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/Optional.h"
18 #include "llvm/ADT/STLExtras.h"
19 #include "llvm/ADT/ScopeExit.h"
20 #include "llvm/ADT/Sequence.h"
21 #include "llvm/ADT/SetOperations.h"
22 #include "llvm/ADT/SetVector.h"
23 #include "llvm/ADT/SmallPtrSet.h"
24 #include "llvm/ADT/SmallVector.h"
25 #include "llvm/ADT/Statistic.h"
26 #include "llvm/ADT/StringRef.h"
27 #include "llvm/Analysis/AssumptionCache.h"
28 #include "llvm/Analysis/ConstantFolding.h"
29 #include "llvm/Analysis/EHPersonalities.h"
30 #include "llvm/Analysis/GuardUtils.h"
31 #include "llvm/Analysis/InstructionSimplify.h"
32 #include "llvm/Analysis/MemorySSA.h"
33 #include "llvm/Analysis/MemorySSAUpdater.h"
34 #include "llvm/Analysis/TargetTransformInfo.h"
35 #include "llvm/Analysis/ValueTracking.h"
36 #include "llvm/IR/Attributes.h"
37 #include "llvm/IR/BasicBlock.h"
38 #include "llvm/IR/CFG.h"
39 #include "llvm/IR/Constant.h"
40 #include "llvm/IR/ConstantRange.h"
41 #include "llvm/IR/Constants.h"
42 #include "llvm/IR/DataLayout.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/Intrinsics.h"
53 #include "llvm/IR/LLVMContext.h"
54 #include "llvm/IR/MDBuilder.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/Type.h"
61 #include "llvm/IR/Use.h"
62 #include "llvm/IR/User.h"
63 #include "llvm/IR/Value.h"
64 #include "llvm/IR/ValueHandle.h"
65 #include "llvm/Support/Casting.h"
66 #include "llvm/Support/CommandLine.h"
67 #include "llvm/Support/Debug.h"
68 #include "llvm/Support/ErrorHandling.h"
69 #include "llvm/Support/KnownBits.h"
70 #include "llvm/Support/MathExtras.h"
71 #include "llvm/Support/raw_ostream.h"
72 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
73 #include "llvm/Transforms/Utils/Local.h"
74 #include "llvm/Transforms/Utils/SSAUpdater.h"
75 #include "llvm/Transforms/Utils/ValueMapper.h"
76 #include <algorithm>
77 #include <cassert>
78 #include <climits>
79 #include <cstddef>
80 #include <cstdint>
81 #include <iterator>
82 #include <map>
83 #include <set>
84 #include <tuple>
85 #include <utility>
86 #include <vector>
87
88 using namespace llvm;
89 using namespace PatternMatch;
90
91 #define DEBUG_TYPE "simplifycfg"
92
93 cl::opt<bool> llvm::RequireAndPreserveDomTree(
94 "simplifycfg-require-and-preserve-domtree", cl::Hidden, cl::ZeroOrMore,
95 cl::init(false),
96 cl::desc("Temorary development switch used to gradually uplift SimplifyCFG "
97 "into preserving DomTree,"));
98
99 // Chosen as 2 so as to be cheap, but still to have enough power to fold
100 // a select, so the "clamp" idiom (of a min followed by a max) will be caught.
101 // To catch this, we need to fold a compare and a select, hence '2' being the
102 // minimum reasonable default.
103 static cl::opt<unsigned> PHINodeFoldingThreshold(
104 "phi-node-folding-threshold", cl::Hidden, cl::init(2),
105 cl::desc(
106 "Control the amount of phi node folding to perform (default = 2)"));
107
108 static cl::opt<unsigned> TwoEntryPHINodeFoldingThreshold(
109 "two-entry-phi-node-folding-threshold", cl::Hidden, cl::init(4),
110 cl::desc("Control the maximal total instruction cost that we are willing "
111 "to speculatively execute to fold a 2-entry PHI node into a "
112 "select (default = 4)"));
113
114 static cl::opt<bool> DupRet(
115 "simplifycfg-dup-ret", cl::Hidden, cl::init(false),
116 cl::desc("Duplicate return instructions into unconditional branches"));
117
118 static cl::opt<bool>
119 HoistCommon("simplifycfg-hoist-common", cl::Hidden, cl::init(true),
120 cl::desc("Hoist common instructions up to the parent block"));
121
122 static cl::opt<bool>
123 SinkCommon("simplifycfg-sink-common", cl::Hidden, cl::init(true),
124 cl::desc("Sink common instructions down to the end block"));
125
126 static cl::opt<bool> HoistCondStores(
127 "simplifycfg-hoist-cond-stores", cl::Hidden, cl::init(true),
128 cl::desc("Hoist conditional stores if an unconditional store precedes"));
129
130 static cl::opt<bool> MergeCondStores(
131 "simplifycfg-merge-cond-stores", cl::Hidden, cl::init(true),
132 cl::desc("Hoist conditional stores even if an unconditional store does not "
133 "precede - hoist multiple conditional stores into a single "
134 "predicated store"));
135
136 static cl::opt<bool> MergeCondStoresAggressively(
137 "simplifycfg-merge-cond-stores-aggressively", cl::Hidden, cl::init(false),
138 cl::desc("When merging conditional stores, do so even if the resultant "
139 "basic blocks are unlikely to be if-converted as a result"));
140
141 static cl::opt<bool> SpeculateOneExpensiveInst(
142 "speculate-one-expensive-inst", cl::Hidden, cl::init(true),
143 cl::desc("Allow exactly one expensive instruction to be speculatively "
144 "executed"));
145
146 static cl::opt<unsigned> MaxSpeculationDepth(
147 "max-speculation-depth", cl::Hidden, cl::init(10),
148 cl::desc("Limit maximum recursion depth when calculating costs of "
149 "speculatively executed instructions"));
150
151 static cl::opt<int>
152 MaxSmallBlockSize("simplifycfg-max-small-block-size", cl::Hidden, cl::init(10),
153 cl::desc("Max size of a block which is still considered "
154 "small enough to thread through"));
155
156 // Two is chosen to allow one negation and a logical combine.
157 static cl::opt<unsigned>
158 BranchFoldThreshold("simplifycfg-branch-fold-threshold", cl::Hidden,
159 cl::init(2),
160 cl::desc("Maximum cost of combining conditions when "
161 "folding branches"));
162
163 STATISTIC(NumBitMaps, "Number of switch instructions turned into bitmaps");
164 STATISTIC(NumLinearMaps,
165 "Number of switch instructions turned into linear mapping");
166 STATISTIC(NumLookupTables,
167 "Number of switch instructions turned into lookup tables");
168 STATISTIC(
169 NumLookupTablesHoles,
170 "Number of switch instructions turned into lookup tables (holes checked)");
171 STATISTIC(NumTableCmpReuses, "Number of reused switch table lookup compares");
172 STATISTIC(NumFoldValueComparisonIntoPredecessors,
173 "Number of value comparisons folded into predecessor basic blocks");
174 STATISTIC(NumFoldBranchToCommonDest,
175 "Number of branches folded into predecessor basic block");
176 STATISTIC(
177 NumHoistCommonCode,
178 "Number of common instruction 'blocks' hoisted up to the begin block");
179 STATISTIC(NumHoistCommonInstrs,
180 "Number of common instructions hoisted up to the begin block");
181 STATISTIC(NumSinkCommonCode,
182 "Number of common instruction 'blocks' sunk down to the end block");
183 STATISTIC(NumSinkCommonInstrs,
184 "Number of common instructions sunk down to the end block");
185 STATISTIC(NumSpeculations, "Number of speculative executed instructions");
186 STATISTIC(NumInvokes,
187 "Number of invokes with empty resume blocks simplified into calls");
188
189 namespace {
190
191 // The first field contains the value that the switch produces when a certain
192 // case group is selected, and the second field is a vector containing the
193 // cases composing the case group.
194 using SwitchCaseResultVectorTy =
195 SmallVector<std::pair<Constant *, SmallVector<ConstantInt *, 4>>, 2>;
196
197 // The first field contains the phi node that generates a result of the switch
198 // and the second field contains the value generated for a certain case in the
199 // switch for that PHI.
200 using SwitchCaseResultsTy = SmallVector<std::pair<PHINode *, Constant *>, 4>;
201
202 /// ValueEqualityComparisonCase - Represents a case of a switch.
203 struct ValueEqualityComparisonCase {
204 ConstantInt *Value;
205 BasicBlock *Dest;
206
ValueEqualityComparisonCase__anon5fcf8e030111::ValueEqualityComparisonCase207 ValueEqualityComparisonCase(ConstantInt *Value, BasicBlock *Dest)
208 : Value(Value), Dest(Dest) {}
209
operator <__anon5fcf8e030111::ValueEqualityComparisonCase210 bool operator<(ValueEqualityComparisonCase RHS) const {
211 // Comparing pointers is ok as we only rely on the order for uniquing.
212 return Value < RHS.Value;
213 }
214
operator ==__anon5fcf8e030111::ValueEqualityComparisonCase215 bool operator==(BasicBlock *RHSDest) const { return Dest == RHSDest; }
216 };
217
218 class SimplifyCFGOpt {
219 const TargetTransformInfo &TTI;
220 DomTreeUpdater *DTU;
221 const DataLayout &DL;
222 ArrayRef<WeakVH> LoopHeaders;
223 const SimplifyCFGOptions &Options;
224 bool Resimplify;
225
226 Value *isValueEqualityComparison(Instruction *TI);
227 BasicBlock *GetValueEqualityComparisonCases(
228 Instruction *TI, std::vector<ValueEqualityComparisonCase> &Cases);
229 bool SimplifyEqualityComparisonWithOnlyPredecessor(Instruction *TI,
230 BasicBlock *Pred,
231 IRBuilder<> &Builder);
232 bool PerformValueComparisonIntoPredecessorFolding(Instruction *TI, Value *&CV,
233 Instruction *PTI,
234 IRBuilder<> &Builder);
235 bool FoldValueComparisonIntoPredecessors(Instruction *TI,
236 IRBuilder<> &Builder);
237
238 bool simplifyReturn(ReturnInst *RI, IRBuilder<> &Builder);
239 bool simplifyResume(ResumeInst *RI, IRBuilder<> &Builder);
240 bool simplifySingleResume(ResumeInst *RI);
241 bool simplifyCommonResume(ResumeInst *RI);
242 bool simplifyCleanupReturn(CleanupReturnInst *RI);
243 bool simplifyUnreachable(UnreachableInst *UI);
244 bool simplifySwitch(SwitchInst *SI, IRBuilder<> &Builder);
245 bool simplifyIndirectBr(IndirectBrInst *IBI);
246 bool simplifyBranch(BranchInst *Branch, IRBuilder<> &Builder);
247 bool simplifyUncondBranch(BranchInst *BI, IRBuilder<> &Builder);
248 bool simplifyCondBranch(BranchInst *BI, IRBuilder<> &Builder);
249 bool SimplifyCondBranchToTwoReturns(BranchInst *BI, IRBuilder<> &Builder);
250
251 bool tryToSimplifyUncondBranchWithICmpInIt(ICmpInst *ICI,
252 IRBuilder<> &Builder);
253
254 bool HoistThenElseCodeToIf(BranchInst *BI, const TargetTransformInfo &TTI);
255 bool SpeculativelyExecuteBB(BranchInst *BI, BasicBlock *ThenBB,
256 const TargetTransformInfo &TTI);
257 bool SimplifyTerminatorOnSelect(Instruction *OldTerm, Value *Cond,
258 BasicBlock *TrueBB, BasicBlock *FalseBB,
259 uint32_t TrueWeight, uint32_t FalseWeight);
260 bool SimplifyBranchOnICmpChain(BranchInst *BI, IRBuilder<> &Builder,
261 const DataLayout &DL);
262 bool SimplifySwitchOnSelect(SwitchInst *SI, SelectInst *Select);
263 bool SimplifyIndirectBrOnSelect(IndirectBrInst *IBI, SelectInst *SI);
264 bool TurnSwitchRangeIntoICmp(SwitchInst *SI, IRBuilder<> &Builder);
265
266 public:
SimplifyCFGOpt(const TargetTransformInfo & TTI,DomTreeUpdater * DTU,const DataLayout & DL,ArrayRef<WeakVH> LoopHeaders,const SimplifyCFGOptions & Opts)267 SimplifyCFGOpt(const TargetTransformInfo &TTI, DomTreeUpdater *DTU,
268 const DataLayout &DL, ArrayRef<WeakVH> LoopHeaders,
269 const SimplifyCFGOptions &Opts)
270 : TTI(TTI), DTU(DTU), DL(DL), LoopHeaders(LoopHeaders), Options(Opts) {
271 assert((!DTU || !DTU->hasPostDomTree()) &&
272 "SimplifyCFG is not yet capable of maintaining validity of a "
273 "PostDomTree, so don't ask for it.");
274 }
275
276 bool simplifyOnce(BasicBlock *BB);
277 bool simplifyOnceImpl(BasicBlock *BB);
278 bool run(BasicBlock *BB);
279
280 // Helper to set Resimplify and return change indication.
requestResimplify()281 bool requestResimplify() {
282 Resimplify = true;
283 return true;
284 }
285 };
286
287 } // end anonymous namespace
288
289 /// Return true if it is safe to merge these two
290 /// terminator instructions together.
291 static bool
SafeToMergeTerminators(Instruction * SI1,Instruction * SI2,SmallSetVector<BasicBlock *,4> * FailBlocks=nullptr)292 SafeToMergeTerminators(Instruction *SI1, Instruction *SI2,
293 SmallSetVector<BasicBlock *, 4> *FailBlocks = nullptr) {
294 if (SI1 == SI2)
295 return false; // Can't merge with self!
296
297 // It is not safe to merge these two switch instructions if they have a common
298 // successor, and if that successor has a PHI node, and if *that* PHI node has
299 // conflicting incoming values from the two switch blocks.
300 BasicBlock *SI1BB = SI1->getParent();
301 BasicBlock *SI2BB = SI2->getParent();
302
303 SmallPtrSet<BasicBlock *, 16> SI1Succs(succ_begin(SI1BB), succ_end(SI1BB));
304 bool Fail = false;
305 for (BasicBlock *Succ : successors(SI2BB))
306 if (SI1Succs.count(Succ))
307 for (BasicBlock::iterator BBI = Succ->begin(); isa<PHINode>(BBI); ++BBI) {
308 PHINode *PN = cast<PHINode>(BBI);
309 if (PN->getIncomingValueForBlock(SI1BB) !=
310 PN->getIncomingValueForBlock(SI2BB)) {
311 if (FailBlocks)
312 FailBlocks->insert(Succ);
313 Fail = true;
314 }
315 }
316
317 return !Fail;
318 }
319
320 /// Update PHI nodes in Succ to indicate that there will now be entries in it
321 /// from the 'NewPred' block. The values that will be flowing into the PHI nodes
322 /// will be the same as those coming in from ExistPred, an existing predecessor
323 /// of Succ.
AddPredecessorToBlock(BasicBlock * Succ,BasicBlock * NewPred,BasicBlock * ExistPred,MemorySSAUpdater * MSSAU=nullptr)324 static void AddPredecessorToBlock(BasicBlock *Succ, BasicBlock *NewPred,
325 BasicBlock *ExistPred,
326 MemorySSAUpdater *MSSAU = nullptr) {
327 for (PHINode &PN : Succ->phis())
328 PN.addIncoming(PN.getIncomingValueForBlock(ExistPred), NewPred);
329 if (MSSAU)
330 if (auto *MPhi = MSSAU->getMemorySSA()->getMemoryAccess(Succ))
331 MPhi->addIncoming(MPhi->getIncomingValueForBlock(ExistPred), NewPred);
332 }
333
334 /// Compute an abstract "cost" of speculating the given instruction,
335 /// which is assumed to be safe to speculate. TCC_Free means cheap,
336 /// TCC_Basic means less cheap, and TCC_Expensive means prohibitively
337 /// expensive.
ComputeSpeculationCost(const User * I,const TargetTransformInfo & TTI)338 static unsigned ComputeSpeculationCost(const User *I,
339 const TargetTransformInfo &TTI) {
340 assert(isSafeToSpeculativelyExecute(I) &&
341 "Instruction is not safe to speculatively execute!");
342 return TTI.getUserCost(I, TargetTransformInfo::TCK_SizeAndLatency);
343 }
344
345 /// If we have a merge point of an "if condition" as accepted above,
346 /// return true if the specified value dominates the block. We
347 /// don't handle the true generality of domination here, just a special case
348 /// which works well enough for us.
349 ///
350 /// If AggressiveInsts is non-null, and if V does not dominate BB, we check to
351 /// see if V (which must be an instruction) and its recursive operands
352 /// that do not dominate BB have a combined cost lower than CostRemaining and
353 /// are non-trapping. If both are true, the instruction is inserted into the
354 /// set and true is returned.
355 ///
356 /// The cost for most non-trapping instructions is defined as 1 except for
357 /// Select whose cost is 2.
358 ///
359 /// After this function returns, CostRemaining is decreased by the cost of
360 /// V plus its non-dominating operands. If that cost is greater than
361 /// CostRemaining, false is returned and CostRemaining is undefined.
DominatesMergePoint(Value * V,BasicBlock * BB,SmallPtrSetImpl<Instruction * > & AggressiveInsts,int & BudgetRemaining,const TargetTransformInfo & TTI,unsigned Depth=0)362 static bool DominatesMergePoint(Value *V, BasicBlock *BB,
363 SmallPtrSetImpl<Instruction *> &AggressiveInsts,
364 int &BudgetRemaining,
365 const TargetTransformInfo &TTI,
366 unsigned Depth = 0) {
367 // It is possible to hit a zero-cost cycle (phi/gep instructions for example),
368 // so limit the recursion depth.
369 // TODO: While this recursion limit does prevent pathological behavior, it
370 // would be better to track visited instructions to avoid cycles.
371 if (Depth == MaxSpeculationDepth)
372 return false;
373
374 Instruction *I = dyn_cast<Instruction>(V);
375 if (!I) {
376 // Non-instructions all dominate instructions, but not all constantexprs
377 // can be executed unconditionally.
378 if (ConstantExpr *C = dyn_cast<ConstantExpr>(V))
379 if (C->canTrap())
380 return false;
381 return true;
382 }
383 BasicBlock *PBB = I->getParent();
384
385 // We don't want to allow weird loops that might have the "if condition" in
386 // the bottom of this block.
387 if (PBB == BB)
388 return false;
389
390 // If this instruction is defined in a block that contains an unconditional
391 // branch to BB, then it must be in the 'conditional' part of the "if
392 // statement". If not, it definitely dominates the region.
393 BranchInst *BI = dyn_cast<BranchInst>(PBB->getTerminator());
394 if (!BI || BI->isConditional() || BI->getSuccessor(0) != BB)
395 return true;
396
397 // If we have seen this instruction before, don't count it again.
398 if (AggressiveInsts.count(I))
399 return true;
400
401 // Okay, it looks like the instruction IS in the "condition". Check to
402 // see if it's a cheap instruction to unconditionally compute, and if it
403 // only uses stuff defined outside of the condition. If so, hoist it out.
404 if (!isSafeToSpeculativelyExecute(I))
405 return false;
406
407 BudgetRemaining -= ComputeSpeculationCost(I, TTI);
408
409 // Allow exactly one instruction to be speculated regardless of its cost
410 // (as long as it is safe to do so).
411 // This is intended to flatten the CFG even if the instruction is a division
412 // or other expensive operation. The speculation of an expensive instruction
413 // is expected to be undone in CodeGenPrepare if the speculation has not
414 // enabled further IR optimizations.
415 if (BudgetRemaining < 0 &&
416 (!SpeculateOneExpensiveInst || !AggressiveInsts.empty() || Depth > 0))
417 return false;
418
419 // Okay, we can only really hoist these out if their operands do
420 // not take us over the cost threshold.
421 for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i)
422 if (!DominatesMergePoint(*i, BB, AggressiveInsts, BudgetRemaining, TTI,
423 Depth + 1))
424 return false;
425 // Okay, it's safe to do this! Remember this instruction.
426 AggressiveInsts.insert(I);
427 return true;
428 }
429
430 /// Extract ConstantInt from value, looking through IntToPtr
431 /// and PointerNullValue. Return NULL if value is not a constant int.
GetConstantInt(Value * V,const DataLayout & DL)432 static ConstantInt *GetConstantInt(Value *V, const DataLayout &DL) {
433 // Normal constant int.
434 ConstantInt *CI = dyn_cast<ConstantInt>(V);
435 if (CI || !isa<Constant>(V) || !V->getType()->isPointerTy())
436 return CI;
437
438 // This is some kind of pointer constant. Turn it into a pointer-sized
439 // ConstantInt if possible.
440 IntegerType *PtrTy = cast<IntegerType>(DL.getIntPtrType(V->getType()));
441
442 // Null pointer means 0, see SelectionDAGBuilder::getValue(const Value*).
443 if (isa<ConstantPointerNull>(V))
444 return ConstantInt::get(PtrTy, 0);
445
446 // IntToPtr const int.
447 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
448 if (CE->getOpcode() == Instruction::IntToPtr)
449 if (ConstantInt *CI = dyn_cast<ConstantInt>(CE->getOperand(0))) {
450 // The constant is very likely to have the right type already.
451 if (CI->getType() == PtrTy)
452 return CI;
453 else
454 return cast<ConstantInt>(
455 ConstantExpr::getIntegerCast(CI, PtrTy, /*isSigned=*/false));
456 }
457 return nullptr;
458 }
459
460 namespace {
461
462 /// Given a chain of or (||) or and (&&) comparison of a value against a
463 /// constant, this will try to recover the information required for a switch
464 /// structure.
465 /// It will depth-first traverse the chain of comparison, seeking for patterns
466 /// like %a == 12 or %a < 4 and combine them to produce a set of integer
467 /// representing the different cases for the switch.
468 /// Note that if the chain is composed of '||' it will build the set of elements
469 /// that matches the comparisons (i.e. any of this value validate the chain)
470 /// while for a chain of '&&' it will build the set elements that make the test
471 /// fail.
472 struct ConstantComparesGatherer {
473 const DataLayout &DL;
474
475 /// Value found for the switch comparison
476 Value *CompValue = nullptr;
477
478 /// Extra clause to be checked before the switch
479 Value *Extra = nullptr;
480
481 /// Set of integers to match in switch
482 SmallVector<ConstantInt *, 8> Vals;
483
484 /// Number of comparisons matched in the and/or chain
485 unsigned UsedICmps = 0;
486
487 /// Construct and compute the result for the comparison instruction Cond
ConstantComparesGatherer__anon5fcf8e030211::ConstantComparesGatherer488 ConstantComparesGatherer(Instruction *Cond, const DataLayout &DL) : DL(DL) {
489 gather(Cond);
490 }
491
492 ConstantComparesGatherer(const ConstantComparesGatherer &) = delete;
493 ConstantComparesGatherer &
494 operator=(const ConstantComparesGatherer &) = delete;
495
496 private:
497 /// Try to set the current value used for the comparison, it succeeds only if
498 /// it wasn't set before or if the new value is the same as the old one
setValueOnce__anon5fcf8e030211::ConstantComparesGatherer499 bool setValueOnce(Value *NewVal) {
500 if (CompValue && CompValue != NewVal)
501 return false;
502 CompValue = NewVal;
503 return (CompValue != nullptr);
504 }
505
506 /// Try to match Instruction "I" as a comparison against a constant and
507 /// populates the array Vals with the set of values that match (or do not
508 /// match depending on isEQ).
509 /// Return false on failure. On success, the Value the comparison matched
510 /// against is placed in CompValue.
511 /// If CompValue is already set, the function is expected to fail if a match
512 /// is found but the value compared to is different.
matchInstruction__anon5fcf8e030211::ConstantComparesGatherer513 bool matchInstruction(Instruction *I, bool isEQ) {
514 // If this is an icmp against a constant, handle this as one of the cases.
515 ICmpInst *ICI;
516 ConstantInt *C;
517 if (!((ICI = dyn_cast<ICmpInst>(I)) &&
518 (C = GetConstantInt(I->getOperand(1), DL)))) {
519 return false;
520 }
521
522 Value *RHSVal;
523 const APInt *RHSC;
524
525 // Pattern match a special case
526 // (x & ~2^z) == y --> x == y || x == y|2^z
527 // This undoes a transformation done by instcombine to fuse 2 compares.
528 if (ICI->getPredicate() == (isEQ ? ICmpInst::ICMP_EQ : ICmpInst::ICMP_NE)) {
529 // It's a little bit hard to see why the following transformations are
530 // correct. Here is a CVC3 program to verify them for 64-bit values:
531
532 /*
533 ONE : BITVECTOR(64) = BVZEROEXTEND(0bin1, 63);
534 x : BITVECTOR(64);
535 y : BITVECTOR(64);
536 z : BITVECTOR(64);
537 mask : BITVECTOR(64) = BVSHL(ONE, z);
538 QUERY( (y & ~mask = y) =>
539 ((x & ~mask = y) <=> (x = y OR x = (y | mask)))
540 );
541 QUERY( (y | mask = y) =>
542 ((x | mask = y) <=> (x = y OR x = (y & ~mask)))
543 );
544 */
545
546 // Please note that each pattern must be a dual implication (<--> or
547 // iff). One directional implication can create spurious matches. If the
548 // implication is only one-way, an unsatisfiable condition on the left
549 // side can imply a satisfiable condition on the right side. Dual
550 // implication ensures that satisfiable conditions are transformed to
551 // other satisfiable conditions and unsatisfiable conditions are
552 // transformed to other unsatisfiable conditions.
553
554 // Here is a concrete example of a unsatisfiable condition on the left
555 // implying a satisfiable condition on the right:
556 //
557 // mask = (1 << z)
558 // (x & ~mask) == y --> (x == y || x == (y | mask))
559 //
560 // Substituting y = 3, z = 0 yields:
561 // (x & -2) == 3 --> (x == 3 || x == 2)
562
563 // Pattern match a special case:
564 /*
565 QUERY( (y & ~mask = y) =>
566 ((x & ~mask = y) <=> (x = y OR x = (y | mask)))
567 );
568 */
569 if (match(ICI->getOperand(0),
570 m_And(m_Value(RHSVal), m_APInt(RHSC)))) {
571 APInt Mask = ~*RHSC;
572 if (Mask.isPowerOf2() && (C->getValue() & ~Mask) == C->getValue()) {
573 // If we already have a value for the switch, it has to match!
574 if (!setValueOnce(RHSVal))
575 return false;
576
577 Vals.push_back(C);
578 Vals.push_back(
579 ConstantInt::get(C->getContext(),
580 C->getValue() | Mask));
581 UsedICmps++;
582 return true;
583 }
584 }
585
586 // Pattern match a special case:
587 /*
588 QUERY( (y | mask = y) =>
589 ((x | mask = y) <=> (x = y OR x = (y & ~mask)))
590 );
591 */
592 if (match(ICI->getOperand(0),
593 m_Or(m_Value(RHSVal), m_APInt(RHSC)))) {
594 APInt Mask = *RHSC;
595 if (Mask.isPowerOf2() && (C->getValue() | Mask) == C->getValue()) {
596 // If we already have a value for the switch, it has to match!
597 if (!setValueOnce(RHSVal))
598 return false;
599
600 Vals.push_back(C);
601 Vals.push_back(ConstantInt::get(C->getContext(),
602 C->getValue() & ~Mask));
603 UsedICmps++;
604 return true;
605 }
606 }
607
608 // If we already have a value for the switch, it has to match!
609 if (!setValueOnce(ICI->getOperand(0)))
610 return false;
611
612 UsedICmps++;
613 Vals.push_back(C);
614 return ICI->getOperand(0);
615 }
616
617 // If we have "x ult 3", for example, then we can add 0,1,2 to the set.
618 ConstantRange Span = ConstantRange::makeAllowedICmpRegion(
619 ICI->getPredicate(), C->getValue());
620
621 // Shift the range if the compare is fed by an add. This is the range
622 // compare idiom as emitted by instcombine.
623 Value *CandidateVal = I->getOperand(0);
624 if (match(I->getOperand(0), m_Add(m_Value(RHSVal), m_APInt(RHSC)))) {
625 Span = Span.subtract(*RHSC);
626 CandidateVal = RHSVal;
627 }
628
629 // If this is an and/!= check, then we are looking to build the set of
630 // value that *don't* pass the and chain. I.e. to turn "x ugt 2" into
631 // x != 0 && x != 1.
632 if (!isEQ)
633 Span = Span.inverse();
634
635 // If there are a ton of values, we don't want to make a ginormous switch.
636 if (Span.isSizeLargerThan(8) || Span.isEmptySet()) {
637 return false;
638 }
639
640 // If we already have a value for the switch, it has to match!
641 if (!setValueOnce(CandidateVal))
642 return false;
643
644 // Add all values from the range to the set
645 for (APInt Tmp = Span.getLower(); Tmp != Span.getUpper(); ++Tmp)
646 Vals.push_back(ConstantInt::get(I->getContext(), Tmp));
647
648 UsedICmps++;
649 return true;
650 }
651
652 /// Given a potentially 'or'd or 'and'd together collection of icmp
653 /// eq/ne/lt/gt instructions that compare a value against a constant, extract
654 /// the value being compared, and stick the list constants into the Vals
655 /// vector.
656 /// One "Extra" case is allowed to differ from the other.
gather__anon5fcf8e030211::ConstantComparesGatherer657 void gather(Value *V) {
658 bool isEQ = match(V, m_LogicalOr(m_Value(), m_Value()));
659
660 // Keep a stack (SmallVector for efficiency) for depth-first traversal
661 SmallVector<Value *, 8> DFT;
662 SmallPtrSet<Value *, 8> Visited;
663
664 // Initialize
665 Visited.insert(V);
666 DFT.push_back(V);
667
668 while (!DFT.empty()) {
669 V = DFT.pop_back_val();
670
671 if (Instruction *I = dyn_cast<Instruction>(V)) {
672 // If it is a || (or && depending on isEQ), process the operands.
673 Value *Op0, *Op1;
674 if (isEQ ? match(I, m_LogicalOr(m_Value(Op0), m_Value(Op1)))
675 : match(I, m_LogicalAnd(m_Value(Op0), m_Value(Op1)))) {
676 if (Visited.insert(Op1).second)
677 DFT.push_back(Op1);
678 if (Visited.insert(Op0).second)
679 DFT.push_back(Op0);
680
681 continue;
682 }
683
684 // Try to match the current instruction
685 if (matchInstruction(I, isEQ))
686 // Match succeed, continue the loop
687 continue;
688 }
689
690 // One element of the sequence of || (or &&) could not be match as a
691 // comparison against the same value as the others.
692 // We allow only one "Extra" case to be checked before the switch
693 if (!Extra) {
694 Extra = V;
695 continue;
696 }
697 // Failed to parse a proper sequence, abort now
698 CompValue = nullptr;
699 break;
700 }
701 }
702 };
703
704 } // end anonymous namespace
705
EraseTerminatorAndDCECond(Instruction * TI,MemorySSAUpdater * MSSAU=nullptr)706 static void EraseTerminatorAndDCECond(Instruction *TI,
707 MemorySSAUpdater *MSSAU = nullptr) {
708 Instruction *Cond = nullptr;
709 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
710 Cond = dyn_cast<Instruction>(SI->getCondition());
711 } else if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
712 if (BI->isConditional())
713 Cond = dyn_cast<Instruction>(BI->getCondition());
714 } else if (IndirectBrInst *IBI = dyn_cast<IndirectBrInst>(TI)) {
715 Cond = dyn_cast<Instruction>(IBI->getAddress());
716 }
717
718 TI->eraseFromParent();
719 if (Cond)
720 RecursivelyDeleteTriviallyDeadInstructions(Cond, nullptr, MSSAU);
721 }
722
723 /// Return true if the specified terminator checks
724 /// to see if a value is equal to constant integer value.
isValueEqualityComparison(Instruction * TI)725 Value *SimplifyCFGOpt::isValueEqualityComparison(Instruction *TI) {
726 Value *CV = nullptr;
727 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
728 // Do not permit merging of large switch instructions into their
729 // predecessors unless there is only one predecessor.
730 if (!SI->getParent()->hasNPredecessorsOrMore(128 / SI->getNumSuccessors()))
731 CV = SI->getCondition();
732 } else if (BranchInst *BI = dyn_cast<BranchInst>(TI))
733 if (BI->isConditional() && BI->getCondition()->hasOneUse())
734 if (ICmpInst *ICI = dyn_cast<ICmpInst>(BI->getCondition())) {
735 if (ICI->isEquality() && GetConstantInt(ICI->getOperand(1), DL))
736 CV = ICI->getOperand(0);
737 }
738
739 // Unwrap any lossless ptrtoint cast.
740 if (CV) {
741 if (PtrToIntInst *PTII = dyn_cast<PtrToIntInst>(CV)) {
742 Value *Ptr = PTII->getPointerOperand();
743 if (PTII->getType() == DL.getIntPtrType(Ptr->getType()))
744 CV = Ptr;
745 }
746 }
747 return CV;
748 }
749
750 /// Given a value comparison instruction,
751 /// decode all of the 'cases' that it represents and return the 'default' block.
GetValueEqualityComparisonCases(Instruction * TI,std::vector<ValueEqualityComparisonCase> & Cases)752 BasicBlock *SimplifyCFGOpt::GetValueEqualityComparisonCases(
753 Instruction *TI, std::vector<ValueEqualityComparisonCase> &Cases) {
754 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
755 Cases.reserve(SI->getNumCases());
756 for (auto Case : SI->cases())
757 Cases.push_back(ValueEqualityComparisonCase(Case.getCaseValue(),
758 Case.getCaseSuccessor()));
759 return SI->getDefaultDest();
760 }
761
762 BranchInst *BI = cast<BranchInst>(TI);
763 ICmpInst *ICI = cast<ICmpInst>(BI->getCondition());
764 BasicBlock *Succ = BI->getSuccessor(ICI->getPredicate() == ICmpInst::ICMP_NE);
765 Cases.push_back(ValueEqualityComparisonCase(
766 GetConstantInt(ICI->getOperand(1), DL), Succ));
767 return BI->getSuccessor(ICI->getPredicate() == ICmpInst::ICMP_EQ);
768 }
769
770 /// Given a vector of bb/value pairs, remove any entries
771 /// in the list that match the specified block.
772 static void
EliminateBlockCases(BasicBlock * BB,std::vector<ValueEqualityComparisonCase> & Cases)773 EliminateBlockCases(BasicBlock *BB,
774 std::vector<ValueEqualityComparisonCase> &Cases) {
775 llvm::erase_value(Cases, BB);
776 }
777
778 /// Return true if there are any keys in C1 that exist in C2 as well.
ValuesOverlap(std::vector<ValueEqualityComparisonCase> & C1,std::vector<ValueEqualityComparisonCase> & C2)779 static bool ValuesOverlap(std::vector<ValueEqualityComparisonCase> &C1,
780 std::vector<ValueEqualityComparisonCase> &C2) {
781 std::vector<ValueEqualityComparisonCase> *V1 = &C1, *V2 = &C2;
782
783 // Make V1 be smaller than V2.
784 if (V1->size() > V2->size())
785 std::swap(V1, V2);
786
787 if (V1->empty())
788 return false;
789 if (V1->size() == 1) {
790 // Just scan V2.
791 ConstantInt *TheVal = (*V1)[0].Value;
792 for (unsigned i = 0, e = V2->size(); i != e; ++i)
793 if (TheVal == (*V2)[i].Value)
794 return true;
795 }
796
797 // Otherwise, just sort both lists and compare element by element.
798 array_pod_sort(V1->begin(), V1->end());
799 array_pod_sort(V2->begin(), V2->end());
800 unsigned i1 = 0, i2 = 0, e1 = V1->size(), e2 = V2->size();
801 while (i1 != e1 && i2 != e2) {
802 if ((*V1)[i1].Value == (*V2)[i2].Value)
803 return true;
804 if ((*V1)[i1].Value < (*V2)[i2].Value)
805 ++i1;
806 else
807 ++i2;
808 }
809 return false;
810 }
811
812 // Set branch weights on SwitchInst. This sets the metadata if there is at
813 // least one non-zero weight.
setBranchWeights(SwitchInst * SI,ArrayRef<uint32_t> Weights)814 static void setBranchWeights(SwitchInst *SI, ArrayRef<uint32_t> Weights) {
815 // Check that there is at least one non-zero weight. Otherwise, pass
816 // nullptr to setMetadata which will erase the existing metadata.
817 MDNode *N = nullptr;
818 if (llvm::any_of(Weights, [](uint32_t W) { return W != 0; }))
819 N = MDBuilder(SI->getParent()->getContext()).createBranchWeights(Weights);
820 SI->setMetadata(LLVMContext::MD_prof, N);
821 }
822
823 // Similar to the above, but for branch and select instructions that take
824 // exactly 2 weights.
setBranchWeights(Instruction * I,uint32_t TrueWeight,uint32_t FalseWeight)825 static void setBranchWeights(Instruction *I, uint32_t TrueWeight,
826 uint32_t FalseWeight) {
827 assert(isa<BranchInst>(I) || isa<SelectInst>(I));
828 // Check that there is at least one non-zero weight. Otherwise, pass
829 // nullptr to setMetadata which will erase the existing metadata.
830 MDNode *N = nullptr;
831 if (TrueWeight || FalseWeight)
832 N = MDBuilder(I->getParent()->getContext())
833 .createBranchWeights(TrueWeight, FalseWeight);
834 I->setMetadata(LLVMContext::MD_prof, N);
835 }
836
837 /// If TI is known to be a terminator instruction and its block is known to
838 /// only have a single predecessor block, check to see if that predecessor is
839 /// also a value comparison with the same value, and if that comparison
840 /// determines the outcome of this comparison. If so, simplify TI. This does a
841 /// very limited form of jump threading.
SimplifyEqualityComparisonWithOnlyPredecessor(Instruction * TI,BasicBlock * Pred,IRBuilder<> & Builder)842 bool SimplifyCFGOpt::SimplifyEqualityComparisonWithOnlyPredecessor(
843 Instruction *TI, BasicBlock *Pred, IRBuilder<> &Builder) {
844 Value *PredVal = isValueEqualityComparison(Pred->getTerminator());
845 if (!PredVal)
846 return false; // Not a value comparison in predecessor.
847
848 Value *ThisVal = isValueEqualityComparison(TI);
849 assert(ThisVal && "This isn't a value comparison!!");
850 if (ThisVal != PredVal)
851 return false; // Different predicates.
852
853 // TODO: Preserve branch weight metadata, similarly to how
854 // FoldValueComparisonIntoPredecessors preserves it.
855
856 // Find out information about when control will move from Pred to TI's block.
857 std::vector<ValueEqualityComparisonCase> PredCases;
858 BasicBlock *PredDef =
859 GetValueEqualityComparisonCases(Pred->getTerminator(), PredCases);
860 EliminateBlockCases(PredDef, PredCases); // Remove default from cases.
861
862 // Find information about how control leaves this block.
