[DAGCombiner] Add target hook function to decide folding (mul (add x, c1), c2)
[llvm-project.git] / llvm / lib / CodeGen / TargetInstrInfo.cpp
blob1eab8e7443a73701768f962db9bc9282a13d257d
1 //===-- TargetInstrInfo.cpp - Target Instruction Information --------------===//
2 //
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
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 // This file implements the TargetInstrInfo class.
11 //===----------------------------------------------------------------------===//
13 #include "llvm/CodeGen/TargetInstrInfo.h"
14 #include "llvm/ADT/StringExtras.h"
15 #include "llvm/CodeGen/MachineFrameInfo.h"
16 #include "llvm/CodeGen/MachineInstrBuilder.h"
17 #include "llvm/CodeGen/MachineMemOperand.h"
18 #include "llvm/CodeGen/MachineRegisterInfo.h"
19 #include "llvm/CodeGen/MachineScheduler.h"
20 #include "llvm/CodeGen/PseudoSourceValue.h"
21 #include "llvm/CodeGen/ScoreboardHazardRecognizer.h"
22 #include "llvm/CodeGen/StackMaps.h"
23 #include "llvm/CodeGen/TargetFrameLowering.h"
24 #include "llvm/CodeGen/TargetLowering.h"
25 #include "llvm/CodeGen/TargetRegisterInfo.h"
26 #include "llvm/CodeGen/TargetSchedule.h"
27 #include "llvm/IR/DataLayout.h"
28 #include "llvm/IR/DebugInfoMetadata.h"
29 #include "llvm/MC/MCAsmInfo.h"
30 #include "llvm/MC/MCInstrItineraries.h"
31 #include "llvm/Support/CommandLine.h"
32 #include "llvm/Support/ErrorHandling.h"
33 #include "llvm/Support/raw_ostream.h"
34 #include "llvm/Target/TargetMachine.h"
35 #include <cctype>
37 using namespace llvm;
39 static cl::opt<bool> DisableHazardRecognizer(
40 "disable-sched-hazard", cl::Hidden, cl::init(false),
41 cl::desc("Disable hazard detection during preRA scheduling"));
43 TargetInstrInfo::~TargetInstrInfo() {
46 const TargetRegisterClass*
47 TargetInstrInfo::getRegClass(const MCInstrDesc &MCID, unsigned OpNum,
48 const TargetRegisterInfo *TRI,
49 const MachineFunction &MF) const {
50 if (OpNum >= MCID.getNumOperands())
51 return nullptr;
53 short RegClass = MCID.OpInfo[OpNum].RegClass;
54 if (MCID.OpInfo[OpNum].isLookupPtrRegClass())
55 return TRI->getPointerRegClass(MF, RegClass);
57 // Instructions like INSERT_SUBREG do not have fixed register classes.
58 if (RegClass < 0)
59 return nullptr;
61 // Otherwise just look it up normally.
62 return TRI->getRegClass(RegClass);
65 /// insertNoop - Insert a noop into the instruction stream at the specified
66 /// point.
67 void TargetInstrInfo::insertNoop(MachineBasicBlock &MBB,
68 MachineBasicBlock::iterator MI) const {
69 llvm_unreachable("Target didn't implement insertNoop!");
72 /// insertNoops - Insert noops into the instruction stream at the specified
73 /// point.
74 void TargetInstrInfo::insertNoops(MachineBasicBlock &MBB,
75 MachineBasicBlock::iterator MI,
76 unsigned Quantity) const {
77 for (unsigned i = 0; i < Quantity; ++i)
78 insertNoop(MBB, MI);
81 static bool isAsmComment(const char *Str, const MCAsmInfo &MAI) {
82 return strncmp(Str, MAI.getCommentString().data(),
83 MAI.getCommentString().size()) == 0;
86 /// Measure the specified inline asm to determine an approximation of its
87 /// length.
88 /// Comments (which run till the next SeparatorString or newline) do not
89 /// count as an instruction.
90 /// Any other non-whitespace text is considered an instruction, with
91 /// multiple instructions separated by SeparatorString or newlines.
92 /// Variable-length instructions are not handled here; this function
93 /// may be overloaded in the target code to do that.
94 /// We implement a special case of the .space directive which takes only a
95 /// single integer argument in base 10 that is the size in bytes. This is a
96 /// restricted form of the GAS directive in that we only interpret
97 /// simple--i.e. not a logical or arithmetic expression--size values without
98 /// the optional fill value. This is primarily used for creating arbitrary
99 /// sized inline asm blocks for testing purposes.
100 unsigned TargetInstrInfo::getInlineAsmLength(
101 const char *Str,
102 const MCAsmInfo &MAI, const TargetSubtargetInfo *STI) const {
103 // Count the number of instructions in the asm.
104 bool AtInsnStart = true;
105 unsigned Length = 0;
106 const unsigned MaxInstLength = MAI.getMaxInstLength(STI);
107 for (; *Str; ++Str) {
108 if (*Str == '\n' || strncmp(Str, MAI.getSeparatorString(),
109 strlen(MAI.getSeparatorString())) == 0) {
110 AtInsnStart = true;
111 } else if (isAsmComment(Str, MAI)) {
112 // Stop counting as an instruction after a comment until the next
113 // separator.
114 AtInsnStart = false;
117 if (AtInsnStart && !isSpace(static_cast<unsigned char>(*Str))) {
118 unsigned AddLength = MaxInstLength;
119 if (strncmp(Str, ".space", 6) == 0) {
120 char *EStr;
121 int SpaceSize;
122 SpaceSize = strtol(Str + 6, &EStr, 10);
123 SpaceSize = SpaceSize < 0 ? 0 : SpaceSize;
124 while (*EStr != '\n' && isSpace(static_cast<unsigned char>(*EStr)))
125 ++EStr;
126 if (*EStr == '\0' || *EStr == '\n' ||
127 isAsmComment(EStr, MAI)) // Successfully parsed .space argument
128 AddLength = SpaceSize;
130 Length += AddLength;
131 AtInsnStart = false;
135 return Length;
138 /// ReplaceTailWithBranchTo - Delete the instruction OldInst and everything
139 /// after it, replacing it with an unconditional branch to NewDest.
140 void
141 TargetInstrInfo::ReplaceTailWithBranchTo(MachineBasicBlock::iterator Tail,
142 MachineBasicBlock *NewDest) const {
143 MachineBasicBlock *MBB = Tail->getParent();
145 // Remove all the old successors of MBB from the CFG.
146 while (!MBB->succ_empty())
147 MBB->removeSuccessor(MBB->succ_begin());
149 // Save off the debug loc before erasing the instruction.
150 DebugLoc DL = Tail->getDebugLoc();
152 // Update call site info and remove all the dead instructions
153 // from the end of MBB.
154 while (Tail != MBB->end()) {
155 auto MI = Tail++;
156 if (MI->shouldUpdateCallSiteInfo())
157 MBB->getParent()->eraseCallSiteInfo(&*MI);
158 MBB->erase(MI);
161 // If MBB isn't immediately before MBB, insert a branch to it.
162 if (++MachineFunction::iterator(MBB) != MachineFunction::iterator(NewDest))
163 insertBranch(*MBB, NewDest, nullptr, SmallVector<MachineOperand, 0>(), DL);
164 MBB->addSuccessor(NewDest);
167 MachineInstr *TargetInstrInfo::commuteInstructionImpl(MachineInstr &MI,
168 bool NewMI, unsigned Idx1,
169 unsigned Idx2) const {
170 const MCInstrDesc &MCID = MI.getDesc();
171 bool HasDef = MCID.getNumDefs();
172 if (HasDef && !MI.getOperand(0).isReg())
173 // No idea how to commute this instruction. Target should implement its own.
