• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 //===- DetectDeadLanes.cpp - SubRegister Lane Usage Analysis --*- C++ -*---===//
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 /// \file
11 /// Analysis that tracks defined/used subregister lanes across COPY instructions
12 /// and instructions that get lowered to a COPY (PHI, REG_SEQUENCE,
13 /// INSERT_SUBREG, EXTRACT_SUBREG).
14 /// The information is used to detect dead definitions and the usage of
15 /// (completely) undefined values and mark the operands as such.
16 /// This pass is necessary because the dead/undef status is not obvious anymore
17 /// when subregisters are involved.
18 ///
19 /// Example:
20 ///    %vreg0 = some definition
21 ///    %vreg1 = IMPLICIT_DEF
22 ///    %vreg2 = REG_SEQUENCE %vreg0, sub0, %vreg1, sub1
23 ///    %vreg3 = EXTRACT_SUBREG %vreg2, sub1
24 ///           = use %vreg3
25 /// The %vreg0 definition is dead and %vreg3 contains an undefined value.
26 //
27 //===----------------------------------------------------------------------===//
28 
29 #include <deque>
30 #include <vector>
31 
32 #include "llvm/ADT/BitVector.h"
33 #include "llvm/ADT/SetVector.h"
34 #include "llvm/CodeGen/MachineFunctionPass.h"
35 #include "llvm/CodeGen/MachineRegisterInfo.h"
36 #include "llvm/CodeGen/Passes.h"
37 #include "llvm/InitializePasses.h"
38 #include "llvm/Pass.h"
39 #include "llvm/PassRegistry.h"
40 #include "llvm/Support/Debug.h"
41 #include "llvm/Support/raw_ostream.h"
42 #include "llvm/Target/TargetInstrInfo.h"
43 #include "llvm/Target/TargetRegisterInfo.h"
44 #include "llvm/Target/TargetSubtargetInfo.h"
45 
46 using namespace llvm;
47 
48 #define DEBUG_TYPE "detect-dead-lanes"
49 
50 namespace {
51 
52 /// Contains a bitmask of which lanes of a given virtual register are
53 /// defined and which ones are actually used.
54 struct VRegInfo {
55   LaneBitmask UsedLanes;
56   LaneBitmask DefinedLanes;
57 };
58 
59 class DetectDeadLanes : public MachineFunctionPass {
60 public:
61   bool runOnMachineFunction(MachineFunction &MF) override;
62 
63   static char ID;
DetectDeadLanes()64   DetectDeadLanes() : MachineFunctionPass(ID) {}
65 
getPassName() const66   const char *getPassName() const override { return "Detect Dead Lanes"; }
67 
getAnalysisUsage(AnalysisUsage & AU) const68   void getAnalysisUsage(AnalysisUsage &AU) const override {
69     AU.setPreservesCFG();
70     MachineFunctionPass::getAnalysisUsage(AU);
71   }
72 
73 private:
74   /// Add used lane bits on the register used by operand \p MO. This translates
75   /// the bitmask based on the operands subregister, and puts the register into
76   /// the worklist if any new bits were added.
77   void addUsedLanesOnOperand(const MachineOperand &MO, LaneBitmask UsedLanes);
78 
79   /// Given a bitmask \p UsedLanes for the used lanes on a def output of a
80   /// COPY-like instruction determine the lanes used on the use operands
81   /// and call addUsedLanesOnOperand() for them.
82   void transferUsedLanesStep(const MachineInstr &MI, LaneBitmask UsedLanes);
83 
84   /// Given a use regiser operand \p Use and a mask of defined lanes, check
85   /// if the operand belongs to a lowersToCopies() instruction, transfer the
86   /// mask to the def and put the instruction into the worklist.
87   void transferDefinedLanesStep(const MachineOperand &Use,
88                                 LaneBitmask DefinedLanes);
89 
90   /// Given a mask \p DefinedLanes of lanes defined at operand \p OpNum
91   /// of COPY-like instruction, determine which lanes are defined at the output
92   /// operand \p Def.
