1 //===---- MachineCombiner.cpp - Instcombining on SSA form machine code ----===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 // The machine combiner pass uses machine trace metrics to ensure the combined
10 // instructions do not lengthen the critical path or the resource depth.
11 //===----------------------------------------------------------------------===//
12 
13 #include "llvm/ADT/DenseMap.h"
14 #include "llvm/ADT/Statistic.h"
15 #include "llvm/Analysis/ProfileSummaryInfo.h"
16 #include "llvm/CodeGen/LazyMachineBlockFrequencyInfo.h"
17 #include "llvm/CodeGen/MachineCombinerPattern.h"
18 #include "llvm/CodeGen/MachineDominators.h"
19 #include "llvm/CodeGen/MachineFunction.h"
20 #include "llvm/CodeGen/MachineFunctionPass.h"
21 #include "llvm/CodeGen/MachineLoopInfo.h"
22 #include "llvm/CodeGen/MachineRegisterInfo.h"
23 #include "llvm/CodeGen/MachineSizeOpts.h"
24 #include "llvm/CodeGen/MachineTraceMetrics.h"
25 #include "llvm/CodeGen/RegisterClassInfo.h"
26 #include "llvm/CodeGen/TargetInstrInfo.h"
27 #include "llvm/CodeGen/TargetRegisterInfo.h"
28 #include "llvm/CodeGen/TargetSchedule.h"
29 #include "llvm/CodeGen/TargetSubtargetInfo.h"
30 #include "llvm/InitializePasses.h"
31 #include "llvm/Support/CommandLine.h"
32 #include "llvm/Support/Debug.h"
33 #include "llvm/Support/raw_ostream.h"
34 
35 using namespace llvm;
36 
37 #define DEBUG_TYPE "machine-combiner"
38 
39 STATISTIC(NumInstCombined, "Number of machineinst combined");
40 
41 static cl::opt<unsigned>
42 inc_threshold("machine-combiner-inc-threshold", cl::Hidden,
43               cl::desc("Incremental depth computation will be used for basic "
44                        "blocks with more instructions."), cl::init(500));
45 
46 static cl::opt<bool> dump_intrs("machine-combiner-dump-subst-intrs", cl::Hidden,
47                                 cl::desc("Dump all substituted intrs"),
48                                 cl::init(false));
49 
50 #ifdef EXPENSIVE_CHECKS
51 static cl::opt<bool> VerifyPatternOrder(
52     "machine-combiner-verify-pattern-order", cl::Hidden,
53     cl::desc(
54         "Verify that the generated patterns are ordered by increasing latency"),
55     cl::init(true));
56 #else
57 static cl::opt<bool> VerifyPatternOrder(
58     "machine-combiner-verify-pattern-order", cl::Hidden,
59     cl::desc(
60         "Verify that the generated patterns are ordered by increasing latency"),
61     cl::init(false));
62 #endif
63 
64 namespace {
65 class MachineCombiner : public MachineFunctionPass {
66   const TargetSubtargetInfo *STI = nullptr;
67   const TargetInstrInfo *TII = nullptr;
68   const TargetRegisterInfo *TRI = nullptr;
69   MCSchedModel SchedModel;
70   MachineRegisterInfo *MRI = nullptr;
71   MachineLoopInfo *MLI = nullptr; // Current MachineLoopInfo
72   MachineTraceMetrics *Traces = nullptr;
73   MachineTraceMetrics::Ensemble *TraceEnsemble = nullptr;
74   MachineBlockFrequencyInfo *MBFI = nullptr;
75   ProfileSummaryInfo *PSI = nullptr;
76   RegisterClassInfo RegClassInfo;
77 
78   TargetSchedModel TSchedModel;
79 
80   /// True if optimizing for code size.
