xref: /freebsd/contrib/llvm-project/llvm/lib/CodeGen/SelectOptimize.cpp (revision f126890ac5386406dadf7c4cfa9566cbb56537c5)
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/ADT/SmallVector.h"
14 #include "llvm/ADT/Statistic.h"
15 #include "llvm/Analysis/BlockFrequencyInfo.h"
16 #include "llvm/Analysis/BranchProbabilityInfo.h"
17 #include "llvm/Analysis/LoopInfo.h"
18 #include "llvm/Analysis/OptimizationRemarkEmitter.h"
19 #include "llvm/Analysis/ProfileSummaryInfo.h"
20 #include "llvm/Analysis/TargetTransformInfo.h"
21 #include "llvm/CodeGen/Passes.h"
22 #include "llvm/CodeGen/TargetLowering.h"
23 #include "llvm/CodeGen/TargetPassConfig.h"
24 #include "llvm/CodeGen/TargetSchedule.h"
25 #include "llvm/CodeGen/TargetSubtargetInfo.h"
26 #include "llvm/IR/BasicBlock.h"
27 #include "llvm/IR/Dominators.h"
28 #include "llvm/IR/Function.h"
29 #include "llvm/IR/IRBuilder.h"
30 #include "llvm/IR/Instruction.h"
31 #include "llvm/IR/PatternMatch.h"
32 #include "llvm/IR/ProfDataUtils.h"
33 #include "llvm/InitializePasses.h"
34 #include "llvm/Pass.h"
35 #include "llvm/Support/ScaledNumber.h"
36 #include "llvm/Target/TargetMachine.h"
37 #include "llvm/Transforms/Utils/SizeOpts.h"
38 #include <algorithm>
39 #include <memory>
40 #include <queue>
41 #include <stack>
42 #include <string>
43 
44 using namespace llvm;
45 
46 #define DEBUG_TYPE "select-optimize"
47 
48 STATISTIC(NumSelectOptAnalyzed,
49           "Number of select groups considered for conversion to branch");
50 STATISTIC(NumSelectConvertedExpColdOperand,
51           "Number of select groups converted due to expensive cold operand");
52 STATISTIC(NumSelectConvertedHighPred,
53           "Number of select groups converted due to high-predictability");
54 STATISTIC(NumSelectUnPred,
55           "Number of select groups not converted due to unpredictability");
56 STATISTIC(NumSelectColdBB,
57           "Number of select groups not converted due to cold basic block");
58 STATISTIC(NumSelectConvertedLoop,
59           "Number of select groups converted due to loop-level analysis");
60 STATISTIC(NumSelectsConverted, "Number of selects converted");
61 
62 static cl::opt<unsigned> ColdOperandThreshold(
63     "cold-operand-threshold",
64     cl::desc("Maximum frequency of path for an operand to be considered cold."),
65     cl::init(20), cl::Hidden);
66 
67 static cl::opt<unsigned> ColdOperandMaxCostMultiplier(
68     "cold-operand-max-cost-multiplier",
69     cl::desc("Maximum cost multiplier of TCC_expensive for the dependence "
70              "slice of a cold operand to be considered inexpensive."),
71     cl::init(1), cl::Hidden);
72 
73 static cl::opt<unsigned>
74     GainGradientThreshold("select-opti-loop-gradient-gain-threshold",
75                           cl::desc("Gradient gain threshold (%)."),
76                           cl::init(25), cl::Hidden);
77 
78 static cl::opt<unsigned>
79     GainCycleThreshold("select-opti-loop-cycle-gain-threshold",
80                        cl::desc("Minimum gain per loop (in cycles) threshold."),
81                        cl::init(4), cl::Hidden);
82 
83 static cl::opt<unsigned> GainRelativeThreshold(
84     "select-opti-loop-relative-gain-threshold",
85     cl::desc(
86         "Minimum relative gain per loop threshold (1/X). Defaults to 12.5%"),
87     cl::init(8), cl::Hidden);
88 
89 static cl::opt<unsigned> MispredictDefaultRate(
90     "mispredict-default-rate", cl::Hidden, cl::init(25),
91     cl::desc("Default mispredict rate (initialized to 25%)."));
92 
93 static cl::opt<bool>
94     DisableLoopLevelHeuristics("disable-loop-level-heuristics", cl::Hidden,
95                                cl::init(false),
96                                cl::desc("Disable loop-level heuristics."));
97 
98 namespace {
99 
100 class SelectOptimize : public FunctionPass {
101   const TargetMachine *TM = nullptr;
102   const TargetSubtargetInfo *TSI = nullptr;
103   const TargetLowering *TLI = nullptr;
104   const TargetTransformInfo *TTI = nullptr;
105   const LoopInfo *LI = nullptr;
106   DominatorTree *DT = nullptr;
107   std::unique_ptr<BlockFrequencyInfo> BFI;
108   std::unique_ptr<BranchProbabilityInfo> BPI;
109   ProfileSummaryInfo *PSI = nullptr;
110   OptimizationRemarkEmitter *ORE = nullptr;
111   TargetSchedModel TSchedModel;
112 
113 public:
114   static char ID;
115 
116   SelectOptimize() : FunctionPass(ID) {
117     initializeSelectOptimizePass(*PassRegistry::getPassRegistry());
118   }
119 
120   bool runOnFunction(Function &F) override;
121 
122   void getAnalysisUsage(AnalysisUsage &AU) const override {
123     AU.addRequired<ProfileSummaryInfoWrapperPass>();
124     AU.addRequired<TargetPassConfig>();
125     AU.addRequired<TargetTransformInfoWrapperPass>();
126     AU.addRequired<DominatorTreeWrapperPass>();
127     AU.addRequired<LoopInfoWrapperPass>();
128     AU.addRequired<OptimizationRemarkEmitterWrapperPass>();
129   }
130 
131 private:
132   // Select groups consist of consecutive select instructions with the same
133   // condition.
134   using SelectGroup = SmallVector<SelectInst *, 2>;
135   using SelectGroups = SmallVector<SelectGroup, 2>;
136 
137   using Scaled64 = ScaledNumber<uint64_t>;
138 
139   struct CostInfo {
140     /// Predicated cost (with selects as conditional moves).
