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
35 using namespace llvm;
36
37 static cl::opt<bool> DisableHazardRecognizer(
38 "disable-sched-hazard", cl::Hidden, cl::init(false),
39 cl::desc("Disable hazard detection during preRA scheduling"));
40
~TargetInstrInfo()41 TargetInstrInfo::~TargetInstrInfo() {
42 }
43
44 const TargetRegisterClass*
getRegClass(const MCInstrDesc & MCID,unsigned OpNum,const TargetRegisterInfo * TRI,const MachineFunction & MF) const45 TargetInstrInfo::getRegClass(const MCInstrDesc &MCID, unsigned OpNum,
46 const TargetRegisterInfo *TRI,
47 const MachineFunction &MF) const {
48 if (OpNum >= MCID.getNumOperands())
49 return nullptr;
50
51 short RegClass = MCID.OpInfo[OpNum].RegClass;
52 if (MCID.OpInfo[OpNum].isLookupPtrRegClass())
53 return TRI->getPointerRegClass(MF, RegClass);
54
55 // Instructions like INSERT_SUBREG do not have fixed register classes.
56 if (RegClass < 0)
57 return nullptr;
58
59 // Otherwise just look it up normally.
60 return TRI->getRegClass(RegClass);
61 }
62
63 /// insertNoop - Insert a noop into the instruction stream at the specified
64 /// point.
insertNoop(MachineBasicBlock & MBB,MachineBasicBlock::iterator MI) const65 void TargetInstrInfo::insertNoop(MachineBasicBlock &MBB,
66 MachineBasicBlock::iterator MI) const {
67 llvm_unreachable("Target didn't implement insertNoop!");
68 }
69
70 /// Measure the specified inline asm to determine an approximation of its
71 /// length.
72 /// Comments (which run till the next SeparatorString or newline) do not
73 /// count as an instruction.
74 /// Any other non-whitespace text is considered an instruction, with
75 /// multiple instructions separated by SeparatorString or newlines.
76 /// Variable-length instructions are not handled here; this function
77 /// may be overloaded in the target code to do that.
getInlineAsmLength(const char * Str,const MCAsmInfo & MAI) const78 unsigned TargetInstrInfo::getInlineAsmLength(const char *Str,
79 const MCAsmInfo &MAI) const {
80 // Count the number of instructions in the asm.
81 bool atInsnStart = true;
82 unsigned InstCount = 0;
83 for (; *Str; ++Str) {
84 if (*Str == '\n' || strncmp(Str, MAI.getSeparatorString(),
85 strlen(MAI.getSeparatorString())) == 0) {
86 atInsnStart = true;
87 } else if (strncmp(Str, MAI.getCommentString(),
88 strlen(MAI.getCommentString())) == 0) {
89 // Stop counting as an instruction after a comment until the next
90 // separator.
91 atInsnStart = false;
92 }
93
94 if (atInsnStart && !std::isspace(static_cast<unsigned char>(*Str))) {
95 ++InstCount;
96 atInsnStart = false;
97 }
98 }
99
100 return InstCount * MAI.getMaxInstLength();
101 }
102
103 /// ReplaceTailWithBranchTo - Delete the instruction OldInst and everything
104 /// after it, replacing it with an unconditional branch to NewDest.
105 void
ReplaceTailWithBranchTo(MachineBasicBlock::iterator Tail,MachineBasicBlock * NewDest) const106 TargetInstrInfo::ReplaceTailWithBranchTo(MachineBasicBlock::iterator Tail,
107 MachineBasicBlock *NewDest) const {
108 MachineBasicBlock *MBB = Tail->getParent();
109
110 // Remove all the old successors of MBB from the CFG.
111 while (!MBB->succ_empty())
112 MBB->removeSuccessor(MBB->succ_begin());
113
114 // Save off the debug loc before erasing the instruction.
115 DebugLoc DL = Tail->getDebugLoc();
116
117 // Remove all the dead instructions from the end of MBB.
118 MBB->erase(Tail, MBB->end());
119
120 // If MBB isn't immediately before MBB, insert a branch to it.
121 if (++MachineFunction::iterator(MBB) != MachineFunction::iterator(NewDest))
122 InsertBranch(*MBB, NewDest, nullptr, SmallVector<MachineOperand, 0>(), DL);
123 MBB->addSuccessor(NewDest);
124 }
125
commuteInstructionImpl(MachineInstr & MI,bool NewMI,unsigned Idx1,unsigned Idx2) const126 MachineInstr *TargetInstrInfo::commuteInstructionImpl(MachineInstr &MI,
127 bool NewMI, unsigned Idx1,
128 unsigned Idx2) const {
129 const MCInstrDesc &MCID = MI.getDesc();
130 bool HasDef = MCID.getNumDefs();
131 if (HasDef && !MI.getOperand(0).isReg())
132 // No idea how to commute this instruction. Target should implement its own.
133 return nullptr;
134
135 unsigned CommutableOpIdx1 = Idx1; (void)CommutableOpIdx1;
136 unsigned CommutableOpIdx2 = Idx2; (void)CommutableOpIdx2;
137 assert(findCommutedOpIndices(MI, CommutableOpIdx1, CommutableOpIdx2) &&
138 CommutableOpIdx1 == Idx1 && CommutableOpIdx2 == Idx2 &&
139 "TargetInstrInfo::CommuteInstructionImpl(): not commutable operands.");
140 assert(MI.getOperand(Idx1).isReg() && MI.getOperand(Idx2).isReg() &&
141 "This only knows how to commute register operands so far");
142
143 unsigned Reg0 = HasDef ? MI.getOperand(0).getReg() : 0;
144 unsigned Reg1 = MI.getOperand(Idx1).getReg();
145 unsigned Reg2 = MI.getOperand(Idx2).getReg();
146 unsigned SubReg0 = HasDef ? MI.getOperand(0).getSubReg() : 0;
147 unsigned SubReg1 = MI.getOperand(Idx1).getSubReg();
148 unsigned SubReg2 = MI.getOperand(Idx2).getSubReg();
149 bool Reg1IsKill = MI.getOperand(Idx1).isKill();
150 bool Reg2IsKill = MI.getOperand(Idx2).isKill();
151 bool Reg1IsUndef = MI.getOperand(Idx1).isUndef();
152 bool Reg2IsUndef = MI.getOperand(Idx2).isUndef();
153 bool Reg1IsInternal = MI.getOperand(Idx1).isInternalRead();
154 bool Reg2IsInternal = MI.getOperand(Idx2).isInternalRead();
155 // If destination is tied to either of the commuted source register, then
156 // it must be updated.
157 if (HasDef && Reg0 == Reg1 &&
158 MI.getDesc().getOperandConstraint(Idx1, MCOI::TIED_TO) == 0) {
159 Reg2IsKill = false;
160 Reg0 = Reg2;
161 SubReg0 = SubReg2;
162 } else if (HasDef && Reg0 == Reg2 &&
163 MI.getDesc().getOperandConstraint(Idx2, MCOI::TIED_TO) == 0) {
164 Reg1IsKill = false;
165 Reg0 = Reg1;
166 SubReg0 = SubReg1;
167 }
168
169 MachineInstr *CommutedMI = nullptr;
170 if (NewMI) {
171 // Create a new instruction.