81   bool OptSize = false;
82 
83 public:
84   static char ID;
85   MachineCombiner() : MachineFunctionPass(ID) {
86     initializeMachineCombinerPass(*PassRegistry::getPassRegistry());
87   }
88   void getAnalysisUsage(AnalysisUsage &AU) const override;
89   bool runOnMachineFunction(MachineFunction &MF) override;
90   StringRef getPassName() const override { return "Machine InstCombiner"; }
91 
92 private:
93   bool combineInstructions(MachineBasicBlock *);
94   MachineInstr *getOperandDef(const MachineOperand &MO);
95   bool isTransientMI(const MachineInstr *MI);
96   unsigned getDepth(SmallVectorImpl<MachineInstr *> &InsInstrs,
97                     DenseMap<unsigned, unsigned> &InstrIdxForVirtReg,
98                     MachineTraceMetrics::Trace BlockTrace,
99                     const MachineBasicBlock &MBB);
100   unsigned getLatency(MachineInstr *Root, MachineInstr *NewRoot,
101                       MachineTraceMetrics::Trace BlockTrace);
102   bool
103   improvesCriticalPathLen(MachineBasicBlock *MBB, MachineInstr *Root,
104                           MachineTraceMetrics::Trace BlockTrace,
105                           SmallVectorImpl<MachineInstr *> &InsInstrs,
106                           SmallVectorImpl<MachineInstr *> &DelInstrs,
107                           DenseMap<unsigned, unsigned> &InstrIdxForVirtReg,
108                           MachineCombinerPattern Pattern, bool SlackIsAccurate);
109   bool reduceRegisterPressure(MachineInstr &Root, MachineBasicBlock *MBB,
110                               SmallVectorImpl<MachineInstr *> &InsInstrs,
111                               SmallVectorImpl<MachineInstr *> &DelInstrs,
112                               MachineCombinerPattern Pattern);
113   bool preservesResourceLen(MachineBasicBlock *MBB,
114                             MachineTraceMetrics::Trace BlockTrace,
115                             SmallVectorImpl<MachineInstr *> &InsInstrs,
116                             SmallVectorImpl<MachineInstr *> &DelInstrs);
117   void instr2instrSC(SmallVectorImpl<MachineInstr *> &Instrs,
118                      SmallVectorImpl<const MCSchedClassDesc *> &InstrsSC);
119   std::pair<unsigned, unsigned>
120   getLatenciesForInstrSequences(MachineInstr &MI,
121                                 SmallVectorImpl<MachineInstr *> &InsInstrs,
122                                 SmallVectorImpl<MachineInstr *> &DelInstrs,
123                                 MachineTraceMetrics::Trace BlockTrace);
124 
125   void verifyPatternOrder(MachineBasicBlock *MBB, MachineInstr &Root,
126                           SmallVector<MachineCombinerPattern, 16> &Patterns);
127 };
128 }
129 
130 char MachineCombiner::ID = 0;
131 char &llvm::MachineCombinerID = MachineCombiner::ID;
132 
133 INITIALIZE_PASS_BEGIN(MachineCombiner, DEBUG_TYPE,
134                       "Machine InstCombiner", false, false)
135 INITIALIZE_PASS_DEPENDENCY(MachineLoopInfo)
136 INITIALIZE_PASS_DEPENDENCY(MachineTraceMetrics)
137 INITIALIZE_PASS_END(MachineCombiner, DEBUG_TYPE, "Machine InstCombiner",
138                     false, false)
139 
140 void MachineCombiner::getAnalysisUsage(AnalysisUsage &AU) const {
141   AU.setPreservesCFG();
142   AU.addPreserved<MachineDominatorTree>();
143   AU.addRequired<MachineLoopInfo>();
144   AU.addPreserved<MachineLoopInfo>();
145   AU.addRequired<MachineTraceMetrics>();
146   AU.addPreserved<MachineTraceMetrics>();
147   AU.addRequired<LazyMachineBlockFrequencyInfoPass>();
148   AU.addRequired<ProfileSummaryInfoWrapperPass>();
149   MachineFunctionPass::getAnalysisUsage(AU);
150 }
151 
152 MachineInstr *
153 MachineCombiner::getOperandDef(const MachineOperand &MO) {
154   MachineInstr *DefInstr = nullptr;
155   // We need a virtual register definition.
156   if (MO.isReg() && MO.getReg().isVirtual())
157     DefInstr = MRI->getUniqueVRegDef(MO.getReg());
158   // PHI's have no depth etc.
159   if (DefInstr && DefInstr->isPHI())
160     DefInstr = nullptr;
161   return DefInstr;
162 }
163 
164 /// Return true if MI is unlikely to generate an actual target instruction.
165 bool MachineCombiner::isTransientMI(const MachineInstr *MI) {
166   if (!MI->isCopy())
167     return MI->isTransient();
168 
169   // If MI is a COPY, check if its src and dst registers can be coalesced.
170   Register Dst = MI->getOperand(0).getReg();
171   Register Src = MI->getOperand(1).getReg();
172 
173   if (!MI->isFullCopy()) {
174     // If src RC contains super registers of dst RC, it can also be coalesced.
175     if (MI->getOperand(0).getSubReg() || Src.isPhysical() || Dst.isPhysical())
176       return false;
177 
178     auto SrcSub = MI->getOperand(1).getSubReg();
179     auto SrcRC = MRI->getRegClass(Src);
180     auto DstRC = MRI->getRegClass(Dst);
181     return TRI->getMatchingSuperRegClass(SrcRC, DstRC, SrcSub) != nullptr;
182   }
183 
184   if (Src.isPhysical() && Dst.isPhysical())
185     return Src == Dst;
186 
187   if (Src.isVirtual() && Dst.isVirtual()) {
188     auto SrcRC = MRI->getRegClass(Src);
189     auto DstRC = MRI->getRegClass(Dst);
190     return SrcRC->hasSuperClassEq(DstRC) || SrcRC->hasSubClassEq(DstRC);
191   }
192 
193   if (Src.isVirtual())
194     std::swap(Src, Dst);
195 
196   // Now Src is physical register, Dst is virtual register.
