xref: /freebsd/contrib/llvm-project/llvm/lib/CodeGen/TargetInstrInfo.cpp (revision d8096b2df282d7a50e56eddba523bcdda1676106)
1 //===-- TargetInstrInfo.cpp - Target Instruction Information --------------===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 // This file implements the TargetInstrInfo class.
10 //
11 //===----------------------------------------------------------------------===//
12 
13 #include "llvm/CodeGen/TargetInstrInfo.h"
14 #include "llvm/ADT/StringExtras.h"
15 #include "llvm/CodeGen/MachineFrameInfo.h"
16 #include "llvm/CodeGen/MachineInstrBuilder.h"
17 #include "llvm/CodeGen/MachineMemOperand.h"
18 #include "llvm/CodeGen/MachineRegisterInfo.h"
19 #include "llvm/CodeGen/MachineScheduler.h"
20 #include "llvm/CodeGen/PseudoSourceValue.h"
21 #include "llvm/CodeGen/ScoreboardHazardRecognizer.h"
22 #include "llvm/CodeGen/StackMaps.h"
23 #include "llvm/CodeGen/TargetFrameLowering.h"
24 #include "llvm/CodeGen/TargetLowering.h"
25 #include "llvm/CodeGen/TargetRegisterInfo.h"
26 #include "llvm/CodeGen/TargetSchedule.h"
27 #include "llvm/IR/DataLayout.h"
28 #include "llvm/IR/DebugInfoMetadata.h"
29 #include "llvm/MC/MCAsmInfo.h"
30 #include "llvm/MC/MCInstrItineraries.h"
31 #include "llvm/Support/CommandLine.h"
32 #include "llvm/Support/ErrorHandling.h"
33 #include "llvm/Support/raw_ostream.h"
34 #include "llvm/Target/TargetMachine.h"
35 #include <cctype>
36 
37 using namespace llvm;
38 
39 static cl::opt<bool> DisableHazardRecognizer(
40   "disable-sched-hazard", cl::Hidden, cl::init(false),
41   cl::desc("Disable hazard detection during preRA scheduling"));
42 
43 TargetInstrInfo::~TargetInstrInfo() {
44 }
45 
46 const TargetRegisterClass*
47 TargetInstrInfo::getRegClass(const MCInstrDesc &MCID, unsigned OpNum,
48                              const TargetRegisterInfo *TRI,
49                              const MachineFunction &MF) const {
50   if (OpNum >= MCID.getNumOperands())
51     return nullptr;
52 
53   short RegClass = MCID.OpInfo[OpNum].RegClass;
54   if (MCID.OpInfo[OpNum].isLookupPtrRegClass())
55     return TRI->getPointerRegClass(MF, RegClass);
56 
57   // Instructions like INSERT_SUBREG do not have fixed register classes.
58   if (RegClass < 0)
59     return nullptr;
60 
61   // Otherwise just look it up normally.
62   return TRI->getRegClass(RegClass);
63 }
64 
65 /// insertNoop - Insert a noop into the instruction stream at the specified
66 /// point.
67 void TargetInstrInfo::insertNoop(MachineBasicBlock &MBB,
68                                  MachineBasicBlock::iterator MI) const {
69   llvm_unreachable("Target didn't implement insertNoop!");
70 }
71 
72 /// insertNoops - Insert noops into the instruction stream at the specified
73 /// point.
74 void TargetInstrInfo::insertNoops(MachineBasicBlock &MBB,
75                                   MachineBasicBlock::iterator MI,
76                                   unsigned Quantity) const {
77   for (unsigned i = 0; i < Quantity; ++i)
78     insertNoop(MBB, MI);
79 }
80 
81 static bool isAsmComment(const char *Str, const MCAsmInfo &MAI) {
82   return strncmp(Str, MAI.getCommentString().data(),
83                  MAI.getCommentString().size()) == 0;
84 }
85 
86 /// Measure the specified inline asm to determine an approximation of its
87 /// length.
88 /// Comments (which run till the next SeparatorString or newline) do not
89 /// count as an instruction.
90 /// Any other non-whitespace text is considered an instruction, with
91 /// multiple instructions separated by SeparatorString or newlines.
92 /// Variable-length instructions are not handled here; this function
93 /// may be overloaded in the target code to do that.
94 /// We implement a special case of the .space directive which takes only a
95 /// single integer argument in base 10 that is the size in bytes. This is a
96 /// restricted form of the GAS directive in that we only interpret
97 /// simple--i.e. not a logical or arithmetic expression--size values without
98 /// the optional fill value. This is primarily used for creating arbitrary
99 /// sized inline asm blocks for testing purposes.
100 unsigned TargetInstrInfo::getInlineAsmLength(
101   const char *Str,
102   const MCAsmInfo &MAI, const TargetSubtargetInfo *STI) const {
103   // Count the number of instructions in the asm.
104   bool AtInsnStart = true;
105   unsigned Length = 0;
106   const unsigned MaxInstLength = MAI.getMaxInstLength(STI);
107   for (; *Str; ++Str) {
108     if (*Str == '\n' || strncmp(Str, MAI.getSeparatorString(),
109                                 strlen(MAI.getSeparatorString())) == 0) {
110       AtInsnStart = true;
111     } else if (isAsmComment(Str, MAI)) {
112       // Stop counting as an instruction after a comment until the next
113       // separator.
114       AtInsnStart = false;
115     }
116 
117     if (AtInsnStart && !isSpace(static_cast<unsigned char>(*Str))) {
118       unsigned AddLength = MaxInstLength;
119       if (strncmp(Str, ".space", 6) == 0) {
120         char *EStr;
121         int SpaceSize;
122         SpaceSize = strtol(Str + 6, &EStr, 10);
123         SpaceSize = SpaceSize < 0 ? 0 : SpaceSize;
124         while (*EStr != '\n' && isSpace(static_cast<unsigned char>(*EStr)))
125           ++EStr;
126         if (*EStr == '\0' || *EStr == '\n' ||
127             isAsmComment(EStr, MAI)) // Successfully parsed .space argument
128           AddLength = SpaceSize;
129       }
130       Length += AddLength;
131       AtInsnStart = false;
132     }
133   }
134 
135   return Length;
136 }
137 
138 /// ReplaceTailWithBranchTo - Delete the instruction OldInst and everything
139 /// after it, replacing it with an unconditional branch to NewDest.
140 void
141 TargetInstrInfo::ReplaceTailWithBranchTo(MachineBasicBlock::iterator Tail,
142                                          MachineBasicBlock *NewDest) const {
143   MachineBasicBlock *MBB = Tail->getParent();
144 
145   // Remove all the old successors of MBB from the CFG.
146   while (!MBB->succ_empty())
147     MBB->removeSuccessor(MBB->succ_begin());
148 
149   // Save off the debug loc before erasing the instruction.
150   DebugLoc DL = Tail->getDebugLoc();
151 
152   // Update call site info and remove all the dead instructions
153   // from the end of MBB.
154   while (Tail != MBB->end()) {
155     auto MI = Tail++;
156     if (MI->shouldUpdateCallSiteInfo())
157       MBB->getParent()->eraseCallSiteInfo(&*MI);
158     MBB->erase(MI);
159   }
160 
161   // If MBB isn't immediately before MBB, insert a branch to it.
162   if (++MachineFunction::iterator(MBB) != MachineFunction::iterator(NewDest))
163     insertBranch(*MBB, NewDest, nullptr, SmallVector<MachineOperand, 0>(), DL);
164   MBB->addSuccessor(NewDest);
165 }
166 
167 MachineInstr *TargetInstrInfo::commuteInstructionImpl(MachineInstr &MI,
168                                                       bool NewMI, unsigned Idx1,
169                                                       unsigned Idx2) const {
170   const MCInstrDesc &MCID = MI.getDesc();
171   bool HasDef = MCID.getNumDefs();
172   if (HasDef && !MI.getOperand(0).isReg())
173     // No idea how to commute this instruction. Target should implement its own.
