1 | //===- SimplifyCFG.cpp - Code to perform CFG simplification ---------------===// |
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 | // Peephole optimize the CFG. |
10 | // |
11 | //===----------------------------------------------------------------------===// |
12 | |
13 | #include "llvm/ADT/APInt.h" |
14 | #include "llvm/ADT/ArrayRef.h" |
15 | #include "llvm/ADT/DenseMap.h" |
16 | #include "llvm/ADT/MapVector.h" |
17 | #include "llvm/ADT/STLExtras.h" |
18 | #include "llvm/ADT/Sequence.h" |
19 | #include "llvm/ADT/SetOperations.h" |
20 | #include "llvm/ADT/SetVector.h" |
21 | #include "llvm/ADT/SmallPtrSet.h" |
22 | #include "llvm/ADT/SmallVector.h" |
23 | #include "llvm/ADT/Statistic.h" |
24 | #include "llvm/ADT/StringRef.h" |
25 | #include "llvm/Analysis/AssumptionCache.h" |
26 | #include "llvm/Analysis/CaptureTracking.h" |
27 | #include "llvm/Analysis/ConstantFolding.h" |
28 | #include "llvm/Analysis/DomTreeUpdater.h" |
29 | #include "llvm/Analysis/GuardUtils.h" |
30 | #include "llvm/Analysis/InstructionSimplify.h" |
31 | #include "llvm/Analysis/MemorySSA.h" |
32 | #include "llvm/Analysis/MemorySSAUpdater.h" |
33 | #include "llvm/Analysis/TargetTransformInfo.h" |
34 | #include "llvm/Analysis/ValueTracking.h" |
35 | #include "llvm/IR/Attributes.h" |
36 | #include "llvm/IR/BasicBlock.h" |
37 | #include "llvm/IR/CFG.h" |
38 | #include "llvm/IR/Constant.h" |
39 | #include "llvm/IR/ConstantRange.h" |
40 | #include "llvm/IR/Constants.h" |
41 | #include "llvm/IR/DataLayout.h" |
42 | #include "llvm/IR/DebugInfo.h" |
43 | #include "llvm/IR/DerivedTypes.h" |
44 | #include "llvm/IR/Function.h" |
45 | #include "llvm/IR/GlobalValue.h" |
46 | #include "llvm/IR/GlobalVariable.h" |
47 | #include "llvm/IR/IRBuilder.h" |
48 | #include "llvm/IR/InstrTypes.h" |
49 | #include "llvm/IR/Instruction.h" |
50 | #include "llvm/IR/Instructions.h" |
51 | #include "llvm/IR/IntrinsicInst.h" |
52 | #include "llvm/IR/LLVMContext.h" |
53 | #include "llvm/IR/MDBuilder.h" |
54 | #include "llvm/IR/MemoryModelRelaxationAnnotations.h" |
55 | #include "llvm/IR/Metadata.h" |
56 | #include "llvm/IR/Module.h" |
57 | #include "llvm/IR/NoFolder.h" |
58 | #include "llvm/IR/Operator.h" |
59 | #include "llvm/IR/PatternMatch.h" |
60 | #include "llvm/IR/ProfDataUtils.h" |
61 | #include "llvm/IR/Type.h" |
62 | #include "llvm/IR/Use.h" |
63 | #include "llvm/IR/User.h" |
64 | #include "llvm/IR/Value.h" |
65 | #include "llvm/IR/ValueHandle.h" |
66 | #include "llvm/Support/BranchProbability.h" |
67 | #include "llvm/Support/Casting.h" |
68 | #include "llvm/Support/CommandLine.h" |
69 | #include "llvm/Support/Debug.h" |
70 | #include "llvm/Support/ErrorHandling.h" |
71 | #include "llvm/Support/KnownBits.h" |
72 | #include "llvm/Support/MathExtras.h" |
73 | #include "llvm/Support/raw_ostream.h" |
74 | #include "llvm/Transforms/Utils/BasicBlockUtils.h" |
75 | #include "llvm/Transforms/Utils/Local.h" |
76 | #include "llvm/Transforms/Utils/ValueMapper.h" |
77 | #include <algorithm> |
78 | #include <cassert> |
79 | #include <climits> |
80 | #include <cstddef> |
81 | #include <cstdint> |
82 | #include <iterator> |
83 | #include <map> |
84 | #include <optional> |
85 | #include <set> |
86 | #include <tuple> |
87 | #include <utility> |
88 | #include <vector> |
89 | |
90 | using namespace llvm; |
91 | using namespace PatternMatch; |
92 | |
93 | #define DEBUG_TYPE "simplifycfg" |
94 | |
95 | cl::opt<bool> llvm::RequireAndPreserveDomTree( |
96 | "simplifycfg-require-and-preserve-domtree" , cl::Hidden, |
97 | |
98 | cl::desc("Temorary development switch used to gradually uplift SimplifyCFG " |
99 | "into preserving DomTree," )); |
100 | |
101 | // Chosen as 2 so as to be cheap, but still to have enough power to fold |
102 | // a select, so the "clamp" idiom (of a min followed by a max) will be caught. |
103 | // To catch this, we need to fold a compare and a select, hence '2' being the |
104 | // minimum reasonable default. |
105 | static cl::opt<unsigned> PHINodeFoldingThreshold( |
106 | "phi-node-folding-threshold" , cl::Hidden, cl::init(Val: 2), |
107 | cl::desc( |
108 | "Control the amount of phi node folding to perform (default = 2)" )); |
109 | |
110 | static cl::opt<unsigned> TwoEntryPHINodeFoldingThreshold( |
111 | "two-entry-phi-node-folding-threshold" , cl::Hidden, cl::init(Val: 4), |
112 | cl::desc("Control the maximal total instruction cost that we are willing " |
113 | "to speculatively execute to fold a 2-entry PHI node into a " |
114 | "select (default = 4)" )); |
115 | |
116 | static cl::opt<bool> |
117 | HoistCommon("simplifycfg-hoist-common" , cl::Hidden, cl::init(Val: true), |
118 | cl::desc("Hoist common instructions up to the parent block" )); |
119 | |
120 | static cl::opt<unsigned> |
121 | HoistCommonSkipLimit("simplifycfg-hoist-common-skip-limit" , cl::Hidden, |
122 | cl::init(Val: 20), |
123 | cl::desc("Allow reordering across at most this many " |
124 | "instructions when hoisting" )); |
125 | |
126 | static cl::opt<bool> |
127 | SinkCommon("simplifycfg-sink-common" , cl::Hidden, cl::init(Val: true), |
128 | cl::desc("Sink common instructions down to the end block" )); |
129 | |
130 | static cl::opt<bool> HoistCondStores( |
131 | "simplifycfg-hoist-cond-stores" , cl::Hidden, cl::init(Val: true), |
132 | cl::desc("Hoist conditional stores if an unconditional store precedes" )); |
133 | |
134 | static cl::opt<bool> MergeCondStores( |
135 | "simplifycfg-merge-cond-stores" , cl::Hidden, cl::init(Val: true), |
136 | cl::desc("Hoist conditional stores even if an unconditional store does not " |
137 | "precede - hoist multiple conditional stores into a single " |
138 | "predicated store" )); |
139 | |
140 | static cl::opt<bool> MergeCondStoresAggressively( |
141 | "simplifycfg-merge-cond-stores-aggressively" , cl::Hidden, cl::init(Val: false), |
142 | cl::desc("When merging conditional stores, do so even if the resultant " |
143 | "basic blocks are unlikely to be if-converted as a result" )); |
144 | |
145 | static cl::opt<bool> SpeculateOneExpensiveInst( |
146 | "speculate-one-expensive-inst" , cl::Hidden, cl::init(Val: true), |
147 | cl::desc("Allow exactly one expensive instruction to be speculatively " |
148 | "executed" )); |
149 | |
150 | static cl::opt<unsigned> MaxSpeculationDepth( |
151 | "max-speculation-depth" , cl::Hidden, cl::init(Val: 10), |
152 | cl::desc("Limit maximum recursion depth when calculating costs of " |
153 | "speculatively executed instructions" )); |
154 | |
155 | static cl::opt<int> |
156 | MaxSmallBlockSize("simplifycfg-max-small-block-size" , cl::Hidden, |
157 | cl::init(Val: 10), |
158 | cl::desc("Max size of a block which is still considered " |
159 | "small enough to thread through" )); |
160 | |
161 | // Two is chosen to allow one negation and a logical combine. |
162 | static cl::opt<unsigned> |
163 | BranchFoldThreshold("simplifycfg-branch-fold-threshold" , cl::Hidden, |
164 | cl::init(Val: 2), |
165 | cl::desc("Maximum cost of combining conditions when " |
166 | "folding branches" )); |
167 | |
168 | static cl::opt<unsigned> BranchFoldToCommonDestVectorMultiplier( |
169 | "simplifycfg-branch-fold-common-dest-vector-multiplier" , cl::Hidden, |
170 | cl::init(Val: 2), |
171 | cl::desc("Multiplier to apply to threshold when determining whether or not " |
172 | "to fold branch to common destination when vector operations are " |
173 | "present" )); |
174 | |
175 | static cl::opt<bool> EnableMergeCompatibleInvokes( |
176 | "simplifycfg-merge-compatible-invokes" , cl::Hidden, cl::init(Val: true), |
177 | cl::desc("Allow SimplifyCFG to merge invokes together when appropriate" )); |
178 | |
179 | static cl::opt<unsigned> MaxSwitchCasesPerResult( |
180 | "max-switch-cases-per-result" , cl::Hidden, cl::init(Val: 16), |
181 | cl::desc("Limit cases to analyze when converting a switch to select" )); |
182 | |
183 | STATISTIC(NumBitMaps, "Number of switch instructions turned into bitmaps" ); |
184 | STATISTIC(NumLinearMaps, |
185 | "Number of switch instructions turned into linear mapping" ); |
186 | STATISTIC(NumLookupTables, |
187 | "Number of switch instructions turned into lookup tables" ); |
188 | STATISTIC( |
189 | NumLookupTablesHoles, |
190 | "Number of switch instructions turned into lookup tables (holes checked)" ); |
191 | STATISTIC(NumTableCmpReuses, "Number of reused switch table lookup compares" ); |
192 | STATISTIC(NumFoldValueComparisonIntoPredecessors, |
193 | "Number of value comparisons folded into predecessor basic blocks" ); |
194 | STATISTIC(NumFoldBranchToCommonDest, |
195 | "Number of branches folded into predecessor basic block" ); |
196 | STATISTIC( |
197 | NumHoistCommonCode, |
198 | "Number of common instruction 'blocks' hoisted up to the begin block" ); |
199 | STATISTIC(NumHoistCommonInstrs, |
200 | "Number of common instructions hoisted up to the begin block" ); |
201 | STATISTIC(NumSinkCommonCode, |
202 | "Number of common instruction 'blocks' sunk down to the end block" ); |
203 | STATISTIC(NumSinkCommonInstrs, |
204 | "Number of common instructions sunk down to the end block" ); |
205 | STATISTIC(NumSpeculations, "Number of speculative executed instructions" ); |
206 | STATISTIC(NumInvokes, |
207 | "Number of invokes with empty resume blocks simplified into calls" ); |
208 | STATISTIC(NumInvokesMerged, "Number of invokes that were merged together" ); |
209 | STATISTIC(NumInvokeSetsFormed, "Number of invoke sets that were formed" ); |
210 | |
211 | namespace { |
212 | |
213 | // The first field contains the value that the switch produces when a certain |
214 | // case group is selected, and the second field is a vector containing the |
215 | // cases composing the case group. |
216 | using SwitchCaseResultVectorTy = |
217 | SmallVector<std::pair<Constant *, SmallVector<ConstantInt *, 4>>, 2>; |
218 | |
219 | // The first field contains the phi node that generates a result of the switch |
220 | // and the second field contains the value generated for a certain case in the |
221 | // switch for that PHI. |
222 | using SwitchCaseResultsTy = SmallVector<std::pair<PHINode *, Constant *>, 4>; |
223 | |
224 | /// ValueEqualityComparisonCase - Represents a case of a switch. |
225 | struct ValueEqualityComparisonCase { |
226 | ConstantInt *Value; |
227 | BasicBlock *Dest; |
228 | |
229 | ValueEqualityComparisonCase(ConstantInt *Value, BasicBlock *Dest) |
230 | : Value(Value), Dest(Dest) {} |
231 | |
232 | bool operator<(ValueEqualityComparisonCase RHS) const { |
233 | // Comparing pointers is ok as we only rely on the order for uniquing. |
234 | return Value < RHS.Value; |
235 | } |
236 | |
237 | bool operator==(BasicBlock *RHSDest) const { return Dest == RHSDest; } |
238 | }; |
239 | |
240 | class SimplifyCFGOpt { |
241 | const TargetTransformInfo &TTI; |
242 | DomTreeUpdater *DTU; |
243 | const DataLayout &DL; |
244 | ArrayRef<WeakVH> ; |
245 | const SimplifyCFGOptions &Options; |
246 | bool Resimplify; |
247 | |
248 | Value *isValueEqualityComparison(Instruction *TI); |
249 | BasicBlock *GetValueEqualityComparisonCases( |
250 | Instruction *TI, std::vector<ValueEqualityComparisonCase> &Cases); |
251 | bool SimplifyEqualityComparisonWithOnlyPredecessor(Instruction *TI, |
252 | BasicBlock *Pred, |
253 | IRBuilder<> &Builder); |
254 | bool PerformValueComparisonIntoPredecessorFolding(Instruction *TI, Value *&CV, |
255 | Instruction *PTI, |
256 | IRBuilder<> &Builder); |
257 | bool FoldValueComparisonIntoPredecessors(Instruction *TI, |
258 | IRBuilder<> &Builder); |
259 | |
260 | bool simplifyResume(ResumeInst *RI, IRBuilder<> &Builder); |
261 | bool simplifySingleResume(ResumeInst *RI); |
262 | bool simplifyCommonResume(ResumeInst *RI); |
263 | bool simplifyCleanupReturn(CleanupReturnInst *RI); |
264 | bool simplifyUnreachable(UnreachableInst *UI); |
265 | bool simplifySwitch(SwitchInst *SI, IRBuilder<> &Builder); |
266 | bool simplifyIndirectBr(IndirectBrInst *IBI); |
267 | bool simplifyBranch(BranchInst *Branch, IRBuilder<> &Builder); |
268 | bool simplifyUncondBranch(BranchInst *BI, IRBuilder<> &Builder); |
269 | bool simplifyCondBranch(BranchInst *BI, IRBuilder<> &Builder); |
270 | |
271 | bool tryToSimplifyUncondBranchWithICmpInIt(ICmpInst *ICI, |
272 | IRBuilder<> &Builder); |
273 | |
274 | bool hoistCommonCodeFromSuccessors(BasicBlock *BB, bool EqTermsOnly); |
275 | bool hoistSuccIdenticalTerminatorToSwitchOrIf( |
276 | Instruction *TI, Instruction *I1, |
277 | SmallVectorImpl<Instruction *> &OtherSuccTIs); |
278 | bool SpeculativelyExecuteBB(BranchInst *BI, BasicBlock *ThenBB); |
279 | bool SimplifyTerminatorOnSelect(Instruction *OldTerm, Value *Cond, |
280 | BasicBlock *TrueBB, BasicBlock *FalseBB, |
281 | uint32_t TrueWeight, uint32_t FalseWeight); |
282 | bool SimplifyBranchOnICmpChain(BranchInst *BI, IRBuilder<> &Builder, |
283 | const DataLayout &DL); |
284 | bool SimplifySwitchOnSelect(SwitchInst *SI, SelectInst *Select); |
285 | bool SimplifyIndirectBrOnSelect(IndirectBrInst *IBI, SelectInst *SI); |
286 | bool TurnSwitchRangeIntoICmp(SwitchInst *SI, IRBuilder<> &Builder); |
287 | |
288 | public: |
289 | SimplifyCFGOpt(const TargetTransformInfo &TTI, DomTreeUpdater *DTU, |
290 | const DataLayout &DL, ArrayRef<WeakVH> , |
291 | const SimplifyCFGOptions &Opts) |
292 | : TTI(TTI), DTU(DTU), DL(DL), LoopHeaders(LoopHeaders), Options(Opts) { |
293 | assert((!DTU || !DTU->hasPostDomTree()) && |
294 | "SimplifyCFG is not yet capable of maintaining validity of a " |
295 | "PostDomTree, so don't ask for it." ); |
296 | } |
297 | |
298 | bool simplifyOnce(BasicBlock *BB); |
299 | bool run(BasicBlock *BB); |
300 | |
301 | // Helper to set Resimplify and return change indication. |
302 | bool requestResimplify() { |
303 | Resimplify = true; |
304 | return true; |
305 | } |
306 | }; |
307 | |
308 | } // end anonymous namespace |
309 | |
310 | /// Return true if all the PHI nodes in the basic block \p BB |
311 | /// receive compatible (identical) incoming values when coming from |
312 | /// all of the predecessor blocks that are specified in \p IncomingBlocks. |
313 | /// |
314 | /// Note that if the values aren't exactly identical, but \p EquivalenceSet |
315 | /// is provided, and *both* of the values are present in the set, |
316 | /// then they are considered equal. |
317 | static bool IncomingValuesAreCompatible( |
318 | BasicBlock *BB, ArrayRef<BasicBlock *> IncomingBlocks, |
319 | SmallPtrSetImpl<Value *> *EquivalenceSet = nullptr) { |
320 | assert(IncomingBlocks.size() == 2 && |
321 | "Only for a pair of incoming blocks at the time!" ); |
322 | |
323 | // FIXME: it is okay if one of the incoming values is an `undef` value, |
324 | // iff the other incoming value is guaranteed to be a non-poison value. |
325 | // FIXME: it is okay if one of the incoming values is a `poison` value. |
326 | return all_of(Range: BB->phis(), P: [IncomingBlocks, EquivalenceSet](PHINode &PN) { |
327 | Value *IV0 = PN.getIncomingValueForBlock(BB: IncomingBlocks[0]); |
328 | Value *IV1 = PN.getIncomingValueForBlock(BB: IncomingBlocks[1]); |
329 | if (IV0 == IV1) |
330 | return true; |
331 | if (EquivalenceSet && EquivalenceSet->contains(Ptr: IV0) && |
332 | EquivalenceSet->contains(Ptr: IV1)) |
333 | return true; |
334 | return false; |
335 | }); |
336 | } |
337 | |
338 | /// Return true if it is safe to merge these two |
339 | /// terminator instructions together. |
340 | static bool |
341 | SafeToMergeTerminators(Instruction *SI1, Instruction *SI2, |
342 | SmallSetVector<BasicBlock *, 4> *FailBlocks = nullptr) { |
343 | if (SI1 == SI2) |
344 | return false; // Can't merge with self! |
345 | |
346 | // It is not safe to merge these two switch instructions if they have a common |
347 | // successor, and if that successor has a PHI node, and if *that* PHI node has |
348 | // conflicting incoming values from the two switch blocks. |
349 | BasicBlock *SI1BB = SI1->getParent(); |
350 | BasicBlock *SI2BB = SI2->getParent(); |
351 | |
352 | SmallPtrSet<BasicBlock *, 16> SI1Succs(succ_begin(BB: SI1BB), succ_end(BB: SI1BB)); |
353 | bool Fail = false; |
354 | for (BasicBlock *Succ : successors(BB: SI2BB)) { |
355 | if (!SI1Succs.count(Ptr: Succ)) |
356 | continue; |
357 | if (IncomingValuesAreCompatible(BB: Succ, IncomingBlocks: {SI1BB, SI2BB})) |
358 | continue; |
359 | Fail = true; |
360 | if (FailBlocks) |
361 | FailBlocks->insert(X: Succ); |
362 | else |
363 | break; |
364 | } |
365 | |
366 | return !Fail; |
367 | } |
368 | |
369 | /// Update PHI nodes in Succ to indicate that there will now be entries in it |
370 | /// from the 'NewPred' block. The values that will be flowing into the PHI nodes |
371 | /// will be the same as those coming in from ExistPred, an existing predecessor |
372 | /// of Succ. |
373 | static void AddPredecessorToBlock(BasicBlock *Succ, BasicBlock *NewPred, |
374 | BasicBlock *ExistPred, |
375 | MemorySSAUpdater *MSSAU = nullptr) { |
376 | for (PHINode &PN : Succ->phis()) |
377 | PN.addIncoming(V: PN.getIncomingValueForBlock(BB: ExistPred), BB: NewPred); |
378 | if (MSSAU) |
379 | if (auto *MPhi = MSSAU->getMemorySSA()->getMemoryAccess(BB: Succ)) |
380 | MPhi->addIncoming(V: MPhi->getIncomingValueForBlock(BB: ExistPred), BB: NewPred); |
381 | } |
382 | |
383 | /// Compute an abstract "cost" of speculating the given instruction, |
384 | /// which is assumed to be safe to speculate. TCC_Free means cheap, |
385 | /// TCC_Basic means less cheap, and TCC_Expensive means prohibitively |
386 | /// expensive. |
387 | static InstructionCost computeSpeculationCost(const User *I, |
388 | const TargetTransformInfo &TTI) { |
389 | assert((!isa<Instruction>(I) || |
390 | isSafeToSpeculativelyExecute(cast<Instruction>(I))) && |
391 | "Instruction is not safe to speculatively execute!" ); |
392 | return TTI.getInstructionCost(U: I, CostKind: TargetTransformInfo::TCK_SizeAndLatency); |
393 | } |
394 | |
395 | /// If we have a merge point of an "if condition" as accepted above, |
396 | /// return true if the specified value dominates the block. We |
397 | /// don't handle the true generality of domination here, just a special case |
398 | /// which works well enough for us. |
399 | /// |
400 | /// If AggressiveInsts is non-null, and if V does not dominate BB, we check to |
401 | /// see if V (which must be an instruction) and its recursive operands |
402 | /// that do not dominate BB have a combined cost lower than Budget and |
403 | /// are non-trapping. If both are true, the instruction is inserted into the |
404 | /// set and true is returned. |
405 | /// |
406 | /// The cost for most non-trapping instructions is defined as 1 except for |
407 | /// Select whose cost is 2. |
408 | /// |
409 | /// After this function returns, Cost is increased by the cost of |
410 | /// V plus its non-dominating operands. If that cost is greater than |
411 | /// Budget, false is returned and Cost is undefined. |
412 | static bool dominatesMergePoint(Value *V, BasicBlock *BB, |
413 | SmallPtrSetImpl<Instruction *> &AggressiveInsts, |
414 | InstructionCost &Cost, |
415 | InstructionCost Budget, |
416 | const TargetTransformInfo &TTI, |
417 | unsigned Depth = 0) { |
418 | // It is possible to hit a zero-cost cycle (phi/gep instructions for example), |
419 | // so limit the recursion depth. |
420 | // TODO: While this recursion limit does prevent pathological behavior, it |
421 | // would be better to track visited instructions to avoid cycles. |
422 | if (Depth == MaxSpeculationDepth) |
423 | return false; |
424 | |
425 | Instruction *I = dyn_cast<Instruction>(Val: V); |
426 | if (!I) { |
427 | // Non-instructions dominate all instructions and can be executed |
428 | // unconditionally. |
429 | return true; |
430 | } |
431 | BasicBlock *PBB = I->getParent(); |
432 | |
433 | // We don't want to allow weird loops that might have the "if condition" in |
434 | // the bottom of this block. |
435 | if (PBB == BB) |
436 | return false; |
437 | |
438 | // If this instruction is defined in a block that contains an unconditional |
439 | // branch to BB, then it must be in the 'conditional' part of the "if |
440 | // statement". If not, it definitely dominates the region. |
441 | BranchInst *BI = dyn_cast<BranchInst>(Val: PBB->getTerminator()); |
442 | if (!BI || BI->isConditional() || BI->getSuccessor(i: 0) != BB) |
443 | return true; |
444 | |
445 | // If we have seen this instruction before, don't count it again. |
446 | if (AggressiveInsts.count(Ptr: I)) |
447 | return true; |
448 | |
449 | // Okay, it looks like the instruction IS in the "condition". Check to |
450 | // see if it's a cheap instruction to unconditionally compute, and if it |
451 | // only uses stuff defined outside of the condition. If so, hoist it out. |
452 | if (!isSafeToSpeculativelyExecute(I)) |
453 | return false; |
454 | |
455 | Cost += computeSpeculationCost(I, TTI); |
456 | |
457 | // Allow exactly one instruction to be speculated regardless of its cost |
458 | // (as long as it is safe to do so). |
459 | // This is intended to flatten the CFG even if the instruction is a division |
460 | // or other expensive operation. The speculation of an expensive instruction |
461 | // is expected to be undone in CodeGenPrepare if the speculation has not |
462 | // enabled further IR optimizations. |
463 | if (Cost > Budget && |
464 | (!SpeculateOneExpensiveInst || !AggressiveInsts.empty() || Depth > 0 || |
465 | !Cost.isValid())) |
466 | return false; |
467 | |
468 | // Okay, we can only really hoist these out if their operands do |
469 | // not take us over the cost threshold. |
470 | for (Use &Op : I->operands()) |
471 | if (!dominatesMergePoint(V: Op, BB, AggressiveInsts, Cost, Budget, TTI, |
472 | Depth: Depth + 1)) |
473 | return false; |
474 | // Okay, it's safe to do this! Remember this instruction. |
475 | AggressiveInsts.insert(Ptr: I); |
476 | return true; |
477 | } |
478 | |
479 | /// Extract ConstantInt from value, looking through IntToPtr |
480 | /// and PointerNullValue. Return NULL if value is not a constant int. |
481 | static ConstantInt *GetConstantInt(Value *V, const DataLayout &DL) { |
482 | // Normal constant int. |
483 | ConstantInt *CI = dyn_cast<ConstantInt>(Val: V); |
484 | if (CI || !isa<Constant>(Val: V) || !V->getType()->isPointerTy() || |
485 | DL.isNonIntegralPointerType(Ty: V->getType())) |
486 | return CI; |
487 | |
488 | // This is some kind of pointer constant. Turn it into a pointer-sized |
489 | // ConstantInt if possible. |
490 | IntegerType *PtrTy = cast<IntegerType>(Val: DL.getIntPtrType(V->getType())); |
491 | |
492 | // Null pointer means 0, see SelectionDAGBuilder::getValue(const Value*). |
493 | if (isa<ConstantPointerNull>(Val: V)) |
494 | return ConstantInt::get(Ty: PtrTy, V: 0); |
495 | |
496 | // IntToPtr const int. |
497 | if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Val: V)) |
498 | if (CE->getOpcode() == Instruction::IntToPtr) |
499 | if (ConstantInt *CI = dyn_cast<ConstantInt>(Val: CE->getOperand(i_nocapture: 0))) { |
500 | // The constant is very likely to have the right type already. |
501 | if (CI->getType() == PtrTy) |
502 | return CI; |
503 | else |
504 | return cast<ConstantInt>( |
505 | Val: ConstantFoldIntegerCast(C: CI, DestTy: PtrTy, /*isSigned=*/IsSigned: false, DL)); |
506 | } |
507 | return nullptr; |
508 | } |
509 | |
510 | namespace { |
511 | |
512 | /// Given a chain of or (||) or and (&&) comparison of a value against a |
513 | /// constant, this will try to recover the information required for a switch |
514 | /// structure. |
515 | /// It will depth-first traverse the chain of comparison, seeking for patterns |
516 | /// like %a == 12 or %a < 4 and combine them to produce a set of integer |
517 | /// representing the different cases for the switch. |
518 | /// Note that if the chain is composed of '||' it will build the set of elements |
519 | /// that matches the comparisons (i.e. any of this value validate the chain) |
520 | /// while for a chain of '&&' it will build the set elements that make the test |
521 | /// fail. |
522 | struct ConstantComparesGatherer { |
523 | const DataLayout &DL; |
524 | |
525 | /// Value found for the switch comparison |
526 | Value *CompValue = nullptr; |
527 | |
528 | /// Extra clause to be checked before the switch |
529 | Value * = nullptr; |
530 | |
531 | /// Set of integers to match in switch |
532 | SmallVector<ConstantInt *, 8> Vals; |
533 | |
534 | /// Number of comparisons matched in the and/or chain |
535 | unsigned UsedICmps = 0; |
536 | |
537 | /// Construct and compute the result for the comparison instruction Cond |
538 | ConstantComparesGatherer(Instruction *Cond, const DataLayout &DL) : DL(DL) { |
539 | gather(V: Cond); |
540 | } |
541 | |
542 | ConstantComparesGatherer(const ConstantComparesGatherer &) = delete; |
543 | ConstantComparesGatherer & |
544 | operator=(const ConstantComparesGatherer &) = delete; |
545 | |
546 | private: |
547 | /// Try to set the current value used for the comparison, it succeeds only if |
548 | /// it wasn't set before or if the new value is the same as the old one |
549 | bool setValueOnce(Value *NewVal) { |
550 | if (CompValue && CompValue != NewVal) |
551 | return false; |
552 | CompValue = NewVal; |
553 | return (CompValue != nullptr); |
554 | } |
555 | |
556 | /// Try to match Instruction "I" as a comparison against a constant and |
557 | /// populates the array Vals with the set of values that match (or do not |
558 | /// match depending on isEQ). |
559 | /// Return false on failure. On success, the Value the comparison matched |
560 | /// against is placed in CompValue. |
561 | /// If CompValue is already set, the function is expected to fail if a match |
562 | /// is found but the value compared to is different. |
563 | bool matchInstruction(Instruction *I, bool isEQ) { |
564 | // If this is an icmp against a constant, handle this as one of the cases. |
565 | ICmpInst *ICI; |
566 | ConstantInt *C; |
567 | if (!((ICI = dyn_cast<ICmpInst>(Val: I)) && |
568 | (C = GetConstantInt(V: I->getOperand(i: 1), DL)))) { |
569 | return false; |
570 | } |
571 | |
572 | Value *RHSVal; |
573 | const APInt *RHSC; |
574 | |
575 | // Pattern match a special case |
576 | // (x & ~2^z) == y --> x == y || x == y|2^z |
577 | // This undoes a transformation done by instcombine to fuse 2 compares. |
578 | if (ICI->getPredicate() == (isEQ ? ICmpInst::ICMP_EQ : ICmpInst::ICMP_NE)) { |
579 | // It's a little bit hard to see why the following transformations are |
580 | // correct. Here is a CVC3 program to verify them for 64-bit values: |
581 | |
582 | /* |
583 | ONE : BITVECTOR(64) = BVZEROEXTEND(0bin1, 63); |
584 | x : BITVECTOR(64); |
585 | y : BITVECTOR(64); |
586 | z : BITVECTOR(64); |
587 | mask : BITVECTOR(64) = BVSHL(ONE, z); |
588 | QUERY( (y & ~mask = y) => |
589 | ((x & ~mask = y) <=> (x = y OR x = (y | mask))) |
590 | ); |
591 | QUERY( (y | mask = y) => |
592 | ((x | mask = y) <=> (x = y OR x = (y & ~mask))) |
593 | ); |
594 | */ |
595 | |
596 | // Please note that each pattern must be a dual implication (<--> or |
597 | // iff). One directional implication can create spurious matches. If the |
598 | // implication is only one-way, an unsatisfiable condition on the left |
599 | // side can imply a satisfiable condition on the right side. Dual |
600 | // implication ensures that satisfiable conditions are transformed to |
601 | // other satisfiable conditions and unsatisfiable conditions are |
602 | // transformed to other unsatisfiable conditions. |
603 | |
604 | // Here is a concrete example of a unsatisfiable condition on the left |
605 | // implying a satisfiable condition on the right: |
606 | // |
607 | // mask = (1 << z) |
608 | // (x & ~mask) == y --> (x == y || x == (y | mask)) |
609 | // |
610 | // Substituting y = 3, z = 0 yields: |
611 | // (x & -2) == 3 --> (x == 3 || x == 2) |
612 | |
613 | // Pattern match a special case: |
614 | /* |
615 | QUERY( (y & ~mask = y) => |
616 | ((x & ~mask = y) <=> (x = y OR x = (y | mask))) |
617 | ); |
618 | */ |
619 | if (match(V: ICI->getOperand(i_nocapture: 0), |
620 | P: m_And(L: m_Value(V&: RHSVal), R: m_APInt(Res&: RHSC)))) { |
621 | APInt Mask = ~*RHSC; |
622 | if (Mask.isPowerOf2() && (C->getValue() & ~Mask) == C->getValue()) { |
623 | // If we already have a value for the switch, it has to match! |
624 | if (!setValueOnce(RHSVal)) |
625 | return false; |
626 | |
627 | Vals.push_back(Elt: C); |
628 | Vals.push_back( |
629 | Elt: ConstantInt::get(Context&: C->getContext(), |
630 | V: C->getValue() | Mask)); |
631 | UsedICmps++; |
632 | return true; |
633 | } |
634 | } |
635 | |
636 | // Pattern match a special case: |
637 | /* |
638 | QUERY( (y | mask = y) => |
639 | ((x | mask = y) <=> (x = y OR x = (y & ~mask))) |
640 | ); |
641 | */ |
642 | if (match(V: ICI->getOperand(i_nocapture: 0), |
643 | P: m_Or(L: m_Value(V&: RHSVal), R: m_APInt(Res&: RHSC)))) { |
644 | APInt Mask = *RHSC; |
645 | if (Mask.isPowerOf2() && (C->getValue() | Mask) == C->getValue()) { |
646 | // If we already have a value for the switch, it has to match! |
647 | if (!setValueOnce(RHSVal)) |
648 | return false; |
649 | |
650 | Vals.push_back(Elt: C); |
651 | Vals.push_back(Elt: ConstantInt::get(Context&: C->getContext(), |
652 | V: C->getValue() & ~Mask)); |
653 | UsedICmps++; |
654 | return true; |
655 | } |
656 | } |
657 | |
658 | // If we already have a value for the switch, it has to match! |
659 | if (!setValueOnce(ICI->getOperand(i_nocapture: 0))) |
660 | return false; |
661 | |
662 | UsedICmps++; |
663 | Vals.push_back(Elt: C); |
664 | return ICI->getOperand(i_nocapture: 0); |
665 | } |
666 | |
667 | // If we have "x ult 3", for example, then we can add 0,1,2 to the set. |
668 | ConstantRange Span = |
669 | ConstantRange::makeExactICmpRegion(Pred: ICI->getPredicate(), Other: C->getValue()); |
670 | |
671 | // Shift the range if the compare is fed by an add. This is the range |
672 | // compare idiom as emitted by instcombine. |
673 | Value *CandidateVal = I->getOperand(i: 0); |
674 | if (match(V: I->getOperand(i: 0), P: m_Add(L: m_Value(V&: RHSVal), R: m_APInt(Res&: RHSC)))) { |
675 | Span = Span.subtract(CI: *RHSC); |
676 | CandidateVal = RHSVal; |
677 | } |
678 | |
679 | // If this is an and/!= check, then we are looking to build the set of |
680 | // value that *don't* pass the and chain. I.e. to turn "x ugt 2" into |
681 | // x != 0 && x != 1. |
682 | if (!isEQ) |
683 | Span = Span.inverse(); |
684 | |
685 | // If there are a ton of values, we don't want to make a ginormous switch. |
686 | if (Span.isSizeLargerThan(MaxSize: 8) || Span.isEmptySet()) { |
687 | return false; |
688 | } |
689 | |
690 | // If we already have a value for the switch, it has to match! |
691 | if (!setValueOnce(CandidateVal)) |
692 | return false; |
693 | |
694 | // Add all values from the range to the set |
695 | for (APInt Tmp = Span.getLower(); Tmp != Span.getUpper(); ++Tmp) |
696 | Vals.push_back(Elt: ConstantInt::get(Context&: I->getContext(), V: Tmp)); |
697 | |
698 | UsedICmps++; |
699 | return true; |
700 | } |
701 | |
702 | /// Given a potentially 'or'd or 'and'd together collection of icmp |
703 | /// eq/ne/lt/gt instructions that compare a value against a constant, extract |
704 | /// the value being compared, and stick the list constants into the Vals |
705 | /// vector. |
706 | /// One "Extra" case is allowed to differ from the other. |
707 | void gather(Value *V) { |
708 | bool isEQ = match(V, P: m_LogicalOr(L: m_Value(), R: m_Value())); |
709 | |
710 | // Keep a stack (SmallVector for efficiency) for depth-first traversal |
711 | SmallVector<Value *, 8> DFT; |
712 | SmallPtrSet<Value *, 8> Visited; |
713 | |
714 | // Initialize |
715 | Visited.insert(Ptr: V); |
716 | DFT.push_back(Elt: V); |
717 | |
718 | while (!DFT.empty()) { |
719 | V = DFT.pop_back_val(); |
720 | |
721 | if (Instruction *I = dyn_cast<Instruction>(Val: V)) { |
722 | // If it is a || (or && depending on isEQ), process the operands. |
723 | Value *Op0, *Op1; |
724 | if (isEQ ? match(V: I, P: m_LogicalOr(L: m_Value(V&: Op0), R: m_Value(V&: Op1))) |
725 | : match(V: I, P: m_LogicalAnd(L: m_Value(V&: Op0), R: m_Value(V&: Op1)))) { |
726 | if (Visited.insert(Ptr: Op1).second) |
727 | DFT.push_back(Elt: Op1); |
728 | if (Visited.insert(Ptr: Op0).second) |
729 | DFT.push_back(Elt: Op0); |
730 | |
731 | continue; |
732 | } |
733 | |
734 | // Try to match the current instruction |
735 | if (matchInstruction(I, isEQ)) |
736 | // Match succeed, continue the loop |
737 | continue; |
738 | } |
739 | |
740 | // One element of the sequence of || (or &&) could not be match as a |
741 | // comparison against the same value as the others. |
742 | // We allow only one "Extra" case to be checked before the switch |
743 | if (!Extra) { |
744 | Extra = V; |
745 | continue; |
746 | } |
747 | // Failed to parse a proper sequence, abort now |
748 | CompValue = nullptr; |
749 | break; |
750 | } |
751 | } |
752 | }; |
753 | |
754 | } // end anonymous namespace |
755 | |
756 | static void EraseTerminatorAndDCECond(Instruction *TI, |
757 | MemorySSAUpdater *MSSAU = nullptr) { |
758 | Instruction *Cond = nullptr; |
759 | if (SwitchInst *SI = dyn_cast<SwitchInst>(Val: TI)) { |
760 | Cond = dyn_cast<Instruction>(Val: SI->getCondition()); |
761 | } else if (BranchInst *BI = dyn_cast<BranchInst>(Val: TI)) { |
762 | if (BI->isConditional()) |
763 | Cond = dyn_cast<Instruction>(Val: BI->getCondition()); |
764 | } else if (IndirectBrInst *IBI = dyn_cast<IndirectBrInst>(Val: TI)) { |
765 | Cond = dyn_cast<Instruction>(Val: IBI->getAddress()); |
766 | } |
767 | |
768 | TI->eraseFromParent(); |
769 | if (Cond) |
770 | RecursivelyDeleteTriviallyDeadInstructions(V: Cond, TLI: nullptr, MSSAU); |
771 | } |
772 | |
773 | /// Return true if the specified terminator checks |
774 | /// to see if a value is equal to constant integer value. |
775 | Value *SimplifyCFGOpt::isValueEqualityComparison(Instruction *TI) { |
776 | Value *CV = nullptr; |
777 | if (SwitchInst *SI = dyn_cast<SwitchInst>(Val: TI)) { |
778 | // Do not permit merging of large switch instructions into their |
779 | // predecessors unless there is only one predecessor. |
780 | if (!SI->getParent()->hasNPredecessorsOrMore(N: 128 / SI->getNumSuccessors())) |
781 | CV = SI->getCondition(); |
782 | } else if (BranchInst *BI = dyn_cast<BranchInst>(Val: TI)) |
783 | if (BI->isConditional() && BI->getCondition()->hasOneUse()) |
784 | if (ICmpInst *ICI = dyn_cast<ICmpInst>(Val: BI->getCondition())) { |
785 | if (ICI->isEquality() && GetConstantInt(V: ICI->getOperand(i_nocapture: 1), DL)) |
786 | CV = ICI->getOperand(i_nocapture: 0); |
787 | } |
788 | |
789 | // Unwrap any lossless ptrtoint cast. |
790 | if (CV) { |
791 | if (PtrToIntInst *PTII = dyn_cast<PtrToIntInst>(Val: CV)) { |
792 | Value *Ptr = PTII->getPointerOperand(); |
793 | if (PTII->getType() == DL.getIntPtrType(Ptr->getType())) |
794 | CV = Ptr; |
795 | } |
796 | } |
797 | return CV; |
798 | } |
799 | |
800 | /// Given a value comparison instruction, |
801 | /// decode all of the 'cases' that it represents and return the 'default' block. |
802 | BasicBlock *SimplifyCFGOpt::GetValueEqualityComparisonCases( |
803 | Instruction *TI, std::vector<ValueEqualityComparisonCase> &Cases) { |
804 | if (SwitchInst *SI = dyn_cast<SwitchInst>(Val: TI)) { |
805 | Cases.reserve(n: SI->getNumCases()); |
806 | for (auto Case : SI->cases()) |
807 | Cases.push_back(x: ValueEqualityComparisonCase(Case.getCaseValue(), |
808 | Case.getCaseSuccessor())); |
809 | return SI->getDefaultDest(); |
810 | } |
811 | |
812 | BranchInst *BI = cast<BranchInst>(Val: TI); |
813 | ICmpInst *ICI = cast<ICmpInst>(Val: BI->getCondition()); |
814 | BasicBlock *Succ = BI->getSuccessor(i: ICI->getPredicate() == ICmpInst::ICMP_NE); |
815 | Cases.push_back(x: ValueEqualityComparisonCase( |
816 | GetConstantInt(V: ICI->getOperand(i_nocapture: 1), DL), Succ)); |
817 | return BI->getSuccessor(i: ICI->getPredicate() == ICmpInst::ICMP_EQ); |
818 | } |
819 | |
820 | /// Given a vector of bb/value pairs, remove any entries |
821 | /// in the list that match the specified block. |
822 | static void |
823 | EliminateBlockCases(BasicBlock *BB, |
824 | std::vector<ValueEqualityComparisonCase> &Cases) { |
825 | llvm::erase(C&: Cases, V: BB); |
826 | } |
827 | |
828 | /// Return true if there are any keys in C1 that exist in C2 as well. |
829 | static bool ValuesOverlap(std::vector<ValueEqualityComparisonCase> &C1, |
830 | std::vector<ValueEqualityComparisonCase> &C2) { |
831 | std::vector<ValueEqualityComparisonCase> *V1 = &C1, *V2 = &C2; |
832 | |
833 | // Make V1 be smaller than V2. |
834 | if (V1->size() > V2->size()) |
835 | std::swap(a&: V1, b&: V2); |
836 | |
837 | if (V1->empty()) |
838 | return false; |
839 | if (V1->size() == 1) { |
840 | // Just scan V2. |
841 | ConstantInt *TheVal = (*V1)[0].Value; |
842 | for (const ValueEqualityComparisonCase &VECC : *V2) |
843 | if (TheVal == VECC.Value) |
844 | return true; |
845 | } |
846 | |
847 | // Otherwise, just sort both lists and compare element by element. |
848 | array_pod_sort(Start: V1->begin(), End: V1->end()); |
849 | array_pod_sort(Start: V2->begin(), End: V2->end()); |
850 | unsigned i1 = 0, i2 = 0, e1 = V1->size(), e2 = V2->size(); |
851 | while (i1 != e1 && i2 != e2) { |
852 | if ((*V1)[i1].Value == (*V2)[i2].Value) |
853 | return true; |
854 | if ((*V1)[i1].Value < (*V2)[i2].Value) |
855 | ++i1; |
856 | else |
857 | ++i2; |
858 | } |
859 | return false; |
860 | } |
861 | |
862 | // Set branch weights on SwitchInst. This sets the metadata if there is at |
863 | // least one non-zero weight. |
864 | static void setBranchWeights(SwitchInst *SI, ArrayRef<uint32_t> Weights, |
865 | bool IsExpected) { |
866 | // Check that there is at least one non-zero weight. Otherwise, pass |
867 | // nullptr to setMetadata which will erase the existing metadata. |
868 | MDNode *N = nullptr; |
869 | if (llvm::any_of(Range&: Weights, P: [](uint32_t W) { return W != 0; })) |
870 | N = MDBuilder(SI->getParent()->getContext()) |
871 | .createBranchWeights(Weights, IsExpected); |
872 | SI->setMetadata(KindID: LLVMContext::MD_prof, Node: N); |
873 | } |
874 | |
875 | // Similar to the above, but for branch and select instructions that take |
876 | // exactly 2 weights. |
877 | static void setBranchWeights(Instruction *I, uint32_t TrueWeight, |
878 | uint32_t FalseWeight, bool IsExpected) { |
879 | assert(isa<BranchInst>(I) || isa<SelectInst>(I)); |
880 | // Check that there is at least one non-zero weight. Otherwise, pass |
881 | // nullptr to setMetadata which will erase the existing metadata. |
882 | MDNode *N = nullptr; |
883 | if (TrueWeight || FalseWeight) |
884 | N = MDBuilder(I->getParent()->getContext()) |
885 | .createBranchWeights(TrueWeight, FalseWeight, IsExpected); |
886 | I->setMetadata(KindID: LLVMContext::MD_prof, Node: N); |
887 | } |
888 | |
889 | /// If TI is known to be a terminator instruction and its block is known to |
890 | /// only have a single predecessor block, check to see if that predecessor is |
891 | /// also a value comparison with the same value, and if that comparison |
892 | /// determines the outcome of this comparison. If so, simplify TI. This does a |
893 | /// very limited form of jump threading. |
894 | bool SimplifyCFGOpt::SimplifyEqualityComparisonWithOnlyPredecessor( |
895 | Instruction *TI, BasicBlock *Pred, IRBuilder<> &Builder) { |
896 | Value *PredVal = isValueEqualityComparison(TI: Pred->getTerminator()); |
897 | if (!PredVal) |
898 | return false; // Not a value comparison in predecessor. |
899 | |
900 | Value *ThisVal = isValueEqualityComparison(TI); |
901 | assert(ThisVal && "This isn't a value comparison!!" ); |
902 | if (ThisVal != PredVal) |
903 | return false; // Different predicates. |
904 | |
905 | // TODO: Preserve branch weight metadata, similarly to how |
906 | // FoldValueComparisonIntoPredecessors preserves it. |
907 | |
908 | // Find out information about when control will move from Pred to TI's block. |
909 | std::vector<ValueEqualityComparisonCase> PredCases; |
910 | BasicBlock *PredDef = |
911 | GetValueEqualityComparisonCases(TI: Pred->getTerminator(), Cases&: PredCases); |
912 | EliminateBlockCases(BB: PredDef, Cases&: PredCases); // Remove default from cases. |
913 | |
914 | // Find information about how control leaves this block. |
915 | std::vector<ValueEqualityComparisonCase> ThisCases; |
916 | BasicBlock *ThisDef = GetValueEqualityComparisonCases(TI, Cases&: ThisCases); |
917 | EliminateBlockCases(BB: ThisDef, Cases&: ThisCases); // Remove default from cases. |
918 | |
919 | // If TI's block is the default block from Pred's comparison, potentially |
920 | // simplify TI based on this knowledge. |
921 | if (PredDef == TI->getParent()) { |
922 | // If we are here, we know that the value is none of those cases listed in |
923 | // PredCases. If there are any cases in ThisCases that are in PredCases, we |
924 | // can simplify TI. |
925 | if (!ValuesOverlap(C1&: PredCases, C2&: ThisCases)) |
926 | return false; |
927 | |
928 | if (isa<BranchInst>(Val: TI)) { |
929 | // Okay, one of the successors of this condbr is dead. Convert it to a |
930 | // uncond br. |
931 | assert(ThisCases.size() == 1 && "Branch can only have one case!" ); |
932 | // Insert the new branch. |
933 | Instruction *NI = Builder.CreateBr(Dest: ThisDef); |
934 | (void)NI; |
935 | |
936 | // Remove PHI node entries for the dead edge. |
937 | ThisCases[0].Dest->removePredecessor(Pred: PredDef); |
938 | |
939 | LLVM_DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator() |
940 | << "Through successor TI: " << *TI << "Leaving: " << *NI |
941 | << "\n" ); |
942 | |
943 | EraseTerminatorAndDCECond(TI); |
944 | |
945 | if (DTU) |
946 | DTU->applyUpdates( |
947 | Updates: {{DominatorTree::Delete, PredDef, ThisCases[0].Dest}}); |
948 | |
949 | return true; |
950 | } |
951 | |
952 | SwitchInstProfUpdateWrapper SI = *cast<SwitchInst>(Val: TI); |
953 | // Okay, TI has cases that are statically dead, prune them away. |
954 | SmallPtrSet<Constant *, 16> DeadCases; |
955 | for (unsigned i = 0, e = PredCases.size(); i != e; ++i) |
956 | DeadCases.insert(Ptr: PredCases[i].Value); |
957 | |
958 | LLVM_DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator() |
959 | << "Through successor TI: " << *TI); |
960 | |
961 | SmallDenseMap<BasicBlock *, int, 8> NumPerSuccessorCases; |
962 | for (SwitchInst::CaseIt i = SI->case_end(), e = SI->case_begin(); i != e;) { |
963 | --i; |
964 | auto *Successor = i->getCaseSuccessor(); |
965 | if (DTU) |
966 | ++NumPerSuccessorCases[Successor]; |
967 | if (DeadCases.count(Ptr: i->getCaseValue())) { |
968 | Successor->removePredecessor(Pred: PredDef); |
969 | SI.removeCase(I: i); |
970 | if (DTU) |
971 | --NumPerSuccessorCases[Successor]; |
972 | } |
973 | } |
974 | |
975 | if (DTU) { |
976 | std::vector<DominatorTree::UpdateType> Updates; |
977 | for (const std::pair<BasicBlock *, int> &I : NumPerSuccessorCases) |
978 | if (I.second == 0) |
979 | Updates.push_back(x: {DominatorTree::Delete, PredDef, I.first}); |
980 | DTU->applyUpdates(Updates); |
981 | } |
982 | |
983 | LLVM_DEBUG(dbgs() << "Leaving: " << *TI << "\n" ); |
984 | return true; |
985 | } |
986 | |
987 | // Otherwise, TI's block must correspond to some matched value. Find out |
988 | // which value (or set of values) this is. |
989 | ConstantInt *TIV = nullptr; |
990 | BasicBlock *TIBB = TI->getParent(); |
991 | for (unsigned i = 0, e = PredCases.size(); i != e; ++i) |
992 | if (PredCases[i].Dest == TIBB) { |
993 | if (TIV) |
994 | return false; // Cannot handle multiple values coming to this block. |
995 | TIV = PredCases[i].Value; |
996 | } |
997 | assert(TIV && "No edge from pred to succ?" ); |
998 | |
999 | // Okay, we found the one constant that our value can be if we get into TI's |
1000 | // BB. Find out which successor will unconditionally be branched to. |
1001 | BasicBlock *TheRealDest = nullptr; |