197   auto DstRC = MRI->getRegClass(Dst);
198   return DstRC->contains(Src);
199 }
200 
201 /// Computes depth of instructions in vector \InsInstr.
202 ///
203 /// \param InsInstrs is a vector of machine instructions
204 /// \param InstrIdxForVirtReg is a dense map of virtual register to index
205 /// of defining machine instruction in \p InsInstrs
206 /// \param BlockTrace is a trace of machine instructions
207 ///
208 /// \returns Depth of last instruction in \InsInstrs ("NewRoot")
209 unsigned
210 MachineCombiner::getDepth(SmallVectorImpl<MachineInstr *> &InsInstrs,
211                           DenseMap<unsigned, unsigned> &InstrIdxForVirtReg,
212                           MachineTraceMetrics::Trace BlockTrace,
213                           const MachineBasicBlock &MBB) {
214   SmallVector<unsigned, 16> InstrDepth;
215   // For each instruction in the new sequence compute the depth based on the
216   // operands. Use the trace information when possible. For new operands which
217   // are tracked in the InstrIdxForVirtReg map depth is looked up in InstrDepth
218   for (auto *InstrPtr : InsInstrs) { // for each Use
219     unsigned IDepth = 0;
220     for (const MachineOperand &MO : InstrPtr->all_uses()) {
221       // Check for virtual register operand.
222       if (!MO.getReg().isVirtual())
223         continue;
224       unsigned DepthOp = 0;
225       unsigned LatencyOp = 0;
226       DenseMap<unsigned, unsigned>::iterator II =
227           InstrIdxForVirtReg.find(MO.getReg());
228       if (II != InstrIdxForVirtReg.end()) {
229         // Operand is new virtual register not in trace
230         assert(II->second < InstrDepth.size() && "Bad Index");
231         MachineInstr *DefInstr = InsInstrs[II->second];
232         assert(DefInstr &&
233                "There must be a definition for a new virtual register");
234         DepthOp = InstrDepth[II->second];
235         int DefIdx = DefInstr->findRegisterDefOperandIdx(MO.getReg());
236         int UseIdx = InstrPtr->findRegisterUseOperandIdx(MO.getReg());
237         LatencyOp = TSchedModel.computeOperandLatency(DefInstr, DefIdx,
238                                                       InstrPtr, UseIdx);
239       } else {
240         MachineInstr *DefInstr = getOperandDef(MO);
241         if (DefInstr && (TII->getMachineCombinerTraceStrategy() !=
242                              MachineTraceStrategy::TS_Local ||
243                          DefInstr->getParent() == &MBB)) {
244           DepthOp = BlockTrace.getInstrCycles(*DefInstr).Depth;
245           if (!isTransientMI(DefInstr))
246             LatencyOp = TSchedModel.computeOperandLatency(
247                 DefInstr, DefInstr->findRegisterDefOperandIdx(MO.getReg()),
248                 InstrPtr, InstrPtr->findRegisterUseOperandIdx(MO.getReg()));
249         }
250       }
251       IDepth = std::max(IDepth, DepthOp + LatencyOp);
252     }
253     InstrDepth.push_back(IDepth);
254   }
255   unsigned NewRootIdx = InsInstrs.size() - 1;
256   return InstrDepth[NewRootIdx];
257 }
258 
259 /// Computes instruction latency as max of latency of defined operands.
260 ///
261 /// \param Root is a machine instruction that could be replaced by NewRoot.
262 /// It is used to compute a more accurate latency information for NewRoot in
263 /// case there is a dependent instruction in the same trace (\p BlockTrace)
264 /// \param NewRoot is the instruction for which the latency is computed
265 /// \param BlockTrace is a trace of machine instructions
266 ///
267 /// \returns Latency of \p NewRoot
268 unsigned MachineCombiner::getLatency(MachineInstr *Root, MachineInstr *NewRoot,
269                                      MachineTraceMetrics::Trace BlockTrace) {
270   // Check each definition in NewRoot and compute the latency
271   unsigned NewRootLatency = 0;
272 
273   for (const MachineOperand &MO : NewRoot->all_defs()) {
274     // Check for virtual register operand.
275     if (!MO.getReg().isVirtual())
276       continue;
277     // Get the first instruction that uses MO
278     MachineRegisterInfo::reg_iterator RI = MRI->reg_begin(MO.getReg());
279     RI++;
280     if (RI == MRI->reg_end())
281       continue;
282     MachineInstr *UseMO = RI->getParent();
283     unsigned LatencyOp = 0;
284     if (UseMO && BlockTrace.isDepInTrace(*Root, *UseMO)) {
285       LatencyOp = TSchedModel.computeOperandLatency(
286           NewRoot, NewRoot->findRegisterDefOperandIdx(MO.getReg()), UseMO,
287           UseMO->findRegisterUseOperandIdx(MO.getReg()));
288     } else {
289       LatencyOp = TSchedModel.computeInstrLatency(NewRoot);
290     }
291     NewRootLatency = std::max(NewRootLatency, LatencyOp);
292   }
293   return NewRootLatency;
294 }
295 
296 /// The combiner's goal may differ based on which pattern it is attempting
297 /// to optimize.