172 MachineFunction &MF = *MI.getParent()->getParent();
173 CommutedMI = MF.CloneMachineInstr(&MI);
174 } else {
175 CommutedMI = &MI;
176 }
177
178 if (HasDef) {
179 CommutedMI->getOperand(0).setReg(Reg0);
180 CommutedMI->getOperand(0).setSubReg(SubReg0);
181 }
182 CommutedMI->getOperand(Idx2).setReg(Reg1);
183 CommutedMI->getOperand(Idx1).setReg(Reg2);
184 CommutedMI->getOperand(Idx2).setSubReg(SubReg1);
185 CommutedMI->getOperand(Idx1).setSubReg(SubReg2);
186 CommutedMI->getOperand(Idx2).setIsKill(Reg1IsKill);
187 CommutedMI->getOperand(Idx1).setIsKill(Reg2IsKill);
188 CommutedMI->getOperand(Idx2).setIsUndef(Reg1IsUndef);
189 CommutedMI->getOperand(Idx1).setIsUndef(Reg2IsUndef);
190 CommutedMI->getOperand(Idx2).setIsInternalRead(Reg1IsInternal);
191 CommutedMI->getOperand(Idx1).setIsInternalRead(Reg2IsInternal);
192 return CommutedMI;
193 }
194
commuteInstruction(MachineInstr & MI,bool NewMI,unsigned OpIdx1,unsigned OpIdx2) const195 MachineInstr *TargetInstrInfo::commuteInstruction(MachineInstr &MI, bool NewMI,
196 unsigned OpIdx1,
197 unsigned OpIdx2) const {
198 // If OpIdx1 or OpIdx2 is not specified, then this method is free to choose
199 // any commutable operand, which is done in findCommutedOpIndices() method
200 // called below.
201 if ((OpIdx1 == CommuteAnyOperandIndex || OpIdx2 == CommuteAnyOperandIndex) &&
202 !findCommutedOpIndices(MI, OpIdx1, OpIdx2)) {
203 assert(MI.isCommutable() &&
204 "Precondition violation: MI must be commutable.");
205 return nullptr;
206 }
207 return commuteInstructionImpl(MI, NewMI, OpIdx1, OpIdx2);
208 }
209
fixCommutedOpIndices(unsigned & ResultIdx1,unsigned & ResultIdx2,unsigned CommutableOpIdx1,unsigned CommutableOpIdx2)210 bool TargetInstrInfo::fixCommutedOpIndices(unsigned &ResultIdx1,
211 unsigned &ResultIdx2,
212 unsigned CommutableOpIdx1,
213 unsigned CommutableOpIdx2) {
214 if (ResultIdx1 == CommuteAnyOperandIndex &&
215 ResultIdx2 == CommuteAnyOperandIndex) {
216 ResultIdx1 = CommutableOpIdx1;
217 ResultIdx2 = CommutableOpIdx2;
218 } else if (ResultIdx1 == CommuteAnyOperandIndex) {
219 if (ResultIdx2 == CommutableOpIdx1)
220 ResultIdx1 = CommutableOpIdx2;
221 else if (ResultIdx2 == CommutableOpIdx2)
222 ResultIdx1 = CommutableOpIdx1;
223 else
224 return false;
225 } else if (ResultIdx2 == CommuteAnyOperandIndex) {
226 if (ResultIdx1 == CommutableOpIdx1)
227 ResultIdx2 = CommutableOpIdx2;
228 else if (ResultIdx1 == CommutableOpIdx2)
229 ResultIdx2 = CommutableOpIdx1;
230 else
231 return false;
232 } else
233 // Check that the result operand indices match the given commutable
234 // operand indices.
235 return (ResultIdx1 == CommutableOpIdx1 && ResultIdx2 == CommutableOpIdx2) ||
236 (ResultIdx1 == CommutableOpIdx2 && ResultIdx2 == CommutableOpIdx1);
237
238 return true;
239 }
240
findCommutedOpIndices(MachineInstr & MI,unsigned & SrcOpIdx1,unsigned & SrcOpIdx2) const241 bool TargetInstrInfo::findCommutedOpIndices(MachineInstr &MI,
242 unsigned &SrcOpIdx1,
243 unsigned &SrcOpIdx2) const {
244 assert(!MI.isBundle() &&
245 "TargetInstrInfo::findCommutedOpIndices() can't handle bundles");
246
247 const MCInstrDesc &MCID = MI.getDesc();
248 if (!MCID.isCommutable())
249 return false;
250
251 // This assumes v0 = op v1, v2 and commuting would swap v1 and v2. If this
252 // is not true, then the target must implement this.
253 unsigned CommutableOpIdx1 = MCID.getNumDefs();
254 unsigned CommutableOpIdx2 = CommutableOpIdx1 + 1;
255 if (!fixCommutedOpIndices(SrcOpIdx1, SrcOpIdx2,
256 CommutableOpIdx1, CommutableOpIdx2))
257 return false;
258
259 if (!MI.getOperand(SrcOpIdx1).isReg() || !MI.getOperand(SrcOpIdx2).isReg())
260 // No idea.
261 return false;
262 return true;
263 }
264
isUnpredicatedTerminator(const MachineInstr & MI) const265 bool TargetInstrInfo::isUnpredicatedTerminator(const MachineInstr &MI) const {
266 if (!MI.isTerminator()) return false;
267
268 // Conditional branch is a special case.
