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/Optional.h"
17 #include "llvm/ADT/STLExtras.h"
18 #include "llvm/ADT/SetOperations.h"
19 #include "llvm/ADT/SetVector.h"
20 #include "llvm/ADT/SmallPtrSet.h"
21 #include "llvm/ADT/SmallVector.h"
22 #include "llvm/ADT/Statistic.h"
23 #include "llvm/ADT/StringRef.h"
24 #include "llvm/Analysis/AssumptionCache.h"
25 #include "llvm/Analysis/ConstantFolding.h"
26 #include "llvm/Analysis/EHPersonalities.h"
27 #include "llvm/Analysis/GuardUtils.h"
28 #include "llvm/Analysis/InstructionSimplify.h"
29 #include "llvm/Analysis/MemorySSA.h"
30 #include "llvm/Analysis/MemorySSAUpdater.h"
31 #include "llvm/Analysis/TargetTransformInfo.h"
32 #include "llvm/Analysis/ValueTracking.h"
33 #include "llvm/IR/Attributes.h"
34 #include "llvm/IR/BasicBlock.h"
35 #include "llvm/IR/CFG.h"
36 #include "llvm/IR/Constant.h"
37 #include "llvm/IR/ConstantRange.h"
38 #include "llvm/IR/Constants.h"
39 #include "llvm/IR/DataLayout.h"
40 #include "llvm/IR/DerivedTypes.h"
41 #include "llvm/IR/Function.h"
42 #include "llvm/IR/GlobalValue.h"
43 #include "llvm/IR/GlobalVariable.h"
44 #include "llvm/IR/IRBuilder.h"
45 #include "llvm/IR/InstrTypes.h"
46 #include "llvm/IR/Instruction.h"
47 #include "llvm/IR/Instructions.h"
48 #include "llvm/IR/IntrinsicInst.h"
49 #include "llvm/IR/Intrinsics.h"
50 #include "llvm/IR/LLVMContext.h"
51 #include "llvm/IR/MDBuilder.h"
52 #include "llvm/IR/Metadata.h"
53 #include "llvm/IR/Module.h"
54 #include "llvm/IR/NoFolder.h"
55 #include "llvm/IR/Operator.h"
56 #include "llvm/IR/PatternMatch.h"
57 #include "llvm/IR/Type.h"
58 #include "llvm/IR/Use.h"
59 #include "llvm/IR/User.h"
60 #include "llvm/IR/Value.h"
61 #include "llvm/Support/Casting.h"
62 #include "llvm/Support/CommandLine.h"
63 #include "llvm/Support/Debug.h"
64 #include "llvm/Support/ErrorHandling.h"
65 #include "llvm/Support/KnownBits.h"
66 #include "llvm/Support/MathExtras.h"
67 #include "llvm/Support/raw_ostream.h"
68 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
69 #include "llvm/Transforms/Utils/Local.h"
70 #include "llvm/Transforms/Utils/ValueMapper.h"
71 #include <algorithm>
72 #include <cassert>
73 #include <climits>
74 #include <cstddef>
75 #include <cstdint>
76 #include <iterator>
77 #include <map>
78 #include <set>
79 #include <tuple>
80 #include <utility>
81 #include <vector>
82
83 using namespace llvm;
84 using namespace PatternMatch;
85
86 #define DEBUG_TYPE "simplifycfg"
87
88 // Chosen as 2 so as to be cheap, but still to have enough power to fold
89 // a select, so the "clamp" idiom (of a min followed by a max) will be caught.
90 // To catch this, we need to fold a compare and a select, hence '2' being the
91 // minimum reasonable default.
92 static cl::opt<unsigned> PHINodeFoldingThreshold(
93 "phi-node-folding-threshold", cl::Hidden, cl::init(2),
94 cl::desc(
95 "Control the amount of phi node folding to perform (default = 2)"));
96
97 static cl::opt<unsigned> TwoEntryPHINodeFoldingThreshold(
98 "two-entry-phi-node-folding-threshold", cl::Hidden, cl::init(4),
99 cl::desc("Control the maximal total instruction cost that we are willing "
100 "to speculatively execute to fold a 2-entry PHI node into a "
101 "select (default = 4)"));
102
103 static cl::opt<bool> DupRet(
104 "simplifycfg-dup-ret", cl::Hidden, cl::init(false),
105 cl::desc("Duplicate return instructions into unconditional branches"));
106
107 static cl::opt<bool>
108 SinkCommon("simplifycfg-sink-common", cl::Hidden, cl::init(true),
109 cl::desc("Sink common instructions down to the end block"));
110
111 static cl::opt<bool> HoistCondStores(
112 "simplifycfg-hoist-cond-stores", cl::Hidden, cl::init(true),
113 cl::desc("Hoist conditional stores if an unconditional store precedes"));
114
115 static cl::opt<bool> MergeCondStores(
116 "simplifycfg-merge-cond-stores", cl::Hidden, cl::init(true),
117 cl::desc("Hoist conditional stores even if an unconditional store does not "
118 "precede - hoist multiple conditional stores into a single "
119 "predicated store"));
120
121 static cl::opt<bool> MergeCondStoresAggressively(
122 "simplifycfg-merge-cond-stores-aggressively", cl::Hidden, cl::init(false),
123 cl::desc("When merging conditional stores, do so even if the resultant "
124 "basic blocks are unlikely to be if-converted as a result"));
125
126 static cl::opt<bool> SpeculateOneExpensiveInst(
127 "speculate-one-expensive-inst", cl::Hidden, cl::init(true),
128 cl::desc("Allow exactly one expensive instruction to be speculatively "
129 "executed"));
130
131 static cl::opt<unsigned> MaxSpeculationDepth(
132 "max-speculation-depth", cl::Hidden, cl::init(10),
133 cl::desc("Limit maximum recursion depth when calculating costs of "
134 "speculatively executed instructions"));
135
136 static cl::opt<int>
137 MaxSmallBlockSize("simplifycfg-max-small-block-size", cl::Hidden, cl::init(10),
138 cl::desc("Max size of a block which is still considered "
139 "small enough to thread through"));
140
141 STATISTIC(NumBitMaps, "Number of switch instructions turned into bitmaps");
142 STATISTIC(NumLinearMaps,
143 "Number of switch instructions turned into linear mapping");
144 STATISTIC(NumLookupTables,
145 "Number of switch instructions turned into lookup tables");
146 STATISTIC(
147 NumLookupTablesHoles,
148 "Number of switch instructions turned into lookup tables (holes checked)");
149 STATISTIC(NumTableCmpReuses, "Number of reused switch table lookup compares");
150 STATISTIC(NumSinkCommons,
151 "Number of common instructions sunk down to the end block");
152 STATISTIC(NumSpeculations, "Number of speculative executed instructions");
153
154 namespace {
155
156 // The first field contains the value that the switch produces when a certain
157 // case group is selected, and the second field is a vector containing the
158 // cases composing the case group.
159 using SwitchCaseResultVectorTy =
160 SmallVector<std::pair<Constant *, SmallVector<ConstantInt *, 4>>, 2>;
161
162 // The first field contains the phi node that generates a result of the switch
163 // and the second field contains the value generated for a certain case in the
164 // switch for that PHI.
165 using SwitchCaseResultsTy = SmallVector<std::pair<PHINode *, Constant *>, 4>;
166
167 /// ValueEqualityComparisonCase - Represents a case of a switch.
168 struct ValueEqualityComparisonCase {
169 ConstantInt *Value;
170 BasicBlock *Dest;
171
ValueEqualityComparisonCase__anon122b05330111::ValueEqualityComparisonCase172 ValueEqualityComparisonCase(ConstantInt *Value, BasicBlock *Dest)
173 : Value(Value), Dest(Dest) {}
174
operator <__anon122b05330111::ValueEqualityComparisonCase175 bool operator<(ValueEqualityComparisonCase RHS) const {
176 // Comparing pointers is ok as we only rely on the order for uniquing.
177 return Value < RHS.Value;
178 }
179
operator ==__anon122b05330111::ValueEqualityComparisonCase180 bool operator==(BasicBlock *RHSDest) const { return Dest == RHSDest; }
181 };
182
183 class SimplifyCFGOpt {
184 const TargetTransformInfo &TTI;
185 const DataLayout &DL;
186 SmallPtrSetImpl<BasicBlock *> *LoopHeaders;
187 const SimplifyCFGOptions &Options;
188 bool Resimplify;
189
190 Value *isValueEqualityComparison(Instruction *TI);
191 BasicBlock *GetValueEqualityComparisonCases(
192 Instruction *TI, std::vector<ValueEqualityComparisonCase> &Cases);
193 bool SimplifyEqualityComparisonWithOnlyPredecessor(Instruction *TI,
194 BasicBlock *Pred,
195 IRBuilder<> &Builder);
196 bool FoldValueComparisonIntoPredecessors(Instruction *TI,
197 IRBuilder<> &Builder);
198
199 bool simplifyReturn(ReturnInst *RI, IRBuilder<> &Builder);
200 bool simplifyResume(ResumeInst *RI, IRBuilder<> &Builder);
201 bool simplifySingleResume(ResumeInst *RI);
202 bool simplifyCommonResume(ResumeInst *RI);
203 bool simplifyCleanupReturn(CleanupReturnInst *RI);
204 bool simplifyUnreachable(UnreachableInst *UI);
205 bool simplifySwitch(SwitchInst *SI, IRBuilder<> &Builder);
206 bool simplifyIndirectBr(IndirectBrInst *IBI);
207 bool simplifyBranch(BranchInst *Branch, IRBuilder<> &Builder);
208 bool simplifyUncondBranch(BranchInst *BI, IRBuilder<> &Builder);
209 bool simplifyCondBranch(BranchInst *BI, IRBuilder<> &Builder);
210 bool SimplifyCondBranchToTwoReturns(BranchInst *BI, IRBuilder<> &Builder);
211
212 bool tryToSimplifyUncondBranchWithICmpInIt(ICmpInst *ICI,
213 IRBuilder<> &Builder);
214
215 bool HoistThenElseCodeToIf(BranchInst *BI, const TargetTransformInfo &TTI);
216 bool SpeculativelyExecuteBB(BranchInst *BI, BasicBlock *ThenBB,
217 const TargetTransformInfo &TTI);
218 bool SimplifyTerminatorOnSelect(Instruction *OldTerm, Value *Cond,
219 BasicBlock *TrueBB, BasicBlock *FalseBB,
220 uint32_t TrueWeight, uint32_t FalseWeight);
221 bool SimplifyBranchOnICmpChain(BranchInst *BI, IRBuilder<> &Builder,
222 const DataLayout &DL);
223 bool SimplifySwitchOnSelect(SwitchInst *SI, SelectInst *Select);
224 bool SimplifyIndirectBrOnSelect(IndirectBrInst *IBI, SelectInst *SI);
225 bool TurnSwitchRangeIntoICmp(SwitchInst *SI, IRBuilder<> &Builder);
226
227 public:
SimplifyCFGOpt(const TargetTransformInfo & TTI,const DataLayout & DL,SmallPtrSetImpl<BasicBlock * > * LoopHeaders,const SimplifyCFGOptions & Opts)228 SimplifyCFGOpt(const TargetTransformInfo &TTI, const DataLayout &DL,
229 SmallPtrSetImpl<BasicBlock *> *LoopHeaders,
230 const SimplifyCFGOptions &Opts)
231 : TTI(TTI), DL(DL), LoopHeaders(LoopHeaders), Options(Opts) {}
232
233 bool run(BasicBlock *BB);
234 bool simplifyOnce(BasicBlock *BB);
235
236 // Helper to set Resimplify and return change indication.
requestResimplify()237 bool requestResimplify() {
238 Resimplify = true;
239 return true;
240 }
241 };
242
243 } // end anonymous namespace
244
245 /// Return true if it is safe to merge these two
246 /// terminator instructions together.
247 static bool
SafeToMergeTerminators(Instruction * SI1,Instruction * SI2,SmallSetVector<BasicBlock *,4> * FailBlocks=nullptr)248 SafeToMergeTerminators(Instruction *SI1, Instruction *SI2,
249 SmallSetVector<BasicBlock *, 4> *FailBlocks = nullptr) {
250 if (SI1 == SI2)
251 return false; // Can't merge with self!
252
253 // It is not safe to merge these two switch instructions if they have a common
254 // successor, and if that successor has a PHI node, and if *that* PHI node has
255 // conflicting incoming values from the two switch blocks.
256 BasicBlock *SI1BB = SI1->getParent();
257 BasicBlock *SI2BB = SI2->getParent();
258
259 SmallPtrSet<BasicBlock *, 16> SI1Succs(succ_begin(SI1BB), succ_end(SI1BB));
260 bool Fail = false;
261 for (BasicBlock *Succ : successors(SI2BB))
262 if (SI1Succs.count(Succ))
263 for (BasicBlock::iterator BBI = Succ->begin(); isa<PHINode>(BBI); ++BBI) {
264 PHINode *PN = cast<PHINode>(BBI);
265 if (PN->getIncomingValueForBlock(SI1BB) !=
266 PN->getIncomingValueForBlock(SI2BB)) {
267 if (FailBlocks)
268 FailBlocks->insert(Succ);
269 Fail = true;
270 }
271 }
272
273 return !Fail;
274 }
275
276 /// Return true if it is safe and profitable to merge these two terminator
277 /// instructions together, where SI1 is an unconditional branch. PhiNodes will
278 /// store all PHI nodes in common successors.
279 static bool
isProfitableToFoldUnconditional(BranchInst * SI1,BranchInst * SI2,Instruction * Cond,SmallVectorImpl<PHINode * > & PhiNodes)280 isProfitableToFoldUnconditional(BranchInst *SI1, BranchInst *SI2,
281 Instruction *Cond,
282 SmallVectorImpl<PHINode *> &PhiNodes) {
283 if (SI1 == SI2)
284 return false; // Can't merge with self!
285 assert(SI1->isUnconditional() && SI2->isConditional());
286
287 // We fold the unconditional branch if we can easily update all PHI nodes in
288 // common successors:
289 // 1> We have a constant incoming value for the conditional branch;
290 // 2> We have "Cond" as the incoming value for the unconditional branch;
291 // 3> SI2->getCondition() and Cond have same operands.
292 CmpInst *Ci2 = dyn_cast<CmpInst>(SI2->getCondition());
293 if (!Ci2)
294 return false;
295 if (!(Cond->getOperand(0) == Ci2->getOperand(0) &&
296 Cond->getOperand(1) == Ci2->getOperand(1)) &&
297 !(Cond->getOperand(0) == Ci2->getOperand(1) &&
298 Cond->getOperand(1) == Ci2->getOperand(0)))
299 return false;
300
301 BasicBlock *SI1BB = SI1->getParent();
302 BasicBlock *SI2BB = SI2->getParent();
303 SmallPtrSet<BasicBlock *, 16> SI1Succs(succ_begin(SI1BB), succ_end(SI1BB));
304 for (BasicBlock *Succ : successors(SI2BB))
305 if (SI1Succs.count(Succ))
306 for (BasicBlock::iterator BBI = Succ->begin(); isa<PHINode>(BBI); ++BBI) {
307 PHINode *PN = cast<PHINode>(BBI);
308 if (PN->getIncomingValueForBlock(SI1BB) != Cond ||
309 !isa<ConstantInt>(PN->getIncomingValueForBlock(SI2BB)))
310 return false;
311 PhiNodes.push_back(PN);
312 }
313 return true;
314 }
315
316 /// Update PHI nodes in Succ to indicate that there will now be entries in it
317 /// from the 'NewPred' block. The values that will be flowing into the PHI nodes
318 /// will be the same as those coming in from ExistPred, an existing predecessor
319 /// of Succ.
AddPredecessorToBlock(BasicBlock * Succ,BasicBlock * NewPred,BasicBlock * ExistPred,MemorySSAUpdater * MSSAU=nullptr)320 static void AddPredecessorToBlock(BasicBlock *Succ, BasicBlock *NewPred,
321 BasicBlock *ExistPred,
322 MemorySSAUpdater *MSSAU = nullptr) {
323 for (PHINode &PN : Succ->phis())
324 PN.addIncoming(PN.getIncomingValueForBlock(ExistPred), NewPred);
325 if (MSSAU)
326 if (auto *MPhi = MSSAU->getMemorySSA()->getMemoryAccess(Succ))
327 MPhi->addIncoming(MPhi->getIncomingValueForBlock(ExistPred), NewPred);
328 }
329
330 /// Compute an abstract "cost" of speculating the given instruction,
331 /// which is assumed to be safe to speculate. TCC_Free means cheap,
332 /// TCC_Basic means less cheap, and TCC_Expensive means prohibitively
333 /// expensive.
ComputeSpeculationCost(const User * I,const TargetTransformInfo & TTI)334 static unsigned ComputeSpeculationCost(const User *I,
335 const TargetTransformInfo &TTI) {
336 assert(isSafeToSpeculativelyExecute(I) &&
337 "Instruction is not safe to speculatively execute!");
338 return TTI.getUserCost(I, TargetTransformInfo::TCK_SizeAndLatency);
339 }
340
341 /// If we have a merge point of an "if condition" as accepted above,
342 /// return true if the specified value dominates the block. We
343 /// don't handle the true generality of domination here, just a special case
344 /// which works well enough for us.
345 ///
346 /// If AggressiveInsts is non-null, and if V does not dominate BB, we check to
347 /// see if V (which must be an instruction) and its recursive operands
348 /// that do not dominate BB have a combined cost lower than CostRemaining and
349 /// are non-trapping. If both are true, the instruction is inserted into the
350 /// set and true is returned.
351 ///
352 /// The cost for most non-trapping instructions is defined as 1 except for
353 /// Select whose cost is 2.
354 ///
355 /// After this function returns, CostRemaining is decreased by the cost of
356 /// V plus its non-dominating operands. If that cost is greater than
357 /// CostRemaining, false is returned and CostRemaining is undefined.
DominatesMergePoint(Value * V,BasicBlock * BB,SmallPtrSetImpl<Instruction * > & AggressiveInsts,int & BudgetRemaining,const TargetTransformInfo & TTI,unsigned Depth=0)358 static bool DominatesMergePoint(Value *V, BasicBlock *BB,
359 SmallPtrSetImpl<Instruction *> &AggressiveInsts,
360 int &BudgetRemaining,
361 const TargetTransformInfo &TTI,
362 unsigned Depth = 0) {
363 // It is possible to hit a zero-cost cycle (phi/gep instructions for example),
364 // so limit the recursion depth.
365 // TODO: While this recursion limit does prevent pathological behavior, it
366 // would be better to track visited instructions to avoid cycles.
367 if (Depth == MaxSpeculationDepth)
368 return false;
369
370 Instruction *I = dyn_cast<Instruction>(V);
371 if (!I) {
372 // Non-instructions all dominate instructions, but not all constantexprs
373 // can be executed unconditionally.
374 if (ConstantExpr *C = dyn_cast<ConstantExpr>(V))
375 if (C->canTrap())
376 return false;
377 return true;
378 }
379 BasicBlock *PBB = I->getParent();
380
381 // We don't want to allow weird loops that might have the "if condition" in
382 // the bottom of this block.
383 if (PBB == BB)
384 return false;
385
386 // If this instruction is defined in a block that contains an unconditional
387 // branch to BB, then it must be in the 'conditional' part of the "if
388 // statement". If not, it definitely dominates the region.