174 return nullptr;
176 unsigned CommutableOpIdx1 = Idx1; (void)CommutableOpIdx1;
177 unsigned CommutableOpIdx2 = Idx2; (void)CommutableOpIdx2;
178 assert(findCommutedOpIndices(MI, CommutableOpIdx1, CommutableOpIdx2) &&
179 CommutableOpIdx1 == Idx1 && CommutableOpIdx2 == Idx2 &&
180 "TargetInstrInfo::CommuteInstructionImpl(): not commutable operands.");
181 assert(MI.getOperand(Idx1).isReg() && MI.getOperand(Idx2).isReg() &&
182 "This only knows how to commute register operands so far");
184 Register Reg0 = HasDef ? MI.getOperand(0).getReg() : Register();
185 Register Reg1 = MI.getOperand(Idx1).getReg();
186 Register Reg2 = MI.getOperand(Idx2).getReg();
187 unsigned SubReg0 = HasDef ? MI.getOperand(0).getSubReg() : 0;
188 unsigned SubReg1 = MI.getOperand(Idx1).getSubReg();
189 unsigned SubReg2 = MI.getOperand(Idx2).getSubReg();
190 bool Reg1IsKill = MI.getOperand(Idx1).isKill();
191 bool Reg2IsKill = MI.getOperand(Idx2).isKill();
192 bool Reg1IsUndef = MI.getOperand(Idx1).isUndef();
193 bool Reg2IsUndef = MI.getOperand(Idx2).isUndef();
194 bool Reg1IsInternal = MI.getOperand(Idx1).isInternalRead();
195 bool Reg2IsInternal = MI.getOperand(Idx2).isInternalRead();
196 // Avoid calling isRenamable for virtual registers since we assert that
197 // renamable property is only queried/set for physical registers.
198 bool Reg1IsRenamable = Register::isPhysicalRegister(Reg1)
199 ? MI.getOperand(Idx1).isRenamable()
200 : false;
201 bool Reg2IsRenamable = Register::isPhysicalRegister(Reg2)
202 ? MI.getOperand(Idx2).isRenamable()
203 : false;
204 // If destination is tied to either of the commuted source register, then
205 // it must be updated.
206 if (HasDef && Reg0 == Reg1 &&
207 MI.getDesc().getOperandConstraint(Idx1, MCOI::TIED_TO) == 0) {
208 Reg2IsKill = false;
209 Reg0 = Reg2;
210 SubReg0 = SubReg2;
211 } else if (HasDef && Reg0 == Reg2 &&
212 MI.getDesc().getOperandConstraint(Idx2, MCOI::TIED_TO) == 0) {
213 Reg1IsKill = false;
214 Reg0 = Reg1;
215 SubReg0 = SubReg1;
218 MachineInstr *CommutedMI = nullptr;
219 if (NewMI) {
220 // Create a new instruction.
221 MachineFunction &MF = *MI.getMF();
222 CommutedMI = MF.CloneMachineInstr(&MI);
223 } else {
224 CommutedMI = &MI;
227 if (HasDef) {
228 CommutedMI->getOperand(0).setReg(Reg0);
229 CommutedMI->getOperand(0).setSubReg(SubReg0);
231 CommutedMI->getOperand(Idx2).setReg(Reg1);
232 CommutedMI->getOperand(Idx1).setReg(Reg2);
233 CommutedMI->getOperand(Idx2).setSubReg(SubReg1);
234 CommutedMI->getOperand(Idx1).setSubReg(SubReg2);
235 CommutedMI->getOperand(Idx2).setIsKill(Reg1IsKill);
236 CommutedMI->getOperand(Idx1).setIsKill(Reg2IsKill);
237 CommutedMI->getOperand(Idx2).setIsUndef(Reg1IsUndef);
238 CommutedMI->getOperand(Idx1).setIsUndef(Reg2IsUndef);
239 CommutedMI->getOperand(Idx2).setIsInternalRead(Reg1IsInternal);
240 CommutedMI->getOperand(Idx1).setIsInternalRead(Reg2IsInternal);
241 // Avoid calling setIsRenamable for virtual registers since we assert that
242 // renamable property is only queried/set for physical registers.
243 if (Register::isPhysicalRegister(Reg1))
244 CommutedMI->getOperand(Idx2).setIsRenamable(Reg1IsRenamable);
245 if (Register::isPhysicalRegister(Reg2))
246 CommutedMI->getOperand(Idx1).setIsRenamable(Reg2IsRenamable);
247 return CommutedMI;
250 MachineInstr *TargetInstrInfo::commuteInstruction(MachineInstr &MI, bool NewMI,
251 unsigned OpIdx1,
252 unsigned OpIdx2) const {
253 // If OpIdx1 or OpIdx2 is not specified, then this method is free to choose
254 // any commutable operand, which is done in findCommutedOpIndices() method
255 // called below.
256 if ((OpIdx1 == CommuteAnyOperandIndex || OpIdx2 == CommuteAnyOperandIndex) &&
257 !findCommutedOpIndices(MI, OpIdx1, OpIdx2)) {
258 assert(MI.isCommutable() &&
259 "Precondition violation: MI must be commutable.");
260 return nullptr;
262 return commuteInstructionImpl(MI, NewMI, OpIdx1, OpIdx2);
265 bool TargetInstrInfo::fixCommutedOpIndices(unsigned &ResultIdx1,
266 unsigned &ResultIdx2,
267 unsigned CommutableOpIdx1,
268 unsigned CommutableOpIdx2) {
269 if (ResultIdx1 == CommuteAnyOperandIndex &&
270 ResultIdx2 == CommuteAnyOperandIndex) {
271 ResultIdx1 = CommutableOpIdx1;
272 ResultIdx2 = CommutableOpIdx2;
273 } else if (ResultIdx1 == CommuteAnyOperandIndex) {
274 if (ResultIdx2 == CommutableOpIdx1)
275 ResultIdx1 = CommutableOpIdx2;
276 else if (ResultIdx2 == CommutableOpIdx2)
277 ResultIdx1 = CommutableOpIdx1;
278 else
279 return false;
280 } else if (ResultIdx2 == CommuteAnyOperandIndex) {
281 if (ResultIdx1 == CommutableOpIdx1)
282 ResultIdx2 = CommutableOpIdx2;
283 else if (ResultIdx1 == CommutableOpIdx2)
284 ResultIdx2 = CommutableOpIdx1;
285 else
286 return false;
287 } else
288 // Check that the result operand indices match the given commutable
289 // operand indices.
290 return (ResultIdx1 == CommutableOpIdx1 && ResultIdx2 == CommutableOpIdx2) ||
291 (ResultIdx1 == CommutableOpIdx2 && ResultIdx2 == CommutableOpIdx1);
293 return true;
296 bool TargetInstrInfo::findCommutedOpIndices(const MachineInstr &MI,
297 unsigned &SrcOpIdx1,
298 unsigned &SrcOpIdx2) const {
299 assert(!MI.isBundle() &&
300 "TargetInstrInfo::findCommutedOpIndices() can't handle bundles");
302 const MCInstrDesc &MCID = MI.getDesc();
303 if (!MCID.isCommutable())
304 return false;
306 // This assumes v0 = op v1, v2 and commuting would swap v1 and v2. If this
307 // is not true, then the target must implement this.
308 unsigned CommutableOpIdx1 = MCID.getNumDefs();
309 unsigned CommutableOpIdx2 = CommutableOpIdx1 + 1;
310 if (!fixCommutedOpIndices(SrcOpIdx1, SrcOpIdx2,
311 CommutableOpIdx1, CommutableOpIdx2))
312 return false;
314 if (!MI.getOperand(SrcOpIdx1).isReg() || !MI.getOperand(SrcOpIdx2).isReg())
315 // No idea.
316 return false;
317 return true;
320 bool TargetInstrInfo::isUnpredicatedTerminator(const MachineInstr &MI) const {
321 if (!MI.isTerminator()) return false;
323 // Conditional branch is a special case.