141     Scaled64 PredCost;
142     /// Non-predicated cost (with selects converted to branches).
143     Scaled64 NonPredCost;
144   };
145 
146   // Converts select instructions of a function to conditional jumps when deemed
147   // profitable. Returns true if at least one select was converted.
148   bool optimizeSelects(Function &F);
149 
150   // Heuristics for determining which select instructions can be profitably
151   // conveted to branches. Separate heuristics for selects in inner-most loops
152   // and the rest of code regions (base heuristics for non-inner-most loop
153   // regions).
154   void optimizeSelectsBase(Function &F, SelectGroups &ProfSIGroups);
155   void optimizeSelectsInnerLoops(Function &F, SelectGroups &ProfSIGroups);
156 
157   // Converts to branches the select groups that were deemed
158   // profitable-to-convert.
159   void convertProfitableSIGroups(SelectGroups &ProfSIGroups);
160 
161   // Splits selects of a given basic block into select groups.
162   void collectSelectGroups(BasicBlock &BB, SelectGroups &SIGroups);
163 
164   // Determines for which select groups it is profitable converting to branches
165   // (base and inner-most-loop heuristics).
166   void findProfitableSIGroupsBase(SelectGroups &SIGroups,
167                                   SelectGroups &ProfSIGroups);
168   void findProfitableSIGroupsInnerLoops(const Loop *L, SelectGroups &SIGroups,
169                                         SelectGroups &ProfSIGroups);
170 
171   // Determines if a select group should be converted to a branch (base
172   // heuristics).
173   bool isConvertToBranchProfitableBase(const SmallVector<SelectInst *, 2> &ASI);
174 
175   // Returns true if there are expensive instructions in the cold value
176   // operand's (if any) dependence slice of any of the selects of the given
177   // group.
178   bool hasExpensiveColdOperand(const SmallVector<SelectInst *, 2> &ASI);
179 
180   // For a given source instruction, collect its backwards dependence slice
181   // consisting of instructions exclusively computed for producing the operands
182   // of the source instruction.
183   void getExclBackwardsSlice(Instruction *I, std::stack<Instruction *> &Slice,
184                              Instruction *SI, bool ForSinking = false);
185 
186   // Returns true if the condition of the select is highly predictable.
187   bool isSelectHighlyPredictable(const SelectInst *SI);
188 
189   // Loop-level checks to determine if a non-predicated version (with branches)
190   // of the given loop is more profitable than its predicated version.
191   bool checkLoopHeuristics(const Loop *L, const CostInfo LoopDepth[2]);
192 
193   // Computes instruction and loop-critical-path costs for both the predicated
194   // and non-predicated version of the given loop.
195   bool computeLoopCosts(const Loop *L, const SelectGroups &SIGroups,
196                         DenseMap<const Instruction *, CostInfo> &InstCostMap,
197                         CostInfo *LoopCost);
198 
199   // Returns a set of all the select instructions in the given select groups.
200   SmallPtrSet<const Instruction *, 2> getSIset(const SelectGroups &SIGroups);
201 
202   // Returns the latency cost of a given instruction.
203   std::optional<uint64_t> computeInstCost(const Instruction *I);
204 
205   // Returns the misprediction cost of a given select when converted to branch.
206   Scaled64 getMispredictionCost(const SelectInst *SI, const Scaled64 CondCost);
207 
208   // Returns the cost of a branch when the prediction is correct.
209   Scaled64 getPredictedPathCost(Scaled64 TrueCost, Scaled64 FalseCost,
210                                 const SelectInst *SI);
211 
212   // Returns true if the target architecture supports lowering a given select.
213   bool isSelectKindSupported(SelectInst *SI);
214 };
215 } // namespace
216 
217 char SelectOptimize::ID = 0;
218 
219 INITIALIZE_PASS_BEGIN(SelectOptimize, DEBUG_TYPE, "Optimize selects", false,
220                       false)
221 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
222 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
223 INITIALIZE_PASS_DEPENDENCY(ProfileSummaryInfoWrapperPass)
224 INITIALIZE_PASS_DEPENDENCY(TargetPassConfig)
225 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
226 INITIALIZE_PASS_DEPENDENCY(OptimizationRemarkEmitterWrapperPass)
227 INITIALIZE_PASS_END(SelectOptimize, DEBUG_TYPE, "Optimize selects", false,
228                     false)
229 
230 FunctionPass *llvm::createSelectOptimizePass() { return new SelectOptimize(); }
231 
232 bool SelectOptimize::runOnFunction(Function &F) {
233   TM = &getAnalysis<TargetPassConfig>().getTM<TargetMachine>();
234   TSI = TM->getSubtargetImpl(F);
235   TLI = TSI->getTargetLowering();
236 
237   // If none of the select types is supported then skip this pass.
238   // This is an optimization pass. Legality issues will be handled by
239   // instruction selection.
240   if (!TLI->isSelectSupported(TargetLowering::ScalarValSelect) &&
241       !TLI->isSelectSupported(TargetLowering::ScalarCondVectorVal) &&
242       !TLI->isSelectSupported(TargetLowering::VectorMaskSelect))
243     return false;
244 
245   TTI = &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
246 
247   if (!TTI->enableSelectOptimize())
248     return false;
249 
250   DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
251   LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
252   BPI.reset(new BranchProbabilityInfo(F, *LI));
253   BFI.reset(new BlockFrequencyInfo(F, *BPI, *LI));
254   PSI = &getAnalysis<ProfileSummaryInfoWrapperPass>().getPSI();
255   ORE = &getAnalysis<OptimizationRemarkEmitterWrapperPass>().getORE();
256   TSchedModel.init(TSI);
257 
258   // When optimizing for size, selects are preferable over branches.
259   if (F.hasOptSize() || llvm::shouldOptimizeForSize(&F, PSI, BFI.get()))
260     return false;
261 
262   return optimizeSelects(F);
263 }
264 
265 bool SelectOptimize::optimizeSelects(Function &F) {
266   // Determine for which select groups it is profitable converting to branches.
267   SelectGroups ProfSIGroups;
268   // Base heuristics apply only to non-loops and outer loops.
