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/IR/DataLayout.h"
23 #include "llvm/MC/MCAsmInfo.h"
24 #include "llvm/MC/MCInstrItineraries.h"
25 #include "llvm/Support/CommandLine.h"
26 #include "llvm/Support/ErrorHandling.h"
27 #include "llvm/Support/raw_ostream.h"
28 #include "llvm/Target/TargetFrameLowering.h"
29 #include "llvm/Target/TargetLowering.h"
30 #include "llvm/Target/TargetMachine.h"
31 #include "llvm/Target/TargetRegisterInfo.h"
32 #include <cctype>
33 using namespace llvm;
34 
35 static cl::opt<bool> DisableHazardRecognizer(
36   "disable-sched-hazard", cl::Hidden, cl::init(false),
37   cl::desc("Disable hazard detection during preRA scheduling"));
38 
~TargetInstrInfo()39 TargetInstrInfo::~TargetInstrInfo() {
40 }
41 
42 const TargetRegisterClass*
getRegClass(const MCInstrDesc & MCID,unsigned OpNum,const TargetRegisterInfo * TRI,const MachineFunction & MF) const43 TargetInstrInfo::getRegClass(const MCInstrDesc &MCID, unsigned OpNum,
44                              const TargetRegisterInfo *TRI,
45                              const MachineFunction &MF) const {
46   if (OpNum >= MCID.getNumOperands())
47     return nullptr;
48 
49   short RegClass = MCID.OpInfo[OpNum].RegClass;
50   if (MCID.OpInfo[OpNum].isLookupPtrRegClass())
51     return TRI->getPointerRegClass(MF, RegClass);
52 
53   // Instructions like INSERT_SUBREG do not have fixed register classes.
54   if (RegClass < 0)
55     return nullptr;
56 
57   // Otherwise just look it up normally.
58   return TRI->getRegClass(RegClass);
59 }
60 
61 /// insertNoop - Insert a noop into the instruction stream at the specified
62 /// point.
insertNoop(MachineBasicBlock & MBB,MachineBasicBlock::iterator MI) const63 void TargetInstrInfo::insertNoop(MachineBasicBlock &MBB,
64                                  MachineBasicBlock::iterator MI) const {
65   llvm_unreachable("Target didn't implement insertNoop!");
66 }
67 
68 /// Measure the specified inline asm to determine an approximation of its
69 /// length.
70 /// Comments (which run till the next SeparatorString or newline) do not
71 /// count as an instruction.
72 /// Any other non-whitespace text is considered an instruction, with
73 /// multiple instructions separated by SeparatorString or newlines.
74 /// Variable-length instructions are not handled here; this function
75 /// may be overloaded in the target code to do that.
getInlineAsmLength(const char * Str,const MCAsmInfo & MAI) const76 unsigned TargetInstrInfo::getInlineAsmLength(const char *Str,
77                                              const MCAsmInfo &MAI) const {
78 
79 
80   // Count the number of instructions in the asm.
81   bool atInsnStart = true;
82   unsigned Length = 0;
83   for (; *Str; ++Str) {
84     if (*Str == '\n' || strncmp(Str, MAI.getSeparatorString(),
85                                 strlen(MAI.getSeparatorString())) == 0)
86       atInsnStart = true;
87     if (atInsnStart && !std::isspace(static_cast<unsigned char>(*Str))) {
88       Length += MAI.getMaxInstLength();
89       atInsnStart = false;
90     }
91     if (atInsnStart && strncmp(Str, MAI.getCommentString(),
92                                strlen(MAI.getCommentString())) == 0)
93       atInsnStart = false;
94   }
95 
96   return Length;
97 }
98 
99 /// ReplaceTailWithBranchTo - Delete the instruction OldInst and everything
100 /// after it, replacing it with an unconditional branch to NewDest.
101 void
ReplaceTailWithBranchTo(MachineBasicBlock::iterator Tail,MachineBasicBlock * NewDest) const102 TargetInstrInfo::ReplaceTailWithBranchTo(MachineBasicBlock::iterator Tail,
103                                          MachineBasicBlock *NewDest) const {
104   MachineBasicBlock *MBB = Tail->getParent();
105 
106   // Remove all the old successors of MBB from the CFG.
107   while (!MBB->succ_empty())
108     MBB->removeSuccessor(MBB->succ_begin());
109 
110   // Remove all the dead instructions from the end of MBB.
111   MBB->erase(Tail, MBB->end());
112 
113   // If MBB isn't immediately before MBB, insert a branch to it.
114   if (++MachineFunction::iterator(MBB) != MachineFunction::iterator(NewDest))
115     InsertBranch(*MBB, NewDest, nullptr, SmallVector<MachineOperand, 0>(),
116                  Tail->getDebugLoc());
117   MBB->addSuccessor(NewDest);
118 }
119 
120 // commuteInstruction - The default implementation of this method just exchanges
121 // the two operands returned by findCommutedOpIndices.
commuteInstruction(MachineInstr * MI,bool NewMI) const122 MachineInstr *TargetInstrInfo::commuteInstruction(MachineInstr *MI,
123                                                   bool NewMI) const {
124   const MCInstrDesc &MCID = MI->getDesc();
125   bool HasDef = MCID.getNumDefs();
126   if (HasDef && !MI->getOperand(0).isReg())
127     // No idea how to commute this instruction. Target should implement its own.