93   LaneBitmask transferDefinedLanes(const MachineOperand &Def, unsigned OpNum,
94                                    LaneBitmask DefinedLanes) const;
95 
96   /// Given a mask \p UsedLanes used from the output of instruction \p MI
97   /// determine which lanes are used from operand \p MO of this instruction.
98   LaneBitmask transferUsedLanes(const MachineInstr &MI, LaneBitmask UsedLanes,
99                                 const MachineOperand &MO) const;
100 
101   bool runOnce(MachineFunction &MF);
102 
103   LaneBitmask determineInitialDefinedLanes(unsigned Reg);
104   LaneBitmask determineInitialUsedLanes(unsigned Reg);
105 
106   bool isUndefRegAtInput(const MachineOperand &MO,
107                          const VRegInfo &RegInfo) const;
108 
109   bool isUndefInput(const MachineOperand &MO, bool *CrossCopy) const;
110 
111   const MachineRegisterInfo *MRI;
112   const TargetRegisterInfo *TRI;
113 
PutInWorklist(unsigned RegIdx)114   void PutInWorklist(unsigned RegIdx) {
115     if (WorklistMembers.test(RegIdx))
116       return;
117     WorklistMembers.set(RegIdx);
118     Worklist.push_back(RegIdx);
119   }
120 
121   VRegInfo *VRegInfos;
122   /// Worklist containing virtreg indexes.
123   std::deque<unsigned> Worklist;
124   BitVector WorklistMembers;
125   /// This bitvector is set for each vreg index where the vreg is defined
126   /// by an instruction where lowersToCopies()==true.
127   BitVector DefinedByCopy;
128 };
129 
130 } // end anonymous namespace
131 
132 char DetectDeadLanes::ID = 0;
133 char &llvm::DetectDeadLanesID = DetectDeadLanes::ID;
134 
135 INITIALIZE_PASS(DetectDeadLanes, "detect-dead-lanes", "Detect Dead Lanes",
136                 false, false)
137 
138 /// Returns true if \p MI will get lowered to a series of COPY instructions.
139 /// We call this a COPY-like instruction.
lowersToCopies(const MachineInstr & MI)140 static bool lowersToCopies(const MachineInstr &MI) {
141   // Note: We could support instructions with MCInstrDesc::isRegSequenceLike(),
142   // isExtractSubRegLike(), isInsertSubregLike() in the future even though they
143   // are not lowered to a COPY.
144   switch (MI.getOpcode()) {
145   case TargetOpcode::COPY:
146   case TargetOpcode::PHI:
147   case TargetOpcode::INSERT_SUBREG:
148   case TargetOpcode::REG_SEQUENCE:
149   case TargetOpcode::EXTRACT_SUBREG:
150     return true;
151   }
152   return false;
153 }
154 
isCrossCopy(const MachineRegisterInfo & MRI,const MachineInstr & MI,const TargetRegisterClass * DstRC,const MachineOperand & MO)155 static bool isCrossCopy(const MachineRegisterInfo &MRI,
156                         const MachineInstr &MI,
157                         const TargetRegisterClass *DstRC,
158                         const MachineOperand &MO) {
159   assert(lowersToCopies(MI));
160   unsigned SrcReg = MO.getReg();
161   const TargetRegisterClass *SrcRC = MRI.getRegClass(SrcReg);
162   if (DstRC == SrcRC)
163     return false;
164 
165   unsigned SrcSubIdx = MO.getSubReg();
166 
167   const TargetRegisterInfo &TRI = *MRI.getTargetRegisterInfo();
168   unsigned DstSubIdx = 0;
169   switch (MI.getOpcode()) {
170   case TargetOpcode::INSERT_SUBREG:
171     if (MI.getOperandNo(&MO) == 2)
172       DstSubIdx = MI.getOperand(3).getImm();
173     break;
174   case TargetOpcode::REG_SEQUENCE: {
175     unsigned OpNum = MI.getOperandNo(&MO);
176     DstSubIdx = MI.getOperand(OpNum+1).getImm();
177     break;
178   }
179   case TargetOpcode::EXTRACT_SUBREG: {
180     unsigned SubReg = MI.getOperand(2).getImm();
181     SrcSubIdx = TRI.composeSubRegIndices(SubReg, SrcSubIdx);
182   }
183   }
184 
185   unsigned PreA, PreB; // Unused.