174     return nullptr;
175 
176   unsigned CommutableOpIdx1 = Idx1; (void)CommutableOpIdx1;
177   unsigned CommutableOpIdx2 = Idx2; (void)CommutableOpIdx2;
178   assert(findCommutedOpIndices(MI, CommutableOpIdx1, CommutableOpIdx2) &&
179          CommutableOpIdx1 == Idx1 && CommutableOpIdx2 == Idx2 &&
180          "TargetInstrInfo::CommuteInstructionImpl(): not commutable operands.");
181   assert(MI.getOperand(Idx1).isReg() && MI.getOperand(Idx2).isReg() &&
182          "This only knows how to commute register operands so far");
183 
184   Register Reg0 = HasDef ? MI.getOperand(0).getReg() : Register();
185   Register Reg1 = MI.getOperand(Idx1).getReg();
186   Register Reg2 = MI.getOperand(Idx2).getReg();
187   unsigned SubReg0 = HasDef ? MI.getOperand(0).getSubReg() : 0;
188   unsigned SubReg1 = MI.getOperand(Idx1).getSubReg();
189   unsigned SubReg2 = MI.getOperand(Idx2).getSubReg();
190   bool Reg1IsKill = MI.getOperand(Idx1).isKill();
191   bool Reg2IsKill = MI.getOperand(Idx2).isKill();
192   bool Reg1IsUndef = MI.getOperand(Idx1).isUndef();
193   bool Reg2IsUndef = MI.getOperand(Idx2).isUndef();
194   bool Reg1IsInternal = MI.getOperand(Idx1).isInternalRead();
195   bool Reg2IsInternal = MI.getOperand(Idx2).isInternalRead();
196   // Avoid calling isRenamable for virtual registers since we assert that
197   // renamable property is only queried/set for physical registers.
198   bool Reg1IsRenamable = Register::isPhysicalRegister(Reg1)
199                              ? MI.getOperand(Idx1).isRenamable()
200                              : false;
201   bool Reg2IsRenamable = Register::isPhysicalRegister(Reg2)
202                              ? MI.getOperand(Idx2).isRenamable()
203                              : false;
204   // If destination is tied to either of the commuted source register, then
205   // it must be updated.
206   if (HasDef && Reg0 == Reg1 &&
207       MI.getDesc().getOperandConstraint(Idx1, MCOI::TIED_TO) == 0) {
208     Reg2IsKill = false;
209     Reg0 = Reg2;
210     SubReg0 = SubReg2;
211   } else if (HasDef && Reg0 == Reg2 &&
212              MI.getDesc().getOperandConstraint(Idx2, MCOI::TIED_TO) == 0) {
213     Reg1IsKill = false;
214     Reg0 = Reg1;
215     SubReg0 = SubReg1;
216   }
217 
218   MachineInstr *CommutedMI = nullptr;
219   if (NewMI) {
220     // Create a new instruction.
221     MachineFunction &MF = *MI.getMF();
222     CommutedMI = MF.CloneMachineInstr(&MI);
223   } else {
224     CommutedMI = &MI;
225   }
226 
227   if (HasDef) {
228     CommutedMI->getOperand(0).setReg(Reg0);
229     CommutedMI->getOperand(0).setSubReg(SubReg0);
230   }
231   CommutedMI->getOperand(Idx2).setReg(Reg1);
232   CommutedMI->getOperand(Idx1).setReg(Reg2);
233   CommutedMI->getOperand(Idx2).setSubReg(SubReg1);
234   CommutedMI->getOperand(Idx1).setSubReg(SubReg2);
235   CommutedMI->getOperand(Idx2).setIsKill(Reg1IsKill);
236   CommutedMI->getOperand(Idx1).setIsKill(Reg2IsKill);
237   CommutedMI->getOperand(Idx2).setIsUndef(Reg1IsUndef);
238   CommutedMI->getOperand(Idx1).setIsUndef(Reg2IsUndef);
239   CommutedMI->getOperand(Idx2).setIsInternalRead(Reg1IsInternal);
240   CommutedMI->getOperand(Idx1).setIsInternalRead(Reg2IsInternal);
241   // Avoid calling setIsRenamable for virtual registers since we assert that
242   // renamable property is only queried/set for physical registers.
243   if (Register::isPhysicalRegister(Reg1))
244     CommutedMI->getOperand(Idx2).setIsRenamable(Reg1IsRenamable);
245   if (Register::isPhysicalRegister(Reg2))
246     CommutedMI->getOperand(Idx1).setIsRenamable(Reg2IsRenamable);
247   return CommutedMI;
248 }
249 
250 MachineInstr *TargetInstrInfo::commuteInstruction(MachineInstr &MI, bool NewMI,
251                                                   unsigned OpIdx1,
252                                                   unsigned OpIdx2) const {
253   // If OpIdx1 or OpIdx2 is not specified, then this method is free to choose
254   // any commutable operand, which is done in findCommutedOpIndices() method
255   // called below.
256   if ((OpIdx1 == CommuteAnyOperandIndex || OpIdx2 == CommuteAnyOperandIndex) &&
257       !findCommutedOpIndices(MI, OpIdx1, OpIdx2)) {
258     assert(MI.isCommutable() &&
259            "Precondition violation: MI must be commutable.");
260     return nullptr;
261   }
262   return commuteInstructionImpl(MI, NewMI, OpIdx1, OpIdx2);
263 }
264 
265 bool TargetInstrInfo::fixCommutedOpIndices(unsigned &ResultIdx1,
266                                            unsigned &ResultIdx2,
267                                            unsigned CommutableOpIdx1,
268                                            unsigned CommutableOpIdx2) {
269   if (ResultIdx1 == CommuteAnyOperandIndex &&
270       ResultIdx2 == CommuteAnyOperandIndex) {
271     ResultIdx1 = CommutableOpIdx1;
272     ResultIdx2 = CommutableOpIdx2;
273   } else if (ResultIdx1 == CommuteAnyOperandIndex) {
274     if (ResultIdx2 == CommutableOpIdx1)
275       ResultIdx1 = CommutableOpIdx2;
276     else if (ResultIdx2 == CommutableOpIdx2)
277       ResultIdx1 = CommutableOpIdx1;
278     else
279       return false;
280   } else if (ResultIdx2 == CommuteAnyOperandIndex) {
281     if (ResultIdx1 == CommutableOpIdx1)
282       ResultIdx2 = CommutableOpIdx2;
283     else if (ResultIdx1 == CommutableOpIdx2)
284       ResultIdx2 = CommutableOpIdx1;
285     else
286       return false;
287   } else
288     // Check that the result operand indices match the given commutable
289     // operand indices.
290     return (ResultIdx1 == CommutableOpIdx1 && ResultIdx2 == CommutableOpIdx2) ||
291            (ResultIdx1 == CommutableOpIdx2 && ResultIdx2 == CommutableOpIdx1);
292 
293   return true;
294 }
295 
296 bool TargetInstrInfo::findCommutedOpIndices(const MachineInstr &MI,
297                                             unsigned &SrcOpIdx1,
298                                             unsigned &SrcOpIdx2) const {
299   assert(!MI.isBundle() &&
300          "TargetInstrInfo::findCommutedOpIndices() can't handle bundles");
301 
302   const MCInstrDesc &MCID = MI.getDesc();
303   if (!MCID.isCommutable())
304     return false;
305 
306   // This assumes v0 = op v1, v2 and commuting would swap v1 and v2. If this
307   // is not true, then the target must implement this.
308   unsigned CommutableOpIdx1 = MCID.getNumDefs();
309   unsigned CommutableOpIdx2 = CommutableOpIdx1 + 1;
310   if (!fixCommutedOpIndices(SrcOpIdx1, SrcOpIdx2,
311                             CommutableOpIdx1, CommutableOpIdx2))
312     return false;
313 
314   if (!MI.getOperand(SrcOpIdx1).isReg() || !MI.getOperand(SrcOpIdx2).isReg())
315     // No idea.
316     return false;
317   return true;
318 }
319 
320 bool TargetInstrInfo::isUnpredicatedTerminator(const MachineInstr &MI) const {
321   if (!MI.isTerminator()) return false;
322 
323   // Conditional branch is a special case.
