xref: /freebsd/contrib/llvm-project/llvm/lib/CodeGen/SelectOptimize.cpp (revision c57c26179033f64c2011a2d2a904ee3fa62e826a)
1 //===--- SelectOptimize.cpp - Convert select to branches if profitable ---===//
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 // This pass converts selects to conditional jumps when profitable.
10 //
11 //===----------------------------------------------------------------------===//
12 
13 #include "llvm/CodeGen/SelectOptimize.h"
14 #include "llvm/ADT/SmallVector.h"
15 #include "llvm/ADT/Statistic.h"
16 #include "llvm/Analysis/BlockFrequencyInfo.h"
17 #include "llvm/Analysis/BranchProbabilityInfo.h"
18 #include "llvm/Analysis/LoopInfo.h"
19 #include "llvm/Analysis/OptimizationRemarkEmitter.h"
20 #include "llvm/Analysis/ProfileSummaryInfo.h"
21 #include "llvm/Analysis/TargetTransformInfo.h"
22 #include "llvm/CodeGen/Passes.h"
23 #include "llvm/CodeGen/TargetLowering.h"
24 #include "llvm/CodeGen/TargetPassConfig.h"
25 #include "llvm/CodeGen/TargetSchedule.h"
26 #include "llvm/CodeGen/TargetSubtargetInfo.h"
27 #include "llvm/IR/BasicBlock.h"
28 #include "llvm/IR/Dominators.h"
29 #include "llvm/IR/Function.h"
30 #include "llvm/IR/IRBuilder.h"
31 #include "llvm/IR/Instruction.h"
32 #include "llvm/IR/PatternMatch.h"
33 #include "llvm/IR/ProfDataUtils.h"
34 #include "llvm/InitializePasses.h"
35 #include "llvm/Pass.h"
36 #include "llvm/Support/ScaledNumber.h"
37 #include "llvm/Target/TargetMachine.h"
38 #include "llvm/Transforms/Utils/SizeOpts.h"
39 #include <algorithm>
40 #include <memory>
41 #include <queue>
42 #include <stack>
43 
44 using namespace llvm;
45 using namespace llvm::PatternMatch;
46 
47 #define DEBUG_TYPE "select-optimize"
48 
49 STATISTIC(NumSelectOptAnalyzed,
50           "Number of select groups considered for conversion to branch");
51 STATISTIC(NumSelectConvertedExpColdOperand,
52           "Number of select groups converted due to expensive cold operand");
53 STATISTIC(NumSelectConvertedHighPred,
54           "Number of select groups converted due to high-predictability");
55 STATISTIC(NumSelectUnPred,
56           "Number of select groups not converted due to unpredictability");
57 STATISTIC(NumSelectColdBB,
58           "Number of select groups not converted due to cold basic block");
59 STATISTIC(NumSelectConvertedLoop,
60           "Number of select groups converted due to loop-level analysis");
61 STATISTIC(NumSelectsConverted, "Number of selects converted");
62 
63 static cl::opt<unsigned> ColdOperandThreshold(
64     "cold-operand-threshold",
65     cl::desc("Maximum frequency of path for an operand to be considered cold."),
66     cl::init(20), cl::Hidden);
67 
68 static cl::opt<unsigned> ColdOperandMaxCostMultiplier(
69     "cold-operand-max-cost-multiplier",
70     cl::desc("Maximum cost multiplier of TCC_expensive for the dependence "
71              "slice of a cold operand to be considered inexpensive."),
72     cl::init(1), cl::Hidden);
73 
74 static cl::opt<unsigned>
75     GainGradientThreshold("select-opti-loop-gradient-gain-threshold",
76                           cl::desc("Gradient gain threshold (%)."),
77                           cl::init(25), cl::Hidden);
78 
79 static cl::opt<unsigned>
80     GainCycleThreshold("select-opti-loop-cycle-gain-threshold",
81                        cl::desc("Minimum gain per loop (in cycles) threshold."),
82                        cl::init(4), cl::Hidden);
83 
84 static cl::opt<unsigned> GainRelativeThreshold(
85     "select-opti-loop-relative-gain-threshold",
86     cl::desc(
87         "Minimum relative gain per loop threshold (1/X). Defaults to 12.5%"),
88     cl::init(8), cl::Hidden);
89 
90 static cl::opt<unsigned> MispredictDefaultRate(
91     "mispredict-default-rate", cl::Hidden, cl::init(25),
92     cl::desc("Default mispredict rate (initialized to 25%)."));
93 
94 static cl::opt<bool>
95     DisableLoopLevelHeuristics("disable-loop-level-heuristics", cl::Hidden,
96                                cl::init(false),
97                                cl::desc("Disable loop-level heuristics."));
98 
99 namespace {
100 
101 class SelectOptimizeImpl {
102   const TargetMachine *TM = nullptr;
103   const TargetSubtargetInfo *TSI = nullptr;
104   const TargetLowering *TLI = nullptr;
105   const TargetTransformInfo *TTI = nullptr;
106   const LoopInfo *LI = nullptr;
107   BlockFrequencyInfo *BFI;
108   ProfileSummaryInfo *PSI = nullptr;
109   OptimizationRemarkEmitter *ORE = nullptr;
110   TargetSchedModel TSchedModel;
111 
112 public:
113   SelectOptimizeImpl() = default;
114   SelectOptimizeImpl(const TargetMachine *TM) : TM(TM){};
115   PreservedAnalyses run(Function &F, FunctionAnalysisManager &FAM);
116   bool runOnFunction(Function &F, Pass &P);
117 
118   using Scaled64 = ScaledNumber<uint64_t>;
119 
120   struct CostInfo {
121     /// Predicated cost (with selects as conditional moves).
122     Scaled64 PredCost;
123     /// Non-predicated cost (with selects converted to branches).
124     Scaled64 NonPredCost;
125   };
126 
127   /// SelectLike is an abstraction over SelectInst and other operations that can
128   /// act like selects. For example Or(Zext(icmp), X) can be treated like
129   /// select(icmp, X|1, X).
130   class SelectLike {
131     SelectLike(Instruction *I) : I(I) {}
132 
133     Instruction *I;
134 
135   public:
136     /// Match a select or select-like instruction, returning a SelectLike.
137     static SelectLike match(Instruction *I) {
138       // Select instruction are what we are usually looking for.
139       if (isa<SelectInst>(I))
140         return SelectLike(I);
141 
142       // An Or(zext(i1 X), Y) can also be treated like a select, with condition
143       // C and values Y|1 and Y.
144       Value *X;
145       if (PatternMatch::match(
146               I, m_c_Or(m_OneUse(m_ZExt(m_Value(X))), m_Value())) &&
147           X->getType()->isIntegerTy(1))
148         return SelectLike(I);
149 
150       return SelectLike(nullptr);
151     }
152 
153     bool isValid() { return I; }
154     operator bool() { return isValid(); }
155 
156     Instruction *getI() { return I; }
157     const Instruction *getI() const { return I; }
158 
159     Type *getType() const { return I->getType(); }
160 
161     /// Return the condition for the SelectLike instruction. For example the
162     /// condition of a select or c in `or(zext(c), x)`
163     Value *getCondition() const {
164       if (auto *Sel = dyn_cast<SelectInst>(I))
165         return Sel->getCondition();
166       // Or(zext) case
167       if (auto *BO = dyn_cast<BinaryOperator>(I)) {
168         Value *X;
169         if (PatternMatch::match(BO->getOperand(0),
170                                 m_OneUse(m_ZExt(m_Value(X)))))
171           return X;
172         if (PatternMatch::match(BO->getOperand(1),
173                                 m_OneUse(m_ZExt(m_Value(X)))))
174           return X;
175       }
176 
177       llvm_unreachable("Unhandled case in getCondition");
178     }
179 
180     /// Return the true value for the SelectLike instruction. Note this may not
181     /// exist for all SelectLike instructions. For example, for `or(zext(c), x)`
182     /// the true value would be `or(x,1)`. As this value does not exist, nullptr
183     /// is returned.