389 BranchInst *BI = dyn_cast<BranchInst>(PBB->getTerminator());
390 if (!BI || BI->isConditional() || BI->getSuccessor(0) != BB)
391 return true;
392
393 // If we have seen this instruction before, don't count it again.
394 if (AggressiveInsts.count(I))
395 return true;
396
397 // Okay, it looks like the instruction IS in the "condition". Check to
398 // see if it's a cheap instruction to unconditionally compute, and if it
399 // only uses stuff defined outside of the condition. If so, hoist it out.
400 if (!isSafeToSpeculativelyExecute(I))
401 return false;
402
403 BudgetRemaining -= ComputeSpeculationCost(I, TTI);
404
405 // Allow exactly one instruction to be speculated regardless of its cost
406 // (as long as it is safe to do so).
407 // This is intended to flatten the CFG even if the instruction is a division
408 // or other expensive operation. The speculation of an expensive instruction
409 // is expected to be undone in CodeGenPrepare if the speculation has not
410 // enabled further IR optimizations.
411 if (BudgetRemaining < 0 &&
412 (!SpeculateOneExpensiveInst || !AggressiveInsts.empty() || Depth > 0))
413 return false;
414
415 // Okay, we can only really hoist these out if their operands do
416 // not take us over the cost threshold.
417 for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i)
418 if (!DominatesMergePoint(*i, BB, AggressiveInsts, BudgetRemaining, TTI,
419 Depth + 1))
420 return false;
421 // Okay, it's safe to do this! Remember this instruction.
422 AggressiveInsts.insert(I);
423 return true;
424 }
425
426 /// Extract ConstantInt from value, looking through IntToPtr
427 /// and PointerNullValue. Return NULL if value is not a constant int.
GetConstantInt(Value * V,const DataLayout & DL)428 static ConstantInt *GetConstantInt(Value *V, const DataLayout &DL) {
429 // Normal constant int.
430 ConstantInt *CI = dyn_cast<ConstantInt>(V);
431 if (CI || !isa<Constant>(V) || !V->getType()->isPointerTy())
432 return CI;
433
434 // This is some kind of pointer constant. Turn it into a pointer-sized
435 // ConstantInt if possible.
436 IntegerType *PtrTy = cast<IntegerType>(DL.getIntPtrType(V->getType()));
437
438 // Null pointer means 0, see SelectionDAGBuilder::getValue(const Value*).
439 if (isa<ConstantPointerNull>(V))
440 return ConstantInt::get(PtrTy, 0);
441
442 // IntToPtr const int.
443 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
444 if (CE->getOpcode() == Instruction::IntToPtr)
445 if (ConstantInt *CI = dyn_cast<ConstantInt>(CE->getOperand(0))) {
446 // The constant is very likely to have the right type already.
447 if (CI->getType() == PtrTy)
448 return CI;
449 else
450 return cast<ConstantInt>(
451 ConstantExpr::getIntegerCast(CI, PtrTy, /*isSigned=*/false));
452 }
453 return nullptr;
454 }
455
456 namespace {
457
458 /// Given a chain of or (||) or and (&&) comparison of a value against a
459 /// constant, this will try to recover the information required for a switch
460 /// structure.
461 /// It will depth-first traverse the chain of comparison, seeking for patterns
462 /// like %a == 12 or %a < 4 and combine them to produce a set of integer
463 /// representing the different cases for the switch.
464 /// Note that if the chain is composed of '||' it will build the set of elements
465 /// that matches the comparisons (i.e. any of this value validate the chain)
466 /// while for a chain of '&&' it will build the set elements that make the test
467 /// fail.
468 struct ConstantComparesGatherer {
469 const DataLayout &DL;
470
471 /// Value found for the switch comparison
472 Value *CompValue = nullptr;
473
474 /// Extra clause to be checked before the switch
475 Value *Extra = nullptr;
476
477 /// Set of integers to match in switch
478 SmallVector<ConstantInt *, 8> Vals;
479
480 /// Number of comparisons matched in the and/or chain
481 unsigned UsedICmps = 0;
482
483 /// Construct and compute the result for the comparison instruction Cond
ConstantComparesGatherer__anon122b05330211::ConstantComparesGatherer484 ConstantComparesGatherer(Instruction *Cond, const DataLayout &DL) : DL(DL) {
485 gather(Cond);
486 }
487
488 ConstantComparesGatherer(const ConstantComparesGatherer &) = delete;
489 ConstantComparesGatherer &
490 operator=(const ConstantComparesGatherer &) = delete;
491
492 private:
493 /// Try to set the current value used for the comparison, it succeeds only if
494 /// it wasn't set before or if the new value is the same as the old one
setValueOnce__anon122b05330211::ConstantComparesGatherer495 bool setValueOnce(Value *NewVal) {
496 if (CompValue && CompValue != NewVal)
497 return false;
498 CompValue = NewVal;
499 return (CompValue != nullptr);
500 }
501
502 /// Try to match Instruction "I" as a comparison against a constant and
503 /// populates the array Vals with the set of values that match (or do not
504 /// match depending on isEQ).
505 /// Return false on failure. On success, the Value the comparison matched
506 /// against is placed in CompValue.
507 /// If CompValue is already set, the function is expected to fail if a match
508 /// is found but the value compared to is different.
matchInstruction__anon122b05330211::ConstantComparesGatherer509 bool matchInstruction(Instruction *I, bool isEQ) {
510 // If this is an icmp against a constant, handle this as one of the cases.
511 ICmpInst *ICI;
512 ConstantInt *C;
513 if (!((ICI = dyn_cast<ICmpInst>(I)) &&
514 (C = GetConstantInt(I->getOperand(1), DL)))) {
515 return false;
516 }
517
518 Value *RHSVal;
519 const APInt *RHSC;
520
521 // Pattern match a special case
522 // (x & ~2^z) == y --> x == y || x == y|2^z
523 // This undoes a transformation done by instcombine to fuse 2 compares.
524 if (ICI->getPredicate() == (isEQ ? ICmpInst::ICMP_EQ : ICmpInst::ICMP_NE)) {
525 // It's a little bit hard to see why the following transformations are
526 // correct. Here is a CVC3 program to verify them for 64-bit values:
527
528 /*
529 ONE : BITVECTOR(64) = BVZEROEXTEND(0bin1, 63);
530 x : BITVECTOR(64);
531 y : BITVECTOR(64);
532 z : BITVECTOR(64);
533 mask : BITVECTOR(64) = BVSHL(ONE, z);
534 QUERY( (y & ~mask = y) =>
535 ((x & ~mask = y) <=> (x = y OR x = (y | mask)))
536 );
537 QUERY( (y | mask = y) =>
538 ((x | mask = y) <=> (x = y OR x = (y & ~mask)))
539 );
540 */
541
542 // Please note that each pattern must be a dual implication (<--> or
543 // iff). One directional implication can create spurious matches. If the
544 // implication is only one-way, an unsatisfiable condition on the left
545 // side can imply a satisfiable condition on the right side. Dual
546 // implication ensures that satisfiable conditions are transformed to
547 // other satisfiable conditions and unsatisfiable conditions are
548 // transformed to other unsatisfiable conditions.
549
550 // Here is a concrete example of a unsatisfiable condition on the left
551 // implying a satisfiable condition on the right:
552 //
553 // mask = (1 << z)
554 // (x & ~mask) == y --> (x == y || x == (y | mask))
555 //
556 // Substituting y = 3, z = 0 yields:
557 // (x & -2) == 3 --> (x == 3 || x == 2)
558
559 // Pattern match a special case:
560 /*
561 QUERY( (y & ~mask = y) =>
562 ((x & ~mask = y) <=> (x = y OR x = (y | mask)))
563 );
564 */
565 if (match(ICI->getOperand(0),
566 m_And(m_Value(RHSVal), m_APInt(RHSC)))) {
567 APInt Mask = ~*RHSC;
568 if (Mask.isPowerOf2() && (C->getValue() & ~Mask) == C->getValue()) {
569 // If we already have a value for the switch, it has to match!
570 if (!setValueOnce(RHSVal))
571 return false;
572
573 Vals.push_back(C);
574 Vals.push_back(
575 ConstantInt::get(C->getContext(),
576 C->getValue() | Mask));
577 UsedICmps++;
578 return true;
579 }
580 }
581
582 // Pattern match a special case:
583 /*
584 QUERY( (y | mask = y) =>
585 ((x | mask = y) <=> (x = y OR x = (y & ~mask)))
586 );
587 */
588 if (match(ICI->getOperand(0),
589 m_Or(m_Value(RHSVal), m_APInt(RHSC)))) {
590 APInt Mask = *RHSC;
591 if (Mask.isPowerOf2() && (C->getValue() | Mask) == C->getValue()) {
592 // If we already have a value for the switch, it has to match!
593 if (!setValueOnce(RHSVal))
594 return false;
595
596 Vals.push_back(C);
597 Vals.push_back(ConstantInt::get(C->getContext(),
598 C->getValue() & ~Mask));
599 UsedICmps++;
600 return true;
601 }
602 }
603
604 // If we already have a value for the switch, it has to match!
605 if (!setValueOnce(ICI->getOperand(0)))
606 return false;
607
608 UsedICmps++;
609 Vals.push_back(C);
610 return ICI->getOperand(0);
611 }
612
613 // If we have "x ult 3", for example, then we can add 0,1,2 to the set.
614 ConstantRange Span = ConstantRange::makeAllowedICmpRegion(
615 ICI->getPredicate(), C->getValue());
616
617 // Shift the range if the compare is fed by an add. This is the range
618 // compare idiom as emitted by instcombine.
619 Value *CandidateVal = I->getOperand(0);
620 if (match(I->getOperand(0), m_Add(m_Value(RHSVal), m_APInt(RHSC)))) {
621 Span = Span.subtract(*RHSC);
622 CandidateVal = RHSVal;
623 }
624
625 // If this is an and/!= check, then we are looking to build the set of
626 // value that *don't* pass the and chain. I.e. to turn "x ugt 2" into
627 // x != 0 && x != 1.
628 if (!isEQ)
629 Span = Span.inverse();
630
631 // If there are a ton of values, we don't want to make a ginormous switch.
632 if (Span.isSizeLargerThan(8) || Span.isEmptySet()) {
633 return false;
634 }
635
636 // If we already have a value for the switch, it has to match!
637 if (!setValueOnce(CandidateVal))
638 return false;
639
640 // Add all values from the range to the set
641 for (APInt Tmp = Span.getLower(); Tmp != Span.getUpper(); ++Tmp)
642 Vals.push_back(ConstantInt::get(I->getContext(), Tmp));
643
644 UsedICmps++;
645 return true;
646 }
647
648 /// Given a potentially 'or'd or 'and'd together collection of icmp
649 /// eq/ne/lt/gt instructions that compare a value against a constant, extract
650 /// the value being compared, and stick the list constants into the Vals
651 /// vector.
652 /// One "Extra" case is allowed to differ from the other.
gather__anon122b05330211::ConstantComparesGatherer653 void gather(Value *V) {
654 bool isEQ = (cast<Instruction>(V)->getOpcode() == Instruction::Or);
655
656 // Keep a stack (SmallVector for efficiency) for depth-first traversal
657 SmallVector<Value *, 8> DFT;
658 SmallPtrSet<Value *, 8> Visited;
659
660 // Initialize
661 Visited.insert(V);
662 DFT.push_back(V);
663
664 while (!DFT.empty()) {
665 V = DFT.pop_back_val();
666
667 if (Instruction *I = dyn_cast<Instruction>(V)) {
668 // If it is a || (or && depending on isEQ), process the operands.
669 if (I->getOpcode() == (isEQ ? Instruction::Or : Instruction::And)) {
670 if (Visited.insert(I->getOperand(1)).second)
671 DFT.push_back(I->getOperand(1));
672 if (Visited.insert(I->getOperand(0)).second)
673 DFT.push_back(I->getOperand(0));
674 continue;
675 }
676
677 // Try to match the current instruction
678 if (matchInstruction(I, isEQ))
679 // Match succeed, continue the loop
680 continue;
681 }
682
683 // One element of the sequence of || (or &&) could not be match as a
684 // comparison against the same value as the others.
685 // We allow only one "Extra" case to be checked before the switch
686 if (!Extra) {
687 Extra = V;
688 continue;
689 }
690 // Failed to parse a proper sequence, abort now
691 CompValue = nullptr;
692 break;
693 }
694 }
695 };
696
697 } // end anonymous namespace
698
EraseTerminatorAndDCECond(Instruction * TI,MemorySSAUpdater * MSSAU=nullptr)699 static void EraseTerminatorAndDCECond(Instruction *TI,
700 MemorySSAUpdater *MSSAU = nullptr) {
701 Instruction *Cond = nullptr;
702 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
703 Cond = dyn_cast<Instruction>(SI->getCondition());
704 } else if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
705 if (BI->isConditional())
706 Cond = dyn_cast<Instruction>(BI->getCondition());
707 } else if (IndirectBrInst *IBI = dyn_cast<IndirectBrInst>(TI)) {
708 Cond = dyn_cast<Instruction>(IBI->getAddress());
709 }
710
711 TI->eraseFromParent();
712 if (Cond)
713 RecursivelyDeleteTriviallyDeadInstructions(Cond, nullptr, MSSAU);
714 }
715
716 /// Return true if the specified terminator checks
717 /// to see if a value is equal to constant integer value.
isValueEqualityComparison(Instruction * TI)718 Value *SimplifyCFGOpt::isValueEqualityComparison(Instruction *TI) {
719 Value *CV = nullptr;
720 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
721 // Do not permit merging of large switch instructions into their
722 // predecessors unless there is only one predecessor.
723 if (!SI->getParent()->hasNPredecessorsOrMore(128 / SI->getNumSuccessors()))
724 CV = SI->getCondition();
725 } else if (BranchInst *BI = dyn_cast<BranchInst>(TI))
726 if (BI->isConditional() && BI->getCondition()->hasOneUse())
727 if (ICmpInst *ICI = dyn_cast<ICmpInst>(BI->getCondition())) {
728 if (ICI->isEquality() && GetConstantInt(ICI->getOperand(1), DL))
729 CV = ICI->getOperand(0);
730 }
731
732 // Unwrap any lossless ptrtoint cast.
733 if (CV) {
734 if (PtrToIntInst *PTII = dyn_cast<PtrToIntInst>(CV)) {
735 Value *Ptr = PTII->getPointerOperand();
736 if (PTII->getType() == DL.getIntPtrType(Ptr->getType()))
737 CV = Ptr;
738 }
739 }
740 return CV;
741 }
742
743 /// Given a value comparison instruction,
744 /// decode all of the 'cases' that it represents and return the 'default' block.
GetValueEqualityComparisonCases(Instruction * TI,std::vector<ValueEqualityComparisonCase> & Cases)745 BasicBlock *SimplifyCFGOpt::GetValueEqualityComparisonCases(
746 Instruction *TI, std::vector<ValueEqualityComparisonCase> &Cases) {
747 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
748 Cases.reserve(SI->getNumCases());
749 for (auto Case : SI->cases())
750 Cases.push_back(ValueEqualityComparisonCase(Case.getCaseValue(),
751 Case.getCaseSuccessor()));
752 return SI->getDefaultDest();
753 }
754
755 BranchInst *BI = cast<BranchInst>(TI);
756 ICmpInst *ICI = cast<ICmpInst>(BI->getCondition());
757 BasicBlock *Succ = BI->getSuccessor(ICI->getPredicate() == ICmpInst::ICMP_NE);
758 Cases.push_back(ValueEqualityComparisonCase(
759 GetConstantInt(ICI->getOperand(1), DL), Succ));
760 return BI->getSuccessor(ICI->getPredicate() == ICmpInst::ICMP_EQ);
761 }
762
763 /// Given a vector of bb/value pairs, remove any entries
764 /// in the list that match the specified block.
765 static void
EliminateBlockCases(BasicBlock * BB,std::vector<ValueEqualityComparisonCase> & Cases)766 EliminateBlockCases(BasicBlock *BB,
767 std::vector<ValueEqualityComparisonCase> &Cases) {
768 Cases.erase(std::remove(Cases.begin(), Cases.end(), BB), Cases.end());
769 }
770
771 /// Return true if there are any keys in C1 that exist in C2 as well.
ValuesOverlap(std::vector<ValueEqualityComparisonCase> & C1,std::vector<ValueEqualityComparisonCase> & C2)772 static bool ValuesOverlap(std::vector<ValueEqualityComparisonCase> &C1,
773 std::vector<ValueEqualityComparisonCase> &C2) {
774 std::vector<ValueEqualityComparisonCase> *V1 = &C1, *V2 = &C2;
775
776 // Make V1 be smaller than V2.
777 if (V1->size() > V2->size())
778 std::swap(V1, V2);
779
780 if (V1->empty())
781 return false;
782 if (V1->size() == 1) {
783 // Just scan V2.
784 ConstantInt *TheVal = (*V1)[0].Value;
785 for (unsigned i = 0, e = V2->size(); i != e; ++i)
786 if (TheVal == (*V2)[i].Value)
787 return true;
788 }
789
790 // Otherwise, just sort both lists and compare element by element.
791 array_pod_sort(V1->begin(), V1->end());
792 array_pod_sort(V2->begin(), V2->end());
793 unsigned i1 = 0, i2 = 0, e1 = V1->size(), e2 = V2->size();
794 while (i1 != e1 && i2 != e2) {
795 if ((*V1)[i1].Value == (*V2)[i2].Value)
796 return true;
797 if ((*V1)[i1].Value < (*V2)[i2].Value)
798 ++i1;
799 else
800 ++i2;
801 }
802 return false;
803 }
804
805 // Set branch weights on SwitchInst. This sets the metadata if there is at
806 // least one non-zero weight.
setBranchWeights(SwitchInst * SI,ArrayRef<uint32_t> Weights)807 static void setBranchWeights(SwitchInst *SI, ArrayRef<uint32_t> Weights) {
808 // Check that there is at least one non-zero weight. Otherwise, pass
809 // nullptr to setMetadata which will erase the existing metadata.
810 MDNode *N = nullptr;
811 if (llvm::any_of(Weights, [](uint32_t W) { return W != 0; }))
812 N = MDBuilder(SI->getParent()->getContext()).createBranchWeights(Weights);
813 SI->setMetadata(LLVMContext::MD_prof, N);
814 }
815
816 // Similar to the above, but for branch and select instructions that take
817 // exactly 2 weights.
setBranchWeights(Instruction * I,uint32_t TrueWeight,uint32_t FalseWeight)818 static void setBranchWeights(Instruction *I, uint32_t TrueWeight,
819 uint32_t FalseWeight) {
820 assert(isa<BranchInst>(I) || isa<SelectInst>(I));
821 // Check that there is at least one non-zero weight. Otherwise, pass
822 // nullptr to setMetadata which will erase the existing metadata.
823 MDNode *N = nullptr;
824 if (TrueWeight || FalseWeight)
825 N = MDBuilder(I->getParent()->getContext())
826 .createBranchWeights(TrueWeight, FalseWeight);
827 I->setMetadata(LLVMContext::MD_prof, N);
828 }
829
830 /// If TI is known to be a terminator instruction and its block is known to
831 /// only have a single predecessor block, check to see if that predecessor is
832 /// also a value comparison with the same value, and if that comparison
833 /// determines the outcome of this comparison. If so, simplify TI. This does a
834 /// very limited form of jump threading.