324 if (MI.isBranch() && !MI.isBarrier())
325 return true;
326 if (!MI.isPredicable())
327 return true;
328 return !isPredicated(MI);
331 bool TargetInstrInfo::PredicateInstruction(
332 MachineInstr &MI, ArrayRef<MachineOperand> Pred) const {
333 bool MadeChange = false;
335 assert(!MI.isBundle() &&
336 "TargetInstrInfo::PredicateInstruction() can't handle bundles");
338 const MCInstrDesc &MCID = MI.getDesc();
339 if (!MI.isPredicable())
340 return false;
342 for (unsigned j = 0, i = 0, e = MI.getNumOperands(); i != e; ++i) {
343 if (MCID.OpInfo[i].isPredicate()) {
344 MachineOperand &MO = MI.getOperand(i);
345 if (MO.isReg()) {
346 MO.setReg(Pred[j].getReg());
347 MadeChange = true;
348 } else if (MO.isImm()) {
349 MO.setImm(Pred[j].getImm());
350 MadeChange = true;
351 } else if (MO.isMBB()) {
352 MO.setMBB(Pred[j].getMBB());
353 MadeChange = true;
355 ++j;
358 return MadeChange;
361 bool TargetInstrInfo::hasLoadFromStackSlot(
362 const MachineInstr &MI,
363 SmallVectorImpl<const MachineMemOperand *> &Accesses) const {
364 size_t StartSize = Accesses.size();
365 for (MachineInstr::mmo_iterator o = MI.memoperands_begin(),
366 oe = MI.memoperands_end();
367 o != oe; ++o) {
368 if ((*o)->isLoad() &&
369 dyn_cast_or_null<FixedStackPseudoSourceValue>((*o)->getPseudoValue()))
370 Accesses.push_back(*o);
372 return Accesses.size() != StartSize;
375 bool TargetInstrInfo::hasStoreToStackSlot(
376 const MachineInstr &MI,
377 SmallVectorImpl<const MachineMemOperand *> &Accesses) const {
378 size_t StartSize = Accesses.size();
379 for (MachineInstr::mmo_iterator o = MI.memoperands_begin(),
380 oe = MI.memoperands_end();
381 o != oe; ++o) {
382 if ((*o)->isStore() &&
383 dyn_cast_or_null<FixedStackPseudoSourceValue>((*o)->getPseudoValue()))
384 Accesses.push_back(*o);
386 return Accesses.size() != StartSize;
389 bool TargetInstrInfo::getStackSlotRange(const TargetRegisterClass *RC,
390 unsigned SubIdx, unsigned &Size,
391 unsigned &Offset,
392 const MachineFunction &MF) const {
393 const TargetRegisterInfo *TRI = MF.getSubtarget().getRegisterInfo();
394 if (!SubIdx) {
395 Size = TRI->getSpillSize(*RC);
396 Offset = 0;
397 return true;
399 unsigned BitSize = TRI->getSubRegIdxSize(SubIdx);
400 // Convert bit size to byte size.
401 if (BitSize % 8)
402 return false;
404 int BitOffset = TRI->getSubRegIdxOffset(SubIdx);
405 if (BitOffset < 0 || BitOffset % 8)
406 return false;
408 Size = BitSize / 8;
409 Offset = (unsigned)BitOffset / 8;
411 assert(TRI->getSpillSize(*RC) >= (Offset + Size) && "bad subregister range");
413 if (!MF.getDataLayout().isLittleEndian()) {
414 Offset = TRI->getSpillSize(*RC) - (Offset + Size);
416 return true;
419 void TargetInstrInfo::reMaterialize(MachineBasicBlock &MBB,
420 MachineBasicBlock::iterator I,
421 Register DestReg, unsigned SubIdx,
422 const MachineInstr &Orig,
423 const TargetRegisterInfo &TRI) const {
424 MachineInstr *MI = MBB.getParent()->CloneMachineInstr(&Orig);
425 MI->substituteRegister(MI->getOperand(0).getReg(), DestReg, SubIdx, TRI);
426 MBB.insert(I, MI);
429 bool TargetInstrInfo::produceSameValue(const MachineInstr &MI0,
430 const MachineInstr &MI1,
431 const MachineRegisterInfo *MRI) const {
432 return MI0.isIdenticalTo(MI1, MachineInstr::IgnoreVRegDefs);
435 MachineInstr &TargetInstrInfo::duplicate(MachineBasicBlock &MBB,
436 MachineBasicBlock::iterator InsertBefore, const MachineInstr &Orig) const {
437 assert(!Orig.isNotDuplicable() && "Instruction cannot be duplicated");
438 MachineFunction &MF = *MBB.getParent();
439 return MF.CloneMachineInstrBundle(MBB, InsertBefore, Orig);
442 // If the COPY instruction in MI can be folded to a stack operation, return
443 // the register class to use.
444 static const TargetRegisterClass *canFoldCopy(const MachineInstr &MI,
445 unsigned FoldIdx) {
446 assert(MI.isCopy() && "MI must be a COPY instruction");
447 if (MI.getNumOperands() != 2)
448 return nullptr;
449 assert(FoldIdx<2 && "FoldIdx refers no nonexistent operand");
451 const MachineOperand &FoldOp = MI.getOperand(FoldIdx);
452 const MachineOperand &LiveOp = MI.getOperand(1 - FoldIdx);
454 if (FoldOp.getSubReg() || LiveOp.getSubReg())
455 return nullptr;
457 Register FoldReg = FoldOp.getReg();
458 Register LiveReg = LiveOp.getReg();
460 assert(Register::isVirtualRegister(FoldReg) && "Cannot fold physregs");
462 const MachineRegisterInfo &MRI = MI.getMF()->getRegInfo();
463 const TargetRegisterClass *RC = MRI.getRegClass(FoldReg);
465 if (Register::isPhysicalRegister(LiveOp.getReg()))
466 return RC->contains(LiveOp.getReg()) ? RC : nullptr;
468 if (RC->hasSubClassEq(MRI.getRegClass(LiveReg)))
469 return RC;
471 // FIXME: Allow folding when register classes are memory compatible.
472 return nullptr;
475 MCInst TargetInstrInfo::getNop() const { llvm_unreachable("Not implemented"); }
477 std::pair<unsigned, unsigned>
478 TargetInstrInfo::getPatchpointUnfoldableRange(const MachineInstr &MI) const {
479 switch (MI.getOpcode()) {
480 case TargetOpcode::STACKMAP:
481 // StackMapLiveValues are foldable
482 return std::make_pair(0, StackMapOpers(&MI).getVarIdx());
483 case TargetOpcode::PATCHPOINT:
484 // For PatchPoint, the call args are not foldable (even if reported in the
485 // stackmap e.g. via anyregcc).
486 return std::make_pair(0, PatchPointOpers(&MI).getVarIdx());
487 case TargetOpcode::STATEPOINT:
488 // For statepoints, fold deopt and gc arguments, but not call arguments.
489 return std::make_pair(MI.getNumDefs(), StatepointOpers(&MI).getVarIdx());
490 default:
491 llvm_unreachable("unexpected stackmap opcode");
495 static MachineInstr *foldPatchpoint(MachineFunction &MF, MachineInstr &MI,
496 ArrayRef<unsigned> Ops, int FrameIndex,
497 const TargetInstrInfo &TII) {
498 unsigned StartIdx = 0;
499 unsigned NumDefs = 0;
500 // getPatchpointUnfoldableRange throws guarantee if MI is not a patchpoint.
501 std::tie(NumDefs, StartIdx) = TII.getPatchpointUnfoldableRange(MI);
503 unsigned DefToFoldIdx = MI.getNumOperands();
505 // Return false if any operands requested for folding are not foldable (not
506 // part of the stackmap's live values).
