case X86::VMOVDQA32Z128rm:
5731 case X86::VMOVDQU32Z128rm:
5732 case X86::VMOVDQA64Z128rm:
5733 case X86::VMOVDQU64Z128rm:
5734 case X86::VMOVAPSZ256rm:
5735 case X86::VMOVUPSZ256rm:
5736 case X86::VMOVAPSZ256rm_NOVLX:
5737 case X86::VMOVUPSZ256rm_NOVLX:
5738 case X86::VMOVAPDZ256rm:
5739 case X86::VMOVUPDZ256rm:
5740 case X86::VMOVDQU8Z256rm:
5741 case X86::VMOVDQU16Z256rm:
5742 case X86::VMOVDQA32Z256rm:
5743 case X86::VMOVDQU32Z256rm:
5744 case X86::VMOVDQA64Z256rm:
5745 case X86::VMOVDQU64Z256rm:
5746 case X86::VMOVAPSZrm:
5747 case X86::VMOVUPSZrm:
5748 case X86::VMOVAPDZrm:
5749 case X86::VMOVUPDZrm:
5750 case X86::VMOVDQU8Zrm:
5751 case X86::VMOVDQU16Zrm:
5752 case X86::VMOVDQA32Zrm:
5753 case X86::VMOVDQU32Zrm:
5754 case X86::VMOVDQA64Zrm:
5755 case X86::VMOVDQU64Zrm:
5763 // Lambda to check if both the loads have the same value for an operand index.
5764 auto HasSameOp = [&](int I) {
5765 return Load1->getOperand(I) == Load2->getOperand(I);
5768 // All operands except the displacement should match.
5769 if (!HasSameOp(X86::AddrBaseReg) || !HasSameOp(X86::AddrScaleAmt) ||
5770 !HasSameOp(X86::AddrIndexReg) || !HasSameOp(X86::AddrSegmentReg))
5773 // Chain Operand must be the same.
5777 // Now let's examine if the displacements are constants.
5778 auto Disp1 = dyn_cast<ConstantSDNode>(Load1->getOperand(X86::AddrDisp));
5779 auto Disp2 = dyn_cast<ConstantSDNode>(Load2->getOperand(X86::AddrDisp));
5780 if (!Disp1 || !Disp2)
5783 Offset1 = Disp1->getSExtValue();
5784 Offset2 = Disp2->getSExtValue();
5788 bool X86InstrInfo::shouldScheduleLoadsNear(SDNode *Load1, SDNode *Load2,
5789 int64_t Offset1, int64_t Offset2,
5790 unsigned NumLoads) const {
5791 assert(Offset2 > Offset1);
5792 if ((Offset2 - Offset1) / 8 > 64)
5795 unsigned Opc1 = Load1->getMachineOpcode();
5796 unsigned Opc2 = Load2->getMachineOpcode();
5798 return false; // FIXME: overly conservative?
5805 case X86::MMX_MOVD64rm:
5806 case X86::MMX_MOVQ64rm:
5810 EVT VT = Load1->getValueType(0);
5811 switch (VT.getSimpleVT().SimpleTy) {
5813 // XMM registers. In 64-bit mode we can be a bit more aggressive since we
5814 // have 16 of them to play with.
5815 if (Subtarget.is64Bit()) {
5818 } else if (NumLoads) {
5837 reverseBranchCondition(SmallVectorImpl<MachineOperand> &Cond) const {
5838 assert(Cond.size() == 1 && "Invalid X86 branch condition!");
5839 X86::CondCode CC = static_cast<X86::CondCode>(Cond[0].getImm());
5840 Cond[0].setImm(GetOppositeBranchCondition(CC));
5845 isSafeToMoveRegClassDefs(const TargetRegisterClass *RC) const {
5846 // FIXME: Return false for x87 stack register classes for now. We can't
5847 // allow any loads of these registers before FpGet_ST0_80.
5848 return !(RC == &X86::CCRRegClass || RC == &X86::DFCCRRegClass ||
5849 RC == &X86::RFP32RegClass || RC == &X86::RFP64RegClass ||
5850 RC == &X86::RFP80RegClass);
5853 /// Return a virtual register initialized with the
5854 /// the global base register value. Output instructions required to
5855 /// initialize the register in the function entry block, if necessary.
5857 /// TODO: Eliminate this and move the code to X86MachineFunctionInfo.
5859 unsigned X86InstrInfo::getGlobalBaseReg(MachineFunction *MF) const {
5860 assert((!Subtarget.is64Bit() ||
5861 MF->getTarget().getCodeModel() == CodeModel::Medium ||
5862 MF->getTarget().getCodeModel() == CodeModel::Large) &&
5863 "X86-64 PIC uses RIP relative addressing");
5865 X86MachineFunctionInfo *X86FI = MF->getInfo<X86MachineFunctionInfo>();
5866 unsigned GlobalBaseReg = X86FI->getGlobalBaseReg();
5867 if (GlobalBaseReg != 0)
5868 return GlobalBaseReg;
5870 // Create the register. The code to initialize it is inserted
5871 // later, by the CGBR pass (below).
5872 MachineRegisterInfo &RegInfo = MF->getRegInfo();
5873 GlobalBaseReg = RegInfo.createVirtualRegister(
5874 Subtarget.is64Bit() ? &X86::GR64_NOSPRegClass : &X86::GR32_NOSPRegClass);
5875 X86FI->setGlobalBaseReg(GlobalBaseReg);
5876 return GlobalBaseReg;
5879 // These are the replaceable SSE instructions. Some of these have Int variants
5880 // that we don't include here. We don't want to replace instructions selected
5882 static const uint16_t ReplaceableInstrs[][3] = {
5883 //PackedSingle PackedDouble PackedInt
5884 { X86::MOVAPSmr, X86::MOVAPDmr, X86::MOVDQAmr },
5885 { X86::MOVAPSrm, X86::MOVAPDrm, X86::MOVDQArm },
5886 { X86::MOVAPSrr, X86::MOVAPDrr, X86::MOVDQArr },
5887 { X86::MOVUPSmr, X86::MOVUPDmr, X86::MOVDQUmr },
5888 { X86::MOVUPSrm, X86::MOVUPDrm, X86::MOVDQUrm },
5889 { X86::MOVLPSmr, X86::MOVLPDmr, X86::MOVPQI2QImr },
5890 { X86::MOVSDmr, X86::MOVSDmr, X86::MOVPQI2QImr },
5891 { X86::MOVSSmr, X86::MOVSSmr, X86::MOVPDI2DImr },
5892 { X86::MOVSDrm, X86::MOVSDrm, X86::MOVQI2PQIrm },
5893 { X86::MOVSSrm, X86::MOVSSrm, X86::MOVDI2PDIrm },
5894 { X86::MOVNTPSmr, X86::MOVNTPDmr, X86::MOVNTDQmr },
5895 { X86::ANDNPSrm, X86::ANDNPDrm, X86::PANDNrm },
5896 { X86::ANDNPSrr, X86::ANDNPDrr, X86::PANDNrr },
5897 { X86::ANDPSrm, X86::ANDPDrm, X86::PANDrm },
5898 { X86::ANDPSrr, X86::ANDPDrr, X86::PANDrr },
5899 { X86::ORPSrm, X86::ORPDrm, X86::PORrm },
5900 { X86::ORPSrr, X86::ORPDrr, X86::PORrr },
5901 { X86::XORPSrm, X86::XORPDrm, X86::PXORrm },
5902 { X86::XORPSrr, X86::XORPDrr, X86::PXORrr },
5903 { X86::UNPCKLPDrm, X86::UNPCKLPDrm, X86::PUNPCKLQDQrm },
5904 { X86::MOVLHPSrr, X86::UNPCKLPDrr, X86::PUNPCKLQDQrr },
5905 { X86::UNPCKHPDrm, X86::UNPCKHPDrm, X86::PUNPCKHQDQrm },
5906 { X86::UNPCKHPDrr, X86::UNPCKHPDrr, X86::PUNPCKHQDQrr },
5907 { X86::UNPCKLPSrm, X86::UNPCKLPSrm, X86::PUNPCKLDQrm },
5908 { X86::UNPCKLPSrr, X86::UNPCKLPSrr, X86::PUNPCKLDQrr },
5909 { X86::UNPCKHPSrm, X86::UNPCKHPSrm, X86::PUNPCKHDQrm },
5910 { X86::UNPCKHPSrr, X86::UNPCKHPSrr, X86::PUNPCKHDQrr },
5911 { X86::EXTRACTPSmr, X86::EXTRACTPSmr, X86::PEXTRDmr },
5912 { X86::EXTRACTPSrr, X86::EXTRACTPSrr, X86::PEXTRDrr },
5913 // AVX 128-bit support
5914 { X86::VMOVAPSmr, X86::VMOVAPDmr, X86::VMOVDQAmr },
5915 { X86::VMOVAPSrm, X86::VMOVAPDrm, X86::VMOVDQArm },
5916 { X86::VMOVAPSrr, X86::VMOVAPDrr, X86::VMOVDQArr },
5917 { X86::VMOVUPSmr, X86::VMOVUPDmr, X86::VMOVDQUmr },
5918 { X86::VMOVUPSrm, X86::VMOVUPDrm, X86::VMOVDQUrm },
5919 { X86::VMOVLPSmr, X86::VMOVLPDmr, X86::VMOVPQI2QImr },
5920 { X86::VMOVSDmr, X86::VMOVSDmr, X86::VMOVPQI2QImr },
5921 { X86::VMOVSSmr, X86::VMOVSSmr, X86::VMOVPDI2DImr },
5922 { X86::VMOVSDrm, X86::VMOVSDrm, X86::VMOVQI2PQIrm },
5923 { X86::VMOVSSrm, X86::VMOVSSrm, X86::VMOVDI2PDIrm },
5924 { X86::VMOVNTPSmr, X86::VMOVNTPDmr, X86::VMOVNTDQmr },
5925 { X86::VANDNPSrm, X86::VANDNPDrm, X86::VPANDNrm },
5926 { X86::VANDNPSrr, X86::VANDNPDrr, X86::VPANDNrr },
5927 { X86::VANDPSrm, X86::VANDPDrm, X86::VPANDrm },
5928 { X86::VANDPSrr, X86::VANDPDrr, X86::VPANDrr },
5929 { X86::VORPSrm, X86::VORPDrm, X86::VPORrm },
5930 { X86::VORPSrr, X86::VORPDrr, X86::VPORrr },
5931 { X86::VXORPSrm, X86::VXORPDrm, X86::VPXORrm },
5932 { X86::VXORPSrr, X86::VXORPDrr, X86::VPXORrr },
5933 { X86::VUNPCKLPDrm, X86::VUNPCKLPDrm, X86::VPUNPCKLQDQrm },
5934 { X86::VMOVLHPSrr, X86::VUNPCKLPDrr, X86::VPUNPCKLQDQrr },
5935 { X86::VUNPCKHPDrm, X86::VUNPCKHPDrm, X86::VPUNPCKHQDQrm },
5936 { X86::VUNPCKHPDrr, X86::VUNPCKHPDrr, X86::VPUNPCKHQDQrr },
5937 { X86::VUNPCKLPSrm, X86::VUNPCKLPSrm, X86::VPUNPCKLDQrm },
5938 { X86::VUNPCKLPSrr, X86::VUNPCKLPSrr, X86::VPUNPCKLDQrr },
5939 { X86::VUNPCKHPSrm, X86::VUNPCKHPSrm, X86::VPUNPCKHDQrm },
5940 { X86::VUNPCKHPSrr, X86::VUNPCKHPSrr, X86::VPUNPCKHDQrr },
5941 { X86::VEXTRACTPSmr, X86::VEXTRACTPSmr, X86::VPEXTRDmr },
5942 { X86::VEXTRACTPSrr, X86::VEXTRACTPSrr, X86::VPEXTRDrr },
5943 // AVX 256-bit support
5944 { X86::VMOVAPSYmr, X86::VMOVAPDYmr, X86::VMOVDQAYmr },
5945 { X86::VMOVAPSYrm, X86::VMOVAPDYrm, X86::VMOVDQAYrm },
5946 { X86::VMOVAPSYrr, X86::VMOVAPDYrr, X86::VMOVDQAYrr },
5947 { X86::VMOVUPSYmr, X86::VMOVUPDYmr, X86::VMOVDQUYmr },
5948 { X86::VMOVUPSYrm, X86::VMOVUPDYrm, X86::VMOVDQUYrm },
5949 { X86::VMOVNTPSYmr, X86::VMOVNTPDYmr, X86::VMOVNTDQYmr },
5950 { X86::VPERMPSYrm, X86::VPERMPSYrm, X86::VPERMDYrm },
5951 { X86::VPERMPSYrr, X86::VPERMPSYrr, X86::VPERMDYrr },
5952 { X86::VPERMPDYmi, X86::VPERMPDYmi, X86::VPERMQYmi },
5953 { X86::VPERMPDYri, X86::VPERMPDYri, X86::VPERMQYri },
5955 { X86::VMOVLPSZ128mr, X86::VMOVLPDZ128mr, X86::VMOVPQI2QIZmr },
5956 { X86::VMOVNTPSZ128mr, X86::VMOVNTPDZ128mr, X86::VMOVNTDQZ128mr },
5957 { X86::VMOVNTPSZ256mr, X86::VMOVNTPDZ256mr, X86::VMOVNTDQZ256mr },
5958 { X86::VMOVNTPSZmr, X86::VMOVNTPDZmr, X86::VMOVNTDQZmr },
5959 { X86::VMOVSDZmr, X86::VMOVSDZmr, X86::VMOVPQI2QIZmr },
5960 { X86::VMOVSSZmr, X86::VMOVSSZmr, X86::VMOVPDI2DIZmr },
5961 { X86::VMOVSDZrm, X86::VMOVSDZrm, X86::VMOVQI2PQIZrm },