863 std::vector<ValueEqualityComparisonCase> ThisCases;
864 BasicBlock *ThisDef = GetValueEqualityComparisonCases(TI, ThisCases);
865 EliminateBlockCases(ThisDef, ThisCases); // Remove default from cases.
866
867 // If TI's block is the default block from Pred's comparison, potentially
868 // simplify TI based on this knowledge.
869 if (PredDef == TI->getParent()) {
870 // If we are here, we know that the value is none of those cases listed in
871 // PredCases. If there are any cases in ThisCases that are in PredCases, we
872 // can simplify TI.
873 if (!ValuesOverlap(PredCases, ThisCases))
874 return false;
875
876 if (isa<BranchInst>(TI)) {
877 // Okay, one of the successors of this condbr is dead. Convert it to a
878 // uncond br.
879 assert(ThisCases.size() == 1 && "Branch can only have one case!");
880 // Insert the new branch.
881 Instruction *NI = Builder.CreateBr(ThisDef);
882 (void)NI;
883
884 // Remove PHI node entries for the dead edge.
885 ThisCases[0].Dest->removePredecessor(PredDef);
886
887 LLVM_DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator()
888 << "Through successor TI: " << *TI << "Leaving: " << *NI
889 << "\n");
890
891 EraseTerminatorAndDCECond(TI);
892
893 if (DTU)
894 DTU->applyUpdates(
895 {{DominatorTree::Delete, PredDef, ThisCases[0].Dest}});
896
897 return true;
898 }
899
900 SwitchInstProfUpdateWrapper SI = *cast<SwitchInst>(TI);
901 // Okay, TI has cases that are statically dead, prune them away.
902 SmallPtrSet<Constant *, 16> DeadCases;
903 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
904 DeadCases.insert(PredCases[i].Value);
905
906 LLVM_DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator()
907 << "Through successor TI: " << *TI);
908
909 SmallMapVector<BasicBlock *, int, 8> NumPerSuccessorCases;
910 for (SwitchInst::CaseIt i = SI->case_end(), e = SI->case_begin(); i != e;) {
911 --i;
912 auto *Successor = i->getCaseSuccessor();
913 ++NumPerSuccessorCases[Successor];
914 if (DeadCases.count(i->getCaseValue())) {
915 Successor->removePredecessor(PredDef);
916 SI.removeCase(i);
917 --NumPerSuccessorCases[Successor];
918 }
919 }
920
921 std::vector<DominatorTree::UpdateType> Updates;
922 for (const std::pair<BasicBlock *, int> &I : NumPerSuccessorCases)
923 if (I.second == 0)
924 Updates.push_back({DominatorTree::Delete, PredDef, I.first});
925 if (DTU)
926 DTU->applyUpdates(Updates);
927
928 LLVM_DEBUG(dbgs() << "Leaving: " << *TI << "\n");
929 return true;
930 }
931
932 // Otherwise, TI's block must correspond to some matched value. Find out
933 // which value (or set of values) this is.
934 ConstantInt *TIV = nullptr;
935 BasicBlock *TIBB = TI->getParent();
936 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
937 if (PredCases[i].Dest == TIBB) {
938 if (TIV)
939 return false; // Cannot handle multiple values coming to this block.
940 TIV = PredCases[i].Value;
941 }
942 assert(TIV && "No edge from pred to succ?");
943
944 // Okay, we found the one constant that our value can be if we get into TI's
945 // BB. Find out which successor will unconditionally be branched to.
946 BasicBlock *TheRealDest = nullptr;
947 for (unsigned i = 0, e = ThisCases.size(); i != e; ++i)
948 if (ThisCases[i].Value == TIV) {
949 TheRealDest = ThisCases[i].Dest;
950 break;
951 }
952
953 // If not handled by any explicit cases, it is handled by the default case.
954 if (!TheRealDest)
955 TheRealDest = ThisDef;
956
957 SmallSetVector<BasicBlock *, 2> RemovedSuccs;
958
959 // Remove PHI node entries for dead edges.
960 BasicBlock *CheckEdge = TheRealDest;
961 for (BasicBlock *Succ : successors(TIBB))
962 if (Succ != CheckEdge) {
963 if (Succ != TheRealDest)
964 RemovedSuccs.insert(Succ);
965 Succ->removePredecessor(TIBB);
966 } else
967 CheckEdge = nullptr;
968
969 // Insert the new branch.
970 Instruction *NI = Builder.CreateBr(TheRealDest);
971 (void)NI;
972
973 LLVM_DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator()
974 << "Through successor TI: " << *TI << "Leaving: " << *NI
975 << "\n");
976
977 EraseTerminatorAndDCECond(TI);
978 if (DTU) {
979 SmallVector<DominatorTree::UpdateType, 2> Updates;
980 Updates.reserve(RemovedSuccs.size());
981 for (auto *RemovedSucc : RemovedSuccs)
982 Updates.push_back({DominatorTree::Delete, TIBB, RemovedSucc});
983 DTU->applyUpdates(Updates);
984 }
985 return true;
986 }
987
988 namespace {
989
990 /// This class implements a stable ordering of constant
991 /// integers that does not depend on their address. This is important for
992 /// applications that sort ConstantInt's to ensure uniqueness.
993 struct ConstantIntOrdering {
operator ()__anon5fcf8e030411::ConstantIntOrdering994 bool operator()(const ConstantInt *LHS, const ConstantInt *RHS) const {
995 return LHS->getValue().ult(RHS->getValue());
996 }
997 };
998
999 } // end anonymous namespace
1000
ConstantIntSortPredicate(ConstantInt * const * P1,ConstantInt * const * P2)1001 static int ConstantIntSortPredicate(ConstantInt *const *P1,
1002 ConstantInt *const *P2) {
1003 const ConstantInt *LHS = *P1;
1004 const ConstantInt *RHS = *P2;
1005 if (LHS == RHS)
1006 return 0;
1007 return LHS->getValue().ult(RHS->getValue()) ? 1 : -1;
1008 }
1009
HasBranchWeights(const Instruction * I)1010 static inline bool HasBranchWeights(const Instruction *I) {
1011 MDNode *ProfMD = I->getMetadata(LLVMContext::MD_prof);
1012 if (ProfMD && ProfMD->getOperand(0))
1013 if (MDString *MDS = dyn_cast<MDString>(ProfMD->getOperand(0)))
1014 return MDS->getString().equals("branch_weights");
1015
1016 return false;
1017 }
1018
1019 /// Get Weights of a given terminator, the default weight is at the front
1020 /// of the vector. If TI is a conditional eq, we need to swap the branch-weight
1021 /// metadata.
GetBranchWeights(Instruction * TI,SmallVectorImpl<uint64_t> & Weights)1022 static void GetBranchWeights(Instruction *TI,
1023 SmallVectorImpl<uint64_t> &Weights) {
1024 MDNode *MD = TI->getMetadata(LLVMContext::MD_prof);
1025 assert(MD);
1026 for (unsigned i = 1, e = MD->getNumOperands(); i < e; ++i) {
1027 ConstantInt *CI = mdconst::extract<ConstantInt>(MD->getOperand(i));
1028 Weights.push_back(CI->getValue().getZExtValue());
1029 }
1030
1031 // If TI is a conditional eq, the default case is the false case,
1032 // and the corresponding branch-weight data is at index 2. We swap the
1033 // default weight to be the first entry.
1034 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
1035 assert(Weights.size() == 2);
1036 ICmpInst *ICI = cast<ICmpInst>(BI->getCondition());
1037 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
1038 std::swap(Weights.front(), Weights.back());
1039 }
1040 }
1041
1042 /// Keep halving the weights until all can fit in uint32_t.
FitWeights(MutableArrayRef<uint64_t> Weights)1043 static void FitWeights(MutableArrayRef<uint64_t> Weights) {
1044 uint64_t Max = *std::max_element(Weights.begin(), Weights.end());
1045 if (Max > UINT_MAX) {
1046 unsigned Offset = 32 - countLeadingZeros(Max);
1047 for (uint64_t &I : Weights)
1048 I >>= Offset;
1049 }
1050 }
1051
CloneInstructionsIntoPredecessorBlockAndUpdateSSAUses(BasicBlock * BB,BasicBlock * PredBlock,ValueToValueMapTy & VMap)1052 static void CloneInstructionsIntoPredecessorBlockAndUpdateSSAUses(
1053 BasicBlock *BB, BasicBlock *PredBlock, ValueToValueMapTy &VMap) {
1054 Instruction *PTI = PredBlock->getTerminator();
1055
1056 // If we have bonus instructions, clone them into the predecessor block.
1057 // Note that there may be multiple predecessor blocks, so we cannot move
1058 // bonus instructions to a predecessor block.
1059 for (Instruction &BonusInst : *BB) {
1060 if (isa<DbgInfoIntrinsic>(BonusInst) || BonusInst.isTerminator())
1061 continue;
1062
1063 Instruction *NewBonusInst = BonusInst.clone();
1064
1065 if (PTI->getDebugLoc() != NewBonusInst->getDebugLoc()) {
1066 // Unless the instruction has the same !dbg location as the original
1067 // branch, drop it. When we fold the bonus instructions we want to make
1068 // sure we reset their debug locations in order to avoid stepping on
1069 // dead code caused by folding dead branches.
1070 NewBonusInst->setDebugLoc(DebugLoc());
1071 }
1072
1073 RemapInstruction(NewBonusInst, VMap,
1074 RF_NoModuleLevelChanges | RF_IgnoreMissingLocals);
1075 VMap[&BonusInst] = NewBonusInst;
1076
1077 // If we moved a load, we cannot any longer claim any knowledge about
1078 // its potential value. The previous information might have been valid
1079 // only given the branch precondition.
1080 // For an analogous reason, we must also drop all the metadata whose
1081 // semantics we don't understand. We *can* preserve !annotation, because
1082 // it is tied to the instruction itself, not the value or position.
1083 NewBonusInst->dropUnknownNonDebugMetadata(LLVMContext::MD_annotation);
1084
1085 PredBlock->getInstList().insert(PTI->getIterator(), NewBonusInst);
1086 NewBonusInst->takeName(&BonusInst);
1087 BonusInst.setName(NewBonusInst->getName() + ".old");
1088
1089 // Update (liveout) uses of bonus instructions,
1090 // now that the bonus instruction has been cloned into predecessor.
1091 SSAUpdater SSAUpdate;
1092 SSAUpdate.Initialize(BonusInst.getType(),
1093 (NewBonusInst->getName() + ".merge").str());
1094 SSAUpdate.AddAvailableValue(BB, &BonusInst);
1095 SSAUpdate.AddAvailableValue(PredBlock, NewBonusInst);
1096 for (Use &U : make_early_inc_range(BonusInst.uses()))
1097 SSAUpdate.RewriteUseAfterInsertions(U);
1098 }
1099 }
1100
PerformValueComparisonIntoPredecessorFolding(Instruction * TI,Value * & CV,Instruction * PTI,IRBuilder<> & Builder)1101 bool SimplifyCFGOpt::PerformValueComparisonIntoPredecessorFolding(
1102 Instruction *TI, Value *&CV, Instruction *PTI, IRBuilder<> &Builder) {
1103 BasicBlock *BB = TI->getParent();
1104 BasicBlock *Pred = PTI->getParent();
1105
1106 std::vector<DominatorTree::UpdateType> Updates;
1107
1108 // Figure out which 'cases' to copy from SI to PSI.
1109 std::vector<ValueEqualityComparisonCase> BBCases;
1110 BasicBlock *BBDefault = GetValueEqualityComparisonCases(TI, BBCases);
1111
1112 std::vector<ValueEqualityComparisonCase> PredCases;
1113 BasicBlock *PredDefault = GetValueEqualityComparisonCases(PTI, PredCases);
1114
1115 // Based on whether the default edge from PTI goes to BB or not, fill in
1116 // PredCases and PredDefault with the new switch cases we would like to
1117 // build.
1118 SmallMapVector<BasicBlock *, int, 8> NewSuccessors;
1119
1120 // Update the branch weight metadata along the way
1121 SmallVector<uint64_t, 8> Weights;
1122 bool PredHasWeights = HasBranchWeights(PTI);
1123 bool SuccHasWeights = HasBranchWeights(TI);
1124
1125 if (PredHasWeights) {
1126 GetBranchWeights(PTI, Weights);
1127 // branch-weight metadata is inconsistent here.
1128 if (Weights.size() != 1 + PredCases.size())
1129 PredHasWeights = SuccHasWeights = false;
1130 } else if (SuccHasWeights)
1131 // If there are no predecessor weights but there are successor weights,
1132 // populate Weights with 1, which will later be scaled to the sum of
1133 // successor's weights
1134 Weights.assign(1 + PredCases.size(), 1);
1135
1136 SmallVector<uint64_t, 8> SuccWeights;
1137 if (SuccHasWeights) {
1138 GetBranchWeights(TI, SuccWeights);
1139 // branch-weight metadata is inconsistent here.
1140 if (SuccWeights.size() != 1 + BBCases.size())
1141 PredHasWeights = SuccHasWeights = false;
1142 } else if (PredHasWeights)
1143 SuccWeights.assign(1 + BBCases.size(), 1);
1144
1145 if (PredDefault == BB) {
1146 // If this is the default destination from PTI, only the edges in TI
1147 // that don't occur in PTI, or that branch to BB will be activated.
1148 std::set<ConstantInt *, ConstantIntOrdering> PTIHandled;
1149 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
1150 if (PredCases[i].Dest != BB)
1151 PTIHandled.insert(PredCases[i].Value);
1152 else {
1153 // The default destination is BB, we don't need explicit targets.
1154 std::swap(PredCases[i], PredCases.back());
1155
1156 if (PredHasWeights || SuccHasWeights) {
1157 // Increase weight for the default case.
1158 Weights[0] += Weights[i + 1];
1159 std::swap(Weights[i + 1], Weights.back());
1160 Weights.pop_back();
1161 }
1162
1163 PredCases.pop_back();
1164 --i;
1165 --e;
1166 }
1167
1168 // Reconstruct the new switch statement we will be building.
1169 if (PredDefault != BBDefault) {
1170 PredDefault->removePredecessor(Pred);
1171 if (PredDefault != BB)
1172 Updates.push_back({DominatorTree::Delete, Pred, PredDefault});
1173 PredDefault = BBDefault;
1174 ++NewSuccessors[BBDefault];
1175 }
1176
1177 unsigned CasesFromPred = Weights.size();
1178 uint64_t ValidTotalSuccWeight = 0;
1179 for (unsigned i = 0, e = BBCases.size(); i != e; ++i)
1180 if (!PTIHandled.count(BBCases[i].Value) && BBCases[i].Dest != BBDefault) {
1181 PredCases.push_back(BBCases[i]);
1182 ++NewSuccessors[BBCases[i].Dest];
1183 if (SuccHasWeights || PredHasWeights) {
1184 // The default weight is at index 0, so weight for the ith case
1185 // should be at index i+1. Scale the cases from successor by
1186 // PredDefaultWeight (Weights[0]).
1187 Weights.push_back(Weights[0] * SuccWeights[i + 1]);
1188 ValidTotalSuccWeight += SuccWeights[i + 1];
1189 }
1190 }
1191
1192 if (SuccHasWeights || PredHasWeights) {
1193 ValidTotalSuccWeight += SuccWeights[0];
1194 // Scale the cases from predecessor by ValidTotalSuccWeight.
1195 for (unsigned i = 1; i < CasesFromPred; ++i)
1196 Weights[i] *= ValidTotalSuccWeight;
1197 // Scale the default weight by SuccDefaultWeight (SuccWeights[0]).
1198 Weights[0] *= SuccWeights[0];
1199 }
1200 } else {
1201 // If this is not the default destination from PSI, only the edges
1202 // in SI that occur in PSI with a destination of BB will be
1203 // activated.
1204 std::set<ConstantInt *, ConstantIntOrdering> PTIHandled;
1205 std::map<ConstantInt *, uint64_t> WeightsForHandled;
1206 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
1207 if (PredCases[i].Dest == BB) {
1208 PTIHandled.insert(PredCases[i].Value);
1209
1210 if (PredHasWeights || SuccHasWeights) {
1211 WeightsForHandled[PredCases[i].Value] = Weights[i + 1];
1212 std::swap(Weights[i + 1], Weights.back());
1213 Weights.pop_back();
1214 }
1215
1216 std::swap(PredCases[i], PredCases.back());
1217 PredCases.pop_back();
1218 --i;
1219 --e;
1220 }
1221
1222 // Okay, now we know which constants were sent to BB from the
1223 // predecessor. Figure out where they will all go now.
1224 for (unsigned i = 0, e = BBCases.size(); i != e; ++i)
1225 if (PTIHandled.count(BBCases[i].Value)) {
1226 // If this is one we are capable of getting...
1227 if (PredHasWeights || SuccHasWeights)
1228 Weights.push_back(WeightsForHandled[BBCases[i].Value]);
1229 PredCases.push_back(BBCases[i]);
1230 ++NewSuccessors[BBCases[i].Dest];
1231 PTIHandled.erase(BBCases[i].Value); // This constant is taken care of
1232 }
1233
1234 // If there are any constants vectored to BB that TI doesn't handle,
1235 // they must go to the default destination of TI.
1236 for (ConstantInt *I : PTIHandled) {
1237 if (PredHasWeights || SuccHasWeights)
1238 Weights.push_back(WeightsForHandled[I]);
1239 PredCases.push_back(ValueEqualityComparisonCase(I, BBDefault));
1240 ++NewSuccessors[BBDefault];
1241 }
1242 }
1243
1244 // Okay, at this point, we know which new successor Pred will get. Make
1245 // sure we update the number of entries in the PHI nodes for these
1246 // successors.
1247 for (const std::pair<BasicBlock *, int /*Num*/> &NewSuccessor :
1248 NewSuccessors) {
1249 for (auto I : seq(0, NewSuccessor.second)) {
1250 (void)I;
1251 AddPredecessorToBlock(NewSuccessor.first, Pred, BB);
1252 }
1253 if (!is_contained(successors(Pred), NewSuccessor.first))
1254 Updates.push_back({DominatorTree::Insert, Pred, NewSuccessor.first});
1255 }
1256
1257 Builder.SetInsertPoint(PTI);
1258 // Convert pointer to int before we switch.
1259 if (CV->getType()->isPointerTy()) {
1260 CV =
1261 Builder.CreatePtrToInt(CV, DL.getIntPtrType(CV->getType()), "magicptr");
1262 }
1263
1264 // Now that the successors are updated, create the new Switch instruction.
1265 SwitchInst *NewSI = Builder.CreateSwitch(CV, PredDefault, PredCases.size());
1266 NewSI->setDebugLoc(PTI->getDebugLoc());
1267 for (ValueEqualityComparisonCase &V : PredCases)
1268 NewSI->addCase(V.Value, V.Dest);
1269
1270 if (PredHasWeights || SuccHasWeights) {
1271 // Halve the weights if any of them cannot fit in an uint32_t
1272 FitWeights(Weights);
1273
1274 SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
1275
1276 setBranchWeights(NewSI, MDWeights);
1277 }
1278
1279 EraseTerminatorAndDCECond(PTI);
1280
1281 // Okay, last check. If BB is still a successor of PSI, then we must
1282 // have an infinite loop case. If so, add an infinitely looping block
1283 // to handle the case to preserve the behavior of the code.
1284 BasicBlock *InfLoopBlock = nullptr;
1285 for (unsigned i = 0, e = NewSI->getNumSuccessors(); i != e; ++i)
1286 if (NewSI->getSuccessor(i) == BB) {
1287 if (!InfLoopBlock) {
1288 // Insert it at the end of the function, because it's either code,
1289 // or it won't matter if it's hot. :)
1290 InfLoopBlock =
1291 BasicBlock::Create(BB->getContext(), "infloop", BB->getParent());
1292 BranchInst::Create(InfLoopBlock, InfLoopBlock);
1293 Updates.push_back({DominatorTree::Insert, InfLoopBlock, InfLoopBlock});
1294 }
1295 NewSI->setSuccessor(i, InfLoopBlock);
1296 }
1297
1298 if (InfLoopBlock)
1299 Updates.push_back({DominatorTree::Insert, Pred, InfLoopBlock});
1300
1301 Updates.push_back({DominatorTree::Delete, Pred, BB});
1302
1303 if (DTU)
1304 DTU->applyUpdates(Updates);
1305
1306 ++NumFoldValueComparisonIntoPredecessors;
1307 return true;
1308 }
1309
1310 /// The specified terminator is a value equality comparison instruction
1311 /// (either a switch or a branch on "X == c").
1312 /// See if any of the predecessors of the terminator block are value comparisons
1313 /// on the same value. If so, and if safe to do so, fold them together.
FoldValueComparisonIntoPredecessors(Instruction * TI,IRBuilder<> & Builder)1314 bool SimplifyCFGOpt::FoldValueComparisonIntoPredecessors(Instruction *TI,
1315 IRBuilder<> &Builder) {
1316 BasicBlock *BB = TI->getParent();
1317 Value *CV = isValueEqualityComparison(TI); // CondVal
1318 assert(CV && "Not a comparison?");
1319
1320 bool Changed = false;
1321
1322 SmallSetVector<BasicBlock *, 16> Preds(pred_begin(BB), pred_end(BB));
1323 while (!Preds.empty()) {
1324 BasicBlock *Pred = Preds.pop_back_val();
1325 Instruction *PTI = Pred->getTerminator();
1326
1327 // Don't try to fold into itself.
1328 if (Pred == BB)
1329 continue;
1330
1331 // See if the predecessor is a comparison with the same value.
1332 Value *PCV = isValueEqualityComparison(PTI); // PredCondVal
1333 if (PCV != CV)
1334 continue;
1335
1336 SmallSetVector<BasicBlock *, 4> FailBlocks;
1337 if (!SafeToMergeTerminators(TI, PTI, &FailBlocks)) {
1338 for (auto *Succ : FailBlocks) {
1339 if (!SplitBlockPredecessors(Succ, TI->getParent(), ".fold.split", DTU))
1340 return false;
1341 }
1342 }
1343
1344 PerformValueComparisonIntoPredecessorFolding(TI, CV, PTI, Builder);
1345 Changed = true;
1346 }
1347 return Changed;
1348 }
1349
1350 // If we would need to insert a select that uses the value of this invoke
1351 // (comments in HoistThenElseCodeToIf explain why we would need to do this), we
1352 // can't hoist the invoke, as there is nowhere to put the select in this case.
isSafeToHoistInvoke(BasicBlock * BB1,BasicBlock * BB2,Instruction * I1,Instruction * I2)1353 static bool isSafeToHoistInvoke(BasicBlock *BB1, BasicBlock *BB2,
1354 Instruction *I1, Instruction *I2) {
1355 for (BasicBlock *Succ : successors(BB1)) {
1356 for (const PHINode &PN : Succ->phis()) {
1357 Value *BB1V = PN.getIncomingValueForBlock(BB1);
1358 Value *BB2V = PN.getIncomingValueForBlock(BB2);
1359 if (BB1V != BB2V && (BB1V == I1 || BB2V == I2)) {
1360 return false;
1361 }
1362 }
1363 }
1364 return true;
1365 }
1366
1367 static bool passingValueIsAlwaysUndefined(Value *V, Instruction *I, bool PtrValueMayBeModified = false);
1368
1369 /// Given a conditional branch that goes to BB1 and BB2, hoist any common code
1370 /// in the two blocks up into the branch block. The caller of this function
1371 /// guarantees that BI's block dominates BB1 and BB2.
HoistThenElseCodeToIf(BranchInst * BI,const TargetTransformInfo & TTI)1372 bool SimplifyCFGOpt::HoistThenElseCodeToIf(BranchInst *BI,
1373 const TargetTransformInfo &TTI) {
1374 // This does very trivial matching, with limited scanning, to find identical
1375 // instructions in the two blocks. In particular, we don't want to get into
1376 // O(M*N) situations here where M and N are the sizes of BB1 and BB2. As
1377 // such, we currently just scan for obviously identical instructions in an
1378 // identical order.
1379 BasicBlock *BB1 = BI->getSuccessor(0); // The true destination.
1380 BasicBlock *BB2 = BI->getSuccessor(1); // The false destination
1381
1382 BasicBlock::iterator BB1_Itr = BB1->begin();
1383 BasicBlock::iterator BB2_Itr = BB2->begin();
1384
1385 Instruction *I1 = &*BB1_Itr++, *I2 = &*BB2_Itr++;
1386 // Skip debug info if it is not identical.
1387 DbgInfoIntrinsic *DBI1 = dyn_cast<DbgInfoIntrinsic>(I1);
1388 DbgInfoIntrinsic *DBI2 = dyn_cast<DbgInfoIntrinsic>(I2);
1389 if (!DBI1 || !DBI2 || !DBI1->isIdenticalToWhenDefined(DBI2)) {
1390 while (isa<DbgInfoIntrinsic>(I1))
1391 I1 = &*BB1_Itr++;
1392 while (isa<DbgInfoIntrinsic>(I2))
1393 I2 = &*BB2_Itr++;
1394 }
1395 // FIXME: Can we define a safety predicate for CallBr?
1396 if (isa<PHINode>(I1) || !I1->isIdenticalToWhenDefined(I2) ||
1397 (isa<InvokeInst>(I1) && !isSafeToHoistInvoke(BB1, BB2, I1, I2)) ||
1398 isa<CallBrInst>(I1))
1399 return false;
1400
1401 BasicBlock *BIParent = BI->getParent();
1402
1403 bool Changed = false;
1404
1405 auto _ = make_scope_exit([&]() {
1406 if (Changed)
1407 ++NumHoistCommonCode;
1408 });
1409
1410 do {
1411 // If we are hoisting the terminator instruction, don't move one (making a
1412 // broken BB), instead clone it, and remove BI.
1413 if (I1->isTerminator())
1414 goto HoistTerminator;
1415
1416 // If we're going to hoist a call, make sure that the two instructions we're
1417 // commoning/hoisting are both marked with musttail, or neither of them is
1418 // marked as such. Otherwise, we might end up in a situation where we hoist
1419 // from a block where the terminator is a `ret` to a block where the terminator
1420 // is a `br`, and `musttail` calls expect to be followed by a return.
1421 auto *C1 = dyn_cast<CallInst>(I1);
1422 auto *C2 = dyn_cast<CallInst>(I2);
1423 if (C1 && C2)
1424 if (C1->isMustTailCall() != C2->isMustTailCall())
1425 return Changed;
1426
1427 if (!TTI.isProfitableToHoist(I1) || !TTI.isProfitableToHoist(I2))
1428 return Changed;
1429
1430 // If any of the two call sites has nomerge attribute, stop hoisting.
1431 if (const auto *CB1 = dyn_cast<CallBase>(I1))
1432 if (CB1->cannotMerge())
1433 return Changed;
1434 if (const auto *CB2 = dyn_cast<CallBase>(I2))
1435 if (CB2->cannotMerge())
1436 return Changed;
1437
1438 if (isa<DbgInfoIntrinsic>(I1) || isa<DbgInfoIntrinsic>(I2)) {
1439 assert (isa<DbgInfoIntrinsic>(I1) && isa<DbgInfoIntrinsic>(I2));
1440 // The debug location is an integral part of a debug info intrinsic
1441 // and can't be separated from it or replaced. Instead of attempting
1442 // to merge locations, simply hoist both copies of the intrinsic.
1443 BIParent->getInstList().splice(BI->getIterator(),
1444 BB1->getInstList(), I1);
1445 BIParent->getInstList().splice(BI->getIterator(),
1446 BB2->getInstList(), I2);
1447 Changed = true;
1448 } else {
1449 // For a normal instruction, we just move one to right before the branch,
1450 // then replace all uses of the other with the first. Finally, we remove
1451 // the now redundant second instruction.
1452 BIParent->getInstList().splice(BI->getIterator(),
1453 BB1->getInstList(), I1);
1454 if (!I2->use_empty())
1455 I2->replaceAllUsesWith(I1);
1456 I1->andIRFlags(I2);
1457 unsigned KnownIDs[] = {LLVMContext::MD_tbaa,
1458 LLVMContext::MD_range,
1459 LLVMContext::MD_fpmath,
1460 LLVMContext::MD_invariant_load,
1461 LLVMContext::MD_nonnull,
1462 LLVMContext::MD_invariant_group,
1463 LLVMContext::MD_align,
1464 LLVMContext::MD_dereferenceable,
1465 LLVMContext::MD_dereferenceable_or_null,
1466 LLVMContext::MD_mem_parallel_loop_access,
1467 LLVMContext::MD_access_group,
1468 LLVMContext::MD_preserve_access_index};
1469 combineMetadata(I1, I2, KnownIDs, true);
1470
1471 // I1 and I2 are being combined into a single instruction. Its debug
1472 // location is the merged locations of the original instructions.
1473 I1->applyMergedLocation(I1->getDebugLoc(), I2->getDebugLoc());
1474
1475 I2->eraseFromParent();
1476 Changed = true;
1477 }
1478 ++NumHoistCommonInstrs;
1479
1480 I1 = &*BB1_Itr++;
1481 I2 = &*BB2_Itr++;
1482 // Skip debug info if it is not identical.
1483 DbgInfoIntrinsic *DBI1 = dyn_cast<DbgInfoIntrinsic>(I1);
1484 DbgInfoIntrinsic *DBI2 = dyn_cast<DbgInfoIntrinsic>(I2);
1485 if (!DBI1 || !DBI2 || !DBI1->isIdenticalToWhenDefined(DBI2)) {
1486 while (isa<DbgInfoIntrinsic>(I1))
1487 I1 = &*BB1_Itr++;
1488 while (isa<DbgInfoIntrinsic>(I2))
1489 I2 = &*BB2_Itr++;
1490 }
1491 } while (I1->isIdenticalToWhenDefined(I2));
1492
1493 return true;
1494
1495 HoistTerminator:
1496 // It may not be possible to hoist an invoke.
1497 // FIXME: Can we define a safety predicate for CallBr?
1498 if (isa<InvokeInst>(I1) && !isSafeToHoistInvoke(BB1, BB2, I1, I2))
1499 return Changed;
1500
1501 // TODO: callbr hoisting currently disabled pending further study.
1502 if (isa<CallBrInst>(I1))
1503 return Changed;
1504
1505 for (BasicBlock *Succ : successors(BB1)) {
1506 for (PHINode &PN : Succ->phis()) {
1507 Value *BB1V = PN.getIncomingValueForBlock(BB1);
1508 Value *BB2V = PN.getIncomingValueForBlock(BB2);
1509 if (BB1V == BB2V)
1510 continue;
1511
1512 // Check for passingValueIsAlwaysUndefined here because we would rather
1513 // eliminate undefined control flow then converting it to a select.
1514 if (passingValueIsAlwaysUndefined(BB1V, &PN) ||
1515 passingValueIsAlwaysUndefined(BB2V, &PN))
1516 return Changed;
1517
1518 if (isa<ConstantExpr>(BB1V) && !isSafeToSpeculativelyExecute(BB1V))
1519 return Changed;
1520 if (isa<ConstantExpr>(BB2V) && !isSafeToSpeculativelyExecute(BB2V))
1521 return Changed;
1522 }
1523 }
1524
1525 // Okay, it is safe to hoist the terminator.
1526 Instruction *NT = I1->clone();
1527 BIParent->getInstList().insert(BI->getIterator(), NT);
1528 if (!NT->getType()->isVoidTy()) {
1529 I1->replaceAllUsesWith(NT);
1530 I2->replaceAllUsesWith(NT);
1531 NT->takeName(I1);
1532 }
1533 Changed = true;
1534 ++NumHoistCommonInstrs;
1535
1536 // Ensure terminator gets a debug location, even an unknown one, in case
1537 // it involves inlinable calls.
1538 NT->applyMergedLocation(I1->getDebugLoc(), I2->getDebugLoc());
1539
1540 // PHIs created below will adopt NT's merged DebugLoc.
1541 IRBuilder<NoFolder> Builder(NT);
1542
1543 // Hoisting one of the terminators from our successor is a great thing.
1544 // Unfortunately, the successors of the if/else blocks may have PHI nodes in
1545 // them. If they do, all PHI entries for BB1/BB2 must agree for all PHI
1546 // nodes, so we insert select instruction to compute the final result.
1547 std::map<std::pair<Value *, Value *>, SelectInst *> InsertedSelects;
1548 for (BasicBlock *Succ : successors(BB1)) {
1549 for (PHINode &PN : Succ->phis()) {
1550 Value *BB1V = PN.getIncomingValueForBlock(BB1);
1551 Value *BB2V = PN.getIncomingValueForBlock(BB2);
1552 if (BB1V == BB2V)
1553 continue;
1554
1555 // These values do not agree. Insert a select instruction before NT
1556 // that determines the right value.
1557 SelectInst *&SI = InsertedSelects[std::make_pair(BB1V, BB2V)];
1558 if (!SI) {
1559 // Propagate fast-math-flags from phi node to its replacement select.
1560 IRBuilder<>::FastMathFlagGuard FMFGuard(Builder);
1561 if (isa<FPMathOperator>(PN))
1562 Builder.setFastMathFlags(PN.getFastMathFlags());
1563
1564 SI = cast<SelectInst>(
1565 Builder.CreateSelect(BI->getCondition(), BB1V, BB2V,
1566 BB1V->getName() + "." + BB2V->getName(), BI));
1567 }
1568
1569 // Make the PHI node use the select for all incoming values for BB1/BB2
1570 for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i)
1571 if (PN.getIncomingBlock(i) == BB1 || PN.getIncomingBlock(i) == BB2)
1572 PN.setIncomingValue(i, SI);
1573 }
1574 }
1575
1576 SmallVector<DominatorTree::UpdateType, 4> Updates;
1577
1578 // Update any PHI nodes in our new successors.
1579 for (BasicBlock *Succ : successors(BB1)) {
1580 AddPredecessorToBlock(Succ, BIParent, BB1);
1581 Updates.push_back({DominatorTree::Insert, BIParent, Succ});
1582 }
1583 for (BasicBlock *Succ : successors(BI))
1584 Updates.push_back({DominatorTree::Delete, BIParent, Succ});
1585
1586 EraseTerminatorAndDCECond(BI);
1587 if (DTU)
1588 DTU->applyUpdates(Updates);
1589 return Changed;
1590 }
1591
1592 // Check lifetime markers.
isLifeTimeMarker(const Instruction * I)1593 static bool isLifeTimeMarker(const Instruction *I) {
1594 if (auto II = dyn_cast<IntrinsicInst>(I)) {
1595 switch (II->getIntrinsicID()) {
1596 default:
1597 break;
1598 case Intrinsic::lifetime_start:
1599 case Intrinsic::lifetime_end:
1600 return true;
1601 }
1602 }
1603 return false;
1604 }
1605
1606 // TODO: Refine this. This should avoid cases like turning constant memcpy sizes
1607 // into variables.
replacingOperandWithVariableIsCheap(const Instruction * I,int OpIdx)1608 static bool replacingOperandWithVariableIsCheap(const Instruction *I,
1609 int OpIdx) {
1610 return !isa<IntrinsicInst>(I);
1611 }
1612
1613 // All instructions in Insts belong to different blocks that all unconditionally
1614 // branch to a common successor. Analyze each instruction and return true if it
1615 // would be possible to sink them into their successor, creating one common
1616 // instruction instead. For every value that would be required to be provided by
1617 // PHI node (because an operand varies in each input block), add to PHIOperands.
canSinkInstructions(ArrayRef<Instruction * > Insts,DenseMap<Instruction *,SmallVector<Value *,4>> & PHIOperands)1618 static bool canSinkInstructions(
1619 ArrayRef<Instruction *> Insts,
1620 DenseMap<Instruction *, SmallVector<Value *, 4>> &PHIOperands) {
1621 // Prune out obviously bad instructions to move. Any non-store instruction
1622 // must have exactly one use, and we check later that use is by a single,
1623 // common PHI instruction in the successor.
1624 for (auto *I : Insts) {
1625 // These instructions may change or break semantics if moved.
1626 if (isa<PHINode>(I) || I->isEHPad() || isa<AllocaInst>(I) ||
1627 I->getType()->isTokenTy())
1628 return false;
1629
1630 // Do not try to sink an instruction in an infinite loop - it can cause
1631 // this algorithm to infinite loop.
1632 if (I->getParent()->getSingleSuccessor() == I->getParent())
1633 return false;
1634
1635 // Conservatively return false if I is an inline-asm instruction. Sinking
1636 // and merging inline-asm instructions can potentially create arguments
1637 // that cannot satisfy the inline-asm constraints.
1638 // If the instruction has nomerge attribute, return false.
1639 if (const auto *C = dyn_cast<CallBase>(I))
1640 if (C->isInlineAsm() || C->cannotMerge())
1641 return false;
1642
1643 // Everything must have only one use too, apart from stores which
1644 // have no uses.
1645 if (!isa<StoreInst>(I) && !I->hasOneUse())
1646 return false;
1647 }
1648
1649 const Instruction *I0 = Insts.front();
1650 for (auto *I : Insts)
1651 if (!I->isSameOperationAs(I0))
1652 return false;
1653
1654 // All instructions in Insts are known to be the same opcode. If they aren't
1655 // stores, check the only user of each is a PHI or in the same block as the
1656 // instruction, because if a user is in the same block as an instruction
1657 // we're contemplating sinking, it must already be determined to be sinkable.
1658 if (!isa<StoreInst>(I0)) {
1659 auto *PNUse = dyn_cast<PHINode>(*I0->user_begin());
1660 auto *Succ = I0->getParent()->getTerminator()->getSuccessor(0);
1661 if (!all_of(Insts, [&PNUse,&Succ](const Instruction *I) -> bool {
1662 auto *U = cast<Instruction>(*I->user_begin());
1663 return (PNUse &&
1664 PNUse->getParent() == Succ &&
1665 PNUse->getIncomingValueForBlock(I->getParent()) == I) ||
1666 U->getParent() == I->getParent();
1667 }))
1668 return false;
1669 }
1670
1671 // Because SROA can't handle speculating stores of selects, try not to sink
1672 // loads, stores or lifetime markers of allocas when we'd have to create a
1673 // PHI for the address operand. Also, because it is likely that loads or
1674 // stores of allocas will disappear when Mem2Reg/SROA is run, don't sink
1675 // them.
1676 // This can cause code churn which can have unintended consequences down
1677 // the line - see https://llvm.org/bugs/show_bug.cgi?id=30244.
1678 // FIXME: This is a workaround for a deficiency in SROA - see
1679 // https://llvm.org/bugs/show_bug.cgi?id=30188
1680 if (isa<StoreInst>(I0) && any_of(Insts, [](const Instruction *I) {
1681 return isa<AllocaInst>(I->getOperand(1)->stripPointerCasts());
1682 }))
1683 return false;
1684 if (isa<LoadInst>(I0) && any_of(Insts, [](const Instruction *I) {
1685 return isa<AllocaInst>(I->getOperand(0)->stripPointerCasts());
1686 }))
1687 return false;
1688 if (isLifeTimeMarker(I0) && any_of(Insts, [](const Instruction *I) {
1689 return isa<AllocaInst>(I->getOperand(1)->stripPointerCasts());
1690 }))
1691 return false;
1692
1693 for (unsigned OI = 0, OE = I0->getNumOperands(); OI != OE; ++OI) {
1694 Value *Op = I0->getOperand(OI);
1695 if (Op->getType()->isTokenTy())
1696 // Don't touch any operand of token type.