SimplifyEqualityComparisonWithOnlyPredecessor(Instruction * TI,BasicBlock * Pred,IRBuilder<> & Builder)835 bool SimplifyCFGOpt::SimplifyEqualityComparisonWithOnlyPredecessor(
836 Instruction *TI, BasicBlock *Pred, IRBuilder<> &Builder) {
837 Value *PredVal = isValueEqualityComparison(Pred->getTerminator());
838 if (!PredVal)
839 return false; // Not a value comparison in predecessor.
840
841 Value *ThisVal = isValueEqualityComparison(TI);
842 assert(ThisVal && "This isn't a value comparison!!");
843 if (ThisVal != PredVal)
844 return false; // Different predicates.
845
846 // TODO: Preserve branch weight metadata, similarly to how
847 // FoldValueComparisonIntoPredecessors preserves it.
848
849 // Find out information about when control will move from Pred to TI's block.
850 std::vector<ValueEqualityComparisonCase> PredCases;
851 BasicBlock *PredDef =
852 GetValueEqualityComparisonCases(Pred->getTerminator(), PredCases);
853 EliminateBlockCases(PredDef, PredCases); // Remove default from cases.
854
855 // Find information about how control leaves this block.
856 std::vector<ValueEqualityComparisonCase> ThisCases;
857 BasicBlock *ThisDef = GetValueEqualityComparisonCases(TI, ThisCases);
858 EliminateBlockCases(ThisDef, ThisCases); // Remove default from cases.
859
860 // If TI's block is the default block from Pred's comparison, potentially
861 // simplify TI based on this knowledge.
862 if (PredDef == TI->getParent()) {
863 // If we are here, we know that the value is none of those cases listed in
864 // PredCases. If there are any cases in ThisCases that are in PredCases, we
865 // can simplify TI.
866 if (!ValuesOverlap(PredCases, ThisCases))
867 return false;
868
869 if (isa<BranchInst>(TI)) {
870 // Okay, one of the successors of this condbr is dead. Convert it to a
871 // uncond br.
872 assert(ThisCases.size() == 1 && "Branch can only have one case!");
873 // Insert the new branch.
874 Instruction *NI = Builder.CreateBr(ThisDef);
875 (void)NI;
876
877 // Remove PHI node entries for the dead edge.
878 ThisCases[0].Dest->removePredecessor(TI->getParent());
879
880 LLVM_DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator()
881 << "Through successor TI: " << *TI << "Leaving: " << *NI
882 << "\n");
883
884 EraseTerminatorAndDCECond(TI);
885 return true;
886 }
887
888 SwitchInstProfUpdateWrapper SI = *cast<SwitchInst>(TI);
889 // Okay, TI has cases that are statically dead, prune them away.
890 SmallPtrSet<Constant *, 16> DeadCases;
891 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
892 DeadCases.insert(PredCases[i].Value);
893
894 LLVM_DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator()
895 << "Through successor TI: " << *TI);
896
897 for (SwitchInst::CaseIt i = SI->case_end(), e = SI->case_begin(); i != e;) {
898 --i;
899 if (DeadCases.count(i->getCaseValue())) {
900 i->getCaseSuccessor()->removePredecessor(TI->getParent());
901 SI.removeCase(i);
902 }
903 }
904 LLVM_DEBUG(dbgs() << "Leaving: " << *TI << "\n");
905 return true;
906 }
907
908 // Otherwise, TI's block must correspond to some matched value. Find out
909 // which value (or set of values) this is.
910 ConstantInt *TIV = nullptr;
911 BasicBlock *TIBB = TI->getParent();
912 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
913 if (PredCases[i].Dest == TIBB) {
914 if (TIV)
915 return false; // Cannot handle multiple values coming to this block.
916 TIV = PredCases[i].Value;
917 }
918 assert(TIV && "No edge from pred to succ?");
919
920 // Okay, we found the one constant that our value can be if we get into TI's
921 // BB. Find out which successor will unconditionally be branched to.
922 BasicBlock *TheRealDest = nullptr;
923 for (unsigned i = 0, e = ThisCases.size(); i != e; ++i)
924 if (ThisCases[i].Value == TIV) {
925 TheRealDest = ThisCases[i].Dest;
926 break;
927 }
928
929 // If not handled by any explicit cases, it is handled by the default case.
930 if (!TheRealDest)
931 TheRealDest = ThisDef;
932
933 // Remove PHI node entries for dead edges.
934 BasicBlock *CheckEdge = TheRealDest;
935 for (BasicBlock *Succ : successors(TIBB))
936 if (Succ != CheckEdge)
937 Succ->removePredecessor(TIBB);
938 else
939 CheckEdge = nullptr;
940
941 // Insert the new branch.
942 Instruction *NI = Builder.CreateBr(TheRealDest);
943 (void)NI;
944
945 LLVM_DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator()
946 << "Through successor TI: " << *TI << "Leaving: " << *NI
947 << "\n");
948
949 EraseTerminatorAndDCECond(TI);
950 return true;
951 }
952
953 namespace {
954
955 /// This class implements a stable ordering of constant
956 /// integers that does not depend on their address. This is important for
957 /// applications that sort ConstantInt's to ensure uniqueness.
958 struct ConstantIntOrdering {
operator ()__anon122b05330411::ConstantIntOrdering959 bool operator()(const ConstantInt *LHS, const ConstantInt *RHS) const {
960 return LHS->getValue().ult(RHS->getValue());
961 }
962 };
963
964 } // end anonymous namespace
965
ConstantIntSortPredicate(ConstantInt * const * P1,ConstantInt * const * P2)966 static int ConstantIntSortPredicate(ConstantInt *const *P1,
967 ConstantInt *const *P2) {
968 const ConstantInt *LHS = *P1;
969 const ConstantInt *RHS = *P2;
970 if (LHS == RHS)
971 return 0;
972 return LHS->getValue().ult(RHS->getValue()) ? 1 : -1;
973 }
974
HasBranchWeights(const Instruction * I)975 static inline bool HasBranchWeights(const Instruction *I) {
976 MDNode *ProfMD = I->getMetadata(LLVMContext::MD_prof);
977 if (ProfMD && ProfMD->getOperand(0))
978 if (MDString *MDS = dyn_cast<MDString>(ProfMD->getOperand(0)))
979 return MDS->getString().equals("branch_weights");
980
981 return false;
982 }
983
984 /// Get Weights of a given terminator, the default weight is at the front
985 /// of the vector. If TI is a conditional eq, we need to swap the branch-weight
986 /// metadata.
GetBranchWeights(Instruction * TI,SmallVectorImpl<uint64_t> & Weights)987 static void GetBranchWeights(Instruction *TI,
988 SmallVectorImpl<uint64_t> &Weights) {
989 MDNode *MD = TI->getMetadata(LLVMContext::MD_prof);
990 assert(MD);
991 for (unsigned i = 1, e = MD->getNumOperands(); i < e; ++i) {
992 ConstantInt *CI = mdconst::extract<ConstantInt>(MD->getOperand(i));
993 Weights.push_back(CI->getValue().getZExtValue());
994 }
995
996 // If TI is a conditional eq, the default case is the false case,
997 // and the corresponding branch-weight data is at index 2. We swap the
998 // default weight to be the first entry.
999 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
1000 assert(Weights.size() == 2);
1001 ICmpInst *ICI = cast<ICmpInst>(BI->getCondition());
1002 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
1003 std::swap(Weights.front(), Weights.back());
1004 }
1005 }
1006
1007 /// Keep halving the weights until all can fit in uint32_t.
FitWeights(MutableArrayRef<uint64_t> Weights)1008 static void FitWeights(MutableArrayRef<uint64_t> Weights) {
1009 uint64_t Max = *std::max_element(Weights.begin(), Weights.end());
1010 if (Max > UINT_MAX) {
1011 unsigned Offset = 32 - countLeadingZeros(Max);
1012 for (uint64_t &I : Weights)
1013 I >>= Offset;
1014 }
1015 }
1016
1017 /// The specified terminator is a value equality comparison instruction
1018 /// (either a switch or a branch on "X == c").
1019 /// See if any of the predecessors of the terminator block are value comparisons
1020 /// on the same value. If so, and if safe to do so, fold them together.
FoldValueComparisonIntoPredecessors(Instruction * TI,IRBuilder<> & Builder)1021 bool SimplifyCFGOpt::FoldValueComparisonIntoPredecessors(Instruction *TI,
1022 IRBuilder<> &Builder) {
1023 BasicBlock *BB = TI->getParent();
1024 Value *CV = isValueEqualityComparison(TI); // CondVal
1025 assert(CV && "Not a comparison?");
1026 bool Changed = false;
1027
1028 SmallVector<BasicBlock *, 16> Preds(pred_begin(BB), pred_end(BB));
1029 while (!Preds.empty()) {
1030 BasicBlock *Pred = Preds.pop_back_val();
1031
1032 // See if the predecessor is a comparison with the same value.
1033 Instruction *PTI = Pred->getTerminator();
1034 Value *PCV = isValueEqualityComparison(PTI); // PredCondVal
1035
1036 if (PCV == CV && TI != PTI) {
1037 SmallSetVector<BasicBlock*, 4> FailBlocks;
1038 if (!SafeToMergeTerminators(TI, PTI, &FailBlocks)) {
1039 for (auto *Succ : FailBlocks) {
1040 if (!SplitBlockPredecessors(Succ, TI->getParent(), ".fold.split"))
1041 return false;
1042 }
1043 }
1044
1045 // Figure out which 'cases' to copy from SI to PSI.
1046 std::vector<ValueEqualityComparisonCase> BBCases;
1047 BasicBlock *BBDefault = GetValueEqualityComparisonCases(TI, BBCases);
1048
1049 std::vector<ValueEqualityComparisonCase> PredCases;
1050 BasicBlock *PredDefault = GetValueEqualityComparisonCases(PTI, PredCases);
1051
1052 // Based on whether the default edge from PTI goes to BB or not, fill in
1053 // PredCases and PredDefault with the new switch cases we would like to
1054 // build.
1055 SmallVector<BasicBlock *, 8> NewSuccessors;
1056
1057 // Update the branch weight metadata along the way
1058 SmallVector<uint64_t, 8> Weights;
1059 bool PredHasWeights = HasBranchWeights(PTI);
1060 bool SuccHasWeights = HasBranchWeights(TI);
1061
1062 if (PredHasWeights) {
1063 GetBranchWeights(PTI, Weights);
1064 // branch-weight metadata is inconsistent here.
1065 if (Weights.size() != 1 + PredCases.size())
1066 PredHasWeights = SuccHasWeights = false;
1067 } else if (SuccHasWeights)
1068 // If there are no predecessor weights but there are successor weights,
1069 // populate Weights with 1, which will later be scaled to the sum of
1070 // successor's weights
1071 Weights.assign(1 + PredCases.size(), 1);
1072
1073 SmallVector<uint64_t, 8> SuccWeights;
1074 if (SuccHasWeights) {
1075 GetBranchWeights(TI, SuccWeights);
1076 // branch-weight metadata is inconsistent here.
1077 if (SuccWeights.size() != 1 + BBCases.size())
1078 PredHasWeights = SuccHasWeights = false;
1079 } else if (PredHasWeights)
1080 SuccWeights.assign(1 + BBCases.size(), 1);
1081
1082 if (PredDefault == BB) {
1083 // If this is the default destination from PTI, only the edges in TI
1084 // that don't occur in PTI, or that branch to BB will be activated.
1085 std::set<ConstantInt *, ConstantIntOrdering> PTIHandled;
1086 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
1087 if (PredCases[i].Dest != BB)
1088 PTIHandled.insert(PredCases[i].Value);
1089 else {
1090 // The default destination is BB, we don't need explicit targets.
1091 std::swap(PredCases[i], PredCases.back());
1092
1093 if (PredHasWeights || SuccHasWeights) {
1094 // Increase weight for the default case.
1095 Weights[0] += Weights[i + 1];
1096 std::swap(Weights[i + 1], Weights.back());
1097 Weights.pop_back();
1098 }
1099
1100 PredCases.pop_back();
1101 --i;
1102 --e;
1103 }
1104
1105 // Reconstruct the new switch statement we will be building.
1106 if (PredDefault != BBDefault) {
1107 PredDefault->removePredecessor(Pred);
1108 PredDefault = BBDefault;
1109 NewSuccessors.push_back(BBDefault);
1110 }
1111
1112 unsigned CasesFromPred = Weights.size();
1113 uint64_t ValidTotalSuccWeight = 0;
1114 for (unsigned i = 0, e = BBCases.size(); i != e; ++i)
1115 if (!PTIHandled.count(BBCases[i].Value) &&
1116 BBCases[i].Dest != BBDefault) {
1117 PredCases.push_back(BBCases[i]);
1118 NewSuccessors.push_back(BBCases[i].Dest);
1119 if (SuccHasWeights || PredHasWeights) {
1120 // The default weight is at index 0, so weight for the ith case
1121 // should be at index i+1. Scale the cases from successor by
1122 // PredDefaultWeight (Weights[0]).
1123 Weights.push_back(Weights[0] * SuccWeights[i + 1]);
1124 ValidTotalSuccWeight += SuccWeights[i + 1];
1125 }
1126 }
1127
1128 if (SuccHasWeights || PredHasWeights) {
1129 ValidTotalSuccWeight += SuccWeights[0];
1130 // Scale the cases from predecessor by ValidTotalSuccWeight.
1131 for (unsigned i = 1; i < CasesFromPred; ++i)
1132 Weights[i] *= ValidTotalSuccWeight;
1133 // Scale the default weight by SuccDefaultWeight (SuccWeights[0]).
1134 Weights[0] *= SuccWeights[0];
1135 }
1136 } else {
1137 // If this is not the default destination from PSI, only the edges
1138 // in SI that occur in PSI with a destination of BB will be
1139 // activated.
1140 std::set<ConstantInt *, ConstantIntOrdering> PTIHandled;
1141 std::map<ConstantInt *, uint64_t> WeightsForHandled;
1142 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
1143 if (PredCases[i].Dest == BB) {
1144 PTIHandled.insert(PredCases[i].Value);
1145
1146 if (PredHasWeights || SuccHasWeights) {
1147 WeightsForHandled[PredCases[i].Value] = Weights[i + 1];
1148 std::swap(Weights[i + 1], Weights.back());
1149 Weights.pop_back();
1150 }
1151
1152 std::swap(PredCases[i], PredCases.back());
1153 PredCases.pop_back();
1154 --i;
1155 --e;
1156 }
1157
1158 // Okay, now we know which constants were sent to BB from the
1159 // predecessor. Figure out where they will all go now.
1160 for (unsigned i = 0, e = BBCases.size(); i != e; ++i)
1161 if (PTIHandled.count(BBCases[i].Value)) {
1162 // If this is one we are capable of getting...
1163 if (PredHasWeights || SuccHasWeights)
1164 Weights.push_back(WeightsForHandled[BBCases[i].Value]);
1165 PredCases.push_back(BBCases[i]);
1166 NewSuccessors.push_back(BBCases[i].Dest);
1167 PTIHandled.erase(
1168 BBCases[i].Value); // This constant is taken care of
1169 }
1170
1171 // If there are any constants vectored to BB that TI doesn't handle,
1172 // they must go to the default destination of TI.
1173 for (ConstantInt *I : PTIHandled) {
1174 if (PredHasWeights || SuccHasWeights)
1175 Weights.push_back(WeightsForHandled[I]);
1176 PredCases.push_back(ValueEqualityComparisonCase(I, BBDefault));
1177 NewSuccessors.push_back(BBDefault);
1178 }
1179 }
1180
1181 // Okay, at this point, we know which new successor Pred will get. Make
1182 // sure we update the number of entries in the PHI nodes for these
1183 // successors.
1184 for (BasicBlock *NewSuccessor : NewSuccessors)
1185 AddPredecessorToBlock(NewSuccessor, Pred, BB);
1186
1187 Builder.SetInsertPoint(PTI);
1188 // Convert pointer to int before we switch.
1189 if (CV->getType()->isPointerTy()) {
1190 CV = Builder.CreatePtrToInt(CV, DL.getIntPtrType(CV->getType()),
1191 "magicptr");
1192 }
1193
1194 // Now that the successors are updated, create the new Switch instruction.
1195 SwitchInst *NewSI =
1196 Builder.CreateSwitch(CV, PredDefault, PredCases.size());
1197 NewSI->setDebugLoc(PTI->getDebugLoc());
1198 for (ValueEqualityComparisonCase &V : PredCases)
1199 NewSI->addCase(V.Value, V.Dest);
1200
1201 if (PredHasWeights || SuccHasWeights) {
1202 // Halve the weights if any of them cannot fit in an uint32_t
1203 FitWeights(Weights);
1204
1205 SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
1206
1207 setBranchWeights(NewSI, MDWeights);
1208 }
1209
1210 EraseTerminatorAndDCECond(PTI);
1211
1212 // Okay, last check. If BB is still a successor of PSI, then we must
1213 // have an infinite loop case. If so, add an infinitely looping block
1214 // to handle the case to preserve the behavior of the code.
1215 BasicBlock *InfLoopBlock = nullptr;
1216 for (unsigned i = 0, e = NewSI->getNumSuccessors(); i != e; ++i)
1217 if (NewSI->getSuccessor(i) == BB) {
1218 if (!InfLoopBlock) {
1219 // Insert it at the end of the function, because it's either code,
1220 // or it won't matter if it's hot. :)
1221 InfLoopBlock = BasicBlock::Create(BB->getContext(), "infloop",
1222 BB->getParent());
1223 BranchInst::Create(InfLoopBlock, InfLoopBlock);
1224 }
1225 NewSI->setSuccessor(i, InfLoopBlock);
1226 }
1227
1228 Changed = true;
1229 }
1230 }
1231 return Changed;
1232 }
1233
1234 // If we would need to insert a select that uses the value of this invoke
1235 // (comments in HoistThenElseCodeToIf explain why we would need to do this), we
1236 // 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)1237 static bool isSafeToHoistInvoke(BasicBlock *BB1, BasicBlock *BB2,
1238 Instruction *I1, Instruction *I2) {
1239 for (BasicBlock *Succ : successors(BB1)) {
1240 for (const PHINode &PN : Succ->phis()) {
1241 Value *BB1V = PN.getIncomingValueForBlock(BB1);
1242 Value *BB2V = PN.getIncomingValueForBlock(BB2);
1243 if (BB1V != BB2V && (BB1V == I1 || BB2V == I2)) {
1244 return false;
1245 }
1246 }
1247 }
1248 return true;
1249 }
1250
1251 static bool passingValueIsAlwaysUndefined(Value *V, Instruction *I);
1252
1253 /// Given a conditional branch that goes to BB1 and BB2, hoist any common code
1254 /// in the two blocks up into the branch block. The caller of this function
1255 /// guarantees that BI's block dominates BB1 and BB2.
HoistThenElseCodeToIf(BranchInst * BI,const TargetTransformInfo & TTI)1256 bool SimplifyCFGOpt::HoistThenElseCodeToIf(BranchInst *BI,
1257 const TargetTransformInfo &TTI) {
1258 // This does very trivial matching, with limited scanning, to find identical
1259 // instructions in the two blocks. In particular, we don't want to get into
1260 // O(M*N) situations here where M and N are the sizes of BB1 and BB2. As
1261 // such, we currently just scan for obviously identical instructions in an
1262 // identical order.
1263 BasicBlock *BB1 = BI->getSuccessor(0); // The true destination.
1264 BasicBlock *BB2 = BI->getSuccessor(1); // The false destination
1265
1266 BasicBlock::iterator BB1_Itr = BB1->begin();
1267 BasicBlock::iterator BB2_Itr = BB2->begin();
1268
1269 Instruction *I1 = &*BB1_Itr++, *I2 = &*BB2_Itr++;
1270 // Skip debug info if it is not identical.