1002 | for (unsigned i = 0, e = ThisCases.size(); i != e; ++i) |
1003 | if (ThisCases[i].Value == TIV) { |
1004 | TheRealDest = ThisCases[i].Dest; |
1005 | break; |
1006 | } |
1007 | |
1008 | // If not handled by any explicit cases, it is handled by the default case. |
1009 | if (!TheRealDest) |
1010 | TheRealDest = ThisDef; |
1011 | |
1012 | SmallPtrSet<BasicBlock *, 2> RemovedSuccs; |
1013 | |
1014 | // Remove PHI node entries for dead edges. |
1015 | BasicBlock *CheckEdge = TheRealDest; |
1016 | for (BasicBlock *Succ : successors(BB: TIBB)) |
1017 | if (Succ != CheckEdge) { |
1018 | if (Succ != TheRealDest) |
1019 | RemovedSuccs.insert(Ptr: Succ); |
1020 | Succ->removePredecessor(Pred: TIBB); |
1021 | } else |
1022 | CheckEdge = nullptr; |
1023 | |
1024 | // Insert the new branch. |
1025 | Instruction *NI = Builder.CreateBr(Dest: TheRealDest); |
1026 | (void)NI; |
1027 | |
1028 | LLVM_DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator() |
1029 | << "Through successor TI: " << *TI << "Leaving: " << *NI |
1030 | << "\n" ); |
1031 | |
1032 | EraseTerminatorAndDCECond(TI); |
1033 | if (DTU) { |
1034 | SmallVector<DominatorTree::UpdateType, 2> Updates; |
1035 | Updates.reserve(N: RemovedSuccs.size()); |
1036 | for (auto *RemovedSucc : RemovedSuccs) |
1037 | Updates.push_back(Elt: {DominatorTree::Delete, TIBB, RemovedSucc}); |
1038 | DTU->applyUpdates(Updates); |
1039 | } |
1040 | return true; |
1041 | } |
1042 | |
1043 | namespace { |
1044 | |
1045 | /// This class implements a stable ordering of constant |
1046 | /// integers that does not depend on their address. This is important for |
1047 | /// applications that sort ConstantInt's to ensure uniqueness. |
1048 | struct ConstantIntOrdering { |
1049 | bool operator()(const ConstantInt *LHS, const ConstantInt *RHS) const { |
1050 | return LHS->getValue().ult(RHS: RHS->getValue()); |
1051 | } |
1052 | }; |
1053 | |
1054 | } // end anonymous namespace |
1055 | |
1056 | static int ConstantIntSortPredicate(ConstantInt *const *P1, |
1057 | ConstantInt *const *P2) { |
1058 | const ConstantInt *LHS = *P1; |
1059 | const ConstantInt *RHS = *P2; |
1060 | if (LHS == RHS) |
1061 | return 0; |
1062 | return LHS->getValue().ult(RHS: RHS->getValue()) ? 1 : -1; |
1063 | } |
1064 | |
1065 | /// Get Weights of a given terminator, the default weight is at the front |
1066 | /// of the vector. If TI is a conditional eq, we need to swap the branch-weight |
1067 | /// metadata. |
1068 | static void GetBranchWeights(Instruction *TI, |
1069 | SmallVectorImpl<uint64_t> &Weights) { |
1070 | MDNode *MD = TI->getMetadata(KindID: LLVMContext::MD_prof); |
1071 | assert(MD && "Invalid branch-weight metadata" ); |
1072 | extractFromBranchWeightMD64(ProfileData: MD, Weights); |
1073 | |
1074 | // If TI is a conditional eq, the default case is the false case, |
1075 | // and the corresponding branch-weight data is at index 2. We swap the |
1076 | // default weight to be the first entry. |
1077 | if (BranchInst *BI = dyn_cast<BranchInst>(Val: TI)) { |
1078 | assert(Weights.size() == 2); |
1079 | ICmpInst *ICI = cast<ICmpInst>(Val: BI->getCondition()); |
1080 | if (ICI->getPredicate() == ICmpInst::ICMP_EQ) |
1081 | std::swap(a&: Weights.front(), b&: Weights.back()); |
1082 | } |
1083 | } |
1084 | |
1085 | /// Keep halving the weights until all can fit in uint32_t. |
1086 | static void FitWeights(MutableArrayRef<uint64_t> Weights) { |
1087 | uint64_t Max = *llvm::max_element(Range&: Weights); |
1088 | if (Max > UINT_MAX) { |
1089 | unsigned Offset = 32 - llvm::countl_zero(Val: Max); |
1090 | for (uint64_t &I : Weights) |
1091 | I >>= Offset; |
1092 | } |
1093 | } |
1094 | |
1095 | static void CloneInstructionsIntoPredecessorBlockAndUpdateSSAUses( |
1096 | BasicBlock *BB, BasicBlock *PredBlock, ValueToValueMapTy &VMap) { |
1097 | Instruction *PTI = PredBlock->getTerminator(); |
1098 | |
1099 | // If we have bonus instructions, clone them into the predecessor block. |
1100 | // Note that there may be multiple predecessor blocks, so we cannot move |
1101 | // bonus instructions to a predecessor block. |
1102 | for (Instruction &BonusInst : *BB) { |
1103 | if (BonusInst.isTerminator()) |
1104 | continue; |
1105 | |
1106 | Instruction *NewBonusInst = BonusInst.clone(); |
1107 | |
1108 | if (!isa<DbgInfoIntrinsic>(Val: BonusInst) && |
1109 | PTI->getDebugLoc() != NewBonusInst->getDebugLoc()) { |
1110 | // Unless the instruction has the same !dbg location as the original |
1111 | // branch, drop it. When we fold the bonus instructions we want to make |
1112 | // sure we reset their debug locations in order to avoid stepping on |
1113 | // dead code caused by folding dead branches. |
1114 | NewBonusInst->setDebugLoc(DebugLoc()); |
1115 | } |
1116 | |
1117 | RemapInstruction(I: NewBonusInst, VM&: VMap, |
1118 | Flags: RF_NoModuleLevelChanges | RF_IgnoreMissingLocals); |
1119 | |
1120 | // If we speculated an instruction, we need to drop any metadata that may |
1121 | // result in undefined behavior, as the metadata might have been valid |
1122 | // only given the branch precondition. |
1123 | // Similarly strip attributes on call parameters that may cause UB in |
1124 | // location the call is moved to. |
1125 | NewBonusInst->dropUBImplyingAttrsAndMetadata(); |
1126 | |
1127 | NewBonusInst->insertInto(ParentBB: PredBlock, It: PTI->getIterator()); |
1128 | auto Range = NewBonusInst->cloneDebugInfoFrom(From: &BonusInst); |
1129 | RemapDbgRecordRange(M: NewBonusInst->getModule(), Range, VM&: VMap, |
1130 | Flags: RF_NoModuleLevelChanges | RF_IgnoreMissingLocals); |
1131 | |
1132 | if (isa<DbgInfoIntrinsic>(Val: BonusInst)) |
1133 | continue; |
1134 | |
1135 | NewBonusInst->takeName(V: &BonusInst); |
1136 | BonusInst.setName(NewBonusInst->getName() + ".old" ); |
1137 | VMap[&BonusInst] = NewBonusInst; |
1138 | |
1139 | // Update (liveout) uses of bonus instructions, |
1140 | // now that the bonus instruction has been cloned into predecessor. |
1141 | // Note that we expect to be in a block-closed SSA form for this to work! |
1142 | for (Use &U : make_early_inc_range(Range: BonusInst.uses())) { |
1143 | auto *UI = cast<Instruction>(Val: U.getUser()); |
1144 | auto *PN = dyn_cast<PHINode>(Val: UI); |
1145 | if (!PN) { |
1146 | assert(UI->getParent() == BB && BonusInst.comesBefore(UI) && |
1147 | "If the user is not a PHI node, then it should be in the same " |
1148 | "block as, and come after, the original bonus instruction." ); |
1149 | continue; // Keep using the original bonus instruction. |
1150 | } |
1151 | // Is this the block-closed SSA form PHI node? |
1152 | if (PN->getIncomingBlock(U) == BB) |
1153 | continue; // Great, keep using the original bonus instruction. |
1154 | // The only other alternative is an "use" when coming from |
1155 | // the predecessor block - here we should refer to the cloned bonus instr. |
1156 | assert(PN->getIncomingBlock(U) == PredBlock && |
1157 | "Not in block-closed SSA form?" ); |
1158 | U.set(NewBonusInst); |
1159 | } |
1160 | } |
1161 | } |
1162 | |
1163 | bool SimplifyCFGOpt::PerformValueComparisonIntoPredecessorFolding( |
1164 | Instruction *TI, Value *&CV, Instruction *PTI, IRBuilder<> &Builder) { |
1165 | BasicBlock *BB = TI->getParent(); |
1166 | BasicBlock *Pred = PTI->getParent(); |
1167 | |
1168 | SmallVector<DominatorTree::UpdateType, 32> Updates; |
1169 | |
1170 | // Figure out which 'cases' to copy from SI to PSI. |
1171 | std::vector<ValueEqualityComparisonCase> BBCases; |
1172 | BasicBlock *BBDefault = GetValueEqualityComparisonCases(TI, Cases&: BBCases); |
1173 | |
1174 | std::vector<ValueEqualityComparisonCase> PredCases; |
1175 | BasicBlock *PredDefault = GetValueEqualityComparisonCases(TI: PTI, Cases&: PredCases); |
1176 | |
1177 | // Based on whether the default edge from PTI goes to BB or not, fill in |
1178 | // PredCases and PredDefault with the new switch cases we would like to |
1179 | // build. |
1180 | SmallMapVector<BasicBlock *, int, 8> NewSuccessors; |
1181 | |
1182 | // Update the branch weight metadata along the way |
1183 | SmallVector<uint64_t, 8> Weights; |
1184 | bool PredHasWeights = hasBranchWeightMD(I: *PTI); |
1185 | bool SuccHasWeights = hasBranchWeightMD(I: *TI); |
1186 | |
1187 | if (PredHasWeights) { |
1188 | GetBranchWeights(TI: PTI, Weights); |
1189 | // branch-weight metadata is inconsistent here. |
1190 | if (Weights.size() != 1 + PredCases.size()) |
1191 | PredHasWeights = SuccHasWeights = false; |
1192 | } else if (SuccHasWeights) |
1193 | // If there are no predecessor weights but there are successor weights, |
1194 | // populate Weights with 1, which will later be scaled to the sum of |
1195 | // successor's weights |
1196 | Weights.assign(NumElts: 1 + PredCases.size(), Elt: 1); |
1197 | |
1198 | SmallVector<uint64_t, 8> SuccWeights; |
1199 | if (SuccHasWeights) { |
1200 | GetBranchWeights(TI, Weights&: SuccWeights); |
1201 | // branch-weight metadata is inconsistent here. |
1202 | if (SuccWeights.size() != 1 + BBCases.size()) |
1203 | PredHasWeights = SuccHasWeights = false; |
1204 | } else if (PredHasWeights) |
1205 | SuccWeights.assign(NumElts: 1 + BBCases.size(), Elt: 1); |
1206 | |
1207 | if (PredDefault == BB) { |
1208 | // If this is the default destination from PTI, only the edges in TI |
1209 | // that don't occur in PTI, or that branch to BB will be activated. |
1210 | std::set<ConstantInt *, ConstantIntOrdering> PTIHandled; |
1211 | for (unsigned i = 0, e = PredCases.size(); i != e; ++i) |
1212 | if (PredCases[i].Dest != BB) |
1213 | PTIHandled.insert(x: PredCases[i].Value); |
1214 | else { |
1215 | // The default destination is BB, we don't need explicit targets. |
1216 | std::swap(a&: PredCases[i], b&: PredCases.back()); |
1217 | |
1218 | if (PredHasWeights || SuccHasWeights) { |
1219 | // Increase weight for the default case. |
1220 | Weights[0] += Weights[i + 1]; |
1221 | std::swap(a&: Weights[i + 1], b&: Weights.back()); |
1222 | Weights.pop_back(); |
1223 | } |
1224 | |
1225 | PredCases.pop_back(); |
1226 | --i; |
1227 | --e; |
1228 | } |
1229 | |
1230 | // Reconstruct the new switch statement we will be building. |
1231 | if (PredDefault != BBDefault) { |
1232 | PredDefault->removePredecessor(Pred); |
1233 | if (DTU && PredDefault != BB) |
1234 | Updates.push_back(Elt: {DominatorTree::Delete, Pred, PredDefault}); |
1235 | PredDefault = BBDefault; |
1236 | ++NewSuccessors[BBDefault]; |
1237 | } |
1238 | |
1239 | unsigned CasesFromPred = Weights.size(); |
1240 | uint64_t ValidTotalSuccWeight = 0; |
1241 | for (unsigned i = 0, e = BBCases.size(); i != e; ++i) |
1242 | if (!PTIHandled.count(x: BBCases[i].Value) && BBCases[i].Dest != BBDefault) { |
1243 | PredCases.push_back(x: BBCases[i]); |
1244 | ++NewSuccessors[BBCases[i].Dest]; |
1245 | if (SuccHasWeights || PredHasWeights) { |
1246 | // The default weight is at index 0, so weight for the ith case |
1247 | // should be at index i+1. Scale the cases from successor by |
1248 | // PredDefaultWeight (Weights[0]). |
1249 | Weights.push_back(Elt: Weights[0] * SuccWeights[i + 1]); |
1250 | ValidTotalSuccWeight += SuccWeights[i + 1]; |
1251 | } |
1252 | } |
1253 | |
1254 | if (SuccHasWeights || PredHasWeights) { |
1255 | ValidTotalSuccWeight += SuccWeights[0]; |
1256 | // Scale the cases from predecessor by ValidTotalSuccWeight. |
1257 | for (unsigned i = 1; i < CasesFromPred; ++i) |
1258 | Weights[i] *= ValidTotalSuccWeight; |
1259 | // Scale the default weight by SuccDefaultWeight (SuccWeights[0]). |
1260 | Weights[0] *= SuccWeights[0]; |
1261 | } |
1262 | } else { |
1263 | // If this is not the default destination from PSI, only the edges |
1264 | // in SI that occur in PSI with a destination of BB will be |
1265 | // activated. |
1266 | std::set<ConstantInt *, ConstantIntOrdering> PTIHandled; |
1267 | std::map<ConstantInt *, uint64_t> WeightsForHandled; |
1268 | for (unsigned i = 0, e = PredCases.size(); i != e; ++i) |
1269 | if (PredCases[i].Dest == BB) { |
1270 | PTIHandled.insert(x: PredCases[i].Value); |
1271 | |
1272 | if (PredHasWeights || SuccHasWeights) { |
1273 | WeightsForHandled[PredCases[i].Value] = Weights[i + 1]; |
1274 | std::swap(a&: Weights[i + 1], b&: Weights.back()); |
1275 | Weights.pop_back(); |
1276 | } |
1277 | |
1278 | std::swap(a&: PredCases[i], b&: PredCases.back()); |
1279 | PredCases.pop_back(); |
1280 | --i; |
1281 | --e; |
1282 | } |
1283 | |
1284 | // Okay, now we know which constants were sent to BB from the |
1285 | // predecessor. Figure out where they will all go now. |
1286 | for (unsigned i = 0, e = BBCases.size(); i != e; ++i) |
1287 | if (PTIHandled.count(x: BBCases[i].Value)) { |
1288 | // If this is one we are capable of getting... |
1289 | if (PredHasWeights || SuccHasWeights) |
1290 | Weights.push_back(Elt: WeightsForHandled[BBCases[i].Value]); |
1291 | PredCases.push_back(x: BBCases[i]); |
1292 | ++NewSuccessors[BBCases[i].Dest]; |
1293 | PTIHandled.erase(x: BBCases[i].Value); // This constant is taken care of |
1294 | } |
1295 | |
1296 | // If there are any constants vectored to BB that TI doesn't handle, |
1297 | // they must go to the default destination of TI. |
1298 | for (ConstantInt *I : PTIHandled) { |
1299 | if (PredHasWeights || SuccHasWeights) |
1300 | Weights.push_back(Elt: WeightsForHandled[I]); |
1301 | PredCases.push_back(x: ValueEqualityComparisonCase(I, BBDefault)); |
1302 | ++NewSuccessors[BBDefault]; |
1303 | } |
1304 | } |
1305 | |
1306 | // Okay, at this point, we know which new successor Pred will get. Make |
1307 | // sure we update the number of entries in the PHI nodes for these |
1308 | // successors. |
1309 | SmallPtrSet<BasicBlock *, 2> SuccsOfPred; |
1310 | if (DTU) { |
1311 | SuccsOfPred = {succ_begin(BB: Pred), succ_end(BB: Pred)}; |
1312 | Updates.reserve(N: Updates.size() + NewSuccessors.size()); |
1313 | } |
1314 | for (const std::pair<BasicBlock *, int /*Num*/> &NewSuccessor : |
1315 | NewSuccessors) { |
1316 | for (auto I : seq(Size: NewSuccessor.second)) { |
1317 | (void)I; |
1318 | AddPredecessorToBlock(Succ: NewSuccessor.first, NewPred: Pred, ExistPred: BB); |
1319 | } |
1320 | if (DTU && !SuccsOfPred.contains(Ptr: NewSuccessor.first)) |
1321 | Updates.push_back(Elt: {DominatorTree::Insert, Pred, NewSuccessor.first}); |
1322 | } |
1323 | |
1324 | Builder.SetInsertPoint(PTI); |
1325 | // Convert pointer to int before we switch. |
1326 | if (CV->getType()->isPointerTy()) { |
1327 | CV = |
1328 | Builder.CreatePtrToInt(V: CV, DestTy: DL.getIntPtrType(CV->getType()), Name: "magicptr" ); |
1329 | } |
1330 | |
1331 | // Now that the successors are updated, create the new Switch instruction. |
1332 | SwitchInst *NewSI = Builder.CreateSwitch(V: CV, Dest: PredDefault, NumCases: PredCases.size()); |
1333 | NewSI->setDebugLoc(PTI->getDebugLoc()); |
1334 | for (ValueEqualityComparisonCase &V : PredCases) |
1335 | NewSI->addCase(OnVal: V.Value, Dest: V.Dest); |
1336 | |
1337 | if (PredHasWeights || SuccHasWeights) { |
1338 | // Halve the weights if any of them cannot fit in an uint32_t |
1339 | FitWeights(Weights); |
1340 | |
1341 | SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end()); |
1342 | |
1343 | setBranchWeights(SI: NewSI, Weights: MDWeights, /*IsExpected=*/false); |
1344 | } |
1345 | |
1346 | EraseTerminatorAndDCECond(TI: PTI); |
1347 | |
1348 | // Okay, last check. If BB is still a successor of PSI, then we must |
1349 | // have an infinite loop case. If so, add an infinitely looping block |
1350 | // to handle the case to preserve the behavior of the code. |
1351 | BasicBlock *InfLoopBlock = nullptr; |
1352 | for (unsigned i = 0, e = NewSI->getNumSuccessors(); i != e; ++i) |
1353 | if (NewSI->getSuccessor(idx: i) == BB) { |
1354 | if (!InfLoopBlock) { |
1355 | // Insert it at the end of the function, because it's either code, |
1356 | // or it won't matter if it's hot. :) |
1357 | InfLoopBlock = |
1358 | BasicBlock::Create(Context&: BB->getContext(), Name: "infloop" , Parent: BB->getParent()); |
1359 | BranchInst::Create(IfTrue: InfLoopBlock, InsertBefore: InfLoopBlock); |
1360 | if (DTU) |
1361 | Updates.push_back( |
1362 | Elt: {DominatorTree::Insert, InfLoopBlock, InfLoopBlock}); |
1363 | } |
1364 | NewSI->setSuccessor(idx: i, NewSucc: InfLoopBlock); |
1365 | } |
1366 | |
1367 | if (DTU) { |
1368 | if (InfLoopBlock) |
1369 | Updates.push_back(Elt: {DominatorTree::Insert, Pred, InfLoopBlock}); |
1370 | |
1371 | Updates.push_back(Elt: {DominatorTree::Delete, Pred, BB}); |
1372 | |
1373 | DTU->applyUpdates(Updates); |
1374 | } |
1375 | |
1376 | ++NumFoldValueComparisonIntoPredecessors; |
1377 | return true; |
1378 | } |
1379 | |
1380 | /// The specified terminator is a value equality comparison instruction |
1381 | /// (either a switch or a branch on "X == c"). |
1382 | /// See if any of the predecessors of the terminator block are value comparisons |
1383 | /// on the same value. If so, and if safe to do so, fold them together. |
1384 | bool SimplifyCFGOpt::FoldValueComparisonIntoPredecessors(Instruction *TI, |
1385 | IRBuilder<> &Builder) { |
1386 | BasicBlock *BB = TI->getParent(); |
1387 | Value *CV = isValueEqualityComparison(TI); // CondVal |
1388 | assert(CV && "Not a comparison?" ); |
1389 | |
1390 | bool Changed = false; |
1391 | |
1392 | SmallSetVector<BasicBlock *, 16> Preds(pred_begin(BB), pred_end(BB)); |
1393 | while (!Preds.empty()) { |
1394 | BasicBlock *Pred = Preds.pop_back_val(); |
1395 | Instruction *PTI = Pred->getTerminator(); |
1396 | |
1397 | // Don't try to fold into itself. |
1398 | if (Pred == BB) |
1399 | continue; |
1400 | |
1401 | // See if the predecessor is a comparison with the same value. |
1402 | Value *PCV = isValueEqualityComparison(TI: PTI); // PredCondVal |
1403 | if (PCV != CV) |
1404 | continue; |
1405 | |
1406 | SmallSetVector<BasicBlock *, 4> FailBlocks; |
1407 | if (!SafeToMergeTerminators(SI1: TI, SI2: PTI, FailBlocks: &FailBlocks)) { |
1408 | for (auto *Succ : FailBlocks) { |
1409 | if (!SplitBlockPredecessors(BB: Succ, Preds: TI->getParent(), Suffix: ".fold.split" , DTU)) |
1410 | return false; |
1411 | } |
1412 | } |
1413 | |
1414 | PerformValueComparisonIntoPredecessorFolding(TI, CV, PTI, Builder); |
1415 | Changed = true; |
1416 | } |
1417 | return Changed; |
1418 | } |
1419 | |
1420 | // If we would need to insert a select that uses the value of this invoke |
1421 | // (comments in hoistSuccIdenticalTerminatorToSwitchOrIf explain why we would |
1422 | // need to do this), we can't hoist the invoke, as there is nowhere to put the |
1423 | // select in this case. |
1424 | static bool isSafeToHoistInvoke(BasicBlock *BB1, BasicBlock *BB2, |
1425 | Instruction *I1, Instruction *I2) { |
1426 | for (BasicBlock *Succ : successors(BB: BB1)) { |
1427 | for (const PHINode &PN : Succ->phis()) { |
1428 | Value *BB1V = PN.getIncomingValueForBlock(BB: BB1); |
1429 | Value *BB2V = PN.getIncomingValueForBlock(BB: BB2); |
1430 | if (BB1V != BB2V && (BB1V == I1 || BB2V == I2)) { |
1431 | return false; |
1432 | } |
1433 | } |
1434 | } |
1435 | return true; |
1436 | } |
1437 | |
1438 | // Get interesting characteristics of instructions that |
1439 | // `hoistCommonCodeFromSuccessors` didn't hoist. They restrict what kind of |
1440 | // instructions can be reordered across. |
1441 | enum SkipFlags { |
1442 | SkipReadMem = 1, |
1443 | SkipSideEffect = 2, |
1444 | SkipImplicitControlFlow = 4 |
1445 | }; |
1446 | |
1447 | static unsigned skippedInstrFlags(Instruction *I) { |
1448 | unsigned Flags = 0; |
1449 | if (I->mayReadFromMemory()) |
1450 | Flags |= SkipReadMem; |
1451 | // We can't arbitrarily move around allocas, e.g. moving allocas (especially |
1452 | // inalloca) across stacksave/stackrestore boundaries. |
1453 | if (I->mayHaveSideEffects() || isa<AllocaInst>(Val: I)) |
1454 | Flags |= SkipSideEffect; |
1455 | if (!isGuaranteedToTransferExecutionToSuccessor(I)) |
1456 | Flags |= SkipImplicitControlFlow; |
1457 | return Flags; |
1458 | } |
1459 | |
1460 | // Returns true if it is safe to reorder an instruction across preceding |
1461 | // instructions in a basic block. |
1462 | static bool isSafeToHoistInstr(Instruction *I, unsigned Flags) { |
1463 | // Don't reorder a store over a load. |
1464 | if ((Flags & SkipReadMem) && I->mayWriteToMemory()) |
1465 | return false; |
1466 | |
1467 | // If we have seen an instruction with side effects, it's unsafe to reorder an |
1468 | // instruction which reads memory or itself has side effects. |
1469 | if ((Flags & SkipSideEffect) && |
1470 | (I->mayReadFromMemory() || I->mayHaveSideEffects() || isa<AllocaInst>(Val: I))) |
1471 | return false; |
1472 | |
1473 | // Reordering across an instruction which does not necessarily transfer |
1474 | // control to the next instruction is speculation. |
1475 | if ((Flags & SkipImplicitControlFlow) && !isSafeToSpeculativelyExecute(I)) |
1476 | return false; |
1477 | |
1478 | // Hoisting of llvm.deoptimize is only legal together with the next return |
1479 | // instruction, which this pass is not always able to do. |
1480 | if (auto *CB = dyn_cast<CallBase>(Val: I)) |
1481 | if (CB->getIntrinsicID() == Intrinsic::experimental_deoptimize) |
1482 | return false; |
1483 | |
1484 | // It's also unsafe/illegal to hoist an instruction above its instruction |
1485 | // operands |
1486 | BasicBlock *BB = I->getParent(); |
1487 | for (Value *Op : I->operands()) { |
1488 | if (auto *J = dyn_cast<Instruction>(Val: Op)) |
1489 | if (J->getParent() == BB) |
1490 | return false; |
1491 | } |
1492 | |
1493 | return true; |
1494 | } |
1495 | |
1496 | static bool passingValueIsAlwaysUndefined(Value *V, Instruction *I, bool PtrValueMayBeModified = false); |
1497 | |
1498 | /// Helper function for hoistCommonCodeFromSuccessors. Return true if identical |
1499 | /// instructions \p I1 and \p I2 can and should be hoisted. |
1500 | static bool shouldHoistCommonInstructions(Instruction *I1, Instruction *I2, |
1501 | const TargetTransformInfo &TTI) { |
1502 | // If we're going to hoist a call, make sure that the two instructions |
1503 | // we're commoning/hoisting are both marked with musttail, or neither of |
1504 | // them is marked as such. Otherwise, we might end up in a situation where |
1505 | // we hoist from a block where the terminator is a `ret` to a block where |
1506 | // the terminator is a `br`, and `musttail` calls expect to be followed by |
1507 | // a return. |
1508 | auto *C1 = dyn_cast<CallInst>(Val: I1); |
1509 | auto *C2 = dyn_cast<CallInst>(Val: I2); |
1510 | if (C1 && C2) |
1511 | if (C1->isMustTailCall() != C2->isMustTailCall()) |
1512 | return false; |
1513 | |
1514 | if (!TTI.isProfitableToHoist(I: I1) || !TTI.isProfitableToHoist(I: I2)) |
1515 | return false; |
1516 | |
1517 | // If any of the two call sites has nomerge or convergent attribute, stop |
1518 | // hoisting. |
1519 | if (const auto *CB1 = dyn_cast<CallBase>(Val: I1)) |
1520 | if (CB1->cannotMerge() || CB1->isConvergent()) |
1521 | return false; |
1522 | if (const auto *CB2 = dyn_cast<CallBase>(Val: I2)) |
1523 | if (CB2->cannotMerge() || CB2->isConvergent()) |
1524 | return false; |
1525 | |
1526 | return true; |
1527 | } |
1528 | |
1529 | /// Hoists DbgVariableRecords from \p I1 and \p OtherInstrs that are identical |
1530 | /// in lock-step to \p TI. This matches how dbg.* intrinsics are hoisting in |
1531 | /// hoistCommonCodeFromSuccessors. e.g. The input: |
1532 | /// I1 DVRs: { x, z }, |
1533 | /// OtherInsts: { I2 DVRs: { x, y, z } } |
1534 | /// would result in hoisting only DbgVariableRecord x. |
1535 | static void hoistLockstepIdenticalDbgVariableRecords( |
1536 | Instruction *TI, Instruction *I1, |
1537 | SmallVectorImpl<Instruction *> &OtherInsts) { |
1538 | if (!I1->hasDbgRecords()) |
1539 | return; |
1540 | using CurrentAndEndIt = |
1541 | std::pair<DbgRecord::self_iterator, DbgRecord::self_iterator>; |
1542 | // Vector of {Current, End} iterators. |
1543 | SmallVector<CurrentAndEndIt> Itrs; |
1544 | Itrs.reserve(N: OtherInsts.size() + 1); |
1545 | // Helper lambdas for lock-step checks: |
1546 | // Return true if this Current == End. |
1547 | auto atEnd = [](const CurrentAndEndIt &Pair) { |
1548 | return Pair.first == Pair.second; |
1549 | }; |
1550 | // Return true if all Current are identical. |
1551 | auto allIdentical = [](const SmallVector<CurrentAndEndIt> &Itrs) { |
1552 | return all_of(Range: make_first_range(c: ArrayRef(Itrs).drop_front()), |
1553 | P: [&](DbgRecord::self_iterator I) { |
1554 | return Itrs[0].first->isIdenticalToWhenDefined(R: *I); |
1555 | }); |
1556 | }; |
1557 | |
1558 | // Collect the iterators. |
1559 | Itrs.push_back( |
1560 | Elt: {I1->getDbgRecordRange().begin(), I1->getDbgRecordRange().end()}); |
1561 | for (Instruction *Other : OtherInsts) { |
1562 | if (!Other->hasDbgRecords()) |
1563 | return; |
1564 | Itrs.push_back( |
1565 | Elt: {Other->getDbgRecordRange().begin(), Other->getDbgRecordRange().end()}); |
1566 | } |
1567 | |
1568 | // Iterate in lock-step until any of the DbgRecord lists are exausted. If |
1569 | // the lock-step DbgRecord are identical, hoist all of them to TI. |
1570 | // This replicates the dbg.* intrinsic behaviour in |
1571 | // hoistCommonCodeFromSuccessors. |
1572 | while (none_of(Range&: Itrs, P: atEnd)) { |
1573 | bool HoistDVRs = allIdentical(Itrs); |
1574 | for (CurrentAndEndIt &Pair : Itrs) { |
1575 | // Increment Current iterator now as we may be about to move the |
1576 | // DbgRecord. |
1577 | DbgRecord &DR = *Pair.first++; |
1578 | if (HoistDVRs) { |
1579 | DR.removeFromParent(); |
1580 | TI->getParent()->insertDbgRecordBefore(DR: &DR, Here: TI->getIterator()); |
1581 | } |
1582 | } |
1583 | } |
1584 | } |
1585 | |
1586 | /// Hoist any common code in the successor blocks up into the block. This |
1587 | /// function guarantees that BB dominates all successors. If EqTermsOnly is |
1588 | /// given, only perform hoisting in case both blocks only contain a terminator. |
1589 | /// In that case, only the original BI will be replaced and selects for PHIs are |
1590 | /// added. |
1591 | bool SimplifyCFGOpt::hoistCommonCodeFromSuccessors(BasicBlock *BB, |
1592 | bool EqTermsOnly) { |
1593 | // This does very trivial matching, with limited scanning, to find identical |
1594 | // instructions in the two blocks. In particular, we don't want to get into |
1595 | // O(N1*N2*...) situations here where Ni are the sizes of these successors. As |
1596 | // such, we currently just scan for obviously identical instructions in an |
1597 | // identical order, possibly separated by the same number of non-identical |
1598 | // instructions. |
1599 | unsigned int SuccSize = succ_size(BB); |
1600 | if (SuccSize < 2) |
1601 | return false; |
1602 | |
1603 | // If either of the blocks has it's address taken, then we can't do this fold, |
1604 | // because the code we'd hoist would no longer run when we jump into the block |
1605 | // by it's address. |
1606 | for (auto *Succ : successors(BB)) |
1607 | if (Succ->hasAddressTaken() || !Succ->getSinglePredecessor()) |
1608 | return false; |
1609 | |
1610 | auto *TI = BB->getTerminator(); |
1611 | |
1612 | // The second of pair is a SkipFlags bitmask. |
1613 | using SuccIterPair = std::pair<BasicBlock::iterator, unsigned>; |
1614 | SmallVector<SuccIterPair, 8> SuccIterPairs; |
1615 | for (auto *Succ : successors(BB)) { |
1616 | BasicBlock::iterator SuccItr = Succ->begin(); |
1617 | if (isa<PHINode>(Val: *SuccItr)) |
1618 | return false; |
1619 | SuccIterPairs.push_back(Elt: SuccIterPair(SuccItr, 0)); |
1620 | } |
1621 | |
1622 | // Check if only hoisting terminators is allowed. This does not add new |
1623 | // instructions to the hoist location. |
1624 | if (EqTermsOnly) { |
1625 | // Skip any debug intrinsics, as they are free to hoist. |
1626 | for (auto &SuccIter : make_first_range(c&: SuccIterPairs)) { |
1627 | auto *INonDbg = &*skipDebugIntrinsics(It: SuccIter); |
1628 | if (!INonDbg->isTerminator()) |
1629 | return false; |
1630 | } |
1631 | // Now we know that we only need to hoist debug intrinsics and the |
1632 | // terminator. Let the loop below handle those 2 cases. |
1633 | } |
1634 | |
1635 | // Count how many instructions were not hoisted so far. There's a limit on how |
1636 | // many instructions we skip, serving as a compilation time control as well as |
1637 | // preventing excessive increase of life ranges. |
1638 | unsigned NumSkipped = 0; |
1639 | // If we find an unreachable instruction at the beginning of a basic block, we |
1640 | // can still hoist instructions from the rest of the basic blocks. |
1641 | if (SuccIterPairs.size() > 2) { |
1642 | erase_if(C&: SuccIterPairs, |
1643 | P: [](const auto &Pair) { return isa<UnreachableInst>(Pair.first); }); |
1644 | if (SuccIterPairs.size() < 2) |
1645 | return false; |
1646 | } |
1647 | |
1648 | bool Changed = false; |
1649 | |
1650 | for (;;) { |
1651 | auto *SuccIterPairBegin = SuccIterPairs.begin(); |
1652 | auto &BB1ItrPair = *SuccIterPairBegin++; |
1653 | auto OtherSuccIterPairRange = |
1654 | iterator_range(SuccIterPairBegin, SuccIterPairs.end()); |
1655 | auto OtherSuccIterRange = make_first_range(c&: OtherSuccIterPairRange); |
1656 | |
1657 | Instruction *I1 = &*BB1ItrPair.first; |
1658 | |
1659 | // Skip debug info if it is not identical. |
1660 | bool AllDbgInstsAreIdentical = all_of(Range&: OtherSuccIterRange, P: [I1](auto &Iter) { |
1661 | Instruction *I2 = &*Iter; |
1662 | return I1->isIdenticalToWhenDefined(I: I2); |
1663 | }); |
1664 | if (!AllDbgInstsAreIdentical) { |
1665 | while (isa<DbgInfoIntrinsic>(Val: I1)) |
1666 | I1 = &*++BB1ItrPair.first; |
1667 | for (auto &SuccIter : OtherSuccIterRange) { |
1668 | Instruction *I2 = &*SuccIter; |
1669 | while (isa<DbgInfoIntrinsic>(Val: I2)) |
1670 | I2 = &*++SuccIter; |
1671 | } |
1672 | } |
1673 | |
1674 | bool AllInstsAreIdentical = true; |
1675 | bool HasTerminator = I1->isTerminator(); |
1676 | for (auto &SuccIter : OtherSuccIterRange) { |
1677 | Instruction *I2 = &*SuccIter; |
1678 | HasTerminator |= I2->isTerminator(); |
1679 | if (AllInstsAreIdentical && (!I1->isIdenticalToWhenDefined(I: I2) || |
1680 | MMRAMetadata(*I1) != MMRAMetadata(*I2))) |
1681 | AllInstsAreIdentical = false; |
1682 | } |
1683 | |
1684 | SmallVector<Instruction *, 8> OtherInsts; |
1685 | for (auto &SuccIter : OtherSuccIterRange) |
1686 | OtherInsts.push_back(Elt: &*SuccIter); |
1687 | |
1688 | // If we are hoisting the terminator instruction, don't move one (making a |
1689 | // broken BB), instead clone it, and remove BI. |
1690 | if (HasTerminator) { |
1691 | // Even if BB, which contains only one unreachable instruction, is ignored |
1692 | // at the beginning of the loop, we can hoist the terminator instruction. |
1693 | // If any instructions remain in the block, we cannot hoist terminators. |
1694 | if (NumSkipped || !AllInstsAreIdentical) { |
1695 | hoistLockstepIdenticalDbgVariableRecords(TI, I1, OtherInsts); |
1696 | return Changed; |
1697 | } |
1698 | |
1699 | return hoistSuccIdenticalTerminatorToSwitchOrIf(TI, I1, OtherSuccTIs&: OtherInsts) || |
1700 | Changed; |
1701 | } |
1702 | |
1703 | if (AllInstsAreIdentical) { |
1704 | unsigned SkipFlagsBB1 = BB1ItrPair.second; |
1705 | AllInstsAreIdentical = |
1706 | isSafeToHoistInstr(I: I1, Flags: SkipFlagsBB1) && |
1707 | all_of(Range&: OtherSuccIterPairRange, P: [=](const auto &Pair) { |
1708 | Instruction *I2 = &*Pair.first; |
1709 | unsigned SkipFlagsBB2 = Pair.second; |
1710 | // Even if the instructions are identical, it may not |
1711 | // be safe to hoist them if we have skipped over |
1712 | // instructions with side effects or their operands |
1713 | // weren't hoisted. |
1714 | return isSafeToHoistInstr(I: I2, Flags: SkipFlagsBB2) && |
1715 | shouldHoistCommonInstructions(I1, I2, TTI); |
1716 | }); |
1717 | } |
1718 | |
1719 | if (AllInstsAreIdentical) { |
1720 | BB1ItrPair.first++; |
1721 | if (isa<DbgInfoIntrinsic>(Val: I1)) { |
1722 | // The debug location is an integral part of a debug info intrinsic |
1723 | // and can't be separated from it or replaced. Instead of attempting |
1724 | // to merge locations, simply hoist both copies of the intrinsic. |
1725 | hoistLockstepIdenticalDbgVariableRecords(TI, I1, OtherInsts); |
1726 | // We've just hoisted DbgVariableRecords; move I1 after them (before TI) |
1727 | // and leave any that were not hoisted behind (by calling moveBefore |
1728 | // rather than moveBeforePreserving). |
1729 | I1->moveBefore(MovePos: TI); |
1730 | for (auto &SuccIter : OtherSuccIterRange) { |
1731 | auto *I2 = &*SuccIter++; |
1732 | assert(isa<DbgInfoIntrinsic>(I2)); |
1733 | I2->moveBefore(MovePos: TI); |
1734 | } |
1735 | } else { |
1736 | // For a normal instruction, we just move one to right before the |
1737 | // branch, then replace all uses of the other with the first. Finally, |
1738 | // we remove the now redundant second instruction. |
1739 | hoistLockstepIdenticalDbgVariableRecords(TI, I1, OtherInsts); |
1740 | // We've just hoisted DbgVariableRecords; move I1 after them (before TI) |
1741 | // and leave any that were not hoisted behind (by calling moveBefore |
1742 | // rather than moveBeforePreserving). |
1743 | I1->moveBefore(MovePos: TI); |
1744 | for (auto &SuccIter : OtherSuccIterRange) { |
1745 | Instruction *I2 = &*SuccIter++; |
1746 | assert(I2 != I1); |
1747 | if (!I2->use_empty()) |
1748 | I2->replaceAllUsesWith(V: I1); |
1749 | I1->andIRFlags(V: I2); |
1750 | combineMetadataForCSE(K: I1, J: I2, DoesKMove: true); |
1751 | // I1 and I2 are being combined into a single instruction. Its debug |
1752 | // location is the merged locations of the original instructions. |
1753 | I1->applyMergedLocation(LocA: I1->getDebugLoc(), LocB: I2->getDebugLoc()); |
1754 | I2->eraseFromParent(); |
1755 | } |
1756 | } |
1757 | if (!Changed) |
1758 | NumHoistCommonCode += SuccIterPairs.size(); |
1759 | Changed = true; |
1760 | NumHoistCommonInstrs += SuccIterPairs.size(); |
1761 | } else { |
1762 | if (NumSkipped >= HoistCommonSkipLimit) { |
1763 | hoistLockstepIdenticalDbgVariableRecords(TI, I1, OtherInsts); |
1764 | return Changed; |
1765 | } |
1766 | // We are about to skip over a pair of non-identical instructions. Record |
1767 | // if any have characteristics that would prevent reordering instructions |
1768 | // across them. |
1769 | for (auto &SuccIterPair : SuccIterPairs) { |
1770 | Instruction *I = &*SuccIterPair.first++; |
1771 | SuccIterPair.second |= skippedInstrFlags(I); |
1772 | } |
1773 | ++NumSkipped; |
1774 | } |
1775 | } |
1776 | } |
1777 | |
1778 | bool SimplifyCFGOpt::hoistSuccIdenticalTerminatorToSwitchOrIf( |
1779 | Instruction *TI, Instruction *I1, |
1780 | SmallVectorImpl<Instruction *> &OtherSuccTIs) { |
1781 | |
1782 | auto *BI = dyn_cast<BranchInst>(Val: TI); |
1783 | |
1784 | bool Changed = false; |
1785 | BasicBlock *TIParent = TI->getParent(); |
1786 | BasicBlock *BB1 = I1->getParent(); |
1787 | |
1788 | // Use only for an if statement. |
1789 | auto *I2 = *OtherSuccTIs.begin(); |
1790 | auto *BB2 = I2->getParent(); |
1791 | if (BI) { |
1792 | assert(OtherSuccTIs.size() == 1); |
1793 | assert(BI->getSuccessor(0) == I1->getParent()); |
1794 | assert(BI->getSuccessor(1) == I2->getParent()); |
1795 | } |
1796 | |
1797 | // In the case of an if statement, we try to hoist an invoke. |
1798 | // FIXME: Can we define a safety predicate for CallBr? |
1799 | // FIXME: Test case llvm/test/Transforms/SimplifyCFG/2009-06-15-InvokeCrash.ll |
1800 | // removed in 4c923b3b3fd0ac1edebf0603265ca3ba51724937 commit? |
1801 | if (isa<InvokeInst>(Val: I1) && (!BI || !isSafeToHoistInvoke(BB1, BB2, I1, I2))) |
1802 | return false; |
1803 | |
1804 | // TODO: callbr hoisting currently disabled pending further study. |
1805 | if (isa<CallBrInst>(Val: I1)) |
1806 | return false; |
1807 | |
1808 | for (BasicBlock *Succ : successors(BB: BB1)) { |
1809 | for (PHINode &PN : Succ->phis()) { |
1810 | Value *BB1V = PN.getIncomingValueForBlock(BB: BB1); |
1811 | for (Instruction *OtherSuccTI : OtherSuccTIs) { |
1812 | Value *BB2V = PN.getIncomingValueForBlock(BB: OtherSuccTI->getParent()); |
1813 | if (BB1V == BB2V) |
1814 | continue; |
1815 | |
1816 | // In the case of an if statement, check for |
1817 | // passingValueIsAlwaysUndefined here because we would rather eliminate |
1818 | // undefined control flow then converting it to a select. |
1819 | if (!BI || passingValueIsAlwaysUndefined(V: BB1V, I: &PN) || |
1820 | passingValueIsAlwaysUndefined(V: BB2V, I: &PN)) |
1821 | return false; |
1822 | } |
1823 | } |
1824 | } |
1825 | |
1826 | // Hoist DbgVariableRecords attached to the terminator to match dbg.* |
1827 | // intrinsic hoisting behaviour in hoistCommonCodeFromSuccessors. |
1828 | hoistLockstepIdenticalDbgVariableRecords(TI, I1, OtherInsts&: OtherSuccTIs); |
1829 | // Clone the terminator and hoist it into the pred, without any debug info. |
1830 | Instruction *NT = I1->clone(); |
1831 | NT->insertInto(ParentBB: TIParent, It: TI->getIterator()); |
1832 | if (!NT->getType()->isVoidTy()) { |
1833 | I1->replaceAllUsesWith(V: NT); |
1834 | for (Instruction *OtherSuccTI : OtherSuccTIs) |
1835 | OtherSuccTI->replaceAllUsesWith(V: NT); |
1836 | NT->takeName(V: I1); |
1837 | } |
1838 | Changed = true; |
1839 | NumHoistCommonInstrs += OtherSuccTIs.size() + 1; |
1840 | |
1841 | // Ensure terminator gets a debug location, even an unknown one, in case |
1842 | // it involves inlinable calls. |
1843 | SmallVector<DILocation *, 4> Locs; |
1844 | Locs.push_back(Elt: I1->getDebugLoc()); |
1845 | for (auto *OtherSuccTI : OtherSuccTIs) |
1846 | Locs.push_back(Elt: OtherSuccTI->getDebugLoc()); |
1847 | NT->setDebugLoc(DILocation::getMergedLocations(Locs)); |
1848 | |
1849 | // PHIs created below will adopt NT's merged DebugLoc. |
1850 | IRBuilder<NoFolder> Builder(NT); |
1851 | |
1852 | // In the case of an if statement, hoisting one of the terminators from our |
1853 | // successor is a great thing. Unfortunately, the successors of the if/else |
1854 | // blocks may have PHI nodes in them. If they do, all PHI entries for BB1/BB2 |
1855 | // must agree for all PHI nodes, so we insert select instruction to compute |
1856 | // the final result. |
1857 | if (BI) { |
1858 | std::map<std::pair<Value *, Value *>, SelectInst *> InsertedSelects; |
1859 | for (BasicBlock *Succ : successors(BB: BB1)) { |
1860 | for (PHINode &PN : Succ->phis()) { |
1861 | Value *BB1V = PN.getIncomingValueForBlock(BB: BB1); |
1862 | Value *BB2V = PN.getIncomingValueForBlock(BB: BB2); |
1863 | if (BB1V == BB2V) |
1864 | continue; |
1865 | |
1866 | // These values do not agree. Insert a select instruction before NT |
1867 | // that determines the right value. |
1868 | SelectInst *&SI = InsertedSelects[std::make_pair(x&: BB1V, y&: BB2V)]; |
1869 | if (!SI) { |
1870 | // Propagate fast-math-flags from phi node to its replacement select. |
1871 | IRBuilder<>::FastMathFlagGuard FMFGuard(Builder); |
1872 | if (isa<FPMathOperator>(Val: PN)) |
1873 | Builder.setFastMathFlags(PN.getFastMathFlags()); |
1874 | |
1875 | SI = cast<SelectInst>(Val: Builder.CreateSelect( |
1876 | C: BI->getCondition(), True: BB1V, False: BB2V, |
1877 | Name: BB1V->getName() + "." + BB2V->getName(), MDFrom: BI)); |
1878 | } |
1879 | |
1880 | // Make the PHI node use the select for all incoming values for BB1/BB2 |
1881 | for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i) |
1882 | if (PN.getIncomingBlock(i) == BB1 || PN.getIncomingBlock(i) == BB2) |
1883 | PN.setIncomingValue(i, V: SI); |
1884 | } |
1885 | } |
1886 | } |
1887 | |
1888 | SmallVector<DominatorTree::UpdateType, 4> Updates; |
1889 | |
1890 | // Update any PHI nodes in our new successors. |
1891 | for (BasicBlock *Succ : successors(BB: BB1)) { |
1892 | AddPredecessorToBlock(Succ, NewPred: TIParent, ExistPred: BB1); |
1893 | if (DTU) |
1894 | Updates.push_back(Elt: {DominatorTree::Insert, TIParent, Succ}); |
1895 | } |
1896 | |
1897 | if (DTU) |
1898 | for (BasicBlock *Succ : successors(I: TI)) |
1899 | Updates.push_back(Elt: {DominatorTree::Delete, TIParent, Succ}); |
1900 | |
1901 | EraseTerminatorAndDCECond(TI); |
1902 | if (DTU) |
1903 | DTU->applyUpdates(Updates); |
1904 | return Changed; |
1905 | } |
1906 | |
1907 | // Check lifetime markers. |
1908 | static bool isLifeTimeMarker(const Instruction *I) { |
1909 | if (auto II = dyn_cast<IntrinsicInst>(Val: I)) { |
1910 | switch (II->getIntrinsicID()) { |
1911 | default: |
1912 | break; |
1913 | case Intrinsic::lifetime_start: |
1914 | case Intrinsic::lifetime_end: |
1915 | return true; |
1916 | } |
1917 | } |
1918 | return false; |
1919 | } |
1920 | |
1921 | // TODO: Refine this. This should avoid cases like turning constant memcpy sizes |
1922 | // into variables. |
1923 | static bool replacingOperandWithVariableIsCheap(const Instruction *I, |
1924 | int OpIdx) { |
1925 | return !isa<IntrinsicInst>(Val: I); |
1926 | } |
1927 | |
1928 | // All instructions in Insts belong to different blocks that all unconditionally |
1929 | // branch to a common successor. Analyze each instruction and return true if it |
1930 | // would be possible to sink them into their successor, creating one common |
1931 | // instruction instead. For every value that would be required to be provided by |
1932 | // PHI node (because an operand varies in each input block), add to PHIOperands. |
1933 | static bool canSinkInstructions( |
1934 | ArrayRef<Instruction *> Insts, |
1935 | DenseMap<const Use *, SmallVector<Value *, 4>> &PHIOperands) { |
1936 | // Prune out obviously bad instructions to move. Each instruction must have |
1937 | // the same number of uses, and we check later that the uses are consistent. |
1938 | std::optional<unsigned> NumUses; |
1939 | for (auto *I : Insts) { |
1940 | // These instructions may change or break semantics if moved. |
1941 | if (isa<PHINode>(Val: I) || I->isEHPad() || isa<AllocaInst>(Val: I) || |
1942 | I->getType()->isTokenTy()) |
1943 | return false; |
1944 | |
1945 | // Do not try to sink an instruction in an infinite loop - it can cause |
1946 | // this algorithm to infinite loop. |
1947 | if (I->getParent()->getSingleSuccessor() == I->getParent()) |
1948 | return false; |
1949 | |
1950 | // Conservatively return false if I is an inline-asm instruction. Sinking |
1951 | // and merging inline-asm instructions can potentially create arguments |
1952 | // that cannot satisfy the inline-asm constraints. |
1953 | // If the instruction has nomerge or convergent attribute, return false. |
1954 | if (const auto *C = dyn_cast<CallBase>(Val: I)) |
1955 | if (C->isInlineAsm() || C->cannotMerge() || C->isConvergent()) |
1956 | return false; |
1957 | |
1958 | if (!NumUses) |
1959 | NumUses = I->getNumUses(); |
1960 | else if (NumUses != I->getNumUses()) |
1961 | return false; |
1962 | } |
1963 | |
1964 | const Instruction *I0 = Insts.front(); |
1965 | const auto I0MMRA = MMRAMetadata(*I0); |
1966 | for (auto *I : Insts) { |
1967 | if (!I->isSameOperationAs(I: I0)) |
1968 | return false; |
1969 | |
1970 | // swifterror pointers can only be used by a load or store; sinking a load |
1971 | // or store would require introducing a select for the pointer operand, |
1972 | // which isn't allowed for swifterror pointers. |
1973 | if (isa<StoreInst>(Val: I) && I->getOperand(i: 1)->isSwiftError()) |
1974 | return false; |
1975 | if (isa<LoadInst>(Val: I) && I->getOperand(i: 0)->isSwiftError()) |
1976 | return false; |
1977 | |
1978 | // Treat MMRAs conservatively. This pass can be quite aggressive and |
1979 | // could drop a lot of MMRAs otherwise. |
1980 | if (MMRAMetadata(*I) != I0MMRA) |
1981 | return false; |
1982 | } |
1983 | |
1984 | // Uses must be consistent: If I0 is used in a phi node in the sink target, |
1985 | // then the other phi operands must match the instructions from Insts. This |
1986 | // also has to hold true for any phi nodes that would be created as a result |
1987 | // of sinking. Both of these cases are represented by PhiOperands. |
1988 | for (const Use &U : I0->uses()) { |
1989 | auto It = PHIOperands.find(Val: &U); |