5962 { X86::VMOVSSZrm, X86::VMOVSSZrm, X86::VMOVDI2PDIZrm },
5963 { X86::VBROADCASTSSZ128r, X86::VBROADCASTSSZ128r, X86::VPBROADCASTDZ128r },
5964 { X86::VBROADCASTSSZ128m, X86::VBROADCASTSSZ128m, X86::VPBROADCASTDZ128m },
5965 { X86::VBROADCASTSSZ256r, X86::VBROADCASTSSZ256r, X86::VPBROADCASTDZ256r },
5966 { X86::VBROADCASTSSZ256m, X86::VBROADCASTSSZ256m, X86::VPBROADCASTDZ256m },
5967 { X86::VBROADCASTSSZr, X86::VBROADCASTSSZr, X86::VPBROADCASTDZr },
5968 { X86::VBROADCASTSSZm, X86::VBROADCASTSSZm, X86::VPBROADCASTDZm },
5969 { X86::VMOVDDUPZ128rr, X86::VMOVDDUPZ128rr, X86::VPBROADCASTQZ128r },
5970 { X86::VMOVDDUPZ128rm, X86::VMOVDDUPZ128rm, X86::VPBROADCASTQZ128m },
5971 { X86::VBROADCASTSDZ256r, X86::VBROADCASTSDZ256r, X86::VPBROADCASTQZ256r },
5972 { X86::VBROADCASTSDZ256m, X86::VBROADCASTSDZ256m, X86::VPBROADCASTQZ256m },
5973 { X86::VBROADCASTSDZr, X86::VBROADCASTSDZr, X86::VPBROADCASTQZr },
5974 { X86::VBROADCASTSDZm, X86::VBROADCASTSDZm, X86::VPBROADCASTQZm },
5975 { X86::VINSERTF32x4Zrr, X86::VINSERTF32x4Zrr, X86::VINSERTI32x4Zrr },
5976 { X86::VINSERTF32x4Zrm, X86::VINSERTF32x4Zrm, X86::VINSERTI32x4Zrm },
5977 { X86::VINSERTF32x8Zrr, X86::VINSERTF32x8Zrr, X86::VINSERTI32x8Zrr },
5978 { X86::VINSERTF32x8Zrm, X86::VINSERTF32x8Zrm, X86::VINSERTI32x8Zrm },
5979 { X86::VINSERTF64x2Zrr, X86::VINSERTF64x2Zrr, X86::VINSERTI64x2Zrr },
5980 { X86::VINSERTF64x2Zrm, X86::VINSERTF64x2Zrm, X86::VINSERTI64x2Zrm },
5981 { X86::VINSERTF64x4Zrr, X86::VINSERTF64x4Zrr, X86::VINSERTI64x4Zrr },
5982 { X86::VINSERTF64x4Zrm, X86::VINSERTF64x4Zrm, X86::VINSERTI64x4Zrm },
5983 { X86::VINSERTF32x4Z256rr,X86::VINSERTF32x4Z256rr,X86::VINSERTI32x4Z256rr },
5984 { X86::VINSERTF32x4Z256rm,X86::VINSERTF32x4Z256rm,X86::VINSERTI32x4Z256rm },
5985 { X86::VINSERTF64x2Z256rr,X86::VINSERTF64x2Z256rr,X86::VINSERTI64x2Z256rr },
5986 { X86::VINSERTF64x2Z256rm,X86::VINSERTF64x2Z256rm,X86::VINSERTI64x2Z256rm },
5987 { X86::VEXTRACTF32x4Zrr, X86::VEXTRACTF32x4Zrr, X86::VEXTRACTI32x4Zrr },
5988 { X86::VEXTRACTF32x4Zmr, X86::VEXTRACTF32x4Zmr, X86::VEXTRACTI32x4Zmr },
5989 { X86::VEXTRACTF32x8Zrr, X86::VEXTRACTF32x8Zrr, X86::VEXTRACTI32x8Zrr },
5990 { X86::VEXTRACTF32x8Zmr, X86::VEXTRACTF32x8Zmr, X86::VEXTRACTI32x8Zmr },
5991 { X86::VEXTRACTF64x2Zrr, X86::VEXTRACTF64x2Zrr, X86::VEXTRACTI64x2Zrr },
5992 { X86::VEXTRACTF64x2Zmr, X86::VEXTRACTF64x2Zmr, X86::VEXTRACTI64x2Zmr },
5993 { X86::VEXTRACTF64x4Zrr, X86::VEXTRACTF64x4Zrr, X86::VEXTRACTI64x4Zrr },
5994 { X86::VEXTRACTF64x4Zmr, X86::VEXTRACTF64x4Zmr, X86::VEXTRACTI64x4Zmr },
5995 { X86::VEXTRACTF32x4Z256rr,X86::VEXTRACTF32x4Z256rr,X86::VEXTRACTI32x4Z256rr },
5996 { X86::VEXTRACTF32x4Z256mr,X86::VEXTRACTF32x4Z256mr,X86::VEXTRACTI32x4Z256mr },
5997 { X86::VEXTRACTF64x2Z256rr,X86::VEXTRACTF64x2Z256rr,X86::VEXTRACTI64x2Z256rr },
5998 { X86::VEXTRACTF64x2Z256mr,X86::VEXTRACTF64x2Z256mr,X86::VEXTRACTI64x2Z256mr },
5999 { X86::VPERMILPSmi, X86::VPERMILPSmi, X86::VPSHUFDmi },
6000 { X86::VPERMILPSri, X86::VPERMILPSri, X86::VPSHUFDri },
6001 { X86::VPERMILPSZ128mi, X86::VPERMILPSZ128mi, X86::VPSHUFDZ128mi },
6002 { X86::VPERMILPSZ128ri, X86::VPERMILPSZ128ri, X86::VPSHUFDZ128ri },
6003 { X86::VPERMILPSZ256mi, X86::VPERMILPSZ256mi, X86::VPSHUFDZ256mi },
6004 { X86::VPERMILPSZ256ri, X86::VPERMILPSZ256ri, X86::VPSHUFDZ256ri },
6005 { X86::VPERMILPSZmi, X86::VPERMILPSZmi, X86::VPSHUFDZmi },
6006 { X86::VPERMILPSZri, X86::VPERMILPSZri, X86::VPSHUFDZri },
6007 { X86::VPERMPSZ256rm, X86::VPERMPSZ256rm, X86::VPERMDZ256rm },
6008 { X86::VPERMPSZ256rr, X86::VPERMPSZ256rr, X86::VPERMDZ256rr },
6009 { X86::VPERMPDZ256mi, X86::VPERMPDZ256mi, X86::VPERMQZ256mi },
6010 { X86::VPERMPDZ256ri, X86::VPERMPDZ256ri, X86::VPERMQZ256ri },
6011 { X86::VPERMPDZ256rm, X86::VPERMPDZ256rm, X86::VPERMQZ256rm },
6012 { X86::VPERMPDZ256rr, X86::VPERMPDZ256rr, X86::VPERMQZ256rr },
6013 { X86::VPERMPSZrm, X86::VPERMPSZrm, X86::VPERMDZrm },
6014 { X86::VPERMPSZrr, X86::VPERMPSZrr, X86::VPERMDZrr },
6015 { X86::VPERMPDZmi, X86::VPERMPDZmi, X86::VPERMQZmi },
6016 { X86::VPERMPDZri, X86::VPERMPDZri, X86::VPERMQZri },
6017 { X86::VPERMPDZrm, X86::VPERMPDZrm, X86::VPERMQZrm },
6018 { X86::VPERMPDZrr, X86::VPERMPDZrr, X86::VPERMQZrr },
6019 { X86::VUNPCKLPDZ256rm, X86::VUNPCKLPDZ256rm, X86::VPUNPCKLQDQZ256rm },
6020 { X86::VUNPCKLPDZ256rr, X86::VUNPCKLPDZ256rr, X86::VPUNPCKLQDQZ256rr },
6021 { X86::VUNPCKHPDZ256rm, X86::VUNPCKHPDZ256rm, X86::VPUNPCKHQDQZ256rm },
6022 { X86::VUNPCKHPDZ256rr, X86::VUNPCKHPDZ256rr, X86::VPUNPCKHQDQZ256rr },
6023 { X86::VUNPCKLPSZ256rm, X86::VUNPCKLPSZ256rm, X86::VPUNPCKLDQZ256rm },
6024 { X86::VUNPCKLPSZ256rr, X86::VUNPCKLPSZ256rr, X86::VPUNPCKLDQZ256rr },
6025 { X86::VUNPCKHPSZ256rm, X86::VUNPCKHPSZ256rm, X86::VPUNPCKHDQZ256rm },
6026 { X86::VUNPCKHPSZ256rr, X86::VUNPCKHPSZ256rr, X86::VPUNPCKHDQZ256rr },
6027 { X86::VUNPCKLPDZ128rm, X86::VUNPCKLPDZ128rm, X86::VPUNPCKLQDQZ128rm },
6028 { X86::VMOVLHPSZrr, X86::VUNPCKLPDZ128rr, X86::VPUNPCKLQDQZ128rr },
6029 { X86::VUNPCKHPDZ128rm, X86::VUNPCKHPDZ128rm, X86::VPUNPCKHQDQZ128rm },
6030 { X86::VUNPCKHPDZ128rr, X86::VUNPCKHPDZ128rr, X86::VPUNPCKHQDQZ128rr },
6031 { X86::VUNPCKLPSZ128rm, X86::VUNPCKLPSZ128rm, X86::VPUNPCKLDQZ128rm },
6032 { X86::VUNPCKLPSZ128rr, X86::VUNPCKLPSZ128rr, X86::VPUNPCKLDQZ128rr },
6033 { X86::VUNPCKHPSZ128rm, X86::VUNPCKHPSZ128rm, X86::VPUNPCKHDQZ128rm },
6034 { X86::VUNPCKHPSZ128rr, X86::VUNPCKHPSZ128rr, X86::VPUNPCKHDQZ128rr },
6035 { X86::VUNPCKLPDZrm, X86::VUNPCKLPDZrm, X86::VPUNPCKLQDQZrm },
6036 { X86::VUNPCKLPDZrr, X86::VUNPCKLPDZrr, X86::VPUNPCKLQDQZrr },
6037 { X86::VUNPCKHPDZrm, X86::VUNPCKHPDZrm, X86::VPUNPCKHQDQZrm },
6038 { X86::VUNPCKHPDZrr, X86::VUNPCKHPDZrr, X86::VPUNPCKHQDQZrr },
6039 { X86::VUNPCKLPSZrm, X86::VUNPCKLPSZrm, X86::VPUNPCKLDQZrm },
6040 { X86::VUNPCKLPSZrr, X86::VUNPCKLPSZrr, X86::VPUNPCKLDQZrr },
6041 { X86::VUNPCKHPSZrm, X86::VUNPCKHPSZrm, X86::VPUNPCKHDQZrm },
6042 { X86::VUNPCKHPSZrr, X86::VUNPCKHPSZrr, X86::VPUNPCKHDQZrr },
6043 { X86::VEXTRACTPSZmr, X86::VEXTRACTPSZmr, X86::VPEXTRDZmr },
6044 { X86::VEXTRACTPSZrr, X86::VEXTRACTPSZrr, X86::VPEXTRDZrr },
6047 static const uint16_t ReplaceableInstrsAVX2[][3] = {
6048 //PackedSingle PackedDouble PackedInt
6049 { X86::VANDNPSYrm, X86::VANDNPDYrm, X86::VPANDNYrm },
6050 { X86::VANDNPSYrr, X86::VANDNPDYrr, X86::VPANDNYrr },
6051 { X86::VANDPSYrm, X86::VANDPDYrm, X86::VPANDYrm },
6052 { X86::VANDPSYrr, X86::VANDPDYrr, X86::VPANDYrr },
6053 { X86::VORPSYrm, X86::VORPDYrm, X86::VPORYrm },
6054 { X86::VORPSYrr, X86::VORPDYrr, X86::VPORYrr },
6055 { X86::VXORPSYrm, X86::VXORPDYrm, X86::VPXORYrm },
6056 { X86::VXORPSYrr, X86::VXORPDYrr, X86::VPXORYrr },
6057 { X86::VPERM2F128rm, X86::VPERM2F128rm, X86::VPERM2I128rm },
6058 { X86::VPERM2F128rr, X86::VPERM2F128rr, X86::VPERM2I128rr },
6059 { X86::VBROADCASTSSrm, X86::VBROADCASTSSrm, X86::VPBROADCASTDrm},
6060 { X86::VBROADCASTSSrr, X86::VBROADCASTSSrr, X86::VPBROADCASTDrr},
6061 { X86::VMOVDDUPrm, X86::VMOVDDUPrm, X86::VPBROADCASTQrm},
6062 { X86::VMOVDDUPrr, X86::VMOVDDUPrr, X86::VPBROADCASTQrr},
6063 { X86::VBROADCASTSSYrr, X86::VBROADCASTSSYrr, X86::VPBROADCASTDYrr},
6064 { X86::VBROADCASTSSYrm, X86::VBROADCASTSSYrm, X86::VPBROADCASTDYrm},
6065 { X86::VBROADCASTSDYrr, X86::VBROADCASTSDYrr, X86::VPBROADCASTQYrr},
6066 { X86::VBROADCASTSDYrm, X86::VBROADCASTSDYrm, X86::VPBROADCASTQYrm},
6067 { X86::VBROADCASTF128, X86::VBROADCASTF128, X86::VBROADCASTI128 },
6068 { X86::VBLENDPSYrri, X86::VBLENDPSYrri, X86::VPBLENDDYrri },
6069 { X86::VBLENDPSYrmi, X86::VBLENDPSYrmi, X86::VPBLENDDYrmi },
6070 { X86::VPERMILPSYmi, X86::VPERMILPSYmi, X86::VPSHUFDYmi },
6071 { X86::VPERMILPSYri, X86::VPERMILPSYri, X86::VPSHUFDYri },
6072 { X86::VUNPCKLPDYrm, X86::VUNPCKLPDYrm, X86::VPUNPCKLQDQYrm },
6073 { X86::VUNPCKLPDYrr, X86::VUNPCKLPDYrr, X86::VPUNPCKLQDQYrr },
6074 { X86::VUNPCKHPDYrm, X86::VUNPCKHPDYrm, X86::VPUNPCKHQDQYrm },
6075 { X86::VUNPCKHPDYrr, X86::VUNPCKHPDYrr, X86::VPUNPCKHQDQYrr },
6076 { X86::VUNPCKLPSYrm, X86::VUNPCKLPSYrm, X86::VPUNPCKLDQYrm },
6077 { X86::VUNPCKLPSYrr, X86::VUNPCKLPSYrr, X86::VPUNPCKLDQYrr },
6078 { X86::VUNPCKHPSYrm, X86::VUNPCKHPSYrm, X86::VPUNPCKHDQYrm },
6079 { X86::VUNPCKHPSYrr, X86::VUNPCKHPSYrr, X86::VPUNPCKHDQYrr },
6082 static const uint16_t ReplaceableInstrsAVX2InsertExtract[][3] = {
6083 //PackedSingle PackedDouble PackedInt
6084 { X86::VEXTRACTF128mr, X86::VEXTRACTF128mr, X86::VEXTRACTI128mr },
6085 { X86::VEXTRACTF128rr, X86::VEXTRACTF128rr, X86::VEXTRACTI128rr },
6086 { X86::VINSERTF128rm, X86::VINSERTF128rm, X86::VINSERTI128rm },
6087 { X86::VINSERTF128rr, X86::VINSERTF128rr, X86::VINSERTI128rr },
6090 static const uint16_t ReplaceableInstrsAVX512[][4] = {
6091 // Two integer columns for 64-bit and 32-bit elements.