1271 DbgInfoIntrinsic *DBI1 = dyn_cast<DbgInfoIntrinsic>(I1);
1272 DbgInfoIntrinsic *DBI2 = dyn_cast<DbgInfoIntrinsic>(I2);
1273 if (!DBI1 || !DBI2 || !DBI1->isIdenticalToWhenDefined(DBI2)) {
1274 while (isa<DbgInfoIntrinsic>(I1))
1275 I1 = &*BB1_Itr++;
1276 while (isa<DbgInfoIntrinsic>(I2))
1277 I2 = &*BB2_Itr++;
1278 }
1279 // FIXME: Can we define a safety predicate for CallBr?
1280 if (isa<PHINode>(I1) || !I1->isIdenticalToWhenDefined(I2) ||
1281 (isa<InvokeInst>(I1) && !isSafeToHoistInvoke(BB1, BB2, I1, I2)) ||
1282 isa<CallBrInst>(I1))
1283 return false;
1284
1285 BasicBlock *BIParent = BI->getParent();
1286
1287 bool Changed = false;
1288 do {
1289 // If we are hoisting the terminator instruction, don't move one (making a
1290 // broken BB), instead clone it, and remove BI.
1291 if (I1->isTerminator())
1292 goto HoistTerminator;
1293
1294 // If we're going to hoist a call, make sure that the two instructions we're
1295 // commoning/hoisting are both marked with musttail, or neither of them is
1296 // marked as such. Otherwise, we might end up in a situation where we hoist
1297 // from a block where the terminator is a `ret` to a block where the terminator
1298 // is a `br`, and `musttail` calls expect to be followed by a return.
1299 auto *C1 = dyn_cast<CallInst>(I1);
1300 auto *C2 = dyn_cast<CallInst>(I2);
1301 if (C1 && C2)
1302 if (C1->isMustTailCall() != C2->isMustTailCall())
1303 return Changed;
1304
1305 if (!TTI.isProfitableToHoist(I1) || !TTI.isProfitableToHoist(I2))
1306 return Changed;
1307
1308 // If any of the two call sites has nomerge attribute, stop hoisting.
1309 if (const auto *CB1 = dyn_cast<CallBase>(I1))
1310 if (CB1->cannotMerge())
1311 return Changed;
1312 if (const auto *CB2 = dyn_cast<CallBase>(I2))
1313 if (CB2->cannotMerge())
1314 return Changed;
1315
1316 if (isa<DbgInfoIntrinsic>(I1) || isa<DbgInfoIntrinsic>(I2)) {
1317 assert (isa<DbgInfoIntrinsic>(I1) && isa<DbgInfoIntrinsic>(I2));
1318 // The debug location is an integral part of a debug info intrinsic
1319 // and can't be separated from it or replaced. Instead of attempting
1320 // to merge locations, simply hoist both copies of the intrinsic.
1321 BIParent->getInstList().splice(BI->getIterator(),
1322 BB1->getInstList(), I1);
1323 BIParent->getInstList().splice(BI->getIterator(),
1324 BB2->getInstList(), I2);
1325 Changed = true;
1326 } else {
1327 // For a normal instruction, we just move one to right before the branch,
1328 // then replace all uses of the other with the first. Finally, we remove
1329 // the now redundant second instruction.
1330 BIParent->getInstList().splice(BI->getIterator(),
1331 BB1->getInstList(), I1);
1332 if (!I2->use_empty())
1333 I2->replaceAllUsesWith(I1);
1334 I1->andIRFlags(I2);
1335 unsigned KnownIDs[] = {LLVMContext::MD_tbaa,
1336 LLVMContext::MD_range,
1337 LLVMContext::MD_fpmath,
1338 LLVMContext::MD_invariant_load,
1339 LLVMContext::MD_nonnull,
1340 LLVMContext::MD_invariant_group,
1341 LLVMContext::MD_align,
1342 LLVMContext::MD_dereferenceable,
1343 LLVMContext::MD_dereferenceable_or_null,
1344 LLVMContext::MD_mem_parallel_loop_access,
1345 LLVMContext::MD_access_group,
1346 LLVMContext::MD_preserve_access_index};
1347 combineMetadata(I1, I2, KnownIDs, true);
1348
1349 // I1 and I2 are being combined into a single instruction. Its debug
1350 // location is the merged locations of the original instructions.
1351 I1->applyMergedLocation(I1->getDebugLoc(), I2->getDebugLoc());
1352
1353 I2->eraseFromParent();
1354 Changed = true;
1355 }
1356
1357 I1 = &*BB1_Itr++;
1358 I2 = &*BB2_Itr++;
1359 // Skip debug info if it is not identical.
1360 DbgInfoIntrinsic *DBI1 = dyn_cast<DbgInfoIntrinsic>(I1);
1361 DbgInfoIntrinsic *DBI2 = dyn_cast<DbgInfoIntrinsic>(I2);
1362 if (!DBI1 || !DBI2 || !DBI1->isIdenticalToWhenDefined(DBI2)) {
1363 while (isa<DbgInfoIntrinsic>(I1))
1364 I1 = &*BB1_Itr++;
1365 while (isa<DbgInfoIntrinsic>(I2))
1366 I2 = &*BB2_Itr++;
1367 }
1368 } while (I1->isIdenticalToWhenDefined(I2));
1369
1370 return true;
1371
1372 HoistTerminator:
1373 // It may not be possible to hoist an invoke.
1374 // FIXME: Can we define a safety predicate for CallBr?
1375 if (isa<InvokeInst>(I1) && !isSafeToHoistInvoke(BB1, BB2, I1, I2))
1376 return Changed;
1377
1378 // TODO: callbr hoisting currently disabled pending further study.
1379 if (isa<CallBrInst>(I1))
1380 return Changed;
1381
1382 for (BasicBlock *Succ : successors(BB1)) {
1383 for (PHINode &PN : Succ->phis()) {
1384 Value *BB1V = PN.getIncomingValueForBlock(BB1);
1385 Value *BB2V = PN.getIncomingValueForBlock(BB2);
1386 if (BB1V == BB2V)
1387 continue;
1388
1389 // Check for passingValueIsAlwaysUndefined here because we would rather
1390 // eliminate undefined control flow then converting it to a select.
1391 if (passingValueIsAlwaysUndefined(BB1V, &PN) ||
1392 passingValueIsAlwaysUndefined(BB2V, &PN))
1393 return Changed;
1394
1395 if (isa<ConstantExpr>(BB1V) && !isSafeToSpeculativelyExecute(BB1V))
1396 return Changed;
1397 if (isa<ConstantExpr>(BB2V) && !isSafeToSpeculativelyExecute(BB2V))
1398 return Changed;
1399 }
1400 }
1401
1402 // Okay, it is safe to hoist the terminator.
1403 Instruction *NT = I1->clone();
1404 BIParent->getInstList().insert(BI->getIterator(), NT);
1405 if (!NT->getType()->isVoidTy()) {
1406 I1->replaceAllUsesWith(NT);
1407 I2->replaceAllUsesWith(NT);
1408 NT->takeName(I1);
1409 }
1410
1411 // Ensure terminator gets a debug location, even an unknown one, in case
1412 // it involves inlinable calls.
1413 NT->applyMergedLocation(I1->getDebugLoc(), I2->getDebugLoc());
1414
1415 // PHIs created below will adopt NT's merged DebugLoc.
1416 IRBuilder<NoFolder> Builder(NT);
1417
1418 // Hoisting one of the terminators from our successor is a great thing.
1419 // Unfortunately, the successors of the if/else blocks may have PHI nodes in
1420 // them. If they do, all PHI entries for BB1/BB2 must agree for all PHI
1421 // nodes, so we insert select instruction to compute the final result.
1422 std::map<std::pair<Value *, Value *>, SelectInst *> InsertedSelects;
1423 for (BasicBlock *Succ : successors(BB1)) {
1424 for (PHINode &PN : Succ->phis()) {
1425 Value *BB1V = PN.getIncomingValueForBlock(BB1);
1426 Value *BB2V = PN.getIncomingValueForBlock(BB2);
1427 if (BB1V == BB2V)
1428 continue;
1429
1430 // These values do not agree. Insert a select instruction before NT
1431 // that determines the right value.
1432 SelectInst *&SI = InsertedSelects[std::make_pair(BB1V, BB2V)];
1433 if (!SI) {
1434 // Propagate fast-math-flags from phi node to its replacement select.
1435 IRBuilder<>::FastMathFlagGuard FMFGuard(Builder);
1436 if (isa<FPMathOperator>(PN))
1437 Builder.setFastMathFlags(PN.getFastMathFlags());
1438
1439 SI = cast<SelectInst>(
1440 Builder.CreateSelect(BI->getCondition(), BB1V, BB2V,
1441 BB1V->getName() + "." + BB2V->getName(), BI));
1442 }
1443
1444 // Make the PHI node use the select for all incoming values for BB1/BB2
1445 for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i)
1446 if (PN.getIncomingBlock(i) == BB1 || PN.getIncomingBlock(i) == BB2)
1447 PN.setIncomingValue(i, SI);
1448 }
1449 }
1450
1451 // Update any PHI nodes in our new successors.
1452 for (BasicBlock *Succ : successors(BB1))
1453 AddPredecessorToBlock(Succ, BIParent, BB1);
1454
1455 EraseTerminatorAndDCECond(BI);
1456 return true;
1457 }
1458
1459 // Check lifetime markers.
isLifeTimeMarker(const Instruction * I)1460 static bool isLifeTimeMarker(const Instruction *I) {
1461 if (auto II = dyn_cast<IntrinsicInst>(I)) {
1462 switch (II->getIntrinsicID()) {
1463 default:
1464 break;
1465 case Intrinsic::lifetime_start:
1466 case Intrinsic::lifetime_end:
1467 return true;
1468 }
1469 }
1470 return false;
1471 }
1472
1473 // TODO: Refine this. This should avoid cases like turning constant memcpy sizes
1474 // into variables.
replacingOperandWithVariableIsCheap(const Instruction * I,int OpIdx)1475 static bool replacingOperandWithVariableIsCheap(const Instruction *I,
1476 int OpIdx) {
1477 return !isa<IntrinsicInst>(I);
1478 }
1479
1480 // All instructions in Insts belong to different blocks that all unconditionally
1481 // branch to a common successor. Analyze each instruction and return true if it
1482 // would be possible to sink them into their successor, creating one common
1483 // instruction instead. For every value that would be required to be provided by
1484 // PHI node (because an operand varies in each input block), add to PHIOperands.
canSinkInstructions(ArrayRef<Instruction * > Insts,DenseMap<Instruction *,SmallVector<Value *,4>> & PHIOperands)1485 static bool canSinkInstructions(
1486 ArrayRef<Instruction *> Insts,
1487 DenseMap<Instruction *, SmallVector<Value *, 4>> &PHIOperands) {
1488 // Prune out obviously bad instructions to move. Any non-store instruction
1489 // must have exactly one use, and we check later that use is by a single,
1490 // common PHI instruction in the successor.
1491 for (auto *I : Insts) {
1492 // These instructions may change or break semantics if moved.
1493 if (isa<PHINode>(I) || I->isEHPad() || isa<AllocaInst>(I) ||
1494 I->getType()->isTokenTy())
1495 return false;
1496
1497 // Conservatively return false if I is an inline-asm instruction. Sinking
1498 // and merging inline-asm instructions can potentially create arguments
1499 // that cannot satisfy the inline-asm constraints.
1500 // If the instruction has nomerge attribute, return false.
1501 if (const auto *C = dyn_cast<CallBase>(I))
1502 if (C->isInlineAsm() || C->cannotMerge())
1503 return false;
1504
1505 // Everything must have only one use too, apart from stores which
1506 // have no uses.
1507 if (!isa<StoreInst>(I) && !I->hasOneUse())
1508 return false;
1509 }
1510
1511 const Instruction *I0 = Insts.front();
1512 for (auto *I : Insts)
1513 if (!I->isSameOperationAs(I0))
1514 return false;
1515
1516 // All instructions in Insts are known to be the same opcode. If they aren't
1517 // stores, check the only user of each is a PHI or in the same block as the
1518 // instruction, because if a user is in the same block as an instruction
1519 // we're contemplating sinking, it must already be determined to be sinkable.
1520 if (!isa<StoreInst>(I0)) {
1521 auto *PNUse = dyn_cast<PHINode>(*I0->user_begin());
1522 auto *Succ = I0->getParent()->getTerminator()->getSuccessor(0);
1523 if (!all_of(Insts, [&PNUse,&Succ](const Instruction *I) -> bool {
1524 auto *U = cast<Instruction>(*I->user_begin());
1525 return (PNUse &&
1526 PNUse->getParent() == Succ &&
1527 PNUse->getIncomingValueForBlock(I->getParent()) == I) ||
1528 U->getParent() == I->getParent();
1529 }))
1530 return false;
1531 }
1532
1533 // Because SROA can't handle speculating stores of selects, try not to sink
1534 // loads, stores or lifetime markers of allocas when we'd have to create a
1535 // PHI for the address operand. Also, because it is likely that loads or
1536 // stores of allocas will disappear when Mem2Reg/SROA is run, don't sink
1537 // them.
1538 // This can cause code churn which can have unintended consequences down
1539 // the line - see https://llvm.org/bugs/show_bug.cgi?id=30244.
1540 // FIXME: This is a workaround for a deficiency in SROA - see
1541 // https://llvm.org/bugs/show_bug.cgi?id=30188
1542 if (isa<StoreInst>(I0) && any_of(Insts, [](const Instruction *I) {
1543 return isa<AllocaInst>(I->getOperand(1)->stripPointerCasts());
1544 }))
1545 return false;
1546 if (isa<LoadInst>(I0) && any_of(Insts, [](const Instruction *I) {
1547 return isa<AllocaInst>(I->getOperand(0)->stripPointerCasts());
1548 }))
1549 return false;
1550 if (isLifeTimeMarker(I0) && any_of(Insts, [](const Instruction *I) {
1551 return isa<AllocaInst>(I->getOperand(1)->stripPointerCasts());
1552 }))
1553 return false;
1554
1555 for (unsigned OI = 0, OE = I0->getNumOperands(); OI != OE; ++OI) {
1556 Value *Op = I0->getOperand(OI);
1557 if (Op->getType()->isTokenTy())
1558 // Don't touch any operand of token type.
1559 return false;
1560
1561 auto SameAsI0 = [&I0, OI](const Instruction *I) {
1562 assert(I->getNumOperands() == I0->getNumOperands());
1563 return I->getOperand(OI) == I0->getOperand(OI);
1564 };
1565 if (!all_of(Insts, SameAsI0)) {
1566 if ((isa<Constant>(Op) && !replacingOperandWithVariableIsCheap(I0, OI)) ||
1567 !canReplaceOperandWithVariable(I0, OI))
1568 // We can't create a PHI from this GEP.
1569 return false;
1570 // Don't create indirect calls! The called value is the final operand.
1571 if (isa<CallBase>(I0) && OI == OE - 1) {
1572 // FIXME: if the call was *already* indirect, we should do this.
1573 return false;
1574 }
1575 for (auto *I : Insts)
1576 PHIOperands[I].push_back(I->getOperand(OI));
1577 }
1578 }
1579 return true;
1580 }
1581
1582 // Assuming canSinkLastInstruction(Blocks) has returned true, sink the last
1583 // instruction of every block in Blocks to their common successor, commoning
1584 // into one instruction.
sinkLastInstruction(ArrayRef<BasicBlock * > Blocks)1585 static bool sinkLastInstruction(ArrayRef<BasicBlock*> Blocks) {
1586 auto *BBEnd = Blocks[0]->getTerminator()->getSuccessor(0);
1587
1588 // canSinkLastInstruction returning true guarantees that every block has at
1589 // least one non-terminator instruction.
1590 SmallVector<Instruction*,4> Insts;
1591 for (auto *BB : Blocks) {
1592 Instruction *I = BB->getTerminator();
1593 do {
1594 I = I->getPrevNode();
1595 } while (isa<DbgInfoIntrinsic>(I) && I != &BB->front());
1596 if (!isa<DbgInfoIntrinsic>(I))
1597 Insts.push_back(I);
1598 }
1599
1600 // The only checking we need to do now is that all users of all instructions
1601 // are the same PHI node. canSinkLastInstruction should have checked this but
1602 // it is slightly over-aggressive - it gets confused by commutative instructions
1603 // so double-check it here.
1604 Instruction *I0 = Insts.front();
1605 if (!isa<StoreInst>(I0)) {
1606 auto *PNUse = dyn_cast<PHINode>(*I0->user_begin());
1607 if (!all_of(Insts, [&PNUse](const Instruction *I) -> bool {
1608 auto *U = cast<Instruction>(*I->user_begin());
1609 return U == PNUse;
1610 }))
1611 return false;
1612 }
1613
1614 // We don't need to do any more checking here; canSinkLastInstruction should
1615 // have done it all for us.
1616 SmallVector<Value*, 4> NewOperands;
1617 for (unsigned O = 0, E = I0->getNumOperands(); O != E; ++O) {
1618 // This check is different to that in canSinkLastInstruction. There, we
1619 // cared about the global view once simplifycfg (and instcombine) have
1620 // completed - it takes into account PHIs that become trivially
1621 // simplifiable. However here we need a more local view; if an operand
1622 // differs we create a PHI and rely on instcombine to clean up the very
1623 // small mess we may make.
1624 bool NeedPHI = any_of(Insts, [&I0, O](const Instruction *I) {
1625 return I->getOperand(O) != I0->getOperand(O);
1626 });
1627 if (!NeedPHI) {
1628 NewOperands.push_back(I0->getOperand(O));
1629 continue;
1630 }
1631
1632 // Create a new PHI in the successor block and populate it.
1633 auto *Op = I0->getOperand(O);
1634 assert(!Op->getType()->isTokenTy() && "Can't PHI tokens!");
1635 auto *PN = PHINode::Create(Op->getType(), Insts.size(),
1636 Op->getName() + ".sink", &BBEnd->front());
1637 for (auto *I : Insts)
1638 PN->addIncoming(I->getOperand(O), I->getParent());
1639 NewOperands.push_back(PN);
1640 }
1641
1642 // Arbitrarily use I0 as the new "common" instruction; remap its operands
1643 // and move it to the start of the successor block.
1644 for (unsigned O = 0, E = I0->getNumOperands(); O != E; ++O)
1645 I0->getOperandUse(O).set(NewOperands[O]);
1646 I0->moveBefore(&*BBEnd->getFirstInsertionPt());
1647
1648 // Update metadata and IR flags, and merge debug locations.
1649 for (auto *I : Insts)
1650 if (I != I0) {
1651 // The debug location for the "common" instruction is the merged locations
1652 // of all the commoned instructions. We start with the original location
1653 // of the "common" instruction and iteratively merge each location in the
1654 // loop below.
1655 // This is an N-way merge, which will be inefficient if I0 is a CallInst.
1656 // However, as N-way merge for CallInst is rare, so we use simplified API
1657 // instead of using complex API for N-way merge.
1658 I0->applyMergedLocation(I0->getDebugLoc(), I->getDebugLoc());
1659 combineMetadataForCSE(I0, I, true);
1660 I0->andIRFlags(I);
1661 }
1662
1663 if (!isa<StoreInst>(I0)) {
1664 // canSinkLastInstruction checked that all instructions were used by
1665 // one and only one PHI node. Find that now, RAUW it to our common
1666 // instruction and nuke it.
