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/CodeGen/TargetInstrInfo.h"
15 #include "llvm/CodeGen/MachineFrameInfo.h"
16 #include "llvm/CodeGen/MachineInstrBuilder.h"
17 #include "llvm/CodeGen/MachineMemOperand.h"
18 #include "llvm/CodeGen/MachineRegisterInfo.h"
19 #include "llvm/CodeGen/PseudoSourceValue.h"
20 #include "llvm/CodeGen/ScoreboardHazardRecognizer.h"
21 #include "llvm/CodeGen/StackMaps.h"
22 #include "llvm/CodeGen/TargetFrameLowering.h"
23 #include "llvm/CodeGen/TargetLowering.h"
24 #include "llvm/CodeGen/TargetRegisterInfo.h"
25 #include "llvm/CodeGen/TargetSchedule.h"
26 #include "llvm/IR/DataLayout.h"
27 #include "llvm/MC/MCAsmInfo.h"
28 #include "llvm/MC/MCInstrItineraries.h"
29 #include "llvm/Support/CommandLine.h"
30 #include "llvm/Support/ErrorHandling.h"
31 #include "llvm/Support/raw_ostream.h"
32 #include "llvm/Target/TargetMachine.h"
33 #include <cctype>
34 
35 using namespace llvm;
36 
37 static cl::opt<bool> DisableHazardRecognizer(
38   "disable-sched-hazard", cl::Hidden, cl::init(false),
39   cl::desc("Disable hazard detection during preRA scheduling"));
40 
~TargetInstrInfo()41 TargetInstrInfo::~TargetInstrInfo() {
42 }
43 
44 const TargetRegisterClass*
getRegClass(const MCInstrDesc & MCID,unsigned OpNum,const TargetRegisterInfo * TRI,const MachineFunction & MF) const45 TargetInstrInfo::getRegClass(const MCInstrDesc &MCID, unsigned OpNum,
46                              const TargetRegisterInfo *TRI,
47                              const MachineFunction &MF) const {
48   if (OpNum >= MCID.getNumOperands())
49     return nullptr;
50 
51   short RegClass = MCID.OpInfo[OpNum].RegClass;
52   if (MCID.OpInfo[OpNum].isLookupPtrRegClass())
53     return TRI->getPointerRegClass(MF, RegClass);
54 
55   // Instructions like INSERT_SUBREG do not have fixed register classes.
56   if (RegClass < 0)
57     return nullptr;
58 
59   // Otherwise just look it up normally.
60   return TRI->getRegClass(RegClass);
61 }
62 
63 /// insertNoop - Insert a noop into the instruction stream at the specified
64 /// point.
insertNoop(MachineBasicBlock & MBB,MachineBasicBlock::iterator MI) const65 void TargetInstrInfo::insertNoop(MachineBasicBlock &MBB,
66                                  MachineBasicBlock::iterator MI) const {
67   llvm_unreachable("Target didn't implement insertNoop!");
68 }
69 
isAsmComment(const char * Str,const MCAsmInfo & MAI)70 static bool isAsmComment(const char *Str, const MCAsmInfo &MAI) {
71   return strncmp(Str, MAI.getCommentString().data(),
72                  MAI.getCommentString().size()) == 0;
73 }
74 
75 /// Measure the specified inline asm to determine an approximation of its
76 /// length.
77 /// Comments (which run till the next SeparatorString or newline) do not
78 /// count as an instruction.
79 /// Any other non-whitespace text is considered an instruction, with
80 /// multiple instructions separated by SeparatorString or newlines.
81 /// Variable-length instructions are not handled here; this function
82 /// may be overloaded in the target code to do that.
83 /// We implement a special case of the .space directive which takes only a
84 /// single integer argument in base 10 that is the size in bytes. This is a
85 /// restricted form of the GAS directive in that we only interpret
86 /// simple--i.e. not a logical or arithmetic expression--size values without
87 /// the optional fill value. This is primarily used for creating arbitrary
88 /// sized inline asm blocks for testing purposes.
getInlineAsmLength(const char * Str,const MCAsmInfo & MAI) const89 unsigned TargetInstrInfo::getInlineAsmLength(const char *Str,
90                                              const MCAsmInfo &MAI) const {
91   // Count the number of instructions in the asm.
92   bool AtInsnStart = true;
93   unsigned Length = 0;
94   for (; *Str; ++Str) {
95     if (*Str == '\n' || strncmp(Str, MAI.getSeparatorString(),
96                                 strlen(MAI.getSeparatorString())) == 0) {
97       AtInsnStart = true;
98     } else if (isAsmComment(Str, MAI)) {
99       // Stop counting as an instruction after a comment until the next
100       // separator.
101       AtInsnStart = false;
102     }
103 
104     if (AtInsnStart && !std::isspace(static_cast<unsigned char>(*Str))) {
105       unsigned AddLength = MAI.getMaxInstLength();
106       if (strncmp(Str, ".space", 6) == 0) {
107         char *EStr;
108         int SpaceSize;
109         SpaceSize = strtol(Str + 6, &EStr, 10);
110         SpaceSize = SpaceSize < 0 ? 0 : SpaceSize;
111         while (*EStr != '\n' && std::isspace(static_cast<unsigned char>(*EStr)))
112           ++EStr;
113         if (*EStr == '\0' || *EStr == '\n' ||
114             isAsmComment(EStr, MAI)) // Successfully parsed .space argument
115           AddLength = SpaceSize;
116       }
117       Length += AddLength;
118       AtInsnStart = false;
119     }
120   }
121 
122   return Length;
123 }
124 
125 /// ReplaceTailWithBranchTo - Delete the instruction OldInst and everything
126 /// after it, replacing it with an unconditional branch to NewDest.
127 void
ReplaceTailWithBranchTo(MachineBasicBlock::iterator Tail,MachineBasicBlock * NewDest) const128 TargetInstrInfo::ReplaceTailWithBranchTo(MachineBasicBlock::iterator Tail,
129                                          MachineBasicBlock *NewDest) const {
130   MachineBasicBlock *MBB = Tail->getParent();
131 
132   // Remove all the old successors of MBB from the CFG.
133   while (!MBB->succ_empty())
134     MBB->removeSuccessor(MBB->succ_begin());
135 
136   // Save off the debug loc before erasing the instruction.
137   DebugLoc DL = Tail->getDebugLoc();
138 
139   // Remove all the dead instructions from the end of MBB.
140   MBB->erase(Tail, MBB->end());
141 
142   // If MBB isn't immediately before MBB, insert a branch to it.
143   if (++MachineFunction::iterator(MBB) != MachineFunction::iterator(NewDest))
144     insertBranch(*MBB, NewDest, nullptr, SmallVector<MachineOperand, 0>(), DL);
145   MBB->addSuccessor(NewDest);
146 }
147 
commuteInstructionImpl(MachineInstr & MI,bool NewMI,unsigned Idx1,unsigned Idx2) const148 MachineInstr *TargetInstrInfo::commuteInstructionImpl(MachineInstr &MI,
149                                                       bool NewMI, unsigned Idx1,
150                                                       unsigned Idx2) const {
151   const MCInstrDesc &MCID = MI.getDesc();
152   bool HasDef = MCID.getNumDefs();
153   if (HasDef && !MI.getOperand(0).isReg())
154     // No idea how to commute this instruction. Target should implement its own.