324   if (MI.isBranch() && !MI.isBarrier())
325     return true;
326   if (!MI.isPredicable())
327     return true;
328   return !isPredicated(MI);
329 }
330 
331 bool TargetInstrInfo::PredicateInstruction(
332     MachineInstr &MI, ArrayRef<MachineOperand> Pred) const {
333   bool MadeChange = false;
334 
335   assert(!MI.isBundle() &&
336          "TargetInstrInfo::PredicateInstruction() can't handle bundles");
337 
338   const MCInstrDesc &MCID = MI.getDesc();
339   if (!MI.isPredicable())
340     return false;
341 
342   for (unsigned j = 0, i = 0, e = MI.getNumOperands(); i != e; ++i) {
343     if (MCID.OpInfo[i].isPredicate()) {
344       MachineOperand &MO = MI.getOperand(i);
345       if (MO.isReg()) {
346         MO.setReg(Pred[j].getReg());
347         MadeChange = true;
348       } else if (MO.isImm()) {
349         MO.setImm(Pred[j].getImm());
350         MadeChange = true;
351       } else if (MO.isMBB()) {
352         MO.setMBB(Pred[j].getMBB());
353         MadeChange = true;
354       }
355       ++j;
356     }
357   }
358   return MadeChange;
359 }
360 
361 bool TargetInstrInfo::hasLoadFromStackSlot(
362     const MachineInstr &MI,
363     SmallVectorImpl<const MachineMemOperand *> &Accesses) const {
364   size_t StartSize = Accesses.size();
365   for (MachineInstr::mmo_iterator o = MI.memoperands_begin(),
366                                   oe = MI.memoperands_end();
367        o != oe; ++o) {
368     if ((*o)->isLoad() &&
369         isa_and_nonnull<FixedStackPseudoSourceValue>((*o)->getPseudoValue()))
370       Accesses.push_back(*o);
371   }
372   return Accesses.size() != StartSize;
373 }
374 
375 bool TargetInstrInfo::hasStoreToStackSlot(
376     const MachineInstr &MI,
377     SmallVectorImpl<const MachineMemOperand *> &Accesses) const {
378   size_t StartSize = Accesses.size();
379   for (MachineInstr::mmo_iterator o = MI.memoperands_begin(),
380                                   oe = MI.memoperands_end();
381        o != oe; ++o) {
382     if ((*o)->isStore() &&
383         isa_and_nonnull<FixedStackPseudoSourceValue>((*o)->getPseudoValue()))
384       Accesses.push_back(*o);
385   }
386   return Accesses.size() != StartSize;
387 }
388 
389 bool TargetInstrInfo::getStackSlotRange(const TargetRegisterClass *RC,
390                                         unsigned SubIdx, unsigned &Size,
391                                         unsigned &Offset,
392                                         const MachineFunction &MF) const {
393   const TargetRegisterInfo *TRI = MF.getSubtarget().getRegisterInfo();
394   if (!SubIdx) {
395     Size = TRI->getSpillSize(*RC);
396     Offset = 0;
397     return true;
398   }
399   unsigned BitSize = TRI->getSubRegIdxSize(SubIdx);
400   // Convert bit size to byte size.
401   if (BitSize % 8)
402     return false;
403 
404   int BitOffset = TRI->getSubRegIdxOffset(SubIdx);
405   if (BitOffset < 0 || BitOffset % 8)
406     return false;
407 
408   Size = BitSize / 8;
409   Offset = (unsigned)BitOffset / 8;
410 
411   assert(TRI->getSpillSize(*RC) >= (Offset + Size) && "bad subregister range");
412 
413   if (!MF.getDataLayout().isLittleEndian()) {
414     Offset = TRI->getSpillSize(*RC) - (Offset + Size);
415   }
416   return true;
417 }
418 
419 void TargetInstrInfo::reMaterialize(MachineBasicBlock &MBB,
420                                     MachineBasicBlock::iterator I,
421                                     Register DestReg, unsigned SubIdx,
422                                     const MachineInstr &Orig,
423                                     const TargetRegisterInfo &TRI) const {
424   MachineInstr *MI = MBB.getParent()->CloneMachineInstr(&Orig);
425   MI->substituteRegister(MI->getOperand(0).getReg(), DestReg, SubIdx, TRI);
426   MBB.insert(I, MI);
427 }
428 
429 bool TargetInstrInfo::produceSameValue(const MachineInstr &MI0,
430                                        const MachineInstr &MI1,
431                                        const MachineRegisterInfo *MRI) const {
432   return MI0.isIdenticalTo(MI1, MachineInstr::IgnoreVRegDefs);
433 }
434 
435 MachineInstr &TargetInstrInfo::duplicate(MachineBasicBlock &MBB,
436     MachineBasicBlock::iterator InsertBefore, const MachineInstr &Orig) const {
437   assert(!Orig.isNotDuplicable() && "Instruction cannot be duplicated");
438   MachineFunction &MF = *MBB.getParent();
439   return MF.cloneMachineInstrBundle(MBB, InsertBefore, Orig);
440 }
441 
442 // If the COPY instruction in MI can be folded to a stack operation, return
443 // the register class to use.
444 static const TargetRegisterClass *canFoldCopy(const MachineInstr &MI,
445                                               unsigned FoldIdx) {
446   assert(MI.isCopy() && "MI must be a COPY instruction");
447   if (MI.getNumOperands() != 2)
448     return nullptr;
449   assert(FoldIdx<2 && "FoldIdx refers no nonexistent operand");
450 
451   const MachineOperand &FoldOp = MI.getOperand(FoldIdx);
452   const MachineOperand &LiveOp = MI.getOperand(1 - FoldIdx);
453 
454   if (FoldOp.getSubReg() || LiveOp.getSubReg())
455     return nullptr;
456 
457   Register FoldReg = FoldOp.getReg();
458   Register LiveReg = LiveOp.getReg();
459 
460   assert(Register::isVirtualRegister(FoldReg) && "Cannot fold physregs");
461 
462   const MachineRegisterInfo &MRI = MI.getMF()->getRegInfo();
463   const TargetRegisterClass *RC = MRI.getRegClass(FoldReg);
464 
465   if (Register::isPhysicalRegister(LiveOp.getReg()))
466     return RC->contains(LiveOp.getReg()) ? RC : nullptr;
467 
468   if (RC->hasSubClassEq(MRI.getRegClass(LiveReg)))
469     return RC;
470 
471   // FIXME: Allow folding when register classes are memory compatible.
472   return nullptr;
473 }
474 
475 MCInst TargetInstrInfo::getNop() const { llvm_unreachable("Not implemented"); }
476 
477 std::pair<unsigned, unsigned>
478 TargetInstrInfo::getPatchpointUnfoldableRange(const MachineInstr &MI) const {
479   switch (MI.getOpcode()) {
480   case TargetOpcode::STACKMAP:
481     // StackMapLiveValues are foldable
482     return std::make_pair(0, StackMapOpers(&MI).getVarIdx());
483   case TargetOpcode::PATCHPOINT:
484     // For PatchPoint, the call args are not foldable (even if reported in the
485     // stackmap e.g. via anyregcc).
486     return std::make_pair(0, PatchPointOpers(&MI).getVarIdx());
487   case TargetOpcode::STATEPOINT:
488     // For statepoints, fold deopt and gc arguments, but not call arguments.
489     return std::make_pair(MI.getNumDefs(), StatepointOpers(&MI).getVarIdx());
490   default:
491     llvm_unreachable("unexpected stackmap opcode");
492   }
493 }
494 
495 static MachineInstr *foldPatchpoint(MachineFunction &MF, MachineInstr &MI,
496                                     ArrayRef<unsigned> Ops, int FrameIndex,
497                                     const TargetInstrInfo &TII) {
498   unsigned StartIdx = 0;
499   unsigned NumDefs = 0;
500   // getPatchpointUnfoldableRange throws guarantee if MI is not a patchpoint.
501   std::tie(NumDefs, StartIdx) = TII.getPatchpointUnfoldableRange(MI);
502 
503   unsigned DefToFoldIdx = MI.getNumOperands();
504 
505   // Return false if any operands requested for folding are not foldable (not
506   // part of the stackmap's live values).
507   for (unsigned Op : Ops) {
508     if (Op < NumDefs) {
509       assert(DefToFoldIdx == MI.getNumOperands() && "Folding multiple defs");
510       DefToFoldIdx = Op;
511     } else if (Op < StartIdx) {
512       return nullptr;
513     }
514     if (MI.getOperand(Op).isTied())
515       return nullptr;
516   }
517 
518   MachineInstr *NewMI =
519       MF.CreateMachineInstr(TII.get(MI.getOpcode()), MI.getDebugLoc(), true);
520   MachineInstrBuilder MIB(MF, NewMI);
521 
522   // No need to fold return, the meta data, and function arguments
523   for (unsigned i = 0; i < StartIdx; ++i)
524     if (i != DefToFoldIdx)
525       MIB.add(MI.getOperand(i));
526 
527   for (unsigned i = StartIdx, e = MI.getNumOperands(); i < e; ++i) {
528     MachineOperand &MO = MI.getOperand(i);
529     unsigned TiedTo = e;
530     (void)MI.isRegTiedToDefOperand(i, &TiedTo);
531 
532     if (is_contained(Ops, i)) {
533       assert(TiedTo == e && "Cannot fold tied operands");
534       unsigned SpillSize;
535       unsigned SpillOffset;
536       // Compute the spill slot size and offset.
