1 //===-- TargetInstrInfo.cpp - Target Instruction Information --------------===//
2 //
3 // The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 // This file implements the TargetInstrInfo class.
11 //
12 //===----------------------------------------------------------------------===//
13
14 #include "llvm/Target/TargetInstrInfo.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/PseudoSourceValue.h"
20 #include "llvm/CodeGen/ScoreboardHazardRecognizer.h"
21 #include "llvm/CodeGen/StackMaps.h"
22 #include "llvm/CodeGen/TargetSchedule.h"
23 #include "llvm/IR/DataLayout.h"
24 #include "llvm/MC/MCAsmInfo.h"
25 #include "llvm/MC/MCInstrItineraries.h"
26 #include "llvm/Support/CommandLine.h"
27 #include "llvm/Support/ErrorHandling.h"
28 #include "llvm/Support/raw_ostream.h"
29 #include "llvm/Target/TargetFrameLowering.h"
30 #include "llvm/Target/TargetLowering.h"
31 #include "llvm/Target/TargetMachine.h"
32 #include "llvm/Target/TargetRegisterInfo.h"
33 #include <cctype>
34 using namespace llvm;
35
36 static cl::opt<bool> DisableHazardRecognizer(
37 "disable-sched-hazard", cl::Hidden, cl::init(false),
38 cl::desc("Disable hazard detection during preRA scheduling"));
39
~TargetInstrInfo()40 TargetInstrInfo::~TargetInstrInfo() {
41 }
42
43 const TargetRegisterClass*
getRegClass(const MCInstrDesc & MCID,unsigned OpNum,const TargetRegisterInfo * TRI,const MachineFunction & MF) const44 TargetInstrInfo::getRegClass(const MCInstrDesc &MCID, unsigned OpNum,
45 const TargetRegisterInfo *TRI,
46 const MachineFunction &MF) const {
47 if (OpNum >= MCID.getNumOperands())
48 return nullptr;
49
50 short RegClass = MCID.OpInfo[OpNum].RegClass;
51 if (MCID.OpInfo[OpNum].isLookupPtrRegClass())
52 return TRI->getPointerRegClass(MF, RegClass);
53
54 // Instructions like INSERT_SUBREG do not have fixed register classes.
55 if (RegClass < 0)
56 return nullptr;
57
58 // Otherwise just look it up normally.
59 return TRI->getRegClass(RegClass);
60 }
61
62 /// insertNoop - Insert a noop into the instruction stream at the specified
63 /// point.
insertNoop(MachineBasicBlock & MBB,MachineBasicBlock::iterator MI) const64 void TargetInstrInfo::insertNoop(MachineBasicBlock &MBB,
65 MachineBasicBlock::iterator MI) const {
66 llvm_unreachable("Target didn't implement insertNoop!");
67 }
68
69 /// Measure the specified inline asm to determine an approximation of its
70 /// length.
71 /// Comments (which run till the next SeparatorString or newline) do not
72 /// count as an instruction.
73 /// Any other non-whitespace text is considered an instruction, with
74 /// multiple instructions separated by SeparatorString or newlines.
75 /// Variable-length instructions are not handled here; this function
76 /// may be overloaded in the target code to do that.
getInlineAsmLength(const char * Str,const MCAsmInfo & MAI) const77 unsigned TargetInstrInfo::getInlineAsmLength(const char *Str,
78 const MCAsmInfo &MAI) const {
79
80
81 // Count the number of instructions in the asm.
82 bool atInsnStart = true;
83 unsigned Length = 0;
84 for (; *Str; ++Str) {
85 if (*Str == '\n' || strncmp(Str, MAI.getSeparatorString(),
86 strlen(MAI.getSeparatorString())) == 0)
87 atInsnStart = true;
88 if (atInsnStart && !std::isspace(static_cast<unsigned char>(*Str))) {
89 Length += MAI.getMaxInstLength();
90 atInsnStart = false;
91 }
92 if (atInsnStart && strncmp(Str, MAI.getCommentString(),
93 strlen(MAI.getCommentString())) == 0)
94 atInsnStart = false;
95 }
96
97 return Length;
98 }
99
100 /// ReplaceTailWithBranchTo - Delete the instruction OldInst and everything
101 /// after it, replacing it with an unconditional branch to NewDest.
102 void
ReplaceTailWithBranchTo(MachineBasicBlock::iterator Tail,MachineBasicBlock * NewDest) const103 TargetInstrInfo::ReplaceTailWithBranchTo(MachineBasicBlock::iterator Tail,
104 MachineBasicBlock *NewDest) const {
105 MachineBasicBlock *MBB = Tail->getParent();
106
107 // Remove all the old successors of MBB from the CFG.
108 while (!MBB->succ_empty())
109 MBB->removeSuccessor(MBB->succ_begin());
110
111 // Remove all the dead instructions from the end of MBB.
112 MBB->erase(Tail, MBB->end());
113
114 // If MBB isn't immediately before MBB, insert a branch to it.
115 if (++MachineFunction::iterator(MBB) != MachineFunction::iterator(NewDest))
116 InsertBranch(*MBB, NewDest, nullptr, SmallVector<MachineOperand, 0>(),
117 Tail->getDebugLoc());
118 MBB->addSuccessor(NewDest);
119 }
120
commuteInstructionImpl(MachineInstr * MI,bool NewMI,unsigned Idx1,unsigned Idx2) const121 MachineInstr *TargetInstrInfo::commuteInstructionImpl(MachineInstr *MI,
122 bool NewMI,
123 unsigned Idx1,
124 unsigned Idx2) const {
125 const MCInstrDesc &MCID = MI->getDesc();
126 bool HasDef = MCID.getNumDefs();
127 if (HasDef && !MI->getOperand(0).isReg())
128 // No idea how to commute this instruction. Target should implement its own.
129 return nullptr;
130
131 unsigned CommutableOpIdx1 = Idx1; (void)CommutableOpIdx1;
132 unsigned CommutableOpIdx2 = Idx2; (void)CommutableOpIdx2;
133 assert(findCommutedOpIndices(MI, CommutableOpIdx1, CommutableOpIdx2) &&
134 CommutableOpIdx1 == Idx1 && CommutableOpIdx2 == Idx2 &&
135 "TargetInstrInfo::CommuteInstructionImpl(): not commutable operands.");
136 assert(MI->getOperand(Idx1).isReg() && MI->getOperand(Idx2).isReg() &&
137 "This only knows how to commute register operands so far");
138
139 unsigned Reg0 = HasDef ? MI->getOperand(0).getReg() : 0;
140 unsigned Reg1 = MI->getOperand(Idx1).getReg();
141 unsigned Reg2 = MI->getOperand(Idx2).getReg();
142 unsigned SubReg0 = HasDef ? MI->getOperand(0).getSubReg() : 0;
143 unsigned SubReg1 = MI->getOperand(Idx1).getSubReg();
144 unsigned SubReg2 = MI->getOperand(Idx2).getSubReg();
145 bool Reg1IsKill = MI->getOperand(Idx1).isKill();
146 bool Reg2IsKill = MI->getOperand(Idx2).isKill();
147 bool Reg1IsUndef = MI->getOperand(Idx1).isUndef();
148 bool Reg2IsUndef = MI->getOperand(Idx2).isUndef();
149 bool Reg1IsInternal = MI->getOperand(Idx1).isInternalRead();
150 bool Reg2IsInternal = MI->getOperand(Idx2).isInternalRead();
151 // If destination is tied to either of the commuted source register, then
152 // it must be updated.