128     return nullptr;
129   unsigned Idx1, Idx2;
130   if (!findCommutedOpIndices(MI, Idx1, Idx2)) {
131     assert(MI->isCommutable() && "Precondition violation: MI must be commutable.");
132     return nullptr;
133   }
134 
135   assert(MI->getOperand(Idx1).isReg() && MI->getOperand(Idx2).isReg() &&
136          "This only knows how to commute register operands so far");
137   unsigned Reg0 = HasDef ? MI->getOperand(0).getReg() : 0;
138   unsigned Reg1 = MI->getOperand(Idx1).getReg();
139   unsigned Reg2 = MI->getOperand(Idx2).getReg();
140   unsigned SubReg0 = HasDef ? MI->getOperand(0).getSubReg() : 0;
141   unsigned SubReg1 = MI->getOperand(Idx1).getSubReg();
142   unsigned SubReg2 = MI->getOperand(Idx2).getSubReg();
143   bool Reg1IsKill = MI->getOperand(Idx1).isKill();
144   bool Reg2IsKill = MI->getOperand(Idx2).isKill();
145   // If destination is tied to either of the commuted source register, then
146   // it must be updated.
147   if (HasDef && Reg0 == Reg1 &&
148       MI->getDesc().getOperandConstraint(Idx1, MCOI::TIED_TO) == 0) {
149     Reg2IsKill = false;
150     Reg0 = Reg2;
151     SubReg0 = SubReg2;
152   } else if (HasDef && Reg0 == Reg2 &&
153              MI->getDesc().getOperandConstraint(Idx2, MCOI::TIED_TO) == 0) {
154     Reg1IsKill = false;
155     Reg0 = Reg1;
156     SubReg0 = SubReg1;
157   }
158 
159   if (NewMI) {
160     // Create a new instruction.
161     MachineFunction &MF = *MI->getParent()->getParent();
162     MI = MF.CloneMachineInstr(MI);
163   }
164 
165   if (HasDef) {
166     MI->getOperand(0).setReg(Reg0);
167     MI->getOperand(0).setSubReg(SubReg0);
168   }
169   MI->getOperand(Idx2).setReg(Reg1);
170   MI->getOperand(Idx1).setReg(Reg2);
171   MI->getOperand(Idx2).setSubReg(SubReg1);
172   MI->getOperand(Idx1).setSubReg(SubReg2);
173   MI->getOperand(Idx2).setIsKill(Reg1IsKill);
174   MI->getOperand(Idx1).setIsKill(Reg2IsKill);
175   return MI;
176 }
177 
178 /// findCommutedOpIndices - If specified MI is commutable, return the two
179 /// operand indices that would swap value. Return true if the instruction
180 /// is not in a form which this routine understands.
findCommutedOpIndices(MachineInstr * MI,unsigned & SrcOpIdx1,unsigned & SrcOpIdx2) const181 bool TargetInstrInfo::findCommutedOpIndices(MachineInstr *MI,
182                                             unsigned &SrcOpIdx1,
183                                             unsigned &SrcOpIdx2) const {
184   assert(!MI->isBundle() &&
185          "TargetInstrInfo::findCommutedOpIndices() can't handle bundles");
186 
187   const MCInstrDesc &MCID = MI->getDesc();
188   if (!MCID.isCommutable())
189     return false;
190   // This assumes v0 = op v1, v2 and commuting would swap v1 and v2. If this
191   // is not true, then the target must implement this.
192   SrcOpIdx1 = MCID.getNumDefs();
193   SrcOpIdx2 = SrcOpIdx1 + 1;
194   if (!MI->getOperand(SrcOpIdx1).isReg() ||
195       !MI->getOperand(SrcOpIdx2).isReg())
196     // No idea.
197     return false;
198   return true;
199 }
200 
201 
202 bool
isUnpredicatedTerminator(const MachineInstr * MI) const203 TargetInstrInfo::isUnpredicatedTerminator(const MachineInstr *MI) const {
204   if (!MI->isTerminator()) return false;
205 
206   // Conditional branch is a special case.
207   if (MI->isBranch() && !MI->isBarrier())
208     return true;
209   if (!MI->isPredicable())
210     return true;
211   return !isPredicated(MI);
212 }
213 
214 
PredicateInstruction(MachineInstr * MI,const SmallVectorImpl<MachineOperand> & Pred) const215 bool TargetInstrInfo::PredicateInstruction(MachineInstr *MI,
216                             const SmallVectorImpl<MachineOperand> &Pred) const {
217   bool MadeChange = false;
218 
219   assert(!MI->isBundle() &&
220          "TargetInstrInfo::PredicateInstruction() can't handle bundles");
221 
222   const MCInstrDesc &MCID = MI->getDesc();
223   if (!MI->isPredicable())
224     return false;
225 
226   for (unsigned j = 0, i = 0, e = MI->getNumOperands(); i != e; ++i) {
227     if (MCID.OpInfo[i].isPredicate()) {
228       MachineOperand &MO = MI->getOperand(i);
229       if (MO.isReg()) {
230         MO.setReg(Pred[j].getReg());
231         MadeChange = true;
232       } else if (MO.isImm()) {
233         MO.setImm(Pred[j].getImm());
234         MadeChange = true;
235       } else if (MO.isMBB()) {
236         MO.setMBB(Pred[j].getMBB());
237         MadeChange = true;
238       }
239       ++j;
240     }
241   }
242   return MadeChange;
243 }
244 
hasLoadFromStackSlot(const MachineInstr * MI,const MachineMemOperand * & MMO,int & FrameIndex) const245 bool TargetInstrInfo::hasLoadFromStackSlot(const MachineInstr *MI,
246                                            const MachineMemOperand *&MMO,
247                                            int &FrameIndex) const {
248   for (MachineInstr::mmo_iterator o = MI->memoperands_begin(),
249          oe = MI->memoperands_end();
250        o != oe;
251        ++o) {
252     if ((*o)->isLoad()) {
253       if (const FixedStackPseudoSourceValue *Value =
254           dyn_cast_or_null<FixedStackPseudoSourceValue>(