184     Value *getTrueValue() const {
185       if (auto *Sel = dyn_cast<SelectInst>(I))
186         return Sel->getTrueValue();
187       // Or(zext) case - The true value is Or(X), so return nullptr as the value
188       // does not yet exist.
189       if (isa<BinaryOperator>(I))
190         return nullptr;
191 
192       llvm_unreachable("Unhandled case in getTrueValue");
193     }
194 
195     /// Return the false value for the SelectLike instruction. For example the
196     /// getFalseValue of a select or `x` in `or(zext(c), x)` (which is
197     /// `select(c, x|1, x)`)
198     Value *getFalseValue() const {
199       if (auto *Sel = dyn_cast<SelectInst>(I))
200         return Sel->getFalseValue();
201       // Or(zext) case - return the operand which is not the zext.
202       if (auto *BO = dyn_cast<BinaryOperator>(I)) {
203         Value *X;
204         if (PatternMatch::match(BO->getOperand(0),
205                                 m_OneUse(m_ZExt(m_Value(X)))))
206           return BO->getOperand(1);
207         if (PatternMatch::match(BO->getOperand(1),
208                                 m_OneUse(m_ZExt(m_Value(X)))))
209           return BO->getOperand(0);
210       }
211 
212       llvm_unreachable("Unhandled case in getFalseValue");
213     }
214 
215     /// Return the NonPredCost cost of the true op, given the costs in
216     /// InstCostMap. This may need to be generated for select-like instructions.
217     Scaled64 getTrueOpCost(DenseMap<const Instruction *, CostInfo> &InstCostMap,
218                            const TargetTransformInfo *TTI) {
219       if (auto *Sel = dyn_cast<SelectInst>(I))
220         if (auto *I = dyn_cast<Instruction>(Sel->getTrueValue()))
221           return InstCostMap.contains(I) ? InstCostMap[I].NonPredCost
222                                          : Scaled64::getZero();
223 
224       // Or case - add the cost of an extra Or to the cost of the False case.
225       if (isa<BinaryOperator>(I))
226         if (auto I = dyn_cast<Instruction>(getFalseValue()))
227           if (InstCostMap.contains(I)) {
228             InstructionCost OrCost = TTI->getArithmeticInstrCost(
229                 Instruction::Or, I->getType(), TargetTransformInfo::TCK_Latency,
230                 {TargetTransformInfo::OK_AnyValue,
231                  TargetTransformInfo::OP_None},
232                 {TTI::OK_UniformConstantValue, TTI::OP_PowerOf2});
233             return InstCostMap[I].NonPredCost +
234                    Scaled64::get(*OrCost.getValue());
235           }
236 
237       return Scaled64::getZero();
238     }
239 
240     /// Return the NonPredCost cost of the false op, given the costs in
241     /// InstCostMap. This may need to be generated for select-like instructions.
242     Scaled64
243     getFalseOpCost(DenseMap<const Instruction *, CostInfo> &InstCostMap,
244                    const TargetTransformInfo *TTI) {
245       if (auto *Sel = dyn_cast<SelectInst>(I))
246         if (auto *I = dyn_cast<Instruction>(Sel->getFalseValue()))
247           return InstCostMap.contains(I) ? InstCostMap[I].NonPredCost
248                                          : Scaled64::getZero();
249 
250       // Or case - return the cost of the false case
251       if (isa<BinaryOperator>(I))
252         if (auto I = dyn_cast<Instruction>(getFalseValue()))
253           if (InstCostMap.contains(I))
254             return InstCostMap[I].NonPredCost;
255 
256       return Scaled64::getZero();
257     }
258   };
259 
260 private:
261   // Select groups consist of consecutive select instructions with the same
262   // condition.
263   using SelectGroup = SmallVector<SelectLike, 2>;
264   using SelectGroups = SmallVector<SelectGroup, 2>;
265 
266   // Converts select instructions of a function to conditional jumps when deemed
267   // profitable. Returns true if at least one select was converted.
268   bool optimizeSelects(Function &F);
269 
270   // Heuristics for determining which select instructions can be profitably
271   // conveted to branches. Separate heuristics for selects in inner-most loops
272   // and the rest of code regions (base heuristics for non-inner-most loop
273   // regions).
274   void optimizeSelectsBase(Function &F, SelectGroups &ProfSIGroups);
275   void optimizeSelectsInnerLoops(Function &F, SelectGroups &ProfSIGroups);
276 
277   // Converts to branches the select groups that were deemed
278   // profitable-to-convert.
279   void convertProfitableSIGroups(SelectGroups &ProfSIGroups);
280 
281   // Splits selects of a given basic block into select groups.
282   void collectSelectGroups(BasicBlock &BB, SelectGroups &SIGroups);
283 
284   // Determines for which select groups it is profitable converting to branches
285   // (base and inner-most-loop heuristics).
286   void findProfitableSIGroupsBase(SelectGroups &SIGroups,
287                                   SelectGroups &ProfSIGroups);
288   void findProfitableSIGroupsInnerLoops(const Loop *L, SelectGroups &SIGroups,
289                                         SelectGroups &ProfSIGroups);
290 
291   // Determines if a select group should be converted to a branch (base
292   // heuristics).
293   bool isConvertToBranchProfitableBase(const SelectGroup &ASI);
294 
295   // Returns true if there are expensive instructions in the cold value
296   // operand's (if any) dependence slice of any of the selects of the given
297   // group.
298   bool hasExpensiveColdOperand(const SelectGroup &ASI);
299 
300   // For a given source instruction, collect its backwards dependence slice
301   // consisting of instructions exclusively computed for producing the operands
302   // of the source instruction.
303   void getExclBackwardsSlice(Instruction *I, std::stack<Instruction *> &Slice,
304                              Instruction *SI, bool ForSinking = false);
305 
306   // Returns true if the condition of the select is highly predictable.
307   bool isSelectHighlyPredictable(const SelectLike SI);
308 
309   // Loop-level checks to determine if a non-predicated version (with branches)
310   // of the given loop is more profitable than its predicated version.
311   bool checkLoopHeuristics(const Loop *L, const CostInfo LoopDepth[2]);
312 
313   // Computes instruction and loop-critical-path costs for both the predicated
314   // and non-predicated version of the given loop.
315   bool computeLoopCosts(const Loop *L, const SelectGroups &SIGroups,
316                         DenseMap<const Instruction *, CostInfo> &InstCostMap,
317                         CostInfo *LoopCost);
318 
319   // Returns a set of all the select instructions in the given select groups.
320   SmallDenseMap<const Instruction *, SelectLike, 2>
321   getSImap(const SelectGroups &SIGroups);
322 
323   // Returns the latency cost of a given instruction.
324   std::optional<uint64_t> computeInstCost(const Instruction *I);
325 
326   // Returns the misprediction cost of a given select when converted to branch.
327   Scaled64 getMispredictionCost(const SelectLike SI, const Scaled64 CondCost);
328 
329   // Returns the cost of a branch when the prediction is correct.