186   if (SrcSubIdx && DstSubIdx)
187     return !TRI.getCommonSuperRegClass(SrcRC, SrcSubIdx, DstRC, DstSubIdx, PreA,
188                                        PreB);
189   if (SrcSubIdx)
190     return !TRI.getMatchingSuperRegClass(SrcRC, DstRC, SrcSubIdx);
191   if (DstSubIdx)
192     return !TRI.getMatchingSuperRegClass(DstRC, SrcRC, DstSubIdx);
193   return !TRI.getCommonSubClass(SrcRC, DstRC);
194 }
195 
addUsedLanesOnOperand(const MachineOperand & MO,LaneBitmask UsedLanes)196 void DetectDeadLanes::addUsedLanesOnOperand(const MachineOperand &MO,
197                                             LaneBitmask UsedLanes) {
198   if (!MO.readsReg())
199     return;
200   unsigned MOReg = MO.getReg();
201   if (!TargetRegisterInfo::isVirtualRegister(MOReg))
202     return;
203 
204   unsigned MOSubReg = MO.getSubReg();
205   if (MOSubReg != 0)
206     UsedLanes = TRI->composeSubRegIndexLaneMask(MOSubReg, UsedLanes);
207   UsedLanes &= MRI->getMaxLaneMaskForVReg(MOReg);
208 
209   unsigned MORegIdx = TargetRegisterInfo::virtReg2Index(MOReg);
210   VRegInfo &MORegInfo = VRegInfos[MORegIdx];
211   LaneBitmask PrevUsedLanes = MORegInfo.UsedLanes;
212   // Any change at all?
213   if ((UsedLanes & ~PrevUsedLanes) == 0)
214     return;
215 
216   // Set UsedLanes and remember instruction for further propagation.
217   MORegInfo.UsedLanes = PrevUsedLanes | UsedLanes;
218   if (DefinedByCopy.test(MORegIdx))
219     PutInWorklist(MORegIdx);
220 }
221 
transferUsedLanesStep(const MachineInstr & MI,LaneBitmask UsedLanes)222 void DetectDeadLanes::transferUsedLanesStep(const MachineInstr &MI,
223                                             LaneBitmask UsedLanes) {
224   for (const MachineOperand &MO : MI.uses()) {
225     if (!MO.isReg() || !TargetRegisterInfo::isVirtualRegister(MO.getReg()))
226       continue;
227     LaneBitmask UsedOnMO = transferUsedLanes(MI, UsedLanes, MO);
228     addUsedLanesOnOperand(MO, UsedOnMO);
229   }
230 }
231 
transferUsedLanes(const MachineInstr & MI,LaneBitmask UsedLanes,const MachineOperand & MO) const232 LaneBitmask DetectDeadLanes::transferUsedLanes(const MachineInstr &MI,
233                                                LaneBitmask UsedLanes,
234                                                const MachineOperand &MO) const {
235   unsigned OpNum = MI.getOperandNo(&MO);
236   assert(lowersToCopies(MI) && DefinedByCopy[
237            TargetRegisterInfo::virtReg2Index(MI.getOperand(0).getReg())]);
238 
239   switch (MI.getOpcode()) {
240   case TargetOpcode::COPY:
241   case TargetOpcode::PHI:
242     return UsedLanes;
243   case TargetOpcode::REG_SEQUENCE: {
244     assert(OpNum % 2 == 1);
245     unsigned SubIdx = MI.getOperand(OpNum + 1).getImm();
246     return TRI->reverseComposeSubRegIndexLaneMask(SubIdx, UsedLanes);
247   }
248   case TargetOpcode::INSERT_SUBREG: {
249     unsigned SubIdx = MI.getOperand(3).getImm();
250     LaneBitmask MO2UsedLanes =
251         TRI->reverseComposeSubRegIndexLaneMask(SubIdx, UsedLanes);
252     if (OpNum == 2)
253       return MO2UsedLanes;
254 
255     const MachineOperand &Def = MI.getOperand(0);
256     unsigned DefReg = Def.getReg();
257     const TargetRegisterClass *RC = MRI->getRegClass(DefReg);
258     LaneBitmask MO1UsedLanes;
259     if (RC->CoveredBySubRegs)
260       MO1UsedLanes = UsedLanes & ~TRI->getSubRegIndexLaneMask(SubIdx);
261     else
262       MO1UsedLanes = RC->LaneMask;
263 
264     assert(OpNum == 1);
265     return MO1UsedLanes;
266   }
267   case TargetOpcode::EXTRACT_SUBREG: {
268     assert(OpNum == 1);
269     unsigned SubIdx = MI.getOperand(2).getImm();
270     return TRI->composeSubRegIndexLaneMask(SubIdx, UsedLanes);
271   }
272   default:
273     llvm_unreachable("function must be called with COPY-like instruction");
274   }
275 }
276 
transferDefinedLanesStep(const MachineOperand & Use,LaneBitmask DefinedLanes)277 void DetectDeadLanes::transferDefinedLanesStep(const MachineOperand &Use,
278                                                LaneBitmask DefinedLanes) {
279   if (!Use.readsReg())
280     return;
281   // Check whether the operand writes a vreg and is part of a COPY-like
282   // instruction.