255               (*o)->getPseudoValue())) {
256         FrameIndex = Value->getFrameIndex();
257         MMO = *o;
258         return true;
259       }
260     }
261   }
262   return false;
263 }
264 
hasStoreToStackSlot(const MachineInstr * MI,const MachineMemOperand * & MMO,int & FrameIndex) const265 bool TargetInstrInfo::hasStoreToStackSlot(const MachineInstr *MI,
266                                           const MachineMemOperand *&MMO,
267                                           int &FrameIndex) const {
268   for (MachineInstr::mmo_iterator o = MI->memoperands_begin(),
269          oe = MI->memoperands_end();
270        o != oe;
271        ++o) {
272     if ((*o)->isStore()) {
273       if (const FixedStackPseudoSourceValue *Value =
274           dyn_cast_or_null<FixedStackPseudoSourceValue>(
275               (*o)->getPseudoValue())) {
276         FrameIndex = Value->getFrameIndex();
277         MMO = *o;
278         return true;
279       }
280     }
281   }
282   return false;
283 }
284 
getStackSlotRange(const TargetRegisterClass * RC,unsigned SubIdx,unsigned & Size,unsigned & Offset,const MachineFunction & MF) const285 bool TargetInstrInfo::getStackSlotRange(const TargetRegisterClass *RC,
286                                         unsigned SubIdx, unsigned &Size,
287                                         unsigned &Offset,
288                                         const MachineFunction &MF) const {
289   if (!SubIdx) {
290     Size = RC->getSize();
291     Offset = 0;
292     return true;
293   }
294   const TargetRegisterInfo *TRI = MF.getSubtarget().getRegisterInfo();
295   unsigned BitSize = TRI->getSubRegIdxSize(SubIdx);
296   // Convert bit size to byte size to be consistent with
297   // MCRegisterClass::getSize().
298   if (BitSize % 8)
299     return false;
300 
301   int BitOffset = TRI->getSubRegIdxOffset(SubIdx);
302   if (BitOffset < 0 || BitOffset % 8)
303     return false;
304 
305   Size = BitSize /= 8;
306   Offset = (unsigned)BitOffset / 8;
307 
308   assert(RC->getSize() >= (Offset + Size) && "bad subregister range");
309 
310   if (!MF.getTarget().getDataLayout()->isLittleEndian()) {
311     Offset = RC->getSize() - (Offset + Size);
312   }
313   return true;
314 }
315 
reMaterialize(MachineBasicBlock & MBB,MachineBasicBlock::iterator I,unsigned DestReg,unsigned SubIdx,const MachineInstr * Orig,const TargetRegisterInfo & TRI) const316 void TargetInstrInfo::reMaterialize(MachineBasicBlock &MBB,
317                                     MachineBasicBlock::iterator I,
318                                     unsigned DestReg,
319                                     unsigned SubIdx,
320                                     const MachineInstr *Orig,
321                                     const TargetRegisterInfo &TRI) const {
322   MachineInstr *MI = MBB.getParent()->CloneMachineInstr(Orig);
323   MI->substituteRegister(MI->getOperand(0).getReg(), DestReg, SubIdx, TRI);
324   MBB.insert(I, MI);
325 }
326 
327 bool
produceSameValue(const MachineInstr * MI0,const MachineInstr * MI1,const MachineRegisterInfo * MRI) const328 TargetInstrInfo::produceSameValue(const MachineInstr *MI0,
329                                   const MachineInstr *MI1,
330                                   const MachineRegisterInfo *MRI) const {
331   return MI0->isIdenticalTo(MI1, MachineInstr::IgnoreVRegDefs);
332 }
333 
duplicate(MachineInstr * Orig,MachineFunction & MF) const334 MachineInstr *TargetInstrInfo::duplicate(MachineInstr *Orig,
335                                          MachineFunction &MF) const {
336   assert(!Orig->isNotDuplicable() &&
337          "Instruction cannot be duplicated");
338   return MF.CloneMachineInstr(Orig);
339 }
340 
341 // If the COPY instruction in MI can be folded to a stack operation, return
342 // the register class to use.
canFoldCopy(const MachineInstr * MI,unsigned FoldIdx)343 static const TargetRegisterClass *canFoldCopy(const MachineInstr *MI,
344                                               unsigned FoldIdx) {
345   assert(MI->isCopy() && "MI must be a COPY instruction");
346   if (MI->getNumOperands() != 2)
347     return nullptr;
348   assert(FoldIdx<2 && "FoldIdx refers no nonexistent operand");
349 
350   const MachineOperand &FoldOp = MI->getOperand(FoldIdx);
351   const MachineOperand &LiveOp = MI->getOperand(1-FoldIdx);
352 
353   if (FoldOp.getSubReg() || LiveOp.getSubReg())
354     return nullptr;
355 
356   unsigned FoldReg = FoldOp.getReg();
357   unsigned LiveReg = LiveOp.getReg();
358 
359   assert(TargetRegisterInfo::isVirtualRegister(FoldReg) &&
360          "Cannot fold physregs");
361 
362   const MachineRegisterInfo &MRI = MI->getParent()->getParent()->getRegInfo();
363   const TargetRegisterClass *RC = MRI.getRegClass(FoldReg);
364 
365   if (TargetRegisterInfo::isPhysicalRegister(LiveOp.getReg()))
366     return RC->contains(LiveOp.getReg()) ? RC : nullptr;
367 
368   if (RC->hasSubClassEq(MRI.getRegClass(LiveReg)))
369     return RC;
370 
371   // FIXME: Allow folding when register classes are memory compatible.