537       const TargetRegisterClass *RC =
538         MF.getRegInfo().getRegClass(MO.getReg());
539       bool Valid =
540           TII.getStackSlotRange(RC, MO.getSubReg(), SpillSize, SpillOffset, MF);
541       if (!Valid)
542         report_fatal_error("cannot spill patchpoint subregister operand");
543       MIB.addImm(StackMaps::IndirectMemRefOp);
544       MIB.addImm(SpillSize);
545       MIB.addFrameIndex(FrameIndex);
546       MIB.addImm(SpillOffset);
547     } else {
548       MIB.add(MO);
549       if (TiedTo < e) {
550         assert(TiedTo < NumDefs && "Bad tied operand");
551         if (TiedTo > DefToFoldIdx)
552           --TiedTo;
553         NewMI->tieOperands(TiedTo, NewMI->getNumOperands() - 1);
554       }
555     }
556   }
557   return NewMI;
558 }
559 
560 MachineInstr *TargetInstrInfo::foldMemoryOperand(MachineInstr &MI,
561                                                  ArrayRef<unsigned> Ops, int FI,
562                                                  LiveIntervals *LIS,
563                                                  VirtRegMap *VRM) const {
564   auto Flags = MachineMemOperand::MONone;
565   for (unsigned OpIdx : Ops)
566     Flags |= MI.getOperand(OpIdx).isDef() ? MachineMemOperand::MOStore
567                                           : MachineMemOperand::MOLoad;
568 
569   MachineBasicBlock *MBB = MI.getParent();
570   assert(MBB && "foldMemoryOperand needs an inserted instruction");
571   MachineFunction &MF = *MBB->getParent();
572 
573   // If we're not folding a load into a subreg, the size of the load is the
574   // size of the spill slot. But if we are, we need to figure out what the
575   // actual load size is.
576   int64_t MemSize = 0;
577   const MachineFrameInfo &MFI = MF.getFrameInfo();
578   const TargetRegisterInfo *TRI = MF.getSubtarget().getRegisterInfo();
579 
580   if (Flags & MachineMemOperand::MOStore) {
581     MemSize = MFI.getObjectSize(FI);
582   } else {
583     for (unsigned OpIdx : Ops) {
584       int64_t OpSize = MFI.getObjectSize(FI);
585 
586       if (auto SubReg = MI.getOperand(OpIdx).getSubReg()) {
587         unsigned SubRegSize = TRI->getSubRegIdxSize(SubReg);
588         if (SubRegSize > 0 && !(SubRegSize % 8))
589           OpSize = SubRegSize / 8;
590       }
591 
592       MemSize = std::max(MemSize, OpSize);
593     }
594   }
595 
596   assert(MemSize && "Did not expect a zero-sized stack slot");
597 
598   MachineInstr *NewMI = nullptr;
599 
600   if (MI.getOpcode() == TargetOpcode::STACKMAP ||
601       MI.getOpcode() == TargetOpcode::PATCHPOINT ||
602       MI.getOpcode() == TargetOpcode::STATEPOINT) {
603     // Fold stackmap/patchpoint.
604     NewMI = foldPatchpoint(MF, MI, Ops, FI, *this);
605     if (NewMI)
606       MBB->insert(MI, NewMI);
607   } else {
608     // Ask the target to do the actual folding.
609     NewMI = foldMemoryOperandImpl(MF, MI, Ops, MI, FI, LIS, VRM);
610   }
611 
612   if (NewMI) {
613     NewMI->setMemRefs(MF, MI.memoperands());
614     // Add a memory operand, foldMemoryOperandImpl doesn't do that.
615     assert((!(Flags & MachineMemOperand::MOStore) ||
616             NewMI->mayStore()) &&
617            "Folded a def to a non-store!");
618     assert((!(Flags & MachineMemOperand::MOLoad) ||
619             NewMI->mayLoad()) &&
620            "Folded a use to a non-load!");
621     assert(MFI.getObjectOffset(FI) != -1);
622     MachineMemOperand *MMO =
623         MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(MF, FI),
624                                 Flags, MemSize, MFI.getObjectAlign(FI));
625     NewMI->addMemOperand(MF, MMO);
626 
627     // The pass "x86 speculative load hardening" always attaches symbols to
628     // call instructions. We need copy it form old instruction.
629     NewMI->cloneInstrSymbols(MF, MI);
630 
631     return NewMI;
632   }
633 
634   // Straight COPY may fold as load/store.
635   if (!MI.isCopy() || Ops.size() != 1)
636     return nullptr;
637 
638   const TargetRegisterClass *RC = canFoldCopy(MI, Ops[0]);
639   if (!RC)
640     return nullptr;
641 
642   const MachineOperand &MO = MI.getOperand(1 - Ops[0]);
643   MachineBasicBlock::iterator Pos = MI;
644 
645   if (Flags == MachineMemOperand::MOStore)
646     storeRegToStackSlot(*MBB, Pos, MO.getReg(), MO.isKill(), FI, RC, TRI);
647   else
648     loadRegFromStackSlot(*MBB, Pos, MO.getReg(), FI, RC, TRI);
649   return &*--Pos;
650 }
651 
652 MachineInstr *TargetInstrInfo::foldMemoryOperand(MachineInstr &MI,
653                                                  ArrayRef<unsigned> Ops,
654                                                  MachineInstr &LoadMI,
655                                                  LiveIntervals *LIS) const {
656   assert(LoadMI.canFoldAsLoad() && "LoadMI isn't foldable!");
657 #ifndef NDEBUG
658   for (unsigned OpIdx : Ops)
659     assert(MI.getOperand(OpIdx).isUse() && "Folding load into def!");
660 #endif
661 
662   MachineBasicBlock &MBB = *MI.getParent();
663   MachineFunction &MF = *MBB.getParent();
664 
665   // Ask the target to do the actual folding.
666   MachineInstr *NewMI = nullptr;
667   int FrameIndex = 0;
668 
669   if ((MI.getOpcode() == TargetOpcode::STACKMAP ||
670        MI.getOpcode() == TargetOpcode::PATCHPOINT ||
671        MI.getOpcode() == TargetOpcode::STATEPOINT) &&
672       isLoadFromStackSlot(LoadMI, FrameIndex)) {
673     // Fold stackmap/patchpoint.
674     NewMI = foldPatchpoint(MF, MI, Ops, FrameIndex, *this);
675     if (NewMI)
676       NewMI = &*MBB.insert(MI, NewMI);
677   } else {
678     // Ask the target to do the actual folding.
679     NewMI = foldMemoryOperandImpl(MF, MI, Ops, MI, LoadMI, LIS);
680   }
681 
682   if (!NewMI)
683     return nullptr;
684 
685   // Copy the memoperands from the load to the folded instruction.
686   if (MI.memoperands_empty()) {
687     NewMI->setMemRefs(MF, LoadMI.memoperands());
688   } else {
689     // Handle the rare case of folding multiple loads.
690     NewMI->setMemRefs(MF, MI.memoperands());
691     for (MachineInstr::mmo_iterator I = LoadMI.memoperands_begin(),
692                                     E = LoadMI.memoperands_end();
693          I != E; ++I) {
694       NewMI->addMemOperand(MF, *I);
695     }
696   }
697   return NewMI;
698 }
699 
700 bool TargetInstrInfo::hasReassociableOperands(
701     const MachineInstr &Inst, const MachineBasicBlock *MBB) const {
702   const MachineOperand &Op1 = Inst.getOperand(1);
703   const MachineOperand &Op2 = Inst.getOperand(2);
704   const MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo();
705 
706   // We need virtual register definitions for the operands that we will
707   // reassociate.
708   MachineInstr *MI1 = nullptr;
709   MachineInstr *MI2 = nullptr;
710   if (Op1.isReg() && Register::isVirtualRegister(Op1.getReg()))
711     MI1 = MRI.getUniqueVRegDef(Op1.getReg());
712   if (Op2.isReg() && Register::isVirtualRegister(Op2.getReg()))
713     MI2 = MRI.getUniqueVRegDef(Op2.getReg());
714 
715   // And they need to be in the trace (otherwise, they won't have a depth).