269   optimizeSelectsBase(F, ProfSIGroups);
270   // Separate heuristics for inner-most loops.
271   optimizeSelectsInnerLoops(F, ProfSIGroups);
272 
273   // Convert to branches the select groups that were deemed
274   // profitable-to-convert.
275   convertProfitableSIGroups(ProfSIGroups);
276 
277   // Code modified if at least one select group was converted.
278   return !ProfSIGroups.empty();
279 }
280 
281 void SelectOptimize::optimizeSelectsBase(Function &F,
282                                          SelectGroups &ProfSIGroups) {
283   // Collect all the select groups.
284   SelectGroups SIGroups;
285   for (BasicBlock &BB : F) {
286     // Base heuristics apply only to non-loops and outer loops.
287     Loop *L = LI->getLoopFor(&BB);
288     if (L && L->isInnermost())
289       continue;
290     collectSelectGroups(BB, SIGroups);
291   }
292 
293   // Determine for which select groups it is profitable converting to branches.
294   findProfitableSIGroupsBase(SIGroups, ProfSIGroups);
295 }
296 
297 void SelectOptimize::optimizeSelectsInnerLoops(Function &F,
298                                                SelectGroups &ProfSIGroups) {
299   SmallVector<Loop *, 4> Loops(LI->begin(), LI->end());
300   // Need to check size on each iteration as we accumulate child loops.
301   for (unsigned long i = 0; i < Loops.size(); ++i)
302     for (Loop *ChildL : Loops[i]->getSubLoops())
303       Loops.push_back(ChildL);
304 
305   for (Loop *L : Loops) {
306     if (!L->isInnermost())
307       continue;
308 
309     SelectGroups SIGroups;
310     for (BasicBlock *BB : L->getBlocks())
311       collectSelectGroups(*BB, SIGroups);
312 
313     findProfitableSIGroupsInnerLoops(L, SIGroups, ProfSIGroups);
314   }
315 }
316 
317 /// If \p isTrue is true, return the true value of \p SI, otherwise return
318 /// false value of \p SI. If the true/false value of \p SI is defined by any
319 /// select instructions in \p Selects, look through the defining select
320 /// instruction until the true/false value is not defined in \p Selects.
321 static Value *
322 getTrueOrFalseValue(SelectInst *SI, bool isTrue,
323                     const SmallPtrSet<const Instruction *, 2> &Selects) {
324   Value *V = nullptr;
325   for (SelectInst *DefSI = SI; DefSI != nullptr && Selects.count(DefSI);
326        DefSI = dyn_cast<SelectInst>(V)) {
327     assert(DefSI->getCondition() == SI->getCondition() &&
328            "The condition of DefSI does not match with SI");
329     V = (isTrue ? DefSI->getTrueValue() : DefSI->getFalseValue());
330   }
331   assert(V && "Failed to get select true/false value");
332   return V;
333 }
334 
335 void SelectOptimize::convertProfitableSIGroups(SelectGroups &ProfSIGroups) {
336   for (SelectGroup &ASI : ProfSIGroups) {
337     // The code transformation here is a modified version of the sinking
338     // transformation in CodeGenPrepare::optimizeSelectInst with a more
339     // aggressive strategy of which instructions to sink.
340     //
341     // TODO: eliminate the redundancy of logic transforming selects to branches
342     // by removing CodeGenPrepare::optimizeSelectInst and optimizing here
343     // selects for all cases (with and without profile information).
344 
345     // Transform a sequence like this:
346     //    start:
347     //       %cmp = cmp uge i32 %a, %b
348     //       %sel = select i1 %cmp, i32 %c, i32 %d
349     //
350     // Into:
351     //    start:
352     //       %cmp = cmp uge i32 %a, %b
353     //       %cmp.frozen = freeze %cmp
354     //       br i1 %cmp.frozen, label %select.true, label %select.false
355     //    select.true:
356     //       br label %select.end
357     //    select.false:
358     //       br label %select.end
359     //    select.end:
360     //       %sel = phi i32 [ %c, %select.true ], [ %d, %select.false ]
361     //
362     // %cmp should be frozen, otherwise it may introduce undefined behavior.
363     // In addition, we may sink instructions that produce %c or %d into the
364     // destination(s) of the new branch.
365     // If the true or false blocks do not contain a sunken instruction, that
366     // block and its branch may be optimized away. In that case, one side of the
367     // first branch will point directly to select.end, and the corresponding PHI
368     // predecessor block will be the start block.
369 
370     // Find all the instructions that can be soundly sunk to the true/false
371     // blocks. These are instructions that are computed solely for producing the
372     // operands of the select instructions in the group and can be sunk without
373     // breaking the semantics of the LLVM IR (e.g., cannot sink instructions
374     // with side effects).
375     SmallVector<std::stack<Instruction *>, 2> TrueSlices, FalseSlices;
376     typedef std::stack<Instruction *>::size_type StackSizeType;
377     StackSizeType maxTrueSliceLen = 0, maxFalseSliceLen = 0;
378     for (SelectInst *SI : ASI) {
379       // For each select, compute the sinkable dependence chains of the true and
380       // false operands.
381       if (auto *TI = dyn_cast<Instruction>(SI->getTrueValue())) {
382         std::stack<Instruction *> TrueSlice;
383         getExclBackwardsSlice(TI, TrueSlice, SI, true);
384         maxTrueSliceLen = std::max(maxTrueSliceLen, TrueSlice.size());
385         TrueSlices.push_back(TrueSlice);
386       }
387       if (auto *FI = dyn_cast<Instruction>(SI->getFalseValue())) {
388         std::stack<Instruction *> FalseSlice;
389         getExclBackwardsSlice(FI, FalseSlice, SI, true);
390         maxFalseSliceLen = std::max(maxFalseSliceLen, FalseSlice.size());
391         FalseSlices.push_back(FalseSlice);
392       }
393     }
394     // In the case of multiple select instructions in the same group, the order
395     // of non-dependent instructions (instructions of different dependence
396     // slices) in the true/false blocks appears to affect performance.
397     // Interleaving the slices seems to experimentally be the optimal approach.