372   return nullptr;
373 }
374 
getNoopForMachoTarget(MCInst & NopInst) const375 void TargetInstrInfo::getNoopForMachoTarget(MCInst &NopInst) const {
376   llvm_unreachable("Not a MachO target");
377 }
378 
canFoldMemoryOperand(const MachineInstr * MI,ArrayRef<unsigned> Ops) const379 bool TargetInstrInfo::canFoldMemoryOperand(const MachineInstr *MI,
380                                            ArrayRef<unsigned> Ops) const {
381   return MI->isCopy() && Ops.size() == 1 && canFoldCopy(MI, Ops[0]);
382 }
383 
foldPatchpoint(MachineFunction & MF,MachineInstr * MI,ArrayRef<unsigned> Ops,int FrameIndex,const TargetInstrInfo & TII)384 static MachineInstr *foldPatchpoint(MachineFunction &MF, MachineInstr *MI,
385                                     ArrayRef<unsigned> Ops, int FrameIndex,
386                                     const TargetInstrInfo &TII) {
387   unsigned StartIdx = 0;
388   switch (MI->getOpcode()) {
389   case TargetOpcode::STACKMAP:
390     StartIdx = 2; // Skip ID, nShadowBytes.
391     break;
392   case TargetOpcode::PATCHPOINT: {
393     // For PatchPoint, the call args are not foldable.
394     PatchPointOpers opers(MI);
395     StartIdx = opers.getVarIdx();
396     break;
397   }
398   default:
399     llvm_unreachable("unexpected stackmap opcode");
400   }
401 
402   // Return false if any operands requested for folding are not foldable (not
403   // part of the stackmap's live values).
404   for (unsigned Op : Ops) {
405     if (Op < StartIdx)
406       return nullptr;
407   }
408 
409   MachineInstr *NewMI =
410     MF.CreateMachineInstr(TII.get(MI->getOpcode()), MI->getDebugLoc(), true);
411   MachineInstrBuilder MIB(MF, NewMI);
412 
413   // No need to fold return, the meta data, and function arguments
414   for (unsigned i = 0; i < StartIdx; ++i)
415     MIB.addOperand(MI->getOperand(i));
416 
417   for (unsigned i = StartIdx; i < MI->getNumOperands(); ++i) {
418     MachineOperand &MO = MI->getOperand(i);
419     if (std::find(Ops.begin(), Ops.end(), i) != Ops.end()) {
420       unsigned SpillSize;
421       unsigned SpillOffset;
422       // Compute the spill slot size and offset.
423       const TargetRegisterClass *RC =
424         MF.getRegInfo().getRegClass(MO.getReg());
425       bool Valid =
426           TII.getStackSlotRange(RC, MO.getSubReg(), SpillSize, SpillOffset, MF);
427       if (!Valid)
428         report_fatal_error("cannot spill patchpoint subregister operand");
429       MIB.addImm(StackMaps::IndirectMemRefOp);
430       MIB.addImm(SpillSize);
431       MIB.addFrameIndex(FrameIndex);
432       MIB.addImm(SpillOffset);
433     }
434     else
435       MIB.addOperand(MO);
436   }
437   return NewMI;
438 }
439 
440 /// foldMemoryOperand - Attempt to fold a load or store of the specified stack
441 /// slot into the specified machine instruction for the specified operand(s).
442 /// If this is possible, a new instruction is returned with the specified
443 /// operand folded, otherwise NULL is returned. The client is responsible for
444 /// removing the old instruction and adding the new one in the instruction
445 /// stream.
foldMemoryOperand(MachineBasicBlock::iterator MI,ArrayRef<unsigned> Ops,int FI) const446 MachineInstr *TargetInstrInfo::foldMemoryOperand(MachineBasicBlock::iterator MI,
447                                                  ArrayRef<unsigned> Ops,
448                                                  int FI) const {
449   unsigned Flags = 0;
450   for (unsigned i = 0, e = Ops.size(); i != e; ++i)
451     if (MI->getOperand(Ops[i]).isDef())
452       Flags |= MachineMemOperand::MOStore;
453     else
454       Flags |= MachineMemOperand::MOLoad;
455 
456   MachineBasicBlock *MBB = MI->getParent();
457   assert(MBB && "foldMemoryOperand needs an inserted instruction");
458   MachineFunction &MF = *MBB->getParent();
459 
460   MachineInstr *NewMI = nullptr;
461 
462   if (MI->getOpcode() == TargetOpcode::STACKMAP ||
463       MI->getOpcode() == TargetOpcode::PATCHPOINT) {
464     // Fold stackmap/patchpoint.
465     NewMI = foldPatchpoint(MF, MI, Ops, FI, *this);
466   } else {
467     // Ask the target to do the actual folding.
468     NewMI =foldMemoryOperandImpl(MF, MI, Ops, FI);
469   }
470 
471   if (NewMI) {
472     NewMI->setMemRefs(MI->memoperands_begin(), MI->memoperands_end());
473     // Add a memory operand, foldMemoryOperandImpl doesn't do that.
474     assert((!(Flags & MachineMemOperand::MOStore) ||
475             NewMI->mayStore()) &&
476            "Folded a def to a non-store!");
477     assert((!(Flags & MachineMemOperand::MOLoad) ||
478             NewMI->mayLoad()) &&
479            "Folded a use to a non-load!");
480     const MachineFrameInfo &MFI = *MF.getFrameInfo();
481     assert(MFI.getObjectOffset(FI) != -1);
482     MachineMemOperand *MMO =
483       MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(FI),
484                               Flags, MFI.getObjectSize(FI),
485                               MFI.getObjectAlignment(FI));
486     NewMI->addMemOperand(MF, MMO);
487 
488     // FIXME: change foldMemoryOperandImpl semantics to also insert NewMI.
489     return MBB->insert(MI, NewMI);
490   }
491 
492   // Straight COPY may fold as load/store.