330   Scaled64 getPredictedPathCost(Scaled64 TrueCost, Scaled64 FalseCost,
331                                 const SelectLike SI);
332 
333   // Returns true if the target architecture supports lowering a given select.
334   bool isSelectKindSupported(const SelectLike SI);
335 };
336 
337 class SelectOptimize : public FunctionPass {
338   SelectOptimizeImpl Impl;
339 
340 public:
341   static char ID;
342 
343   SelectOptimize() : FunctionPass(ID) {
344     initializeSelectOptimizePass(*PassRegistry::getPassRegistry());
345   }
346 
347   bool runOnFunction(Function &F) override {
348     return Impl.runOnFunction(F, *this);
349   }
350 
351   void getAnalysisUsage(AnalysisUsage &AU) const override {
352     AU.addRequired<ProfileSummaryInfoWrapperPass>();
353     AU.addRequired<TargetPassConfig>();
354     AU.addRequired<TargetTransformInfoWrapperPass>();
355     AU.addRequired<LoopInfoWrapperPass>();
356     AU.addRequired<BlockFrequencyInfoWrapperPass>();
357     AU.addRequired<OptimizationRemarkEmitterWrapperPass>();
358   }
359 };
360 
361 } // namespace
362 
363 PreservedAnalyses SelectOptimizePass::run(Function &F,
364                                           FunctionAnalysisManager &FAM) {
365   SelectOptimizeImpl Impl(TM);
366   return Impl.run(F, FAM);
367 }
368 
369 char SelectOptimize::ID = 0;
370 
371 INITIALIZE_PASS_BEGIN(SelectOptimize, DEBUG_TYPE, "Optimize selects", false,
372                       false)
373 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
374 INITIALIZE_PASS_DEPENDENCY(ProfileSummaryInfoWrapperPass)
375 INITIALIZE_PASS_DEPENDENCY(TargetPassConfig)
376 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
377 INITIALIZE_PASS_DEPENDENCY(BlockFrequencyInfoWrapperPass)
378 INITIALIZE_PASS_DEPENDENCY(OptimizationRemarkEmitterWrapperPass)
379 INITIALIZE_PASS_END(SelectOptimize, DEBUG_TYPE, "Optimize selects", false,
380                     false)
381 
382 FunctionPass *llvm::createSelectOptimizePass() { return new SelectOptimize(); }
383 
384 PreservedAnalyses SelectOptimizeImpl::run(Function &F,
385                                           FunctionAnalysisManager &FAM) {
386   TSI = TM->getSubtargetImpl(F);
387   TLI = TSI->getTargetLowering();
388 
389   // If none of the select types are supported then skip this pass.
390   // This is an optimization pass. Legality issues will be handled by
391   // instruction selection.
392   if (!TLI->isSelectSupported(TargetLowering::ScalarValSelect) &&
393       !TLI->isSelectSupported(TargetLowering::ScalarCondVectorVal) &&
394       !TLI->isSelectSupported(TargetLowering::VectorMaskSelect))
395     return PreservedAnalyses::all();
396 
397   TTI = &FAM.getResult<TargetIRAnalysis>(F);
398   if (!TTI->enableSelectOptimize())
399     return PreservedAnalyses::all();
400 
401   PSI = FAM.getResult<ModuleAnalysisManagerFunctionProxy>(F)
402             .getCachedResult<ProfileSummaryAnalysis>(*F.getParent());
403   assert(PSI && "This pass requires module analysis pass `profile-summary`!");
404   BFI = &FAM.getResult<BlockFrequencyAnalysis>(F);
405 
406   // When optimizing for size, selects are preferable over branches.
407   if (F.hasOptSize() || llvm::shouldOptimizeForSize(&F, PSI, BFI))
408     return PreservedAnalyses::all();
409 
410   LI = &FAM.getResult<LoopAnalysis>(F);
411   ORE = &FAM.getResult<OptimizationRemarkEmitterAnalysis>(F);
412   TSchedModel.init(TSI);
413 
414   bool Changed = optimizeSelects(F);
415   return Changed ? PreservedAnalyses::none() : PreservedAnalyses::all();
416 }
417 
418 bool SelectOptimizeImpl::runOnFunction(Function &F, Pass &P) {
419   TM = &P.getAnalysis<TargetPassConfig>().getTM<TargetMachine>();
420   TSI = TM->getSubtargetImpl(F);
421   TLI = TSI->getTargetLowering();
422 
423   // If none of the select types are supported then skip this pass.
424   // This is an optimization pass. Legality issues will be handled by
425   // instruction selection.
426   if (!TLI->isSelectSupported(TargetLowering::ScalarValSelect) &&
427       !TLI->isSelectSupported(TargetLowering::ScalarCondVectorVal) &&
428       !TLI->isSelectSupported(TargetLowering::VectorMaskSelect))
429     return false;
430 
431   TTI = &P.getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
432 
433   if (!TTI->enableSelectOptimize())
434     return false;
435 
436   LI = &P.getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
437   BFI = &P.getAnalysis<BlockFrequencyInfoWrapperPass>().getBFI();
438   PSI = &P.getAnalysis<ProfileSummaryInfoWrapperPass>().getPSI();
439   ORE = &P.getAnalysis<OptimizationRemarkEmitterWrapperPass>().getORE();
440   TSchedModel.init(TSI);
441 
442   // When optimizing for size, selects are preferable over branches.
443   if (F.hasOptSize() || llvm::shouldOptimizeForSize(&F, PSI, BFI))
444     return false;
445 
446   return optimizeSelects(F);
447 }
448 
449 bool SelectOptimizeImpl::optimizeSelects(Function &F) {
450   // Determine for which select groups it is profitable converting to branches.
451   SelectGroups ProfSIGroups;
452   // Base heuristics apply only to non-loops and outer loops.
453   optimizeSelectsBase(F, ProfSIGroups);
454   // Separate heuristics for inner-most loops.
455   optimizeSelectsInnerLoops(F, ProfSIGroups);
456 
457   // Convert to branches the select groups that were deemed
458   // profitable-to-convert.
459   convertProfitableSIGroups(ProfSIGroups);
460 
461   // Code modified if at least one select group was converted.
462   return !ProfSIGroups.empty();
463 }
464 
465 void SelectOptimizeImpl::optimizeSelectsBase(Function &F,
466                                              SelectGroups &ProfSIGroups) {
467   // Collect all the select groups.
468   SelectGroups SIGroups;
469   for (BasicBlock &BB : F) {
470     // Base heuristics apply only to non-loops and outer loops.
471     Loop *L = LI->getLoopFor(&BB);
472     if (L && L->isInnermost())
473       continue;
474     collectSelectGroups(BB, SIGroups);
475   }
476 
477   // Determine for which select groups it is profitable converting to branches.
478   findProfitableSIGroupsBase(SIGroups, ProfSIGroups);
479 }
480 
481 void SelectOptimizeImpl::optimizeSelectsInnerLoops(Function &F,
482                                                    SelectGroups &ProfSIGroups) {
483   SmallVector<Loop *, 4> Loops(LI->begin(), LI->end());
484   // Need to check size on each iteration as we accumulate child loops.