507 for (unsigned Op : Ops) {
508 if (Op < NumDefs) {
509 assert(DefToFoldIdx == MI.getNumOperands() && "Folding multiple defs");
510 DefToFoldIdx = Op;
511 } else if (Op < StartIdx) {
512 return nullptr;
514 if (MI.getOperand(Op).isTied())
515 return nullptr;
518 MachineInstr *NewMI =
519 MF.CreateMachineInstr(TII.get(MI.getOpcode()), MI.getDebugLoc(), true);
520 MachineInstrBuilder MIB(MF, NewMI);
522 // No need to fold return, the meta data, and function arguments
523 for (unsigned i = 0; i < StartIdx; ++i)
524 if (i != DefToFoldIdx)
525 MIB.add(MI.getOperand(i));
527 for (unsigned i = StartIdx, e = MI.getNumOperands(); i < e; ++i) {
528 MachineOperand &MO = MI.getOperand(i);
529 unsigned TiedTo = e;
530 (void)MI.isRegTiedToDefOperand(i, &TiedTo);
532 if (is_contained(Ops, i)) {
533 assert(TiedTo == e && "Cannot fold tied operands");
534 unsigned SpillSize;
535 unsigned SpillOffset;
536 // Compute the spill slot size and offset.
537 const TargetRegisterClass *RC =
538 MF.getRegInfo().getRegClass(MO.getReg());
539 bool Valid =
540 TII.getStackSlotRange(RC, MO.getSubReg(), SpillSize, SpillOffset, MF);
541 if (!Valid)
542 report_fatal_error("cannot spill patchpoint subregister operand");
543 MIB.addImm(StackMaps::IndirectMemRefOp);
544 MIB.addImm(SpillSize);
545 MIB.addFrameIndex(FrameIndex);
546 MIB.addImm(SpillOffset);
547 } else {
548 MIB.add(MO);
549 if (TiedTo < e) {
550 assert(TiedTo < NumDefs && "Bad tied operand");
551 if (TiedTo > DefToFoldIdx)
552 --TiedTo;
553 NewMI->tieOperands(TiedTo, NewMI->getNumOperands() - 1);
557 return NewMI;
560 MachineInstr *TargetInstrInfo::foldMemoryOperand(MachineInstr &MI,
561 ArrayRef<unsigned> Ops, int FI,
562 LiveIntervals *LIS,
563 VirtRegMap *VRM) const {
564 auto Flags = MachineMemOperand::MONone;
565 for (unsigned OpIdx : Ops)
566 Flags |= MI.getOperand(OpIdx).isDef() ? MachineMemOperand::MOStore
567 : MachineMemOperand::MOLoad;
569 MachineBasicBlock *MBB = MI.getParent();
570 assert(MBB && "foldMemoryOperand needs an inserted instruction");
571 MachineFunction &MF = *MBB->getParent();
573 // If we're not folding a load into a subreg, the size of the load is the
574 // size of the spill slot. But if we are, we need to figure out what the
575 // actual load size is.
576 int64_t MemSize = 0;
577 const MachineFrameInfo &MFI = MF.getFrameInfo();
578 const TargetRegisterInfo *TRI = MF.getSubtarget().getRegisterInfo();
580 if (Flags & MachineMemOperand::MOStore) {
581 MemSize = MFI.getObjectSize(FI);
582 } else {
583 for (unsigned OpIdx : Ops) {
584 int64_t OpSize = MFI.getObjectSize(FI);
586 if (auto SubReg = MI.getOperand(OpIdx).getSubReg()) {
587 unsigned SubRegSize = TRI->getSubRegIdxSize(SubReg);
588 if (SubRegSize > 0 && !(SubRegSize % 8))
589 OpSize = SubRegSize / 8;
592 MemSize = std::max(MemSize, OpSize);
596 assert(MemSize && "Did not expect a zero-sized stack slot");
598 MachineInstr *NewMI = nullptr;
600 if (MI.getOpcode() == TargetOpcode::STACKMAP ||
601 MI.getOpcode() == TargetOpcode::PATCHPOINT ||
602 MI.getOpcode() == TargetOpcode::STATEPOINT) {
603 // Fold stackmap/patchpoint.
604 NewMI = foldPatchpoint(MF, MI, Ops, FI, *this);
605 if (NewMI)
606 MBB->insert(MI, NewMI);
607 } else {
608 // Ask the target to do the actual folding.
609 NewMI = foldMemoryOperandImpl(MF, MI, Ops, MI, FI, LIS, VRM);
612 if (NewMI) {
613 NewMI->setMemRefs(MF, MI.memoperands());
614 // Add a memory operand, foldMemoryOperandImpl doesn't do that.
615 assert((!(Flags & MachineMemOperand::MOStore) ||
616 NewMI->mayStore()) &&
617 "Folded a def to a non-store!");
618 assert((!(Flags & MachineMemOperand::MOLoad) ||
619 NewMI->mayLoad()) &&
620 "Folded a use to a non-load!");
621 assert(MFI.getObjectOffset(FI) != -1);
622 MachineMemOperand *MMO =
623 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(MF, FI),
624 Flags, MemSize, MFI.getObjectAlign(FI));
625 NewMI->addMemOperand(MF, MMO);
627 // The pass "x86 speculative load hardening" always attaches symbols to
628 // call instructions. We need copy it form old instruction.
629 NewMI->cloneInstrSymbols(MF, MI);
631 return NewMI;
634 // Straight COPY may fold as load/store.
635 if (!MI.isCopy() || Ops.size() != 1)
636 return nullptr;
638 const TargetRegisterClass *RC = canFoldCopy(MI, Ops[0]);
639 if (!RC)
640 return nullptr;
642 const MachineOperand &MO = MI.getOperand(1 - Ops[0]);
643 MachineBasicBlock::iterator Pos = MI;
645 if (Flags == MachineMemOperand::MOStore)
646 storeRegToStackSlot(*MBB, Pos, MO.getReg(), MO.isKill(), FI, RC, TRI);
647 else
648 loadRegFromStackSlot(*MBB, Pos, MO.getReg(), FI, RC, TRI);
649 return &*--Pos;
652 MachineInstr *TargetInstrInfo::foldMemoryOperand(MachineInstr &MI,
653 ArrayRef<unsigned> Ops,
654 MachineInstr &LoadMI,
655 LiveIntervals *LIS) const {
656 assert(LoadMI.canFoldAsLoad() && "LoadMI isn't foldable!");
657 #ifndef NDEBUG
658 for (unsigned OpIdx : Ops)
659 assert(MI.getOperand(OpIdx).isUse() && "Folding load into def!");
660 #endif
662 MachineBasicBlock &MBB = *MI.getParent();
663 MachineFunction &MF = *MBB.getParent();
665 // Ask the target to do the actual folding.
666 MachineInstr *NewMI = nullptr;
667 int FrameIndex = 0;
669 if ((MI.getOpcode() == TargetOpcode::STACKMAP ||
670 MI.getOpcode() == TargetOpcode::PATCHPOINT ||
671 MI.getOpcode() == TargetOpcode::STATEPOINT) &&
672 isLoadFromStackSlot(LoadMI, FrameIndex)) {
673 // Fold stackmap/patchpoint.
674 NewMI = foldPatchpoint(MF, MI, Ops, FrameIndex, *this);
675 if (NewMI)
676 NewMI = &*MBB.insert(MI, NewMI);
677 } else {
678 // Ask the target to do the actual folding.
679 NewMI = foldMemoryOperandImpl(MF, MI, Ops, MI, LoadMI, LIS);
682 if (!NewMI)
683 return nullptr;
685 // Copy the memoperands from the load to the folded instruction.
686 if (MI.memoperands_empty()) {
687 NewMI->setMemRefs(MF, LoadMI.memoperands());
688 } else {
689 // Handle the rare case of folding multiple loads.
690 NewMI->setMemRefs(MF, MI.memoperands());
691 for (MachineInstr::mmo_iterator I = LoadMI.memoperands_begin(),
692 E = LoadMI.memoperands_end();
693 I != E; ++I) {
694 NewMI->addMemOperand(MF, *I);
697 return NewMI;
700 bool TargetInstrInfo::hasReassociableOperands(
701 const MachineInstr &Inst, const MachineBasicBlock *MBB) const {
702 const MachineOperand &Op1 = Inst.getOperand(1);
703 const MachineOperand &Op2 = Inst.getOperand(2);
704 const MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo();
706 // We need virtual register definitions for the operands that we will
707 // reassociate.