493   if (!MI->isCopy() || Ops.size() != 1)
494     return nullptr;
495 
496   const TargetRegisterClass *RC = canFoldCopy(MI, Ops[0]);
497   if (!RC)
498     return nullptr;
499 
500   const MachineOperand &MO = MI->getOperand(1-Ops[0]);
501   MachineBasicBlock::iterator Pos = MI;
502   const TargetRegisterInfo *TRI = MF.getSubtarget().getRegisterInfo();
503 
504   if (Flags == MachineMemOperand::MOStore)
505     storeRegToStackSlot(*MBB, Pos, MO.getReg(), MO.isKill(), FI, RC, TRI);
506   else
507     loadRegFromStackSlot(*MBB, Pos, MO.getReg(), FI, RC, TRI);
508   return --Pos;
509 }
510 
511 /// foldMemoryOperand - Same as the previous version except it allows folding
512 /// of any load and store from / to any address, not just from a specific
513 /// stack slot.
foldMemoryOperand(MachineBasicBlock::iterator MI,ArrayRef<unsigned> Ops,MachineInstr * LoadMI) const514 MachineInstr *TargetInstrInfo::foldMemoryOperand(MachineBasicBlock::iterator MI,
515                                                  ArrayRef<unsigned> Ops,
516                                                  MachineInstr *LoadMI) const {
517   assert(LoadMI->canFoldAsLoad() && "LoadMI isn't foldable!");
518 #ifndef NDEBUG
519   for (unsigned i = 0, e = Ops.size(); i != e; ++i)
520     assert(MI->getOperand(Ops[i]).isUse() && "Folding load into def!");
521 #endif
522   MachineBasicBlock &MBB = *MI->getParent();
523   MachineFunction &MF = *MBB.getParent();
524 
525   // Ask the target to do the actual folding.
526   MachineInstr *NewMI = nullptr;
527   int FrameIndex = 0;
528 
529   if ((MI->getOpcode() == TargetOpcode::STACKMAP ||
530        MI->getOpcode() == TargetOpcode::PATCHPOINT) &&
531       isLoadFromStackSlot(LoadMI, FrameIndex)) {
532     // Fold stackmap/patchpoint.
533     NewMI = foldPatchpoint(MF, MI, Ops, FrameIndex, *this);
534   } else {
535     // Ask the target to do the actual folding.
536     NewMI = foldMemoryOperandImpl(MF, MI, Ops, LoadMI);
537   }
538 
539   if (!NewMI) return nullptr;
540 
541   NewMI = MBB.insert(MI, NewMI);
542 
543   // Copy the memoperands from the load to the folded instruction.
544   if (MI->memoperands_empty()) {
545     NewMI->setMemRefs(LoadMI->memoperands_begin(),
546                       LoadMI->memoperands_end());
547   }
548   else {
549     // Handle the rare case of folding multiple loads.
550     NewMI->setMemRefs(MI->memoperands_begin(),
551                       MI->memoperands_end());
552     for (MachineInstr::mmo_iterator I = LoadMI->memoperands_begin(),
553            E = LoadMI->memoperands_end(); I != E; ++I) {
554       NewMI->addMemOperand(MF, *I);
555     }
556   }
557   return NewMI;
558 }
559 
560 bool TargetInstrInfo::
isReallyTriviallyReMaterializableGeneric(const MachineInstr * MI,AliasAnalysis * AA) const561 isReallyTriviallyReMaterializableGeneric(const MachineInstr *MI,
562                                          AliasAnalysis *AA) const {
563   const MachineFunction &MF = *MI->getParent()->getParent();
564   const MachineRegisterInfo &MRI = MF.getRegInfo();
565 
566   // Remat clients assume operand 0 is the defined register.
567   if (!MI->getNumOperands() || !MI->getOperand(0).isReg())
568     return false;
569   unsigned DefReg = MI->getOperand(0).getReg();
570 
571   // A sub-register definition can only be rematerialized if the instruction
572   // doesn't read the other parts of the register.  Otherwise it is really a
573   // read-modify-write operation on the full virtual register which cannot be
574   // moved safely.
575   if (TargetRegisterInfo::isVirtualRegister(DefReg) &&
576       MI->getOperand(0).getSubReg() && MI->readsVirtualRegister(DefReg))
577     return false;
578 
579   // A load from a fixed stack slot can be rematerialized. This may be
580   // redundant with subsequent checks, but it's target-independent,
581   // simple, and a common case.
582   int FrameIdx = 0;
583   if (isLoadFromStackSlot(MI, FrameIdx) &&
584       MF.getFrameInfo()->isImmutableObjectIndex(FrameIdx))
585     return true;
586 
587   // Avoid instructions obviously unsafe for remat.
588   if (MI->isNotDuplicable() || MI->mayStore() ||
589       MI->hasUnmodeledSideEffects())
590     return false;
591 
592   // Don't remat inline asm. We have no idea how expensive it is
593   // even if it's side effect free.
594   if (MI->isInlineAsm())
595     return false;
596 
597   // Avoid instructions which load from potentially varying memory.
598   if (MI->mayLoad() && !MI->isInvariantLoad(AA))
599     return false;
600 
601   // If any of the registers accessed are non-constant, conservatively assume
602   // the instruction is not rematerializable.
603   for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
604     const MachineOperand &MO = MI->getOperand(i);
605     if (!MO.isReg()) continue;
606     unsigned Reg = MO.getReg();
607     if (Reg == 0)
608       continue;
609 
610     // Check for a well-behaved physical register.
611     if (TargetRegisterInfo::isPhysicalRegister(Reg)) {
612       if (MO.isUse()) {
613         // If the physreg has no defs anywhere, it's just an ambient register
614         // and we can freely move its uses. Alternatively, if it's allocatable,
615         // it could get allocated to something with a def during allocation.