398     // This interleaving scheduling allows for more ILP (with a natural downside
399     // of increasing a bit register pressure) compared to a simple ordering of
400     // one whole chain after another. One would expect that this ordering would
401     // not matter since the scheduling in the backend of the compiler  would
402     // take care of it, but apparently the scheduler fails to deliver optimal
403     // ILP with a naive ordering here.
404     SmallVector<Instruction *, 2> TrueSlicesInterleaved, FalseSlicesInterleaved;
405     for (StackSizeType IS = 0; IS < maxTrueSliceLen; ++IS) {
406       for (auto &S : TrueSlices) {
407         if (!S.empty()) {
408           TrueSlicesInterleaved.push_back(S.top());
409           S.pop();
410         }
411       }
412     }
413     for (StackSizeType IS = 0; IS < maxFalseSliceLen; ++IS) {
414       for (auto &S : FalseSlices) {
415         if (!S.empty()) {
416           FalseSlicesInterleaved.push_back(S.top());
417           S.pop();
418         }
419       }
420     }
421 
422     // We split the block containing the select(s) into two blocks.
423     SelectInst *SI = ASI.front();
424     SelectInst *LastSI = ASI.back();
425     BasicBlock *StartBlock = SI->getParent();
426     BasicBlock::iterator SplitPt = ++(BasicBlock::iterator(LastSI));
427     BasicBlock *EndBlock = StartBlock->splitBasicBlock(SplitPt, "select.end");
428     BFI->setBlockFreq(EndBlock, BFI->getBlockFreq(StartBlock).getFrequency());
429     // Delete the unconditional branch that was just created by the split.
430     StartBlock->getTerminator()->eraseFromParent();
431 
432     // Move any debug/pseudo instructions that were in-between the select
433     // group to the newly-created end block.
434     SmallVector<Instruction *, 2> DebugPseudoINS;
435     auto DIt = SI->getIterator();
436     while (&*DIt != LastSI) {
437       if (DIt->isDebugOrPseudoInst())
438         DebugPseudoINS.push_back(&*DIt);
439       DIt++;
440     }
441     for (auto *DI : DebugPseudoINS) {
442       DI->moveBefore(&*EndBlock->getFirstInsertionPt());
443     }
444 
445     // These are the new basic blocks for the conditional branch.
446     // At least one will become an actual new basic block.
447     BasicBlock *TrueBlock = nullptr, *FalseBlock = nullptr;
448     BranchInst *TrueBranch = nullptr, *FalseBranch = nullptr;
449     if (!TrueSlicesInterleaved.empty()) {
450       TrueBlock = BasicBlock::Create(LastSI->getContext(), "select.true.sink",
451                                      EndBlock->getParent(), EndBlock);
452       TrueBranch = BranchInst::Create(EndBlock, TrueBlock);
453       TrueBranch->setDebugLoc(LastSI->getDebugLoc());
454       for (Instruction *TrueInst : TrueSlicesInterleaved)
455         TrueInst->moveBefore(TrueBranch);
456     }
457     if (!FalseSlicesInterleaved.empty()) {
458       FalseBlock = BasicBlock::Create(LastSI->getContext(), "select.false.sink",
459                                       EndBlock->getParent(), EndBlock);
460       FalseBranch = BranchInst::Create(EndBlock, FalseBlock);
461       FalseBranch->setDebugLoc(LastSI->getDebugLoc());
462       for (Instruction *FalseInst : FalseSlicesInterleaved)
463         FalseInst->moveBefore(FalseBranch);
464     }
465     // If there was nothing to sink, then arbitrarily choose the 'false' side
466     // for a new input value to the PHI.
467     if (TrueBlock == FalseBlock) {
468       assert(TrueBlock == nullptr &&
469              "Unexpected basic block transform while optimizing select");
470 
471       FalseBlock = BasicBlock::Create(SI->getContext(), "select.false",
472                                       EndBlock->getParent(), EndBlock);
473       auto *FalseBranch = BranchInst::Create(EndBlock, FalseBlock);
474       FalseBranch->setDebugLoc(SI->getDebugLoc());
475     }
476 
477     // Insert the real conditional branch based on the original condition.
478     // If we did not create a new block for one of the 'true' or 'false' paths
479     // of the condition, it means that side of the branch goes to the end block
480     // directly and the path originates from the start block from the point of
481     // view of the new PHI.
482     BasicBlock *TT, *FT;
483     if (TrueBlock == nullptr) {
484       TT = EndBlock;
485       FT = FalseBlock;
486       TrueBlock = StartBlock;
487     } else if (FalseBlock == nullptr) {
488       TT = TrueBlock;
489       FT = EndBlock;
490       FalseBlock = StartBlock;
491     } else {
492       TT = TrueBlock;
493       FT = FalseBlock;
494     }
495     IRBuilder<> IB(SI);
496     auto *CondFr =
497         IB.CreateFreeze(SI->getCondition(), SI->getName() + ".frozen");
498     IB.CreateCondBr(CondFr, TT, FT, SI);
499 
500     SmallPtrSet<const Instruction *, 2> INS;
501     INS.insert(ASI.begin(), ASI.end());
502     // Use reverse iterator because later select may use the value of the
503     // earlier select, and we need to propagate value through earlier select
504     // to get the PHI operand.
505     for (auto It = ASI.rbegin(); It != ASI.rend(); ++It) {
506       SelectInst *SI = *It;
507       // The select itself is replaced with a PHI Node.
508       PHINode *PN = PHINode::Create(SI->getType(), 2, "", &EndBlock->front());
509       PN->takeName(SI);
510       PN->addIncoming(getTrueOrFalseValue(SI, true, INS), TrueBlock);
511       PN->addIncoming(getTrueOrFalseValue(SI, false, INS), FalseBlock);
512       PN->setDebugLoc(SI->getDebugLoc());
513 
514       SI->replaceAllUsesWith(PN);
515       SI->eraseFromParent();
516       INS.erase(SI);
517       ++NumSelectsConverted;
518     }
519   }
520 }
521 
522 static bool isSpecialSelect(SelectInst *SI) {
523   using namespace llvm::PatternMatch;
524 
525   // If the select is a logical-and/logical-or then it is better treated as a
526   // and/or by the backend.