155     return nullptr;
156 
157   unsigned CommutableOpIdx1 = Idx1; (void)CommutableOpIdx1;
158   unsigned CommutableOpIdx2 = Idx2; (void)CommutableOpIdx2;
159   assert(findCommutedOpIndices(MI, CommutableOpIdx1, CommutableOpIdx2) &&
160          CommutableOpIdx1 == Idx1 && CommutableOpIdx2 == Idx2 &&
161          "TargetInstrInfo::CommuteInstructionImpl(): not commutable operands.");
162   assert(MI.getOperand(Idx1).isReg() && MI.getOperand(Idx2).isReg() &&
163          "This only knows how to commute register operands so far");
164 
165   unsigned Reg0 = HasDef ? MI.getOperand(0).getReg() : 0;
166   unsigned Reg1 = MI.getOperand(Idx1).getReg();
167   unsigned Reg2 = MI.getOperand(Idx2).getReg();
168   unsigned SubReg0 = HasDef ? MI.getOperand(0).getSubReg() : 0;
169   unsigned SubReg1 = MI.getOperand(Idx1).getSubReg();
170   unsigned SubReg2 = MI.getOperand(Idx2).getSubReg();
171   bool Reg1IsKill = MI.getOperand(Idx1).isKill();
172   bool Reg2IsKill = MI.getOperand(Idx2).isKill();
173   bool Reg1IsUndef = MI.getOperand(Idx1).isUndef();
174   bool Reg2IsUndef = MI.getOperand(Idx2).isUndef();
175   bool Reg1IsInternal = MI.getOperand(Idx1).isInternalRead();
176   bool Reg2IsInternal = MI.getOperand(Idx2).isInternalRead();
177   // Avoid calling isRenamable for virtual registers since we assert that
178   // renamable property is only queried/set for physical registers.
179   bool Reg1IsRenamable = TargetRegisterInfo::isPhysicalRegister(Reg1)
180                              ? MI.getOperand(Idx1).isRenamable()
181                              : false;
182   bool Reg2IsRenamable = TargetRegisterInfo::isPhysicalRegister(Reg2)
183                              ? MI.getOperand(Idx2).isRenamable()
184                              : false;
185   // If destination is tied to either of the commuted source register, then
186   // it must be updated.
187   if (HasDef && Reg0 == Reg1 &&
188       MI.getDesc().getOperandConstraint(Idx1, MCOI::TIED_TO) == 0) {
189     Reg2IsKill = false;
190     Reg0 = Reg2;
191     SubReg0 = SubReg2;
192   } else if (HasDef && Reg0 == Reg2 &&
193              MI.getDesc().getOperandConstraint(Idx2, MCOI::TIED_TO) == 0) {
194     Reg1IsKill = false;
195     Reg0 = Reg1;
196     SubReg0 = SubReg1;
197   }
198 
199   MachineInstr *CommutedMI = nullptr;
200   if (NewMI) {
201     // Create a new instruction.
202     MachineFunction &MF = *MI.getMF();
203     CommutedMI = MF.CloneMachineInstr(&MI);
204   } else {
205     CommutedMI = &MI;
206   }
207 
208   if (HasDef) {
209     CommutedMI->getOperand(0).setReg(Reg0);
210     CommutedMI->getOperand(0).setSubReg(SubReg0);
211   }
212   CommutedMI->getOperand(Idx2).setReg(Reg1);
213   CommutedMI->getOperand(Idx1).setReg(Reg2);
214   CommutedMI->getOperand(Idx2).setSubReg(SubReg1);
215   CommutedMI->getOperand(Idx1).setSubReg(SubReg2);
216   CommutedMI->getOperand(Idx2).setIsKill(Reg1IsKill);
217   CommutedMI->getOperand(Idx1).setIsKill(Reg2IsKill);
218   CommutedMI->getOperand(Idx2).setIsUndef(Reg1IsUndef);
219   CommutedMI->getOperand(Idx1).setIsUndef(Reg2IsUndef);
220   CommutedMI->getOperand(Idx2).setIsInternalRead(Reg1IsInternal);
221   CommutedMI->getOperand(Idx1).setIsInternalRead(Reg2IsInternal);
222   // Avoid calling setIsRenamable for virtual registers since we assert that
223   // renamable property is only queried/set for physical registers.
224   if (TargetRegisterInfo::isPhysicalRegister(Reg1))
225     CommutedMI->getOperand(Idx2).setIsRenamable(Reg1IsRenamable);
226   if (TargetRegisterInfo::isPhysicalRegister(Reg2))
227     CommutedMI->getOperand(Idx1).setIsRenamable(Reg2IsRenamable);
228   return CommutedMI;
229 }
230 
commuteInstruction(MachineInstr & MI,bool NewMI,unsigned OpIdx1,unsigned OpIdx2) const231 MachineInstr *TargetInstrInfo::commuteInstruction(MachineInstr &MI, bool NewMI,
232                                                   unsigned OpIdx1,
233                                                   unsigned OpIdx2) const {
234   // If OpIdx1 or OpIdx2 is not specified, then this method is free to choose
235   // any commutable operand, which is done in findCommutedOpIndices() method
236   // called below.
237   if ((OpIdx1 == CommuteAnyOperandIndex || OpIdx2 == CommuteAnyOperandIndex) &&
238       !findCommutedOpIndices(MI, OpIdx1, OpIdx2)) {
239     assert(MI.isCommutable() &&
240            "Precondition violation: MI must be commutable.");
241     return nullptr;
242   }
243   return commuteInstructionImpl(MI, NewMI, OpIdx1, OpIdx2);
244 }
245 
fixCommutedOpIndices(unsigned & ResultIdx1,unsigned & ResultIdx2,unsigned CommutableOpIdx1,unsigned CommutableOpIdx2)246 bool TargetInstrInfo::fixCommutedOpIndices(unsigned &ResultIdx1,
247                                            unsigned &ResultIdx2,
248                                            unsigned CommutableOpIdx1,
249                                            unsigned CommutableOpIdx2) {
250   if (ResultIdx1 == CommuteAnyOperandIndex &&
251       ResultIdx2 == CommuteAnyOperandIndex) {
252     ResultIdx1 = CommutableOpIdx1;
253     ResultIdx2 = CommutableOpIdx2;
254   } else if (ResultIdx1 == CommuteAnyOperandIndex) {
255     if (ResultIdx2 == CommutableOpIdx1)
256       ResultIdx1 = CommutableOpIdx2;
257     else if (ResultIdx2 == CommutableOpIdx2)
258       ResultIdx1 = CommutableOpIdx1;
259     else
260       return false;
261   } else if (ResultIdx2 == CommuteAnyOperandIndex) {
262     if (ResultIdx1 == CommutableOpIdx1)
263       ResultIdx2 = CommutableOpIdx2;
264     else if (ResultIdx1 == CommutableOpIdx2)
265       ResultIdx2 = CommutableOpIdx1;
266     else
267       return false;
268   } else
269     // Check that the result operand indices match the given commutable
270     // operand indices.
271     return (ResultIdx1 == CommutableOpIdx1 && ResultIdx2 == CommutableOpIdx2) ||
272            (ResultIdx1 == CommutableOpIdx2 && ResultIdx2 == CommutableOpIdx1);
273 
274   return true;
275 }
276 
findCommutedOpIndices(MachineInstr & MI,unsigned & SrcOpIdx1,unsigned & SrcOpIdx2) const277 bool TargetInstrInfo::findCommutedOpIndices(MachineInstr &MI,
278                                             unsigned &SrcOpIdx1,
279                                             unsigned &SrcOpIdx2) const {
280   assert(!MI.isBundle() &&
281          "TargetInstrInfo::findCommutedOpIndices() can't handle bundles");
282 
283   const MCInstrDesc &MCID = MI.getDesc();
284   if (!MCID.isCommutable())
285     return false;
286 
287   // This assumes v0 = op v1, v2 and commuting would swap v1 and v2. If this
288   // is not true, then the target must implement this.
289   unsigned CommutableOpIdx1 = MCID.getNumDefs();
290   unsigned CommutableOpIdx2 = CommutableOpIdx1 + 1;
291   if (!fixCommutedOpIndices(SrcOpIdx1, SrcOpIdx2,
292                             CommutableOpIdx1, CommutableOpIdx2))
293     return false;
294 
295   if (!MI.getOperand(SrcOpIdx1).isReg() || !MI.getOperand(SrcOpIdx2).isReg())
296     // No idea.