1990 | if (It == PHIOperands.end()) |
1991 | // There may be uses in other blocks when sinking into a loop header. |
1992 | return false; |
1993 | if (!equal(LRange&: Insts, RRange&: It->second)) |
1994 | return false; |
1995 | } |
1996 | |
1997 | // Because SROA can't handle speculating stores of selects, try not to sink |
1998 | // loads, stores or lifetime markers of allocas when we'd have to create a |
1999 | // PHI for the address operand. Also, because it is likely that loads or |
2000 | // stores of allocas will disappear when Mem2Reg/SROA is run, don't sink |
2001 | // them. |
2002 | // This can cause code churn which can have unintended consequences down |
2003 | // the line - see https://llvm.org/bugs/show_bug.cgi?id=30244. |
2004 | // FIXME: This is a workaround for a deficiency in SROA - see |
2005 | // https://llvm.org/bugs/show_bug.cgi?id=30188 |
2006 | if (isa<StoreInst>(Val: I0) && any_of(Range&: Insts, P: [](const Instruction *I) { |
2007 | return isa<AllocaInst>(Val: I->getOperand(i: 1)->stripPointerCasts()); |
2008 | })) |
2009 | return false; |
2010 | if (isa<LoadInst>(Val: I0) && any_of(Range&: Insts, P: [](const Instruction *I) { |
2011 | return isa<AllocaInst>(Val: I->getOperand(i: 0)->stripPointerCasts()); |
2012 | })) |
2013 | return false; |
2014 | if (isLifeTimeMarker(I: I0) && any_of(Range&: Insts, P: [](const Instruction *I) { |
2015 | return isa<AllocaInst>(Val: I->getOperand(i: 1)->stripPointerCasts()); |
2016 | })) |
2017 | return false; |
2018 | |
2019 | // For calls to be sinkable, they must all be indirect, or have same callee. |
2020 | // I.e. if we have two direct calls to different callees, we don't want to |
2021 | // turn that into an indirect call. Likewise, if we have an indirect call, |
2022 | // and a direct call, we don't actually want to have a single indirect call. |
2023 | if (isa<CallBase>(Val: I0)) { |
2024 | auto IsIndirectCall = [](const Instruction *I) { |
2025 | return cast<CallBase>(Val: I)->isIndirectCall(); |
2026 | }; |
2027 | bool HaveIndirectCalls = any_of(Range&: Insts, P: IsIndirectCall); |
2028 | bool AllCallsAreIndirect = all_of(Range&: Insts, P: IsIndirectCall); |
2029 | if (HaveIndirectCalls) { |
2030 | if (!AllCallsAreIndirect) |
2031 | return false; |
2032 | } else { |
2033 | // All callees must be identical. |
2034 | Value *Callee = nullptr; |
2035 | for (const Instruction *I : Insts) { |
2036 | Value *CurrCallee = cast<CallBase>(Val: I)->getCalledOperand(); |
2037 | if (!Callee) |
2038 | Callee = CurrCallee; |
2039 | else if (Callee != CurrCallee) |
2040 | return false; |
2041 | } |
2042 | } |
2043 | } |
2044 | |
2045 | for (unsigned OI = 0, OE = I0->getNumOperands(); OI != OE; ++OI) { |
2046 | Value *Op = I0->getOperand(i: OI); |
2047 | if (Op->getType()->isTokenTy()) |
2048 | // Don't touch any operand of token type. |
2049 | return false; |
2050 | |
2051 | auto SameAsI0 = [&I0, OI](const Instruction *I) { |
2052 | assert(I->getNumOperands() == I0->getNumOperands()); |
2053 | return I->getOperand(i: OI) == I0->getOperand(i: OI); |
2054 | }; |
2055 | if (!all_of(Range&: Insts, P: SameAsI0)) { |
2056 | if ((isa<Constant>(Val: Op) && !replacingOperandWithVariableIsCheap(I: I0, OpIdx: OI)) || |
2057 | !canReplaceOperandWithVariable(I: I0, OpIdx: OI)) |
2058 | // We can't create a PHI from this GEP. |
2059 | return false; |
2060 | auto &Ops = PHIOperands[&I0->getOperandUse(i: OI)]; |
2061 | for (auto *I : Insts) |
2062 | Ops.push_back(Elt: I->getOperand(i: OI)); |
2063 | } |
2064 | } |
2065 | return true; |
2066 | } |
2067 | |
2068 | // Assuming canSinkInstructions(Blocks) has returned true, sink the last |
2069 | // instruction of every block in Blocks to their common successor, commoning |
2070 | // into one instruction. |
2071 | static void sinkLastInstruction(ArrayRef<BasicBlock*> Blocks) { |
2072 | auto *BBEnd = Blocks[0]->getTerminator()->getSuccessor(Idx: 0); |
2073 | |
2074 | // canSinkInstructions returning true guarantees that every block has at |
2075 | // least one non-terminator instruction. |
2076 | SmallVector<Instruction*,4> Insts; |
2077 | for (auto *BB : Blocks) { |
2078 | Instruction *I = BB->getTerminator(); |
2079 | do { |
2080 | I = I->getPrevNode(); |
2081 | } while (isa<DbgInfoIntrinsic>(Val: I) && I != &BB->front()); |
2082 | if (!isa<DbgInfoIntrinsic>(Val: I)) |
2083 | Insts.push_back(Elt: I); |
2084 | } |
2085 | |
2086 | // We don't need to do any more checking here; canSinkInstructions should |
2087 | // have done it all for us. |
2088 | SmallVector<Value*, 4> NewOperands; |
2089 | Instruction *I0 = Insts.front(); |
2090 | for (unsigned O = 0, E = I0->getNumOperands(); O != E; ++O) { |
2091 | // This check is different to that in canSinkInstructions. There, we |
2092 | // cared about the global view once simplifycfg (and instcombine) have |
2093 | // completed - it takes into account PHIs that become trivially |
2094 | // simplifiable. However here we need a more local view; if an operand |
2095 | // differs we create a PHI and rely on instcombine to clean up the very |
2096 | // small mess we may make. |
2097 | bool NeedPHI = any_of(Range&: Insts, P: [&I0, O](const Instruction *I) { |
2098 | return I->getOperand(i: O) != I0->getOperand(i: O); |
2099 | }); |
2100 | if (!NeedPHI) { |
2101 | NewOperands.push_back(Elt: I0->getOperand(i: O)); |
2102 | continue; |
2103 | } |
2104 | |
2105 | // Create a new PHI in the successor block and populate it. |
2106 | auto *Op = I0->getOperand(i: O); |
2107 | assert(!Op->getType()->isTokenTy() && "Can't PHI tokens!" ); |
2108 | auto *PN = |
2109 | PHINode::Create(Ty: Op->getType(), NumReservedValues: Insts.size(), NameStr: Op->getName() + ".sink" ); |
2110 | PN->insertBefore(InsertPos: BBEnd->begin()); |
2111 | for (auto *I : Insts) |
2112 | PN->addIncoming(V: I->getOperand(i: O), BB: I->getParent()); |
2113 | NewOperands.push_back(Elt: PN); |
2114 | } |
2115 | |
2116 | // Arbitrarily use I0 as the new "common" instruction; remap its operands |
2117 | // and move it to the start of the successor block. |
2118 | for (unsigned O = 0, E = I0->getNumOperands(); O != E; ++O) |
2119 | I0->getOperandUse(i: O).set(NewOperands[O]); |
2120 | |
2121 | I0->moveBefore(BB&: *BBEnd, I: BBEnd->getFirstInsertionPt()); |
2122 | |
2123 | // Update metadata and IR flags, and merge debug locations. |
2124 | for (auto *I : Insts) |
2125 | if (I != I0) { |
2126 | // The debug location for the "common" instruction is the merged locations |
2127 | // of all the commoned instructions. We start with the original location |
2128 | // of the "common" instruction and iteratively merge each location in the |
2129 | // loop below. |
2130 | // This is an N-way merge, which will be inefficient if I0 is a CallInst. |
2131 | // However, as N-way merge for CallInst is rare, so we use simplified API |
2132 | // instead of using complex API for N-way merge. |
2133 | I0->applyMergedLocation(LocA: I0->getDebugLoc(), LocB: I->getDebugLoc()); |
2134 | combineMetadataForCSE(K: I0, J: I, DoesKMove: true); |
2135 | I0->andIRFlags(V: I); |
2136 | } |
2137 | |
2138 | for (User *U : make_early_inc_range(Range: I0->users())) { |
2139 | // canSinkLastInstruction checked that all instructions are only used by |
2140 | // phi nodes in a way that allows replacing the phi node with the common |
2141 | // instruction. |
2142 | auto *PN = cast<PHINode>(Val: U); |
2143 | PN->replaceAllUsesWith(V: I0); |
2144 | PN->eraseFromParent(); |
2145 | } |
2146 | |
2147 | // Finally nuke all instructions apart from the common instruction. |
2148 | for (auto *I : Insts) { |
2149 | if (I == I0) |
2150 | continue; |
2151 | // The remaining uses are debug users, replace those with the common inst. |
2152 | // In most (all?) cases this just introduces a use-before-def. |
2153 | assert(I->user_empty() && "Inst unexpectedly still has non-dbg users" ); |
2154 | I->replaceAllUsesWith(V: I0); |
2155 | I->eraseFromParent(); |
2156 | } |
2157 | } |
2158 | |
2159 | namespace { |
2160 | |
2161 | // LockstepReverseIterator - Iterates through instructions |
2162 | // in a set of blocks in reverse order from the first non-terminator. |
2163 | // For example (assume all blocks have size n): |
2164 | // LockstepReverseIterator I([B1, B2, B3]); |
2165 | // *I-- = [B1[n], B2[n], B3[n]]; |
2166 | // *I-- = [B1[n-1], B2[n-1], B3[n-1]]; |
2167 | // *I-- = [B1[n-2], B2[n-2], B3[n-2]]; |
2168 | // ... |
2169 | class LockstepReverseIterator { |
2170 | ArrayRef<BasicBlock*> Blocks; |
2171 | SmallVector<Instruction*,4> Insts; |
2172 | bool Fail; |
2173 | |
2174 | public: |
2175 | LockstepReverseIterator(ArrayRef<BasicBlock*> Blocks) : Blocks(Blocks) { |
2176 | reset(); |
2177 | } |
2178 | |
2179 | void reset() { |
2180 | Fail = false; |
2181 | Insts.clear(); |
2182 | for (auto *BB : Blocks) { |
2183 | Instruction *Inst = BB->getTerminator(); |
2184 | for (Inst = Inst->getPrevNode(); Inst && isa<DbgInfoIntrinsic>(Val: Inst);) |
2185 | Inst = Inst->getPrevNode(); |
2186 | if (!Inst) { |
2187 | // Block wasn't big enough. |
2188 | Fail = true; |
2189 | return; |
2190 | } |
2191 | Insts.push_back(Elt: Inst); |
2192 | } |
2193 | } |
2194 | |
2195 | bool isValid() const { |
2196 | return !Fail; |
2197 | } |
2198 | |
2199 | void operator--() { |
2200 | if (Fail) |
2201 | return; |
2202 | for (auto *&Inst : Insts) { |
2203 | for (Inst = Inst->getPrevNode(); Inst && isa<DbgInfoIntrinsic>(Val: Inst);) |
2204 | Inst = Inst->getPrevNode(); |
2205 | // Already at beginning of block. |
2206 | if (!Inst) { |
2207 | Fail = true; |
2208 | return; |
2209 | } |
2210 | } |
2211 | } |
2212 | |
2213 | void operator++() { |
2214 | if (Fail) |
2215 | return; |
2216 | for (auto *&Inst : Insts) { |
2217 | for (Inst = Inst->getNextNode(); Inst && isa<DbgInfoIntrinsic>(Val: Inst);) |
2218 | Inst = Inst->getNextNode(); |
2219 | // Already at end of block. |
2220 | if (!Inst) { |
2221 | Fail = true; |
2222 | return; |
2223 | } |
2224 | } |
2225 | } |
2226 | |
2227 | ArrayRef<Instruction*> operator * () const { |
2228 | return Insts; |
2229 | } |
2230 | }; |
2231 | |
2232 | } // end anonymous namespace |
2233 | |
2234 | /// Check whether BB's predecessors end with unconditional branches. If it is |
2235 | /// true, sink any common code from the predecessors to BB. |
2236 | static bool SinkCommonCodeFromPredecessors(BasicBlock *BB, |
2237 | DomTreeUpdater *DTU) { |
2238 | // We support two situations: |
2239 | // (1) all incoming arcs are unconditional |
2240 | // (2) there are non-unconditional incoming arcs |
2241 | // |
2242 | // (2) is very common in switch defaults and |
2243 | // else-if patterns; |
2244 | // |
2245 | // if (a) f(1); |
2246 | // else if (b) f(2); |
2247 | // |
2248 | // produces: |
2249 | // |
2250 | // [if] |
2251 | // / \ |
2252 | // [f(1)] [if] |
2253 | // | | \ |
2254 | // | | | |
2255 | // | [f(2)]| |
2256 | // \ | / |
2257 | // [ end ] |
2258 | // |
2259 | // [end] has two unconditional predecessor arcs and one conditional. The |
2260 | // conditional refers to the implicit empty 'else' arc. This conditional |
2261 | // arc can also be caused by an empty default block in a switch. |
2262 | // |
2263 | // In this case, we attempt to sink code from all *unconditional* arcs. |
2264 | // If we can sink instructions from these arcs (determined during the scan |
2265 | // phase below) we insert a common successor for all unconditional arcs and |
2266 | // connect that to [end], to enable sinking: |
2267 | // |
2268 | // [if] |
2269 | // / \ |
2270 | // [x(1)] [if] |
2271 | // | | \ |
2272 | // | | \ |
2273 | // | [x(2)] | |
2274 | // \ / | |
2275 | // [sink.split] | |
2276 | // \ / |
2277 | // [ end ] |
2278 | // |
2279 | SmallVector<BasicBlock*,4> UnconditionalPreds; |
2280 | bool HaveNonUnconditionalPredecessors = false; |
2281 | for (auto *PredBB : predecessors(BB)) { |
2282 | auto *PredBr = dyn_cast<BranchInst>(Val: PredBB->getTerminator()); |
2283 | if (PredBr && PredBr->isUnconditional()) |
2284 | UnconditionalPreds.push_back(Elt: PredBB); |
2285 | else |
2286 | HaveNonUnconditionalPredecessors = true; |
2287 | } |
2288 | if (UnconditionalPreds.size() < 2) |
2289 | return false; |
2290 | |
2291 | // We take a two-step approach to tail sinking. First we scan from the end of |
2292 | // each block upwards in lockstep. If the n'th instruction from the end of each |
2293 | // block can be sunk, those instructions are added to ValuesToSink and we |
2294 | // carry on. If we can sink an instruction but need to PHI-merge some operands |
2295 | // (because they're not identical in each instruction) we add these to |
2296 | // PHIOperands. |
2297 | // We prepopulate PHIOperands with the phis that already exist in BB. |
2298 | DenseMap<const Use *, SmallVector<Value *, 4>> PHIOperands; |
2299 | for (PHINode &PN : BB->phis()) { |
2300 | SmallDenseMap<BasicBlock *, const Use *, 4> IncomingVals; |
2301 | for (const Use &U : PN.incoming_values()) |
2302 | IncomingVals.insert(KV: {PN.getIncomingBlock(U), &U}); |
2303 | auto &Ops = PHIOperands[IncomingVals[UnconditionalPreds[0]]]; |
2304 | for (BasicBlock *Pred : UnconditionalPreds) |
2305 | Ops.push_back(Elt: *IncomingVals[Pred]); |
2306 | } |
2307 | |
2308 | int ScanIdx = 0; |
2309 | SmallPtrSet<Value*,4> InstructionsToSink; |
2310 | LockstepReverseIterator LRI(UnconditionalPreds); |
2311 | while (LRI.isValid() && |
2312 | canSinkInstructions(Insts: *LRI, PHIOperands)) { |
2313 | LLVM_DEBUG(dbgs() << "SINK: instruction can be sunk: " << *(*LRI)[0] |
2314 | << "\n" ); |
2315 | InstructionsToSink.insert(I: (*LRI).begin(), E: (*LRI).end()); |
2316 | ++ScanIdx; |
2317 | --LRI; |
2318 | } |
2319 | |
2320 | // If no instructions can be sunk, early-return. |
2321 | if (ScanIdx == 0) |
2322 | return false; |
2323 | |
2324 | bool followedByDeoptOrUnreachable = IsBlockFollowedByDeoptOrUnreachable(BB); |
2325 | |
2326 | if (!followedByDeoptOrUnreachable) { |
2327 | // Okay, we *could* sink last ScanIdx instructions. But how many can we |
2328 | // actually sink before encountering instruction that is unprofitable to |
2329 | // sink? |
2330 | auto ProfitableToSinkInstruction = [&](LockstepReverseIterator &LRI) { |
2331 | unsigned NumPHIInsts = 0; |
2332 | for (Use &U : (*LRI)[0]->operands()) { |
2333 | auto It = PHIOperands.find(Val: &U); |
2334 | if (It != PHIOperands.end() && !all_of(Range&: It->second, P: [&](Value *V) { |
2335 | return InstructionsToSink.contains(Ptr: V); |
2336 | })) { |
2337 | ++NumPHIInsts; |
2338 | // FIXME: this check is overly optimistic. We may end up not sinking |
2339 | // said instruction, due to the very same profitability check. |
2340 | // See @creating_too_many_phis in sink-common-code.ll. |
2341 | } |
2342 | } |
2343 | LLVM_DEBUG(dbgs() << "SINK: #phi insts: " << NumPHIInsts << "\n" ); |
2344 | return NumPHIInsts <= 1; |
2345 | }; |
2346 | |
2347 | // We've determined that we are going to sink last ScanIdx instructions, |
2348 | // and recorded them in InstructionsToSink. Now, some instructions may be |
2349 | // unprofitable to sink. But that determination depends on the instructions |
2350 | // that we are going to sink. |
2351 | |
2352 | // First, forward scan: find the first instruction unprofitable to sink, |
2353 | // recording all the ones that are profitable to sink. |
2354 | // FIXME: would it be better, after we detect that not all are profitable. |
2355 | // to either record the profitable ones, or erase the unprofitable ones? |
2356 | // Maybe we need to choose (at runtime) the one that will touch least |
2357 | // instrs? |
2358 | LRI.reset(); |
2359 | int Idx = 0; |
2360 | SmallPtrSet<Value *, 4> InstructionsProfitableToSink; |
2361 | while (Idx < ScanIdx) { |
2362 | if (!ProfitableToSinkInstruction(LRI)) { |
2363 | // Too many PHIs would be created. |
2364 | LLVM_DEBUG( |
2365 | dbgs() << "SINK: stopping here, too many PHIs would be created!\n" ); |
2366 | break; |
2367 | } |
2368 | InstructionsProfitableToSink.insert(I: (*LRI).begin(), E: (*LRI).end()); |
2369 | --LRI; |
2370 | ++Idx; |
2371 | } |
2372 | |
2373 | // If no instructions can be sunk, early-return. |
2374 | if (Idx == 0) |
2375 | return false; |
2376 | |
2377 | // Did we determine that (only) some instructions are unprofitable to sink? |
2378 | if (Idx < ScanIdx) { |
2379 | // Okay, some instructions are unprofitable. |
2380 | ScanIdx = Idx; |
2381 | InstructionsToSink = InstructionsProfitableToSink; |
2382 | |
2383 | // But, that may make other instructions unprofitable, too. |
2384 | // So, do a backward scan, do any earlier instructions become |
2385 | // unprofitable? |
2386 | assert( |
2387 | !ProfitableToSinkInstruction(LRI) && |
2388 | "We already know that the last instruction is unprofitable to sink" ); |
2389 | ++LRI; |
2390 | --Idx; |
2391 | while (Idx >= 0) { |
2392 | // If we detect that an instruction becomes unprofitable to sink, |
2393 | // all earlier instructions won't be sunk either, |
2394 | // so preemptively keep InstructionsProfitableToSink in sync. |
2395 | // FIXME: is this the most performant approach? |
2396 | for (auto *I : *LRI) |
2397 | InstructionsProfitableToSink.erase(Ptr: I); |
2398 | if (!ProfitableToSinkInstruction(LRI)) { |
2399 | // Everything starting with this instruction won't be sunk. |
2400 | ScanIdx = Idx; |
2401 | InstructionsToSink = InstructionsProfitableToSink; |
2402 | } |
2403 | ++LRI; |
2404 | --Idx; |
2405 | } |
2406 | } |
2407 | |
2408 | // If no instructions can be sunk, early-return. |
2409 | if (ScanIdx == 0) |
2410 | return false; |
2411 | } |
2412 | |
2413 | bool Changed = false; |
2414 | |
2415 | if (HaveNonUnconditionalPredecessors) { |
2416 | if (!followedByDeoptOrUnreachable) { |
2417 | // It is always legal to sink common instructions from unconditional |
2418 | // predecessors. However, if not all predecessors are unconditional, |
2419 | // this transformation might be pessimizing. So as a rule of thumb, |
2420 | // don't do it unless we'd sink at least one non-speculatable instruction. |
2421 | // See https://bugs.llvm.org/show_bug.cgi?id=30244 |
2422 | LRI.reset(); |
2423 | int Idx = 0; |
2424 | bool Profitable = false; |
2425 | while (Idx < ScanIdx) { |
2426 | if (!isSafeToSpeculativelyExecute(I: (*LRI)[0])) { |
2427 | Profitable = true; |
2428 | break; |
2429 | } |
2430 | --LRI; |
2431 | ++Idx; |
2432 | } |
2433 | if (!Profitable) |
2434 | return false; |
2435 | } |
2436 | |
2437 | LLVM_DEBUG(dbgs() << "SINK: Splitting edge\n" ); |
2438 | // We have a conditional edge and we're going to sink some instructions. |
2439 | // Insert a new block postdominating all blocks we're going to sink from. |
2440 | if (!SplitBlockPredecessors(BB, Preds: UnconditionalPreds, Suffix: ".sink.split" , DTU)) |
2441 | // Edges couldn't be split. |
2442 | return false; |
2443 | Changed = true; |
2444 | } |
2445 | |
2446 | // Now that we've analyzed all potential sinking candidates, perform the |
2447 | // actual sink. We iteratively sink the last non-terminator of the source |
2448 | // blocks into their common successor unless doing so would require too |
2449 | // many PHI instructions to be generated (currently only one PHI is allowed |
2450 | // per sunk instruction). |
2451 | // |
2452 | // We can use InstructionsToSink to discount values needing PHI-merging that will |
2453 | // actually be sunk in a later iteration. This allows us to be more |
2454 | // aggressive in what we sink. This does allow a false positive where we |
2455 | // sink presuming a later value will also be sunk, but stop half way through |
2456 | // and never actually sink it which means we produce more PHIs than intended. |
2457 | // This is unlikely in practice though. |
2458 | int SinkIdx = 0; |
2459 | for (; SinkIdx != ScanIdx; ++SinkIdx) { |
2460 | LLVM_DEBUG(dbgs() << "SINK: Sink: " |
2461 | << *UnconditionalPreds[0]->getTerminator()->getPrevNode() |
2462 | << "\n" ); |
2463 | |
2464 | // Because we've sunk every instruction in turn, the current instruction to |
2465 | // sink is always at index 0. |
2466 | LRI.reset(); |
2467 | |
2468 | sinkLastInstruction(Blocks: UnconditionalPreds); |
2469 | NumSinkCommonInstrs++; |
2470 | Changed = true; |
2471 | } |
2472 | if (SinkIdx != 0) |
2473 | ++NumSinkCommonCode; |
2474 | return Changed; |
2475 | } |
2476 | |
2477 | namespace { |
2478 | |
2479 | struct CompatibleSets { |
2480 | using SetTy = SmallVector<InvokeInst *, 2>; |
2481 | |
2482 | SmallVector<SetTy, 1> Sets; |
2483 | |
2484 | static bool shouldBelongToSameSet(ArrayRef<InvokeInst *> Invokes); |
2485 | |
2486 | SetTy &getCompatibleSet(InvokeInst *II); |
2487 | |
2488 | void insert(InvokeInst *II); |
2489 | }; |
2490 | |
2491 | CompatibleSets::SetTy &CompatibleSets::getCompatibleSet(InvokeInst *II) { |
2492 | // Perform a linear scan over all the existing sets, see if the new `invoke` |
2493 | // is compatible with any particular set. Since we know that all the `invokes` |
2494 | // within a set are compatible, only check the first `invoke` in each set. |
2495 | // WARNING: at worst, this has quadratic complexity. |
2496 | for (CompatibleSets::SetTy &Set : Sets) { |
2497 | if (CompatibleSets::shouldBelongToSameSet(Invokes: {Set.front(), II})) |
2498 | return Set; |
2499 | } |
2500 | |
2501 | // Otherwise, we either had no sets yet, or this invoke forms a new set. |
2502 | return Sets.emplace_back(); |
2503 | } |
2504 | |
2505 | void CompatibleSets::insert(InvokeInst *II) { |
2506 | getCompatibleSet(II).emplace_back(Args&: II); |
2507 | } |
2508 | |
2509 | bool CompatibleSets::shouldBelongToSameSet(ArrayRef<InvokeInst *> Invokes) { |
2510 | assert(Invokes.size() == 2 && "Always called with exactly two candidates." ); |
2511 | |
2512 | // Can we theoretically merge these `invoke`s? |
2513 | auto IsIllegalToMerge = [](InvokeInst *II) { |
2514 | return II->cannotMerge() || II->isInlineAsm(); |
2515 | }; |
2516 | if (any_of(Range&: Invokes, P: IsIllegalToMerge)) |
2517 | return false; |
2518 | |
2519 | // Either both `invoke`s must be direct, |
2520 | // or both `invoke`s must be indirect. |
2521 | auto IsIndirectCall = [](InvokeInst *II) { return II->isIndirectCall(); }; |
2522 | bool HaveIndirectCalls = any_of(Range&: Invokes, P: IsIndirectCall); |
2523 | bool AllCallsAreIndirect = all_of(Range&: Invokes, P: IsIndirectCall); |
2524 | if (HaveIndirectCalls) { |
2525 | if (!AllCallsAreIndirect) |
2526 | return false; |
2527 | } else { |
2528 | // All callees must be identical. |
2529 | Value *Callee = nullptr; |
2530 | for (InvokeInst *II : Invokes) { |
2531 | Value *CurrCallee = II->getCalledOperand(); |
2532 | assert(CurrCallee && "There is always a called operand." ); |
2533 | if (!Callee) |
2534 | Callee = CurrCallee; |
2535 | else if (Callee != CurrCallee) |
2536 | return false; |
2537 | } |
2538 | } |
2539 | |
2540 | // Either both `invoke`s must not have a normal destination, |
2541 | // or both `invoke`s must have a normal destination, |
2542 | auto HasNormalDest = [](InvokeInst *II) { |
2543 | return !isa<UnreachableInst>(Val: II->getNormalDest()->getFirstNonPHIOrDbg()); |
2544 | }; |
2545 | if (any_of(Range&: Invokes, P: HasNormalDest)) { |
2546 | // Do not merge `invoke` that does not have a normal destination with one |
2547 | // that does have a normal destination, even though doing so would be legal. |
2548 | if (!all_of(Range&: Invokes, P: HasNormalDest)) |
2549 | return false; |
2550 | |
2551 | // All normal destinations must be identical. |
2552 | BasicBlock *NormalBB = nullptr; |
2553 | for (InvokeInst *II : Invokes) { |
2554 | BasicBlock *CurrNormalBB = II->getNormalDest(); |
2555 | assert(CurrNormalBB && "There is always a 'continue to' basic block." ); |
2556 | if (!NormalBB) |
2557 | NormalBB = CurrNormalBB; |
2558 | else if (NormalBB != CurrNormalBB) |
2559 | return false; |
2560 | } |
2561 | |
2562 | // In the normal destination, the incoming values for these two `invoke`s |
2563 | // must be compatible. |
2564 | SmallPtrSet<Value *, 16> EquivalenceSet(Invokes.begin(), Invokes.end()); |
2565 | if (!IncomingValuesAreCompatible( |
2566 | BB: NormalBB, IncomingBlocks: {Invokes[0]->getParent(), Invokes[1]->getParent()}, |
2567 | EquivalenceSet: &EquivalenceSet)) |
2568 | return false; |
2569 | } |
2570 | |
2571 | #ifndef NDEBUG |
2572 | // All unwind destinations must be identical. |
2573 | // We know that because we have started from said unwind destination. |
2574 | BasicBlock *UnwindBB = nullptr; |
2575 | for (InvokeInst *II : Invokes) { |
2576 | BasicBlock *CurrUnwindBB = II->getUnwindDest(); |
2577 | assert(CurrUnwindBB && "There is always an 'unwind to' basic block." ); |
2578 | if (!UnwindBB) |
2579 | UnwindBB = CurrUnwindBB; |
2580 | else |
2581 | assert(UnwindBB == CurrUnwindBB && "Unexpected unwind destination." ); |
2582 | } |
2583 | #endif |
2584 | |
2585 | // In the unwind destination, the incoming values for these two `invoke`s |
2586 | // must be compatible. |
2587 | if (!IncomingValuesAreCompatible( |
2588 | BB: Invokes.front()->getUnwindDest(), |
2589 | IncomingBlocks: {Invokes[0]->getParent(), Invokes[1]->getParent()})) |
2590 | return false; |
2591 | |
2592 | // Ignoring arguments, these `invoke`s must be identical, |
2593 | // including operand bundles. |
2594 | const InvokeInst *II0 = Invokes.front(); |
2595 | for (auto *II : Invokes.drop_front()) |
2596 | if (!II->isSameOperationAs(I: II0)) |
2597 | return false; |
2598 | |
2599 | // Can we theoretically form the data operands for the merged `invoke`? |
2600 | auto IsIllegalToMergeArguments = [](auto Ops) { |
2601 | Use &U0 = std::get<0>(Ops); |
2602 | Use &U1 = std::get<1>(Ops); |
2603 | if (U0 == U1) |
2604 | return false; |
2605 | return U0->getType()->isTokenTy() || |
2606 | !canReplaceOperandWithVariable(I: cast<Instruction>(Val: U0.getUser()), |
2607 | OpIdx: U0.getOperandNo()); |
2608 | }; |
2609 | assert(Invokes.size() == 2 && "Always called with exactly two candidates." ); |
2610 | if (any_of(Range: zip(t: Invokes[0]->data_ops(), u: Invokes[1]->data_ops()), |
2611 | P: IsIllegalToMergeArguments)) |
2612 | return false; |
2613 | |
2614 | return true; |
2615 | } |
2616 | |
2617 | } // namespace |
2618 | |
2619 | // Merge all invokes in the provided set, all of which are compatible |
2620 | // as per the `CompatibleSets::shouldBelongToSameSet()`. |
2621 | static void MergeCompatibleInvokesImpl(ArrayRef<InvokeInst *> Invokes, |
2622 | DomTreeUpdater *DTU) { |
2623 | assert(Invokes.size() >= 2 && "Must have at least two invokes to merge." ); |
2624 | |
2625 | SmallVector<DominatorTree::UpdateType, 8> Updates; |
2626 | if (DTU) |
2627 | Updates.reserve(N: 2 + 3 * Invokes.size()); |
2628 | |
2629 | bool HasNormalDest = |
2630 | !isa<UnreachableInst>(Val: Invokes[0]->getNormalDest()->getFirstNonPHIOrDbg()); |
2631 | |
2632 | // Clone one of the invokes into a new basic block. |
2633 | // Since they are all compatible, it doesn't matter which invoke is cloned. |
2634 | InvokeInst *MergedInvoke = [&Invokes, HasNormalDest]() { |
2635 | InvokeInst *II0 = Invokes.front(); |
2636 | BasicBlock *II0BB = II0->getParent(); |
2637 | BasicBlock *InsertBeforeBlock = |
2638 | II0->getParent()->getIterator()->getNextNode(); |
2639 | Function *Func = II0BB->getParent(); |
2640 | LLVMContext &Ctx = II0->getContext(); |
2641 | |
2642 | BasicBlock *MergedInvokeBB = BasicBlock::Create( |
2643 | Context&: Ctx, Name: II0BB->getName() + ".invoke" , Parent: Func, InsertBefore: InsertBeforeBlock); |
2644 | |
2645 | auto *MergedInvoke = cast<InvokeInst>(Val: II0->clone()); |
2646 | // NOTE: all invokes have the same attributes, so no handling needed. |
2647 | MergedInvoke->insertInto(ParentBB: MergedInvokeBB, It: MergedInvokeBB->end()); |
2648 | |
2649 | if (!HasNormalDest) { |
2650 | // This set does not have a normal destination, |
2651 | // so just form a new block with unreachable terminator. |
2652 | BasicBlock *MergedNormalDest = BasicBlock::Create( |
2653 | Context&: Ctx, Name: II0BB->getName() + ".cont" , Parent: Func, InsertBefore: InsertBeforeBlock); |
2654 | new UnreachableInst(Ctx, MergedNormalDest); |
2655 | MergedInvoke->setNormalDest(MergedNormalDest); |
2656 | } |
2657 | |
2658 | // The unwind destination, however, remainds identical for all invokes here. |
2659 | |
2660 | return MergedInvoke; |
2661 | }(); |
2662 | |
2663 | if (DTU) { |
2664 | // Predecessor blocks that contained these invokes will now branch to |
2665 | // the new block that contains the merged invoke, ... |
2666 | for (InvokeInst *II : Invokes) |
2667 | Updates.push_back( |
2668 | Elt: {DominatorTree::Insert, II->getParent(), MergedInvoke->getParent()}); |
2669 | |
2670 | // ... which has the new `unreachable` block as normal destination, |
2671 | // or unwinds to the (same for all `invoke`s in this set) `landingpad`, |
2672 | for (BasicBlock *SuccBBOfMergedInvoke : successors(I: MergedInvoke)) |
2673 | Updates.push_back(Elt: {DominatorTree::Insert, MergedInvoke->getParent(), |
2674 | SuccBBOfMergedInvoke}); |
2675 | |
2676 | // Since predecessor blocks now unconditionally branch to a new block, |
2677 | // they no longer branch to their original successors. |
2678 | for (InvokeInst *II : Invokes) |
2679 | for (BasicBlock *SuccOfPredBB : successors(BB: II->getParent())) |
2680 | Updates.push_back( |
2681 | Elt: {DominatorTree::Delete, II->getParent(), SuccOfPredBB}); |
2682 | } |
2683 | |
2684 | bool IsIndirectCall = Invokes[0]->isIndirectCall(); |
2685 | |
2686 | // Form the merged operands for the merged invoke. |
2687 | for (Use &U : MergedInvoke->operands()) { |
2688 | // Only PHI together the indirect callees and data operands. |
2689 | if (MergedInvoke->isCallee(U: &U)) { |
2690 | if (!IsIndirectCall) |
2691 | continue; |
2692 | } else if (!MergedInvoke->isDataOperand(U: &U)) |
2693 | continue; |
2694 | |
2695 | // Don't create trivial PHI's with all-identical incoming values. |
2696 | bool NeedPHI = any_of(Range&: Invokes, P: [&U](InvokeInst *II) { |
2697 | return II->getOperand(i_nocapture: U.getOperandNo()) != U.get(); |
2698 | }); |
2699 | if (!NeedPHI) |
2700 | continue; |
2701 | |
2702 | // Form a PHI out of all the data ops under this index. |
2703 | PHINode *PN = PHINode::Create( |
2704 | Ty: U->getType(), /*NumReservedValues=*/Invokes.size(), NameStr: "" , InsertBefore: MergedInvoke->getIterator()); |
2705 | for (InvokeInst *II : Invokes) |
2706 | PN->addIncoming(V: II->getOperand(i_nocapture: U.getOperandNo()), BB: II->getParent()); |
2707 | |
2708 | U.set(PN); |
2709 | } |
2710 | |
2711 | // We've ensured that each PHI node has compatible (identical) incoming values |
2712 | // when coming from each of the `invoke`s in the current merge set, |
2713 | // so update the PHI nodes accordingly. |
2714 | for (BasicBlock *Succ : successors(I: MergedInvoke)) |
2715 | AddPredecessorToBlock(Succ, /*NewPred=*/MergedInvoke->getParent(), |
2716 | /*ExistPred=*/Invokes.front()->getParent()); |
2717 | |
2718 | // And finally, replace the original `invoke`s with an unconditional branch |
2719 | // to the block with the merged `invoke`. Also, give that merged `invoke` |
2720 | // the merged debugloc of all the original `invoke`s. |
2721 | DILocation *MergedDebugLoc = nullptr; |
2722 | for (InvokeInst *II : Invokes) { |
2723 | // Compute the debug location common to all the original `invoke`s. |
2724 | if (!MergedDebugLoc) |
2725 | MergedDebugLoc = II->getDebugLoc(); |
2726 | else |
2727 | MergedDebugLoc = |
2728 | DILocation::getMergedLocation(LocA: MergedDebugLoc, LocB: II->getDebugLoc()); |
2729 | |
2730 | // And replace the old `invoke` with an unconditionally branch |
2731 | // to the block with the merged `invoke`. |
2732 | for (BasicBlock *OrigSuccBB : successors(BB: II->getParent())) |
2733 | OrigSuccBB->removePredecessor(Pred: II->getParent()); |
2734 | BranchInst::Create(IfTrue: MergedInvoke->getParent(), InsertBefore: II->getParent()); |
2735 | II->replaceAllUsesWith(V: MergedInvoke); |
2736 | II->eraseFromParent(); |
2737 | ++NumInvokesMerged; |
2738 | } |
2739 | MergedInvoke->setDebugLoc(MergedDebugLoc); |
2740 | ++NumInvokeSetsFormed; |
2741 | |
2742 | if (DTU) |
2743 | DTU->applyUpdates(Updates); |
2744 | } |
2745 | |
2746 | /// If this block is a `landingpad` exception handling block, categorize all |
2747 | /// the predecessor `invoke`s into sets, with all `invoke`s in each set |
2748 | /// being "mergeable" together, and then merge invokes in each set together. |
2749 | /// |
2750 | /// This is a weird mix of hoisting and sinking. Visually, it goes from: |
2751 | /// [...] [...] |
2752 | /// | | |
2753 | /// [invoke0] [invoke1] |
2754 | /// / \ / \ |
2755 | /// [cont0] [landingpad] [cont1] |
2756 | /// to: |
2757 | /// [...] [...] |
2758 | /// \ / |
2759 | /// [invoke] |
2760 | /// / \ |
2761 | /// [cont] [landingpad] |
2762 | /// |
2763 | /// But of course we can only do that if the invokes share the `landingpad`, |
2764 | /// edges invoke0->cont0 and invoke1->cont1 are "compatible", |
2765 | /// and the invoked functions are "compatible". |
2766 | static bool MergeCompatibleInvokes(BasicBlock *BB, DomTreeUpdater *DTU) { |
2767 | if (!EnableMergeCompatibleInvokes) |
2768 | return false; |
2769 | |
2770 | bool Changed = false; |
2771 | |
2772 | // FIXME: generalize to all exception handling blocks? |
2773 | if (!BB->isLandingPad()) |
2774 | return Changed; |
2775 | |
2776 | CompatibleSets Grouper; |
2777 | |
2778 | // Record all the predecessors of this `landingpad`. As per verifier, |
2779 | // the only allowed predecessor is the unwind edge of an `invoke`. |
2780 | // We want to group "compatible" `invokes` into the same set to be merged. |
2781 | for (BasicBlock *PredBB : predecessors(BB)) |
2782 | Grouper.insert(II: cast<InvokeInst>(Val: PredBB->getTerminator())); |
2783 | |
2784 | // And now, merge `invoke`s that were grouped togeter. |
2785 | for (ArrayRef<InvokeInst *> Invokes : Grouper.Sets) { |
2786 | if (Invokes.size() < 2) |
2787 | continue; |
2788 | Changed = true; |
2789 | MergeCompatibleInvokesImpl(Invokes, DTU); |
2790 | } |
2791 | |
2792 | return Changed; |
2793 | } |
2794 | |
2795 | namespace { |
2796 | /// Track ephemeral values, which should be ignored for cost-modelling |
2797 | /// purposes. Requires walking instructions in reverse order. |
2798 | class EphemeralValueTracker { |
2799 | SmallPtrSet<const Instruction *, 32> EphValues; |
2800 | |
2801 | bool isEphemeral(const Instruction *I) { |
2802 | if (isa<AssumeInst>(Val: I)) |
2803 | return true; |
2804 | return !I->mayHaveSideEffects() && !I->isTerminator() && |
2805 | all_of(Range: I->users(), P: [&](const User *U) { |
2806 | return EphValues.count(Ptr: cast<Instruction>(Val: U)); |
2807 | }); |
2808 | } |
2809 | |
2810 | public: |
2811 | bool track(const Instruction *I) { |
2812 | if (isEphemeral(I)) { |
2813 | EphValues.insert(Ptr: I); |
2814 | return true; |
2815 | } |
2816 | return false; |
2817 | } |
2818 | |
2819 | bool contains(const Instruction *I) const { return EphValues.contains(Ptr: I); } |
2820 | }; |
2821 | } // namespace |
2822 | |
2823 | /// Determine if we can hoist sink a sole store instruction out of a |
2824 | /// conditional block. |
2825 | /// |
2826 | /// We are looking for code like the following: |
2827 | /// BrBB: |
2828 | /// store i32 %add, i32* %arrayidx2 |
2829 | /// ... // No other stores or function calls (we could be calling a memory |
2830 | /// ... // function). |
2831 | /// %cmp = icmp ult %x, %y |
2832 | /// br i1 %cmp, label %EndBB, label %ThenBB |
2833 | /// ThenBB: |
2834 | /// store i32 %add5, i32* %arrayidx2 |
2835 | /// br label EndBB |
2836 | /// EndBB: |
2837 | /// ... |
2838 | /// We are going to transform this into: |
2839 | /// BrBB: |
2840 | /// store i32 %add, i32* %arrayidx2 |
2841 | /// ... // |
2842 | /// %cmp = icmp ult %x, %y |
2843 | /// %add.add5 = select i1 %cmp, i32 %add, %add5 |
2844 | /// store i32 %add.add5, i32* %arrayidx2 |
2845 | /// ... |
2846 | /// |
2847 | /// \return The pointer to the value of the previous store if the store can be |
2848 | /// hoisted into the predecessor block. 0 otherwise. |
2849 | static Value *isSafeToSpeculateStore(Instruction *I, BasicBlock *BrBB, |
2850 | BasicBlock *StoreBB, BasicBlock *EndBB) { |
2851 | StoreInst *StoreToHoist = dyn_cast<StoreInst>(Val: I); |
2852 | if (!StoreToHoist) |
2853 | return nullptr; |
2854 | |
2855 | // Volatile or atomic. |
2856 | if (!StoreToHoist->isSimple()) |
2857 | return nullptr; |
2858 | |
2859 | Value *StorePtr = StoreToHoist->getPointerOperand(); |
2860 | Type *StoreTy = StoreToHoist->getValueOperand()->getType(); |
2861 | |
2862 | // Look for a store to the same pointer in BrBB. |
2863 | unsigned MaxNumInstToLookAt = 9; |
2864 | // Skip pseudo probe intrinsic calls which are not really killing any memory |
2865 | // accesses. |
2866 | for (Instruction &CurI : reverse(C: BrBB->instructionsWithoutDebug(SkipPseudoOp: true))) { |
2867 | if (!MaxNumInstToLookAt) |
2868 | break; |
2869 | --MaxNumInstToLookAt; |
2870 | |
2871 | // Could be calling an instruction that affects memory like free(). |
2872 | if (CurI.mayWriteToMemory() && !isa<StoreInst>(Val: CurI)) |
2873 | return nullptr; |
2874 | |
2875 | if (auto *SI = dyn_cast<StoreInst>(Val: &CurI)) { |
2876 | // Found the previous store to same location and type. Make sure it is |
2877 | // simple, to avoid introducing a spurious non-atomic write after an |
2878 | // atomic write. |
2879 | if (SI->getPointerOperand() == StorePtr && |
2880 | SI->getValueOperand()->getType() == StoreTy && SI->isSimple() && |
2881 | SI->getAlign() >= StoreToHoist->getAlign()) |
2882 | // Found the previous store, return its value operand. |
2883 | return SI->getValueOperand(); |
2884 | return nullptr; // Unknown store. |
2885 | } |
2886 | |
2887 | if (auto *LI = dyn_cast<LoadInst>(Val: &CurI)) { |
2888 | if (LI->getPointerOperand() == StorePtr && LI->getType() == StoreTy && |
2889 | LI->isSimple() && LI->getAlign() >= StoreToHoist->getAlign()) { |
2890 | // Local objects (created by an `alloca` instruction) are always |
2891 | // writable, so once we are past a read from a location it is valid to |
2892 | // also write to that same location. |
2893 | // If the address of the local object never escapes the function, that |
2894 | // means it's never concurrently read or written, hence moving the store |
2895 | // from under the condition will not introduce a data race. |
2896 | auto *AI = dyn_cast<AllocaInst>(Val: getUnderlyingObject(V: StorePtr)); |
2897 | if (AI && !PointerMayBeCaptured(V: AI, ReturnCaptures: false, StoreCaptures: true)) |
2898 | // Found a previous load, return it. |
2899 | return LI; |
2900 | } |
2901 | // The load didn't work out, but we may still find a store. |
2902 | } |
2903 | } |
2904 | |
2905 | return nullptr; |
2906 | } |
2907 | |
2908 | /// Estimate the cost of the insertion(s) and check that the PHI nodes can be |
2909 | /// converted to selects. |
2910 | static bool validateAndCostRequiredSelects(BasicBlock *BB, BasicBlock *ThenBB, |
2911 | BasicBlock *EndBB, |
2912 | unsigned &SpeculatedInstructions, |
2913 | InstructionCost &Cost, |
2914 | const TargetTransformInfo &TTI) { |
2915 | TargetTransformInfo::TargetCostKind CostKind = |
2916 | BB->getParent()->hasMinSize() |
2917 | ? TargetTransformInfo::TCK_CodeSize |
2918 | : TargetTransformInfo::TCK_SizeAndLatency; |
2919 | |
2920 | bool HaveRewritablePHIs = false; |
2921 | for (PHINode &PN : EndBB->phis()) { |
2922 | Value *OrigV = PN.getIncomingValueForBlock(BB); |
2923 | Value *ThenV = PN.getIncomingValueForBlock(BB: ThenBB); |
2924 | |
2925 | // FIXME: Try to remove some of the duplication with |
2926 | // hoistCommonCodeFromSuccessors. Skip PHIs which are trivial. |
2927 | if (ThenV == OrigV) |
2928 | continue; |
2929 | |
2930 | Cost += TTI.getCmpSelInstrCost(Opcode: Instruction::Select, ValTy: PN.getType(), CondTy: nullptr, |
2931 | VecPred: CmpInst::BAD_ICMP_PREDICATE, CostKind); |
2932 | |
2933 | // Don't convert to selects if we could remove undefined behavior instead. |
2934 | if (passingValueIsAlwaysUndefined(V: OrigV, I: &PN) || |
2935 | passingValueIsAlwaysUndefined(V: ThenV, I: &PN)) |
2936 | return false; |
2937 | |
2938 | HaveRewritablePHIs = true; |
2939 | ConstantExpr *OrigCE = dyn_cast<ConstantExpr>(Val: OrigV); |
2940 | ConstantExpr *ThenCE = dyn_cast<ConstantExpr>(Val: ThenV); |
2941 | if (!OrigCE && !ThenCE) |
2942 | continue; // Known cheap (FIXME: Maybe not true for aggregates). |
2943 | |
2944 | InstructionCost OrigCost = OrigCE ? computeSpeculationCost(I: OrigCE, TTI) : 0; |
2945 | InstructionCost ThenCost = ThenCE ? computeSpeculationCost(I: ThenCE, TTI) : 0; |
2946 | InstructionCost MaxCost = |
2947 | 2 * PHINodeFoldingThreshold * TargetTransformInfo::TCC_Basic; |
2948 | if (OrigCost + ThenCost > MaxCost) |
2949 | return false; |
2950 | |
2951 | // Account for the cost of an unfolded ConstantExpr which could end up |
2952 | // getting expanded into Instructions. |
2953 | // FIXME: This doesn't account for how many operations are combined in the |
2954 | // constant expression. |
2955 | ++SpeculatedInstructions; |
2956 | if (SpeculatedInstructions > 1) |
2957 | return false; |
2958 | } |
2959 | |
2960 | return HaveRewritablePHIs; |
2961 | } |
2962 | |
2963 | /// Speculate a conditional basic block flattening the CFG. |
2964 | /// |
2965 | /// Note that this is a very risky transform currently. Speculating |
2966 | /// instructions like this is most often not desirable. Instead, there is an MI |
2967 | /// pass which can do it with full awareness of the resource constraints. |
2968 | /// However, some cases are "obvious" and we should do directly. An example of |
2969 | /// this is speculating a single, reasonably cheap instruction. |
2970 | /// |
2971 | /// There is only one distinct advantage to flattening the CFG at the IR level: |
2972 | /// it makes very common but simplistic optimizations such as are common in |
2973 | /// instcombine and the DAG combiner more powerful by removing CFG edges and |
2974 | /// modeling their effects with easier to reason about SSA value graphs. |
2975 | /// |
2976 | /// |
2977 | /// An illustration of this transform is turning this IR: |
2978 | /// \code |
2979 | /// BB: |
2980 | /// %cmp = icmp ult %x, %y |
2981 | /// br i1 %cmp, label %EndBB, label %ThenBB |
2982 | /// ThenBB: |
2983 | /// %sub = sub %x, %y |
2984 | /// br label BB2 |
2985 | /// EndBB: |
2986 | /// %phi = phi [ %sub, %ThenBB ], [ 0, %EndBB ] |
2987 | /// ... |
2988 | /// \endcode |
2989 | /// |
2990 | /// Into this IR: |
2991 | /// \code |
2992 | /// BB: |
2993 | /// %cmp = icmp ult %x, %y |
2994 | /// %sub = sub %x, %y |
2995 | /// %cond = select i1 %cmp, 0, %sub |
2996 | /// ... |
2997 | /// \endcode |
2998 | /// |
2999 | /// \returns true if the conditional block is removed. |
3000 | bool SimplifyCFGOpt::SpeculativelyExecuteBB(BranchInst *BI, |
3001 | BasicBlock *ThenBB) { |
3002 | if (!Options.SpeculateBlocks) |
3003 | return false; |
3004 | |
3005 | // Be conservative for now. FP select instruction can often be expensive. |
3006 | Value *BrCond = BI->getCondition(); |
3007 | if (isa<FCmpInst>(Val: BrCond)) |
3008 | return false; |
3009 | |
3010 | BasicBlock *BB = BI->getParent(); |
3011 | BasicBlock *EndBB = ThenBB->getTerminator()->getSuccessor(Idx: 0); |
3012 | InstructionCost Budget = |
3013 | PHINodeFoldingThreshold * TargetTransformInfo::TCC_Basic; |
3014 | |
3015 | // If ThenBB is actually on the false edge of the conditional branch, remember |
3016 | // to swap the select operands later. |
3017 | bool Invert = false; |
3018 | if (ThenBB != BI->getSuccessor(i: 0)) { |
3019 | assert(ThenBB == BI->getSuccessor(1) && "No edge from 'if' block?" ); |
3020 | Invert = true; |
3021 | } |
3022 | assert(EndBB == BI->getSuccessor(!Invert) && "No edge from to end block" ); |
3023 | |
3024 | // If the branch is non-unpredictable, and is predicted to *not* branch to |
3025 | // the `then` block, then avoid speculating it. |
3026 | if (!BI->getMetadata(KindID: LLVMContext::MD_unpredictable)) { |
