xref: /freebsd/contrib/llvm-project/llvm/lib/CodeGen/SelectOptimize.cpp (revision 7fdf597e96a02165cfe22ff357b857d5fa15ed8a)
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     /// The select (/or) instruction.
134     Instruction *I;
135     /// Whether this select is inverted, "not(cond), FalseVal, TrueVal", as
136     /// opposed to the original condition.
137     bool Inverted = false;
138 
139   public:
140     /// Match a select or select-like instruction, returning a SelectLike.
141     static SelectLike match(Instruction *I) {
142       // Select instruction are what we are usually looking for.
143       if (isa<SelectInst>(I))
144         return SelectLike(I);
145 
146       // An Or(zext(i1 X), Y) can also be treated like a select, with condition
147       // C and values Y|1 and Y.
148       Value *X;
149       if (PatternMatch::match(
150               I, m_c_Or(m_OneUse(m_ZExt(m_Value(X))), m_Value())) &&
151           X->getType()->isIntegerTy(1))
152         return SelectLike(I);
153 
154       return SelectLike(nullptr);
155     }
156 
157     bool isValid() { return I; }
158     operator bool() { return isValid(); }
159 
160     /// Invert the select by inverting the condition and switching the operands.
161     void setInverted() {
162       assert(!Inverted && "Trying to invert an inverted SelectLike");
163       assert(isa<Instruction>(getCondition()) &&
164              cast<Instruction>(getCondition())->getOpcode() ==
165                  Instruction::Xor);
166       Inverted = true;
167     }
168     bool isInverted() const { return Inverted; }
169 
170     Instruction *getI() { return I; }
171     const Instruction *getI() const { return I; }
172 
173     Type *getType() const { return I->getType(); }
174 
175     Value *getNonInvertedCondition() const {
176       if (auto *Sel = dyn_cast<SelectInst>(I))
177         return Sel->getCondition();
178       // Or(zext) case
179       if (auto *BO = dyn_cast<BinaryOperator>(I)) {
180         Value *X;
181         if (PatternMatch::match(BO->getOperand(0),
182                                 m_OneUse(m_ZExt(m_Value(X)))))
183           return X;
184         if (PatternMatch::match(BO->getOperand(1),
185                                 m_OneUse(m_ZExt(m_Value(X)))))
186           return X;
187       }
188 
189       llvm_unreachable("Unhandled case in getCondition");
190     }
191 
192     /// Return the condition for the SelectLike instruction. For example the
193     /// condition of a select or c in `or(zext(c), x)`
194     Value *getCondition() const {
195       Value *CC = getNonInvertedCondition();
196       // For inverted conditions the CC is checked when created to be a not
197       // (xor) instruction.
198       if (Inverted)
199         return cast<Instruction>(CC)->getOperand(0);
200       return CC;
201     }
202 
203     /// Return the true value for the SelectLike instruction. Note this may not
204     /// exist for all SelectLike instructions. For example, for `or(zext(c), x)`
205     /// the true value would be `or(x,1)`. As this value does not exist, nullptr
206     /// is returned.
207     Value *getTrueValue(bool HonorInverts = true) const {
208       if (Inverted && HonorInverts)
209         return getFalseValue(/*HonorInverts=*/false);
210       if (auto *Sel = dyn_cast<SelectInst>(I))
211         return Sel->getTrueValue();
212       // Or(zext) case - The true value is Or(X), so return nullptr as the value
213       // does not yet exist.
214       if (isa<BinaryOperator>(I))
215         return nullptr;
216 
217       llvm_unreachable("Unhandled case in getTrueValue");
218     }
219 
220     /// Return the false value for the SelectLike instruction. For example the
221     /// getFalseValue of a select or `x` in `or(zext(c), x)` (which is
222     /// `select(c, x|1, x)`)
223     Value *getFalseValue(bool HonorInverts = true) const {
224       if (Inverted && HonorInverts)
225         return getTrueValue(/*HonorInverts=*/false);
226       if (auto *Sel = dyn_cast<SelectInst>(I))
227         return Sel->getFalseValue();
228       // Or(zext) case - return the operand which is not the zext.
229       if (auto *BO = dyn_cast<BinaryOperator>(I)) {
230         Value *X;
231         if (PatternMatch::match(BO->getOperand(0),
232                                 m_OneUse(m_ZExt(m_Value(X)))))
233           return BO->getOperand(1);
234         if (PatternMatch::match(BO->getOperand(1),
235                                 m_OneUse(m_ZExt(m_Value(X)))))
236           return BO->getOperand(0);
237       }
238 
239       llvm_unreachable("Unhandled case in getFalseValue");
240     }
241 
242     /// Return the NonPredCost cost of the true op, given the costs in
243     /// InstCostMap. This may need to be generated for select-like instructions.
244     Scaled64 getTrueOpCost(DenseMap<const Instruction *, CostInfo> &InstCostMap,
245                            const TargetTransformInfo *TTI) {
246       if (isa<SelectInst>(I))
247         if (auto *I = dyn_cast<Instruction>(getTrueValue()))
248           return InstCostMap.contains(I) ? InstCostMap[I].NonPredCost
249                                          : Scaled64::getZero();
250 
251       // Or case - add the cost of an extra Or to the cost of the False case.
252       if (isa<BinaryOperator>(I))
253         if (auto I = dyn_cast<Instruction>(getFalseValue()))
254           if (InstCostMap.contains(I)) {
255             InstructionCost OrCost = TTI->getArithmeticInstrCost(
256                 Instruction::Or, I->getType(), TargetTransformInfo::TCK_Latency,
257                 {TargetTransformInfo::OK_AnyValue,
258                  TargetTransformInfo::OP_None},
259                 {TTI::OK_UniformConstantValue, TTI::OP_PowerOf2});
260             return InstCostMap[I].NonPredCost +
261                    Scaled64::get(*OrCost.getValue());
262           }
263 
264       return Scaled64::getZero();
265     }
266 
267     /// Return the NonPredCost cost of the false op, given the costs in
268     /// InstCostMap. This may need to be generated for select-like instructions.
269     Scaled64
270     getFalseOpCost(DenseMap<const Instruction *, CostInfo> &InstCostMap,
271                    const TargetTransformInfo *TTI) {
272       if (isa<SelectInst>(I))
273         if (auto *I = dyn_cast<Instruction>(getFalseValue()))
274           return InstCostMap.contains(I) ? InstCostMap[I].NonPredCost
275                                          : Scaled64::getZero();
276 
277       // Or case - return the cost of the false case
278       if (isa<BinaryOperator>(I))
279         if (auto I = dyn_cast<Instruction>(getFalseValue()))
280           if (InstCostMap.contains(I))
281             return InstCostMap[I].NonPredCost;
282 
283       return Scaled64::getZero();
284     }
285   };
286 
287 private:
288   // Select groups consist of consecutive select instructions with the same
289   // condition.
290   using SelectGroup = SmallVector<SelectLike, 2>;
291   using SelectGroups = SmallVector<SelectGroup, 2>;
292 
293   // Converts select instructions of a function to conditional jumps when deemed
294   // profitable. Returns true if at least one select was converted.