153 if (HasDef && Reg0 == Reg1 &&
154 MI->getDesc().getOperandConstraint(Idx1, MCOI::TIED_TO) == 0) {
155 Reg2IsKill = false;
156 Reg0 = Reg2;
157 SubReg0 = SubReg2;
158 } else if (HasDef && Reg0 == Reg2 &&
159 MI->getDesc().getOperandConstraint(Idx2, MCOI::TIED_TO) == 0) {
160 Reg1IsKill = false;
161 Reg0 = Reg1;
162 SubReg0 = SubReg1;
163 }
164
165 if (NewMI) {
166 // Create a new instruction.
167 MachineFunction &MF = *MI->getParent()->getParent();
168 MI = MF.CloneMachineInstr(MI);
169 }
170
171 if (HasDef) {
172 MI->getOperand(0).setReg(Reg0);
173 MI->getOperand(0).setSubReg(SubReg0);
174 }
175 MI->getOperand(Idx2).setReg(Reg1);
176 MI->getOperand(Idx1).setReg(Reg2);
177 MI->getOperand(Idx2).setSubReg(SubReg1);
178 MI->getOperand(Idx1).setSubReg(SubReg2);
179 MI->getOperand(Idx2).setIsKill(Reg1IsKill);
180 MI->getOperand(Idx1).setIsKill(Reg2IsKill);
181 MI->getOperand(Idx2).setIsUndef(Reg1IsUndef);
182 MI->getOperand(Idx1).setIsUndef(Reg2IsUndef);
183 MI->getOperand(Idx2).setIsInternalRead(Reg1IsInternal);
184 MI->getOperand(Idx1).setIsInternalRead(Reg2IsInternal);
185 return MI;
186 }
187
commuteInstruction(MachineInstr * MI,bool NewMI,unsigned OpIdx1,unsigned OpIdx2) const188 MachineInstr *TargetInstrInfo::commuteInstruction(MachineInstr *MI,
189 bool NewMI,
190 unsigned OpIdx1,
191 unsigned OpIdx2) const {
192 // If OpIdx1 or OpIdx2 is not specified, then this method is free to choose
193 // any commutable operand, which is done in findCommutedOpIndices() method
194 // called below.
195 if ((OpIdx1 == CommuteAnyOperandIndex || OpIdx2 == CommuteAnyOperandIndex) &&
196 !findCommutedOpIndices(MI, OpIdx1, OpIdx2)) {
197 assert(MI->isCommutable() &&
198 "Precondition violation: MI must be commutable.");
199 return nullptr;
200 }
201 return commuteInstructionImpl(MI, NewMI, OpIdx1, OpIdx2);
202 }
203
fixCommutedOpIndices(unsigned & ResultIdx1,unsigned & ResultIdx2,unsigned CommutableOpIdx1,unsigned CommutableOpIdx2)204 bool TargetInstrInfo::fixCommutedOpIndices(unsigned &ResultIdx1,
205 unsigned &ResultIdx2,
206 unsigned CommutableOpIdx1,
207 unsigned CommutableOpIdx2) {
208 if (ResultIdx1 == CommuteAnyOperandIndex &&
209 ResultIdx2 == CommuteAnyOperandIndex) {
210 ResultIdx1 = CommutableOpIdx1;
211 ResultIdx2 = CommutableOpIdx2;
212 } else if (ResultIdx1 == CommuteAnyOperandIndex) {
213 if (ResultIdx2 == CommutableOpIdx1)
214 ResultIdx1 = CommutableOpIdx2;
215 else if (ResultIdx2 == CommutableOpIdx2)
216 ResultIdx1 = CommutableOpIdx1;
217 else
218 return false;
219 } else if (ResultIdx2 == CommuteAnyOperandIndex) {
220 if (ResultIdx1 == CommutableOpIdx1)
221 ResultIdx2 = CommutableOpIdx2;
222 else if (ResultIdx1 == CommutableOpIdx2)
223 ResultIdx2 = CommutableOpIdx1;
224 else
225 return false;
226 } else
227 // Check that the result operand indices match the given commutable
228 // operand indices.
229 return (ResultIdx1 == CommutableOpIdx1 && ResultIdx2 == CommutableOpIdx2) ||
230 (ResultIdx1 == CommutableOpIdx2 && ResultIdx2 == CommutableOpIdx1);
231
232 return true;
233 }
234
findCommutedOpIndices(MachineInstr * MI,unsigned & SrcOpIdx1,unsigned & SrcOpIdx2) const235 bool TargetInstrInfo::findCommutedOpIndices(MachineInstr *MI,
236 unsigned &SrcOpIdx1,
237 unsigned &SrcOpIdx2) const {
238 assert(!MI->isBundle() &&
239 "TargetInstrInfo::findCommutedOpIndices() can't handle bundles");
240
241 const MCInstrDesc &MCID = MI->getDesc();
242 if (!MCID.isCommutable())
243 return false;
244
245 // This assumes v0 = op v1, v2 and commuting would swap v1 and v2. If this
246 // is not true, then the target must implement this.
247 unsigned CommutableOpIdx1 = MCID.getNumDefs();
248 unsigned CommutableOpIdx2 = CommutableOpIdx1 + 1;
249 if (!fixCommutedOpIndices(SrcOpIdx1, SrcOpIdx2,
250 CommutableOpIdx1, CommutableOpIdx2))
251 return false;
252
253 if (!MI->getOperand(SrcOpIdx1).isReg() ||
254 !MI->getOperand(SrcOpIdx2).isReg())
255 // No idea.
256 return false;
257 return true;
258 }
259
260 bool
isUnpredicatedTerminator(const MachineInstr * MI) const261 TargetInstrInfo::isUnpredicatedTerminator(const MachineInstr *MI) const {
262 if (!MI->isTerminator()) return false;
263
264 // Conditional branch is a special case.