716   return MI1 && MI2 && MI1->getParent() == MBB && MI2->getParent() == MBB;
717 }
718 
719 bool TargetInstrInfo::hasReassociableSibling(const MachineInstr &Inst,
720                                              bool &Commuted) const {
721   const MachineBasicBlock *MBB = Inst.getParent();
722   const MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo();
723   MachineInstr *MI1 = MRI.getUniqueVRegDef(Inst.getOperand(1).getReg());
724   MachineInstr *MI2 = MRI.getUniqueVRegDef(Inst.getOperand(2).getReg());
725   unsigned AssocOpcode = Inst.getOpcode();
726 
727   // If only one operand has the same opcode and it's the second source operand,
728   // the operands must be commuted.
729   Commuted = MI1->getOpcode() != AssocOpcode && MI2->getOpcode() == AssocOpcode;
730   if (Commuted)
731     std::swap(MI1, MI2);
732 
733   // 1. The previous instruction must be the same type as Inst.
734   // 2. The previous instruction must also be associative/commutative (this can
735   //    be different even for instructions with the same opcode if traits like
736   //    fast-math-flags are included).
737   // 3. The previous instruction must have virtual register definitions for its
738   //    operands in the same basic block as Inst.
739   // 4. The previous instruction's result must only be used by Inst.
740   return MI1->getOpcode() == AssocOpcode && isAssociativeAndCommutative(*MI1) &&
741          hasReassociableOperands(*MI1, MBB) &&
742          MRI.hasOneNonDBGUse(MI1->getOperand(0).getReg());
743 }
744 
745 // 1. The operation must be associative and commutative.
746 // 2. The instruction must have virtual register definitions for its
747 //    operands in the same basic block.
748 // 3. The instruction must have a reassociable sibling.
749 bool TargetInstrInfo::isReassociationCandidate(const MachineInstr &Inst,
750                                                bool &Commuted) const {
751   return isAssociativeAndCommutative(Inst) &&
752          hasReassociableOperands(Inst, Inst.getParent()) &&
753          hasReassociableSibling(Inst, Commuted);
754 }
755 
756 // The concept of the reassociation pass is that these operations can benefit
757 // from this kind of transformation:
758 //
759 // A = ? op ?
760 // B = A op X (Prev)
761 // C = B op Y (Root)
762 // -->
763 // A = ? op ?
764 // B = X op Y
765 // C = A op B
766 //
767 // breaking the dependency between A and B, allowing them to be executed in
768 // parallel (or back-to-back in a pipeline) instead of depending on each other.
769 
770 // FIXME: This has the potential to be expensive (compile time) while not
771 // improving the code at all. Some ways to limit the overhead:
772 // 1. Track successful transforms; bail out if hit rate gets too low.
773 // 2. Only enable at -O3 or some other non-default optimization level.
774 // 3. Pre-screen pattern candidates here: if an operand of the previous
775 //    instruction is known to not increase the critical path, then don't match
776 //    that pattern.
777 bool TargetInstrInfo::getMachineCombinerPatterns(
778     MachineInstr &Root, SmallVectorImpl<MachineCombinerPattern> &Patterns,
779     bool DoRegPressureReduce) const {
780   bool Commute;
781   if (isReassociationCandidate(Root, Commute)) {
782     // We found a sequence of instructions that may be suitable for a
783     // reassociation of operands to increase ILP. Specify each commutation
784     // possibility for the Prev instruction in the sequence and let the
785     // machine combiner decide if changing the operands is worthwhile.
786     if (Commute) {
787       Patterns.push_back(MachineCombinerPattern::REASSOC_AX_YB);
788       Patterns.push_back(MachineCombinerPattern::REASSOC_XA_YB);
789     } else {
790       Patterns.push_back(MachineCombinerPattern::REASSOC_AX_BY);
791       Patterns.push_back(MachineCombinerPattern::REASSOC_XA_BY);
792     }
793     return true;
794   }
795 
796   return false;
797 }
798 
799 /// Return true when a code sequence can improve loop throughput.
800 bool
801 TargetInstrInfo::isThroughputPattern(MachineCombinerPattern Pattern) const {
802   return false;
803 }
804 
805 /// Attempt the reassociation transformation to reduce critical path length.
806 /// See the above comments before getMachineCombinerPatterns().
807 void TargetInstrInfo::reassociateOps(
808     MachineInstr &Root, MachineInstr &Prev,
809     MachineCombinerPattern Pattern,
810     SmallVectorImpl<MachineInstr *> &InsInstrs,
811     SmallVectorImpl<MachineInstr *> &DelInstrs,
812     DenseMap<unsigned, unsigned> &InstrIdxForVirtReg) const {
813   MachineFunction *MF = Root.getMF();
814   MachineRegisterInfo &MRI = MF->getRegInfo();
815   const TargetInstrInfo *TII = MF->getSubtarget().getInstrInfo();
816   const TargetRegisterInfo *TRI = MF->getSubtarget().getRegisterInfo();
817   const TargetRegisterClass *RC = Root.getRegClassConstraint(0, TII, TRI);
818 
819   // This array encodes the operand index for each parameter because the
820   // operands may be commuted. Each row corresponds to a pattern value,
821   // and each column specifies the index of A, B, X, Y.
822   unsigned OpIdx[4][4] = {
823     { 1, 1, 2, 2 },
824     { 1, 2, 2, 1 },
825     { 2, 1, 1, 2 },
826     { 2, 2, 1, 1 }
827   };
828 
829   int Row;
830   switch (Pattern) {
831   case MachineCombinerPattern::REASSOC_AX_BY: Row = 0; break;
832   case MachineCombinerPattern::REASSOC_AX_YB: Row = 1; break;
833   case MachineCombinerPattern::REASSOC_XA_BY: Row = 2; break;
834   case MachineCombinerPattern::REASSOC_XA_YB: Row = 3; break;
835   default: llvm_unreachable("unexpected MachineCombinerPattern");
836   }
837 
838   MachineOperand &OpA = Prev.getOperand(OpIdx[Row][0]);
839   MachineOperand &OpB = Root.getOperand(OpIdx[Row][1]);
840   MachineOperand &OpX = Prev.getOperand(OpIdx[Row][2]);
841   MachineOperand &OpY = Root.getOperand(OpIdx[Row][3]);
842   MachineOperand &OpC = Root.getOperand(0);
843 
844   Register RegA = OpA.getReg();
845   Register RegB = OpB.getReg();
846   Register RegX = OpX.getReg();
847   Register RegY = OpY.getReg();
848   Register RegC = OpC.getReg();
849 
850   if (Register::isVirtualRegister(RegA))
851     MRI.constrainRegClass(RegA, RC);
852   if (Register::isVirtualRegister(RegB))
853     MRI.constrainRegClass(RegB, RC);
854   if (Register::isVirtualRegister(RegX))
855     MRI.constrainRegClass(RegX, RC);
856   if (Register::isVirtualRegister(RegY))
857     MRI.constrainRegClass(RegY, RC);
858   if (Register::isVirtualRegister(RegC))
859     MRI.constrainRegClass(RegC, RC);
860 
861   // Create a new virtual register for the result of (X op Y) instead of
862   // recycling RegB because the MachineCombiner's computation of the critical
863   // path requires a new register definition rather than an existing one.
864   Register NewVR = MRI.createVirtualRegister(RC);
865   InstrIdxForVirtReg.insert(std::make_pair(NewVR, 0));
866 
867   unsigned Opcode = Root.getOpcode();
868   bool KillA = OpA.isKill();
869   bool KillX = OpX.isKill();
870   bool KillY = OpY.isKill();
871 
872   // Create new instructions for insertion.
873   MachineInstrBuilder MIB1 =
874       BuildMI(*MF, Prev.getDebugLoc(), TII->get(Opcode), NewVR)
875           .addReg(RegX, getKillRegState(KillX))
876           .addReg(RegY, getKillRegState(KillY));
877   MachineInstrBuilder MIB2 =
878       BuildMI(*MF, Root.getDebugLoc(), TII->get(Opcode), RegC)
879           .addReg(RegA, getKillRegState(KillA))
880           .addReg(NewVR, getKillRegState(true));
881 
882   setSpecialOperandAttr(Root, Prev, *MIB1, *MIB2);
883 
884   // Record new instructions for insertion and old instructions for deletion.
885   InsInstrs.push_back(MIB1);
886   InsInstrs.push_back(MIB2);
887   DelInstrs.push_back(&Prev);
888   DelInstrs.push_back(&Root);
889 }
890 
891 void TargetInstrInfo::genAlternativeCodeSequence(
892     MachineInstr &Root, MachineCombinerPattern Pattern,
893     SmallVectorImpl<MachineInstr *> &InsInstrs,
894     SmallVectorImpl<MachineInstr *> &DelInstrs,
895     DenseMap<unsigned, unsigned> &InstIdxForVirtReg) const {
896   MachineRegisterInfo &MRI = Root.getMF()->getRegInfo();
897 
898   // Select the previous instruction in the sequence based on the input pattern.