708 MachineInstr *MI1 = nullptr;
709 MachineInstr *MI2 = nullptr;
710 if (Op1.isReg() && Register::isVirtualRegister(Op1.getReg()))
711 MI1 = MRI.getUniqueVRegDef(Op1.getReg());
712 if (Op2.isReg() && Register::isVirtualRegister(Op2.getReg()))
713 MI2 = MRI.getUniqueVRegDef(Op2.getReg());
715 // And they need to be in the trace (otherwise, they won't have a depth).
716 return MI1 && MI2 && MI1->getParent() == MBB && MI2->getParent() == MBB;
719 bool TargetInstrInfo::hasReassociableSibling(const MachineInstr &Inst,
720 bool &Commuted) const {
721 const MachineBasicBlock *MBB = Inst.getParent();
722 const MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo();
723 MachineInstr *MI1 = MRI.getUniqueVRegDef(Inst.getOperand(1).getReg());
724 MachineInstr *MI2 = MRI.getUniqueVRegDef(Inst.getOperand(2).getReg());
725 unsigned AssocOpcode = Inst.getOpcode();
727 // If only one operand has the same opcode and it's the second source operand,
728 // the operands must be commuted.
729 Commuted = MI1->getOpcode() != AssocOpcode && MI2->getOpcode() == AssocOpcode;
730 if (Commuted)
731 std::swap(MI1, MI2);
733 // 1. The previous instruction must be the same type as Inst.
734 // 2. The previous instruction must also be associative/commutative (this can
735 // be different even for instructions with the same opcode if traits like
736 // fast-math-flags are included).
737 // 3. The previous instruction must have virtual register definitions for its
738 // operands in the same basic block as Inst.
739 // 4. The previous instruction's result must only be used by Inst.
740 return MI1->getOpcode() == AssocOpcode && isAssociativeAndCommutative(*MI1) &&
741 hasReassociableOperands(*MI1, MBB) &&
742 MRI.hasOneNonDBGUse(MI1->getOperand(0).getReg());
745 // 1. The operation must be associative and commutative.
746 // 2. The instruction must have virtual register definitions for its
747 // operands in the same basic block.
748 // 3. The instruction must have a reassociable sibling.
749 bool TargetInstrInfo::isReassociationCandidate(const MachineInstr &Inst,
750 bool &Commuted) const {
751 return isAssociativeAndCommutative(Inst) &&
752 hasReassociableOperands(Inst, Inst.getParent()) &&
753 hasReassociableSibling(Inst, Commuted);
756 // The concept of the reassociation pass is that these operations can benefit
757 // from this kind of transformation:
759 // A = ? op ?
760 // B = A op X (Prev)
761 // C = B op Y (Root)
762 // -->
763 // A = ? op ?
764 // B = X op Y
765 // C = A op B
767 // breaking the dependency between A and B, allowing them to be executed in
768 // parallel (or back-to-back in a pipeline) instead of depending on each other.
770 // FIXME: This has the potential to be expensive (compile time) while not
771 // improving the code at all. Some ways to limit the overhead:
772 // 1. Track successful transforms; bail out if hit rate gets too low.
773 // 2. Only enable at -O3 or some other non-default optimization level.
774 // 3. Pre-screen pattern candidates here: if an operand of the previous
775 // instruction is known to not increase the critical path, then don't match
776 // that pattern.
777 bool TargetInstrInfo::getMachineCombinerPatterns(
778 MachineInstr &Root, SmallVectorImpl<MachineCombinerPattern> &Patterns,
779 bool DoRegPressureReduce) const {
780 bool Commute;
781 if (isReassociationCandidate(Root, Commute)) {
782 // We found a sequence of instructions that may be suitable for a
783 // reassociation of operands to increase ILP. Specify each commutation
784 // possibility for the Prev instruction in the sequence and let the
785 // machine combiner decide if changing the operands is worthwhile.
786 if (Commute) {
787 Patterns.push_back(MachineCombinerPattern::REASSOC_AX_YB);
788 Patterns.push_back(MachineCombinerPattern::REASSOC_XA_YB);
789 } else {
790 Patterns.push_back(MachineCombinerPattern::REASSOC_AX_BY);
791 Patterns.push_back(MachineCombinerPattern::REASSOC_XA_BY);
793 return true;
796 return false;
799 /// Return true when a code sequence can improve loop throughput.
800 bool
801 TargetInstrInfo::isThroughputPattern(MachineCombinerPattern Pattern) const {
802 return false;
805 /// Attempt the reassociation transformation to reduce critical path length.
806 /// See the above comments before getMachineCombinerPatterns().
807 void TargetInstrInfo::reassociateOps(
808 MachineInstr &Root, MachineInstr &Prev,
809 MachineCombinerPattern Pattern,
810 SmallVectorImpl<MachineInstr *> &InsInstrs,
811 SmallVectorImpl<MachineInstr *> &DelInstrs,
812 DenseMap<unsigned, unsigned> &InstrIdxForVirtReg) const {
813 MachineFunction *MF = Root.getMF();
814 MachineRegisterInfo &MRI = MF->getRegInfo();
815 const TargetInstrInfo *TII = MF->getSubtarget().getInstrInfo();
816 const TargetRegisterInfo *TRI = MF->getSubtarget().getRegisterInfo();
817 const TargetRegisterClass *RC = Root.getRegClassConstraint(0, TII, TRI);
819 // This array encodes the operand index for each parameter because the
820 // operands may be commuted. Each row corresponds to a pattern value,
821 // and each column specifies the index of A, B, X, Y.
822 unsigned OpIdx[4][4] = {
823 { 1, 1, 2, 2 },
824 { 1, 2, 2, 1 },
825 { 2, 1, 1, 2 },
826 { 2, 2, 1, 1 }
829 int Row;
830 switch (Pattern) {
831 case MachineCombinerPattern::REASSOC_AX_BY: Row = 0; break;
832 case MachineCombinerPattern::REASSOC_AX_YB: Row = 1; break;
833 case MachineCombinerPattern::REASSOC_XA_BY: Row = 2; break;
834 case MachineCombinerPattern::REASSOC_XA_YB: Row = 3; break;
835 default: llvm_unreachable("unexpected MachineCombinerPattern");
838 MachineOperand &OpA = Prev.getOperand(OpIdx[Row][0]);
839 MachineOperand &OpB = Root.getOperand(OpIdx[Row][1]);
840 MachineOperand &OpX = Prev.getOperand(OpIdx[Row][2]);
841 MachineOperand &OpY = Root.getOperand(OpIdx[Row][3]);
842 MachineOperand &OpC = Root.getOperand(0);
844 Register RegA = OpA.getReg();
845 Register RegB = OpB.getReg();
846 Register RegX = OpX.getReg();
847 Register RegY = OpY.getReg();
848 Register RegC = OpC.getReg();
850 if (Register::isVirtualRegister(RegA))
851 MRI.constrainRegClass(RegA, RC);
852 if (Register::isVirtualRegister(RegB))
853 MRI.constrainRegClass(RegB, RC);
854 if (Register::isVirtualRegister(RegX))
855 MRI.constrainRegClass(RegX, RC);
856 if (Register::isVirtualRegister(RegY))
857 MRI.constrainRegClass(RegY, RC);
858 if (Register::isVirtualRegister(RegC))
859 MRI.constrainRegClass(RegC, RC);
861 // Create a new virtual register for the result of (X op Y) instead of
862 // recycling RegB because the MachineCombiner's computation of the critical
863 // path requires a new register definition rather than an existing one.
864 Register NewVR = MRI.createVirtualRegister(RC);
865 InstrIdxForVirtReg.insert(std::make_pair(NewVR, 0));
867 unsigned Opcode = Root.getOpcode();
868 bool KillA = OpA.isKill();
869 bool KillX = OpX.isKill();
870 bool KillY = OpY.isKill();
872 // Create new instructions for insertion.