1697 return false;
1698
1699 auto SameAsI0 = [&I0, OI](const Instruction *I) {
1700 assert(I->getNumOperands() == I0->getNumOperands());
1701 return I->getOperand(OI) == I0->getOperand(OI);
1702 };
1703 if (!all_of(Insts, SameAsI0)) {
1704 if ((isa<Constant>(Op) && !replacingOperandWithVariableIsCheap(I0, OI)) ||
1705 !canReplaceOperandWithVariable(I0, OI))
1706 // We can't create a PHI from this GEP.
1707 return false;
1708 // Don't create indirect calls! The called value is the final operand.
1709 if (isa<CallBase>(I0) && OI == OE - 1) {
1710 // FIXME: if the call was *already* indirect, we should do this.
1711 return false;
1712 }
1713 for (auto *I : Insts)
1714 PHIOperands[I].push_back(I->getOperand(OI));
1715 }
1716 }
1717 return true;
1718 }
1719
1720 // Assuming canSinkInstructions(Blocks) has returned true, sink the last
1721 // instruction of every block in Blocks to their common successor, commoning
1722 // into one instruction.
sinkLastInstruction(ArrayRef<BasicBlock * > Blocks)1723 static bool sinkLastInstruction(ArrayRef<BasicBlock*> Blocks) {
1724 auto *BBEnd = Blocks[0]->getTerminator()->getSuccessor(0);
1725
1726 // canSinkInstructions returning true guarantees that every block has at
1727 // least one non-terminator instruction.
1728 SmallVector<Instruction*,4> Insts;
1729 for (auto *BB : Blocks) {
1730 Instruction *I = BB->getTerminator();
1731 do {
1732 I = I->getPrevNode();
1733 } while (isa<DbgInfoIntrinsic>(I) && I != &BB->front());
1734 if (!isa<DbgInfoIntrinsic>(I))
1735 Insts.push_back(I);
1736 }
1737
1738 // The only checking we need to do now is that all users of all instructions
1739 // are the same PHI node. canSinkInstructions should have checked this but
1740 // it is slightly over-aggressive - it gets confused by commutative
1741 // instructions so double-check it here.
1742 Instruction *I0 = Insts.front();
1743 if (!isa<StoreInst>(I0)) {
1744 auto *PNUse = dyn_cast<PHINode>(*I0->user_begin());
1745 if (!all_of(Insts, [&PNUse](const Instruction *I) -> bool {
1746 auto *U = cast<Instruction>(*I->user_begin());
1747 return U == PNUse;
1748 }))
1749 return false;
1750 }
1751
1752 // We don't need to do any more checking here; canSinkInstructions should
1753 // have done it all for us.
1754 SmallVector<Value*, 4> NewOperands;
1755 for (unsigned O = 0, E = I0->getNumOperands(); O != E; ++O) {
1756 // This check is different to that in canSinkInstructions. There, we
1757 // cared about the global view once simplifycfg (and instcombine) have
1758 // completed - it takes into account PHIs that become trivially
1759 // simplifiable. However here we need a more local view; if an operand
1760 // differs we create a PHI and rely on instcombine to clean up the very
1761 // small mess we may make.
1762 bool NeedPHI = any_of(Insts, [&I0, O](const Instruction *I) {
1763 return I->getOperand(O) != I0->getOperand(O);
1764 });
1765 if (!NeedPHI) {
1766 NewOperands.push_back(I0->getOperand(O));
1767 continue;
1768 }
1769
1770 // Create a new PHI in the successor block and populate it.
1771 auto *Op = I0->getOperand(O);
1772 assert(!Op->getType()->isTokenTy() && "Can't PHI tokens!");
1773 auto *PN = PHINode::Create(Op->getType(), Insts.size(),
1774 Op->getName() + ".sink", &BBEnd->front());
1775 for (auto *I : Insts)
1776 PN->addIncoming(I->getOperand(O), I->getParent());
1777 NewOperands.push_back(PN);
1778 }
1779
1780 // Arbitrarily use I0 as the new "common" instruction; remap its operands
1781 // and move it to the start of the successor block.
1782 for (unsigned O = 0, E = I0->getNumOperands(); O != E; ++O)
1783 I0->getOperandUse(O).set(NewOperands[O]);
1784 I0->moveBefore(&*BBEnd->getFirstInsertionPt());
1785
1786 // Update metadata and IR flags, and merge debug locations.
1787 for (auto *I : Insts)
1788 if (I != I0) {
1789 // The debug location for the "common" instruction is the merged locations
1790 // of all the commoned instructions. We start with the original location
1791 // of the "common" instruction and iteratively merge each location in the
1792 // loop below.
1793 // This is an N-way merge, which will be inefficient if I0 is a CallInst.
1794 // However, as N-way merge for CallInst is rare, so we use simplified API
1795 // instead of using complex API for N-way merge.
1796 I0->applyMergedLocation(I0->getDebugLoc(), I->getDebugLoc());
1797 combineMetadataForCSE(I0, I, true);
1798 I0->andIRFlags(I);
1799 }
1800
1801 if (!isa<StoreInst>(I0)) {
1802 // canSinkLastInstruction checked that all instructions were used by
1803 // one and only one PHI node. Find that now, RAUW it to our common
1804 // instruction and nuke it.
1805 assert(I0->hasOneUse());
1806 auto *PN = cast<PHINode>(*I0->user_begin());
1807 PN->replaceAllUsesWith(I0);
1808 PN->eraseFromParent();
1809 }
1810
1811 // Finally nuke all instructions apart from the common instruction.
1812 for (auto *I : Insts)
1813 if (I != I0)
1814 I->eraseFromParent();
1815
1816 return true;
1817 }
1818
1819 namespace {
1820
1821 // LockstepReverseIterator - Iterates through instructions
1822 // in a set of blocks in reverse order from the first non-terminator.
1823 // For example (assume all blocks have size n):
1824 // LockstepReverseIterator I([B1, B2, B3]);
1825 // *I-- = [B1[n], B2[n], B3[n]];
1826 // *I-- = [B1[n-1], B2[n-1], B3[n-1]];
1827 // *I-- = [B1[n-2], B2[n-2], B3[n-2]];
1828 // ...
1829 class LockstepReverseIterator {
1830 ArrayRef<BasicBlock*> Blocks;
1831 SmallVector<Instruction*,4> Insts;
1832 bool Fail;
1833
1834 public:
LockstepReverseIterator(ArrayRef<BasicBlock * > Blocks)1835 LockstepReverseIterator(ArrayRef<BasicBlock*> Blocks) : Blocks(Blocks) {
1836 reset();
1837 }
1838
reset()1839 void reset() {
1840 Fail = false;
1841 Insts.clear();
1842 for (auto *BB : Blocks) {
1843 Instruction *Inst = BB->getTerminator();
1844 for (Inst = Inst->getPrevNode(); Inst && isa<DbgInfoIntrinsic>(Inst);)
1845 Inst = Inst->getPrevNode();
1846 if (!Inst) {
1847 // Block wasn't big enough.
1848 Fail = true;
1849 return;
1850 }
1851 Insts.push_back(Inst);
1852 }
1853 }
1854
isValid() const1855 bool isValid() const {
1856 return !Fail;
1857 }
1858
operator --()1859 void operator--() {
1860 if (Fail)
1861 return;
1862 for (auto *&Inst : Insts) {
1863 for (Inst = Inst->getPrevNode(); Inst && isa<DbgInfoIntrinsic>(Inst);)
1864 Inst = Inst->getPrevNode();
1865 // Already at beginning of block.
1866 if (!Inst) {
1867 Fail = true;
1868 return;
1869 }
1870 }
1871 }
1872
operator *() const1873 ArrayRef<Instruction*> operator * () const {
1874 return Insts;
1875 }
1876 };
1877
1878 } // end anonymous namespace
1879
1880 /// Check whether BB's predecessors end with unconditional branches. If it is
1881 /// true, sink any common code from the predecessors to BB.
1882 /// We also allow one predecessor to end with conditional branch (but no more
1883 /// than one).
SinkCommonCodeFromPredecessors(BasicBlock * BB,DomTreeUpdater * DTU)1884 static bool SinkCommonCodeFromPredecessors(BasicBlock *BB,
1885 DomTreeUpdater *DTU) {
1886 // We support two situations:
1887 // (1) all incoming arcs are unconditional
1888 // (2) one incoming arc is conditional
1889 //
1890 // (2) is very common in switch defaults and
1891 // else-if patterns;
1892 //
1893 // if (a) f(1);
1894 // else if (b) f(2);
1895 //
1896 // produces:
1897 //
1898 // [if]
1899 // / \
1900 // [f(1)] [if]
1901 // | | \
1902 // | | |
1903 // | [f(2)]|
1904 // \ | /
1905 // [ end ]
1906 //
1907 // [end] has two unconditional predecessor arcs and one conditional. The
1908 // conditional refers to the implicit empty 'else' arc. This conditional
1909 // arc can also be caused by an empty default block in a switch.
1910 //
1911 // In this case, we attempt to sink code from all *unconditional* arcs.
1912 // If we can sink instructions from these arcs (determined during the scan
1913 // phase below) we insert a common successor for all unconditional arcs and
1914 // connect that to [end], to enable sinking:
1915 //
1916 // [if]
1917 // / \
1918 // [x(1)] [if]
1919 // | | \
1920 // | | \
1921 // | [x(2)] |
1922 // \ / |
1923 // [sink.split] |
1924 // \ /
1925 // [ end ]
1926 //
1927 SmallVector<BasicBlock*,4> UnconditionalPreds;
1928 Instruction *Cond = nullptr;
1929 for (auto *B : predecessors(BB)) {
1930 auto *T = B->getTerminator();
1931 if (isa<BranchInst>(T) && cast<BranchInst>(T)->isUnconditional())
1932 UnconditionalPreds.push_back(B);
1933 else if ((isa<BranchInst>(T) || isa<SwitchInst>(T)) && !Cond)
1934 Cond = T;
1935 else
1936 return false;
1937 }
1938 if (UnconditionalPreds.size() < 2)
1939 return false;
1940
1941 // We take a two-step approach to tail sinking. First we scan from the end of
1942 // each block upwards in lockstep. If the n'th instruction from the end of each
1943 // block can be sunk, those instructions are added to ValuesToSink and we
1944 // carry on. If we can sink an instruction but need to PHI-merge some operands
1945 // (because they're not identical in each instruction) we add these to
1946 // PHIOperands.
1947 unsigned ScanIdx = 0;
1948 SmallPtrSet<Value*,4> InstructionsToSink;
1949 DenseMap<Instruction*, SmallVector<Value*,4>> PHIOperands;
1950 LockstepReverseIterator LRI(UnconditionalPreds);
1951 while (LRI.isValid() &&
1952 canSinkInstructions(*LRI, PHIOperands)) {
1953 LLVM_DEBUG(dbgs() << "SINK: instruction can be sunk: " << *(*LRI)[0]
1954 << "\n");
1955 InstructionsToSink.insert((*LRI).begin(), (*LRI).end());
1956 ++ScanIdx;
1957 --LRI;
1958 }
1959
1960 // If no instructions can be sunk, early-return.
1961 if (ScanIdx == 0)
1962 return false;
1963
1964 bool Changed = false;
1965
1966 auto ProfitableToSinkInstruction = [&](LockstepReverseIterator &LRI) {
1967 unsigned NumPHIdValues = 0;
1968 for (auto *I : *LRI)
1969 for (auto *V : PHIOperands[I])
1970 if (InstructionsToSink.count(V) == 0)
1971 ++NumPHIdValues;
1972 LLVM_DEBUG(dbgs() << "SINK: #phid values: " << NumPHIdValues << "\n");
1973 unsigned NumPHIInsts = NumPHIdValues / UnconditionalPreds.size();
1974 if ((NumPHIdValues % UnconditionalPreds.size()) != 0)
1975 NumPHIInsts++;
1976
1977 return NumPHIInsts <= 1;
1978 };
1979
1980 if (Cond) {
1981 // Check if we would actually sink anything first! This mutates the CFG and
1982 // adds an extra block. The goal in doing this is to allow instructions that
1983 // couldn't be sunk before to be sunk - obviously, speculatable instructions
1984 // (such as trunc, add) can be sunk and predicated already. So we check that
1985 // we're going to sink at least one non-speculatable instruction.
1986 LRI.reset();
1987 unsigned Idx = 0;
1988 bool Profitable = false;
1989 while (ProfitableToSinkInstruction(LRI) && Idx < ScanIdx) {
1990 if (!isSafeToSpeculativelyExecute((*LRI)[0])) {
1991 Profitable = true;
1992 break;
1993 }
1994 --LRI;
1995 ++Idx;
1996 }
1997 if (!Profitable)
1998 return false;
1999
2000 LLVM_DEBUG(dbgs() << "SINK: Splitting edge\n");
2001 // We have a conditional edge and we're going to sink some instructions.
2002 // Insert a new block postdominating all blocks we're going to sink from.
2003 if (!SplitBlockPredecessors(BB, UnconditionalPreds, ".sink.split", DTU))
2004 // Edges couldn't be split.
2005 return false;
2006 Changed = true;
2007 }
2008
2009 // Now that we've analyzed all potential sinking candidates, perform the
2010 // actual sink. We iteratively sink the last non-terminator of the source
2011 // blocks into their common successor unless doing so would require too
2012 // many PHI instructions to be generated (currently only one PHI is allowed
2013 // per sunk instruction).
2014 //
2015 // We can use InstructionsToSink to discount values needing PHI-merging that will
2016 // actually be sunk in a later iteration. This allows us to be more
2017 // aggressive in what we sink. This does allow a false positive where we
2018 // sink presuming a later value will also be sunk, but stop half way through
2019 // and never actually sink it which means we produce more PHIs than intended.
2020 // This is unlikely in practice though.
2021 unsigned SinkIdx = 0;
2022 for (; SinkIdx != ScanIdx; ++SinkIdx) {
2023 LLVM_DEBUG(dbgs() << "SINK: Sink: "
2024 << *UnconditionalPreds[0]->getTerminator()->getPrevNode()
2025 << "\n");
2026
2027 // Because we've sunk every instruction in turn, the current instruction to
2028 // sink is always at index 0.
2029 LRI.reset();
2030 if (!ProfitableToSinkInstruction(LRI)) {
2031 // Too many PHIs would be created.
2032 LLVM_DEBUG(
2033 dbgs() << "SINK: stopping here, too many PHIs would be created!\n");
2034 break;
2035 }
2036
2037 if (!sinkLastInstruction(UnconditionalPreds)) {
2038 LLVM_DEBUG(
2039 dbgs()
2040 << "SINK: stopping here, failed to actually sink instruction!\n");
2041 break;
2042 }
2043
2044 NumSinkCommonInstrs++;
2045 Changed = true;
2046 }
2047 if (SinkIdx != 0)
2048 ++NumSinkCommonCode;
2049 return Changed;
2050 }
2051
2052 /// Determine if we can hoist sink a sole store instruction out of a
2053 /// conditional block.
2054 ///
2055 /// We are looking for code like the following:
2056 /// BrBB:
2057 /// store i32 %add, i32* %arrayidx2
2058 /// ... // No other stores or function calls (we could be calling a memory
2059 /// ... // function).
2060 /// %cmp = icmp ult %x, %y
2061 /// br i1 %cmp, label %EndBB, label %ThenBB
2062 /// ThenBB:
2063 /// store i32 %add5, i32* %arrayidx2
2064 /// br label EndBB
2065 /// EndBB:
2066 /// ...
2067 /// We are going to transform this into:
2068 /// BrBB:
2069 /// store i32 %add, i32* %arrayidx2
2070 /// ... //
2071 /// %cmp = icmp ult %x, %y
2072 /// %add.add5 = select i1 %cmp, i32 %add, %add5
2073 /// store i32 %add.add5, i32* %arrayidx2
2074 /// ...
2075 ///
2076 /// \return The pointer to the value of the previous store if the store can be
2077 /// hoisted into the predecessor block. 0 otherwise.
isSafeToSpeculateStore(Instruction * I,BasicBlock * BrBB,BasicBlock * StoreBB,BasicBlock * EndBB)2078 static Value *isSafeToSpeculateStore(Instruction *I, BasicBlock *BrBB,
2079 BasicBlock *StoreBB, BasicBlock *EndBB) {
2080 StoreInst *StoreToHoist = dyn_cast<StoreInst>(I);
2081 if (!StoreToHoist)
2082 return nullptr;
2083
2084 // Volatile or atomic.
2085 if (!StoreToHoist->isSimple())
2086 return nullptr;
2087
2088 Value *StorePtr = StoreToHoist->getPointerOperand();
2089
2090 // Look for a store to the same pointer in BrBB.
2091 unsigned MaxNumInstToLookAt = 9;
2092 // Skip pseudo probe intrinsic calls which are not really killing any memory
2093 // accesses.
2094 for (Instruction &CurI : reverse(BrBB->instructionsWithoutDebug(true))) {
2095 if (!MaxNumInstToLookAt)
2096 break;
2097 --MaxNumInstToLookAt;
2098
2099 // Could be calling an instruction that affects memory like free().
2100 if (CurI.mayHaveSideEffects() && !isa<StoreInst>(CurI))
2101 return nullptr;
2102
2103 if (auto *SI = dyn_cast<StoreInst>(&CurI)) {
2104 // Found the previous store make sure it stores to the same location.
2105 if (SI->getPointerOperand() == StorePtr)
2106 // Found the previous store, return its value operand.
2107 return SI->getValueOperand();
2108 return nullptr; // Unknown store.
2109 }
2110 }
2111
2112 return nullptr;
2113 }
2114
2115 /// Estimate the cost of the insertion(s) and check that the PHI nodes can be
2116 /// converted to selects.
validateAndCostRequiredSelects(BasicBlock * BB,BasicBlock * ThenBB,BasicBlock * EndBB,unsigned & SpeculatedInstructions,int & BudgetRemaining,const TargetTransformInfo & TTI)2117 static bool validateAndCostRequiredSelects(BasicBlock *BB, BasicBlock *ThenBB,
2118 BasicBlock *EndBB,
2119 unsigned &SpeculatedInstructions,
2120 int &BudgetRemaining,
2121 const TargetTransformInfo &TTI) {
2122 TargetTransformInfo::TargetCostKind CostKind =
2123 BB->getParent()->hasMinSize()
2124 ? TargetTransformInfo::TCK_CodeSize
2125 : TargetTransformInfo::TCK_SizeAndLatency;
2126
2127 bool HaveRewritablePHIs = false;
2128 for (PHINode &PN : EndBB->phis()) {
2129 Value *OrigV = PN.getIncomingValueForBlock(BB);
2130 Value *ThenV = PN.getIncomingValueForBlock(ThenBB);
2131
2132 // FIXME: Try to remove some of the duplication with HoistThenElseCodeToIf.
2133 // Skip PHIs which are trivial.
2134 if (ThenV == OrigV)
2135 continue;
2136
2137 BudgetRemaining -=
2138 TTI.getCmpSelInstrCost(Instruction::Select, PN.getType(), nullptr,
2139 CmpInst::BAD_ICMP_PREDICATE, CostKind);
2140
2141 // Don't convert to selects if we could remove undefined behavior instead.
2142 if (passingValueIsAlwaysUndefined(OrigV, &PN) ||
2143 passingValueIsAlwaysUndefined(ThenV, &PN))
2144 return false;
2145
2146 HaveRewritablePHIs = true;
2147 ConstantExpr *OrigCE = dyn_cast<ConstantExpr>(OrigV);
2148 ConstantExpr *ThenCE = dyn_cast<ConstantExpr>(ThenV);
2149 if (!OrigCE && !ThenCE)
2150 continue; // Known safe and cheap.
2151
2152 if ((ThenCE && !isSafeToSpeculativelyExecute(ThenCE)) ||
2153 (OrigCE && !isSafeToSpeculativelyExecute(OrigCE)))
2154 return false;
2155 unsigned OrigCost = OrigCE ? ComputeSpeculationCost(OrigCE, TTI) : 0;
2156 unsigned ThenCost = ThenCE ? ComputeSpeculationCost(ThenCE, TTI) : 0;
2157 unsigned MaxCost =
2158 2 * PHINodeFoldingThreshold * TargetTransformInfo::TCC_Basic;
2159 if (OrigCost + ThenCost > MaxCost)
2160 return false;
2161
2162 // Account for the cost of an unfolded ConstantExpr which could end up
2163 // getting expanded into Instructions.
2164 // FIXME: This doesn't account for how many operations are combined in the
2165 // constant expression.
2166 ++SpeculatedInstructions;
2167 if (SpeculatedInstructions > 1)
2168 return false;
2169 }
2170
2171 return HaveRewritablePHIs;
2172 }
2173
2174 /// Speculate a conditional basic block flattening the CFG.
2175 ///
2176 /// Note that this is a very risky transform currently. Speculating
2177 /// instructions like this is most often not desirable. Instead, there is an MI
2178 /// pass which can do it with full awareness of the resource constraints.
2179 /// However, some cases are "obvious" and we should do directly. An example of
2180 /// this is speculating a single, reasonably cheap instruction.
2181 ///
2182 /// There is only one distinct advantage to flattening the CFG at the IR level:
2183 /// it makes very common but simplistic optimizations such as are common in
2184 /// instcombine and the DAG combiner more powerful by removing CFG edges and
2185 /// modeling their effects with easier to reason about SSA value graphs.
2186 ///
2187 ///
2188 /// An illustration of this transform is turning this IR:
2189 /// \code
2190 /// BB:
2191 /// %cmp = icmp ult %x, %y
2192 /// br i1 %cmp, label %EndBB, label %ThenBB
2193 /// ThenBB:
2194 /// %sub = sub %x, %y
2195 /// br label BB2
2196 /// EndBB:
2197 /// %phi = phi [ %sub, %ThenBB ], [ 0, %EndBB ]
2198 /// ...
2199 /// \endcode
2200 ///
2201 /// Into this IR:
2202 /// \code
2203 /// BB:
2204 /// %cmp = icmp ult %x, %y
2205 /// %sub = sub %x, %y
2206 /// %cond = select i1 %cmp, 0, %sub
2207 /// ...
2208 /// \endcode
2209 ///
2210 /// \returns true if the conditional block is removed.
SpeculativelyExecuteBB(BranchInst * BI,BasicBlock * ThenBB,const TargetTransformInfo & TTI)2211 bool SimplifyCFGOpt::SpeculativelyExecuteBB(BranchInst *BI, BasicBlock *ThenBB,
2212 const TargetTransformInfo &TTI) {
2213 // Be conservative for now. FP select instruction can often be expensive.
2214 Value *BrCond = BI->getCondition();
2215 if (isa<FCmpInst>(BrCond))
2216 return false;
2217
2218 BasicBlock *BB = BI->getParent();
2219 BasicBlock *EndBB = ThenBB->getTerminator()->getSuccessor(0);
2220 int BudgetRemaining =
2221 PHINodeFoldingThreshold * TargetTransformInfo::TCC_Basic;
2222
2223 // If ThenBB is actually on the false edge of the conditional branch, remember
2224 // to swap the select operands later.
2225 bool Invert = false;
2226 if (ThenBB != BI->getSuccessor(0)) {
2227 assert(ThenBB == BI->getSuccessor(1) && "No edge from 'if' block?");
2228 Invert = true;
2229 }
2230 assert(EndBB == BI->getSuccessor(!Invert) && "No edge from to end block");
2231
2232 // Keep a count of how many times instructions are used within ThenBB when
2233 // they are candidates for sinking into ThenBB. Specifically:
2234 // - They are defined in BB, and
2235 // - They have no side effects, and
2236 // - All of their uses are in ThenBB.
2237 SmallDenseMap<Instruction *, unsigned, 4> SinkCandidateUseCounts;
2238
2239 SmallVector<Instruction *, 4> SpeculatedDbgIntrinsics;
2240
2241 unsigned SpeculatedInstructions = 0;
2242 Value *SpeculatedStoreValue = nullptr;
2243 StoreInst *SpeculatedStore = nullptr;
2244 for (BasicBlock::iterator BBI = ThenBB->begin(),
2245 BBE = std::prev(ThenBB->end());
2246 BBI != BBE; ++BBI) {
2247 Instruction *I = &*BBI;
2248 // Skip debug info.
2249 if (isa<DbgInfoIntrinsic>(I)) {
2250 SpeculatedDbgIntrinsics.push_back(I);
2251 continue;
2252 }
2253
2254 // Skip pseudo probes. The consequence is we lose track of the branch
2255 // probability for ThenBB, which is fine since the optimization here takes
2256 // place regardless of the branch probability.
2257 if (isa<PseudoProbeInst>(I)) {
2258 SpeculatedDbgIntrinsics.push_back(I);
2259 continue;
2260 }
2261
2262 // Only speculatively execute a single instruction (not counting the
2263 // terminator) for now.
2264 ++SpeculatedInstructions;
2265 if (SpeculatedInstructions > 1)
2266 return false;
2267
2268 // Don't hoist the instruction if it's unsafe or expensive.
2269 if (!isSafeToSpeculativelyExecute(I) &&
2270 !(HoistCondStores && (SpeculatedStoreValue = isSafeToSpeculateStore(
2271 I, BB, ThenBB, EndBB))))
2272 return false;
2273 if (!SpeculatedStoreValue &&
2274 ComputeSpeculationCost(I, TTI) >
2275 PHINodeFoldingThreshold * TargetTransformInfo::TCC_Basic)
2276 return false;
2277
2278 // Store the store speculation candidate.
2279 if (SpeculatedStoreValue)
2280 SpeculatedStore = cast<StoreInst>(I);
2281
2282 // Do not hoist the instruction if any of its operands are defined but not
2283 // used in BB. The transformation will prevent the operand from
2284 // being sunk into the use block.
2285 for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i) {
2286 Instruction *OpI = dyn_cast<Instruction>(*i);
2287 if (!OpI || OpI->getParent() != BB || OpI->mayHaveSideEffects())
2288 continue; // Not a candidate for sinking.
2289
2290 ++SinkCandidateUseCounts[OpI];
2291 }
2292 }
2293
2294 // Consider any sink candidates which are only used in ThenBB as costs for
2295 // speculation. Note, while we iterate over a DenseMap here, we are summing
2296 // and so iteration order isn't significant.
2297 for (SmallDenseMap<Instruction *, unsigned, 4>::iterator
2298 I = SinkCandidateUseCounts.begin(),
2299 E = SinkCandidateUseCounts.end();
2300 I != E; ++I)
2301 if (I->first->hasNUses(I->second)) {
2302 ++SpeculatedInstructions;
2303 if (SpeculatedInstructions > 1)
2304 return false;
2305 }
2306
2307 // Check that we can insert the selects and that it's not too expensive to do
2308 // so.
2309 bool Convert = SpeculatedStore != nullptr;
2310 Convert |= validateAndCostRequiredSelects(BB, ThenBB, EndBB,
2311 SpeculatedInstructions,
2312 BudgetRemaining, TTI);
2313 if (!Convert || BudgetRemaining < 0)
2314 return false;
2315
2316 // If we get here, we can hoist the instruction and if-convert.
2317 LLVM_DEBUG(dbgs() << "SPECULATIVELY EXECUTING BB" << *ThenBB << "\n";);
2318
2319 // Insert a select of the value of the speculated store.
2320 if (SpeculatedStoreValue) {
2321 IRBuilder<NoFolder> Builder(BI);
2322 Value *TrueV = SpeculatedStore->getValueOperand();
2323 Value *FalseV = SpeculatedStoreValue;
2324 if (Invert)
2325 std::swap(TrueV, FalseV);
2326 Value *S = Builder.CreateSelect(
2327 BrCond, TrueV, FalseV, "spec.store.select", BI);
2328 SpeculatedStore->setOperand(0, S);
2329 SpeculatedStore->applyMergedLocation(BI->getDebugLoc(),
2330 SpeculatedStore->getDebugLoc());
2331 }
2332
2333 // Metadata can be dependent on the condition we are hoisting above.
2334 // Conservatively strip all metadata on the instruction. Drop the debug loc
2335 // to avoid making it appear as if the condition is a constant, which would
2336 // be misleading while debugging.
2337 for (auto &I : *ThenBB) {
2338 if (!SpeculatedStoreValue || &I != SpeculatedStore)
2339 I.setDebugLoc(DebugLoc());
2340 I.dropUnknownNonDebugMetadata();
2341 }
2342
2343 // Hoist the instructions.
2344 BB->getInstList().splice(BI->getIterator(), ThenBB->getInstList(),
2345 ThenBB->begin(), std::prev(ThenBB->end()));
2346
2347 // Insert selects and rewrite the PHI operands.
2348 IRBuilder<NoFolder> Builder(BI);
2349 for (PHINode &PN : EndBB->phis()) {
2350 unsigned OrigI = PN.getBasicBlockIndex(BB);
2351 unsigned ThenI = PN.getBasicBlockIndex(ThenBB);
2352 Value *OrigV = PN.getIncomingValue(OrigI);
2353 Value *ThenV = PN.getIncomingValue(ThenI);
2354
2355 // Skip PHIs which are trivial.
2356 if (OrigV == ThenV)
2357 continue;
2358
2359 // Create a select whose true value is the speculatively executed value and
2360 // false value is the pre-existing value. Swap them if the branch
2361 // destinations were inverted.
2362 Value *TrueV = ThenV, *FalseV = OrigV;
2363 if (Invert)
2364 std::swap(TrueV, FalseV);
2365 Value *V = Builder.CreateSelect(BrCond, TrueV, FalseV, "spec.select", BI);
2366 PN.setIncomingValue(OrigI, V);
2367 PN.setIncomingValue(ThenI, V);
2368 }
2369
2370 // Remove speculated dbg intrinsics.
2371 // FIXME: Is it possible to do this in a more elegant way? Moving/merging the
2372 // dbg value for the different flows and inserting it after the select.
2373 for (Instruction *I : SpeculatedDbgIntrinsics)
2374 I->eraseFromParent();
2375
2376 ++NumSpeculations;
2377 return true;
2378 }
2379
2380 /// Return true if we can thread a branch across this block.
BlockIsSimpleEnoughToThreadThrough(BasicBlock * BB)2381 static bool BlockIsSimpleEnoughToThreadThrough(BasicBlock *BB) {
2382 int Size = 0;
2383
2384 for (Instruction &I : BB->instructionsWithoutDebug()) {
2385 if (Size > MaxSmallBlockSize)
2386 return false; // Don't clone large BB's.
2387
2388 // Can't fold blocks that contain noduplicate or convergent calls.
2389 if (CallInst *CI = dyn_cast<CallInst>(&I))
2390 if (CI->cannotDuplicate() || CI->isConvergent())
2391 return false;
2392
2393 // We will delete Phis while threading, so Phis should not be accounted in
2394 // block's size
2395 if (!isa<PHINode>(I))
2396 ++Size;
2397
2398 // We can only support instructions that do not define values that are
2399 // live outside of the current basic block.
2400 for (User *U : I.users()) {
2401 Instruction *UI = cast<Instruction>(U);
2402 if (UI->getParent() != BB || isa<PHINode>(UI))
2403 return false;
2404 }
2405
2406 // Looks ok, continue checking.
2407 }
2408
2409 return true;
2410 }
2411
2412 /// If we have a conditional branch on a PHI node value that is defined in the
2413 /// same block as the branch and if any PHI entries are constants, thread edges
2414 /// corresponding to that entry to be branches to their ultimate destination.
FoldCondBranchOnPHI(BranchInst * BI,DomTreeUpdater * DTU,const DataLayout & DL,AssumptionCache * AC)2415 static bool FoldCondBranchOnPHI(BranchInst *BI, DomTreeUpdater *DTU,
2416 const DataLayout &DL, AssumptionCache *AC) {
2417 BasicBlock *BB = BI->getParent();
2418 PHINode *PN = dyn_cast<PHINode>(BI->getCondition());
2419 // NOTE: we currently cannot transform this case if the PHI node is used
2420 // outside of the block.
2421 if (!PN || PN->getParent() != BB || !PN->hasOneUse())
2422 return false;
2423
2424 // Degenerate case of a single entry PHI.
2425 if (PN->getNumIncomingValues() == 1) {
2426 FoldSingleEntryPHINodes(PN->getParent());
2427 return true;
2428 }
2429
2430 // Now we know that this block has multiple preds and two succs.
2431 if (!BlockIsSimpleEnoughToThreadThrough(BB))
2432 return false;
2433
2434 // Okay, this is a simple enough basic block. See if any phi values are
2435 // constants.
2436 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
2437 ConstantInt *CB = dyn_cast<ConstantInt>(PN->getIncomingValue(i));
2438 if (!CB || !CB->getType()->isIntegerTy(1))
2439 continue;
2440
2441 // Okay, we now know that all edges from PredBB should be revectored to
2442 // branch to RealDest.
2443 BasicBlock *PredBB = PN->getIncomingBlock(i);
2444 BasicBlock *RealDest = BI->getSuccessor(!CB->getZExtValue());
2445
2446 if (RealDest == BB)
2447 continue; // Skip self loops.
2448 // Skip if the predecessor's terminator is an indirect branch.
2449 if (isa<IndirectBrInst>(PredBB->getTerminator()))
2450 continue;
2451
2452 SmallVector<DominatorTree::UpdateType, 3> Updates;
2453
2454 // The dest block might have PHI nodes, other predecessors and other
2455 // difficult cases. Instead of being smart about this, just insert a new
2456 // block that jumps to the destination block, effectively splitting
2457 // the edge we are about to create.
2458 BasicBlock *EdgeBB =
2459 BasicBlock::Create(BB->getContext(), RealDest->getName() + ".critedge",
2460 RealDest->getParent(), RealDest);
2461 BranchInst *CritEdgeBranch = BranchInst::Create(RealDest, EdgeBB);
2462 Updates.push_back({DominatorTree::Insert, EdgeBB, RealDest});
2463 CritEdgeBranch->setDebugLoc(BI->getDebugLoc());
2464
2465 // Update PHI nodes.
2466 AddPredecessorToBlock(RealDest, EdgeBB, BB);
2467
2468 // BB may have instructions that are being threaded over. Clone these
2469 // instructions into EdgeBB. We know that there will be no uses of the
2470 // cloned instructions outside of EdgeBB.
2471 BasicBlock::iterator InsertPt = EdgeBB->begin();
2472 DenseMap<Value *, Value *> TranslateMap; // Track translated values.
2473 for (BasicBlock::iterator BBI = BB->begin(); &*BBI != BI; ++BBI) {
2474 if (PHINode *PN = dyn_cast<PHINode>(BBI)) {
2475 TranslateMap[PN] = PN->getIncomingValueForBlock(PredBB);
2476 continue;
2477 }
2478 // Clone the instruction.
2479 Instruction *N = BBI->clone();
2480 if (BBI->hasName())
2481 N->setName(BBI->getName() + ".c");
2482
2483 // Update operands due to translation.
2484 for (User::op_iterator i = N->op_begin(), e = N->op_end(); i != e; ++i) {
2485 DenseMap<Value *, Value *>::iterator PI = TranslateMap.find(*i);
2486 if (PI != TranslateMap.end())
2487 *i = PI->second;
2488 }
2489
2490 // Check for trivial simplification.
2491 if (Value *V = SimplifyInstruction(N, {DL, nullptr, nullptr, AC})) {
2492 if (!BBI->use_empty())
2493 TranslateMap[&*BBI] = V;
2494 if (!N->mayHaveSideEffects()) {
2495 N->deleteValue(); // Instruction folded away, don't need actual inst
2496 N = nullptr;
2497 }
2498 } else {
2499 if (!BBI->use_empty())
2500 TranslateMap[&*BBI] = N;
2501 }
2502 if (N) {
2503 // Insert the new instruction into its new home.
2504 EdgeBB->getInstList().insert(InsertPt, N);
2505
2506 // Register the new instruction with the assumption cache if necessary.
2507 if (AC && match(N, m_Intrinsic<Intrinsic::assume>()))
2508 AC->registerAssumption(cast<IntrinsicInst>(N));
2509 }
2510 }
2511
2512 // Loop over all of the edges from PredBB to BB, changing them to branch
2513 // to EdgeBB instead.
2514 Instruction *PredBBTI = PredBB->getTerminator();
2515 for (unsigned i = 0, e = PredBBTI->getNumSuccessors(); i != e; ++i)
2516 if (PredBBTI->getSuccessor(i) == BB) {
2517 BB->removePredecessor(PredBB);
2518 PredBBTI->setSuccessor(i, EdgeBB);
2519 }
2520
2521 Updates.push_back({DominatorTree::Insert, PredBB, EdgeBB});
2522 Updates.push_back({DominatorTree::Delete, PredBB, BB});
2523
2524 if (DTU)
2525 DTU->applyUpdates(Updates);
2526
2527 // Recurse, simplifying any other constants.
2528 return FoldCondBranchOnPHI(BI, DTU, DL, AC) || true;
2529 }
2530
2531 return false;
2532 }
2533
2534 /// Given a BB that starts with the specified two-entry PHI node,
2535 /// see if we can eliminate it.
FoldTwoEntryPHINode(PHINode * PN,const TargetTransformInfo & TTI,DomTreeUpdater * DTU,const DataLayout & DL)2536 static bool FoldTwoEntryPHINode(PHINode *PN, const TargetTransformInfo &TTI,
2537 DomTreeUpdater *DTU, const DataLayout &DL) {
2538 // Ok, this is a two entry PHI node. Check to see if this is a simple "if
2539 // statement", which has a very simple dominance structure. Basically, we
2540 // are trying to find the condition that is being branched on, which
2541 // subsequently causes this merge to happen. We really want control
2542 // dependence information for this check, but simplifycfg can't keep it up
2543 // to date, and this catches most of the cases we care about anyway.
2544 BasicBlock *BB = PN->getParent();
2545
2546 BasicBlock *IfTrue, *IfFalse;
2547 Value *IfCond = GetIfCondition(BB, IfTrue, IfFalse);
2548 if (!IfCond ||
2549 // Don't bother if the branch will be constant folded trivially.
2550 isa<ConstantInt>(IfCond))
2551 return false;
2552
2553 // Okay, we found that we can merge this two-entry phi node into a select.
2554 // Doing so would require us to fold *all* two entry phi nodes in this block.
2555 // At some point this becomes non-profitable (particularly if the target
2556 // doesn't support cmov's). Only do this transformation if there are two or
2557 // fewer PHI nodes in this block.
2558 unsigned NumPhis = 0;
2559 for (BasicBlock::iterator I = BB->begin(); isa<PHINode>(I); ++NumPhis, ++I)
2560 if (NumPhis > 2)
2561 return false;
2562
2563 // Loop over the PHI's seeing if we can promote them all to select
2564 // instructions. While we are at it, keep track of the instructions
2565 // that need to be moved to the dominating block.