269 if (MI.isBranch() && !MI.isBarrier())
270 return true;
271 if (!MI.isPredicable())
272 return true;
273 return !isPredicated(MI);
274 }
275
PredicateInstruction(MachineInstr & MI,ArrayRef<MachineOperand> Pred) const276 bool TargetInstrInfo::PredicateInstruction(
277 MachineInstr &MI, ArrayRef<MachineOperand> Pred) const {
278 bool MadeChange = false;
279
280 assert(!MI.isBundle() &&
281 "TargetInstrInfo::PredicateInstruction() can't handle bundles");
282
283 const MCInstrDesc &MCID = MI.getDesc();
284 if (!MI.isPredicable())
285 return false;
286
287 for (unsigned j = 0, i = 0, e = MI.getNumOperands(); i != e; ++i) {
288 if (MCID.OpInfo[i].isPredicate()) {
289 MachineOperand &MO = MI.getOperand(i);
290 if (MO.isReg()) {
291 MO.setReg(Pred[j].getReg());
292 MadeChange = true;
293 } else if (MO.isImm()) {
294 MO.setImm(Pred[j].getImm());
295 MadeChange = true;
296 } else if (MO.isMBB()) {
297 MO.setMBB(Pred[j].getMBB());
298 MadeChange = true;
299 }
300 ++j;
301 }
302 }
303 return MadeChange;
304 }
305
hasLoadFromStackSlot(const MachineInstr & MI,const MachineMemOperand * & MMO,int & FrameIndex) const306 bool TargetInstrInfo::hasLoadFromStackSlot(const MachineInstr &MI,
307 const MachineMemOperand *&MMO,
308 int &FrameIndex) const {
309 for (MachineInstr::mmo_iterator o = MI.memoperands_begin(),
310 oe = MI.memoperands_end();
311 o != oe; ++o) {
312 if ((*o)->isLoad()) {
313 if (const FixedStackPseudoSourceValue *Value =
314 dyn_cast_or_null<FixedStackPseudoSourceValue>(
315 (*o)->getPseudoValue())) {
316 FrameIndex = Value->getFrameIndex();
317 MMO = *o;
318 return true;
319 }
320 }
321 }
322 return false;
323 }
324
hasStoreToStackSlot(const MachineInstr & MI,const MachineMemOperand * & MMO,int & FrameIndex) const325 bool TargetInstrInfo::hasStoreToStackSlot(const MachineInstr &MI,
326 const MachineMemOperand *&MMO,
327 int &FrameIndex) const {
328 for (MachineInstr::mmo_iterator o = MI.memoperands_begin(),
329 oe = MI.memoperands_end();
330 o != oe; ++o) {
331 if ((*o)->isStore()) {
332 if (const FixedStackPseudoSourceValue *Value =
333 dyn_cast_or_null<FixedStackPseudoSourceValue>(
334 (*o)->getPseudoValue())) {
335 FrameIndex = Value->getFrameIndex();
336 MMO = *o;
337 return true;
338 }
339 }
340 }
341 return false;
342 }
343
getStackSlotRange(const TargetRegisterClass * RC,unsigned SubIdx,unsigned & Size,unsigned & Offset,const MachineFunction & MF) const344 bool TargetInstrInfo::getStackSlotRange(const TargetRegisterClass *RC,
345 unsigned SubIdx, unsigned &Size,
346 unsigned &Offset,
347 const MachineFunction &MF) const {
348 if (!SubIdx) {
349 Size = RC->getSize();
350 Offset = 0;
351 return true;
352 }
353 const TargetRegisterInfo *TRI = MF.getSubtarget().getRegisterInfo();
354 unsigned BitSize = TRI->getSubRegIdxSize(SubIdx);
355 // Convert bit size to byte size to be consistent with
356 // MCRegisterClass::getSize().
357 if (BitSize % 8)
358 return false;
359
360 int BitOffset = TRI->getSubRegIdxOffset(SubIdx);
361 if (BitOffset < 0 || BitOffset % 8)
362 return false;
363
364 Size = BitSize /= 8;
365 Offset = (unsigned)BitOffset / 8;
366
367 assert(RC->getSize() >= (Offset + Size) && "bad subregister range");
368
369 if (!MF.getDataLayout().isLittleEndian()) {
370 Offset = RC->getSize() - (Offset + Size);
371 }
372 return true;
373 }
374
reMaterialize(MachineBasicBlock & MBB,MachineBasicBlock::iterator I,unsigned DestReg,unsigned SubIdx,const MachineInstr & Orig,const TargetRegisterInfo & TRI) const375 void TargetInstrInfo::reMaterialize(MachineBasicBlock &MBB,
376 MachineBasicBlock::iterator I,
377 unsigned DestReg, unsigned SubIdx,
378 const MachineInstr &Orig,
379 const TargetRegisterInfo &TRI) const {
380 MachineInstr *MI = MBB.getParent()->CloneMachineInstr(&Orig);
381 MI->substituteRegister(MI->getOperand(0).getReg(), DestReg, SubIdx, TRI);
382 MBB.insert(I, MI);
383 }
384
produceSameValue(const MachineInstr & MI0,const MachineInstr & MI1,const MachineRegisterInfo * MRI) const385 bool 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() && "Instruction cannot be duplicated");
394 return MF.CloneMachineInstr(&Orig);
395 }
396
397 // If the COPY instruction in MI can be folded to a stack operation, return
398 // the register class to use.
canFoldCopy(const MachineInstr & MI,unsigned FoldIdx)399 static const TargetRegisterClass *canFoldCopy(const MachineInstr &MI,
400 unsigned FoldIdx) {
401 assert(MI.isCopy() && "MI must be a COPY instruction");
402 if (MI.getNumOperands() != 2)
403 return nullptr;
404 assert(FoldIdx<2 && "FoldIdx refers no nonexistent operand");
405
406 const MachineOperand &FoldOp = MI.getOperand(FoldIdx);
407 const MachineOperand &LiveOp = MI.getOperand(1 - FoldIdx);
408
409 if (FoldOp.getSubReg() || LiveOp.getSubReg())
410 return nullptr;
411
412 unsigned FoldReg = FoldOp.getReg();
413 unsigned LiveReg = LiveOp.getReg();
414
415 assert(TargetRegisterInfo::isVirtualRegister(FoldReg) &&
416 "Cannot fold physregs");
417
418 const MachineRegisterInfo &MRI = MI.getParent()->getParent()->getRegInfo();
419 const TargetRegisterClass *RC = MRI.getRegClass(FoldReg);
420
421 if (TargetRegisterInfo::isPhysicalRegister(LiveOp.getReg()))
422 return RC->contains(LiveOp.getReg()) ? RC : nullptr;
423
424 if (RC->hasSubClassEq(MRI.getRegClass(LiveReg)))
425 return RC;
426
427 // FIXME: Allow folding when register classes are memory compatible.
428 return nullptr;
429 }
430
getNoopForMachoTarget(MCInst & NopInst) const431 void TargetInstrInfo::getNoopForMachoTarget(MCInst &NopInst) const {
432 llvm_unreachable("Not a MachO target");
433 }
434
foldPatchpoint(MachineFunction & MF,MachineInstr & MI,ArrayRef<unsigned> Ops,int FrameIndex,const TargetInstrInfo & TII)435 static MachineInstr *foldPatchpoint(MachineFunction &MF, MachineInstr &MI,
436 ArrayRef<unsigned> Ops, int FrameIndex,
437 const TargetInstrInfo &TII) {
438 unsigned StartIdx = 0;
439 switch (MI.getOpcode()) {
440 case TargetOpcode::STACKMAP:
441 StartIdx = 2; // Skip ID, nShadowBytes.
442 break;
443 case TargetOpcode::PATCHPOINT: {
444 // For PatchPoint, the call args are not foldable.
445 PatchPointOpers opers(&MI);
446 StartIdx = opers.getVarIdx();
447 break;
448 }
449 default:
450 llvm_unreachable("unexpected stackmap opcode");
451 }
452
453 // Return false if any operands requested for folding are not foldable (not
454 // part of the stackmap's live values).
455 for (unsigned Op : Ops) {
456 if (Op < StartIdx)
457 return nullptr;
458 }
459
460 MachineInstr *NewMI =
461 MF.CreateMachineInstr(TII.get(MI.getOpcode()), MI.getDebugLoc(), true);
462 MachineInstrBuilder MIB(MF, NewMI);
463
464 // No need to fold return, the meta data, and function arguments
465 for (unsigned i = 0; i < StartIdx; ++i)
466 MIB.addOperand(MI.getOperand(i));
467
468 for (unsigned i = StartIdx; i < MI.getNumOperands(); ++i) {
469 MachineOperand &MO = MI.getOperand(i);
470 if (std::find(Ops.begin(), Ops.end(), i) != Ops.end()) {
471 unsigned SpillSize;
472 unsigned SpillOffset;
473 // Compute the spill slot size and offset.