265 if (MI->isBranch() && !MI->isBarrier())
266 return true;
267 if (!MI->isPredicable())
268 return true;
269 return !isPredicated(MI);
270 }
271
PredicateInstruction(MachineInstr * MI,ArrayRef<MachineOperand> Pred) const272 bool TargetInstrInfo::PredicateInstruction(
273 MachineInstr *MI, ArrayRef<MachineOperand> Pred) const {
274 bool MadeChange = false;
275
276 assert(!MI->isBundle() &&
277 "TargetInstrInfo::PredicateInstruction() can't handle bundles");
278
279 const MCInstrDesc &MCID = MI->getDesc();
280 if (!MI->isPredicable())
281 return false;
282
283 for (unsigned j = 0, i = 0, e = MI->getNumOperands(); i != e; ++i) {
284 if (MCID.OpInfo[i].isPredicate()) {
285 MachineOperand &MO = MI->getOperand(i);
286 if (MO.isReg()) {
287 MO.setReg(Pred[j].getReg());
288 MadeChange = true;
289 } else if (MO.isImm()) {
290 MO.setImm(Pred[j].getImm());
291 MadeChange = true;
292 } else if (MO.isMBB()) {
293 MO.setMBB(Pred[j].getMBB());
294 MadeChange = true;
295 }
296 ++j;
297 }
298 }
299 return MadeChange;
300 }
301
hasLoadFromStackSlot(const MachineInstr * MI,const MachineMemOperand * & MMO,int & FrameIndex) const302 bool TargetInstrInfo::hasLoadFromStackSlot(const MachineInstr *MI,
303 const MachineMemOperand *&MMO,
304 int &FrameIndex) const {
305 for (MachineInstr::mmo_iterator o = MI->memoperands_begin(),
306 oe = MI->memoperands_end();
307 o != oe;
308 ++o) {
309 if ((*o)->isLoad()) {
310 if (const FixedStackPseudoSourceValue *Value =
311 dyn_cast_or_null<FixedStackPseudoSourceValue>(
312 (*o)->getPseudoValue())) {
313 FrameIndex = Value->getFrameIndex();
314 MMO = *o;
315 return true;
316 }
317 }
318 }
319 return false;
320 }
321
hasStoreToStackSlot(const MachineInstr * MI,const MachineMemOperand * & MMO,int & FrameIndex) const322 bool TargetInstrInfo::hasStoreToStackSlot(const MachineInstr *MI,
323 const MachineMemOperand *&MMO,
324 int &FrameIndex) const {
325 for (MachineInstr::mmo_iterator o = MI->memoperands_begin(),
326 oe = MI->memoperands_end();
327 o != oe;
328 ++o) {
329 if ((*o)->isStore()) {
330 if (const FixedStackPseudoSourceValue *Value =
331 dyn_cast_or_null<FixedStackPseudoSourceValue>(
332 (*o)->getPseudoValue())) {
333 FrameIndex = Value->getFrameIndex();
334 MMO = *o;
335 return true;
336 }
337 }
338 }
339 return false;
340 }
341
getStackSlotRange(const TargetRegisterClass * RC,unsigned SubIdx,unsigned & Size,unsigned & Offset,const MachineFunction & MF) const342 bool TargetInstrInfo::getStackSlotRange(const TargetRegisterClass *RC,
343 unsigned SubIdx, unsigned &Size,
344 unsigned &Offset,
345 const MachineFunction &MF) const {
346 if (!SubIdx) {
347 Size = RC->getSize();
348 Offset = 0;
349 return true;
350 }
351 const TargetRegisterInfo *TRI = MF.getSubtarget().getRegisterInfo();
352 unsigned BitSize = TRI->getSubRegIdxSize(SubIdx);
353 // Convert bit size to byte size to be consistent with
354 // MCRegisterClass::getSize().
355 if (BitSize % 8)
356 return false;
357
358 int BitOffset = TRI->getSubRegIdxOffset(SubIdx);
359 if (BitOffset < 0 || BitOffset % 8)
360 return false;
361
362 Size = BitSize /= 8;
363 Offset = (unsigned)BitOffset / 8;
364
365 assert(RC->getSize() >= (Offset + Size) && "bad subregister range");
366
367 if (!MF.getDataLayout().isLittleEndian()) {
368 Offset = RC->getSize() - (Offset + Size);
369 }
370 return true;
371 }
372
reMaterialize(MachineBasicBlock & MBB,MachineBasicBlock::iterator I,unsigned DestReg,unsigned SubIdx,const MachineInstr * Orig,const TargetRegisterInfo & TRI) const373 void TargetInstrInfo::reMaterialize(MachineBasicBlock &MBB,
374 MachineBasicBlock::iterator I,
375 unsigned DestReg,
376 unsigned SubIdx,
377 const MachineInstr *Orig,
378 const TargetRegisterInfo &TRI) const {
379 MachineInstr *MI = MBB.getParent()->CloneMachineInstr(Orig);
380 MI->substituteRegister(MI->getOperand(0).getReg(), DestReg, SubIdx, TRI);
381 MBB.insert(I, MI);
382 }
383
384 bool
produceSameValue(const MachineInstr * MI0,const MachineInstr * MI1,const MachineRegisterInfo * MRI) const385 TargetInstrInfo::produceSameValue(const MachineInstr *MI0,
386 const MachineInstr *MI1,
387 const MachineRegisterInfo *MRI) const {
388 return MI0->isIdenticalTo(MI1, MachineInstr::IgnoreVRegDefs);
389 }
390
duplicate(MachineInstr * Orig,MachineFunction & MF) const391 MachineInstr *TargetInstrInfo::duplicate(MachineInstr *Orig,
392 MachineFunction &MF) const {
393 assert(!Orig->isNotDuplicable() &&
394 "Instruction cannot be duplicated");
395 return MF.CloneMachineInstr(Orig);
396 }
397
398 // If the COPY instruction in MI can be folded to a stack operation, return
399 // the register class to use.
canFoldCopy(const MachineInstr * MI,unsigned FoldIdx)400 static const TargetRegisterClass *canFoldCopy(const MachineInstr *MI,
401 unsigned FoldIdx) {
402 assert(MI->isCopy() && "MI must be a COPY instruction");
403 if (MI->getNumOperands() != 2)
404 return nullptr;
405 assert(FoldIdx<2 && "FoldIdx refers no nonexistent operand");
406
407 const MachineOperand &FoldOp = MI->getOperand(FoldIdx);
408 const MachineOperand &LiveOp = MI->getOperand(1-FoldIdx);
409
410 if (FoldOp.getSubReg() || LiveOp.getSubReg())
411 return nullptr;
412
413 unsigned FoldReg = FoldOp.getReg();
414 unsigned LiveReg = LiveOp.getReg();
415
416 assert(TargetRegisterInfo::isVirtualRegister(FoldReg) &&
417 "Cannot fold physregs");
418
419 const MachineRegisterInfo &MRI = MI->getParent()->getParent()->getRegInfo();
420 const TargetRegisterClass *RC = MRI.getRegClass(FoldReg);
421
422 if (TargetRegisterInfo::isPhysicalRegister(LiveOp.getReg()))
423 return RC->contains(LiveOp.getReg()) ? RC : nullptr;
424
425 if (RC->hasSubClassEq(MRI.getRegClass(LiveReg)))
426 return RC;
427
428 // FIXME: Allow folding when register classes are memory compatible.
429 return nullptr;
430 }
431
getNoopForMachoTarget(MCInst & NopInst) const432 void TargetInstrInfo::getNoopForMachoTarget(MCInst &NopInst) const {
433 llvm_unreachable("Not a MachO target");
434 }
435
foldPatchpoint(MachineFunction & MF,MachineInstr * MI,ArrayRef<unsigned> Ops,int FrameIndex,const TargetInstrInfo & TII)436 static MachineInstr *foldPatchpoint(MachineFunction &MF, MachineInstr *MI,
437 ArrayRef<unsigned> Ops, int FrameIndex,
438 const TargetInstrInfo &TII) {
439 unsigned StartIdx = 0;
440 switch (MI->getOpcode()) {
441 case TargetOpcode::STACKMAP:
442 StartIdx = 2; // Skip ID, nShadowBytes.
443 break;
444 case TargetOpcode::PATCHPOINT: {
445 // For PatchPoint, the call args are not foldable.
446 PatchPointOpers opers(MI);
447 StartIdx = opers.getVarIdx();
448 break;
449 }
450 default:
451 llvm_unreachable("unexpected stackmap opcode");
452 }
453
454 // Return false if any operands requested for folding are not foldable (not
455 // part of the stackmap's live values).
456 for (unsigned Op : Ops) {
457 if (Op < StartIdx)
458 return nullptr;
459 }
460
461 MachineInstr *NewMI =
462 MF.CreateMachineInstr(TII.get(MI->getOpcode()), MI->getDebugLoc(), true);
463 MachineInstrBuilder MIB(MF, NewMI);
464
465 // No need to fold return, the meta data, and function arguments
466 for (unsigned i = 0; i < StartIdx; ++i)
467 MIB.addOperand(MI->getOperand(i));
468
469 for (unsigned i = StartIdx; i < MI->getNumOperands(); ++i) {
470 MachineOperand &MO = MI->getOperand(i);
471 if (std::find(Ops.begin(), Ops.end(), i) != Ops.end()) {
472 unsigned SpillSize;
473 unsigned SpillOffset;
474 // Compute the spill slot size and offset.