899   MachineInstr *Prev = nullptr;
900   switch (Pattern) {
901   case MachineCombinerPattern::REASSOC_AX_BY:
902   case MachineCombinerPattern::REASSOC_XA_BY:
903     Prev = MRI.getUniqueVRegDef(Root.getOperand(1).getReg());
904     break;
905   case MachineCombinerPattern::REASSOC_AX_YB:
906   case MachineCombinerPattern::REASSOC_XA_YB:
907     Prev = MRI.getUniqueVRegDef(Root.getOperand(2).getReg());
908     break;
909   default:
910     break;
911   }
912 
913   assert(Prev && "Unknown pattern for machine combiner");
914 
915   reassociateOps(Root, *Prev, Pattern, InsInstrs, DelInstrs, InstIdxForVirtReg);
916 }
917 
918 bool TargetInstrInfo::isReallyTriviallyReMaterializableGeneric(
919     const MachineInstr &MI, AAResults *AA) const {
920   const MachineFunction &MF = *MI.getMF();
921   const MachineRegisterInfo &MRI = MF.getRegInfo();
922 
923   // Remat clients assume operand 0 is the defined register.
924   if (!MI.getNumOperands() || !MI.getOperand(0).isReg())
925     return false;
926   Register DefReg = MI.getOperand(0).getReg();
927 
928   // A sub-register definition can only be rematerialized if the instruction
929   // doesn't read the other parts of the register.  Otherwise it is really a
930   // read-modify-write operation on the full virtual register which cannot be
931   // moved safely.
932   if (Register::isVirtualRegister(DefReg) && MI.getOperand(0).getSubReg() &&
933       MI.readsVirtualRegister(DefReg))
934     return false;
935 
936   // A load from a fixed stack slot can be rematerialized. This may be
937   // redundant with subsequent checks, but it's target-independent,
938   // simple, and a common case.
939   int FrameIdx = 0;
940   if (isLoadFromStackSlot(MI, FrameIdx) &&
941       MF.getFrameInfo().isImmutableObjectIndex(FrameIdx))
942     return true;
943 
944   // Avoid instructions obviously unsafe for remat.
945   if (MI.isNotDuplicable() || MI.mayStore() || MI.mayRaiseFPException() ||
946       MI.hasUnmodeledSideEffects())
947     return false;
948 
949   // Don't remat inline asm. We have no idea how expensive it is
950   // even if it's side effect free.
951   if (MI.isInlineAsm())
952     return false;
953 
954   // Avoid instructions which load from potentially varying memory.
955   if (MI.mayLoad() && !MI.isDereferenceableInvariantLoad(AA))
956     return false;
957 
958   // If any of the registers accessed are non-constant, conservatively assume
959   // the instruction is not rematerializable.
960   for (const MachineOperand &MO : MI.operands()) {
961     if (!MO.isReg()) continue;
962     Register Reg = MO.getReg();
963     if (Reg == 0)
964       continue;
965 
966     // Check for a well-behaved physical register.
967     if (Register::isPhysicalRegister(Reg)) {
968       if (MO.isUse()) {
969         // If the physreg has no defs anywhere, it's just an ambient register
970         // and we can freely move its uses. Alternatively, if it's allocatable,
971         // it could get allocated to something with a def during allocation.
972         if (!MRI.isConstantPhysReg(Reg))
973           return false;
974       } else {
975         // A physreg def. We can't remat it.
976         return false;
977       }
978       continue;
979     }
980 
981     // Only allow one virtual-register def.  There may be multiple defs of the
982     // same virtual register, though.
983     if (MO.isDef() && Reg != DefReg)
984       return false;
985 
986     // Don't allow any virtual-register uses. Rematting an instruction with
987     // virtual register uses would length the live ranges of the uses, which
988     // is not necessarily a good idea, certainly not "trivial".
989     if (MO.isUse())
990       return false;
991   }
992 
993   // Everything checked out.
994   return true;
995 }
996 
997 int TargetInstrInfo::getSPAdjust(const MachineInstr &MI) const {
998   const MachineFunction *MF = MI.getMF();
999   const TargetFrameLowering *TFI = MF->getSubtarget().getFrameLowering();
1000   bool StackGrowsDown =
1001     TFI->getStackGrowthDirection() == TargetFrameLowering::StackGrowsDown;
1002 
1003   unsigned FrameSetupOpcode = getCallFrameSetupOpcode();
1004   unsigned FrameDestroyOpcode = getCallFrameDestroyOpcode();
1005 
1006   if (!isFrameInstr(MI))
1007     return 0;
1008 
1009   int SPAdj = TFI->alignSPAdjust(getFrameSize(MI));
1010 
1011   if ((!StackGrowsDown && MI.getOpcode() == FrameSetupOpcode) ||
1012       (StackGrowsDown && MI.getOpcode() == FrameDestroyOpcode))
1013     SPAdj = -SPAdj;
1014 
1015   return SPAdj;
1016 }
1017 
1018 /// isSchedulingBoundary - Test if the given instruction should be
1019 /// considered a scheduling boundary. This primarily includes labels
1020 /// and terminators.
1021 bool TargetInstrInfo::isSchedulingBoundary(const MachineInstr &MI,
1022                                            const MachineBasicBlock *MBB,
1023                                            const MachineFunction &MF) const {
1024   // Terminators and labels can't be scheduled around.
1025   if (MI.isTerminator() || MI.isPosition())
1026     return true;
1027 
1028   // INLINEASM_BR can jump to another block
1029   if (MI.getOpcode() == TargetOpcode::INLINEASM_BR)
1030     return true;
1031 
1032   // Don't attempt to schedule around any instruction that defines
1033   // a stack-oriented pointer, as it's unlikely to be profitable. This
1034   // saves compile time, because it doesn't require every single
1035   // stack slot reference to depend on the instruction that does the
1036   // modification.
1037   const TargetLowering &TLI = *MF.getSubtarget().getTargetLowering();
1038   const TargetRegisterInfo *TRI = MF.getSubtarget().getRegisterInfo();
1039   return MI.modifiesRegister(TLI.getStackPointerRegisterToSaveRestore(), TRI);
1040 }
1041 
1042 // Provide a global flag for disabling the PreRA hazard recognizer that targets
1043 // may choose to honor.
1044 bool TargetInstrInfo::usePreRAHazardRecognizer() const {
1045   return !DisableHazardRecognizer;
1046 }
1047 
1048 // Default implementation of CreateTargetRAHazardRecognizer.
1049 ScheduleHazardRecognizer *TargetInstrInfo::
1050 CreateTargetHazardRecognizer(const TargetSubtargetInfo *STI,
1051                              const ScheduleDAG *DAG) const {
1052   // Dummy hazard recognizer allows all instructions to issue.
1053   return new ScheduleHazardRecognizer();
1054 }
1055 
1056 // Default implementation of CreateTargetMIHazardRecognizer.
1057 ScheduleHazardRecognizer *TargetInstrInfo::CreateTargetMIHazardRecognizer(
1058     const InstrItineraryData *II, const ScheduleDAGMI *DAG) const {
1059   return new ScoreboardHazardRecognizer(II, DAG, "machine-scheduler");
1060 }
1061 
1062 // Default implementation of CreateTargetPostRAHazardRecognizer.
1063 ScheduleHazardRecognizer *TargetInstrInfo::
1064 CreateTargetPostRAHazardRecognizer(const InstrItineraryData *II,
1065                                    const ScheduleDAG *DAG) const {
1066   return new ScoreboardHazardRecognizer(II, DAG, "post-RA-sched");
1067 }
1068 
1069 // Default implementation of getMemOperandWithOffset.
1070 bool TargetInstrInfo::getMemOperandWithOffset(
1071     const MachineInstr &MI, const MachineOperand *&BaseOp, int64_t &Offset,
1072     bool &OffsetIsScalable, const TargetRegisterInfo *TRI) const {
1073   SmallVector<const MachineOperand *, 4> BaseOps;
1074   unsigned Width;
1075   if (!getMemOperandsWithOffsetWidth(MI, BaseOps, Offset, OffsetIsScalable,
1076                                      Width, TRI) ||
1077       BaseOps.size() != 1)
1078     return false;
1079   BaseOp = BaseOps.front();
1080   return true;
1081 }
1082 
1083 //===----------------------------------------------------------------------===//
1084 //  SelectionDAG latency interface.