3027 | uint64_t TWeight, FWeight; |
3028 | if (extractBranchWeights(I: *BI, TrueVal&: TWeight, FalseVal&: FWeight) && |
3029 | (TWeight + FWeight) != 0) { |
3030 | uint64_t EndWeight = Invert ? TWeight : FWeight; |
3031 | BranchProbability BIEndProb = |
3032 | BranchProbability::getBranchProbability(Numerator: EndWeight, Denominator: TWeight + FWeight); |
3033 | BranchProbability Likely = TTI.getPredictableBranchThreshold(); |
3034 | if (BIEndProb >= Likely) |
3035 | return false; |
3036 | } |
3037 | } |
3038 | |
3039 | // Keep a count of how many times instructions are used within ThenBB when |
3040 | // they are candidates for sinking into ThenBB. Specifically: |
3041 | // - They are defined in BB, and |
3042 | // - They have no side effects, and |
3043 | // - All of their uses are in ThenBB. |
3044 | SmallDenseMap<Instruction *, unsigned, 4> SinkCandidateUseCounts; |
3045 | |
3046 | SmallVector<Instruction *, 4> SpeculatedDbgIntrinsics; |
3047 | |
3048 | unsigned SpeculatedInstructions = 0; |
3049 | Value *SpeculatedStoreValue = nullptr; |
3050 | StoreInst *SpeculatedStore = nullptr; |
3051 | EphemeralValueTracker EphTracker; |
3052 | for (Instruction &I : reverse(C: drop_end(RangeOrContainer&: *ThenBB))) { |
3053 | // Skip debug info. |
3054 | if (isa<DbgInfoIntrinsic>(Val: I)) { |
3055 | SpeculatedDbgIntrinsics.push_back(Elt: &I); |
3056 | continue; |
3057 | } |
3058 | |
3059 | // Skip pseudo probes. The consequence is we lose track of the branch |
3060 | // probability for ThenBB, which is fine since the optimization here takes |
3061 | // place regardless of the branch probability. |
3062 | if (isa<PseudoProbeInst>(Val: I)) { |
3063 | // The probe should be deleted so that it will not be over-counted when |
3064 | // the samples collected on the non-conditional path are counted towards |
3065 | // the conditional path. We leave it for the counts inference algorithm to |
3066 | // figure out a proper count for an unknown probe. |
3067 | SpeculatedDbgIntrinsics.push_back(Elt: &I); |
3068 | continue; |
3069 | } |
3070 | |
3071 | // Ignore ephemeral values, they will be dropped by the transform. |
3072 | if (EphTracker.track(I: &I)) |
3073 | continue; |
3074 | |
3075 | // Only speculatively execute a single instruction (not counting the |
3076 | // terminator) for now. |
3077 | ++SpeculatedInstructions; |
3078 | if (SpeculatedInstructions > 1) |
3079 | return false; |
3080 | |
3081 | // Don't hoist the instruction if it's unsafe or expensive. |
3082 | if (!isSafeToSpeculativelyExecute(I: &I) && |
3083 | !(HoistCondStores && (SpeculatedStoreValue = isSafeToSpeculateStore( |
3084 | I: &I, BrBB: BB, StoreBB: ThenBB, EndBB)))) |
3085 | return false; |
3086 | if (!SpeculatedStoreValue && |
3087 | computeSpeculationCost(I: &I, TTI) > |
3088 | PHINodeFoldingThreshold * TargetTransformInfo::TCC_Basic) |
3089 | return false; |
3090 | |
3091 | // Store the store speculation candidate. |
3092 | if (SpeculatedStoreValue) |
3093 | SpeculatedStore = cast<StoreInst>(Val: &I); |
3094 | |
3095 | // Do not hoist the instruction if any of its operands are defined but not |
3096 | // used in BB. The transformation will prevent the operand from |
3097 | // being sunk into the use block. |
3098 | for (Use &Op : I.operands()) { |
3099 | Instruction *OpI = dyn_cast<Instruction>(Val&: Op); |
3100 | if (!OpI || OpI->getParent() != BB || OpI->mayHaveSideEffects()) |
3101 | continue; // Not a candidate for sinking. |
3102 | |
3103 | ++SinkCandidateUseCounts[OpI]; |
3104 | } |
3105 | } |
3106 | |
3107 | // Consider any sink candidates which are only used in ThenBB as costs for |
3108 | // speculation. Note, while we iterate over a DenseMap here, we are summing |
3109 | // and so iteration order isn't significant. |
3110 | for (const auto &[Inst, Count] : SinkCandidateUseCounts) |
3111 | if (Inst->hasNUses(N: Count)) { |
3112 | ++SpeculatedInstructions; |
3113 | if (SpeculatedInstructions > 1) |
3114 | return false; |
3115 | } |
3116 | |
3117 | // Check that we can insert the selects and that it's not too expensive to do |
3118 | // so. |
3119 | bool Convert = SpeculatedStore != nullptr; |
3120 | InstructionCost Cost = 0; |
3121 | Convert |= validateAndCostRequiredSelects(BB, ThenBB, EndBB, |
3122 | SpeculatedInstructions, |
3123 | Cost, TTI); |
3124 | if (!Convert || Cost > Budget) |
3125 | return false; |
3126 | |
3127 | // If we get here, we can hoist the instruction and if-convert. |
3128 | LLVM_DEBUG(dbgs() << "SPECULATIVELY EXECUTING BB" << *ThenBB << "\n" ;); |
3129 | |
3130 | // Insert a select of the value of the speculated store. |
3131 | if (SpeculatedStoreValue) { |
3132 | IRBuilder<NoFolder> Builder(BI); |
3133 | Value *OrigV = SpeculatedStore->getValueOperand(); |
3134 | Value *TrueV = SpeculatedStore->getValueOperand(); |
3135 | Value *FalseV = SpeculatedStoreValue; |
3136 | if (Invert) |
3137 | std::swap(a&: TrueV, b&: FalseV); |
3138 | Value *S = Builder.CreateSelect( |
3139 | C: BrCond, True: TrueV, False: FalseV, Name: "spec.store.select" , MDFrom: BI); |
3140 | SpeculatedStore->setOperand(i_nocapture: 0, Val_nocapture: S); |
3141 | SpeculatedStore->applyMergedLocation(LocA: BI->getDebugLoc(), |
3142 | LocB: SpeculatedStore->getDebugLoc()); |
3143 | // The value stored is still conditional, but the store itself is now |
3144 | // unconditonally executed, so we must be sure that any linked dbg.assign |
3145 | // intrinsics are tracking the new stored value (the result of the |
3146 | // select). If we don't, and the store were to be removed by another pass |
3147 | // (e.g. DSE), then we'd eventually end up emitting a location describing |
3148 | // the conditional value, unconditionally. |
3149 | // |
3150 | // === Before this transformation === |
3151 | // pred: |
3152 | // store %one, %x.dest, !DIAssignID !1 |
3153 | // dbg.assign %one, "x", ..., !1, ... |
3154 | // br %cond if.then |
3155 | // |
3156 | // if.then: |
3157 | // store %two, %x.dest, !DIAssignID !2 |
3158 | // dbg.assign %two, "x", ..., !2, ... |
3159 | // |
3160 | // === After this transformation === |
3161 | // pred: |
3162 | // store %one, %x.dest, !DIAssignID !1 |
3163 | // dbg.assign %one, "x", ..., !1 |
3164 | /// ... |
3165 | // %merge = select %cond, %two, %one |
3166 | // store %merge, %x.dest, !DIAssignID !2 |
3167 | // dbg.assign %merge, "x", ..., !2 |
3168 | auto replaceVariable = [OrigV, S](auto *DbgAssign) { |
3169 | if (llvm::is_contained(DbgAssign->location_ops(), OrigV)) |
3170 | DbgAssign->replaceVariableLocationOp(OrigV, S); |
3171 | }; |
3172 | for_each(Range: at::getAssignmentMarkers(Inst: SpeculatedStore), F: replaceVariable); |
3173 | for_each(Range: at::getDVRAssignmentMarkers(Inst: SpeculatedStore), F: replaceVariable); |
3174 | } |
3175 | |
3176 | // Metadata can be dependent on the condition we are hoisting above. |
3177 | // Strip all UB-implying metadata on the instruction. Drop the debug loc |
3178 | // to avoid making it appear as if the condition is a constant, which would |
3179 | // be misleading while debugging. |
3180 | // Similarly strip attributes that maybe dependent on condition we are |
3181 | // hoisting above. |
3182 | for (auto &I : make_early_inc_range(Range&: *ThenBB)) { |
3183 | if (!SpeculatedStoreValue || &I != SpeculatedStore) { |
3184 | // Don't update the DILocation of dbg.assign intrinsics. |
3185 | if (!isa<DbgAssignIntrinsic>(Val: &I)) |
3186 | I.setDebugLoc(DebugLoc()); |
3187 | } |
3188 | I.dropUBImplyingAttrsAndMetadata(); |
3189 | |
3190 | // Drop ephemeral values. |
3191 | if (EphTracker.contains(I: &I)) { |
3192 | I.replaceAllUsesWith(V: PoisonValue::get(T: I.getType())); |
3193 | I.eraseFromParent(); |
3194 | } |
3195 | } |
3196 | |
3197 | // Hoist the instructions. |
3198 | // In "RemoveDIs" non-instr debug-info mode, drop DbgVariableRecords attached |
3199 | // to these instructions, in the same way that dbg.value intrinsics are |
3200 | // dropped at the end of this block. |
3201 | for (auto &It : make_range(x: ThenBB->begin(), y: ThenBB->end())) |
3202 | for (DbgRecord &DR : make_early_inc_range(Range: It.getDbgRecordRange())) |
3203 | // Drop all records except assign-kind DbgVariableRecords (dbg.assign |
3204 | // equivalent). |
3205 | if (DbgVariableRecord *DVR = dyn_cast<DbgVariableRecord>(Val: &DR); |
3206 | !DVR || !DVR->isDbgAssign()) |
3207 | It.dropOneDbgRecord(I: &DR); |
3208 | BB->splice(ToIt: BI->getIterator(), FromBB: ThenBB, FromBeginIt: ThenBB->begin(), |
3209 | FromEndIt: std::prev(x: ThenBB->end())); |
3210 | |
3211 | // Insert selects and rewrite the PHI operands. |
3212 | IRBuilder<NoFolder> Builder(BI); |
3213 | for (PHINode &PN : EndBB->phis()) { |
3214 | unsigned OrigI = PN.getBasicBlockIndex(BB); |
3215 | unsigned ThenI = PN.getBasicBlockIndex(BB: ThenBB); |
3216 | Value *OrigV = PN.getIncomingValue(i: OrigI); |
3217 | Value *ThenV = PN.getIncomingValue(i: ThenI); |
3218 | |
3219 | // Skip PHIs which are trivial. |
3220 | if (OrigV == ThenV) |
3221 | continue; |
3222 | |
3223 | // Create a select whose true value is the speculatively executed value and |
3224 | // false value is the pre-existing value. Swap them if the branch |
3225 | // destinations were inverted. |
3226 | Value *TrueV = ThenV, *FalseV = OrigV; |
3227 | if (Invert) |
3228 | std::swap(a&: TrueV, b&: FalseV); |
3229 | Value *V = Builder.CreateSelect(C: BrCond, True: TrueV, False: FalseV, Name: "spec.select" , MDFrom: BI); |
3230 | PN.setIncomingValue(i: OrigI, V); |
3231 | PN.setIncomingValue(i: ThenI, V); |
3232 | } |
3233 | |
3234 | // Remove speculated dbg intrinsics. |
3235 | // FIXME: Is it possible to do this in a more elegant way? Moving/merging the |
3236 | // dbg value for the different flows and inserting it after the select. |
3237 | for (Instruction *I : SpeculatedDbgIntrinsics) { |
3238 | // We still want to know that an assignment took place so don't remove |
3239 | // dbg.assign intrinsics. |
3240 | if (!isa<DbgAssignIntrinsic>(Val: I)) |
3241 | I->eraseFromParent(); |
3242 | } |
3243 | |
3244 | ++NumSpeculations; |
3245 | return true; |
3246 | } |
3247 | |
3248 | /// Return true if we can thread a branch across this block. |
3249 | static bool BlockIsSimpleEnoughToThreadThrough(BasicBlock *BB) { |
3250 | int Size = 0; |
3251 | EphemeralValueTracker EphTracker; |
3252 | |
3253 | // Walk the loop in reverse so that we can identify ephemeral values properly |
3254 | // (values only feeding assumes). |
3255 | for (Instruction &I : reverse(C: BB->instructionsWithoutDebug(SkipPseudoOp: false))) { |
3256 | // Can't fold blocks that contain noduplicate or convergent calls. |
3257 | if (CallInst *CI = dyn_cast<CallInst>(Val: &I)) |
3258 | if (CI->cannotDuplicate() || CI->isConvergent()) |
3259 | return false; |
3260 | |
3261 | // Ignore ephemeral values which are deleted during codegen. |
3262 | // We will delete Phis while threading, so Phis should not be accounted in |
3263 | // block's size. |
3264 | if (!EphTracker.track(I: &I) && !isa<PHINode>(Val: I)) { |
3265 | if (Size++ > MaxSmallBlockSize) |
3266 | return false; // Don't clone large BB's. |
3267 | } |
3268 | |
3269 | // We can only support instructions that do not define values that are |
3270 | // live outside of the current basic block. |
3271 | for (User *U : I.users()) { |
3272 | Instruction *UI = cast<Instruction>(Val: U); |
3273 | if (UI->getParent() != BB || isa<PHINode>(Val: UI)) |
3274 | return false; |
3275 | } |
3276 | |
3277 | // Looks ok, continue checking. |
3278 | } |
3279 | |
3280 | return true; |
3281 | } |
3282 | |
3283 | static ConstantInt *getKnownValueOnEdge(Value *V, BasicBlock *From, |
3284 | BasicBlock *To) { |
3285 | // Don't look past the block defining the value, we might get the value from |
3286 | // a previous loop iteration. |
3287 | auto *I = dyn_cast<Instruction>(Val: V); |
3288 | if (I && I->getParent() == To) |
3289 | return nullptr; |
3290 | |
3291 | // We know the value if the From block branches on it. |
3292 | auto *BI = dyn_cast<BranchInst>(Val: From->getTerminator()); |
3293 | if (BI && BI->isConditional() && BI->getCondition() == V && |
3294 | BI->getSuccessor(i: 0) != BI->getSuccessor(i: 1)) |
3295 | return BI->getSuccessor(i: 0) == To ? ConstantInt::getTrue(Context&: BI->getContext()) |
3296 | : ConstantInt::getFalse(Context&: BI->getContext()); |
3297 | |
3298 | return nullptr; |
3299 | } |
3300 | |
3301 | /// If we have a conditional branch on something for which we know the constant |
3302 | /// value in predecessors (e.g. a phi node in the current block), thread edges |
3303 | /// from the predecessor to their ultimate destination. |
3304 | static std::optional<bool> |
3305 | FoldCondBranchOnValueKnownInPredecessorImpl(BranchInst *BI, DomTreeUpdater *DTU, |
3306 | const DataLayout &DL, |
3307 | AssumptionCache *AC) { |
3308 | SmallMapVector<ConstantInt *, SmallSetVector<BasicBlock *, 2>, 2> KnownValues; |
3309 | BasicBlock *BB = BI->getParent(); |
3310 | Value *Cond = BI->getCondition(); |
3311 | PHINode *PN = dyn_cast<PHINode>(Val: Cond); |
3312 | if (PN && PN->getParent() == BB) { |
3313 | // Degenerate case of a single entry PHI. |
3314 | if (PN->getNumIncomingValues() == 1) { |
3315 | FoldSingleEntryPHINodes(BB: PN->getParent()); |
3316 | return true; |
3317 | } |
3318 | |
3319 | for (Use &U : PN->incoming_values()) |
3320 | if (auto *CB = dyn_cast<ConstantInt>(Val&: U)) |
3321 | KnownValues[CB].insert(X: PN->getIncomingBlock(U)); |
3322 | } else { |
3323 | for (BasicBlock *Pred : predecessors(BB)) { |
3324 | if (ConstantInt *CB = getKnownValueOnEdge(V: Cond, From: Pred, To: BB)) |
3325 | KnownValues[CB].insert(X: Pred); |
3326 | } |
3327 | } |
3328 | |
3329 | if (KnownValues.empty()) |
3330 | return false; |
3331 | |
3332 | // Now we know that this block has multiple preds and two succs. |
3333 | // Check that the block is small enough and values defined in the block are |
3334 | // not used outside of it. |
3335 | if (!BlockIsSimpleEnoughToThreadThrough(BB)) |
3336 | return false; |
3337 | |
3338 | for (const auto &Pair : KnownValues) { |
3339 | // Okay, we now know that all edges from PredBB should be revectored to |
3340 | // branch to RealDest. |
3341 | ConstantInt *CB = Pair.first; |
3342 | ArrayRef<BasicBlock *> PredBBs = Pair.second.getArrayRef(); |
3343 | BasicBlock *RealDest = BI->getSuccessor(i: !CB->getZExtValue()); |
3344 | |
3345 | if (RealDest == BB) |
3346 | continue; // Skip self loops. |
3347 | |
3348 | // Skip if the predecessor's terminator is an indirect branch. |
3349 | if (any_of(Range&: PredBBs, P: [](BasicBlock *PredBB) { |
3350 | return isa<IndirectBrInst>(Val: PredBB->getTerminator()); |
3351 | })) |
3352 | continue; |
3353 | |
3354 | LLVM_DEBUG({ |
3355 | dbgs() << "Condition " << *Cond << " in " << BB->getName() |
3356 | << " has value " << *Pair.first << " in predecessors:\n" ; |
3357 | for (const BasicBlock *PredBB : Pair.second) |
3358 | dbgs() << " " << PredBB->getName() << "\n" ; |
3359 | dbgs() << "Threading to destination " << RealDest->getName() << ".\n" ; |
3360 | }); |
3361 | |
3362 | // Split the predecessors we are threading into a new edge block. We'll |
3363 | // clone the instructions into this block, and then redirect it to RealDest. |
3364 | BasicBlock *EdgeBB = SplitBlockPredecessors(BB, Preds: PredBBs, Suffix: ".critedge" , DTU); |
3365 | |
3366 | // TODO: These just exist to reduce test diff, we can drop them if we like. |
3367 | EdgeBB->setName(RealDest->getName() + ".critedge" ); |
3368 | EdgeBB->moveBefore(MovePos: RealDest); |
3369 | |
3370 | // Update PHI nodes. |
3371 | AddPredecessorToBlock(Succ: RealDest, NewPred: EdgeBB, ExistPred: BB); |
3372 | |
3373 | // BB may have instructions that are being threaded over. Clone these |
3374 | // instructions into EdgeBB. We know that there will be no uses of the |
3375 | // cloned instructions outside of EdgeBB. |
3376 | BasicBlock::iterator InsertPt = EdgeBB->getFirstInsertionPt(); |
3377 | DenseMap<Value *, Value *> TranslateMap; // Track translated values. |
3378 | TranslateMap[Cond] = CB; |
3379 | |
3380 | // RemoveDIs: track instructions that we optimise away while folding, so |
3381 | // that we can copy DbgVariableRecords from them later. |
3382 | BasicBlock::iterator SrcDbgCursor = BB->begin(); |
3383 | for (BasicBlock::iterator BBI = BB->begin(); &*BBI != BI; ++BBI) { |
3384 | if (PHINode *PN = dyn_cast<PHINode>(Val&: BBI)) { |
3385 | TranslateMap[PN] = PN->getIncomingValueForBlock(BB: EdgeBB); |
3386 | continue; |
3387 | } |
3388 | // Clone the instruction. |
3389 | Instruction *N = BBI->clone(); |
3390 | // Insert the new instruction into its new home. |
3391 | N->insertInto(ParentBB: EdgeBB, It: InsertPt); |
3392 | |
3393 | if (BBI->hasName()) |
3394 | N->setName(BBI->getName() + ".c" ); |
3395 | |
3396 | // Update operands due to translation. |
3397 | for (Use &Op : N->operands()) { |
3398 | DenseMap<Value *, Value *>::iterator PI = TranslateMap.find(Val: Op); |
3399 | if (PI != TranslateMap.end()) |
3400 | Op = PI->second; |
3401 | } |
3402 | |
3403 | // Check for trivial simplification. |
3404 | if (Value *V = simplifyInstruction(I: N, Q: {DL, nullptr, nullptr, AC})) { |
3405 | if (!BBI->use_empty()) |
3406 | TranslateMap[&*BBI] = V; |
3407 | if (!N->mayHaveSideEffects()) { |
3408 | N->eraseFromParent(); // Instruction folded away, don't need actual |
3409 | // inst |
3410 | N = nullptr; |
3411 | } |
3412 | } else { |
3413 | if (!BBI->use_empty()) |
3414 | TranslateMap[&*BBI] = N; |
3415 | } |
3416 | if (N) { |
3417 | // Copy all debug-info attached to instructions from the last we |
3418 | // successfully clone, up to this instruction (they might have been |
3419 | // folded away). |
3420 | for (; SrcDbgCursor != BBI; ++SrcDbgCursor) |
3421 | N->cloneDebugInfoFrom(From: &*SrcDbgCursor); |
3422 | SrcDbgCursor = std::next(x: BBI); |
3423 | // Clone debug-info on this instruction too. |
3424 | N->cloneDebugInfoFrom(From: &*BBI); |
3425 | |
3426 | // Register the new instruction with the assumption cache if necessary. |
3427 | if (auto *Assume = dyn_cast<AssumeInst>(Val: N)) |
3428 | if (AC) |
3429 | AC->registerAssumption(CI: Assume); |
3430 | } |
3431 | } |
3432 | |
3433 | for (; &*SrcDbgCursor != BI; ++SrcDbgCursor) |
3434 | InsertPt->cloneDebugInfoFrom(From: &*SrcDbgCursor); |
3435 | InsertPt->cloneDebugInfoFrom(From: BI); |
3436 | |
3437 | BB->removePredecessor(Pred: EdgeBB); |
3438 | BranchInst *EdgeBI = cast<BranchInst>(Val: EdgeBB->getTerminator()); |
3439 | EdgeBI->setSuccessor(idx: 0, NewSucc: RealDest); |
3440 | EdgeBI->setDebugLoc(BI->getDebugLoc()); |
3441 | |
3442 | if (DTU) { |
3443 | SmallVector<DominatorTree::UpdateType, 2> Updates; |
3444 | Updates.push_back(Elt: {DominatorTree::Delete, EdgeBB, BB}); |
3445 | Updates.push_back(Elt: {DominatorTree::Insert, EdgeBB, RealDest}); |
3446 | DTU->applyUpdates(Updates); |
3447 | } |
3448 | |
3449 | // For simplicity, we created a separate basic block for the edge. Merge |
3450 | // it back into the predecessor if possible. This not only avoids |
3451 | // unnecessary SimplifyCFG iterations, but also makes sure that we don't |
3452 | // bypass the check for trivial cycles above. |
3453 | MergeBlockIntoPredecessor(BB: EdgeBB, DTU); |
3454 | |
3455 | // Signal repeat, simplifying any other constants. |
3456 | return std::nullopt; |
3457 | } |
3458 | |
3459 | return false; |
3460 | } |
3461 | |
3462 | static bool FoldCondBranchOnValueKnownInPredecessor(BranchInst *BI, |
3463 | DomTreeUpdater *DTU, |
3464 | const DataLayout &DL, |
3465 | AssumptionCache *AC) { |
3466 | std::optional<bool> Result; |
3467 | bool EverChanged = false; |
3468 | do { |
3469 | // Note that None means "we changed things, but recurse further." |
3470 | Result = FoldCondBranchOnValueKnownInPredecessorImpl(BI, DTU, DL, AC); |
3471 | EverChanged |= Result == std::nullopt || *Result; |
3472 | } while (Result == std::nullopt); |
3473 | return EverChanged; |
3474 | } |
3475 | |
3476 | /// Given a BB that starts with the specified two-entry PHI node, |
3477 | /// see if we can eliminate it. |
3478 | static bool FoldTwoEntryPHINode(PHINode *PN, const TargetTransformInfo &TTI, |
3479 | DomTreeUpdater *DTU, const DataLayout &DL, |
3480 | bool SpeculateUnpredictables) { |
3481 | // Ok, this is a two entry PHI node. Check to see if this is a simple "if |
3482 | // statement", which has a very simple dominance structure. Basically, we |
3483 | // are trying to find the condition that is being branched on, which |
3484 | // subsequently causes this merge to happen. We really want control |
3485 | // dependence information for this check, but simplifycfg can't keep it up |
3486 | // to date, and this catches most of the cases we care about anyway. |
3487 | BasicBlock *BB = PN->getParent(); |
3488 | |
3489 | BasicBlock *IfTrue, *IfFalse; |
3490 | BranchInst *DomBI = GetIfCondition(BB, IfTrue, IfFalse); |
3491 | if (!DomBI) |
3492 | return false; |
3493 | Value *IfCond = DomBI->getCondition(); |
3494 | // Don't bother if the branch will be constant folded trivially. |
3495 | if (isa<ConstantInt>(Val: IfCond)) |
3496 | return false; |
3497 | |
3498 | BasicBlock *DomBlock = DomBI->getParent(); |
3499 | SmallVector<BasicBlock *, 2> IfBlocks; |
3500 | llvm::copy_if( |
3501 | Range: PN->blocks(), Out: std::back_inserter(x&: IfBlocks), P: [](BasicBlock *IfBlock) { |
3502 | return cast<BranchInst>(Val: IfBlock->getTerminator())->isUnconditional(); |
3503 | }); |
3504 | assert((IfBlocks.size() == 1 || IfBlocks.size() == 2) && |
3505 | "Will have either one or two blocks to speculate." ); |
3506 | |
3507 | // If the branch is non-unpredictable, see if we either predictably jump to |
3508 | // the merge bb (if we have only a single 'then' block), or if we predictably |
3509 | // jump to one specific 'then' block (if we have two of them). |
3510 | // It isn't beneficial to speculatively execute the code |
3511 | // from the block that we know is predictably not entered. |
3512 | bool IsUnpredictable = DomBI->getMetadata(KindID: LLVMContext::MD_unpredictable); |
3513 | if (!IsUnpredictable) { |
3514 | uint64_t TWeight, FWeight; |
3515 | if (extractBranchWeights(I: *DomBI, TrueVal&: TWeight, FalseVal&: FWeight) && |
3516 | (TWeight + FWeight) != 0) { |
3517 | BranchProbability BITrueProb = |
3518 | BranchProbability::getBranchProbability(Numerator: TWeight, Denominator: TWeight + FWeight); |
3519 | BranchProbability Likely = TTI.getPredictableBranchThreshold(); |
3520 | BranchProbability BIFalseProb = BITrueProb.getCompl(); |
3521 | if (IfBlocks.size() == 1) { |
3522 | BranchProbability BIBBProb = |
3523 | DomBI->getSuccessor(i: 0) == BB ? BITrueProb : BIFalseProb; |
3524 | if (BIBBProb >= Likely) |
3525 | return false; |
3526 | } else { |
3527 | if (BITrueProb >= Likely || BIFalseProb >= Likely) |
3528 | return false; |
3529 | } |
3530 | } |
3531 | } |
3532 | |
3533 | // Don't try to fold an unreachable block. For example, the phi node itself |
3534 | // can't be the candidate if-condition for a select that we want to form. |
3535 | if (auto *IfCondPhiInst = dyn_cast<PHINode>(Val: IfCond)) |
3536 | if (IfCondPhiInst->getParent() == BB) |
3537 | return false; |
3538 | |
3539 | // Okay, we found that we can merge this two-entry phi node into a select. |
3540 | // Doing so would require us to fold *all* two entry phi nodes in this block. |
3541 | // At some point this becomes non-profitable (particularly if the target |
3542 | // doesn't support cmov's). Only do this transformation if there are two or |
3543 | // fewer PHI nodes in this block. |
3544 | unsigned NumPhis = 0; |
3545 | for (BasicBlock::iterator I = BB->begin(); isa<PHINode>(Val: I); ++NumPhis, ++I) |
3546 | if (NumPhis > 2) |
3547 | return false; |
3548 | |
3549 | // Loop over the PHI's seeing if we can promote them all to select |
3550 | // instructions. While we are at it, keep track of the instructions |
3551 | // that need to be moved to the dominating block. |
3552 | SmallPtrSet<Instruction *, 4> AggressiveInsts; |
3553 | InstructionCost Cost = 0; |
3554 | InstructionCost Budget = |
3555 | TwoEntryPHINodeFoldingThreshold * TargetTransformInfo::TCC_Basic; |
3556 | if (SpeculateUnpredictables && IsUnpredictable) |
3557 | Budget += TTI.getBranchMispredictPenalty(); |
3558 | |
3559 | bool Changed = false; |
3560 | for (BasicBlock::iterator II = BB->begin(); isa<PHINode>(Val: II);) { |
3561 | PHINode *PN = cast<PHINode>(Val: II++); |
3562 | if (Value *V = simplifyInstruction(I: PN, Q: {DL, PN})) { |
3563 | PN->replaceAllUsesWith(V); |
3564 | PN->eraseFromParent(); |
3565 | Changed = true; |
3566 | continue; |
3567 | } |
3568 | |
3569 | if (!dominatesMergePoint(V: PN->getIncomingValue(i: 0), BB, AggressiveInsts, |
3570 | Cost, Budget, TTI) || |
3571 | !dominatesMergePoint(V: PN->getIncomingValue(i: 1), BB, AggressiveInsts, |
3572 | Cost, Budget, TTI)) |
3573 | return Changed; |
3574 | } |
3575 | |
3576 | // If we folded the first phi, PN dangles at this point. Refresh it. If |
3577 | // we ran out of PHIs then we simplified them all. |
3578 | PN = dyn_cast<PHINode>(Val: BB->begin()); |
3579 | if (!PN) |
3580 | return true; |
3581 | |
3582 | // Return true if at least one of these is a 'not', and another is either |
3583 | // a 'not' too, or a constant. |
3584 | auto CanHoistNotFromBothValues = [](Value *V0, Value *V1) { |
3585 | if (!match(V: V0, P: m_Not(V: m_Value()))) |
3586 | std::swap(a&: V0, b&: V1); |
3587 | auto Invertible = m_CombineOr(L: m_Not(V: m_Value()), R: m_AnyIntegralConstant()); |
3588 | return match(V: V0, P: m_Not(V: m_Value())) && match(V: V1, P: Invertible); |
3589 | }; |
3590 | |
3591 | // Don't fold i1 branches on PHIs which contain binary operators or |
3592 | // (possibly inverted) select form of or/ands, unless one of |
3593 | // the incoming values is an 'not' and another one is freely invertible. |
3594 | // These can often be turned into switches and other things. |
3595 | auto IsBinOpOrAnd = [](Value *V) { |
3596 | return match( |
3597 | V, P: m_CombineOr( |
3598 | L: m_BinOp(), |
3599 | R: m_CombineOr(L: m_Select(C: m_Value(), L: m_ImmConstant(), R: m_Value()), |
3600 | R: m_Select(C: m_Value(), L: m_Value(), R: m_ImmConstant())))); |
3601 | }; |
3602 | if (PN->getType()->isIntegerTy(Bitwidth: 1) && |
3603 | (IsBinOpOrAnd(PN->getIncomingValue(i: 0)) || |
3604 | IsBinOpOrAnd(PN->getIncomingValue(i: 1)) || IsBinOpOrAnd(IfCond)) && |
3605 | !CanHoistNotFromBothValues(PN->getIncomingValue(i: 0), |
3606 | PN->getIncomingValue(i: 1))) |
3607 | return Changed; |
3608 | |
3609 | // If all PHI nodes are promotable, check to make sure that all instructions |
3610 | // in the predecessor blocks can be promoted as well. If not, we won't be able |
3611 | // to get rid of the control flow, so it's not worth promoting to select |
3612 | // instructions. |
3613 | for (BasicBlock *IfBlock : IfBlocks) |
3614 | for (BasicBlock::iterator I = IfBlock->begin(); !I->isTerminator(); ++I) |
3615 | if (!AggressiveInsts.count(Ptr: &*I) && !I->isDebugOrPseudoInst()) { |
3616 | // This is not an aggressive instruction that we can promote. |
3617 | // Because of this, we won't be able to get rid of the control flow, so |
3618 | // the xform is not worth it. |
3619 | return Changed; |
3620 | } |
3621 | |
3622 | // If either of the blocks has it's address taken, we can't do this fold. |
3623 | if (any_of(Range&: IfBlocks, |
3624 | P: [](BasicBlock *IfBlock) { return IfBlock->hasAddressTaken(); })) |
3625 | return Changed; |
3626 | |
3627 | LLVM_DEBUG(dbgs() << "FOUND IF CONDITION! " << *IfCond; |
3628 | if (IsUnpredictable) dbgs() << " (unpredictable)" ; |
3629 | dbgs() << " T: " << IfTrue->getName() |
3630 | << " F: " << IfFalse->getName() << "\n" ); |
3631 | |
3632 | // If we can still promote the PHI nodes after this gauntlet of tests, |
3633 | // do all of the PHI's now. |
3634 | |
3635 | // Move all 'aggressive' instructions, which are defined in the |
3636 | // conditional parts of the if's up to the dominating block. |
3637 | for (BasicBlock *IfBlock : IfBlocks) |
3638 | hoistAllInstructionsInto(DomBlock, InsertPt: DomBI, BB: IfBlock); |
3639 | |
3640 | IRBuilder<NoFolder> Builder(DomBI); |
3641 | // Propagate fast-math-flags from phi nodes to replacement selects. |
3642 | IRBuilder<>::FastMathFlagGuard FMFGuard(Builder); |
3643 | while (PHINode *PN = dyn_cast<PHINode>(Val: BB->begin())) { |
3644 | if (isa<FPMathOperator>(Val: PN)) |
3645 | Builder.setFastMathFlags(PN->getFastMathFlags()); |
3646 | |
3647 | // Change the PHI node into a select instruction. |
3648 | Value *TrueVal = PN->getIncomingValueForBlock(BB: IfTrue); |
3649 | Value *FalseVal = PN->getIncomingValueForBlock(BB: IfFalse); |
3650 | |
3651 | Value *Sel = Builder.CreateSelect(C: IfCond, True: TrueVal, False: FalseVal, Name: "" , MDFrom: DomBI); |
3652 | PN->replaceAllUsesWith(V: Sel); |
3653 | Sel->takeName(V: PN); |
3654 | PN->eraseFromParent(); |
3655 | } |
3656 | |
3657 | // At this point, all IfBlocks are empty, so our if statement |
3658 | // has been flattened. Change DomBlock to jump directly to our new block to |
3659 | // avoid other simplifycfg's kicking in on the diamond. |
3660 | Builder.CreateBr(Dest: BB); |
3661 | |
3662 | SmallVector<DominatorTree::UpdateType, 3> Updates; |
3663 | if (DTU) { |
3664 | Updates.push_back(Elt: {DominatorTree::Insert, DomBlock, BB}); |
3665 | for (auto *Successor : successors(BB: DomBlock)) |
3666 | Updates.push_back(Elt: {DominatorTree::Delete, DomBlock, Successor}); |
3667 | } |
3668 | |
3669 | DomBI->eraseFromParent(); |
3670 | if (DTU) |
3671 | DTU->applyUpdates(Updates); |
3672 | |
3673 | return true; |
3674 | } |
3675 | |
3676 | static Value *createLogicalOp(IRBuilderBase &Builder, |
3677 | Instruction::BinaryOps Opc, Value *LHS, |
3678 | Value *RHS, const Twine &Name = "" ) { |
3679 | // Try to relax logical op to binary op. |
3680 | if (impliesPoison(ValAssumedPoison: RHS, V: LHS)) |
3681 | return Builder.CreateBinOp(Opc, LHS, RHS, Name); |
3682 | if (Opc == Instruction::And) |
3683 | return Builder.CreateLogicalAnd(Cond1: LHS, Cond2: RHS, Name); |
3684 | if (Opc == Instruction::Or) |
3685 | return Builder.CreateLogicalOr(Cond1: LHS, Cond2: RHS, Name); |
3686 | llvm_unreachable("Invalid logical opcode" ); |
3687 | } |
3688 | |
3689 | /// Return true if either PBI or BI has branch weight available, and store |
3690 | /// the weights in {Pred|Succ}{True|False}Weight. If one of PBI and BI does |
3691 | /// not have branch weight, use 1:1 as its weight. |
3692 | static bool (BranchInst *PBI, BranchInst *BI, |
3693 | uint64_t &PredTrueWeight, |
3694 | uint64_t &PredFalseWeight, |
3695 | uint64_t &SuccTrueWeight, |
3696 | uint64_t &SuccFalseWeight) { |
3697 | bool PredHasWeights = |
3698 | extractBranchWeights(I: *PBI, TrueVal&: PredTrueWeight, FalseVal&: PredFalseWeight); |
3699 | bool SuccHasWeights = |
3700 | extractBranchWeights(I: *BI, TrueVal&: SuccTrueWeight, FalseVal&: SuccFalseWeight); |
3701 | if (PredHasWeights || SuccHasWeights) { |
3702 | if (!PredHasWeights) |
3703 | PredTrueWeight = PredFalseWeight = 1; |
3704 | if (!SuccHasWeights) |
3705 | SuccTrueWeight = SuccFalseWeight = 1; |
3706 | return true; |
3707 | } else { |
3708 | return false; |
3709 | } |
3710 | } |
3711 | |
3712 | /// Determine if the two branches share a common destination and deduce a glue |
3713 | /// that joins the branches' conditions to arrive at the common destination if |
3714 | /// that would be profitable. |
3715 | static std::optional<std::tuple<BasicBlock *, Instruction::BinaryOps, bool>> |
3716 | shouldFoldCondBranchesToCommonDestination(BranchInst *BI, BranchInst *PBI, |
3717 | const TargetTransformInfo *TTI) { |
3718 | assert(BI && PBI && BI->isConditional() && PBI->isConditional() && |
3719 | "Both blocks must end with a conditional branches." ); |
3720 | assert(is_contained(predecessors(BI->getParent()), PBI->getParent()) && |
3721 | "PredBB must be a predecessor of BB." ); |
3722 | |
3723 | // We have the potential to fold the conditions together, but if the |
3724 | // predecessor branch is predictable, we may not want to merge them. |
3725 | uint64_t PTWeight, PFWeight; |
3726 | BranchProbability PBITrueProb, Likely; |
3727 | if (TTI && !PBI->getMetadata(KindID: LLVMContext::MD_unpredictable) && |
3728 | extractBranchWeights(I: *PBI, TrueVal&: PTWeight, FalseVal&: PFWeight) && |
3729 | (PTWeight + PFWeight) != 0) { |
3730 | PBITrueProb = |
3731 | BranchProbability::getBranchProbability(Numerator: PTWeight, Denominator: PTWeight + PFWeight); |
3732 | Likely = TTI->getPredictableBranchThreshold(); |
3733 | } |
3734 | |
3735 | if (PBI->getSuccessor(i: 0) == BI->getSuccessor(i: 0)) { |
3736 | // Speculate the 2nd condition unless the 1st is probably true. |
3737 | if (PBITrueProb.isUnknown() || PBITrueProb < Likely) |
3738 | return {{BI->getSuccessor(i: 0), Instruction::Or, false}}; |
3739 | } else if (PBI->getSuccessor(i: 1) == BI->getSuccessor(i: 1)) { |
3740 | // Speculate the 2nd condition unless the 1st is probably false. |
3741 | if (PBITrueProb.isUnknown() || PBITrueProb.getCompl() < Likely) |
3742 | return {{BI->getSuccessor(i: 1), Instruction::And, false}}; |
3743 | } else if (PBI->getSuccessor(i: 0) == BI->getSuccessor(i: 1)) { |
3744 | // Speculate the 2nd condition unless the 1st is probably true. |
3745 | if (PBITrueProb.isUnknown() || PBITrueProb < Likely) |
3746 | return {{BI->getSuccessor(i: 1), Instruction::And, true}}; |
3747 | } else if (PBI->getSuccessor(i: 1) == BI->getSuccessor(i: 0)) { |
3748 | // Speculate the 2nd condition unless the 1st is probably false. |
3749 | if (PBITrueProb.isUnknown() || PBITrueProb.getCompl() < Likely) |
3750 | return {{BI->getSuccessor(i: 0), Instruction::Or, true}}; |
3751 | } |
3752 | return std::nullopt; |
3753 | } |
3754 | |
3755 | static bool performBranchToCommonDestFolding(BranchInst *BI, BranchInst *PBI, |
3756 | DomTreeUpdater *DTU, |
3757 | MemorySSAUpdater *MSSAU, |
3758 | const TargetTransformInfo *TTI) { |
3759 | BasicBlock *BB = BI->getParent(); |
3760 | BasicBlock *PredBlock = PBI->getParent(); |
3761 | |
3762 | // Determine if the two branches share a common destination. |
3763 | BasicBlock *CommonSucc; |
3764 | Instruction::BinaryOps Opc; |
3765 | bool InvertPredCond; |
3766 | std::tie(args&: CommonSucc, args&: Opc, args&: InvertPredCond) = |
3767 | *shouldFoldCondBranchesToCommonDestination(BI, PBI, TTI); |
3768 | |
3769 | LLVM_DEBUG(dbgs() << "FOLDING BRANCH TO COMMON DEST:\n" << *PBI << *BB); |
3770 | |
3771 | IRBuilder<> Builder(PBI); |
3772 | // The builder is used to create instructions to eliminate the branch in BB. |
3773 | // If BB's terminator has !annotation metadata, add it to the new |
3774 | // instructions. |
3775 | Builder.CollectMetadataToCopy(Src: BB->getTerminator(), |
3776 | MetadataKinds: {LLVMContext::MD_annotation}); |
3777 | |
3778 | // If we need to invert the condition in the pred block to match, do so now. |
3779 | if (InvertPredCond) { |
3780 | InvertBranch(PBI, Builder); |
3781 | } |
3782 | |
3783 | BasicBlock *UniqueSucc = |
3784 | PBI->getSuccessor(i: 0) == BB ? BI->getSuccessor(i: 0) : BI->getSuccessor(i: 1); |
3785 | |
3786 | // Before cloning instructions, notify the successor basic block that it |
3787 | // is about to have a new predecessor. This will update PHI nodes, |
3788 | // which will allow us to update live-out uses of bonus instructions. |
3789 | AddPredecessorToBlock(Succ: UniqueSucc, NewPred: PredBlock, ExistPred: BB, MSSAU); |
3790 | |
3791 | // Try to update branch weights. |
3792 | uint64_t PredTrueWeight, PredFalseWeight, SuccTrueWeight, SuccFalseWeight; |
3793 | if (extractPredSuccWeights(PBI, BI, PredTrueWeight, PredFalseWeight, |
3794 | SuccTrueWeight, SuccFalseWeight)) { |
3795 | SmallVector<uint64_t, 8> NewWeights; |
3796 | |
3797 | if (PBI->getSuccessor(i: 0) == BB) { |
3798 | // PBI: br i1 %x, BB, FalseDest |
3799 | // BI: br i1 %y, UniqueSucc, FalseDest |
3800 | // TrueWeight is TrueWeight for PBI * TrueWeight for BI. |
3801 | NewWeights.push_back(Elt: PredTrueWeight * SuccTrueWeight); |
3802 | // FalseWeight is FalseWeight for PBI * TotalWeight for BI + |
3803 | // TrueWeight for PBI * FalseWeight for BI. |
3804 | // We assume that total weights of a BranchInst can fit into 32 bits. |
3805 | // Therefore, we will not have overflow using 64-bit arithmetic. |
3806 | NewWeights.push_back(Elt: PredFalseWeight * |
3807 | (SuccFalseWeight + SuccTrueWeight) + |
3808 | PredTrueWeight * SuccFalseWeight); |
3809 | } else { |
3810 | // PBI: br i1 %x, TrueDest, BB |
3811 | // BI: br i1 %y, TrueDest, UniqueSucc |
3812 | // TrueWeight is TrueWeight for PBI * TotalWeight for BI + |
3813 | // FalseWeight for PBI * TrueWeight for BI. |
3814 | NewWeights.push_back(Elt: PredTrueWeight * (SuccFalseWeight + SuccTrueWeight) + |
3815 | PredFalseWeight * SuccTrueWeight); |
3816 | // FalseWeight is FalseWeight for PBI * FalseWeight for BI. |
3817 | NewWeights.push_back(Elt: PredFalseWeight * SuccFalseWeight); |
3818 | } |
3819 | |
3820 | // Halve the weights if any of them cannot fit in an uint32_t |
3821 | FitWeights(Weights: NewWeights); |
3822 | |
3823 | SmallVector<uint32_t, 8> MDWeights(NewWeights.begin(), NewWeights.end()); |
3824 | setBranchWeights(I: PBI, TrueWeight: MDWeights[0], FalseWeight: MDWeights[1], /*IsExpected=*/false); |
3825 | |
3826 | // TODO: If BB is reachable from all paths through PredBlock, then we |
3827 | // could replace PBI's branch probabilities with BI's. |
3828 | } else |
3829 | PBI->setMetadata(KindID: LLVMContext::MD_prof, Node: nullptr); |
3830 | |
3831 | // Now, update the CFG. |
3832 | PBI->setSuccessor(idx: PBI->getSuccessor(i: 0) != BB, NewSucc: UniqueSucc); |
3833 | |
3834 | if (DTU) |
3835 | DTU->applyUpdates(Updates: {{DominatorTree::Insert, PredBlock, UniqueSucc}, |
3836 | {DominatorTree::Delete, PredBlock, BB}}); |
3837 | |
3838 | // If BI was a loop latch, it may have had associated loop metadata. |
3839 | // We need to copy it to the new latch, that is, PBI. |
3840 | if (MDNode *LoopMD = BI->getMetadata(KindID: LLVMContext::MD_loop)) |
3841 | PBI->setMetadata(KindID: LLVMContext::MD_loop, Node: LoopMD); |
3842 | |
3843 | ValueToValueMapTy VMap; // maps original values to cloned values |
3844 | CloneInstructionsIntoPredecessorBlockAndUpdateSSAUses(BB, PredBlock, VMap); |
3845 | |
3846 | Module *M = BB->getModule(); |
3847 | |
3848 | if (PredBlock->IsNewDbgInfoFormat) { |
3849 | PredBlock->getTerminator()->cloneDebugInfoFrom(From: BB->getTerminator()); |
3850 | for (DbgVariableRecord &DVR : |
3851 | filterDbgVars(R: PredBlock->getTerminator()->getDbgRecordRange())) { |
3852 | RemapDbgRecord(M, DR: &DVR, VM&: VMap, |
3853 | Flags: RF_NoModuleLevelChanges | RF_IgnoreMissingLocals); |
3854 | } |
3855 | } |
3856 | |
3857 | // Now that the Cond was cloned into the predecessor basic block, |
3858 | // or/and the two conditions together. |
3859 | Value *BICond = VMap[BI->getCondition()]; |
3860 | PBI->setCondition( |
3861 | createLogicalOp(Builder, Opc, LHS: PBI->getCondition(), RHS: BICond, Name: "or.cond" )); |
3862 | |
3863 | ++NumFoldBranchToCommonDest; |
3864 | return true; |
3865 | } |
3866 | |
3867 | /// Return if an instruction's type or any of its operands' types are a vector |
3868 | /// type. |
3869 | static bool isVectorOp(Instruction &I) { |
3870 | return I.getType()->isVectorTy() || any_of(Range: I.operands(), P: [](Use &U) { |
3871 | return U->getType()->isVectorTy(); |
3872 | }); |
3873 | } |
3874 | |
3875 | /// If this basic block is simple enough, and if a predecessor branches to us |
3876 | /// and one of our successors, fold the block into the predecessor and use |
3877 | /// logical operations to pick the right destination. |
3878 | bool llvm::FoldBranchToCommonDest(BranchInst *BI, DomTreeUpdater *DTU, |
3879 | MemorySSAUpdater *MSSAU, |
3880 | const TargetTransformInfo *TTI, |
3881 | unsigned BonusInstThreshold) { |
3882 | // If this block ends with an unconditional branch, |
3883 | // let SpeculativelyExecuteBB() deal with it. |
3884 | if (!BI->isConditional()) |
3885 | return false; |
3886 | |
3887 | BasicBlock *BB = BI->getParent(); |
3888 | TargetTransformInfo::TargetCostKind CostKind = |
3889 | BB->getParent()->hasMinSize() ? TargetTransformInfo::TCK_CodeSize |
3890 | : TargetTransformInfo::TCK_SizeAndLatency; |
3891 | |
3892 | Instruction *Cond = dyn_cast<Instruction>(Val: BI->getCondition()); |
3893 | |
3894 | if (!Cond || |
3895 | (!isa<CmpInst>(Val: Cond) && !isa<BinaryOperator>(Val: Cond) && |
3896 | !isa<SelectInst>(Val: Cond)) || |
3897 | Cond->getParent() != BB || !Cond->hasOneUse()) |
3898 | return false; |
3899 | |
3900 | // Finally, don't infinitely unroll conditional loops. |
3901 | if (is_contained(Range: successors(BB), Element: BB)) |
3902 | return false; |
3903 | |
3904 | // With which predecessors will we want to deal with? |
3905 | SmallVector<BasicBlock *, 8> Preds; |
3906 | for (BasicBlock *PredBlock : predecessors(BB)) { |
3907 | BranchInst *PBI = dyn_cast<BranchInst>(Val: PredBlock->getTerminator()); |
3908 | |
3909 | // Check that we have two conditional branches. If there is a PHI node in |
3910 | // the common successor, verify that the same value flows in from both |
3911 | // blocks. |
3912 | if (!PBI || PBI->isUnconditional() || !SafeToMergeTerminators(SI1: BI, SI2: PBI)) |
3913 | continue; |
3914 | |
3915 | // Determine if the two branches share a common destination. |
3916 | BasicBlock *CommonSucc; |
3917 | Instruction::BinaryOps Opc; |
3918 | bool InvertPredCond; |
3919 | if (auto Recipe = shouldFoldCondBranchesToCommonDestination(BI, PBI, TTI)) |
3920 | std::tie(args&: CommonSucc, args&: Opc, args&: InvertPredCond) = *Recipe; |
3921 | else |
3922 | continue; |
3923 | |
3924 | // Check the cost of inserting the necessary logic before performing the |
3925 | // transformation. |
3926 | if (TTI) { |
3927 | Type *Ty = BI->getCondition()->getType(); |
3928 | InstructionCost Cost = TTI->getArithmeticInstrCost(Opcode: Opc, Ty, CostKind); |
3929 | if (InvertPredCond && (!PBI->getCondition()->hasOneUse() || |
3930 | !isa<CmpInst>(Val: PBI->getCondition()))) |
3931 | Cost += TTI->getArithmeticInstrCost(Opcode: Instruction::Xor, Ty, CostKind); |
3932 | |
3933 | if (Cost > BranchFoldThreshold) |
3934 | continue; |
3935 | } |
3936 | |
3937 | // Ok, we do want to deal with this predecessor. Record it. |
3938 | Preds.emplace_back(Args&: PredBlock); |
3939 | } |
3940 | |
3941 | // If there aren't any predecessors into which we can fold, |
3942 | // don't bother checking the cost. |
3943 | if (Preds.empty()) |
3944 | return false; |
3945 | |
3946 | // Only allow this transformation if computing the condition doesn't involve |
3947 | // too many instructions and these involved instructions can be executed |
3948 | // unconditionally. We denote all involved instructions except the condition |
3949 | // as "bonus instructions", and only allow this transformation when the |
3950 | // number of the bonus instructions we'll need to create when cloning into |
3951 | // each predecessor does not exceed a certain threshold. |
3952 | unsigned NumBonusInsts = 0; |
3953 | bool SawVectorOp = false; |
3954 | const unsigned PredCount = Preds.size(); |
3955 | for (Instruction &I : *BB) { |
3956 | // Don't check the branch condition comparison itself. |
3957 | if (&I == Cond) |
3958 | continue; |
3959 | // Ignore dbg intrinsics, and the terminator. |
3960 | if (isa<DbgInfoIntrinsic>(Val: I) || isa<BranchInst>(Val: I)) |
3961 | continue; |
3962 | // I must be safe to execute unconditionally. |
3963 | if (!isSafeToSpeculativelyExecute(I: &I)) |
3964 | return false; |
3965 | SawVectorOp |= isVectorOp(I); |
3966 | |
3967 | // Account for the cost of duplicating this instruction into each |
3968 | // predecessor. Ignore free instructions. |
3969 | if (!TTI || TTI->getInstructionCost(U: &I, CostKind) != |
3970 | TargetTransformInfo::TCC_Free) { |
3971 | NumBonusInsts += PredCount; |
3972 | |
3973 | // Early exits once we reach the limit. |