485   for (unsigned long i = 0; i < Loops.size(); ++i)
486     for (Loop *ChildL : Loops[i]->getSubLoops())
487       Loops.push_back(ChildL);
488 
489   for (Loop *L : Loops) {
490     if (!L->isInnermost())
491       continue;
492 
493     SelectGroups SIGroups;
494     for (BasicBlock *BB : L->getBlocks())
495       collectSelectGroups(*BB, SIGroups);
496 
497     findProfitableSIGroupsInnerLoops(L, SIGroups, ProfSIGroups);
498   }
499 }
500 
501 /// If \p isTrue is true, return the true value of \p SI, otherwise return
502 /// false value of \p SI. If the true/false value of \p SI is defined by any
503 /// select instructions in \p Selects, look through the defining select
504 /// instruction until the true/false value is not defined in \p Selects.
505 static Value *
506 getTrueOrFalseValue(SelectOptimizeImpl::SelectLike SI, bool isTrue,
507                     const SmallPtrSet<const Instruction *, 2> &Selects,
508                     IRBuilder<> &IB) {
509   Value *V = nullptr;
510   for (SelectInst *DefSI = dyn_cast<SelectInst>(SI.getI());
511        DefSI != nullptr && Selects.count(DefSI);
512        DefSI = dyn_cast<SelectInst>(V)) {
513     assert(DefSI->getCondition() == SI.getCondition() &&
514            "The condition of DefSI does not match with SI");
515     V = (isTrue ? DefSI->getTrueValue() : DefSI->getFalseValue());
516   }
517 
518   if (isa<BinaryOperator>(SI.getI())) {
519     assert(SI.getI()->getOpcode() == Instruction::Or &&
520            "Only currently handling Or instructions.");
521     V = SI.getFalseValue();
522     if (isTrue)
523       V = IB.CreateOr(V, ConstantInt::get(V->getType(), 1));
524   }
525 
526   assert(V && "Failed to get select true/false value");
527   return V;
528 }
529 
530 void SelectOptimizeImpl::convertProfitableSIGroups(SelectGroups &ProfSIGroups) {
531   for (SelectGroup &ASI : ProfSIGroups) {
532     // The code transformation here is a modified version of the sinking
533     // transformation in CodeGenPrepare::optimizeSelectInst with a more
534     // aggressive strategy of which instructions to sink.
535     //
536     // TODO: eliminate the redundancy of logic transforming selects to branches
537     // by removing CodeGenPrepare::optimizeSelectInst and optimizing here
538     // selects for all cases (with and without profile information).
539 
540     // Transform a sequence like this:
541     //    start:
542     //       %cmp = cmp uge i32 %a, %b
543     //       %sel = select i1 %cmp, i32 %c, i32 %d
544     //
545     // Into:
546     //    start:
547     //       %cmp = cmp uge i32 %a, %b
548     //       %cmp.frozen = freeze %cmp
549     //       br i1 %cmp.frozen, label %select.true, label %select.false
550     //    select.true:
551     //       br label %select.end
552     //    select.false:
553     //       br label %select.end
554     //    select.end:
555     //       %sel = phi i32 [ %c, %select.true ], [ %d, %select.false ]
556     //
557     // %cmp should be frozen, otherwise it may introduce undefined behavior.
558     // In addition, we may sink instructions that produce %c or %d into the
559     // destination(s) of the new branch.
560     // If the true or false blocks do not contain a sunken instruction, that
561     // block and its branch may be optimized away. In that case, one side of the
562     // first branch will point directly to select.end, and the corresponding PHI
563     // predecessor block will be the start block.
564 
565     // Find all the instructions that can be soundly sunk to the true/false
566     // blocks. These are instructions that are computed solely for producing the
567     // operands of the select instructions in the group and can be sunk without
568     // breaking the semantics of the LLVM IR (e.g., cannot sink instructions
569     // with side effects).
570     SmallVector<std::stack<Instruction *>, 2> TrueSlices, FalseSlices;
571     typedef std::stack<Instruction *>::size_type StackSizeType;
572     StackSizeType maxTrueSliceLen = 0, maxFalseSliceLen = 0;
573     for (SelectLike SI : ASI) {
574       // For each select, compute the sinkable dependence chains of the true and
575       // false operands.
576       if (auto *TI = dyn_cast_or_null<Instruction>(SI.getTrueValue())) {
577         std::stack<Instruction *> TrueSlice;
578         getExclBackwardsSlice(TI, TrueSlice, SI.getI(), true);
579         maxTrueSliceLen = std::max(maxTrueSliceLen, TrueSlice.size());
580         TrueSlices.push_back(TrueSlice);
581       }
582       if (auto *FI = dyn_cast_or_null<Instruction>(SI.getFalseValue())) {
583         if (isa<SelectInst>(SI.getI()) || !FI->hasOneUse()) {
584           std::stack<Instruction *> FalseSlice;
585           getExclBackwardsSlice(FI, FalseSlice, SI.getI(), true);
586           maxFalseSliceLen = std::max(maxFalseSliceLen, FalseSlice.size());
587           FalseSlices.push_back(FalseSlice);
588         }
589       }
590     }
591     // In the case of multiple select instructions in the same group, the order
592     // of non-dependent instructions (instructions of different dependence
593     // slices) in the true/false blocks appears to affect performance.
594     // Interleaving the slices seems to experimentally be the optimal approach.
595     // This interleaving scheduling allows for more ILP (with a natural downside
596     // of increasing a bit register pressure) compared to a simple ordering of
597     // one whole chain after another. One would expect that this ordering would
598     // not matter since the scheduling in the backend of the compiler  would
599     // take care of it, but apparently the scheduler fails to deliver optimal
600     // ILP with a naive ordering here.
601     SmallVector<Instruction *, 2> TrueSlicesInterleaved, FalseSlicesInterleaved;
602     for (StackSizeType IS = 0; IS < maxTrueSliceLen; ++IS) {
603       for (auto &S : TrueSlices) {
604         if (!S.empty()) {
605           TrueSlicesInterleaved.push_back(S.top());
606           S.pop();
607         }
608       }
609     }
610     for (StackSizeType IS = 0; IS < maxFalseSliceLen; ++IS) {
611       for (auto &S : FalseSlices) {
612         if (!S.empty()) {
613           FalseSlicesInterleaved.push_back(S.top());
614           S.pop();
615         }
616       }
617     }
618 
619     // We split the block containing the select(s) into two blocks.
620     SelectLike SI = ASI.front();
621     SelectLike LastSI = ASI.back();
622     BasicBlock *StartBlock = SI.getI()->getParent();
623     BasicBlock::iterator SplitPt = ++(BasicBlock::iterator(LastSI.getI()));
624     BasicBlock *EndBlock = StartBlock->splitBasicBlock(SplitPt, "select.end");
625     BFI->setBlockFreq(EndBlock, BFI->getBlockFreq(StartBlock));
626     // Delete the unconditional branch that was just created by the split.
627     StartBlock->getTerminator()->eraseFromParent();
628 
629     // Move any debug/pseudo instructions that were in-between the select
630     // group to the newly-created end block.
631     SmallVector<Instruction *, 2> DebugPseudoINS;
632     auto DIt = SI.getI()->getIterator();
633     while (&*DIt != LastSI.getI()) {
634       if (DIt->isDebugOrPseudoInst())
635         DebugPseudoINS.push_back(&*DIt);
636       DIt++;
637     }
638     for (auto *DI : DebugPseudoINS) {
639       DI->moveBeforePreserving(&*EndBlock->getFirstInsertionPt());
640     }
641 
642     // Duplicate implementation for DPValues, the non-instruction debug-info
643     // record. Helper lambda for moving DPValues to the end block.