295   bool optimizeSelects(Function &F);
296 
297   // Heuristics for determining which select instructions can be profitably
298   // conveted to branches. Separate heuristics for selects in inner-most loops
299   // and the rest of code regions (base heuristics for non-inner-most loop
300   // regions).
301   void optimizeSelectsBase(Function &F, SelectGroups &ProfSIGroups);
302   void optimizeSelectsInnerLoops(Function &F, SelectGroups &ProfSIGroups);
303 
304   // Converts to branches the select groups that were deemed
305   // profitable-to-convert.
306   void convertProfitableSIGroups(SelectGroups &ProfSIGroups);
307 
308   // Splits selects of a given basic block into select groups.
309   void collectSelectGroups(BasicBlock &BB, SelectGroups &SIGroups);
310 
311   // Determines for which select groups it is profitable converting to branches
312   // (base and inner-most-loop heuristics).
313   void findProfitableSIGroupsBase(SelectGroups &SIGroups,
314                                   SelectGroups &ProfSIGroups);
315   void findProfitableSIGroupsInnerLoops(const Loop *L, SelectGroups &SIGroups,
316                                         SelectGroups &ProfSIGroups);
317 
318   // Determines if a select group should be converted to a branch (base
319   // heuristics).
320   bool isConvertToBranchProfitableBase(const SelectGroup &ASI);
321 
322   // Returns true if there are expensive instructions in the cold value
323   // operand's (if any) dependence slice of any of the selects of the given
324   // group.
325   bool hasExpensiveColdOperand(const SelectGroup &ASI);
326 
327   // For a given source instruction, collect its backwards dependence slice
328   // consisting of instructions exclusively computed for producing the operands
329   // of the source instruction.
330   void getExclBackwardsSlice(Instruction *I, std::stack<Instruction *> &Slice,
331                              Instruction *SI, bool ForSinking = false);
332 
333   // Returns true if the condition of the select is highly predictable.
334   bool isSelectHighlyPredictable(const SelectLike SI);
335 
336   // Loop-level checks to determine if a non-predicated version (with branches)
337   // of the given loop is more profitable than its predicated version.
338   bool checkLoopHeuristics(const Loop *L, const CostInfo LoopDepth[2]);
339 
340   // Computes instruction and loop-critical-path costs for both the predicated
341   // and non-predicated version of the given loop.
342   bool computeLoopCosts(const Loop *L, const SelectGroups &SIGroups,
343                         DenseMap<const Instruction *, CostInfo> &InstCostMap,
344                         CostInfo *LoopCost);
345 
346   // Returns a set of all the select instructions in the given select groups.
347   SmallDenseMap<const Instruction *, SelectLike, 2>
348   getSImap(const SelectGroups &SIGroups);
349 
350   // Returns the latency cost of a given instruction.
351   std::optional<uint64_t> computeInstCost(const Instruction *I);
352 
353   // Returns the misprediction cost of a given select when converted to branch.
354   Scaled64 getMispredictionCost(const SelectLike SI, const Scaled64 CondCost);
355 
356   // Returns the cost of a branch when the prediction is correct.
357   Scaled64 getPredictedPathCost(Scaled64 TrueCost, Scaled64 FalseCost,
358                                 const SelectLike SI);
359 
360   // Returns true if the target architecture supports lowering a given select.
361   bool isSelectKindSupported(const SelectLike SI);
362 };
363 
364 class SelectOptimize : public FunctionPass {
365   SelectOptimizeImpl Impl;
366 
367 public:
368   static char ID;
369 
370   SelectOptimize() : FunctionPass(ID) {
371     initializeSelectOptimizePass(*PassRegistry::getPassRegistry());
372   }
373 
374   bool runOnFunction(Function &F) override {
375     return Impl.runOnFunction(F, *this);
376   }
377 
378   void getAnalysisUsage(AnalysisUsage &AU) const override {
379     AU.addRequired<ProfileSummaryInfoWrapperPass>();
380     AU.addRequired<TargetPassConfig>();
381     AU.addRequired<TargetTransformInfoWrapperPass>();
382     AU.addRequired<LoopInfoWrapperPass>();
383     AU.addRequired<BlockFrequencyInfoWrapperPass>();
384     AU.addRequired<OptimizationRemarkEmitterWrapperPass>();
385   }
386 };
387 
388 } // namespace
389 
390 PreservedAnalyses SelectOptimizePass::run(Function &F,
391                                           FunctionAnalysisManager &FAM) {
392   SelectOptimizeImpl Impl(TM);
393   return Impl.run(F, FAM);
394 }
395 
396 char SelectOptimize::ID = 0;
397 
398 INITIALIZE_PASS_BEGIN(SelectOptimize, DEBUG_TYPE, "Optimize selects", false,
399                       false)
400 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
401 INITIALIZE_PASS_DEPENDENCY(ProfileSummaryInfoWrapperPass)
402 INITIALIZE_PASS_DEPENDENCY(TargetPassConfig)
403 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
404 INITIALIZE_PASS_DEPENDENCY(BlockFrequencyInfoWrapperPass)
405 INITIALIZE_PASS_DEPENDENCY(OptimizationRemarkEmitterWrapperPass)
406 INITIALIZE_PASS_END(SelectOptimize, DEBUG_TYPE, "Optimize selects", false,
407                     false)
408 
409 FunctionPass *llvm::createSelectOptimizePass() { return new SelectOptimize(); }
410 
411 PreservedAnalyses SelectOptimizeImpl::run(Function &F,
412                                           FunctionAnalysisManager &FAM) {
413   TSI = TM->getSubtargetImpl(F);
414   TLI = TSI->getTargetLowering();
415 
416   // If none of the select types are supported then skip this pass.
417   // This is an optimization pass. Legality issues will be handled by
418   // instruction selection.
419   if (!TLI->isSelectSupported(TargetLowering::ScalarValSelect) &&
420       !TLI->isSelectSupported(TargetLowering::ScalarCondVectorVal) &&
421       !TLI->isSelectSupported(TargetLowering::VectorMaskSelect))
422     return PreservedAnalyses::all();
423 
424   TTI = &FAM.getResult<TargetIRAnalysis>(F);
425   if (!TTI->enableSelectOptimize())
426     return PreservedAnalyses::all();
427 
428   PSI = FAM.getResult<ModuleAnalysisManagerFunctionProxy>(F)
429             .getCachedResult<ProfileSummaryAnalysis>(*F.getParent());
430   assert(PSI && "This pass requires module analysis pass `profile-summary`!");
431   BFI = &FAM.getResult<BlockFrequencyAnalysis>(F);
432 
433   // When optimizing for size, selects are preferable over branches.