1667 assert(I0->hasOneUse());
1668 auto *PN = cast<PHINode>(*I0->user_begin());
1669 PN->replaceAllUsesWith(I0);
1670 PN->eraseFromParent();
1671 }
1672
1673 // Finally nuke all instructions apart from the common instruction.
1674 for (auto *I : Insts)
1675 if (I != I0)
1676 I->eraseFromParent();
1677
1678 return true;
1679 }
1680
1681 namespace {
1682
1683 // LockstepReverseIterator - Iterates through instructions
1684 // in a set of blocks in reverse order from the first non-terminator.
1685 // For example (assume all blocks have size n):
1686 // LockstepReverseIterator I([B1, B2, B3]);
1687 // *I-- = [B1[n], B2[n], B3[n]];
1688 // *I-- = [B1[n-1], B2[n-1], B3[n-1]];
1689 // *I-- = [B1[n-2], B2[n-2], B3[n-2]];
1690 // ...
1691 class LockstepReverseIterator {
1692 ArrayRef<BasicBlock*> Blocks;
1693 SmallVector<Instruction*,4> Insts;
1694 bool Fail;
1695
1696 public:
LockstepReverseIterator(ArrayRef<BasicBlock * > Blocks)1697 LockstepReverseIterator(ArrayRef<BasicBlock*> Blocks) : Blocks(Blocks) {
1698 reset();
1699 }
1700
reset()1701 void reset() {
1702 Fail = false;
1703 Insts.clear();
1704 for (auto *BB : Blocks) {
1705 Instruction *Inst = BB->getTerminator();
1706 for (Inst = Inst->getPrevNode(); Inst && isa<DbgInfoIntrinsic>(Inst);)
1707 Inst = Inst->getPrevNode();
1708 if (!Inst) {
1709 // Block wasn't big enough.
1710 Fail = true;
1711 return;
1712 }
1713 Insts.push_back(Inst);
1714 }
1715 }
1716
isValid() const1717 bool isValid() const {
1718 return !Fail;
1719 }
1720
operator --()1721 void operator--() {
1722 if (Fail)
1723 return;
1724 for (auto *&Inst : Insts) {
1725 for (Inst = Inst->getPrevNode(); Inst && isa<DbgInfoIntrinsic>(Inst);)
1726 Inst = Inst->getPrevNode();
1727 // Already at beginning of block.
1728 if (!Inst) {
1729 Fail = true;
1730 return;
1731 }
1732 }
1733 }
1734
operator *() const1735 ArrayRef<Instruction*> operator * () const {
1736 return Insts;
1737 }
1738 };
1739
1740 } // end anonymous namespace
1741
1742 /// Check whether BB's predecessors end with unconditional branches. If it is
1743 /// true, sink any common code from the predecessors to BB.
1744 /// We also allow one predecessor to end with conditional branch (but no more
1745 /// than one).
SinkCommonCodeFromPredecessors(BasicBlock * BB)1746 static bool SinkCommonCodeFromPredecessors(BasicBlock *BB) {
1747 // We support two situations:
1748 // (1) all incoming arcs are unconditional
1749 // (2) one incoming arc is conditional
1750 //
1751 // (2) is very common in switch defaults and
1752 // else-if patterns;
1753 //
1754 // if (a) f(1);
1755 // else if (b) f(2);
1756 //
1757 // produces:
1758 //
1759 // [if]
1760 // / \
1761 // [f(1)] [if]
1762 // | | \
1763 // | | |
1764 // | [f(2)]|
1765 // \ | /
1766 // [ end ]
1767 //
1768 // [end] has two unconditional predecessor arcs and one conditional. The
1769 // conditional refers to the implicit empty 'else' arc. This conditional
1770 // arc can also be caused by an empty default block in a switch.
1771 //
1772 // In this case, we attempt to sink code from all *unconditional* arcs.
1773 // If we can sink instructions from these arcs (determined during the scan
1774 // phase below) we insert a common successor for all unconditional arcs and
1775 // connect that to [end], to enable sinking:
1776 //
1777 // [if]
1778 // / \
1779 // [x(1)] [if]
1780 // | | \
1781 // | | \
1782 // | [x(2)] |
1783 // \ / |
1784 // [sink.split] |
1785 // \ /
1786 // [ end ]
1787 //
1788 SmallVector<BasicBlock*,4> UnconditionalPreds;
1789 Instruction *Cond = nullptr;
1790 for (auto *B : predecessors(BB)) {
1791 auto *T = B->getTerminator();
1792 if (isa<BranchInst>(T) && cast<BranchInst>(T)->isUnconditional())
1793 UnconditionalPreds.push_back(B);
1794 else if ((isa<BranchInst>(T) || isa<SwitchInst>(T)) && !Cond)
1795 Cond = T;
1796 else
1797 return false;
1798 }
1799 if (UnconditionalPreds.size() < 2)
1800 return false;
1801
1802 bool Changed = false;
1803 // We take a two-step approach to tail sinking. First we scan from the end of
1804 // each block upwards in lockstep. If the n'th instruction from the end of each
1805 // block can be sunk, those instructions are added to ValuesToSink and we
1806 // carry on. If we can sink an instruction but need to PHI-merge some operands
1807 // (because they're not identical in each instruction) we add these to
1808 // PHIOperands.
1809 unsigned ScanIdx = 0;
1810 SmallPtrSet<Value*,4> InstructionsToSink;
1811 DenseMap<Instruction*, SmallVector<Value*,4>> PHIOperands;
1812 LockstepReverseIterator LRI(UnconditionalPreds);
1813 while (LRI.isValid() &&
1814 canSinkInstructions(*LRI, PHIOperands)) {
1815 LLVM_DEBUG(dbgs() << "SINK: instruction can be sunk: " << *(*LRI)[0]
1816 << "\n");
1817 InstructionsToSink.insert((*LRI).begin(), (*LRI).end());
1818 ++ScanIdx;
1819 --LRI;
1820 }
1821
1822 auto ProfitableToSinkInstruction = [&](LockstepReverseIterator &LRI) {
1823 unsigned NumPHIdValues = 0;
1824 for (auto *I : *LRI)
1825 for (auto *V : PHIOperands[I])
1826 if (InstructionsToSink.count(V) == 0)
1827 ++NumPHIdValues;
1828 LLVM_DEBUG(dbgs() << "SINK: #phid values: " << NumPHIdValues << "\n");
1829 unsigned NumPHIInsts = NumPHIdValues / UnconditionalPreds.size();
1830 if ((NumPHIdValues % UnconditionalPreds.size()) != 0)
1831 NumPHIInsts++;
1832
1833 return NumPHIInsts <= 1;
1834 };
1835
1836 if (ScanIdx > 0 && Cond) {
1837 // Check if we would actually sink anything first! This mutates the CFG and
1838 // adds an extra block. The goal in doing this is to allow instructions that
1839 // couldn't be sunk before to be sunk - obviously, speculatable instructions
1840 // (such as trunc, add) can be sunk and predicated already. So we check that
1841 // we're going to sink at least one non-speculatable instruction.
1842 LRI.reset();
1843 unsigned Idx = 0;
1844 bool Profitable = false;
1845 while (ProfitableToSinkInstruction(LRI) && Idx < ScanIdx) {
1846 if (!isSafeToSpeculativelyExecute((*LRI)[0])) {
1847 Profitable = true;
1848 break;
1849 }
1850 --LRI;
1851 ++Idx;
1852 }
1853 if (!Profitable)
1854 return false;
1855
1856 LLVM_DEBUG(dbgs() << "SINK: Splitting edge\n");
1857 // We have a conditional edge and we're going to sink some instructions.
1858 // Insert a new block postdominating all blocks we're going to sink from.
1859 if (!SplitBlockPredecessors(BB, UnconditionalPreds, ".sink.split"))
1860 // Edges couldn't be split.
1861 return false;
1862 Changed = true;
1863 }
1864
1865 // Now that we've analyzed all potential sinking candidates, perform the
1866 // actual sink. We iteratively sink the last non-terminator of the source
1867 // blocks into their common successor unless doing so would require too
1868 // many PHI instructions to be generated (currently only one PHI is allowed
1869 // per sunk instruction).
1870 //
1871 // We can use InstructionsToSink to discount values needing PHI-merging that will
1872 // actually be sunk in a later iteration. This allows us to be more
1873 // aggressive in what we sink. This does allow a false positive where we
1874 // sink presuming a later value will also be sunk, but stop half way through
1875 // and never actually sink it which means we produce more PHIs than intended.
1876 // This is unlikely in practice though.
1877 for (unsigned SinkIdx = 0; SinkIdx != ScanIdx; ++SinkIdx) {
1878 LLVM_DEBUG(dbgs() << "SINK: Sink: "
1879 << *UnconditionalPreds[0]->getTerminator()->getPrevNode()
1880 << "\n");
1881
1882 // Because we've sunk every instruction in turn, the current instruction to
1883 // sink is always at index 0.
1884 LRI.reset();
1885 if (!ProfitableToSinkInstruction(LRI)) {
1886 // Too many PHIs would be created.
1887 LLVM_DEBUG(
1888 dbgs() << "SINK: stopping here, too many PHIs would be created!\n");
1889 break;
1890 }
1891
1892 if (!sinkLastInstruction(UnconditionalPreds))
1893 return Changed;
1894 NumSinkCommons++;
1895 Changed = true;
1896 }
1897 return Changed;
1898 }
1899
1900 /// Determine if we can hoist sink a sole store instruction out of a
1901 /// conditional block.
1902 ///
1903 /// We are looking for code like the following:
1904 /// BrBB:
1905 /// store i32 %add, i32* %arrayidx2
1906 /// ... // No other stores or function calls (we could be calling a memory
1907 /// ... // function).
1908 /// %cmp = icmp ult %x, %y
1909 /// br i1 %cmp, label %EndBB, label %ThenBB
1910 /// ThenBB:
1911 /// store i32 %add5, i32* %arrayidx2
1912 /// br label EndBB
1913 /// EndBB:
1914 /// ...
1915 /// We are going to transform this into:
1916 /// BrBB:
1917 /// store i32 %add, i32* %arrayidx2
1918 /// ... //
1919 /// %cmp = icmp ult %x, %y
1920 /// %add.add5 = select i1 %cmp, i32 %add, %add5
1921 /// store i32 %add.add5, i32* %arrayidx2
1922 /// ...
1923 ///
1924 /// \return The pointer to the value of the previous store if the store can be
1925 /// hoisted into the predecessor block. 0 otherwise.
isSafeToSpeculateStore(Instruction * I,BasicBlock * BrBB,BasicBlock * StoreBB,BasicBlock * EndBB)1926 static Value *isSafeToSpeculateStore(Instruction *I, BasicBlock *BrBB,
1927 BasicBlock *StoreBB, BasicBlock *EndBB) {
1928 StoreInst *StoreToHoist = dyn_cast<StoreInst>(I);
1929 if (!StoreToHoist)
1930 return nullptr;
1931
1932 // Volatile or atomic.
1933 if (!StoreToHoist->isSimple())
1934 return nullptr;
1935
1936 Value *StorePtr = StoreToHoist->getPointerOperand();
1937
1938 // Look for a store to the same pointer in BrBB.
1939 unsigned MaxNumInstToLookAt = 9;
1940 for (Instruction &CurI : reverse(BrBB->instructionsWithoutDebug())) {
1941 if (!MaxNumInstToLookAt)
1942 break;
1943 --MaxNumInstToLookAt;
1944
1945 // Could be calling an instruction that affects memory like free().
1946 if (CurI.mayHaveSideEffects() && !isa<StoreInst>(CurI))
1947 return nullptr;
1948
1949 if (auto *SI = dyn_cast<StoreInst>(&CurI)) {
1950 // Found the previous store make sure it stores to the same location.
1951 if (SI->getPointerOperand() == StorePtr)
1952 // Found the previous store, return its value operand.
1953 return SI->getValueOperand();
1954 return nullptr; // Unknown store.
1955 }
1956 }
1957
1958 return nullptr;
1959 }
1960
1961 /// Speculate a conditional basic block flattening the CFG.
1962 ///
1963 /// Note that this is a very risky transform currently. Speculating
1964 /// instructions like this is most often not desirable. Instead, there is an MI
1965 /// pass which can do it with full awareness of the resource constraints.
1966 /// However, some cases are "obvious" and we should do directly. An example of
1967 /// this is speculating a single, reasonably cheap instruction.
1968 ///
1969 /// There is only one distinct advantage to flattening the CFG at the IR level:
1970 /// it makes very common but simplistic optimizations such as are common in
1971 /// instcombine and the DAG combiner more powerful by removing CFG edges and
1972 /// modeling their effects with easier to reason about SSA value graphs.
1973 ///
1974 ///
1975 /// An illustration of this transform is turning this IR:
1976 /// \code
1977 /// BB:
1978 /// %cmp = icmp ult %x, %y
1979 /// br i1 %cmp, label %EndBB, label %ThenBB
1980 /// ThenBB:
1981 /// %sub = sub %x, %y
1982 /// br label BB2
1983 /// EndBB:
1984 /// %phi = phi [ %sub, %ThenBB ], [ 0, %EndBB ]
1985 /// ...
1986 /// \endcode
1987 ///
1988 /// Into this IR:
1989 /// \code
1990 /// BB:
1991 /// %cmp = icmp ult %x, %y
1992 /// %sub = sub %x, %y
1993 /// %cond = select i1 %cmp, 0, %sub
1994 /// ...
1995 /// \endcode
1996 ///
1997 /// \returns true if the conditional block is removed.
SpeculativelyExecuteBB(BranchInst * BI,BasicBlock * ThenBB,const TargetTransformInfo & TTI)1998 bool SimplifyCFGOpt::SpeculativelyExecuteBB(BranchInst *BI, BasicBlock *ThenBB,
1999 const TargetTransformInfo &TTI) {
2000 // Be conservative for now. FP select instruction can often be expensive.
2001 Value *BrCond = BI->getCondition();
2002 if (isa<FCmpInst>(BrCond))
2003 return false;
2004
2005 BasicBlock *BB = BI->getParent();
2006 BasicBlock *EndBB = ThenBB->getTerminator()->getSuccessor(0);
2007
2008 // If ThenBB is actually on the false edge of the conditional branch, remember
2009 // to swap the select operands later.
2010 bool Invert = false;
2011 if (ThenBB != BI->getSuccessor(0)) {
2012 assert(ThenBB == BI->getSuccessor(1) && "No edge from 'if' block?");
2013 Invert = true;
2014 }
2015 assert(EndBB == BI->getSuccessor(!Invert) && "No edge from to end block");
2016
2017 // Keep a count of how many times instructions are used within ThenBB when
2018 // they are candidates for sinking into ThenBB. Specifically:
2019 // - They are defined in BB, and
2020 // - They have no side effects, and
2021 // - All of their uses are in ThenBB.
2022 SmallDenseMap<Instruction *, unsigned, 4> SinkCandidateUseCounts;
2023
2024 SmallVector<Instruction *, 4> SpeculatedDbgIntrinsics;
2025
2026 unsigned SpeculatedInstructions = 0;
2027 Value *SpeculatedStoreValue = nullptr;
2028 StoreInst *SpeculatedStore = nullptr;
2029 for (BasicBlock::iterator BBI = ThenBB->begin(),
2030 BBE = std::prev(ThenBB->end());
2031 BBI != BBE; ++BBI) {
2032 Instruction *I = &*BBI;
2033 // Skip debug info.
2034 if (isa<DbgInfoIntrinsic>(I)) {
2035 SpeculatedDbgIntrinsics.push_back(I);
2036 continue;
2037 }
2038
2039 // Only speculatively execute a single instruction (not counting the
2040 // terminator) for now.
2041 ++SpeculatedInstructions;
2042 if (SpeculatedInstructions > 1)
2043 return false;
2044
2045 // Don't hoist the instruction if it's unsafe or expensive.
2046 if (!isSafeToSpeculativelyExecute(I) &&
2047 !(HoistCondStores && (SpeculatedStoreValue = isSafeToSpeculateStore(
2048 I, BB, ThenBB, EndBB))))
2049 return false;
2050 if (!SpeculatedStoreValue &&
2051 ComputeSpeculationCost(I, TTI) >
2052 PHINodeFoldingThreshold * TargetTransformInfo::TCC_Basic)
2053 return false;
2054
2055 // Store the store speculation candidate.
2056 if (SpeculatedStoreValue)
2057 SpeculatedStore = cast<StoreInst>(I);
2058
2059 // Do not hoist the instruction if any of its operands are defined but not
2060 // used in BB. The transformation will prevent the operand from
2061 // being sunk into the use block.
2062 for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i) {
2063 Instruction *OpI = dyn_cast<Instruction>(*i);
2064 if (!OpI || OpI->getParent() != BB || OpI->mayHaveSideEffects())
2065 continue; // Not a candidate for sinking.
2066
2067 ++SinkCandidateUseCounts[OpI];
2068 }
2069 }
2070
2071 // Consider any sink candidates which are only used in ThenBB as costs for
2072 // speculation. Note, while we iterate over a DenseMap here, we are summing
2073 // and so iteration order isn't significant.
2074 for (SmallDenseMap<Instruction *, unsigned, 4>::iterator
2075 I = SinkCandidateUseCounts.begin(),
2076 E = SinkCandidateUseCounts.end();
2077 I != E; ++I)
2078 if (I->first->hasNUses(I->second)) {
2079 ++SpeculatedInstructions;
2080 if (SpeculatedInstructions > 1)
2081 return false;
2082 }
2083
2084 // Check that the PHI nodes can be converted to selects.
2085 bool HaveRewritablePHIs = false;
2086 for (PHINode &PN : EndBB->phis()) {
2087 Value *OrigV = PN.getIncomingValueForBlock(BB);
2088 Value *ThenV = PN.getIncomingValueForBlock(ThenBB);
2089
2090 // FIXME: Try to remove some of the duplication with HoistThenElseCodeToIf.
2091 // Skip PHIs which are trivial.
2092 if (ThenV == OrigV)
2093 continue;
2094
2095 // Don't convert to selects if we could remove undefined behavior instead.
2096 if (passingValueIsAlwaysUndefined(OrigV, &PN) ||
2097 passingValueIsAlwaysUndefined(ThenV, &PN))
2098 return false;
2099
2100 HaveRewritablePHIs = true;
2101 ConstantExpr *OrigCE = dyn_cast<ConstantExpr>(OrigV);
2102 ConstantExpr *ThenCE = dyn_cast<ConstantExpr>(ThenV);
2103 if (!OrigCE && !ThenCE)
2104 continue; // Known safe and cheap.
2105
2106 if ((ThenCE && !isSafeToSpeculativelyExecute(ThenCE)) ||
2107 (OrigCE && !isSafeToSpeculativelyExecute(OrigCE)))
2108 return false;
2109 unsigned OrigCost = OrigCE ? ComputeSpeculationCost(OrigCE, TTI) : 0;
2110 unsigned ThenCost = ThenCE ? ComputeSpeculationCost(ThenCE, TTI) : 0;
2111 unsigned MaxCost =
2112 2 * PHINodeFoldingThreshold * TargetTransformInfo::TCC_Basic;
2113 if (OrigCost + ThenCost > MaxCost)
2114 return false;
2115
2116 // Account for the cost of an unfolded ConstantExpr which could end up
2117 // getting expanded into Instructions.