283   const MachineInstr &MI = *Use.getParent();
284   if (MI.getDesc().getNumDefs() != 1)
285     return;
286   // FIXME: PATCHPOINT instructions announce a Def that does not always exist,
287   // they really need to be modeled differently!
288   if (MI.getOpcode() == TargetOpcode::PATCHPOINT)
289     return;
290   const MachineOperand &Def = *MI.defs().begin();
291   unsigned DefReg = Def.getReg();
292   if (!TargetRegisterInfo::isVirtualRegister(DefReg))
293     return;
294   unsigned DefRegIdx = TargetRegisterInfo::virtReg2Index(DefReg);
295   if (!DefinedByCopy.test(DefRegIdx))
296     return;
297 
298   unsigned OpNum = MI.getOperandNo(&Use);
299   DefinedLanes =
300       TRI->reverseComposeSubRegIndexLaneMask(Use.getSubReg(), DefinedLanes);
301   DefinedLanes = transferDefinedLanes(Def, OpNum, DefinedLanes);
302 
303   VRegInfo &RegInfo = VRegInfos[DefRegIdx];
304   LaneBitmask PrevDefinedLanes = RegInfo.DefinedLanes;
305   // Any change at all?
306   if ((DefinedLanes & ~PrevDefinedLanes) == 0)
307     return;
308 
309   RegInfo.DefinedLanes = PrevDefinedLanes | DefinedLanes;
310   PutInWorklist(DefRegIdx);
311 }
312 
transferDefinedLanes(const MachineOperand & Def,unsigned OpNum,LaneBitmask DefinedLanes) const313 LaneBitmask DetectDeadLanes::transferDefinedLanes(const MachineOperand &Def,
314     unsigned OpNum, LaneBitmask DefinedLanes) const {
315   const MachineInstr &MI = *Def.getParent();
316   // Translate DefinedLanes if necessary.
317   switch (MI.getOpcode()) {
318   case TargetOpcode::REG_SEQUENCE: {
319     unsigned SubIdx = MI.getOperand(OpNum + 1).getImm();
320     DefinedLanes = TRI->composeSubRegIndexLaneMask(SubIdx, DefinedLanes);
321     DefinedLanes &= TRI->getSubRegIndexLaneMask(SubIdx);
322     break;
323   }
324   case TargetOpcode::INSERT_SUBREG: {
325     unsigned SubIdx = MI.getOperand(3).getImm();
326     if (OpNum == 2) {
327       DefinedLanes = TRI->composeSubRegIndexLaneMask(SubIdx, DefinedLanes);
328       DefinedLanes &= TRI->getSubRegIndexLaneMask(SubIdx);
329     } else {
330       assert(OpNum == 1 && "INSERT_SUBREG must have two operands");
331       // Ignore lanes defined by operand 2.