474 const TargetRegisterClass *RC =
475 MF.getRegInfo().getRegClass(MO.getReg());
476 bool Valid =
477 TII.getStackSlotRange(RC, MO.getSubReg(), SpillSize, SpillOffset, MF);
478 if (!Valid)
479 report_fatal_error("cannot spill patchpoint subregister operand");
480 MIB.addImm(StackMaps::IndirectMemRefOp);
481 MIB.addImm(SpillSize);
482 MIB.addFrameIndex(FrameIndex);
483 MIB.addImm(SpillOffset);
484 }
485 else
486 MIB.addOperand(MO);
487 }
488 return NewMI;
489 }
490
491 /// foldMemoryOperand - Attempt to fold a load or store of the specified stack
492 /// slot into the specified machine instruction for the specified operand(s).
493 /// If this is possible, a new instruction is returned with the specified
494 /// operand folded, otherwise NULL is returned. The client is responsible for
495 /// removing the old instruction and adding the new one in the instruction
496 /// stream.
foldMemoryOperand(MachineInstr & MI,ArrayRef<unsigned> Ops,int FI,LiveIntervals * LIS) const497 MachineInstr *TargetInstrInfo::foldMemoryOperand(MachineInstr &MI,
498 ArrayRef<unsigned> Ops, int FI,
499 LiveIntervals *LIS) const {
500 unsigned Flags = 0;
501 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
502 if (MI.getOperand(Ops[i]).isDef())
503 Flags |= MachineMemOperand::MOStore;
504 else
505 Flags |= MachineMemOperand::MOLoad;
506
507 MachineBasicBlock *MBB = MI.getParent();
508 assert(MBB && "foldMemoryOperand needs an inserted instruction");
509 MachineFunction &MF = *MBB->getParent();
510
511 MachineInstr *NewMI = nullptr;
512
513 if (MI.getOpcode() == TargetOpcode::STACKMAP ||
514 MI.getOpcode() == TargetOpcode::PATCHPOINT) {
515 // Fold stackmap/patchpoint.
516 NewMI = foldPatchpoint(MF, MI, Ops, FI, *this);
517 if (NewMI)
518 MBB->insert(MI, NewMI);
519 } else {
520 // Ask the target to do the actual folding.
521 NewMI = foldMemoryOperandImpl(MF, MI, Ops, MI, FI, LIS);
522 }
523
524 if (NewMI) {
525 NewMI->setMemRefs(MI.memoperands_begin(), MI.memoperands_end());
526 // Add a memory operand, foldMemoryOperandImpl doesn't do that.
527 assert((!(Flags & MachineMemOperand::MOStore) ||
528 NewMI->mayStore()) &&
529 "Folded a def to a non-store!");
530 assert((!(Flags & MachineMemOperand::MOLoad) ||
531 NewMI->mayLoad()) &&
532 "Folded a use to a non-load!");
533 const MachineFrameInfo &MFI = *MF.getFrameInfo();
534 assert(MFI.getObjectOffset(FI) != -1);
535 MachineMemOperand *MMO = MF.getMachineMemOperand(
536 MachinePointerInfo::getFixedStack(MF, FI), Flags, MFI.getObjectSize(FI),
537 MFI.getObjectAlignment(FI));
538 NewMI->addMemOperand(MF, MMO);
539
540 return NewMI;
541 }
542
543 // Straight COPY may fold as load/store.
544 if (!MI.isCopy() || Ops.size() != 1)
545 return nullptr;
546
547 const TargetRegisterClass *RC = canFoldCopy(MI, Ops[0]);
548 if (!RC)
549 return nullptr;
550
551 const MachineOperand &MO = MI.getOperand(1 - Ops[0]);
552 MachineBasicBlock::iterator Pos = MI;
553 const TargetRegisterInfo *TRI = MF.getSubtarget().getRegisterInfo();
554
555 if (Flags == MachineMemOperand::MOStore)
556 storeRegToStackSlot(*MBB, Pos, MO.getReg(), MO.isKill(), FI, RC, TRI);
557 else
558 loadRegFromStackSlot(*MBB, Pos, MO.getReg(), FI, RC, TRI);
559 return &*--Pos;
560 }
561
hasReassociableOperands(const MachineInstr & Inst,const MachineBasicBlock * MBB) const562 bool TargetInstrInfo::hasReassociableOperands(
563 const MachineInstr &Inst, const MachineBasicBlock *MBB) const {
564 const MachineOperand &Op1 = Inst.getOperand(1);
565 const MachineOperand &Op2 = Inst.getOperand(2);
566 const MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo();
567
568 // We need virtual register definitions for the operands that we will
569 // reassociate.
570 MachineInstr *MI1 = nullptr;
571 MachineInstr *MI2 = nullptr;
572 if (Op1.isReg() && TargetRegisterInfo::isVirtualRegister(Op1.getReg()))
573 MI1 = MRI.getUniqueVRegDef(Op1.getReg());
574 if (Op2.isReg() && TargetRegisterInfo::isVirtualRegister(Op2.getReg()))
575 MI2 = MRI.getUniqueVRegDef(Op2.getReg());
576
577 // And they need to be in the trace (otherwise, they won't have a depth).
578 return MI1 && MI2 && MI1->getParent() == MBB && MI2->getParent() == MBB;
579 }
580
hasReassociableSibling(const MachineInstr & Inst,bool & Commuted) const581 bool TargetInstrInfo::hasReassociableSibling(const MachineInstr &Inst,
582 bool &Commuted) const {
583 const MachineBasicBlock *MBB = Inst.getParent();
584 const MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo();
585 MachineInstr *MI1 = MRI.getUniqueVRegDef(Inst.getOperand(1).getReg());
586 MachineInstr *MI2 = MRI.getUniqueVRegDef(Inst.getOperand(2).getReg());
587 unsigned AssocOpcode = Inst.getOpcode();
588
589 // If only one operand has the same opcode and it's the second source operand,
590 // the operands must be commuted.
591 Commuted = MI1->getOpcode() != AssocOpcode && MI2->getOpcode() == AssocOpcode;
592 if (Commuted)
593 std::swap(MI1, MI2);
594
595 // 1. The previous instruction must be the same type as Inst.
596 // 2. The previous instruction must have virtual register definitions for its
597 // operands in the same basic block as Inst.
598 // 3. The previous instruction's result must only be used by Inst.
599 return MI1->getOpcode() == AssocOpcode &&
600 hasReassociableOperands(*MI1, MBB) &&
601 MRI.hasOneNonDBGUse(MI1->getOperand(0).getReg());
602 }
603
604 // 1. The operation must be associative and commutative.
605 // 2. The instruction must have virtual register definitions for its
606 // operands in the same basic block.
607 // 3. The instruction must have a reassociable sibling.
isReassociationCandidate(const MachineInstr & Inst,bool & Commuted) const608 bool TargetInstrInfo::isReassociationCandidate(const MachineInstr &Inst,
609 bool &Commuted) const {
610 return isAssociativeAndCommutative(Inst) &&
611 hasReassociableOperands(Inst, Inst.getParent()) &&
612 hasReassociableSibling(Inst, Commuted);
613 }
614
615 // The concept of the reassociation pass is that these operations can benefit
616 // from this kind of transformation:
617 //
618 // A = ? op ?
619 // B = A op X (Prev)
620 // C = B op Y (Root)
621 // -->
622 // A = ? op ?