3974 | if (NumBonusInsts > |
3975 | BonusInstThreshold * BranchFoldToCommonDestVectorMultiplier) |
3976 | return false; |
3977 | } |
3978 | |
3979 | auto IsBCSSAUse = [BB, &I](Use &U) { |
3980 | auto *UI = cast<Instruction>(Val: U.getUser()); |
3981 | if (auto *PN = dyn_cast<PHINode>(Val: UI)) |
3982 | return PN->getIncomingBlock(U) == BB; |
3983 | return UI->getParent() == BB && I.comesBefore(Other: UI); |
3984 | }; |
3985 | |
3986 | // Does this instruction require rewriting of uses? |
3987 | if (!all_of(Range: I.uses(), P: IsBCSSAUse)) |
3988 | return false; |
3989 | } |
3990 | if (NumBonusInsts > |
3991 | BonusInstThreshold * |
3992 | (SawVectorOp ? BranchFoldToCommonDestVectorMultiplier : 1)) |
3993 | return false; |
3994 | |
3995 | // Ok, we have the budget. Perform the transformation. |
3996 | for (BasicBlock *PredBlock : Preds) { |
3997 | auto *PBI = cast<BranchInst>(Val: PredBlock->getTerminator()); |
3998 | return performBranchToCommonDestFolding(BI, PBI, DTU, MSSAU, TTI); |
3999 | } |
4000 | return false; |
4001 | } |
4002 | |
4003 | // If there is only one store in BB1 and BB2, return it, otherwise return |
4004 | // nullptr. |
4005 | static StoreInst *findUniqueStoreInBlocks(BasicBlock *BB1, BasicBlock *BB2) { |
4006 | StoreInst *S = nullptr; |
4007 | for (auto *BB : {BB1, BB2}) { |
4008 | if (!BB) |
4009 | continue; |
4010 | for (auto &I : *BB) |
4011 | if (auto *SI = dyn_cast<StoreInst>(Val: &I)) { |
4012 | if (S) |
4013 | // Multiple stores seen. |
4014 | return nullptr; |
4015 | else |
4016 | S = SI; |
4017 | } |
4018 | } |
4019 | return S; |
4020 | } |
4021 | |
4022 | static Value *ensureValueAvailableInSuccessor(Value *V, BasicBlock *BB, |
4023 | Value *AlternativeV = nullptr) { |
4024 | // PHI is going to be a PHI node that allows the value V that is defined in |
4025 | // BB to be referenced in BB's only successor. |
4026 | // |
4027 | // If AlternativeV is nullptr, the only value we care about in PHI is V. It |
4028 | // doesn't matter to us what the other operand is (it'll never get used). We |
4029 | // could just create a new PHI with an undef incoming value, but that could |
4030 | // increase register pressure if EarlyCSE/InstCombine can't fold it with some |
4031 | // other PHI. So here we directly look for some PHI in BB's successor with V |
4032 | // as an incoming operand. If we find one, we use it, else we create a new |
4033 | // one. |
4034 | // |
4035 | // If AlternativeV is not nullptr, we care about both incoming values in PHI. |
4036 | // PHI must be exactly: phi <ty> [ %BB, %V ], [ %OtherBB, %AlternativeV] |
4037 | // where OtherBB is the single other predecessor of BB's only successor. |
4038 | PHINode *PHI = nullptr; |
4039 | BasicBlock *Succ = BB->getSingleSuccessor(); |
4040 | |
4041 | for (auto I = Succ->begin(); isa<PHINode>(Val: I); ++I) |
4042 | if (cast<PHINode>(Val&: I)->getIncomingValueForBlock(BB) == V) { |
4043 | PHI = cast<PHINode>(Val&: I); |
4044 | if (!AlternativeV) |
4045 | break; |
4046 | |
4047 | assert(Succ->hasNPredecessors(2)); |
4048 | auto PredI = pred_begin(BB: Succ); |
4049 | BasicBlock *OtherPredBB = *PredI == BB ? *++PredI : *PredI; |
4050 | if (PHI->getIncomingValueForBlock(BB: OtherPredBB) == AlternativeV) |
4051 | break; |
4052 | PHI = nullptr; |
4053 | } |
4054 | if (PHI) |
4055 | return PHI; |
4056 | |
4057 | // If V is not an instruction defined in BB, just return it. |
4058 | if (!AlternativeV && |
4059 | (!isa<Instruction>(Val: V) || cast<Instruction>(Val: V)->getParent() != BB)) |
4060 | return V; |
4061 | |
4062 | PHI = PHINode::Create(Ty: V->getType(), NumReservedValues: 2, NameStr: "simplifycfg.merge" ); |
4063 | PHI->insertBefore(InsertPos: Succ->begin()); |
4064 | PHI->addIncoming(V, BB); |
4065 | for (BasicBlock *PredBB : predecessors(BB: Succ)) |
4066 | if (PredBB != BB) |
4067 | PHI->addIncoming( |
4068 | V: AlternativeV ? AlternativeV : PoisonValue::get(T: V->getType()), BB: PredBB); |
4069 | return PHI; |
4070 | } |
4071 | |
4072 | static bool mergeConditionalStoreToAddress( |
4073 | BasicBlock *PTB, BasicBlock *PFB, BasicBlock *QTB, BasicBlock *QFB, |
4074 | BasicBlock *PostBB, Value *Address, bool InvertPCond, bool InvertQCond, |
4075 | DomTreeUpdater *DTU, const DataLayout &DL, const TargetTransformInfo &TTI) { |
4076 | // For every pointer, there must be exactly two stores, one coming from |
4077 | // PTB or PFB, and the other from QTB or QFB. We don't support more than one |
4078 | // store (to any address) in PTB,PFB or QTB,QFB. |
4079 | // FIXME: We could relax this restriction with a bit more work and performance |
4080 | // testing. |
4081 | StoreInst *PStore = findUniqueStoreInBlocks(BB1: PTB, BB2: PFB); |
4082 | StoreInst *QStore = findUniqueStoreInBlocks(BB1: QTB, BB2: QFB); |
4083 | if (!PStore || !QStore) |
4084 | return false; |
4085 | |
4086 | // Now check the stores are compatible. |
4087 | if (!QStore->isUnordered() || !PStore->isUnordered() || |
4088 | PStore->getValueOperand()->getType() != |
4089 | QStore->getValueOperand()->getType()) |
4090 | return false; |
4091 | |
4092 | // Check that sinking the store won't cause program behavior changes. Sinking |
4093 | // the store out of the Q blocks won't change any behavior as we're sinking |
4094 | // from a block to its unconditional successor. But we're moving a store from |
4095 | // the P blocks down through the middle block (QBI) and past both QFB and QTB. |
4096 | // So we need to check that there are no aliasing loads or stores in |
4097 | // QBI, QTB and QFB. We also need to check there are no conflicting memory |
4098 | // operations between PStore and the end of its parent block. |
4099 | // |
4100 | // The ideal way to do this is to query AliasAnalysis, but we don't |
4101 | // preserve AA currently so that is dangerous. Be super safe and just |
4102 | // check there are no other memory operations at all. |
4103 | for (auto &I : *QFB->getSinglePredecessor()) |
4104 | if (I.mayReadOrWriteMemory()) |
4105 | return false; |
4106 | for (auto &I : *QFB) |
4107 | if (&I != QStore && I.mayReadOrWriteMemory()) |
4108 | return false; |
4109 | if (QTB) |
4110 | for (auto &I : *QTB) |
4111 | if (&I != QStore && I.mayReadOrWriteMemory()) |
4112 | return false; |
4113 | for (auto I = BasicBlock::iterator(PStore), E = PStore->getParent()->end(); |
4114 | I != E; ++I) |
4115 | if (&*I != PStore && I->mayReadOrWriteMemory()) |
4116 | return false; |
4117 | |
4118 | // If we're not in aggressive mode, we only optimize if we have some |
4119 | // confidence that by optimizing we'll allow P and/or Q to be if-converted. |
4120 | auto IsWorthwhile = [&](BasicBlock *BB, ArrayRef<StoreInst *> FreeStores) { |
4121 | if (!BB) |
4122 | return true; |
4123 | // Heuristic: if the block can be if-converted/phi-folded and the |
4124 | // instructions inside are all cheap (arithmetic/GEPs), it's worthwhile to |
4125 | // thread this store. |
4126 | InstructionCost Cost = 0; |
4127 | InstructionCost Budget = |
4128 | PHINodeFoldingThreshold * TargetTransformInfo::TCC_Basic; |
4129 | for (auto &I : BB->instructionsWithoutDebug(SkipPseudoOp: false)) { |
4130 | // Consider terminator instruction to be free. |
4131 | if (I.isTerminator()) |
4132 | continue; |
4133 | // If this is one the stores that we want to speculate out of this BB, |
4134 | // then don't count it's cost, consider it to be free. |
4135 | if (auto *S = dyn_cast<StoreInst>(Val: &I)) |
4136 | if (llvm::find(Range&: FreeStores, Val: S)) |
4137 | continue; |
4138 | // Else, we have a white-list of instructions that we are ak speculating. |
4139 | if (!isa<BinaryOperator>(Val: I) && !isa<GetElementPtrInst>(Val: I)) |
4140 | return false; // Not in white-list - not worthwhile folding. |
4141 | // And finally, if this is a non-free instruction that we are okay |
4142 | // speculating, ensure that we consider the speculation budget. |
4143 | Cost += |
4144 | TTI.getInstructionCost(U: &I, CostKind: TargetTransformInfo::TCK_SizeAndLatency); |
4145 | if (Cost > Budget) |
4146 | return false; // Eagerly refuse to fold as soon as we're out of budget. |
4147 | } |
4148 | assert(Cost <= Budget && |
4149 | "When we run out of budget we will eagerly return from within the " |
4150 | "per-instruction loop." ); |
4151 | return true; |
4152 | }; |
4153 | |
4154 | const std::array<StoreInst *, 2> FreeStores = {PStore, QStore}; |
4155 | if (!MergeCondStoresAggressively && |
4156 | (!IsWorthwhile(PTB, FreeStores) || !IsWorthwhile(PFB, FreeStores) || |
4157 | !IsWorthwhile(QTB, FreeStores) || !IsWorthwhile(QFB, FreeStores))) |
4158 | return false; |
4159 | |
4160 | // If PostBB has more than two predecessors, we need to split it so we can |
4161 | // sink the store. |
4162 | if (std::next(x: pred_begin(BB: PostBB), n: 2) != pred_end(BB: PostBB)) { |
4163 | // We know that QFB's only successor is PostBB. And QFB has a single |
4164 | // predecessor. If QTB exists, then its only successor is also PostBB. |
4165 | // If QTB does not exist, then QFB's only predecessor has a conditional |
4166 | // branch to QFB and PostBB. |
4167 | BasicBlock *TruePred = QTB ? QTB : QFB->getSinglePredecessor(); |
4168 | BasicBlock *NewBB = |
4169 | SplitBlockPredecessors(BB: PostBB, Preds: {QFB, TruePred}, Suffix: "condstore.split" , DTU); |
4170 | if (!NewBB) |
4171 | return false; |
4172 | PostBB = NewBB; |
4173 | } |
4174 | |
4175 | // OK, we're going to sink the stores to PostBB. The store has to be |
4176 | // conditional though, so first create the predicate. |
4177 | Value *PCond = cast<BranchInst>(Val: PFB->getSinglePredecessor()->getTerminator()) |
4178 | ->getCondition(); |
4179 | Value *QCond = cast<BranchInst>(Val: QFB->getSinglePredecessor()->getTerminator()) |
4180 | ->getCondition(); |
4181 | |
4182 | Value *PPHI = ensureValueAvailableInSuccessor(V: PStore->getValueOperand(), |
4183 | BB: PStore->getParent()); |
4184 | Value *QPHI = ensureValueAvailableInSuccessor(V: QStore->getValueOperand(), |
4185 | BB: QStore->getParent(), AlternativeV: PPHI); |
4186 | |
4187 | BasicBlock::iterator PostBBFirst = PostBB->getFirstInsertionPt(); |
4188 | IRBuilder<> QB(PostBB, PostBBFirst); |
4189 | QB.SetCurrentDebugLocation(PostBBFirst->getStableDebugLoc()); |
4190 | |
4191 | Value *PPred = PStore->getParent() == PTB ? PCond : QB.CreateNot(V: PCond); |
4192 | Value *QPred = QStore->getParent() == QTB ? QCond : QB.CreateNot(V: QCond); |
4193 | |
4194 | if (InvertPCond) |
4195 | PPred = QB.CreateNot(V: PPred); |
4196 | if (InvertQCond) |
4197 | QPred = QB.CreateNot(V: QPred); |
4198 | Value *CombinedPred = QB.CreateOr(LHS: PPred, RHS: QPred); |
4199 | |
4200 | BasicBlock::iterator InsertPt = QB.GetInsertPoint(); |
4201 | auto *T = SplitBlockAndInsertIfThen(Cond: CombinedPred, SplitBefore: InsertPt, |
4202 | /*Unreachable=*/false, |
4203 | /*BranchWeights=*/nullptr, DTU); |
4204 | |
4205 | QB.SetInsertPoint(T); |
4206 | StoreInst *SI = cast<StoreInst>(Val: QB.CreateStore(Val: QPHI, Ptr: Address)); |
4207 | SI->setAAMetadata(PStore->getAAMetadata().merge(Other: QStore->getAAMetadata())); |
4208 | // Choose the minimum alignment. If we could prove both stores execute, we |
4209 | // could use biggest one. In this case, though, we only know that one of the |
4210 | // stores executes. And we don't know it's safe to take the alignment from a |
4211 | // store that doesn't execute. |
4212 | SI->setAlignment(std::min(a: PStore->getAlign(), b: QStore->getAlign())); |
4213 | |
4214 | QStore->eraseFromParent(); |
4215 | PStore->eraseFromParent(); |
4216 | |
4217 | return true; |
4218 | } |
4219 | |
4220 | static bool mergeConditionalStores(BranchInst *PBI, BranchInst *QBI, |
4221 | DomTreeUpdater *DTU, const DataLayout &DL, |
4222 | const TargetTransformInfo &TTI) { |
4223 | // The intention here is to find diamonds or triangles (see below) where each |
4224 | // conditional block contains a store to the same address. Both of these |
4225 | // stores are conditional, so they can't be unconditionally sunk. But it may |
4226 | // be profitable to speculatively sink the stores into one merged store at the |
4227 | // end, and predicate the merged store on the union of the two conditions of |
4228 | // PBI and QBI. |
4229 | // |
4230 | // This can reduce the number of stores executed if both of the conditions are |
4231 | // true, and can allow the blocks to become small enough to be if-converted. |
4232 | // This optimization will also chain, so that ladders of test-and-set |
4233 | // sequences can be if-converted away. |
4234 | // |
4235 | // We only deal with simple diamonds or triangles: |
4236 | // |
4237 | // PBI or PBI or a combination of the two |
4238 | // / \ | \ |
4239 | // PTB PFB | PFB |
4240 | // \ / | / |
4241 | // QBI QBI |
4242 | // / \ | \ |
4243 | // QTB QFB | QFB |
4244 | // \ / | / |
4245 | // PostBB PostBB |
4246 | // |
4247 | // We model triangles as a type of diamond with a nullptr "true" block. |
4248 | // Triangles are canonicalized so that the fallthrough edge is represented by |
4249 | // a true condition, as in the diagram above. |
4250 | BasicBlock *PTB = PBI->getSuccessor(i: 0); |
4251 | BasicBlock *PFB = PBI->getSuccessor(i: 1); |
4252 | BasicBlock *QTB = QBI->getSuccessor(i: 0); |
4253 | BasicBlock *QFB = QBI->getSuccessor(i: 1); |
4254 | BasicBlock *PostBB = QFB->getSingleSuccessor(); |
4255 | |
4256 | // Make sure we have a good guess for PostBB. If QTB's only successor is |
4257 | // QFB, then QFB is a better PostBB. |
4258 | if (QTB->getSingleSuccessor() == QFB) |
4259 | PostBB = QFB; |
4260 | |
4261 | // If we couldn't find a good PostBB, stop. |
4262 | if (!PostBB) |
4263 | return false; |
4264 | |
4265 | bool InvertPCond = false, InvertQCond = false; |
4266 | // Canonicalize fallthroughs to the true branches. |
4267 | if (PFB == QBI->getParent()) { |
4268 | std::swap(a&: PFB, b&: PTB); |
4269 | InvertPCond = true; |
4270 | } |
4271 | if (QFB == PostBB) { |
4272 | std::swap(a&: QFB, b&: QTB); |
4273 | InvertQCond = true; |
4274 | } |
4275 | |
4276 | // From this point on we can assume PTB or QTB may be fallthroughs but PFB |
4277 | // and QFB may not. Model fallthroughs as a nullptr block. |
4278 | if (PTB == QBI->getParent()) |
4279 | PTB = nullptr; |
4280 | if (QTB == PostBB) |
4281 | QTB = nullptr; |
4282 | |
4283 | // Legality bailouts. We must have at least the non-fallthrough blocks and |
4284 | // the post-dominating block, and the non-fallthroughs must only have one |
4285 | // predecessor. |
4286 | auto HasOnePredAndOneSucc = [](BasicBlock *BB, BasicBlock *P, BasicBlock *S) { |
4287 | return BB->getSinglePredecessor() == P && BB->getSingleSuccessor() == S; |
4288 | }; |
4289 | if (!HasOnePredAndOneSucc(PFB, PBI->getParent(), QBI->getParent()) || |
4290 | !HasOnePredAndOneSucc(QFB, QBI->getParent(), PostBB)) |
4291 | return false; |
4292 | if ((PTB && !HasOnePredAndOneSucc(PTB, PBI->getParent(), QBI->getParent())) || |
4293 | (QTB && !HasOnePredAndOneSucc(QTB, QBI->getParent(), PostBB))) |
4294 | return false; |
4295 | if (!QBI->getParent()->hasNUses(N: 2)) |
4296 | return false; |
4297 | |
4298 | // OK, this is a sequence of two diamonds or triangles. |
4299 | // Check if there are stores in PTB or PFB that are repeated in QTB or QFB. |
4300 | SmallPtrSet<Value *, 4> PStoreAddresses, QStoreAddresses; |
4301 | for (auto *BB : {PTB, PFB}) { |
4302 | if (!BB) |
4303 | continue; |
4304 | for (auto &I : *BB) |
4305 | if (StoreInst *SI = dyn_cast<StoreInst>(Val: &I)) |
4306 | PStoreAddresses.insert(Ptr: SI->getPointerOperand()); |
4307 | } |
4308 | for (auto *BB : {QTB, QFB}) { |
4309 | if (!BB) |
4310 | continue; |
4311 | for (auto &I : *BB) |
4312 | if (StoreInst *SI = dyn_cast<StoreInst>(Val: &I)) |
4313 | QStoreAddresses.insert(Ptr: SI->getPointerOperand()); |
4314 | } |
4315 | |
4316 | set_intersect(S1&: PStoreAddresses, S2: QStoreAddresses); |
4317 | // set_intersect mutates PStoreAddresses in place. Rename it here to make it |
4318 | // clear what it contains. |
4319 | auto &CommonAddresses = PStoreAddresses; |
4320 | |
4321 | bool Changed = false; |
4322 | for (auto *Address : CommonAddresses) |
4323 | Changed |= |
4324 | mergeConditionalStoreToAddress(PTB, PFB, QTB, QFB, PostBB, Address, |
4325 | InvertPCond, InvertQCond, DTU, DL, TTI); |
4326 | return Changed; |
4327 | } |
4328 | |
4329 | /// If the previous block ended with a widenable branch, determine if reusing |
4330 | /// the target block is profitable and legal. This will have the effect of |
4331 | /// "widening" PBI, but doesn't require us to reason about hosting safety. |
4332 | static bool tryWidenCondBranchToCondBranch(BranchInst *PBI, BranchInst *BI, |
4333 | DomTreeUpdater *DTU) { |
4334 | // TODO: This can be generalized in two important ways: |
4335 | // 1) We can allow phi nodes in IfFalseBB and simply reuse all the input |
4336 | // values from the PBI edge. |
4337 | // 2) We can sink side effecting instructions into BI's fallthrough |
4338 | // successor provided they doesn't contribute to computation of |
4339 | // BI's condition. |
4340 | BasicBlock *IfTrueBB = PBI->getSuccessor(i: 0); |
4341 | BasicBlock *IfFalseBB = PBI->getSuccessor(i: 1); |
4342 | if (!isWidenableBranch(U: PBI) || IfTrueBB != BI->getParent() || |
4343 | !BI->getParent()->getSinglePredecessor()) |
4344 | return false; |
4345 | if (!IfFalseBB->phis().empty()) |
4346 | return false; // TODO |
4347 | // This helps avoid infinite loop with SimplifyCondBranchToCondBranch which |
4348 | // may undo the transform done here. |
4349 | // TODO: There might be a more fine-grained solution to this. |
4350 | if (!llvm::succ_empty(BB: IfFalseBB)) |
4351 | return false; |
4352 | // Use lambda to lazily compute expensive condition after cheap ones. |
4353 | auto NoSideEffects = [](BasicBlock &BB) { |
4354 | return llvm::none_of(Range&: BB, P: [](const Instruction &I) { |
4355 | return I.mayWriteToMemory() || I.mayHaveSideEffects(); |
4356 | }); |
4357 | }; |
4358 | if (BI->getSuccessor(i: 1) != IfFalseBB && // no inf looping |
4359 | BI->getSuccessor(i: 1)->getTerminatingDeoptimizeCall() && // profitability |
4360 | NoSideEffects(*BI->getParent())) { |
4361 | auto *OldSuccessor = BI->getSuccessor(i: 1); |
4362 | OldSuccessor->removePredecessor(Pred: BI->getParent()); |
4363 | BI->setSuccessor(idx: 1, NewSucc: IfFalseBB); |
4364 | if (DTU) |
4365 | DTU->applyUpdates( |
4366 | Updates: {{DominatorTree::Insert, BI->getParent(), IfFalseBB}, |
4367 | {DominatorTree::Delete, BI->getParent(), OldSuccessor}}); |
4368 | return true; |
4369 | } |
4370 | if (BI->getSuccessor(i: 0) != IfFalseBB && // no inf looping |
4371 | BI->getSuccessor(i: 0)->getTerminatingDeoptimizeCall() && // profitability |
4372 | NoSideEffects(*BI->getParent())) { |
4373 | auto *OldSuccessor = BI->getSuccessor(i: 0); |
4374 | OldSuccessor->removePredecessor(Pred: BI->getParent()); |
4375 | BI->setSuccessor(idx: 0, NewSucc: IfFalseBB); |
4376 | if (DTU) |
4377 | DTU->applyUpdates( |
4378 | Updates: {{DominatorTree::Insert, BI->getParent(), IfFalseBB}, |
4379 | {DominatorTree::Delete, BI->getParent(), OldSuccessor}}); |
4380 | return true; |
4381 | } |
4382 | return false; |
4383 | } |
4384 | |
4385 | /// If we have a conditional branch as a predecessor of another block, |
4386 | /// this function tries to simplify it. We know |
4387 | /// that PBI and BI are both conditional branches, and BI is in one of the |
4388 | /// successor blocks of PBI - PBI branches to BI. |
4389 | static bool SimplifyCondBranchToCondBranch(BranchInst *PBI, BranchInst *BI, |
4390 | DomTreeUpdater *DTU, |
4391 | const DataLayout &DL, |
4392 | const TargetTransformInfo &TTI) { |
4393 | assert(PBI->isConditional() && BI->isConditional()); |
4394 | BasicBlock *BB = BI->getParent(); |
4395 | |
4396 | // If this block ends with a branch instruction, and if there is a |
4397 | // predecessor that ends on a branch of the same condition, make |
4398 | // this conditional branch redundant. |
4399 | if (PBI->getCondition() == BI->getCondition() && |
4400 | PBI->getSuccessor(i: 0) != PBI->getSuccessor(i: 1)) { |
4401 | // Okay, the outcome of this conditional branch is statically |
4402 | // knowable. If this block had a single pred, handle specially, otherwise |
4403 | // FoldCondBranchOnValueKnownInPredecessor() will handle it. |
4404 | if (BB->getSinglePredecessor()) { |
4405 | // Turn this into a branch on constant. |
4406 | bool CondIsTrue = PBI->getSuccessor(i: 0) == BB; |
4407 | BI->setCondition( |
4408 | ConstantInt::get(Ty: Type::getInt1Ty(C&: BB->getContext()), V: CondIsTrue)); |
4409 | return true; // Nuke the branch on constant. |
4410 | } |
4411 | } |
4412 | |
4413 | // If the previous block ended with a widenable branch, determine if reusing |
4414 | // the target block is profitable and legal. This will have the effect of |
4415 | // "widening" PBI, but doesn't require us to reason about hosting safety. |
4416 | if (tryWidenCondBranchToCondBranch(PBI, BI, DTU)) |
4417 | return true; |
4418 | |
4419 | // If both branches are conditional and both contain stores to the same |
4420 | // address, remove the stores from the conditionals and create a conditional |
4421 | // merged store at the end. |
4422 | if (MergeCondStores && mergeConditionalStores(PBI, QBI: BI, DTU, DL, TTI)) |
4423 | return true; |
4424 | |
4425 | // If this is a conditional branch in an empty block, and if any |
4426 | // predecessors are a conditional branch to one of our destinations, |
4427 | // fold the conditions into logical ops and one cond br. |
4428 | |
4429 | // Ignore dbg intrinsics. |
4430 | if (&*BB->instructionsWithoutDebug(SkipPseudoOp: false).begin() != BI) |
4431 | return false; |
4432 | |
4433 | int PBIOp, BIOp; |
4434 | if (PBI->getSuccessor(i: 0) == BI->getSuccessor(i: 0)) { |
4435 | PBIOp = 0; |
4436 | BIOp = 0; |
4437 | } else if (PBI->getSuccessor(i: 0) == BI->getSuccessor(i: 1)) { |
4438 | PBIOp = 0; |
4439 | BIOp = 1; |
4440 | } else if (PBI->getSuccessor(i: 1) == BI->getSuccessor(i: 0)) { |
4441 | PBIOp = 1; |
4442 | BIOp = 0; |
4443 | } else if (PBI->getSuccessor(i: 1) == BI->getSuccessor(i: 1)) { |
4444 | PBIOp = 1; |
4445 | BIOp = 1; |
4446 | } else { |
4447 | return false; |
4448 | } |
4449 | |
4450 | // Check to make sure that the other destination of this branch |
4451 | // isn't BB itself. If so, this is an infinite loop that will |
4452 | // keep getting unwound. |
4453 | if (PBI->getSuccessor(i: PBIOp) == BB) |
4454 | return false; |
4455 | |
4456 | // If predecessor's branch probability to BB is too low don't merge branches. |
4457 | SmallVector<uint32_t, 2> PredWeights; |
4458 | if (!PBI->getMetadata(KindID: LLVMContext::MD_unpredictable) && |
4459 | extractBranchWeights(I: *PBI, Weights&: PredWeights) && |
4460 | (static_cast<uint64_t>(PredWeights[0]) + PredWeights[1]) != 0) { |
4461 | |
4462 | BranchProbability CommonDestProb = BranchProbability::getBranchProbability( |
4463 | Numerator: PredWeights[PBIOp], |
4464 | Denominator: static_cast<uint64_t>(PredWeights[0]) + PredWeights[1]); |
4465 | |
4466 | BranchProbability Likely = TTI.getPredictableBranchThreshold(); |
4467 | if (CommonDestProb >= Likely) |
4468 | return false; |
4469 | } |
4470 | |
4471 | // Do not perform this transformation if it would require |
4472 | // insertion of a large number of select instructions. For targets |
4473 | // without predication/cmovs, this is a big pessimization. |
4474 | |
4475 | BasicBlock *CommonDest = PBI->getSuccessor(i: PBIOp); |
4476 | BasicBlock *RemovedDest = PBI->getSuccessor(i: PBIOp ^ 1); |
4477 | unsigned NumPhis = 0; |
4478 | for (BasicBlock::iterator II = CommonDest->begin(); isa<PHINode>(Val: II); |
4479 | ++II, ++NumPhis) { |
4480 | if (NumPhis > 2) // Disable this xform. |
4481 | return false; |
4482 | } |
4483 | |
4484 | // Finally, if everything is ok, fold the branches to logical ops. |
4485 | BasicBlock *OtherDest = BI->getSuccessor(i: BIOp ^ 1); |
4486 | |
4487 | LLVM_DEBUG(dbgs() << "FOLDING BRs:" << *PBI->getParent() |
4488 | << "AND: " << *BI->getParent()); |
4489 | |
4490 | SmallVector<DominatorTree::UpdateType, 5> Updates; |
4491 | |
4492 | // If OtherDest *is* BB, then BB is a basic block with a single conditional |
4493 | // branch in it, where one edge (OtherDest) goes back to itself but the other |
4494 | // exits. We don't *know* that the program avoids the infinite loop |
4495 | // (even though that seems likely). If we do this xform naively, we'll end up |
4496 | // recursively unpeeling the loop. Since we know that (after the xform is |
4497 | // done) that the block *is* infinite if reached, we just make it an obviously |
4498 | // infinite loop with no cond branch. |
4499 | if (OtherDest == BB) { |
4500 | // Insert it at the end of the function, because it's either code, |
4501 | // or it won't matter if it's hot. :) |
4502 | BasicBlock *InfLoopBlock = |
4503 | BasicBlock::Create(Context&: BB->getContext(), Name: "infloop" , Parent: BB->getParent()); |
4504 | BranchInst::Create(IfTrue: InfLoopBlock, InsertBefore: InfLoopBlock); |
4505 | if (DTU) |
4506 | Updates.push_back(Elt: {DominatorTree::Insert, InfLoopBlock, InfLoopBlock}); |
4507 | OtherDest = InfLoopBlock; |
4508 | } |
4509 | |
4510 | LLVM_DEBUG(dbgs() << *PBI->getParent()->getParent()); |
4511 | |
4512 | // BI may have other predecessors. Because of this, we leave |
4513 | // it alone, but modify PBI. |
4514 | |
4515 | // Make sure we get to CommonDest on True&True directions. |
4516 | Value *PBICond = PBI->getCondition(); |
4517 | IRBuilder<NoFolder> Builder(PBI); |
4518 | if (PBIOp) |
4519 | PBICond = Builder.CreateNot(V: PBICond, Name: PBICond->getName() + ".not" ); |
4520 | |
4521 | Value *BICond = BI->getCondition(); |
4522 | if (BIOp) |
4523 | BICond = Builder.CreateNot(V: BICond, Name: BICond->getName() + ".not" ); |
4524 | |
4525 | // Merge the conditions. |
4526 | Value *Cond = |
4527 | createLogicalOp(Builder, Opc: Instruction::Or, LHS: PBICond, RHS: BICond, Name: "brmerge" ); |
4528 | |
4529 | // Modify PBI to branch on the new condition to the new dests. |
4530 | PBI->setCondition(Cond); |
4531 | PBI->setSuccessor(idx: 0, NewSucc: CommonDest); |
4532 | PBI->setSuccessor(idx: 1, NewSucc: OtherDest); |
4533 | |
4534 | if (DTU) { |
4535 | Updates.push_back(Elt: {DominatorTree::Insert, PBI->getParent(), OtherDest}); |
4536 | Updates.push_back(Elt: {DominatorTree::Delete, PBI->getParent(), RemovedDest}); |
4537 | |
4538 | DTU->applyUpdates(Updates); |
4539 | } |
4540 | |
4541 | // Update branch weight for PBI. |
4542 | uint64_t PredTrueWeight, PredFalseWeight, SuccTrueWeight, SuccFalseWeight; |
4543 | uint64_t PredCommon, PredOther, SuccCommon, SuccOther; |
4544 | bool HasWeights = |
4545 | extractPredSuccWeights(PBI, BI, PredTrueWeight, PredFalseWeight, |
4546 | SuccTrueWeight, SuccFalseWeight); |
4547 | if (HasWeights) { |
4548 | PredCommon = PBIOp ? PredFalseWeight : PredTrueWeight; |
4549 | PredOther = PBIOp ? PredTrueWeight : PredFalseWeight; |
4550 | SuccCommon = BIOp ? SuccFalseWeight : SuccTrueWeight; |
4551 | SuccOther = BIOp ? SuccTrueWeight : SuccFalseWeight; |
4552 | // The weight to CommonDest should be PredCommon * SuccTotal + |
4553 | // PredOther * SuccCommon. |
4554 | // The weight to OtherDest should be PredOther * SuccOther. |
4555 | uint64_t NewWeights[2] = {PredCommon * (SuccCommon + SuccOther) + |
4556 | PredOther * SuccCommon, |
4557 | PredOther * SuccOther}; |
4558 | // Halve the weights if any of them cannot fit in an uint32_t |
4559 | FitWeights(Weights: NewWeights); |
4560 | |
4561 | setBranchWeights(I: PBI, TrueWeight: NewWeights[0], FalseWeight: NewWeights[1], /*IsExpected=*/false); |
4562 | } |
4563 | |
4564 | // OtherDest may have phi nodes. If so, add an entry from PBI's |
4565 | // block that are identical to the entries for BI's block. |
4566 | AddPredecessorToBlock(Succ: OtherDest, NewPred: PBI->getParent(), ExistPred: BB); |
4567 | |
4568 | // We know that the CommonDest already had an edge from PBI to |
4569 | // it. If it has PHIs though, the PHIs may have different |
4570 | // entries for BB and PBI's BB. If so, insert a select to make |
4571 | // them agree. |
4572 | for (PHINode &PN : CommonDest->phis()) { |
4573 | Value *BIV = PN.getIncomingValueForBlock(BB); |
4574 | unsigned PBBIdx = PN.getBasicBlockIndex(BB: PBI->getParent()); |
4575 | Value *PBIV = PN.getIncomingValue(i: PBBIdx); |
4576 | if (BIV != PBIV) { |
4577 | // Insert a select in PBI to pick the right value. |
4578 | SelectInst *NV = cast<SelectInst>( |
4579 | Val: Builder.CreateSelect(C: PBICond, True: PBIV, False: BIV, Name: PBIV->getName() + ".mux" )); |
4580 | PN.setIncomingValue(i: PBBIdx, V: NV); |
4581 | // Although the select has the same condition as PBI, the original branch |
4582 | // weights for PBI do not apply to the new select because the select's |
4583 | // 'logical' edges are incoming edges of the phi that is eliminated, not |
4584 | // the outgoing edges of PBI. |
4585 | if (HasWeights) { |
4586 | uint64_t PredCommon = PBIOp ? PredFalseWeight : PredTrueWeight; |
4587 | uint64_t PredOther = PBIOp ? PredTrueWeight : PredFalseWeight; |
4588 | uint64_t SuccCommon = BIOp ? SuccFalseWeight : SuccTrueWeight; |
4589 | uint64_t SuccOther = BIOp ? SuccTrueWeight : SuccFalseWeight; |
4590 | // The weight to PredCommonDest should be PredCommon * SuccTotal. |
4591 | // The weight to PredOtherDest should be PredOther * SuccCommon. |
4592 | uint64_t NewWeights[2] = {PredCommon * (SuccCommon + SuccOther), |
4593 | PredOther * SuccCommon}; |
4594 | |
4595 | FitWeights(Weights: NewWeights); |
4596 | |
4597 | setBranchWeights(I: NV, TrueWeight: NewWeights[0], FalseWeight: NewWeights[1], |
4598 | /*IsExpected=*/false); |
4599 | } |
4600 | } |
4601 | } |
4602 | |
4603 | LLVM_DEBUG(dbgs() << "INTO: " << *PBI->getParent()); |
4604 | LLVM_DEBUG(dbgs() << *PBI->getParent()->getParent()); |
4605 | |
4606 | // This basic block is probably dead. We know it has at least |
4607 | // one fewer predecessor. |
4608 | return true; |
4609 | } |
4610 | |
4611 | // Simplifies a terminator by replacing it with a branch to TrueBB if Cond is |
4612 | // true or to FalseBB if Cond is false. |
4613 | // Takes care of updating the successors and removing the old terminator. |
4614 | // Also makes sure not to introduce new successors by assuming that edges to |
4615 | // non-successor TrueBBs and FalseBBs aren't reachable. |
4616 | bool SimplifyCFGOpt::SimplifyTerminatorOnSelect(Instruction *OldTerm, |
4617 | Value *Cond, BasicBlock *TrueBB, |
4618 | BasicBlock *FalseBB, |
4619 | uint32_t TrueWeight, |
4620 | uint32_t FalseWeight) { |
4621 | auto *BB = OldTerm->getParent(); |
4622 | // Remove any superfluous successor edges from the CFG. |
4623 | // First, figure out which successors to preserve. |
4624 | // If TrueBB and FalseBB are equal, only try to preserve one copy of that |
4625 | // successor. |
4626 | BasicBlock *KeepEdge1 = TrueBB; |
4627 | BasicBlock *KeepEdge2 = TrueBB != FalseBB ? FalseBB : nullptr; |
4628 | |
4629 | SmallSetVector<BasicBlock *, 2> RemovedSuccessors; |
4630 | |
4631 | // Then remove the rest. |
4632 | for (BasicBlock *Succ : successors(I: OldTerm)) { |
4633 | // Make sure only to keep exactly one copy of each edge. |
4634 | if (Succ == KeepEdge1) |
4635 | KeepEdge1 = nullptr; |
4636 | else if (Succ == KeepEdge2) |
4637 | KeepEdge2 = nullptr; |
4638 | else { |
4639 | Succ->removePredecessor(Pred: BB, |
4640 | /*KeepOneInputPHIs=*/true); |
4641 | |
4642 | if (Succ != TrueBB && Succ != FalseBB) |
4643 | RemovedSuccessors.insert(X: Succ); |
4644 | } |
4645 | } |
4646 | |
4647 | IRBuilder<> Builder(OldTerm); |
4648 | Builder.SetCurrentDebugLocation(OldTerm->getDebugLoc()); |
4649 | |
4650 | // Insert an appropriate new terminator. |
4651 | if (!KeepEdge1 && !KeepEdge2) { |
4652 | if (TrueBB == FalseBB) { |
4653 | // We were only looking for one successor, and it was present. |
4654 | // Create an unconditional branch to it. |
4655 | Builder.CreateBr(Dest: TrueBB); |
4656 | } else { |
4657 | // We found both of the successors we were looking for. |
4658 | // Create a conditional branch sharing the condition of the select. |
4659 | BranchInst *NewBI = Builder.CreateCondBr(Cond, True: TrueBB, False: FalseBB); |
4660 | if (TrueWeight != FalseWeight) |
4661 | setBranchWeights(I: NewBI, TrueWeight, FalseWeight, /*IsExpected=*/false); |
4662 | } |
4663 | } else if (KeepEdge1 && (KeepEdge2 || TrueBB == FalseBB)) { |
4664 | // Neither of the selected blocks were successors, so this |
4665 | // terminator must be unreachable. |
4666 | new UnreachableInst(OldTerm->getContext(), OldTerm->getIterator()); |
4667 | } else { |
4668 | // One of the selected values was a successor, but the other wasn't. |
4669 | // Insert an unconditional branch to the one that was found; |
4670 | // the edge to the one that wasn't must be unreachable. |
4671 | if (!KeepEdge1) { |
4672 | // Only TrueBB was found. |
4673 | Builder.CreateBr(Dest: TrueBB); |
4674 | } else { |
4675 | // Only FalseBB was found. |
4676 | Builder.CreateBr(Dest: FalseBB); |
4677 | } |
4678 | } |
4679 | |
4680 | EraseTerminatorAndDCECond(TI: OldTerm); |
4681 | |
4682 | if (DTU) { |
4683 | SmallVector<DominatorTree::UpdateType, 2> Updates; |
4684 | Updates.reserve(N: RemovedSuccessors.size()); |
4685 | for (auto *RemovedSuccessor : RemovedSuccessors) |
4686 | Updates.push_back(Elt: {DominatorTree::Delete, BB, RemovedSuccessor}); |
4687 | DTU->applyUpdates(Updates); |
4688 | } |
4689 | |
4690 | return true; |
4691 | } |
4692 | |
4693 | // Replaces |
4694 | // (switch (select cond, X, Y)) on constant X, Y |
4695 | // with a branch - conditional if X and Y lead to distinct BBs, |
4696 | // unconditional otherwise. |
4697 | bool SimplifyCFGOpt::SimplifySwitchOnSelect(SwitchInst *SI, |
4698 | SelectInst *Select) { |
4699 | // Check for constant integer values in the select. |
4700 | ConstantInt *TrueVal = dyn_cast<ConstantInt>(Val: Select->getTrueValue()); |
4701 | ConstantInt *FalseVal = dyn_cast<ConstantInt>(Val: Select->getFalseValue()); |
4702 | if (!TrueVal || !FalseVal) |
4703 | return false; |
4704 | |
4705 | // Find the relevant condition and destinations. |
4706 | Value *Condition = Select->getCondition(); |
4707 | BasicBlock *TrueBB = SI->findCaseValue(C: TrueVal)->getCaseSuccessor(); |
4708 | BasicBlock *FalseBB = SI->findCaseValue(C: FalseVal)->getCaseSuccessor(); |
4709 | |
4710 | // Get weight for TrueBB and FalseBB. |
4711 | uint32_t TrueWeight = 0, FalseWeight = 0; |
4712 | SmallVector<uint64_t, 8> Weights; |
4713 | bool HasWeights = hasBranchWeightMD(I: *SI); |
4714 | if (HasWeights) { |
4715 | GetBranchWeights(TI: SI, Weights); |
4716 | if (Weights.size() == 1 + SI->getNumCases()) { |
4717 | TrueWeight = |
4718 | (uint32_t)Weights[SI->findCaseValue(C: TrueVal)->getSuccessorIndex()]; |
4719 | FalseWeight = |
4720 | (uint32_t)Weights[SI->findCaseValue(C: FalseVal)->getSuccessorIndex()]; |
4721 | } |
4722 | } |
4723 | |
4724 | // Perform the actual simplification. |
4725 | return SimplifyTerminatorOnSelect(OldTerm: SI, Cond: Condition, TrueBB, FalseBB, TrueWeight, |
4726 | FalseWeight); |
4727 | } |
4728 | |
4729 | // Replaces |
4730 | // (indirectbr (select cond, blockaddress(@fn, BlockA), |
4731 | // blockaddress(@fn, BlockB))) |
4732 | // with |
4733 | // (br cond, BlockA, BlockB). |
4734 | bool SimplifyCFGOpt::SimplifyIndirectBrOnSelect(IndirectBrInst *IBI, |
4735 | SelectInst *SI) { |
4736 | // Check that both operands of the select are block addresses. |
4737 | BlockAddress *TBA = dyn_cast<BlockAddress>(Val: SI->getTrueValue()); |
4738 | BlockAddress *FBA = dyn_cast<BlockAddress>(Val: SI->getFalseValue()); |
4739 | if (!TBA || !FBA) |
4740 | return false; |
4741 | |
4742 | // Extract the actual blocks. |
4743 | BasicBlock *TrueBB = TBA->getBasicBlock(); |
4744 | BasicBlock *FalseBB = FBA->getBasicBlock(); |
4745 | |
4746 | // Perform the actual simplification. |
4747 | return SimplifyTerminatorOnSelect(OldTerm: IBI, Cond: SI->getCondition(), TrueBB, FalseBB, TrueWeight: 0, |
4748 | FalseWeight: 0); |
4749 | } |
4750 | |
4751 | /// This is called when we find an icmp instruction |
4752 | /// (a seteq/setne with a constant) as the only instruction in a |
4753 | /// block that ends with an uncond branch. We are looking for a very specific |
4754 | /// pattern that occurs when "A == 1 || A == 2 || A == 3" gets simplified. In |
4755 | /// this case, we merge the first two "or's of icmp" into a switch, but then the |
4756 | /// default value goes to an uncond block with a seteq in it, we get something |
4757 | /// like: |
4758 | /// |
4759 | /// switch i8 %A, label %DEFAULT [ i8 1, label %end i8 2, label %end ] |
4760 | /// DEFAULT: |
4761 | /// %tmp = icmp eq i8 %A, 92 |
4762 | /// br label %end |
4763 | /// end: |
4764 | /// ... = phi i1 [ true, %entry ], [ %tmp, %DEFAULT ], [ true, %entry ] |
4765 | /// |
4766 | /// We prefer to split the edge to 'end' so that there is a true/false entry to |
4767 | /// the PHI, merging the third icmp into the switch. |
4768 | bool SimplifyCFGOpt::tryToSimplifyUncondBranchWithICmpInIt( |
4769 | ICmpInst *ICI, IRBuilder<> &Builder) { |
4770 | BasicBlock *BB = ICI->getParent(); |
4771 | |
4772 | // If the block has any PHIs in it or the icmp has multiple uses, it is too |
4773 | // complex. |
4774 | if (isa<PHINode>(Val: BB->begin()) || !ICI->hasOneUse()) |
4775 | return false; |
4776 | |
4777 | Value *V = ICI->getOperand(i_nocapture: 0); |
4778 | ConstantInt *Cst = cast<ConstantInt>(Val: ICI->getOperand(i_nocapture: 1)); |
4779 | |
4780 | // The pattern we're looking for is where our only predecessor is a switch on |
4781 | // 'V' and this block is the default case for the switch. In this case we can |
4782 | // fold the compared value into the switch to simplify things. |
4783 | BasicBlock *Pred = BB->getSinglePredecessor(); |
4784 | if (!Pred || !isa<SwitchInst>(Val: Pred->getTerminator())) |
4785 | return false; |
4786 | |
4787 | SwitchInst *SI = cast<SwitchInst>(Val: Pred->getTerminator()); |
4788 | if (SI->getCondition() != V) |
4789 | return false; |
4790 | |
4791 | // If BB is reachable on a non-default case, then we simply know the value of |
4792 | // V in this block. Substitute it and constant fold the icmp instruction |
4793 | // away. |
4794 | if (SI->getDefaultDest() != BB) { |
4795 | ConstantInt *VVal = SI->findCaseDest(BB); |
4796 | assert(VVal && "Should have a unique destination value" ); |
4797 | ICI->setOperand(i_nocapture: 0, Val_nocapture: VVal); |
4798 | |
4799 | if (Value *V = simplifyInstruction(I: ICI, Q: {DL, ICI})) { |
4800 | ICI->replaceAllUsesWith(V); |
4801 | ICI->eraseFromParent(); |
4802 | } |
4803 | // BB is now empty, so it is likely to simplify away. |
4804 | return requestResimplify(); |
4805 | } |
4806 | |
4807 | // Ok, the block is reachable from the default dest. If the constant we're |
4808 | // comparing exists in one of the other edges, then we can constant fold ICI |
4809 | // and zap it. |
4810 | if (SI->findCaseValue(C: Cst) != SI->case_default()) { |
4811 | Value *V; |
4812 | if (ICI->getPredicate() == ICmpInst::ICMP_EQ) |
4813 | V = ConstantInt::getFalse(Context&: BB->getContext()); |
4814 | else |
4815 | V = ConstantInt::getTrue(Context&: BB->getContext()); |
4816 | |
4817 | ICI->replaceAllUsesWith(V); |
4818 | ICI->eraseFromParent(); |
4819 | // BB is now empty, so it is likely to simplify away. |
4820 | return requestResimplify(); |
4821 | } |
4822 | |
4823 | // The use of the icmp has to be in the 'end' block, by the only PHI node in |
4824 | // the block. |
4825 | BasicBlock *SuccBlock = BB->getTerminator()->getSuccessor(Idx: 0); |
4826 | PHINode *PHIUse = dyn_cast<PHINode>(Val: ICI->user_back()); |
4827 | if (PHIUse == nullptr || PHIUse != &SuccBlock->front() || |
4828 | isa<PHINode>(Val: ++BasicBlock::iterator(PHIUse))) |
4829 | return false; |
4830 | |
4831 | // If the icmp is a SETEQ, then the default dest gets false, the new edge gets |
4832 | // true in the PHI. |
4833 | Constant *DefaultCst = ConstantInt::getTrue(Context&: BB->getContext()); |
4834 | Constant *NewCst = ConstantInt::getFalse(Context&: BB->getContext()); |
4835 | |
4836 | if (ICI->getPredicate() == ICmpInst::ICMP_EQ) |
4837 | std::swap(a&: DefaultCst, b&: NewCst); |
4838 | |
4839 | // Replace ICI (which is used by the PHI for the default value) with true or |
4840 | // false depending on if it is EQ or NE. |
4841 | ICI->replaceAllUsesWith(V: DefaultCst); |
4842 | ICI->eraseFromParent(); |
4843 | |
4844 | SmallVector<DominatorTree::UpdateType, 2> Updates; |
4845 | |
4846 | // Okay, the switch goes to this block on a default value. Add an edge from |
4847 | // the switch to the merge point on the compared value. |
4848 | BasicBlock *NewBB = |
4849 | BasicBlock::Create(Context&: BB->getContext(), Name: "switch.edge" , Parent: BB->getParent(), InsertBefore: BB); |
4850 | { |
4851 | SwitchInstProfUpdateWrapper SIW(*SI); |
4852 | auto W0 = SIW.getSuccessorWeight(idx: 0); |
4853 | SwitchInstProfUpdateWrapper::CaseWeightOpt NewW; |
4854 | if (W0) { |
4855 | NewW = ((uint64_t(*W0) + 1) >> 1); |
4856 | SIW.setSuccessorWeight(idx: 0, W: *NewW); |
4857 | } |
4858 | SIW.addCase(OnVal: Cst, Dest: NewBB, W: NewW); |
4859 | if (DTU) |
4860 | Updates.push_back(Elt: {DominatorTree::Insert, Pred, NewBB}); |
4861 | } |
4862 | |
4863 | // NewBB branches to the phi block, add the uncond branch and the phi entry. |
4864 | Builder.SetInsertPoint(NewBB); |
4865 | Builder.SetCurrentDebugLocation(SI->getDebugLoc()); |
4866 | Builder.CreateBr(Dest: SuccBlock); |
4867 | PHIUse->addIncoming(V: NewCst, BB: NewBB); |
4868 | if (DTU) { |
4869 | Updates.push_back(Elt: {DominatorTree::Insert, NewBB, SuccBlock}); |
4870 | DTU->applyUpdates(Updates); |
4871 | } |
4872 | return true; |
4873 | } |
4874 | |
4875 | /// The specified branch is a conditional branch. |
4876 | /// Check to see if it is branching on an or/and chain of icmp instructions, and |
4877 | /// fold it into a switch instruction if so. |
4878 | bool SimplifyCFGOpt::SimplifyBranchOnICmpChain(BranchInst *BI, |
4879 | IRBuilder<> &Builder, |
4880 | const DataLayout &DL) { |
4881 | Instruction *Cond = dyn_cast<Instruction>(Val: BI->getCondition()); |
4882 | if (!Cond) |
4883 | return false; |
4884 | |
4885 | // Change br (X == 0 | X == 1), T, F into a switch instruction. |
4886 | // If this is a bunch of seteq's or'd together, or if it's a bunch of |
4887 | // 'setne's and'ed together, collect them. |
4888 | |
4889 | // Try to gather values from a chain of and/or to be turned into a switch |
4890 | ConstantComparesGatherer ConstantCompare(Cond, DL); |
4891 | // Unpack the result |
4892 | SmallVectorImpl<ConstantInt *> &Values = ConstantCompare.Vals; |
4893 | Value *CompVal = ConstantCompare.CompValue; |
4894 | unsigned UsedICmps = ConstantCompare.UsedICmps; |
4895 | Value * = ConstantCompare.Extra; |
4896 | |
4897 | // If we didn't have a multiply compared value, fail. |
4898 | if (!CompVal) |
4899 | return false; |
4900 | |
4901 | // Avoid turning single icmps into a switch. |
4902 | if (UsedICmps <= 1) |
4903 | return false; |
4904 | |
4905 | bool TrueWhenEqual = match(V: Cond, P: m_LogicalOr(L: m_Value(), R: m_Value())); |
4906 | |
4907 | // There might be duplicate constants in the list, which the switch |
4908 | // instruction can't handle, remove them now. |
4909 | array_pod_sort(Start: Values.begin(), End: Values.end(), Compare: ConstantIntSortPredicate); |
4910 | Values.erase(CS: llvm::unique(R&: Values), CE: Values.end()); |
4911 | |
4912 | // If Extra was used, we require at least two switch values to do the |
4913 | // transformation. A switch with one value is just a conditional branch. |
4914 | if (ExtraCase && Values.size() < 2) |
4915 | return false; |
4916 | |
4917 | // TODO: Preserve branch weight metadata, similarly to how |
4918 | // FoldValueComparisonIntoPredecessors preserves it. |
4919 | |
4920 | // Figure out which block is which destination. |
4921 | BasicBlock *DefaultBB = BI->getSuccessor(i: 1); |