616         if (!MRI.isConstantPhysReg(Reg, MF))
617           return false;
618       } else {
619         // A physreg def. We can't remat it.
620         return false;
621       }
622       continue;
623     }
624 
625     // Only allow one virtual-register def.  There may be multiple defs of the
626     // same virtual register, though.
627     if (MO.isDef() && Reg != DefReg)
628       return false;
629 
630     // Don't allow any virtual-register uses. Rematting an instruction with
631     // virtual register uses would length the live ranges of the uses, which
632     // is not necessarily a good idea, certainly not "trivial".
633     if (MO.isUse())
634       return false;
635   }
636 
637   // Everything checked out.
638   return true;
639 }
640 
getSPAdjust(const MachineInstr * MI) const641 int TargetInstrInfo::getSPAdjust(const MachineInstr *MI) const {
642   const MachineFunction *MF = MI->getParent()->getParent();
643   const TargetFrameLowering *TFI = MF->getSubtarget().getFrameLowering();
644   bool StackGrowsDown =
645     TFI->getStackGrowthDirection() == TargetFrameLowering::StackGrowsDown;
646 
647   int FrameSetupOpcode = getCallFrameSetupOpcode();
648   int FrameDestroyOpcode = getCallFrameDestroyOpcode();
649 
650   if (MI->getOpcode() != FrameSetupOpcode &&
651       MI->getOpcode() != FrameDestroyOpcode)
652     return 0;
653 
654   int SPAdj = MI->getOperand(0).getImm();
655 
656   if ((!StackGrowsDown && MI->getOpcode() == FrameSetupOpcode) ||
657        (StackGrowsDown && MI->getOpcode() == FrameDestroyOpcode))
658     SPAdj = -SPAdj;
659 
660   return SPAdj;
661 }
662 
663 /// isSchedulingBoundary - Test if the given instruction should be
664 /// considered a scheduling boundary. This primarily includes labels
665 /// and terminators.
isSchedulingBoundary(const MachineInstr * MI,const MachineBasicBlock * MBB,const MachineFunction & MF) const666 bool TargetInstrInfo::isSchedulingBoundary(const MachineInstr *MI,
667                                            const MachineBasicBlock *MBB,
668                                            const MachineFunction &MF) const {
669   // Terminators and labels can't be scheduled around.
670   if (MI->isTerminator() || MI->isPosition())
671     return true;
672 
673   // Don't attempt to schedule around any instruction that defines
674   // a stack-oriented pointer, as it's unlikely to be profitable. This
675   // saves compile time, because it doesn't require every single
676   // stack slot reference to depend on the instruction that does the
677   // modification.
678   const TargetLowering &TLI = *MF.getSubtarget().getTargetLowering();
679   const TargetRegisterInfo *TRI = MF.getSubtarget().getRegisterInfo();
680   if (MI->modifiesRegister(TLI.getStackPointerRegisterToSaveRestore(), TRI))
681     return true;
682 
683   return false;
684 }
685 
686 // Provide a global flag for disabling the PreRA hazard recognizer that targets
687 // may choose to honor.
usePreRAHazardRecognizer() const688 bool TargetInstrInfo::usePreRAHazardRecognizer() const {
689   return !DisableHazardRecognizer;
690 }
691 
692 // Default implementation of CreateTargetRAHazardRecognizer.
693 ScheduleHazardRecognizer *TargetInstrInfo::
CreateTargetHazardRecognizer(const TargetSubtargetInfo * STI,const ScheduleDAG * DAG) const694 CreateTargetHazardRecognizer(const TargetSubtargetInfo *STI,
695                              const ScheduleDAG *DAG) const {
696   // Dummy hazard recognizer allows all instructions to issue.
697   return new ScheduleHazardRecognizer();
698 }
699 
700 // Default implementation of CreateTargetMIHazardRecognizer.
701 ScheduleHazardRecognizer *TargetInstrInfo::
CreateTargetMIHazardRecognizer(const InstrItineraryData * II,const ScheduleDAG * DAG) const702 CreateTargetMIHazardRecognizer(const InstrItineraryData *II,
703                                const ScheduleDAG *DAG) const {
704   return (ScheduleHazardRecognizer *)
705     new ScoreboardHazardRecognizer(II, DAG, "misched");
706 }
707 
708 // Default implementation of CreateTargetPostRAHazardRecognizer.
709 ScheduleHazardRecognizer *TargetInstrInfo::
CreateTargetPostRAHazardRecognizer(const InstrItineraryData * II,const ScheduleDAG * DAG) const710 CreateTargetPostRAHazardRecognizer(const InstrItineraryData *II,
711                                    const ScheduleDAG *DAG) const {
712   return (ScheduleHazardRecognizer *)
713     new ScoreboardHazardRecognizer(II, DAG, "post-RA-sched");
714 }
715 
716 //===----------------------------------------------------------------------===//
717 //  SelectionDAG latency interface.