298 enum class CombinerObjective {
299   MustReduceDepth,            // The data dependency chain must be improved.
300   MustReduceRegisterPressure, // The register pressure must be reduced.
301   Default                     // The critical path must not be lengthened.
302 };
303 
304 static CombinerObjective getCombinerObjective(MachineCombinerPattern P) {
305   // TODO: If C++ ever gets a real enum class, make this part of the
306   // MachineCombinerPattern class.
307   switch (P) {
308   case MachineCombinerPattern::REASSOC_AX_BY:
309   case MachineCombinerPattern::REASSOC_AX_YB:
310   case MachineCombinerPattern::REASSOC_XA_BY:
311   case MachineCombinerPattern::REASSOC_XA_YB:
312   case MachineCombinerPattern::REASSOC_XY_AMM_BMM:
313   case MachineCombinerPattern::REASSOC_XMM_AMM_BMM:
314   case MachineCombinerPattern::SUBADD_OP1:
315   case MachineCombinerPattern::SUBADD_OP2:
316   case MachineCombinerPattern::FMADD_AX:
317   case MachineCombinerPattern::FMADD_XA:
318   case MachineCombinerPattern::FMSUB:
319   case MachineCombinerPattern::FNMSUB:
320     return CombinerObjective::MustReduceDepth;
321   case MachineCombinerPattern::REASSOC_XY_BCA:
322   case MachineCombinerPattern::REASSOC_XY_BAC:
323     return CombinerObjective::MustReduceRegisterPressure;
324   default:
325     return CombinerObjective::Default;
326   }
327 }
328 
329 /// Estimate the latency of the new and original instruction sequence by summing
330 /// up the latencies of the inserted and deleted instructions. This assumes
331 /// that the inserted and deleted instructions are dependent instruction chains,
332 /// which might not hold in all cases.
333 std::pair<unsigned, unsigned> MachineCombiner::getLatenciesForInstrSequences(
334     MachineInstr &MI, SmallVectorImpl<MachineInstr *> &InsInstrs,
335     SmallVectorImpl<MachineInstr *> &DelInstrs,
336     MachineTraceMetrics::Trace BlockTrace) {
337   assert(!InsInstrs.empty() && "Only support sequences that insert instrs.");
338   unsigned NewRootLatency = 0;
339   // NewRoot is the last instruction in the \p InsInstrs vector.
340   MachineInstr *NewRoot = InsInstrs.back();
341   for (unsigned i = 0; i < InsInstrs.size() - 1; i++)
342     NewRootLatency += TSchedModel.computeInstrLatency(InsInstrs[i]);
343   NewRootLatency += getLatency(&MI, NewRoot, BlockTrace);
344 
345   unsigned RootLatency = 0;
346   for (auto *I : DelInstrs)
347     RootLatency += TSchedModel.computeInstrLatency(I);
348 
349   return {NewRootLatency, RootLatency};
350 }
351 
352 bool MachineCombiner::reduceRegisterPressure(
353     MachineInstr &Root, MachineBasicBlock *MBB,
354     SmallVectorImpl<MachineInstr *> &InsInstrs,
355     SmallVectorImpl<MachineInstr *> &DelInstrs,
356     MachineCombinerPattern Pattern) {
357   // FIXME: for now, we don't do any check for the register pressure patterns.
358   // We treat them as always profitable. But we can do better if we make
359   // RegPressureTracker class be aware of TIE attribute. Then we can get an
360   // accurate compare of register pressure with DelInstrs or InsInstrs.
361   return true;
362 }
363 
364 /// The DAGCombine code sequence ends in MI (Machine Instruction) Root.
365 /// The new code sequence ends in MI NewRoot. A necessary condition for the new
366 /// sequence to replace the old sequence is that it cannot lengthen the critical
367 /// path. The definition of "improve" may be restricted by specifying that the
368 /// new path improves the data dependency chain (MustReduceDepth).
369 bool MachineCombiner::improvesCriticalPathLen(
370     MachineBasicBlock *MBB, MachineInstr *Root,
371     MachineTraceMetrics::Trace BlockTrace,
372     SmallVectorImpl<MachineInstr *> &InsInstrs,
373     SmallVectorImpl<MachineInstr *> &DelInstrs,
374     DenseMap<unsigned, unsigned> &InstrIdxForVirtReg,
375     MachineCombinerPattern Pattern,
376     bool SlackIsAccurate) {
377   // Get depth and latency of NewRoot and Root.