6092 //PackedSingle PackedDouble PackedInt PackedInt
6093 { X86::VMOVAPSZ128mr, X86::VMOVAPDZ128mr, X86::VMOVDQA64Z128mr, X86::VMOVDQA32Z128mr },
6094 { X86::VMOVAPSZ128rm, X86::VMOVAPDZ128rm, X86::VMOVDQA64Z128rm, X86::VMOVDQA32Z128rm },
6095 { X86::VMOVAPSZ128rr, X86::VMOVAPDZ128rr, X86::VMOVDQA64Z128rr, X86::VMOVDQA32Z128rr },
6096 { X86::VMOVUPSZ128mr, X86::VMOVUPDZ128mr, X86::VMOVDQU64Z128mr, X86::VMOVDQU32Z128mr },
6097 { X86::VMOVUPSZ128rm, X86::VMOVUPDZ128rm, X86::VMOVDQU64Z128rm, X86::VMOVDQU32Z128rm },
6098 { X86::VMOVAPSZ256mr, X86::VMOVAPDZ256mr, X86::VMOVDQA64Z256mr, X86::VMOVDQA32Z256mr },
6099 { X86::VMOVAPSZ256rm, X86::VMOVAPDZ256rm, X86::VMOVDQA64Z256rm, X86::VMOVDQA32Z256rm },
6100 { X86::VMOVAPSZ256rr, X86::VMOVAPDZ256rr, X86::VMOVDQA64Z256rr, X86::VMOVDQA32Z256rr },
6101 { X86::VMOVUPSZ256mr, X86::VMOVUPDZ256mr, X86::VMOVDQU64Z256mr, X86::VMOVDQU32Z256mr },
6102 { X86::VMOVUPSZ256rm, X86::VMOVUPDZ256rm, X86::VMOVDQU64Z256rm, X86::VMOVDQU32Z256rm },
6103 { X86::VMOVAPSZmr, X86::VMOVAPDZmr, X86::VMOVDQA64Zmr, X86::VMOVDQA32Zmr },
6104 { X86::VMOVAPSZrm, X86::VMOVAPDZrm, X86::VMOVDQA64Zrm, X86::VMOVDQA32Zrm },
6105 { X86::VMOVAPSZrr, X86::VMOVAPDZrr, X86::VMOVDQA64Zrr, X86::VMOVDQA32Zrr },
6106 { X86::VMOVUPSZmr, X86::VMOVUPDZmr, X86::VMOVDQU64Zmr, X86::VMOVDQU32Zmr },
6107 { X86::VMOVUPSZrm, X86::VMOVUPDZrm, X86::VMOVDQU64Zrm, X86::VMOVDQU32Zrm },
6110 static const uint16_t ReplaceableInstrsAVX512DQ[][4] = {
6111 // Two integer columns for 64-bit and 32-bit elements.
6112 //PackedSingle PackedDouble PackedInt PackedInt
6113 { X86::VANDNPSZ128rm, X86::VANDNPDZ128rm, X86::VPANDNQZ128rm, X86::VPANDNDZ128rm },
6114 { X86::VANDNPSZ128rr, X86::VANDNPDZ128rr, X86::VPANDNQZ128rr, X86::VPANDNDZ128rr },
6115 { X86::VANDPSZ128rm, X86::VANDPDZ128rm, X86::VPANDQZ128rm, X86::VPANDDZ128rm },
6116 { X86::VANDPSZ128rr, X86::VANDPDZ128rr, X86::VPANDQZ128rr, X86::VPANDDZ128rr },
6117 { X86::VORPSZ128rm, X86::VORPDZ128rm, X86::VPORQZ128rm, X86::VPORDZ128rm },
6118 { X86::VORPSZ128rr, X86::VORPDZ128rr, X86::VPORQZ128rr, X86::VPORDZ128rr },
6119 { X86::VXORPSZ128rm, X86::VXORPDZ128rm, X86::VPXORQZ128rm, X86::VPXORDZ128rm },
6120 { X86::VXORPSZ128rr, X86::VXORPDZ128rr, X86::VPXORQZ128rr, X86::VPXORDZ128rr },
6121 { X86::VANDNPSZ256rm, X86::VANDNPDZ256rm, X86::VPANDNQZ256rm, X86::VPANDNDZ256rm },
6122 { X86::VANDNPSZ256rr, X86::VANDNPDZ256rr, X86::VPANDNQZ256rr, X86::VPANDNDZ256rr },
6123 { X86::VANDPSZ256rm, X86::VANDPDZ256rm, X86::VPANDQZ256rm, X86::VPANDDZ256rm },
6124 { X86::VANDPSZ256rr, X86::VANDPDZ256rr, X86::VPANDQZ256rr, X86::VPANDDZ256rr },
6125 { X86::VORPSZ256rm, X86::VORPDZ256rm, X86::VPORQZ256rm, X86::VPORDZ256rm },
6126 { X86::VORPSZ256rr, X86::VORPDZ256rr, X86::VPORQZ256rr, X86::VPORDZ256rr },
6127 { X86::VXORPSZ256rm, X86::VXORPDZ256rm, X86::VPXORQZ256rm, X86::VPXORDZ256rm },
6128 { X86::VXORPSZ256rr, X86::VXORPDZ256rr, X86::VPXORQZ256rr, X86::VPXORDZ256rr },
6129 { X86::VANDNPSZrm, X86::VANDNPDZrm, X86::VPANDNQZrm, X86::VPANDNDZrm },
6130 { X86::VANDNPSZrr, X86::VANDNPDZrr, X86::VPANDNQZrr, X86::VPANDNDZrr },
6131 { X86::VANDPSZrm, X86::VANDPDZrm, X86::VPANDQZrm, X86::VPANDDZrm },
6132 { X86::VANDPSZrr, X86::VANDPDZrr, X86::VPANDQZrr, X86::VPANDDZrr },
6133 { X86::VORPSZrm, X86::VORPDZrm, X86::VPORQZrm, X86::VPORDZrm },
6134 { X86::VORPSZrr, X86::VORPDZrr, X86::VPORQZrr, X86::VPORDZrr },
6135 { X86::VXORPSZrm, X86::VXORPDZrm, X86::VPXORQZrm, X86::VPXORDZrm },
6136 { X86::VXORPSZrr, X86::VXORPDZrr, X86::VPXORQZrr, X86::VPXORDZrr },
6139 static const uint16_t ReplaceableInstrsAVX512DQMasked[][4] = {
6140 // Two integer columns for 64-bit and 32-bit elements.
6141 //PackedSingle PackedDouble
6142 //PackedInt PackedInt
6143 { X86::VANDNPSZ128rmk, X86::VANDNPDZ128rmk,
6144 X86::VPANDNQZ128rmk, X86::VPANDNDZ128rmk },
6145 { X86::VANDNPSZ128rmkz, X86::VANDNPDZ128rmkz,
6146 X86::VPANDNQZ128rmkz, X86::VPANDNDZ128rmkz },
6147 { X86::VANDNPSZ128rrk, X86::VANDNPDZ128rrk,
6148 X86::VPANDNQZ128rrk, X86::VPANDNDZ128rrk },
6149 { X86::VANDNPSZ128rrkz, X86::VANDNPDZ128rrkz,
6150 X86::VPANDNQZ128rrkz, X86::VPANDNDZ128rrkz },
6151 { X86::VANDPSZ128rmk, X86::VANDPDZ128rmk,
6152 X86::VPANDQZ128rmk, X86::VPANDDZ128rmk },
6153 { X86::VANDPSZ128rmkz, X86::VANDPDZ128rmkz,
6154 X86::VPANDQZ128rmkz, X86::VPANDDZ128rmkz },
6155 { X86::VANDPSZ128rrk, X86::VANDPDZ128rrk,
6156 X86::VPANDQZ128rrk, X86::VPANDDZ128rrk },
6157 { X86::VANDPSZ128rrkz, X86::VANDPDZ128rrkz,
6158 X86::VPANDQZ128rrkz, X86::VPANDDZ128rrkz },
6159 { X86::VORPSZ128rmk, X86::VORPDZ128rmk,
6160 X86::VPORQZ128rmk, X86::VPORDZ128rmk },
6161 { X86::VORPSZ128rmkz, X86::VORPDZ128rmkz,
6162 X86::VPORQZ128rmkz, X86::VPORDZ128rmkz },
6163 { X86::VORPSZ128rrk, X86::VORPDZ128rrk,
6164 X86::VPORQZ128rrk, X86::VPORDZ128rrk },
6165 { X86::VORPSZ128rrkz, X86::VORPDZ128rrkz,
6166 X86::VPORQZ128rrkz, X86::VPORDZ128rrkz },
6167 { X86::VXORPSZ128rmk, X86::VXORPDZ128rmk,
6168 X86::VPXORQZ128rmk, X86::VPXORDZ128rmk },
6169 { X86::VXORPSZ128rmkz, X86::VXORPDZ128rmkz,
6170 X86::VPXORQZ128rmkz, X86::VPXORDZ128rmkz },
6171 { X86::VXORPSZ128rrk, X86::VXORPDZ128rrk,
6172 X86::VPXORQZ128rrk, X86::VPXORDZ128rrk },
6173 { X86::VXORPSZ128rrkz, X86::VXORPDZ128rrkz,
6174 X86::VPXORQZ128rrkz, X86::VPXORDZ128rrkz },
6175 { X86::VANDNPSZ256rmk, X86::VANDNPDZ256rmk,
6176 X86::VPANDNQZ256rmk, X86::VPANDNDZ256rmk },
6177 { X86::VANDNPSZ256rmkz, X86::VANDNPDZ256rmkz,
6178 X86::VPANDNQZ256rmkz, X86::VPANDNDZ256rmkz },
6179 { X86::VANDNPSZ256rrk, X86::VANDNPDZ256rrk,
6180 X86::VPANDNQZ256rrk, X86::VPANDNDZ256rrk },
6181 { X86::VANDNPSZ256rrkz, X86::VANDNPDZ256rrkz,
6182 X86::VPANDNQZ256rrkz, X86::VPANDNDZ256rrkz },
6183 { X86::VANDPSZ256rmk, X86::VANDPDZ256rmk,
6184 X86::VPANDQZ256rmk, X86::VPANDDZ256rmk },
6185 { X86::VANDPSZ256rmkz, X86::VANDPDZ256rmkz,
6186 X86::VPANDQZ256rmkz, X86::VPANDDZ256rmkz },
6187 { X86::VANDPSZ256rrk, X86::VANDPDZ256rrk,
6188 X86::VPANDQZ256rrk, X86::VPANDDZ256rrk },
6189 { X86::VANDPSZ256rrkz, X86::VANDPDZ256rrkz,
6190 X86::VPANDQZ256rrkz, X86::VPANDDZ256rrkz },
6191 { X86::VORPSZ256rmk, X86::VORPDZ256rmk,
6192 X86::VPORQZ256rmk, X86::VPORDZ256rmk },
6193 { X86::VORPSZ256rmkz, X86::VORPDZ256rmkz,
6194 X86::VPORQZ256rmkz, X86::VPORDZ256rmkz },
6195 { X86::VORPSZ256rrk, X86::VORPDZ256rrk,
6196 X86::VPORQZ256rrk, X86::VPORDZ256rrk },
6197 { X86::VORPSZ256rrkz, X86::VORPDZ256rrkz,
6198 X86::VPORQZ256rrkz, X86::VPORDZ256rrkz },
6199 { X86::VXORPSZ256rmk, X86::VXORPDZ256rmk,
6200 X86::VPXORQZ256rmk, X86::VPXORDZ256rmk },
6201 { X86::VXORPSZ256rmkz, X86::VXORPDZ256rmkz,
6202 X86::VPXORQZ256rmkz, X86::VPXORDZ256rmkz },
6203 { X86::VXORPSZ256rrk, X86::VXORPDZ256rrk,
6204 X86::VPXORQZ256rrk, X86::VPXORDZ256rrk },
6205 { X86::VXORPSZ256rrkz, X86::VXORPDZ256rrkz,
6206 X86::VPXORQZ256rrkz, X86::VPXORDZ256rrkz },
6207 { X86::VANDNPSZrmk, X86::VANDNPDZrmk,
6208 X86::VPANDNQZrmk, X86::VPANDNDZrmk },
6209 { X86::VANDNPSZrmkz, X86::VANDNPDZrmkz,
6210 X86::VPANDNQZrmkz, X86::VPANDNDZrmkz },
6211 { X86::VANDNPSZrrk, X86::VANDNPDZrrk,
6212 X86::VPANDNQZrrk, X86::VPANDNDZrrk },
6213 { X86::VANDNPSZrrkz, X86::VANDNPDZrrkz,
6214 X86::VPANDNQZrrkz, X86::VPANDNDZrrkz },
6215 { X86::VANDPSZrmk, X86::VANDPDZrmk,
6216 X86::VPANDQZrmk, X86::VPANDDZrmk },
6217 { X86::VANDPSZrmkz, X86::VANDPDZrmkz,
6218 X86::VPANDQZrmkz, X86::VPANDDZrmkz },
6219 { X86::VANDPSZrrk, X86::VANDPDZrrk,
6220 X86::VPANDQZrrk, X86::VPANDDZrrk },
6221 { X86::VANDPSZrrkz, X86::VANDPDZrrkz,
6222 X86::VPANDQZrrkz, X86::VPANDDZrrkz },
6223 { X86::VORPSZrmk, X86::VORPDZrmk,
6224 X86::VPORQZrmk, X86::VPORDZrmk },
6225 { X86::VORPSZrmkz, X86::VORPDZrmkz,
6226 X86::VPORQZrmkz, X86::VPORDZrmkz },
6227 { X86::VORPSZrrk, X86::VORPDZrrk,
6228 X86::VPORQZrrk, X86::VPORDZrrk },
6229 { X86::VORPSZrrkz, X86::VORPDZrrkz,
6230 X86::VPORQZrrkz, X86::VPORDZrrkz },
6231 { X86::VXORPSZrmk, X86::VXORPDZrmk,
6232 X86::VPXORQZrmk, X86::VPXORDZrmk },
6233 { X86::VXORPSZrmkz, X86::VXORPDZrmkz,
6234 X86::VPXORQZrmkz, X86::VPXORDZrmkz },
6235 { X86::VXORPSZrrk, X86::VXORPDZrrk,
6236 X86::VPXORQZrrk, X86::VPXORDZrrk },
6237 { X86::VXORPSZrrkz, X86::VXORPDZrrkz,
6238 X86::VPXORQZrrkz, X86::VPXORDZrrkz },
6239 // Broadcast loads can be handled the same as masked operations to avoid
6240 // changing element size.