4922 | BasicBlock *EdgeBB = BI->getSuccessor(i: 0); |
4923 | if (!TrueWhenEqual) |
4924 | std::swap(a&: DefaultBB, b&: EdgeBB); |
4925 | |
4926 | BasicBlock *BB = BI->getParent(); |
4927 | |
4928 | LLVM_DEBUG(dbgs() << "Converting 'icmp' chain with " << Values.size() |
4929 | << " cases into SWITCH. BB is:\n" |
4930 | << *BB); |
4931 | |
4932 | SmallVector<DominatorTree::UpdateType, 2> Updates; |
4933 | |
4934 | // If there are any extra values that couldn't be folded into the switch |
4935 | // then we evaluate them with an explicit branch first. Split the block |
4936 | // right before the condbr to handle it. |
4937 | if (ExtraCase) { |
4938 | BasicBlock *NewBB = SplitBlock(Old: BB, SplitPt: BI, DTU, /*LI=*/nullptr, |
4939 | /*MSSAU=*/nullptr, BBName: "switch.early.test" ); |
4940 | |
4941 | // Remove the uncond branch added to the old block. |
4942 | Instruction *OldTI = BB->getTerminator(); |
4943 | Builder.SetInsertPoint(OldTI); |
4944 | |
4945 | // There can be an unintended UB if extra values are Poison. Before the |
4946 | // transformation, extra values may not be evaluated according to the |
4947 | // condition, and it will not raise UB. But after transformation, we are |
4948 | // evaluating extra values before checking the condition, and it will raise |
4949 | // UB. It can be solved by adding freeze instruction to extra values. |
4950 | AssumptionCache *AC = Options.AC; |
4951 | |
4952 | if (!isGuaranteedNotToBeUndefOrPoison(V: ExtraCase, AC, CtxI: BI, DT: nullptr)) |
4953 | ExtraCase = Builder.CreateFreeze(V: ExtraCase); |
4954 | |
4955 | if (TrueWhenEqual) |
4956 | Builder.CreateCondBr(Cond: ExtraCase, True: EdgeBB, False: NewBB); |
4957 | else |
4958 | Builder.CreateCondBr(Cond: ExtraCase, True: NewBB, False: EdgeBB); |
4959 | |
4960 | OldTI->eraseFromParent(); |
4961 | |
4962 | if (DTU) |
4963 | Updates.push_back(Elt: {DominatorTree::Insert, BB, EdgeBB}); |
4964 | |
4965 | // If there are PHI nodes in EdgeBB, then we need to add a new entry to them |
4966 | // for the edge we just added. |
4967 | AddPredecessorToBlock(Succ: EdgeBB, NewPred: BB, ExistPred: NewBB); |
4968 | |
4969 | LLVM_DEBUG(dbgs() << " ** 'icmp' chain unhandled condition: " << *ExtraCase |
4970 | << "\nEXTRABB = " << *BB); |
4971 | BB = NewBB; |
4972 | } |
4973 | |
4974 | Builder.SetInsertPoint(BI); |
4975 | // Convert pointer to int before we switch. |
4976 | if (CompVal->getType()->isPointerTy()) { |
4977 | CompVal = Builder.CreatePtrToInt( |
4978 | V: CompVal, DestTy: DL.getIntPtrType(CompVal->getType()), Name: "magicptr" ); |
4979 | } |
4980 | |
4981 | // Create the new switch instruction now. |
4982 | SwitchInst *New = Builder.CreateSwitch(V: CompVal, Dest: DefaultBB, NumCases: Values.size()); |
4983 | |
4984 | // Add all of the 'cases' to the switch instruction. |
4985 | for (unsigned i = 0, e = Values.size(); i != e; ++i) |
4986 | New->addCase(OnVal: Values[i], Dest: EdgeBB); |
4987 | |
4988 | // We added edges from PI to the EdgeBB. As such, if there were any |
4989 | // PHI nodes in EdgeBB, they need entries to be added corresponding to |
4990 | // the number of edges added. |
4991 | for (BasicBlock::iterator BBI = EdgeBB->begin(); isa<PHINode>(Val: BBI); ++BBI) { |
4992 | PHINode *PN = cast<PHINode>(Val&: BBI); |
4993 | Value *InVal = PN->getIncomingValueForBlock(BB); |
4994 | for (unsigned i = 0, e = Values.size() - 1; i != e; ++i) |
4995 | PN->addIncoming(V: InVal, BB); |
4996 | } |
4997 | |
4998 | // Erase the old branch instruction. |
4999 | EraseTerminatorAndDCECond(TI: BI); |
5000 | if (DTU) |
5001 | DTU->applyUpdates(Updates); |
5002 | |
5003 | LLVM_DEBUG(dbgs() << " ** 'icmp' chain result is:\n" << *BB << '\n'); |
5004 | return true; |
5005 | } |
5006 | |
5007 | bool SimplifyCFGOpt::simplifyResume(ResumeInst *RI, IRBuilder<> &Builder) { |
5008 | if (isa<PHINode>(Val: RI->getValue())) |
5009 | return simplifyCommonResume(RI); |
5010 | else if (isa<LandingPadInst>(Val: RI->getParent()->getFirstNonPHI()) && |
5011 | RI->getValue() == RI->getParent()->getFirstNonPHI()) |
5012 | // The resume must unwind the exception that caused control to branch here. |
5013 | return simplifySingleResume(RI); |
5014 | |
5015 | return false; |
5016 | } |
5017 | |
5018 | // Check if cleanup block is empty |
5019 | static bool isCleanupBlockEmpty(iterator_range<BasicBlock::iterator> R) { |
5020 | for (Instruction &I : R) { |
5021 | auto *II = dyn_cast<IntrinsicInst>(Val: &I); |
5022 | if (!II) |
5023 | return false; |
5024 | |
5025 | Intrinsic::ID IntrinsicID = II->getIntrinsicID(); |
5026 | switch (IntrinsicID) { |
5027 | case Intrinsic::dbg_declare: |
5028 | case Intrinsic::dbg_value: |
5029 | case Intrinsic::dbg_label: |
5030 | case Intrinsic::lifetime_end: |
5031 | break; |
5032 | default: |
5033 | return false; |
5034 | } |
5035 | } |
5036 | return true; |
5037 | } |
5038 | |
5039 | // Simplify resume that is shared by several landing pads (phi of landing pad). |
5040 | bool SimplifyCFGOpt::simplifyCommonResume(ResumeInst *RI) { |
5041 | BasicBlock *BB = RI->getParent(); |
5042 | |
5043 | // Check that there are no other instructions except for debug and lifetime |
5044 | // intrinsics between the phi's and resume instruction. |
5045 | if (!isCleanupBlockEmpty( |
5046 | R: make_range(x: RI->getParent()->getFirstNonPHI(), y: BB->getTerminator()))) |
5047 | return false; |
5048 | |
5049 | SmallSetVector<BasicBlock *, 4> TrivialUnwindBlocks; |
5050 | auto *PhiLPInst = cast<PHINode>(Val: RI->getValue()); |
5051 | |
5052 | // Check incoming blocks to see if any of them are trivial. |
5053 | for (unsigned Idx = 0, End = PhiLPInst->getNumIncomingValues(); Idx != End; |
5054 | Idx++) { |
5055 | auto *IncomingBB = PhiLPInst->getIncomingBlock(i: Idx); |
5056 | auto *IncomingValue = PhiLPInst->getIncomingValue(i: Idx); |
5057 | |
5058 | // If the block has other successors, we can not delete it because |
5059 | // it has other dependents. |
5060 | if (IncomingBB->getUniqueSuccessor() != BB) |
5061 | continue; |
5062 | |
5063 | auto *LandingPad = dyn_cast<LandingPadInst>(Val: IncomingBB->getFirstNonPHI()); |
5064 | // Not the landing pad that caused the control to branch here. |
5065 | if (IncomingValue != LandingPad) |
5066 | continue; |
5067 | |
5068 | if (isCleanupBlockEmpty( |
5069 | R: make_range(x: LandingPad->getNextNode(), y: IncomingBB->getTerminator()))) |
5070 | TrivialUnwindBlocks.insert(X: IncomingBB); |
5071 | } |
5072 | |
5073 | // If no trivial unwind blocks, don't do any simplifications. |
5074 | if (TrivialUnwindBlocks.empty()) |
5075 | return false; |
5076 | |
5077 | // Turn all invokes that unwind here into calls. |
5078 | for (auto *TrivialBB : TrivialUnwindBlocks) { |
5079 | // Blocks that will be simplified should be removed from the phi node. |
5080 | // Note there could be multiple edges to the resume block, and we need |
5081 | // to remove them all. |
5082 | while (PhiLPInst->getBasicBlockIndex(BB: TrivialBB) != -1) |
5083 | BB->removePredecessor(Pred: TrivialBB, KeepOneInputPHIs: true); |
5084 | |
5085 | for (BasicBlock *Pred : |
5086 | llvm::make_early_inc_range(Range: predecessors(BB: TrivialBB))) { |
5087 | removeUnwindEdge(BB: Pred, DTU); |
5088 | ++NumInvokes; |
5089 | } |
5090 | |
5091 | // In each SimplifyCFG run, only the current processed block can be erased. |
5092 | // Otherwise, it will break the iteration of SimplifyCFG pass. So instead |
5093 | // of erasing TrivialBB, we only remove the branch to the common resume |
5094 | // block so that we can later erase the resume block since it has no |
5095 | // predecessors. |
5096 | TrivialBB->getTerminator()->eraseFromParent(); |
5097 | new UnreachableInst(RI->getContext(), TrivialBB); |
5098 | if (DTU) |
5099 | DTU->applyUpdates(Updates: {{DominatorTree::Delete, TrivialBB, BB}}); |
5100 | } |
5101 | |
5102 | // Delete the resume block if all its predecessors have been removed. |
5103 | if (pred_empty(BB)) |
5104 | DeleteDeadBlock(BB, DTU); |
5105 | |
5106 | return !TrivialUnwindBlocks.empty(); |
5107 | } |
5108 | |
5109 | // Simplify resume that is only used by a single (non-phi) landing pad. |
5110 | bool SimplifyCFGOpt::simplifySingleResume(ResumeInst *RI) { |
5111 | BasicBlock *BB = RI->getParent(); |
5112 | auto *LPInst = cast<LandingPadInst>(Val: BB->getFirstNonPHI()); |
5113 | assert(RI->getValue() == LPInst && |
5114 | "Resume must unwind the exception that caused control to here" ); |
5115 | |
5116 | // Check that there are no other instructions except for debug intrinsics. |
5117 | if (!isCleanupBlockEmpty( |
5118 | R: make_range<Instruction *>(x: LPInst->getNextNode(), y: RI))) |
5119 | return false; |
5120 | |
5121 | // Turn all invokes that unwind here into calls and delete the basic block. |
5122 | for (BasicBlock *Pred : llvm::make_early_inc_range(Range: predecessors(BB))) { |
5123 | removeUnwindEdge(BB: Pred, DTU); |
5124 | ++NumInvokes; |
5125 | } |
5126 | |
5127 | // The landingpad is now unreachable. Zap it. |
5128 | DeleteDeadBlock(BB, DTU); |
5129 | return true; |
5130 | } |
5131 | |
5132 | static bool removeEmptyCleanup(CleanupReturnInst *RI, DomTreeUpdater *DTU) { |
5133 | // If this is a trivial cleanup pad that executes no instructions, it can be |
5134 | // eliminated. If the cleanup pad continues to the caller, any predecessor |
5135 | // that is an EH pad will be updated to continue to the caller and any |
5136 | // predecessor that terminates with an invoke instruction will have its invoke |
5137 | // instruction converted to a call instruction. If the cleanup pad being |
5138 | // simplified does not continue to the caller, each predecessor will be |
5139 | // updated to continue to the unwind destination of the cleanup pad being |
5140 | // simplified. |
5141 | BasicBlock *BB = RI->getParent(); |
5142 | CleanupPadInst *CPInst = RI->getCleanupPad(); |
5143 | if (CPInst->getParent() != BB) |
5144 | // This isn't an empty cleanup. |
5145 | return false; |
5146 | |
5147 | // We cannot kill the pad if it has multiple uses. This typically arises |
5148 | // from unreachable basic blocks. |
5149 | if (!CPInst->hasOneUse()) |
5150 | return false; |
5151 | |
5152 | // Check that there are no other instructions except for benign intrinsics. |
5153 | if (!isCleanupBlockEmpty( |
5154 | R: make_range<Instruction *>(x: CPInst->getNextNode(), y: RI))) |
5155 | return false; |
5156 | |
5157 | // If the cleanup return we are simplifying unwinds to the caller, this will |
5158 | // set UnwindDest to nullptr. |
5159 | BasicBlock *UnwindDest = RI->getUnwindDest(); |
5160 | Instruction *DestEHPad = UnwindDest ? UnwindDest->getFirstNonPHI() : nullptr; |
5161 | |
5162 | // We're about to remove BB from the control flow. Before we do, sink any |
5163 | // PHINodes into the unwind destination. Doing this before changing the |
5164 | // control flow avoids some potentially slow checks, since we can currently |
5165 | // be certain that UnwindDest and BB have no common predecessors (since they |
5166 | // are both EH pads). |
5167 | if (UnwindDest) { |
5168 | // First, go through the PHI nodes in UnwindDest and update any nodes that |
5169 | // reference the block we are removing |
5170 | for (PHINode &DestPN : UnwindDest->phis()) { |
5171 | int Idx = DestPN.getBasicBlockIndex(BB); |
5172 | // Since BB unwinds to UnwindDest, it has to be in the PHI node. |
5173 | assert(Idx != -1); |
5174 | // This PHI node has an incoming value that corresponds to a control |
5175 | // path through the cleanup pad we are removing. If the incoming |
5176 | // value is in the cleanup pad, it must be a PHINode (because we |
5177 | // verified above that the block is otherwise empty). Otherwise, the |
5178 | // value is either a constant or a value that dominates the cleanup |
5179 | // pad being removed. |
5180 | // |
5181 | // Because BB and UnwindDest are both EH pads, all of their |
5182 | // predecessors must unwind to these blocks, and since no instruction |
5183 | // can have multiple unwind destinations, there will be no overlap in |
5184 | // incoming blocks between SrcPN and DestPN. |
5185 | Value *SrcVal = DestPN.getIncomingValue(i: Idx); |
5186 | PHINode *SrcPN = dyn_cast<PHINode>(Val: SrcVal); |
5187 | |
5188 | bool NeedPHITranslation = SrcPN && SrcPN->getParent() == BB; |
5189 | for (auto *Pred : predecessors(BB)) { |
5190 | Value *Incoming = |
5191 | NeedPHITranslation ? SrcPN->getIncomingValueForBlock(BB: Pred) : SrcVal; |
5192 | DestPN.addIncoming(V: Incoming, BB: Pred); |
5193 | } |
5194 | } |
5195 | |
5196 | // Sink any remaining PHI nodes directly into UnwindDest. |
5197 | Instruction *InsertPt = DestEHPad; |
5198 | for (PHINode &PN : make_early_inc_range(Range: BB->phis())) { |
5199 | if (PN.use_empty() || !PN.isUsedOutsideOfBlock(BB)) |
5200 | // If the PHI node has no uses or all of its uses are in this basic |
5201 | // block (meaning they are debug or lifetime intrinsics), just leave |
5202 | // it. It will be erased when we erase BB below. |
5203 | continue; |
5204 | |
5205 | // Otherwise, sink this PHI node into UnwindDest. |
5206 | // Any predecessors to UnwindDest which are not already represented |
5207 | // must be back edges which inherit the value from the path through |
5208 | // BB. In this case, the PHI value must reference itself. |
5209 | for (auto *pred : predecessors(BB: UnwindDest)) |
5210 | if (pred != BB) |
5211 | PN.addIncoming(V: &PN, BB: pred); |
5212 | PN.moveBefore(MovePos: InsertPt); |
5213 | // Also, add a dummy incoming value for the original BB itself, |
5214 | // so that the PHI is well-formed until we drop said predecessor. |
5215 | PN.addIncoming(V: PoisonValue::get(T: PN.getType()), BB); |
5216 | } |
5217 | } |
5218 | |
5219 | std::vector<DominatorTree::UpdateType> Updates; |
5220 | |
5221 | // We use make_early_inc_range here because we will remove all predecessors. |
5222 | for (BasicBlock *PredBB : llvm::make_early_inc_range(Range: predecessors(BB))) { |
5223 | if (UnwindDest == nullptr) { |
5224 | if (DTU) { |
5225 | DTU->applyUpdates(Updates); |
5226 | Updates.clear(); |
5227 | } |
5228 | removeUnwindEdge(BB: PredBB, DTU); |
5229 | ++NumInvokes; |
5230 | } else { |
5231 | BB->removePredecessor(Pred: PredBB); |
5232 | Instruction *TI = PredBB->getTerminator(); |
5233 | TI->replaceUsesOfWith(From: BB, To: UnwindDest); |
5234 | if (DTU) { |
5235 | Updates.push_back(x: {DominatorTree::Insert, PredBB, UnwindDest}); |
5236 | Updates.push_back(x: {DominatorTree::Delete, PredBB, BB}); |
5237 | } |
5238 | } |
5239 | } |
5240 | |
5241 | if (DTU) |
5242 | DTU->applyUpdates(Updates); |
5243 | |
5244 | DeleteDeadBlock(BB, DTU); |
5245 | |
5246 | return true; |
5247 | } |
5248 | |
5249 | // Try to merge two cleanuppads together. |
5250 | static bool mergeCleanupPad(CleanupReturnInst *RI) { |
5251 | // Skip any cleanuprets which unwind to caller, there is nothing to merge |
5252 | // with. |
5253 | BasicBlock *UnwindDest = RI->getUnwindDest(); |
5254 | if (!UnwindDest) |
5255 | return false; |
5256 | |
5257 | // This cleanupret isn't the only predecessor of this cleanuppad, it wouldn't |
5258 | // be safe to merge without code duplication. |
5259 | if (UnwindDest->getSinglePredecessor() != RI->getParent()) |
5260 | return false; |
5261 | |
5262 | // Verify that our cleanuppad's unwind destination is another cleanuppad. |
5263 | auto *SuccessorCleanupPad = dyn_cast<CleanupPadInst>(Val: &UnwindDest->front()); |
5264 | if (!SuccessorCleanupPad) |
5265 | return false; |
5266 | |
5267 | CleanupPadInst *PredecessorCleanupPad = RI->getCleanupPad(); |
5268 | // Replace any uses of the successor cleanupad with the predecessor pad |
5269 | // The only cleanuppad uses should be this cleanupret, it's cleanupret and |
5270 | // funclet bundle operands. |
5271 | SuccessorCleanupPad->replaceAllUsesWith(V: PredecessorCleanupPad); |
5272 | // Remove the old cleanuppad. |
5273 | SuccessorCleanupPad->eraseFromParent(); |
5274 | // Now, we simply replace the cleanupret with a branch to the unwind |
5275 | // destination. |
5276 | BranchInst::Create(IfTrue: UnwindDest, InsertBefore: RI->getParent()); |
5277 | RI->eraseFromParent(); |
5278 | |
5279 | return true; |
5280 | } |
5281 | |
5282 | bool SimplifyCFGOpt::simplifyCleanupReturn(CleanupReturnInst *RI) { |
5283 | // It is possible to transiantly have an undef cleanuppad operand because we |
5284 | // have deleted some, but not all, dead blocks. |
5285 | // Eventually, this block will be deleted. |
5286 | if (isa<UndefValue>(Val: RI->getOperand(i_nocapture: 0))) |
5287 | return false; |
5288 | |
5289 | if (mergeCleanupPad(RI)) |
5290 | return true; |
5291 | |
5292 | if (removeEmptyCleanup(RI, DTU)) |
5293 | return true; |
5294 | |
5295 | return false; |
5296 | } |
5297 | |
5298 | // WARNING: keep in sync with InstCombinerImpl::visitUnreachableInst()! |
5299 | bool SimplifyCFGOpt::simplifyUnreachable(UnreachableInst *UI) { |
5300 | BasicBlock *BB = UI->getParent(); |
5301 | |
5302 | bool Changed = false; |
5303 | |
5304 | // Ensure that any debug-info records that used to occur after the Unreachable |
5305 | // are moved to in front of it -- otherwise they'll "dangle" at the end of |
5306 | // the block. |
5307 | BB->flushTerminatorDbgRecords(); |
5308 | |
5309 | // Debug-info records on the unreachable inst itself should be deleted, as |
5310 | // below we delete everything past the final executable instruction. |
5311 | UI->dropDbgRecords(); |
5312 | |
5313 | // If there are any instructions immediately before the unreachable that can |
5314 | // be removed, do so. |
5315 | while (UI->getIterator() != BB->begin()) { |
5316 | BasicBlock::iterator BBI = UI->getIterator(); |
5317 | --BBI; |
5318 | |
5319 | if (!isGuaranteedToTransferExecutionToSuccessor(I: &*BBI)) |
5320 | break; // Can not drop any more instructions. We're done here. |
5321 | // Otherwise, this instruction can be freely erased, |
5322 | // even if it is not side-effect free. |
5323 | |
5324 | // Note that deleting EH's here is in fact okay, although it involves a bit |
5325 | // of subtle reasoning. If this inst is an EH, all the predecessors of this |
5326 | // block will be the unwind edges of Invoke/CatchSwitch/CleanupReturn, |
5327 | // and we can therefore guarantee this block will be erased. |
5328 | |
5329 | // If we're deleting this, we're deleting any subsequent debug info, so |
5330 | // delete DbgRecords. |
5331 | BBI->dropDbgRecords(); |
5332 | |
5333 | // Delete this instruction (any uses are guaranteed to be dead) |
5334 | BBI->replaceAllUsesWith(V: PoisonValue::get(T: BBI->getType())); |
5335 | BBI->eraseFromParent(); |
5336 | Changed = true; |
5337 | } |
5338 | |
5339 | // If the unreachable instruction is the first in the block, take a gander |
5340 | // at all of the predecessors of this instruction, and simplify them. |
5341 | if (&BB->front() != UI) |
5342 | return Changed; |
5343 | |
5344 | std::vector<DominatorTree::UpdateType> Updates; |
5345 | |
5346 | SmallSetVector<BasicBlock *, 8> Preds(pred_begin(BB), pred_end(BB)); |
5347 | for (BasicBlock *Predecessor : Preds) { |
5348 | Instruction *TI = Predecessor->getTerminator(); |
5349 | IRBuilder<> Builder(TI); |
5350 | if (auto *BI = dyn_cast<BranchInst>(Val: TI)) { |
5351 | // We could either have a proper unconditional branch, |
5352 | // or a degenerate conditional branch with matching destinations. |
5353 | if (all_of(Range: BI->successors(), |
5354 | P: [BB](auto *Successor) { return Successor == BB; })) { |
5355 | new UnreachableInst(TI->getContext(), TI->getIterator()); |
5356 | TI->eraseFromParent(); |
5357 | Changed = true; |
5358 | } else { |
5359 | assert(BI->isConditional() && "Can't get here with an uncond branch." ); |
5360 | Value* Cond = BI->getCondition(); |
5361 | assert(BI->getSuccessor(0) != BI->getSuccessor(1) && |
5362 | "The destinations are guaranteed to be different here." ); |
5363 | CallInst *Assumption; |
5364 | if (BI->getSuccessor(i: 0) == BB) { |
5365 | Assumption = Builder.CreateAssumption(Cond: Builder.CreateNot(V: Cond)); |
5366 | Builder.CreateBr(Dest: BI->getSuccessor(i: 1)); |
5367 | } else { |
5368 | assert(BI->getSuccessor(1) == BB && "Incorrect CFG" ); |
5369 | Assumption = Builder.CreateAssumption(Cond); |
5370 | Builder.CreateBr(Dest: BI->getSuccessor(i: 0)); |
5371 | } |
5372 | if (Options.AC) |
5373 | Options.AC->registerAssumption(CI: cast<AssumeInst>(Val: Assumption)); |
5374 | |
5375 | EraseTerminatorAndDCECond(TI: BI); |
5376 | Changed = true; |
5377 | } |
5378 | if (DTU) |
5379 | Updates.push_back(x: {DominatorTree::Delete, Predecessor, BB}); |
5380 | } else if (auto *SI = dyn_cast<SwitchInst>(Val: TI)) { |
5381 | SwitchInstProfUpdateWrapper SU(*SI); |
5382 | for (auto i = SU->case_begin(), e = SU->case_end(); i != e;) { |
5383 | if (i->getCaseSuccessor() != BB) { |
5384 | ++i; |
5385 | continue; |
5386 | } |
5387 | BB->removePredecessor(Pred: SU->getParent()); |
5388 | i = SU.removeCase(I: i); |
5389 | e = SU->case_end(); |
5390 | Changed = true; |
5391 | } |
5392 | // Note that the default destination can't be removed! |
5393 | if (DTU && SI->getDefaultDest() != BB) |
5394 | Updates.push_back(x: {DominatorTree::Delete, Predecessor, BB}); |
5395 | } else if (auto *II = dyn_cast<InvokeInst>(Val: TI)) { |
5396 | if (II->getUnwindDest() == BB) { |
5397 | if (DTU) { |
5398 | DTU->applyUpdates(Updates); |
5399 | Updates.clear(); |
5400 | } |
5401 | auto *CI = cast<CallInst>(Val: removeUnwindEdge(BB: TI->getParent(), DTU)); |
5402 | if (!CI->doesNotThrow()) |
5403 | CI->setDoesNotThrow(); |
5404 | Changed = true; |
5405 | } |
5406 | } else if (auto *CSI = dyn_cast<CatchSwitchInst>(Val: TI)) { |
5407 | if (CSI->getUnwindDest() == BB) { |
5408 | if (DTU) { |
5409 | DTU->applyUpdates(Updates); |
5410 | Updates.clear(); |
5411 | } |
5412 | removeUnwindEdge(BB: TI->getParent(), DTU); |
5413 | Changed = true; |
5414 | continue; |
5415 | } |
5416 | |
5417 | for (CatchSwitchInst::handler_iterator I = CSI->handler_begin(), |
5418 | E = CSI->handler_end(); |
5419 | I != E; ++I) { |
5420 | if (*I == BB) { |
5421 | CSI->removeHandler(HI: I); |
5422 | --I; |
5423 | --E; |
5424 | Changed = true; |
5425 | } |
5426 | } |
5427 | if (DTU) |
5428 | Updates.push_back(x: {DominatorTree::Delete, Predecessor, BB}); |
5429 | if (CSI->getNumHandlers() == 0) { |
5430 | if (CSI->hasUnwindDest()) { |
5431 | // Redirect all predecessors of the block containing CatchSwitchInst |
5432 | // to instead branch to the CatchSwitchInst's unwind destination. |
5433 | if (DTU) { |
5434 | for (auto *PredecessorOfPredecessor : predecessors(BB: Predecessor)) { |
5435 | Updates.push_back(x: {DominatorTree::Insert, |
5436 | PredecessorOfPredecessor, |
5437 | CSI->getUnwindDest()}); |
5438 | Updates.push_back(x: {DominatorTree::Delete, |
5439 | PredecessorOfPredecessor, Predecessor}); |
5440 | } |
5441 | } |
5442 | Predecessor->replaceAllUsesWith(V: CSI->getUnwindDest()); |
5443 | } else { |
5444 | // Rewrite all preds to unwind to caller (or from invoke to call). |
5445 | if (DTU) { |
5446 | DTU->applyUpdates(Updates); |
5447 | Updates.clear(); |
5448 | } |
5449 | SmallVector<BasicBlock *, 8> EHPreds(predecessors(BB: Predecessor)); |
5450 | for (BasicBlock *EHPred : EHPreds) |
5451 | removeUnwindEdge(BB: EHPred, DTU); |
5452 | } |
5453 | // The catchswitch is no longer reachable. |
5454 | new UnreachableInst(CSI->getContext(), CSI->getIterator()); |
5455 | CSI->eraseFromParent(); |
5456 | Changed = true; |
5457 | } |
5458 | } else if (auto *CRI = dyn_cast<CleanupReturnInst>(Val: TI)) { |
5459 | (void)CRI; |
5460 | assert(CRI->hasUnwindDest() && CRI->getUnwindDest() == BB && |
5461 | "Expected to always have an unwind to BB." ); |
5462 | if (DTU) |
5463 | Updates.push_back(x: {DominatorTree::Delete, Predecessor, BB}); |
5464 | new UnreachableInst(TI->getContext(), TI->getIterator()); |
5465 | TI->eraseFromParent(); |
5466 | Changed = true; |
5467 | } |
5468 | } |
5469 | |
5470 | if (DTU) |
5471 | DTU->applyUpdates(Updates); |
5472 | |
5473 | // If this block is now dead, remove it. |
5474 | if (pred_empty(BB) && BB != &BB->getParent()->getEntryBlock()) { |
5475 | DeleteDeadBlock(BB, DTU); |
5476 | return true; |
5477 | } |
5478 | |
5479 | return Changed; |
5480 | } |
5481 | |
5482 | static bool CasesAreContiguous(SmallVectorImpl<ConstantInt *> &Cases) { |
5483 | assert(Cases.size() >= 1); |
5484 | |
5485 | array_pod_sort(Start: Cases.begin(), End: Cases.end(), Compare: ConstantIntSortPredicate); |
5486 | for (size_t I = 1, E = Cases.size(); I != E; ++I) { |
5487 | if (Cases[I - 1]->getValue() != Cases[I]->getValue() + 1) |
5488 | return false; |
5489 | } |
5490 | return true; |
5491 | } |
5492 | |
5493 | static void createUnreachableSwitchDefault(SwitchInst *Switch, |
5494 | DomTreeUpdater *DTU, |
5495 | bool RemoveOrigDefaultBlock = true) { |
5496 | LLVM_DEBUG(dbgs() << "SimplifyCFG: switch default is dead.\n" ); |
5497 | auto *BB = Switch->getParent(); |
5498 | auto *OrigDefaultBlock = Switch->getDefaultDest(); |
5499 | if (RemoveOrigDefaultBlock) |
5500 | OrigDefaultBlock->removePredecessor(Pred: BB); |
5501 | BasicBlock *NewDefaultBlock = BasicBlock::Create( |
5502 | Context&: BB->getContext(), Name: BB->getName() + ".unreachabledefault" , Parent: BB->getParent(), |
5503 | InsertBefore: OrigDefaultBlock); |
5504 | new UnreachableInst(Switch->getContext(), NewDefaultBlock); |
5505 | Switch->setDefaultDest(&*NewDefaultBlock); |
5506 | if (DTU) { |
5507 | SmallVector<DominatorTree::UpdateType, 2> Updates; |
5508 | Updates.push_back(Elt: {DominatorTree::Insert, BB, &*NewDefaultBlock}); |
5509 | if (RemoveOrigDefaultBlock && |
5510 | !is_contained(Range: successors(BB), Element: OrigDefaultBlock)) |
5511 | Updates.push_back(Elt: {DominatorTree::Delete, BB, &*OrigDefaultBlock}); |
5512 | DTU->applyUpdates(Updates); |
5513 | } |
5514 | } |
5515 | |
5516 | /// Turn a switch into an integer range comparison and branch. |
5517 | /// Switches with more than 2 destinations are ignored. |
5518 | /// Switches with 1 destination are also ignored. |
5519 | bool SimplifyCFGOpt::TurnSwitchRangeIntoICmp(SwitchInst *SI, |
5520 | IRBuilder<> &Builder) { |
5521 | assert(SI->getNumCases() > 1 && "Degenerate switch?" ); |
5522 | |
5523 | bool HasDefault = |
5524 | !isa<UnreachableInst>(Val: SI->getDefaultDest()->getFirstNonPHIOrDbg()); |
5525 | |
5526 | auto *BB = SI->getParent(); |
5527 | |
5528 | // Partition the cases into two sets with different destinations. |
5529 | BasicBlock *DestA = HasDefault ? SI->getDefaultDest() : nullptr; |
5530 | BasicBlock *DestB = nullptr; |
5531 | SmallVector<ConstantInt *, 16> CasesA; |
5532 | SmallVector<ConstantInt *, 16> CasesB; |
5533 | |
5534 | for (auto Case : SI->cases()) { |
5535 | BasicBlock *Dest = Case.getCaseSuccessor(); |
5536 | if (!DestA) |
5537 | DestA = Dest; |
5538 | if (Dest == DestA) { |
5539 | CasesA.push_back(Elt: Case.getCaseValue()); |
5540 | continue; |
5541 | } |
5542 | if (!DestB) |
5543 | DestB = Dest; |
5544 | if (Dest == DestB) { |
5545 | CasesB.push_back(Elt: Case.getCaseValue()); |
5546 | continue; |
5547 | } |
5548 | return false; // More than two destinations. |
5549 | } |
5550 | if (!DestB) |
5551 | return false; // All destinations are the same and the default is unreachable |
5552 | |
5553 | assert(DestA && DestB && |
5554 | "Single-destination switch should have been folded." ); |
5555 | assert(DestA != DestB); |
5556 | assert(DestB != SI->getDefaultDest()); |
5557 | assert(!CasesB.empty() && "There must be non-default cases." ); |
5558 | assert(!CasesA.empty() || HasDefault); |
5559 | |
5560 | // Figure out if one of the sets of cases form a contiguous range. |
5561 | SmallVectorImpl<ConstantInt *> *ContiguousCases = nullptr; |
5562 | BasicBlock *ContiguousDest = nullptr; |
5563 | BasicBlock *OtherDest = nullptr; |
5564 | if (!CasesA.empty() && CasesAreContiguous(Cases&: CasesA)) { |
5565 | ContiguousCases = &CasesA; |
5566 | ContiguousDest = DestA; |
5567 | OtherDest = DestB; |
5568 | } else if (CasesAreContiguous(Cases&: CasesB)) { |
5569 | ContiguousCases = &CasesB; |
5570 | ContiguousDest = DestB; |
5571 | OtherDest = DestA; |
5572 | } else |
5573 | return false; |
5574 | |
5575 | // Start building the compare and branch. |
5576 | |
5577 | Constant *Offset = ConstantExpr::getNeg(C: ContiguousCases->back()); |
5578 | Constant *NumCases = |
5579 | ConstantInt::get(Ty: Offset->getType(), V: ContiguousCases->size()); |
5580 | |
5581 | Value *Sub = SI->getCondition(); |
5582 | if (!Offset->isNullValue()) |
5583 | Sub = Builder.CreateAdd(LHS: Sub, RHS: Offset, Name: Sub->getName() + ".off" ); |
5584 | |
5585 | Value *Cmp; |
5586 | // If NumCases overflowed, then all possible values jump to the successor. |
5587 | if (NumCases->isNullValue() && !ContiguousCases->empty()) |
5588 | Cmp = ConstantInt::getTrue(Context&: SI->getContext()); |
5589 | else |
5590 | Cmp = Builder.CreateICmpULT(LHS: Sub, RHS: NumCases, Name: "switch" ); |
5591 | BranchInst *NewBI = Builder.CreateCondBr(Cond: Cmp, True: ContiguousDest, False: OtherDest); |
5592 | |
5593 | // Update weight for the newly-created conditional branch. |
5594 | if (hasBranchWeightMD(I: *SI)) { |
5595 | SmallVector<uint64_t, 8> Weights; |
5596 | GetBranchWeights(TI: SI, Weights); |
5597 | if (Weights.size() == 1 + SI->getNumCases()) { |
5598 | uint64_t TrueWeight = 0; |
5599 | uint64_t FalseWeight = 0; |
5600 | for (size_t I = 0, E = Weights.size(); I != E; ++I) { |
5601 | if (SI->getSuccessor(idx: I) == ContiguousDest) |
5602 | TrueWeight += Weights[I]; |
5603 | else |
5604 | FalseWeight += Weights[I]; |
5605 | } |
5606 | while (TrueWeight > UINT32_MAX || FalseWeight > UINT32_MAX) { |
5607 | TrueWeight /= 2; |
5608 | FalseWeight /= 2; |
5609 | } |
5610 | setBranchWeights(I: NewBI, TrueWeight, FalseWeight, /*IsExpected=*/false); |
5611 | } |
5612 | } |
5613 | |
5614 | // Prune obsolete incoming values off the successors' PHI nodes. |
5615 | for (auto BBI = ContiguousDest->begin(); isa<PHINode>(Val: BBI); ++BBI) { |
5616 | unsigned PreviousEdges = ContiguousCases->size(); |
5617 | if (ContiguousDest == SI->getDefaultDest()) |
5618 | ++PreviousEdges; |
5619 | for (unsigned I = 0, E = PreviousEdges - 1; I != E; ++I) |
5620 | cast<PHINode>(Val&: BBI)->removeIncomingValue(BB: SI->getParent()); |
5621 | } |
5622 | for (auto BBI = OtherDest->begin(); isa<PHINode>(Val: BBI); ++BBI) { |
5623 | unsigned PreviousEdges = SI->getNumCases() - ContiguousCases->size(); |
5624 | if (OtherDest == SI->getDefaultDest()) |
5625 | ++PreviousEdges; |
5626 | for (unsigned I = 0, E = PreviousEdges - 1; I != E; ++I) |
5627 | cast<PHINode>(Val&: BBI)->removeIncomingValue(BB: SI->getParent()); |
5628 | } |
5629 | |
5630 | // Clean up the default block - it may have phis or other instructions before |
5631 | // the unreachable terminator. |
5632 | if (!HasDefault) |
5633 | createUnreachableSwitchDefault(Switch: SI, DTU); |
5634 | |
5635 | auto *UnreachableDefault = SI->getDefaultDest(); |
5636 | |
5637 | // Drop the switch. |
5638 | SI->eraseFromParent(); |
5639 | |
5640 | if (!HasDefault && DTU) |
5641 | DTU->applyUpdates(Updates: {{DominatorTree::Delete, BB, UnreachableDefault}}); |
5642 | |
5643 | return true; |
5644 | } |
5645 | |
5646 | /// Compute masked bits for the condition of a switch |
5647 | /// and use it to remove dead cases. |
5648 | static bool eliminateDeadSwitchCases(SwitchInst *SI, DomTreeUpdater *DTU, |
5649 | AssumptionCache *AC, |
5650 | const DataLayout &DL) { |
5651 | Value *Cond = SI->getCondition(); |
5652 | KnownBits Known = computeKnownBits(V: Cond, DL, Depth: 0, AC, CxtI: SI); |
5653 | |
5654 | // We can also eliminate cases by determining that their values are outside of |
5655 | // the limited range of the condition based on how many significant (non-sign) |
5656 | // bits are in the condition value. |
5657 | unsigned MaxSignificantBitsInCond = |
5658 | ComputeMaxSignificantBits(Op: Cond, DL, Depth: 0, AC, CxtI: SI); |
5659 | |
5660 | // Gather dead cases. |
5661 | SmallVector<ConstantInt *, 8> DeadCases; |
5662 | SmallDenseMap<BasicBlock *, int, 8> NumPerSuccessorCases; |
5663 | SmallVector<BasicBlock *, 8> UniqueSuccessors; |
5664 | for (const auto &Case : SI->cases()) { |
5665 | auto *Successor = Case.getCaseSuccessor(); |
5666 | if (DTU) { |
5667 | if (!NumPerSuccessorCases.count(Val: Successor)) |
5668 | UniqueSuccessors.push_back(Elt: Successor); |
5669 | ++NumPerSuccessorCases[Successor]; |
5670 | } |
5671 | const APInt &CaseVal = Case.getCaseValue()->getValue(); |
5672 | if (Known.Zero.intersects(RHS: CaseVal) || !Known.One.isSubsetOf(RHS: CaseVal) || |
5673 | (CaseVal.getSignificantBits() > MaxSignificantBitsInCond)) { |
5674 | DeadCases.push_back(Elt: Case.getCaseValue()); |
5675 | if (DTU) |
5676 | --NumPerSuccessorCases[Successor]; |
5677 | LLVM_DEBUG(dbgs() << "SimplifyCFG: switch case " << CaseVal |
5678 | << " is dead.\n" ); |
5679 | } |
5680 | } |
5681 | |
5682 | // If we can prove that the cases must cover all possible values, the |
5683 | // default destination becomes dead and we can remove it. If we know some |
5684 | // of the bits in the value, we can use that to more precisely compute the |
5685 | // number of possible unique case values. |
5686 | bool HasDefault = |
5687 | !isa<UnreachableInst>(Val: SI->getDefaultDest()->getFirstNonPHIOrDbg()); |
5688 | const unsigned NumUnknownBits = |
5689 | Known.getBitWidth() - (Known.Zero | Known.One).popcount(); |
5690 | assert(NumUnknownBits <= Known.getBitWidth()); |
5691 | if (HasDefault && DeadCases.empty() && |
5692 | NumUnknownBits < 64 /* avoid overflow */) { |
5693 | uint64_t AllNumCases = 1ULL << NumUnknownBits; |
5694 | if (SI->getNumCases() == AllNumCases) { |
5695 | createUnreachableSwitchDefault(Switch: SI, DTU); |
5696 | return true; |
5697 | } |
5698 | // When only one case value is missing, replace default with that case. |
5699 | // Eliminating the default branch will provide more opportunities for |
5700 | // optimization, such as lookup tables. |
5701 | if (SI->getNumCases() == AllNumCases - 1) { |
5702 | assert(NumUnknownBits > 1 && "Should be canonicalized to a branch" ); |
5703 | IntegerType *CondTy = cast<IntegerType>(Val: Cond->getType()); |
5704 | if (CondTy->getIntegerBitWidth() > 64 || |
5705 | !DL.fitsInLegalInteger(Width: CondTy->getIntegerBitWidth())) |
5706 | return false; |
5707 | |
5708 | uint64_t MissingCaseVal = 0; |
5709 | for (const auto &Case : SI->cases()) |
5710 | MissingCaseVal ^= Case.getCaseValue()->getValue().getLimitedValue(); |
5711 | auto *MissingCase = |
5712 | cast<ConstantInt>(Val: ConstantInt::get(Ty: Cond->getType(), V: MissingCaseVal)); |
5713 | SwitchInstProfUpdateWrapper SIW(*SI); |
5714 | SIW.addCase(OnVal: MissingCase, Dest: SI->getDefaultDest(), W: SIW.getSuccessorWeight(idx: 0)); |
5715 | createUnreachableSwitchDefault(Switch: SI, DTU, /*RemoveOrigDefaultBlock*/ false); |
5716 | SIW.setSuccessorWeight(idx: 0, W: 0); |
5717 | return true; |
5718 | } |
5719 | } |
5720 | |
5721 | if (DeadCases.empty()) |
5722 | return false; |
5723 | |
5724 | SwitchInstProfUpdateWrapper SIW(*SI); |
5725 | for (ConstantInt *DeadCase : DeadCases) { |
5726 | SwitchInst::CaseIt CaseI = SI->findCaseValue(C: DeadCase); |
5727 | assert(CaseI != SI->case_default() && |
5728 | "Case was not found. Probably mistake in DeadCases forming." ); |
5729 | // Prune unused values from PHI nodes. |
5730 | CaseI->getCaseSuccessor()->removePredecessor(Pred: SI->getParent()); |
5731 | SIW.removeCase(I: CaseI); |
5732 | } |
5733 | |
5734 | if (DTU) { |
5735 | std::vector<DominatorTree::UpdateType> Updates; |
5736 | for (auto *Successor : UniqueSuccessors) |
5737 | if (NumPerSuccessorCases[Successor] == 0) |
5738 | Updates.push_back(x: {DominatorTree::Delete, SI->getParent(), Successor}); |
5739 | DTU->applyUpdates(Updates); |
5740 | } |
5741 | |
5742 | return true; |
5743 | } |
5744 | |
5745 | /// If BB would be eligible for simplification by |
5746 | /// TryToSimplifyUncondBranchFromEmptyBlock (i.e. it is empty and terminated |
5747 | /// by an unconditional branch), look at the phi node for BB in the successor |
5748 | /// block and see if the incoming value is equal to CaseValue. If so, return |
5749 | /// the phi node, and set PhiIndex to BB's index in the phi node. |
5750 | static PHINode *FindPHIForConditionForwarding(ConstantInt *CaseValue, |
5751 | BasicBlock *BB, int *PhiIndex) { |
5752 | if (BB->getFirstNonPHIOrDbg() != BB->getTerminator()) |
5753 | return nullptr; // BB must be empty to be a candidate for simplification. |
5754 | if (!BB->getSinglePredecessor()) |
5755 | return nullptr; // BB must be dominated by the switch. |
5756 | |
5757 | BranchInst *Branch = dyn_cast<BranchInst>(Val: BB->getTerminator()); |
5758 | if (!Branch || !Branch->isUnconditional()) |
5759 | return nullptr; // Terminator must be unconditional branch. |
5760 | |
5761 | BasicBlock *Succ = Branch->getSuccessor(i: 0); |
5762 | |
5763 | for (PHINode &PHI : Succ->phis()) { |
5764 | int Idx = PHI.getBasicBlockIndex(BB); |
5765 | assert(Idx >= 0 && "PHI has no entry for predecessor?" ); |
5766 | |
5767 | Value *InValue = PHI.getIncomingValue(i: Idx); |
5768 | if (InValue != CaseValue) |
5769 | continue; |
5770 | |
5771 | *PhiIndex = Idx; |
5772 | return &PHI; |
5773 | } |
5774 | |
5775 | return nullptr; |
5776 | } |
5777 | |
5778 | /// Try to forward the condition of a switch instruction to a phi node |
5779 | /// dominated by the switch, if that would mean that some of the destination |
5780 | /// blocks of the switch can be folded away. Return true if a change is made. |
5781 | static bool ForwardSwitchConditionToPHI(SwitchInst *SI) { |
5782 | using ForwardingNodesMap = DenseMap<PHINode *, SmallVector<int, 4>>; |
5783 | |
5784 | ForwardingNodesMap ForwardingNodes; |
5785 | BasicBlock *SwitchBlock = SI->getParent(); |
5786 | bool Changed = false; |
5787 | for (const auto &Case : SI->cases()) { |
5788 | ConstantInt *CaseValue = Case.getCaseValue(); |
5789 | BasicBlock *CaseDest = Case.getCaseSuccessor(); |
5790 | |
5791 | // Replace phi operands in successor blocks that are using the constant case |
5792 | // value rather than the switch condition variable: |
5793 | // switchbb: |
5794 | // switch i32 %x, label %default [ |
5795 | // i32 17, label %succ |
5796 | // ... |
5797 | // succ: |
5798 | // %r = phi i32 ... [ 17, %switchbb ] ... |
5799 | // --> |
5800 | // %r = phi i32 ... [ %x, %switchbb ] ... |
5801 | |
5802 | for (PHINode &Phi : CaseDest->phis()) { |
5803 | // This only works if there is exactly 1 incoming edge from the switch to |
5804 | // a phi. If there is >1, that means multiple cases of the switch map to 1 |
5805 | // value in the phi, and that phi value is not the switch condition. Thus, |
5806 | // this transform would not make sense (the phi would be invalid because |
5807 | // a phi can't have different incoming values from the same block). |
5808 | int SwitchBBIdx = Phi.getBasicBlockIndex(BB: SwitchBlock); |
5809 | if (Phi.getIncomingValue(i: SwitchBBIdx) == CaseValue && |
5810 | count(Range: Phi.blocks(), Element: SwitchBlock) == 1) { |
5811 | Phi.setIncomingValue(i: SwitchBBIdx, V: SI->getCondition()); |
5812 | Changed = true; |
5813 | } |
5814 | } |
5815 | |
5816 | // Collect phi nodes that are indirectly using this switch's case constants. |
5817 | int PhiIdx; |
5818 | if (auto *Phi = FindPHIForConditionForwarding(CaseValue, BB: CaseDest, PhiIndex: &PhiIdx)) |
5819 | ForwardingNodes[Phi].push_back(Elt: PhiIdx); |
5820 | } |
5821 | |
5822 | for (auto &ForwardingNode : ForwardingNodes) { |
5823 | PHINode *Phi = ForwardingNode.first; |
5824 | SmallVectorImpl<int> &Indexes = ForwardingNode.second; |
5825 | // Check if it helps to fold PHI. |
5826 | if (Indexes.size() < 2 && !llvm::is_contained(Range: Phi->incoming_values(), Element: SI->getCondition())) |
5827 | continue; |
5828 | |
5829 | for (int Index : Indexes) |
5830 | Phi->setIncomingValue(i: Index, V: SI->getCondition()); |
5831 | Changed = true; |
5832 | } |
5833 | |
5834 | return Changed; |
5835 | } |
5836 | |
5837 | /// Return true if the backend will be able to handle |
5838 | /// initializing an array of constants like C. |
5839 | static bool ValidLookupTableConstant(Constant *C, const TargetTransformInfo &TTI) { |
5840 | if (C->isThreadDependent()) |
5841 | return false; |
5842 | if (C->isDLLImportDependent()) |
5843 | return false; |
5844 | |
5845 | if (!isa<ConstantFP>(Val: C) && !isa<ConstantInt>(Val: C) && |
5846 | !isa<ConstantPointerNull>(Val: C) && !isa<GlobalValue>(Val: C) && |
5847 | !isa<UndefValue>(Val: C) && !isa<ConstantExpr>(Val: C)) |
5848 | return false; |
5849 | |
5850 | if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Val: C)) { |
5851 | // Pointer casts and in-bounds GEPs will not prohibit the backend from |
5852 | // materializing the array of constants. |
5853 | Constant *StrippedC = cast<Constant>(Val: CE->stripInBoundsConstantOffsets()); |
5854 | if (StrippedC == C || !ValidLookupTableConstant(C: StrippedC, TTI)) |
5855 | return false; |
5856 | } |
5857 | |
5858 | if (!TTI.shouldBuildLookupTablesForConstant(C)) |
5859 | return false; |
5860 | |
5861 | return true; |
5862 | } |
5863 | |
5864 | /// If V is a Constant, return it. Otherwise, try to look up |
5865 | /// its constant value in ConstantPool, returning 0 if it's not there. |
5866 | static Constant * |
5867 | LookupConstant(Value *V, |
5868 | const SmallDenseMap<Value *, Constant *> &ConstantPool) { |
5869 | if (Constant *C = dyn_cast<Constant>(Val: V)) |
5870 | return C; |
5871 | return ConstantPool.lookup(Val: V); |
5872 | } |
5873 | |
5874 | /// Try to fold instruction I into a constant. This works for |
5875 | /// simple instructions such as binary operations where both operands are |
5876 | /// constant or can be replaced by constants from the ConstantPool. Returns the |
5877 | /// resulting constant on success, 0 otherwise. |
5878 | static Constant * |
5879 | ConstantFold(Instruction *I, const DataLayout &DL, |
5880 | const SmallDenseMap<Value *, Constant *> &ConstantPool) { |
5881 | if (SelectInst *Select = dyn_cast<SelectInst>(Val: I)) { |
5882 | Constant *A = LookupConstant(V: Select->getCondition(), ConstantPool); |
5883 | if (!A) |
5884 | return nullptr; |
5885 | if (A->isAllOnesValue()) |
5886 | return LookupConstant(V: Select->getTrueValue(), ConstantPool); |
5887 | if (A->isNullValue()) |
5888 | return LookupConstant(V: Select->getFalseValue(), ConstantPool); |
5889 | return nullptr; |
5890 | } |
5891 | |
5892 | SmallVector<Constant *, 4> COps; |
5893 | for (unsigned N = 0, E = I->getNumOperands(); N != E; ++N) { |
5894 | if (Constant *A = LookupConstant(V: I->getOperand(i: N), ConstantPool)) |
5895 | COps.push_back(Elt: A); |
5896 | else |
5897 | return nullptr; |
5898 | } |
5899 | |
5900 | return ConstantFoldInstOperands(I, Ops: COps, DL); |
5901 | } |
5902 | |
5903 | /// Try to determine the resulting constant values in phi nodes |
5904 | /// at the common destination basic block, *CommonDest, for one of the case |
5905 | /// destionations CaseDest corresponding to value CaseVal (0 for the default |
5906 | /// case), of a switch instruction SI. |
5907 | static bool |
5908 | getCaseResults(SwitchInst *SI, ConstantInt *CaseVal, BasicBlock *CaseDest, |
5909 | BasicBlock **CommonDest, |
5910 | SmallVectorImpl<std::pair<PHINode *, Constant *>> &Res, |
5911 | const DataLayout &DL, const TargetTransformInfo &TTI) { |
5912 | // The block from which we enter the common destination. |
5913 | BasicBlock *Pred = SI->getParent(); |
5914 | |
5915 | // If CaseDest is empty except for some side-effect free instructions through |
5916 | // which we can constant-propagate the CaseVal, continue to its successor. |
5917 | SmallDenseMap<Value *, Constant *> ConstantPool; |
5918 | ConstantPool.insert(KV: std::make_pair(x: SI->getCondition(), y&: CaseVal)); |
5919 | for (Instruction &I : CaseDest->instructionsWithoutDebug(SkipPseudoOp: false)) { |
5920 | if (I.isTerminator()) { |