434   if (F.hasOptSize() || llvm::shouldOptimizeForSize(&F, PSI, BFI))
435     return PreservedAnalyses::all();
436 
437   LI = &FAM.getResult<LoopAnalysis>(F);
438   ORE = &FAM.getResult<OptimizationRemarkEmitterAnalysis>(F);
439   TSchedModel.init(TSI);
440 
441   bool Changed = optimizeSelects(F);
442   return Changed ? PreservedAnalyses::none() : PreservedAnalyses::all();
443 }
444 
445 bool SelectOptimizeImpl::runOnFunction(Function &F, Pass &P) {
446   TM = &P.getAnalysis<TargetPassConfig>().getTM<TargetMachine>();
447   TSI = TM->getSubtargetImpl(F);
448   TLI = TSI->getTargetLowering();
449 
450   // If none of the select types are supported then skip this pass.
451   // This is an optimization pass. Legality issues will be handled by
452   // instruction selection.
453   if (!TLI->isSelectSupported(TargetLowering::ScalarValSelect) &&
454       !TLI->isSelectSupported(TargetLowering::ScalarCondVectorVal) &&
455       !TLI->isSelectSupported(TargetLowering::VectorMaskSelect))
456     return false;
457 
458   TTI = &P.getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
459 
460   if (!TTI->enableSelectOptimize())
461     return false;
462 
463   LI = &P.getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
464   BFI = &P.getAnalysis<BlockFrequencyInfoWrapperPass>().getBFI();
465   PSI = &P.getAnalysis<ProfileSummaryInfoWrapperPass>().getPSI();
466   ORE = &P.getAnalysis<OptimizationRemarkEmitterWrapperPass>().getORE();
467   TSchedModel.init(TSI);
468 
469   // When optimizing for size, selects are preferable over branches.
470   if (F.hasOptSize() || llvm::shouldOptimizeForSize(&F, PSI, BFI))
471     return false;
472 
473   return optimizeSelects(F);
474 }
475 
476 bool SelectOptimizeImpl::optimizeSelects(Function &F) {
477   // Determine for which select groups it is profitable converting to branches.
478   SelectGroups ProfSIGroups;
479   // Base heuristics apply only to non-loops and outer loops.
480   optimizeSelectsBase(F, ProfSIGroups);
481   // Separate heuristics for inner-most loops.
482   optimizeSelectsInnerLoops(F, ProfSIGroups);
483 
484   // Convert to branches the select groups that were deemed
485   // profitable-to-convert.
486   convertProfitableSIGroups(ProfSIGroups);
487 
488   // Code modified if at least one select group was converted.
489   return !ProfSIGroups.empty();
490 }
491 
492 void SelectOptimizeImpl::optimizeSelectsBase(Function &F,
493                                              SelectGroups &ProfSIGroups) {
494   // Collect all the select groups.
495   SelectGroups SIGroups;
496   for (BasicBlock &BB : F) {
497     // Base heuristics apply only to non-loops and outer loops.
498     Loop *L = LI->getLoopFor(&BB);
499     if (L && L->isInnermost())
500       continue;
501     collectSelectGroups(BB, SIGroups);
502   }
503 
504   // Determine for which select groups it is profitable converting to branches.
505   findProfitableSIGroupsBase(SIGroups, ProfSIGroups);
506 }
507 
508 void SelectOptimizeImpl::optimizeSelectsInnerLoops(Function &F,
509                                                    SelectGroups &ProfSIGroups) {
510   SmallVector<Loop *, 4> Loops(LI->begin(), LI->end());
511   // Need to check size on each iteration as we accumulate child loops.
512   for (unsigned long i = 0; i < Loops.size(); ++i)
513     for (Loop *ChildL : Loops[i]->getSubLoops())
514       Loops.push_back(ChildL);
515 
516   for (Loop *L : Loops) {
517     if (!L->isInnermost())
518       continue;
519 
520     SelectGroups SIGroups;
521     for (BasicBlock *BB : L->getBlocks())
522       collectSelectGroups(*BB, SIGroups);
523 
524     findProfitableSIGroupsInnerLoops(L, SIGroups, ProfSIGroups);
525   }
526 }
527 
528 /// If \p isTrue is true, return the true value of \p SI, otherwise return
529 /// false value of \p SI. If the true/false value of \p SI is defined by any
530 /// select instructions in \p Selects, look through the defining select
531 /// instruction until the true/false value is not defined in \p Selects.
532 static Value *
533 getTrueOrFalseValue(SelectOptimizeImpl::SelectLike SI, bool isTrue,
534                     const SmallPtrSet<const Instruction *, 2> &Selects,
535                     IRBuilder<> &IB) {
536   Value *V = nullptr;
537   for (SelectInst *DefSI = dyn_cast<SelectInst>(SI.getI());
538        DefSI != nullptr && Selects.count(DefSI);
539        DefSI = dyn_cast<SelectInst>(V)) {
540     if (DefSI->getCondition() == SI.getCondition())
541       V = (isTrue ? DefSI->getTrueValue() : DefSI->getFalseValue());
542     else // Handle inverted SI
543       V = (!isTrue ? DefSI->getTrueValue() : DefSI->getFalseValue());
544   }
545 
546   if (isa<BinaryOperator>(SI.getI())) {
547     assert(SI.getI()->getOpcode() == Instruction::Or &&
548            "Only currently handling Or instructions.");
549     V = SI.getFalseValue();
550     if (isTrue)
551       V = IB.CreateOr(V, ConstantInt::get(V->getType(), 1));
552   }
553 
554   assert(V && "Failed to get select true/false value");
555   return V;
556 }
557 
558 void SelectOptimizeImpl::convertProfitableSIGroups(SelectGroups &ProfSIGroups) {
559   for (SelectGroup &ASI : ProfSIGroups) {
560     // The code transformation here is a modified version of the sinking
561     // transformation in CodeGenPrepare::optimizeSelectInst with a more
562     // aggressive strategy of which instructions to sink.
563     //
564     // TODO: eliminate the redundancy of logic transforming selects to branches
565     // by removing CodeGenPrepare::optimizeSelectInst and optimizing here
566     // selects for all cases (with and without profile information).
567 
568     // Transform a sequence like this:
569     //    start:
570     //       %cmp = cmp uge i32 %a, %b
571     //       %sel = select i1 %cmp, i32 %c, i32 %d
572     //
573     // Into:
574     //    start:
575     //       %cmp = cmp uge i32 %a, %b
576     //       %cmp.frozen = freeze %cmp
577     //       br i1 %cmp.frozen, label %select.true, label %select.false
578     //    select.true:
579     //       br label %select.end
580     //    select.false:
581     //       br label %select.end
582     //    select.end:
583     //       %sel = phi i32 [ %c, %select.true ], [ %d, %select.false ]
584     //
585     // %cmp should be frozen, otherwise it may introduce undefined behavior.
586     // In addition, we may sink instructions that produce %c or %d into the
587     // destination(s) of the new branch.
588     // If the true or false blocks do not contain a sunken instruction, that
589     // block and its branch may be optimized away. In that case, one side of the
590     // first branch will point directly to select.end, and the corresponding PHI
591     // predecessor block will be the start block.
592 
593     // Find all the instructions that can be soundly sunk to the true/false
594     // blocks. These are instructions that are computed solely for producing the
595     // operands of the select instructions in the group and can be sunk without
596     // breaking the semantics of the LLVM IR (e.g., cannot sink instructions
597     // with side effects).