332       DefinedLanes &= ~TRI->getSubRegIndexLaneMask(SubIdx);
333     }
334     break;
335   }
336   case TargetOpcode::EXTRACT_SUBREG: {
337     unsigned SubIdx = MI.getOperand(2).getImm();
338     assert(OpNum == 1 && "EXTRACT_SUBREG must have one register operand only");
339     DefinedLanes = TRI->reverseComposeSubRegIndexLaneMask(SubIdx, DefinedLanes);
340     break;
341   }
342   case TargetOpcode::COPY:
343   case TargetOpcode::PHI:
344     break;
345   default:
346     llvm_unreachable("function must be called with COPY-like instruction");
347   }
348 
349   assert(Def.getSubReg() == 0 &&
350          "Should not have subregister defs in machine SSA phase");
351   DefinedLanes &= MRI->getMaxLaneMaskForVReg(Def.getReg());
352   return DefinedLanes;
353 }
354 
determineInitialDefinedLanes(unsigned Reg)355 LaneBitmask DetectDeadLanes::determineInitialDefinedLanes(unsigned Reg) {
356   // Live-In or unused registers have no definition but are considered fully
357   // defined.
358   if (!MRI->hasOneDef(Reg))
359     return ~0u;
360 
361   const MachineOperand &Def = *MRI->def_begin(Reg);
362   const MachineInstr &DefMI = *Def.getParent();
363   if (lowersToCopies(DefMI)) {
364     // Start optimisatically with no used or defined lanes for copy
365     // instructions. The following dataflow analysis will add more bits.
366     unsigned RegIdx = TargetRegisterInfo::virtReg2Index(Reg);
367     DefinedByCopy.set(RegIdx);
368     PutInWorklist(RegIdx);
369 
370     if (Def.isDead())
371       return 0;
372 
373     // COPY/PHI can copy across unrelated register classes (example: float/int)
374     // with incompatible subregister structure. Do not include these in the
375     // dataflow analysis since we cannot transfer lanemasks in a meaningful way.
376     const TargetRegisterClass *DefRC = MRI->getRegClass(Reg);
377 
378     // Determine initially DefinedLanes.
379     LaneBitmask DefinedLanes = 0;
380     for (const MachineOperand &MO : DefMI.uses()) {
381       if (!MO.isReg() || !MO.readsReg())
382         continue;
383       unsigned MOReg = MO.getReg();
384       if (!MOReg)
385         continue;
386 
387       LaneBitmask MODefinedLanes;
388       if (TargetRegisterInfo::isPhysicalRegister(MOReg)) {
389         MODefinedLanes = ~0u;
390       } else if (isCrossCopy(*MRI, DefMI, DefRC, MO)) {
391         MODefinedLanes = ~0u;
392       } else {
393         assert(TargetRegisterInfo::isVirtualRegister(MOReg));
394         if (MRI->hasOneDef(MOReg)) {
395           const MachineOperand &MODef = *MRI->def_begin(MOReg);
396           const MachineInstr &MODefMI = *MODef.getParent();
397           // Bits from copy-like operations will be added later.
398           if (lowersToCopies(MODefMI) || MODefMI.isImplicitDef())
399             continue;
400         }
401         unsigned MOSubReg = MO.getSubReg();
402         MODefinedLanes = MRI->getMaxLaneMaskForVReg(MOReg);
403         MODefinedLanes = TRI->reverseComposeSubRegIndexLaneMask(
404             MOSubReg, MODefinedLanes);
405       }
406 
407       unsigned OpNum = DefMI.getOperandNo(&MO);
408       DefinedLanes |= transferDefinedLanes(Def, OpNum, MODefinedLanes);
409     }
410     return DefinedLanes;
411   }
412   if (DefMI.isImplicitDef() || Def.isDead())
413     return 0;
414 
415   assert(Def.getSubReg() == 0 &&
416          "Should not have subregister defs in machine SSA phase");
417   return MRI->getMaxLaneMaskForVReg(Reg);
418 }
419 
determineInitialUsedLanes(unsigned Reg)420 LaneBitmask DetectDeadLanes::determineInitialUsedLanes(unsigned Reg) {
421   LaneBitmask UsedLanes = 0;
422   for (const MachineOperand &MO : MRI->use_nodbg_operands(Reg)) {
423     if (!MO.readsReg())
424       continue;
425 
426     const MachineInstr &UseMI = *MO.getParent();
427     if (UseMI.isKill())
428       continue;
429 
430     unsigned SubReg = MO.getSubReg();
431     if (lowersToCopies(UseMI)) {
432       assert(UseMI.getDesc().getNumDefs() == 1);
433       const MachineOperand &Def = *UseMI.defs().begin();
434       unsigned DefReg = Def.getReg();
435       // The used lanes of COPY-like instruction operands are determined by the
436       // following dataflow analysis.