718 //===----------------------------------------------------------------------===//
719 
720 int
getOperandLatency(const InstrItineraryData * ItinData,SDNode * DefNode,unsigned DefIdx,SDNode * UseNode,unsigned UseIdx) const721 TargetInstrInfo::getOperandLatency(const InstrItineraryData *ItinData,
722                                    SDNode *DefNode, unsigned DefIdx,
723                                    SDNode *UseNode, unsigned UseIdx) const {
724   if (!ItinData || ItinData->isEmpty())
725     return -1;
726 
727   if (!DefNode->isMachineOpcode())
728     return -1;
729 
730   unsigned DefClass = get(DefNode->getMachineOpcode()).getSchedClass();
731   if (!UseNode->isMachineOpcode())
732     return ItinData->getOperandCycle(DefClass, DefIdx);
733   unsigned UseClass = get(UseNode->getMachineOpcode()).getSchedClass();
734   return ItinData->getOperandLatency(DefClass, DefIdx, UseClass, UseIdx);
735 }
736 
getInstrLatency(const InstrItineraryData * ItinData,SDNode * N) const737 int TargetInstrInfo::getInstrLatency(const InstrItineraryData *ItinData,
738                                      SDNode *N) const {
739   if (!ItinData || ItinData->isEmpty())
740     return 1;
741 
742   if (!N->isMachineOpcode())
743     return 1;
744 
745   return ItinData->getStageLatency(get(N->getMachineOpcode()).getSchedClass());
746 }
747 
748 //===----------------------------------------------------------------------===//
749 //  MachineInstr latency interface.
750 //===----------------------------------------------------------------------===//
751 
752 unsigned
getNumMicroOps(const InstrItineraryData * ItinData,const MachineInstr * MI) const753 TargetInstrInfo::getNumMicroOps(const InstrItineraryData *ItinData,
754                                 const MachineInstr *MI) const {
755   if (!ItinData || ItinData->isEmpty())
756     return 1;
757 
758   unsigned Class = MI->getDesc().getSchedClass();
759   int UOps = ItinData->Itineraries[Class].NumMicroOps;
760   if (UOps >= 0)
761     return UOps;
762 
763   // The # of u-ops is dynamically determined. The specific target should
764   // override this function to return the right number.
765   return 1;
766 }
767 
768 /// Return the default expected latency for a def based on it's opcode.
defaultDefLatency(const MCSchedModel & SchedModel,const MachineInstr * DefMI) const769 unsigned TargetInstrInfo::defaultDefLatency(const MCSchedModel &SchedModel,
770                                             const MachineInstr *DefMI) const {
771   if (DefMI->isTransient())
772     return 0;
773   if (DefMI->mayLoad())
774     return SchedModel.LoadLatency;
775   if (isHighLatencyDef(DefMI->getOpcode()))
776     return SchedModel.HighLatency;
777   return 1;
778 }
779 
getPredicationCost(const MachineInstr *) const780 unsigned TargetInstrInfo::getPredicationCost(const MachineInstr *) const {
781   return 0;
782 }
783 
784 unsigned TargetInstrInfo::
getInstrLatency(const InstrItineraryData * ItinData,const MachineInstr * MI,unsigned * PredCost) const785 getInstrLatency(const InstrItineraryData *ItinData,
786                 const MachineInstr *MI,
787                 unsigned *PredCost) const {
788   // Default to one cycle for no itinerary. However, an "empty" itinerary may
789   // still have a MinLatency property, which getStageLatency checks.
790   if (!ItinData)
791     return MI->mayLoad() ? 2 : 1;
792 
793   return ItinData->getStageLatency(MI->getDesc().getSchedClass());
794 }
795 
hasLowDefLatency(const InstrItineraryData * ItinData,const MachineInstr * DefMI,unsigned DefIdx) const796 bool TargetInstrInfo::hasLowDefLatency(const InstrItineraryData *ItinData,
797                                        const MachineInstr *DefMI,
798                                        unsigned DefIdx) const {
799   if (!ItinData || ItinData->isEmpty())
800     return false;
801 
802   unsigned DefClass = DefMI->getDesc().getSchedClass();
803   int DefCycle = ItinData->getOperandCycle(DefClass, DefIdx);
804   return (DefCycle != -1 && DefCycle <= 1);
805 }
806 
807 /// Both DefMI and UseMI must be valid.  By default, call directly to the
808 /// itinerary. This may be overriden by the target.
809 int TargetInstrInfo::
getOperandLatency(const InstrItineraryData * ItinData,const MachineInstr * DefMI,unsigned DefIdx,const MachineInstr * UseMI,unsigned UseIdx) const810 getOperandLatency(const InstrItineraryData *ItinData,
811                   const MachineInstr *DefMI, unsigned DefIdx,
812                   const MachineInstr *UseMI, unsigned UseIdx) const {
813   unsigned DefClass = DefMI->getDesc().getSchedClass();
814   unsigned UseClass = UseMI->getDesc().getSchedClass();
815   return ItinData->getOperandLatency(DefClass, DefIdx, UseClass, UseIdx);
816 }
817 
818 /// If we can determine the operand latency from the def only, without itinerary
819 /// lookup, do so. Otherwise return -1.
computeDefOperandLatency(const InstrItineraryData * ItinData,const MachineInstr * DefMI) const820 int TargetInstrInfo::computeDefOperandLatency(
821   const InstrItineraryData *ItinData,
822   const MachineInstr *DefMI) const {
823 
824   // Let the target hook getInstrLatency handle missing itineraries.
825   if (!ItinData)
826     return getInstrLatency(ItinData, DefMI);
827 
828   if(ItinData->isEmpty())
829     return defaultDefLatency(ItinData->SchedModel, DefMI);
830 
831   // ...operand lookup required
832   return -1;
833 }
834 
835 /// computeOperandLatency - Compute and return the latency of the given data
836 /// dependent def and use when the operand indices are already known. UseMI may
837 /// be NULL for an unknown use.
838 ///
839 /// FindMin may be set to get the minimum vs. expected latency. Minimum
840 /// latency is used for scheduling groups, while expected latency is for
841 /// instruction cost and critical path.
842 ///
843 /// Depending on the subtarget's itinerary properties, this may or may not need
844 /// to call getOperandLatency(). For most subtargets, we don't need DefIdx or
845 /// UseIdx to compute min latency.