527   if (match(SI, m_CombineOr(m_LogicalAnd(m_Value(), m_Value()),
528                             m_LogicalOr(m_Value(), m_Value()))))
529     return true;
530 
531   return false;
532 }
533 
534 void SelectOptimize::collectSelectGroups(BasicBlock &BB,
535                                          SelectGroups &SIGroups) {
536   BasicBlock::iterator BBIt = BB.begin();
537   while (BBIt != BB.end()) {
538     Instruction *I = &*BBIt++;
539     if (SelectInst *SI = dyn_cast<SelectInst>(I)) {
540       if (isSpecialSelect(SI))
541         continue;
542 
543       SelectGroup SIGroup;
544       SIGroup.push_back(SI);
545       while (BBIt != BB.end()) {
546         Instruction *NI = &*BBIt;
547         SelectInst *NSI = dyn_cast<SelectInst>(NI);
548         if (NSI && SI->getCondition() == NSI->getCondition()) {
549           SIGroup.push_back(NSI);
550         } else if (!NI->isDebugOrPseudoInst()) {
551           // Debug/pseudo instructions should be skipped and not prevent the
552           // formation of a select group.
553           break;
554         }
555         ++BBIt;
556       }
557 
558       // If the select type is not supported, no point optimizing it.
559       // Instruction selection will take care of it.
560       if (!isSelectKindSupported(SI))
561         continue;
562 
563       SIGroups.push_back(SIGroup);
564     }
565   }
566 }
567 
568 void SelectOptimize::findProfitableSIGroupsBase(SelectGroups &SIGroups,
569                                                 SelectGroups &ProfSIGroups) {
570   for (SelectGroup &ASI : SIGroups) {
571     ++NumSelectOptAnalyzed;
572     if (isConvertToBranchProfitableBase(ASI))
573       ProfSIGroups.push_back(ASI);
574   }
575 }
576 
577 static void EmitAndPrintRemark(OptimizationRemarkEmitter *ORE,
578                                DiagnosticInfoOptimizationBase &Rem) {
579   LLVM_DEBUG(dbgs() << Rem.getMsg() << "\n");
580   ORE->emit(Rem);
581 }
582 
583 void SelectOptimize::findProfitableSIGroupsInnerLoops(
584     const Loop *L, SelectGroups &SIGroups, SelectGroups &ProfSIGroups) {
585   NumSelectOptAnalyzed += SIGroups.size();
586   // For each select group in an inner-most loop,
587   // a branch is more preferable than a select/conditional-move if:
588   // i) conversion to branches for all the select groups of the loop satisfies
589   //    loop-level heuristics including reducing the loop's critical path by
590   //    some threshold (see SelectOptimize::checkLoopHeuristics); and
591   // ii) the total cost of the select group is cheaper with a branch compared
592   //     to its predicated version. The cost is in terms of latency and the cost
593   //     of a select group is the cost of its most expensive select instruction
594   //     (assuming infinite resources and thus fully leveraging available ILP).
595 
596   DenseMap<const Instruction *, CostInfo> InstCostMap;
597   CostInfo LoopCost[2] = {{Scaled64::getZero(), Scaled64::getZero()},
598                           {Scaled64::getZero(), Scaled64::getZero()}};
599   if (!computeLoopCosts(L, SIGroups, InstCostMap, LoopCost) ||
600       !checkLoopHeuristics(L, LoopCost)) {
601     return;
602   }
603 
604   for (SelectGroup &ASI : SIGroups) {
605     // Assuming infinite resources, the cost of a group of instructions is the
606     // cost of the most expensive instruction of the group.
607     Scaled64 SelectCost = Scaled64::getZero(), BranchCost = Scaled64::getZero();
608     for (SelectInst *SI : ASI) {
609       SelectCost = std::max(SelectCost, InstCostMap[SI].PredCost);
610       BranchCost = std::max(BranchCost, InstCostMap[SI].NonPredCost);
611     }
612     if (BranchCost < SelectCost) {
613       OptimizationRemark OR(DEBUG_TYPE, "SelectOpti", ASI.front());
614       OR << "Profitable to convert to branch (loop analysis). BranchCost="
615          << BranchCost.toString() << ", SelectCost=" << SelectCost.toString()
616          << ". ";
617       EmitAndPrintRemark(ORE, OR);
618       ++NumSelectConvertedLoop;
619       ProfSIGroups.push_back(ASI);
620     } else {
621       OptimizationRemarkMissed ORmiss(DEBUG_TYPE, "SelectOpti", ASI.front());
622       ORmiss << "Select is more profitable (loop analysis). BranchCost="
623              << BranchCost.toString()
624              << ", SelectCost=" << SelectCost.toString() << ". ";
625       EmitAndPrintRemark(ORE, ORmiss);
626     }
627   }
628 }
629 
630 bool SelectOptimize::isConvertToBranchProfitableBase(
631     const SmallVector<SelectInst *, 2> &ASI) {
632   SelectInst *SI = ASI.front();
633   LLVM_DEBUG(dbgs() << "Analyzing select group containing " << *SI << "\n");
634   OptimizationRemark OR(DEBUG_TYPE, "SelectOpti", SI);
635   OptimizationRemarkMissed ORmiss(DEBUG_TYPE, "SelectOpti", SI);
636 
637   // Skip cold basic blocks. Better to optimize for size for cold blocks.
638   if (PSI->isColdBlock(SI->getParent(), BFI.get())) {
639     ++NumSelectColdBB;
640     ORmiss << "Not converted to branch because of cold basic block. ";
641     EmitAndPrintRemark(ORE, ORmiss);
642     return false;
643   }
644 
645   // If unpredictable, branch form is less profitable.
646   if (SI->getMetadata(LLVMContext::MD_unpredictable)) {
647     ++NumSelectUnPred;
648     ORmiss << "Not converted to branch because of unpredictable branch. ";
649     EmitAndPrintRemark(ORE, ORmiss);
650     return false;
651   }
652 
653   // If highly predictable, branch form is more profitable, unless a
654   // predictable select is inexpensive in the target architecture.