297     return false;
298   return true;
299 }
300 
isUnpredicatedTerminator(const MachineInstr & MI) const301 bool TargetInstrInfo::isUnpredicatedTerminator(const MachineInstr &MI) const {
302   if (!MI.isTerminator()) return false;
303 
304   // Conditional branch is a special case.
305   if (MI.isBranch() && !MI.isBarrier())
306     return true;
307   if (!MI.isPredicable())
308     return true;
309   return !isPredicated(MI);
310 }
311 
PredicateInstruction(MachineInstr & MI,ArrayRef<MachineOperand> Pred) const312 bool TargetInstrInfo::PredicateInstruction(
313     MachineInstr &MI, ArrayRef<MachineOperand> Pred) const {
314   bool MadeChange = false;
315 
316   assert(!MI.isBundle() &&
317          "TargetInstrInfo::PredicateInstruction() can't handle bundles");
318 
319   const MCInstrDesc &MCID = MI.getDesc();
320   if (!MI.isPredicable())
321     return false;
322 
323   for (unsigned j = 0, i = 0, e = MI.getNumOperands(); i != e; ++i) {
324     if (MCID.OpInfo[i].isPredicate()) {
325       MachineOperand &MO = MI.getOperand(i);
326       if (MO.isReg()) {
327         MO.setReg(Pred[j].getReg());
328         MadeChange = true;
329       } else if (MO.isImm()) {
330         MO.setImm(Pred[j].getImm());
331         MadeChange = true;
332       } else if (MO.isMBB()) {
333         MO.setMBB(Pred[j].getMBB());
334         MadeChange = true;
335       }
336       ++j;
337     }
338   }
339   return MadeChange;
340 }
341 
hasLoadFromStackSlot(const MachineInstr & MI,SmallVectorImpl<const MachineMemOperand * > & Accesses) const342 bool TargetInstrInfo::hasLoadFromStackSlot(
343     const MachineInstr &MI,
344     SmallVectorImpl<const MachineMemOperand *> &Accesses) const {
345   size_t StartSize = Accesses.size();
346   for (MachineInstr::mmo_iterator o = MI.memoperands_begin(),
347                                   oe = MI.memoperands_end();
348        o != oe; ++o) {
349     if ((*o)->isLoad() &&
350         dyn_cast_or_null<FixedStackPseudoSourceValue>((*o)->getPseudoValue()))
351       Accesses.push_back(*o);
352   }
353   return Accesses.size() != StartSize;
354 }
355 
hasStoreToStackSlot(const MachineInstr & MI,SmallVectorImpl<const MachineMemOperand * > & Accesses) const356 bool TargetInstrInfo::hasStoreToStackSlot(
357     const MachineInstr &MI,
358     SmallVectorImpl<const MachineMemOperand *> &Accesses) const {
359   size_t StartSize = Accesses.size();
360   for (MachineInstr::mmo_iterator o = MI.memoperands_begin(),
361                                   oe = MI.memoperands_end();
362        o != oe; ++o) {
363     if ((*o)->isStore() &&
364         dyn_cast_or_null<FixedStackPseudoSourceValue>((*o)->getPseudoValue()))
365       Accesses.push_back(*o);
366   }
367   return Accesses.size() != StartSize;
368 }
369 
getStackSlotRange(const TargetRegisterClass * RC,unsigned SubIdx,unsigned & Size,unsigned & Offset,const MachineFunction & MF) const370 bool TargetInstrInfo::getStackSlotRange(const TargetRegisterClass *RC,
371                                         unsigned SubIdx, unsigned &Size,
372                                         unsigned &Offset,
373                                         const MachineFunction &MF) const {
374   const TargetRegisterInfo *TRI = MF.getSubtarget().getRegisterInfo();
375   if (!SubIdx) {
376     Size = TRI->getSpillSize(*RC);
377     Offset = 0;
378     return true;
379   }
380   unsigned BitSize = TRI->getSubRegIdxSize(SubIdx);
381   // Convert bit size to byte size.
382   if (BitSize % 8)
383     return false;
384 
385   int BitOffset = TRI->getSubRegIdxOffset(SubIdx);
386   if (BitOffset < 0 || BitOffset % 8)
387     return false;
388 
389   Size = BitSize /= 8;
390   Offset = (unsigned)BitOffset / 8;
391 
392   assert(TRI->getSpillSize(*RC) >= (Offset + Size) && "bad subregister range");
393 
394   if (!MF.getDataLayout().isLittleEndian()) {
395     Offset = TRI->getSpillSize(*RC) - (Offset + Size);
396   }
397   return true;
398 }
399 
reMaterialize(MachineBasicBlock & MBB,MachineBasicBlock::iterator I,unsigned DestReg,unsigned SubIdx,const MachineInstr & Orig,const TargetRegisterInfo & TRI) const400 void TargetInstrInfo::reMaterialize(MachineBasicBlock &MBB,
401                                     MachineBasicBlock::iterator I,
402                                     unsigned DestReg, unsigned SubIdx,
403                                     const MachineInstr &Orig,
404                                     const TargetRegisterInfo &TRI) const {
405   MachineInstr *MI = MBB.getParent()->CloneMachineInstr(&Orig);
406   MI->substituteRegister(MI->getOperand(0).getReg(), DestReg, SubIdx, TRI);
407   MBB.insert(I, MI);
408 }
409 
produceSameValue(const MachineInstr & MI0,const MachineInstr & MI1,const MachineRegisterInfo * MRI) const410 bool TargetInstrInfo::produceSameValue(const MachineInstr &MI0,
411                                        const MachineInstr &MI1,
412                                        const MachineRegisterInfo *MRI) const {
413   return MI0.isIdenticalTo(MI1, MachineInstr::IgnoreVRegDefs);
414 }
415 
duplicate(MachineBasicBlock & MBB,MachineBasicBlock::iterator InsertBefore,const MachineInstr & Orig) const416 MachineInstr &TargetInstrInfo::duplicate(MachineBasicBlock &MBB,
417     MachineBasicBlock::iterator InsertBefore, const MachineInstr &Orig) const {
418   assert(!Orig.isNotDuplicable() && "Instruction cannot be duplicated");
419   MachineFunction &MF = *MBB.getParent();
420   return MF.CloneMachineInstrBundle(MBB, InsertBefore, Orig);
421 }
422 
423 // If the COPY instruction in MI can be folded to a stack operation, return
424 // the register class to use.
canFoldCopy(const MachineInstr & MI,unsigned FoldIdx)425 static const TargetRegisterClass *canFoldCopy(const MachineInstr &MI,
426                                               unsigned FoldIdx) {
427   assert(MI.isCopy() && "MI must be a COPY instruction");
428   if (MI.getNumOperands() != 2)
429     return nullptr;
430   assert(FoldIdx<2 && "FoldIdx refers no nonexistent operand");
431 
432   const MachineOperand &FoldOp = MI.getOperand(FoldIdx);
433   const MachineOperand &LiveOp = MI.getOperand(1 - FoldIdx);
434 
435   if (FoldOp.getSubReg() || LiveOp.getSubReg())
436     return nullptr;
437 
438   unsigned FoldReg = FoldOp.getReg();
439   unsigned LiveReg = LiveOp.getReg();
440 
441   assert(TargetRegisterInfo::isVirtualRegister(FoldReg) &&
442          "Cannot fold physregs");
443 
444   const MachineRegisterInfo &MRI = MI.getMF()->getRegInfo();
445   const TargetRegisterClass *RC = MRI.getRegClass(FoldReg);
446 
447   if (TargetRegisterInfo::isPhysicalRegister(LiveOp.getReg()))
448     return RC->contains(LiveOp.getReg()) ? RC : nullptr;
449 
450   if (RC->hasSubClassEq(MRI.getRegClass(LiveReg)))
451     return RC;
452 
453   // FIXME: Allow folding when register classes are memory compatible.