437       if (TargetRegisterInfo::isVirtualRegister(DefReg)) {
438         // But ignore copies across incompatible register classes.
439         bool CrossCopy = false;
440         if (lowersToCopies(UseMI)) {
441           const TargetRegisterClass *DstRC = MRI->getRegClass(DefReg);
442           CrossCopy = isCrossCopy(*MRI, UseMI, DstRC, MO);
443           if (CrossCopy)
444             DEBUG(dbgs() << "Copy accross incompatible classes: " << UseMI);
445         }
446 
447         if (!CrossCopy)
448           continue;
449       }
450     }
451 
452     // Shortcut: All lanes are used.
453     if (SubReg == 0)
454       return MRI->getMaxLaneMaskForVReg(Reg);
455 
456     UsedLanes |= TRI->getSubRegIndexLaneMask(SubReg);
457   }
458   return UsedLanes;
459 }
460 
isUndefRegAtInput(const MachineOperand & MO,const VRegInfo & RegInfo) const461 bool DetectDeadLanes::isUndefRegAtInput(const MachineOperand &MO,
462                                         const VRegInfo &RegInfo) const {
463   unsigned SubReg = MO.getSubReg();
464   LaneBitmask Mask = TRI->getSubRegIndexLaneMask(SubReg);
465   return (RegInfo.DefinedLanes & RegInfo.UsedLanes & Mask) == 0;
466 }
467 
isUndefInput(const MachineOperand & MO,bool * CrossCopy) const468 bool DetectDeadLanes::isUndefInput(const MachineOperand &MO,
469                                    bool *CrossCopy) const {
470   if (!MO.isUse())
471     return false;
472   const MachineInstr &MI = *MO.getParent();
473   if (!lowersToCopies(MI))
474     return false;
475   const MachineOperand &Def = MI.getOperand(0);
476   unsigned DefReg = Def.getReg();
477   if (!TargetRegisterInfo::isVirtualRegister(DefReg))
478     return false;
479   unsigned DefRegIdx = TargetRegisterInfo::virtReg2Index(DefReg);
480   if (!DefinedByCopy.test(DefRegIdx))
481     return false;
482 
483   const VRegInfo &DefRegInfo = VRegInfos[DefRegIdx];
484   LaneBitmask UsedLanes = transferUsedLanes(MI, DefRegInfo.UsedLanes, MO);
485   if (UsedLanes != 0)
486     return false;
487 
488   unsigned MOReg = MO.getReg();
489   if (TargetRegisterInfo::isVirtualRegister(MOReg)) {
490     const TargetRegisterClass *DstRC = MRI->getRegClass(DefReg);
491     *CrossCopy = isCrossCopy(*MRI, MI, DstRC, MO);
492   }
493   return true;
494 }
495 
runOnce(MachineFunction & MF)496 bool DetectDeadLanes::runOnce(MachineFunction &MF) {
497   // First pass: Populate defs/uses of vregs with initial values
498   unsigned NumVirtRegs = MRI->getNumVirtRegs();
499   for (unsigned RegIdx = 0; RegIdx < NumVirtRegs; ++RegIdx) {
500     unsigned Reg = TargetRegisterInfo::index2VirtReg(RegIdx);
501 
502     // Determine used/defined lanes and add copy instructions to worklist.
503     VRegInfo &Info = VRegInfos[RegIdx];
504     Info.DefinedLanes = determineInitialDefinedLanes(Reg);
505     Info.UsedLanes = determineInitialUsedLanes(Reg);
506   }
507 
508   // Iterate as long as defined lanes/used lanes keep changing.