378   unsigned NewRootDepth =
379       getDepth(InsInstrs, InstrIdxForVirtReg, BlockTrace, *MBB);
380   unsigned RootDepth = BlockTrace.getInstrCycles(*Root).Depth;
381 
382   LLVM_DEBUG(dbgs() << "  Dependence data for " << *Root << "\tNewRootDepth: "
383                     << NewRootDepth << "\tRootDepth: " << RootDepth);
384 
385   // For a transform such as reassociation, the cost equation is
386   // conservatively calculated so that we must improve the depth (data
387   // dependency cycles) in the critical path to proceed with the transform.
388   // Being conservative also protects against inaccuracies in the underlying
389   // machine trace metrics and CPU models.
390   if (getCombinerObjective(Pattern) == CombinerObjective::MustReduceDepth) {
391     LLVM_DEBUG(dbgs() << "\tIt MustReduceDepth ");
392     LLVM_DEBUG(NewRootDepth < RootDepth
393                    ? dbgs() << "\t  and it does it\n"
394                    : dbgs() << "\t  but it does NOT do it\n");
395     return NewRootDepth < RootDepth;
396   }
397 
398   // A more flexible cost calculation for the critical path includes the slack
399   // of the original code sequence. This may allow the transform to proceed
400   // even if the instruction depths (data dependency cycles) become worse.
401 
402   // Account for the latency of the inserted and deleted instructions by
403   unsigned NewRootLatency, RootLatency;
404   if (TII->accumulateInstrSeqToRootLatency(*Root)) {
405     std::tie(NewRootLatency, RootLatency) =
406         getLatenciesForInstrSequences(*Root, InsInstrs, DelInstrs, BlockTrace);
407   } else {
408     NewRootLatency = TSchedModel.computeInstrLatency(InsInstrs.back());
409     RootLatency = TSchedModel.computeInstrLatency(Root);
410   }
411 
412   unsigned RootSlack = BlockTrace.getInstrSlack(*Root);
413   unsigned NewCycleCount = NewRootDepth + NewRootLatency;
414   unsigned OldCycleCount =
415       RootDepth + RootLatency + (SlackIsAccurate ? RootSlack : 0);
416   LLVM_DEBUG(dbgs() << "\n\tNewRootLatency: " << NewRootLatency
417                     << "\tRootLatency: " << RootLatency << "\n\tRootSlack: "
418                     << RootSlack << " SlackIsAccurate=" << SlackIsAccurate
419                     << "\n\tNewRootDepth + NewRootLatency = " << NewCycleCount
420                     << "\n\tRootDepth + RootLatency + RootSlack = "
421                     << OldCycleCount;);
422   LLVM_DEBUG(NewCycleCount <= OldCycleCount
423                  ? dbgs() << "\n\t  It IMPROVES PathLen because"
424                  : dbgs() << "\n\t  It DOES NOT improve PathLen because");
425   LLVM_DEBUG(dbgs() << "\n\t\tNewCycleCount = " << NewCycleCount
426                     << ", OldCycleCount = " << OldCycleCount << "\n");
427 
428   return NewCycleCount <= OldCycleCount;
429 }
430 
431 /// helper routine to convert instructions into SC
432 void MachineCombiner::instr2instrSC(
433     SmallVectorImpl<MachineInstr *> &Instrs,
434     SmallVectorImpl<const MCSchedClassDesc *> &InstrsSC) {
435   for (auto *InstrPtr : Instrs) {
436     unsigned Opc = InstrPtr->getOpcode();
437     unsigned Idx = TII->get(Opc).getSchedClass();
438     const MCSchedClassDesc *SC = SchedModel.getSchedClassDesc(Idx);
439     InstrsSC.push_back(SC);
440   }
441 }
442 
443 /// True when the new instructions do not increase resource length
444 bool MachineCombiner::preservesResourceLen(
445     MachineBasicBlock *MBB, MachineTraceMetrics::Trace BlockTrace,
446     SmallVectorImpl<MachineInstr *> &InsInstrs,
447     SmallVectorImpl<MachineInstr *> &DelInstrs) {
448   if (!TSchedModel.hasInstrSchedModel())
449     return true;
450 
451   // Compute current resource length
452 
453   //ArrayRef<const MachineBasicBlock *> MBBarr(MBB);
454   SmallVector <const MachineBasicBlock *, 1> MBBarr;
455   MBBarr.push_back(MBB);
456   unsigned ResLenBeforeCombine = BlockTrace.getResourceLength(MBBarr);
457 
458   // Deal with SC rather than Instructions.
459   SmallVector<const MCSchedClassDesc *, 16> InsInstrsSC;
460   SmallVector<const MCSchedClassDesc *, 16> DelInstrsSC;
461 
462   instr2instrSC(InsInstrs, InsInstrsSC);
463   instr2instrSC(DelInstrs, DelInstrsSC);
464 
465   ArrayRef<const MCSchedClassDesc *> MSCInsArr{InsInstrsSC};
466   ArrayRef<const MCSchedClassDesc *> MSCDelArr{DelInstrsSC};
467 
468   // Compute new resource length.