6241 { X86::VANDNPSZ128rmb, X86::VANDNPDZ128rmb,
6242 X86::VPANDNQZ128rmb, X86::VPANDNDZ128rmb },
6243 { X86::VANDPSZ128rmb, X86::VANDPDZ128rmb,
6244 X86::VPANDQZ128rmb, X86::VPANDDZ128rmb },
6245 { X86::VORPSZ128rmb, X86::VORPDZ128rmb,
6246 X86::VPORQZ128rmb, X86::VPORDZ128rmb },
6247 { X86::VXORPSZ128rmb, X86::VXORPDZ128rmb,
6248 X86::VPXORQZ128rmb, X86::VPXORDZ128rmb },
6249 { X86::VANDNPSZ256rmb, X86::VANDNPDZ256rmb,
6250 X86::VPANDNQZ256rmb, X86::VPANDNDZ256rmb },
6251 { X86::VANDPSZ256rmb, X86::VANDPDZ256rmb,
6252 X86::VPANDQZ256rmb, X86::VPANDDZ256rmb },
6253 { X86::VORPSZ256rmb, X86::VORPDZ256rmb,
6254 X86::VPORQZ256rmb, X86::VPORDZ256rmb },
6255 { X86::VXORPSZ256rmb, X86::VXORPDZ256rmb,
6256 X86::VPXORQZ256rmb, X86::VPXORDZ256rmb },
6257 { X86::VANDNPSZrmb, X86::VANDNPDZrmb,
6258 X86::VPANDNQZrmb, X86::VPANDNDZrmb },
6259 { X86::VANDPSZrmb, X86::VANDPDZrmb,
6260 X86::VPANDQZrmb, X86::VPANDDZrmb },
6261 { X86::VANDPSZrmb, X86::VANDPDZrmb,
6262 X86::VPANDQZrmb, X86::VPANDDZrmb },
6263 { X86::VORPSZrmb, X86::VORPDZrmb,
6264 X86::VPORQZrmb, X86::VPORDZrmb },
6265 { X86::VXORPSZrmb, X86::VXORPDZrmb,
6266 X86::VPXORQZrmb, X86::VPXORDZrmb },
6267 { X86::VANDNPSZ128rmbk, X86::VANDNPDZ128rmbk,
6268 X86::VPANDNQZ128rmbk, X86::VPANDNDZ128rmbk },
6269 { X86::VANDPSZ128rmbk, X86::VANDPDZ128rmbk,
6270 X86::VPANDQZ128rmbk, X86::VPANDDZ128rmbk },
6271 { X86::VORPSZ128rmbk, X86::VORPDZ128rmbk,
6272 X86::VPORQZ128rmbk, X86::VPORDZ128rmbk },
6273 { X86::VXORPSZ128rmbk, X86::VXORPDZ128rmbk,
6274 X86::VPXORQZ128rmbk, X86::VPXORDZ128rmbk },
6275 { X86::VANDNPSZ256rmbk, X86::VANDNPDZ256rmbk,
6276 X86::VPANDNQZ256rmbk, X86::VPANDNDZ256rmbk },
6277 { X86::VANDPSZ256rmbk, X86::VANDPDZ256rmbk,
6278 X86::VPANDQZ256rmbk, X86::VPANDDZ256rmbk },
6279 { X86::VORPSZ256rmbk, X86::VORPDZ256rmbk,
6280 X86::VPORQZ256rmbk, X86::VPORDZ256rmbk },
6281 { X86::VXORPSZ256rmbk, X86::VXORPDZ256rmbk,
6282 X86::VPXORQZ256rmbk, X86::VPXORDZ256rmbk },
6283 { X86::VANDNPSZrmbk, X86::VANDNPDZrmbk,
6284 X86::VPANDNQZrmbk, X86::VPANDNDZrmbk },
6285 { X86::VANDPSZrmbk, X86::VANDPDZrmbk,
6286 X86::VPANDQZrmbk, X86::VPANDDZrmbk },
6287 { X86::VANDPSZrmbk, X86::VANDPDZrmbk,
6288 X86::VPANDQZrmbk, X86::VPANDDZrmbk },
6289 { X86::VORPSZrmbk, X86::VORPDZrmbk,
6290 X86::VPORQZrmbk, X86::VPORDZrmbk },
6291 { X86::VXORPSZrmbk, X86::VXORPDZrmbk,
6292 X86::VPXORQZrmbk, X86::VPXORDZrmbk },
6293 { X86::VANDNPSZ128rmbkz,X86::VANDNPDZ128rmbkz,
6294 X86::VPANDNQZ128rmbkz,X86::VPANDNDZ128rmbkz},
6295 { X86::VANDPSZ128rmbkz, X86::VANDPDZ128rmbkz,
6296 X86::VPANDQZ128rmbkz, X86::VPANDDZ128rmbkz },
6297 { X86::VORPSZ128rmbkz, X86::VORPDZ128rmbkz,
6298 X86::VPORQZ128rmbkz, X86::VPORDZ128rmbkz },
6299 { X86::VXORPSZ128rmbkz, X86::VXORPDZ128rmbkz,
6300 X86::VPXORQZ128rmbkz, X86::VPXORDZ128rmbkz },
6301 { X86::VANDNPSZ256rmbkz,X86::VANDNPDZ256rmbkz,
6302 X86::VPANDNQZ256rmbkz,X86::VPANDNDZ256rmbkz},
6303 { X86::VANDPSZ256rmbkz, X86::VANDPDZ256rmbkz,
6304 X86::VPANDQZ256rmbkz, X86::VPANDDZ256rmbkz },
6305 { X86::VORPSZ256rmbkz, X86::VORPDZ256rmbkz,
6306 X86::VPORQZ256rmbkz, X86::VPORDZ256rmbkz },
6307 { X86::VXORPSZ256rmbkz, X86::VXORPDZ256rmbkz,
6308 X86::VPXORQZ256rmbkz, X86::VPXORDZ256rmbkz },
6309 { X86::VANDNPSZrmbkz, X86::VANDNPDZrmbkz,
6310 X86::VPANDNQZrmbkz, X86::VPANDNDZrmbkz },
6311 { X86::VANDPSZrmbkz, X86::VANDPDZrmbkz,
6312 X86::VPANDQZrmbkz, X86::VPANDDZrmbkz },
6313 { X86::VANDPSZrmbkz, X86::VANDPDZrmbkz,
6314 X86::VPANDQZrmbkz, X86::VPANDDZrmbkz },
6315 { X86::VORPSZrmbkz, X86::VORPDZrmbkz,
6316 X86::VPORQZrmbkz, X86::VPORDZrmbkz },
6317 { X86::VXORPSZrmbkz, X86::VXORPDZrmbkz,
6318 X86::VPXORQZrmbkz, X86::VPXORDZrmbkz },
6321 // NOTE: These should only be used by the custom domain methods.
6322 static const uint16_t ReplaceableCustomInstrs[][3] = {
6323 //PackedSingle PackedDouble PackedInt
6324 { X86::BLENDPSrmi, X86::BLENDPDrmi, X86::PBLENDWrmi },
6325 { X86::BLENDPSrri, X86::BLENDPDrri, X86::PBLENDWrri },
6326 { X86::VBLENDPSrmi, X86::VBLENDPDrmi, X86::VPBLENDWrmi },
6327 { X86::VBLENDPSrri, X86::VBLENDPDrri, X86::VPBLENDWrri },
6328 { X86::VBLENDPSYrmi, X86::VBLENDPDYrmi, X86::VPBLENDWYrmi },
6329 { X86::VBLENDPSYrri, X86::VBLENDPDYrri, X86::VPBLENDWYrri },
6331 static const uint16_t ReplaceableCustomAVX2Instrs[][3] = {
6332 //PackedSingle PackedDouble PackedInt
6333 { X86::VBLENDPSrmi, X86::VBLENDPDrmi, X86::VPBLENDDrmi },
6334 { X86::VBLENDPSrri, X86::VBLENDPDrri, X86::VPBLENDDrri },
6335 { X86::VBLENDPSYrmi, X86::VBLENDPDYrmi, X86::VPBLENDDYrmi },
6336 { X86::VBLENDPSYrri, X86::VBLENDPDYrri, X86::VPBLENDDYrri },
6339 // Special table for changing EVEX logic instructions to VEX.
6340 // TODO: Should we run EVEX->VEX earlier?
6341 static const uint16_t ReplaceableCustomAVX512LogicInstrs[][4] = {
6342 // Two integer columns for 64-bit and 32-bit elements.
6343 //PackedSingle PackedDouble PackedInt PackedInt
6344 { X86::VANDNPSrm, X86::VANDNPDrm, X86::VPANDNQZ128rm, X86::VPANDNDZ128rm },
6345 { X86::VANDNPSrr, X86::VANDNPDrr, X86::VPANDNQZ128rr, X86::VPANDNDZ128rr },
6346 { X86::VANDPSrm, X86::VANDPDrm, X86::VPANDQZ128rm, X86::VPANDDZ128rm },
6347 { X86::VANDPSrr, X86::VANDPDrr, X86::VPANDQZ128rr, X86::VPANDDZ128rr },
6348 { X86::VORPSrm, X86::VORPDrm, X86::VPORQZ128rm, X86::VPORDZ128rm },
6349 { X86::VORPSrr, X86::VORPDrr, X86::VPORQZ128rr, X86::VPORDZ128rr },
6350 { X86::VXORPSrm, X86::VXORPDrm, X86::VPXORQZ128rm, X86::VPXORDZ128rm },
6351 { X86::VXORPSrr, X86::VXORPDrr, X86::VPXORQZ128rr, X86::VPXORDZ128rr },
6352 { X86::VANDNPSYrm, X86::VANDNPDYrm, X86::VPANDNQZ256rm, X86::VPANDNDZ256rm },
6353 { X86::VANDNPSYrr, X86::VANDNPDYrr, X86::VPANDNQZ256rr, X86::VPANDNDZ256rr },
6354 { X86::VANDPSYrm, X86::VANDPDYrm, X86::VPANDQZ256rm, X86::VPANDDZ256rm },
6355 { X86::VANDPSYrr, X86::VANDPDYrr, X86::VPANDQZ256rr, X86::VPANDDZ256rr },
6356 { X86::VORPSYrm, X86::VORPDYrm, X86::VPORQZ256rm, X86::VPORDZ256rm },
6357 { X86::VORPSYrr, X86::VORPDYrr, X86::VPORQZ256rr, X86::VPORDZ256rr },
6358 { X86::VXORPSYrm, X86::VXORPDYrm, X86::VPXORQZ256rm, X86::VPXORDZ256rm },
6359 { X86::VXORPSYrr, X86::VXORPDYrr, X86::VPXORQZ256rr, X86::VPXORDZ256rr },
6362 // FIXME: Some shuffle and unpack instructions have equivalents in different
6363 // domains, but they require a bit more work than just switching opcodes.
6365 static const uint16_t *lookup(unsigned opcode, unsigned domain,
6366 ArrayRef<uint16_t[3]> Table) {
6367 for (const uint16_t (&Row)[3] : Table)
6368 if (Row[domain-1] == opcode)
6373 static const uint16_t *lookupAVX512(unsigned opcode, unsigned domain,
6374 ArrayRef<uint16_t[4]> Table) {
6375 // If this is the integer domain make sure to check both integer columns.
6376 for (const uint16_t (&Row)[4] : Table)
6377 if (Row[domain-1] == opcode || (domain == 3 && Row[3] == opcode))
6382 // Helper to attempt to widen/narrow blend masks.