454   return nullptr;
455 }
456 
getNoop(MCInst & NopInst) const457 void TargetInstrInfo::getNoop(MCInst &NopInst) const {
458   llvm_unreachable("Not implemented");
459 }
460 
foldPatchpoint(MachineFunction & MF,MachineInstr & MI,ArrayRef<unsigned> Ops,int FrameIndex,const TargetInstrInfo & TII)461 static MachineInstr *foldPatchpoint(MachineFunction &MF, MachineInstr &MI,
462                                     ArrayRef<unsigned> Ops, int FrameIndex,
463                                     const TargetInstrInfo &TII) {
464   unsigned StartIdx = 0;
465   switch (MI.getOpcode()) {
466   case TargetOpcode::STACKMAP: {
467     // StackMapLiveValues are foldable
468     StartIdx = StackMapOpers(&MI).getVarIdx();
469     break;
470   }
471   case TargetOpcode::PATCHPOINT: {
472     // For PatchPoint, the call args are not foldable (even if reported in the
473     // stackmap e.g. via anyregcc).
474     StartIdx = PatchPointOpers(&MI).getVarIdx();
475     break;
476   }
477   case TargetOpcode::STATEPOINT: {
478     // For statepoints, fold deopt and gc arguments, but not call arguments.
479     StartIdx = StatepointOpers(&MI).getVarIdx();
480     break;
481   }
482   default:
483     llvm_unreachable("unexpected stackmap opcode");
484   }
485 
486   // Return false if any operands requested for folding are not foldable (not
487   // part of the stackmap's live values).
488   for (unsigned Op : Ops) {
489     if (Op < StartIdx)
490       return nullptr;
491   }
492 
493   MachineInstr *NewMI =
494       MF.CreateMachineInstr(TII.get(MI.getOpcode()), MI.getDebugLoc(), true);
495   MachineInstrBuilder MIB(MF, NewMI);
496 
497   // No need to fold return, the meta data, and function arguments
498   for (unsigned i = 0; i < StartIdx; ++i)
499     MIB.add(MI.getOperand(i));
500 
501   for (unsigned i = StartIdx; i < MI.getNumOperands(); ++i) {
502     MachineOperand &MO = MI.getOperand(i);
503     if (is_contained(Ops, i)) {
504       unsigned SpillSize;
505       unsigned SpillOffset;
506       // Compute the spill slot size and offset.
507       const TargetRegisterClass *RC =
508         MF.getRegInfo().getRegClass(MO.getReg());
509       bool Valid =
510           TII.getStackSlotRange(RC, MO.getSubReg(), SpillSize, SpillOffset, MF);
511       if (!Valid)
512         report_fatal_error("cannot spill patchpoint subregister operand");
513       MIB.addImm(StackMaps::IndirectMemRefOp);
514       MIB.addImm(SpillSize);
515       MIB.addFrameIndex(FrameIndex);
516       MIB.addImm(SpillOffset);
517     }
518     else
519       MIB.add(MO);
520   }
521   return NewMI;
522 }
523 
foldMemoryOperand(MachineInstr & MI,ArrayRef<unsigned> Ops,int FI,LiveIntervals * LIS) const524 MachineInstr *TargetInstrInfo::foldMemoryOperand(MachineInstr &MI,
525                                                  ArrayRef<unsigned> Ops, int FI,
526                                                  LiveIntervals *LIS) const {
527   auto Flags = MachineMemOperand::MONone;
528   for (unsigned OpIdx : Ops)
529     Flags |= MI.getOperand(OpIdx).isDef() ? MachineMemOperand::MOStore
530                                           : MachineMemOperand::MOLoad;
531 
532   MachineBasicBlock *MBB = MI.getParent();
533   assert(MBB && "foldMemoryOperand needs an inserted instruction");
534   MachineFunction &MF = *MBB->getParent();
535 
536   // If we're not folding a load into a subreg, the size of the load is the
537   // size of the spill slot. But if we are, we need to figure out what the
538   // actual load size is.
539   int64_t MemSize = 0;
540   const MachineFrameInfo &MFI = MF.getFrameInfo();
541   const TargetRegisterInfo *TRI = MF.getSubtarget().getRegisterInfo();
542 
543   if (Flags & MachineMemOperand::MOStore) {
544     MemSize = MFI.getObjectSize(FI);
545   } else {
546     for (unsigned OpIdx : Ops) {
547       int64_t OpSize = MFI.getObjectSize(FI);
548 
549       if (auto SubReg = MI.getOperand(OpIdx).getSubReg()) {
550         unsigned SubRegSize = TRI->getSubRegIdxSize(SubReg);
551         if (SubRegSize > 0 && !(SubRegSize % 8))
552           OpSize = SubRegSize / 8;
553       }
554 
555       MemSize = std::max(MemSize, OpSize);
556     }
557   }
558 
559   assert(MemSize && "Did not expect a zero-sized stack slot");
560 
561   MachineInstr *NewMI = nullptr;
562 
563   if (MI.getOpcode() == TargetOpcode::STACKMAP ||
564       MI.getOpcode() == TargetOpcode::PATCHPOINT ||
565       MI.getOpcode() == TargetOpcode::STATEPOINT) {
566     // Fold stackmap/patchpoint.
567     NewMI = foldPatchpoint(MF, MI, Ops, FI, *this);
568     if (NewMI)
569       MBB->insert(MI, NewMI);
570   } else {
571     // Ask the target to do the actual folding.
572     NewMI = foldMemoryOperandImpl(MF, MI, Ops, MI, FI, LIS);
573   }
574 
575   if (NewMI) {
576     NewMI->setMemRefs(MF, MI.memoperands());
577     // Add a memory operand, foldMemoryOperandImpl doesn't do that.
578     assert((!(Flags & MachineMemOperand::MOStore) ||
579             NewMI->mayStore()) &&
580            "Folded a def to a non-store!");
581     assert((!(Flags & MachineMemOperand::MOLoad) ||
582             NewMI->mayLoad()) &&
583            "Folded a use to a non-load!");
584     assert(MFI.getObjectOffset(FI) != -1);
585     MachineMemOperand *MMO = MF.getMachineMemOperand(
586         MachinePointerInfo::getFixedStack(MF, FI), Flags, MemSize,
587         MFI.getObjectAlignment(FI));
588     NewMI->addMemOperand(MF, MMO);
589 
590     return NewMI;
591   }
592 
593   // Straight COPY may fold as load/store.
594   if (!MI.isCopy() || Ops.size() != 1)
595     return nullptr;
596 
597   const TargetRegisterClass *RC = canFoldCopy(MI, Ops[0]);
598   if (!RC)
599     return nullptr;
600 
601   const MachineOperand &MO = MI.getOperand(1 - Ops[0]);
602   MachineBasicBlock::iterator Pos = MI;
603 
604   if (Flags == MachineMemOperand::MOStore)
605     storeRegToStackSlot(*MBB, Pos, MO.getReg(), MO.isKill(), FI, RC, TRI);
606   else
607     loadRegFromStackSlot(*MBB, Pos, MO.getReg(), FI, RC, TRI);
608   return &*--Pos;
609 }
610 
foldMemoryOperand(MachineInstr & MI,ArrayRef<unsigned> Ops,MachineInstr & LoadMI,LiveIntervals * LIS) const611 MachineInstr *TargetInstrInfo::foldMemoryOperand(MachineInstr &MI,
612                                                  ArrayRef<unsigned> Ops,
613                                                  MachineInstr &LoadMI,
614                                                  LiveIntervals *LIS) const {
615   assert(LoadMI.canFoldAsLoad() && "LoadMI isn't foldable!");
616 #ifndef NDEBUG
617   for (unsigned OpIdx : Ops)
618     assert(MI.getOperand(OpIdx).isUse() && "Folding load into def!");
619 #endif
620 
621   MachineBasicBlock &MBB = *MI.getParent();
622   MachineFunction &MF = *MBB.getParent();
623 
624   // Ask the target to do the actual folding.