469   unsigned ResLenAfterCombine =
470       BlockTrace.getResourceLength(MBBarr, MSCInsArr, MSCDelArr);
471 
472   LLVM_DEBUG(dbgs() << "\t\tResource length before replacement: "
473                     << ResLenBeforeCombine
474                     << " and after: " << ResLenAfterCombine << "\n";);
475   LLVM_DEBUG(
476       ResLenAfterCombine <=
477       ResLenBeforeCombine + TII->getExtendResourceLenLimit()
478           ? dbgs() << "\t\t  As result it IMPROVES/PRESERVES Resource Length\n"
479           : dbgs() << "\t\t  As result it DOES NOT improve/preserve Resource "
480                       "Length\n");
481 
482   return ResLenAfterCombine <=
483          ResLenBeforeCombine + TII->getExtendResourceLenLimit();
484 }
485 
486 /// Inserts InsInstrs and deletes DelInstrs. Incrementally updates instruction
487 /// depths if requested.
488 ///
489 /// \param MBB basic block to insert instructions in
490 /// \param MI current machine instruction
491 /// \param InsInstrs new instructions to insert in \p MBB
492 /// \param DelInstrs instruction to delete from \p MBB
493 /// \param TraceEnsemble is a pointer to the machine trace information
494 /// \param RegUnits set of live registers, needed to compute instruction depths
495 /// \param TII is target instruction info, used to call target hook
496 /// \param Pattern is used to call target hook finalizeInsInstrs
497 /// \param IncrementalUpdate if true, compute instruction depths incrementally,
498 ///                          otherwise invalidate the trace
499 static void insertDeleteInstructions(
500     MachineBasicBlock *MBB, MachineInstr &MI,
501     SmallVectorImpl<MachineInstr *> &InsInstrs,
502     SmallVectorImpl<MachineInstr *> &DelInstrs,
503     MachineTraceMetrics::Ensemble *TraceEnsemble,
504     SparseSet<LiveRegUnit> &RegUnits, const TargetInstrInfo *TII,
505     MachineCombinerPattern Pattern, bool IncrementalUpdate) {
506   // If we want to fix up some placeholder for some target, do it now.
507   // We need this because in genAlternativeCodeSequence, we have not decided the
508   // better pattern InsInstrs or DelInstrs, so we don't want generate some
509   // sideeffect to the function. For example we need to delay the constant pool
510   // entry creation here after InsInstrs is selected as better pattern.
511   // Otherwise the constant pool entry created for InsInstrs will not be deleted
512   // even if InsInstrs is not the better pattern.
513   TII->finalizeInsInstrs(MI, Pattern, InsInstrs);
514 
515   for (auto *InstrPtr : InsInstrs)
516     MBB->insert((MachineBasicBlock::iterator)&MI, InstrPtr);
517 
518   for (auto *InstrPtr : DelInstrs) {
519     InstrPtr->eraseFromParent();
520     // Erase all LiveRegs defined by the removed instruction
521     for (auto *I = RegUnits.begin(); I != RegUnits.end();) {
522       if (I->MI == InstrPtr)
523         I = RegUnits.erase(I);
524       else
525         I++;
526     }
527   }
528 
529   if (IncrementalUpdate)
530     for (auto *InstrPtr : InsInstrs)
531       TraceEnsemble->updateDepth(MBB, *InstrPtr, RegUnits);
532   else
533     TraceEnsemble->invalidate(MBB);
534 
535   NumInstCombined++;
536 }
537 
538 // Check that the difference between original and new latency is decreasing for
539 // later patterns. This helps to discover sub-optimal pattern orderings.
540 void MachineCombiner::verifyPatternOrder(
541     MachineBasicBlock *MBB, MachineInstr &Root,
542     SmallVector<MachineCombinerPattern, 16> &Patterns) {
543   long PrevLatencyDiff = std::numeric_limits<long>::max();
544   (void)PrevLatencyDiff; // Variable is used in assert only.
545   for (auto P : Patterns) {
546     SmallVector<MachineInstr *, 16> InsInstrs;
547     SmallVector<MachineInstr *, 16> DelInstrs;
548     DenseMap<unsigned, unsigned> InstrIdxForVirtReg;
549     TII->genAlternativeCodeSequence(Root, P, InsInstrs, DelInstrs,
550                                     InstrIdxForVirtReg);
551     // Found pattern, but did not generate alternative sequence.
552     // This can happen e.g. when an immediate could not be materialized
553     // in a single instruction.
554     if (InsInstrs.empty() || !TSchedModel.hasInstrSchedModelOrItineraries())
555       continue;
556 
557     unsigned NewRootLatency, RootLatency;
558     std::tie(NewRootLatency, RootLatency) = getLatenciesForInstrSequences(
559         Root, InsInstrs, DelInstrs, TraceEnsemble->getTrace(MBB));
560     long CurrentLatencyDiff = ((long)RootLatency) - ((long)NewRootLatency);
561     assert(CurrentLatencyDiff <= PrevLatencyDiff &&
562            "Current pattern is better than previous pattern.");
563     PrevLatencyDiff = CurrentLatencyDiff;
564   }
565 }
566 
567 /// Substitute a slow code sequence with a faster one by
568 /// evaluating instruction combining pattern.