6383 static bool AdjustBlendMask(unsigned OldMask, unsigned OldWidth,
6384 unsigned NewWidth, unsigned *pNewMask = nullptr) {
6385 assert(((OldWidth % NewWidth) == 0 || (NewWidth % OldWidth) == 0) &&
6386 "Illegal blend mask scale");
6387 unsigned NewMask = 0;
6389 if ((OldWidth % NewWidth) == 0) {
6390 unsigned Scale = OldWidth / NewWidth;
6391 unsigned SubMask = (1u << Scale) - 1;
6392 for (unsigned i = 0; i != NewWidth; ++i) {
6393 unsigned Sub = (OldMask >> (i * Scale)) & SubMask;
6395 NewMask |= (1u << i);
6396 else if (Sub != 0x0)
6400 unsigned Scale = NewWidth / OldWidth;
6401 unsigned SubMask = (1u << Scale) - 1;
6402 for (unsigned i = 0; i != OldWidth; ++i) {
6403 if (OldMask & (1 << i)) {
6404 NewMask |= (SubMask << (i * Scale));
6410 *pNewMask = NewMask;
6414 uint16_t X86InstrInfo::getExecutionDomainCustom(const MachineInstr &MI) const {
6415 unsigned Opcode = MI.getOpcode();
6416 unsigned NumOperands = MI.getDesc().getNumOperands();
6418 auto GetBlendDomains = [&](unsigned ImmWidth, bool Is256) {
6419 uint16_t validDomains = 0;
6420 if (MI.getOperand(NumOperands - 1).isImm()) {
6421 unsigned Imm = MI.getOperand(NumOperands - 1).getImm();
6422 if (AdjustBlendMask(Imm, ImmWidth, Is256 ? 8 : 4))
6423 validDomains |= 0x2; // PackedSingle
6424 if (AdjustBlendMask(Imm, ImmWidth, Is256 ? 4 : 2))
6425 validDomains |= 0x4; // PackedDouble
6426 if (!Is256 || Subtarget.hasAVX2())
6427 validDomains |= 0x8; // PackedInt
6429 return validDomains;
6433 case X86::BLENDPDrmi:
6434 case X86::BLENDPDrri:
6435 case X86::VBLENDPDrmi:
6436 case X86::VBLENDPDrri:
6437 return GetBlendDomains(2, false);
6438 case X86::VBLENDPDYrmi:
6439 case X86::VBLENDPDYrri:
6440 return GetBlendDomains(4, true);
6441 case X86::BLENDPSrmi:
6442 case X86::BLENDPSrri:
6443 case X86::VBLENDPSrmi:
6444 case X86::VBLENDPSrri:
6445 case X86::VPBLENDDrmi:
6446 case X86::VPBLENDDrri:
6447 return GetBlendDomains(4, false);
6448 case X86::VBLENDPSYrmi:
6449 case X86::VBLENDPSYrri:
6450 case X86::VPBLENDDYrmi:
6451 case X86::VPBLENDDYrri:
6452 return GetBlendDomains(8, true);
6453 case X86::PBLENDWrmi:
6454 case X86::PBLENDWrri:
6455 case X86::VPBLENDWrmi:
6456 case X86::VPBLENDWrri:
6457 // Treat VPBLENDWY as a 128-bit vector as it repeats the lo/hi masks.
6458 case X86::VPBLENDWYrmi:
6459 case X86::VPBLENDWYrri:
6460 return GetBlendDomains(8, false);
6461 case X86::VPANDDZ128rr: case X86::VPANDDZ128rm:
6462 case X86::VPANDDZ256rr: case X86::VPANDDZ256rm:
6463 case X86::VPANDQZ128rr: case X86::VPANDQZ128rm:
6464 case X86::VPANDQZ256rr: case X86::VPANDQZ256rm:
6465 case X86::VPANDNDZ128rr: case X86::VPANDNDZ128rm:
6466 case X86::VPANDNDZ256rr: case X86::VPANDNDZ256rm:
6467 case X86::VPANDNQZ128rr: case X86::VPANDNQZ128rm:
6468 case X86::VPANDNQZ256rr: case X86::VPANDNQZ256rm:
6469 case X86::VPORDZ128rr: case X86::VPORDZ128rm:
6470 case X86::VPORDZ256rr: case X86::VPORDZ256rm:
6471 case X86::VPORQZ128rr: case X86::VPORQZ128rm:
6472 case X86::VPORQZ256rr: case X86::VPORQZ256rm:
6473 case X86::VPXORDZ128rr: case X86::VPXORDZ128rm:
6474 case X86::VPXORDZ256rr: case X86::VPXORDZ256rm:
6475 case X86::VPXORQZ128rr: case X86::VPXORQZ128rm:
6476 case X86::VPXORQZ256rr: case X86::VPXORQZ256rm:
6477 // If we don't have DQI see if we can still switch from an EVEX integer
6478 // instruction to a VEX floating point instruction.
6479 if (Subtarget.hasDQI())
6482 if (RI.getEncodingValue(MI.getOperand(0).getReg()) >= 16)
6484 if (RI.getEncodingValue(MI.getOperand(1).getReg()) >= 16)
6486 // Register forms will have 3 operands. Memory form will have more.
6487 if (NumOperands == 3 &&
6488 RI.getEncodingValue(MI.getOperand(2).getReg()) >= 16)
6491 // All domains are valid.
6493 case X86::MOVHLPSrr:
6494 // We can swap domains when both inputs are the same register.
6495 // FIXME: This doesn't catch all the cases we would like. If the input
6496 // register isn't KILLed by the instruction, the two address instruction
6497 // pass puts a COPY on one input. The other input uses the original
6498 // register. This prevents the same physical register from being used by
6500 if (MI.getOperand(1).getReg() == MI.getOperand(2).getReg() &&
6501 MI.getOperand(0).getSubReg() == 0 &&
6502 MI.getOperand(1).getSubReg() == 0 &&
6503 MI.getOperand(2).getSubReg() == 0)
6510 bool X86InstrInfo::setExecutionDomainCustom(MachineInstr &MI,
6511 unsigned Domain) const {
6512 assert(Domain > 0 && Domain < 4 && "Invalid execution domain");
6513 uint16_t dom = (MI.getDesc().TSFlags >> X86II::SSEDomainShift) & 3;
6514 assert(dom && "Not an SSE instruction");
6516 unsigned Opcode = MI.getOpcode();
6517 unsigned NumOperands = MI.getDesc().getNumOperands();
6519 auto SetBlendDomain = [&](unsigned ImmWidth, bool Is256) {
6520 if (MI.getOperand(NumOperands - 1).isImm()) {
6521 unsigned Imm = MI.getOperand(NumOperands - 1).getImm() & 255;
6522 Imm = (ImmWidth == 16 ? ((Imm << 8) | Imm) : Imm);
6523 unsigned NewImm = Imm;
6525 const uint16_t *table = lookup(Opcode, dom, ReplaceableCustomInstrs);
6527 table = lookup(Opcode, dom, ReplaceableCustomAVX2Instrs);
6529 if (Domain == 1) { // PackedSingle
6530 AdjustBlendMask(Imm, ImmWidth, Is256 ? 8 : 4, &NewImm);
6531 } else if (Domain == 2) { // PackedDouble
6532 AdjustBlendMask(Imm, ImmWidth, Is256 ? 4 : 2, &NewImm);
6533 } else if (Domain == 3) { // PackedInt
6534 if (Subtarget.hasAVX2()) {
6535 // If we are already VPBLENDW use that, else use VPBLENDD.
6536 if ((ImmWidth / (Is256 ? 2 : 1)) != 8) {
6537 table = lookup(Opcode, dom, ReplaceableCustomAVX2Instrs);
6538 AdjustBlendMask(Imm, ImmWidth, Is256 ? 8 : 4, &NewImm);
6541 assert(!Is256 && "128-bit vector expected");
6542 AdjustBlendMask(Imm, ImmWidth, 8, &NewImm);
6546 assert(table && table[Domain - 1] && "Unknown domain op");
6547 MI.setDesc(get(table[Domain - 1]));
6548 MI.getOperand(NumOperands - 1).setImm(NewImm & 255);
6554 case X86::BLENDPDrmi:
6555 case X86::BLENDPDrri:
6556 case X86::VBLENDPDrmi:
6557 case X86::VBLENDPDrri:
6558 return SetBlendDomain(2, false);
6559 case X86::VBLENDPDYrmi:
6560 case X86::VBLENDPDYrri:
6561 return SetBlendDomain(4, true);
6562 case X86::BLENDPSrmi:
6563 case X86::BLENDPSrri:
6564 case X86::VBLENDPSrmi:
6565 case X86::VBLENDPSrri:
6566 case X86::VPBLENDDrmi:
6567 case X86::VPBLENDDrri:
6568 return SetBlendDomain(4, false);
6569 case X86::VBLENDPSYrmi:
6570 case X86::VBLENDPSYrri:
6571 case X86::VPBLENDDYrmi:
6572 case X86::VPBLENDDYrri:
6573 return SetBlendDomain(8, true);
6574 case X86::PBLENDWrmi:
6575 case X86::PBLENDWrri:
6576 case X86::VPBLENDWrmi:
6577 case X86::VPBLENDWrri:
6578 return SetBlendDomain(8, false);
6579 case X86::VPBLENDWYrmi:
6580 case X86::VPBLENDWYrri:
6581 return SetBlendDomain(16, true);
6582 case X86::VPANDDZ128rr: case X86::VPANDDZ128rm:
6583 case X86::VPANDDZ256rr: case X86::VPANDDZ256rm:
6584 case X86::VPANDQZ128rr: case X86::VPANDQZ128rm:
6585 case X86::VPANDQZ256rr: case X86::VPANDQZ256rm:
6586 case X86::VPANDNDZ128rr: case X86::VPANDNDZ128rm:
6587 case X86::VPANDNDZ256rr: case X86::VPANDNDZ256rm:
6588 case X86::VPANDNQZ128rr: case X86::VPANDNQZ128rm:
6589 case X86::VPANDNQZ256rr: case X86::VPANDNQZ256rm:
6590 case X86::VPORDZ128rr: case X86::VPORDZ128rm:
6591 case X86::VPORDZ256rr: case X86::VPORDZ256rm:
6592 case X86::VPORQZ128rr: case X86::VPORQZ128rm:
6593 case X86::VPORQZ256rr: case X86::VPORQZ256rm:
6594 case X86::VPXORDZ128rr: case X86::VPXORDZ128rm:
6595 case X86::VPXORDZ256rr: case X86::VPXORDZ256rm:
6596 case X86::VPXORQZ128rr: case X86::VPXORQZ128rm:
6597 case X86::VPXORQZ256rr: case X86::VPXORQZ256rm: {
6598 // Without DQI, convert EVEX instructions to VEX instructions.
6599 if (Subtarget.hasDQI())
6602 const uint16_t *table = lookupAVX512(MI.getOpcode(), dom,
6603 ReplaceableCustomAVX512LogicInstrs);
6604 assert(table && "Instruction not found in table?");
6605 // Don't change integer Q instructions to D instructions and
6606 // use D intructions if we started with a PS instruction.
6607 if (Domain == 3 && (dom == 1 || table[3] == MI.getOpcode()))
6609 MI.setDesc(get(table[Domain - 1]));
6612 case X86::UNPCKHPDrr:
6613 case X86::MOVHLPSrr:
6614 // We just need to commute the instruction which will switch the domains.
6615 if (Domain != dom && Domain != 3 &&
6616 MI.getOperand(1).getReg() == MI.getOperand(2).getReg() &&
6617 MI.getOperand(0).getSubReg() == 0 &&
6618 MI.getOperand(1).getSubReg() == 0 &&
6619 MI.getOperand(2).getSubReg() == 0) {
6620 commuteInstruction(MI, false);
6623 // We must always return true for MOVHLPSrr.
6624 if (Opcode == X86::MOVHLPSrr)
6630 std::pair<uint16_t, uint16_t>
6631 X86InstrInfo::getExecutionDomain(const MachineInstr &MI) const {
6632 uint16_t domain = (MI.getDesc().TSFlags >> X86II::SSEDomainShift) & 3;
6633 unsigned opcode = MI.getOpcode();
6634 uint16_t validDomains = 0;
6636 // Attempt to match for custom instructions.
6637 validDomains = getExecutionDomainCustom(MI);
6639 return std::make_pair(domain, validDomains);
6641 if (lookup(opcode, domain, ReplaceableInstrs)) {
6643 } else if (lookup(opcode, domain, ReplaceableInstrsAVX2)) {
6644 validDomains = Subtarget.hasAVX2() ? 0xe : 0x6;
6645 } else if (lookup(opcode, domain, ReplaceableInstrsAVX2InsertExtract)) {
6646 // Insert/extract instructions should only effect domain if AVX2
6648 if (!Subtarget.hasAVX2())
6649 return std::make_pair(0, 0);
6651 } else if (lookupAVX512(opcode, domain, ReplaceableInstrsAVX512)) {
6653 } else if (Subtarget.hasDQI() && lookupAVX512(opcode, domain,
6654 ReplaceableInstrsAVX512DQ)) {
6656 } else if (Subtarget.hasDQI()) {
6657 if (const uint16_t *table = lookupAVX512(opcode, domain,
6658 ReplaceableInstrsAVX512DQMasked)) {
6659 if (domain == 1 || (domain == 3 && table[3] == opcode))
6666 return std::make_pair(domain, validDomains);
6669 void X86InstrInfo::setExecutionDomain(MachineInstr &MI, unsigned Domain) const {
6670 assert(Domain>0 && Domain<4 && "Invalid execution domain");
6671 uint16_t dom = (MI.getDesc().TSFlags >> X86II::SSEDomainShift) & 3;
6672 assert(dom && "Not an SSE instruction");
6674 // Attempt to match for custom instructions.
6675 if (setExecutionDomainCustom(MI, Domain))
6678 const uint16_t *table = lookup(MI.getOpcode(), dom, ReplaceableInstrs);
6679 if (!table) { // try the other table
6680 assert((Subtarget.hasAVX2() || Domain < 3) &&
6681 "256-bit vector operations only available in AVX2");
6682 table = lookup(MI.getOpcode(), dom, ReplaceableInstrsAVX2);
6684 if (!table) { // try the other table
6685 assert(Subtarget.hasAVX2() &&
6686 "256-bit insert/extract only available in AVX2");
6687 table = lookup(MI.getOpcode(), dom, ReplaceableInstrsAVX2InsertExtract);
6689 if (!table) { // try the AVX512 table
6690 assert(Subtarget.hasAVX512() && "Requires AVX-512");
6691 table = lookupAVX512(MI.getOpcode(), dom, ReplaceableInstrsAVX512);
6692 // Don't change integer Q instructions to D instructions.