625   MachineInstr *NewMI = nullptr;
626   int FrameIndex = 0;
627 
628   if ((MI.getOpcode() == TargetOpcode::STACKMAP ||
629        MI.getOpcode() == TargetOpcode::PATCHPOINT ||
630        MI.getOpcode() == TargetOpcode::STATEPOINT) &&
631       isLoadFromStackSlot(LoadMI, FrameIndex)) {
632     // Fold stackmap/patchpoint.
633     NewMI = foldPatchpoint(MF, MI, Ops, FrameIndex, *this);
634     if (NewMI)
635       NewMI = &*MBB.insert(MI, NewMI);
636   } else {
637     // Ask the target to do the actual folding.
638     NewMI = foldMemoryOperandImpl(MF, MI, Ops, MI, LoadMI, LIS);
639   }
640 
641   if (!NewMI)
642     return nullptr;
643 
644   // Copy the memoperands from the load to the folded instruction.
645   if (MI.memoperands_empty()) {
646     NewMI->setMemRefs(MF, LoadMI.memoperands());
647   } else {
648     // Handle the rare case of folding multiple loads.
649     NewMI->setMemRefs(MF, MI.memoperands());
650     for (MachineInstr::mmo_iterator I = LoadMI.memoperands_begin(),
651                                     E = LoadMI.memoperands_end();
652          I != E; ++I) {
653       NewMI->addMemOperand(MF, *I);
654     }
655   }
656   return NewMI;
657 }
658 
hasReassociableOperands(const MachineInstr & Inst,const MachineBasicBlock * MBB) const659 bool TargetInstrInfo::hasReassociableOperands(
660     const MachineInstr &Inst, const MachineBasicBlock *MBB) const {
661   const MachineOperand &Op1 = Inst.getOperand(1);
662   const MachineOperand &Op2 = Inst.getOperand(2);
663   const MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo();
664 
665   // We need virtual register definitions for the operands that we will
666   // reassociate.
667   MachineInstr *MI1 = nullptr;
668   MachineInstr *MI2 = nullptr;
669   if (Op1.isReg() && TargetRegisterInfo::isVirtualRegister(Op1.getReg()))
670     MI1 = MRI.getUniqueVRegDef(Op1.getReg());
671   if (Op2.isReg() && TargetRegisterInfo::isVirtualRegister(Op2.getReg()))
672     MI2 = MRI.getUniqueVRegDef(Op2.getReg());
673 
674   // And they need to be in the trace (otherwise, they won't have a depth).
675   return MI1 && MI2 && MI1->getParent() == MBB && MI2->getParent() == MBB;
676 }
677 
hasReassociableSibling(const MachineInstr & Inst,bool & Commuted) const678 bool TargetInstrInfo::hasReassociableSibling(const MachineInstr &Inst,
679                                              bool &Commuted) const {
680   const MachineBasicBlock *MBB = Inst.getParent();
681   const MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo();
682   MachineInstr *MI1 = MRI.getUniqueVRegDef(Inst.getOperand(1).getReg());
683   MachineInstr *MI2 = MRI.getUniqueVRegDef(Inst.getOperand(2).getReg());
684   unsigned AssocOpcode = Inst.getOpcode();
685 
686   // If only one operand has the same opcode and it's the second source operand,
687   // the operands must be commuted.
688   Commuted = MI1->getOpcode() != AssocOpcode && MI2->getOpcode() == AssocOpcode;
689   if (Commuted)
690     std::swap(MI1, MI2);
691 
692   // 1. The previous instruction must be the same type as Inst.
693   // 2. The previous instruction must have virtual register definitions for its
694   //    operands in the same basic block as Inst.
695   // 3. The previous instruction's result must only be used by Inst.
696   return MI1->getOpcode() == AssocOpcode &&
697          hasReassociableOperands(*MI1, MBB) &&
698          MRI.hasOneNonDBGUse(MI1->getOperand(0).getReg());
699 }
700 
701 // 1. The operation must be associative and commutative.
702 // 2. The instruction must have virtual register definitions for its
703 //    operands in the same basic block.
704 // 3. The instruction must have a reassociable sibling.
isReassociationCandidate(const MachineInstr & Inst,bool & Commuted) const705 bool TargetInstrInfo::isReassociationCandidate(const MachineInstr &Inst,
706                                                bool &Commuted) const {
707   return isAssociativeAndCommutative(Inst) &&
708          hasReassociableOperands(Inst, Inst.getParent()) &&
709          hasReassociableSibling(Inst, Commuted);
710 }
711 
712 // The concept of the reassociation pass is that these operations can benefit
713 // from this kind of transformation:
714 //
715 // A = ? op ?
716 // B = A op X (Prev)
717 // C = B op Y (Root)
718 // -->
719 // A = ? op ?
720 // B = X op Y
721 // C = A op B
722 //
723 // breaking the dependency between A and B, allowing them to be executed in
724 // parallel (or back-to-back in a pipeline) instead of depending on each other.
725 
726 // FIXME: This has the potential to be expensive (compile time) while not
727 // improving the code at all. Some ways to limit the overhead:
728 // 1. Track successful transforms; bail out if hit rate gets too low.
729 // 2. Only enable at -O3 or some other non-default optimization level.
730 // 3. Pre-screen pattern candidates here: if an operand of the previous
731 //    instruction is known to not increase the critical path, then don't match
732 //    that pattern.
getMachineCombinerPatterns(MachineInstr & Root,SmallVectorImpl<MachineCombinerPattern> & Patterns) const733 bool TargetInstrInfo::getMachineCombinerPatterns(
734     MachineInstr &Root,
735     SmallVectorImpl<MachineCombinerPattern> &Patterns) const {
736   bool Commute;
737   if (isReassociationCandidate(Root, Commute)) {
738     // We found a sequence of instructions that may be suitable for a
739     // reassociation of operands to increase ILP. Specify each commutation
740     // possibility for the Prev instruction in the sequence and let the
741     // machine combiner decide if changing the operands is worthwhile.
742     if (Commute) {
743       Patterns.push_back(MachineCombinerPattern::REASSOC_AX_YB);
744       Patterns.push_back(MachineCombinerPattern::REASSOC_XA_YB);
745     } else {
746       Patterns.push_back(MachineCombinerPattern::REASSOC_AX_BY);
747       Patterns.push_back(MachineCombinerPattern::REASSOC_XA_BY);
748     }
749     return true;
750   }
751 
752   return false;
753 }
754 
755 /// Return true when a code sequence can improve loop throughput.
756 bool
isThroughputPattern(MachineCombinerPattern Pattern) const757 TargetInstrInfo::isThroughputPattern(MachineCombinerPattern Pattern) const {
758   return false;
759 }
760 
761 /// Attempt the reassociation transformation to reduce critical path length.
762 /// See the above comments before getMachineCombinerPatterns().
reassociateOps(MachineInstr & Root,MachineInstr & Prev,MachineCombinerPattern Pattern,SmallVectorImpl<MachineInstr * > & InsInstrs,SmallVectorImpl<MachineInstr * > & DelInstrs,DenseMap<unsigned,unsigned> & InstrIdxForVirtReg) const763 void TargetInstrInfo::reassociateOps(
764     MachineInstr &Root, MachineInstr &Prev,
765     MachineCombinerPattern Pattern,
766     SmallVectorImpl<MachineInstr *> &InsInstrs,
767     SmallVectorImpl<MachineInstr *> &DelInstrs,
768     DenseMap<unsigned, unsigned> &InstrIdxForVirtReg) const {
769   MachineFunction *MF = Root.getMF();
770   MachineRegisterInfo &MRI = MF->getRegInfo();
771   const TargetInstrInfo *TII = MF->getSubtarget().getInstrInfo();
772   const TargetRegisterInfo *TRI = MF->getSubtarget().getRegisterInfo();
773   const TargetRegisterClass *RC = Root.getRegClassConstraint(0, TII, TRI);
774 
775   // This array encodes the operand index for each parameter because the
776   // operands may be commuted. Each row corresponds to a pattern value,
777   // and each column specifies the index of A, B, X, Y.
