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