6693 if (table && Domain == 3 && table[3] == MI.getOpcode())
6696 if (!table) { // try the AVX512DQ table
6697 assert((Subtarget.hasDQI() || Domain >= 3) && "Requires AVX-512DQ");
6698 table = lookupAVX512(MI.getOpcode(), dom, ReplaceableInstrsAVX512DQ);
6699 // Don't change integer Q instructions to D instructions and
6700 // use D intructions if we started with a PS instruction.
6701 if (table && Domain == 3 && (dom == 1 || table[3] == MI.getOpcode()))
6704 if (!table) { // try the AVX512DQMasked table
6705 assert((Subtarget.hasDQI() || Domain >= 3) && "Requires AVX-512DQ");
6706 table = lookupAVX512(MI.getOpcode(), dom, ReplaceableInstrsAVX512DQMasked);
6707 if (table && Domain == 3 && (dom == 1 || table[3] == MI.getOpcode()))
6710 assert(table && "Cannot change domain");
6711 MI.setDesc(get(table[Domain - 1]));
6714 /// Return the noop instruction to use for a noop.
6715 void X86InstrInfo::getNoop(MCInst &NopInst) const {
6716 NopInst.setOpcode(X86::NOOP);
6719 bool X86InstrInfo::isHighLatencyDef(int opc) const {
6721 default: return false;
6727 case X86::DIVSDrm_Int:
6729 case X86::DIVSDrr_Int:
6731 case X86::DIVSSrm_Int:
6733 case X86::DIVSSrr_Int:
6739 case X86::SQRTSDm_Int:
6741 case X86::SQRTSDr_Int:
6743 case X86::SQRTSSm_Int:
6745 case X86::SQRTSSr_Int:
6746 // AVX instructions with high latency
6749 case X86::VDIVPDYrm:
6750 case X86::VDIVPDYrr:
6753 case X86::VDIVPSYrm:
6754 case X86::VDIVPSYrr:
6756 case X86::VDIVSDrm_Int:
6758 case X86::VDIVSDrr_Int:
6760 case X86::VDIVSSrm_Int:
6762 case X86::VDIVSSrr_Int:
6765 case X86::VSQRTPDYm:
6766 case X86::VSQRTPDYr:
6769 case X86::VSQRTPSYm:
6770 case X86::VSQRTPSYr:
6772 case X86::VSQRTSDm_Int:
6774 case X86::VSQRTSDr_Int:
6776 case X86::VSQRTSSm_Int:
6778 case X86::VSQRTSSr_Int:
6779 // AVX512 instructions with high latency
6780 case X86::VDIVPDZ128rm:
6781 case X86::VDIVPDZ128rmb:
6782 case X86::VDIVPDZ128rmbk:
6783 case X86::VDIVPDZ128rmbkz:
6784 case X86::VDIVPDZ128rmk:
6785 case X86::VDIVPDZ128rmkz:
6786 case X86::VDIVPDZ128rr:
6787 case X86::VDIVPDZ128rrk:
6788 case X86::VDIVPDZ128rrkz:
6789 case X86::VDIVPDZ256rm:
6790 case X86::VDIVPDZ256rmb:
6791 case X86::VDIVPDZ256rmbk:
6792 case X86::VDIVPDZ256rmbkz:
6793 case X86::VDIVPDZ256rmk:
6794 case X86::VDIVPDZ256rmkz:
6795 case X86::VDIVPDZ256rr:
6796 case X86::VDIVPDZ256rrk:
6797 case X86::VDIVPDZ256rrkz:
6798 case X86::VDIVPDZrrb:
6799 case X86::VDIVPDZrrbk:
6800 case X86::VDIVPDZrrbkz:
6801 case X86::VDIVPDZrm:
6802 case X86::VDIVPDZrmb:
6803 case X86::VDIVPDZrmbk:
6804 case X86::VDIVPDZrmbkz:
6805 case X86::VDIVPDZrmk:
6806 case X86::VDIVPDZrmkz:
6807 case X86::VDIVPDZrr:
6808 case X86::VDIVPDZrrk:
6809 case X86::VDIVPDZrrkz:
6810 case X86::VDIVPSZ128rm:
6811 case X86::VDIVPSZ128rmb:
6812 case X86::VDIVPSZ128rmbk:
6813 case X86::VDIVPSZ128rmbkz:
6814 case X86::VDIVPSZ128rmk:
6815 case X86::VDIVPSZ128rmkz:
6816 case X86::VDIVPSZ128rr:
6817 case X86::VDIVPSZ128rrk:
6818 case X86::VDIVPSZ128rrkz:
6819 case X86::VDIVPSZ256rm:
6820 case X86::VDIVPSZ256rmb:
6821 case X86::VDIVPSZ256rmbk:
6822 case X86::VDIVPSZ256rmbkz:
6823 case X86::VDIVPSZ256rmk:
6824 case X86::VDIVPSZ256rmkz:
6825 case X86::VDIVPSZ256rr:
6826 case X86::VDIVPSZ256rrk:
6827 case X86::VDIVPSZ256rrkz:
6828 case X86::VDIVPSZrrb:
6829 case X86::VDIVPSZrrbk:
6830 case X86::VDIVPSZrrbkz:
6831 case X86::VDIVPSZrm:
6832 case X86::VDIVPSZrmb:
6833 case X86::VDIVPSZrmbk:
6834 case X86::VDIVPSZrmbkz:
6835 case X86::VDIVPSZrmk:
6836 case X86::VDIVPSZrmkz:
6837 case X86::VDIVPSZrr:
6838 case X86::VDIVPSZrrk:
6839 case X86::VDIVPSZrrkz:
6840 case X86::VDIVSDZrm:
6841 case X86::VDIVSDZrr:
6842 case X86::VDIVSDZrm_Int:
6843 case X86::VDIVSDZrm_Intk:
6844 case X86::VDIVSDZrm_Intkz:
6845 case X86::VDIVSDZrr_Int:
6846 case X86::VDIVSDZrr_Intk:
6847 case X86::VDIVSDZrr_Intkz:
6848 case X86::VDIVSDZrrb_Int:
6849 case X86::VDIVSDZrrb_Intk:
6850 case X86::VDIVSDZrrb_Intkz:
6851 case X86::VDIVSSZrm:
6852 case X86::VDIVSSZrr:
6853 case X86::VDIVSSZrm_Int:
6854 case X86::VDIVSSZrm_Intk:
6855 case X86::VDIVSSZrm_Intkz:
6856 case X86::VDIVSSZrr_Int:
6857 case X86::VDIVSSZrr_Intk:
6858 case X86::VDIVSSZrr_Intkz:
6859 case X86::VDIVSSZrrb_Int:
6860 case X86::VDIVSSZrrb_Intk:
6861 case X86::VDIVSSZrrb_Intkz:
6862 case X86::VSQRTPDZ128m:
6863 case X86::VSQRTPDZ128mb:
6864 case X86::VSQRTPDZ128mbk:
6865 case X86::VSQRTPDZ128mbkz:
6866 case X86::VSQRTPDZ128mk:
6867 case X86::VSQRTPDZ128mkz:
6868 case X86::VSQRTPDZ128r:
6869 case X86::VSQRTPDZ128rk:
6870 case X86::VSQRTPDZ128rkz:
6871 case X86::VSQRTPDZ256m:
6872 case X86::VSQRTPDZ256mb:
6873 case X86::VSQRTPDZ256mbk:
6874 case X86::VSQRTPDZ256mbkz:
6875 case X86::VSQRTPDZ256mk:
6876 case X86::VSQRTPDZ256mkz:
6877 case X86::VSQRTPDZ256r:
6878 case X86::VSQRTPDZ256rk:
6879 case X86::VSQRTPDZ256rkz:
6880 case X86::VSQRTPDZm:
6881 case X86::VSQRTPDZmb:
6882 case X86::VSQRTPDZmbk:
6883 case X86::VSQRTPDZmbkz:
6884 case X86::VSQRTPDZmk:
6885 case X86::VSQRTPDZmkz:
6886 case X86::VSQRTPDZr:
6887 case X86::VSQRTPDZrb:
6888 case X86::VSQRTPDZrbk:
6889 case X86::VSQRTPDZrbkz:
6890 case X86::VSQRTPDZrk:
6891 case X86::VSQRTPDZrkz:
6892 case X86::VSQRTPSZ128m:
6893 case X86::VSQRTPSZ128mb:
6894 case X86::VSQRTPSZ128mbk:
6895 case X86::VSQRTPSZ128mbkz:
6896 case X86::VSQRTPSZ128mk:
6897 case X86::VSQRTPSZ128mkz:
6898 case X86::VSQRTPSZ128r:
6899 case X86::VSQRTPSZ128rk:
6900 case X86::VSQRTPSZ128rkz:
6901 case X86::VSQRTPSZ256m:
6902 case X86::VSQRTPSZ256mb:
6903 case X86::VSQRTPSZ256mbk:
6904 case X86::VSQRTPSZ256mbkz:
6905 case X86::VSQRTPSZ256mk:
6906 case X86::VSQRTPSZ256mkz:
6907 case X86::VSQRTPSZ256r:
6908 case X86::VSQRTPSZ256rk:
6909 case X86::VSQRTPSZ256rkz:
6910 case X86::VSQRTPSZm:
6911 case X86::VSQRTPSZmb:
6912 case X86::VSQRTPSZmbk:
6913 case X86::VSQRTPSZmbkz:
6914 case X86::VSQRTPSZmk:
6915 case X86::VSQRTPSZmkz:
6916 case X86::VSQRTPSZr:
6917 case X86::VSQRTPSZrb:
6918 case X86::VSQRTPSZrbk:
6919 case X86::VSQRTPSZrbkz:
6920 case X86::VSQRTPSZrk:
6921 case X86::VSQRTPSZrkz:
6922 case X86::VSQRTSDZm:
6923 case X86::VSQRTSDZm_Int:
6924 case X86::VSQRTSDZm_Intk:
6925 case X86::VSQRTSDZm_Intkz:
6926 case X86::VSQRTSDZr:
6927 case X86::VSQRTSDZr_Int:
6928 case X86::VSQRTSDZr_Intk:
6929 case X86::VSQRTSDZr_Intkz:
6930 case X86::VSQRTSDZrb_Int:
6931 case X86::VSQRTSDZrb_Intk:
6932 case X86::VSQRTSDZrb_Intkz:
6933 case X86::VSQRTSSZm:
6934 case X86::VSQRTSSZm_Int:
6935 case X86::VSQRTSSZm_Intk:
6936 case X86::VSQRTSSZm_Intkz:
6937 case X86::VSQRTSSZr:
6938 case X86::VSQRTSSZr_Int:
6939 case X86::VSQRTSSZr_Intk:
6940 case X86::VSQRTSSZr_Intkz:
6941 case X86::VSQRTSSZrb_Int:
6942 case X86::VSQRTSSZrb_Intk:
6943 case X86::VSQRTSSZrb_Intkz:
6945 case X86::VGATHERDPDYrm:
6946 case X86::VGATHERDPDZ128rm:
6947 case X86::VGATHERDPDZ256rm:
6948 case X86::VGATHERDPDZrm:
6949 case X86::VGATHERDPDrm:
6950 case X86::VGATHERDPSYrm:
6951 case X86::VGATHERDPSZ128rm:
6952 case X86::VGATHERDPSZ256rm:
6953 case X86::VGATHERDPSZrm:
6954 case X86::VGATHERDPSrm:
6955 case X86::VGATHERPF0DPDm:
6956 case X86::VGATHERPF0DPSm:
6957 case X86::VGATHERPF0QPDm:
6958 case X86::VGATHERPF0QPSm:
6959 case X86::VGATHERPF1DPDm:
6960 case X86::VGATHERPF1DPSm:
6961 case X86::VGATHERPF1QPDm:
6962 case X86::VGATHERPF1QPSm:
6963 case X86::VGATHERQPDYrm:
6964 case X86::VGATHERQPDZ128rm:
6965 case X86::VGATHERQPDZ256rm:
6966 case X86::VGATHERQPDZrm:
6967 case X86::VGATHERQPDrm:
6968 case X86::VGATHERQPSYrm:
6969 case X86::VGATHERQPSZ128rm:
6970 case X86::VGATHERQPSZ256rm:
6971 case X86::VGATHERQPSZrm:
6972 case X86::VGATHERQPSrm:
6973 case X86::VPGATHERDDYrm:
6974 case X86::VPGATHERDDZ128rm:
6975 case X86::VPGATHERDDZ256rm:
6976 case X86::VPGATHERDDZrm:
6977 case X86::VPGATHERDDrm:
6978 case X86::VPGATHERDQYrm:
6979 case X86::VPGATHERDQZ128rm:
6980 case X86::VPGATHERDQZ256rm:
6981 case X86::VPGATHERDQZrm:
6982 case X86::VPGATHERDQrm:
6983 case X86::VPGATHERQDYrm:
6984 case X86::VPGATHERQDZ128rm:
6985 case X86::VPGATHERQDZ256rm:
6986 case X86::VPGATHERQDZrm:
6987 case X86::VPGATHERQDrm:
6988 case X86::VPGATHERQQYrm:
6989 case X86::VPGATHERQQZ128rm:
6990 case X86::VPGATHERQQZ256rm:
6991 case X86::VPGATHERQQZrm:
6992 case X86::VPGATHERQQrm:
6993 case X86::VSCATTERDPDZ128mr:
6994 case X86::VSCATTERDPDZ256mr:
6995 case X86::VSCATTERDPDZmr:
6996 case X86::VSCATTERDPSZ128mr:
6997 case X86::VSCATTERDPSZ256mr:
6998 case X86::VSCATTERDPSZmr:
6999 case X86::VSCATTERPF0DPDm:
7000 case X86::VSCATTERPF0DPSm:
7001 case X86::VSCATTERPF0QPDm:
7002 case X86::VSCATTERPF0QPSm:
7003 case X86::VSCATTERPF1DPDm:
7004 case X86::VSCATTERPF1DPSm:
7005 case X86::VSCATTERPF1QPDm:
7006 case X86::VSCATTERPF1QPSm:
7007 case X86::VSCATTERQPDZ128mr:
7008 case X86::VSCATTERQPDZ256mr:
7009 case X86::VSCATTERQPDZmr:
7010 case X86::VSCATTERQPSZ128mr:
7011 case X86::VSCATTERQPSZ256mr:
7012 case X86::VSCATTERQPSZmr:
7013 case X86::VPSCATTERDDZ128mr:
7014 case X86::VPSCATTERDDZ256mr:
7015 case X86::VPSCATTERDDZmr:
7016 case X86::VPSCATTERDQZ128mr:
7017 case X86::VPSCATTERDQZ256mr:
7018 case X86::VPSCATTERDQZmr:
7019 case X86::VPSCATTERQDZ128mr:
7020 case X86::VPSCATTERQDZ256mr:
7021 case X86::VPSCATTERQDZmr:
7022 case X86::VPSCATTERQQZ128mr:
7023 case X86::VPSCATTERQQZ256mr:
7024 case X86::VPSCATTERQQZmr:
7029 bool X86InstrInfo::hasHighOperandLatency(const TargetSchedModel &SchedModel,
7030 const MachineRegisterInfo *MRI,
7031 const MachineInstr &DefMI,
7033 const MachineInstr &UseMI,
7034 unsigned UseIdx) const {
7035 return isHighLatencyDef(DefMI.getOpcode());
7038 bool X86InstrInfo::hasReassociableOperands(const MachineInstr &Inst,
7039 const MachineBasicBlock *MBB) const {
7040 assert((Inst.getNumOperands() == 3 || Inst.getNumOperands() == 4) &&
7041 "Reassociation needs binary operators");
7043 // Integer binary math/logic instructions have a third source operand:
7044 // the EFLAGS register. That operand must be both defined here and never
7045 // used; ie, it must be dead. If the EFLAGS operand is live, then we can
7046 // not change anything because rearranging the operands could affect other
7047 // instructions that depend on the exact status flags (zero, sign, etc.)