623 // B = X op Y
624 // C = A op B
625 //
626 // breaking the dependency between A and B, allowing them to be executed in
627 // parallel (or back-to-back in a pipeline) instead of depending on each other.
628
629 // FIXME: This has the potential to be expensive (compile time) while not
630 // improving the code at all. Some ways to limit the overhead:
631 // 1. Track successful transforms; bail out if hit rate gets too low.
632 // 2. Only enable at -O3 or some other non-default optimization level.
633 // 3. Pre-screen pattern candidates here: if an operand of the previous
634 // instruction is known to not increase the critical path, then don't match
635 // that pattern.
getMachineCombinerPatterns(MachineInstr & Root,SmallVectorImpl<MachineCombinerPattern> & Patterns) const636 bool TargetInstrInfo::getMachineCombinerPatterns(
637 MachineInstr &Root,
638 SmallVectorImpl<MachineCombinerPattern> &Patterns) const {
639 bool Commute;
640 if (isReassociationCandidate(Root, Commute)) {
641 // We found a sequence of instructions that may be suitable for a
642 // reassociation of operands to increase ILP. Specify each commutation
643 // possibility for the Prev instruction in the sequence and let the
644 // machine combiner decide if changing the operands is worthwhile.
645 if (Commute) {
646 Patterns.push_back(MachineCombinerPattern::REASSOC_AX_YB);
647 Patterns.push_back(MachineCombinerPattern::REASSOC_XA_YB);
648 } else {
649 Patterns.push_back(MachineCombinerPattern::REASSOC_AX_BY);
650 Patterns.push_back(MachineCombinerPattern::REASSOC_XA_BY);
651 }
652 return true;
653 }
654
655 return false;
656 }
657 /// Return true when a code sequence can improve loop throughput.
658 bool
isThroughputPattern(MachineCombinerPattern Pattern) const659 TargetInstrInfo::isThroughputPattern(MachineCombinerPattern Pattern) const {
660 return false;
661 }
662 /// Attempt the reassociation transformation to reduce critical path length.
663 /// See the above comments before getMachineCombinerPatterns().
reassociateOps(MachineInstr & Root,MachineInstr & Prev,MachineCombinerPattern Pattern,SmallVectorImpl<MachineInstr * > & InsInstrs,SmallVectorImpl<MachineInstr * > & DelInstrs,DenseMap<unsigned,unsigned> & InstrIdxForVirtReg) const664 void TargetInstrInfo::reassociateOps(
665 MachineInstr &Root, MachineInstr &Prev,
666 MachineCombinerPattern Pattern,
667 SmallVectorImpl<MachineInstr *> &InsInstrs,
668 SmallVectorImpl<MachineInstr *> &DelInstrs,
669 DenseMap<unsigned, unsigned> &InstrIdxForVirtReg) const {
670 MachineFunction *MF = Root.getParent()->getParent();
671 MachineRegisterInfo &MRI = MF->getRegInfo();
672 const TargetInstrInfo *TII = MF->getSubtarget().getInstrInfo();
673 const TargetRegisterInfo *TRI = MF->getSubtarget().getRegisterInfo();
674 const TargetRegisterClass *RC = Root.getRegClassConstraint(0, TII, TRI);
675
676 // This array encodes the operand index for each parameter because the
677 // operands may be commuted. Each row corresponds to a pattern value,
678 // and each column specifies the index of A, B, X, Y.
679 unsigned OpIdx[4][4] = {
680 { 1, 1, 2, 2 },
681 { 1, 2, 2, 1 },
682 { 2, 1, 1, 2 },
683 { 2, 2, 1, 1 }
684 };
685
686 int Row;
687 switch (Pattern) {
688 case MachineCombinerPattern::REASSOC_AX_BY: Row = 0; break;
689 case MachineCombinerPattern::REASSOC_AX_YB: Row = 1; break;
690 case MachineCombinerPattern::REASSOC_XA_BY: Row = 2; break;
691 case MachineCombinerPattern::REASSOC_XA_YB: Row = 3; break;
692 default: llvm_unreachable("unexpected MachineCombinerPattern");
693 }
694
695 MachineOperand &OpA = Prev.getOperand(OpIdx[Row][0]);
696 MachineOperand &OpB = Root.getOperand(OpIdx[Row][1]);
697 MachineOperand &OpX = Prev.getOperand(OpIdx[Row][2]);
698 MachineOperand &OpY = Root.getOperand(OpIdx[Row][3]);
699 MachineOperand &OpC = Root.getOperand(0);
700
701 unsigned RegA = OpA.getReg();
702 unsigned RegB = OpB.getReg();
703 unsigned RegX = OpX.getReg();
704 unsigned RegY = OpY.getReg();
705 unsigned RegC = OpC.getReg();
706
707 if (TargetRegisterInfo::isVirtualRegister(RegA))
708 MRI.constrainRegClass(RegA, RC);
709 if (TargetRegisterInfo::isVirtualRegister(RegB))
710 MRI.constrainRegClass(RegB, RC);
711 if (TargetRegisterInfo::isVirtualRegister(RegX))
712 MRI.constrainRegClass(RegX, RC);
713 if (TargetRegisterInfo::isVirtualRegister(RegY))
714 MRI.constrainRegClass(RegY, RC);
715 if (TargetRegisterInfo::isVirtualRegister(RegC))
716 MRI.constrainRegClass(RegC, RC);
717
718 // Create a new virtual register for the result of (X op Y) instead of
719 // recycling RegB because the MachineCombiner's computation of the critical
720 // path requires a new register definition rather than an existing one.
721 unsigned NewVR = MRI.createVirtualRegister(RC);
722 InstrIdxForVirtReg.insert(std::make_pair(NewVR, 0));
723
724 unsigned Opcode = Root.getOpcode();
725 bool KillA = OpA.isKill();
726 bool KillX = OpX.isKill();
727 bool KillY = OpY.isKill();
728
729 // Create new instructions for insertion.
730 MachineInstrBuilder MIB1 =
731 BuildMI(*MF, Prev.getDebugLoc(), TII->get(Opcode), NewVR)
732 .addReg(RegX, getKillRegState(KillX))
733 .addReg(RegY, getKillRegState(KillY));
734 MachineInstrBuilder MIB2 =
735 BuildMI(*MF, Root.getDebugLoc(), TII->get(Opcode), RegC)
736 .addReg(RegA, getKillRegState(KillA))
737 .addReg(NewVR, getKillRegState(true));
738
739 setSpecialOperandAttr(Root, Prev, *MIB1, *MIB2);
740
741 // Record new instructions for insertion and old instructions for deletion.
742 InsInstrs.push_back(MIB1);
743 InsInstrs.push_back(MIB2);
744 DelInstrs.push_back(&Prev);
745 DelInstrs.push_back(&Root);
746 }
747
genAlternativeCodeSequence(MachineInstr & Root,MachineCombinerPattern Pattern,SmallVectorImpl<MachineInstr * > & InsInstrs,SmallVectorImpl<MachineInstr * > & DelInstrs,DenseMap<unsigned,unsigned> & InstIdxForVirtReg) const748 void TargetInstrInfo::genAlternativeCodeSequence(
749 MachineInstr &Root, MachineCombinerPattern Pattern,
750 SmallVectorImpl<MachineInstr *> &InsInstrs,
751 SmallVectorImpl<MachineInstr *> &DelInstrs,
752 DenseMap<unsigned, unsigned> &InstIdxForVirtReg) const {
753 MachineRegisterInfo &MRI = Root.getParent()->getParent()->getRegInfo();
754
755 // Select the previous instruction in the sequence based on the input pattern.