1085 //===----------------------------------------------------------------------===//
1086 
1087 int
1088 TargetInstrInfo::getOperandLatency(const InstrItineraryData *ItinData,
1089                                    SDNode *DefNode, unsigned DefIdx,
1090                                    SDNode *UseNode, unsigned UseIdx) const {
1091   if (!ItinData || ItinData->isEmpty())
1092     return -1;
1093 
1094   if (!DefNode->isMachineOpcode())
1095     return -1;
1096 
1097   unsigned DefClass = get(DefNode->getMachineOpcode()).getSchedClass();
1098   if (!UseNode->isMachineOpcode())
1099     return ItinData->getOperandCycle(DefClass, DefIdx);
1100   unsigned UseClass = get(UseNode->getMachineOpcode()).getSchedClass();
1101   return ItinData->getOperandLatency(DefClass, DefIdx, UseClass, UseIdx);
1102 }
1103 
1104 int TargetInstrInfo::getInstrLatency(const InstrItineraryData *ItinData,
1105                                      SDNode *N) const {
1106   if (!ItinData || ItinData->isEmpty())
1107     return 1;
1108 
1109   if (!N->isMachineOpcode())
1110     return 1;
1111 
1112   return ItinData->getStageLatency(get(N->getMachineOpcode()).getSchedClass());
1113 }
1114 
1115 //===----------------------------------------------------------------------===//
1116 //  MachineInstr latency interface.
1117 //===----------------------------------------------------------------------===//
1118 
1119 unsigned TargetInstrInfo::getNumMicroOps(const InstrItineraryData *ItinData,
1120                                          const MachineInstr &MI) const {
1121   if (!ItinData || ItinData->isEmpty())
1122     return 1;
1123 
1124   unsigned Class = MI.getDesc().getSchedClass();
1125   int UOps = ItinData->Itineraries[Class].NumMicroOps;
1126   if (UOps >= 0)
1127     return UOps;
1128 
1129   // The # of u-ops is dynamically determined. The specific target should
1130   // override this function to return the right number.
1131   return 1;
1132 }
1133 
1134 /// Return the default expected latency for a def based on it's opcode.
1135 unsigned TargetInstrInfo::defaultDefLatency(const MCSchedModel &SchedModel,
1136                                             const MachineInstr &DefMI) const {
1137   if (DefMI.isTransient())
1138     return 0;
1139   if (DefMI.mayLoad())
1140     return SchedModel.LoadLatency;
1141   if (isHighLatencyDef(DefMI.getOpcode()))
1142     return SchedModel.HighLatency;
1143   return 1;
1144 }
1145 
1146 unsigned TargetInstrInfo::getPredicationCost(const MachineInstr &) const {
1147   return 0;
1148 }
1149 
1150 unsigned TargetInstrInfo::getInstrLatency(const InstrItineraryData *ItinData,
1151                                           const MachineInstr &MI,
1152                                           unsigned *PredCost) const {
1153   // Default to one cycle for no itinerary. However, an "empty" itinerary may
1154   // still have a MinLatency property, which getStageLatency checks.
1155   if (!ItinData)
1156     return MI.mayLoad() ? 2 : 1;
1157 
1158   return ItinData->getStageLatency(MI.getDesc().getSchedClass());
1159 }
1160 
1161 bool TargetInstrInfo::hasLowDefLatency(const TargetSchedModel &SchedModel,
1162                                        const MachineInstr &DefMI,
1163                                        unsigned DefIdx) const {
1164   const InstrItineraryData *ItinData = SchedModel.getInstrItineraries();
1165   if (!ItinData || ItinData->isEmpty())
1166     return false;
1167 
1168   unsigned DefClass = DefMI.getDesc().getSchedClass();
1169   int DefCycle = ItinData->getOperandCycle(DefClass, DefIdx);
1170   return (DefCycle != -1 && DefCycle <= 1);
1171 }
1172 
1173 Optional<ParamLoadedValue>
1174 TargetInstrInfo::describeLoadedValue(const MachineInstr &MI,
1175                                      Register Reg) const {
1176   const MachineFunction *MF = MI.getMF();
1177   const TargetRegisterInfo *TRI = MF->getSubtarget().getRegisterInfo();
1178   DIExpression *Expr = DIExpression::get(MF->getFunction().getContext(), {});
1179   int64_t Offset;
1180   bool OffsetIsScalable;
1181 
1182   // To simplify the sub-register handling, verify that we only need to
1183   // consider physical registers.
1184   assert(MF->getProperties().hasProperty(
1185       MachineFunctionProperties::Property::NoVRegs));
1186 
1187   if (auto DestSrc = isCopyInstr(MI)) {
1188     Register DestReg = DestSrc->Destination->getReg();
1189 
1190     // If the copy destination is the forwarding reg, describe the forwarding
1191     // reg using the copy source as the backup location. Example:
1192     //
1193     //   x0 = MOV x7
1194     //   call callee(x0)      ; x0 described as x7
1195     if (Reg == DestReg)
1196       return ParamLoadedValue(*DestSrc->Source, Expr);
1197 
1198     // Cases where super- or sub-registers needs to be described should
1199     // be handled by the target's hook implementation.
1200     assert(!TRI->isSuperOrSubRegisterEq(Reg, DestReg) &&
1201            "TargetInstrInfo::describeLoadedValue can't describe super- or "
1202            "sub-regs for copy instructions");
1203     return None;
1204   } else if (auto RegImm = isAddImmediate(MI, Reg)) {
1205     Register SrcReg = RegImm->Reg;
1206     Offset = RegImm->Imm;
1207     Expr = DIExpression::prepend(Expr, DIExpression::ApplyOffset, Offset);
1208     return ParamLoadedValue(MachineOperand::CreateReg(SrcReg, false), Expr);
1209   } else if (MI.hasOneMemOperand()) {
1210     // Only describe memory which provably does not escape the function. As
1211     // described in llvm.org/PR43343, escaped memory may be clobbered by the
1212     // callee (or by another thread).
1213     const auto &TII = MF->getSubtarget().getInstrInfo();
1214     const MachineFrameInfo &MFI = MF->getFrameInfo();
1215     const MachineMemOperand *MMO = MI.memoperands()[0];
1216     const PseudoSourceValue *PSV = MMO->getPseudoValue();
1217 
1218     // If the address points to "special" memory (e.g. a spill slot), it's
1219     // sufficient to check that it isn't aliased by any high-level IR value.
1220     if (!PSV || PSV->mayAlias(&MFI))
1221       return None;
1222 
1223     const MachineOperand *BaseOp;
1224     if (!TII->getMemOperandWithOffset(MI, BaseOp, Offset, OffsetIsScalable,
1225                                       TRI))
1226       return None;
1227 
1228     // FIXME: Scalable offsets are not yet handled in the offset code below.
1229     if (OffsetIsScalable)
1230       return None;
1231 
1232     // TODO: Can currently only handle mem instructions with a single define.
1233     // An example from the x86 target:
1234     //    ...
1235     //    DIV64m $rsp, 1, $noreg, 24, $noreg, implicit-def dead $rax, implicit-def $rdx
1236     //    ...
1237     //
1238     if (MI.getNumExplicitDefs() != 1)
1239       return None;
1240 
1241     // TODO: In what way do we need to take Reg into consideration here?
1242 
1243     SmallVector<uint64_t, 8> Ops;
1244     DIExpression::appendOffset(Ops, Offset);
1245     Ops.push_back(dwarf::DW_OP_deref_size);
1246     Ops.push_back(MMO->getSize());
1247     Expr = DIExpression::prependOpcodes(Expr, Ops);
1248     return ParamLoadedValue(*BaseOp, Expr);
1249   }
1250 
1251   return None;
1252 }
1253 
1254 /// Both DefMI and UseMI must be valid.  By default, call directly to the
1255 /// itinerary. This may be overriden by the target.