778   unsigned OpIdx[4][4] = {
779     { 1, 1, 2, 2 },
780     { 1, 2, 2, 1 },
781     { 2, 1, 1, 2 },
782     { 2, 2, 1, 1 }
783   };
784 
785   int Row;
786   switch (Pattern) {
787   case MachineCombinerPattern::REASSOC_AX_BY: Row = 0; break;
788   case MachineCombinerPattern::REASSOC_AX_YB: Row = 1; break;
789   case MachineCombinerPattern::REASSOC_XA_BY: Row = 2; break;
790   case MachineCombinerPattern::REASSOC_XA_YB: Row = 3; break;
791   default: llvm_unreachable("unexpected MachineCombinerPattern");
792   }
793 
794   MachineOperand &OpA = Prev.getOperand(OpIdx[Row][0]);
795   MachineOperand &OpB = Root.getOperand(OpIdx[Row][1]);
796   MachineOperand &OpX = Prev.getOperand(OpIdx[Row][2]);
797   MachineOperand &OpY = Root.getOperand(OpIdx[Row][3]);
798   MachineOperand &OpC = Root.getOperand(0);
799 
800   unsigned RegA = OpA.getReg();
801   unsigned RegB = OpB.getReg();
802   unsigned RegX = OpX.getReg();
803   unsigned RegY = OpY.getReg();
804   unsigned RegC = OpC.getReg();
805 
806   if (TargetRegisterInfo::isVirtualRegister(RegA))
807     MRI.constrainRegClass(RegA, RC);
808   if (TargetRegisterInfo::isVirtualRegister(RegB))
809     MRI.constrainRegClass(RegB, RC);
810   if (TargetRegisterInfo::isVirtualRegister(RegX))
811     MRI.constrainRegClass(RegX, RC);
812   if (TargetRegisterInfo::isVirtualRegister(RegY))
813     MRI.constrainRegClass(RegY, RC);
814   if (TargetRegisterInfo::isVirtualRegister(RegC))
815     MRI.constrainRegClass(RegC, RC);
816 
817   // Create a new virtual register for the result of (X op Y) instead of
818   // recycling RegB because the MachineCombiner's computation of the critical
819   // path requires a new register definition rather than an existing one.
820   unsigned NewVR = MRI.createVirtualRegister(RC);
821   InstrIdxForVirtReg.insert(std::make_pair(NewVR, 0));
822 
823   unsigned Opcode = Root.getOpcode();
824   bool KillA = OpA.isKill();
825   bool KillX = OpX.isKill();
826   bool KillY = OpY.isKill();
827 
828   // Create new instructions for insertion.
829   MachineInstrBuilder MIB1 =
830       BuildMI(*MF, Prev.getDebugLoc(), TII->get(Opcode), NewVR)
831           .addReg(RegX, getKillRegState(KillX))
832           .addReg(RegY, getKillRegState(KillY));
833   MachineInstrBuilder MIB2 =
834       BuildMI(*MF, Root.getDebugLoc(), TII->get(Opcode), RegC)
835           .addReg(RegA, getKillRegState(KillA))
836           .addReg(NewVR, getKillRegState(true));
837 
838   setSpecialOperandAttr(Root, Prev, *MIB1, *MIB2);
839 
840   // Record new instructions for insertion and old instructions for deletion.
841   InsInstrs.push_back(MIB1);
842   InsInstrs.push_back(MIB2);
843   DelInstrs.push_back(&Prev);
844   DelInstrs.push_back(&Root);
845 }
846 
genAlternativeCodeSequence(MachineInstr & Root,MachineCombinerPattern Pattern,SmallVectorImpl<MachineInstr * > & InsInstrs,SmallVectorImpl<MachineInstr * > & DelInstrs,DenseMap<unsigned,unsigned> & InstIdxForVirtReg) const847 void TargetInstrInfo::genAlternativeCodeSequence(
848     MachineInstr &Root, MachineCombinerPattern Pattern,
849     SmallVectorImpl<MachineInstr *> &InsInstrs,
850     SmallVectorImpl<MachineInstr *> &DelInstrs,
851     DenseMap<unsigned, unsigned> &InstIdxForVirtReg) const {
852   MachineRegisterInfo &MRI = Root.getMF()->getRegInfo();
853 
854   // Select the previous instruction in the sequence based on the input pattern.
855   MachineInstr *Prev = nullptr;
856   switch (Pattern) {
857   case MachineCombinerPattern::REASSOC_AX_BY:
858   case MachineCombinerPattern::REASSOC_XA_BY:
859     Prev = MRI.getUniqueVRegDef(Root.getOperand(1).getReg());
860     break;
861   case MachineCombinerPattern::REASSOC_AX_YB:
862   case MachineCombinerPattern::REASSOC_XA_YB:
863     Prev = MRI.getUniqueVRegDef(Root.getOperand(2).getReg());
864     break;
865   default:
866     break;
867   }
868 
869   assert(Prev && "Unknown pattern for machine combiner");
870 
871   reassociateOps(Root, *Prev, Pattern, InsInstrs, DelInstrs, InstIdxForVirtReg);
872 }
873 
isReallyTriviallyReMaterializableGeneric(const MachineInstr & MI,AliasAnalysis * AA) const874 bool TargetInstrInfo::isReallyTriviallyReMaterializableGeneric(
875     const MachineInstr &MI, AliasAnalysis *AA) const {
876   const MachineFunction &MF = *MI.getMF();
877   const MachineRegisterInfo &MRI = MF.getRegInfo();
878 
879   // Remat clients assume operand 0 is the defined register.
880   if (!MI.getNumOperands() || !MI.getOperand(0).isReg())
881     return false;
882   unsigned DefReg = MI.getOperand(0).getReg();
883 
884   // A sub-register definition can only be rematerialized if the instruction
885   // doesn't read the other parts of the register.  Otherwise it is really a
886   // read-modify-write operation on the full virtual register which cannot be
887   // moved safely.
888   if (TargetRegisterInfo::isVirtualRegister(DefReg) &&
889       MI.getOperand(0).getSubReg() && MI.readsVirtualRegister(DefReg))
890     return false;
891 
892   // A load from a fixed stack slot can be rematerialized. This may be
893   // redundant with subsequent checks, but it's target-independent,
894   // simple, and a common case.
895   int FrameIdx = 0;
896   if (isLoadFromStackSlot(MI, FrameIdx) &&
897       MF.getFrameInfo().isImmutableObjectIndex(FrameIdx))
898     return true;
899 
900   // Avoid instructions obviously unsafe for remat.
901   if (MI.isNotDuplicable() || MI.mayStore() || MI.hasUnmodeledSideEffects())
902     return false;
903 
904   // Don't remat inline asm. We have no idea how expensive it is
905   // even if it's side effect free.
906   if (MI.isInlineAsm())
907     return false;
908 
909   // Avoid instructions which load from potentially varying memory.
910   if (MI.mayLoad() && !MI.isDereferenceableInvariantLoad(AA))
911     return false;
912 
913   // If any of the registers accessed are non-constant, conservatively assume
914   // the instruction is not rematerializable.
915   for (unsigned i = 0, e = MI.getNumOperands(); i != e; ++i) {
916     const MachineOperand &MO = MI.getOperand(i);
917     if (!MO.isReg()) continue;
918     unsigned Reg = MO.getReg();
919     if (Reg == 0)
920       continue;
921 
922     // Check for a well-behaved physical register.
923     if (TargetRegisterInfo::isPhysicalRegister(Reg)) {
924       if (MO.isUse()) {
925         // If the physreg has no defs anywhere, it's just an ambient register
926         // and we can freely move its uses. Alternatively, if it's allocatable,
927         // it could get allocated to something with a def during allocation.