598     SmallVector<std::stack<Instruction *>, 2> TrueSlices, FalseSlices;
599     typedef std::stack<Instruction *>::size_type StackSizeType;
600     StackSizeType maxTrueSliceLen = 0, maxFalseSliceLen = 0;
601     for (SelectLike SI : ASI) {
602       // For each select, compute the sinkable dependence chains of the true and
603       // false operands.
604       if (auto *TI = dyn_cast_or_null<Instruction>(SI.getTrueValue())) {
605         std::stack<Instruction *> TrueSlice;
606         getExclBackwardsSlice(TI, TrueSlice, SI.getI(), true);
607         maxTrueSliceLen = std::max(maxTrueSliceLen, TrueSlice.size());
608         TrueSlices.push_back(TrueSlice);
609       }
610       if (auto *FI = dyn_cast_or_null<Instruction>(SI.getFalseValue())) {
611         if (isa<SelectInst>(SI.getI()) || !FI->hasOneUse()) {
612           std::stack<Instruction *> FalseSlice;
613           getExclBackwardsSlice(FI, FalseSlice, SI.getI(), true);
614           maxFalseSliceLen = std::max(maxFalseSliceLen, FalseSlice.size());
615           FalseSlices.push_back(FalseSlice);
616         }
617       }
618     }
619     // In the case of multiple select instructions in the same group, the order
620     // of non-dependent instructions (instructions of different dependence
621     // slices) in the true/false blocks appears to affect performance.
622     // Interleaving the slices seems to experimentally be the optimal approach.
623     // This interleaving scheduling allows for more ILP (with a natural downside
624     // of increasing a bit register pressure) compared to a simple ordering of
625     // one whole chain after another. One would expect that this ordering would
626     // not matter since the scheduling in the backend of the compiler  would
627     // take care of it, but apparently the scheduler fails to deliver optimal
628     // ILP with a naive ordering here.
629     SmallVector<Instruction *, 2> TrueSlicesInterleaved, FalseSlicesInterleaved;
630     for (StackSizeType IS = 0; IS < maxTrueSliceLen; ++IS) {
631       for (auto &S : TrueSlices) {
632         if (!S.empty()) {
633           TrueSlicesInterleaved.push_back(S.top());
634           S.pop();
635         }
636       }
637     }
638     for (StackSizeType IS = 0; IS < maxFalseSliceLen; ++IS) {
639       for (auto &S : FalseSlices) {
640         if (!S.empty()) {
641           FalseSlicesInterleaved.push_back(S.top());
642           S.pop();
643         }
644       }
645     }
646 
647     // We split the block containing the select(s) into two blocks.
648     SelectLike SI = ASI.front();
649     SelectLike LastSI = ASI.back();
650     BasicBlock *StartBlock = SI.getI()->getParent();
651     BasicBlock::iterator SplitPt = ++(BasicBlock::iterator(LastSI.getI()));
652     // With RemoveDIs turned off, SplitPt can be a dbg.* intrinsic. With
653     // RemoveDIs turned on, SplitPt would instead point to the next
654     // instruction. To match existing dbg.* intrinsic behaviour with RemoveDIs,
655     // tell splitBasicBlock that we want to include any DbgVariableRecords
656     // attached to SplitPt in the splice.
657     SplitPt.setHeadBit(true);
658     BasicBlock *EndBlock = StartBlock->splitBasicBlock(SplitPt, "select.end");
659     BFI->setBlockFreq(EndBlock, BFI->getBlockFreq(StartBlock));
660     // Delete the unconditional branch that was just created by the split.
661     StartBlock->getTerminator()->eraseFromParent();
662 
663     // Move any debug/pseudo instructions and not's that were in-between the
664     // select group to the newly-created end block.
665     SmallVector<Instruction *, 2> SinkInstrs;
666     auto DIt = SI.getI()->getIterator();
667     while (&*DIt != LastSI.getI()) {
668       if (DIt->isDebugOrPseudoInst())
669         SinkInstrs.push_back(&*DIt);
670       if (match(&*DIt, m_Not(m_Specific(SI.getCondition()))))
671         SinkInstrs.push_back(&*DIt);
672       DIt++;
673     }
674     for (auto *DI : SinkInstrs)
675       DI->moveBeforePreserving(&*EndBlock->getFirstInsertionPt());
676 
677     // Duplicate implementation for DbgRecords, the non-instruction debug-info
678     // format. Helper lambda for moving DbgRecords to the end block.
679     auto TransferDbgRecords = [&](Instruction &I) {
680       for (auto &DbgRecord :
681            llvm::make_early_inc_range(I.getDbgRecordRange())) {
682         DbgRecord.removeFromParent();
683         EndBlock->insertDbgRecordBefore(&DbgRecord,
684                                         EndBlock->getFirstInsertionPt());
685       }
686     };
687 
688     // Iterate over all instructions in between SI and LastSI, not including
689     // SI itself. These are all the variable assignments that happen "in the
690     // middle" of the select group.
691     auto R = make_range(std::next(SI.getI()->getIterator()),
692                         std::next(LastSI.getI()->getIterator()));
693     llvm::for_each(R, TransferDbgRecords);
694 
695     // These are the new basic blocks for the conditional branch.
696     // At least one will become an actual new basic block.
697     BasicBlock *TrueBlock = nullptr, *FalseBlock = nullptr;
698     BranchInst *TrueBranch = nullptr, *FalseBranch = nullptr;
699     if (!TrueSlicesInterleaved.empty()) {
700       TrueBlock = BasicBlock::Create(EndBlock->getContext(), "select.true.sink",
701                                      EndBlock->getParent(), EndBlock);
702       TrueBranch = BranchInst::Create(EndBlock, TrueBlock);
703       TrueBranch->setDebugLoc(LastSI.getI()->getDebugLoc());
704       for (Instruction *TrueInst : TrueSlicesInterleaved)
705         TrueInst->moveBefore(TrueBranch);
706     }
707     if (!FalseSlicesInterleaved.empty()) {
708       FalseBlock =
709           BasicBlock::Create(EndBlock->getContext(), "select.false.sink",
710                              EndBlock->getParent(), EndBlock);
711       FalseBranch = BranchInst::Create(EndBlock, FalseBlock);
712       FalseBranch->setDebugLoc(LastSI.getI()->getDebugLoc());
713       for (Instruction *FalseInst : FalseSlicesInterleaved)
714         FalseInst->moveBefore(FalseBranch);
715     }
716     // If there was nothing to sink, then arbitrarily choose the 'false' side
717     // for a new input value to the PHI.
718     if (TrueBlock == FalseBlock) {
719       assert(TrueBlock == nullptr &&
720              "Unexpected basic block transform while optimizing select");
721 
722       FalseBlock = BasicBlock::Create(StartBlock->getContext(), "select.false",
723                                       EndBlock->getParent(), EndBlock);
724       auto *FalseBranch = BranchInst::Create(EndBlock, FalseBlock);
725       FalseBranch->setDebugLoc(SI.getI()->getDebugLoc());
726     }
727 
728     // Insert the real conditional branch based on the original condition.