873 MachineInstrBuilder MIB1 =
874 BuildMI(*MF, Prev.getDebugLoc(), TII->get(Opcode), NewVR)
875 .addReg(RegX, getKillRegState(KillX))
876 .addReg(RegY, getKillRegState(KillY));
877 MachineInstrBuilder MIB2 =
878 BuildMI(*MF, Root.getDebugLoc(), TII->get(Opcode), RegC)
879 .addReg(RegA, getKillRegState(KillA))
880 .addReg(NewVR, getKillRegState(true));
882 setSpecialOperandAttr(Root, Prev, *MIB1, *MIB2);
884 // Record new instructions for insertion and old instructions for deletion.
885 InsInstrs.push_back(MIB1);
886 InsInstrs.push_back(MIB2);
887 DelInstrs.push_back(&Prev);
888 DelInstrs.push_back(&Root);
891 void TargetInstrInfo::genAlternativeCodeSequence(
892 MachineInstr &Root, MachineCombinerPattern Pattern,
893 SmallVectorImpl<MachineInstr *> &InsInstrs,
894 SmallVectorImpl<MachineInstr *> &DelInstrs,
895 DenseMap<unsigned, unsigned> &InstIdxForVirtReg) const {
896 MachineRegisterInfo &MRI = Root.getMF()->getRegInfo();
898 // Select the previous instruction in the sequence based on the input pattern.
899 MachineInstr *Prev = nullptr;
900 switch (Pattern) {
901 case MachineCombinerPattern::REASSOC_AX_BY:
902 case MachineCombinerPattern::REASSOC_XA_BY:
903 Prev = MRI.getUniqueVRegDef(Root.getOperand(1).getReg());
904 break;
905 case MachineCombinerPattern::REASSOC_AX_YB:
906 case MachineCombinerPattern::REASSOC_XA_YB:
907 Prev = MRI.getUniqueVRegDef(Root.getOperand(2).getReg());
908 break;
909 default:
910 break;
913 assert(Prev && "Unknown pattern for machine combiner");
915 reassociateOps(Root, *Prev, Pattern, InsInstrs, DelInstrs, InstIdxForVirtReg);
918 bool TargetInstrInfo::isReallyTriviallyReMaterializableGeneric(
919 const MachineInstr &MI, AAResults *AA) const {
920 const MachineFunction &MF = *MI.getMF();
921 const MachineRegisterInfo &MRI = MF.getRegInfo();
923 // Remat clients assume operand 0 is the defined register.
924 if (!MI.getNumOperands() || !MI.getOperand(0).isReg())
925 return false;
926 Register DefReg = MI.getOperand(0).getReg();
928 // A sub-register definition can only be rematerialized if the instruction
929 // doesn't read the other parts of the register. Otherwise it is really a
930 // read-modify-write operation on the full virtual register which cannot be
931 // moved safely.
932 if (Register::isVirtualRegister(DefReg) && MI.getOperand(0).getSubReg() &&
933 MI.readsVirtualRegister(DefReg))
934 return false;
936 // A load from a fixed stack slot can be rematerialized. This may be
937 // redundant with subsequent checks, but it's target-independent,
938 // simple, and a common case.
939 int FrameIdx = 0;
940 if (isLoadFromStackSlot(MI, FrameIdx) &&
941 MF.getFrameInfo().isImmutableObjectIndex(FrameIdx))
942 return true;
944 // Avoid instructions obviously unsafe for remat.
945 if (MI.isNotDuplicable() || MI.mayStore() || MI.mayRaiseFPException() ||
946 MI.hasUnmodeledSideEffects())
947 return false;
949 // Don't remat inline asm. We have no idea how expensive it is
950 // even if it's side effect free.
951 if (MI.isInlineAsm())
952 return false;
954 // Avoid instructions which load from potentially varying memory.
955 if (MI.mayLoad() && !MI.isDereferenceableInvariantLoad(AA))
956 return false;
958 // If any of the registers accessed are non-constant, conservatively assume
959 // the instruction is not rematerializable.
960 for (unsigned i = 0, e = MI.getNumOperands(); i != e; ++i) {
961 const MachineOperand &MO = MI.getOperand(i);
962 if (!MO.isReg()) continue;
963 Register Reg = MO.getReg();
964 if (Reg == 0)
965 continue;
967 // Check for a well-behaved physical register.
968 if (Register::isPhysicalRegister(Reg)) {
969 if (MO.isUse()) {
970 // If the physreg has no defs anywhere, it's just an ambient register
971 // and we can freely move its uses. Alternatively, if it's allocatable,
972 // it could get allocated to something with a def during allocation.
973 if (!MRI.isConstantPhysReg(Reg))
974 return false;
975 } else {
976 // A physreg def. We can't remat it.
977 return false;
979 continue;
982 // Only allow one virtual-register def. There may be multiple defs of the
983 // same virtual register, though.
984 if (MO.isDef() && Reg != DefReg)
985 return false;
987 // Don't allow any virtual-register uses. Rematting an instruction with
988 // virtual register uses would length the live ranges of the uses, which
989 // is not necessarily a good idea, certainly not "trivial".
990 if (MO.isUse())
991 return false;
994 // Everything checked out.
995 return true;
998 int TargetInstrInfo::getSPAdjust(const MachineInstr &MI) const {
999 const MachineFunction *MF = MI.getMF();
1000 const TargetFrameLowering *TFI = MF->getSubtarget().getFrameLowering();
1001 bool StackGrowsDown =
1002 TFI->getStackGrowthDirection() == TargetFrameLowering::StackGrowsDown;
1004 unsigned FrameSetupOpcode = getCallFrameSetupOpcode();
1005 unsigned FrameDestroyOpcode = getCallFrameDestroyOpcode();
1007 if (!isFrameInstr(MI))
1008 return 0;
1010 int SPAdj = TFI->alignSPAdjust(getFrameSize(MI));
1012 if ((!StackGrowsDown && MI.getOpcode() == FrameSetupOpcode) ||
1013 (StackGrowsDown && MI.getOpcode() == FrameDestroyOpcode))
1014 SPAdj = -SPAdj;
1016 return SPAdj;
1019 /// isSchedulingBoundary - Test if the given instruction should be
1020 /// considered a scheduling boundary. This primarily includes labels
1021 /// and terminators.
1022 bool TargetInstrInfo::isSchedulingBoundary(const MachineInstr &MI,
1023 const MachineBasicBlock *MBB,
1024 const MachineFunction &MF) const {
1025 // Terminators and labels can't be scheduled around.
1026 if (MI.isTerminator() || MI.isPosition())
1027 return true;
1029 // INLINEASM_BR can jump to another block
1030 if (MI.getOpcode() == TargetOpcode::INLINEASM_BR)
1031 return true;
1033 // Don't attempt to schedule around any instruction that defines
1034 // a stack-oriented pointer, as it's unlikely to be profitable. This
1035 // saves compile time, because it doesn't require every single
1036 // stack slot reference to depend on the instruction that does the
1037 // modification.
1038 const TargetLowering &TLI = *MF.getSubtarget().getTargetLowering();
1039 const TargetRegisterInfo *TRI = MF.getSubtarget().getRegisterInfo();
1040 return MI.modifiesRegister(TLI.getStackPointerRegisterToSaveRestore(), TRI);
1043 // Provide a global flag for disabling the PreRA hazard recognizer that targets
1044 // may choose to honor.
1045 bool TargetInstrInfo::usePreRAHazardRecognizer() const {
1046 return !DisableHazardRecognizer;
1049 // Default implementation of CreateTargetRAHazardRecognizer.
1050 ScheduleHazardRecognizer *TargetInstrInfo::
1051 CreateTargetHazardRecognizer(const TargetSubtargetInfo *STI,
1052 const ScheduleDAG *DAG) const {
1053 // Dummy hazard recognizer allows all instructions to issue.