644     auto TransferDPValues = [&](Instruction &I) {
645       for (auto &DPValue : llvm::make_early_inc_range(I.getDbgValueRange())) {
646         DPValue.removeFromParent();
647         EndBlock->insertDPValueBefore(&DPValue,
648                                       EndBlock->getFirstInsertionPt());
649       }
650     };
651 
652     // Iterate over all instructions in between SI and LastSI, not including
653     // SI itself. These are all the variable assignments that happen "in the
654     // middle" of the select group.
655     auto R = make_range(std::next(SI.getI()->getIterator()),
656                         std::next(LastSI.getI()->getIterator()));
657     llvm::for_each(R, TransferDPValues);
658 
659     // These are the new basic blocks for the conditional branch.
660     // At least one will become an actual new basic block.
661     BasicBlock *TrueBlock = nullptr, *FalseBlock = nullptr;
662     BranchInst *TrueBranch = nullptr, *FalseBranch = nullptr;
663     if (!TrueSlicesInterleaved.empty()) {
664       TrueBlock = BasicBlock::Create(EndBlock->getContext(), "select.true.sink",
665                                      EndBlock->getParent(), EndBlock);
666       TrueBranch = BranchInst::Create(EndBlock, TrueBlock);
667       TrueBranch->setDebugLoc(LastSI.getI()->getDebugLoc());
668       for (Instruction *TrueInst : TrueSlicesInterleaved)
669         TrueInst->moveBefore(TrueBranch);
670     }
671     if (!FalseSlicesInterleaved.empty()) {
672       FalseBlock =
673           BasicBlock::Create(EndBlock->getContext(), "select.false.sink",
674                              EndBlock->getParent(), EndBlock);
675       FalseBranch = BranchInst::Create(EndBlock, FalseBlock);
676       FalseBranch->setDebugLoc(LastSI.getI()->getDebugLoc());
677       for (Instruction *FalseInst : FalseSlicesInterleaved)
678         FalseInst->moveBefore(FalseBranch);
679     }
680     // If there was nothing to sink, then arbitrarily choose the 'false' side
681     // for a new input value to the PHI.
682     if (TrueBlock == FalseBlock) {
683       assert(TrueBlock == nullptr &&
684              "Unexpected basic block transform while optimizing select");
685 
686       FalseBlock = BasicBlock::Create(StartBlock->getContext(), "select.false",
687                                       EndBlock->getParent(), EndBlock);
688       auto *FalseBranch = BranchInst::Create(EndBlock, FalseBlock);
689       FalseBranch->setDebugLoc(SI.getI()->getDebugLoc());
690     }
691 
692     // Insert the real conditional branch based on the original condition.
693     // If we did not create a new block for one of the 'true' or 'false' paths
694     // of the condition, it means that side of the branch goes to the end block
695     // directly and the path originates from the start block from the point of
696     // view of the new PHI.
697     BasicBlock *TT, *FT;
698     if (TrueBlock == nullptr) {
699       TT = EndBlock;
700       FT = FalseBlock;
701       TrueBlock = StartBlock;
702     } else if (FalseBlock == nullptr) {
703       TT = TrueBlock;
704       FT = EndBlock;
705       FalseBlock = StartBlock;
706     } else {
707       TT = TrueBlock;
708       FT = FalseBlock;
709     }
710     IRBuilder<> IB(SI.getI());
711     auto *CondFr = IB.CreateFreeze(SI.getCondition(),
712                                    SI.getCondition()->getName() + ".frozen");
713 
714     SmallPtrSet<const Instruction *, 2> INS;
715     for (auto SI : ASI)
716       INS.insert(SI.getI());
717 
718     // Use reverse iterator because later select may use the value of the
719     // earlier select, and we need to propagate value through earlier select
720     // to get the PHI operand.
721     for (auto It = ASI.rbegin(); It != ASI.rend(); ++It) {
722       SelectLike SI = *It;
723       // The select itself is replaced with a PHI Node.
724       PHINode *PN = PHINode::Create(SI.getType(), 2, "");
725       PN->insertBefore(EndBlock->begin());
726       PN->takeName(SI.getI());
727       PN->addIncoming(getTrueOrFalseValue(SI, true, INS, IB), TrueBlock);
728       PN->addIncoming(getTrueOrFalseValue(SI, false, INS, IB), FalseBlock);
729       PN->setDebugLoc(SI.getI()->getDebugLoc());
730       SI.getI()->replaceAllUsesWith(PN);
731       INS.erase(SI.getI());
732       ++NumSelectsConverted;
733     }
734     IB.CreateCondBr(CondFr, TT, FT, SI.getI());
735 
736     // Remove the old select instructions, now that they are not longer used.
737     for (auto SI : ASI)
738       SI.getI()->eraseFromParent();
739   }
740 }
741 
742 void SelectOptimizeImpl::collectSelectGroups(BasicBlock &BB,
743                                              SelectGroups &SIGroups) {
744   BasicBlock::iterator BBIt = BB.begin();
745   while (BBIt != BB.end()) {
746     Instruction *I = &*BBIt++;
747     if (SelectLike SI = SelectLike::match(I)) {
748       if (!TTI->shouldTreatInstructionLikeSelect(I))
749         continue;
750 
751       SelectGroup SIGroup;
752       SIGroup.push_back(SI);
753       while (BBIt != BB.end()) {
754         Instruction *NI = &*BBIt;
755         // Debug/pseudo instructions should be skipped and not prevent the
756         // formation of a select group.
757         if (NI->isDebugOrPseudoInst()) {
758           ++BBIt;
759           continue;
760         }
761         // We only allow selects in the same group, not other select-like
762         // instructions.
763         if (!isa<SelectInst>(NI))
764           break;
765 
766         SelectLike NSI = SelectLike::match(NI);
767         if (NSI && SI.getCondition() == NSI.getCondition()) {
768           SIGroup.push_back(NSI);
769         } else
770           break;
771         ++BBIt;
772       }
773 
774       // If the select type is not supported, no point optimizing it.
775       // Instruction selection will take care of it.
776       if (!isSelectKindSupported(SI))
777         continue;
778 
779       SIGroups.push_back(SIGroup);
780     }
781   }
782 }
783 
784 void SelectOptimizeImpl::findProfitableSIGroupsBase(
785     SelectGroups &SIGroups, SelectGroups &ProfSIGroups) {
786   for (SelectGroup &ASI : SIGroups) {
787     ++NumSelectOptAnalyzed;
788     if (isConvertToBranchProfitableBase(ASI))
789       ProfSIGroups.push_back(ASI);
790   }
791 }
792 
793 static void EmitAndPrintRemark(OptimizationRemarkEmitter *ORE,
794                                DiagnosticInfoOptimizationBase &Rem) {
795   LLVM_DEBUG(dbgs() << Rem.getMsg() << "\n");
796   ORE->emit(Rem);
797 }
798 
799 void SelectOptimizeImpl::findProfitableSIGroupsInnerLoops(
800     const Loop *L, SelectGroups &SIGroups, SelectGroups &ProfSIGroups) {
801   NumSelectOptAnalyzed += SIGroups.size();
802   // For each select group in an inner-most loop,
803   // a branch is more preferable than a select/conditional-move if:
804   // i) conversion to branches for all the select groups of the loop satisfies
805   //    loop-level heuristics including reducing the loop's critical path by
806   //    some threshold (see SelectOptimizeImpl::checkLoopHeuristics); and
807   // ii) the total cost of the select group is cheaper with a branch compared
808   //     to its predicated version. The cost is in terms of latency and the cost
809   //     of a select group is the cost of its most expensive select instruction
810   //     (assuming infinite resources and thus fully leveraging available ILP).