7048 // that are set by using these particular operands with this operation.
7049 if (Inst.getNumOperands() == 4) {
7050 assert(Inst.getOperand(3).isReg() &&
7051 Inst.getOperand(3).getReg() == X86::EFLAGS &&
7052 "Unexpected operand in reassociable instruction");
7053 if (!Inst.getOperand(3).isDead())
7057 return TargetInstrInfo::hasReassociableOperands(Inst, MBB);
7060 // TODO: There are many more machine instruction opcodes to match:
7061 // 1. Other data types (integer, vectors)
7062 // 2. Other math / logic operations (xor, or)
7063 // 3. Other forms of the same operation (intrinsics and other variants)
7064 bool X86InstrInfo::isAssociativeAndCommutative(const MachineInstr &Inst) const {
7065 switch (Inst.getOpcode()) {
7096 case X86::VPANDDZ128rr:
7097 case X86::VPANDDZ256rr:
7098 case X86::VPANDDZrr:
7099 case X86::VPANDQZ128rr:
7100 case X86::VPANDQZ256rr:
7101 case X86::VPANDQZrr:
7104 case X86::VPORDZ128rr:
7105 case X86::VPORDZ256rr:
7107 case X86::VPORQZ128rr:
7108 case X86::VPORQZ256rr:
7112 case X86::VPXORDZ128rr:
7113 case X86::VPXORDZ256rr:
7114 case X86::VPXORDZrr:
7115 case X86::VPXORQZ128rr:
7116 case X86::VPXORQZ256rr:
7117 case X86::VPXORQZrr:
7120 case X86::VANDPDYrr:
7121 case X86::VANDPSYrr:
7122 case X86::VANDPDZ128rr:
7123 case X86::VANDPSZ128rr:
7124 case X86::VANDPDZ256rr:
7125 case X86::VANDPSZ256rr:
7126 case X86::VANDPDZrr:
7127 case X86::VANDPSZrr:
7132 case X86::VORPDZ128rr:
7133 case X86::VORPSZ128rr:
7134 case X86::VORPDZ256rr:
7135 case X86::VORPSZ256rr:
7140 case X86::VXORPDYrr:
7141 case X86::VXORPSYrr:
7142 case X86::VXORPDZ128rr:
7143 case X86::VXORPSZ128rr:
7144 case X86::VXORPDZ256rr:
7145 case X86::VXORPSZ256rr:
7146 case X86::VXORPDZrr:
7147 case X86::VXORPSZrr:
7168 case X86::VPADDBYrr:
7169 case X86::VPADDWYrr:
7170 case X86::VPADDDYrr:
7171 case X86::VPADDQYrr:
7172 case X86::VPADDBZ128rr:
7173 case X86::VPADDWZ128rr:
7174 case X86::VPADDDZ128rr:
7175 case X86::VPADDQZ128rr:
7176 case X86::VPADDBZ256rr:
7177 case X86::VPADDWZ256rr:
7178 case X86::VPADDDZ256rr:
7179 case X86::VPADDQZ256rr:
7180 case X86::VPADDBZrr:
7181 case X86::VPADDWZrr:
7182 case X86::VPADDDZrr:
7183 case X86::VPADDQZrr:
7184 case X86::VPMULLWrr:
7185 case X86::VPMULLWYrr:
7186 case X86::VPMULLWZ128rr:
7187 case X86::VPMULLWZ256rr:
7188 case X86::VPMULLWZrr:
7189 case X86::VPMULLDrr:
7190 case X86::VPMULLDYrr:
7191 case X86::VPMULLDZ128rr:
7192 case X86::VPMULLDZ256rr:
7193 case X86::VPMULLDZrr:
7194 case X86::VPMULLQZ128rr:
7195 case X86::VPMULLQZ256rr:
7196 case X86::VPMULLQZrr:
7197 // Normal min/max instructions are not commutative because of NaN and signed
7198 // zero semantics, but these are. Thus, there's no need to check for global
7199 // relaxed math; the instructions themselves have the properties we need.
7208 case X86::VMAXCPDrr:
7209 case X86::VMAXCPSrr:
7210 case X86::VMAXCPDYrr:
7211 case X86::VMAXCPSYrr:
7212 case X86::VMAXCPDZ128rr:
7213 case X86::VMAXCPSZ128rr:
7214 case X86::VMAXCPDZ256rr:
7215 case X86::VMAXCPSZ256rr:
7216 case X86::VMAXCPDZrr:
7217 case X86::VMAXCPSZrr:
7218 case X86::VMAXCSDrr:
7219 case X86::VMAXCSSrr:
7220 case X86::VMAXCSDZrr:
7221 case X86::VMAXCSSZrr:
7222 case X86::VMINCPDrr:
7223 case X86::VMINCPSrr:
7224 case X86::VMINCPDYrr:
7225 case X86::VMINCPSYrr:
7226 case X86::VMINCPDZ128rr:
7227 case X86::VMINCPSZ128rr:
7228 case X86::VMINCPDZ256rr:
7229 case X86::VMINCPSZ256rr:
7230 case X86::VMINCPDZrr:
7231 case X86::VMINCPSZrr:
7232 case X86::VMINCSDrr:
7233 case X86::VMINCSSrr:
7234 case X86::VMINCSDZrr:
7235 case X86::VMINCSSZrr:
7247 case X86::VADDPDYrr:
7248 case X86::VADDPSYrr:
7249 case X86::VADDPDZ128rr:
7250 case X86::VADDPSZ128rr:
7251 case X86::VADDPDZ256rr:
7252 case X86::VADDPSZ256rr:
7253 case X86::VADDPDZrr:
7254 case X86::VADDPSZrr:
7257 case X86::VADDSDZrr:
7258 case X86::VADDSSZrr:
7261 case X86::VMULPDYrr:
7262 case X86::VMULPSYrr:
7263 case X86::VMULPDZ128rr:
7264 case X86::VMULPSZ128rr:
7265 case X86::VMULPDZ256rr:
7266 case X86::VMULPSZ256rr:
7267 case X86::VMULPDZrr:
7268 case X86::VMULPSZrr:
7271 case X86::VMULSDZrr:
7272 case X86::VMULSSZrr:
7273 return Inst.getParent()->getParent()->getTarget().Options.UnsafeFPMath;
7279 /// This is an architecture-specific helper function of reassociateOps.
7280 /// Set special operand attributes for new instructions after reassociation.
7281 void X86InstrInfo::setSpecialOperandAttr(MachineInstr &OldMI1,
7282 MachineInstr &OldMI2,
7283 MachineInstr &NewMI1,
7284 MachineInstr &NewMI2) const {
7285 // Integer instructions define an implicit EFLAGS source register operand as
7286 // the third source (fourth total) operand.
7287 if (OldMI1.getNumOperands() != 4 || OldMI2.getNumOperands() != 4)
7290 assert(NewMI1.getNumOperands() == 4 && NewMI2.getNumOperands() == 4 &&
7291 "Unexpected instruction type for reassociation");
7293 MachineOperand &OldOp1 = OldMI1.getOperand(3);
7294 MachineOperand &OldOp2 = OldMI2.getOperand(3);
7295 MachineOperand &NewOp1 = NewMI1.getOperand(3);
7296 MachineOperand &NewOp2 = NewMI2.getOperand(3);
7298 assert(OldOp1.isReg() && OldOp1.getReg() == X86::EFLAGS && OldOp1.isDead() &&
7299 "Must have dead EFLAGS operand in reassociable instruction");
7300 assert(OldOp2.isReg() && OldOp2.getReg() == X86::EFLAGS && OldOp2.isDead() &&
7301 "Must have dead EFLAGS operand in reassociable instruction");
7306 assert(NewOp1.isReg() && NewOp1.getReg() == X86::EFLAGS &&
7307 "Unexpected operand in reassociable instruction");
7308 assert(NewOp2.isReg() && NewOp2.getReg() == X86::EFLAGS &&
7309 "Unexpected operand in reassociable instruction");
7311 // Mark the new EFLAGS operands as dead to be helpful to subsequent iterations
7312 // of this pass or other passes. The EFLAGS operands must be dead in these new
7313 // instructions because the EFLAGS operands in the original instructions must
7314 // be dead in order for reassociation to occur.
7319 std::pair<unsigned, unsigned>
7320 X86InstrInfo::decomposeMachineOperandsTargetFlags(unsigned TF) const {
7321 return std::make_pair(TF, 0u);
7324 ArrayRef<std::pair<unsigned, const char *>>
7325 X86InstrInfo::getSerializableDirectMachineOperandTargetFlags() const {
7326 using namespace X86II;
7327 static const std::pair<unsigned, const char *> TargetFlags[] = {
7328 {MO_GOT_ABSOLUTE_ADDRESS, "x86-got-absolute-address"},
7329 {MO_PIC_BASE_OFFSET, "x86-pic-base-offset"},
7330 {MO_GOT, "x86-got"},
7331 {MO_GOTOFF, "x86-gotoff"},
7332 {MO_GOTPCREL, "x86-gotpcrel"},
7333 {MO_PLT, "x86-plt"},
7334 {MO_TLSGD, "x86-tlsgd"},
7335 {MO_TLSLD, "x86-tlsld"},
7336 {MO_TLSLDM, "x86-tlsldm"},
7337 {MO_GOTTPOFF, "x86-gottpoff"},
7338 {MO_INDNTPOFF, "x86-indntpoff"},
7339 {MO_TPOFF, "x86-tpoff"},
7340 {MO_DTPOFF, "x86-dtpoff"},
7341 {MO_NTPOFF, "x86-ntpoff"},
7342 {MO_GOTNTPOFF, "x86-gotntpoff"},
7343 {MO_DLLIMPORT, "x86-dllimport"},
7344 {MO_DARWIN_NONLAZY, "x86-darwin-nonlazy"},
7345 {MO_DARWIN_NONLAZY_PIC_BASE, "x86-darwin-nonlazy-pic-base"},
7346 {MO_TLVP, "x86-tlvp"},
7347 {MO_TLVP_PIC_BASE, "x86-tlvp-pic-base"},
7348 {MO_SECREL, "x86-secrel"},
7349 {MO_COFFSTUB, "x86-coffstub"}};
7350 return makeArrayRef(TargetFlags);
7354 /// Create Global Base Reg pass. This initializes the PIC
7355 /// global base register for x86-32.
7356 struct CGBR : public MachineFunctionPass {
7358 CGBR() : MachineFunctionPass(ID) {}
7360 bool runOnMachineFunction(MachineFunction &MF) override {
7361 const X86TargetMachine *TM =
7362 static_cast<const X86TargetMachine *>(&MF.getTarget());
7363 const X86Subtarget &STI = MF.getSubtarget<X86Subtarget>();
7365 // Don't do anything in the 64-bit small and kernel code models. They use
7366 // RIP-relative addressing for everything.
7367 if (STI.is64Bit() && (TM->getCodeModel() == CodeModel::Small ||
7368 TM->getCodeModel() == CodeModel::Kernel))
7371 // Only emit a global base reg in PIC mode.
7372 if (!TM->isPositionIndependent())
7375 X86MachineFunctionInfo *X86FI = MF.getInfo<X86MachineFunctionInfo>();
7376 unsigned GlobalBaseReg = X86FI->getGlobalBaseReg();
7378 // If we didn't need a GlobalBaseReg, don't insert code.
7379 if (GlobalBaseReg == 0)
7382 // Insert the set of GlobalBaseReg into the first MBB of the function
7383 MachineBasicBlock &FirstMBB = MF.front();
7384 MachineBasicBlock::iterator MBBI = FirstMBB.begin();
7385 DebugLoc DL = FirstMBB.findDebugLoc(MBBI);
7386 MachineRegisterInfo &RegInfo = MF.getRegInfo();
7387 const X86InstrInfo *TII = STI.getInstrInfo();
7390 if (STI.isPICStyleGOT())
7391 PC = RegInfo.createVirtualRegister(&X86::GR32RegClass);
7395 if (STI.is64Bit()) {
7396 if (TM->getCodeModel() == CodeModel::Medium) {
7397 // In the medium code model, use a RIP-relative LEA to materialize the
7399 BuildMI(FirstMBB, MBBI, DL, TII->get(X86::LEA64r), PC)
7403 .addExternalSymbol("_GLOBAL_OFFSET_TABLE_")
7405 } else if (TM->getCodeModel() == CodeModel::Large) {
7406 // In the large code model, we are aiming for this code, though the
7407 // register allocation may vary:
7408 // leaq .LN$pb(%rip), %rax
7409 // movq $_GLOBAL_OFFSET_TABLE_ - .LN$pb, %rcx
7411 // RAX now holds address of _GLOBAL_OFFSET_TABLE_.