569 /// The prototype of such a pattern is MUl + ADD -> MADD. Performs instruction
570 /// combining based on machine trace metrics. Only combine a sequence of
571 /// instructions  when this neither lengthens the critical path nor increases
572 /// resource pressure. When optimizing for codesize always combine when the new
573 /// sequence is shorter.
574 bool MachineCombiner::combineInstructions(MachineBasicBlock *MBB) {
575   bool Changed = false;
576   LLVM_DEBUG(dbgs() << "Combining MBB " << MBB->getName() << "\n");
577 
578   bool IncrementalUpdate = false;
579   auto BlockIter = MBB->begin();
580   decltype(BlockIter) LastUpdate;
581   // Check if the block is in a loop.
582   const MachineLoop *ML = MLI->getLoopFor(MBB);
583   if (!TraceEnsemble)
584     TraceEnsemble = Traces->getEnsemble(TII->getMachineCombinerTraceStrategy());
585 
586   SparseSet<LiveRegUnit> RegUnits;
587   RegUnits.setUniverse(TRI->getNumRegUnits());
588 
589   bool OptForSize = OptSize || llvm::shouldOptimizeForSize(MBB, PSI, MBFI);
590 
591   bool DoRegPressureReduce =
592       TII->shouldReduceRegisterPressure(MBB, &RegClassInfo);
593 
594   while (BlockIter != MBB->end()) {
595     auto &MI = *BlockIter++;
596     SmallVector<MachineCombinerPattern, 16> Patterns;
597     // The motivating example is:
598     //
599     //     MUL  Other        MUL_op1 MUL_op2  Other
600     //      \    /               \      |    /
601     //      ADD/SUB      =>        MADD/MSUB
602     //      (=Root)                (=NewRoot)
603 
604     // The DAGCombine code always replaced MUL + ADD/SUB by MADD. While this is
605     // usually beneficial for code size it unfortunately can hurt performance
606     // when the ADD is on the critical path, but the MUL is not. With the
607     // substitution the MUL becomes part of the critical path (in form of the
608     // MADD) and can lengthen it on architectures where the MADD latency is
609     // longer than the ADD latency.
610     //
611     // For each instruction we check if it can be the root of a combiner
612     // pattern. Then for each pattern the new code sequence in form of MI is
613     // generated and evaluated. When the efficiency criteria (don't lengthen
614     // critical path, don't use more resources) is met the new sequence gets
615     // hooked up into the basic block before the old sequence is removed.
616     //
617     // The algorithm does not try to evaluate all patterns and pick the best.
618     // This is only an artificial restriction though. In practice there is
619     // mostly one pattern, and getMachineCombinerPatterns() can order patterns
620     // based on an internal cost heuristic. If
621     // machine-combiner-verify-pattern-order is enabled, all patterns are
622     // checked to ensure later patterns do not provide better latency savings.
623 
624     if (!TII->getMachineCombinerPatterns(MI, Patterns, DoRegPressureReduce))
625       continue;
626 
627     if (VerifyPatternOrder)
628       verifyPatternOrder(MBB, MI, Patterns);
629 
630     for (const auto P : Patterns) {
631       SmallVector<MachineInstr *, 16> InsInstrs;
632       SmallVector<MachineInstr *, 16> DelInstrs;
633       DenseMap<unsigned, unsigned> InstrIdxForVirtReg;
634       TII->genAlternativeCodeSequence(MI, P, InsInstrs, DelInstrs,
635                                       InstrIdxForVirtReg);
636       // Found pattern, but did not generate alternative sequence.
637       // This can happen e.g. when an immediate could not be materialized
638       // in a single instruction.
639       if (InsInstrs.empty())
640         continue;
641 
642       LLVM_DEBUG(if (dump_intrs) {
643         dbgs() << "\tFor the Pattern (" << (int)P
644                << ") these instructions could be removed\n";
645         for (auto const *InstrPtr : DelInstrs)
646           InstrPtr->print(dbgs(), /*IsStandalone*/false, /*SkipOpers*/false,
647                           /*SkipDebugLoc*/false, /*AddNewLine*/true, TII);
648         dbgs() << "\tThese instructions could replace the removed ones\n";
649         for (auto const *InstrPtr : InsInstrs)
650           InstrPtr->print(dbgs(), /*IsStandalone*/false, /*SkipOpers*/false,
651                           /*SkipDebugLoc*/false, /*AddNewLine*/true, TII);
652       });
653 
654       if (IncrementalUpdate && LastUpdate != BlockIter) {
655         // Update depths since the last incremental update.