811 
812   DenseMap<const Instruction *, CostInfo> InstCostMap;
813   CostInfo LoopCost[2] = {{Scaled64::getZero(), Scaled64::getZero()},
814                           {Scaled64::getZero(), Scaled64::getZero()}};
815   if (!computeLoopCosts(L, SIGroups, InstCostMap, LoopCost) ||
816       !checkLoopHeuristics(L, LoopCost)) {
817     return;
818   }
819 
820   for (SelectGroup &ASI : SIGroups) {
821     // Assuming infinite resources, the cost of a group of instructions is the
822     // cost of the most expensive instruction of the group.
823     Scaled64 SelectCost = Scaled64::getZero(), BranchCost = Scaled64::getZero();
824     for (SelectLike SI : ASI) {
825       SelectCost = std::max(SelectCost, InstCostMap[SI.getI()].PredCost);
826       BranchCost = std::max(BranchCost, InstCostMap[SI.getI()].NonPredCost);
827     }
828     if (BranchCost < SelectCost) {
829       OptimizationRemark OR(DEBUG_TYPE, "SelectOpti", ASI.front().getI());
830       OR << "Profitable to convert to branch (loop analysis). BranchCost="
831          << BranchCost.toString() << ", SelectCost=" << SelectCost.toString()
832          << ". ";
833       EmitAndPrintRemark(ORE, OR);
834       ++NumSelectConvertedLoop;
835       ProfSIGroups.push_back(ASI);
836     } else {
837       OptimizationRemarkMissed ORmiss(DEBUG_TYPE, "SelectOpti",
838                                       ASI.front().getI());
839       ORmiss << "Select is more profitable (loop analysis). BranchCost="
840              << BranchCost.toString()
841              << ", SelectCost=" << SelectCost.toString() << ". ";
842       EmitAndPrintRemark(ORE, ORmiss);
843     }
844   }
845 }
846 
847 bool SelectOptimizeImpl::isConvertToBranchProfitableBase(
848     const SelectGroup &ASI) {
849   SelectLike SI = ASI.front();
850   LLVM_DEBUG(dbgs() << "Analyzing select group containing " << SI.getI()
851                     << "\n");
852   OptimizationRemark OR(DEBUG_TYPE, "SelectOpti", SI.getI());
853   OptimizationRemarkMissed ORmiss(DEBUG_TYPE, "SelectOpti", SI.getI());
854 
855   // Skip cold basic blocks. Better to optimize for size for cold blocks.
856   if (PSI->isColdBlock(SI.getI()->getParent(), BFI)) {
857     ++NumSelectColdBB;
858     ORmiss << "Not converted to branch because of cold basic block. ";
859     EmitAndPrintRemark(ORE, ORmiss);
860     return false;
861   }
862 
863   // If unpredictable, branch form is less profitable.
864   if (SI.getI()->getMetadata(LLVMContext::MD_unpredictable)) {
865     ++NumSelectUnPred;
866     ORmiss << "Not converted to branch because of unpredictable branch. ";
867     EmitAndPrintRemark(ORE, ORmiss);
868     return false;
869   }
870 
871   // If highly predictable, branch form is more profitable, unless a
872   // predictable select is inexpensive in the target architecture.
873   if (isSelectHighlyPredictable(SI) && TLI->isPredictableSelectExpensive()) {
874     ++NumSelectConvertedHighPred;
875     OR << "Converted to branch because of highly predictable branch. ";
876     EmitAndPrintRemark(ORE, OR);
877     return true;
878   }
879 
880   // Look for expensive instructions in the cold operand's (if any) dependence
881   // slice of any of the selects in the group.
882   if (hasExpensiveColdOperand(ASI)) {
883     ++NumSelectConvertedExpColdOperand;
884     OR << "Converted to branch because of expensive cold operand.";
885     EmitAndPrintRemark(ORE, OR);
886     return true;
887   }
888 
889   ORmiss << "Not profitable to convert to branch (base heuristic).";
890   EmitAndPrintRemark(ORE, ORmiss);
891   return false;
892 }
893 
894 static InstructionCost divideNearest(InstructionCost Numerator,
895                                      uint64_t Denominator) {
896   return (Numerator + (Denominator / 2)) / Denominator;
897 }
898 
899 static bool extractBranchWeights(const SelectOptimizeImpl::SelectLike SI,
900                                  uint64_t &TrueVal, uint64_t &FalseVal) {
901   if (isa<SelectInst>(SI.getI()))
902     return extractBranchWeights(*SI.getI(), TrueVal, FalseVal);
903   return false;
904 }
905 
906 bool SelectOptimizeImpl::hasExpensiveColdOperand(const SelectGroup &ASI) {
907   bool ColdOperand = false;
908   uint64_t TrueWeight, FalseWeight, TotalWeight;
909   if (extractBranchWeights(ASI.front(), TrueWeight, FalseWeight)) {
910     uint64_t MinWeight = std::min(TrueWeight, FalseWeight);
911     TotalWeight = TrueWeight + FalseWeight;
912     // Is there a path with frequency <ColdOperandThreshold% (default:20%) ?
913     ColdOperand = TotalWeight * ColdOperandThreshold > 100 * MinWeight;
914   } else if (PSI->hasProfileSummary()) {
915     OptimizationRemarkMissed ORmiss(DEBUG_TYPE, "SelectOpti",
916                                     ASI.front().getI());
917     ORmiss << "Profile data available but missing branch-weights metadata for "
918               "select instruction. ";
919     EmitAndPrintRemark(ORE, ORmiss);
920   }
921   if (!ColdOperand)
922     return false;
923   // Check if the cold path's dependence slice is expensive for any of the
924   // selects of the group.
925   for (SelectLike SI : ASI) {
926     Instruction *ColdI = nullptr;
927     uint64_t HotWeight;
928     if (TrueWeight < FalseWeight) {
929       ColdI = dyn_cast_or_null<Instruction>(SI.getTrueValue());
930       HotWeight = FalseWeight;
931     } else {
932       ColdI = dyn_cast_or_null<Instruction>(SI.getFalseValue());
933       HotWeight = TrueWeight;
934     }
935     if (ColdI) {
936       std::stack<Instruction *> ColdSlice;
937       getExclBackwardsSlice(ColdI, ColdSlice, SI.getI());
938       InstructionCost SliceCost = 0;
939       while (!ColdSlice.empty()) {
940         SliceCost += TTI->getInstructionCost(ColdSlice.top(),
941                                              TargetTransformInfo::TCK_Latency);
942         ColdSlice.pop();
943       }
944       // The colder the cold value operand of the select is the more expensive
945       // the cmov becomes for computing the cold value operand every time. Thus,
946       // the colder the cold operand is the more its cost counts.
947       // Get nearest integer cost adjusted for coldness.
948       InstructionCost AdjSliceCost =
949           divideNearest(SliceCost * HotWeight, TotalWeight);
950       if (AdjSliceCost >=
951           ColdOperandMaxCostMultiplier * TargetTransformInfo::TCC_Expensive)
952         return true;
953     }
954   }
955   return false;
956 }
957 
958 // Check if it is safe to move LoadI next to the SI.
959 // Conservatively assume it is safe only if there is no instruction
960 // modifying memory in-between the load and the select instruction.
961 static bool isSafeToSinkLoad(Instruction *LoadI, Instruction *SI) {
962   // Assume loads from different basic blocks are unsafe to move.