729     // If we did not create a new block for one of the 'true' or 'false' paths
730     // of the condition, it means that side of the branch goes to the end block
731     // directly and the path originates from the start block from the point of
732     // view of the new PHI.
733     BasicBlock *TT, *FT;
734     if (TrueBlock == nullptr) {
735       TT = EndBlock;
736       FT = FalseBlock;
737       TrueBlock = StartBlock;
738     } else if (FalseBlock == nullptr) {
739       TT = TrueBlock;
740       FT = EndBlock;
741       FalseBlock = StartBlock;
742     } else {
743       TT = TrueBlock;
744       FT = FalseBlock;
745     }
746     IRBuilder<> IB(SI.getI());
747     auto *CondFr = IB.CreateFreeze(SI.getCondition(),
748                                    SI.getCondition()->getName() + ".frozen");
749 
750     SmallPtrSet<const Instruction *, 2> INS;
751     for (auto SI : ASI)
752       INS.insert(SI.getI());
753 
754     // Use reverse iterator because later select may use the value of the
755     // earlier select, and we need to propagate value through earlier select
756     // to get the PHI operand.
757     for (auto It = ASI.rbegin(); It != ASI.rend(); ++It) {
758       SelectLike SI = *It;
759       // The select itself is replaced with a PHI Node.
760       PHINode *PN = PHINode::Create(SI.getType(), 2, "");
761       PN->insertBefore(EndBlock->begin());
762       PN->takeName(SI.getI());
763       PN->addIncoming(getTrueOrFalseValue(SI, true, INS, IB), TrueBlock);
764       PN->addIncoming(getTrueOrFalseValue(SI, false, INS, IB), FalseBlock);
765       PN->setDebugLoc(SI.getI()->getDebugLoc());
766       SI.getI()->replaceAllUsesWith(PN);
767       INS.erase(SI.getI());
768       ++NumSelectsConverted;
769     }
770     IB.CreateCondBr(CondFr, TT, FT, SI.getI());
771 
772     // Remove the old select instructions, now that they are not longer used.
773     for (auto SI : ASI)
774       SI.getI()->eraseFromParent();
775   }
776 }
777 
778 void SelectOptimizeImpl::collectSelectGroups(BasicBlock &BB,
779                                              SelectGroups &SIGroups) {
780   BasicBlock::iterator BBIt = BB.begin();
781   while (BBIt != BB.end()) {
782     Instruction *I = &*BBIt++;
783     if (SelectLike SI = SelectLike::match(I)) {
784       if (!TTI->shouldTreatInstructionLikeSelect(I))
785         continue;
786 
787       SelectGroup SIGroup;
788       SIGroup.push_back(SI);
789       while (BBIt != BB.end()) {
790         Instruction *NI = &*BBIt;
791         // Debug/pseudo instructions should be skipped and not prevent the
792         // formation of a select group.
793         if (NI->isDebugOrPseudoInst()) {
794           ++BBIt;
795           continue;
796         }
797 
798         // Skip not(select(..)), if the not is part of the same select group
799         if (match(NI, m_Not(m_Specific(SI.getCondition())))) {
800           ++BBIt;
801           continue;
802         }
803 
804         // We only allow selects in the same group, not other select-like
805         // instructions.
806         if (!isa<SelectInst>(NI))
807           break;
808 
809         SelectLike NSI = SelectLike::match(NI);
810         if (NSI && SI.getCondition() == NSI.getCondition()) {
811           SIGroup.push_back(NSI);
812         } else if (NSI && match(NSI.getCondition(),
813                                 m_Not(m_Specific(SI.getCondition())))) {
814           NSI.setInverted();
815           SIGroup.push_back(NSI);
816         } else
817           break;
818         ++BBIt;
819       }
820 
821       // If the select type is not supported, no point optimizing it.
822       // Instruction selection will take care of it.
823       if (!isSelectKindSupported(SI))
824         continue;
825 
826       LLVM_DEBUG({
827         dbgs() << "New Select group with\n";
828         for (auto SI : SIGroup)
829           dbgs() << "  " << *SI.getI() << "\n";
830       });
831 
832       SIGroups.push_back(SIGroup);
833     }
834   }
835 }
836 
837 void SelectOptimizeImpl::findProfitableSIGroupsBase(
838     SelectGroups &SIGroups, SelectGroups &ProfSIGroups) {
839   for (SelectGroup &ASI : SIGroups) {
840     ++NumSelectOptAnalyzed;
841     if (isConvertToBranchProfitableBase(ASI))
842       ProfSIGroups.push_back(ASI);
843   }
844 }
845 
846 static void EmitAndPrintRemark(OptimizationRemarkEmitter *ORE,
847                                DiagnosticInfoOptimizationBase &Rem) {
848   LLVM_DEBUG(dbgs() << Rem.getMsg() << "\n");
849   ORE->emit(Rem);
850 }
851 
852 void SelectOptimizeImpl::findProfitableSIGroupsInnerLoops(
853     const Loop *L, SelectGroups &SIGroups, SelectGroups &ProfSIGroups) {
854   NumSelectOptAnalyzed += SIGroups.size();
855   // For each select group in an inner-most loop,
856   // a branch is more preferable than a select/conditional-move if:
857   // i) conversion to branches for all the select groups of the loop satisfies
858   //    loop-level heuristics including reducing the loop's critical path by
859   //    some threshold (see SelectOptimizeImpl::checkLoopHeuristics); and
860   // ii) the total cost of the select group is cheaper with a branch compared
861   //     to its predicated version. The cost is in terms of latency and the cost
862   //     of a select group is the cost of its most expensive select instruction
863   //     (assuming infinite resources and thus fully leveraging available ILP).
864 
865   DenseMap<const Instruction *, CostInfo> InstCostMap;
866   CostInfo LoopCost[2] = {{Scaled64::getZero(), Scaled64::getZero()},
867                           {Scaled64::getZero(), Scaled64::getZero()}};
868   if (!computeLoopCosts(L, SIGroups, InstCostMap, LoopCost) ||
869       !checkLoopHeuristics(L, LoopCost)) {
870     return;
871   }
872 
873   for (SelectGroup &ASI : SIGroups) {
874     // Assuming infinite resources, the cost of a group of instructions is the
875     // cost of the most expensive instruction of the group.