2566 SmallPtrSet<Instruction *, 4> AggressiveInsts;
2567 int BudgetRemaining =
2568 TwoEntryPHINodeFoldingThreshold * TargetTransformInfo::TCC_Basic;
2569
2570 bool Changed = false;
2571 for (BasicBlock::iterator II = BB->begin(); isa<PHINode>(II);) {
2572 PHINode *PN = cast<PHINode>(II++);
2573 if (Value *V = SimplifyInstruction(PN, {DL, PN})) {
2574 PN->replaceAllUsesWith(V);
2575 PN->eraseFromParent();
2576 Changed = true;
2577 continue;
2578 }
2579
2580 if (!DominatesMergePoint(PN->getIncomingValue(0), BB, AggressiveInsts,
2581 BudgetRemaining, TTI) ||
2582 !DominatesMergePoint(PN->getIncomingValue(1), BB, AggressiveInsts,
2583 BudgetRemaining, TTI))
2584 return Changed;
2585 }
2586
2587 // If we folded the first phi, PN dangles at this point. Refresh it. If
2588 // we ran out of PHIs then we simplified them all.
2589 PN = dyn_cast<PHINode>(BB->begin());
2590 if (!PN)
2591 return true;
2592
2593 // Return true if at least one of these is a 'not', and another is either
2594 // a 'not' too, or a constant.
2595 auto CanHoistNotFromBothValues = [](Value *V0, Value *V1) {
2596 if (!match(V0, m_Not(m_Value())))
2597 std::swap(V0, V1);
2598 auto Invertible = m_CombineOr(m_Not(m_Value()), m_AnyIntegralConstant());
2599 return match(V0, m_Not(m_Value())) && match(V1, Invertible);
2600 };
2601
2602 // Don't fold i1 branches on PHIs which contain binary operators, unless one
2603 // of the incoming values is an 'not' and another one is freely invertible.
2604 // These can often be turned into switches and other things.
2605 if (PN->getType()->isIntegerTy(1) &&
2606 (isa<BinaryOperator>(PN->getIncomingValue(0)) ||
2607 isa<BinaryOperator>(PN->getIncomingValue(1)) ||
2608 isa<BinaryOperator>(IfCond)) &&
2609 !CanHoistNotFromBothValues(PN->getIncomingValue(0),
2610 PN->getIncomingValue(1)))
2611 return Changed;
2612
2613 // If all PHI nodes are promotable, check to make sure that all instructions
2614 // in the predecessor blocks can be promoted as well. If not, we won't be able
2615 // to get rid of the control flow, so it's not worth promoting to select
2616 // instructions.
2617 BasicBlock *DomBlock = nullptr;
2618 BasicBlock *IfBlock1 = PN->getIncomingBlock(0);
2619 BasicBlock *IfBlock2 = PN->getIncomingBlock(1);
2620 if (cast<BranchInst>(IfBlock1->getTerminator())->isConditional()) {
2621 IfBlock1 = nullptr;
2622 } else {
2623 DomBlock = *pred_begin(IfBlock1);
2624 for (BasicBlock::iterator I = IfBlock1->begin(); !I->isTerminator(); ++I)
2625 if (!AggressiveInsts.count(&*I) && !isa<DbgInfoIntrinsic>(I) &&
2626 !isa<PseudoProbeInst>(I)) {
2627 // This is not an aggressive instruction that we can promote.
2628 // Because of this, we won't be able to get rid of the control flow, so
2629 // the xform is not worth it.
2630 return Changed;
2631 }
2632 }
2633
2634 if (cast<BranchInst>(IfBlock2->getTerminator())->isConditional()) {
2635 IfBlock2 = nullptr;
2636 } else {
2637 DomBlock = *pred_begin(IfBlock2);
2638 for (BasicBlock::iterator I = IfBlock2->begin(); !I->isTerminator(); ++I)
2639 if (!AggressiveInsts.count(&*I) && !isa<DbgInfoIntrinsic>(I) &&
2640 !isa<PseudoProbeInst>(I)) {
2641 // This is not an aggressive instruction that we can promote.
2642 // Because of this, we won't be able to get rid of the control flow, so
2643 // the xform is not worth it.
2644 return Changed;
2645 }
2646 }
2647 assert(DomBlock && "Failed to find root DomBlock");
2648
2649 LLVM_DEBUG(dbgs() << "FOUND IF CONDITION! " << *IfCond
2650 << " T: " << IfTrue->getName()
2651 << " F: " << IfFalse->getName() << "\n");
2652
2653 // If we can still promote the PHI nodes after this gauntlet of tests,
2654 // do all of the PHI's now.
2655 Instruction *InsertPt = DomBlock->getTerminator();
2656 IRBuilder<NoFolder> Builder(InsertPt);
2657
2658 // Move all 'aggressive' instructions, which are defined in the
2659 // conditional parts of the if's up to the dominating block.
2660 if (IfBlock1)
2661 hoistAllInstructionsInto(DomBlock, InsertPt, IfBlock1);
2662 if (IfBlock2)
2663 hoistAllInstructionsInto(DomBlock, InsertPt, IfBlock2);
2664
2665 // Propagate fast-math-flags from phi nodes to replacement selects.
2666 IRBuilder<>::FastMathFlagGuard FMFGuard(Builder);
2667 while (PHINode *PN = dyn_cast<PHINode>(BB->begin())) {
2668 if (isa<FPMathOperator>(PN))
2669 Builder.setFastMathFlags(PN->getFastMathFlags());
2670
2671 // Change the PHI node into a select instruction.
2672 Value *TrueVal = PN->getIncomingValue(PN->getIncomingBlock(0) == IfFalse);
2673 Value *FalseVal = PN->getIncomingValue(PN->getIncomingBlock(0) == IfTrue);
2674
2675 Value *Sel = Builder.CreateSelect(IfCond, TrueVal, FalseVal, "", InsertPt);
2676 PN->replaceAllUsesWith(Sel);
2677 Sel->takeName(PN);
2678 PN->eraseFromParent();
2679 }
2680
2681 // At this point, IfBlock1 and IfBlock2 are both empty, so our if statement
2682 // has been flattened. Change DomBlock to jump directly to our new block to
2683 // avoid other simplifycfg's kicking in on the diamond.
2684 Instruction *OldTI = DomBlock->getTerminator();
2685 Builder.SetInsertPoint(OldTI);
2686 Builder.CreateBr(BB);
2687
2688 SmallVector<DominatorTree::UpdateType, 3> Updates;
2689 if (DTU) {
2690 Updates.push_back({DominatorTree::Insert, DomBlock, BB});
2691 for (auto *Successor : successors(DomBlock))
2692 Updates.push_back({DominatorTree::Delete, DomBlock, Successor});
2693 }
2694
2695 OldTI->eraseFromParent();
2696 if (DTU)
2697 DTU->applyUpdates(Updates);
2698
2699 return true;
2700 }
2701
2702 /// If we found a conditional branch that goes to two returning blocks,
2703 /// try to merge them together into one return,
2704 /// introducing a select if the return values disagree.
SimplifyCondBranchToTwoReturns(BranchInst * BI,IRBuilder<> & Builder)2705 bool SimplifyCFGOpt::SimplifyCondBranchToTwoReturns(BranchInst *BI,
2706 IRBuilder<> &Builder) {
2707 auto *BB = BI->getParent();
2708 assert(BI->isConditional() && "Must be a conditional branch");
2709 BasicBlock *TrueSucc = BI->getSuccessor(0);
2710 BasicBlock *FalseSucc = BI->getSuccessor(1);
2711 // NOTE: destinations may match, this could be degenerate uncond branch.
2712 ReturnInst *TrueRet = cast<ReturnInst>(TrueSucc->getTerminator());
2713 ReturnInst *FalseRet = cast<ReturnInst>(FalseSucc->getTerminator());
2714
2715 // Check to ensure both blocks are empty (just a return) or optionally empty
2716 // with PHI nodes. If there are other instructions, merging would cause extra
2717 // computation on one path or the other.
2718 if (!TrueSucc->getFirstNonPHIOrDbg()->isTerminator())
2719 return false;
2720 if (!FalseSucc->getFirstNonPHIOrDbg()->isTerminator())
2721 return false;
2722
2723 Builder.SetInsertPoint(BI);
2724 // Okay, we found a branch that is going to two return nodes. If
2725 // there is no return value for this function, just change the
2726 // branch into a return.
2727 if (FalseRet->getNumOperands() == 0) {
2728 TrueSucc->removePredecessor(BB);
2729 FalseSucc->removePredecessor(BB);
2730 Builder.CreateRetVoid();
2731 EraseTerminatorAndDCECond(BI);
2732 if (DTU) {
2733 SmallVector<DominatorTree::UpdateType, 2> Updates;
2734 Updates.push_back({DominatorTree::Delete, BB, TrueSucc});
2735 if (TrueSucc != FalseSucc)
2736 Updates.push_back({DominatorTree::Delete, BB, FalseSucc});
2737 DTU->applyUpdates(Updates);
2738 }
2739 return true;
2740 }
2741
2742 // Otherwise, figure out what the true and false return values are
2743 // so we can insert a new select instruction.
2744 Value *TrueValue = TrueRet->getReturnValue();
2745 Value *FalseValue = FalseRet->getReturnValue();
2746
2747 // Unwrap any PHI nodes in the return blocks.
2748 if (PHINode *TVPN = dyn_cast_or_null<PHINode>(TrueValue))
2749 if (TVPN->getParent() == TrueSucc)
2750 TrueValue = TVPN->getIncomingValueForBlock(BB);
2751 if (PHINode *FVPN = dyn_cast_or_null<PHINode>(FalseValue))
2752 if (FVPN->getParent() == FalseSucc)
2753 FalseValue = FVPN->getIncomingValueForBlock(BB);
2754
2755 // In order for this transformation to be safe, we must be able to
2756 // unconditionally execute both operands to the return. This is
2757 // normally the case, but we could have a potentially-trapping
2758 // constant expression that prevents this transformation from being
2759 // safe.
2760 if (ConstantExpr *TCV = dyn_cast_or_null<ConstantExpr>(TrueValue))
2761 if (TCV->canTrap())
2762 return false;
2763 if (ConstantExpr *FCV = dyn_cast_or_null<ConstantExpr>(FalseValue))
2764 if (FCV->canTrap())
2765 return false;
2766
2767 // Okay, we collected all the mapped values and checked them for sanity, and
2768 // defined to really do this transformation. First, update the CFG.
2769 TrueSucc->removePredecessor(BB);
2770 FalseSucc->removePredecessor(BB);
2771
2772 // Insert select instructions where needed.
2773 Value *BrCond = BI->getCondition();
2774 if (TrueValue) {
2775 // Insert a select if the results differ.
2776 if (TrueValue == FalseValue || isa<UndefValue>(FalseValue)) {
2777 } else if (isa<UndefValue>(TrueValue)) {
2778 TrueValue = FalseValue;
2779 } else {
2780 TrueValue =
2781 Builder.CreateSelect(BrCond, TrueValue, FalseValue, "retval", BI);
2782 }
2783 }
2784
2785 Value *RI =
2786 !TrueValue ? Builder.CreateRetVoid() : Builder.CreateRet(TrueValue);
2787
2788 (void)RI;
2789
2790 LLVM_DEBUG(dbgs() << "\nCHANGING BRANCH TO TWO RETURNS INTO SELECT:"
2791 << "\n " << *BI << "\nNewRet = " << *RI << "\nTRUEBLOCK: "
2792 << *TrueSucc << "\nFALSEBLOCK: " << *FalseSucc);
2793
2794 EraseTerminatorAndDCECond(BI);
2795 if (DTU) {
2796 SmallVector<DominatorTree::UpdateType, 2> Updates;
2797 Updates.push_back({DominatorTree::Delete, BB, TrueSucc});
2798 if (TrueSucc != FalseSucc)
2799 Updates.push_back({DominatorTree::Delete, BB, FalseSucc});
2800 DTU->applyUpdates(Updates);
2801 }
2802
2803 return true;
2804 }
2805
2806 /// Return true if either PBI or BI has branch weight available, and store
2807 /// the weights in {Pred|Succ}{True|False}Weight. If one of PBI and BI does
2808 /// not have branch weight, use 1:1 as its weight.
extractPredSuccWeights(BranchInst * PBI,BranchInst * BI,uint64_t & PredTrueWeight,uint64_t & PredFalseWeight,uint64_t & SuccTrueWeight,uint64_t & SuccFalseWeight)2809 static bool extractPredSuccWeights(BranchInst *PBI, BranchInst *BI,
2810 uint64_t &PredTrueWeight,
2811 uint64_t &PredFalseWeight,
2812 uint64_t &SuccTrueWeight,
2813 uint64_t &SuccFalseWeight) {
2814 bool PredHasWeights =
2815 PBI->extractProfMetadata(PredTrueWeight, PredFalseWeight);
2816 bool SuccHasWeights =
2817 BI->extractProfMetadata(SuccTrueWeight, SuccFalseWeight);
2818 if (PredHasWeights || SuccHasWeights) {
2819 if (!PredHasWeights)
2820 PredTrueWeight = PredFalseWeight = 1;
2821 if (!SuccHasWeights)
2822 SuccTrueWeight = SuccFalseWeight = 1;
2823 return true;
2824 } else {
2825 return false;
2826 }
2827 }
2828
2829 // Determine if the two branches share a common destination,
2830 // and deduce a glue that we need to use to join branch's conditions
2831 // to arrive at the common destination.
2832 static Optional<std::pair<Instruction::BinaryOps, bool>>
CheckIfCondBranchesShareCommonDestination(BranchInst * BI,BranchInst * PBI)2833 CheckIfCondBranchesShareCommonDestination(BranchInst *BI, BranchInst *PBI) {
2834 assert(BI && PBI && BI->isConditional() && PBI->isConditional() &&
2835 "Both blocks must end with a conditional branches.");
2836 assert(is_contained(predecessors(BI->getParent()), PBI->getParent()) &&
2837 "PredBB must be a predecessor of BB.");
2838
2839 if (PBI->getSuccessor(0) == BI->getSuccessor(0))
2840 return {{Instruction::Or, false}};
2841 else if (PBI->getSuccessor(1) == BI->getSuccessor(1))
2842 return {{Instruction::And, false}};
2843 else if (PBI->getSuccessor(0) == BI->getSuccessor(1))
2844 return {{Instruction::And, true}};
2845 else if (PBI->getSuccessor(1) == BI->getSuccessor(0))
2846 return {{Instruction::Or, true}};
2847 return None;
2848 }
2849
PerformBranchToCommonDestFolding(BranchInst * BI,BranchInst * PBI,DomTreeUpdater * DTU,MemorySSAUpdater * MSSAU)2850 static bool PerformBranchToCommonDestFolding(BranchInst *BI, BranchInst *PBI,
2851 DomTreeUpdater *DTU,
2852 MemorySSAUpdater *MSSAU) {
2853 BasicBlock *BB = BI->getParent();
2854 BasicBlock *PredBlock = PBI->getParent();
2855
2856 // Determine if the two branches share a common destination.
2857 Instruction::BinaryOps Opc;
2858 bool InvertPredCond;
2859 std::tie(Opc, InvertPredCond) =
2860 *CheckIfCondBranchesShareCommonDestination(BI, PBI);
2861
2862 LLVM_DEBUG(dbgs() << "FOLDING BRANCH TO COMMON DEST:\n" << *PBI << *BB);
2863
2864 IRBuilder<> Builder(PBI);
2865 // The builder is used to create instructions to eliminate the branch in BB.
2866 // If BB's terminator has !annotation metadata, add it to the new
2867 // instructions.
2868 Builder.CollectMetadataToCopy(BB->getTerminator(),
2869 {LLVMContext::MD_annotation});
2870
2871 // If we need to invert the condition in the pred block to match, do so now.
2872 if (InvertPredCond) {
2873 Value *NewCond = PBI->getCondition();
2874 if (NewCond->hasOneUse() && isa<CmpInst>(NewCond)) {
2875 CmpInst *CI = cast<CmpInst>(NewCond);
2876 CI->setPredicate(CI->getInversePredicate());
2877 } else {
2878 NewCond =
2879 Builder.CreateNot(NewCond, PBI->getCondition()->getName() + ".not");
2880 }
2881
2882 PBI->setCondition(NewCond);
2883 PBI->swapSuccessors();
2884 }
2885
2886 BasicBlock *UniqueSucc =
2887 PBI->getSuccessor(0) == BB ? BI->getSuccessor(0) : BI->getSuccessor(1);
2888
2889 // Before cloning instructions, notify the successor basic block that it
2890 // is about to have a new predecessor. This will update PHI nodes,
2891 // which will allow us to update live-out uses of bonus instructions.
2892 AddPredecessorToBlock(UniqueSucc, PredBlock, BB, MSSAU);
2893
2894 // Try to update branch weights.
2895 uint64_t PredTrueWeight, PredFalseWeight, SuccTrueWeight, SuccFalseWeight;
2896 if (extractPredSuccWeights(PBI, BI, PredTrueWeight, PredFalseWeight,
2897 SuccTrueWeight, SuccFalseWeight)) {
2898 SmallVector<uint64_t, 8> NewWeights;
2899
2900 if (PBI->getSuccessor(0) == BB) {
2901 // PBI: br i1 %x, BB, FalseDest
2902 // BI: br i1 %y, UniqueSucc, FalseDest
2903 // TrueWeight is TrueWeight for PBI * TrueWeight for BI.
2904 NewWeights.push_back(PredTrueWeight * SuccTrueWeight);
2905 // FalseWeight is FalseWeight for PBI * TotalWeight for BI +
2906 // TrueWeight for PBI * FalseWeight for BI.
2907 // We assume that total weights of a BranchInst can fit into 32 bits.
2908 // Therefore, we will not have overflow using 64-bit arithmetic.
2909 NewWeights.push_back(PredFalseWeight *
2910 (SuccFalseWeight + SuccTrueWeight) +
2911 PredTrueWeight * SuccFalseWeight);
2912 } else {
2913 // PBI: br i1 %x, TrueDest, BB
2914 // BI: br i1 %y, TrueDest, UniqueSucc
2915 // TrueWeight is TrueWeight for PBI * TotalWeight for BI +
2916 // FalseWeight for PBI * TrueWeight for BI.
2917 NewWeights.push_back(PredTrueWeight * (SuccFalseWeight + SuccTrueWeight) +
2918 PredFalseWeight * SuccTrueWeight);
2919 // FalseWeight is FalseWeight for PBI * FalseWeight for BI.
2920 NewWeights.push_back(PredFalseWeight * SuccFalseWeight);
2921 }
2922
2923 // Halve the weights if any of them cannot fit in an uint32_t
2924 FitWeights(NewWeights);
2925
2926 SmallVector<uint32_t, 8> MDWeights(NewWeights.begin(), NewWeights.end());
2927 setBranchWeights(PBI, MDWeights[0], MDWeights[1]);
2928
2929 // TODO: If BB is reachable from all paths through PredBlock, then we
2930 // could replace PBI's branch probabilities with BI's.
2931 } else
2932 PBI->setMetadata(LLVMContext::MD_prof, nullptr);
2933
2934 // Now, update the CFG.
2935 PBI->setSuccessor(PBI->getSuccessor(0) != BB, UniqueSucc);
2936
2937 if (DTU)
2938 DTU->applyUpdates({{DominatorTree::Insert, PredBlock, UniqueSucc},
2939 {DominatorTree::Delete, PredBlock, BB}});
2940
2941 // If BI was a loop latch, it may have had associated loop metadata.
2942 // We need to copy it to the new latch, that is, PBI.
2943 if (MDNode *LoopMD = BI->getMetadata(LLVMContext::MD_loop))
2944 PBI->setMetadata(LLVMContext::MD_loop, LoopMD);
2945
2946 ValueToValueMapTy VMap; // maps original values to cloned values
2947 CloneInstructionsIntoPredecessorBlockAndUpdateSSAUses(BB, PredBlock, VMap);
2948
2949 // Now that the Cond was cloned into the predecessor basic block,
2950 // or/and the two conditions together.
2951 Instruction *NewCond = cast<Instruction>(Builder.CreateBinOp(
2952 Opc, PBI->getCondition(), VMap[BI->getCondition()], "or.cond"));
2953 PBI->setCondition(NewCond);
2954
2955 // Copy any debug value intrinsics into the end of PredBlock.
2956 for (Instruction &I : *BB) {
2957 if (isa<DbgInfoIntrinsic>(I)) {
2958 Instruction *NewI = I.clone();
2959 RemapInstruction(NewI, VMap,
2960 RF_NoModuleLevelChanges | RF_IgnoreMissingLocals);
2961 NewI->insertBefore(PBI);
2962 }
2963 }
2964
2965 ++NumFoldBranchToCommonDest;
2966 return true;
2967 }
2968
2969 /// If this basic block is simple enough, and if a predecessor branches to us
2970 /// and one of our successors, fold the block into the predecessor and use
2971 /// logical operations to pick the right destination.
FoldBranchToCommonDest(BranchInst * BI,DomTreeUpdater * DTU,MemorySSAUpdater * MSSAU,const TargetTransformInfo * TTI,unsigned BonusInstThreshold)2972 bool llvm::FoldBranchToCommonDest(BranchInst *BI, DomTreeUpdater *DTU,
2973 MemorySSAUpdater *MSSAU,
2974 const TargetTransformInfo *TTI,
2975 unsigned BonusInstThreshold) {
2976 // If this block ends with an unconditional branch,
2977 // let SpeculativelyExecuteBB() deal with it.
2978 if (!BI->isConditional())
2979 return false;
2980
2981 BasicBlock *BB = BI->getParent();
2982
2983 const unsigned PredCount = pred_size(BB);
2984
2985 bool Changed = false;
2986
2987 TargetTransformInfo::TargetCostKind CostKind =
2988 BB->getParent()->hasMinSize() ? TargetTransformInfo::TCK_CodeSize
2989 : TargetTransformInfo::TCK_SizeAndLatency;
2990
2991 Instruction *Cond = dyn_cast<Instruction>(BI->getCondition());
2992
2993 if (!Cond || (!isa<CmpInst>(Cond) && !isa<BinaryOperator>(Cond)) ||
2994 Cond->getParent() != BB || !Cond->hasOneUse())
2995 return Changed;
2996
2997 // Only allow this transformation if computing the condition doesn't involve
2998 // too many instructions and these involved instructions can be executed
2999 // unconditionally. We denote all involved instructions except the condition
3000 // as "bonus instructions", and only allow this transformation when the
3001 // number of the bonus instructions we'll need to create when cloning into
3002 // each predecessor does not exceed a certain threshold.
3003 unsigned NumBonusInsts = 0;
3004 for (Instruction &I : *BB) {
3005 // Don't check the branch condition comparison itself.
3006 if (&I == Cond)
3007 continue;
3008 // Ignore dbg intrinsics, and the terminator.
3009 if (isa<DbgInfoIntrinsic>(I) || isa<BranchInst>(I))
3010 continue;
3011 // I must be safe to execute unconditionally.
3012 if (!isSafeToSpeculativelyExecute(&I))
3013 return Changed;
3014
3015 // Account for the cost of duplicating this instruction into each
3016 // predecessor.
3017 NumBonusInsts += PredCount;
3018 // Early exits once we reach the limit.
3019 if (NumBonusInsts > BonusInstThreshold)
3020 return Changed;
3021 }
3022
3023 // Cond is known to be a compare or binary operator. Check to make sure that
3024 // neither operand is a potentially-trapping constant expression.
3025 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(0)))
3026 if (CE->canTrap())
3027 return Changed;
3028 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(1)))
3029 if (CE->canTrap())
3030 return Changed;
3031
3032 // Finally, don't infinitely unroll conditional loops.
3033 if (is_contained(successors(BB), BB))
3034 return Changed;
3035
3036 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
3037 BasicBlock *PredBlock = *PI;
3038 BranchInst *PBI = dyn_cast<BranchInst>(PredBlock->getTerminator());
3039
3040 // Check that we have two conditional branches. If there is a PHI node in
3041 // the common successor, verify that the same value flows in from both
3042 // blocks.
3043 if (!PBI || PBI->isUnconditional() || !SafeToMergeTerminators(BI, PBI))
3044 continue;
3045
3046 // Determine if the two branches share a common destination.
3047 Instruction::BinaryOps Opc;
3048 bool InvertPredCond;
3049 if (auto Recepie = CheckIfCondBranchesShareCommonDestination(BI, PBI))
3050 std::tie(Opc, InvertPredCond) = *Recepie;
3051 else
3052 continue;
3053
3054 // Check the cost of inserting the necessary logic before performing the
3055 // transformation.
3056 if (TTI) {
3057 Type *Ty = BI->getCondition()->getType();
3058 unsigned Cost = TTI->getArithmeticInstrCost(Opc, Ty, CostKind);
3059 if (InvertPredCond && (!PBI->getCondition()->hasOneUse() ||
3060 !isa<CmpInst>(PBI->getCondition())))
3061 Cost += TTI->getArithmeticInstrCost(Instruction::Xor, Ty, CostKind);
3062
3063 if (Cost > BranchFoldThreshold)
3064 continue;
3065 }
3066
3067 return PerformBranchToCommonDestFolding(BI, PBI, DTU, MSSAU);
3068 }
3069 return Changed;
3070 }
3071
3072 // If there is only one store in BB1 and BB2, return it, otherwise return
3073 // nullptr.
findUniqueStoreInBlocks(BasicBlock * BB1,BasicBlock * BB2)3074 static StoreInst *findUniqueStoreInBlocks(BasicBlock *BB1, BasicBlock *BB2) {
3075 StoreInst *S = nullptr;
3076 for (auto *BB : {BB1, BB2}) {
3077 if (!BB)
3078 continue;
3079 for (auto &I : *BB)
3080 if (auto *SI = dyn_cast<StoreInst>(&I)) {
3081 if (S)
3082 // Multiple stores seen.
3083 return nullptr;
3084 else
3085 S = SI;
3086 }
3087 }
3088 return S;
3089 }
3090
ensureValueAvailableInSuccessor(Value * V,BasicBlock * BB,Value * AlternativeV=nullptr)3091 static Value *ensureValueAvailableInSuccessor(Value *V, BasicBlock *BB,
3092 Value *AlternativeV = nullptr) {
3093 // PHI is going to be a PHI node that allows the value V that is defined in
3094 // BB to be referenced in BB's only successor.
3095 //
3096 // If AlternativeV is nullptr, the only value we care about in PHI is V. It
3097 // doesn't matter to us what the other operand is (it'll never get used). We
3098 // could just create a new PHI with an undef incoming value, but that could
3099 // increase register pressure if EarlyCSE/InstCombine can't fold it with some
3100 // other PHI. So here we directly look for some PHI in BB's successor with V
3101 // as an incoming operand. If we find one, we use it, else we create a new
3102 // one.
3103 //
3104 // If AlternativeV is not nullptr, we care about both incoming values in PHI.
3105 // PHI must be exactly: phi <ty> [ %BB, %V ], [ %OtherBB, %AlternativeV]
3106 // where OtherBB is the single other predecessor of BB's only successor.
3107 PHINode *PHI = nullptr;
3108 BasicBlock *Succ = BB->getSingleSuccessor();
3109
3110 for (auto I = Succ->begin(); isa<PHINode>(I); ++I)
3111 if (cast<PHINode>(I)->getIncomingValueForBlock(BB) == V) {
3112 PHI = cast<PHINode>(I);
3113 if (!AlternativeV)
3114 break;
3115
3116 assert(Succ->hasNPredecessors(2));
3117 auto PredI = pred_begin(Succ);
3118 BasicBlock *OtherPredBB = *PredI == BB ? *++PredI : *PredI;
3119 if (PHI->getIncomingValueForBlock(OtherPredBB) == AlternativeV)
3120 break;
3121 PHI = nullptr;
3122 }
3123 if (PHI)
3124 return PHI;
3125
3126 // If V is not an instruction defined in BB, just return it.
3127 if (!AlternativeV &&
3128 (!isa<Instruction>(V) || cast<Instruction>(V)->getParent() != BB))
3129 return V;
3130
3131 PHI = PHINode::Create(V->getType(), 2, "simplifycfg.merge", &Succ->front());
3132 PHI->addIncoming(V, BB);
3133 for (BasicBlock *PredBB : predecessors(Succ))
3134 if (PredBB != BB)
3135 PHI->addIncoming(
3136 AlternativeV ? AlternativeV : UndefValue::get(V->getType()), PredBB);
3137 return PHI;
3138 }
3139
mergeConditionalStoreToAddress(BasicBlock * PTB,BasicBlock * PFB,BasicBlock * QTB,BasicBlock * QFB,BasicBlock * PostBB,Value * Address,bool InvertPCond,bool InvertQCond,DomTreeUpdater * DTU,const DataLayout & DL,const TargetTransformInfo & TTI)3140 static bool mergeConditionalStoreToAddress(
3141 BasicBlock *PTB, BasicBlock *PFB, BasicBlock *QTB, BasicBlock *QFB,
3142 BasicBlock *PostBB, Value *Address, bool InvertPCond, bool InvertQCond,
3143 DomTreeUpdater *DTU, const DataLayout &DL, const TargetTransformInfo &TTI) {
3144 // For every pointer, there must be exactly two stores, one coming from
3145 // PTB or PFB, and the other from QTB or QFB. We don't support more than one
3146 // store (to any address) in PTB,PFB or QTB,QFB.
3147 // FIXME: We could relax this restriction with a bit more work and performance
3148 // testing.
3149 StoreInst *PStore = findUniqueStoreInBlocks(PTB, PFB);
3150 StoreInst *QStore = findUniqueStoreInBlocks(QTB, QFB);
3151 if (!PStore || !QStore)
3152 return false;
3153
3154 // Now check the stores are compatible.
3155 if (!QStore->isUnordered() || !PStore->isUnordered())
3156 return false;
3157
3158 // Check that sinking the store won't cause program behavior changes. Sinking
3159 // the store out of the Q blocks won't change any behavior as we're sinking
3160 // from a block to its unconditional successor. But we're moving a store from
3161 // the P blocks down through the middle block (QBI) and past both QFB and QTB.
3162 // So we need to check that there are no aliasing loads or stores in
3163 // QBI, QTB and QFB. We also need to check there are no conflicting memory
3164 // operations between PStore and the end of its parent block.
3165 //
3166 // The ideal way to do this is to query AliasAnalysis, but we don't
3167 // preserve AA currently so that is dangerous. Be super safe and just
3168 // check there are no other memory operations at all.
3169 for (auto &I : *QFB->getSinglePredecessor())
3170 if (I.mayReadOrWriteMemory())
3171 return false;
3172 for (auto &I : *QFB)
3173 if (&I != QStore && I.mayReadOrWriteMemory())
3174 return false;
3175 if (QTB)
3176 for (auto &I : *QTB)
3177 if (&I != QStore && I.mayReadOrWriteMemory())
3178 return false;
3179 for (auto I = BasicBlock::iterator(PStore), E = PStore->getParent()->end();
3180 I != E; ++I)
3181 if (&*I != PStore && I->mayReadOrWriteMemory())
3182 return false;
3183
3184 // If we're not in aggressive mode, we only optimize if we have some
3185 // confidence that by optimizing we'll allow P and/or Q to be if-converted.
3186 auto IsWorthwhile = [&](BasicBlock *BB, ArrayRef<StoreInst *> FreeStores) {
3187 if (!BB)
3188 return true;
3189 // Heuristic: if the block can be if-converted/phi-folded and the
3190 // instructions inside are all cheap (arithmetic/GEPs), it's worthwhile to
3191 // thread this store.
3192 int BudgetRemaining =
3193 PHINodeFoldingThreshold * TargetTransformInfo::TCC_Basic;
3194 for (auto &I : BB->instructionsWithoutDebug()) {
3195 // Consider terminator instruction to be free.
3196 if (I.isTerminator())
3197 continue;
3198 // If this is one the stores that we want to speculate out of this BB,
3199 // then don't count it's cost, consider it to be free.
3200 if (auto *S = dyn_cast<StoreInst>(&I))
3201 if (llvm::find(FreeStores, S))
3202 continue;
3203 // Else, we have a white-list of instructions that we are ak speculating.
3204 if (!isa<BinaryOperator>(I) && !isa<GetElementPtrInst>(I))
3205 return false; // Not in white-list - not worthwhile folding.
3206 // And finally, if this is a non-free instruction that we are okay
3207 // speculating, ensure that we consider the speculation budget.
3208 BudgetRemaining -= TTI.getUserCost(&I, TargetTransformInfo::TCK_SizeAndLatency);
3209 if (BudgetRemaining < 0)
3210 return false; // Eagerly refuse to fold as soon as we're out of budget.
3211 }
3212 assert(BudgetRemaining >= 0 &&
3213 "When we run out of budget we will eagerly return from within the "
3214 "per-instruction loop.");
3215 return true;
3216 };
3217
3218 const std::array<StoreInst *, 2> FreeStores = {PStore, QStore};
3219 if (!MergeCondStoresAggressively &&
3220 (!IsWorthwhile(PTB, FreeStores) || !IsWorthwhile(PFB, FreeStores) ||
3221 !IsWorthwhile(QTB, FreeStores) || !IsWorthwhile(QFB, FreeStores)))
3222 return false;
3223
3224 // If PostBB has more than two predecessors, we need to split it so we can
3225 // sink the store.
3226 if (std::next(pred_begin(PostBB), 2) != pred_end(PostBB)) {
3227 // We know that QFB's only successor is PostBB. And QFB has a single
3228 // predecessor. If QTB exists, then its only successor is also PostBB.
3229 // If QTB does not exist, then QFB's only predecessor has a conditional
3230 // branch to QFB and PostBB.
3231 BasicBlock *TruePred = QTB ? QTB : QFB->getSinglePredecessor();
3232 BasicBlock *NewBB =
3233 SplitBlockPredecessors(PostBB, {QFB, TruePred}, "condstore.split", DTU);
3234 if (!NewBB)
3235 return false;
3236 PostBB = NewBB;
3237 }
3238
3239 // OK, we're going to sink the stores to PostBB. The store has to be
3240 // conditional though, so first create the predicate.
3241 Value *PCond = cast<BranchInst>(PFB->getSinglePredecessor()->getTerminator())
3242 ->getCondition();
3243 Value *QCond = cast<BranchInst>(QFB->getSinglePredecessor()->getTerminator())
3244 ->getCondition();
3245
3246 Value *PPHI = ensureValueAvailableInSuccessor(PStore->getValueOperand(),
3247 PStore->getParent());
3248 Value *QPHI = ensureValueAvailableInSuccessor(QStore->getValueOperand(),
3249 QStore->getParent(), PPHI);
3250
3251 IRBuilder<> QB(&*PostBB->getFirstInsertionPt());
3252
3253 Value *PPred = PStore->getParent() == PTB ? PCond : QB.CreateNot(PCond);
3254 Value *QPred = QStore->getParent() == QTB ? QCond : QB.CreateNot(QCond);
3255
3256 if (InvertPCond)
3257 PPred = QB.CreateNot(PPred);
3258 if (InvertQCond)
3259 QPred = QB.CreateNot(QPred);
3260 Value *CombinedPred = QB.CreateOr(PPred, QPred);
3261
3262 auto *T = SplitBlockAndInsertIfThen(CombinedPred, &*QB.GetInsertPoint(),
3263 /*Unreachable=*/false,
3264 /*BranchWeights=*/nullptr, DTU);
3265 QB.SetInsertPoint(T);
3266 StoreInst *SI = cast<StoreInst>(QB.CreateStore(QPHI, Address));
3267 AAMDNodes AAMD;
3268 PStore->getAAMetadata(AAMD, /*Merge=*/false);
3269 PStore->getAAMetadata(AAMD, /*Merge=*/true);
3270 SI->setAAMetadata(AAMD);
3271 // Choose the minimum alignment. If we could prove both stores execute, we
3272 // could use biggest one. In this case, though, we only know that one of the
3273 // stores executes. And we don't know it's safe to take the alignment from a
3274 // store that doesn't execute.
3275 SI->setAlignment(std::min(PStore->getAlign(), QStore->getAlign()));
3276
3277 QStore->eraseFromParent();
3278 PStore->eraseFromParent();
3279
3280 return true;
3281 }
3282
mergeConditionalStores(BranchInst * PBI,BranchInst * QBI,DomTreeUpdater * DTU,const DataLayout & DL,const TargetTransformInfo & TTI)3283 static bool mergeConditionalStores(BranchInst *PBI, BranchInst *QBI,
3284 DomTreeUpdater *DTU, const DataLayout &DL,
3285 const TargetTransformInfo &TTI) {
3286 // The intention here is to find diamonds or triangles (see below) where each
3287 // conditional block contains a store to the same address. Both of these
3288 // stores are conditional, so they can't be unconditionally sunk. But it may
3289 // be profitable to speculatively sink the stores into one merged store at the
3290 // end, and predicate the merged store on the union of the two conditions of
3291 // PBI and QBI.
3292 //
3293 // This can reduce the number of stores executed if both of the conditions are
3294 // true, and can allow the blocks to become small enough to be if-converted.
3295 // This optimization will also chain, so that ladders of test-and-set
3296 // sequences can be if-converted away.
3297 //
3298 // We only deal with simple diamonds or triangles:
3299 //
3300 // PBI or PBI or a combination of the two
3301 // / \ | \
3302 // PTB PFB | PFB
3303 // \ / | /
3304 // QBI QBI
3305 // / \ | \
3306 // QTB QFB | QFB
3307 // \ / | /
3308 // PostBB PostBB
3309 //
3310 // We model triangles as a type of diamond with a nullptr "true" block.
3311 // Triangles are canonicalized so that the fallthrough edge is represented by
3312 // a true condition, as in the diagram above.
3313 BasicBlock *PTB = PBI->getSuccessor(0);
3314 BasicBlock *PFB = PBI->getSuccessor(1);
3315 BasicBlock *QTB = QBI->getSuccessor(0);
3316 BasicBlock *QFB = QBI->getSuccessor(1);
3317 BasicBlock *PostBB = QFB->getSingleSuccessor();
3318
3319 // Make sure we have a good guess for PostBB. If QTB's only successor is
3320 // QFB, then QFB is a better PostBB.
3321 if (QTB->getSingleSuccessor() == QFB)
3322 PostBB = QFB;
3323
3324 // If we couldn't find a good PostBB, stop.
3325 if (!PostBB)
3326 return false;
3327
3328 bool InvertPCond = false, InvertQCond = false;
3329 // Canonicalize fallthroughs to the true branches.
3330 if (PFB == QBI->getParent()) {
3331 std::swap(PFB, PTB);
3332 InvertPCond = true;
3333 }
3334 if (QFB == PostBB) {
3335 std::swap(QFB, QTB);
3336 InvertQCond = true;
3337 }
3338
3339 // From this point on we can assume PTB or QTB may be fallthroughs but PFB
3340 // and QFB may not. Model fallthroughs as a nullptr block.