963   if (LoadI->getParent() != SI->getParent())
964     return false;
965   auto It = LoadI->getIterator();
966   while (&*It != SI) {
967     if (It->mayWriteToMemory())
968       return false;
969     It++;
970   }
971   return true;
972 }
973 
974 // For a given source instruction, collect its backwards dependence slice
975 // consisting of instructions exclusively computed for the purpose of producing
976 // the operands of the source instruction. As an approximation
977 // (sufficiently-accurate in practice), we populate this set with the
978 // instructions of the backwards dependence slice that only have one-use and
979 // form an one-use chain that leads to the source instruction.
980 void SelectOptimizeImpl::getExclBackwardsSlice(Instruction *I,
981                                                std::stack<Instruction *> &Slice,
982                                                Instruction *SI,
983                                                bool ForSinking) {
984   SmallPtrSet<Instruction *, 2> Visited;
985   std::queue<Instruction *> Worklist;
986   Worklist.push(I);
987   while (!Worklist.empty()) {
988     Instruction *II = Worklist.front();
989     Worklist.pop();
990 
991     // Avoid cycles.
992     if (!Visited.insert(II).second)
993       continue;
994 
995     if (!II->hasOneUse())
996       continue;
997 
998     // Cannot soundly sink instructions with side-effects.
999     // Terminator or phi instructions cannot be sunk.
1000     // Avoid sinking other select instructions (should be handled separetely).
1001     if (ForSinking && (II->isTerminator() || II->mayHaveSideEffects() ||
1002                        isa<SelectInst>(II) || isa<PHINode>(II)))
1003       continue;
1004 
1005     // Avoid sinking loads in order not to skip state-modifying instructions,
1006     // that may alias with the loaded address.
1007     // Only allow sinking of loads within the same basic block that are
1008     // conservatively proven to be safe.
1009     if (ForSinking && II->mayReadFromMemory() && !isSafeToSinkLoad(II, SI))
1010       continue;
1011 
1012     // Avoid considering instructions with less frequency than the source
1013     // instruction (i.e., avoid colder code regions of the dependence slice).
1014     if (BFI->getBlockFreq(II->getParent()) < BFI->getBlockFreq(I->getParent()))
1015       continue;
1016 
1017     // Eligible one-use instruction added to the dependence slice.
1018     Slice.push(II);
1019 
1020     // Explore all the operands of the current instruction to expand the slice.
1021     for (unsigned k = 0; k < II->getNumOperands(); ++k)
1022       if (auto *OpI = dyn_cast<Instruction>(II->getOperand(k)))
1023         Worklist.push(OpI);
1024   }
1025 }
1026 
1027 bool SelectOptimizeImpl::isSelectHighlyPredictable(const SelectLike SI) {
1028   uint64_t TrueWeight, FalseWeight;
1029   if (extractBranchWeights(SI, TrueWeight, FalseWeight)) {
1030     uint64_t Max = std::max(TrueWeight, FalseWeight);
1031     uint64_t Sum = TrueWeight + FalseWeight;
1032     if (Sum != 0) {
1033       auto Probability = BranchProbability::getBranchProbability(Max, Sum);
1034       if (Probability > TTI->getPredictableBranchThreshold())
1035         return true;
1036     }
1037   }
1038   return false;
1039 }
1040 
1041 bool SelectOptimizeImpl::checkLoopHeuristics(const Loop *L,
1042                                              const CostInfo LoopCost[2]) {
1043   // Loop-level checks to determine if a non-predicated version (with branches)
1044   // of the loop is more profitable than its predicated version.
1045 
1046   if (DisableLoopLevelHeuristics)
1047     return true;
1048 
1049   OptimizationRemarkMissed ORmissL(DEBUG_TYPE, "SelectOpti",
1050                                    L->getHeader()->getFirstNonPHI());
1051 
1052   if (LoopCost[0].NonPredCost > LoopCost[0].PredCost ||
1053       LoopCost[1].NonPredCost >= LoopCost[1].PredCost) {
1054     ORmissL << "No select conversion in the loop due to no reduction of loop's "
1055                "critical path. ";
1056     EmitAndPrintRemark(ORE, ORmissL);
1057     return false;
1058   }
1059 
1060   Scaled64 Gain[2] = {LoopCost[0].PredCost - LoopCost[0].NonPredCost,
1061                       LoopCost[1].PredCost - LoopCost[1].NonPredCost};
1062 
1063   // Profitably converting to branches need to reduce the loop's critical path
1064   // by at least some threshold (absolute gain of GainCycleThreshold cycles and
1065   // relative gain of 12.5%).
1066   if (Gain[1] < Scaled64::get(GainCycleThreshold) ||
1067       Gain[1] * Scaled64::get(GainRelativeThreshold) < LoopCost[1].PredCost) {
1068     Scaled64 RelativeGain = Scaled64::get(100) * Gain[1] / LoopCost[1].PredCost;
1069     ORmissL << "No select conversion in the loop due to small reduction of "
1070                "loop's critical path. Gain="
1071             << Gain[1].toString()
1072             << ", RelativeGain=" << RelativeGain.toString() << "%. ";
1073     EmitAndPrintRemark(ORE, ORmissL);
1074     return false;
1075   }
1076 
1077   // If the loop's critical path involves loop-carried dependences, the gradient
1078   // of the gain needs to be at least GainGradientThreshold% (defaults to 25%).
1079   // This check ensures that the latency reduction for the loop's critical path
1080   // keeps decreasing with sufficient rate beyond the two analyzed loop
1081   // iterations.
1082   if (Gain[1] > Gain[0]) {
1083     Scaled64 GradientGain = Scaled64::get(100) * (Gain[1] - Gain[0]) /
1084                             (LoopCost[1].PredCost - LoopCost[0].PredCost);
1085     if (GradientGain < Scaled64::get(GainGradientThreshold)) {
1086       ORmissL << "No select conversion in the loop due to small gradient gain. "
1087                  "GradientGain="
1088               << GradientGain.toString() << "%. ";
1089       EmitAndPrintRemark(ORE, ORmissL);
1090       return false;
1091     }
1092   }
1093   // If the gain decreases it is not profitable to convert.
1094   else if (Gain[1] < Gain[0]) {
1095     ORmissL
1096         << "No select conversion in the loop due to negative gradient gain. ";
1097     EmitAndPrintRemark(ORE, ORmissL);
1098     return false;
1099   }
1100 
1101   // Non-predicated version of the loop is more profitable than its
1102   // predicated version.
1103   return true;
1104 }
1105 
1106 // Computes instruction and loop-critical-path costs for both the predicated
1107 // and non-predicated version of the given loop.
1108 // Returns false if unable to compute these costs due to invalid cost of loop
1109 // instruction(s).
1110 bool SelectOptimizeImpl::computeLoopCosts(
1111     const Loop *L, const SelectGroups &SIGroups,
1112     DenseMap<const Instruction *, CostInfo> &InstCostMap, CostInfo *LoopCost) {
1113   LLVM_DEBUG(dbgs() << "Calculating Latency / IPredCost / INonPredCost of loop "
1114                     << L->getHeader()->getName() << "\n");
1115   const auto &SImap = getSImap(SIGroups);
1116   // Compute instruction and loop-critical-path costs across two iterations for
1117   // both predicated and non-predicated version.
1118   const unsigned Iterations = 2;
1119   for (unsigned Iter = 0; Iter < Iterations; ++Iter) {
1120     // Cost of the loop's critical path.