2118 // FIXME: This doesn't account for how many operations are combined in the
2119 // constant expression.
2120 ++SpeculatedInstructions;
2121 if (SpeculatedInstructions > 1)
2122 return false;
2123 }
2124
2125 // If there are no PHIs to process, bail early. This helps ensure idempotence
2126 // as well.
2127 if (!HaveRewritablePHIs && !(HoistCondStores && SpeculatedStoreValue))
2128 return false;
2129
2130 // If we get here, we can hoist the instruction and if-convert.
2131 LLVM_DEBUG(dbgs() << "SPECULATIVELY EXECUTING BB" << *ThenBB << "\n";);
2132
2133 // Insert a select of the value of the speculated store.
2134 if (SpeculatedStoreValue) {
2135 IRBuilder<NoFolder> Builder(BI);
2136 Value *TrueV = SpeculatedStore->getValueOperand();
2137 Value *FalseV = SpeculatedStoreValue;
2138 if (Invert)
2139 std::swap(TrueV, FalseV);
2140 Value *S = Builder.CreateSelect(
2141 BrCond, TrueV, FalseV, "spec.store.select", BI);
2142 SpeculatedStore->setOperand(0, S);
2143 SpeculatedStore->applyMergedLocation(BI->getDebugLoc(),
2144 SpeculatedStore->getDebugLoc());
2145 }
2146
2147 // Metadata can be dependent on the condition we are hoisting above.
2148 // Conservatively strip all metadata on the instruction. Drop the debug loc
2149 // to avoid making it appear as if the condition is a constant, which would
2150 // be misleading while debugging.
2151 for (auto &I : *ThenBB) {
2152 if (!SpeculatedStoreValue || &I != SpeculatedStore)
2153 I.setDebugLoc(DebugLoc());
2154 I.dropUnknownNonDebugMetadata();
2155 }
2156
2157 // Hoist the instructions.
2158 BB->getInstList().splice(BI->getIterator(), ThenBB->getInstList(),
2159 ThenBB->begin(), std::prev(ThenBB->end()));
2160
2161 // Insert selects and rewrite the PHI operands.
2162 IRBuilder<NoFolder> Builder(BI);
2163 for (PHINode &PN : EndBB->phis()) {
2164 unsigned OrigI = PN.getBasicBlockIndex(BB);
2165 unsigned ThenI = PN.getBasicBlockIndex(ThenBB);
2166 Value *OrigV = PN.getIncomingValue(OrigI);
2167 Value *ThenV = PN.getIncomingValue(ThenI);
2168
2169 // Skip PHIs which are trivial.
2170 if (OrigV == ThenV)
2171 continue;
2172
2173 // Create a select whose true value is the speculatively executed value and
2174 // false value is the pre-existing value. Swap them if the branch
2175 // destinations were inverted.
2176 Value *TrueV = ThenV, *FalseV = OrigV;
2177 if (Invert)
2178 std::swap(TrueV, FalseV);
2179 Value *V = Builder.CreateSelect(BrCond, TrueV, FalseV, "spec.select", BI);
2180 PN.setIncomingValue(OrigI, V);
2181 PN.setIncomingValue(ThenI, V);
2182 }
2183
2184 // Remove speculated dbg intrinsics.
2185 // FIXME: Is it possible to do this in a more elegant way? Moving/merging the
2186 // dbg value for the different flows and inserting it after the select.
2187 for (Instruction *I : SpeculatedDbgIntrinsics)
2188 I->eraseFromParent();
2189
2190 ++NumSpeculations;
2191 return true;
2192 }
2193
2194 /// Return true if we can thread a branch across this block.
BlockIsSimpleEnoughToThreadThrough(BasicBlock * BB)2195 static bool BlockIsSimpleEnoughToThreadThrough(BasicBlock *BB) {
2196 int Size = 0;
2197
2198 for (Instruction &I : BB->instructionsWithoutDebug()) {
2199 if (Size > MaxSmallBlockSize)
2200 return false; // Don't clone large BB's.
2201 // We will delete Phis while threading, so Phis should not be accounted in
2202 // block's size
2203 if (!isa<PHINode>(I))
2204 ++Size;
2205
2206 // We can only support instructions that do not define values that are
2207 // live outside of the current basic block.
2208 for (User *U : I.users()) {
2209 Instruction *UI = cast<Instruction>(U);
2210 if (UI->getParent() != BB || isa<PHINode>(UI))
2211 return false;
2212 }
2213
2214 // Looks ok, continue checking.
2215 }
2216
2217 return true;
2218 }
2219
2220 /// If we have a conditional branch on a PHI node value that is defined in the
2221 /// same block as the branch and if any PHI entries are constants, thread edges
2222 /// corresponding to that entry to be branches to their ultimate destination.
FoldCondBranchOnPHI(BranchInst * BI,const DataLayout & DL,AssumptionCache * AC)2223 static bool FoldCondBranchOnPHI(BranchInst *BI, const DataLayout &DL,
2224 AssumptionCache *AC) {
2225 BasicBlock *BB = BI->getParent();
2226 PHINode *PN = dyn_cast<PHINode>(BI->getCondition());
2227 // NOTE: we currently cannot transform this case if the PHI node is used
2228 // outside of the block.
2229 if (!PN || PN->getParent() != BB || !PN->hasOneUse())
2230 return false;
2231
2232 // Degenerate case of a single entry PHI.
2233 if (PN->getNumIncomingValues() == 1) {
2234 FoldSingleEntryPHINodes(PN->getParent());
2235 return true;
2236 }
2237
2238 // Now we know that this block has multiple preds and two succs.
2239 if (!BlockIsSimpleEnoughToThreadThrough(BB))
2240 return false;
2241
2242 // Can't fold blocks that contain noduplicate or convergent calls.
2243 if (any_of(*BB, [](const Instruction &I) {
2244 const CallInst *CI = dyn_cast<CallInst>(&I);
2245 return CI && (CI->cannotDuplicate() || CI->isConvergent());
2246 }))
2247 return false;
2248
2249 // Okay, this is a simple enough basic block. See if any phi values are
2250 // constants.
2251 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
2252 ConstantInt *CB = dyn_cast<ConstantInt>(PN->getIncomingValue(i));
2253 if (!CB || !CB->getType()->isIntegerTy(1))
2254 continue;
2255
2256 // Okay, we now know that all edges from PredBB should be revectored to
2257 // branch to RealDest.
2258 BasicBlock *PredBB = PN->getIncomingBlock(i);
2259 BasicBlock *RealDest = BI->getSuccessor(!CB->getZExtValue());
2260
2261 if (RealDest == BB)
2262 continue; // Skip self loops.
2263 // Skip if the predecessor's terminator is an indirect branch.
2264 if (isa<IndirectBrInst>(PredBB->getTerminator()))
2265 continue;
2266
2267 // The dest block might have PHI nodes, other predecessors and other
2268 // difficult cases. Instead of being smart about this, just insert a new
2269 // block that jumps to the destination block, effectively splitting
2270 // the edge we are about to create.
2271 BasicBlock *EdgeBB =
2272 BasicBlock::Create(BB->getContext(), RealDest->getName() + ".critedge",
2273 RealDest->getParent(), RealDest);
2274 BranchInst *CritEdgeBranch = BranchInst::Create(RealDest, EdgeBB);
2275 CritEdgeBranch->setDebugLoc(BI->getDebugLoc());
2276
2277 // Update PHI nodes.
2278 AddPredecessorToBlock(RealDest, EdgeBB, BB);
2279
2280 // BB may have instructions that are being threaded over. Clone these
2281 // instructions into EdgeBB. We know that there will be no uses of the
2282 // cloned instructions outside of EdgeBB.
2283 BasicBlock::iterator InsertPt = EdgeBB->begin();
2284 DenseMap<Value *, Value *> TranslateMap; // Track translated values.
2285 for (BasicBlock::iterator BBI = BB->begin(); &*BBI != BI; ++BBI) {
2286 if (PHINode *PN = dyn_cast<PHINode>(BBI)) {
2287 TranslateMap[PN] = PN->getIncomingValueForBlock(PredBB);
2288 continue;
2289 }
2290 // Clone the instruction.
2291 Instruction *N = BBI->clone();
2292 if (BBI->hasName())
2293 N->setName(BBI->getName() + ".c");
2294
2295 // Update operands due to translation.
2296 for (User::op_iterator i = N->op_begin(), e = N->op_end(); i != e; ++i) {
2297 DenseMap<Value *, Value *>::iterator PI = TranslateMap.find(*i);
2298 if (PI != TranslateMap.end())
2299 *i = PI->second;
2300 }
2301
2302 // Check for trivial simplification.
2303 if (Value *V = SimplifyInstruction(N, {DL, nullptr, nullptr, AC})) {
2304 if (!BBI->use_empty())
2305 TranslateMap[&*BBI] = V;
2306 if (!N->mayHaveSideEffects()) {
2307 N->deleteValue(); // Instruction folded away, don't need actual inst
2308 N = nullptr;
2309 }
2310 } else {
2311 if (!BBI->use_empty())
2312 TranslateMap[&*BBI] = N;
2313 }
2314 if (N) {
2315 // Insert the new instruction into its new home.
2316 EdgeBB->getInstList().insert(InsertPt, N);
2317
2318 // Register the new instruction with the assumption cache if necessary.
2319 if (AC && match(N, m_Intrinsic<Intrinsic::assume>()))
2320 AC->registerAssumption(cast<IntrinsicInst>(N));
2321 }
2322 }
2323
2324 // Loop over all of the edges from PredBB to BB, changing them to branch
2325 // to EdgeBB instead.
2326 Instruction *PredBBTI = PredBB->getTerminator();
2327 for (unsigned i = 0, e = PredBBTI->getNumSuccessors(); i != e; ++i)
2328 if (PredBBTI->getSuccessor(i) == BB) {
2329 BB->removePredecessor(PredBB);
2330 PredBBTI->setSuccessor(i, EdgeBB);
2331 }
2332
2333 // Recurse, simplifying any other constants.
2334 return FoldCondBranchOnPHI(BI, DL, AC) || true;
2335 }
2336
2337 return false;
2338 }
2339
2340 /// Given a BB that starts with the specified two-entry PHI node,
2341 /// see if we can eliminate it.
FoldTwoEntryPHINode(PHINode * PN,const TargetTransformInfo & TTI,const DataLayout & DL)2342 static bool FoldTwoEntryPHINode(PHINode *PN, const TargetTransformInfo &TTI,
2343 const DataLayout &DL) {
2344 // Ok, this is a two entry PHI node. Check to see if this is a simple "if
2345 // statement", which has a very simple dominance structure. Basically, we
2346 // are trying to find the condition that is being branched on, which
2347 // subsequently causes this merge to happen. We really want control
2348 // dependence information for this check, but simplifycfg can't keep it up
2349 // to date, and this catches most of the cases we care about anyway.
2350 BasicBlock *BB = PN->getParent();
2351
2352 BasicBlock *IfTrue, *IfFalse;
2353 Value *IfCond = GetIfCondition(BB, IfTrue, IfFalse);
2354 if (!IfCond ||
2355 // Don't bother if the branch will be constant folded trivially.
2356 isa<ConstantInt>(IfCond))
2357 return false;
2358
2359 // Okay, we found that we can merge this two-entry phi node into a select.
2360 // Doing so would require us to fold *all* two entry phi nodes in this block.
2361 // At some point this becomes non-profitable (particularly if the target
2362 // doesn't support cmov's). Only do this transformation if there are two or
2363 // fewer PHI nodes in this block.
2364 unsigned NumPhis = 0;
2365 for (BasicBlock::iterator I = BB->begin(); isa<PHINode>(I); ++NumPhis, ++I)
2366 if (NumPhis > 2)
2367 return false;
2368
2369 // Loop over the PHI's seeing if we can promote them all to select
2370 // instructions. While we are at it, keep track of the instructions
2371 // that need to be moved to the dominating block.
2372 SmallPtrSet<Instruction *, 4> AggressiveInsts;
2373 int BudgetRemaining =
2374 TwoEntryPHINodeFoldingThreshold * TargetTransformInfo::TCC_Basic;
2375
2376 for (BasicBlock::iterator II = BB->begin(); isa<PHINode>(II);) {
2377 PHINode *PN = cast<PHINode>(II++);
2378 if (Value *V = SimplifyInstruction(PN, {DL, PN})) {
2379 PN->replaceAllUsesWith(V);
2380 PN->eraseFromParent();
2381 continue;
2382 }
2383
2384 if (!DominatesMergePoint(PN->getIncomingValue(0), BB, AggressiveInsts,
2385 BudgetRemaining, TTI) ||
2386 !DominatesMergePoint(PN->getIncomingValue(1), BB, AggressiveInsts,
2387 BudgetRemaining, TTI))
2388 return false;
2389 }
2390
2391 // If we folded the first phi, PN dangles at this point. Refresh it. If
2392 // we ran out of PHIs then we simplified them all.
2393 PN = dyn_cast<PHINode>(BB->begin());
2394 if (!PN)
2395 return true;
2396
2397 // Return true if at least one of these is a 'not', and another is either
2398 // a 'not' too, or a constant.
2399 auto CanHoistNotFromBothValues = [](Value *V0, Value *V1) {
2400 if (!match(V0, m_Not(m_Value())))
2401 std::swap(V0, V1);
2402 auto Invertible = m_CombineOr(m_Not(m_Value()), m_AnyIntegralConstant());
2403 return match(V0, m_Not(m_Value())) && match(V1, Invertible);
2404 };
2405
2406 // Don't fold i1 branches on PHIs which contain binary operators, unless one
2407 // of the incoming values is an 'not' and another one is freely invertible.
2408 // These can often be turned into switches and other things.
2409 if (PN->getType()->isIntegerTy(1) &&
2410 (isa<BinaryOperator>(PN->getIncomingValue(0)) ||
2411 isa<BinaryOperator>(PN->getIncomingValue(1)) ||
2412 isa<BinaryOperator>(IfCond)) &&
2413 !CanHoistNotFromBothValues(PN->getIncomingValue(0),
2414 PN->getIncomingValue(1)))
2415 return false;
2416
2417 // If all PHI nodes are promotable, check to make sure that all instructions
2418 // in the predecessor blocks can be promoted as well. If not, we won't be able
2419 // to get rid of the control flow, so it's not worth promoting to select
2420 // instructions.
2421 BasicBlock *DomBlock = nullptr;
2422 BasicBlock *IfBlock1 = PN->getIncomingBlock(0);
2423 BasicBlock *IfBlock2 = PN->getIncomingBlock(1);
2424 if (cast<BranchInst>(IfBlock1->getTerminator())->isConditional()) {
2425 IfBlock1 = nullptr;
2426 } else {
2427 DomBlock = *pred_begin(IfBlock1);
2428 for (BasicBlock::iterator I = IfBlock1->begin(); !I->isTerminator(); ++I)
2429 if (!AggressiveInsts.count(&*I) && !isa<DbgInfoIntrinsic>(I)) {
2430 // This is not an aggressive instruction that we can promote.
2431 // Because of this, we won't be able to get rid of the control flow, so
2432 // the xform is not worth it.
2433 return false;
2434 }
2435 }
2436
2437 if (cast<BranchInst>(IfBlock2->getTerminator())->isConditional()) {
2438 IfBlock2 = nullptr;
2439 } else {
2440 DomBlock = *pred_begin(IfBlock2);
2441 for (BasicBlock::iterator I = IfBlock2->begin(); !I->isTerminator(); ++I)
2442 if (!AggressiveInsts.count(&*I) && !isa<DbgInfoIntrinsic>(I)) {
2443 // This is not an aggressive instruction that we can promote.
2444 // Because of this, we won't be able to get rid of the control flow, so
2445 // the xform is not worth it.
2446 return false;
2447 }
2448 }
2449 assert(DomBlock && "Failed to find root DomBlock");
2450
2451 LLVM_DEBUG(dbgs() << "FOUND IF CONDITION! " << *IfCond
2452 << " T: " << IfTrue->getName()
2453 << " F: " << IfFalse->getName() << "\n");
2454
2455 // If we can still promote the PHI nodes after this gauntlet of tests,
2456 // do all of the PHI's now.
2457 Instruction *InsertPt = DomBlock->getTerminator();
2458 IRBuilder<NoFolder> Builder(InsertPt);
2459
2460 // Move all 'aggressive' instructions, which are defined in the
2461 // conditional parts of the if's up to the dominating block.
2462 if (IfBlock1)
2463 hoistAllInstructionsInto(DomBlock, InsertPt, IfBlock1);
2464 if (IfBlock2)
2465 hoistAllInstructionsInto(DomBlock, InsertPt, IfBlock2);
2466
2467 // Propagate fast-math-flags from phi nodes to replacement selects.
2468 IRBuilder<>::FastMathFlagGuard FMFGuard(Builder);
2469 while (PHINode *PN = dyn_cast<PHINode>(BB->begin())) {
2470 if (isa<FPMathOperator>(PN))
2471 Builder.setFastMathFlags(PN->getFastMathFlags());
2472
2473 // Change the PHI node into a select instruction.
2474 Value *TrueVal = PN->getIncomingValue(PN->getIncomingBlock(0) == IfFalse);
2475 Value *FalseVal = PN->getIncomingValue(PN->getIncomingBlock(0) == IfTrue);
2476
2477 Value *Sel = Builder.CreateSelect(IfCond, TrueVal, FalseVal, "", InsertPt);
2478 PN->replaceAllUsesWith(Sel);
2479 Sel->takeName(PN);
2480 PN->eraseFromParent();
2481 }
2482
2483 // At this point, IfBlock1 and IfBlock2 are both empty, so our if statement
2484 // has been flattened. Change DomBlock to jump directly to our new block to
2485 // avoid other simplifycfg's kicking in on the diamond.
2486 Instruction *OldTI = DomBlock->getTerminator();
2487 Builder.SetInsertPoint(OldTI);
2488 Builder.CreateBr(BB);
2489 OldTI->eraseFromParent();
2490 return true;
2491 }
2492
2493 /// If we found a conditional branch that goes to two returning blocks,
2494 /// try to merge them together into one return,
2495 /// introducing a select if the return values disagree.
SimplifyCondBranchToTwoReturns(BranchInst * BI,IRBuilder<> & Builder)2496 bool SimplifyCFGOpt::SimplifyCondBranchToTwoReturns(BranchInst *BI,
2497 IRBuilder<> &Builder) {
2498 assert(BI->isConditional() && "Must be a conditional branch");
2499 BasicBlock *TrueSucc = BI->getSuccessor(0);
2500 BasicBlock *FalseSucc = BI->getSuccessor(1);
2501 ReturnInst *TrueRet = cast<ReturnInst>(TrueSucc->getTerminator());
2502 ReturnInst *FalseRet = cast<ReturnInst>(FalseSucc->getTerminator());
2503
2504 // Check to ensure both blocks are empty (just a return) or optionally empty
2505 // with PHI nodes. If there are other instructions, merging would cause extra
2506 // computation on one path or the other.
2507 if (!TrueSucc->getFirstNonPHIOrDbg()->isTerminator())
2508 return false;
2509 if (!FalseSucc->getFirstNonPHIOrDbg()->isTerminator())
2510 return false;
2511
2512 Builder.SetInsertPoint(BI);
2513 // Okay, we found a branch that is going to two return nodes. If
2514 // there is no return value for this function, just change the
2515 // branch into a return.