475 const TargetRegisterClass *RC =
476 MF.getRegInfo().getRegClass(MO.getReg());
477 bool Valid =
478 TII.getStackSlotRange(RC, MO.getSubReg(), SpillSize, SpillOffset, MF);
479 if (!Valid)
480 report_fatal_error("cannot spill patchpoint subregister operand");
481 MIB.addImm(StackMaps::IndirectMemRefOp);
482 MIB.addImm(SpillSize);
483 MIB.addFrameIndex(FrameIndex);
484 MIB.addImm(SpillOffset);
485 }
486 else
487 MIB.addOperand(MO);
488 }
489 return NewMI;
490 }
491
492 /// foldMemoryOperand - Attempt to fold a load or store of the specified stack
493 /// slot into the specified machine instruction for the specified operand(s).
494 /// If this is possible, a new instruction is returned with the specified
495 /// operand folded, otherwise NULL is returned. The client is responsible for
496 /// removing the old instruction and adding the new one in the instruction
497 /// stream.
foldMemoryOperand(MachineBasicBlock::iterator MI,ArrayRef<unsigned> Ops,int FI) const498 MachineInstr *TargetInstrInfo::foldMemoryOperand(MachineBasicBlock::iterator MI,
499 ArrayRef<unsigned> Ops,
500 int FI) const {
501 unsigned Flags = 0;
502 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
503 if (MI->getOperand(Ops[i]).isDef())
504 Flags |= MachineMemOperand::MOStore;
505 else
506 Flags |= MachineMemOperand::MOLoad;
507
508 MachineBasicBlock *MBB = MI->getParent();
509 assert(MBB && "foldMemoryOperand needs an inserted instruction");
510 MachineFunction &MF = *MBB->getParent();
511
512 MachineInstr *NewMI = nullptr;
513
514 if (MI->getOpcode() == TargetOpcode::STACKMAP ||
515 MI->getOpcode() == TargetOpcode::PATCHPOINT) {
516 // Fold stackmap/patchpoint.
517 NewMI = foldPatchpoint(MF, MI, Ops, FI, *this);
518 if (NewMI)
519 MBB->insert(MI, NewMI);
520 } else {
521 // Ask the target to do the actual folding.
522 NewMI = foldMemoryOperandImpl(MF, MI, Ops, MI, FI);
523 }
524
525 if (NewMI) {
526 NewMI->setMemRefs(MI->memoperands_begin(), MI->memoperands_end());
527 // Add a memory operand, foldMemoryOperandImpl doesn't do that.
528 assert((!(Flags & MachineMemOperand::MOStore) ||
529 NewMI->mayStore()) &&
530 "Folded a def to a non-store!");
531 assert((!(Flags & MachineMemOperand::MOLoad) ||
532 NewMI->mayLoad()) &&
533 "Folded a use to a non-load!");
534 const MachineFrameInfo &MFI = *MF.getFrameInfo();
535 assert(MFI.getObjectOffset(FI) != -1);
536 MachineMemOperand *MMO = MF.getMachineMemOperand(
537 MachinePointerInfo::getFixedStack(MF, FI), Flags, MFI.getObjectSize(FI),
538 MFI.getObjectAlignment(FI));
539 NewMI->addMemOperand(MF, MMO);
540
541 return NewMI;
542 }
543
544 // Straight COPY may fold as load/store.
545 if (!MI->isCopy() || Ops.size() != 1)
546 return nullptr;
547
548 const TargetRegisterClass *RC = canFoldCopy(MI, Ops[0]);
549 if (!RC)
550 return nullptr;
551
552 const MachineOperand &MO = MI->getOperand(1-Ops[0]);
553 MachineBasicBlock::iterator Pos = MI;
554 const TargetRegisterInfo *TRI = MF.getSubtarget().getRegisterInfo();
555
556 if (Flags == MachineMemOperand::MOStore)
557 storeRegToStackSlot(*MBB, Pos, MO.getReg(), MO.isKill(), FI, RC, TRI);
558 else
559 loadRegFromStackSlot(*MBB, Pos, MO.getReg(), FI, RC, TRI);
560 return --Pos;
561 }
562
hasReassociableOperands(const MachineInstr & Inst,const MachineBasicBlock * MBB) const563 bool TargetInstrInfo::hasReassociableOperands(
564 const MachineInstr &Inst, const MachineBasicBlock *MBB) const {
565 const MachineOperand &Op1 = Inst.getOperand(1);
566 const MachineOperand &Op2 = Inst.getOperand(2);
567 const MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo();
568
569 // We need virtual register definitions for the operands that we will
570 // reassociate.
571 MachineInstr *MI1 = nullptr;
572 MachineInstr *MI2 = nullptr;
573 if (Op1.isReg() && TargetRegisterInfo::isVirtualRegister(Op1.getReg()))
574 MI1 = MRI.getUniqueVRegDef(Op1.getReg());
575 if (Op2.isReg() && TargetRegisterInfo::isVirtualRegister(Op2.getReg()))
576 MI2 = MRI.getUniqueVRegDef(Op2.getReg());
577
578 // And they need to be in the trace (otherwise, they won't have a depth).
579 return MI1 && MI2 && MI1->getParent() == MBB && MI2->getParent() == MBB;
580 }
581
hasReassociableSibling(const MachineInstr & Inst,bool & Commuted) const582 bool TargetInstrInfo::hasReassociableSibling(const MachineInstr &Inst,
583 bool &Commuted) const {
584 const MachineBasicBlock *MBB = Inst.getParent();
585 const MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo();
586 MachineInstr *MI1 = MRI.getUniqueVRegDef(Inst.getOperand(1).getReg());
587 MachineInstr *MI2 = MRI.getUniqueVRegDef(Inst.getOperand(2).getReg());
588 unsigned AssocOpcode = Inst.getOpcode();
589
590 // If only one operand has the same opcode and it's the second source operand,
591 // the operands must be commuted.
592 Commuted = MI1->getOpcode() != AssocOpcode && MI2->getOpcode() == AssocOpcode;
593 if (Commuted)
594 std::swap(MI1, MI2);
595
596 // 1. The previous instruction must be the same type as Inst.
597 // 2. The previous instruction must have virtual register definitions for its
598 // operands in the same basic block as Inst.
599 // 3. The previous instruction's result must only be used by Inst.
600 return MI1->getOpcode() == AssocOpcode &&
601 hasReassociableOperands(*MI1, MBB) &&
602 MRI.hasOneNonDBGUse(MI1->getOperand(0).getReg());
603 }
604
605 // 1. The operation must be associative and commutative.
606 // 2. The instruction must have virtual register definitions for its
607 // operands in the same basic block.
608 // 3. The instruction must have a reassociable sibling.
isReassociationCandidate(const MachineInstr & Inst,bool & Commuted) const609 bool TargetInstrInfo::isReassociationCandidate(const MachineInstr &Inst,
610 bool &Commuted) const {
611 return isAssociativeAndCommutative(Inst) &&
612 hasReassociableOperands(Inst, Inst.getParent()) &&
613 hasReassociableSibling(Inst, Commuted);
614 }
615
616 // The concept of the reassociation pass is that these operations can benefit
617 // from this kind of transformation:
618 //
619 // A = ? op ?
620 // B = A op X (Prev)
621 // C = B op Y (Root)
622 // -->
623 // A = ? op ?
624 // B = X op Y
625 // C = A op B
626 //
627 // breaking the dependency between A and B, allowing them to be executed in
628 // parallel (or back-to-back in a pipeline) instead of depending on each other.