1121     CostInfo &MaxCost = LoopCost[Iter];
1122     for (BasicBlock *BB : L->getBlocks()) {
1123       for (const Instruction &I : *BB) {
1124         if (I.isDebugOrPseudoInst())
1125           continue;
1126         // Compute the predicated and non-predicated cost of the instruction.
1127         Scaled64 IPredCost = Scaled64::getZero(),
1128                  INonPredCost = Scaled64::getZero();
1129 
1130         // Assume infinite resources that allow to fully exploit the available
1131         // instruction-level parallelism.
1132         // InstCost = InstLatency + max(Op1Cost, Op2Cost, … OpNCost)
1133         for (const Use &U : I.operands()) {
1134           auto UI = dyn_cast<Instruction>(U.get());
1135           if (!UI)
1136             continue;
1137           if (InstCostMap.count(UI)) {
1138             IPredCost = std::max(IPredCost, InstCostMap[UI].PredCost);
1139             INonPredCost = std::max(INonPredCost, InstCostMap[UI].NonPredCost);
1140           }
1141         }
1142         auto ILatency = computeInstCost(&I);
1143         if (!ILatency) {
1144           OptimizationRemarkMissed ORmissL(DEBUG_TYPE, "SelectOpti", &I);
1145           ORmissL << "Invalid instruction cost preventing analysis and "
1146                      "optimization of the inner-most loop containing this "
1147                      "instruction. ";
1148           EmitAndPrintRemark(ORE, ORmissL);
1149           return false;
1150         }
1151         IPredCost += Scaled64::get(*ILatency);
1152         INonPredCost += Scaled64::get(*ILatency);
1153 
1154         // For a select that can be converted to branch,
1155         // compute its cost as a branch (non-predicated cost).
1156         //
1157         // BranchCost = PredictedPathCost + MispredictCost
1158         // PredictedPathCost = TrueOpCost * TrueProb + FalseOpCost * FalseProb
1159         // MispredictCost = max(MispredictPenalty, CondCost) * MispredictRate
1160         if (SImap.contains(&I)) {
1161           auto SI = SImap.at(&I);
1162           Scaled64 TrueOpCost = SI.getTrueOpCost(InstCostMap, TTI);
1163           Scaled64 FalseOpCost = SI.getFalseOpCost(InstCostMap, TTI);
1164           Scaled64 PredictedPathCost =
1165               getPredictedPathCost(TrueOpCost, FalseOpCost, SI);
1166 
1167           Scaled64 CondCost = Scaled64::getZero();
1168           if (auto *CI = dyn_cast<Instruction>(SI.getCondition()))
1169             if (InstCostMap.count(CI))
1170               CondCost = InstCostMap[CI].NonPredCost;
1171           Scaled64 MispredictCost = getMispredictionCost(SI, CondCost);
1172 
1173           INonPredCost = PredictedPathCost + MispredictCost;
1174         }
1175         LLVM_DEBUG(dbgs() << " " << ILatency << "/" << IPredCost << "/"
1176                           << INonPredCost << " for " << I << "\n");
1177 
1178         InstCostMap[&I] = {IPredCost, INonPredCost};
1179         MaxCost.PredCost = std::max(MaxCost.PredCost, IPredCost);
1180         MaxCost.NonPredCost = std::max(MaxCost.NonPredCost, INonPredCost);
1181       }
1182     }
1183     LLVM_DEBUG(dbgs() << "Iteration " << Iter + 1
1184                       << " MaxCost = " << MaxCost.PredCost << " "
1185                       << MaxCost.NonPredCost << "\n");
1186   }
1187   return true;
1188 }
1189 
1190 SmallDenseMap<const Instruction *, SelectOptimizeImpl::SelectLike, 2>
1191 SelectOptimizeImpl::getSImap(const SelectGroups &SIGroups) {
1192   SmallDenseMap<const Instruction *, SelectLike, 2> SImap;
1193   for (const SelectGroup &ASI : SIGroups)
1194     for (SelectLike SI : ASI)
1195       SImap.try_emplace(SI.getI(), SI);
1196   return SImap;
1197 }
1198 
1199 std::optional<uint64_t>
1200 SelectOptimizeImpl::computeInstCost(const Instruction *I) {
1201   InstructionCost ICost =
1202       TTI->getInstructionCost(I, TargetTransformInfo::TCK_Latency);
1203   if (auto OC = ICost.getValue())
1204     return std::optional<uint64_t>(*OC);
1205   return std::nullopt;
1206 }
1207 
1208 ScaledNumber<uint64_t>
1209 SelectOptimizeImpl::getMispredictionCost(const SelectLike SI,
1210                                          const Scaled64 CondCost) {
1211   uint64_t MispredictPenalty = TSchedModel.getMCSchedModel()->MispredictPenalty;
1212 
1213   // Account for the default misprediction rate when using a branch
1214   // (conservatively set to 25% by default).
1215   uint64_t MispredictRate = MispredictDefaultRate;
1216   // If the select condition is obviously predictable, then the misprediction
1217   // rate is zero.
1218   if (isSelectHighlyPredictable(SI))
1219     MispredictRate = 0;
1220 
1221   // CondCost is included to account for cases where the computation of the
1222   // condition is part of a long dependence chain (potentially loop-carried)
1223   // that would delay detection of a misprediction and increase its cost.
1224   Scaled64 MispredictCost =
1225       std::max(Scaled64::get(MispredictPenalty), CondCost) *
1226       Scaled64::get(MispredictRate);
1227   MispredictCost /= Scaled64::get(100);
1228 
1229   return MispredictCost;
1230 }
1231 
1232 // Returns the cost of a branch when the prediction is correct.
1233 // TrueCost * TrueProbability + FalseCost * FalseProbability.
1234 ScaledNumber<uint64_t>
1235 SelectOptimizeImpl::getPredictedPathCost(Scaled64 TrueCost, Scaled64 FalseCost,
1236                                          const SelectLike SI) {
1237   Scaled64 PredPathCost;
1238   uint64_t TrueWeight, FalseWeight;
1239   if (extractBranchWeights(SI, TrueWeight, FalseWeight)) {
1240     uint64_t SumWeight = TrueWeight + FalseWeight;
1241     if (SumWeight != 0) {
1242       PredPathCost = TrueCost * Scaled64::get(TrueWeight) +
1243                      FalseCost * Scaled64::get(FalseWeight);
1244       PredPathCost /= Scaled64::get(SumWeight);
1245       return PredPathCost;
1246     }
1247   }
1248   // Without branch weight metadata, we assume 75% for the one path and 25% for
1249   // the other, and pick the result with the biggest cost.
1250   PredPathCost = std::max(TrueCost * Scaled64::get(3) + FalseCost,
1251                           FalseCost * Scaled64::get(3) + TrueCost);
1252   PredPathCost /= Scaled64::get(4);
1253   return PredPathCost;
1254 }
1255 
1256 bool SelectOptimizeImpl::isSelectKindSupported(const SelectLike SI) {
1257   bool VectorCond = !SI.getCondition()->getType()->isIntegerTy(1);
1258   if (VectorCond)
1259     return false;
1260   TargetLowering::SelectSupportKind SelectKind;
1261   if (SI.getType()->isVectorTy())
1262     SelectKind = TargetLowering::ScalarCondVectorVal;
1263   else
1264     SelectKind = TargetLowering::ScalarValSelect;
1265   return TLI->isSelectSupported(SelectKind);
1266 }
1267