876     Scaled64 SelectCost = Scaled64::getZero(), BranchCost = Scaled64::getZero();
877     for (SelectLike SI : ASI) {
878       SelectCost = std::max(SelectCost, InstCostMap[SI.getI()].PredCost);
879       BranchCost = std::max(BranchCost, InstCostMap[SI.getI()].NonPredCost);
880     }
881     if (BranchCost < SelectCost) {
882       OptimizationRemark OR(DEBUG_TYPE, "SelectOpti", ASI.front().getI());
883       OR << "Profitable to convert to branch (loop analysis). BranchCost="
884          << BranchCost.toString() << ", SelectCost=" << SelectCost.toString()
885          << ". ";
886       EmitAndPrintRemark(ORE, OR);
887       ++NumSelectConvertedLoop;
888       ProfSIGroups.push_back(ASI);
889     } else {
890       OptimizationRemarkMissed ORmiss(DEBUG_TYPE, "SelectOpti",
891                                       ASI.front().getI());
892       ORmiss << "Select is more profitable (loop analysis). BranchCost="
893              << BranchCost.toString()
894              << ", SelectCost=" << SelectCost.toString() << ". ";
895       EmitAndPrintRemark(ORE, ORmiss);
896     }
897   }
898 }
899 
900 bool SelectOptimizeImpl::isConvertToBranchProfitableBase(
901     const SelectGroup &ASI) {
902   SelectLike SI = ASI.front();
903   LLVM_DEBUG(dbgs() << "Analyzing select group containing " << *SI.getI()
904                     << "\n");
905   OptimizationRemark OR(DEBUG_TYPE, "SelectOpti", SI.getI());
906   OptimizationRemarkMissed ORmiss(DEBUG_TYPE, "SelectOpti", SI.getI());
907 
908   // Skip cold basic blocks. Better to optimize for size for cold blocks.
909   if (PSI->isColdBlock(SI.getI()->getParent(), BFI)) {
910     ++NumSelectColdBB;
911     ORmiss << "Not converted to branch because of cold basic block. ";
912     EmitAndPrintRemark(ORE, ORmiss);
913     return false;
914   }
915 
916   // If unpredictable, branch form is less profitable.
917   if (SI.getI()->getMetadata(LLVMContext::MD_unpredictable)) {
918     ++NumSelectUnPred;
919     ORmiss << "Not converted to branch because of unpredictable branch. ";
920     EmitAndPrintRemark(ORE, ORmiss);
921     return false;
922   }
923 
924   // If highly predictable, branch form is more profitable, unless a
925   // predictable select is inexpensive in the target architecture.
926   if (isSelectHighlyPredictable(SI) && TLI->isPredictableSelectExpensive()) {
927     ++NumSelectConvertedHighPred;
928     OR << "Converted to branch because of highly predictable branch. ";
929     EmitAndPrintRemark(ORE, OR);
930     return true;
931   }
932 
933   // Look for expensive instructions in the cold operand's (if any) dependence
934   // slice of any of the selects in the group.
935   if (hasExpensiveColdOperand(ASI)) {
936     ++NumSelectConvertedExpColdOperand;
937     OR << "Converted to branch because of expensive cold operand.";
938     EmitAndPrintRemark(ORE, OR);
939     return true;
940   }
941 
942   ORmiss << "Not profitable to convert to branch (base heuristic).";
943   EmitAndPrintRemark(ORE, ORmiss);
944   return false;
945 }
946 
947 static InstructionCost divideNearest(InstructionCost Numerator,
948                                      uint64_t Denominator) {
949   return (Numerator + (Denominator / 2)) / Denominator;
950 }
951 
952 static bool extractBranchWeights(const SelectOptimizeImpl::SelectLike SI,
953                                  uint64_t &TrueVal, uint64_t &FalseVal) {
954   if (isa<SelectInst>(SI.getI()))
955     return extractBranchWeights(*SI.getI(), TrueVal, FalseVal);
956   return false;
957 }
958 
959 bool SelectOptimizeImpl::hasExpensiveColdOperand(const SelectGroup &ASI) {
960   bool ColdOperand = false;
961   uint64_t TrueWeight, FalseWeight, TotalWeight;
962   if (extractBranchWeights(ASI.front(), TrueWeight, FalseWeight)) {
963     uint64_t MinWeight = std::min(TrueWeight, FalseWeight);
964     TotalWeight = TrueWeight + FalseWeight;
965     // Is there a path with frequency <ColdOperandThreshold% (default:20%) ?
966     ColdOperand = TotalWeight * ColdOperandThreshold > 100 * MinWeight;
967   } else if (PSI->hasProfileSummary()) {
968     OptimizationRemarkMissed ORmiss(DEBUG_TYPE, "SelectOpti",
969                                     ASI.front().getI());
970     ORmiss << "Profile data available but missing branch-weights metadata for "
971               "select instruction. ";
972     EmitAndPrintRemark(ORE, ORmiss);
973   }
974   if (!ColdOperand)
975     return false;
976   // Check if the cold path's dependence slice is expensive for any of the
977   // selects of the group.
978   for (SelectLike SI : ASI) {
979     Instruction *ColdI = nullptr;
980     uint64_t HotWeight;
981     if (TrueWeight < FalseWeight) {
982       ColdI = dyn_cast_or_null<Instruction>(SI.getTrueValue());
983       HotWeight = FalseWeight;
984     } else {
985       ColdI = dyn_cast_or_null<Instruction>(SI.getFalseValue());
986       HotWeight = TrueWeight;
987     }
988     if (ColdI) {
989       std::stack<Instruction *> ColdSlice;
990       getExclBackwardsSlice(ColdI, ColdSlice, SI.getI());
991       InstructionCost SliceCost = 0;
992       while (!ColdSlice.empty()) {
993         SliceCost += TTI->getInstructionCost(ColdSlice.top(),
994                                              TargetTransformInfo::TCK_Latency);
995         ColdSlice.pop();
996       }
997       // The colder the cold value operand of the select is the more expensive
998       // the cmov becomes for computing the cold value operand every time. Thus,
999       // the colder the cold operand is the more its cost counts.
1000       // Get nearest integer cost adjusted for coldness.
1001       InstructionCost AdjSliceCost =
1002           divideNearest(SliceCost * HotWeight, TotalWeight);
1003       if (AdjSliceCost >=
1004           ColdOperandMaxCostMultiplier * TargetTransformInfo::TCC_Expensive)
1005         return true;
1006     }
1007   }
1008   return false;
1009 }
1010 
1011 // Check if it is safe to move LoadI next to the SI.
1012 // Conservatively assume it is safe only if there is no instruction
1013 // modifying memory in-between the load and the select instruction.
1014 static bool isSafeToSinkLoad(Instruction *LoadI, Instruction *SI) {
1015   // Assume loads from different basic blocks are unsafe to move.
1016   if (LoadI->getParent() != SI->getParent())
1017     return false;
1018   auto It = LoadI->getIterator();
1019   while (&*It != SI) {
1020     if (It->mayWriteToMemory())
1021       return false;
1022     It++;
1023   }
1024   return true;
1025 }
1026 
1027 // For a given source instruction, collect its backwards dependence slice
1028 // consisting of instructions exclusively computed for the purpose of producing
1029 // the operands of the source instruction. As an approximation
1030 // (sufficiently-accurate in practice), we populate this set with the
1031 // instructions of the backwards dependence slice that only have one-use and
1032 // form an one-use chain that leads to the source instruction.