629
630 // FIXME: This has the potential to be expensive (compile time) while not
631 // improving the code at all. Some ways to limit the overhead:
632 // 1. Track successful transforms; bail out if hit rate gets too low.
633 // 2. Only enable at -O3 or some other non-default optimization level.
634 // 3. Pre-screen pattern candidates here: if an operand of the previous
635 // instruction is known to not increase the critical path, then don't match
636 // that pattern.
getMachineCombinerPatterns(MachineInstr & Root,SmallVectorImpl<MachineCombinerPattern> & Patterns) const637 bool TargetInstrInfo::getMachineCombinerPatterns(
638 MachineInstr &Root,
639 SmallVectorImpl<MachineCombinerPattern> &Patterns) const {
640
641 bool Commute;
642 if (isReassociationCandidate(Root, Commute)) {
643 // We found a sequence of instructions that may be suitable for a
644 // reassociation of operands to increase ILP. Specify each commutation
645 // possibility for the Prev instruction in the sequence and let the
646 // machine combiner decide if changing the operands is worthwhile.
647 if (Commute) {
648 Patterns.push_back(MachineCombinerPattern::REASSOC_AX_YB);
649 Patterns.push_back(MachineCombinerPattern::REASSOC_XA_YB);
650 } else {
651 Patterns.push_back(MachineCombinerPattern::REASSOC_AX_BY);
652 Patterns.push_back(MachineCombinerPattern::REASSOC_XA_BY);
653 }
654 return true;
655 }
656
657 return false;
658 }
659
660 /// Attempt the reassociation transformation to reduce critical path length.
661 /// See the above comments before getMachineCombinerPatterns().
reassociateOps(MachineInstr & Root,MachineInstr & Prev,MachineCombinerPattern Pattern,SmallVectorImpl<MachineInstr * > & InsInstrs,SmallVectorImpl<MachineInstr * > & DelInstrs,DenseMap<unsigned,unsigned> & InstrIdxForVirtReg) const662 void TargetInstrInfo::reassociateOps(
663 MachineInstr &Root, MachineInstr &Prev,
664 MachineCombinerPattern Pattern,
665 SmallVectorImpl<MachineInstr *> &InsInstrs,
666 SmallVectorImpl<MachineInstr *> &DelInstrs,
667 DenseMap<unsigned, unsigned> &InstrIdxForVirtReg) const {
668 MachineFunction *MF = Root.getParent()->getParent();
669 MachineRegisterInfo &MRI = MF->getRegInfo();
670 const TargetInstrInfo *TII = MF->getSubtarget().getInstrInfo();
671 const TargetRegisterInfo *TRI = MF->getSubtarget().getRegisterInfo();
672 const TargetRegisterClass *RC = Root.getRegClassConstraint(0, TII, TRI);
673
674 // This array encodes the operand index for each parameter because the
675 // operands may be commuted. Each row corresponds to a pattern value,
676 // and each column specifies the index of A, B, X, Y.
677 unsigned OpIdx[4][4] = {
678 { 1, 1, 2, 2 },
679 { 1, 2, 2, 1 },
680 { 2, 1, 1, 2 },
681 { 2, 2, 1, 1 }
682 };
683
684 int Row;
685 switch (Pattern) {
686 case MachineCombinerPattern::REASSOC_AX_BY: Row = 0; break;
687 case MachineCombinerPattern::REASSOC_AX_YB: Row = 1; break;
688 case MachineCombinerPattern::REASSOC_XA_BY: Row = 2; break;
689 case MachineCombinerPattern::REASSOC_XA_YB: Row = 3; break;
690 default: llvm_unreachable("unexpected MachineCombinerPattern");
691 }
692
693 MachineOperand &OpA = Prev.getOperand(OpIdx[Row][0]);
694 MachineOperand &OpB = Root.getOperand(OpIdx[Row][1]);
695 MachineOperand &OpX = Prev.getOperand(OpIdx[Row][2]);
696 MachineOperand &OpY = Root.getOperand(OpIdx[Row][3]);
697 MachineOperand &OpC = Root.getOperand(0);
698
699 unsigned RegA = OpA.getReg();
700 unsigned RegB = OpB.getReg();
701 unsigned RegX = OpX.getReg();
702 unsigned RegY = OpY.getReg();
703 unsigned RegC = OpC.getReg();
704
705 if (TargetRegisterInfo::isVirtualRegister(RegA))
706 MRI.constrainRegClass(RegA, RC);
707 if (TargetRegisterInfo::isVirtualRegister(RegB))
708 MRI.constrainRegClass(RegB, RC);
709 if (TargetRegisterInfo::isVirtualRegister(RegX))
710 MRI.constrainRegClass(RegX, RC);
711 if (TargetRegisterInfo::isVirtualRegister(RegY))
712 MRI.constrainRegClass(RegY, RC);
713 if (TargetRegisterInfo::isVirtualRegister(RegC))
714 MRI.constrainRegClass(RegC, RC);
715
716 // Create a new virtual register for the result of (X op Y) instead of
717 // recycling RegB because the MachineCombiner's computation of the critical
718 // path requires a new register definition rather than an existing one.
719 unsigned NewVR = MRI.createVirtualRegister(RC);
720 InstrIdxForVirtReg.insert(std::make_pair(NewVR, 0));
721
722 unsigned Opcode = Root.getOpcode();
723 bool KillA = OpA.isKill();
724 bool KillX = OpX.isKill();
725 bool KillY = OpY.isKill();
726
727 // Create new instructions for insertion.
728 MachineInstrBuilder MIB1 =
729 BuildMI(*MF, Prev.getDebugLoc(), TII->get(Opcode), NewVR)
730 .addReg(RegX, getKillRegState(KillX))
731 .addReg(RegY, getKillRegState(KillY));
732 MachineInstrBuilder MIB2 =
733 BuildMI(*MF, Root.getDebugLoc(), TII->get(Opcode), RegC)
734 .addReg(RegA, getKillRegState(KillA))
735 .addReg(NewVR, getKillRegState(true));
736
737 setSpecialOperandAttr(Root, Prev, *MIB1, *MIB2);
738
739 // Record new instructions for insertion and old instructions for deletion.
740 InsInstrs.push_back(MIB1);
741 InsInstrs.push_back(MIB2);
742 DelInstrs.push_back(&Prev);
743 DelInstrs.push_back(&Root);
744 }
745
genAlternativeCodeSequence(MachineInstr & Root,MachineCombinerPattern Pattern,SmallVectorImpl<MachineInstr * > & InsInstrs,SmallVectorImpl<MachineInstr * > & DelInstrs,DenseMap<unsigned,unsigned> & InstIdxForVirtReg) const746 void TargetInstrInfo::genAlternativeCodeSequence(
747 MachineInstr &Root, MachineCombinerPattern Pattern,
748 SmallVectorImpl<MachineInstr *> &InsInstrs,
749 SmallVectorImpl<MachineInstr *> &DelInstrs,
750 DenseMap<unsigned, unsigned> &InstIdxForVirtReg) const {
751 MachineRegisterInfo &MRI = Root.getParent()->getParent()->getRegInfo();
752
753 // Select the previous instruction in the sequence based on the input pattern.