846 unsigned TargetInstrInfo::
computeOperandLatency(const InstrItineraryData * ItinData,const MachineInstr * DefMI,unsigned DefIdx,const MachineInstr * UseMI,unsigned UseIdx) const847 computeOperandLatency(const InstrItineraryData *ItinData,
848                       const MachineInstr *DefMI, unsigned DefIdx,
849                       const MachineInstr *UseMI, unsigned UseIdx) const {
850 
851   int DefLatency = computeDefOperandLatency(ItinData, DefMI);
852   if (DefLatency >= 0)
853     return DefLatency;
854 
855   assert(ItinData && !ItinData->isEmpty() && "computeDefOperandLatency fail");
856 
857   int OperLatency = 0;
858   if (UseMI)
859     OperLatency = getOperandLatency(ItinData, DefMI, DefIdx, UseMI, UseIdx);
860   else {
861     unsigned DefClass = DefMI->getDesc().getSchedClass();
862     OperLatency = ItinData->getOperandCycle(DefClass, DefIdx);
863   }
864   if (OperLatency >= 0)
865     return OperLatency;
866 
867   // No operand latency was found.
868   unsigned InstrLatency = getInstrLatency(ItinData, DefMI);
869 
870   // Expected latency is the max of the stage latency and itinerary props.
871   InstrLatency = std::max(InstrLatency,
872                           defaultDefLatency(ItinData->SchedModel, DefMI));
873   return InstrLatency;
874 }
875 
getRegSequenceInputs(const MachineInstr & MI,unsigned DefIdx,SmallVectorImpl<RegSubRegPairAndIdx> & InputRegs) const876 bool TargetInstrInfo::getRegSequenceInputs(
877     const MachineInstr &MI, unsigned DefIdx,
878     SmallVectorImpl<RegSubRegPairAndIdx> &InputRegs) const {
879   assert((MI.isRegSequence() ||
880           MI.isRegSequenceLike()) && "Instruction do not have the proper type");
881 
882   if (!MI.isRegSequence())
883     return getRegSequenceLikeInputs(MI, DefIdx, InputRegs);
884 
885   // We are looking at:
886   // Def = REG_SEQUENCE v0, sub0, v1, sub1, ...
887   assert(DefIdx == 0 && "REG_SEQUENCE only has one def");
888   for (unsigned OpIdx = 1, EndOpIdx = MI.getNumOperands(); OpIdx != EndOpIdx;
889        OpIdx += 2) {
890     const MachineOperand &MOReg = MI.getOperand(OpIdx);
891     const MachineOperand &MOSubIdx = MI.getOperand(OpIdx + 1);
892     assert(MOSubIdx.isImm() &&
893            "One of the subindex of the reg_sequence is not an immediate");
894     // Record Reg:SubReg, SubIdx.
895     InputRegs.push_back(RegSubRegPairAndIdx(MOReg.getReg(), MOReg.getSubReg(),
896                                             (unsigned)MOSubIdx.getImm()));
897   }
898   return true;
899 }
900 
getExtractSubregInputs(const MachineInstr & MI,unsigned DefIdx,RegSubRegPairAndIdx & InputReg) const901 bool TargetInstrInfo::getExtractSubregInputs(
902     const MachineInstr &MI, unsigned DefIdx,
903     RegSubRegPairAndIdx &InputReg) const {
904   assert((MI.isExtractSubreg() ||
905       MI.isExtractSubregLike()) && "Instruction do not have the proper type");
906 
907   if (!MI.isExtractSubreg())
908     return getExtractSubregLikeInputs(MI, DefIdx, InputReg);
909 
910   // We are looking at:
911   // Def = EXTRACT_SUBREG v0.sub1, sub0.
912   assert(DefIdx == 0 && "EXTRACT_SUBREG only has one def");
913   const MachineOperand &MOReg = MI.getOperand(1);
914   const MachineOperand &MOSubIdx = MI.getOperand(2);
915   assert(MOSubIdx.isImm() &&
916          "The subindex of the extract_subreg is not an immediate");
917 
918   InputReg.Reg = MOReg.getReg();
919   InputReg.SubReg = MOReg.getSubReg();
920   InputReg.SubIdx = (unsigned)MOSubIdx.getImm();
921   return true;
922 }
923 
getInsertSubregInputs(const MachineInstr & MI,unsigned DefIdx,RegSubRegPair & BaseReg,RegSubRegPairAndIdx & InsertedReg) const924 bool TargetInstrInfo::getInsertSubregInputs(
925     const MachineInstr &MI, unsigned DefIdx,
926     RegSubRegPair &BaseReg, RegSubRegPairAndIdx &InsertedReg) const {
927   assert((MI.isInsertSubreg() ||
928       MI.isInsertSubregLike()) && "Instruction do not have the proper type");
929 
930   if (!MI.isInsertSubreg())
931     return getInsertSubregLikeInputs(MI, DefIdx, BaseReg, InsertedReg);
932 
933   // We are looking at:
934   // Def = INSERT_SEQUENCE v0, v1, sub0.
935   assert(DefIdx == 0 && "INSERT_SUBREG only has one def");
936   const MachineOperand &MOBaseReg = MI.getOperand(1);
937   const MachineOperand &MOInsertedReg = MI.getOperand(2);
938   const MachineOperand &MOSubIdx = MI.getOperand(3);
939   assert(MOSubIdx.isImm() &&
940          "One of the subindex of the reg_sequence is not an immediate");
941   BaseReg.Reg = MOBaseReg.getReg();
942   BaseReg.SubReg = MOBaseReg.getSubReg();
943 
944   InsertedReg.Reg = MOInsertedReg.getReg();
945   InsertedReg.SubReg = MOInsertedReg.getSubReg();
946   InsertedReg.SubIdx = (unsigned)MOSubIdx.getImm();
947   return true;
948 }
949