1033 void SelectOptimizeImpl::getExclBackwardsSlice(Instruction *I,
1034                                                std::stack<Instruction *> &Slice,
1035                                                Instruction *SI,
1036                                                bool ForSinking) {
1037   SmallPtrSet<Instruction *, 2> Visited;
1038   std::queue<Instruction *> Worklist;
1039   Worklist.push(I);
1040   while (!Worklist.empty()) {
1041     Instruction *II = Worklist.front();
1042     Worklist.pop();
1043 
1044     // Avoid cycles.
1045     if (!Visited.insert(II).second)
1046       continue;
1047 
1048     if (!II->hasOneUse())
1049       continue;
1050 
1051     // Cannot soundly sink instructions with side-effects.
1052     // Terminator or phi instructions cannot be sunk.
1053     // Avoid sinking other select instructions (should be handled separetely).
1054     if (ForSinking && (II->isTerminator() || II->mayHaveSideEffects() ||
1055                        isa<SelectInst>(II) || isa<PHINode>(II)))
1056       continue;
1057 
1058     // Avoid sinking loads in order not to skip state-modifying instructions,
1059     // that may alias with the loaded address.
1060     // Only allow sinking of loads within the same basic block that are
1061     // conservatively proven to be safe.
1062     if (ForSinking && II->mayReadFromMemory() && !isSafeToSinkLoad(II, SI))
1063       continue;
1064 
1065     // Avoid considering instructions with less frequency than the source
1066     // instruction (i.e., avoid colder code regions of the dependence slice).
1067     if (BFI->getBlockFreq(II->getParent()) < BFI->getBlockFreq(I->getParent()))
1068       continue;
1069 
1070     // Eligible one-use instruction added to the dependence slice.
1071     Slice.push(II);
1072 
1073     // Explore all the operands of the current instruction to expand the slice.
1074     for (Value *Op : II->operand_values())
1075       if (auto *OpI = dyn_cast<Instruction>(Op))
1076         Worklist.push(OpI);
1077   }
1078 }
1079 
1080 bool SelectOptimizeImpl::isSelectHighlyPredictable(const SelectLike SI) {
1081   uint64_t TrueWeight, FalseWeight;
1082   if (extractBranchWeights(SI, TrueWeight, FalseWeight)) {
1083     uint64_t Max = std::max(TrueWeight, FalseWeight);
1084     uint64_t Sum = TrueWeight + FalseWeight;
1085     if (Sum != 0) {
1086       auto Probability = BranchProbability::getBranchProbability(Max, Sum);
1087       if (Probability > TTI->getPredictableBranchThreshold())
1088         return true;
1089     }
1090   }
1091   return false;
1092 }
1093 
1094 bool SelectOptimizeImpl::checkLoopHeuristics(const Loop *L,
1095                                              const CostInfo LoopCost[2]) {
1096   // Loop-level checks to determine if a non-predicated version (with branches)
1097   // of the loop is more profitable than its predicated version.
1098 
1099   if (DisableLoopLevelHeuristics)
1100     return true;
1101 
1102   OptimizationRemarkMissed ORmissL(DEBUG_TYPE, "SelectOpti",
1103                                    L->getHeader()->getFirstNonPHI());
1104 
1105   if (LoopCost[0].NonPredCost > LoopCost[0].PredCost ||
1106       LoopCost[1].NonPredCost >= LoopCost[1].PredCost) {
1107     ORmissL << "No select conversion in the loop due to no reduction of loop's "
1108                "critical path. ";
1109     EmitAndPrintRemark(ORE, ORmissL);
1110     return false;
1111   }
1112 
1113   Scaled64 Gain[2] = {LoopCost[0].PredCost - LoopCost[0].NonPredCost,
1114                       LoopCost[1].PredCost - LoopCost[1].NonPredCost};
1115 
1116   // Profitably converting to branches need to reduce the loop's critical path
1117   // by at least some threshold (absolute gain of GainCycleThreshold cycles and
1118   // relative gain of 12.5%).
1119   if (Gain[1] < Scaled64::get(GainCycleThreshold) ||
1120       Gain[1] * Scaled64::get(GainRelativeThreshold) < LoopCost[1].PredCost) {
1121     Scaled64 RelativeGain = Scaled64::get(100) * Gain[1] / LoopCost[1].PredCost;
1122     ORmissL << "No select conversion in the loop due to small reduction of "
1123                "loop's critical path. Gain="
1124             << Gain[1].toString()
1125             << ", RelativeGain=" << RelativeGain.toString() << "%. ";
1126     EmitAndPrintRemark(ORE, ORmissL);
1127     return false;
1128   }
1129 
1130   // If the loop's critical path involves loop-carried dependences, the gradient
1131   // of the gain needs to be at least GainGradientThreshold% (defaults to 25%).
1132   // This check ensures that the latency reduction for the loop's critical path
1133   // keeps decreasing with sufficient rate beyond the two analyzed loop
1134   // iterations.
1135   if (Gain[1] > Gain[0]) {
1136     Scaled64 GradientGain = Scaled64::get(100) * (Gain[1] - Gain[0]) /
1137                             (LoopCost[1].PredCost - LoopCost[0].PredCost);
1138     if (GradientGain < Scaled64::get(GainGradientThreshold)) {
1139       ORmissL << "No select conversion in the loop due to small gradient gain. "
1140                  "GradientGain="
1141               << GradientGain.toString() << "%. ";
1142       EmitAndPrintRemark(ORE, ORmissL);
1143       return false;
1144     }
1145   }
1146   // If the gain decreases it is not profitable to convert.
1147   else if (Gain[1] < Gain[0]) {
1148     ORmissL
1149         << "No select conversion in the loop due to negative gradient gain. ";
1150     EmitAndPrintRemark(ORE, ORmissL);
1151     return false;
1152   }
1153 
1154   // Non-predicated version of the loop is more profitable than its
1155   // predicated version.
1156   return true;
1157 }
1158 
1159 // Computes instruction and loop-critical-path costs for both the predicated
1160 // and non-predicated version of the given loop.
1161 // Returns false if unable to compute these costs due to invalid cost of loop
1162 // instruction(s).
1163 bool SelectOptimizeImpl::computeLoopCosts(
1164     const Loop *L, const SelectGroups &SIGroups,
1165     DenseMap<const Instruction *, CostInfo> &InstCostMap, CostInfo *LoopCost) {
1166   LLVM_DEBUG(dbgs() << "Calculating Latency / IPredCost / INonPredCost of loop "
1167                     << L->getHeader()->getName() << "\n");
1168   const auto &SImap = getSImap(SIGroups);
1169   // Compute instruction and loop-critical-path costs across two iterations for
1170   // both predicated and non-predicated version.
1171   const unsigned Iterations = 2;
1172   for (unsigned Iter = 0; Iter < Iterations; ++Iter) {
1173     // Cost of the loop's critical path.
1174     CostInfo &MaxCost = LoopCost[Iter];
1175     for (BasicBlock *BB : L->getBlocks()) {
1176       for (const Instruction &I : *BB) {
1177         if (I.isDebugOrPseudoInst())
1178           continue;
1179         // Compute the predicated and non-predicated cost of the instruction.