656         TraceEnsemble->updateDepths(LastUpdate, BlockIter, RegUnits);
657         LastUpdate = BlockIter;
658       }
659 
660       if (DoRegPressureReduce &&
661           getCombinerObjective(P) ==
662               CombinerObjective::MustReduceRegisterPressure) {
663         if (MBB->size() > inc_threshold) {
664           // Use incremental depth updates for basic blocks above threshold
665           IncrementalUpdate = true;
666           LastUpdate = BlockIter;
667         }
668         if (reduceRegisterPressure(MI, MBB, InsInstrs, DelInstrs, P)) {
669           // Replace DelInstrs with InsInstrs.
670           insertDeleteInstructions(MBB, MI, InsInstrs, DelInstrs, TraceEnsemble,
671                                    RegUnits, TII, P, IncrementalUpdate);
672           Changed |= true;
673 
674           // Go back to previous instruction as it may have ILP reassociation
675           // opportunity.
676           BlockIter--;
677           break;
678         }
679       }
680 
681       if (ML && TII->isThroughputPattern(P)) {
682         LLVM_DEBUG(dbgs() << "\t Replacing due to throughput pattern in loop\n");
683         insertDeleteInstructions(MBB, MI, InsInstrs, DelInstrs, TraceEnsemble,
684                                  RegUnits, TII, P, IncrementalUpdate);
685         // Eagerly stop after the first pattern fires.
686         Changed = true;
687         break;
688       } else if (OptForSize && InsInstrs.size() < DelInstrs.size()) {
689         LLVM_DEBUG(dbgs() << "\t Replacing due to OptForSize ("
690                           << InsInstrs.size() << " < "
691                           << DelInstrs.size() << ")\n");
692         insertDeleteInstructions(MBB, MI, InsInstrs, DelInstrs, TraceEnsemble,
693                                  RegUnits, TII, P, IncrementalUpdate);
694         // Eagerly stop after the first pattern fires.
695         Changed = true;
696         break;
697       } else {
698         // For big basic blocks, we only compute the full trace the first time
699         // we hit this. We do not invalidate the trace, but instead update the
700         // instruction depths incrementally.
701         // NOTE: Only the instruction depths up to MI are accurate. All other
702         // trace information is not updated.
703         MachineTraceMetrics::Trace BlockTrace = TraceEnsemble->getTrace(MBB);
704         Traces->verifyAnalysis();
705         if (improvesCriticalPathLen(MBB, &MI, BlockTrace, InsInstrs, DelInstrs,
706                                     InstrIdxForVirtReg, P,
707                                     !IncrementalUpdate) &&
708             preservesResourceLen(MBB, BlockTrace, InsInstrs, DelInstrs)) {
709           if (MBB->size() > inc_threshold) {
710             // Use incremental depth updates for basic blocks above treshold
711             IncrementalUpdate = true;
712             LastUpdate = BlockIter;
713           }
714 
715           insertDeleteInstructions(MBB, MI, InsInstrs, DelInstrs, TraceEnsemble,
716                                    RegUnits, TII, P, IncrementalUpdate);
717 
718           // Eagerly stop after the first pattern fires.
719           Changed = true;
720           break;
721         }
722         // Cleanup instructions of the alternative code sequence. There is no
723         // use for them.
724         MachineFunction *MF = MBB->getParent();
725         for (auto *InstrPtr : InsInstrs)
726           MF->deleteMachineInstr(InstrPtr);
727       }
728       InstrIdxForVirtReg.clear();
729     }
730   }
731 
732   if (Changed && IncrementalUpdate)
733     Traces->invalidate(MBB);
734   return Changed;
735 }
736 
737 bool MachineCombiner::runOnMachineFunction(MachineFunction &MF) {
738   STI = &MF.getSubtarget();
739   TII = STI->getInstrInfo();
740   TRI = STI->getRegisterInfo();
741   SchedModel = STI->getSchedModel();
742   TSchedModel.init(STI);
743   MRI = &MF.getRegInfo();
744   MLI = &getAnalysis<MachineLoopInfo>();
745   Traces = &getAnalysis<MachineTraceMetrics>();
746   PSI = &getAnalysis<ProfileSummaryInfoWrapperPass>().getPSI();
747   MBFI = (PSI && PSI->hasProfileSummary()) ?
748          &getAnalysis<LazyMachineBlockFrequencyInfoPass>().getBFI() :
749          nullptr;
750   TraceEnsemble = nullptr;
751   OptSize = MF.getFunction().hasOptSize();
752   RegClassInfo.runOnMachineFunction(MF);
753 
754   LLVM_DEBUG(dbgs() << getPassName() << ": " << MF.getName() << '\n');
755   if (!TII->useMachineCombiner()) {
756     LLVM_DEBUG(
757         dbgs()
758         << "  Skipping pass: Target does not support machine combiner\n");
759     return false;
760   }
761 
762   bool Changed = false;
763 
764   // Try to combine instructions.
765   for (auto &MBB : MF)
766     Changed |= combineInstructions(&MBB);
767 
768   return Changed;
769 }
770