3341 if (PTB == QBI->getParent())
3342 PTB = nullptr;
3343 if (QTB == PostBB)
3344 QTB = nullptr;
3345
3346 // Legality bailouts. We must have at least the non-fallthrough blocks and
3347 // the post-dominating block, and the non-fallthroughs must only have one
3348 // predecessor.
3349 auto HasOnePredAndOneSucc = [](BasicBlock *BB, BasicBlock *P, BasicBlock *S) {
3350 return BB->getSinglePredecessor() == P && BB->getSingleSuccessor() == S;
3351 };
3352 if (!HasOnePredAndOneSucc(PFB, PBI->getParent(), QBI->getParent()) ||
3353 !HasOnePredAndOneSucc(QFB, QBI->getParent(), PostBB))
3354 return false;
3355 if ((PTB && !HasOnePredAndOneSucc(PTB, PBI->getParent(), QBI->getParent())) ||
3356 (QTB && !HasOnePredAndOneSucc(QTB, QBI->getParent(), PostBB)))
3357 return false;
3358 if (!QBI->getParent()->hasNUses(2))
3359 return false;
3360
3361 // OK, this is a sequence of two diamonds or triangles.
3362 // Check if there are stores in PTB or PFB that are repeated in QTB or QFB.
3363 SmallPtrSet<Value *, 4> PStoreAddresses, QStoreAddresses;
3364 for (auto *BB : {PTB, PFB}) {
3365 if (!BB)
3366 continue;
3367 for (auto &I : *BB)
3368 if (StoreInst *SI = dyn_cast<StoreInst>(&I))
3369 PStoreAddresses.insert(SI->getPointerOperand());
3370 }
3371 for (auto *BB : {QTB, QFB}) {
3372 if (!BB)
3373 continue;
3374 for (auto &I : *BB)
3375 if (StoreInst *SI = dyn_cast<StoreInst>(&I))
3376 QStoreAddresses.insert(SI->getPointerOperand());
3377 }
3378
3379 set_intersect(PStoreAddresses, QStoreAddresses);
3380 // set_intersect mutates PStoreAddresses in place. Rename it here to make it
3381 // clear what it contains.
3382 auto &CommonAddresses = PStoreAddresses;
3383
3384 bool Changed = false;
3385 for (auto *Address : CommonAddresses)
3386 Changed |=
3387 mergeConditionalStoreToAddress(PTB, PFB, QTB, QFB, PostBB, Address,
3388 InvertPCond, InvertQCond, DTU, DL, TTI);
3389 return Changed;
3390 }
3391
3392 /// If the previous block ended with a widenable branch, determine if reusing
3393 /// the target block is profitable and legal. This will have the effect of
3394 /// "widening" PBI, but doesn't require us to reason about hosting safety.
tryWidenCondBranchToCondBranch(BranchInst * PBI,BranchInst * BI,DomTreeUpdater * DTU)3395 static bool tryWidenCondBranchToCondBranch(BranchInst *PBI, BranchInst *BI,
3396 DomTreeUpdater *DTU) {
3397 // TODO: This can be generalized in two important ways:
3398 // 1) We can allow phi nodes in IfFalseBB and simply reuse all the input
3399 // values from the PBI edge.
3400 // 2) We can sink side effecting instructions into BI's fallthrough
3401 // successor provided they doesn't contribute to computation of
3402 // BI's condition.
3403 Value *CondWB, *WC;
3404 BasicBlock *IfTrueBB, *IfFalseBB;
3405 if (!parseWidenableBranch(PBI, CondWB, WC, IfTrueBB, IfFalseBB) ||
3406 IfTrueBB != BI->getParent() || !BI->getParent()->getSinglePredecessor())
3407 return false;
3408 if (!IfFalseBB->phis().empty())
3409 return false; // TODO
3410 // Use lambda to lazily compute expensive condition after cheap ones.
3411 auto NoSideEffects = [](BasicBlock &BB) {
3412 return !llvm::any_of(BB, [](const Instruction &I) {
3413 return I.mayWriteToMemory() || I.mayHaveSideEffects();
3414 });
3415 };
3416 if (BI->getSuccessor(1) != IfFalseBB && // no inf looping
3417 BI->getSuccessor(1)->getTerminatingDeoptimizeCall() && // profitability
3418 NoSideEffects(*BI->getParent())) {
3419 auto *OldSuccessor = BI->getSuccessor(1);
3420 OldSuccessor->removePredecessor(BI->getParent());
3421 BI->setSuccessor(1, IfFalseBB);
3422 if (DTU)
3423 DTU->applyUpdates(
3424 {{DominatorTree::Insert, BI->getParent(), IfFalseBB},
3425 {DominatorTree::Delete, BI->getParent(), OldSuccessor}});
3426 return true;
3427 }
3428 if (BI->getSuccessor(0) != IfFalseBB && // no inf looping
3429 BI->getSuccessor(0)->getTerminatingDeoptimizeCall() && // profitability
3430 NoSideEffects(*BI->getParent())) {
3431 auto *OldSuccessor = BI->getSuccessor(0);
3432 OldSuccessor->removePredecessor(BI->getParent());
3433 BI->setSuccessor(0, IfFalseBB);
3434 if (DTU)
3435 DTU->applyUpdates(
3436 {{DominatorTree::Insert, BI->getParent(), IfFalseBB},
3437 {DominatorTree::Delete, BI->getParent(), OldSuccessor}});
3438 return true;
3439 }
3440 return false;
3441 }
3442
3443 /// If we have a conditional branch as a predecessor of another block,
3444 /// this function tries to simplify it. We know
3445 /// that PBI and BI are both conditional branches, and BI is in one of the
3446 /// successor blocks of PBI - PBI branches to BI.
SimplifyCondBranchToCondBranch(BranchInst * PBI,BranchInst * BI,DomTreeUpdater * DTU,const DataLayout & DL,const TargetTransformInfo & TTI)3447 static bool SimplifyCondBranchToCondBranch(BranchInst *PBI, BranchInst *BI,
3448 DomTreeUpdater *DTU,
3449 const DataLayout &DL,
3450 const TargetTransformInfo &TTI) {
3451 assert(PBI->isConditional() && BI->isConditional());
3452 BasicBlock *BB = BI->getParent();
3453
3454 // If this block ends with a branch instruction, and if there is a
3455 // predecessor that ends on a branch of the same condition, make
3456 // this conditional branch redundant.
3457 if (PBI->getCondition() == BI->getCondition() &&
3458 PBI->getSuccessor(0) != PBI->getSuccessor(1)) {
3459 // Okay, the outcome of this conditional branch is statically
3460 // knowable. If this block had a single pred, handle specially.
3461 if (BB->getSinglePredecessor()) {
3462 // Turn this into a branch on constant.
3463 bool CondIsTrue = PBI->getSuccessor(0) == BB;
3464 BI->setCondition(
3465 ConstantInt::get(Type::getInt1Ty(BB->getContext()), CondIsTrue));
3466 return true; // Nuke the branch on constant.
3467 }
3468
3469 // Otherwise, if there are multiple predecessors, insert a PHI that merges
3470 // in the constant and simplify the block result. Subsequent passes of
3471 // simplifycfg will thread the block.
3472 if (BlockIsSimpleEnoughToThreadThrough(BB)) {
3473 pred_iterator PB = pred_begin(BB), PE = pred_end(BB);
3474 PHINode *NewPN = PHINode::Create(
3475 Type::getInt1Ty(BB->getContext()), std::distance(PB, PE),
3476 BI->getCondition()->getName() + ".pr", &BB->front());
3477 // Okay, we're going to insert the PHI node. Since PBI is not the only
3478 // predecessor, compute the PHI'd conditional value for all of the preds.
3479 // Any predecessor where the condition is not computable we keep symbolic.
3480 for (pred_iterator PI = PB; PI != PE; ++PI) {
3481 BasicBlock *P = *PI;
3482 if ((PBI = dyn_cast<BranchInst>(P->getTerminator())) && PBI != BI &&
3483 PBI->isConditional() && PBI->getCondition() == BI->getCondition() &&
3484 PBI->getSuccessor(0) != PBI->getSuccessor(1)) {
3485 bool CondIsTrue = PBI->getSuccessor(0) == BB;
3486 NewPN->addIncoming(
3487 ConstantInt::get(Type::getInt1Ty(BB->getContext()), CondIsTrue),
3488 P);
3489 } else {
3490 NewPN->addIncoming(BI->getCondition(), P);
3491 }
3492 }
3493
3494 BI->setCondition(NewPN);
3495 return true;
3496 }
3497 }
3498
3499 // If the previous block ended with a widenable branch, determine if reusing
3500 // the target block is profitable and legal. This will have the effect of
3501 // "widening" PBI, but doesn't require us to reason about hosting safety.
3502 if (tryWidenCondBranchToCondBranch(PBI, BI, DTU))
3503 return true;
3504
3505 if (auto *CE = dyn_cast<ConstantExpr>(BI->getCondition()))
3506 if (CE->canTrap())
3507 return false;
3508
3509 // If both branches are conditional and both contain stores to the same
3510 // address, remove the stores from the conditionals and create a conditional
3511 // merged store at the end.
3512 if (MergeCondStores && mergeConditionalStores(PBI, BI, DTU, DL, TTI))
3513 return true;
3514
3515 // If this is a conditional branch in an empty block, and if any
3516 // predecessors are a conditional branch to one of our destinations,
3517 // fold the conditions into logical ops and one cond br.
3518
3519 // Ignore dbg intrinsics.
3520 if (&*BB->instructionsWithoutDebug().begin() != BI)
3521 return false;
3522
3523 int PBIOp, BIOp;
3524 if (PBI->getSuccessor(0) == BI->getSuccessor(0)) {
3525 PBIOp = 0;
3526 BIOp = 0;
3527 } else if (PBI->getSuccessor(0) == BI->getSuccessor(1)) {
3528 PBIOp = 0;
3529 BIOp = 1;
3530 } else if (PBI->getSuccessor(1) == BI->getSuccessor(0)) {
3531 PBIOp = 1;
3532 BIOp = 0;
3533 } else if (PBI->getSuccessor(1) == BI->getSuccessor(1)) {
3534 PBIOp = 1;
3535 BIOp = 1;
3536 } else {
3537 return false;
3538 }
3539
3540 // Check to make sure that the other destination of this branch
3541 // isn't BB itself. If so, this is an infinite loop that will
3542 // keep getting unwound.
3543 if (PBI->getSuccessor(PBIOp) == BB)
3544 return false;
3545
3546 // Do not perform this transformation if it would require
3547 // insertion of a large number of select instructions. For targets
3548 // without predication/cmovs, this is a big pessimization.
3549
3550 // Also do not perform this transformation if any phi node in the common
3551 // destination block can trap when reached by BB or PBB (PR17073). In that
3552 // case, it would be unsafe to hoist the operation into a select instruction.
3553
3554 BasicBlock *CommonDest = PBI->getSuccessor(PBIOp);
3555 BasicBlock *RemovedDest = PBI->getSuccessor(PBIOp ^ 1);
3556 unsigned NumPhis = 0;
3557 for (BasicBlock::iterator II = CommonDest->begin(); isa<PHINode>(II);
3558 ++II, ++NumPhis) {
3559 if (NumPhis > 2) // Disable this xform.
3560 return false;
3561
3562 PHINode *PN = cast<PHINode>(II);
3563 Value *BIV = PN->getIncomingValueForBlock(BB);
3564 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(BIV))
3565 if (CE->canTrap())
3566 return false;
3567
3568 unsigned PBBIdx = PN->getBasicBlockIndex(PBI->getParent());
3569 Value *PBIV = PN->getIncomingValue(PBBIdx);
3570 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(PBIV))
3571 if (CE->canTrap())
3572 return false;
3573 }
3574
3575 // Finally, if everything is ok, fold the branches to logical ops.
3576 BasicBlock *OtherDest = BI->getSuccessor(BIOp ^ 1);
3577
3578 LLVM_DEBUG(dbgs() << "FOLDING BRs:" << *PBI->getParent()
3579 << "AND: " << *BI->getParent());
3580
3581 SmallVector<DominatorTree::UpdateType, 5> Updates;
3582
3583 // If OtherDest *is* BB, then BB is a basic block with a single conditional
3584 // branch in it, where one edge (OtherDest) goes back to itself but the other
3585 // exits. We don't *know* that the program avoids the infinite loop
3586 // (even though that seems likely). If we do this xform naively, we'll end up
3587 // recursively unpeeling the loop. Since we know that (after the xform is
3588 // done) that the block *is* infinite if reached, we just make it an obviously
3589 // infinite loop with no cond branch.
3590 if (OtherDest == BB) {
3591 // Insert it at the end of the function, because it's either code,
3592 // or it won't matter if it's hot. :)
3593 BasicBlock *InfLoopBlock =
3594 BasicBlock::Create(BB->getContext(), "infloop", BB->getParent());
3595 BranchInst::Create(InfLoopBlock, InfLoopBlock);
3596 Updates.push_back({DominatorTree::Insert, InfLoopBlock, InfLoopBlock});
3597 OtherDest = InfLoopBlock;
3598 }
3599
3600 LLVM_DEBUG(dbgs() << *PBI->getParent()->getParent());
3601
3602 // BI may have other predecessors. Because of this, we leave
3603 // it alone, but modify PBI.
3604
3605 // Make sure we get to CommonDest on True&True directions.
3606 Value *PBICond = PBI->getCondition();
3607 IRBuilder<NoFolder> Builder(PBI);
3608 if (PBIOp)
3609 PBICond = Builder.CreateNot(PBICond, PBICond->getName() + ".not");
3610
3611 Value *BICond = BI->getCondition();
3612 if (BIOp)
3613 BICond = Builder.CreateNot(BICond, BICond->getName() + ".not");
3614
3615 // Merge the conditions.
3616 Value *Cond = Builder.CreateOr(PBICond, BICond, "brmerge");
3617
3618 // Modify PBI to branch on the new condition to the new dests.
3619 PBI->setCondition(Cond);
3620 PBI->setSuccessor(0, CommonDest);
3621 PBI->setSuccessor(1, OtherDest);
3622
3623 Updates.push_back({DominatorTree::Insert, PBI->getParent(), OtherDest});
3624 Updates.push_back({DominatorTree::Delete, PBI->getParent(), RemovedDest});
3625
3626 if (DTU)
3627 DTU->applyUpdates(Updates);
3628
3629 // Update branch weight for PBI.
3630 uint64_t PredTrueWeight, PredFalseWeight, SuccTrueWeight, SuccFalseWeight;
3631 uint64_t PredCommon, PredOther, SuccCommon, SuccOther;
3632 bool HasWeights =
3633 extractPredSuccWeights(PBI, BI, PredTrueWeight, PredFalseWeight,
3634 SuccTrueWeight, SuccFalseWeight);
3635 if (HasWeights) {
3636 PredCommon = PBIOp ? PredFalseWeight : PredTrueWeight;
3637 PredOther = PBIOp ? PredTrueWeight : PredFalseWeight;
3638 SuccCommon = BIOp ? SuccFalseWeight : SuccTrueWeight;
3639 SuccOther = BIOp ? SuccTrueWeight : SuccFalseWeight;
3640 // The weight to CommonDest should be PredCommon * SuccTotal +
3641 // PredOther * SuccCommon.
3642 // The weight to OtherDest should be PredOther * SuccOther.
3643 uint64_t NewWeights[2] = {PredCommon * (SuccCommon + SuccOther) +
3644 PredOther * SuccCommon,
3645 PredOther * SuccOther};
3646 // Halve the weights if any of them cannot fit in an uint32_t
3647 FitWeights(NewWeights);
3648
3649 setBranchWeights(PBI, NewWeights[0], NewWeights[1]);
3650 }
3651
3652 // OtherDest may have phi nodes. If so, add an entry from PBI's
3653 // block that are identical to the entries for BI's block.
3654 AddPredecessorToBlock(OtherDest, PBI->getParent(), BB);
3655
3656 // We know that the CommonDest already had an edge from PBI to
3657 // it. If it has PHIs though, the PHIs may have different
3658 // entries for BB and PBI's BB. If so, insert a select to make
3659 // them agree.
3660 for (PHINode &PN : CommonDest->phis()) {
3661 Value *BIV = PN.getIncomingValueForBlock(BB);
3662 unsigned PBBIdx = PN.getBasicBlockIndex(PBI->getParent());
3663 Value *PBIV = PN.getIncomingValue(PBBIdx);
3664 if (BIV != PBIV) {
3665 // Insert a select in PBI to pick the right value.
3666 SelectInst *NV = cast<SelectInst>(
3667 Builder.CreateSelect(PBICond, PBIV, BIV, PBIV->getName() + ".mux"));
3668 PN.setIncomingValue(PBBIdx, NV);
3669 // Although the select has the same condition as PBI, the original branch
3670 // weights for PBI do not apply to the new select because the select's
3671 // 'logical' edges are incoming edges of the phi that is eliminated, not
3672 // the outgoing edges of PBI.
3673 if (HasWeights) {
3674 uint64_t PredCommon = PBIOp ? PredFalseWeight : PredTrueWeight;
3675 uint64_t PredOther = PBIOp ? PredTrueWeight : PredFalseWeight;
3676 uint64_t SuccCommon = BIOp ? SuccFalseWeight : SuccTrueWeight;
3677 uint64_t SuccOther = BIOp ? SuccTrueWeight : SuccFalseWeight;
3678 // The weight to PredCommonDest should be PredCommon * SuccTotal.
3679 // The weight to PredOtherDest should be PredOther * SuccCommon.
3680 uint64_t NewWeights[2] = {PredCommon * (SuccCommon + SuccOther),
3681 PredOther * SuccCommon};
3682
3683 FitWeights(NewWeights);
3684
3685 setBranchWeights(NV, NewWeights[0], NewWeights[1]);
3686 }
3687 }
3688 }
3689
3690 LLVM_DEBUG(dbgs() << "INTO: " << *PBI->getParent());
3691 LLVM_DEBUG(dbgs() << *PBI->getParent()->getParent());
3692
3693 // This basic block is probably dead. We know it has at least
3694 // one fewer predecessor.
3695 return true;
3696 }
3697
3698 // Simplifies a terminator by replacing it with a branch to TrueBB if Cond is
3699 // true or to FalseBB if Cond is false.
3700 // Takes care of updating the successors and removing the old terminator.
3701 // Also makes sure not to introduce new successors by assuming that edges to
3702 // non-successor TrueBBs and FalseBBs aren't reachable.
SimplifyTerminatorOnSelect(Instruction * OldTerm,Value * Cond,BasicBlock * TrueBB,BasicBlock * FalseBB,uint32_t TrueWeight,uint32_t FalseWeight)3703 bool SimplifyCFGOpt::SimplifyTerminatorOnSelect(Instruction *OldTerm,
3704 Value *Cond, BasicBlock *TrueBB,
3705 BasicBlock *FalseBB,
3706 uint32_t TrueWeight,
3707 uint32_t FalseWeight) {
3708 auto *BB = OldTerm->getParent();
3709 // Remove any superfluous successor edges from the CFG.
3710 // First, figure out which successors to preserve.
3711 // If TrueBB and FalseBB are equal, only try to preserve one copy of that
3712 // successor.
3713 BasicBlock *KeepEdge1 = TrueBB;
3714 BasicBlock *KeepEdge2 = TrueBB != FalseBB ? FalseBB : nullptr;
3715
3716 SmallSetVector<BasicBlock *, 2> RemovedSuccessors;
3717
3718 // Then remove the rest.
3719 for (BasicBlock *Succ : successors(OldTerm)) {
3720 // Make sure only to keep exactly one copy of each edge.
3721 if (Succ == KeepEdge1)
3722 KeepEdge1 = nullptr;
3723 else if (Succ == KeepEdge2)
3724 KeepEdge2 = nullptr;
3725 else {
3726 Succ->removePredecessor(BB,
3727 /*KeepOneInputPHIs=*/true);
3728
3729 if (Succ != TrueBB && Succ != FalseBB)
3730 RemovedSuccessors.insert(Succ);
3731 }
3732 }
3733
3734 IRBuilder<> Builder(OldTerm);
3735 Builder.SetCurrentDebugLocation(OldTerm->getDebugLoc());
3736
3737 // Insert an appropriate new terminator.
3738 if (!KeepEdge1 && !KeepEdge2) {
3739 if (TrueBB == FalseBB) {
3740 // We were only looking for one successor, and it was present.
3741 // Create an unconditional branch to it.
3742 Builder.CreateBr(TrueBB);
3743 } else {
3744 // We found both of the successors we were looking for.
3745 // Create a conditional branch sharing the condition of the select.
3746 BranchInst *NewBI = Builder.CreateCondBr(Cond, TrueBB, FalseBB);
3747 if (TrueWeight != FalseWeight)
3748 setBranchWeights(NewBI, TrueWeight, FalseWeight);
3749 }
3750 } else if (KeepEdge1 && (KeepEdge2 || TrueBB == FalseBB)) {
3751 // Neither of the selected blocks were successors, so this
3752 // terminator must be unreachable.
3753 new UnreachableInst(OldTerm->getContext(), OldTerm);
3754 } else {
3755 // One of the selected values was a successor, but the other wasn't.
3756 // Insert an unconditional branch to the one that was found;
3757 // the edge to the one that wasn't must be unreachable.
3758 if (!KeepEdge1) {
3759 // Only TrueBB was found.
3760 Builder.CreateBr(TrueBB);
3761 } else {
3762 // Only FalseBB was found.
3763 Builder.CreateBr(FalseBB);
3764 }
3765 }
3766
3767 EraseTerminatorAndDCECond(OldTerm);
3768
3769 if (DTU) {
3770 SmallVector<DominatorTree::UpdateType, 2> Updates;
3771 Updates.reserve(RemovedSuccessors.size());
3772 for (auto *RemovedSuccessor : RemovedSuccessors)
3773 Updates.push_back({DominatorTree::Delete, BB, RemovedSuccessor});
3774 DTU->applyUpdates(Updates);
3775 }
3776
3777 return true;
3778 }
3779
3780 // Replaces
3781 // (switch (select cond, X, Y)) on constant X, Y
3782 // with a branch - conditional if X and Y lead to distinct BBs,
3783 // unconditional otherwise.
SimplifySwitchOnSelect(SwitchInst * SI,SelectInst * Select)3784 bool SimplifyCFGOpt::SimplifySwitchOnSelect(SwitchInst *SI,
3785 SelectInst *Select) {
3786 // Check for constant integer values in the select.
3787 ConstantInt *TrueVal = dyn_cast<ConstantInt>(Select->getTrueValue());
3788 ConstantInt *FalseVal = dyn_cast<ConstantInt>(Select->getFalseValue());
3789 if (!TrueVal || !FalseVal)
3790 return false;
3791
3792 // Find the relevant condition and destinations.
3793 Value *Condition = Select->getCondition();
3794 BasicBlock *TrueBB = SI->findCaseValue(TrueVal)->getCaseSuccessor();
3795 BasicBlock *FalseBB = SI->findCaseValue(FalseVal)->getCaseSuccessor();
3796
3797 // Get weight for TrueBB and FalseBB.
3798 uint32_t TrueWeight = 0, FalseWeight = 0;
3799 SmallVector<uint64_t, 8> Weights;
3800 bool HasWeights = HasBranchWeights(SI);
3801 if (HasWeights) {
3802 GetBranchWeights(SI, Weights);
3803 if (Weights.size() == 1 + SI->getNumCases()) {
3804 TrueWeight =
3805 (uint32_t)Weights[SI->findCaseValue(TrueVal)->getSuccessorIndex()];
3806 FalseWeight =
3807 (uint32_t)Weights[SI->findCaseValue(FalseVal)->getSuccessorIndex()];
3808 }
3809 }
3810
3811 // Perform the actual simplification.
3812 return SimplifyTerminatorOnSelect(SI, Condition, TrueBB, FalseBB, TrueWeight,
3813 FalseWeight);
3814 }
3815
3816 // Replaces
3817 // (indirectbr (select cond, blockaddress(@fn, BlockA),
3818 // blockaddress(@fn, BlockB)))
3819 // with
3820 // (br cond, BlockA, BlockB).
SimplifyIndirectBrOnSelect(IndirectBrInst * IBI,SelectInst * SI)3821 bool SimplifyCFGOpt::SimplifyIndirectBrOnSelect(IndirectBrInst *IBI,
3822 SelectInst *SI) {
3823 // Check that both operands of the select are block addresses.
3824 BlockAddress *TBA = dyn_cast<BlockAddress>(SI->getTrueValue());
3825 BlockAddress *FBA = dyn_cast<BlockAddress>(SI->getFalseValue());
3826 if (!TBA || !FBA)
3827 return false;
3828
3829 // Extract the actual blocks.
3830 BasicBlock *TrueBB = TBA->getBasicBlock();
3831 BasicBlock *FalseBB = FBA->getBasicBlock();
3832
3833 // Perform the actual simplification.
3834 return SimplifyTerminatorOnSelect(IBI, SI->getCondition(), TrueBB, FalseBB, 0,
3835 0);
3836 }
3837
3838 /// This is called when we find an icmp instruction
3839 /// (a seteq/setne with a constant) as the only instruction in a
3840 /// block that ends with an uncond branch. We are looking for a very specific
3841 /// pattern that occurs when "A == 1 || A == 2 || A == 3" gets simplified. In
3842 /// this case, we merge the first two "or's of icmp" into a switch, but then the
3843 /// default value goes to an uncond block with a seteq in it, we get something
3844 /// like:
3845 ///
3846 /// switch i8 %A, label %DEFAULT [ i8 1, label %end i8 2, label %end ]
3847 /// DEFAULT:
3848 /// %tmp = icmp eq i8 %A, 92
3849 /// br label %end
3850 /// end:
3851 /// ... = phi i1 [ true, %entry ], [ %tmp, %DEFAULT ], [ true, %entry ]
3852 ///
3853 /// We prefer to split the edge to 'end' so that there is a true/false entry to
3854 /// the PHI, merging the third icmp into the switch.
tryToSimplifyUncondBranchWithICmpInIt(ICmpInst * ICI,IRBuilder<> & Builder)3855 bool SimplifyCFGOpt::tryToSimplifyUncondBranchWithICmpInIt(
3856 ICmpInst *ICI, IRBuilder<> &Builder) {
3857 BasicBlock *BB = ICI->getParent();
3858
3859 // If the block has any PHIs in it or the icmp has multiple uses, it is too
3860 // complex.
3861 if (isa<PHINode>(BB->begin()) || !ICI->hasOneUse())
3862 return false;
3863
3864 Value *V = ICI->getOperand(0);
3865 ConstantInt *Cst = cast<ConstantInt>(ICI->getOperand(1));
3866
3867 // The pattern we're looking for is where our only predecessor is a switch on
3868 // 'V' and this block is the default case for the switch. In this case we can
3869 // fold the compared value into the switch to simplify things.
3870 BasicBlock *Pred = BB->getSinglePredecessor();
3871 if (!Pred || !isa<SwitchInst>(Pred->getTerminator()))
3872 return false;
3873
3874 SwitchInst *SI = cast<SwitchInst>(Pred->getTerminator());
3875 if (SI->getCondition() != V)
3876 return false;
3877
3878 // If BB is reachable on a non-default case, then we simply know the value of
3879 // V in this block. Substitute it and constant fold the icmp instruction
3880 // away.
3881 if (SI->getDefaultDest() != BB) {
3882 ConstantInt *VVal = SI->findCaseDest(BB);
3883 assert(VVal && "Should have a unique destination value");
3884 ICI->setOperand(0, VVal);
3885
3886 if (Value *V = SimplifyInstruction(ICI, {DL, ICI})) {
3887 ICI->replaceAllUsesWith(V);
3888 ICI->eraseFromParent();
3889 }
3890 // BB is now empty, so it is likely to simplify away.
3891 return requestResimplify();
3892 }
3893
3894 // Ok, the block is reachable from the default dest. If the constant we're
3895 // comparing exists in one of the other edges, then we can constant fold ICI
3896 // and zap it.
3897 if (SI->findCaseValue(Cst) != SI->case_default()) {
3898 Value *V;
3899 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
3900 V = ConstantInt::getFalse(BB->getContext());
3901 else
3902 V = ConstantInt::getTrue(BB->getContext());
3903
3904 ICI->replaceAllUsesWith(V);
3905 ICI->eraseFromParent();
3906 // BB is now empty, so it is likely to simplify away.
3907 return requestResimplify();
3908 }
3909
3910 // The use of the icmp has to be in the 'end' block, by the only PHI node in
3911 // the block.
3912 BasicBlock *SuccBlock = BB->getTerminator()->getSuccessor(0);
3913 PHINode *PHIUse = dyn_cast<PHINode>(ICI->user_back());
3914 if (PHIUse == nullptr || PHIUse != &SuccBlock->front() ||
3915 isa<PHINode>(++BasicBlock::iterator(PHIUse)))
3916 return false;
3917
3918 // If the icmp is a SETEQ, then the default dest gets false, the new edge gets
3919 // true in the PHI.
3920 Constant *DefaultCst = ConstantInt::getTrue(BB->getContext());
3921 Constant *NewCst = ConstantInt::getFalse(BB->getContext());
3922
3923 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
3924 std::swap(DefaultCst, NewCst);
3925
3926 // Replace ICI (which is used by the PHI for the default value) with true or
3927 // false depending on if it is EQ or NE.
3928 ICI->replaceAllUsesWith(DefaultCst);
3929 ICI->eraseFromParent();
3930
3931 SmallVector<DominatorTree::UpdateType, 2> Updates;
3932
3933 // Okay, the switch goes to this block on a default value. Add an edge from
3934 // the switch to the merge point on the compared value.
3935 BasicBlock *NewBB =
3936 BasicBlock::Create(BB->getContext(), "switch.edge", BB->getParent(), BB);
3937 {
3938 SwitchInstProfUpdateWrapper SIW(*SI);
3939 auto W0 = SIW.getSuccessorWeight(0);
3940 SwitchInstProfUpdateWrapper::CaseWeightOpt NewW;
3941 if (W0) {
3942 NewW = ((uint64_t(*W0) + 1) >> 1);
3943 SIW.setSuccessorWeight(0, *NewW);
3944 }
3945 SIW.addCase(Cst, NewBB, NewW);
3946 Updates.push_back({DominatorTree::Insert, Pred, NewBB});
3947 }
3948
3949 // NewBB branches to the phi block, add the uncond branch and the phi entry.
3950 Builder.SetInsertPoint(NewBB);
3951 Builder.SetCurrentDebugLocation(SI->getDebugLoc());
3952 Builder.CreateBr(SuccBlock);
3953 Updates.push_back({DominatorTree::Insert, NewBB, SuccBlock});
3954 PHIUse->addIncoming(NewCst, NewBB);
3955 if (DTU)
3956 DTU->applyUpdates(Updates);
3957 return true;
3958 }
3959
3960 /// The specified branch is a conditional branch.
3961 /// Check to see if it is branching on an or/and chain of icmp instructions, and
3962 /// fold it into a switch instruction if so.
SimplifyBranchOnICmpChain(BranchInst * BI,IRBuilder<> & Builder,const DataLayout & DL)3963 bool SimplifyCFGOpt::SimplifyBranchOnICmpChain(BranchInst *BI,
3964 IRBuilder<> &Builder,
3965 const DataLayout &DL) {
3966 Instruction *Cond = dyn_cast<Instruction>(BI->getCondition());
3967 if (!Cond)
3968 return false;
3969
3970 // Change br (X == 0 | X == 1), T, F into a switch instruction.
3971 // If this is a bunch of seteq's or'd together, or if it's a bunch of
3972 // 'setne's and'ed together, collect them.
3973
3974 // Try to gather values from a chain of and/or to be turned into a switch
3975 ConstantComparesGatherer ConstantCompare(Cond, DL);
3976 // Unpack the result
3977 SmallVectorImpl<ConstantInt *> &Values = ConstantCompare.Vals;
3978 Value *CompVal = ConstantCompare.CompValue;
3979 unsigned UsedICmps = ConstantCompare.UsedICmps;
3980 Value *ExtraCase = ConstantCompare.Extra;
3981
3982 // If we didn't have a multiply compared value, fail.
3983 if (!CompVal)
3984 return false;
3985
3986 // Avoid turning single icmps into a switch.
3987 if (UsedICmps <= 1)
3988 return false;
3989
3990 bool TrueWhenEqual = match(Cond, m_LogicalOr(m_Value(), m_Value()));
3991
3992 // There might be duplicate constants in the list, which the switch
3993 // instruction can't handle, remove them now.
3994 array_pod_sort(Values.begin(), Values.end(), ConstantIntSortPredicate);
3995 Values.erase(std::unique(Values.begin(), Values.end()), Values.end());
3996
3997 // If Extra was used, we require at least two switch values to do the
3998 // transformation. A switch with one value is just a conditional branch.
3999 if (ExtraCase && Values.size() < 2)
4000 return false;
4001
4002 // TODO: Preserve branch weight metadata, similarly to how
4003 // FoldValueComparisonIntoPredecessors preserves it.
4004
4005 // Figure out which block is which destination.
4006 BasicBlock *DefaultBB = BI->getSuccessor(1);
4007 BasicBlock *EdgeBB = BI->getSuccessor(0);
4008 if (!TrueWhenEqual)
4009 std::swap(DefaultBB, EdgeBB);
4010
4011 BasicBlock *BB = BI->getParent();
4012
4013 // MSAN does not like undefs as branch condition which can be introduced
4014 // with "explicit branch".
4015 if (ExtraCase && BB->getParent()->hasFnAttribute(Attribute::SanitizeMemory))
4016 return false;
4017
4018 LLVM_DEBUG(dbgs() << "Converting 'icmp' chain with " << Values.size()
4019 << " cases into SWITCH. BB is:\n"
4020 << *BB);
4021
4022 SmallVector<DominatorTree::UpdateType, 2> Updates;
4023
4024 // If there are any extra values that couldn't be folded into the switch
4025 // then we evaluate them with an explicit branch first. Split the block
4026 // right before the condbr to handle it.
4027 if (ExtraCase) {
4028 BasicBlock *NewBB = SplitBlock(BB, BI, DTU, /*LI=*/nullptr,
4029 /*MSSAU=*/nullptr, "switch.early.test");
4030
4031 // Remove the uncond branch added to the old block.
4032 Instruction *OldTI = BB->getTerminator();
4033 Builder.SetInsertPoint(OldTI);
4034
4035 if (TrueWhenEqual)
4036 Builder.CreateCondBr(ExtraCase, EdgeBB, NewBB);
4037 else
4038 Builder.CreateCondBr(ExtraCase, NewBB, EdgeBB);
4039
4040 OldTI->eraseFromParent();
4041
4042 Updates.push_back({DominatorTree::Insert, BB, EdgeBB});
4043
4044 // If there are PHI nodes in EdgeBB, then we need to add a new entry to them
4045 // for the edge we just added.
4046 AddPredecessorToBlock(EdgeBB, BB, NewBB);
4047
4048 LLVM_DEBUG(dbgs() << " ** 'icmp' chain unhandled condition: " << *ExtraCase
4049 << "\nEXTRABB = " << *BB);
4050 BB = NewBB;
4051 }
4052
4053 Builder.SetInsertPoint(BI);
4054 // Convert pointer to int before we switch.
4055 if (CompVal->getType()->isPointerTy()) {
4056 CompVal = Builder.CreatePtrToInt(
4057 CompVal, DL.getIntPtrType(CompVal->getType()), "magicptr");
4058 }
4059
4060 // Create the new switch instruction now.
4061 SwitchInst *New = Builder.CreateSwitch(CompVal, DefaultBB, Values.size());
4062
4063 // Add all of the 'cases' to the switch instruction.
4064 for (unsigned i = 0, e = Values.size(); i != e; ++i)
4065 New->addCase(Values[i], EdgeBB);
4066
4067 // We added edges from PI to the EdgeBB. As such, if there were any
4068 // PHI nodes in EdgeBB, they need entries to be added corresponding to
4069 // the number of edges added.
4070 for (BasicBlock::iterator BBI = EdgeBB->begin(); isa<PHINode>(BBI); ++BBI) {
4071 PHINode *PN = cast<PHINode>(BBI);
4072 Value *InVal = PN->getIncomingValueForBlock(BB);
4073 for (unsigned i = 0, e = Values.size() - 1; i != e; ++i)
4074 PN->addIncoming(InVal, BB);
4075 }
4076
4077 // Erase the old branch instruction.
4078 EraseTerminatorAndDCECond(BI);
4079 if (DTU)
4080 DTU->applyUpdates(Updates);
4081
4082 LLVM_DEBUG(dbgs() << " ** 'icmp' chain result is:\n" << *BB << '\n');
4083 return true;
4084 }
4085
simplifyResume(ResumeInst * RI,IRBuilder<> & Builder)4086 bool SimplifyCFGOpt::simplifyResume(ResumeInst *RI, IRBuilder<> &Builder) {
4087 if (isa<PHINode>(RI->getValue()))
4088 return simplifyCommonResume(RI);
4089 else if (isa<LandingPadInst>(RI->getParent()->getFirstNonPHI()) &&
4090 RI->getValue() == RI->getParent()->getFirstNonPHI())
4091 // The resume must unwind the exception that caused control to branch here.
4092 return simplifySingleResume(RI);
4093
4094 return false;
4095 }
4096
4097 // Check if cleanup block is empty
isCleanupBlockEmpty(iterator_range<BasicBlock::iterator> R)4098 static bool isCleanupBlockEmpty(iterator_range<BasicBlock::iterator> R) {
4099 for (Instruction &I : R) {
4100 auto *II = dyn_cast<IntrinsicInst>(&I);
4101 if (!II)
4102 return false;
4103
4104 Intrinsic::ID IntrinsicID = II->getIntrinsicID();
4105 switch (IntrinsicID) {
4106 case Intrinsic::dbg_declare:
4107 case Intrinsic::dbg_value:
4108 case Intrinsic::dbg_label:
4109 case Intrinsic::lifetime_end:
4110 break;
4111 default:
4112 return false;
4113 }
4114 }
4115 return true;
4116 }
4117
4118 // Simplify resume that is shared by several landing pads (phi of landing pad).
simplifyCommonResume(ResumeInst * RI)4119 bool SimplifyCFGOpt::simplifyCommonResume(ResumeInst *RI) {
4120 BasicBlock *BB = RI->getParent();
4121
4122 // Check that there are no other instructions except for debug and lifetime
4123 // intrinsics between the phi's and resume instruction.
4124 if (!isCleanupBlockEmpty(
4125 make_range(RI->getParent()->getFirstNonPHI(), BB->getTerminator())))
4126 return false;
4127
4128 SmallSetVector<BasicBlock *, 4> TrivialUnwindBlocks;
4129 auto *PhiLPInst = cast<PHINode>(RI->getValue());
4130
4131 // Check incoming blocks to see if any of them are trivial.