754 MachineInstr *Prev = nullptr;
755 switch (Pattern) {
756 case MachineCombinerPattern::REASSOC_AX_BY:
757 case MachineCombinerPattern::REASSOC_XA_BY:
758 Prev = MRI.getUniqueVRegDef(Root.getOperand(1).getReg());
759 break;
760 case MachineCombinerPattern::REASSOC_AX_YB:
761 case MachineCombinerPattern::REASSOC_XA_YB:
762 Prev = MRI.getUniqueVRegDef(Root.getOperand(2).getReg());
763 break;
764 default:
765 break;
766 }
767
768 assert(Prev && "Unknown pattern for machine combiner");
769
770 reassociateOps(Root, *Prev, Pattern, InsInstrs, DelInstrs, InstIdxForVirtReg);
771 return;
772 }
773
774 /// foldMemoryOperand - Same as the previous version except it allows folding
775 /// of any load and store from / to any address, not just from a specific
776 /// stack slot.
foldMemoryOperand(MachineBasicBlock::iterator MI,ArrayRef<unsigned> Ops,MachineInstr * LoadMI) const777 MachineInstr *TargetInstrInfo::foldMemoryOperand(MachineBasicBlock::iterator MI,
778 ArrayRef<unsigned> Ops,
779 MachineInstr *LoadMI) const {
780 assert(LoadMI->canFoldAsLoad() && "LoadMI isn't foldable!");
781 #ifndef NDEBUG
782 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
783 assert(MI->getOperand(Ops[i]).isUse() && "Folding load into def!");
784 #endif
785 MachineBasicBlock &MBB = *MI->getParent();
786 MachineFunction &MF = *MBB.getParent();
787
788 // Ask the target to do the actual folding.
789 MachineInstr *NewMI = nullptr;
790 int FrameIndex = 0;
791
792 if ((MI->getOpcode() == TargetOpcode::STACKMAP ||
793 MI->getOpcode() == TargetOpcode::PATCHPOINT) &&
794 isLoadFromStackSlot(LoadMI, FrameIndex)) {
795 // Fold stackmap/patchpoint.
796 NewMI = foldPatchpoint(MF, MI, Ops, FrameIndex, *this);
797 if (NewMI)
798 NewMI = MBB.insert(MI, NewMI);
799 } else {
800 // Ask the target to do the actual folding.
801 NewMI = foldMemoryOperandImpl(MF, MI, Ops, MI, LoadMI);
802 }
803
804 if (!NewMI) return nullptr;
805
806 // Copy the memoperands from the load to the folded instruction.
807 if (MI->memoperands_empty()) {
808 NewMI->setMemRefs(LoadMI->memoperands_begin(),
809 LoadMI->memoperands_end());
810 }
811 else {
812 // Handle the rare case of folding multiple loads.
813 NewMI->setMemRefs(MI->memoperands_begin(),
814 MI->memoperands_end());
815 for (MachineInstr::mmo_iterator I = LoadMI->memoperands_begin(),
816 E = LoadMI->memoperands_end(); I != E; ++I) {
817 NewMI->addMemOperand(MF, *I);
818 }
819 }
820 return NewMI;
821 }
822
823 bool TargetInstrInfo::
isReallyTriviallyReMaterializableGeneric(const MachineInstr * MI,AliasAnalysis * AA) const824 isReallyTriviallyReMaterializableGeneric(const MachineInstr *MI,
825 AliasAnalysis *AA) const {
826 const MachineFunction &MF = *MI->getParent()->getParent();
827 const MachineRegisterInfo &MRI = MF.getRegInfo();
828
829 // Remat clients assume operand 0 is the defined register.
830 if (!MI->getNumOperands() || !MI->getOperand(0).isReg())
831 return false;
832 unsigned DefReg = MI->getOperand(0).getReg();
833
834 // A sub-register definition can only be rematerialized if the instruction
835 // doesn't read the other parts of the register. Otherwise it is really a
836 // read-modify-write operation on the full virtual register which cannot be
837 // moved safely.
838 if (TargetRegisterInfo::isVirtualRegister(DefReg) &&
839 MI->getOperand(0).getSubReg() && MI->readsVirtualRegister(DefReg))
840 return false;
841
842 // A load from a fixed stack slot can be rematerialized. This may be
843 // redundant with subsequent checks, but it's target-independent,
844 // simple, and a common case.
845 int FrameIdx = 0;
846 if (isLoadFromStackSlot(MI, FrameIdx) &&
847 MF.getFrameInfo()->isImmutableObjectIndex(FrameIdx))
848 return true;
849
850 // Avoid instructions obviously unsafe for remat.
851 if (MI->isNotDuplicable() || MI->mayStore() ||
852 MI->hasUnmodeledSideEffects())
853 return false;
854
855 // Don't remat inline asm. We have no idea how expensive it is
856 // even if it's side effect free.
857 if (MI->isInlineAsm())
858 return false;
859
860 // Avoid instructions which load from potentially varying memory.
861 if (MI->mayLoad() && !MI->isInvariantLoad(AA))
862 return false;
863
864 // If any of the registers accessed are non-constant, conservatively assume
865 // the instruction is not rematerializable.
866 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
867 const MachineOperand &MO = MI->getOperand(i);
868 if (!MO.isReg()) continue;
869 unsigned Reg = MO.getReg();
870 if (Reg == 0)
871 continue;
872
873 // Check for a well-behaved physical register.
874 if (TargetRegisterInfo::isPhysicalRegister(Reg)) {
875 if (MO.isUse()) {
876 // If the physreg has no defs anywhere, it's just an ambient register
877 // and we can freely move its uses. Alternatively, if it's allocatable,
878 // it could get allocated to something with a def during allocation.
879 if (!MRI.isConstantPhysReg(Reg, MF))
880 return false;
881 } else {
882 // A physreg def. We can't remat it.
883 return false;
884 }
885 continue;
886 }
887
888 // Only allow one virtual-register def. There may be multiple defs of the
889 // same virtual register, though.
890 if (MO.isDef() && Reg != DefReg)
891 return false;
892
893 // Don't allow any virtual-register uses. Rematting an instruction with
894 // virtual register uses would length the live ranges of the uses, which
895 // is not necessarily a good idea, certainly not "trivial".
896 if (MO.isUse())
897 return false;
898 }
899
900 // Everything checked out.
901 return true;
902 }
903
getSPAdjust(const MachineInstr * MI) const904 int TargetInstrInfo::getSPAdjust(const MachineInstr *MI) const {
905 const MachineFunction *MF = MI->getParent()->getParent();
906 const TargetFrameLowering *TFI = MF->getSubtarget().getFrameLowering();
907 bool StackGrowsDown =
908 TFI->getStackGrowthDirection() == TargetFrameLowering::StackGrowsDown;
909
910 unsigned FrameSetupOpcode = getCallFrameSetupOpcode();
911 unsigned FrameDestroyOpcode = getCallFrameDestroyOpcode();
912
913 if (MI->getOpcode() != FrameSetupOpcode &&
914 MI->getOpcode() != FrameDestroyOpcode)
915 return 0;
916
917 int SPAdj = MI->getOperand(0).getImm();
918 SPAdj = TFI->alignSPAdjust(SPAdj);
919
920 if ((!StackGrowsDown && MI->getOpcode() == FrameSetupOpcode) ||
921 (StackGrowsDown && MI->getOpcode() == FrameDestroyOpcode))
922 SPAdj = -SPAdj;
923
924 return SPAdj;
925 }
926
927 /// isSchedulingBoundary - Test if the given instruction should be
928 /// considered a scheduling boundary. This primarily includes labels
929 /// and terminators.
isSchedulingBoundary(const MachineInstr * MI,const MachineBasicBlock * MBB,const MachineFunction & MF) const930 bool TargetInstrInfo::isSchedulingBoundary(const MachineInstr *MI,
931 const MachineBasicBlock *MBB,
932 const MachineFunction &MF) const {
933 // Terminators and labels can't be scheduled around.