1256 int TargetInstrInfo::getOperandLatency(const InstrItineraryData *ItinData,
1257                                        const MachineInstr &DefMI,
1258                                        unsigned DefIdx,
1259                                        const MachineInstr &UseMI,
1260                                        unsigned UseIdx) const {
1261   unsigned DefClass = DefMI.getDesc().getSchedClass();
1262   unsigned UseClass = UseMI.getDesc().getSchedClass();
1263   return ItinData->getOperandLatency(DefClass, DefIdx, UseClass, UseIdx);
1264 }
1265 
1266 bool TargetInstrInfo::getRegSequenceInputs(
1267     const MachineInstr &MI, unsigned DefIdx,
1268     SmallVectorImpl<RegSubRegPairAndIdx> &InputRegs) const {
1269   assert((MI.isRegSequence() ||
1270           MI.isRegSequenceLike()) && "Instruction do not have the proper type");
1271 
1272   if (!MI.isRegSequence())
1273     return getRegSequenceLikeInputs(MI, DefIdx, InputRegs);
1274 
1275   // We are looking at:
1276   // Def = REG_SEQUENCE v0, sub0, v1, sub1, ...
1277   assert(DefIdx == 0 && "REG_SEQUENCE only has one def");
1278   for (unsigned OpIdx = 1, EndOpIdx = MI.getNumOperands(); OpIdx != EndOpIdx;
1279        OpIdx += 2) {
1280     const MachineOperand &MOReg = MI.getOperand(OpIdx);
1281     if (MOReg.isUndef())
1282       continue;
1283     const MachineOperand &MOSubIdx = MI.getOperand(OpIdx + 1);
1284     assert(MOSubIdx.isImm() &&
1285            "One of the subindex of the reg_sequence is not an immediate");
1286     // Record Reg:SubReg, SubIdx.
1287     InputRegs.push_back(RegSubRegPairAndIdx(MOReg.getReg(), MOReg.getSubReg(),
1288                                             (unsigned)MOSubIdx.getImm()));
1289   }
1290   return true;
1291 }
1292 
1293 bool TargetInstrInfo::getExtractSubregInputs(
1294     const MachineInstr &MI, unsigned DefIdx,
1295     RegSubRegPairAndIdx &InputReg) const {
1296   assert((MI.isExtractSubreg() ||
1297       MI.isExtractSubregLike()) && "Instruction do not have the proper type");
1298 
1299   if (!MI.isExtractSubreg())
1300     return getExtractSubregLikeInputs(MI, DefIdx, InputReg);
1301 
1302   // We are looking at:
1303   // Def = EXTRACT_SUBREG v0.sub1, sub0.
1304   assert(DefIdx == 0 && "EXTRACT_SUBREG only has one def");
1305   const MachineOperand &MOReg = MI.getOperand(1);
1306   if (MOReg.isUndef())
1307     return false;
1308   const MachineOperand &MOSubIdx = MI.getOperand(2);
1309   assert(MOSubIdx.isImm() &&
1310          "The subindex of the extract_subreg is not an immediate");
1311 
1312   InputReg.Reg = MOReg.getReg();
1313   InputReg.SubReg = MOReg.getSubReg();
1314   InputReg.SubIdx = (unsigned)MOSubIdx.getImm();
1315   return true;
1316 }
1317 
1318 bool TargetInstrInfo::getInsertSubregInputs(
1319     const MachineInstr &MI, unsigned DefIdx,
1320     RegSubRegPair &BaseReg, RegSubRegPairAndIdx &InsertedReg) const {
1321   assert((MI.isInsertSubreg() ||
1322       MI.isInsertSubregLike()) && "Instruction do not have the proper type");
1323 
1324   if (!MI.isInsertSubreg())
1325     return getInsertSubregLikeInputs(MI, DefIdx, BaseReg, InsertedReg);
1326 
1327   // We are looking at:
1328   // Def = INSERT_SEQUENCE v0, v1, sub0.
1329   assert(DefIdx == 0 && "INSERT_SUBREG only has one def");
1330   const MachineOperand &MOBaseReg = MI.getOperand(1);
1331   const MachineOperand &MOInsertedReg = MI.getOperand(2);
1332   if (MOInsertedReg.isUndef())
1333     return false;
1334   const MachineOperand &MOSubIdx = MI.getOperand(3);
1335   assert(MOSubIdx.isImm() &&
1336          "One of the subindex of the reg_sequence is not an immediate");
1337   BaseReg.Reg = MOBaseReg.getReg();
1338   BaseReg.SubReg = MOBaseReg.getSubReg();
1339 
1340   InsertedReg.Reg = MOInsertedReg.getReg();
1341   InsertedReg.SubReg = MOInsertedReg.getSubReg();
1342   InsertedReg.SubIdx = (unsigned)MOSubIdx.getImm();
1343   return true;
1344 }
1345 
1346 // Returns a MIRPrinter comment for this machine operand.
1347 std::string TargetInstrInfo::createMIROperandComment(
1348     const MachineInstr &MI, const MachineOperand &Op, unsigned OpIdx,
1349     const TargetRegisterInfo *TRI) const {
1350 
1351   if (!MI.isInlineAsm())
1352     return "";
1353 
1354   std::string Flags;
1355   raw_string_ostream OS(Flags);
1356 
1357   if (OpIdx == InlineAsm::MIOp_ExtraInfo) {
1358     // Print HasSideEffects, MayLoad, MayStore, IsAlignStack
1359     unsigned ExtraInfo = Op.getImm();
1360     bool First = true;
1361     for (StringRef Info : InlineAsm::getExtraInfoNames(ExtraInfo)) {
1362       if (!First)
1363         OS << " ";
1364       First = false;
1365       OS << Info;
1366     }
1367 
1368     return OS.str();
1369   }
1370 
1371   int FlagIdx = MI.findInlineAsmFlagIdx(OpIdx);
1372   if (FlagIdx < 0 || (unsigned)FlagIdx != OpIdx)
1373     return "";
1374 
1375   assert(Op.isImm() && "Expected flag operand to be an immediate");
1376   // Pretty print the inline asm operand descriptor.
1377   unsigned Flag = Op.getImm();
1378   unsigned Kind = InlineAsm::getKind(Flag);
1379   OS << InlineAsm::getKindName(Kind);
1380 
1381   unsigned RCID = 0;
1382   if (!InlineAsm::isImmKind(Flag) && !InlineAsm::isMemKind(Flag) &&
1383       InlineAsm::hasRegClassConstraint(Flag, RCID)) {
1384     if (TRI) {
1385       OS << ':' << TRI->getRegClassName(TRI->getRegClass(RCID));
1386     } else
1387       OS << ":RC" << RCID;
1388   }
1389 
1390   if (InlineAsm::isMemKind(Flag)) {
1391     unsigned MCID = InlineAsm::getMemoryConstraintID(Flag);
1392     OS << ":" << InlineAsm::getMemConstraintName(MCID);
1393   }
1394 
1395   unsigned TiedTo = 0;
1396   if (InlineAsm::isUseOperandTiedToDef(Flag, TiedTo))
1397     OS << " tiedto:$" << TiedTo;
1398 
1399   return OS.str();
1400 }
1401 
1402 TargetInstrInfo::PipelinerLoopInfo::~PipelinerLoopInfo() {}
1403 
1404 void TargetInstrInfo::mergeOutliningCandidateAttributes(
1405     Function &F, std::vector<outliner::Candidate> &Candidates) const {
1406   // Include target features from an arbitrary candidate for the outlined
1407   // function. This makes sure the outlined function knows what kinds of
1408   // instructions are going into it. This is fine, since all parent functions
1409   // must necessarily support the instructions that are in the outlined region.
1410   outliner::Candidate &FirstCand = Candidates.front();
1411   const Function &ParentFn = FirstCand.getMF()->getFunction();
1412   if (ParentFn.hasFnAttribute("target-features"))
1413     F.addFnAttr(ParentFn.getFnAttribute("target-features"));
1414 
1415   // Set nounwind, so we don't generate eh_frame.
1416   if (llvm::all_of(Candidates, [](const outliner::Candidate &C) {
1417         return C.getMF()->getFunction().hasFnAttribute(Attribute::NoUnwind);
1418       }))
1419     F.addFnAttr(Attribute::NoUnwind);
1420 }
1421 
1422 bool TargetInstrInfo::isMBBSafeToOutlineFrom(MachineBasicBlock &MBB,
1423                                              unsigned &Flags) const {
1424   // Some instrumentations create special TargetOpcode at the start which
1425   // expands to special code sequences which must be present.
1426   auto First = MBB.getFirstNonDebugInstr();
1427   if (First != MBB.end() &&
1428       (First->getOpcode() == TargetOpcode::FENTRY_CALL ||
1429        First->getOpcode() == TargetOpcode::PATCHABLE_FUNCTION_ENTER))
1430     return false;
1431 
1432   return true;
1433 }
1434