756 MachineInstr *Prev = nullptr;
757 switch (Pattern) {
758 case MachineCombinerPattern::REASSOC_AX_BY:
759 case MachineCombinerPattern::REASSOC_XA_BY:
760 Prev = MRI.getUniqueVRegDef(Root.getOperand(1).getReg());
761 break;
762 case MachineCombinerPattern::REASSOC_AX_YB:
763 case MachineCombinerPattern::REASSOC_XA_YB:
764 Prev = MRI.getUniqueVRegDef(Root.getOperand(2).getReg());
765 break;
766 default:
767 break;
768 }
769
770 assert(Prev && "Unknown pattern for machine combiner");
771
772 reassociateOps(Root, *Prev, Pattern, InsInstrs, DelInstrs, InstIdxForVirtReg);
773 }
774
775 /// foldMemoryOperand - Same as the previous version except it allows folding
776 /// of any load and store from / to any address, not just from a specific
777 /// stack slot.
foldMemoryOperand(MachineInstr & MI,ArrayRef<unsigned> Ops,MachineInstr & LoadMI,LiveIntervals * LIS) const778 MachineInstr *TargetInstrInfo::foldMemoryOperand(MachineInstr &MI,
779 ArrayRef<unsigned> Ops,
780 MachineInstr &LoadMI,
781 LiveIntervals *LIS) const {
782 assert(LoadMI.canFoldAsLoad() && "LoadMI isn't foldable!");
783 #ifndef NDEBUG
784 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
785 assert(MI.getOperand(Ops[i]).isUse() && "Folding load into def!");
786 #endif
787 MachineBasicBlock &MBB = *MI.getParent();
788 MachineFunction &MF = *MBB.getParent();
789
790 // Ask the target to do the actual folding.
791 MachineInstr *NewMI = nullptr;
792 int FrameIndex = 0;
793
794 if ((MI.getOpcode() == TargetOpcode::STACKMAP ||
795 MI.getOpcode() == TargetOpcode::PATCHPOINT) &&
796 isLoadFromStackSlot(LoadMI, FrameIndex)) {
797 // Fold stackmap/patchpoint.
798 NewMI = foldPatchpoint(MF, MI, Ops, FrameIndex, *this);
799 if (NewMI)
800 NewMI = &*MBB.insert(MI, NewMI);
801 } else {
802 // Ask the target to do the actual folding.
803 NewMI = foldMemoryOperandImpl(MF, MI, Ops, MI, LoadMI, LIS);
804 }
805
806 if (!NewMI) return nullptr;
807
808 // Copy the memoperands from the load to the folded instruction.
809 if (MI.memoperands_empty()) {
810 NewMI->setMemRefs(LoadMI.memoperands_begin(), LoadMI.memoperands_end());
811 }
812 else {
813 // Handle the rare case of folding multiple loads.
814 NewMI->setMemRefs(MI.memoperands_begin(), MI.memoperands_end());
815 for (MachineInstr::mmo_iterator I = LoadMI.memoperands_begin(),
816 E = LoadMI.memoperands_end();
817 I != E; ++I) {
818 NewMI->addMemOperand(MF, *I);
819 }
820 }
821 return NewMI;
822 }
823
isReallyTriviallyReMaterializableGeneric(const MachineInstr & MI,AliasAnalysis * AA) const824 bool TargetInstrInfo::isReallyTriviallyReMaterializableGeneric(
825 const MachineInstr &MI, 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() || MI.hasUnmodeledSideEffects())
852 return false;
853
854 // Don't remat inline asm. We have no idea how expensive it is
855 // even if it's side effect free.
856 if (MI.isInlineAsm())
857 return false;
858
859 // Avoid instructions which load from potentially varying memory.
860 if (MI.mayLoad() && !MI.isInvariantLoad(AA))
861 return false;
862
863 // If any of the registers accessed are non-constant, conservatively assume
864 // the instruction is not rematerializable.
865 for (unsigned i = 0, e = MI.getNumOperands(); i != e; ++i) {
866 const MachineOperand &MO = MI.getOperand(i);
867 if (!MO.isReg()) continue;
868 unsigned Reg = MO.getReg();
869 if (Reg == 0)
870 continue;
871
872 // Check for a well-behaved physical register.
873 if (TargetRegisterInfo::isPhysicalRegister(Reg)) {
874 if (MO.isUse()) {
875 // If the physreg has no defs anywhere, it's just an ambient register
876 // and we can freely move its uses. Alternatively, if it's allocatable,
877 // it could get allocated to something with a def during allocation.
878 if (!MRI.isConstantPhysReg(Reg, MF))
879 return false;
880 } else {
881 // A physreg def. We can't remat it.
882 return false;
883 }
884 continue;
885 }
886
887 // Only allow one virtual-register def. There may be multiple defs of the
888 // same virtual register, though.
889 if (MO.isDef() && Reg != DefReg)
890 return false;
891
892 // Don't allow any virtual-register uses. Rematting an instruction with
893 // virtual register uses would length the live ranges of the uses, which
894 // is not necessarily a good idea, certainly not "trivial".
895 if (MO.isUse())
896 return false;
897 }
898
899 // Everything checked out.
900 return true;
901 }
902
getSPAdjust(const MachineInstr & MI) const903 int TargetInstrInfo::getSPAdjust(const MachineInstr &MI) const {
904 const MachineFunction *MF = MI.getParent()->getParent();
905 const TargetFrameLowering *TFI = MF->getSubtarget().getFrameLowering();
906 bool StackGrowsDown =
907 TFI->getStackGrowthDirection() == TargetFrameLowering::StackGrowsDown;
908
909 unsigned FrameSetupOpcode = getCallFrameSetupOpcode();
910 unsigned FrameDestroyOpcode = getCallFrameDestroyOpcode();
911
912 if (MI.getOpcode() != FrameSetupOpcode &&
913 MI.getOpcode() != FrameDestroyOpcode)
914 return 0;
915
916 int SPAdj = MI.getOperand(0).getImm();
917 SPAdj = TFI->alignSPAdjust(SPAdj);
918
919 if ((!StackGrowsDown && MI.getOpcode() == FrameSetupOpcode) ||
920 (StackGrowsDown && MI.getOpcode() == FrameDestroyOpcode))
921 SPAdj = -SPAdj;
922
923 return SPAdj;
924 }
925
926 /// isSchedulingBoundary - Test if the given instruction should be
927 /// considered a scheduling boundary. This primarily includes labels
928 /// and terminators.
isSchedulingBoundary(const MachineInstr & MI,const MachineBasicBlock * MBB,const MachineFunction & MF) const929 bool TargetInstrInfo::isSchedulingBoundary(const MachineInstr &MI,
930 const MachineBasicBlock *MBB,
931 const MachineFunction &MF) const {
932 // Terminators and labels can't be scheduled around.