2516 if (FalseRet->getNumOperands() == 0) {
2517 TrueSucc->removePredecessor(BI->getParent());
2518 FalseSucc->removePredecessor(BI->getParent());
2519 Builder.CreateRetVoid();
2520 EraseTerminatorAndDCECond(BI);
2521 return true;
2522 }
2523
2524 // Otherwise, figure out what the true and false return values are
2525 // so we can insert a new select instruction.
2526 Value *TrueValue = TrueRet->getReturnValue();
2527 Value *FalseValue = FalseRet->getReturnValue();
2528
2529 // Unwrap any PHI nodes in the return blocks.
2530 if (PHINode *TVPN = dyn_cast_or_null<PHINode>(TrueValue))
2531 if (TVPN->getParent() == TrueSucc)
2532 TrueValue = TVPN->getIncomingValueForBlock(BI->getParent());
2533 if (PHINode *FVPN = dyn_cast_or_null<PHINode>(FalseValue))
2534 if (FVPN->getParent() == FalseSucc)
2535 FalseValue = FVPN->getIncomingValueForBlock(BI->getParent());
2536
2537 // In order for this transformation to be safe, we must be able to
2538 // unconditionally execute both operands to the return. This is
2539 // normally the case, but we could have a potentially-trapping
2540 // constant expression that prevents this transformation from being
2541 // safe.
2542 if (ConstantExpr *TCV = dyn_cast_or_null<ConstantExpr>(TrueValue))
2543 if (TCV->canTrap())
2544 return false;
2545 if (ConstantExpr *FCV = dyn_cast_or_null<ConstantExpr>(FalseValue))
2546 if (FCV->canTrap())
2547 return false;
2548
2549 // Okay, we collected all the mapped values and checked them for sanity, and
2550 // defined to really do this transformation. First, update the CFG.
2551 TrueSucc->removePredecessor(BI->getParent());
2552 FalseSucc->removePredecessor(BI->getParent());
2553
2554 // Insert select instructions where needed.
2555 Value *BrCond = BI->getCondition();
2556 if (TrueValue) {
2557 // Insert a select if the results differ.
2558 if (TrueValue == FalseValue || isa<UndefValue>(FalseValue)) {
2559 } else if (isa<UndefValue>(TrueValue)) {
2560 TrueValue = FalseValue;
2561 } else {
2562 TrueValue =
2563 Builder.CreateSelect(BrCond, TrueValue, FalseValue, "retval", BI);
2564 }
2565 }
2566
2567 Value *RI =
2568 !TrueValue ? Builder.CreateRetVoid() : Builder.CreateRet(TrueValue);
2569
2570 (void)RI;
2571
2572 LLVM_DEBUG(dbgs() << "\nCHANGING BRANCH TO TWO RETURNS INTO SELECT:"
2573 << "\n " << *BI << "\nNewRet = " << *RI << "\nTRUEBLOCK: "
2574 << *TrueSucc << "\nFALSEBLOCK: " << *FalseSucc);
2575
2576 EraseTerminatorAndDCECond(BI);
2577
2578 return true;
2579 }
2580
2581 /// Return true if the given instruction is available
2582 /// in its predecessor block. If yes, the instruction will be removed.
tryCSEWithPredecessor(Instruction * Inst,BasicBlock * PB)2583 static bool tryCSEWithPredecessor(Instruction *Inst, BasicBlock *PB) {
2584 if (!isa<BinaryOperator>(Inst) && !isa<CmpInst>(Inst))
2585 return false;
2586 for (Instruction &I : *PB) {
2587 Instruction *PBI = &I;
2588 // Check whether Inst and PBI generate the same value.
2589 if (Inst->isIdenticalTo(PBI)) {
2590 Inst->replaceAllUsesWith(PBI);
2591 Inst->eraseFromParent();
2592 return true;
2593 }
2594 }
2595 return false;
2596 }
2597
2598 /// Return true if either PBI or BI has branch weight available, and store
2599 /// the weights in {Pred|Succ}{True|False}Weight. If one of PBI and BI does
2600 /// 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)2601 static bool extractPredSuccWeights(BranchInst *PBI, BranchInst *BI,
2602 uint64_t &PredTrueWeight,
2603 uint64_t &PredFalseWeight,
2604 uint64_t &SuccTrueWeight,
2605 uint64_t &SuccFalseWeight) {
2606 bool PredHasWeights =
2607 PBI->extractProfMetadata(PredTrueWeight, PredFalseWeight);
2608 bool SuccHasWeights =
2609 BI->extractProfMetadata(SuccTrueWeight, SuccFalseWeight);
2610 if (PredHasWeights || SuccHasWeights) {
2611 if (!PredHasWeights)
2612 PredTrueWeight = PredFalseWeight = 1;
2613 if (!SuccHasWeights)
2614 SuccTrueWeight = SuccFalseWeight = 1;
2615 return true;
2616 } else {
2617 return false;
2618 }
2619 }
2620
2621 /// If this basic block is simple enough, and if a predecessor branches to us
2622 /// and one of our successors, fold the block into the predecessor and use
2623 /// logical operations to pick the right destination.
FoldBranchToCommonDest(BranchInst * BI,MemorySSAUpdater * MSSAU,unsigned BonusInstThreshold)2624 bool llvm::FoldBranchToCommonDest(BranchInst *BI, MemorySSAUpdater *MSSAU,
2625 unsigned BonusInstThreshold) {
2626 BasicBlock *BB = BI->getParent();
2627
2628 const unsigned PredCount = pred_size(BB);
2629
2630 bool Changed = false;
2631
2632 Instruction *Cond = nullptr;
2633 if (BI->isConditional())
2634 Cond = dyn_cast<Instruction>(BI->getCondition());
2635 else {
2636 // For unconditional branch, check for a simple CFG pattern, where
2637 // BB has a single predecessor and BB's successor is also its predecessor's
2638 // successor. If such pattern exists, check for CSE between BB and its
2639 // predecessor.
2640 if (BasicBlock *PB = BB->getSinglePredecessor())
2641 if (BranchInst *PBI = dyn_cast<BranchInst>(PB->getTerminator()))
2642 if (PBI->isConditional() &&
2643 (BI->getSuccessor(0) == PBI->getSuccessor(0) ||
2644 BI->getSuccessor(0) == PBI->getSuccessor(1))) {
2645 for (auto I = BB->instructionsWithoutDebug().begin(),
2646 E = BB->instructionsWithoutDebug().end();
2647 I != E;) {
2648 Instruction *Curr = &*I++;
2649 if (isa<CmpInst>(Curr)) {
2650 Cond = Curr;
2651 break;
2652 }
2653 // Quit if we can't remove this instruction.
2654 if (!tryCSEWithPredecessor(Curr, PB))
2655 return Changed;
2656 Changed = true;
2657 }
2658 }
2659
2660 if (!Cond)
2661 return Changed;
2662 }
2663
2664 if (!Cond || (!isa<CmpInst>(Cond) && !isa<BinaryOperator>(Cond)) ||
2665 Cond->getParent() != BB || !Cond->hasOneUse())
2666 return Changed;
2667
2668 // Make sure the instruction after the condition is the cond branch.
2669 BasicBlock::iterator CondIt = ++Cond->getIterator();
2670
2671 // Ignore dbg intrinsics.
2672 while (isa<DbgInfoIntrinsic>(CondIt))
2673 ++CondIt;
2674
2675 if (&*CondIt != BI)
2676 return Changed;
2677
2678 // Only allow this transformation if computing the condition doesn't involve
2679 // too many instructions and these involved instructions can be executed
2680 // unconditionally. We denote all involved instructions except the condition
2681 // as "bonus instructions", and only allow this transformation when the
2682 // number of the bonus instructions we'll need to create when cloning into
2683 // each predecessor does not exceed a certain threshold.
2684 unsigned NumBonusInsts = 0;
2685 for (auto I = BB->begin(); Cond != &*I; ++I) {
2686 // Ignore dbg intrinsics.
2687 if (isa<DbgInfoIntrinsic>(I))
2688 continue;
2689 if (!I->hasOneUse() || !isSafeToSpeculativelyExecute(&*I))
2690 return Changed;
2691 // I has only one use and can be executed unconditionally.
2692 Instruction *User = dyn_cast<Instruction>(I->user_back());
2693 if (User == nullptr || User->getParent() != BB)
2694 return Changed;
2695 // I is used in the same BB. Since BI uses Cond and doesn't have more slots
2696 // to use any other instruction, User must be an instruction between next(I)
2697 // and Cond.
2698
2699 // Account for the cost of duplicating this instruction into each
2700 // predecessor.
2701 NumBonusInsts += PredCount;
2702 // Early exits once we reach the limit.
2703 if (NumBonusInsts > BonusInstThreshold)
2704 return Changed;
2705 }
2706
2707 // Cond is known to be a compare or binary operator. Check to make sure that
2708 // neither operand is a potentially-trapping constant expression.
2709 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(0)))
2710 if (CE->canTrap())
2711 return Changed;
2712 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(1)))
2713 if (CE->canTrap())
2714 return Changed;
2715
2716 // Finally, don't infinitely unroll conditional loops.
2717 BasicBlock *TrueDest = BI->getSuccessor(0);
2718 BasicBlock *FalseDest = (BI->isConditional()) ? BI->getSuccessor(1) : nullptr;
2719 if (TrueDest == BB || FalseDest == BB)
2720 return Changed;
2721
2722 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
2723 BasicBlock *PredBlock = *PI;
2724 BranchInst *PBI = dyn_cast<BranchInst>(PredBlock->getTerminator());
2725
2726 // Check that we have two conditional branches. If there is a PHI node in
2727 // the common successor, verify that the same value flows in from both
2728 // blocks.
2729 SmallVector<PHINode *, 4> PHIs;
2730 if (!PBI || PBI->isUnconditional() ||
2731 (BI->isConditional() && !SafeToMergeTerminators(BI, PBI)) ||
2732 (!BI->isConditional() &&
2733 !isProfitableToFoldUnconditional(BI, PBI, Cond, PHIs)))
2734 continue;
2735
2736 // Determine if the two branches share a common destination.
2737 Instruction::BinaryOps Opc = Instruction::BinaryOpsEnd;
2738 bool InvertPredCond = false;
2739
2740 if (BI->isConditional()) {
2741 if (PBI->getSuccessor(0) == TrueDest) {
2742 Opc = Instruction::Or;
2743 } else if (PBI->getSuccessor(1) == FalseDest) {
2744 Opc = Instruction::And;
2745 } else if (PBI->getSuccessor(0) == FalseDest) {
2746 Opc = Instruction::And;
2747 InvertPredCond = true;
2748 } else if (PBI->getSuccessor(1) == TrueDest) {
2749 Opc = Instruction::Or;
2750 InvertPredCond = true;
2751 } else {
2752 continue;
2753 }
2754 } else {
2755 if (PBI->getSuccessor(0) != TrueDest && PBI->getSuccessor(1) != TrueDest)
2756 continue;
2757 }
2758
2759 LLVM_DEBUG(dbgs() << "FOLDING BRANCH TO COMMON DEST:\n" << *PBI << *BB);
2760 Changed = true;
2761
2762 IRBuilder<> Builder(PBI);
2763
2764 // If we need to invert the condition in the pred block to match, do so now.
2765 if (InvertPredCond) {
2766 Value *NewCond = PBI->getCondition();
2767
2768 if (NewCond->hasOneUse() && isa<CmpInst>(NewCond)) {
2769 CmpInst *CI = cast<CmpInst>(NewCond);
2770 CI->setPredicate(CI->getInversePredicate());
2771 } else {
2772 NewCond =
2773 Builder.CreateNot(NewCond, PBI->getCondition()->getName() + ".not");
2774 }
2775
2776 PBI->setCondition(NewCond);
2777 PBI->swapSuccessors();
2778 }
2779
2780 // If we have bonus instructions, clone them into the predecessor block.
2781 // Note that there may be multiple predecessor blocks, so we cannot move
2782 // bonus instructions to a predecessor block.
2783 ValueToValueMapTy VMap; // maps original values to cloned values
2784 // We already make sure Cond is the last instruction before BI. Therefore,
2785 // all instructions before Cond other than DbgInfoIntrinsic are bonus
2786 // instructions.
2787 for (auto BonusInst = BB->begin(); Cond != &*BonusInst; ++BonusInst) {
2788 if (isa<DbgInfoIntrinsic>(BonusInst))
2789 continue;
2790 Instruction *NewBonusInst = BonusInst->clone();
2791
2792 // When we fold the bonus instructions we want to make sure we
2793 // reset their debug locations in order to avoid stepping on dead
2794 // code caused by folding dead branches.
2795 NewBonusInst->setDebugLoc(DebugLoc());
2796
2797 RemapInstruction(NewBonusInst, VMap,
2798 RF_NoModuleLevelChanges | RF_IgnoreMissingLocals);
2799 VMap[&*BonusInst] = NewBonusInst;
2800
2801 // If we moved a load, we cannot any longer claim any knowledge about
2802 // its potential value. The previous information might have been valid
2803 // only given the branch precondition.
2804 // For an analogous reason, we must also drop all the metadata whose
2805 // semantics we don't understand.
2806 NewBonusInst->dropUnknownNonDebugMetadata();
2807
2808 PredBlock->getInstList().insert(PBI->getIterator(), NewBonusInst);
2809 NewBonusInst->takeName(&*BonusInst);
2810 BonusInst->setName(BonusInst->getName() + ".old");
2811 }
2812
2813 // Clone Cond into the predecessor basic block, and or/and the
2814 // two conditions together.
2815 Instruction *CondInPred = Cond->clone();
2816
2817 // Reset the condition debug location to avoid jumping on dead code
2818 // as the result of folding dead branches.
2819 CondInPred->setDebugLoc(DebugLoc());
2820
2821 RemapInstruction(CondInPred, VMap,
2822 RF_NoModuleLevelChanges | RF_IgnoreMissingLocals);
2823 PredBlock->getInstList().insert(PBI->getIterator(), CondInPred);
2824 CondInPred->takeName(Cond);
2825 Cond->setName(CondInPred->getName() + ".old");
2826
2827 if (BI->isConditional()) {
2828 Instruction *NewCond = cast<Instruction>(
2829 Builder.CreateBinOp(Opc, PBI->getCondition(), CondInPred, "or.cond"));
2830 PBI->setCondition(NewCond);
2831
2832 uint64_t PredTrueWeight, PredFalseWeight, SuccTrueWeight, SuccFalseWeight;
2833 bool HasWeights =
2834 extractPredSuccWeights(PBI, BI, PredTrueWeight, PredFalseWeight,
2835 SuccTrueWeight, SuccFalseWeight);
2836 SmallVector<uint64_t, 8> NewWeights;
2837
2838 if (PBI->getSuccessor(0) == BB) {
2839 if (HasWeights) {
2840 // PBI: br i1 %x, BB, FalseDest
2841 // BI: br i1 %y, TrueDest, FalseDest
2842 // TrueWeight is TrueWeight for PBI * TrueWeight for BI.
2843 NewWeights.push_back(PredTrueWeight * SuccTrueWeight);
2844 // FalseWeight is FalseWeight for PBI * TotalWeight for BI +
2845 // TrueWeight for PBI * FalseWeight for BI.
2846 // We assume that total weights of a BranchInst can fit into 32 bits.
2847 // Therefore, we will not have overflow using 64-bit arithmetic.
2848 NewWeights.push_back(PredFalseWeight *
2849 (SuccFalseWeight + SuccTrueWeight) +
2850 PredTrueWeight * SuccFalseWeight);
2851 }
2852 AddPredecessorToBlock(TrueDest, PredBlock, BB, MSSAU);
2853 PBI->setSuccessor(0, TrueDest);
2854 }
2855 if (PBI->getSuccessor(1) == BB) {
2856 if (HasWeights) {
2857 // PBI: br i1 %x, TrueDest, BB
2858 // BI: br i1 %y, TrueDest, FalseDest
2859 // TrueWeight is TrueWeight for PBI * TotalWeight for BI +
2860 // FalseWeight for PBI * TrueWeight for BI.
2861 NewWeights.push_back(PredTrueWeight *
2862 (SuccFalseWeight + SuccTrueWeight) +
2863 PredFalseWeight * SuccTrueWeight);
2864 // FalseWeight is FalseWeight for PBI * FalseWeight for BI.
2865 NewWeights.push_back(PredFalseWeight * SuccFalseWeight);
2866 }
2867 AddPredecessorToBlock(FalseDest, PredBlock, BB, MSSAU);
2868 PBI->setSuccessor(1, FalseDest);
2869 }
2870 if (NewWeights.size() == 2) {
2871 // Halve the weights if any of them cannot fit in an uint32_t
2872 FitWeights(NewWeights);
2873
2874 SmallVector<uint32_t, 8> MDWeights(NewWeights.begin(),
2875 NewWeights.end());
2876 setBranchWeights(PBI, MDWeights[0], MDWeights[1]);
2877 } else
2878 PBI->setMetadata(LLVMContext::MD_prof, nullptr);
2879 } else {
2880 // Update PHI nodes in the common successors.
2881 for (unsigned i = 0, e = PHIs.size(); i != e; ++i) {
2882 ConstantInt *PBI_C = cast<ConstantInt>(
2883 PHIs[i]->getIncomingValueForBlock(PBI->getParent()));
2884 assert(PBI_C->getType()->isIntegerTy(1));
2885 Instruction *MergedCond = nullptr;
2886 if (PBI->getSuccessor(0) == TrueDest) {
2887 // Create (PBI_Cond and PBI_C) or (!PBI_Cond and BI_Value)
2888 // PBI_C is true: PBI_Cond or (!PBI_Cond and BI_Value)
2889 // is false: !PBI_Cond and BI_Value
2890 Instruction *NotCond = cast<Instruction>(
2891 Builder.CreateNot(PBI->getCondition(), "not.cond"));
2892 MergedCond = cast<Instruction>(
2893 Builder.CreateBinOp(Instruction::And, NotCond, CondInPred,
2894 "and.cond"));
2895 if (PBI_C->isOne())
2896 MergedCond = cast<Instruction>(Builder.CreateBinOp(
2897 Instruction::Or, PBI->getCondition(), MergedCond, "or.cond"));
2898 } else {
2899 // Create (PBI_Cond and BI_Value) or (!PBI_Cond and PBI_C)
2900 // PBI_C is true: (PBI_Cond and BI_Value) or (!PBI_Cond)
2901 // is false: PBI_Cond and BI_Value
2902 MergedCond = cast<Instruction>(Builder.CreateBinOp(
2903 Instruction::And, PBI->getCondition(), CondInPred, "and.cond"));
2904 if (PBI_C->isOne()) {
2905 Instruction *NotCond = cast<Instruction>(
2906 Builder.CreateNot(PBI->getCondition(), "not.cond"));
2907 MergedCond = cast<Instruction>(Builder.CreateBinOp(
2908 Instruction::Or, NotCond, MergedCond, "or.cond"));
2909 }
2910 }
2911 // Update PHI Node.
2912 PHIs[i]->setIncomingValueForBlock(PBI->getParent(), MergedCond);
2913 }
2914
2915 // PBI is changed to branch to TrueDest below. Remove itself from
2916 // potential phis from all other successors.
2917 if (MSSAU)
2918 MSSAU->changeCondBranchToUnconditionalTo(PBI, TrueDest);
2919
2920 // Change PBI from Conditional to Unconditional.