4132 for (unsigned Idx = 0, End = PhiLPInst->getNumIncomingValues(); Idx != End;
4133 Idx++) {
4134 auto *IncomingBB = PhiLPInst->getIncomingBlock(Idx);
4135 auto *IncomingValue = PhiLPInst->getIncomingValue(Idx);
4136
4137 // If the block has other successors, we can not delete it because
4138 // it has other dependents.
4139 if (IncomingBB->getUniqueSuccessor() != BB)
4140 continue;
4141
4142 auto *LandingPad = dyn_cast<LandingPadInst>(IncomingBB->getFirstNonPHI());
4143 // Not the landing pad that caused the control to branch here.
4144 if (IncomingValue != LandingPad)
4145 continue;
4146
4147 if (isCleanupBlockEmpty(
4148 make_range(LandingPad->getNextNode(), IncomingBB->getTerminator())))
4149 TrivialUnwindBlocks.insert(IncomingBB);
4150 }
4151
4152 // If no trivial unwind blocks, don't do any simplifications.
4153 if (TrivialUnwindBlocks.empty())
4154 return false;
4155
4156 // Turn all invokes that unwind here into calls.
4157 for (auto *TrivialBB : TrivialUnwindBlocks) {
4158 // Blocks that will be simplified should be removed from the phi node.
4159 // Note there could be multiple edges to the resume block, and we need
4160 // to remove them all.
4161 while (PhiLPInst->getBasicBlockIndex(TrivialBB) != -1)
4162 BB->removePredecessor(TrivialBB, true);
4163
4164 for (pred_iterator PI = pred_begin(TrivialBB), PE = pred_end(TrivialBB);
4165 PI != PE;) {
4166 BasicBlock *Pred = *PI++;
4167 removeUnwindEdge(Pred, DTU);
4168 ++NumInvokes;
4169 }
4170
4171 // In each SimplifyCFG run, only the current processed block can be erased.
4172 // Otherwise, it will break the iteration of SimplifyCFG pass. So instead
4173 // of erasing TrivialBB, we only remove the branch to the common resume
4174 // block so that we can later erase the resume block since it has no
4175 // predecessors.
4176 TrivialBB->getTerminator()->eraseFromParent();
4177 new UnreachableInst(RI->getContext(), TrivialBB);
4178 if (DTU)
4179 DTU->applyUpdates({{DominatorTree::Delete, TrivialBB, BB}});
4180 }
4181
4182 // Delete the resume block if all its predecessors have been removed.
4183 if (pred_empty(BB)) {
4184 if (DTU)
4185 DTU->deleteBB(BB);
4186 else
4187 BB->eraseFromParent();
4188 }
4189
4190 return !TrivialUnwindBlocks.empty();
4191 }
4192
4193 // Simplify resume that is only used by a single (non-phi) landing pad.
simplifySingleResume(ResumeInst * RI)4194 bool SimplifyCFGOpt::simplifySingleResume(ResumeInst *RI) {
4195 BasicBlock *BB = RI->getParent();
4196 auto *LPInst = cast<LandingPadInst>(BB->getFirstNonPHI());
4197 assert(RI->getValue() == LPInst &&
4198 "Resume must unwind the exception that caused control to here");
4199
4200 // Check that there are no other instructions except for debug intrinsics.
4201 if (!isCleanupBlockEmpty(
4202 make_range<Instruction *>(LPInst->getNextNode(), RI)))
4203 return false;
4204
4205 // Turn all invokes that unwind here into calls and delete the basic block.
4206 for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE;) {
4207 BasicBlock *Pred = *PI++;
4208 removeUnwindEdge(Pred, DTU);
4209 ++NumInvokes;
4210 }
4211
4212 // The landingpad is now unreachable. Zap it.
4213 if (DTU)
4214 DTU->deleteBB(BB);
4215 else
4216 BB->eraseFromParent();
4217 return true;
4218 }
4219
removeEmptyCleanup(CleanupReturnInst * RI,DomTreeUpdater * DTU)4220 static bool removeEmptyCleanup(CleanupReturnInst *RI, DomTreeUpdater *DTU) {
4221 // If this is a trivial cleanup pad that executes no instructions, it can be
4222 // eliminated. If the cleanup pad continues to the caller, any predecessor
4223 // that is an EH pad will be updated to continue to the caller and any
4224 // predecessor that terminates with an invoke instruction will have its invoke
4225 // instruction converted to a call instruction. If the cleanup pad being
4226 // simplified does not continue to the caller, each predecessor will be
4227 // updated to continue to the unwind destination of the cleanup pad being
4228 // simplified.
4229 BasicBlock *BB = RI->getParent();
4230 CleanupPadInst *CPInst = RI->getCleanupPad();
4231 if (CPInst->getParent() != BB)
4232 // This isn't an empty cleanup.
4233 return false;
4234
4235 // We cannot kill the pad if it has multiple uses. This typically arises
4236 // from unreachable basic blocks.
4237 if (!CPInst->hasOneUse())
4238 return false;
4239
4240 // Check that there are no other instructions except for benign intrinsics.
4241 if (!isCleanupBlockEmpty(
4242 make_range<Instruction *>(CPInst->getNextNode(), RI)))
4243 return false;
4244
4245 // If the cleanup return we are simplifying unwinds to the caller, this will
4246 // set UnwindDest to nullptr.
4247 BasicBlock *UnwindDest = RI->getUnwindDest();
4248 Instruction *DestEHPad = UnwindDest ? UnwindDest->getFirstNonPHI() : nullptr;
4249
4250 // We're about to remove BB from the control flow. Before we do, sink any
4251 // PHINodes into the unwind destination. Doing this before changing the
4252 // control flow avoids some potentially slow checks, since we can currently
4253 // be certain that UnwindDest and BB have no common predecessors (since they
4254 // are both EH pads).
4255 if (UnwindDest) {
4256 // First, go through the PHI nodes in UnwindDest and update any nodes that
4257 // reference the block we are removing
4258 for (BasicBlock::iterator I = UnwindDest->begin(),
4259 IE = DestEHPad->getIterator();
4260 I != IE; ++I) {
4261 PHINode *DestPN = cast<PHINode>(I);
4262
4263 int Idx = DestPN->getBasicBlockIndex(BB);
4264 // Since BB unwinds to UnwindDest, it has to be in the PHI node.
4265 assert(Idx != -1);
4266 // This PHI node has an incoming value that corresponds to a control
4267 // path through the cleanup pad we are removing. If the incoming
4268 // value is in the cleanup pad, it must be a PHINode (because we
4269 // verified above that the block is otherwise empty). Otherwise, the
4270 // value is either a constant or a value that dominates the cleanup
4271 // pad being removed.
4272 //
4273 // Because BB and UnwindDest are both EH pads, all of their
4274 // predecessors must unwind to these blocks, and since no instruction
4275 // can have multiple unwind destinations, there will be no overlap in
4276 // incoming blocks between SrcPN and DestPN.
4277 Value *SrcVal = DestPN->getIncomingValue(Idx);
4278 PHINode *SrcPN = dyn_cast<PHINode>(SrcVal);
4279
4280 // Remove the entry for the block we are deleting.
4281 DestPN->removeIncomingValue(Idx, false);
4282
4283 if (SrcPN && SrcPN->getParent() == BB) {
4284 // If the incoming value was a PHI node in the cleanup pad we are
4285 // removing, we need to merge that PHI node's incoming values into
4286 // DestPN.
4287 for (unsigned SrcIdx = 0, SrcE = SrcPN->getNumIncomingValues();
4288 SrcIdx != SrcE; ++SrcIdx) {
4289 DestPN->addIncoming(SrcPN->getIncomingValue(SrcIdx),
4290 SrcPN->getIncomingBlock(SrcIdx));
4291 }
4292 } else {
4293 // Otherwise, the incoming value came from above BB and
4294 // so we can just reuse it. We must associate all of BB's
4295 // predecessors with this value.
4296 for (auto *pred : predecessors(BB)) {
4297 DestPN->addIncoming(SrcVal, pred);
4298 }
4299 }
4300 }
4301
4302 // Sink any remaining PHI nodes directly into UnwindDest.
4303 Instruction *InsertPt = DestEHPad;
4304 for (BasicBlock::iterator I = BB->begin(),
4305 IE = BB->getFirstNonPHI()->getIterator();
4306 I != IE;) {
4307 // The iterator must be incremented here because the instructions are
4308 // being moved to another block.
4309 PHINode *PN = cast<PHINode>(I++);
4310 if (PN->use_empty() || !PN->isUsedOutsideOfBlock(BB))
4311 // If the PHI node has no uses or all of its uses are in this basic
4312 // block (meaning they are debug or lifetime intrinsics), just leave
4313 // it. It will be erased when we erase BB below.
4314 continue;
4315
4316 // Otherwise, sink this PHI node into UnwindDest.
4317 // Any predecessors to UnwindDest which are not already represented
4318 // must be back edges which inherit the value from the path through
4319 // BB. In this case, the PHI value must reference itself.
4320 for (auto *pred : predecessors(UnwindDest))
4321 if (pred != BB)
4322 PN->addIncoming(PN, pred);
4323 PN->moveBefore(InsertPt);
4324 }
4325 }
4326
4327 std::vector<DominatorTree::UpdateType> Updates;
4328
4329 for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE;) {
4330 // The iterator must be updated here because we are removing this pred.
4331 BasicBlock *PredBB = *PI++;
4332 if (UnwindDest == nullptr) {
4333 if (DTU)
4334 DTU->applyUpdates(Updates);
4335 Updates.clear();
4336 removeUnwindEdge(PredBB, DTU);
4337 ++NumInvokes;
4338 } else {
4339 Instruction *TI = PredBB->getTerminator();
4340 TI->replaceUsesOfWith(BB, UnwindDest);
4341 Updates.push_back({DominatorTree::Insert, PredBB, UnwindDest});
4342 Updates.push_back({DominatorTree::Delete, PredBB, BB});
4343 }
4344 }
4345
4346 if (DTU) {
4347 DTU->applyUpdates(Updates);
4348 DTU->deleteBB(BB);
4349 } else
4350 // The cleanup pad is now unreachable. Zap it.
4351 BB->eraseFromParent();
4352
4353 return true;
4354 }
4355
4356 // Try to merge two cleanuppads together.
mergeCleanupPad(CleanupReturnInst * RI)4357 static bool mergeCleanupPad(CleanupReturnInst *RI) {
4358 // Skip any cleanuprets which unwind to caller, there is nothing to merge
4359 // with.
4360 BasicBlock *UnwindDest = RI->getUnwindDest();
4361 if (!UnwindDest)
4362 return false;
4363
4364 // This cleanupret isn't the only predecessor of this cleanuppad, it wouldn't
4365 // be safe to merge without code duplication.
4366 if (UnwindDest->getSinglePredecessor() != RI->getParent())
4367 return false;
4368
4369 // Verify that our cleanuppad's unwind destination is another cleanuppad.
4370 auto *SuccessorCleanupPad = dyn_cast<CleanupPadInst>(&UnwindDest->front());
4371 if (!SuccessorCleanupPad)
4372 return false;
4373
4374 CleanupPadInst *PredecessorCleanupPad = RI->getCleanupPad();
4375 // Replace any uses of the successor cleanupad with the predecessor pad
4376 // The only cleanuppad uses should be this cleanupret, it's cleanupret and
4377 // funclet bundle operands.
4378 SuccessorCleanupPad->replaceAllUsesWith(PredecessorCleanupPad);
4379 // Remove the old cleanuppad.
4380 SuccessorCleanupPad->eraseFromParent();
4381 // Now, we simply replace the cleanupret with a branch to the unwind
4382 // destination.
4383 BranchInst::Create(UnwindDest, RI->getParent());
4384 RI->eraseFromParent();
4385
4386 return true;
4387 }
4388
simplifyCleanupReturn(CleanupReturnInst * RI)4389 bool SimplifyCFGOpt::simplifyCleanupReturn(CleanupReturnInst *RI) {
4390 // It is possible to transiantly have an undef cleanuppad operand because we
4391 // have deleted some, but not all, dead blocks.
4392 // Eventually, this block will be deleted.
4393 if (isa<UndefValue>(RI->getOperand(0)))
4394 return false;
4395
4396 if (mergeCleanupPad(RI))
4397 return true;
4398
4399 if (removeEmptyCleanup(RI, DTU))
4400 return true;
4401
4402 return false;
4403 }
4404
simplifyReturn(ReturnInst * RI,IRBuilder<> & Builder)4405 bool SimplifyCFGOpt::simplifyReturn(ReturnInst *RI, IRBuilder<> &Builder) {
4406 BasicBlock *BB = RI->getParent();
4407 if (!BB->getFirstNonPHIOrDbg()->isTerminator())
4408 return false;
4409
4410 // Find predecessors that end with branches.
4411 SmallVector<BasicBlock *, 8> UncondBranchPreds;
4412 SmallVector<BranchInst *, 8> CondBranchPreds;
4413 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
4414 BasicBlock *P = *PI;
4415 Instruction *PTI = P->getTerminator();
4416 if (BranchInst *BI = dyn_cast<BranchInst>(PTI)) {
4417 if (BI->isUnconditional())
4418 UncondBranchPreds.push_back(P);
4419 else
4420 CondBranchPreds.push_back(BI);
4421 }
4422 }
4423
4424 // If we found some, do the transformation!
4425 if (!UncondBranchPreds.empty() && DupRet) {
4426 while (!UncondBranchPreds.empty()) {
4427 BasicBlock *Pred = UncondBranchPreds.pop_back_val();
4428 LLVM_DEBUG(dbgs() << "FOLDING: " << *BB
4429 << "INTO UNCOND BRANCH PRED: " << *Pred);
4430 (void)FoldReturnIntoUncondBranch(RI, BB, Pred, DTU);
4431 }
4432
4433 // If we eliminated all predecessors of the block, delete the block now.
4434 if (pred_empty(BB)) {
4435 // We know there are no successors, so just nuke the block.
4436 if (DTU)
4437 DTU->deleteBB(BB);
4438 else
4439 BB->eraseFromParent();
4440 }
4441
4442 return true;
4443 }
4444
4445 // Check out all of the conditional branches going to this return
4446 // instruction. If any of them just select between returns, change the
4447 // branch itself into a select/return pair.
4448 while (!CondBranchPreds.empty()) {
4449 BranchInst *BI = CondBranchPreds.pop_back_val();
4450
4451 // Check to see if the non-BB successor is also a return block.
4452 if (isa<ReturnInst>(BI->getSuccessor(0)->getTerminator()) &&
4453 isa<ReturnInst>(BI->getSuccessor(1)->getTerminator()) &&
4454 SimplifyCondBranchToTwoReturns(BI, Builder))
4455 return true;
4456 }
4457 return false;
4458 }
4459
simplifyUnreachable(UnreachableInst * UI)4460 bool SimplifyCFGOpt::simplifyUnreachable(UnreachableInst *UI) {
4461 BasicBlock *BB = UI->getParent();
4462
4463 bool Changed = false;
4464
4465 // If there are any instructions immediately before the unreachable that can
4466 // be removed, do so.
4467 while (UI->getIterator() != BB->begin()) {
4468 BasicBlock::iterator BBI = UI->getIterator();
4469 --BBI;
4470 // Do not delete instructions that can have side effects which might cause
4471 // the unreachable to not be reachable; specifically, calls and volatile
4472 // operations may have this effect.
4473 if (isa<CallInst>(BBI) && !isa<DbgInfoIntrinsic>(BBI))
4474 break;
4475
4476 if (BBI->mayHaveSideEffects()) {
4477 if (auto *SI = dyn_cast<StoreInst>(BBI)) {
4478 if (SI->isVolatile())
4479 break;
4480 } else if (auto *LI = dyn_cast<LoadInst>(BBI)) {
4481 if (LI->isVolatile())
4482 break;
4483 } else if (auto *RMWI = dyn_cast<AtomicRMWInst>(BBI)) {
4484 if (RMWI->isVolatile())
4485 break;
4486 } else if (auto *CXI = dyn_cast<AtomicCmpXchgInst>(BBI)) {
4487 if (CXI->isVolatile())
4488 break;
4489 } else if (isa<CatchPadInst>(BBI)) {
4490 // A catchpad may invoke exception object constructors and such, which
4491 // in some languages can be arbitrary code, so be conservative by
4492 // default.
4493 // For CoreCLR, it just involves a type test, so can be removed.
4494 if (classifyEHPersonality(BB->getParent()->getPersonalityFn()) !=
4495 EHPersonality::CoreCLR)
4496 break;
4497 } else if (!isa<FenceInst>(BBI) && !isa<VAArgInst>(BBI) &&
4498 !isa<LandingPadInst>(BBI)) {
4499 break;
4500 }
4501 // Note that deleting LandingPad's here is in fact okay, although it
4502 // involves a bit of subtle reasoning. If this inst is a LandingPad,
4503 // all the predecessors of this block will be the unwind edges of Invokes,
4504 // and we can therefore guarantee this block will be erased.
4505 }
4506
4507 // Delete this instruction (any uses are guaranteed to be dead)
4508 if (!BBI->use_empty())
4509 BBI->replaceAllUsesWith(UndefValue::get(BBI->getType()));
4510 BBI->eraseFromParent();
4511 Changed = true;
4512 }
4513
4514 // If the unreachable instruction is the first in the block, take a gander
4515 // at all of the predecessors of this instruction, and simplify them.
4516 if (&BB->front() != UI)
4517 return Changed;
4518
4519 std::vector<DominatorTree::UpdateType> Updates;
4520
4521 SmallSetVector<BasicBlock *, 8> Preds(pred_begin(BB), pred_end(BB));
4522 for (unsigned i = 0, e = Preds.size(); i != e; ++i) {
4523 auto *Predecessor = Preds[i];
4524 Instruction *TI = Predecessor->getTerminator();
4525 IRBuilder<> Builder(TI);
4526 if (auto *BI = dyn_cast<BranchInst>(TI)) {
4527 // We could either have a proper unconditional branch,
4528 // or a degenerate conditional branch with matching destinations.
4529 if (all_of(BI->successors(),
4530 [BB](auto *Successor) { return Successor == BB; })) {
4531 new UnreachableInst(TI->getContext(), TI);
4532 TI->eraseFromParent();
4533 Changed = true;
4534 } else {
4535 assert(BI->isConditional() && "Can't get here with an uncond branch.");
4536 Value* Cond = BI->getCondition();
4537 assert(BI->getSuccessor(0) != BI->getSuccessor(1) &&
4538 "The destinations are guaranteed to be different here.");
4539 if (BI->getSuccessor(0) == BB) {
4540 Builder.CreateAssumption(Builder.CreateNot(Cond));
4541 Builder.CreateBr(BI->getSuccessor(1));
4542 } else {
4543 assert(BI->getSuccessor(1) == BB && "Incorrect CFG");
4544 Builder.CreateAssumption(Cond);
4545 Builder.CreateBr(BI->getSuccessor(0));
4546 }
4547 EraseTerminatorAndDCECond(BI);
4548 Changed = true;
4549 }
4550 Updates.push_back({DominatorTree::Delete, Predecessor, BB});
4551 } else if (auto *SI = dyn_cast<SwitchInst>(TI)) {
4552 SwitchInstProfUpdateWrapper SU(*SI);
4553 for (auto i = SU->case_begin(), e = SU->case_end(); i != e;) {
4554 if (i->getCaseSuccessor() != BB) {
4555 ++i;
4556 continue;
4557 }
4558 BB->removePredecessor(SU->getParent());
4559 i = SU.removeCase(i);
4560 e = SU->case_end();
4561 Changed = true;
4562 }
4563 // Note that the default destination can't be removed!
4564 if (SI->getDefaultDest() != BB)
4565 Updates.push_back({DominatorTree::Delete, Predecessor, BB});
4566 } else if (auto *II = dyn_cast<InvokeInst>(TI)) {
4567 if (II->getUnwindDest() == BB) {
4568 if (DTU)
4569 DTU->applyUpdates(Updates);
4570 Updates.clear();
4571 removeUnwindEdge(TI->getParent(), DTU);
4572 Changed = true;
4573 }
4574 } else if (auto *CSI = dyn_cast<CatchSwitchInst>(TI)) {
4575 if (CSI->getUnwindDest() == BB) {
4576 if (DTU)
4577 DTU->applyUpdates(Updates);
4578 Updates.clear();
4579 removeUnwindEdge(TI->getParent(), DTU);
4580 Changed = true;
4581 continue;
4582 }
4583
4584 for (CatchSwitchInst::handler_iterator I = CSI->handler_begin(),
4585 E = CSI->handler_end();
4586 I != E; ++I) {
4587 if (*I == BB) {
4588 CSI->removeHandler(I);
4589 --I;
4590 --E;
4591 Changed = true;
4592 }
4593 }
4594 Updates.push_back({DominatorTree::Delete, Predecessor, BB});
4595 if (CSI->getNumHandlers() == 0) {
4596 if (CSI->hasUnwindDest()) {
4597 // Redirect all predecessors of the block containing CatchSwitchInst
4598 // to instead branch to the CatchSwitchInst's unwind destination.
4599 for (auto *PredecessorOfPredecessor : predecessors(Predecessor)) {
4600 Updates.push_back({DominatorTree::Insert, PredecessorOfPredecessor,
4601 CSI->getUnwindDest()});
4602 Updates.push_back(
4603 {DominatorTree::Delete, PredecessorOfPredecessor, Predecessor});
4604 }
4605 Predecessor->replaceAllUsesWith(CSI->getUnwindDest());
4606 } else {
4607 // Rewrite all preds to unwind to caller (or from invoke to call).
4608 if (DTU)
4609 DTU->applyUpdates(Updates);
4610 Updates.clear();
4611 SmallVector<BasicBlock *, 8> EHPreds(predecessors(Predecessor));
4612 for (BasicBlock *EHPred : EHPreds)
4613 removeUnwindEdge(EHPred, DTU);
4614 }
4615 // The catchswitch is no longer reachable.
4616 new UnreachableInst(CSI->getContext(), CSI);
4617 CSI->eraseFromParent();
4618 Changed = true;
4619 }
4620 } else if (auto *CRI = dyn_cast<CleanupReturnInst>(TI)) {
4621 (void)CRI;
4622 assert(CRI->hasUnwindDest() && CRI->getUnwindDest() == BB &&
4623 "Expected to always have an unwind to BB.");
4624 Updates.push_back({DominatorTree::Delete, Predecessor, BB});
4625 new UnreachableInst(TI->getContext(), TI);
4626 TI->eraseFromParent();
4627 Changed = true;
4628 }
4629 }
4630
4631 if (DTU)
4632 DTU->applyUpdates(Updates);
4633
4634 // If this block is now dead, remove it.
4635 if (pred_empty(BB) && BB != &BB->getParent()->getEntryBlock()) {
4636 // We know there are no successors, so just nuke the block.
4637 if (DTU)
4638 DTU->deleteBB(BB);
4639 else
4640 BB->eraseFromParent();
4641 return true;
4642 }
4643
4644 return Changed;
4645 }
4646
CasesAreContiguous(SmallVectorImpl<ConstantInt * > & Cases)4647 static bool CasesAreContiguous(SmallVectorImpl<ConstantInt *> &Cases) {
4648 assert(Cases.size() >= 1);
4649
4650 array_pod_sort(Cases.begin(), Cases.end(), ConstantIntSortPredicate);
4651 for (size_t I = 1, E = Cases.size(); I != E; ++I) {
4652 if (Cases[I - 1]->getValue() != Cases[I]->getValue() + 1)
4653 return false;
4654 }
4655 return true;
4656 }
4657
createUnreachableSwitchDefault(SwitchInst * Switch,DomTreeUpdater * DTU)4658 static void createUnreachableSwitchDefault(SwitchInst *Switch,
4659 DomTreeUpdater *DTU) {
4660 LLVM_DEBUG(dbgs() << "SimplifyCFG: switch default is dead.\n");
4661 auto *BB = Switch->getParent();
4662 BasicBlock *NewDefaultBlock = SplitBlockPredecessors(
4663 Switch->getDefaultDest(), Switch->getParent(), "", DTU);
4664 auto *OrigDefaultBlock = Switch->getDefaultDest();
4665 Switch->setDefaultDest(&*NewDefaultBlock);
4666 if (DTU)
4667 DTU->applyUpdates({{DominatorTree::Insert, BB, &*NewDefaultBlock},
4668 {DominatorTree::Delete, BB, OrigDefaultBlock}});
4669 SplitBlock(&*NewDefaultBlock, &NewDefaultBlock->front(), DTU);
4670 SmallVector<DominatorTree::UpdateType, 2> Updates;
4671 for (auto *Successor : successors(NewDefaultBlock))
4672 Updates.push_back({DominatorTree::Delete, NewDefaultBlock, Successor});
4673 auto *NewTerminator = NewDefaultBlock->getTerminator();
4674 new UnreachableInst(Switch->getContext(), NewTerminator);
4675 EraseTerminatorAndDCECond(NewTerminator);
4676 if (DTU)
4677 DTU->applyUpdates(Updates);
4678 }
4679
4680 /// Turn a switch with two reachable destinations into an integer range
4681 /// comparison and branch.
TurnSwitchRangeIntoICmp(SwitchInst * SI,IRBuilder<> & Builder)4682 bool SimplifyCFGOpt::TurnSwitchRangeIntoICmp(SwitchInst *SI,
4683 IRBuilder<> &Builder) {
4684 assert(SI->getNumCases() > 1 && "Degenerate switch?");
4685
4686 bool HasDefault =
4687 !isa<UnreachableInst>(SI->getDefaultDest()->getFirstNonPHIOrDbg());
4688
4689 auto *BB = SI->getParent();
4690
4691 // Partition the cases into two sets with different destinations.
4692 BasicBlock *DestA = HasDefault ? SI->getDefaultDest() : nullptr;
4693 BasicBlock *DestB = nullptr;
4694 SmallVector<ConstantInt *, 16> CasesA;
4695 SmallVector<ConstantInt *, 16> CasesB;
4696
4697 for (auto Case : SI->cases()) {
4698 BasicBlock *Dest = Case.getCaseSuccessor();
4699 if (!DestA)
4700 DestA = Dest;
4701 if (Dest == DestA) {
4702 CasesA.push_back(Case.getCaseValue());
4703 continue;
4704 }
4705 if (!DestB)
4706 DestB = Dest;
4707 if (Dest == DestB) {
4708 CasesB.push_back(Case.getCaseValue());
4709 continue;
4710 }
4711 return false; // More than two destinations.
4712 }
4713
4714 assert(DestA && DestB &&
4715 "Single-destination switch should have been folded.");
4716 assert(DestA != DestB);
4717 assert(DestB != SI->getDefaultDest());
4718 assert(!CasesB.empty() && "There must be non-default cases.");
4719 assert(!CasesA.empty() || HasDefault);
4720
4721 // Figure out if one of the sets of cases form a contiguous range.
4722 SmallVectorImpl<ConstantInt *> *ContiguousCases = nullptr;
4723 BasicBlock *ContiguousDest = nullptr;
4724 BasicBlock *OtherDest = nullptr;
4725 if (!CasesA.empty() && CasesAreContiguous(CasesA)) {
4726 ContiguousCases = &CasesA;
4727 ContiguousDest = DestA;
4728 OtherDest = DestB;
4729 } else if (CasesAreContiguous(CasesB)) {
4730 ContiguousCases = &CasesB;
4731 ContiguousDest = DestB;
4732 OtherDest = DestA;
4733 } else
4734 return false;
4735
4736 // Start building the compare and branch.
4737
4738 Constant *Offset = ConstantExpr::getNeg(ContiguousCases->back());
4739 Constant *NumCases =
4740 ConstantInt::get(Offset->getType(), ContiguousCases->size());
4741
4742 Value *Sub = SI->getCondition();
4743 if (!Offset->isNullValue())
4744 Sub = Builder.CreateAdd(Sub, Offset, Sub->getName() + ".off");
4745
4746 Value *Cmp;
4747 // If NumCases overflowed, then all possible values jump to the successor.
4748 if (NumCases->isNullValue() && !ContiguousCases->empty())
4749 Cmp = ConstantInt::getTrue(SI->getContext());
4750 else
4751 Cmp = Builder.CreateICmpULT(Sub, NumCases, "switch");
4752 BranchInst *NewBI = Builder.CreateCondBr(Cmp, ContiguousDest, OtherDest);
4753
4754 // Update weight for the newly-created conditional branch.
4755 if (HasBranchWeights(SI)) {
4756 SmallVector<uint64_t, 8> Weights;
4757 GetBranchWeights(SI, Weights);
4758 if (Weights.size() == 1 + SI->getNumCases()) {
4759 uint64_t TrueWeight = 0;
4760 uint64_t FalseWeight = 0;
4761 for (size_t I = 0, E = Weights.size(); I != E; ++I) {
4762 if (SI->getSuccessor(I) == ContiguousDest)
4763 TrueWeight += Weights[I];
4764 else
4765 FalseWeight += Weights[I];
4766 }
4767 while (TrueWeight > UINT32_MAX || FalseWeight > UINT32_MAX) {
4768 TrueWeight /= 2;
4769 FalseWeight /= 2;
4770 }
4771 setBranchWeights(NewBI, TrueWeight, FalseWeight);
4772 }
4773 }
4774
4775 // Prune obsolete incoming values off the successors' PHI nodes.
4776 for (auto BBI = ContiguousDest->begin(); isa<PHINode>(BBI); ++BBI) {
4777 unsigned PreviousEdges = ContiguousCases->size();
4778 if (ContiguousDest == SI->getDefaultDest())
4779 ++PreviousEdges;
4780 for (unsigned I = 0, E = PreviousEdges - 1; I != E; ++I)
4781 cast<PHINode>(BBI)->removeIncomingValue(SI->getParent());
4782 }
4783 for (auto BBI = OtherDest->begin(); isa<PHINode>(BBI); ++BBI) {
4784 unsigned PreviousEdges = SI->getNumCases() - ContiguousCases->size();
4785 if (OtherDest == SI->getDefaultDest())
4786 ++PreviousEdges;
4787 for (unsigned I = 0, E = PreviousEdges - 1; I != E; ++I)
4788 cast<PHINode>(BBI)->removeIncomingValue(SI->getParent());
4789 }
4790
4791 // Clean up the default block - it may have phis or other instructions before
4792 // the unreachable terminator.
4793 if (!HasDefault)
4794 createUnreachableSwitchDefault(SI, DTU);
4795
4796 auto *UnreachableDefault = SI->getDefaultDest();
4797
4798 // Drop the switch.
4799 SI->eraseFromParent();
4800
4801 if (!HasDefault && DTU)
4802 DTU->applyUpdates({{DominatorTree::Delete, BB, UnreachableDefault}});
4803
4804 return true;
4805 }
4806
4807 /// Compute masked bits for the condition of a switch
4808 /// and use it to remove dead cases.
eliminateDeadSwitchCases(SwitchInst * SI,DomTreeUpdater * DTU,AssumptionCache * AC,const DataLayout & DL)4809 static bool eliminateDeadSwitchCases(SwitchInst *SI, DomTreeUpdater *DTU,
4810 AssumptionCache *AC,
4811 const DataLayout &DL) {
4812 Value *Cond = SI->getCondition();
4813 unsigned Bits = Cond->getType()->getIntegerBitWidth();
4814 KnownBits Known = computeKnownBits(Cond, DL, 0, AC, SI);
4815
4816 // We can also eliminate cases by determining that their values are outside of
4817 // the limited range of the condition based on how many significant (non-sign)
4818 // bits are in the condition value.
4819 unsigned ExtraSignBits = ComputeNumSignBits(Cond, DL, 0, AC, SI) - 1;
4820 unsigned MaxSignificantBitsInCond = Bits - ExtraSignBits;
4821
4822 // Gather dead cases.
4823 SmallVector<ConstantInt *, 8> DeadCases;
4824 SmallMapVector<BasicBlock *, int, 8> NumPerSuccessorCases;
4825 for (auto &Case : SI->cases()) {
4826 auto *Successor = Case.getCaseSuccessor();
4827 ++NumPerSuccessorCases[Successor];
4828 const APInt &CaseVal = Case.getCaseValue()->getValue();
4829 if (Known.Zero.intersects(CaseVal) || !Known.One.isSubsetOf(CaseVal) ||
4830 (CaseVal.getMinSignedBits() > MaxSignificantBitsInCond)) {
4831 DeadCases.push_back(Case.getCaseValue());
4832 --NumPerSuccessorCases[Successor];
4833 LLVM_DEBUG(dbgs() << "SimplifyCFG: switch case " << CaseVal
4834 << " is dead.\n");
4835 }
4836 }
4837
4838 // If we can prove that the cases must cover all possible values, the
4839 // default destination becomes dead and we can remove it. If we know some
4840 // of the bits in the value, we can use that to more precisely compute the
4841 // number of possible unique case values.
4842 bool HasDefault =
4843 !isa<UnreachableInst>(SI->getDefaultDest()->getFirstNonPHIOrDbg());
4844 const unsigned NumUnknownBits =
4845 Bits - (Known.Zero | Known.One).countPopulation();
4846 assert(NumUnknownBits <= Bits);
4847 if (HasDefault && DeadCases.empty() &&
4848 NumUnknownBits < 64 /* avoid overflow */ &&
4849 SI->getNumCases() == (1ULL << NumUnknownBits)) {
4850 createUnreachableSwitchDefault(SI, DTU);
4851 return true;
4852 }
4853
4854 if (DeadCases.empty())
4855 return false;
4856
4857 SwitchInstProfUpdateWrapper SIW(*SI);
4858 for (ConstantInt *DeadCase : DeadCases) {
4859 SwitchInst::CaseIt CaseI = SI->findCaseValue(DeadCase);
4860 assert(CaseI != SI->case_default() &&
4861 "Case was not found. Probably mistake in DeadCases forming.");
4862 // Prune unused values from PHI nodes.
4863 CaseI->getCaseSuccessor()->removePredecessor(SI->getParent());
4864 SIW.removeCase(CaseI);
4865 }
4866
4867 std::vector<DominatorTree::UpdateType> Updates;
4868 for (const std::pair<BasicBlock *, int> &I : NumPerSuccessorCases)
4869 if (I.second == 0)
4870 Updates.push_back({DominatorTree::Delete, SI->getParent(), I.first});
4871 if (DTU)
4872 DTU->applyUpdates(Updates);
4873
4874 return true;
4875 }
4876
4877 /// If BB would be eligible for simplification by
4878 /// TryToSimplifyUncondBranchFromEmptyBlock (i.e. it is empty and terminated
4879 /// by an unconditional branch), look at the phi node for BB in the successor
4880 /// block and see if the incoming value is equal to CaseValue. If so, return
4881 /// the phi node, and set PhiIndex to BB's index in the phi node.
FindPHIForConditionForwarding(ConstantInt * CaseValue,BasicBlock * BB,int * PhiIndex)4882 static PHINode *FindPHIForConditionForwarding(ConstantInt *CaseValue,
4883 BasicBlock *BB, int *PhiIndex) {
4884 if (BB->getFirstNonPHIOrDbg() != BB->getTerminator())
4885 return nullptr; // BB must be empty to be a candidate for simplification.
4886 if (!BB->getSinglePredecessor())
4887 return nullptr; // BB must be dominated by the switch.
4888
4889 BranchInst *Branch = dyn_cast<BranchInst>(BB->getTerminator());
4890 if (!Branch || !Branch->isUnconditional())
4891 return nullptr; // Terminator must be unconditional branch.
4892
4893 BasicBlock *Succ = Branch->getSuccessor(0);
4894
4895 for (PHINode &PHI : Succ->phis()) {
4896 int Idx = PHI.getBasicBlockIndex(BB);
4897 assert(Idx >= 0 && "PHI has no entry for predecessor?");
4898
4899 Value *InValue = PHI.getIncomingValue(Idx);
4900 if (InValue != CaseValue)
4901 continue;
4902
4903 *PhiIndex = Idx;
4904 return &PHI;
4905 }
4906
4907 return nullptr;
4908 }
4909
4910 /// Try to forward the condition of a switch instruction to a phi node
4911 /// dominated by the switch, if that would mean that some of the destination
4912 /// blocks of the switch can be folded away. Return true if a change is made.
ForwardSwitchConditionToPHI(SwitchInst * SI)4913 static bool ForwardSwitchConditionToPHI(SwitchInst *SI) {
4914 using ForwardingNodesMap = DenseMap<PHINode *, SmallVector<int, 4>>;
4915
4916 ForwardingNodesMap ForwardingNodes;
4917 BasicBlock *SwitchBlock = SI->getParent();
4918 bool Changed = false;
4919 for (auto &Case : SI->cases()) {
4920 ConstantInt *CaseValue = Case.getCaseValue();
4921 BasicBlock *CaseDest = Case.getCaseSuccessor();
4922
4923 // Replace phi operands in successor blocks that are using the constant case
4924 // value rather than the switch condition variable:
4925 // switchbb:
4926 // switch i32 %x, label %default [
4927 // i32 17, label %succ
4928 // ...
4929 // succ:
4930 // %r = phi i32 ... [ 17, %switchbb ] ...
4931 // -->
4932 // %r = phi i32 ... [ %x, %switchbb ] ...
4933
4934 for (PHINode &Phi : CaseDest->phis()) {
4935 // This only works if there is exactly 1 incoming edge from the switch to
4936 // a phi. If there is >1, that means multiple cases of the switch map to 1
4937 // value in the phi, and that phi value is not the switch condition. Thus,
4938 // this transform would not make sense (the phi would be invalid because
4939 // a phi can't have different incoming values from the same block).
4940 int SwitchBBIdx = Phi.getBasicBlockIndex(SwitchBlock);
4941 if (Phi.getIncomingValue(SwitchBBIdx) == CaseValue &&
4942 count(Phi.blocks(), SwitchBlock) == 1) {
4943 Phi.setIncomingValue(SwitchBBIdx, SI->getCondition());
4944 Changed = true;
4945 }
4946 }
4947
4948 // Collect phi nodes that are indirectly using this switch's case constants.
4949 int PhiIdx;
4950 if (auto *Phi = FindPHIForConditionForwarding(CaseValue, CaseDest, &PhiIdx))
4951 ForwardingNodes[Phi].push_back(PhiIdx);
4952 }
4953
4954 for (auto &ForwardingNode : ForwardingNodes) {
4955 PHINode *Phi = ForwardingNode.first;
4956 SmallVectorImpl<int> &Indexes = ForwardingNode.second;
4957 if (Indexes.size() < 2)
4958 continue;
4959
4960 for (int Index : Indexes)
4961 Phi->setIncomingValue(Index, SI->getCondition());
4962 Changed = true;
4963 }
4964
4965 return Changed;
4966 }
4967
4968 /// Return true if the backend will be able to handle
4969 /// initializing an array of constants like C.