5921 | // If the terminator is a simple branch, continue to the next block. |
5922 | if (I.getNumSuccessors() != 1 || I.isSpecialTerminator()) |
5923 | return false; |
5924 | Pred = CaseDest; |
5925 | CaseDest = I.getSuccessor(Idx: 0); |
5926 | } else if (Constant *C = ConstantFold(I: &I, DL, ConstantPool)) { |
5927 | // Instruction is side-effect free and constant. |
5928 | |
5929 | // If the instruction has uses outside this block or a phi node slot for |
5930 | // the block, it is not safe to bypass the instruction since it would then |
5931 | // no longer dominate all its uses. |
5932 | for (auto &Use : I.uses()) { |
5933 | User *User = Use.getUser(); |
5934 | if (Instruction *I = dyn_cast<Instruction>(Val: User)) |
5935 | if (I->getParent() == CaseDest) |
5936 | continue; |
5937 | if (PHINode *Phi = dyn_cast<PHINode>(Val: User)) |
5938 | if (Phi->getIncomingBlock(U: Use) == CaseDest) |
5939 | continue; |
5940 | return false; |
5941 | } |
5942 | |
5943 | ConstantPool.insert(KV: std::make_pair(x: &I, y&: C)); |
5944 | } else { |
5945 | break; |
5946 | } |
5947 | } |
5948 | |
5949 | // If we did not have a CommonDest before, use the current one. |
5950 | if (!*CommonDest) |
5951 | *CommonDest = CaseDest; |
5952 | // If the destination isn't the common one, abort. |
5953 | if (CaseDest != *CommonDest) |
5954 | return false; |
5955 | |
5956 | // Get the values for this case from phi nodes in the destination block. |
5957 | for (PHINode &PHI : (*CommonDest)->phis()) { |
5958 | int Idx = PHI.getBasicBlockIndex(BB: Pred); |
5959 | if (Idx == -1) |
5960 | continue; |
5961 | |
5962 | Constant *ConstVal = |
5963 | LookupConstant(V: PHI.getIncomingValue(i: Idx), ConstantPool); |
5964 | if (!ConstVal) |
5965 | return false; |
5966 | |
5967 | // Be conservative about which kinds of constants we support. |
5968 | if (!ValidLookupTableConstant(C: ConstVal, TTI)) |
5969 | return false; |
5970 | |
5971 | Res.push_back(Elt: std::make_pair(x: &PHI, y&: ConstVal)); |
5972 | } |
5973 | |
5974 | return Res.size() > 0; |
5975 | } |
5976 | |
5977 | // Helper function used to add CaseVal to the list of cases that generate |
5978 | // Result. Returns the updated number of cases that generate this result. |
5979 | static size_t mapCaseToResult(ConstantInt *CaseVal, |
5980 | SwitchCaseResultVectorTy &UniqueResults, |
5981 | Constant *Result) { |
5982 | for (auto &I : UniqueResults) { |
5983 | if (I.first == Result) { |
5984 | I.second.push_back(Elt: CaseVal); |
5985 | return I.second.size(); |
5986 | } |
5987 | } |
5988 | UniqueResults.push_back( |
5989 | Elt: std::make_pair(x&: Result, y: SmallVector<ConstantInt *, 4>(1, CaseVal))); |
5990 | return 1; |
5991 | } |
5992 | |
5993 | // Helper function that initializes a map containing |
5994 | // results for the PHI node of the common destination block for a switch |
5995 | // instruction. Returns false if multiple PHI nodes have been found or if |
5996 | // there is not a common destination block for the switch. |
5997 | static bool initializeUniqueCases(SwitchInst *SI, PHINode *&PHI, |
5998 | BasicBlock *&CommonDest, |
5999 | SwitchCaseResultVectorTy &UniqueResults, |
6000 | Constant *&DefaultResult, |
6001 | const DataLayout &DL, |
6002 | const TargetTransformInfo &TTI, |
6003 | uintptr_t MaxUniqueResults) { |
6004 | for (const auto &I : SI->cases()) { |
6005 | ConstantInt *CaseVal = I.getCaseValue(); |
6006 | |
6007 | // Resulting value at phi nodes for this case value. |
6008 | SwitchCaseResultsTy Results; |
6009 | if (!getCaseResults(SI, CaseVal, CaseDest: I.getCaseSuccessor(), CommonDest: &CommonDest, Res&: Results, |
6010 | DL, TTI)) |
6011 | return false; |
6012 | |
6013 | // Only one value per case is permitted. |
6014 | if (Results.size() > 1) |
6015 | return false; |
6016 | |
6017 | // Add the case->result mapping to UniqueResults. |
6018 | const size_t NumCasesForResult = |
6019 | mapCaseToResult(CaseVal, UniqueResults, Result: Results.begin()->second); |
6020 | |
6021 | // Early out if there are too many cases for this result. |
6022 | if (NumCasesForResult > MaxSwitchCasesPerResult) |
6023 | return false; |
6024 | |
6025 | // Early out if there are too many unique results. |
6026 | if (UniqueResults.size() > MaxUniqueResults) |
6027 | return false; |
6028 | |
6029 | // Check the PHI consistency. |
6030 | if (!PHI) |
6031 | PHI = Results[0].first; |
6032 | else if (PHI != Results[0].first) |
6033 | return false; |
6034 | } |
6035 | // Find the default result value. |
6036 | SmallVector<std::pair<PHINode *, Constant *>, 1> DefaultResults; |
6037 | BasicBlock *DefaultDest = SI->getDefaultDest(); |
6038 | getCaseResults(SI, CaseVal: nullptr, CaseDest: SI->getDefaultDest(), CommonDest: &CommonDest, Res&: DefaultResults, |
6039 | DL, TTI); |
6040 | // If the default value is not found abort unless the default destination |
6041 | // is unreachable. |
6042 | DefaultResult = |
6043 | DefaultResults.size() == 1 ? DefaultResults.begin()->second : nullptr; |
6044 | if ((!DefaultResult && |
6045 | !isa<UnreachableInst>(Val: DefaultDest->getFirstNonPHIOrDbg()))) |
6046 | return false; |
6047 | |
6048 | return true; |
6049 | } |
6050 | |
6051 | // Helper function that checks if it is possible to transform a switch with only |
6052 | // two cases (or two cases + default) that produces a result into a select. |
6053 | // TODO: Handle switches with more than 2 cases that map to the same result. |
6054 | static Value *foldSwitchToSelect(const SwitchCaseResultVectorTy &ResultVector, |
6055 | Constant *DefaultResult, Value *Condition, |
6056 | IRBuilder<> &Builder) { |
6057 | // If we are selecting between only two cases transform into a simple |
6058 | // select or a two-way select if default is possible. |
6059 | // Example: |
6060 | // switch (a) { %0 = icmp eq i32 %a, 10 |
6061 | // case 10: return 42; %1 = select i1 %0, i32 42, i32 4 |
6062 | // case 20: return 2; ----> %2 = icmp eq i32 %a, 20 |
6063 | // default: return 4; %3 = select i1 %2, i32 2, i32 %1 |
6064 | // } |
6065 | if (ResultVector.size() == 2 && ResultVector[0].second.size() == 1 && |
6066 | ResultVector[1].second.size() == 1) { |
6067 | ConstantInt *FirstCase = ResultVector[0].second[0]; |
6068 | ConstantInt *SecondCase = ResultVector[1].second[0]; |
6069 | Value *SelectValue = ResultVector[1].first; |
6070 | if (DefaultResult) { |
6071 | Value *ValueCompare = |
6072 | Builder.CreateICmpEQ(LHS: Condition, RHS: SecondCase, Name: "switch.selectcmp" ); |
6073 | SelectValue = Builder.CreateSelect(C: ValueCompare, True: ResultVector[1].first, |
6074 | False: DefaultResult, Name: "switch.select" ); |
6075 | } |
6076 | Value *ValueCompare = |
6077 | Builder.CreateICmpEQ(LHS: Condition, RHS: FirstCase, Name: "switch.selectcmp" ); |
6078 | return Builder.CreateSelect(C: ValueCompare, True: ResultVector[0].first, |
6079 | False: SelectValue, Name: "switch.select" ); |
6080 | } |
6081 | |
6082 | // Handle the degenerate case where two cases have the same result value. |
6083 | if (ResultVector.size() == 1 && DefaultResult) { |
6084 | ArrayRef<ConstantInt *> CaseValues = ResultVector[0].second; |
6085 | unsigned CaseCount = CaseValues.size(); |
6086 | // n bits group cases map to the same result: |
6087 | // case 0,4 -> Cond & 0b1..1011 == 0 ? result : default |
6088 | // case 0,2,4,6 -> Cond & 0b1..1001 == 0 ? result : default |
6089 | // case 0,2,8,10 -> Cond & 0b1..0101 == 0 ? result : default |
6090 | if (isPowerOf2_32(Value: CaseCount)) { |
6091 | ConstantInt *MinCaseVal = CaseValues[0]; |
6092 | // Find mininal value. |
6093 | for (auto *Case : CaseValues) |
6094 | if (Case->getValue().slt(RHS: MinCaseVal->getValue())) |
6095 | MinCaseVal = Case; |
6096 | |
6097 | // Mark the bits case number touched. |
6098 | APInt BitMask = APInt::getZero(numBits: MinCaseVal->getBitWidth()); |
6099 | for (auto *Case : CaseValues) |
6100 | BitMask |= (Case->getValue() - MinCaseVal->getValue()); |
6101 | |
6102 | // Check if cases with the same result can cover all number |
6103 | // in touched bits. |
6104 | if (BitMask.popcount() == Log2_32(Value: CaseCount)) { |
6105 | if (!MinCaseVal->isNullValue()) |
6106 | Condition = Builder.CreateSub(LHS: Condition, RHS: MinCaseVal); |
6107 | Value *And = Builder.CreateAnd(LHS: Condition, RHS: ~BitMask, Name: "switch.and" ); |
6108 | Value *Cmp = Builder.CreateICmpEQ( |
6109 | LHS: And, RHS: Constant::getNullValue(Ty: And->getType()), Name: "switch.selectcmp" ); |
6110 | return Builder.CreateSelect(C: Cmp, True: ResultVector[0].first, False: DefaultResult); |
6111 | } |
6112 | } |
6113 | |
6114 | // Handle the degenerate case where two cases have the same value. |
6115 | if (CaseValues.size() == 2) { |
6116 | Value *Cmp1 = Builder.CreateICmpEQ(LHS: Condition, RHS: CaseValues[0], |
6117 | Name: "switch.selectcmp.case1" ); |
6118 | Value *Cmp2 = Builder.CreateICmpEQ(LHS: Condition, RHS: CaseValues[1], |
6119 | Name: "switch.selectcmp.case2" ); |
6120 | Value *Cmp = Builder.CreateOr(LHS: Cmp1, RHS: Cmp2, Name: "switch.selectcmp" ); |
6121 | return Builder.CreateSelect(C: Cmp, True: ResultVector[0].first, False: DefaultResult); |
6122 | } |
6123 | } |
6124 | |
6125 | return nullptr; |
6126 | } |
6127 | |
6128 | // Helper function to cleanup a switch instruction that has been converted into |
6129 | // a select, fixing up PHI nodes and basic blocks. |
6130 | static void removeSwitchAfterSelectFold(SwitchInst *SI, PHINode *PHI, |
6131 | Value *SelectValue, |
6132 | IRBuilder<> &Builder, |
6133 | DomTreeUpdater *DTU) { |
6134 | std::vector<DominatorTree::UpdateType> Updates; |
6135 | |
6136 | BasicBlock *SelectBB = SI->getParent(); |
6137 | BasicBlock *DestBB = PHI->getParent(); |
6138 | |
6139 | if (DTU && !is_contained(Range: predecessors(BB: DestBB), Element: SelectBB)) |
6140 | Updates.push_back(x: {DominatorTree::Insert, SelectBB, DestBB}); |
6141 | Builder.CreateBr(Dest: DestBB); |
6142 | |
6143 | // Remove the switch. |
6144 | |
6145 | PHI->removeIncomingValueIf( |
6146 | Predicate: [&](unsigned Idx) { return PHI->getIncomingBlock(i: Idx) == SelectBB; }); |
6147 | PHI->addIncoming(V: SelectValue, BB: SelectBB); |
6148 | |
6149 | SmallPtrSet<BasicBlock *, 4> RemovedSuccessors; |
6150 | for (unsigned i = 0, e = SI->getNumSuccessors(); i < e; ++i) { |
6151 | BasicBlock *Succ = SI->getSuccessor(idx: i); |
6152 | |
6153 | if (Succ == DestBB) |
6154 | continue; |
6155 | Succ->removePredecessor(Pred: SelectBB); |
6156 | if (DTU && RemovedSuccessors.insert(Ptr: Succ).second) |
6157 | Updates.push_back(x: {DominatorTree::Delete, SelectBB, Succ}); |
6158 | } |
6159 | SI->eraseFromParent(); |
6160 | if (DTU) |
6161 | DTU->applyUpdates(Updates); |
6162 | } |
6163 | |
6164 | /// If a switch is only used to initialize one or more phi nodes in a common |
6165 | /// successor block with only two different constant values, try to replace the |
6166 | /// switch with a select. Returns true if the fold was made. |
6167 | static bool trySwitchToSelect(SwitchInst *SI, IRBuilder<> &Builder, |
6168 | DomTreeUpdater *DTU, const DataLayout &DL, |
6169 | const TargetTransformInfo &TTI) { |
6170 | Value *const Cond = SI->getCondition(); |
6171 | PHINode *PHI = nullptr; |
6172 | BasicBlock *CommonDest = nullptr; |
6173 | Constant *DefaultResult; |
6174 | SwitchCaseResultVectorTy UniqueResults; |
6175 | // Collect all the cases that will deliver the same value from the switch. |
6176 | if (!initializeUniqueCases(SI, PHI, CommonDest, UniqueResults, DefaultResult, |
6177 | DL, TTI, /*MaxUniqueResults*/ 2)) |
6178 | return false; |
6179 | |
6180 | assert(PHI != nullptr && "PHI for value select not found" ); |
6181 | Builder.SetInsertPoint(SI); |
6182 | Value *SelectValue = |
6183 | foldSwitchToSelect(ResultVector: UniqueResults, DefaultResult, Condition: Cond, Builder); |
6184 | if (!SelectValue) |
6185 | return false; |
6186 | |
6187 | removeSwitchAfterSelectFold(SI, PHI, SelectValue, Builder, DTU); |
6188 | return true; |
6189 | } |
6190 | |
6191 | namespace { |
6192 | |
6193 | /// This class represents a lookup table that can be used to replace a switch. |
6194 | class SwitchLookupTable { |
6195 | public: |
6196 | /// Create a lookup table to use as a switch replacement with the contents |
6197 | /// of Values, using DefaultValue to fill any holes in the table. |
6198 | SwitchLookupTable( |
6199 | Module &M, uint64_t TableSize, ConstantInt *Offset, |
6200 | const SmallVectorImpl<std::pair<ConstantInt *, Constant *>> &Values, |
6201 | Constant *DefaultValue, const DataLayout &DL, const StringRef &FuncName); |
6202 | |
6203 | /// Build instructions with Builder to retrieve the value at |
6204 | /// the position given by Index in the lookup table. |
6205 | Value *BuildLookup(Value *Index, IRBuilder<> &Builder); |
6206 | |
6207 | /// Return true if a table with TableSize elements of |
6208 | /// type ElementType would fit in a target-legal register. |
6209 | static bool WouldFitInRegister(const DataLayout &DL, uint64_t TableSize, |
6210 | Type *ElementType); |
6211 | |
6212 | private: |
6213 | // Depending on the contents of the table, it can be represented in |
6214 | // different ways. |
6215 | enum { |
6216 | // For tables where each element contains the same value, we just have to |
6217 | // store that single value and return it for each lookup. |
6218 | SingleValueKind, |
6219 | |
6220 | // For tables where there is a linear relationship between table index |
6221 | // and values. We calculate the result with a simple multiplication |
6222 | // and addition instead of a table lookup. |
6223 | LinearMapKind, |
6224 | |
6225 | // For small tables with integer elements, we can pack them into a bitmap |
6226 | // that fits into a target-legal register. Values are retrieved by |
6227 | // shift and mask operations. |
6228 | BitMapKind, |
6229 | |
6230 | // The table is stored as an array of values. Values are retrieved by load |
6231 | // instructions from the table. |
6232 | ArrayKind |
6233 | } Kind; |
6234 | |
6235 | // For SingleValueKind, this is the single value. |
6236 | Constant *SingleValue = nullptr; |
6237 | |
6238 | // For BitMapKind, this is the bitmap. |
6239 | ConstantInt *BitMap = nullptr; |
6240 | IntegerType *BitMapElementTy = nullptr; |
6241 | |
6242 | // For LinearMapKind, these are the constants used to derive the value. |
6243 | ConstantInt *LinearOffset = nullptr; |
6244 | ConstantInt *LinearMultiplier = nullptr; |
6245 | bool LinearMapValWrapped = false; |
6246 | |
6247 | // For ArrayKind, this is the array. |
6248 | GlobalVariable *Array = nullptr; |
6249 | }; |
6250 | |
6251 | } // end anonymous namespace |
6252 | |
6253 | SwitchLookupTable::SwitchLookupTable( |
6254 | Module &M, uint64_t TableSize, ConstantInt *Offset, |
6255 | const SmallVectorImpl<std::pair<ConstantInt *, Constant *>> &Values, |
6256 | Constant *DefaultValue, const DataLayout &DL, const StringRef &FuncName) { |
6257 | assert(Values.size() && "Can't build lookup table without values!" ); |
6258 | assert(TableSize >= Values.size() && "Can't fit values in table!" ); |
6259 | |
6260 | // If all values in the table are equal, this is that value. |
6261 | SingleValue = Values.begin()->second; |
6262 | |
6263 | Type *ValueType = Values.begin()->second->getType(); |
6264 | |
6265 | // Build up the table contents. |
6266 | SmallVector<Constant *, 64> TableContents(TableSize); |
6267 | for (size_t I = 0, E = Values.size(); I != E; ++I) { |
6268 | ConstantInt *CaseVal = Values[I].first; |
6269 | Constant *CaseRes = Values[I].second; |
6270 | assert(CaseRes->getType() == ValueType); |
6271 | |
6272 | uint64_t Idx = (CaseVal->getValue() - Offset->getValue()).getLimitedValue(); |
6273 | TableContents[Idx] = CaseRes; |
6274 | |
6275 | if (CaseRes != SingleValue) |
6276 | SingleValue = nullptr; |
6277 | } |
6278 | |
6279 | // Fill in any holes in the table with the default result. |
6280 | if (Values.size() < TableSize) { |
6281 | assert(DefaultValue && |
6282 | "Need a default value to fill the lookup table holes." ); |
6283 | assert(DefaultValue->getType() == ValueType); |
6284 | for (uint64_t I = 0; I < TableSize; ++I) { |
6285 | if (!TableContents[I]) |
6286 | TableContents[I] = DefaultValue; |
6287 | } |
6288 | |
6289 | if (DefaultValue != SingleValue) |
6290 | SingleValue = nullptr; |
6291 | } |
6292 | |
6293 | // If each element in the table contains the same value, we only need to store |
6294 | // that single value. |
6295 | if (SingleValue) { |
6296 | Kind = SingleValueKind; |
6297 | return; |
6298 | } |
6299 | |
6300 | // Check if we can derive the value with a linear transformation from the |
6301 | // table index. |
6302 | if (isa<IntegerType>(Val: ValueType)) { |
6303 | bool LinearMappingPossible = true; |
6304 | APInt PrevVal; |
6305 | APInt DistToPrev; |
6306 | // When linear map is monotonic and signed overflow doesn't happen on |
6307 | // maximum index, we can attach nsw on Add and Mul. |
6308 | bool NonMonotonic = false; |
6309 | assert(TableSize >= 2 && "Should be a SingleValue table." ); |
6310 | // Check if there is the same distance between two consecutive values. |
6311 | for (uint64_t I = 0; I < TableSize; ++I) { |
6312 | ConstantInt *ConstVal = dyn_cast<ConstantInt>(Val: TableContents[I]); |
6313 | if (!ConstVal) { |
6314 | // This is an undef. We could deal with it, but undefs in lookup tables |
6315 | // are very seldom. It's probably not worth the additional complexity. |
6316 | LinearMappingPossible = false; |
6317 | break; |
6318 | } |
6319 | const APInt &Val = ConstVal->getValue(); |
6320 | if (I != 0) { |
6321 | APInt Dist = Val - PrevVal; |
6322 | if (I == 1) { |
6323 | DistToPrev = Dist; |
6324 | } else if (Dist != DistToPrev) { |
6325 | LinearMappingPossible = false; |
6326 | break; |
6327 | } |
6328 | NonMonotonic |= |
6329 | Dist.isStrictlyPositive() ? Val.sle(RHS: PrevVal) : Val.sgt(RHS: PrevVal); |
6330 | } |
6331 | PrevVal = Val; |
6332 | } |
6333 | if (LinearMappingPossible) { |
6334 | LinearOffset = cast<ConstantInt>(Val: TableContents[0]); |
6335 | LinearMultiplier = ConstantInt::get(Context&: M.getContext(), V: DistToPrev); |
6336 | bool MayWrap = false; |
6337 | APInt M = LinearMultiplier->getValue(); |
6338 | (void)M.smul_ov(RHS: APInt(M.getBitWidth(), TableSize - 1), Overflow&: MayWrap); |
6339 | LinearMapValWrapped = NonMonotonic || MayWrap; |
6340 | Kind = LinearMapKind; |
6341 | ++NumLinearMaps; |
6342 | return; |
6343 | } |
6344 | } |
6345 | |
6346 | // If the type is integer and the table fits in a register, build a bitmap. |
6347 | if (WouldFitInRegister(DL, TableSize, ElementType: ValueType)) { |
6348 | IntegerType *IT = cast<IntegerType>(Val: ValueType); |
6349 | APInt TableInt(TableSize * IT->getBitWidth(), 0); |
6350 | for (uint64_t I = TableSize; I > 0; --I) { |
6351 | TableInt <<= IT->getBitWidth(); |
6352 | // Insert values into the bitmap. Undef values are set to zero. |
6353 | if (!isa<UndefValue>(Val: TableContents[I - 1])) { |
6354 | ConstantInt *Val = cast<ConstantInt>(Val: TableContents[I - 1]); |
6355 | TableInt |= Val->getValue().zext(width: TableInt.getBitWidth()); |
6356 | } |
6357 | } |
6358 | BitMap = ConstantInt::get(Context&: M.getContext(), V: TableInt); |
6359 | BitMapElementTy = IT; |
6360 | Kind = BitMapKind; |
6361 | ++NumBitMaps; |
6362 | return; |
6363 | } |
6364 | |
6365 | // Store the table in an array. |
6366 | ArrayType *ArrayTy = ArrayType::get(ElementType: ValueType, NumElements: TableSize); |
6367 | Constant *Initializer = ConstantArray::get(T: ArrayTy, V: TableContents); |
6368 | |
6369 | Array = new GlobalVariable(M, ArrayTy, /*isConstant=*/true, |
6370 | GlobalVariable::PrivateLinkage, Initializer, |
6371 | "switch.table." + FuncName); |
6372 | Array->setUnnamedAddr(GlobalValue::UnnamedAddr::Global); |
6373 | // Set the alignment to that of an array items. We will be only loading one |
6374 | // value out of it. |
6375 | Array->setAlignment(DL.getPrefTypeAlign(Ty: ValueType)); |
6376 | Kind = ArrayKind; |
6377 | } |
6378 | |
6379 | Value *SwitchLookupTable::BuildLookup(Value *Index, IRBuilder<> &Builder) { |
6380 | switch (Kind) { |
6381 | case SingleValueKind: |
6382 | return SingleValue; |
6383 | case LinearMapKind: { |
6384 | // Derive the result value from the input value. |
6385 | Value *Result = Builder.CreateIntCast(V: Index, DestTy: LinearMultiplier->getType(), |
6386 | isSigned: false, Name: "switch.idx.cast" ); |
6387 | if (!LinearMultiplier->isOne()) |
6388 | Result = Builder.CreateMul(LHS: Result, RHS: LinearMultiplier, Name: "switch.idx.mult" , |
6389 | /*HasNUW = */ false, |
6390 | /*HasNSW = */ !LinearMapValWrapped); |
6391 | |
6392 | if (!LinearOffset->isZero()) |
6393 | Result = Builder.CreateAdd(LHS: Result, RHS: LinearOffset, Name: "switch.offset" , |
6394 | /*HasNUW = */ false, |
6395 | /*HasNSW = */ !LinearMapValWrapped); |
6396 | return Result; |
6397 | } |
6398 | case BitMapKind: { |
6399 | // Type of the bitmap (e.g. i59). |
6400 | IntegerType *MapTy = BitMap->getIntegerType(); |
6401 | |
6402 | // Cast Index to the same type as the bitmap. |
6403 | // Note: The Index is <= the number of elements in the table, so |
6404 | // truncating it to the width of the bitmask is safe. |
6405 | Value *ShiftAmt = Builder.CreateZExtOrTrunc(V: Index, DestTy: MapTy, Name: "switch.cast" ); |
6406 | |
6407 | // Multiply the shift amount by the element width. NUW/NSW can always be |
6408 | // set, because WouldFitInRegister guarantees Index * ShiftAmt is in |
6409 | // BitMap's bit width. |
6410 | ShiftAmt = Builder.CreateMul( |
6411 | LHS: ShiftAmt, RHS: ConstantInt::get(Ty: MapTy, V: BitMapElementTy->getBitWidth()), |
6412 | Name: "switch.shiftamt" ,/*HasNUW =*/true,/*HasNSW =*/true); |
6413 | |
6414 | // Shift down. |
6415 | Value *DownShifted = |
6416 | Builder.CreateLShr(LHS: BitMap, RHS: ShiftAmt, Name: "switch.downshift" ); |
6417 | // Mask off. |
6418 | return Builder.CreateTrunc(V: DownShifted, DestTy: BitMapElementTy, Name: "switch.masked" ); |
6419 | } |
6420 | case ArrayKind: { |
6421 | // Make sure the table index will not overflow when treated as signed. |
6422 | IntegerType *IT = cast<IntegerType>(Val: Index->getType()); |
6423 | uint64_t TableSize = |
6424 | Array->getInitializer()->getType()->getArrayNumElements(); |
6425 | if (TableSize > (1ULL << std::min(a: IT->getBitWidth() - 1, b: 63u))) |
6426 | Index = Builder.CreateZExt( |
6427 | V: Index, DestTy: IntegerType::get(C&: IT->getContext(), NumBits: IT->getBitWidth() + 1), |
6428 | Name: "switch.tableidx.zext" ); |
6429 | |
6430 | Value *GEPIndices[] = {Builder.getInt32(C: 0), Index}; |
6431 | Value *GEP = Builder.CreateInBoundsGEP(Ty: Array->getValueType(), Ptr: Array, |
6432 | IdxList: GEPIndices, Name: "switch.gep" ); |
6433 | return Builder.CreateLoad( |
6434 | Ty: cast<ArrayType>(Val: Array->getValueType())->getElementType(), Ptr: GEP, |
6435 | Name: "switch.load" ); |
6436 | } |
6437 | } |
6438 | llvm_unreachable("Unknown lookup table kind!" ); |
6439 | } |
6440 | |
6441 | bool SwitchLookupTable::WouldFitInRegister(const DataLayout &DL, |
6442 | uint64_t TableSize, |
6443 | Type *ElementType) { |
6444 | auto *IT = dyn_cast<IntegerType>(Val: ElementType); |
6445 | if (!IT) |
6446 | return false; |
6447 | // FIXME: If the type is wider than it needs to be, e.g. i8 but all values |
6448 | // are <= 15, we could try to narrow the type. |
6449 | |
6450 | // Avoid overflow, fitsInLegalInteger uses unsigned int for the width. |
6451 | if (TableSize >= UINT_MAX / IT->getBitWidth()) |
6452 | return false; |
6453 | return DL.fitsInLegalInteger(Width: TableSize * IT->getBitWidth()); |
6454 | } |
6455 | |
6456 | static bool isTypeLegalForLookupTable(Type *Ty, const TargetTransformInfo &TTI, |
6457 | const DataLayout &DL) { |
6458 | // Allow any legal type. |
6459 | if (TTI.isTypeLegal(Ty)) |
6460 | return true; |
6461 | |
6462 | auto *IT = dyn_cast<IntegerType>(Val: Ty); |
6463 | if (!IT) |
6464 | return false; |
6465 | |
6466 | // Also allow power of 2 integer types that have at least 8 bits and fit in |
6467 | // a register. These types are common in frontend languages and targets |
6468 | // usually support loads of these types. |
6469 | // TODO: We could relax this to any integer that fits in a register and rely |
6470 | // on ABI alignment and padding in the table to allow the load to be widened. |
6471 | // Or we could widen the constants and truncate the load. |
6472 | unsigned BitWidth = IT->getBitWidth(); |
6473 | return BitWidth >= 8 && isPowerOf2_32(Value: BitWidth) && |
6474 | DL.fitsInLegalInteger(Width: IT->getBitWidth()); |
6475 | } |
6476 | |
6477 | static bool isSwitchDense(uint64_t NumCases, uint64_t CaseRange) { |
6478 | // 40% is the default density for building a jump table in optsize/minsize |
6479 | // mode. See also TargetLoweringBase::isSuitableForJumpTable(), which this |
6480 | // function was based on. |
6481 | const uint64_t MinDensity = 40; |
6482 | |
6483 | if (CaseRange >= UINT64_MAX / 100) |
6484 | return false; // Avoid multiplication overflows below. |
6485 | |
6486 | return NumCases * 100 >= CaseRange * MinDensity; |
6487 | } |
6488 | |
6489 | static bool isSwitchDense(ArrayRef<int64_t> Values) { |
6490 | uint64_t Diff = (uint64_t)Values.back() - (uint64_t)Values.front(); |
6491 | uint64_t Range = Diff + 1; |
6492 | if (Range < Diff) |
6493 | return false; // Overflow. |
6494 | |
6495 | return isSwitchDense(NumCases: Values.size(), CaseRange: Range); |
6496 | } |
6497 | |
6498 | /// Determine whether a lookup table should be built for this switch, based on |
6499 | /// the number of cases, size of the table, and the types of the results. |
6500 | // TODO: We could support larger than legal types by limiting based on the |
6501 | // number of loads required and/or table size. If the constants are small we |
6502 | // could use smaller table entries and extend after the load. |
6503 | static bool |
6504 | ShouldBuildLookupTable(SwitchInst *SI, uint64_t TableSize, |
6505 | const TargetTransformInfo &TTI, const DataLayout &DL, |
6506 | const SmallDenseMap<PHINode *, Type *> &ResultTypes) { |
6507 | if (SI->getNumCases() > TableSize) |
6508 | return false; // TableSize overflowed. |
6509 | |
6510 | bool AllTablesFitInRegister = true; |
6511 | bool HasIllegalType = false; |
6512 | for (const auto &I : ResultTypes) { |
6513 | Type *Ty = I.second; |
6514 | |
6515 | // Saturate this flag to true. |
6516 | HasIllegalType = HasIllegalType || !isTypeLegalForLookupTable(Ty, TTI, DL); |
6517 | |
6518 | // Saturate this flag to false. |
6519 | AllTablesFitInRegister = |
6520 | AllTablesFitInRegister && |
6521 | SwitchLookupTable::WouldFitInRegister(DL, TableSize, ElementType: Ty); |
6522 | |
6523 | // If both flags saturate, we're done. NOTE: This *only* works with |
6524 | // saturating flags, and all flags have to saturate first due to the |
6525 | // non-deterministic behavior of iterating over a dense map. |
6526 | if (HasIllegalType && !AllTablesFitInRegister) |
6527 | break; |
6528 | } |
6529 | |
6530 | // If each table would fit in a register, we should build it anyway. |
6531 | if (AllTablesFitInRegister) |
6532 | return true; |
6533 | |
6534 | // Don't build a table that doesn't fit in-register if it has illegal types. |
6535 | if (HasIllegalType) |
6536 | return false; |
6537 | |
6538 | return isSwitchDense(NumCases: SI->getNumCases(), CaseRange: TableSize); |
6539 | } |
6540 | |
6541 | static bool ShouldUseSwitchConditionAsTableIndex( |
6542 | ConstantInt &MinCaseVal, const ConstantInt &MaxCaseVal, |
6543 | bool HasDefaultResults, const SmallDenseMap<PHINode *, Type *> &ResultTypes, |
6544 | const DataLayout &DL, const TargetTransformInfo &TTI) { |
6545 | if (MinCaseVal.isNullValue()) |
6546 | return true; |
6547 | if (MinCaseVal.isNegative() || |
6548 | MaxCaseVal.getLimitedValue() == std::numeric_limits<uint64_t>::max() || |
6549 | !HasDefaultResults) |
6550 | return false; |
6551 | return all_of(Range: ResultTypes, P: [&](const auto &KV) { |
6552 | return SwitchLookupTable::WouldFitInRegister( |
6553 | DL, TableSize: MaxCaseVal.getLimitedValue() + 1 /* TableSize */, |
6554 | ElementType: KV.second /* ResultType */); |
6555 | }); |
6556 | } |
6557 | |
6558 | /// Try to reuse the switch table index compare. Following pattern: |
6559 | /// \code |
6560 | /// if (idx < tablesize) |
6561 | /// r = table[idx]; // table does not contain default_value |
6562 | /// else |
6563 | /// r = default_value; |
6564 | /// if (r != default_value) |
6565 | /// ... |
6566 | /// \endcode |
6567 | /// Is optimized to: |
6568 | /// \code |
6569 | /// cond = idx < tablesize; |
6570 | /// if (cond) |
6571 | /// r = table[idx]; |
6572 | /// else |
6573 | /// r = default_value; |
6574 | /// if (cond) |
6575 | /// ... |
6576 | /// \endcode |
6577 | /// Jump threading will then eliminate the second if(cond). |
6578 | static void reuseTableCompare( |
6579 | User *PhiUser, BasicBlock *PhiBlock, BranchInst *RangeCheckBranch, |
6580 | Constant *DefaultValue, |
6581 | const SmallVectorImpl<std::pair<ConstantInt *, Constant *>> &Values) { |
6582 | ICmpInst *CmpInst = dyn_cast<ICmpInst>(Val: PhiUser); |
6583 | if (!CmpInst) |
6584 | return; |
6585 | |
6586 | // We require that the compare is in the same block as the phi so that jump |
6587 | // threading can do its work afterwards. |
6588 | if (CmpInst->getParent() != PhiBlock) |
6589 | return; |
6590 | |
6591 | Constant *CmpOp1 = dyn_cast<Constant>(Val: CmpInst->getOperand(i_nocapture: 1)); |
6592 | if (!CmpOp1) |
6593 | return; |
6594 | |
6595 | Value *RangeCmp = RangeCheckBranch->getCondition(); |
6596 | Constant *TrueConst = ConstantInt::getTrue(Ty: RangeCmp->getType()); |
6597 | Constant *FalseConst = ConstantInt::getFalse(Ty: RangeCmp->getType()); |
6598 | |
6599 | // Check if the compare with the default value is constant true or false. |
6600 | const DataLayout &DL = PhiBlock->getDataLayout(); |
6601 | Constant *DefaultConst = ConstantFoldCompareInstOperands( |
6602 | Predicate: CmpInst->getPredicate(), LHS: DefaultValue, RHS: CmpOp1, DL); |
6603 | if (DefaultConst != TrueConst && DefaultConst != FalseConst) |
6604 | return; |
6605 | |
6606 | // Check if the compare with the case values is distinct from the default |
6607 | // compare result. |
6608 | for (auto ValuePair : Values) { |
6609 | Constant *CaseConst = ConstantFoldCompareInstOperands( |
6610 | Predicate: CmpInst->getPredicate(), LHS: ValuePair.second, RHS: CmpOp1, DL); |
6611 | if (!CaseConst || CaseConst == DefaultConst || |
6612 | (CaseConst != TrueConst && CaseConst != FalseConst)) |
6613 | return; |
6614 | } |
6615 | |
6616 | // Check if the branch instruction dominates the phi node. It's a simple |
6617 | // dominance check, but sufficient for our needs. |
6618 | // Although this check is invariant in the calling loops, it's better to do it |
6619 | // at this late stage. Practically we do it at most once for a switch. |
6620 | BasicBlock *BranchBlock = RangeCheckBranch->getParent(); |
6621 | for (BasicBlock *Pred : predecessors(BB: PhiBlock)) { |
6622 | if (Pred != BranchBlock && Pred->getUniquePredecessor() != BranchBlock) |
6623 | return; |
6624 | } |
6625 | |
6626 | if (DefaultConst == FalseConst) { |
6627 | // The compare yields the same result. We can replace it. |
6628 | CmpInst->replaceAllUsesWith(V: RangeCmp); |
6629 | ++NumTableCmpReuses; |
6630 | } else { |
6631 | // The compare yields the same result, just inverted. We can replace it. |
6632 | Value *InvertedTableCmp = BinaryOperator::CreateXor( |
6633 | V1: RangeCmp, V2: ConstantInt::get(Ty: RangeCmp->getType(), V: 1), Name: "inverted.cmp" , |
6634 | It: RangeCheckBranch->getIterator()); |
6635 | CmpInst->replaceAllUsesWith(V: InvertedTableCmp); |
6636 | ++NumTableCmpReuses; |
6637 | } |
6638 | } |
6639 | |
6640 | /// If the switch is only used to initialize one or more phi nodes in a common |
6641 | /// successor block with different constant values, replace the switch with |
6642 | /// lookup tables. |
6643 | static bool SwitchToLookupTable(SwitchInst *SI, IRBuilder<> &Builder, |
6644 | DomTreeUpdater *DTU, const DataLayout &DL, |
6645 | const TargetTransformInfo &TTI) { |
6646 | assert(SI->getNumCases() > 1 && "Degenerate switch?" ); |
6647 | |
6648 | BasicBlock *BB = SI->getParent(); |
6649 | Function *Fn = BB->getParent(); |
6650 | // Only build lookup table when we have a target that supports it or the |
6651 | // attribute is not set. |
6652 | if (!TTI.shouldBuildLookupTables() || |
6653 | (Fn->getFnAttribute(Kind: "no-jump-tables" ).getValueAsBool())) |
6654 | return false; |
6655 | |
6656 | // FIXME: If the switch is too sparse for a lookup table, perhaps we could |
6657 | // split off a dense part and build a lookup table for that. |
6658 | |
6659 | // FIXME: This creates arrays of GEPs to constant strings, which means each |
6660 | // GEP needs a runtime relocation in PIC code. We should just build one big |
6661 | // string and lookup indices into that. |
6662 | |
6663 | // Ignore switches with less than three cases. Lookup tables will not make |
6664 | // them faster, so we don't analyze them. |
6665 | if (SI->getNumCases() < 3) |
6666 | return false; |
6667 | |
6668 | // Figure out the corresponding result for each case value and phi node in the |
6669 | // common destination, as well as the min and max case values. |
6670 | assert(!SI->cases().empty()); |
6671 | SwitchInst::CaseIt CI = SI->case_begin(); |
6672 | ConstantInt *MinCaseVal = CI->getCaseValue(); |
6673 | ConstantInt *MaxCaseVal = CI->getCaseValue(); |
6674 | |
6675 | BasicBlock *CommonDest = nullptr; |
6676 | |
6677 | using ResultListTy = SmallVector<std::pair<ConstantInt *, Constant *>, 4>; |
6678 | SmallDenseMap<PHINode *, ResultListTy> ResultLists; |
6679 | |
6680 | SmallDenseMap<PHINode *, Constant *> DefaultResults; |
6681 | SmallDenseMap<PHINode *, Type *> ResultTypes; |
6682 | SmallVector<PHINode *, 4> PHIs; |
6683 | |
6684 | for (SwitchInst::CaseIt E = SI->case_end(); CI != E; ++CI) { |
6685 | ConstantInt *CaseVal = CI->getCaseValue(); |
6686 | if (CaseVal->getValue().slt(RHS: MinCaseVal->getValue())) |
6687 | MinCaseVal = CaseVal; |
6688 | if (CaseVal->getValue().sgt(RHS: MaxCaseVal->getValue())) |
6689 | MaxCaseVal = CaseVal; |
6690 | |
6691 | // Resulting value at phi nodes for this case value. |
6692 | using ResultsTy = SmallVector<std::pair<PHINode *, Constant *>, 4>; |
6693 | ResultsTy Results; |
6694 | if (!getCaseResults(SI, CaseVal, CaseDest: CI->getCaseSuccessor(), CommonDest: &CommonDest, |
6695 | Res&: Results, DL, TTI)) |
6696 | return false; |
6697 | |
6698 | // Append the result from this case to the list for each phi. |
6699 | for (const auto &I : Results) { |
6700 | PHINode *PHI = I.first; |
6701 | Constant *Value = I.second; |
6702 | if (!ResultLists.count(Val: PHI)) |
6703 | PHIs.push_back(Elt: PHI); |
6704 | ResultLists[PHI].push_back(Elt: std::make_pair(x&: CaseVal, y&: Value)); |
6705 | } |
6706 | } |
6707 | |
6708 | // Keep track of the result types. |
6709 | for (PHINode *PHI : PHIs) { |
6710 | ResultTypes[PHI] = ResultLists[PHI][0].second->getType(); |
6711 | } |
6712 | |
6713 | uint64_t NumResults = ResultLists[PHIs[0]].size(); |
6714 | |
6715 | // If the table has holes, we need a constant result for the default case |
6716 | // or a bitmask that fits in a register. |
6717 | SmallVector<std::pair<PHINode *, Constant *>, 4> DefaultResultsList; |
6718 | bool HasDefaultResults = |
6719 | getCaseResults(SI, CaseVal: nullptr, CaseDest: SI->getDefaultDest(), CommonDest: &CommonDest, |
6720 | Res&: DefaultResultsList, DL, TTI); |
6721 | |
6722 | for (const auto &I : DefaultResultsList) { |
6723 | PHINode *PHI = I.first; |
6724 | Constant *Result = I.second; |
6725 | DefaultResults[PHI] = Result; |
6726 | } |
6727 | |
6728 | bool UseSwitchConditionAsTableIndex = ShouldUseSwitchConditionAsTableIndex( |
6729 | MinCaseVal&: *MinCaseVal, MaxCaseVal: *MaxCaseVal, HasDefaultResults, ResultTypes, DL, TTI); |
6730 | uint64_t TableSize; |
6731 | if (UseSwitchConditionAsTableIndex) |
6732 | TableSize = MaxCaseVal->getLimitedValue() + 1; |
6733 | else |
6734 | TableSize = |
6735 | (MaxCaseVal->getValue() - MinCaseVal->getValue()).getLimitedValue() + 1; |
6736 | |
6737 | // If the default destination is unreachable, or if the lookup table covers |
6738 | // all values of the conditional variable, branch directly to the lookup table |
6739 | // BB. Otherwise, check that the condition is within the case range. |
6740 | bool DefaultIsReachable = !SI->defaultDestUndefined(); |
6741 | |
6742 | bool TableHasHoles = (NumResults < TableSize); |
6743 | |
6744 | // If the table has holes but the default destination doesn't produce any |
6745 | // constant results, the lookup table entries corresponding to the holes will |
6746 | // contain undefined values. |
6747 | bool AllHolesAreUndefined = TableHasHoles && !HasDefaultResults; |
6748 | |
6749 | // If the default destination doesn't produce a constant result but is still |
6750 | // reachable, and the lookup table has holes, we need to use a mask to |
6751 | // determine if the current index should load from the lookup table or jump |
6752 | // to the default case. |
6753 | // The mask is unnecessary if the table has holes but the default destination |
6754 | // is unreachable, as in that case the holes must also be unreachable. |
6755 | bool NeedMask = AllHolesAreUndefined && DefaultIsReachable; |
6756 | if (NeedMask) { |
6757 | // As an extra penalty for the validity test we require more cases. |
6758 | if (SI->getNumCases() < 4) // FIXME: Find best threshold value (benchmark). |
6759 | return false; |
6760 | if (!DL.fitsInLegalInteger(Width: TableSize)) |
6761 | return false; |
6762 | } |
6763 | |
6764 | if (!ShouldBuildLookupTable(SI, TableSize, TTI, DL, ResultTypes)) |
6765 | return false; |
6766 | |
6767 | std::vector<DominatorTree::UpdateType> Updates; |
6768 | |
6769 | // Compute the maximum table size representable by the integer type we are |
6770 | // switching upon. |
6771 | unsigned CaseSize = MinCaseVal->getType()->getPrimitiveSizeInBits(); |
6772 | uint64_t MaxTableSize = CaseSize > 63 ? UINT64_MAX : 1ULL << CaseSize; |
6773 | assert(MaxTableSize >= TableSize && |
6774 | "It is impossible for a switch to have more entries than the max " |
6775 | "representable value of its input integer type's size." ); |
6776 | |
6777 | // Create the BB that does the lookups. |
6778 | Module &Mod = *CommonDest->getParent()->getParent(); |
6779 | BasicBlock *LookupBB = BasicBlock::Create( |
6780 | Context&: Mod.getContext(), Name: "switch.lookup" , Parent: CommonDest->getParent(), InsertBefore: CommonDest); |
6781 | |
6782 | // Compute the table index value. |
6783 | Builder.SetInsertPoint(SI); |
6784 | Value *TableIndex; |
6785 | ConstantInt *TableIndexOffset; |
6786 | if (UseSwitchConditionAsTableIndex) { |
6787 | TableIndexOffset = ConstantInt::get(Ty: MaxCaseVal->getIntegerType(), V: 0); |
6788 | TableIndex = SI->getCondition(); |
6789 | } else { |
6790 | TableIndexOffset = MinCaseVal; |
6791 | // If the default is unreachable, all case values are s>= MinCaseVal. Then |
6792 | // we can try to attach nsw. |
6793 | bool MayWrap = true; |
6794 | if (!DefaultIsReachable) { |
6795 | APInt Res = MaxCaseVal->getValue().ssub_ov(RHS: MinCaseVal->getValue(), Overflow&: MayWrap); |
6796 | (void)Res; |
6797 | } |
6798 | |
6799 | TableIndex = Builder.CreateSub(LHS: SI->getCondition(), RHS: TableIndexOffset, |
6800 | Name: "switch.tableidx" , /*HasNUW =*/false, |
6801 | /*HasNSW =*/!MayWrap); |
6802 | } |
6803 | |
6804 | BranchInst *RangeCheckBranch = nullptr; |
6805 | |
6806 | // Grow the table to cover all possible index values to avoid the range check. |
6807 | // It will use the default result to fill in the table hole later, so make |
6808 | // sure it exist. |
6809 | if (UseSwitchConditionAsTableIndex && HasDefaultResults) { |
6810 | ConstantRange CR = computeConstantRange(V: TableIndex, /* ForSigned */ false); |
6811 | // Grow the table shouldn't have any size impact by checking |
6812 | // WouldFitInRegister. |
6813 | // TODO: Consider growing the table also when it doesn't fit in a register |
6814 | // if no optsize is specified. |
6815 | const uint64_t UpperBound = CR.getUpper().getLimitedValue(); |
6816 | if (!CR.isUpperWrapped() && all_of(Range&: ResultTypes, P: [&](const auto &KV) { |
6817 | return SwitchLookupTable::WouldFitInRegister( |
6818 | DL, TableSize: UpperBound, ElementType: KV.second /* ResultType */); |
6819 | })) { |
6820 | // There may be some case index larger than the UpperBound (unreachable |
6821 | // case), so make sure the table size does not get smaller. |
6822 | TableSize = std::max(a: UpperBound, b: TableSize); |
6823 | // The default branch is unreachable after we enlarge the lookup table. |
6824 | // Adjust DefaultIsReachable to reuse code path. |
6825 | DefaultIsReachable = false; |
6826 | } |
6827 | } |
6828 | |
6829 | const bool GeneratingCoveredLookupTable = (MaxTableSize == TableSize); |
6830 | if (!DefaultIsReachable || GeneratingCoveredLookupTable) { |
6831 | Builder.CreateBr(Dest: LookupBB); |
6832 | if (DTU) |
6833 | Updates.push_back(x: {DominatorTree::Insert, BB, LookupBB}); |
6834 | // Note: We call removeProdecessor later since we need to be able to get the |
6835 | // PHI value for the default case in case we're using a bit mask. |
6836 | } else { |
6837 | Value *Cmp = Builder.CreateICmpULT( |
6838 | LHS: TableIndex, RHS: ConstantInt::get(Ty: MinCaseVal->getType(), V: TableSize)); |
6839 | RangeCheckBranch = |
6840 | Builder.CreateCondBr(Cond: Cmp, True: LookupBB, False: SI->getDefaultDest()); |
6841 | if (DTU) |
6842 | Updates.push_back(x: {DominatorTree::Insert, BB, LookupBB}); |
6843 | } |
6844 | |
6845 | // Populate the BB that does the lookups. |
6846 | Builder.SetInsertPoint(LookupBB); |
6847 | |
6848 | if (NeedMask) { |
6849 | // Before doing the lookup, we do the hole check. The LookupBB is therefore |
6850 | // re-purposed to do the hole check, and we create a new LookupBB. |
6851 | BasicBlock *MaskBB = LookupBB; |
6852 | MaskBB->setName("switch.hole_check" ); |
6853 | LookupBB = BasicBlock::Create(Context&: Mod.getContext(), Name: "switch.lookup" , |
6854 | Parent: CommonDest->getParent(), InsertBefore: CommonDest); |
6855 | |
6856 | // Make the mask's bitwidth at least 8-bit and a power-of-2 to avoid |
6857 | // unnecessary illegal types. |
6858 | uint64_t TableSizePowOf2 = NextPowerOf2(A: std::max(a: 7ULL, b: TableSize - 1ULL)); |
6859 | APInt MaskInt(TableSizePowOf2, 0); |
6860 | APInt One(TableSizePowOf2, 1); |
6861 | // Build bitmask; fill in a 1 bit for every case. |
6862 | const ResultListTy &ResultList = ResultLists[PHIs[0]]; |
6863 | for (size_t I = 0, E = ResultList.size(); I != E; ++I) { |
6864 | uint64_t Idx = (ResultList[I].first->getValue() - TableIndexOffset->getValue()) |
6865 | .getLimitedValue(); |
6866 | MaskInt |= One << Idx; |
6867 | } |
6868 | ConstantInt *TableMask = ConstantInt::get(Context&: Mod.getContext(), V: MaskInt); |
6869 | |
6870 | // Get the TableIndex'th bit of the bitmask. |
6871 | // If this bit is 0 (meaning hole) jump to the default destination, |
6872 | // else continue with table lookup. |
6873 | IntegerType *MapTy = TableMask->getIntegerType(); |
6874 | Value *MaskIndex = |
6875 | Builder.CreateZExtOrTrunc(V: TableIndex, DestTy: MapTy, Name: "switch.maskindex" ); |
6876 | Value *Shifted = Builder.CreateLShr(LHS: TableMask, RHS: MaskIndex, Name: "switch.shifted" ); |
6877 | Value *LoBit = Builder.CreateTrunc( |
6878 | V: Shifted, DestTy: Type::getInt1Ty(C&: Mod.getContext()), Name: "switch.lobit" ); |
6879 | Builder.CreateCondBr(Cond: LoBit, True: LookupBB, False: SI->getDefaultDest()); |
6880 | if (DTU) { |
6881 | Updates.push_back(x: {DominatorTree::Insert, MaskBB, LookupBB}); |
6882 | Updates.push_back(x: {DominatorTree::Insert, MaskBB, SI->getDefaultDest()}); |
6883 | } |
6884 | Builder.SetInsertPoint(LookupBB); |
6885 | AddPredecessorToBlock(Succ: SI->getDefaultDest(), NewPred: MaskBB, ExistPred: BB); |