934 if (MI->isTerminator() || MI->isPosition())
935 return true;
936
937 // Don't attempt to schedule around any instruction that defines
938 // a stack-oriented pointer, as it's unlikely to be profitable. This
939 // saves compile time, because it doesn't require every single
940 // stack slot reference to depend on the instruction that does the
941 // modification.
942 const TargetLowering &TLI = *MF.getSubtarget().getTargetLowering();
943 const TargetRegisterInfo *TRI = MF.getSubtarget().getRegisterInfo();
944 return MI->modifiesRegister(TLI.getStackPointerRegisterToSaveRestore(), TRI);
945 }
946
947 // Provide a global flag for disabling the PreRA hazard recognizer that targets
948 // may choose to honor.
usePreRAHazardRecognizer() const949 bool TargetInstrInfo::usePreRAHazardRecognizer() const {
950 return !DisableHazardRecognizer;
951 }
952
953 // Default implementation of CreateTargetRAHazardRecognizer.
954 ScheduleHazardRecognizer *TargetInstrInfo::
CreateTargetHazardRecognizer(const TargetSubtargetInfo * STI,const ScheduleDAG * DAG) const955 CreateTargetHazardRecognizer(const TargetSubtargetInfo *STI,
956 const ScheduleDAG *DAG) const {
957 // Dummy hazard recognizer allows all instructions to issue.
958 return new ScheduleHazardRecognizer();
959 }
960
961 // Default implementation of CreateTargetMIHazardRecognizer.
962 ScheduleHazardRecognizer *TargetInstrInfo::
CreateTargetMIHazardRecognizer(const InstrItineraryData * II,const ScheduleDAG * DAG) const963 CreateTargetMIHazardRecognizer(const InstrItineraryData *II,
964 const ScheduleDAG *DAG) const {
965 return (ScheduleHazardRecognizer *)
966 new ScoreboardHazardRecognizer(II, DAG, "misched");
967 }
968
969 // Default implementation of CreateTargetPostRAHazardRecognizer.
970 ScheduleHazardRecognizer *TargetInstrInfo::
CreateTargetPostRAHazardRecognizer(const InstrItineraryData * II,const ScheduleDAG * DAG) const971 CreateTargetPostRAHazardRecognizer(const InstrItineraryData *II,
972 const ScheduleDAG *DAG) const {
973 return (ScheduleHazardRecognizer *)
974 new ScoreboardHazardRecognizer(II, DAG, "post-RA-sched");
975 }
976
977 //===----------------------------------------------------------------------===//
978 // SelectionDAG latency interface.
979 //===----------------------------------------------------------------------===//
980
981 int
getOperandLatency(const InstrItineraryData * ItinData,SDNode * DefNode,unsigned DefIdx,SDNode * UseNode,unsigned UseIdx) const982 TargetInstrInfo::getOperandLatency(const InstrItineraryData *ItinData,
983 SDNode *DefNode, unsigned DefIdx,
984 SDNode *UseNode, unsigned UseIdx) const {
985 if (!ItinData || ItinData->isEmpty())
986 return -1;
987
988 if (!DefNode->isMachineOpcode())
989 return -1;
990
991 unsigned DefClass = get(DefNode->getMachineOpcode()).getSchedClass();
992 if (!UseNode->isMachineOpcode())
993 return ItinData->getOperandCycle(DefClass, DefIdx);
994 unsigned UseClass = get(UseNode->getMachineOpcode()).getSchedClass();
995 return ItinData->getOperandLatency(DefClass, DefIdx, UseClass, UseIdx);
996 }
997
getInstrLatency(const InstrItineraryData * ItinData,SDNode * N) const998 int TargetInstrInfo::getInstrLatency(const InstrItineraryData *ItinData,
999 SDNode *N) const {
1000 if (!ItinData || ItinData->isEmpty())
1001 return 1;
1002
1003 if (!N->isMachineOpcode())
1004 return 1;
1005
1006 return ItinData->getStageLatency(get(N->getMachineOpcode()).getSchedClass());
1007 }
1008
1009 //===----------------------------------------------------------------------===//
1010 // MachineInstr latency interface.
1011 //===----------------------------------------------------------------------===//
1012
1013 unsigned
getNumMicroOps(const InstrItineraryData * ItinData,const MachineInstr * MI) const1014 TargetInstrInfo::getNumMicroOps(const InstrItineraryData *ItinData,
1015 const MachineInstr *MI) const {
1016 if (!ItinData || ItinData->isEmpty())
1017 return 1;
1018
1019 unsigned Class = MI->getDesc().getSchedClass();
1020 int UOps = ItinData->Itineraries[Class].NumMicroOps;
1021 if (UOps >= 0)
1022 return UOps;
1023
1024 // The # of u-ops is dynamically determined. The specific target should
1025 // override this function to return the right number.
1026 return 1;
1027 }
1028
1029 /// Return the default expected latency for a def based on it's opcode.
defaultDefLatency(const MCSchedModel & SchedModel,const MachineInstr * DefMI) const1030 unsigned TargetInstrInfo::defaultDefLatency(const MCSchedModel &SchedModel,
1031 const MachineInstr *DefMI) const {
1032 if (DefMI->isTransient())
1033 return 0;
1034 if (DefMI->mayLoad())
1035 return SchedModel.LoadLatency;
1036 if (isHighLatencyDef(DefMI->getOpcode()))
1037 return SchedModel.HighLatency;
1038 return 1;
1039 }
1040
getPredicationCost(const MachineInstr *) const1041 unsigned TargetInstrInfo::getPredicationCost(const MachineInstr *) const {
1042 return 0;
1043 }
1044
1045 unsigned TargetInstrInfo::
getInstrLatency(const InstrItineraryData * ItinData,const MachineInstr * MI,unsigned * PredCost) const1046 getInstrLatency(const InstrItineraryData *ItinData,
1047 const MachineInstr *MI,
1048 unsigned *PredCost) const {
1049 // Default to one cycle for no itinerary. However, an "empty" itinerary may
1050 // still have a MinLatency property, which getStageLatency checks.
1051 if (!ItinData)
1052 return MI->mayLoad() ? 2 : 1;
1053
1054 return ItinData->getStageLatency(MI->getDesc().getSchedClass());
1055 }
1056
hasLowDefLatency(const TargetSchedModel & SchedModel,const MachineInstr * DefMI,unsigned DefIdx) const1057 bool TargetInstrInfo::hasLowDefLatency(const TargetSchedModel &SchedModel,
1058 const MachineInstr *DefMI,
1059 unsigned DefIdx) const {
1060 const InstrItineraryData *ItinData = SchedModel.getInstrItineraries();
1061 if (!ItinData || ItinData->isEmpty())
1062 return false;
1063
1064 unsigned DefClass = DefMI->getDesc().getSchedClass();
1065 int DefCycle = ItinData->getOperandCycle(DefClass, DefIdx);
1066 return (DefCycle != -1 && DefCycle <= 1);
1067 }
1068
1069 /// Both DefMI and UseMI must be valid. By default, call directly to the
1070 /// itinerary. This may be overriden by the target.