ValidLookupTableConstant(Constant * C,const TargetTransformInfo & TTI)4970 static bool ValidLookupTableConstant(Constant *C, const TargetTransformInfo &TTI) {
4971 if (C->isThreadDependent())
4972 return false;
4973 if (C->isDLLImportDependent())
4974 return false;
4975
4976 if (!isa<ConstantFP>(C) && !isa<ConstantInt>(C) &&
4977 !isa<ConstantPointerNull>(C) && !isa<GlobalValue>(C) &&
4978 !isa<UndefValue>(C) && !isa<ConstantExpr>(C))
4979 return false;
4980
4981 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) {
4982 if (!CE->isGEPWithNoNotionalOverIndexing())
4983 return false;
4984 if (!ValidLookupTableConstant(CE->getOperand(0), TTI))
4985 return false;
4986 }
4987
4988 if (!TTI.shouldBuildLookupTablesForConstant(C))
4989 return false;
4990
4991 return true;
4992 }
4993
4994 /// If V is a Constant, return it. Otherwise, try to look up
4995 /// its constant value in ConstantPool, returning 0 if it's not there.
4996 static Constant *
LookupConstant(Value * V,const SmallDenseMap<Value *,Constant * > & ConstantPool)4997 LookupConstant(Value *V,
4998 const SmallDenseMap<Value *, Constant *> &ConstantPool) {
4999 if (Constant *C = dyn_cast<Constant>(V))
5000 return C;
5001 return ConstantPool.lookup(V);
5002 }
5003
5004 /// Try to fold instruction I into a constant. This works for
5005 /// simple instructions such as binary operations where both operands are
5006 /// constant or can be replaced by constants from the ConstantPool. Returns the
5007 /// resulting constant on success, 0 otherwise.
5008 static Constant *
ConstantFold(Instruction * I,const DataLayout & DL,const SmallDenseMap<Value *,Constant * > & ConstantPool)5009 ConstantFold(Instruction *I, const DataLayout &DL,
5010 const SmallDenseMap<Value *, Constant *> &ConstantPool) {
5011 if (SelectInst *Select = dyn_cast<SelectInst>(I)) {
5012 Constant *A = LookupConstant(Select->getCondition(), ConstantPool);
5013 if (!A)
5014 return nullptr;
5015 if (A->isAllOnesValue())
5016 return LookupConstant(Select->getTrueValue(), ConstantPool);
5017 if (A->isNullValue())
5018 return LookupConstant(Select->getFalseValue(), ConstantPool);
5019 return nullptr;
5020 }
5021
5022 SmallVector<Constant *, 4> COps;
5023 for (unsigned N = 0, E = I->getNumOperands(); N != E; ++N) {
5024 if (Constant *A = LookupConstant(I->getOperand(N), ConstantPool))
5025 COps.push_back(A);
5026 else
5027 return nullptr;
5028 }
5029
5030 if (CmpInst *Cmp = dyn_cast<CmpInst>(I)) {
5031 return ConstantFoldCompareInstOperands(Cmp->getPredicate(), COps[0],
5032 COps[1], DL);
5033 }
5034
5035 return ConstantFoldInstOperands(I, COps, DL);
5036 }
5037
5038 /// Try to determine the resulting constant values in phi nodes
5039 /// at the common destination basic block, *CommonDest, for one of the case
5040 /// destionations CaseDest corresponding to value CaseVal (0 for the default
5041 /// case), of a switch instruction SI.
5042 static bool
GetCaseResults(SwitchInst * SI,ConstantInt * CaseVal,BasicBlock * CaseDest,BasicBlock ** CommonDest,SmallVectorImpl<std::pair<PHINode *,Constant * >> & Res,const DataLayout & DL,const TargetTransformInfo & TTI)5043 GetCaseResults(SwitchInst *SI, ConstantInt *CaseVal, BasicBlock *CaseDest,
5044 BasicBlock **CommonDest,
5045 SmallVectorImpl<std::pair<PHINode *, Constant *>> &Res,
5046 const DataLayout &DL, const TargetTransformInfo &TTI) {
5047 // The block from which we enter the common destination.
5048 BasicBlock *Pred = SI->getParent();
5049
5050 // If CaseDest is empty except for some side-effect free instructions through
5051 // which we can constant-propagate the CaseVal, continue to its successor.
5052 SmallDenseMap<Value *, Constant *> ConstantPool;
5053 ConstantPool.insert(std::make_pair(SI->getCondition(), CaseVal));
5054 for (Instruction &I :CaseDest->instructionsWithoutDebug()) {
5055 if (I.isTerminator()) {
5056 // If the terminator is a simple branch, continue to the next block.
5057 if (I.getNumSuccessors() != 1 || I.isExceptionalTerminator())
5058 return false;
5059 Pred = CaseDest;
5060 CaseDest = I.getSuccessor(0);
5061 } else if (Constant *C = ConstantFold(&I, DL, ConstantPool)) {
5062 // Instruction is side-effect free and constant.
5063
5064 // If the instruction has uses outside this block or a phi node slot for
5065 // the block, it is not safe to bypass the instruction since it would then
5066 // no longer dominate all its uses.
5067 for (auto &Use : I.uses()) {
5068 User *User = Use.getUser();
5069 if (Instruction *I = dyn_cast<Instruction>(User))
5070 if (I->getParent() == CaseDest)
5071 continue;
5072 if (PHINode *Phi = dyn_cast<PHINode>(User))
5073 if (Phi->getIncomingBlock(Use) == CaseDest)
5074 continue;
5075 return false;
5076 }
5077
5078 ConstantPool.insert(std::make_pair(&I, C));
5079 } else {
5080 break;
5081 }
5082 }
5083
5084 // If we did not have a CommonDest before, use the current one.
5085 if (!*CommonDest)
5086 *CommonDest = CaseDest;
5087 // If the destination isn't the common one, abort.
5088 if (CaseDest != *CommonDest)
5089 return false;
5090
5091 // Get the values for this case from phi nodes in the destination block.
5092 for (PHINode &PHI : (*CommonDest)->phis()) {
5093 int Idx = PHI.getBasicBlockIndex(Pred);
5094 if (Idx == -1)
5095 continue;
5096
5097 Constant *ConstVal =
5098 LookupConstant(PHI.getIncomingValue(Idx), ConstantPool);
5099 if (!ConstVal)
5100 return false;
5101
5102 // Be conservative about which kinds of constants we support.
5103 if (!ValidLookupTableConstant(ConstVal, TTI))
5104 return false;
5105
5106 Res.push_back(std::make_pair(&PHI, ConstVal));
5107 }
5108
5109 return Res.size() > 0;
5110 }
5111
5112 // Helper function used to add CaseVal to the list of cases that generate
5113 // Result. Returns the updated number of cases that generate this result.
MapCaseToResult(ConstantInt * CaseVal,SwitchCaseResultVectorTy & UniqueResults,Constant * Result)5114 static uintptr_t MapCaseToResult(ConstantInt *CaseVal,
5115 SwitchCaseResultVectorTy &UniqueResults,
5116 Constant *Result) {
5117 for (auto &I : UniqueResults) {
5118 if (I.first == Result) {
5119 I.second.push_back(CaseVal);
5120 return I.second.size();
5121 }
5122 }
5123 UniqueResults.push_back(
5124 std::make_pair(Result, SmallVector<ConstantInt *, 4>(1, CaseVal)));
5125 return 1;
5126 }
5127
5128 // Helper function that initializes a map containing
5129 // results for the PHI node of the common destination block for a switch
5130 // instruction. Returns false if multiple PHI nodes have been found or if
5131 // there is not a common destination block for the switch.
5132 static bool
InitializeUniqueCases(SwitchInst * SI,PHINode * & PHI,BasicBlock * & CommonDest,SwitchCaseResultVectorTy & UniqueResults,Constant * & DefaultResult,const DataLayout & DL,const TargetTransformInfo & TTI,uintptr_t MaxUniqueResults,uintptr_t MaxCasesPerResult)5133 InitializeUniqueCases(SwitchInst *SI, PHINode *&PHI, BasicBlock *&CommonDest,
5134 SwitchCaseResultVectorTy &UniqueResults,
5135 Constant *&DefaultResult, const DataLayout &DL,
5136 const TargetTransformInfo &TTI,
5137 uintptr_t MaxUniqueResults, uintptr_t MaxCasesPerResult) {
5138 for (auto &I : SI->cases()) {
5139 ConstantInt *CaseVal = I.getCaseValue();
5140
5141 // Resulting value at phi nodes for this case value.
5142 SwitchCaseResultsTy Results;
5143 if (!GetCaseResults(SI, CaseVal, I.getCaseSuccessor(), &CommonDest, Results,
5144 DL, TTI))
5145 return false;
5146
5147 // Only one value per case is permitted.
5148 if (Results.size() > 1)
5149 return false;
5150
5151 // Add the case->result mapping to UniqueResults.
5152 const uintptr_t NumCasesForResult =
5153 MapCaseToResult(CaseVal, UniqueResults, Results.begin()->second);
5154
5155 // Early out if there are too many cases for this result.
5156 if (NumCasesForResult > MaxCasesPerResult)
5157 return false;
5158
5159 // Early out if there are too many unique results.
5160 if (UniqueResults.size() > MaxUniqueResults)
5161 return false;
5162
5163 // Check the PHI consistency.
5164 if (!PHI)
5165 PHI = Results[0].first;
5166 else if (PHI != Results[0].first)
5167 return false;
5168 }
5169 // Find the default result value.
5170 SmallVector<std::pair<PHINode *, Constant *>, 1> DefaultResults;
5171 BasicBlock *DefaultDest = SI->getDefaultDest();
5172 GetCaseResults(SI, nullptr, SI->getDefaultDest(), &CommonDest, DefaultResults,
5173 DL, TTI);
5174 // If the default value is not found abort unless the default destination
5175 // is unreachable.
5176 DefaultResult =
5177 DefaultResults.size() == 1 ? DefaultResults.begin()->second : nullptr;
5178 if ((!DefaultResult &&
5179 !isa<UnreachableInst>(DefaultDest->getFirstNonPHIOrDbg())))
5180 return false;
5181
5182 return true;
5183 }
5184
5185 // Helper function that checks if it is possible to transform a switch with only
5186 // two cases (or two cases + default) that produces a result into a select.
5187 // Example:
5188 // switch (a) {
5189 // case 10: %0 = icmp eq i32 %a, 10
5190 // return 10; %1 = select i1 %0, i32 10, i32 4
5191 // case 20: ----> %2 = icmp eq i32 %a, 20
5192 // return 2; %3 = select i1 %2, i32 2, i32 %1
5193 // default:
5194 // return 4;
5195 // }
ConvertTwoCaseSwitch(const SwitchCaseResultVectorTy & ResultVector,Constant * DefaultResult,Value * Condition,IRBuilder<> & Builder)5196 static Value *ConvertTwoCaseSwitch(const SwitchCaseResultVectorTy &ResultVector,
5197 Constant *DefaultResult, Value *Condition,
5198 IRBuilder<> &Builder) {
5199 assert(ResultVector.size() == 2 &&
5200 "We should have exactly two unique results at this point");
5201 // If we are selecting between only two cases transform into a simple
5202 // select or a two-way select if default is possible.
5203 if (ResultVector[0].second.size() == 1 &&
5204 ResultVector[1].second.size() == 1) {
5205 ConstantInt *const FirstCase = ResultVector[0].second[0];
5206 ConstantInt *const SecondCase = ResultVector[1].second[0];
5207
5208 bool DefaultCanTrigger = DefaultResult;
5209 Value *SelectValue = ResultVector[1].first;
5210 if (DefaultCanTrigger) {
5211 Value *const ValueCompare =
5212 Builder.CreateICmpEQ(Condition, SecondCase, "switch.selectcmp");
5213 SelectValue = Builder.CreateSelect(ValueCompare, ResultVector[1].first,
5214 DefaultResult, "switch.select");
5215 }
5216 Value *const ValueCompare =
5217 Builder.CreateICmpEQ(Condition, FirstCase, "switch.selectcmp");
5218 return Builder.CreateSelect(ValueCompare, ResultVector[0].first,
5219 SelectValue, "switch.select");
5220 }
5221
5222 return nullptr;
5223 }
5224
5225 // Helper function to cleanup a switch instruction that has been converted into
5226 // a select, fixing up PHI nodes and basic blocks.
RemoveSwitchAfterSelectConversion(SwitchInst * SI,PHINode * PHI,Value * SelectValue,IRBuilder<> & Builder,DomTreeUpdater * DTU)5227 static void RemoveSwitchAfterSelectConversion(SwitchInst *SI, PHINode *PHI,
5228 Value *SelectValue,
5229 IRBuilder<> &Builder,
5230 DomTreeUpdater *DTU) {
5231 std::vector<DominatorTree::UpdateType> Updates;
5232
5233 BasicBlock *SelectBB = SI->getParent();
5234 BasicBlock *DestBB = PHI->getParent();
5235
5236 if (!is_contained(predecessors(DestBB), SelectBB))
5237 Updates.push_back({DominatorTree::Insert, SelectBB, DestBB});
5238 Builder.CreateBr(DestBB);
5239
5240 // Remove the switch.
5241
5242 while (PHI->getBasicBlockIndex(SelectBB) >= 0)
5243 PHI->removeIncomingValue(SelectBB);
5244 PHI->addIncoming(SelectValue, SelectBB);
5245
5246 for (unsigned i = 0, e = SI->getNumSuccessors(); i < e; ++i) {
5247 BasicBlock *Succ = SI->getSuccessor(i);
5248
5249 if (Succ == DestBB)
5250 continue;
5251 Succ->removePredecessor(SelectBB);
5252 Updates.push_back({DominatorTree::Delete, SelectBB, Succ});
5253 }
5254 SI->eraseFromParent();
5255 if (DTU)
5256 DTU->applyUpdates(Updates);
5257 }
5258
5259 /// If the switch is only used to initialize one or more
5260 /// phi nodes in a common successor block with only two different
5261 /// constant values, replace the switch with select.
switchToSelect(SwitchInst * SI,IRBuilder<> & Builder,DomTreeUpdater * DTU,const DataLayout & DL,const TargetTransformInfo & TTI)5262 static bool switchToSelect(SwitchInst *SI, IRBuilder<> &Builder,
5263 DomTreeUpdater *DTU, const DataLayout &DL,
5264 const TargetTransformInfo &TTI) {
5265 Value *const Cond = SI->getCondition();
5266 PHINode *PHI = nullptr;
5267 BasicBlock *CommonDest = nullptr;
5268 Constant *DefaultResult;
5269 SwitchCaseResultVectorTy UniqueResults;
5270 // Collect all the cases that will deliver the same value from the switch.
5271 if (!InitializeUniqueCases(SI, PHI, CommonDest, UniqueResults, DefaultResult,
5272 DL, TTI, 2, 1))
5273 return false;
5274 // Selects choose between maximum two values.
5275 if (UniqueResults.size() != 2)
5276 return false;
5277 assert(PHI != nullptr && "PHI for value select not found");
5278
5279 Builder.SetInsertPoint(SI);
5280 Value *SelectValue =
5281 ConvertTwoCaseSwitch(UniqueResults, DefaultResult, Cond, Builder);
5282 if (SelectValue) {
5283 RemoveSwitchAfterSelectConversion(SI, PHI, SelectValue, Builder, DTU);
5284 return true;
5285 }
5286 // The switch couldn't be converted into a select.
5287 return false;
5288 }
5289
5290 namespace {
5291
5292 /// This class represents a lookup table that can be used to replace a switch.
5293 class SwitchLookupTable {
5294 public:
5295 /// Create a lookup table to use as a switch replacement with the contents
5296 /// of Values, using DefaultValue to fill any holes in the table.
5297 SwitchLookupTable(
5298 Module &M, uint64_t TableSize, ConstantInt *Offset,
5299 const SmallVectorImpl<std::pair<ConstantInt *, Constant *>> &Values,
5300 Constant *DefaultValue, const DataLayout &DL, const StringRef &FuncName);
5301
5302 /// Build instructions with Builder to retrieve the value at
5303 /// the position given by Index in the lookup table.
5304 Value *BuildLookup(Value *Index, IRBuilder<> &Builder);
5305
5306 /// Return true if a table with TableSize elements of
5307 /// type ElementType would fit in a target-legal register.
5308 static bool WouldFitInRegister(const DataLayout &DL, uint64_t TableSize,
5309 Type *ElementType);
5310
5311 private:
5312 // Depending on the contents of the table, it can be represented in
5313 // different ways.
5314 enum {
5315 // For tables where each element contains the same value, we just have to
5316 // store that single value and return it for each lookup.
5317 SingleValueKind,
5318
5319 // For tables where there is a linear relationship between table index
5320 // and values. We calculate the result with a simple multiplication
5321 // and addition instead of a table lookup.
5322 LinearMapKind,
5323
5324 // For small tables with integer elements, we can pack them into a bitmap
5325 // that fits into a target-legal register. Values are retrieved by
5326 // shift and mask operations.
5327 BitMapKind,
5328
5329 // The table is stored as an array of values. Values are retrieved by load
5330 // instructions from the table.
5331 ArrayKind
5332 } Kind;
5333
5334 // For SingleValueKind, this is the single value.
5335 Constant *SingleValue = nullptr;
5336
5337 // For BitMapKind, this is the bitmap.
5338 ConstantInt *BitMap = nullptr;
5339 IntegerType *BitMapElementTy = nullptr;
5340
5341 // For LinearMapKind, these are the constants used to derive the value.
5342 ConstantInt *LinearOffset = nullptr;
5343 ConstantInt *LinearMultiplier = nullptr;
5344
5345 // For ArrayKind, this is the array.
5346 GlobalVariable *Array = nullptr;
5347 };
5348
5349 } // end anonymous namespace
5350
SwitchLookupTable(Module & M,uint64_t TableSize,ConstantInt * Offset,const SmallVectorImpl<std::pair<ConstantInt *,Constant * >> & Values,Constant * DefaultValue,const DataLayout & DL,const StringRef & FuncName)5351 SwitchLookupTable::SwitchLookupTable(
5352 Module &M, uint64_t TableSize, ConstantInt *Offset,
5353 const SmallVectorImpl<std::pair<ConstantInt *, Constant *>> &Values,
5354 Constant *DefaultValue, const DataLayout &DL, const StringRef &FuncName) {
5355 assert(Values.size() && "Can't build lookup table without values!");
5356 assert(TableSize >= Values.size() && "Can't fit values in table!");
5357
5358 // If all values in the table are equal, this is that value.
5359 SingleValue = Values.begin()->second;
5360
5361 Type *ValueType = Values.begin()->second->getType();
5362
5363 // Build up the table contents.
5364 SmallVector<Constant *, 64> TableContents(TableSize);
5365 for (size_t I = 0, E = Values.size(); I != E; ++I) {
5366 ConstantInt *CaseVal = Values[I].first;
5367 Constant *CaseRes = Values[I].second;
5368 assert(CaseRes->getType() == ValueType);
5369
5370 uint64_t Idx = (CaseVal->getValue() - Offset->getValue()).getLimitedValue();
5371 TableContents[Idx] = CaseRes;
5372
5373 if (CaseRes != SingleValue)
5374 SingleValue = nullptr;
5375 }
5376
5377 // Fill in any holes in the table with the default result.
5378 if (Values.size() < TableSize) {
5379 assert(DefaultValue &&
5380 "Need a default value to fill the lookup table holes.");
5381 assert(DefaultValue->getType() == ValueType);
5382 for (uint64_t I = 0; I < TableSize; ++I) {
5383 if (!TableContents[I])
5384 TableContents[I] = DefaultValue;
5385 }
5386
5387 if (DefaultValue != SingleValue)
5388 SingleValue = nullptr;
5389 }
5390
5391 // If each element in the table contains the same value, we only need to store
5392 // that single value.
5393 if (SingleValue) {
5394 Kind = SingleValueKind;
5395 return;
5396 }
5397
5398 // Check if we can derive the value with a linear transformation from the
5399 // table index.
5400 if (isa<IntegerType>(ValueType)) {
5401 bool LinearMappingPossible = true;
5402 APInt PrevVal;
5403 APInt DistToPrev;
5404 assert(TableSize >= 2 && "Should be a SingleValue table.");
5405 // Check if there is the same distance between two consecutive values.
5406 for (uint64_t I = 0; I < TableSize; ++I) {
5407 ConstantInt *ConstVal = dyn_cast<ConstantInt>(TableContents[I]);
5408 if (!ConstVal) {
5409 // This is an undef. We could deal with it, but undefs in lookup tables
5410 // are very seldom. It's probably not worth the additional complexity.
5411 LinearMappingPossible = false;
5412 break;
5413 }
5414 const APInt &Val = ConstVal->getValue();
5415 if (I != 0) {
5416 APInt Dist = Val - PrevVal;
5417 if (I == 1) {
5418 DistToPrev = Dist;
5419 } else if (Dist != DistToPrev) {
5420 LinearMappingPossible = false;
5421 break;
5422 }
5423 }
5424 PrevVal = Val;
5425 }
5426 if (LinearMappingPossible) {
5427 LinearOffset = cast<ConstantInt>(TableContents[0]);
5428 LinearMultiplier = ConstantInt::get(M.getContext(), DistToPrev);
5429 Kind = LinearMapKind;
5430 ++NumLinearMaps;
5431 return;
5432 }
5433 }
5434
5435 // If the type is integer and the table fits in a register, build a bitmap.
5436 if (WouldFitInRegister(DL, TableSize, ValueType)) {
5437 IntegerType *IT = cast<IntegerType>(ValueType);
5438 APInt TableInt(TableSize * IT->getBitWidth(), 0);
5439 for (uint64_t I = TableSize; I > 0; --I) {
5440 TableInt <<= IT->getBitWidth();
5441 // Insert values into the bitmap. Undef values are set to zero.
5442 if (!isa<UndefValue>(TableContents[I - 1])) {
5443 ConstantInt *Val = cast<ConstantInt>(TableContents[I - 1]);
5444 TableInt |= Val->getValue().zext(TableInt.getBitWidth());
5445 }
5446 }
5447 BitMap = ConstantInt::get(M.getContext(), TableInt);
5448 BitMapElementTy = IT;
5449 Kind = BitMapKind;
5450 ++NumBitMaps;
5451 return;
5452 }
5453
5454 // Store the table in an array.
5455 ArrayType *ArrayTy = ArrayType::get(ValueType, TableSize);
5456 Constant *Initializer = ConstantArray::get(ArrayTy, TableContents);
5457
5458 Array = new GlobalVariable(M, ArrayTy, /*isConstant=*/true,
5459 GlobalVariable::PrivateLinkage, Initializer,
5460 "switch.table." + FuncName);
5461 Array->setUnnamedAddr(GlobalValue::UnnamedAddr::Global);
5462 // Set the alignment to that of an array items. We will be only loading one
5463 // value out of it.
5464 Array->setAlignment(Align(DL.getPrefTypeAlignment(ValueType)));
5465 Kind = ArrayKind;
5466 }
5467
BuildLookup(Value * Index,IRBuilder<> & Builder)5468 Value *SwitchLookupTable::BuildLookup(Value *Index, IRBuilder<> &Builder) {
5469 switch (Kind) {
5470 case SingleValueKind:
5471 return SingleValue;
5472 case LinearMapKind: {
5473 // Derive the result value from the input value.
5474 Value *Result = Builder.CreateIntCast(Index, LinearMultiplier->getType(),
5475 false, "switch.idx.cast");
5476 if (!LinearMultiplier->isOne())
5477 Result = Builder.CreateMul(Result, LinearMultiplier, "switch.idx.mult");
5478 if (!LinearOffset->isZero())
5479 Result = Builder.CreateAdd(Result, LinearOffset, "switch.offset");
5480 return Result;
5481 }
5482 case BitMapKind: {
5483 // Type of the bitmap (e.g. i59).
5484 IntegerType *MapTy = BitMap->getType();
5485
5486 // Cast Index to the same type as the bitmap.
5487 // Note: The Index is <= the number of elements in the table, so
5488 // truncating it to the width of the bitmask is safe.
5489 Value *ShiftAmt = Builder.CreateZExtOrTrunc(Index, MapTy, "switch.cast");
5490
5491 // Multiply the shift amount by the element width.
5492 ShiftAmt = Builder.CreateMul(
5493 ShiftAmt, ConstantInt::get(MapTy, BitMapElementTy->getBitWidth()),
5494 "switch.shiftamt");
5495
5496 // Shift down.
5497 Value *DownShifted =
5498 Builder.CreateLShr(BitMap, ShiftAmt, "switch.downshift");
5499 // Mask off.
5500 return Builder.CreateTrunc(DownShifted, BitMapElementTy, "switch.masked");
5501 }
5502 case ArrayKind: {
5503 // Make sure the table index will not overflow when treated as signed.
5504 IntegerType *IT = cast<IntegerType>(Index->getType());
5505 uint64_t TableSize =
5506 Array->getInitializer()->getType()->getArrayNumElements();
5507 if (TableSize > (1ULL << (IT->getBitWidth() - 1)))
5508 Index = Builder.CreateZExt(
5509 Index, IntegerType::get(IT->getContext(), IT->getBitWidth() + 1),
5510 "switch.tableidx.zext");
5511
5512 Value *GEPIndices[] = {Builder.getInt32(0), Index};
5513 Value *GEP = Builder.CreateInBoundsGEP(Array->getValueType(), Array,
5514 GEPIndices, "switch.gep");
5515 return Builder.CreateLoad(
5516 cast<ArrayType>(Array->getValueType())->getElementType(), GEP,
5517 "switch.load");
5518 }
5519 }
5520 llvm_unreachable("Unknown lookup table kind!");
5521 }
5522
WouldFitInRegister(const DataLayout & DL,uint64_t TableSize,Type * ElementType)5523 bool SwitchLookupTable::WouldFitInRegister(const DataLayout &DL,
5524 uint64_t TableSize,
5525 Type *ElementType) {
5526 auto *IT = dyn_cast<IntegerType>(ElementType);
5527 if (!IT)
5528 return false;
5529 // FIXME: If the type is wider than it needs to be, e.g. i8 but all values
5530 // are <= 15, we could try to narrow the type.
5531
5532 // Avoid overflow, fitsInLegalInteger uses unsigned int for the width.
5533 if (TableSize >= UINT_MAX / IT->getBitWidth())
5534 return false;
5535 return DL.fitsInLegalInteger(TableSize * IT->getBitWidth());
5536 }
5537
5538 /// Determine whether a lookup table should be built for this switch, based on
5539 /// the number of cases, size of the table, and the types of the results.
5540 static bool
ShouldBuildLookupTable(SwitchInst * SI,uint64_t TableSize,const TargetTransformInfo & TTI,const DataLayout & DL,const SmallDenseMap<PHINode *,Type * > & ResultTypes)5541 ShouldBuildLookupTable(SwitchInst *SI, uint64_t TableSize,
5542 const TargetTransformInfo &TTI, const DataLayout &DL,
5543 const SmallDenseMap<PHINode *, Type *> &ResultTypes) {
5544 if (SI->getNumCases() > TableSize || TableSize >= UINT64_MAX / 10)
5545 return false; // TableSize overflowed, or mul below might overflow.
5546
5547 bool AllTablesFitInRegister = true;
5548 bool HasIllegalType = false;
5549 for (const auto &I : ResultTypes) {
5550 Type *Ty = I.second;
5551
5552 // Saturate this flag to true.
5553 HasIllegalType = HasIllegalType || !TTI.isTypeLegal(Ty);
5554
5555 // Saturate this flag to false.
5556 AllTablesFitInRegister =
5557 AllTablesFitInRegister &&
5558 SwitchLookupTable::WouldFitInRegister(DL, TableSize, Ty);
5559
5560 // If both flags saturate, we're done. NOTE: This *only* works with
5561 // saturating flags, and all flags have to saturate first due to the
5562 // non-deterministic behavior of iterating over a dense map.
5563 if (HasIllegalType && !AllTablesFitInRegister)
5564 break;
5565 }
5566
5567 // If each table would fit in a register, we should build it anyway.
5568 if (AllTablesFitInRegister)
5569 return true;
5570
5571 // Don't build a table that doesn't fit in-register if it has illegal types.
5572 if (HasIllegalType)
5573 return false;
5574
5575 // The table density should be at least 40%. This is the same criterion as for
5576 // jump tables, see SelectionDAGBuilder::handleJTSwitchCase.
5577 // FIXME: Find the best cut-off.
5578 return SI->getNumCases() * 10 >= TableSize * 4;
5579 }
5580
5581 /// Try to reuse the switch table index compare. Following pattern:
5582 /// \code
5583 /// if (idx < tablesize)
5584 /// r = table[idx]; // table does not contain default_value
5585 /// else
5586 /// r = default_value;
5587 /// if (r != default_value)
5588 /// ...
5589 /// \endcode
5590 /// Is optimized to:
5591 /// \code
5592 /// cond = idx < tablesize;
5593 /// if (cond)
5594 /// r = table[idx];
5595 /// else
5596 /// r = default_value;
5597 /// if (cond)
5598 /// ...
5599 /// \endcode
5600 /// Jump threading will then eliminate the second if(cond).
reuseTableCompare(User * PhiUser,BasicBlock * PhiBlock,BranchInst * RangeCheckBranch,Constant * DefaultValue,const SmallVectorImpl<std::pair<ConstantInt *,Constant * >> & Values)5601 static void reuseTableCompare(
5602 User *PhiUser, BasicBlock *PhiBlock, BranchInst *RangeCheckBranch,
5603 Constant *DefaultValue,
5604 const SmallVectorImpl<std::pair<ConstantInt *, Constant *>> &Values) {
5605 ICmpInst *CmpInst = dyn_cast<ICmpInst>(PhiUser);
5606 if (!CmpInst)
5607 return;
5608
5609 // We require that the compare is in the same block as the phi so that jump
5610 // threading can do its work afterwards.
5611 if (CmpInst->getParent() != PhiBlock)
5612 return;
5613
5614 Constant *CmpOp1 = dyn_cast<Constant>(CmpInst->getOperand(1));
5615 if (!CmpOp1)
5616 return;
5617
5618 Value *RangeCmp = RangeCheckBranch->getCondition();
5619 Constant *TrueConst = ConstantInt::getTrue(RangeCmp->getType());
5620 Constant *FalseConst = ConstantInt::getFalse(RangeCmp->getType());
5621
5622 // Check if the compare with the default value is constant true or false.
5623 Constant *DefaultConst = ConstantExpr::getICmp(CmpInst->getPredicate(),
5624 DefaultValue, CmpOp1, true);
5625 if (DefaultConst != TrueConst && DefaultConst != FalseConst)
5626 return;
5627
5628 // Check if the compare with the case values is distinct from the default
5629 // compare result.
5630 for (auto ValuePair : Values) {
5631 Constant *CaseConst = ConstantExpr::getICmp(CmpInst->getPredicate(),
5632 ValuePair.second, CmpOp1, true);
5633 if (!CaseConst || CaseConst == DefaultConst || isa<UndefValue>(CaseConst))
5634 return;
5635 assert((CaseConst == TrueConst || CaseConst == FalseConst) &&
5636 "Expect true or false as compare result.");
5637 }
5638
5639 // Check if the branch instruction dominates the phi node. It's a simple
5640 // dominance check, but sufficient for our needs.
5641 // Although this check is invariant in the calling loops, it's better to do it
5642 // at this late stage. Practically we do it at most once for a switch.
5643 BasicBlock *BranchBlock = RangeCheckBranch->getParent();
5644 for (auto PI = pred_begin(PhiBlock), E = pred_end(PhiBlock); PI != E; ++PI) {
5645 BasicBlock *Pred = *PI;
5646 if (Pred != BranchBlock && Pred->getUniquePredecessor() != BranchBlock)
5647 return;
5648 }
5649
5650 if (DefaultConst == FalseConst) {
5651 // The compare yields the same result. We can replace it.
5652 CmpInst->replaceAllUsesWith(RangeCmp);
5653 ++NumTableCmpReuses;
5654 } else {
5655 // The compare yields the same result, just inverted. We can replace it.
5656 Value *InvertedTableCmp = BinaryOperator::CreateXor(
5657 RangeCmp, ConstantInt::get(RangeCmp->getType(), 1), "inverted.cmp",
5658 RangeCheckBranch);
5659 CmpInst->replaceAllUsesWith(InvertedTableCmp);
5660 ++NumTableCmpReuses;
5661 }
5662 }
5663
5664 /// If the switch is only used to initialize one or more phi nodes in a common
5665 /// successor block with different constant values, replace the switch with
5666 /// lookup tables.
SwitchToLookupTable(SwitchInst * SI,IRBuilder<> & Builder,DomTreeUpdater * DTU,const DataLayout & DL,const TargetTransformInfo & TTI)5667 static bool SwitchToLookupTable(SwitchInst *SI, IRBuilder<> &Builder,
5668 DomTreeUpdater *DTU, const DataLayout &DL,
5669 const TargetTransformInfo &TTI) {
5670 assert(SI->getNumCases() > 1 && "Degenerate switch?");
5671
5672 BasicBlock *BB = SI->getParent();
5673 Function *Fn = BB->getParent();
5674 // Only build lookup table when we have a target that supports it or the
5675 // attribute is not set.
5676 if (!TTI.shouldBuildLookupTables() ||
5677 (Fn->getFnAttribute("no-jump-tables").getValueAsString() == "true"))
5678 return false;
5679
5680 // FIXME: If the switch is too sparse for a lookup table, perhaps we could
5681 // split off a dense part and build a lookup table for that.
5682
5683 // FIXME: This creates arrays of GEPs to constant strings, which means each
5684 // GEP needs a runtime relocation in PIC code. We should just build one big
5685 // string and lookup indices into that.
5686
5687 // Ignore switches with less than three cases. Lookup tables will not make
5688 // them faster, so we don't analyze them.
5689 if (SI->getNumCases() < 3)
5690 return false;
5691
5692 // Figure out the corresponding result for each case value and phi node in the
5693 // common destination, as well as the min and max case values.
5694 assert(!SI->cases().empty());
5695 SwitchInst::CaseIt CI = SI->case_begin();
5696 ConstantInt *MinCaseVal = CI->getCaseValue();
5697 ConstantInt *MaxCaseVal = CI->getCaseValue();
5698
5699 BasicBlock *CommonDest = nullptr;
5700
5701 using ResultListTy = SmallVector<std::pair<ConstantInt *, Constant *>, 4>;
5702 SmallDenseMap<PHINode *, ResultListTy> ResultLists;
5703
5704 SmallDenseMap<PHINode *, Constant *> DefaultResults;
5705 SmallDenseMap<PHINode *, Type *> ResultTypes;
5706 SmallVector<PHINode *, 4> PHIs;
5707
5708 for (SwitchInst::CaseIt E = SI->case_end(); CI != E; ++CI) {
5709 ConstantInt *CaseVal = CI->getCaseValue();
5710 if (CaseVal->getValue().slt(MinCaseVal->getValue()))
5711 MinCaseVal = CaseVal;
5712 if (CaseVal->getValue().sgt(MaxCaseVal->getValue()))
5713 MaxCaseVal = CaseVal;
5714
5715 // Resulting value at phi nodes for this case value.
5716 using ResultsTy = SmallVector<std::pair<PHINode *, Constant *>, 4>;
5717 ResultsTy Results;
5718 if (!GetCaseResults(SI, CaseVal, CI->getCaseSuccessor(), &CommonDest,
5719 Results, DL, TTI))
5720 return false;
5721
5722 // Append the result from this case to the list for each phi.
5723 for (const auto &I : Results) {
5724 PHINode *PHI = I.first;
5725 Constant *Value = I.second;
5726 if (!ResultLists.count(PHI))
5727 PHIs.push_back(PHI);
5728 ResultLists[PHI].push_back(std::make_pair(CaseVal, Value));
5729 }
5730 }
5731
5732 // Keep track of the result types.
5733 for (PHINode *PHI : PHIs) {
5734 ResultTypes[PHI] = ResultLists[PHI][0].second->getType();
5735 }
5736
5737 uint64_t NumResults = ResultLists[PHIs[0]].size();
5738 APInt RangeSpread = MaxCaseVal->getValue() - MinCaseVal->getValue();
5739 uint64_t TableSize = RangeSpread.getLimitedValue() + 1;
5740 bool TableHasHoles = (NumResults < TableSize);
5741
5742 // If the table has holes, we need a constant result for the default case
5743 // or a bitmask that fits in a register.
5744 SmallVector<std::pair<PHINode *, Constant *>, 4> DefaultResultsList;
5745 bool HasDefaultResults =
5746 GetCaseResults(SI, nullptr, SI->getDefaultDest(), &CommonDest,
5747 DefaultResultsList, DL, TTI);
5748
5749 bool NeedMask = (TableHasHoles && !HasDefaultResults);
5750 if (NeedMask) {
5751 // As an extra penalty for the validity test we require more cases.
5752 if (SI->getNumCases() < 4) // FIXME: Find best threshold value (benchmark).
5753 return false;
5754 if (!DL.fitsInLegalInteger(TableSize))
5755 return false;
5756 }
5757
5758 for (const auto &I : DefaultResultsList) {
5759 PHINode *PHI = I.first;
5760 Constant *Result = I.second;
5761 DefaultResults[PHI] = Result;
5762 }
5763
5764 if (!ShouldBuildLookupTable(SI, TableSize, TTI, DL, ResultTypes))
5765 return false;
5766
5767 std::vector<DominatorTree::UpdateType> Updates;
5768
5769 // Create the BB that does the lookups.
5770 Module &Mod = *CommonDest->getParent()->getParent();
5771 BasicBlock *LookupBB = BasicBlock::Create(
5772 Mod.getContext(), "switch.lookup", CommonDest->getParent(), CommonDest);
5773
5774 // Compute the table index value.
5775 Builder.SetInsertPoint(SI);
5776 Value *TableIndex;
5777 if (MinCaseVal->isNullValue())
5778 TableIndex = SI->getCondition();
5779 else
5780 TableIndex = Builder.CreateSub(SI->getCondition(), MinCaseVal,
5781 "switch.tableidx");
5782
5783 // Compute the maximum table size representable by the integer type we are
5784 // switching upon.
5785 unsigned CaseSize = MinCaseVal->getType()->getPrimitiveSizeInBits();
5786 uint64_t MaxTableSize = CaseSize > 63 ? UINT64_MAX : 1ULL << CaseSize;
5787 assert(MaxTableSize >= TableSize &&
5788 "It is impossible for a switch to have more entries than the max "
5789 "representable value of its input integer type's size.");
5790
5791 // If the default destination is unreachable, or if the lookup table covers
5792 // all values of the conditional variable, branch directly to the lookup table
5793 // BB. Otherwise, check that the condition is within the case range.
5794 const bool DefaultIsReachable =
5795 !isa<UnreachableInst>(SI->getDefaultDest()->getFirstNonPHIOrDbg());
5796 const bool GeneratingCoveredLookupTable = (MaxTableSize == TableSize);
5797 BranchInst *RangeCheckBranch = nullptr;
5798
5799 if (!DefaultIsReachable || GeneratingCoveredLookupTable) {
5800 Builder.CreateBr(LookupBB);
5801 Updates.push_back({DominatorTree::Insert, BB, LookupBB});
5802 // Note: We call removeProdecessor later since we need to be able to get the
5803 // PHI value for the default case in case we're using a bit mask.