933 if (MI.isTerminator() || MI.isPosition())
934 return true;
935
936 // Don't attempt to schedule around any instruction that defines
937 // a stack-oriented pointer, as it's unlikely to be profitable. This
938 // saves compile time, because it doesn't require every single
939 // stack slot reference to depend on the instruction that does the
940 // modification.
941 const TargetLowering &TLI = *MF.getSubtarget().getTargetLowering();
942 const TargetRegisterInfo *TRI = MF.getSubtarget().getRegisterInfo();
943 return MI.modifiesRegister(TLI.getStackPointerRegisterToSaveRestore(), TRI);
944 }
945
946 // Provide a global flag for disabling the PreRA hazard recognizer that targets
947 // may choose to honor.
usePreRAHazardRecognizer() const948 bool TargetInstrInfo::usePreRAHazardRecognizer() const {
949 return !DisableHazardRecognizer;
950 }
951
952 // Default implementation of CreateTargetRAHazardRecognizer.
953 ScheduleHazardRecognizer *TargetInstrInfo::
CreateTargetHazardRecognizer(const TargetSubtargetInfo * STI,const ScheduleDAG * DAG) const954 CreateTargetHazardRecognizer(const TargetSubtargetInfo *STI,
955 const ScheduleDAG *DAG) const {
956 // Dummy hazard recognizer allows all instructions to issue.
957 return new ScheduleHazardRecognizer();
958 }
959
960 // Default implementation of CreateTargetMIHazardRecognizer.
961 ScheduleHazardRecognizer *TargetInstrInfo::
CreateTargetMIHazardRecognizer(const InstrItineraryData * II,const ScheduleDAG * DAG) const962 CreateTargetMIHazardRecognizer(const InstrItineraryData *II,
963 const ScheduleDAG *DAG) const {
964 return (ScheduleHazardRecognizer *)
965 new ScoreboardHazardRecognizer(II, DAG, "misched");
966 }
967
968 // Default implementation of CreateTargetPostRAHazardRecognizer.
969 ScheduleHazardRecognizer *TargetInstrInfo::
CreateTargetPostRAHazardRecognizer(const InstrItineraryData * II,const ScheduleDAG * DAG) const970 CreateTargetPostRAHazardRecognizer(const InstrItineraryData *II,
971 const ScheduleDAG *DAG) const {
972 return (ScheduleHazardRecognizer *)
973 new ScoreboardHazardRecognizer(II, DAG, "post-RA-sched");
974 }
975
976 //===----------------------------------------------------------------------===//
977 // SelectionDAG latency interface.
978 //===----------------------------------------------------------------------===//
979
980 int
getOperandLatency(const InstrItineraryData * ItinData,SDNode * DefNode,unsigned DefIdx,SDNode * UseNode,unsigned UseIdx) const981 TargetInstrInfo::getOperandLatency(const InstrItineraryData *ItinData,
982 SDNode *DefNode, unsigned DefIdx,
983 SDNode *UseNode, unsigned UseIdx) const {
984 if (!ItinData || ItinData->isEmpty())
985 return -1;
986
987 if (!DefNode->isMachineOpcode())
988 return -1;
989
990 unsigned DefClass = get(DefNode->getMachineOpcode()).getSchedClass();
991 if (!UseNode->isMachineOpcode())
992 return ItinData->getOperandCycle(DefClass, DefIdx);
993 unsigned UseClass = get(UseNode->getMachineOpcode()).getSchedClass();
994 return ItinData->getOperandLatency(DefClass, DefIdx, UseClass, UseIdx);
995 }
996
getInstrLatency(const InstrItineraryData * ItinData,SDNode * N) const997 int TargetInstrInfo::getInstrLatency(const InstrItineraryData *ItinData,
998 SDNode *N) const {
999 if (!ItinData || ItinData->isEmpty())
1000 return 1;
1001
1002 if (!N->isMachineOpcode())
1003 return 1;
1004
1005 return ItinData->getStageLatency(get(N->getMachineOpcode()).getSchedClass());
1006 }
1007
1008 //===----------------------------------------------------------------------===//
1009 // MachineInstr latency interface.
1010 //===----------------------------------------------------------------------===//
1011
getNumMicroOps(const InstrItineraryData * ItinData,const MachineInstr & MI) const1012 unsigned TargetInstrInfo::getNumMicroOps(const InstrItineraryData *ItinData,
1013 const MachineInstr &MI) const {
1014 if (!ItinData || ItinData->isEmpty())
1015 return 1;
1016
1017 unsigned Class = MI.getDesc().getSchedClass();
1018 int UOps = ItinData->Itineraries[Class].NumMicroOps;
1019 if (UOps >= 0)
1020 return UOps;
1021
1022 // The # of u-ops is dynamically determined. The specific target should
1023 // override this function to return the right number.
1024 return 1;
1025 }
1026
1027 /// Return the default expected latency for a def based on it's opcode.
defaultDefLatency(const MCSchedModel & SchedModel,const MachineInstr & DefMI) const1028 unsigned TargetInstrInfo::defaultDefLatency(const MCSchedModel &SchedModel,
1029 const MachineInstr &DefMI) const {
1030 if (DefMI.isTransient())
1031 return 0;
1032 if (DefMI.mayLoad())
1033 return SchedModel.LoadLatency;
1034 if (isHighLatencyDef(DefMI.getOpcode()))
1035 return SchedModel.HighLatency;
1036 return 1;
1037 }
1038
getPredicationCost(const MachineInstr &) const1039 unsigned TargetInstrInfo::getPredicationCost(const MachineInstr &) const {
1040 return 0;
1041 }
1042
getInstrLatency(const InstrItineraryData * ItinData,const MachineInstr & MI,unsigned * PredCost) const1043 unsigned TargetInstrInfo::getInstrLatency(const InstrItineraryData *ItinData,
1044 const MachineInstr &MI,
1045 unsigned *PredCost) const {
1046 // Default to one cycle for no itinerary. However, an "empty" itinerary may
1047 // still have a MinLatency property, which getStageLatency checks.
1048 if (!ItinData)
1049 return MI.mayLoad() ? 2 : 1;
1050
1051 return ItinData->getStageLatency(MI.getDesc().getSchedClass());
1052 }
1053
hasLowDefLatency(const TargetSchedModel & SchedModel,const MachineInstr & DefMI,unsigned DefIdx) const1054 bool TargetInstrInfo::hasLowDefLatency(const TargetSchedModel &SchedModel,
1055 const MachineInstr &DefMI,
1056 unsigned DefIdx) const {
1057 const InstrItineraryData *ItinData = SchedModel.getInstrItineraries();
1058 if (!ItinData || ItinData->isEmpty())
1059 return false;
1060
1061 unsigned DefClass = DefMI.getDesc().getSchedClass();
1062 int DefCycle = ItinData->getOperandCycle(DefClass, DefIdx);
1063 return (DefCycle != -1 && DefCycle <= 1);
1064 }
1065
1066 /// Both DefMI and UseMI must be valid. By default, call directly to the
1067 /// itinerary. This may be overriden by the target.
getOperandLatency(const InstrItineraryData * ItinData,const MachineInstr & DefMI,unsigned DefIdx,const MachineInstr & UseMI,unsigned UseIdx) const1068 int TargetInstrInfo::getOperandLatency(const InstrItineraryData *ItinData,
1069 const MachineInstr &DefMI,
1070 unsigned DefIdx,
1071 const MachineInstr &UseMI,
1072 unsigned UseIdx) const {
1073 unsigned DefClass = DefMI.getDesc().getSchedClass();
1074 unsigned UseClass = UseMI.getDesc().getSchedClass();
1075 return ItinData->getOperandLatency(DefClass, DefIdx, UseClass, UseIdx);
1076 }
1077
1078 /// If we can determine the operand latency from the def only, without itinerary
1079 /// lookup, do so. Otherwise return -1.
computeDefOperandLatency(const InstrItineraryData * ItinData,const MachineInstr & DefMI) const1080 int TargetInstrInfo::computeDefOperandLatency(
1081 const InstrItineraryData *ItinData, const MachineInstr &DefMI) const {
1082
1083 // Let the target hook getInstrLatency handle missing itineraries.