928         if (!MRI.isConstantPhysReg(Reg))
929           return false;
930       } else {
931         // A physreg def. We can't remat it.
932         return false;
933       }
934       continue;
935     }
936 
937     // Only allow one virtual-register def.  There may be multiple defs of the
938     // same virtual register, though.
939     if (MO.isDef() && Reg != DefReg)
940       return false;
941 
942     // Don't allow any virtual-register uses. Rematting an instruction with
943     // virtual register uses would length the live ranges of the uses, which
944     // is not necessarily a good idea, certainly not "trivial".
945     if (MO.isUse())
946       return false;
947   }
948 
949   // Everything checked out.
950   return true;
951 }
952 
getSPAdjust(const MachineInstr & MI) const953 int TargetInstrInfo::getSPAdjust(const MachineInstr &MI) const {
954   const MachineFunction *MF = MI.getMF();
955   const TargetFrameLowering *TFI = MF->getSubtarget().getFrameLowering();
956   bool StackGrowsDown =
957     TFI->getStackGrowthDirection() == TargetFrameLowering::StackGrowsDown;
958 
959   unsigned FrameSetupOpcode = getCallFrameSetupOpcode();
960   unsigned FrameDestroyOpcode = getCallFrameDestroyOpcode();
961 
962   if (!isFrameInstr(MI))
963     return 0;
964 
965   int SPAdj = TFI->alignSPAdjust(getFrameSize(MI));
966 
967   if ((!StackGrowsDown && MI.getOpcode() == FrameSetupOpcode) ||
968       (StackGrowsDown && MI.getOpcode() == FrameDestroyOpcode))
969     SPAdj = -SPAdj;
970 
971   return SPAdj;
972 }
973 
974 /// isSchedulingBoundary - Test if the given instruction should be
975 /// considered a scheduling boundary. This primarily includes labels
976 /// and terminators.
isSchedulingBoundary(const MachineInstr & MI,const MachineBasicBlock * MBB,const MachineFunction & MF) const977 bool TargetInstrInfo::isSchedulingBoundary(const MachineInstr &MI,
978                                            const MachineBasicBlock *MBB,
979                                            const MachineFunction &MF) const {
980   // Terminators and labels can't be scheduled around.
981   if (MI.isTerminator() || MI.isPosition())
982     return true;
983 
984   // Don't attempt to schedule around any instruction that defines
985   // a stack-oriented pointer, as it's unlikely to be profitable. This
986   // saves compile time, because it doesn't require every single
987   // stack slot reference to depend on the instruction that does the
988   // modification.
989   const TargetLowering &TLI = *MF.getSubtarget().getTargetLowering();
990   const TargetRegisterInfo *TRI = MF.getSubtarget().getRegisterInfo();
991   return MI.modifiesRegister(TLI.getStackPointerRegisterToSaveRestore(), TRI);
992 }
993 
994 // Provide a global flag for disabling the PreRA hazard recognizer that targets
995 // may choose to honor.
usePreRAHazardRecognizer() const996 bool TargetInstrInfo::usePreRAHazardRecognizer() const {
997   return !DisableHazardRecognizer;
998 }
999 
1000 // Default implementation of CreateTargetRAHazardRecognizer.
1001 ScheduleHazardRecognizer *TargetInstrInfo::
CreateTargetHazardRecognizer(const TargetSubtargetInfo * STI,const ScheduleDAG * DAG) const1002 CreateTargetHazardRecognizer(const TargetSubtargetInfo *STI,
1003                              const ScheduleDAG *DAG) const {
1004   // Dummy hazard recognizer allows all instructions to issue.
1005   return new ScheduleHazardRecognizer();
1006 }
1007 
1008 // Default implementation of CreateTargetMIHazardRecognizer.
1009 ScheduleHazardRecognizer *TargetInstrInfo::
CreateTargetMIHazardRecognizer(const InstrItineraryData * II,const ScheduleDAG * DAG) const1010 CreateTargetMIHazardRecognizer(const InstrItineraryData *II,
1011                                const ScheduleDAG *DAG) const {
1012   return (ScheduleHazardRecognizer *)
1013     new ScoreboardHazardRecognizer(II, DAG, "misched");
1014 }
1015 
1016 // Default implementation of CreateTargetPostRAHazardRecognizer.
1017 ScheduleHazardRecognizer *TargetInstrInfo::
CreateTargetPostRAHazardRecognizer(const InstrItineraryData * II,const ScheduleDAG * DAG) const1018 CreateTargetPostRAHazardRecognizer(const InstrItineraryData *II,
1019                                    const ScheduleDAG *DAG) const {
1020   return (ScheduleHazardRecognizer *)
1021     new ScoreboardHazardRecognizer(II, DAG, "post-RA-sched");
1022 }
1023 
1024 //===----------------------------------------------------------------------===//
1025 //  SelectionDAG latency interface.
1026 //===----------------------------------------------------------------------===//
1027 
1028 int
getOperandLatency(const InstrItineraryData * ItinData,SDNode * DefNode,unsigned DefIdx,SDNode * UseNode,unsigned UseIdx) const1029 TargetInstrInfo::getOperandLatency(const InstrItineraryData *ItinData,
1030                                    SDNode *DefNode, unsigned DefIdx,
1031                                    SDNode *UseNode, unsigned UseIdx) const {
1032   if (!ItinData || ItinData->isEmpty())
1033     return -1;
1034 
1035   if (!DefNode->isMachineOpcode())
1036     return -1;
1037 
1038   unsigned DefClass = get(DefNode->getMachineOpcode()).getSchedClass();
1039   if (!UseNode->isMachineOpcode())
1040     return ItinData->getOperandCycle(DefClass, DefIdx);
1041   unsigned UseClass = get(UseNode->getMachineOpcode()).getSchedClass();
1042   return ItinData->getOperandLatency(DefClass, DefIdx, UseClass, UseIdx);
1043 }
1044 
getInstrLatency(const InstrItineraryData * ItinData,SDNode * N) const1045 int TargetInstrInfo::getInstrLatency(const InstrItineraryData *ItinData,
1046                                      SDNode *N) const {
1047   if (!ItinData || ItinData->isEmpty())
1048     return 1;
1049 
1050   if (!N->isMachineOpcode())
1051     return 1;
1052 
1053   return ItinData->getStageLatency(get(N->getMachineOpcode()).getSchedClass());
1054 }
1055 
1056 //===----------------------------------------------------------------------===//
1057 //  MachineInstr latency interface.
1058 //===----------------------------------------------------------------------===//
1059 
getNumMicroOps(const InstrItineraryData * ItinData,const MachineInstr & MI) const1060 unsigned TargetInstrInfo::getNumMicroOps(const InstrItineraryData *ItinData,
1061                                          const MachineInstr &MI) const {
1062   if (!ItinData || ItinData->isEmpty())
1063     return 1;
1064 
1065   unsigned Class = MI.getDesc().getSchedClass();
1066   int UOps = ItinData->Itineraries[Class].NumMicroOps;
1067   if (UOps >= 0)
1068     return UOps;
1069 
1070   // The # of u-ops is dynamically determined. The specific target should
1071   // override this function to return the right number.