655   if (isSelectHighlyPredictable(SI) && TLI->isPredictableSelectExpensive()) {
656     ++NumSelectConvertedHighPred;
657     OR << "Converted to branch because of highly predictable branch. ";
658     EmitAndPrintRemark(ORE, OR);
659     return true;
660   }
661 
662   // Look for expensive instructions in the cold operand's (if any) dependence
663   // slice of any of the selects in the group.
664   if (hasExpensiveColdOperand(ASI)) {
665     ++NumSelectConvertedExpColdOperand;
666     OR << "Converted to branch because of expensive cold operand.";
667     EmitAndPrintRemark(ORE, OR);
668     return true;
669   }
670 
671   ORmiss << "Not profitable to convert to branch (base heuristic).";
672   EmitAndPrintRemark(ORE, ORmiss);
673   return false;
674 }
675 
676 static InstructionCost divideNearest(InstructionCost Numerator,
677                                      uint64_t Denominator) {
678   return (Numerator + (Denominator / 2)) / Denominator;
679 }
680 
681 bool SelectOptimize::hasExpensiveColdOperand(
682     const SmallVector<SelectInst *, 2> &ASI) {
683   bool ColdOperand = false;
684   uint64_t TrueWeight, FalseWeight, TotalWeight;
685   if (extractBranchWeights(*ASI.front(), TrueWeight, FalseWeight)) {
686     uint64_t MinWeight = std::min(TrueWeight, FalseWeight);
687     TotalWeight = TrueWeight + FalseWeight;
688     // Is there a path with frequency <ColdOperandThreshold% (default:20%) ?
689     ColdOperand = TotalWeight * ColdOperandThreshold > 100 * MinWeight;
690   } else if (PSI->hasProfileSummary()) {
691     OptimizationRemarkMissed ORmiss(DEBUG_TYPE, "SelectOpti", ASI.front());
692     ORmiss << "Profile data available but missing branch-weights metadata for "
693               "select instruction. ";
694     EmitAndPrintRemark(ORE, ORmiss);
695   }
696   if (!ColdOperand)
697     return false;
698   // Check if the cold path's dependence slice is expensive for any of the
699   // selects of the group.
700   for (SelectInst *SI : ASI) {
701     Instruction *ColdI = nullptr;
702     uint64_t HotWeight;
703     if (TrueWeight < FalseWeight) {
704       ColdI = dyn_cast<Instruction>(SI->getTrueValue());
705       HotWeight = FalseWeight;
706     } else {
707       ColdI = dyn_cast<Instruction>(SI->getFalseValue());
708       HotWeight = TrueWeight;
709     }
710     if (ColdI) {
711       std::stack<Instruction *> ColdSlice;
712       getExclBackwardsSlice(ColdI, ColdSlice, SI);
713       InstructionCost SliceCost = 0;
714       while (!ColdSlice.empty()) {
715         SliceCost += TTI->getInstructionCost(ColdSlice.top(),
716                                              TargetTransformInfo::TCK_Latency);
717         ColdSlice.pop();
718       }
719       // The colder the cold value operand of the select is the more expensive
720       // the cmov becomes for computing the cold value operand every time. Thus,
721       // the colder the cold operand is the more its cost counts.
722       // Get nearest integer cost adjusted for coldness.
723       InstructionCost AdjSliceCost =
724           divideNearest(SliceCost * HotWeight, TotalWeight);
725       if (AdjSliceCost >=
726           ColdOperandMaxCostMultiplier * TargetTransformInfo::TCC_Expensive)
727         return true;
728     }
729   }
730   return false;
731 }
732 
733 // Check if it is safe to move LoadI next to the SI.
734 // Conservatively assume it is safe only if there is no instruction
735 // modifying memory in-between the load and the select instruction.
736 static bool isSafeToSinkLoad(Instruction *LoadI, Instruction *SI) {
737   // Assume loads from different basic blocks are unsafe to move.
738   if (LoadI->getParent() != SI->getParent())
739     return false;
740   auto It = LoadI->getIterator();
741   while (&*It != SI) {
742     if (It->mayWriteToMemory())
743       return false;
744     It++;
745   }
746   return true;
747 }
748 
749 // For a given source instruction, collect its backwards dependence slice
750 // consisting of instructions exclusively computed for the purpose of producing
751 // the operands of the source instruction. As an approximation
752 // (sufficiently-accurate in practice), we populate this set with the
753 // instructions of the backwards dependence slice that only have one-use and
754 // form an one-use chain that leads to the source instruction.
755 void SelectOptimize::getExclBackwardsSlice(Instruction *I,
756                                            std::stack<Instruction *> &Slice,
757                                            Instruction *SI, bool ForSinking) {
758   SmallPtrSet<Instruction *, 2> Visited;
759   std::queue<Instruction *> Worklist;
760   Worklist.push(I);
761   while (!Worklist.empty()) {
762     Instruction *II = Worklist.front();
763     Worklist.pop();
764 
765     // Avoid cycles.
766     if (!Visited.insert(II).second)
767       continue;
768 
769     if (!II->hasOneUse())
770       continue;
771 
772     // Cannot soundly sink instructions with side-effects.
773     // Terminator or phi instructions cannot be sunk.
774     // Avoid sinking other select instructions (should be handled separetely).
775     if (ForSinking && (II->isTerminator() || II->mayHaveSideEffects() ||
776                        isa<SelectInst>(II) || isa<PHINode>(II)))
777       continue;
778 
779     // Avoid sinking loads in order not to skip state-modifying instructions,
780     // that may alias with the loaded address.
781     // Only allow sinking of loads within the same basic block that are
782     // conservatively proven to be safe.
783     if (ForSinking && II->mayReadFromMemory() && !isSafeToSinkLoad(II, SI))
784       continue;
785 
786     // Avoid considering instructions with less frequency than the source
787     // instruction (i.e., avoid colder code regions of the dependence slice).
788     if (BFI->getBlockFreq(II->getParent()) < BFI->getBlockFreq(I->getParent()))
789       continue;
790 
791     // Eligible one-use instruction added to the dependence slice.
792     Slice.push(II);
793 
794     // Explore all the operands of the current instruction to expand the slice.