5804 } else {
5805 Value *Cmp = Builder.CreateICmpULT(
5806 TableIndex, ConstantInt::get(MinCaseVal->getType(), TableSize));
5807 RangeCheckBranch =
5808 Builder.CreateCondBr(Cmp, LookupBB, SI->getDefaultDest());
5809 Updates.push_back({DominatorTree::Insert, BB, LookupBB});
5810 }
5811
5812 // Populate the BB that does the lookups.
5813 Builder.SetInsertPoint(LookupBB);
5814
5815 if (NeedMask) {
5816 // Before doing the lookup, we do the hole check. The LookupBB is therefore
5817 // re-purposed to do the hole check, and we create a new LookupBB.
5818 BasicBlock *MaskBB = LookupBB;
5819 MaskBB->setName("switch.hole_check");
5820 LookupBB = BasicBlock::Create(Mod.getContext(), "switch.lookup",
5821 CommonDest->getParent(), CommonDest);
5822
5823 // Make the mask's bitwidth at least 8-bit and a power-of-2 to avoid
5824 // unnecessary illegal types.
5825 uint64_t TableSizePowOf2 = NextPowerOf2(std::max(7ULL, TableSize - 1ULL));
5826 APInt MaskInt(TableSizePowOf2, 0);
5827 APInt One(TableSizePowOf2, 1);
5828 // Build bitmask; fill in a 1 bit for every case.
5829 const ResultListTy &ResultList = ResultLists[PHIs[0]];
5830 for (size_t I = 0, E = ResultList.size(); I != E; ++I) {
5831 uint64_t Idx = (ResultList[I].first->getValue() - MinCaseVal->getValue())
5832 .getLimitedValue();
5833 MaskInt |= One << Idx;
5834 }
5835 ConstantInt *TableMask = ConstantInt::get(Mod.getContext(), MaskInt);
5836
5837 // Get the TableIndex'th bit of the bitmask.
5838 // If this bit is 0 (meaning hole) jump to the default destination,
5839 // else continue with table lookup.
5840 IntegerType *MapTy = TableMask->getType();
5841 Value *MaskIndex =
5842 Builder.CreateZExtOrTrunc(TableIndex, MapTy, "switch.maskindex");
5843 Value *Shifted = Builder.CreateLShr(TableMask, MaskIndex, "switch.shifted");
5844 Value *LoBit = Builder.CreateTrunc(
5845 Shifted, Type::getInt1Ty(Mod.getContext()), "switch.lobit");
5846 Builder.CreateCondBr(LoBit, LookupBB, SI->getDefaultDest());
5847 Updates.push_back({DominatorTree::Insert, MaskBB, LookupBB});
5848 Updates.push_back({DominatorTree::Insert, MaskBB, SI->getDefaultDest()});
5849 Builder.SetInsertPoint(LookupBB);
5850 AddPredecessorToBlock(SI->getDefaultDest(), MaskBB, BB);
5851 }
5852
5853 if (!DefaultIsReachable || GeneratingCoveredLookupTable) {
5854 // We cached PHINodes in PHIs. To avoid accessing deleted PHINodes later,
5855 // do not delete PHINodes here.
5856 SI->getDefaultDest()->removePredecessor(BB,
5857 /*KeepOneInputPHIs=*/true);
5858 Updates.push_back({DominatorTree::Delete, BB, SI->getDefaultDest()});
5859 }
5860
5861 bool ReturnedEarly = false;
5862 for (PHINode *PHI : PHIs) {
5863 const ResultListTy &ResultList = ResultLists[PHI];
5864
5865 // If using a bitmask, use any value to fill the lookup table holes.
5866 Constant *DV = NeedMask ? ResultLists[PHI][0].second : DefaultResults[PHI];
5867 StringRef FuncName = Fn->getName();
5868 SwitchLookupTable Table(Mod, TableSize, MinCaseVal, ResultList, DV, DL,
5869 FuncName);
5870
5871 Value *Result = Table.BuildLookup(TableIndex, Builder);
5872
5873 // If the result is used to return immediately from the function, we want to
5874 // do that right here.
5875 if (PHI->hasOneUse() && isa<ReturnInst>(*PHI->user_begin()) &&
5876 PHI->user_back() == CommonDest->getFirstNonPHIOrDbg()) {
5877 Builder.CreateRet(Result);
5878 ReturnedEarly = true;
5879 break;
5880 }
5881
5882 // Do a small peephole optimization: re-use the switch table compare if
5883 // possible.
5884 if (!TableHasHoles && HasDefaultResults && RangeCheckBranch) {
5885 BasicBlock *PhiBlock = PHI->getParent();
5886 // Search for compare instructions which use the phi.
5887 for (auto *User : PHI->users()) {
5888 reuseTableCompare(User, PhiBlock, RangeCheckBranch, DV, ResultList);
5889 }
5890 }
5891
5892 PHI->addIncoming(Result, LookupBB);
5893 }
5894
5895 if (!ReturnedEarly) {
5896 Builder.CreateBr(CommonDest);
5897 Updates.push_back({DominatorTree::Insert, LookupBB, CommonDest});
5898 }
5899
5900 // Remove the switch.
5901 SmallSetVector<BasicBlock *, 8> RemovedSuccessors;
5902 for (unsigned i = 0, e = SI->getNumSuccessors(); i < e; ++i) {
5903 BasicBlock *Succ = SI->getSuccessor(i);
5904
5905 if (Succ == SI->getDefaultDest())
5906 continue;
5907 Succ->removePredecessor(BB);
5908 RemovedSuccessors.insert(Succ);
5909 }
5910 SI->eraseFromParent();
5911
5912 if (DTU) {
5913 for (BasicBlock *RemovedSuccessor : RemovedSuccessors)
5914 Updates.push_back({DominatorTree::Delete, BB, RemovedSuccessor});
5915 DTU->applyUpdates(Updates);
5916 }
5917
5918 ++NumLookupTables;
5919 if (NeedMask)
5920 ++NumLookupTablesHoles;
5921 return true;
5922 }
5923
isSwitchDense(ArrayRef<int64_t> Values)5924 static bool isSwitchDense(ArrayRef<int64_t> Values) {
5925 // See also SelectionDAGBuilder::isDense(), which this function was based on.
5926 uint64_t Diff = (uint64_t)Values.back() - (uint64_t)Values.front();
5927 uint64_t Range = Diff + 1;
5928 uint64_t NumCases = Values.size();
5929 // 40% is the default density for building a jump table in optsize/minsize mode.
5930 uint64_t MinDensity = 40;
5931
5932 return NumCases * 100 >= Range * MinDensity;
5933 }
5934
5935 /// Try to transform a switch that has "holes" in it to a contiguous sequence
5936 /// of cases.
5937 ///
5938 /// A switch such as: switch(i) {case 5: case 9: case 13: case 17:} can be
5939 /// range-reduced to: switch ((i-5) / 4) {case 0: case 1: case 2: case 3:}.
5940 ///
5941 /// This converts a sparse switch into a dense switch which allows better
5942 /// lowering and could also allow transforming into a lookup table.
ReduceSwitchRange(SwitchInst * SI,IRBuilder<> & Builder,const DataLayout & DL,const TargetTransformInfo & TTI)5943 static bool ReduceSwitchRange(SwitchInst *SI, IRBuilder<> &Builder,
5944 const DataLayout &DL,
5945 const TargetTransformInfo &TTI) {
5946 auto *CondTy = cast<IntegerType>(SI->getCondition()->getType());
5947 if (CondTy->getIntegerBitWidth() > 64 ||
5948 !DL.fitsInLegalInteger(CondTy->getIntegerBitWidth()))
5949 return false;
5950 // Only bother with this optimization if there are more than 3 switch cases;
5951 // SDAG will only bother creating jump tables for 4 or more cases.
5952 if (SI->getNumCases() < 4)
5953 return false;
5954
5955 // This transform is agnostic to the signedness of the input or case values. We
5956 // can treat the case values as signed or unsigned. We can optimize more common
5957 // cases such as a sequence crossing zero {-4,0,4,8} if we interpret case values
5958 // as signed.
5959 SmallVector<int64_t,4> Values;
5960 for (auto &C : SI->cases())
5961 Values.push_back(C.getCaseValue()->getValue().getSExtValue());
5962 llvm::sort(Values);
5963
5964 // If the switch is already dense, there's nothing useful to do here.
5965 if (isSwitchDense(Values))
5966 return false;
5967
5968 // First, transform the values such that they start at zero and ascend.
5969 int64_t Base = Values[0];
5970 for (auto &V : Values)
5971 V -= (uint64_t)(Base);
5972
5973 // Now we have signed numbers that have been shifted so that, given enough
5974 // precision, there are no negative values. Since the rest of the transform
5975 // is bitwise only, we switch now to an unsigned representation.
5976
5977 // This transform can be done speculatively because it is so cheap - it
5978 // results in a single rotate operation being inserted.
5979 // FIXME: It's possible that optimizing a switch on powers of two might also
5980 // be beneficial - flag values are often powers of two and we could use a CLZ
5981 // as the key function.
5982
5983 // countTrailingZeros(0) returns 64. As Values is guaranteed to have more than
5984 // one element and LLVM disallows duplicate cases, Shift is guaranteed to be
5985 // less than 64.
5986 unsigned Shift = 64;
5987 for (auto &V : Values)
5988 Shift = std::min(Shift, countTrailingZeros((uint64_t)V));
5989 assert(Shift < 64);
5990 if (Shift > 0)
5991 for (auto &V : Values)
5992 V = (int64_t)((uint64_t)V >> Shift);
5993
5994 if (!isSwitchDense(Values))
5995 // Transform didn't create a dense switch.
5996 return false;
5997
5998 // The obvious transform is to shift the switch condition right and emit a
5999 // check that the condition actually cleanly divided by GCD, i.e.
6000 // C & (1 << Shift - 1) == 0
6001 // inserting a new CFG edge to handle the case where it didn't divide cleanly.
6002 //
6003 // A cheaper way of doing this is a simple ROTR(C, Shift). This performs the
6004 // shift and puts the shifted-off bits in the uppermost bits. If any of these
6005 // are nonzero then the switch condition will be very large and will hit the
6006 // default case.
6007
6008 auto *Ty = cast<IntegerType>(SI->getCondition()->getType());
6009 Builder.SetInsertPoint(SI);
6010 auto *ShiftC = ConstantInt::get(Ty, Shift);
6011 auto *Sub = Builder.CreateSub(SI->getCondition(), ConstantInt::get(Ty, Base));
6012 auto *LShr = Builder.CreateLShr(Sub, ShiftC);
6013 auto *Shl = Builder.CreateShl(Sub, Ty->getBitWidth() - Shift);
6014 auto *Rot = Builder.CreateOr(LShr, Shl);
6015 SI->replaceUsesOfWith(SI->getCondition(), Rot);
6016
6017 for (auto Case : SI->cases()) {
6018 auto *Orig = Case.getCaseValue();
6019 auto Sub = Orig->getValue() - APInt(Ty->getBitWidth(), Base);
6020 Case.setValue(
6021 cast<ConstantInt>(ConstantInt::get(Ty, Sub.lshr(ShiftC->getValue()))));
6022 }
6023 return true;
6024 }
6025
simplifySwitch(SwitchInst * SI,IRBuilder<> & Builder)6026 bool SimplifyCFGOpt::simplifySwitch(SwitchInst *SI, IRBuilder<> &Builder) {
6027 BasicBlock *BB = SI->getParent();
6028
6029 if (isValueEqualityComparison(SI)) {
6030 // If we only have one predecessor, and if it is a branch on this value,
6031 // see if that predecessor totally determines the outcome of this switch.
6032 if (BasicBlock *OnlyPred = BB->getSinglePredecessor())
6033 if (SimplifyEqualityComparisonWithOnlyPredecessor(SI, OnlyPred, Builder))
6034 return requestResimplify();
6035
6036 Value *Cond = SI->getCondition();
6037 if (SelectInst *Select = dyn_cast<SelectInst>(Cond))
6038 if (SimplifySwitchOnSelect(SI, Select))
6039 return requestResimplify();
6040
6041 // If the block only contains the switch, see if we can fold the block
6042 // away into any preds.
6043 if (SI == &*BB->instructionsWithoutDebug().begin())
6044 if (FoldValueComparisonIntoPredecessors(SI, Builder))
6045 return requestResimplify();
6046 }
6047
6048 // Try to transform the switch into an icmp and a branch.
6049 if (TurnSwitchRangeIntoICmp(SI, Builder))
6050 return requestResimplify();
6051
6052 // Remove unreachable cases.
6053 if (eliminateDeadSwitchCases(SI, DTU, Options.AC, DL))
6054 return requestResimplify();
6055
6056 if (switchToSelect(SI, Builder, DTU, DL, TTI))
6057 return requestResimplify();
6058
6059 if (Options.ForwardSwitchCondToPhi && ForwardSwitchConditionToPHI(SI))
6060 return requestResimplify();
6061
6062 // The conversion from switch to lookup tables results in difficult-to-analyze
6063 // code and makes pruning branches much harder. This is a problem if the
6064 // switch expression itself can still be restricted as a result of inlining or
6065 // CVP. Therefore, only apply this transformation during late stages of the
6066 // optimisation pipeline.
6067 if (Options.ConvertSwitchToLookupTable &&
6068 SwitchToLookupTable(SI, Builder, DTU, DL, TTI))
6069 return requestResimplify();
6070
6071 if (ReduceSwitchRange(SI, Builder, DL, TTI))
6072 return requestResimplify();
6073
6074 return false;
6075 }
6076
simplifyIndirectBr(IndirectBrInst * IBI)6077 bool SimplifyCFGOpt::simplifyIndirectBr(IndirectBrInst *IBI) {
6078 BasicBlock *BB = IBI->getParent();
6079 bool Changed = false;
6080
6081 // Eliminate redundant destinations.
6082 SmallPtrSet<Value *, 8> Succs;
6083 SmallSetVector<BasicBlock *, 8> RemovedSuccs;
6084 for (unsigned i = 0, e = IBI->getNumDestinations(); i != e; ++i) {
6085 BasicBlock *Dest = IBI->getDestination(i);
6086 if (!Dest->hasAddressTaken() || !Succs.insert(Dest).second) {
6087 if (!Dest->hasAddressTaken())
6088 RemovedSuccs.insert(Dest);
6089 Dest->removePredecessor(BB);
6090 IBI->removeDestination(i);
6091 --i;
6092 --e;
6093 Changed = true;
6094 }
6095 }
6096
6097 if (DTU) {
6098 std::vector<DominatorTree::UpdateType> Updates;
6099 Updates.reserve(RemovedSuccs.size());
6100 for (auto *RemovedSucc : RemovedSuccs)
6101 Updates.push_back({DominatorTree::Delete, BB, RemovedSucc});
6102 DTU->applyUpdates(Updates);
6103 }
6104
6105 if (IBI->getNumDestinations() == 0) {
6106 // If the indirectbr has no successors, change it to unreachable.
6107 new UnreachableInst(IBI->getContext(), IBI);
6108 EraseTerminatorAndDCECond(IBI);
6109 return true;
6110 }
6111
6112 if (IBI->getNumDestinations() == 1) {
6113 // If the indirectbr has one successor, change it to a direct branch.
6114 BranchInst::Create(IBI->getDestination(0), IBI);
6115 EraseTerminatorAndDCECond(IBI);
6116 return true;
6117 }
6118
6119 if (SelectInst *SI = dyn_cast<SelectInst>(IBI->getAddress())) {
6120 if (SimplifyIndirectBrOnSelect(IBI, SI))
6121 return requestResimplify();
6122 }
6123 return Changed;
6124 }
6125
6126 /// Given an block with only a single landing pad and a unconditional branch
6127 /// try to find another basic block which this one can be merged with. This
6128 /// handles cases where we have multiple invokes with unique landing pads, but
6129 /// a shared handler.
6130 ///
6131 /// We specifically choose to not worry about merging non-empty blocks
6132 /// here. That is a PRE/scheduling problem and is best solved elsewhere. In
6133 /// practice, the optimizer produces empty landing pad blocks quite frequently
6134 /// when dealing with exception dense code. (see: instcombine, gvn, if-else
6135 /// sinking in this file)
6136 ///
6137 /// This is primarily a code size optimization. We need to avoid performing
6138 /// any transform which might inhibit optimization (such as our ability to
6139 /// specialize a particular handler via tail commoning). We do this by not
6140 /// merging any blocks which require us to introduce a phi. Since the same
6141 /// values are flowing through both blocks, we don't lose any ability to
6142 /// specialize. If anything, we make such specialization more likely.
6143 ///
6144 /// TODO - This transformation could remove entries from a phi in the target
6145 /// block when the inputs in the phi are the same for the two blocks being
6146 /// merged. In some cases, this could result in removal of the PHI entirely.
TryToMergeLandingPad(LandingPadInst * LPad,BranchInst * BI,BasicBlock * BB,DomTreeUpdater * DTU)6147 static bool TryToMergeLandingPad(LandingPadInst *LPad, BranchInst *BI,
6148 BasicBlock *BB, DomTreeUpdater *DTU) {
6149 auto Succ = BB->getUniqueSuccessor();
6150 assert(Succ);
6151 // If there's a phi in the successor block, we'd likely have to introduce
6152 // a phi into the merged landing pad block.
6153 if (isa<PHINode>(*Succ->begin()))
6154 return false;
6155
6156 for (BasicBlock *OtherPred : predecessors(Succ)) {
6157 if (BB == OtherPred)
6158 continue;
6159 BasicBlock::iterator I = OtherPred->begin();
6160 LandingPadInst *LPad2 = dyn_cast<LandingPadInst>(I);
6161 if (!LPad2 || !LPad2->isIdenticalTo(LPad))
6162 continue;
6163 for (++I; isa<DbgInfoIntrinsic>(I); ++I)
6164 ;
6165 BranchInst *BI2 = dyn_cast<BranchInst>(I);
6166 if (!BI2 || !BI2->isIdenticalTo(BI))
6167 continue;
6168
6169 std::vector<DominatorTree::UpdateType> Updates;
6170
6171 // We've found an identical block. Update our predecessors to take that
6172 // path instead and make ourselves dead.
6173 SmallPtrSet<BasicBlock *, 16> Preds;
6174 Preds.insert(pred_begin(BB), pred_end(BB));
6175 for (BasicBlock *Pred : Preds) {
6176 InvokeInst *II = cast<InvokeInst>(Pred->getTerminator());
6177 assert(II->getNormalDest() != BB && II->getUnwindDest() == BB &&
6178 "unexpected successor");
6179 II->setUnwindDest(OtherPred);
6180 Updates.push_back({DominatorTree::Insert, Pred, OtherPred});
6181 Updates.push_back({DominatorTree::Delete, Pred, BB});
6182 }
6183
6184 // The debug info in OtherPred doesn't cover the merged control flow that
6185 // used to go through BB. We need to delete it or update it.
6186 for (auto I = OtherPred->begin(), E = OtherPred->end(); I != E;) {
6187 Instruction &Inst = *I;
6188 I++;
6189 if (isa<DbgInfoIntrinsic>(Inst))
6190 Inst.eraseFromParent();
6191 }
6192
6193 SmallPtrSet<BasicBlock *, 16> Succs;
6194 Succs.insert(succ_begin(BB), succ_end(BB));
6195 for (BasicBlock *Succ : Succs) {
6196 Succ->removePredecessor(BB);
6197 Updates.push_back({DominatorTree::Delete, BB, Succ});
6198 }
6199
6200 IRBuilder<> Builder(BI);
6201 Builder.CreateUnreachable();
6202 BI->eraseFromParent();
6203 if (DTU)
6204 DTU->applyUpdates(Updates);
6205 return true;
6206 }
6207 return false;
6208 }
6209
simplifyBranch(BranchInst * Branch,IRBuilder<> & Builder)6210 bool SimplifyCFGOpt::simplifyBranch(BranchInst *Branch, IRBuilder<> &Builder) {
6211 return Branch->isUnconditional() ? simplifyUncondBranch(Branch, Builder)
6212 : simplifyCondBranch(Branch, Builder);
6213 }
6214
simplifyUncondBranch(BranchInst * BI,IRBuilder<> & Builder)6215 bool SimplifyCFGOpt::simplifyUncondBranch(BranchInst *BI,
6216 IRBuilder<> &Builder) {
6217 BasicBlock *BB = BI->getParent();
6218 BasicBlock *Succ = BI->getSuccessor(0);
6219
6220 // If the Terminator is the only non-phi instruction, simplify the block.
6221 // If LoopHeader is provided, check if the block or its successor is a loop
6222 // header. (This is for early invocations before loop simplify and
6223 // vectorization to keep canonical loop forms for nested loops. These blocks
6224 // can be eliminated when the pass is invoked later in the back-end.)
6225 // Note that if BB has only one predecessor then we do not introduce new
6226 // backedge, so we can eliminate BB.
6227 bool NeedCanonicalLoop =
6228 Options.NeedCanonicalLoop &&
6229 (!LoopHeaders.empty() && BB->hasNPredecessorsOrMore(2) &&
6230 (is_contained(LoopHeaders, BB) || is_contained(LoopHeaders, Succ)));
6231 BasicBlock::iterator I = BB->getFirstNonPHIOrDbg()->getIterator();
6232 if (I->isTerminator() && BB != &BB->getParent()->getEntryBlock() &&
6233 !NeedCanonicalLoop && TryToSimplifyUncondBranchFromEmptyBlock(BB, DTU))
6234 return true;
6235
6236 // If the only instruction in the block is a seteq/setne comparison against a
6237 // constant, try to simplify the block.
6238 if (ICmpInst *ICI = dyn_cast<ICmpInst>(I))
6239 if (ICI->isEquality() && isa<ConstantInt>(ICI->getOperand(1))) {
6240 for (++I; isa<DbgInfoIntrinsic>(I); ++I)
6241 ;
6242 if (I->isTerminator() &&
6243 tryToSimplifyUncondBranchWithICmpInIt(ICI, Builder))
6244 return true;
6245 }
6246
6247 // See if we can merge an empty landing pad block with another which is
6248 // equivalent.
6249 if (LandingPadInst *LPad = dyn_cast<LandingPadInst>(I)) {
6250 for (++I; isa<DbgInfoIntrinsic>(I); ++I)
6251 ;
6252 if (I->isTerminator() && TryToMergeLandingPad(LPad, BI, BB, DTU))
6253 return true;
6254 }
6255
6256 // If this basic block is ONLY a compare and a branch, and if a predecessor
6257 // branches to us and our successor, fold the comparison into the
6258 // predecessor and use logical operations to update the incoming value
6259 // for PHI nodes in common successor.
6260 if (FoldBranchToCommonDest(BI, DTU, /*MSSAU=*/nullptr, &TTI,
6261 Options.BonusInstThreshold))
6262 return requestResimplify();
6263 return false;
6264 }
6265
allPredecessorsComeFromSameSource(BasicBlock * BB)6266 static BasicBlock *allPredecessorsComeFromSameSource(BasicBlock *BB) {
6267 BasicBlock *PredPred = nullptr;
6268 for (auto *P : predecessors(BB)) {
6269 BasicBlock *PPred = P->getSinglePredecessor();
6270 if (!PPred || (PredPred && PredPred != PPred))
6271 return nullptr;
6272 PredPred = PPred;
6273 }
6274 return PredPred;
6275 }
6276
simplifyCondBranch(BranchInst * BI,IRBuilder<> & Builder)6277 bool SimplifyCFGOpt::simplifyCondBranch(BranchInst *BI, IRBuilder<> &Builder) {
6278 BasicBlock *BB = BI->getParent();
6279 if (!Options.SimplifyCondBranch)
6280 return false;
6281
6282 // Conditional branch
6283 if (isValueEqualityComparison(BI)) {
6284 // If we only have one predecessor, and if it is a branch on this value,
6285 // see if that predecessor totally determines the outcome of this
6286 // switch.
6287 if (BasicBlock *OnlyPred = BB->getSinglePredecessor())
6288 if (SimplifyEqualityComparisonWithOnlyPredecessor(BI, OnlyPred, Builder))
6289 return requestResimplify();
6290
6291 // This block must be empty, except for the setcond inst, if it exists.
6292 // Ignore dbg intrinsics.
6293 auto I = BB->instructionsWithoutDebug().begin();
6294 if (&*I == BI) {
6295 if (FoldValueComparisonIntoPredecessors(BI, Builder))
6296 return requestResimplify();
6297 } else if (&*I == cast<Instruction>(BI->getCondition())) {
6298 ++I;
6299 if (&*I == BI && FoldValueComparisonIntoPredecessors(BI, Builder))
6300 return requestResimplify();
6301 }
6302 }
6303
6304 // Try to turn "br (X == 0 | X == 1), T, F" into a switch instruction.
6305 if (SimplifyBranchOnICmpChain(BI, Builder, DL))
6306 return true;
6307
6308 // If this basic block has dominating predecessor blocks and the dominating
6309 // blocks' conditions imply BI's condition, we know the direction of BI.
6310 Optional<bool> Imp = isImpliedByDomCondition(BI->getCondition(), BI, DL);
6311 if (Imp) {
6312 // Turn this into a branch on constant.
6313 auto *OldCond = BI->getCondition();
6314 ConstantInt *TorF = *Imp ? ConstantInt::getTrue(BB->getContext())
6315 : ConstantInt::getFalse(BB->getContext());
6316 BI->setCondition(TorF);
6317 RecursivelyDeleteTriviallyDeadInstructions(OldCond);
6318 return requestResimplify();
6319 }
6320
6321 // If this basic block is ONLY a compare and a branch, and if a predecessor
6322 // branches to us and one of our successors, fold the comparison into the
6323 // predecessor and use logical operations to pick the right destination.
6324 if (FoldBranchToCommonDest(BI, DTU, /*MSSAU=*/nullptr, &TTI,
6325 Options.BonusInstThreshold))
6326 return requestResimplify();
6327
6328 // We have a conditional branch to two blocks that are only reachable
6329 // from BI. We know that the condbr dominates the two blocks, so see if
6330 // there is any identical code in the "then" and "else" blocks. If so, we
6331 // can hoist it up to the branching block.
6332 if (BI->getSuccessor(0)->getSinglePredecessor()) {
6333 if (BI->getSuccessor(1)->getSinglePredecessor()) {
6334 if (HoistCommon && Options.HoistCommonInsts)
6335 if (HoistThenElseCodeToIf(BI, TTI))
6336 return requestResimplify();
6337 } else {
6338 // If Successor #1 has multiple preds, we may be able to conditionally
6339 // execute Successor #0 if it branches to Successor #1.
6340 Instruction *Succ0TI = BI->getSuccessor(0)->getTerminator();
6341 if (Succ0TI->getNumSuccessors() == 1 &&
6342 Succ0TI->getSuccessor(0) == BI->getSuccessor(1))
6343 if (SpeculativelyExecuteBB(BI, BI->getSuccessor(0), TTI))
6344 return requestResimplify();
6345 }
6346 } else if (BI->getSuccessor(1)->getSinglePredecessor()) {
6347 // If Successor #0 has multiple preds, we may be able to conditionally
6348 // execute Successor #1 if it branches to Successor #0.
6349 Instruction *Succ1TI = BI->getSuccessor(1)->getTerminator();
6350 if (Succ1TI->getNumSuccessors() == 1 &&
6351 Succ1TI->getSuccessor(0) == BI->getSuccessor(0))
6352 if (SpeculativelyExecuteBB(BI, BI->getSuccessor(1), TTI))
6353 return requestResimplify();
6354 }
6355
6356 // If this is a branch on a phi node in the current block, thread control
6357 // through this block if any PHI node entries are constants.
6358 if (PHINode *PN = dyn_cast<PHINode>(BI->getCondition()))
6359 if (PN->getParent() == BI->getParent())
6360 if (FoldCondBranchOnPHI(BI, DTU, DL, Options.AC))
6361 return requestResimplify();
6362
6363 // Scan predecessor blocks for conditional branches.
6364 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
6365 if (BranchInst *PBI = dyn_cast<BranchInst>((*PI)->getTerminator()))
6366 if (PBI != BI && PBI->isConditional())
6367 if (SimplifyCondBranchToCondBranch(PBI, BI, DTU, DL, TTI))
6368 return requestResimplify();
6369
6370 // Look for diamond patterns.
6371 if (MergeCondStores)
6372 if (BasicBlock *PrevBB = allPredecessorsComeFromSameSource(BB))
6373 if (BranchInst *PBI = dyn_cast<BranchInst>(PrevBB->getTerminator()))
6374 if (PBI != BI && PBI->isConditional())
6375 if (mergeConditionalStores(PBI, BI, DTU, DL, TTI))
6376 return requestResimplify();
6377
6378 return false;
6379 }
6380
6381 /// Check if passing a value to an instruction will cause undefined behavior.
passingValueIsAlwaysUndefined(Value * V,Instruction * I,bool PtrValueMayBeModified)6382 static bool passingValueIsAlwaysUndefined(Value *V, Instruction *I, bool PtrValueMayBeModified) {
6383 Constant *C = dyn_cast<Constant>(V);
6384 if (!C)
6385 return false;
6386
6387 if (I->use_empty())
6388 return false;
6389
6390 if (C->isNullValue() || isa<UndefValue>(C)) {
6391 // Only look at the first use, avoid hurting compile time with long uselists
6392 User *Use = *I->user_begin();
6393
6394 // Now make sure that there are no instructions in between that can alter
6395 // control flow (eg. calls)
6396 for (BasicBlock::iterator
6397 i = ++BasicBlock::iterator(I),
6398 UI = BasicBlock::iterator(dyn_cast<Instruction>(Use));
6399 i != UI; ++i)
6400 if (i == I->getParent()->end() || i->mayHaveSideEffects())
6401 return false;
6402
6403 // Look through GEPs. A load from a GEP derived from NULL is still undefined
6404 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Use))
6405 if (GEP->getPointerOperand() == I) {
6406 if (!GEP->isInBounds() || !GEP->hasAllZeroIndices())
6407 PtrValueMayBeModified = true;
6408 return passingValueIsAlwaysUndefined(V, GEP, PtrValueMayBeModified);
6409 }
6410
6411 // Look through bitcasts.
6412 if (BitCastInst *BC = dyn_cast<BitCastInst>(Use))
6413 return passingValueIsAlwaysUndefined(V, BC, PtrValueMayBeModified);
6414
6415 // Load from null is undefined.
6416 if (LoadInst *LI = dyn_cast<LoadInst>(Use))
6417 if (!LI->isVolatile())
6418 return !NullPointerIsDefined(LI->getFunction(),
6419 LI->getPointerAddressSpace());
6420
6421 // Store to null is undefined.
6422 if (StoreInst *SI = dyn_cast<StoreInst>(Use))
6423 if (!SI->isVolatile())
6424 return (!NullPointerIsDefined(SI->getFunction(),
6425 SI->getPointerAddressSpace())) &&
6426 SI->getPointerOperand() == I;
6427
6428 if (auto *CB = dyn_cast<CallBase>(Use)) {
6429 if (C->isNullValue() && NullPointerIsDefined(CB->getFunction()))
6430 return false;
6431 // A call to null is undefined.
6432 if (CB->getCalledOperand() == I)
6433 return true;
6434
6435 if (C->isNullValue()) {
6436 for (const llvm::Use &Arg : CB->args())
6437 if (Arg == I) {
6438 unsigned ArgIdx = CB->getArgOperandNo(&Arg);
6439 if (CB->paramHasAttr(ArgIdx, Attribute::NonNull) &&
6440 CB->paramHasAttr(ArgIdx, Attribute::NoUndef)) {
6441 // Passing null to a nonnnull+noundef argument is undefined.
6442 return !PtrValueMayBeModified;
6443 }
6444 }
6445 } else if (isa<UndefValue>(C)) {
6446 // Passing undef to a noundef argument is undefined.
6447 for (const llvm::Use &Arg : CB->args())
6448 if (Arg == I) {
6449 unsigned ArgIdx = CB->getArgOperandNo(&Arg);
6450 if (CB->paramHasAttr(ArgIdx, Attribute::NoUndef)) {
6451 // Passing undef to a noundef argument is undefined.
6452 return true;
6453 }
6454 }
6455 }
6456 }
6457 }
6458 return false;
6459 }
6460
6461 /// If BB has an incoming value that will always trigger undefined behavior
6462 /// (eg. null pointer dereference), remove the branch leading here.
removeUndefIntroducingPredecessor(BasicBlock * BB,DomTreeUpdater * DTU)6463 static bool removeUndefIntroducingPredecessor(BasicBlock *BB,
6464 DomTreeUpdater *DTU) {
6465 for (PHINode &PHI : BB->phis())
6466 for (unsigned i = 0, e = PHI.getNumIncomingValues(); i != e; ++i)
6467 if (passingValueIsAlwaysUndefined(PHI.getIncomingValue(i), &PHI)) {
6468 BasicBlock *Predecessor = PHI.getIncomingBlock(i);
6469 Instruction *T = Predecessor->getTerminator();
6470 IRBuilder<> Builder(T);
6471 if (BranchInst *BI = dyn_cast<BranchInst>(T)) {
6472 BB->removePredecessor(Predecessor);
6473 // Turn uncoditional branches into unreachables and remove the dead
6474 // destination from conditional branches.
6475 if (BI->isUnconditional())
6476 Builder.CreateUnreachable();
6477 else
6478 Builder.CreateBr(BI->getSuccessor(0) == BB ? BI->getSuccessor(1)
6479 : BI->getSuccessor(0));
6480 BI->eraseFromParent();
6481 if (DTU)
6482 DTU->applyUpdates({{DominatorTree::Delete, Predecessor, BB}});
6483 return true;
6484 }
6485 // TODO: SwitchInst.
6486 }
6487
6488 return false;
6489 }
6490
simplifyOnceImpl(BasicBlock * BB)6491 bool SimplifyCFGOpt::simplifyOnceImpl(BasicBlock *BB) {
6492 bool Changed = false;
6493
6494 assert(BB && BB->getParent() && "Block not embedded in function!");
6495 assert(BB->getTerminator() && "Degenerate basic block encountered!");
6496
6497 // Remove basic blocks that have no predecessors (except the entry block)...
6498 // or that just have themself as a predecessor. These are unreachable.
6499 if ((pred_empty(BB) && BB != &BB->getParent()->getEntryBlock()) ||
6500 BB->getSinglePredecessor() == BB) {
6501 LLVM_DEBUG(dbgs() << "Removing BB: \n" << *BB);
6502 DeleteDeadBlock(BB, DTU);
6503 return true;
6504 }
6505
6506 // Check to see if we can constant propagate this terminator instruction
6507 // away...
6508 Changed |= ConstantFoldTerminator(BB, /*DeleteDeadConditions=*/true,
6509 /*TLI=*/nullptr, DTU);
6510
6511 // Check for and eliminate duplicate PHI nodes in this block.
6512 Changed |= EliminateDuplicatePHINodes(BB);
6513
6514 // Check for and remove branches that will always cause undefined behavior.
6515 Changed |= removeUndefIntroducingPredecessor(BB, DTU);
6516
6517 // Merge basic blocks into their predecessor if there is only one distinct
6518 // pred, and if there is only one distinct successor of the predecessor, and
6519 // if there are no PHI nodes.
6520 if (MergeBlockIntoPredecessor(BB, DTU))
6521 return true;
6522
6523 if (SinkCommon && Options.SinkCommonInsts)
6524 Changed |= SinkCommonCodeFromPredecessors(BB, DTU);
6525
6526 IRBuilder<> Builder(BB);
6527
6528 if (Options.FoldTwoEntryPHINode) {
6529 // If there is a trivial two-entry PHI node in this basic block, and we can
6530 // eliminate it, do so now.
6531 if (auto *PN = dyn_cast<PHINode>(BB->begin()))
6532 if (PN->getNumIncomingValues() == 2)
6533 Changed |= FoldTwoEntryPHINode(PN, TTI, DTU, DL);
6534 }
6535
6536 Instruction *Terminator = BB->getTerminator();
6537 Builder.SetInsertPoint(Terminator);
6538 switch (Terminator->getOpcode()) {
6539 case Instruction::Br:
6540 Changed |= simplifyBranch(cast<BranchInst>(Terminator), Builder);
6541 break;
6542 case Instruction::Ret:
6543 Changed |= simplifyReturn(cast<ReturnInst>(Terminator), Builder);
6544 break;
6545 case Instruction::Resume:
6546 Changed |= simplifyResume(cast<ResumeInst>(Terminator), Builder);
6547 break;
6548 case Instruction::CleanupRet:
6549 Changed |= simplifyCleanupReturn(cast<CleanupReturnInst>(Terminator));
6550 break;
6551 case Instruction::Switch:
6552 Changed |= simplifySwitch(cast<SwitchInst>(Terminator), Builder);
6553 break;
6554 case Instruction::Unreachable:
6555 Changed |= simplifyUnreachable(cast<UnreachableInst>(Terminator));
6556 break;
6557 case Instruction::IndirectBr:
6558 Changed |= simplifyIndirectBr(cast<IndirectBrInst>(Terminator));
6559 break;
6560 }
6561
6562 return Changed;
6563 }
6564
simplifyOnce(BasicBlock * BB)6565 bool SimplifyCFGOpt::simplifyOnce(BasicBlock *BB) {
6566 bool Changed = simplifyOnceImpl(BB);
6567
6568 assert((!RequireAndPreserveDomTree ||
6569 (DTU &&
6570 DTU->getDomTree().verify(DominatorTree::VerificationLevel::Full))) &&
6571 "Failed to maintain validity of domtree!");
6572
6573 return Changed;
6574 }
6575
run(BasicBlock * BB)6576 bool SimplifyCFGOpt::run(BasicBlock *BB) {
6577 assert((!RequireAndPreserveDomTree ||
6578 (DTU &&
6579 DTU->getDomTree().verify(DominatorTree::VerificationLevel::Full))) &&
6580 "Original domtree is invalid?");
6581
6582 bool Changed = false;
6583
6584 // Repeated simplify BB as long as resimplification is requested.
6585 do {
6586 Resimplify = false;
6587
6588 // Perform one round of simplifcation. Resimplify flag will be set if
6589 // another iteration is requested.
6590 Changed |= simplifyOnce(BB);
6591 } while (Resimplify);
6592
6593 return Changed;
6594 }
6595
simplifyCFG(BasicBlock * BB,const TargetTransformInfo & TTI,DomTreeUpdater * DTU,const SimplifyCFGOptions & Options,ArrayRef<WeakVH> LoopHeaders)6596 bool llvm::simplifyCFG(BasicBlock *BB, const TargetTransformInfo &TTI,
6597 DomTreeUpdater *DTU, const SimplifyCFGOptions &Options,
6598 ArrayRef<WeakVH> LoopHeaders) {
6599 return SimplifyCFGOpt(TTI, RequireAndPreserveDomTree ? DTU : nullptr,
6600 BB->getModule()->getDataLayout(), LoopHeaders, Options)
6601 .run(BB);
6602 }
6603