1072   return 1;
1073 }
1074 
1075 /// Return the default expected latency for a def based on it's opcode.
defaultDefLatency(const MCSchedModel & SchedModel,const MachineInstr & DefMI) const1076 unsigned TargetInstrInfo::defaultDefLatency(const MCSchedModel &SchedModel,
1077                                             const MachineInstr &DefMI) const {
1078   if (DefMI.isTransient())
1079     return 0;
1080   if (DefMI.mayLoad())
1081     return SchedModel.LoadLatency;
1082   if (isHighLatencyDef(DefMI.getOpcode()))
1083     return SchedModel.HighLatency;
1084   return 1;
1085 }
1086 
getPredicationCost(const MachineInstr &) const1087 unsigned TargetInstrInfo::getPredicationCost(const MachineInstr &) const {
1088   return 0;
1089 }
1090 
getInstrLatency(const InstrItineraryData * ItinData,const MachineInstr & MI,unsigned * PredCost) const1091 unsigned TargetInstrInfo::getInstrLatency(const InstrItineraryData *ItinData,
1092                                           const MachineInstr &MI,
1093                                           unsigned *PredCost) const {
1094   // Default to one cycle for no itinerary. However, an "empty" itinerary may
1095   // still have a MinLatency property, which getStageLatency checks.
1096   if (!ItinData)
1097     return MI.mayLoad() ? 2 : 1;
1098 
1099   return ItinData->getStageLatency(MI.getDesc().getSchedClass());
1100 }
1101 
hasLowDefLatency(const TargetSchedModel & SchedModel,const MachineInstr & DefMI,unsigned DefIdx) const1102 bool TargetInstrInfo::hasLowDefLatency(const TargetSchedModel &SchedModel,
1103                                        const MachineInstr &DefMI,
1104                                        unsigned DefIdx) const {
1105   const InstrItineraryData *ItinData = SchedModel.getInstrItineraries();
1106   if (!ItinData || ItinData->isEmpty())
1107     return false;
1108 
1109   unsigned DefClass = DefMI.getDesc().getSchedClass();
1110   int DefCycle = ItinData->getOperandCycle(DefClass, DefIdx);
1111   return (DefCycle != -1 && DefCycle <= 1);
1112 }
1113 
1114 /// Both DefMI and UseMI must be valid.  By default, call directly to the
1115 /// itinerary. This may be overriden by the target.
getOperandLatency(const InstrItineraryData * ItinData,const MachineInstr & DefMI,unsigned DefIdx,const MachineInstr & UseMI,unsigned UseIdx) const1116 int TargetInstrInfo::getOperandLatency(const InstrItineraryData *ItinData,
1117                                        const MachineInstr &DefMI,
1118                                        unsigned DefIdx,
1119                                        const MachineInstr &UseMI,
1120                                        unsigned UseIdx) const {
1121   unsigned DefClass = DefMI.getDesc().getSchedClass();
1122   unsigned UseClass = UseMI.getDesc().getSchedClass();
1123   return ItinData->getOperandLatency(DefClass, DefIdx, UseClass, UseIdx);
1124 }
1125 
1126 /// If we can determine the operand latency from the def only, without itinerary
1127 /// lookup, do so. Otherwise return -1.
computeDefOperandLatency(const InstrItineraryData * ItinData,const MachineInstr & DefMI) const1128 int TargetInstrInfo::computeDefOperandLatency(
1129     const InstrItineraryData *ItinData, const MachineInstr &DefMI) const {
1130 
1131   // Let the target hook getInstrLatency handle missing itineraries.
1132   if (!ItinData)
1133     return getInstrLatency(ItinData, DefMI);
1134 
1135   if(ItinData->isEmpty())
1136     return defaultDefLatency(ItinData->SchedModel, DefMI);
1137 
1138   // ...operand lookup required
1139   return -1;
1140 }
1141 
getRegSequenceInputs(const MachineInstr & MI,unsigned DefIdx,SmallVectorImpl<RegSubRegPairAndIdx> & InputRegs) const1142 bool TargetInstrInfo::getRegSequenceInputs(
1143     const MachineInstr &MI, unsigned DefIdx,
1144     SmallVectorImpl<RegSubRegPairAndIdx> &InputRegs) const {
1145   assert((MI.isRegSequence() ||
1146           MI.isRegSequenceLike()) && "Instruction do not have the proper type");
1147 
1148   if (!MI.isRegSequence())
1149     return getRegSequenceLikeInputs(MI, DefIdx, InputRegs);
1150 
1151   // We are looking at:
1152   // Def = REG_SEQUENCE v0, sub0, v1, sub1, ...
1153   assert(DefIdx == 0 && "REG_SEQUENCE only has one def");
1154   for (unsigned OpIdx = 1, EndOpIdx = MI.getNumOperands(); OpIdx != EndOpIdx;
1155        OpIdx += 2) {
1156     const MachineOperand &MOReg = MI.getOperand(OpIdx);
1157     if (MOReg.isUndef())
1158       continue;
1159     const MachineOperand &MOSubIdx = MI.getOperand(OpIdx + 1);
1160     assert(MOSubIdx.isImm() &&
1161            "One of the subindex of the reg_sequence is not an immediate");
1162     // Record Reg:SubReg, SubIdx.
1163     InputRegs.push_back(RegSubRegPairAndIdx(MOReg.getReg(), MOReg.getSubReg(),
1164                                             (unsigned)MOSubIdx.getImm()));
1165   }
1166   return true;
1167 }
1168 
getExtractSubregInputs(const MachineInstr & MI,unsigned DefIdx,RegSubRegPairAndIdx & InputReg) const1169 bool TargetInstrInfo::getExtractSubregInputs(
1170     const MachineInstr &MI, unsigned DefIdx,
1171     RegSubRegPairAndIdx &InputReg) const {
1172   assert((MI.isExtractSubreg() ||
1173       MI.isExtractSubregLike()) && "Instruction do not have the proper type");
1174 
1175   if (!MI.isExtractSubreg())
1176     return getExtractSubregLikeInputs(MI, DefIdx, InputReg);
1177 
1178   // We are looking at:
1179   // Def = EXTRACT_SUBREG v0.sub1, sub0.
1180   assert(DefIdx == 0 && "EXTRACT_SUBREG only has one def");
1181   const MachineOperand &MOReg = MI.getOperand(1);
1182   if (MOReg.isUndef())
1183     return false;
1184   const MachineOperand &MOSubIdx = MI.getOperand(2);
1185   assert(MOSubIdx.isImm() &&
1186          "The subindex of the extract_subreg is not an immediate");
1187 
1188   InputReg.Reg = MOReg.getReg();
1189   InputReg.SubReg = MOReg.getSubReg();
1190   InputReg.SubIdx = (unsigned)MOSubIdx.getImm();
1191   return true;
1192 }
1193 
getInsertSubregInputs(const MachineInstr & MI,unsigned DefIdx,RegSubRegPair & BaseReg,RegSubRegPairAndIdx & InsertedReg) const1194 bool TargetInstrInfo::getInsertSubregInputs(
1195     const MachineInstr &MI, unsigned DefIdx,
1196     RegSubRegPair &BaseReg, RegSubRegPairAndIdx &InsertedReg) const {
1197   assert((MI.isInsertSubreg() ||
1198       MI.isInsertSubregLike()) && "Instruction do not have the proper type");
1199 
1200   if (!MI.isInsertSubreg())
1201     return getInsertSubregLikeInputs(MI, DefIdx, BaseReg, InsertedReg);
1202 
1203   // We are looking at:
1204   // Def = INSERT_SEQUENCE v0, v1, sub0.
1205   assert(DefIdx == 0 && "INSERT_SUBREG only has one def");
1206   const MachineOperand &MOBaseReg = MI.getOperand(1);
1207   const MachineOperand &MOInsertedReg = MI.getOperand(2);
1208   if (MOInsertedReg.isUndef())
1209     return false;
1210   const MachineOperand &MOSubIdx = MI.getOperand(3);
1211   assert(MOSubIdx.isImm() &&
1212          "One of the subindex of the reg_sequence is not an immediate");
1213   BaseReg.Reg = MOBaseReg.getReg();
1214   BaseReg.SubReg = MOBaseReg.getSubReg();
1215 
1216   InsertedReg.Reg = MOInsertedReg.getReg();
1217   InsertedReg.SubReg = MOInsertedReg.getSubReg();
1218   InsertedReg.SubIdx = (unsigned)MOSubIdx.getImm();
1219   return true;
1220 }
1221