509   while (!Worklist.empty()) {
510     unsigned RegIdx = Worklist.front();
511     Worklist.pop_front();
512     WorklistMembers.reset(RegIdx);
513     VRegInfo &Info = VRegInfos[RegIdx];
514     unsigned Reg = TargetRegisterInfo::index2VirtReg(RegIdx);
515 
516     // Transfer UsedLanes to operands of DefMI (backwards dataflow).
517     MachineOperand &Def = *MRI->def_begin(Reg);
518     const MachineInstr &MI = *Def.getParent();
519     transferUsedLanesStep(MI, Info.UsedLanes);
520     // Transfer DefinedLanes to users of Reg (forward dataflow).
521     for (const MachineOperand &MO : MRI->use_nodbg_operands(Reg))
522       transferDefinedLanesStep(MO, Info.DefinedLanes);
523   }
524 
525   DEBUG(
526     dbgs() << "Defined/Used lanes:\n";
527     for (unsigned RegIdx = 0; RegIdx < NumVirtRegs; ++RegIdx) {
528       unsigned Reg = TargetRegisterInfo::index2VirtReg(RegIdx);
529       const VRegInfo &Info = VRegInfos[RegIdx];
530       dbgs() << PrintReg(Reg, nullptr)
531              << " Used: " << PrintLaneMask(Info.UsedLanes)
532              << " Def: " << PrintLaneMask(Info.DefinedLanes) << '\n';
533     }
534     dbgs() << "\n";
535   );
536 
537   bool Again = false;
538   // Mark operands as dead/unused.
539   for (MachineBasicBlock &MBB : MF) {
540     for (MachineInstr &MI : MBB) {
541       for (MachineOperand &MO : MI.operands()) {
542         if (!MO.isReg())
543           continue;
544         unsigned Reg = MO.getReg();
545         if (!TargetRegisterInfo::isVirtualRegister(Reg))
546           continue;
547         unsigned RegIdx = TargetRegisterInfo::virtReg2Index(Reg);
548         const VRegInfo &RegInfo = VRegInfos[RegIdx];
549         if (MO.isDef() && !MO.isDead() && RegInfo.UsedLanes == 0) {
550           DEBUG(dbgs() << "Marking operand '" << MO << "' as dead in " << MI);
551           MO.setIsDead();
552         }
553         if (MO.readsReg()) {
554           bool CrossCopy = false;
555           if (isUndefRegAtInput(MO, RegInfo)) {
556             DEBUG(dbgs() << "Marking operand '" << MO << "' as undef in "
557                   << MI);
558             MO.setIsUndef();
559           } else if (isUndefInput(MO, &CrossCopy)) {
560             DEBUG(dbgs() << "Marking operand '" << MO << "' as undef in "
561                   << MI);
562             MO.setIsUndef();
563             if (CrossCopy)
564               Again = true;
565           }
566         }
567       }
568     }
569   }
570 
571   return Again;
572 }
573 
runOnMachineFunction(MachineFunction & MF)574 bool DetectDeadLanes::runOnMachineFunction(MachineFunction &MF) {
575   // Don't bother if we won't track subregister liveness later.  This pass is
576   // required for correctness if subregister liveness is enabled because the
577   // register coalescer cannot deal with hidden dead defs. However without
578   // subregister liveness enabled, the expected benefits of this pass are small
579   // so we safe the compile time.
580   if (!MF.getSubtarget().enableSubRegLiveness()) {
581     DEBUG(dbgs() << "Skipping Detect dead lanes pass\n");
582     return false;
583   }
584 
585   MRI = &MF.getRegInfo();
586   TRI = MRI->getTargetRegisterInfo();
587 
588   unsigned NumVirtRegs = MRI->getNumVirtRegs();
589   VRegInfos = new VRegInfo[NumVirtRegs];
590   WorklistMembers.resize(NumVirtRegs);
591   DefinedByCopy.resize(NumVirtRegs);
592 
593   bool Again;
594   do {
595     Again = runOnce(MF);
596   } while(Again);
597 
598   DefinedByCopy.clear();
599   WorklistMembers.clear();
600   delete[] VRegInfos;
601   return true;
602 }
603