1071 int TargetInstrInfo::
getOperandLatency(const InstrItineraryData * ItinData,const MachineInstr * DefMI,unsigned DefIdx,const MachineInstr * UseMI,unsigned UseIdx) const1072 getOperandLatency(const InstrItineraryData *ItinData,
1073 const MachineInstr *DefMI, unsigned DefIdx,
1074 const MachineInstr *UseMI, unsigned UseIdx) const {
1075 unsigned DefClass = DefMI->getDesc().getSchedClass();
1076 unsigned UseClass = UseMI->getDesc().getSchedClass();
1077 return ItinData->getOperandLatency(DefClass, DefIdx, UseClass, UseIdx);
1078 }
1079
1080 /// If we can determine the operand latency from the def only, without itinerary
1081 /// lookup, do so. Otherwise return -1.
computeDefOperandLatency(const InstrItineraryData * ItinData,const MachineInstr * DefMI) const1082 int TargetInstrInfo::computeDefOperandLatency(
1083 const InstrItineraryData *ItinData,
1084 const MachineInstr *DefMI) const {
1085
1086 // Let the target hook getInstrLatency handle missing itineraries.
1087 if (!ItinData)
1088 return getInstrLatency(ItinData, DefMI);
1089
1090 if(ItinData->isEmpty())
1091 return defaultDefLatency(ItinData->SchedModel, DefMI);
1092
1093 // ...operand lookup required
1094 return -1;
1095 }
1096
1097 /// computeOperandLatency - Compute and return the latency of the given data
1098 /// dependent def and use when the operand indices are already known. UseMI may
1099 /// be NULL for an unknown use.
1100 ///
1101 /// FindMin may be set to get the minimum vs. expected latency. Minimum
1102 /// latency is used for scheduling groups, while expected latency is for
1103 /// instruction cost and critical path.
1104 ///
1105 /// Depending on the subtarget's itinerary properties, this may or may not need
1106 /// to call getOperandLatency(). For most subtargets, we don't need DefIdx or
1107 /// UseIdx to compute min latency.
1108 unsigned TargetInstrInfo::
computeOperandLatency(const InstrItineraryData * ItinData,const MachineInstr * DefMI,unsigned DefIdx,const MachineInstr * UseMI,unsigned UseIdx) const1109 computeOperandLatency(const InstrItineraryData *ItinData,
1110 const MachineInstr *DefMI, unsigned DefIdx,
1111 const MachineInstr *UseMI, unsigned UseIdx) const {
1112
1113 int DefLatency = computeDefOperandLatency(ItinData, DefMI);
1114 if (DefLatency >= 0)
1115 return DefLatency;
1116
1117 assert(ItinData && !ItinData->isEmpty() && "computeDefOperandLatency fail");
1118
1119 int OperLatency = 0;
1120 if (UseMI)
1121 OperLatency = getOperandLatency(ItinData, DefMI, DefIdx, UseMI, UseIdx);
1122 else {
1123 unsigned DefClass = DefMI->getDesc().getSchedClass();
1124 OperLatency = ItinData->getOperandCycle(DefClass, DefIdx);
1125 }
1126 if (OperLatency >= 0)
1127 return OperLatency;
1128
1129 // No operand latency was found.
1130 unsigned InstrLatency = getInstrLatency(ItinData, DefMI);
1131
1132 // Expected latency is the max of the stage latency and itinerary props.
1133 InstrLatency = std::max(InstrLatency,
1134 defaultDefLatency(ItinData->SchedModel, DefMI));
1135 return InstrLatency;
1136 }
1137
getRegSequenceInputs(const MachineInstr & MI,unsigned DefIdx,SmallVectorImpl<RegSubRegPairAndIdx> & InputRegs) const1138 bool TargetInstrInfo::getRegSequenceInputs(
1139 const MachineInstr &MI, unsigned DefIdx,
1140 SmallVectorImpl<RegSubRegPairAndIdx> &InputRegs) const {
1141 assert((MI.isRegSequence() ||
1142 MI.isRegSequenceLike()) && "Instruction do not have the proper type");
1143
1144 if (!MI.isRegSequence())
1145 return getRegSequenceLikeInputs(MI, DefIdx, InputRegs);
1146
1147 // We are looking at:
1148 // Def = REG_SEQUENCE v0, sub0, v1, sub1, ...
1149 assert(DefIdx == 0 && "REG_SEQUENCE only has one def");
1150 for (unsigned OpIdx = 1, EndOpIdx = MI.getNumOperands(); OpIdx != EndOpIdx;
1151 OpIdx += 2) {
1152 const MachineOperand &MOReg = MI.getOperand(OpIdx);
1153 const MachineOperand &MOSubIdx = MI.getOperand(OpIdx + 1);
1154 assert(MOSubIdx.isImm() &&
1155 "One of the subindex of the reg_sequence is not an immediate");
1156 // Record Reg:SubReg, SubIdx.
1157 InputRegs.push_back(RegSubRegPairAndIdx(MOReg.getReg(), MOReg.getSubReg(),
1158 (unsigned)MOSubIdx.getImm()));
1159 }
1160 return true;
1161 }
1162
getExtractSubregInputs(const MachineInstr & MI,unsigned DefIdx,RegSubRegPairAndIdx & InputReg) const1163 bool TargetInstrInfo::getExtractSubregInputs(
1164 const MachineInstr &MI, unsigned DefIdx,
1165 RegSubRegPairAndIdx &InputReg) const {
1166 assert((MI.isExtractSubreg() ||
1167 MI.isExtractSubregLike()) && "Instruction do not have the proper type");
1168
1169 if (!MI.isExtractSubreg())
1170 return getExtractSubregLikeInputs(MI, DefIdx, InputReg);
1171
1172 // We are looking at:
1173 // Def = EXTRACT_SUBREG v0.sub1, sub0.
1174 assert(DefIdx == 0 && "EXTRACT_SUBREG only has one def");
1175 const MachineOperand &MOReg = MI.getOperand(1);
1176 const MachineOperand &MOSubIdx = MI.getOperand(2);
1177 assert(MOSubIdx.isImm() &&
1178 "The subindex of the extract_subreg is not an immediate");
1179
1180 InputReg.Reg = MOReg.getReg();
1181 InputReg.SubReg = MOReg.getSubReg();
1182 InputReg.SubIdx = (unsigned)MOSubIdx.getImm();
1183 return true;
1184 }
1185
getInsertSubregInputs(const MachineInstr & MI,unsigned DefIdx,RegSubRegPair & BaseReg,RegSubRegPairAndIdx & InsertedReg) const1186 bool TargetInstrInfo::getInsertSubregInputs(
1187 const MachineInstr &MI, unsigned DefIdx,
1188 RegSubRegPair &BaseReg, RegSubRegPairAndIdx &InsertedReg) const {
1189 assert((MI.isInsertSubreg() ||
1190 MI.isInsertSubregLike()) && "Instruction do not have the proper type");
1191
1192 if (!MI.isInsertSubreg())
1193 return getInsertSubregLikeInputs(MI, DefIdx, BaseReg, InsertedReg);
1194
1195 // We are looking at:
1196 // Def = INSERT_SEQUENCE v0, v1, sub0.
1197 assert(DefIdx == 0 && "INSERT_SUBREG only has one def");
1198 const MachineOperand &MOBaseReg = MI.getOperand(1);
1199 const MachineOperand &MOInsertedReg = MI.getOperand(2);
1200 const MachineOperand &MOSubIdx = MI.getOperand(3);
1201 assert(MOSubIdx.isImm() &&
1202 "One of the subindex of the reg_sequence is not an immediate");
1203 BaseReg.Reg = MOBaseReg.getReg();
1204 BaseReg.SubReg = MOBaseReg.getSubReg();
1205
1206 InsertedReg.Reg = MOInsertedReg.getReg();
1207 InsertedReg.SubReg = MOInsertedReg.getSubReg();
1208 InsertedReg.SubIdx = (unsigned)MOSubIdx.getImm();
1209 return true;
1210 }
1211