6886 | } |
6887 | |
6888 | if (!DefaultIsReachable || GeneratingCoveredLookupTable) { |
6889 | // We cached PHINodes in PHIs. To avoid accessing deleted PHINodes later, |
6890 | // do not delete PHINodes here. |
6891 | SI->getDefaultDest()->removePredecessor(Pred: BB, |
6892 | /*KeepOneInputPHIs=*/true); |
6893 | if (DTU) |
6894 | Updates.push_back(x: {DominatorTree::Delete, BB, SI->getDefaultDest()}); |
6895 | } |
6896 | |
6897 | for (PHINode *PHI : PHIs) { |
6898 | const ResultListTy &ResultList = ResultLists[PHI]; |
6899 | |
6900 | // Use any value to fill the lookup table holes. |
6901 | Constant *DV = |
6902 | AllHolesAreUndefined ? ResultLists[PHI][0].second : DefaultResults[PHI]; |
6903 | StringRef FuncName = Fn->getName(); |
6904 | SwitchLookupTable Table(Mod, TableSize, TableIndexOffset, ResultList, DV, |
6905 | DL, FuncName); |
6906 | |
6907 | Value *Result = Table.BuildLookup(Index: TableIndex, Builder); |
6908 | |
6909 | // Do a small peephole optimization: re-use the switch table compare if |
6910 | // possible. |
6911 | if (!TableHasHoles && HasDefaultResults && RangeCheckBranch) { |
6912 | BasicBlock *PhiBlock = PHI->getParent(); |
6913 | // Search for compare instructions which use the phi. |
6914 | for (auto *User : PHI->users()) { |
6915 | reuseTableCompare(PhiUser: User, PhiBlock, RangeCheckBranch, DefaultValue: DV, Values: ResultList); |
6916 | } |
6917 | } |
6918 | |
6919 | PHI->addIncoming(V: Result, BB: LookupBB); |
6920 | } |
6921 | |
6922 | Builder.CreateBr(Dest: CommonDest); |
6923 | if (DTU) |
6924 | Updates.push_back(x: {DominatorTree::Insert, LookupBB, CommonDest}); |
6925 | |
6926 | // Remove the switch. |
6927 | SmallPtrSet<BasicBlock *, 8> RemovedSuccessors; |
6928 | for (unsigned i = 0, e = SI->getNumSuccessors(); i < e; ++i) { |
6929 | BasicBlock *Succ = SI->getSuccessor(idx: i); |
6930 | |
6931 | if (Succ == SI->getDefaultDest()) |
6932 | continue; |
6933 | Succ->removePredecessor(Pred: BB); |
6934 | if (DTU && RemovedSuccessors.insert(Ptr: Succ).second) |
6935 | Updates.push_back(x: {DominatorTree::Delete, BB, Succ}); |
6936 | } |
6937 | SI->eraseFromParent(); |
6938 | |
6939 | if (DTU) |
6940 | DTU->applyUpdates(Updates); |
6941 | |
6942 | ++NumLookupTables; |
6943 | if (NeedMask) |
6944 | ++NumLookupTablesHoles; |
6945 | return true; |
6946 | } |
6947 | |
6948 | /// Try to transform a switch that has "holes" in it to a contiguous sequence |
6949 | /// of cases. |
6950 | /// |
6951 | /// A switch such as: switch(i) {case 5: case 9: case 13: case 17:} can be |
6952 | /// range-reduced to: switch ((i-5) / 4) {case 0: case 1: case 2: case 3:}. |
6953 | /// |
6954 | /// This converts a sparse switch into a dense switch which allows better |
6955 | /// lowering and could also allow transforming into a lookup table. |
6956 | static bool ReduceSwitchRange(SwitchInst *SI, IRBuilder<> &Builder, |
6957 | const DataLayout &DL, |
6958 | const TargetTransformInfo &TTI) { |
6959 | auto *CondTy = cast<IntegerType>(Val: SI->getCondition()->getType()); |
6960 | if (CondTy->getIntegerBitWidth() > 64 || |
6961 | !DL.fitsInLegalInteger(Width: CondTy->getIntegerBitWidth())) |
6962 | return false; |
6963 | // Only bother with this optimization if there are more than 3 switch cases; |
6964 | // SDAG will only bother creating jump tables for 4 or more cases. |
6965 | if (SI->getNumCases() < 4) |
6966 | return false; |
6967 | |
6968 | // This transform is agnostic to the signedness of the input or case values. We |
6969 | // can treat the case values as signed or unsigned. We can optimize more common |
6970 | // cases such as a sequence crossing zero {-4,0,4,8} if we interpret case values |
6971 | // as signed. |
6972 | SmallVector<int64_t,4> Values; |
6973 | for (const auto &C : SI->cases()) |
6974 | Values.push_back(Elt: C.getCaseValue()->getValue().getSExtValue()); |
6975 | llvm::sort(C&: Values); |
6976 | |
6977 | // If the switch is already dense, there's nothing useful to do here. |
6978 | if (isSwitchDense(Values)) |
6979 | return false; |
6980 | |
6981 | // First, transform the values such that they start at zero and ascend. |
6982 | int64_t Base = Values[0]; |
6983 | for (auto &V : Values) |
6984 | V -= (uint64_t)(Base); |
6985 | |
6986 | // Now we have signed numbers that have been shifted so that, given enough |
6987 | // precision, there are no negative values. Since the rest of the transform |
6988 | // is bitwise only, we switch now to an unsigned representation. |
6989 | |
6990 | // This transform can be done speculatively because it is so cheap - it |
6991 | // results in a single rotate operation being inserted. |
6992 | |
6993 | // countTrailingZeros(0) returns 64. As Values is guaranteed to have more than |
6994 | // one element and LLVM disallows duplicate cases, Shift is guaranteed to be |
6995 | // less than 64. |
6996 | unsigned Shift = 64; |
6997 | for (auto &V : Values) |
6998 | Shift = std::min(a: Shift, b: (unsigned)llvm::countr_zero(Val: (uint64_t)V)); |
6999 | assert(Shift < 64); |
7000 | if (Shift > 0) |
7001 | for (auto &V : Values) |
7002 | V = (int64_t)((uint64_t)V >> Shift); |
7003 | |
7004 | if (!isSwitchDense(Values)) |
7005 | // Transform didn't create a dense switch. |
7006 | return false; |
7007 | |
7008 | // The obvious transform is to shift the switch condition right and emit a |
7009 | // check that the condition actually cleanly divided by GCD, i.e. |
7010 | // C & (1 << Shift - 1) == 0 |
7011 | // inserting a new CFG edge to handle the case where it didn't divide cleanly. |
7012 | // |
7013 | // A cheaper way of doing this is a simple ROTR(C, Shift). This performs the |
7014 | // shift and puts the shifted-off bits in the uppermost bits. If any of these |
7015 | // are nonzero then the switch condition will be very large and will hit the |
7016 | // default case. |
7017 | |
7018 | auto *Ty = cast<IntegerType>(Val: SI->getCondition()->getType()); |
7019 | Builder.SetInsertPoint(SI); |
7020 | Value *Sub = |
7021 | Builder.CreateSub(LHS: SI->getCondition(), RHS: ConstantInt::get(Ty, V: Base)); |
7022 | Value *Rot = Builder.CreateIntrinsic( |
7023 | RetTy: Ty, ID: Intrinsic::fshl, |
7024 | Args: {Sub, Sub, ConstantInt::get(Ty, V: Ty->getBitWidth() - Shift)}); |
7025 | SI->replaceUsesOfWith(From: SI->getCondition(), To: Rot); |
7026 | |
7027 | for (auto Case : SI->cases()) { |
7028 | auto *Orig = Case.getCaseValue(); |
7029 | auto Sub = Orig->getValue() - APInt(Ty->getBitWidth(), Base); |
7030 | Case.setValue(cast<ConstantInt>(Val: ConstantInt::get(Ty, V: Sub.lshr(shiftAmt: Shift)))); |
7031 | } |
7032 | return true; |
7033 | } |
7034 | |
7035 | /// Tries to transform switch of powers of two to reduce switch range. |
7036 | /// For example, switch like: |
7037 | /// switch (C) { case 1: case 2: case 64: case 128: } |
7038 | /// will be transformed to: |
7039 | /// switch (count_trailing_zeros(C)) { case 0: case 1: case 6: case 7: } |
7040 | /// |
7041 | /// This transformation allows better lowering and could allow transforming into |
7042 | /// a lookup table. |
7043 | static bool simplifySwitchOfPowersOfTwo(SwitchInst *SI, IRBuilder<> &Builder, |
7044 | const DataLayout &DL, |
7045 | const TargetTransformInfo &TTI) { |
7046 | Value *Condition = SI->getCondition(); |
7047 | LLVMContext &Context = SI->getContext(); |
7048 | auto *CondTy = cast<IntegerType>(Val: Condition->getType()); |
7049 | |
7050 | if (CondTy->getIntegerBitWidth() > 64 || |
7051 | !DL.fitsInLegalInteger(Width: CondTy->getIntegerBitWidth())) |
7052 | return false; |
7053 | |
7054 | const auto CttzIntrinsicCost = TTI.getIntrinsicInstrCost( |
7055 | ICA: IntrinsicCostAttributes(Intrinsic::cttz, CondTy, |
7056 | {Condition, ConstantInt::getTrue(Context)}), |
7057 | CostKind: TTI::TCK_SizeAndLatency); |
7058 | |
7059 | if (CttzIntrinsicCost > TTI::TCC_Basic) |
7060 | // Inserting intrinsic is too expensive. |
7061 | return false; |
7062 | |
7063 | // Only bother with this optimization if there are more than 3 switch cases. |
7064 | // SDAG will only bother creating jump tables for 4 or more cases. |
7065 | if (SI->getNumCases() < 4) |
7066 | return false; |
7067 | |
7068 | // We perform this optimization only for switches with |
7069 | // unreachable default case. |
7070 | // This assumtion will save us from checking if `Condition` is a power of two. |
7071 | if (!isa<UnreachableInst>(Val: SI->getDefaultDest()->getFirstNonPHIOrDbg())) |
7072 | return false; |
7073 | |
7074 | // Check that switch cases are powers of two. |
7075 | SmallVector<uint64_t, 4> Values; |
7076 | for (const auto &Case : SI->cases()) { |
7077 | uint64_t CaseValue = Case.getCaseValue()->getValue().getZExtValue(); |
7078 | if (llvm::has_single_bit(Value: CaseValue)) |
7079 | Values.push_back(Elt: CaseValue); |
7080 | else |
7081 | return false; |
7082 | } |
7083 | |
7084 | // isSwichDense requires case values to be sorted. |
7085 | llvm::sort(C&: Values); |
7086 | if (!isSwitchDense(NumCases: Values.size(), CaseRange: llvm::countr_zero(Val: Values.back()) - |
7087 | llvm::countr_zero(Val: Values.front()) + 1)) |
7088 | // Transform is unable to generate dense switch. |
7089 | return false; |
7090 | |
7091 | Builder.SetInsertPoint(SI); |
7092 | |
7093 | // Replace each case with its trailing zeros number. |
7094 | for (auto &Case : SI->cases()) { |
7095 | auto *OrigValue = Case.getCaseValue(); |
7096 | Case.setValue(ConstantInt::get(Ty: OrigValue->getIntegerType(), |
7097 | V: OrigValue->getValue().countr_zero())); |
7098 | } |
7099 | |
7100 | // Replace condition with its trailing zeros number. |
7101 | auto *ConditionTrailingZeros = Builder.CreateIntrinsic( |
7102 | ID: Intrinsic::cttz, Types: {CondTy}, Args: {Condition, ConstantInt::getTrue(Context)}); |
7103 | |
7104 | SI->setCondition(ConditionTrailingZeros); |
7105 | |
7106 | return true; |
7107 | } |
7108 | |
7109 | bool SimplifyCFGOpt::simplifySwitch(SwitchInst *SI, IRBuilder<> &Builder) { |
7110 | BasicBlock *BB = SI->getParent(); |
7111 | |
7112 | if (isValueEqualityComparison(TI: SI)) { |
7113 | // If we only have one predecessor, and if it is a branch on this value, |
7114 | // see if that predecessor totally determines the outcome of this switch. |
7115 | if (BasicBlock *OnlyPred = BB->getSinglePredecessor()) |
7116 | if (SimplifyEqualityComparisonWithOnlyPredecessor(TI: SI, Pred: OnlyPred, Builder)) |
7117 | return requestResimplify(); |
7118 | |
7119 | Value *Cond = SI->getCondition(); |
7120 | if (SelectInst *Select = dyn_cast<SelectInst>(Val: Cond)) |
7121 | if (SimplifySwitchOnSelect(SI, Select)) |
7122 | return requestResimplify(); |
7123 | |
7124 | // If the block only contains the switch, see if we can fold the block |
7125 | // away into any preds. |
7126 | if (SI == &*BB->instructionsWithoutDebug(SkipPseudoOp: false).begin()) |
7127 | if (FoldValueComparisonIntoPredecessors(TI: SI, Builder)) |
7128 | return requestResimplify(); |
7129 | } |
7130 | |
7131 | // Try to transform the switch into an icmp and a branch. |
7132 | // The conversion from switch to comparison may lose information on |
7133 | // impossible switch values, so disable it early in the pipeline. |
7134 | if (Options.ConvertSwitchRangeToICmp && TurnSwitchRangeIntoICmp(SI, Builder)) |
7135 | return requestResimplify(); |
7136 | |
7137 | // Remove unreachable cases. |
7138 | if (eliminateDeadSwitchCases(SI, DTU, AC: Options.AC, DL)) |
7139 | return requestResimplify(); |
7140 | |
7141 | if (trySwitchToSelect(SI, Builder, DTU, DL, TTI)) |
7142 | return requestResimplify(); |
7143 | |
7144 | if (Options.ForwardSwitchCondToPhi && ForwardSwitchConditionToPHI(SI)) |
7145 | return requestResimplify(); |
7146 | |
7147 | // The conversion from switch to lookup tables results in difficult-to-analyze |
7148 | // code and makes pruning branches much harder. This is a problem if the |
7149 | // switch expression itself can still be restricted as a result of inlining or |
7150 | // CVP. Therefore, only apply this transformation during late stages of the |
7151 | // optimisation pipeline. |
7152 | if (Options.ConvertSwitchToLookupTable && |
7153 | SwitchToLookupTable(SI, Builder, DTU, DL, TTI)) |
7154 | return requestResimplify(); |
7155 | |
7156 | if (simplifySwitchOfPowersOfTwo(SI, Builder, DL, TTI)) |
7157 | return requestResimplify(); |
7158 | |
7159 | if (ReduceSwitchRange(SI, Builder, DL, TTI)) |
7160 | return requestResimplify(); |
7161 | |
7162 | if (HoistCommon && |
7163 | hoistCommonCodeFromSuccessors(BB: SI->getParent(), EqTermsOnly: !Options.HoistCommonInsts)) |
7164 | return requestResimplify(); |
7165 | |
7166 | return false; |
7167 | } |
7168 | |
7169 | bool SimplifyCFGOpt::simplifyIndirectBr(IndirectBrInst *IBI) { |
7170 | BasicBlock *BB = IBI->getParent(); |
7171 | bool Changed = false; |
7172 | |
7173 | // Eliminate redundant destinations. |
7174 | SmallPtrSet<Value *, 8> Succs; |
7175 | SmallSetVector<BasicBlock *, 8> RemovedSuccs; |
7176 | for (unsigned i = 0, e = IBI->getNumDestinations(); i != e; ++i) { |
7177 | BasicBlock *Dest = IBI->getDestination(i); |
7178 | if (!Dest->hasAddressTaken() || !Succs.insert(Ptr: Dest).second) { |
7179 | if (!Dest->hasAddressTaken()) |
7180 | RemovedSuccs.insert(X: Dest); |
7181 | Dest->removePredecessor(Pred: BB); |
7182 | IBI->removeDestination(i); |
7183 | --i; |
7184 | --e; |
7185 | Changed = true; |
7186 | } |
7187 | } |
7188 | |
7189 | if (DTU) { |
7190 | std::vector<DominatorTree::UpdateType> Updates; |
7191 | Updates.reserve(n: RemovedSuccs.size()); |
7192 | for (auto *RemovedSucc : RemovedSuccs) |
7193 | Updates.push_back(x: {DominatorTree::Delete, BB, RemovedSucc}); |
7194 | DTU->applyUpdates(Updates); |
7195 | } |
7196 | |
7197 | if (IBI->getNumDestinations() == 0) { |
7198 | // If the indirectbr has no successors, change it to unreachable. |
7199 | new UnreachableInst(IBI->getContext(), IBI->getIterator()); |
7200 | EraseTerminatorAndDCECond(TI: IBI); |
7201 | return true; |
7202 | } |
7203 | |
7204 | if (IBI->getNumDestinations() == 1) { |
7205 | // If the indirectbr has one successor, change it to a direct branch. |
7206 | BranchInst::Create(IfTrue: IBI->getDestination(i: 0), InsertBefore: IBI->getIterator()); |
7207 | EraseTerminatorAndDCECond(TI: IBI); |
7208 | return true; |
7209 | } |
7210 | |
7211 | if (SelectInst *SI = dyn_cast<SelectInst>(Val: IBI->getAddress())) { |
7212 | if (SimplifyIndirectBrOnSelect(IBI, SI)) |
7213 | return requestResimplify(); |
7214 | } |
7215 | return Changed; |
7216 | } |
7217 | |
7218 | /// Given an block with only a single landing pad and a unconditional branch |
7219 | /// try to find another basic block which this one can be merged with. This |
7220 | /// handles cases where we have multiple invokes with unique landing pads, but |
7221 | /// a shared handler. |
7222 | /// |
7223 | /// We specifically choose to not worry about merging non-empty blocks |
7224 | /// here. That is a PRE/scheduling problem and is best solved elsewhere. In |
7225 | /// practice, the optimizer produces empty landing pad blocks quite frequently |
7226 | /// when dealing with exception dense code. (see: instcombine, gvn, if-else |
7227 | /// sinking in this file) |
7228 | /// |
7229 | /// This is primarily a code size optimization. We need to avoid performing |
7230 | /// any transform which might inhibit optimization (such as our ability to |
7231 | /// specialize a particular handler via tail commoning). We do this by not |
7232 | /// merging any blocks which require us to introduce a phi. Since the same |
7233 | /// values are flowing through both blocks, we don't lose any ability to |
7234 | /// specialize. If anything, we make such specialization more likely. |
7235 | /// |
7236 | /// TODO - This transformation could remove entries from a phi in the target |
7237 | /// block when the inputs in the phi are the same for the two blocks being |
7238 | /// merged. In some cases, this could result in removal of the PHI entirely. |
7239 | static bool TryToMergeLandingPad(LandingPadInst *LPad, BranchInst *BI, |
7240 | BasicBlock *BB, DomTreeUpdater *DTU) { |
7241 | auto Succ = BB->getUniqueSuccessor(); |
7242 | assert(Succ); |
7243 | // If there's a phi in the successor block, we'd likely have to introduce |
7244 | // a phi into the merged landing pad block. |
7245 | if (isa<PHINode>(Val: *Succ->begin())) |
7246 | return false; |
7247 | |
7248 | for (BasicBlock *OtherPred : predecessors(BB: Succ)) { |
7249 | if (BB == OtherPred) |
7250 | continue; |
7251 | BasicBlock::iterator I = OtherPred->begin(); |
7252 | LandingPadInst *LPad2 = dyn_cast<LandingPadInst>(Val&: I); |
7253 | if (!LPad2 || !LPad2->isIdenticalTo(I: LPad)) |
7254 | continue; |
7255 | for (++I; isa<DbgInfoIntrinsic>(Val: I); ++I) |
7256 | ; |
7257 | BranchInst *BI2 = dyn_cast<BranchInst>(Val&: I); |
7258 | if (!BI2 || !BI2->isIdenticalTo(I: BI)) |
7259 | continue; |
7260 | |
7261 | std::vector<DominatorTree::UpdateType> Updates; |
7262 | |
7263 | // We've found an identical block. Update our predecessors to take that |
7264 | // path instead and make ourselves dead. |
7265 | SmallSetVector<BasicBlock *, 16> UniquePreds(pred_begin(BB), pred_end(BB)); |
7266 | for (BasicBlock *Pred : UniquePreds) { |
7267 | InvokeInst *II = cast<InvokeInst>(Val: Pred->getTerminator()); |
7268 | assert(II->getNormalDest() != BB && II->getUnwindDest() == BB && |
7269 | "unexpected successor" ); |
7270 | II->setUnwindDest(OtherPred); |
7271 | if (DTU) { |
7272 | Updates.push_back(x: {DominatorTree::Insert, Pred, OtherPred}); |
7273 | Updates.push_back(x: {DominatorTree::Delete, Pred, BB}); |
7274 | } |
7275 | } |
7276 | |
7277 | // The debug info in OtherPred doesn't cover the merged control flow that |
7278 | // used to go through BB. We need to delete it or update it. |
7279 | for (Instruction &Inst : llvm::make_early_inc_range(Range&: *OtherPred)) |
7280 | if (isa<DbgInfoIntrinsic>(Val: Inst)) |
7281 | Inst.eraseFromParent(); |
7282 | |
7283 | SmallSetVector<BasicBlock *, 16> UniqueSuccs(succ_begin(BB), succ_end(BB)); |
7284 | for (BasicBlock *Succ : UniqueSuccs) { |
7285 | Succ->removePredecessor(Pred: BB); |
7286 | if (DTU) |
7287 | Updates.push_back(x: {DominatorTree::Delete, BB, Succ}); |
7288 | } |
7289 | |
7290 | IRBuilder<> Builder(BI); |
7291 | Builder.CreateUnreachable(); |
7292 | BI->eraseFromParent(); |
7293 | if (DTU) |
7294 | DTU->applyUpdates(Updates); |
7295 | return true; |
7296 | } |
7297 | return false; |
7298 | } |
7299 | |
7300 | bool SimplifyCFGOpt::simplifyBranch(BranchInst *Branch, IRBuilder<> &Builder) { |
7301 | return Branch->isUnconditional() ? simplifyUncondBranch(BI: Branch, Builder) |
7302 | : simplifyCondBranch(BI: Branch, Builder); |
7303 | } |
7304 | |
7305 | bool SimplifyCFGOpt::simplifyUncondBranch(BranchInst *BI, |
7306 | IRBuilder<> &Builder) { |
7307 | BasicBlock *BB = BI->getParent(); |
7308 | BasicBlock *Succ = BI->getSuccessor(i: 0); |
7309 | |
7310 | // If the Terminator is the only non-phi instruction, simplify the block. |
7311 | // If LoopHeader is provided, check if the block or its successor is a loop |
7312 | // header. (This is for early invocations before loop simplify and |
7313 | // vectorization to keep canonical loop forms for nested loops. These blocks |
7314 | // can be eliminated when the pass is invoked later in the back-end.) |
7315 | // Note that if BB has only one predecessor then we do not introduce new |
7316 | // backedge, so we can eliminate BB. |
7317 | bool NeedCanonicalLoop = |
7318 | Options.NeedCanonicalLoop && |
7319 | (!LoopHeaders.empty() && BB->hasNPredecessorsOrMore(N: 2) && |
7320 | (is_contained(Range&: LoopHeaders, Element: BB) || is_contained(Range&: LoopHeaders, Element: Succ))); |
7321 | BasicBlock::iterator I = BB->getFirstNonPHIOrDbg(SkipPseudoOp: true)->getIterator(); |
7322 | if (I->isTerminator() && BB != &BB->getParent()->getEntryBlock() && |
7323 | !NeedCanonicalLoop && TryToSimplifyUncondBranchFromEmptyBlock(BB, DTU)) |
7324 | return true; |
7325 | |
7326 | // If the only instruction in the block is a seteq/setne comparison against a |
7327 | // constant, try to simplify the block. |
7328 | if (ICmpInst *ICI = dyn_cast<ICmpInst>(Val&: I)) |
7329 | if (ICI->isEquality() && isa<ConstantInt>(Val: ICI->getOperand(i_nocapture: 1))) { |
7330 | for (++I; isa<DbgInfoIntrinsic>(Val: I); ++I) |
7331 | ; |
7332 | if (I->isTerminator() && |
7333 | tryToSimplifyUncondBranchWithICmpInIt(ICI, Builder)) |
7334 | return true; |
7335 | } |
7336 | |
7337 | // See if we can merge an empty landing pad block with another which is |
7338 | // equivalent. |
7339 | if (LandingPadInst *LPad = dyn_cast<LandingPadInst>(Val&: I)) { |
7340 | for (++I; isa<DbgInfoIntrinsic>(Val: I); ++I) |
7341 | ; |
7342 | if (I->isTerminator() && TryToMergeLandingPad(LPad, BI, BB, DTU)) |
7343 | return true; |
7344 | } |
7345 | |
7346 | // If this basic block is ONLY a compare and a branch, and if a predecessor |
7347 | // branches to us and our successor, fold the comparison into the |
7348 | // predecessor and use logical operations to update the incoming value |
7349 | // for PHI nodes in common successor. |
7350 | if (Options.SpeculateBlocks && |
7351 | FoldBranchToCommonDest(BI, DTU, /*MSSAU=*/nullptr, TTI: &TTI, |
7352 | BonusInstThreshold: Options.BonusInstThreshold)) |
7353 | return requestResimplify(); |
7354 | return false; |
7355 | } |
7356 | |
7357 | static BasicBlock *allPredecessorsComeFromSameSource(BasicBlock *BB) { |
7358 | BasicBlock *PredPred = nullptr; |
7359 | for (auto *P : predecessors(BB)) { |
7360 | BasicBlock *PPred = P->getSinglePredecessor(); |
7361 | if (!PPred || (PredPred && PredPred != PPred)) |
7362 | return nullptr; |
7363 | PredPred = PPred; |
7364 | } |
7365 | return PredPred; |
7366 | } |
7367 | |
7368 | /// Fold the following pattern: |
7369 | /// bb0: |
7370 | /// br i1 %cond1, label %bb1, label %bb2 |
7371 | /// bb1: |
7372 | /// br i1 %cond2, label %bb3, label %bb4 |
7373 | /// bb2: |
7374 | /// br i1 %cond2, label %bb4, label %bb3 |
7375 | /// bb3: |
7376 | /// ... |
7377 | /// bb4: |
7378 | /// ... |
7379 | /// into |
7380 | /// bb0: |
7381 | /// %cond = xor i1 %cond1, %cond2 |
7382 | /// br i1 %cond, label %bb4, label %bb3 |
7383 | /// bb3: |
7384 | /// ... |
7385 | /// bb4: |
7386 | /// ... |
7387 | /// NOTE: %cond2 always dominates the terminator of bb0. |
7388 | static bool mergeNestedCondBranch(BranchInst *BI, DomTreeUpdater *DTU) { |
7389 | BasicBlock *BB = BI->getParent(); |
7390 | BasicBlock *BB1 = BI->getSuccessor(i: 0); |
7391 | BasicBlock *BB2 = BI->getSuccessor(i: 1); |
7392 | auto IsSimpleSuccessor = [BB](BasicBlock *Succ, BranchInst *&SuccBI) { |
7393 | if (Succ == BB) |
7394 | return false; |
7395 | if (&Succ->front() != Succ->getTerminator()) |
7396 | return false; |
7397 | SuccBI = dyn_cast<BranchInst>(Val: Succ->getTerminator()); |
7398 | if (!SuccBI || !SuccBI->isConditional()) |
7399 | return false; |
7400 | BasicBlock *Succ1 = SuccBI->getSuccessor(i: 0); |
7401 | BasicBlock *Succ2 = SuccBI->getSuccessor(i: 1); |
7402 | return Succ1 != Succ && Succ2 != Succ && Succ1 != BB && Succ2 != BB && |
7403 | !isa<PHINode>(Val: Succ1->front()) && !isa<PHINode>(Val: Succ2->front()); |
7404 | }; |
7405 | BranchInst *BB1BI, *BB2BI; |
7406 | if (!IsSimpleSuccessor(BB1, BB1BI) || !IsSimpleSuccessor(BB2, BB2BI)) |
7407 | return false; |
7408 | |
7409 | if (BB1BI->getCondition() != BB2BI->getCondition() || |
7410 | BB1BI->getSuccessor(i: 0) != BB2BI->getSuccessor(i: 1) || |
7411 | BB1BI->getSuccessor(i: 1) != BB2BI->getSuccessor(i: 0)) |
7412 | return false; |
7413 | |
7414 | BasicBlock *BB3 = BB1BI->getSuccessor(i: 0); |
7415 | BasicBlock *BB4 = BB1BI->getSuccessor(i: 1); |
7416 | IRBuilder<> Builder(BI); |
7417 | BI->setCondition( |
7418 | Builder.CreateXor(LHS: BI->getCondition(), RHS: BB1BI->getCondition())); |
7419 | BB1->removePredecessor(Pred: BB); |
7420 | BI->setSuccessor(idx: 0, NewSucc: BB4); |
7421 | BB2->removePredecessor(Pred: BB); |
7422 | BI->setSuccessor(idx: 1, NewSucc: BB3); |
7423 | if (DTU) { |
7424 | SmallVector<DominatorTree::UpdateType, 4> Updates; |
7425 | Updates.push_back(Elt: {DominatorTree::Delete, BB, BB1}); |
7426 | Updates.push_back(Elt: {DominatorTree::Insert, BB, BB4}); |
7427 | Updates.push_back(Elt: {DominatorTree::Delete, BB, BB2}); |
7428 | Updates.push_back(Elt: {DominatorTree::Insert, BB, BB3}); |
7429 | |
7430 | DTU->applyUpdates(Updates); |
7431 | } |
7432 | bool HasWeight = false; |
7433 | uint64_t BBTWeight, BBFWeight; |
7434 | if (extractBranchWeights(I: *BI, TrueVal&: BBTWeight, FalseVal&: BBFWeight)) |
7435 | HasWeight = true; |
7436 | else |
7437 | BBTWeight = BBFWeight = 1; |
7438 | uint64_t BB1TWeight, BB1FWeight; |
7439 | if (extractBranchWeights(I: *BB1BI, TrueVal&: BB1TWeight, FalseVal&: BB1FWeight)) |
7440 | HasWeight = true; |
7441 | else |
7442 | BB1TWeight = BB1FWeight = 1; |
7443 | uint64_t BB2TWeight, BB2FWeight; |
7444 | if (extractBranchWeights(I: *BB2BI, TrueVal&: BB2TWeight, FalseVal&: BB2FWeight)) |
7445 | HasWeight = true; |
7446 | else |
7447 | BB2TWeight = BB2FWeight = 1; |
7448 | if (HasWeight) { |
7449 | uint64_t Weights[2] = {BBTWeight * BB1FWeight + BBFWeight * BB2TWeight, |
7450 | BBTWeight * BB1TWeight + BBFWeight * BB2FWeight}; |
7451 | FitWeights(Weights); |
7452 | setBranchWeights(I: BI, TrueWeight: Weights[0], FalseWeight: Weights[1], /*IsExpected=*/false); |
7453 | } |
7454 | return true; |
7455 | } |
7456 | |
7457 | bool SimplifyCFGOpt::simplifyCondBranch(BranchInst *BI, IRBuilder<> &Builder) { |
7458 | assert( |
7459 | !isa<ConstantInt>(BI->getCondition()) && |
7460 | BI->getSuccessor(0) != BI->getSuccessor(1) && |
7461 | "Tautological conditional branch should have been eliminated already." ); |
7462 | |
7463 | BasicBlock *BB = BI->getParent(); |
7464 | if (!Options.SimplifyCondBranch || |
7465 | BI->getFunction()->hasFnAttribute(Kind: Attribute::OptForFuzzing)) |
7466 | return false; |
7467 | |
7468 | // Conditional branch |
7469 | if (isValueEqualityComparison(TI: BI)) { |
7470 | // If we only have one predecessor, and if it is a branch on this value, |
7471 | // see if that predecessor totally determines the outcome of this |
7472 | // switch. |
7473 | if (BasicBlock *OnlyPred = BB->getSinglePredecessor()) |
7474 | if (SimplifyEqualityComparisonWithOnlyPredecessor(TI: BI, Pred: OnlyPred, Builder)) |
7475 | return requestResimplify(); |
7476 | |
7477 | // This block must be empty, except for the setcond inst, if it exists. |
7478 | // Ignore dbg and pseudo intrinsics. |
7479 | auto I = BB->instructionsWithoutDebug(SkipPseudoOp: true).begin(); |
7480 | if (&*I == BI) { |
7481 | if (FoldValueComparisonIntoPredecessors(TI: BI, Builder)) |
7482 | return requestResimplify(); |
7483 | } else if (&*I == cast<Instruction>(Val: BI->getCondition())) { |
7484 | ++I; |
7485 | if (&*I == BI && FoldValueComparisonIntoPredecessors(TI: BI, Builder)) |
7486 | return requestResimplify(); |
7487 | } |
7488 | } |
7489 | |
7490 | // Try to turn "br (X == 0 | X == 1), T, F" into a switch instruction. |
7491 | if (SimplifyBranchOnICmpChain(BI, Builder, DL)) |
7492 | return true; |
7493 | |
7494 | // If this basic block has dominating predecessor blocks and the dominating |
7495 | // blocks' conditions imply BI's condition, we know the direction of BI. |
7496 | std::optional<bool> Imp = isImpliedByDomCondition(Cond: BI->getCondition(), ContextI: BI, DL); |
7497 | if (Imp) { |
7498 | // Turn this into a branch on constant. |
7499 | auto *OldCond = BI->getCondition(); |
7500 | ConstantInt *TorF = *Imp ? ConstantInt::getTrue(Context&: BB->getContext()) |
7501 | : ConstantInt::getFalse(Context&: BB->getContext()); |
7502 | BI->setCondition(TorF); |
7503 | RecursivelyDeleteTriviallyDeadInstructions(V: OldCond); |
7504 | return requestResimplify(); |
7505 | } |
7506 | |
7507 | // If this basic block is ONLY a compare and a branch, and if a predecessor |
7508 | // branches to us and one of our successors, fold the comparison into the |
7509 | // predecessor and use logical operations to pick the right destination. |
7510 | if (Options.SpeculateBlocks && |
7511 | FoldBranchToCommonDest(BI, DTU, /*MSSAU=*/nullptr, TTI: &TTI, |
7512 | BonusInstThreshold: Options.BonusInstThreshold)) |
7513 | return requestResimplify(); |
7514 | |
7515 | // We have a conditional branch to two blocks that are only reachable |
7516 | // from BI. We know that the condbr dominates the two blocks, so see if |
7517 | // there is any identical code in the "then" and "else" blocks. If so, we |
7518 | // can hoist it up to the branching block. |
7519 | if (BI->getSuccessor(i: 0)->getSinglePredecessor()) { |
7520 | if (BI->getSuccessor(i: 1)->getSinglePredecessor()) { |
7521 | if (HoistCommon && hoistCommonCodeFromSuccessors( |
7522 | BB: BI->getParent(), EqTermsOnly: !Options.HoistCommonInsts)) |
7523 | return requestResimplify(); |
7524 | } else { |
7525 | // If Successor #1 has multiple preds, we may be able to conditionally |
7526 | // execute Successor #0 if it branches to Successor #1. |
7527 | Instruction *Succ0TI = BI->getSuccessor(i: 0)->getTerminator(); |
7528 | if (Succ0TI->getNumSuccessors() == 1 && |
7529 | Succ0TI->getSuccessor(Idx: 0) == BI->getSuccessor(i: 1)) |
7530 | if (SpeculativelyExecuteBB(BI, ThenBB: BI->getSuccessor(i: 0))) |
7531 | return requestResimplify(); |
7532 | } |
7533 | } else if (BI->getSuccessor(i: 1)->getSinglePredecessor()) { |
7534 | // If Successor #0 has multiple preds, we may be able to conditionally |
7535 | // execute Successor #1 if it branches to Successor #0. |
7536 | Instruction *Succ1TI = BI->getSuccessor(i: 1)->getTerminator(); |
7537 | if (Succ1TI->getNumSuccessors() == 1 && |
7538 | Succ1TI->getSuccessor(Idx: 0) == BI->getSuccessor(i: 0)) |
7539 | if (SpeculativelyExecuteBB(BI, ThenBB: BI->getSuccessor(i: 1))) |
7540 | return requestResimplify(); |
7541 | } |
7542 | |
7543 | // If this is a branch on something for which we know the constant value in |
7544 | // predecessors (e.g. a phi node in the current block), thread control |
7545 | // through this block. |
7546 | if (FoldCondBranchOnValueKnownInPredecessor(BI, DTU, DL, AC: Options.AC)) |
7547 | return requestResimplify(); |
7548 | |
7549 | // Scan predecessor blocks for conditional branches. |
7550 | for (BasicBlock *Pred : predecessors(BB)) |
7551 | if (BranchInst *PBI = dyn_cast<BranchInst>(Val: Pred->getTerminator())) |
7552 | if (PBI != BI && PBI->isConditional()) |
7553 | if (SimplifyCondBranchToCondBranch(PBI, BI, DTU, DL, TTI)) |
7554 | return requestResimplify(); |
7555 | |
7556 | // Look for diamond patterns. |
7557 | if (MergeCondStores) |
7558 | if (BasicBlock *PrevBB = allPredecessorsComeFromSameSource(BB)) |
7559 | if (BranchInst *PBI = dyn_cast<BranchInst>(Val: PrevBB->getTerminator())) |
7560 | if (PBI != BI && PBI->isConditional()) |
7561 | if (mergeConditionalStores(PBI, QBI: BI, DTU, DL, TTI)) |
7562 | return requestResimplify(); |
7563 | |
7564 | // Look for nested conditional branches. |
7565 | if (mergeNestedCondBranch(BI, DTU)) |
7566 | return requestResimplify(); |
7567 | |
7568 | return false; |
7569 | } |
7570 | |
7571 | /// Check if passing a value to an instruction will cause undefined behavior. |
7572 | static bool passingValueIsAlwaysUndefined(Value *V, Instruction *I, bool PtrValueMayBeModified) { |
7573 | Constant *C = dyn_cast<Constant>(Val: V); |
7574 | if (!C) |
7575 | return false; |
7576 | |
7577 | if (I->use_empty()) |
7578 | return false; |
7579 | |
7580 | if (C->isNullValue() || isa<UndefValue>(Val: C)) { |
7581 | // Only look at the first use we can handle, avoid hurting compile time with |
7582 | // long uselists |
7583 | auto FindUse = llvm::find_if(Range: I->users(), P: [](auto *U) { |
7584 | auto *Use = cast<Instruction>(U); |
7585 | // Change this list when we want to add new instructions. |
7586 | switch (Use->getOpcode()) { |
7587 | default: |
7588 | return false; |
7589 | case Instruction::GetElementPtr: |
7590 | case Instruction::Ret: |
7591 | case Instruction::BitCast: |
7592 | case Instruction::Load: |
7593 | case Instruction::Store: |
7594 | case Instruction::Call: |
7595 | case Instruction::CallBr: |
7596 | case Instruction::Invoke: |
7597 | return true; |
7598 | } |
7599 | }); |
7600 | if (FindUse == I->user_end()) |
7601 | return false; |
7602 | auto *Use = cast<Instruction>(Val: *FindUse); |
7603 | // Bail out if Use is not in the same BB as I or Use == I or Use comes |
7604 | // before I in the block. The latter two can be the case if Use is a |
7605 | // PHI node. |
7606 | if (Use->getParent() != I->getParent() || Use == I || Use->comesBefore(Other: I)) |
7607 | return false; |
7608 | |
7609 | // Now make sure that there are no instructions in between that can alter |
7610 | // control flow (eg. calls) |
7611 | auto InstrRange = |
7612 | make_range(x: std::next(x: I->getIterator()), y: Use->getIterator()); |
7613 | if (any_of(Range&: InstrRange, P: [](Instruction &I) { |
7614 | return !isGuaranteedToTransferExecutionToSuccessor(I: &I); |
7615 | })) |
7616 | return false; |
7617 | |
7618 | // Look through GEPs. A load from a GEP derived from NULL is still undefined |
7619 | if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Val: Use)) |
7620 | if (GEP->getPointerOperand() == I) { |
7621 | // The current base address is null, there are four cases to consider: |
7622 | // getelementptr (TY, null, 0) -> null |
7623 | // getelementptr (TY, null, not zero) -> may be modified |
7624 | // getelementptr inbounds (TY, null, 0) -> null |
7625 | // getelementptr inbounds (TY, null, not zero) -> poison iff null is |
7626 | // undefined? |
7627 | if (!GEP->hasAllZeroIndices() && |
7628 | (!GEP->isInBounds() || |
7629 | NullPointerIsDefined(F: GEP->getFunction(), |
7630 | AS: GEP->getPointerAddressSpace()))) |
7631 | PtrValueMayBeModified = true; |
7632 | return passingValueIsAlwaysUndefined(V, I: GEP, PtrValueMayBeModified); |
7633 | } |
7634 | |
7635 | // Look through return. |
7636 | if (ReturnInst *Ret = dyn_cast<ReturnInst>(Val: Use)) { |
7637 | bool HasNoUndefAttr = |
7638 | Ret->getFunction()->hasRetAttribute(Kind: Attribute::NoUndef); |
7639 | // Return undefined to a noundef return value is undefined. |
7640 | if (isa<UndefValue>(Val: C) && HasNoUndefAttr) |
7641 | return true; |
7642 | // Return null to a nonnull+noundef return value is undefined. |
7643 | if (C->isNullValue() && HasNoUndefAttr && |
7644 | Ret->getFunction()->hasRetAttribute(Kind: Attribute::NonNull)) { |
7645 | return !PtrValueMayBeModified; |
7646 | } |
7647 | } |
7648 | |
7649 | // Look through bitcasts. |
7650 | if (BitCastInst *BC = dyn_cast<BitCastInst>(Val: Use)) |
7651 | return passingValueIsAlwaysUndefined(V, I: BC, PtrValueMayBeModified); |
7652 | |
7653 | // Load from null is undefined. |
7654 | if (LoadInst *LI = dyn_cast<LoadInst>(Val: Use)) |
7655 | if (!LI->isVolatile()) |
7656 | return !NullPointerIsDefined(F: LI->getFunction(), |
7657 | AS: LI->getPointerAddressSpace()); |
7658 | |
7659 | // Store to null is undefined. |
7660 | if (StoreInst *SI = dyn_cast<StoreInst>(Val: Use)) |
7661 | if (!SI->isVolatile()) |
7662 | return (!NullPointerIsDefined(F: SI->getFunction(), |
7663 | AS: SI->getPointerAddressSpace())) && |
7664 | SI->getPointerOperand() == I; |
7665 | |
7666 | // llvm.assume(false/undef) always triggers immediate UB. |
7667 | if (auto *Assume = dyn_cast<AssumeInst>(Val: Use)) { |
7668 | // Ignore assume operand bundles. |
7669 | if (I == Assume->getArgOperand(i: 0)) |
7670 | return true; |
7671 | } |
7672 | |
7673 | if (auto *CB = dyn_cast<CallBase>(Val: Use)) { |
7674 | if (C->isNullValue() && NullPointerIsDefined(F: CB->getFunction())) |
7675 | return false; |
7676 | // A call to null is undefined. |
7677 | if (CB->getCalledOperand() == I) |
7678 | return true; |
7679 | |
7680 | if (C->isNullValue()) { |
7681 | for (const llvm::Use &Arg : CB->args()) |
7682 | if (Arg == I) { |
7683 | unsigned ArgIdx = CB->getArgOperandNo(U: &Arg); |
7684 | if (CB->isPassingUndefUB(ArgNo: ArgIdx) && |
7685 | CB->paramHasAttr(ArgNo: ArgIdx, Kind: Attribute::NonNull)) { |
7686 | // Passing null to a nonnnull+noundef argument is undefined. |
7687 | return !PtrValueMayBeModified; |
7688 | } |
7689 | } |
7690 | } else if (isa<UndefValue>(Val: C)) { |
7691 | // Passing undef to a noundef argument is undefined. |
7692 | for (const llvm::Use &Arg : CB->args()) |
7693 | if (Arg == I) { |
7694 | unsigned ArgIdx = CB->getArgOperandNo(U: &Arg); |
7695 | if (CB->isPassingUndefUB(ArgNo: ArgIdx)) { |
7696 | // Passing undef to a noundef argument is undefined. |
7697 | return true; |
7698 | } |
7699 | } |
7700 | } |
7701 | } |
7702 | } |
7703 | return false; |
7704 | } |
7705 | |
7706 | /// If BB has an incoming value that will always trigger undefined behavior |
7707 | /// (eg. null pointer dereference), remove the branch leading here. |
7708 | static bool removeUndefIntroducingPredecessor(BasicBlock *BB, |
7709 | DomTreeUpdater *DTU, |
7710 | AssumptionCache *AC) { |
7711 | for (PHINode &PHI : BB->phis()) |
7712 | for (unsigned i = 0, e = PHI.getNumIncomingValues(); i != e; ++i) |
7713 | if (passingValueIsAlwaysUndefined(V: PHI.getIncomingValue(i), I: &PHI)) { |
7714 | BasicBlock *Predecessor = PHI.getIncomingBlock(i); |
7715 | Instruction *T = Predecessor->getTerminator(); |
7716 | IRBuilder<> Builder(T); |
7717 | if (BranchInst *BI = dyn_cast<BranchInst>(Val: T)) { |
7718 | BB->removePredecessor(Pred: Predecessor); |
7719 | // Turn unconditional branches into unreachables and remove the dead |
7720 | // destination from conditional branches. |
7721 | if (BI->isUnconditional()) |
7722 | Builder.CreateUnreachable(); |
7723 | else { |
7724 | // Preserve guarding condition in assume, because it might not be |
7725 | // inferrable from any dominating condition. |
7726 | Value *Cond = BI->getCondition(); |
7727 | CallInst *Assumption; |
7728 | if (BI->getSuccessor(i: 0) == BB) |
7729 | Assumption = Builder.CreateAssumption(Cond: Builder.CreateNot(V: Cond)); |
7730 | else |
7731 | Assumption = Builder.CreateAssumption(Cond); |
7732 | if (AC) |
7733 | AC->registerAssumption(CI: cast<AssumeInst>(Val: Assumption)); |
7734 | Builder.CreateBr(Dest: BI->getSuccessor(i: 0) == BB ? BI->getSuccessor(i: 1) |
7735 | : BI->getSuccessor(i: 0)); |
7736 | } |
7737 | BI->eraseFromParent(); |
7738 | if (DTU) |
7739 | DTU->applyUpdates(Updates: {{DominatorTree::Delete, Predecessor, BB}}); |
7740 | return true; |
7741 | } else if (SwitchInst *SI = dyn_cast<SwitchInst>(Val: T)) { |
7742 | // Redirect all branches leading to UB into |
7743 | // a newly created unreachable block. |
7744 | BasicBlock *Unreachable = BasicBlock::Create( |
7745 | Context&: Predecessor->getContext(), Name: "unreachable" , Parent: BB->getParent(), InsertBefore: BB); |
7746 | Builder.SetInsertPoint(Unreachable); |
7747 | // The new block contains only one instruction: Unreachable |
7748 | Builder.CreateUnreachable(); |
7749 | for (const auto &Case : SI->cases()) |
7750 | if (Case.getCaseSuccessor() == BB) { |
7751 | BB->removePredecessor(Pred: Predecessor); |
7752 | Case.setSuccessor(Unreachable); |
7753 | } |
7754 | if (SI->getDefaultDest() == BB) { |
7755 | BB->removePredecessor(Pred: Predecessor); |
7756 | SI->setDefaultDest(Unreachable); |
7757 | } |
7758 | |
7759 | if (DTU) |
7760 | DTU->applyUpdates( |
7761 | Updates: { { DominatorTree::Insert, Predecessor, Unreachable }, |
7762 | { DominatorTree::Delete, Predecessor, BB } }); |
7763 | return true; |
7764 | } |
7765 | } |
7766 | |
7767 | return false; |
7768 | } |
7769 | |
7770 | bool SimplifyCFGOpt::simplifyOnce(BasicBlock *BB) { |
7771 | bool Changed = false; |
7772 | |
7773 | assert(BB && BB->getParent() && "Block not embedded in function!" ); |
7774 | assert(BB->getTerminator() && "Degenerate basic block encountered!" ); |
7775 | |
7776 | // Remove basic blocks that have no predecessors (except the entry block)... |
7777 | // or that just have themself as a predecessor. These are unreachable. |
7778 | if ((pred_empty(BB) && BB != &BB->getParent()->getEntryBlock()) || |
7779 | BB->getSinglePredecessor() == BB) { |
7780 | LLVM_DEBUG(dbgs() << "Removing BB: \n" << *BB); |
7781 | DeleteDeadBlock(BB, DTU); |
7782 | return true; |
7783 | } |
7784 | |
7785 | // Check to see if we can constant propagate this terminator instruction |
7786 | // away... |
7787 | Changed |= ConstantFoldTerminator(BB, /*DeleteDeadConditions=*/true, |
7788 | /*TLI=*/nullptr, DTU); |
7789 | |
7790 | // Check for and eliminate duplicate PHI nodes in this block. |
7791 | Changed |= EliminateDuplicatePHINodes(BB); |
7792 | |
7793 | // Check for and remove branches that will always cause undefined behavior. |
7794 | if (removeUndefIntroducingPredecessor(BB, DTU, AC: Options.AC)) |
7795 | return requestResimplify(); |
7796 | |
7797 | // Merge basic blocks into their predecessor if there is only one distinct |
7798 | // pred, and if there is only one distinct successor of the predecessor, and |
7799 | // if there are no PHI nodes. |
7800 | if (MergeBlockIntoPredecessor(BB, DTU)) |
7801 | return true; |
7802 | |
7803 | if (SinkCommon && Options.SinkCommonInsts) |
7804 | if (SinkCommonCodeFromPredecessors(BB, DTU) || |
7805 | MergeCompatibleInvokes(BB, DTU)) { |
7806 | // SinkCommonCodeFromPredecessors() does not automatically CSE PHI's, |
7807 | // so we may now how duplicate PHI's. |
7808 | // Let's rerun EliminateDuplicatePHINodes() first, |
7809 | // before FoldTwoEntryPHINode() potentially converts them into select's, |
7810 | // after which we'd need a whole EarlyCSE pass run to cleanup them. |
7811 | return true; |
7812 | } |
7813 | |
7814 | IRBuilder<> Builder(BB); |
7815 | |
7816 | if (Options.SpeculateBlocks && |
7817 | !BB->getParent()->hasFnAttribute(Kind: Attribute::OptForFuzzing)) { |
7818 | // If there is a trivial two-entry PHI node in this basic block, and we can |
7819 | // eliminate it, do so now. |
7820 | if (auto *PN = dyn_cast<PHINode>(Val: BB->begin())) |
7821 | if (PN->getNumIncomingValues() == 2) |
7822 | if (FoldTwoEntryPHINode(PN, TTI, DTU, DL, |
7823 | SpeculateUnpredictables: Options.SpeculateUnpredictables)) |
7824 | return true; |
7825 | } |
7826 | |
7827 | Instruction *Terminator = BB->getTerminator(); |
7828 | Builder.SetInsertPoint(Terminator); |
7829 | switch (Terminator->getOpcode()) { |
7830 | case Instruction::Br: |
7831 | Changed |= simplifyBranch(Branch: cast<BranchInst>(Val: Terminator), Builder); |
7832 | break; |
7833 | case Instruction::Resume: |
7834 | Changed |= simplifyResume(RI: cast<ResumeInst>(Val: Terminator), Builder); |
7835 | break; |
7836 | case Instruction::CleanupRet: |
7837 | Changed |= simplifyCleanupReturn(RI: cast<CleanupReturnInst>(Val: Terminator)); |
7838 | break; |
7839 | case Instruction::Switch: |
7840 | Changed |= simplifySwitch(SI: cast<SwitchInst>(Val: Terminator), Builder); |
7841 | break; |
7842 | case Instruction::Unreachable: |
7843 | Changed |= simplifyUnreachable(UI: cast<UnreachableInst>(Val: Terminator)); |
7844 | break; |
7845 | case Instruction::IndirectBr: |
7846 | Changed |= simplifyIndirectBr(IBI: cast<IndirectBrInst>(Val: Terminator)); |
7847 | break; |
7848 | } |
7849 | |
7850 | return Changed; |
7851 | } |
7852 | |
7853 | bool SimplifyCFGOpt::run(BasicBlock *BB) { |
7854 | bool Changed = false; |
7855 | |
7856 | // Repeated simplify BB as long as resimplification is requested. |
7857 | do { |
7858 | Resimplify = false; |
7859 | |
7860 | // Perform one round of simplifcation. Resimplify flag will be set if |
7861 | // another iteration is requested. |
7862 | Changed |= simplifyOnce(BB); |
7863 | } while (Resimplify); |
7864 | |
7865 | return Changed; |
7866 | } |
7867 | |
7868 | bool llvm::simplifyCFG(BasicBlock *BB, const TargetTransformInfo &TTI, |
7869 | DomTreeUpdater *DTU, const SimplifyCFGOptions &Options, |
7870 | ArrayRef<WeakVH> ) { |
7871 | return SimplifyCFGOpt(TTI, DTU, BB->getDataLayout(), LoopHeaders, |
7872 | Options) |
7873 | .run(BB); |
7874 | } |
7875 | |