1054 return new ScheduleHazardRecognizer();
1057 // Default implementation of CreateTargetMIHazardRecognizer.
1058 ScheduleHazardRecognizer *TargetInstrInfo::CreateTargetMIHazardRecognizer(
1059 const InstrItineraryData *II, const ScheduleDAGMI *DAG) const {
1060 return new ScoreboardHazardRecognizer(II, DAG, "machine-scheduler");
1063 // Default implementation of CreateTargetPostRAHazardRecognizer.
1064 ScheduleHazardRecognizer *TargetInstrInfo::
1065 CreateTargetPostRAHazardRecognizer(const InstrItineraryData *II,
1066 const ScheduleDAG *DAG) const {
1067 return new ScoreboardHazardRecognizer(II, DAG, "post-RA-sched");
1070 // Default implementation of getMemOperandWithOffset.
1071 bool TargetInstrInfo::getMemOperandWithOffset(
1072 const MachineInstr &MI, const MachineOperand *&BaseOp, int64_t &Offset,
1073 bool &OffsetIsScalable, const TargetRegisterInfo *TRI) const {
1074 SmallVector<const MachineOperand *, 4> BaseOps;
1075 unsigned Width;
1076 if (!getMemOperandsWithOffsetWidth(MI, BaseOps, Offset, OffsetIsScalable,
1077 Width, TRI) ||
1078 BaseOps.size() != 1)
1079 return false;
1080 BaseOp = BaseOps.front();
1081 return true;
1084 //===----------------------------------------------------------------------===//
1085 // SelectionDAG latency interface.
1086 //===----------------------------------------------------------------------===//
1089 TargetInstrInfo::getOperandLatency(const InstrItineraryData *ItinData,
1090 SDNode *DefNode, unsigned DefIdx,
1091 SDNode *UseNode, unsigned UseIdx) const {
1092 if (!ItinData || ItinData->isEmpty())
1093 return -1;
1095 if (!DefNode->isMachineOpcode())
1096 return -1;
1098 unsigned DefClass = get(DefNode->getMachineOpcode()).getSchedClass();
1099 if (!UseNode->isMachineOpcode())
1100 return ItinData->getOperandCycle(DefClass, DefIdx);
1101 unsigned UseClass = get(UseNode->getMachineOpcode()).getSchedClass();
1102 return ItinData->getOperandLatency(DefClass, DefIdx, UseClass, UseIdx);
1105 int TargetInstrInfo::getInstrLatency(const InstrItineraryData *ItinData,
1106 SDNode *N) const {
1107 if (!ItinData || ItinData->isEmpty())
1108 return 1;
1110 if (!N->isMachineOpcode())
1111 return 1;
1113 return ItinData->getStageLatency(get(N->getMachineOpcode()).getSchedClass());
1116 //===----------------------------------------------------------------------===//
1117 // MachineInstr latency interface.
1118 //===----------------------------------------------------------------------===//
1120 unsigned TargetInstrInfo::getNumMicroOps(const InstrItineraryData *ItinData,
1121 const MachineInstr &MI) const {
1122 if (!ItinData || ItinData->isEmpty())
1123 return 1;
1125 unsigned Class = MI.getDesc().getSchedClass();
1126 int UOps = ItinData->Itineraries[Class].NumMicroOps;
1127 if (UOps >= 0)
1128 return UOps;
1130 // The # of u-ops is dynamically determined. The specific target should
1131 // override this function to return the right number.
1132 return 1;
1135 /// Return the default expected latency for a def based on it's opcode.
1136 unsigned TargetInstrInfo::defaultDefLatency(const MCSchedModel &SchedModel,
1137 const MachineInstr &DefMI) const {
1138 if (DefMI.isTransient())
1139 return 0;
1140 if (DefMI.mayLoad())
1141 return SchedModel.LoadLatency;
1142 if (isHighLatencyDef(DefMI.getOpcode()))
1143 return SchedModel.HighLatency;
1144 return 1;
1147 unsigned TargetInstrInfo::getPredicationCost(const MachineInstr &) const {
1148 return 0;
1151 unsigned TargetInstrInfo::getInstrLatency(const InstrItineraryData *ItinData,
1152 const MachineInstr &MI,
1153 unsigned *PredCost) const {
1154 // Default to one cycle for no itinerary. However, an "empty" itinerary may
1155 // still have a MinLatency property, which getStageLatency checks.
1156 if (!ItinData)
1157 return MI.mayLoad() ? 2 : 1;
1159 return ItinData->getStageLatency(MI.getDesc().getSchedClass());
1162 bool TargetInstrInfo::hasLowDefLatency(const TargetSchedModel &SchedModel,
1163 const MachineInstr &DefMI,
1164 unsigned DefIdx) const {
1165 const InstrItineraryData *ItinData = SchedModel.getInstrItineraries();
1166 if (!ItinData || ItinData->isEmpty())
1167 return false;
1169 unsigned DefClass = DefMI.getDesc().getSchedClass();
1170 int DefCycle = ItinData->getOperandCycle(DefClass, DefIdx);
1171 return (DefCycle != -1 && DefCycle <= 1);
1174 Optional<ParamLoadedValue>
1175 TargetInstrInfo::describeLoadedValue(const MachineInstr &MI,
1176 Register Reg) const {
1177 const MachineFunction *MF = MI.getMF();
1178 const TargetRegisterInfo *TRI = MF->getSubtarget().getRegisterInfo();
1179 DIExpression *Expr = DIExpression::get(MF->getFunction().getContext(), {});
1180 int64_t Offset;
1181 bool OffsetIsScalable;
1183 // To simplify the sub-register handling, verify that we only need to
1184 // consider physical registers.
1185 assert(MF->getProperties().hasProperty(
1186 MachineFunctionProperties::Property::NoVRegs));
1188 if (auto DestSrc = isCopyInstr(MI)) {
1189 Register DestReg = DestSrc->Destination->getReg();
1191 // If the copy destination is the forwarding reg, describe the forwarding
1192 // reg using the copy source as the backup location. Example:
1194 // x0 = MOV x7
1195 // call callee(x0) ; x0 described as x7
1196 if (Reg == DestReg)
1197 return ParamLoadedValue(*DestSrc->Source, Expr);
1199 // Cases where super- or sub-registers needs to be described should
1200 // be handled by the target's hook implementation.
1201 assert(!TRI->isSuperOrSubRegisterEq(Reg, DestReg) &&
1202 "TargetInstrInfo::describeLoadedValue can't describe super- or "
1203 "sub-regs for copy instructions");
1204 return None;
1205 } else if (auto RegImm = isAddImmediate(MI, Reg)) {
1206 Register SrcReg = RegImm->Reg;
1207 Offset = RegImm->Imm;
1208 Expr = DIExpression::prepend(Expr, DIExpression::ApplyOffset, Offset);
1209 return ParamLoadedValue(MachineOperand::CreateReg(SrcReg, false), Expr);
1210 } else if (MI.hasOneMemOperand()) {
1211 // Only describe memory which provably does not escape the function. As
1212 // described in llvm.org/PR43343, escaped memory may be clobbered by the
1213 // callee (or by another thread).
1214 const auto &TII = MF->getSubtarget().getInstrInfo();
1215 const MachineFrameInfo &MFI = MF->getFrameInfo();
1216 const MachineMemOperand *MMO = MI.memoperands()[0];
1217 const PseudoSourceValue *PSV = MMO->getPseudoValue();
1219 // If the address points to "special" memory (e.g. a spill slot), it's
1220 // sufficient to check that it isn't aliased by any high-level IR value.
1221 if (!PSV || PSV->mayAlias(&MFI))
1222 return None;
1224 const MachineOperand *BaseOp;
1225 if (!TII->getMemOperandWithOffset(MI, BaseOp, Offset, OffsetIsScalable,
1226 TRI))
1227 return None;
1229 // FIXME: Scalable offsets are not yet handled in the offset code below.