7412 unsigned PBReg = RegInfo.createVirtualRegister(&X86::GR64RegClass);
7414 RegInfo.createVirtualRegister(&X86::GR64RegClass);
7415 BuildMI(FirstMBB, MBBI, DL, TII->get(X86::LEA64r), PBReg)
7419 .addSym(MF.getPICBaseSymbol())
7421 std::prev(MBBI)->setPreInstrSymbol(MF, MF.getPICBaseSymbol());
7422 BuildMI(FirstMBB, MBBI, DL, TII->get(X86::MOV64ri), GOTReg)
7423 .addExternalSymbol("_GLOBAL_OFFSET_TABLE_",
7424 X86II::MO_PIC_BASE_OFFSET);
7425 BuildMI(FirstMBB, MBBI, DL, TII->get(X86::ADD64rr), PC)
7426 .addReg(PBReg, RegState::Kill)
7427 .addReg(GOTReg, RegState::Kill);
7429 llvm_unreachable("unexpected code model");
7432 // Operand of MovePCtoStack is completely ignored by asm printer. It's
7433 // only used in JIT code emission as displacement to pc.
7434 BuildMI(FirstMBB, MBBI, DL, TII->get(X86::MOVPC32r), PC).addImm(0);
7436 // If we're using vanilla 'GOT' PIC style, we should use relative
7437 // addressing not to pc, but to _GLOBAL_OFFSET_TABLE_ external.
7438 if (STI.isPICStyleGOT()) {
7439 // Generate addl $__GLOBAL_OFFSET_TABLE_ + [.-piclabel],
7441 BuildMI(FirstMBB, MBBI, DL, TII->get(X86::ADD32ri), GlobalBaseReg)
7443 .addExternalSymbol("_GLOBAL_OFFSET_TABLE_",
7444 X86II::MO_GOT_ABSOLUTE_ADDRESS);
7451 StringRef getPassName() const override {
7452 return "X86 PIC Global Base Reg Initialization";
7455 void getAnalysisUsage(AnalysisUsage &AU) const override {
7456 AU.setPreservesCFG();
7457 MachineFunctionPass::getAnalysisUsage(AU);
7464 llvm::createX86GlobalBaseRegPass() { return new CGBR(); }
7467 struct LDTLSCleanup : public MachineFunctionPass {
7469 LDTLSCleanup() : MachineFunctionPass(ID) {}
7471 bool runOnMachineFunction(MachineFunction &MF) override {
7472 if (skipFunction(MF.getFunction()))
7475 X86MachineFunctionInfo *MFI = MF.getInfo<X86MachineFunctionInfo>();
7476 if (MFI->getNumLocalDynamicTLSAccesses() < 2) {
7477 // No point folding accesses if there isn't at least two.
7481 MachineDominatorTree *DT = &getAnalysis<MachineDominatorTree>();
7482 return VisitNode(DT->getRootNode(), 0);
7485 // Visit the dominator subtree rooted at Node in pre-order.
7486 // If TLSBaseAddrReg is non-null, then use that to replace any
7487 // TLS_base_addr instructions. Otherwise, create the register
7488 // when the first such instruction is seen, and then use it
7489 // as we encounter more instructions.
7490 bool VisitNode(MachineDomTreeNode *Node, unsigned TLSBaseAddrReg) {
7491 MachineBasicBlock *BB = Node->getBlock();
7492 bool Changed = false;
7494 // Traverse the current block.
7495 for (MachineBasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;
7497 switch (I->getOpcode()) {
7498 case X86::TLS_base_addr32:
7499 case X86::TLS_base_addr64:
7501 I = ReplaceTLSBaseAddrCall(*I, TLSBaseAddrReg);
7503 I = SetRegister(*I, &TLSBaseAddrReg);
7511 // Visit the children of this block in the dominator tree.
7512 for (MachineDomTreeNode::iterator I = Node->begin(), E = Node->end();
7514 Changed |= VisitNode(*I, TLSBaseAddrReg);
7520 // Replace the TLS_base_addr instruction I with a copy from
7521 // TLSBaseAddrReg, returning the new instruction.
7522 MachineInstr *ReplaceTLSBaseAddrCall(MachineInstr &I,
7523 unsigned TLSBaseAddrReg) {
7524 MachineFunction *MF = I.getParent()->getParent();
7525 const X86Subtarget &STI = MF->getSubtarget<X86Subtarget>();
7526 const bool is64Bit = STI.is64Bit();
7527 const X86InstrInfo *TII = STI.getInstrInfo();
7529 // Insert a Copy from TLSBaseAddrReg to RAX/EAX.
7530 MachineInstr *Copy =
7531 BuildMI(*I.getParent(), I, I.getDebugLoc(),
7532 TII->get(TargetOpcode::COPY), is64Bit ? X86::RAX : X86::EAX)
7533 .addReg(TLSBaseAddrReg);
7535 // Erase the TLS_base_addr instruction.
7536 I.eraseFromParent();
7541 // Create a virtual register in *TLSBaseAddrReg, and populate it by
7542 // inserting a copy instruction after I. Returns the new instruction.
7543 MachineInstr *SetRegister(MachineInstr &I, unsigned *TLSBaseAddrReg) {
7544 MachineFunction *MF = I.getParent()->getParent();
7545 const X86Subtarget &STI = MF->getSubtarget<X86Subtarget>();
7546 const bool is64Bit = STI.is64Bit();
7547 const X86InstrInfo *TII = STI.getInstrInfo();
7549 // Create a virtual register for the TLS base address.
7550 MachineRegisterInfo &RegInfo = MF->getRegInfo();
7551 *TLSBaseAddrReg = RegInfo.createVirtualRegister(is64Bit
7552 ? &X86::GR64RegClass
7553 : &X86::GR32RegClass);
7555 // Insert a copy from RAX/EAX to TLSBaseAddrReg.
7556 MachineInstr *Next = I.getNextNode();
7557 MachineInstr *Copy =
7558 BuildMI(*I.getParent(), Next, I.getDebugLoc(),
7559 TII->get(TargetOpcode::COPY), *TLSBaseAddrReg)
7560 .addReg(is64Bit ? X86::RAX : X86::EAX);
7565 StringRef getPassName() const override {
7566 return "Local Dynamic TLS Access Clean-up";
7569 void getAnalysisUsage(AnalysisUsage &AU) const override {
7570 AU.setPreservesCFG();
7571 AU.addRequired<MachineDominatorTree>();
7572 MachineFunctionPass::getAnalysisUsage(AU);
7577 char LDTLSCleanup::ID = 0;
7579 llvm::createCleanupLocalDynamicTLSPass() { return new LDTLSCleanup(); }
7581 /// Constants defining how certain sequences should be outlined.
7583 /// \p MachineOutlinerDefault implies that the function is called with a call
7584 /// instruction, and a return must be emitted for the outlined function frame.
7588 /// I1 OUTLINED_FUNCTION:
7589 /// I2 --> call OUTLINED_FUNCTION I1
7594 /// * Call construction overhead: 1 (call instruction)
7595 /// * Frame construction overhead: 1 (return instruction)
7597 /// \p MachineOutlinerTailCall implies that the function is being tail called.
7598 /// A jump is emitted instead of a call, and the return is already present in
7599 /// the outlined sequence. That is,
7601 /// I1 OUTLINED_FUNCTION:
7602 /// I2 --> jmp OUTLINED_FUNCTION I1
7606 /// * Call construction overhead: 1 (jump instruction)
7607 /// * Frame construction overhead: 0 (don't need to return)
7609 enum MachineOutlinerClass {
7610 MachineOutlinerDefault,
7611 MachineOutlinerTailCall
7614 outliner::OutlinedFunction X86InstrInfo::getOutliningCandidateInfo(
7615 std::vector<outliner::Candidate> &RepeatedSequenceLocs) const {
7616 unsigned SequenceSize =
7617 std::accumulate(RepeatedSequenceLocs[0].front(),
7618 std::next(RepeatedSequenceLocs[0].back()), 0,
7619 [](unsigned Sum, const MachineInstr &MI) {
7620 // FIXME: x86 doesn't implement getInstSizeInBytes, so
7621 // we can't tell the cost. Just assume each instruction
7623 if (MI.isDebugInstr() || MI.isKill())
7628 // FIXME: Use real size in bytes for call and ret instructions.
7629 if (RepeatedSequenceLocs[0].back()->isTerminator()) {
7630 for (outliner::Candidate &C : RepeatedSequenceLocs)
7631 C.setCallInfo(MachineOutlinerTailCall, 1);
7633 return outliner::OutlinedFunction(RepeatedSequenceLocs, SequenceSize,
7634 0, // Number of bytes to emit frame.
7635 MachineOutlinerTailCall // Type of frame.
7639 for (outliner::Candidate &C : RepeatedSequenceLocs)
7640 C.setCallInfo(MachineOutlinerDefault, 1);
7642 return outliner::OutlinedFunction(RepeatedSequenceLocs, SequenceSize, 1,
7643 MachineOutlinerDefault);
7646 bool X86InstrInfo::isFunctionSafeToOutlineFrom(MachineFunction &MF,
7647 bool OutlineFromLinkOnceODRs) const {
7648 const Function &F = MF.getFunction();
7650 // Does the function use a red zone? If it does, then we can't risk messing
7652 if (!F.hasFnAttribute(Attribute::NoRedZone)) {
7653 // It could have a red zone. If it does, then we don't want to touch it.
7654 const X86MachineFunctionInfo *X86FI = MF.getInfo<X86MachineFunctionInfo>();
7655 if (!X86FI || X86FI->getUsesRedZone())
7659 // If we *don't* want to outline from things that could potentially be deduped
7660 // then return false.
7661 if (!OutlineFromLinkOnceODRs && F.hasLinkOnceODRLinkage())
7664 // This function is viable for outlining, so return true.
7669 X86InstrInfo::getOutliningType(MachineBasicBlock::iterator &MIT, unsigned Flags) const {
7670 MachineInstr &MI = *MIT;
7671 // Don't allow debug values to impact outlining type.
7672 if (MI.isDebugInstr() || MI.isIndirectDebugValue())
7673 return outliner::InstrType::Invisible;
7675 // At this point, KILL instructions don't really tell us much so we can go
7676 // ahead and skip over them.
7678 return outliner::InstrType::Invisible;
7680 // Is this a tail call? If yes, we can outline as a tail call.
7682 return outliner::InstrType::Legal;
7684 // Is this the terminator of a basic block?
7685 if (MI.isTerminator() || MI.isReturn()) {
7687 // Does its parent have any successors in its MachineFunction?
7688 if (MI.getParent()->succ_empty())
7689 return outliner::InstrType::Legal;
7691 // It does, so we can't tail call it.
7692 return outliner::InstrType::Illegal;
7695 // Don't outline anything that modifies or reads from the stack pointer.
7697 // FIXME: There are instructions which are being manually built without
7698 // explicit uses/defs so we also have to check the MCInstrDesc. We should be
7699 // able to remove the extra checks once those are fixed up. For example,
7700 // sometimes we might get something like %rax = POP64r 1. This won't be
7701 // caught by modifiesRegister or readsRegister even though the instruction
7702 // really ought to be formed so that modifiesRegister/readsRegister would
7704 if (MI.modifiesRegister(X86::RSP, &RI) || MI.readsRegister(X86::RSP, &RI) ||
7705 MI.getDesc().hasImplicitUseOfPhysReg(X86::RSP) ||
7706 MI.getDesc().hasImplicitDefOfPhysReg(X86::RSP))
7707 return outliner::InstrType::Illegal;
7709 // Outlined calls change the instruction pointer, so don't read from it.
7710 if (MI.readsRegister(X86::RIP, &RI) ||
7711 MI.getDesc().hasImplicitUseOfPhysReg(X86::RIP) ||
7712 MI.getDesc().hasImplicitDefOfPhysReg(X86::RIP))
7713 return outliner::InstrType::Illegal;
7715 // Positions can't safely be outlined.
7716 if (MI.isPosition())
7717 return outliner::InstrType::Illegal;
7719 // Make sure none of the operands of this instruction do anything tricky.
7720 for (const MachineOperand &MOP : MI.operands())
7721 if (MOP.isCPI() || MOP.isJTI() || MOP.isCFIIndex() || MOP.isFI() ||
7722 MOP.isTargetIndex())
7723 return outliner::InstrType::Illegal;
7725 return outliner::InstrType::Legal;
7728 void X86InstrInfo::buildOutlinedFrame(MachineBasicBlock &MBB,
7729 MachineFunction &MF,
7730 const outliner::OutlinedFunction &OF)
7732 // If we're a tail call, we already have a return, so don't do anything.
7733 if (OF.FrameConstructionID == MachineOutlinerTailCall)
7736 // We're a normal call, so our sequence doesn't have a return instruction.
7738 MachineInstr *retq = BuildMI(MF, DebugLoc(), get(X86::RETQ));
7739 MBB.insert(MBB.end(), retq);
7742 MachineBasicBlock::iterator
7743 X86InstrInfo::insertOutlinedCall(Module &M, MachineBasicBlock &MBB,
7744 MachineBasicBlock::iterator &It,
7745 MachineFunction &MF,
7746 const outliner::Candidate &C) const {
7747 // Is it a tail call?
7748 if (C.CallConstructionID == MachineOutlinerTailCall) {
7749 // Yes, just insert a JMP.
7751 BuildMI(MF, DebugLoc(), get(X86::TAILJMPd64))
7752 .addGlobalAddress(M.getNamedValue(MF.getName())));
7754 // No, insert a call.
7756 BuildMI(MF, DebugLoc(), get(X86::CALL64pcrel32))
7757 .addGlobalAddress(M.getNamedValue(MF.getName())));
7763 #define GET_INSTRINFO_HELPERS
7764 #include "X86GenInstrInfo.inc"