1180         Scaled64 IPredCost = Scaled64::getZero(),
1181                  INonPredCost = Scaled64::getZero();
1182 
1183         // Assume infinite resources that allow to fully exploit the available
1184         // instruction-level parallelism.
1185         // InstCost = InstLatency + max(Op1Cost, Op2Cost, … OpNCost)
1186         for (const Use &U : I.operands()) {
1187           auto UI = dyn_cast<Instruction>(U.get());
1188           if (!UI)
1189             continue;
1190           if (InstCostMap.count(UI)) {
1191             IPredCost = std::max(IPredCost, InstCostMap[UI].PredCost);
1192             INonPredCost = std::max(INonPredCost, InstCostMap[UI].NonPredCost);
1193           }
1194         }
1195         auto ILatency = computeInstCost(&I);
1196         if (!ILatency) {
1197           OptimizationRemarkMissed ORmissL(DEBUG_TYPE, "SelectOpti", &I);
1198           ORmissL << "Invalid instruction cost preventing analysis and "
1199                      "optimization of the inner-most loop containing this "
1200                      "instruction. ";
1201           EmitAndPrintRemark(ORE, ORmissL);
1202           return false;
1203         }
1204         IPredCost += Scaled64::get(*ILatency);
1205         INonPredCost += Scaled64::get(*ILatency);
1206 
1207         // For a select that can be converted to branch,
1208         // compute its cost as a branch (non-predicated cost).
1209         //
1210         // BranchCost = PredictedPathCost + MispredictCost
1211         // PredictedPathCost = TrueOpCost * TrueProb + FalseOpCost * FalseProb
1212         // MispredictCost = max(MispredictPenalty, CondCost) * MispredictRate
1213         if (SImap.contains(&I)) {
1214           auto SI = SImap.at(&I);
1215           Scaled64 TrueOpCost = SI.getTrueOpCost(InstCostMap, TTI);
1216           Scaled64 FalseOpCost = SI.getFalseOpCost(InstCostMap, TTI);
1217           Scaled64 PredictedPathCost =
1218               getPredictedPathCost(TrueOpCost, FalseOpCost, SI);
1219 
1220           Scaled64 CondCost = Scaled64::getZero();
1221           if (auto *CI = dyn_cast<Instruction>(SI.getCondition()))
1222             if (InstCostMap.count(CI))
1223               CondCost = InstCostMap[CI].NonPredCost;
1224           Scaled64 MispredictCost = getMispredictionCost(SI, CondCost);
1225 
1226           INonPredCost = PredictedPathCost + MispredictCost;
1227         }
1228         LLVM_DEBUG(dbgs() << " " << ILatency << "/" << IPredCost << "/"
1229                           << INonPredCost << " for " << I << "\n");
1230 
1231         InstCostMap[&I] = {IPredCost, INonPredCost};
1232         MaxCost.PredCost = std::max(MaxCost.PredCost, IPredCost);
1233         MaxCost.NonPredCost = std::max(MaxCost.NonPredCost, INonPredCost);
1234       }
1235     }
1236     LLVM_DEBUG(dbgs() << "Iteration " << Iter + 1
1237                       << " MaxCost = " << MaxCost.PredCost << " "
1238                       << MaxCost.NonPredCost << "\n");
1239   }
1240   return true;
1241 }
1242 
1243 SmallDenseMap<const Instruction *, SelectOptimizeImpl::SelectLike, 2>
1244 SelectOptimizeImpl::getSImap(const SelectGroups &SIGroups) {
1245   SmallDenseMap<const Instruction *, SelectLike, 2> SImap;
1246   for (const SelectGroup &ASI : SIGroups)
1247     for (SelectLike SI : ASI)
1248       SImap.try_emplace(SI.getI(), SI);
1249   return SImap;
1250 }
1251 
1252 std::optional<uint64_t>
1253 SelectOptimizeImpl::computeInstCost(const Instruction *I) {
1254   InstructionCost ICost =
1255       TTI->getInstructionCost(I, TargetTransformInfo::TCK_Latency);
1256   if (auto OC = ICost.getValue())
1257     return std::optional<uint64_t>(*OC);
1258   return std::nullopt;
1259 }
1260 
1261 ScaledNumber<uint64_t>
1262 SelectOptimizeImpl::getMispredictionCost(const SelectLike SI,
1263                                          const Scaled64 CondCost) {
1264   uint64_t MispredictPenalty = TSchedModel.getMCSchedModel()->MispredictPenalty;
1265 
1266   // Account for the default misprediction rate when using a branch
1267   // (conservatively set to 25% by default).
1268   uint64_t MispredictRate = MispredictDefaultRate;
1269   // If the select condition is obviously predictable, then the misprediction
1270   // rate is zero.
1271   if (isSelectHighlyPredictable(SI))
1272     MispredictRate = 0;
1273 
1274   // CondCost is included to account for cases where the computation of the
1275   // condition is part of a long dependence chain (potentially loop-carried)
1276   // that would delay detection of a misprediction and increase its cost.
1277   Scaled64 MispredictCost =
1278       std::max(Scaled64::get(MispredictPenalty), CondCost) *
1279       Scaled64::get(MispredictRate);
1280   MispredictCost /= Scaled64::get(100);
1281 
1282   return MispredictCost;
1283 }
1284 
1285 // Returns the cost of a branch when the prediction is correct.
1286 // TrueCost * TrueProbability + FalseCost * FalseProbability.
1287 ScaledNumber<uint64_t>
1288 SelectOptimizeImpl::getPredictedPathCost(Scaled64 TrueCost, Scaled64 FalseCost,
1289                                          const SelectLike SI) {
1290   Scaled64 PredPathCost;
1291   uint64_t TrueWeight, FalseWeight;
1292   if (extractBranchWeights(SI, TrueWeight, FalseWeight)) {
1293     uint64_t SumWeight = TrueWeight + FalseWeight;
1294     if (SumWeight != 0) {
1295       PredPathCost = TrueCost * Scaled64::get(TrueWeight) +
1296                      FalseCost * Scaled64::get(FalseWeight);
1297       PredPathCost /= Scaled64::get(SumWeight);
1298       return PredPathCost;
1299     }
1300   }
1301   // Without branch weight metadata, we assume 75% for the one path and 25% for
1302   // the other, and pick the result with the biggest cost.
1303   PredPathCost = std::max(TrueCost * Scaled64::get(3) + FalseCost,
1304                           FalseCost * Scaled64::get(3) + TrueCost);
1305   PredPathCost /= Scaled64::get(4);
1306   return PredPathCost;
1307 }
1308 
1309 bool SelectOptimizeImpl::isSelectKindSupported(const SelectLike SI) {
1310   bool VectorCond = !SI.getCondition()->getType()->isIntegerTy(1);
1311   if (VectorCond)
1312     return false;
1313   TargetLowering::SelectSupportKind SelectKind;
1314   if (SI.getType()->isVectorTy())
1315     SelectKind = TargetLowering::ScalarCondVectorVal;
1316   else
1317     SelectKind = TargetLowering::ScalarValSelect;
1318   return TLI->isSelectSupported(SelectKind);
1319 }
1320