795     for (unsigned k = 0; k < II->getNumOperands(); ++k)
796       if (auto *OpI = dyn_cast<Instruction>(II->getOperand(k)))
797         Worklist.push(OpI);
798   }
799 }
800 
801 bool SelectOptimize::isSelectHighlyPredictable(const SelectInst *SI) {
802   uint64_t TrueWeight, FalseWeight;
803   if (extractBranchWeights(*SI, TrueWeight, FalseWeight)) {
804     uint64_t Max = std::max(TrueWeight, FalseWeight);
805     uint64_t Sum = TrueWeight + FalseWeight;
806     if (Sum != 0) {
807       auto Probability = BranchProbability::getBranchProbability(Max, Sum);
808       if (Probability > TTI->getPredictableBranchThreshold())
809         return true;
810     }
811   }
812   return false;
813 }
814 
815 bool SelectOptimize::checkLoopHeuristics(const Loop *L,
816                                          const CostInfo LoopCost[2]) {
817   // Loop-level checks to determine if a non-predicated version (with branches)
818   // of the loop is more profitable than its predicated version.
819 
820   if (DisableLoopLevelHeuristics)
821     return true;
822 
823   OptimizationRemarkMissed ORmissL(DEBUG_TYPE, "SelectOpti",
824                                    L->getHeader()->getFirstNonPHI());
825 
826   if (LoopCost[0].NonPredCost > LoopCost[0].PredCost ||
827       LoopCost[1].NonPredCost >= LoopCost[1].PredCost) {
828     ORmissL << "No select conversion in the loop due to no reduction of loop's "
829                "critical path. ";
830     EmitAndPrintRemark(ORE, ORmissL);
831     return false;
832   }
833 
834   Scaled64 Gain[2] = {LoopCost[0].PredCost - LoopCost[0].NonPredCost,
835                       LoopCost[1].PredCost - LoopCost[1].NonPredCost};
836 
837   // Profitably converting to branches need to reduce the loop's critical path
838   // by at least some threshold (absolute gain of GainCycleThreshold cycles and
839   // relative gain of 12.5%).
840   if (Gain[1] < Scaled64::get(GainCycleThreshold) ||
841       Gain[1] * Scaled64::get(GainRelativeThreshold) < LoopCost[1].PredCost) {
842     Scaled64 RelativeGain = Scaled64::get(100) * Gain[1] / LoopCost[1].PredCost;
843     ORmissL << "No select conversion in the loop due to small reduction of "
844                "loop's critical path. Gain="
845             << Gain[1].toString()
846             << ", RelativeGain=" << RelativeGain.toString() << "%. ";
847     EmitAndPrintRemark(ORE, ORmissL);
848     return false;
849   }
850 
851   // If the loop's critical path involves loop-carried dependences, the gradient
852   // of the gain needs to be at least GainGradientThreshold% (defaults to 25%).
853   // This check ensures that the latency reduction for the loop's critical path
854   // keeps decreasing with sufficient rate beyond the two analyzed loop
855   // iterations.
856   if (Gain[1] > Gain[0]) {
857     Scaled64 GradientGain = Scaled64::get(100) * (Gain[1] - Gain[0]) /
858                             (LoopCost[1].PredCost - LoopCost[0].PredCost);
859     if (GradientGain < Scaled64::get(GainGradientThreshold)) {
860       ORmissL << "No select conversion in the loop due to small gradient gain. "
861                  "GradientGain="
862               << GradientGain.toString() << "%. ";
863       EmitAndPrintRemark(ORE, ORmissL);
864       return false;
865     }
866   }
867   // If the gain decreases it is not profitable to convert.
868   else if (Gain[1] < Gain[0]) {
869     ORmissL
870         << "No select conversion in the loop due to negative gradient gain. ";
871     EmitAndPrintRemark(ORE, ORmissL);
872     return false;
873   }
874 
875   // Non-predicated version of the loop is more profitable than its
876   // predicated version.
877   return true;
878 }
879 
880 // Computes instruction and loop-critical-path costs for both the predicated
881 // and non-predicated version of the given loop.
882 // Returns false if unable to compute these costs due to invalid cost of loop
883 // instruction(s).
884 bool SelectOptimize::computeLoopCosts(
885     const Loop *L, const SelectGroups &SIGroups,
886     DenseMap<const Instruction *, CostInfo> &InstCostMap, CostInfo *LoopCost) {
887   LLVM_DEBUG(dbgs() << "Calculating Latency / IPredCost / INonPredCost of loop "
888                     << L->getHeader()->getName() << "\n");
889   const auto &SIset = getSIset(SIGroups);
890   // Compute instruction and loop-critical-path costs across two iterations for
891   // both predicated and non-predicated version.
892   const unsigned Iterations = 2;
893   for (unsigned Iter = 0; Iter < Iterations; ++Iter) {
894     // Cost of the loop's critical path.
895     CostInfo &MaxCost = LoopCost[Iter];
896     for (BasicBlock *BB : L->getBlocks()) {
897       for (const Instruction &I : *BB) {
898         if (I.isDebugOrPseudoInst())
899           continue;
900         // Compute the predicated and non-predicated cost of the instruction.
901         Scaled64 IPredCost = Scaled64::getZero(),
902                  INonPredCost = Scaled64::getZero();
903 
904         // Assume infinite resources that allow to fully exploit the available
905         // instruction-level parallelism.
906         // InstCost = InstLatency + max(Op1Cost, Op2Cost, … OpNCost)
907         for (const Use &U : I.operands()) {
908           auto UI = dyn_cast<Instruction>(U.get());
909           if (!UI)
910             continue;
911           if (InstCostMap.count(UI)) {
912             IPredCost = std::max(IPredCost, InstCostMap[UI].PredCost);
913             INonPredCost = std::max(INonPredCost, InstCostMap[UI].NonPredCost);
914           }
915         }
916         auto ILatency = computeInstCost(&I);
917         if (!ILatency) {
918           OptimizationRemarkMissed ORmissL(DEBUG_TYPE, "SelectOpti", &I);
919           ORmissL << "Invalid instruction cost preventing analysis and "
920                      "optimization of the inner-most loop containing this "
921                      "instruction. ";
922           EmitAndPrintRemark(ORE, ORmissL);
923           return false;
924         }
925         IPredCost += Scaled64::get(*ILatency);
926         INonPredCost += Scaled64::get(*ILatency);
927 
928         // For a select that can be converted to branch,
929         // compute its cost as a branch (non-predicated cost).