1084 if (!ItinData)
1085 return getInstrLatency(ItinData, DefMI);
1086
1087 if(ItinData->isEmpty())
1088 return defaultDefLatency(ItinData->SchedModel, DefMI);
1089
1090 // ...operand lookup required
1091 return -1;
1092 }
1093
computeOperandLatency(const InstrItineraryData * ItinData,const MachineInstr & DefMI,unsigned DefIdx,const MachineInstr * UseMI,unsigned UseIdx) const1094 unsigned TargetInstrInfo::computeOperandLatency(
1095 const InstrItineraryData *ItinData, const MachineInstr &DefMI,
1096 unsigned DefIdx, const MachineInstr *UseMI, unsigned UseIdx) const {
1097
1098 int DefLatency = computeDefOperandLatency(ItinData, DefMI);
1099 if (DefLatency >= 0)
1100 return DefLatency;
1101
1102 assert(ItinData && !ItinData->isEmpty() && "computeDefOperandLatency fail");
1103
1104 int OperLatency = 0;
1105 if (UseMI)
1106 OperLatency = getOperandLatency(ItinData, DefMI, DefIdx, *UseMI, UseIdx);
1107 else {
1108 unsigned DefClass = DefMI.getDesc().getSchedClass();
1109 OperLatency = ItinData->getOperandCycle(DefClass, DefIdx);
1110 }
1111 if (OperLatency >= 0)
1112 return OperLatency;
1113
1114 // No operand latency was found.
1115 unsigned InstrLatency = getInstrLatency(ItinData, DefMI);
1116
1117 // Expected latency is the max of the stage latency and itinerary props.
1118 InstrLatency = std::max(InstrLatency,
1119 defaultDefLatency(ItinData->SchedModel, DefMI));
1120 return InstrLatency;
1121 }
1122
getRegSequenceInputs(const MachineInstr & MI,unsigned DefIdx,SmallVectorImpl<RegSubRegPairAndIdx> & InputRegs) const1123 bool TargetInstrInfo::getRegSequenceInputs(
1124 const MachineInstr &MI, unsigned DefIdx,
1125 SmallVectorImpl<RegSubRegPairAndIdx> &InputRegs) const {
1126 assert((MI.isRegSequence() ||
1127 MI.isRegSequenceLike()) && "Instruction do not have the proper type");
1128
1129 if (!MI.isRegSequence())
1130 return getRegSequenceLikeInputs(MI, DefIdx, InputRegs);
1131
1132 // We are looking at:
1133 // Def = REG_SEQUENCE v0, sub0, v1, sub1, ...
1134 assert(DefIdx == 0 && "REG_SEQUENCE only has one def");
1135 for (unsigned OpIdx = 1, EndOpIdx = MI.getNumOperands(); OpIdx != EndOpIdx;
1136 OpIdx += 2) {
1137 const MachineOperand &MOReg = MI.getOperand(OpIdx);
1138 const MachineOperand &MOSubIdx = MI.getOperand(OpIdx + 1);
1139 assert(MOSubIdx.isImm() &&
1140 "One of the subindex of the reg_sequence is not an immediate");
1141 // Record Reg:SubReg, SubIdx.
1142 InputRegs.push_back(RegSubRegPairAndIdx(MOReg.getReg(), MOReg.getSubReg(),
1143 (unsigned)MOSubIdx.getImm()));
1144 }
1145 return true;
1146 }
1147
getExtractSubregInputs(const MachineInstr & MI,unsigned DefIdx,RegSubRegPairAndIdx & InputReg) const1148 bool TargetInstrInfo::getExtractSubregInputs(
1149 const MachineInstr &MI, unsigned DefIdx,
1150 RegSubRegPairAndIdx &InputReg) const {
1151 assert((MI.isExtractSubreg() ||
1152 MI.isExtractSubregLike()) && "Instruction do not have the proper type");
1153
1154 if (!MI.isExtractSubreg())
1155 return getExtractSubregLikeInputs(MI, DefIdx, InputReg);
1156
1157 // We are looking at:
1158 // Def = EXTRACT_SUBREG v0.sub1, sub0.
1159 assert(DefIdx == 0 && "EXTRACT_SUBREG only has one def");
1160 const MachineOperand &MOReg = MI.getOperand(1);
1161 const MachineOperand &MOSubIdx = MI.getOperand(2);
1162 assert(MOSubIdx.isImm() &&
1163 "The subindex of the extract_subreg is not an immediate");
1164
1165 InputReg.Reg = MOReg.getReg();
1166 InputReg.SubReg = MOReg.getSubReg();
1167 InputReg.SubIdx = (unsigned)MOSubIdx.getImm();
1168 return true;
1169 }
1170
getInsertSubregInputs(const MachineInstr & MI,unsigned DefIdx,RegSubRegPair & BaseReg,RegSubRegPairAndIdx & InsertedReg) const1171 bool TargetInstrInfo::getInsertSubregInputs(
1172 const MachineInstr &MI, unsigned DefIdx,
1173 RegSubRegPair &BaseReg, RegSubRegPairAndIdx &InsertedReg) const {
1174 assert((MI.isInsertSubreg() ||
1175 MI.isInsertSubregLike()) && "Instruction do not have the proper type");
1176
1177 if (!MI.isInsertSubreg())
1178 return getInsertSubregLikeInputs(MI, DefIdx, BaseReg, InsertedReg);
1179
1180 // We are looking at:
1181 // Def = INSERT_SEQUENCE v0, v1, sub0.
1182 assert(DefIdx == 0 && "INSERT_SUBREG only has one def");
1183 const MachineOperand &MOBaseReg = MI.getOperand(1);
1184 const MachineOperand &MOInsertedReg = MI.getOperand(2);
1185 const MachineOperand &MOSubIdx = MI.getOperand(3);
1186 assert(MOSubIdx.isImm() &&
1187 "One of the subindex of the reg_sequence is not an immediate");
1188 BaseReg.Reg = MOBaseReg.getReg();
1189 BaseReg.SubReg = MOBaseReg.getSubReg();
1190
1191 InsertedReg.Reg = MOInsertedReg.getReg();
1192 InsertedReg.SubReg = MOInsertedReg.getSubReg();
1193 InsertedReg.SubIdx = (unsigned)MOSubIdx.getImm();
1194 return true;
1195 }
1196