1230 if (OffsetIsScalable)
1231 return None;
1233 // TODO: Can currently only handle mem instructions with a single define.
1234 // An example from the x86 target:
1235 // ...
1236 // DIV64m $rsp, 1, $noreg, 24, $noreg, implicit-def dead $rax, implicit-def $rdx
1237 // ...
1239 if (MI.getNumExplicitDefs() != 1)
1240 return None;
1242 // TODO: In what way do we need to take Reg into consideration here?
1244 SmallVector<uint64_t, 8> Ops;
1245 DIExpression::appendOffset(Ops, Offset);
1246 Ops.push_back(dwarf::DW_OP_deref_size);
1247 Ops.push_back(MMO->getSize());
1248 Expr = DIExpression::prependOpcodes(Expr, Ops);
1249 return ParamLoadedValue(*BaseOp, Expr);
1252 return None;
1255 /// Both DefMI and UseMI must be valid. By default, call directly to the
1256 /// itinerary. This may be overriden by the target.
1257 int TargetInstrInfo::getOperandLatency(const InstrItineraryData *ItinData,
1258 const MachineInstr &DefMI,
1259 unsigned DefIdx,
1260 const MachineInstr &UseMI,
1261 unsigned UseIdx) const {
1262 unsigned DefClass = DefMI.getDesc().getSchedClass();
1263 unsigned UseClass = UseMI.getDesc().getSchedClass();
1264 return ItinData->getOperandLatency(DefClass, DefIdx, UseClass, UseIdx);
1267 bool TargetInstrInfo::getRegSequenceInputs(
1268 const MachineInstr &MI, unsigned DefIdx,
1269 SmallVectorImpl<RegSubRegPairAndIdx> &InputRegs) const {
1270 assert((MI.isRegSequence() ||
1271 MI.isRegSequenceLike()) && "Instruction do not have the proper type");
1273 if (!MI.isRegSequence())
1274 return getRegSequenceLikeInputs(MI, DefIdx, InputRegs);
1276 // We are looking at:
1277 // Def = REG_SEQUENCE v0, sub0, v1, sub1, ...
1278 assert(DefIdx == 0 && "REG_SEQUENCE only has one def");
1279 for (unsigned OpIdx = 1, EndOpIdx = MI.getNumOperands(); OpIdx != EndOpIdx;
1280 OpIdx += 2) {
1281 const MachineOperand &MOReg = MI.getOperand(OpIdx);
1282 if (MOReg.isUndef())
1283 continue;
1284 const MachineOperand &MOSubIdx = MI.getOperand(OpIdx + 1);
1285 assert(MOSubIdx.isImm() &&
1286 "One of the subindex of the reg_sequence is not an immediate");
1287 // Record Reg:SubReg, SubIdx.
1288 InputRegs.push_back(RegSubRegPairAndIdx(MOReg.getReg(), MOReg.getSubReg(),
1289 (unsigned)MOSubIdx.getImm()));
1291 return true;
1294 bool TargetInstrInfo::getExtractSubregInputs(
1295 const MachineInstr &MI, unsigned DefIdx,
1296 RegSubRegPairAndIdx &InputReg) const {
1297 assert((MI.isExtractSubreg() ||
1298 MI.isExtractSubregLike()) && "Instruction do not have the proper type");
1300 if (!MI.isExtractSubreg())
1301 return getExtractSubregLikeInputs(MI, DefIdx, InputReg);
1303 // We are looking at:
1304 // Def = EXTRACT_SUBREG v0.sub1, sub0.
1305 assert(DefIdx == 0 && "EXTRACT_SUBREG only has one def");
1306 const MachineOperand &MOReg = MI.getOperand(1);
1307 if (MOReg.isUndef())
1308 return false;
1309 const MachineOperand &MOSubIdx = MI.getOperand(2);
1310 assert(MOSubIdx.isImm() &&
1311 "The subindex of the extract_subreg is not an immediate");
1313 InputReg.Reg = MOReg.getReg();
1314 InputReg.SubReg = MOReg.getSubReg();
1315 InputReg.SubIdx = (unsigned)MOSubIdx.getImm();
1316 return true;
1319 bool TargetInstrInfo::getInsertSubregInputs(
1320 const MachineInstr &MI, unsigned DefIdx,
1321 RegSubRegPair &BaseReg, RegSubRegPairAndIdx &InsertedReg) const {
1322 assert((MI.isInsertSubreg() ||
1323 MI.isInsertSubregLike()) && "Instruction do not have the proper type");
1325 if (!MI.isInsertSubreg())
1326 return getInsertSubregLikeInputs(MI, DefIdx, BaseReg, InsertedReg);
1328 // We are looking at:
1329 // Def = INSERT_SEQUENCE v0, v1, sub0.
1330 assert(DefIdx == 0 && "INSERT_SUBREG only has one def");
1331 const MachineOperand &MOBaseReg = MI.getOperand(1);
1332 const MachineOperand &MOInsertedReg = MI.getOperand(2);
1333 if (MOInsertedReg.isUndef())
1334 return false;
1335 const MachineOperand &MOSubIdx = MI.getOperand(3);
1336 assert(MOSubIdx.isImm() &&
1337 "One of the subindex of the reg_sequence is not an immediate");
1338 BaseReg.Reg = MOBaseReg.getReg();
1339 BaseReg.SubReg = MOBaseReg.getSubReg();
1341 InsertedReg.Reg = MOInsertedReg.getReg();
1342 InsertedReg.SubReg = MOInsertedReg.getSubReg();
1343 InsertedReg.SubIdx = (unsigned)MOSubIdx.getImm();
1344 return true;
1347 // Returns a MIRPrinter comment for this machine operand.
1348 std::string TargetInstrInfo::createMIROperandComment(
1349 const MachineInstr &MI, const MachineOperand &Op, unsigned OpIdx,
1350 const TargetRegisterInfo *TRI) const {
1352 if (!MI.isInlineAsm())
1353 return "";
1355 std::string Flags;
1356 raw_string_ostream OS(Flags);
1358 if (OpIdx == InlineAsm::MIOp_ExtraInfo) {
1359 // Print HasSideEffects, MayLoad, MayStore, IsAlignStack
1360 unsigned ExtraInfo = Op.getImm();
1361 bool First = true;
1362 for (StringRef Info : InlineAsm::getExtraInfoNames(ExtraInfo)) {
1363 if (!First)
1364 OS << " ";
1365 First = false;
1366 OS << Info;
1369 return OS.str();
1372 int FlagIdx = MI.findInlineAsmFlagIdx(OpIdx);
1373 if (FlagIdx < 0 || (unsigned)FlagIdx != OpIdx)
1374 return "";
1376 assert(Op.isImm() && "Expected flag operand to be an immediate");
1377 // Pretty print the inline asm operand descriptor.
1378 unsigned Flag = Op.getImm();
1379 unsigned Kind = InlineAsm::getKind(Flag);
1380 OS << InlineAsm::getKindName(Kind);
1382 unsigned RCID = 0;
1383 if (!InlineAsm::isImmKind(Flag) && !InlineAsm::isMemKind(Flag) &&
1384 InlineAsm::hasRegClassConstraint(Flag, RCID)) {
1385 if (TRI) {
1386 OS << ':' << TRI->getRegClassName(TRI->getRegClass(RCID));
1387 } else
1388 OS << ":RC" << RCID;
1391 if (InlineAsm::isMemKind(Flag)) {
1392 unsigned MCID = InlineAsm::getMemoryConstraintID(Flag);
1393 OS << ":" << InlineAsm::getMemConstraintName(MCID);
1396 unsigned TiedTo = 0;
1397 if (InlineAsm::isUseOperandTiedToDef(Flag, TiedTo))
1398 OS << " tiedto:$" << TiedTo;
1400 return OS.str();
1403 TargetInstrInfo::PipelinerLoopInfo::~PipelinerLoopInfo() {}