930         //
931         // BranchCost = PredictedPathCost + MispredictCost
932         // PredictedPathCost = TrueOpCost * TrueProb + FalseOpCost * FalseProb
933         // MispredictCost = max(MispredictPenalty, CondCost) * MispredictRate
934         if (SIset.contains(&I)) {
935           auto SI = cast<SelectInst>(&I);
936 
937           Scaled64 TrueOpCost = Scaled64::getZero(),
938                    FalseOpCost = Scaled64::getZero();
939           if (auto *TI = dyn_cast<Instruction>(SI->getTrueValue()))
940             if (InstCostMap.count(TI))
941               TrueOpCost = InstCostMap[TI].NonPredCost;
942           if (auto *FI = dyn_cast<Instruction>(SI->getFalseValue()))
943             if (InstCostMap.count(FI))
944               FalseOpCost = InstCostMap[FI].NonPredCost;
945           Scaled64 PredictedPathCost =
946               getPredictedPathCost(TrueOpCost, FalseOpCost, SI);
947 
948           Scaled64 CondCost = Scaled64::getZero();
949           if (auto *CI = dyn_cast<Instruction>(SI->getCondition()))
950             if (InstCostMap.count(CI))
951               CondCost = InstCostMap[CI].NonPredCost;
952           Scaled64 MispredictCost = getMispredictionCost(SI, CondCost);
953 
954           INonPredCost = PredictedPathCost + MispredictCost;
955         }
956         LLVM_DEBUG(dbgs() << " " << ILatency << "/" << IPredCost << "/"
957                           << INonPredCost << " for " << I << "\n");
958 
959         InstCostMap[&I] = {IPredCost, INonPredCost};
960         MaxCost.PredCost = std::max(MaxCost.PredCost, IPredCost);
961         MaxCost.NonPredCost = std::max(MaxCost.NonPredCost, INonPredCost);
962       }
963     }
964     LLVM_DEBUG(dbgs() << "Iteration " << Iter + 1
965                       << " MaxCost = " << MaxCost.PredCost << " "
966                       << MaxCost.NonPredCost << "\n");
967   }
968   return true;
969 }
970 
971 SmallPtrSet<const Instruction *, 2>
972 SelectOptimize::getSIset(const SelectGroups &SIGroups) {
973   SmallPtrSet<const Instruction *, 2> SIset;
974   for (const SelectGroup &ASI : SIGroups)
975     for (const SelectInst *SI : ASI)
976       SIset.insert(SI);
977   return SIset;
978 }
979 
980 std::optional<uint64_t> SelectOptimize::computeInstCost(const Instruction *I) {
981   InstructionCost ICost =
982       TTI->getInstructionCost(I, TargetTransformInfo::TCK_Latency);
983   if (auto OC = ICost.getValue())
984     return std::optional<uint64_t>(*OC);
985   return std::nullopt;
986 }
987 
988 ScaledNumber<uint64_t>
989 SelectOptimize::getMispredictionCost(const SelectInst *SI,
990                                      const Scaled64 CondCost) {
991   uint64_t MispredictPenalty = TSchedModel.getMCSchedModel()->MispredictPenalty;
992 
993   // Account for the default misprediction rate when using a branch
994   // (conservatively set to 25% by default).
995   uint64_t MispredictRate = MispredictDefaultRate;
996   // If the select condition is obviously predictable, then the misprediction
997   // rate is zero.
998   if (isSelectHighlyPredictable(SI))
999     MispredictRate = 0;
1000 
1001   // CondCost is included to account for cases where the computation of the
1002   // condition is part of a long dependence chain (potentially loop-carried)
1003   // that would delay detection of a misprediction and increase its cost.
1004   Scaled64 MispredictCost =
1005       std::max(Scaled64::get(MispredictPenalty), CondCost) *
1006       Scaled64::get(MispredictRate);
1007   MispredictCost /= Scaled64::get(100);
1008 
1009   return MispredictCost;
1010 }
1011 
1012 // Returns the cost of a branch when the prediction is correct.
1013 // TrueCost * TrueProbability + FalseCost * FalseProbability.
1014 ScaledNumber<uint64_t>
1015 SelectOptimize::getPredictedPathCost(Scaled64 TrueCost, Scaled64 FalseCost,
1016                                      const SelectInst *SI) {
1017   Scaled64 PredPathCost;
1018   uint64_t TrueWeight, FalseWeight;
1019   if (extractBranchWeights(*SI, TrueWeight, FalseWeight)) {
1020     uint64_t SumWeight = TrueWeight + FalseWeight;
1021     if (SumWeight != 0) {
1022       PredPathCost = TrueCost * Scaled64::get(TrueWeight) +
1023                      FalseCost * Scaled64::get(FalseWeight);
1024       PredPathCost /= Scaled64::get(SumWeight);
1025       return PredPathCost;
1026     }
1027   }
1028   // Without branch weight metadata, we assume 75% for the one path and 25% for
1029   // the other, and pick the result with the biggest cost.
1030   PredPathCost = std::max(TrueCost * Scaled64::get(3) + FalseCost,
1031                           FalseCost * Scaled64::get(3) + TrueCost);
1032   PredPathCost /= Scaled64::get(4);
1033   return PredPathCost;
1034 }
1035 
1036 bool SelectOptimize::isSelectKindSupported(SelectInst *SI) {
1037   bool VectorCond = !SI->getCondition()->getType()->isIntegerTy(1);
1038   if (VectorCond)
1039     return false;
1040   TargetLowering::SelectSupportKind SelectKind;
1041   if (SI->getType()->isVectorTy())
1042     SelectKind = TargetLowering::ScalarCondVectorVal;
1043   else
1044     SelectKind = TargetLowering::ScalarValSelect;
1045   return TLI->isSelectSupported(SelectKind);
1046 }
1047