2921 BranchInst *New_PBI = BranchInst::Create(TrueDest, PBI);
2922 EraseTerminatorAndDCECond(PBI, MSSAU);
2923 PBI = New_PBI;
2924 }
2925
2926 // If BI was a loop latch, it may have had associated loop metadata.
2927 // We need to copy it to the new latch, that is, PBI.
2928 if (MDNode *LoopMD = BI->getMetadata(LLVMContext::MD_loop))
2929 PBI->setMetadata(LLVMContext::MD_loop, LoopMD);
2930
2931 // TODO: If BB is reachable from all paths through PredBlock, then we
2932 // could replace PBI's branch probabilities with BI's.
2933
2934 // Copy any debug value intrinsics into the end of PredBlock.
2935 for (Instruction &I : *BB) {
2936 if (isa<DbgInfoIntrinsic>(I)) {
2937 Instruction *NewI = I.clone();
2938 RemapInstruction(NewI, VMap,
2939 RF_NoModuleLevelChanges | RF_IgnoreMissingLocals);
2940 NewI->insertBefore(PBI);
2941 }
2942 }
2943
2944 return Changed;
2945 }
2946 return Changed;
2947 }
2948
2949 // If there is only one store in BB1 and BB2, return it, otherwise return
2950 // nullptr.
findUniqueStoreInBlocks(BasicBlock * BB1,BasicBlock * BB2)2951 static StoreInst *findUniqueStoreInBlocks(BasicBlock *BB1, BasicBlock *BB2) {
2952 StoreInst *S = nullptr;
2953 for (auto *BB : {BB1, BB2}) {
2954 if (!BB)
2955 continue;
2956 for (auto &I : *BB)
2957 if (auto *SI = dyn_cast<StoreInst>(&I)) {
2958 if (S)
2959 // Multiple stores seen.
2960 return nullptr;
2961 else
2962 S = SI;
2963 }
2964 }
2965 return S;
2966 }
2967
ensureValueAvailableInSuccessor(Value * V,BasicBlock * BB,Value * AlternativeV=nullptr)2968 static Value *ensureValueAvailableInSuccessor(Value *V, BasicBlock *BB,
2969 Value *AlternativeV = nullptr) {
2970 // PHI is going to be a PHI node that allows the value V that is defined in
2971 // BB to be referenced in BB's only successor.
2972 //
2973 // If AlternativeV is nullptr, the only value we care about in PHI is V. It
2974 // doesn't matter to us what the other operand is (it'll never get used). We
2975 // could just create a new PHI with an undef incoming value, but that could
2976 // increase register pressure if EarlyCSE/InstCombine can't fold it with some
2977 // other PHI. So here we directly look for some PHI in BB's successor with V
2978 // as an incoming operand. If we find one, we use it, else we create a new
2979 // one.
2980 //
2981 // If AlternativeV is not nullptr, we care about both incoming values in PHI.
2982 // PHI must be exactly: phi <ty> [ %BB, %V ], [ %OtherBB, %AlternativeV]
2983 // where OtherBB is the single other predecessor of BB's only successor.
2984 PHINode *PHI = nullptr;
2985 BasicBlock *Succ = BB->getSingleSuccessor();
2986
2987 for (auto I = Succ->begin(); isa<PHINode>(I); ++I)
2988 if (cast<PHINode>(I)->getIncomingValueForBlock(BB) == V) {
2989 PHI = cast<PHINode>(I);
2990 if (!AlternativeV)
2991 break;
2992
2993 assert(Succ->hasNPredecessors(2));
2994 auto PredI = pred_begin(Succ);
2995 BasicBlock *OtherPredBB = *PredI == BB ? *++PredI : *PredI;
2996 if (PHI->getIncomingValueForBlock(OtherPredBB) == AlternativeV)
2997 break;
2998 PHI = nullptr;
2999 }
3000 if (PHI)
3001 return PHI;
3002
3003 // If V is not an instruction defined in BB, just return it.
3004 if (!AlternativeV &&
3005 (!isa<Instruction>(V) || cast<Instruction>(V)->getParent() != BB))
3006 return V;
3007
3008 PHI = PHINode::Create(V->getType(), 2, "simplifycfg.merge", &Succ->front());
3009 PHI->addIncoming(V, BB);
3010 for (BasicBlock *PredBB : predecessors(Succ))
3011 if (PredBB != BB)
3012 PHI->addIncoming(
3013 AlternativeV ? AlternativeV : UndefValue::get(V->getType()), PredBB);
3014 return PHI;
3015 }
3016
mergeConditionalStoreToAddress(BasicBlock * PTB,BasicBlock * PFB,BasicBlock * QTB,BasicBlock * QFB,BasicBlock * PostBB,Value * Address,bool InvertPCond,bool InvertQCond,const DataLayout & DL,const TargetTransformInfo & TTI)3017 static bool mergeConditionalStoreToAddress(BasicBlock *PTB, BasicBlock *PFB,
3018 BasicBlock *QTB, BasicBlock *QFB,
3019 BasicBlock *PostBB, Value *Address,
3020 bool InvertPCond, bool InvertQCond,
3021 const DataLayout &DL,
3022 const TargetTransformInfo &TTI) {
3023 // For every pointer, there must be exactly two stores, one coming from
3024 // PTB or PFB, and the other from QTB or QFB. We don't support more than one
3025 // store (to any address) in PTB,PFB or QTB,QFB.
3026 // FIXME: We could relax this restriction with a bit more work and performance
3027 // testing.
3028 StoreInst *PStore = findUniqueStoreInBlocks(PTB, PFB);
3029 StoreInst *QStore = findUniqueStoreInBlocks(QTB, QFB);
3030 if (!PStore || !QStore)
3031 return false;
3032
3033 // Now check the stores are compatible.
3034 if (!QStore->isUnordered() || !PStore->isUnordered())
3035 return false;
3036
3037 // Check that sinking the store won't cause program behavior changes. Sinking
3038 // the store out of the Q blocks won't change any behavior as we're sinking
3039 // from a block to its unconditional successor. But we're moving a store from
3040 // the P blocks down through the middle block (QBI) and past both QFB and QTB.
3041 // So we need to check that there are no aliasing loads or stores in
3042 // QBI, QTB and QFB. We also need to check there are no conflicting memory
3043 // operations between PStore and the end of its parent block.
3044 //
3045 // The ideal way to do this is to query AliasAnalysis, but we don't
3046 // preserve AA currently so that is dangerous. Be super safe and just
3047 // check there are no other memory operations at all.
3048 for (auto &I : *QFB->getSinglePredecessor())
3049 if (I.mayReadOrWriteMemory())
3050 return false;
3051 for (auto &I : *QFB)
3052 if (&I != QStore && I.mayReadOrWriteMemory())
3053 return false;
3054 if (QTB)
3055 for (auto &I : *QTB)
3056 if (&I != QStore && I.mayReadOrWriteMemory())
3057 return false;
3058 for (auto I = BasicBlock::iterator(PStore), E = PStore->getParent()->end();
3059 I != E; ++I)
3060 if (&*I != PStore && I->mayReadOrWriteMemory())
3061 return false;
3062
3063 // If we're not in aggressive mode, we only optimize if we have some
3064 // confidence that by optimizing we'll allow P and/or Q to be if-converted.
3065 auto IsWorthwhile = [&](BasicBlock *BB, ArrayRef<StoreInst *> FreeStores) {
3066 if (!BB)
3067 return true;
3068 // Heuristic: if the block can be if-converted/phi-folded and the
3069 // instructions inside are all cheap (arithmetic/GEPs), it's worthwhile to
3070 // thread this store.
3071 int BudgetRemaining =
3072 PHINodeFoldingThreshold * TargetTransformInfo::TCC_Basic;
3073 for (auto &I : BB->instructionsWithoutDebug()) {
3074 // Consider terminator instruction to be free.
3075 if (I.isTerminator())
3076 continue;
3077 // If this is one the stores that we want to speculate out of this BB,
3078 // then don't count it's cost, consider it to be free.
3079 if (auto *S = dyn_cast<StoreInst>(&I))
3080 if (llvm::find(FreeStores, S))
3081 continue;
3082 // Else, we have a white-list of instructions that we are ak speculating.
3083 if (!isa<BinaryOperator>(I) && !isa<GetElementPtrInst>(I))
3084 return false; // Not in white-list - not worthwhile folding.
3085 // And finally, if this is a non-free instruction that we are okay
3086 // speculating, ensure that we consider the speculation budget.
3087 BudgetRemaining -= TTI.getUserCost(&I, TargetTransformInfo::TCK_SizeAndLatency);
3088 if (BudgetRemaining < 0)
3089 return false; // Eagerly refuse to fold as soon as we're out of budget.
3090 }
3091 assert(BudgetRemaining >= 0 &&
3092 "When we run out of budget we will eagerly return from within the "
3093 "per-instruction loop.");
3094 return true;
3095 };
3096
3097 const SmallVector<StoreInst *, 2> FreeStores = {PStore, QStore};
3098 if (!MergeCondStoresAggressively &&
3099 (!IsWorthwhile(PTB, FreeStores) || !IsWorthwhile(PFB, FreeStores) ||
3100 !IsWorthwhile(QTB, FreeStores) || !IsWorthwhile(QFB, FreeStores)))
3101 return false;
3102
3103 // If PostBB has more than two predecessors, we need to split it so we can
3104 // sink the store.
3105 if (std::next(pred_begin(PostBB), 2) != pred_end(PostBB)) {
3106 // We know that QFB's only successor is PostBB. And QFB has a single
3107 // predecessor. If QTB exists, then its only successor is also PostBB.
3108 // If QTB does not exist, then QFB's only predecessor has a conditional
3109 // branch to QFB and PostBB.
3110 BasicBlock *TruePred = QTB ? QTB : QFB->getSinglePredecessor();
3111 BasicBlock *NewBB = SplitBlockPredecessors(PostBB, { QFB, TruePred},
3112 "condstore.split");
3113 if (!NewBB)
3114 return false;
3115 PostBB = NewBB;
3116 }
3117
3118 // OK, we're going to sink the stores to PostBB. The store has to be
3119 // conditional though, so first create the predicate.
3120 Value *PCond = cast<BranchInst>(PFB->getSinglePredecessor()->getTerminator())
3121 ->getCondition();
3122 Value *QCond = cast<BranchInst>(QFB->getSinglePredecessor()->getTerminator())
3123 ->getCondition();
3124
3125 Value *PPHI = ensureValueAvailableInSuccessor(PStore->getValueOperand(),
3126 PStore->getParent());
3127 Value *QPHI = ensureValueAvailableInSuccessor(QStore->getValueOperand(),
3128 QStore->getParent(), PPHI);
3129
3130 IRBuilder<> QB(&*PostBB->getFirstInsertionPt());
3131
3132 Value *PPred = PStore->getParent() == PTB ? PCond : QB.CreateNot(PCond);
3133 Value *QPred = QStore->getParent() == QTB ? QCond : QB.CreateNot(QCond);
3134
3135 if (InvertPCond)
3136 PPred = QB.CreateNot(PPred);
3137 if (InvertQCond)
3138 QPred = QB.CreateNot(QPred);
3139 Value *CombinedPred = QB.CreateOr(PPred, QPred);
3140
3141 auto *T =
3142 SplitBlockAndInsertIfThen(CombinedPred, &*QB.GetInsertPoint(), false);
3143 QB.SetInsertPoint(T);
3144 StoreInst *SI = cast<StoreInst>(QB.CreateStore(QPHI, Address));
3145 AAMDNodes AAMD;
3146 PStore->getAAMetadata(AAMD, /*Merge=*/false);
3147 PStore->getAAMetadata(AAMD, /*Merge=*/true);
3148 SI->setAAMetadata(AAMD);
3149 // Choose the minimum alignment. If we could prove both stores execute, we
3150 // could use biggest one. In this case, though, we only know that one of the
3151 // stores executes. And we don't know it's safe to take the alignment from a
3152 // store that doesn't execute.
3153 SI->setAlignment(std::min(PStore->getAlign(), QStore->getAlign()));
3154
3155 QStore->eraseFromParent();
3156 PStore->eraseFromParent();
3157
3158 return true;
3159 }
3160
mergeConditionalStores(BranchInst * PBI,BranchInst * QBI,const DataLayout & DL,const TargetTransformInfo & TTI)3161 static bool mergeConditionalStores(BranchInst *PBI, BranchInst *QBI,
3162 const DataLayout &DL,
3163 const TargetTransformInfo &TTI) {
3164 // The intention here is to find diamonds or triangles (see below) where each
3165 // conditional block contains a store to the same address. Both of these
3166 // stores are conditional, so they can't be unconditionally sunk. But it may
3167 // be profitable to speculatively sink the stores into one merged store at the
3168 // end, and predicate the merged store on the union of the two conditions of
3169 // PBI and QBI.
3170 //
3171 // This can reduce the number of stores executed if both of the conditions are
3172 // true, and can allow the blocks to become small enough to be if-converted.
3173 // This optimization will also chain, so that ladders of test-and-set
3174 // sequences can be if-converted away.
3175 //
3176 // We only deal with simple diamonds or triangles:
3177 //
3178 // PBI or PBI or a combination of the two
3179 // / \ | \
3180 // PTB PFB | PFB
3181 // \ / | /
3182 // QBI QBI
3183 // / \ | \
3184 // QTB QFB | QFB
3185 // \ / | /
3186 // PostBB PostBB
3187 //
3188 // We model triangles as a type of diamond with a nullptr "true" block.
3189 // Triangles are canonicalized so that the fallthrough edge is represented by
3190 // a true condition, as in the diagram above.
3191 BasicBlock *PTB = PBI->getSuccessor(0);
3192 BasicBlock *PFB = PBI->getSuccessor(1);
3193 BasicBlock *QTB = QBI->getSuccessor(0);
3194 BasicBlock *QFB = QBI->getSuccessor(1);
3195 BasicBlock *PostBB = QFB->getSingleSuccessor();
3196
3197 // Make sure we have a good guess for PostBB. If QTB's only successor is
3198 // QFB, then QFB is a better PostBB.
3199 if (QTB->getSingleSuccessor() == QFB)
3200 PostBB = QFB;
3201
3202 // If we couldn't find a good PostBB, stop.
3203 if (!PostBB)
3204 return false;
3205
3206 bool InvertPCond = false, InvertQCond = false;
3207 // Canonicalize fallthroughs to the true branches.
3208 if (PFB == QBI->getParent()) {
3209 std::swap(PFB, PTB);
3210 InvertPCond = true;
3211 }
3212 if (QFB == PostBB) {
3213 std::swap(QFB, QTB);
3214 InvertQCond = true;
3215 }
3216
3217 // From this point on we can assume PTB or QTB may be fallthroughs but PFB
3218 // and QFB may not. Model fallthroughs as a nullptr block.
3219 if (PTB == QBI->getParent())
3220 PTB = nullptr;
3221 if (QTB == PostBB)
3222 QTB = nullptr;
3223
3224 // Legality bailouts. We must have at least the non-fallthrough blocks and
3225 // the post-dominating block, and the non-fallthroughs must only have one
3226 // predecessor.
3227 auto HasOnePredAndOneSucc = [](BasicBlock *BB, BasicBlock *P, BasicBlock *S) {
3228 return BB->getSinglePredecessor() == P && BB->getSingleSuccessor() == S;
3229 };
3230 if (!HasOnePredAndOneSucc(PFB, PBI->getParent(), QBI->getParent()) ||
3231 !HasOnePredAndOneSucc(QFB, QBI->getParent(), PostBB))
3232 return false;
3233 if ((PTB && !HasOnePredAndOneSucc(PTB, PBI->getParent(), QBI->getParent())) ||
3234 (QTB && !HasOnePredAndOneSucc(QTB, QBI->getParent(), PostBB)))
3235 return false;
3236 if (!QBI->getParent()->hasNUses(2))
3237 return false;
3238
3239 // OK, this is a sequence of two diamonds or triangles.
3240 // Check if there are stores in PTB or PFB that are repeated in QTB or QFB.
3241 SmallPtrSet<Value *, 4> PStoreAddresses, QStoreAddresses;
3242 for (auto *BB : {PTB, PFB}) {
3243 if (!BB)
3244 continue;
3245 for (auto &I : *BB)
3246 if (StoreInst *SI = dyn_cast<StoreInst>(&I))
3247 PStoreAddresses.insert(SI->getPointerOperand());
3248 }
3249 for (auto *BB : {QTB, QFB}) {
3250 if (!BB)
3251 continue;
3252 for (auto &I : *BB)
3253 if (StoreInst *SI = dyn_cast<StoreInst>(&I))
3254 QStoreAddresses.insert(SI->getPointerOperand());
3255 }
3256
3257 set_intersect(PStoreAddresses, QStoreAddresses);
3258 // set_intersect mutates PStoreAddresses in place. Rename it here to make it
3259 // clear what it contains.
3260 auto &CommonAddresses = PStoreAddresses;
3261
3262 bool Changed = false;
3263 for (auto *Address : CommonAddresses)
3264 Changed |= mergeConditionalStoreToAddress(
3265 PTB, PFB, QTB, QFB, PostBB, Address, InvertPCond, InvertQCond, DL, TTI);
3266 return Changed;
3267 }
3268
3269
3270 /// If the previous block ended with a widenable branch, determine if reusing
3271 /// the target block is profitable and legal. This will have the effect of
3272 /// "widening" PBI, but doesn't require us to reason about hosting safety.
tryWidenCondBranchToCondBranch(BranchInst * PBI,BranchInst * BI)3273 static bool tryWidenCondBranchToCondBranch(BranchInst *PBI, BranchInst *BI) {
3274 // TODO: This can be generalized in two important ways:
3275 // 1) We can allow phi nodes in IfFalseBB and simply reuse all the input
3276 // values from the PBI edge.
3277 // 2) We can sink side effecting instructions into BI's fallthrough
3278 // successor provided they doesn't contribute to computation of
3279 // BI's condition.
3280 Value *CondWB, *WC;
3281 BasicBlock *IfTrueBB, *IfFalseBB;
3282 if (!parseWidenableBranch(PBI, CondWB, WC, IfTrueBB, IfFalseBB) ||
3283 IfTrueBB != BI->getParent() || !BI->getParent()->getSinglePredecessor())
3284 return false;
3285 if (!IfFalseBB->phis().empty())
3286 return false; // TODO
3287 // Use lambda to lazily compute expensive condition after cheap ones.
3288 auto NoSideEffects = [](BasicBlock &BB) {
3289 return !llvm::any_of(BB, [](const Instruction &I) {
3290 return I.mayWriteToMemory() || I.mayHaveSideEffects();
3291 });
3292 };
3293 if (BI->getSuccessor(1) != IfFalseBB && // no inf looping
3294 BI->getSuccessor(1)->getTerminatingDeoptimizeCall() && // profitability
3295 NoSideEffects(*BI->getParent())) {
3296 BI->getSuccessor(1)->removePredecessor(BI->getParent());
3297 BI->setSuccessor(1, IfFalseBB);
3298 return true;
3299 }
3300 if (BI->getSuccessor(0) != IfFalseBB && // no inf looping
3301 BI->getSuccessor(0)->getTerminatingDeoptimizeCall() && // profitability
3302 NoSideEffects(*BI->getParent())) {
3303 BI->getSuccessor(0)->removePredecessor(BI->getParent());
3304 BI->setSuccessor(0, IfFalseBB);
3305 return true;
3306 }
3307 return false;
3308 }
3309
3310 /// If we have a conditional branch as a predecessor of another block,
3311 /// this function tries to simplify it. We know
3312 /// that PBI and BI are both conditional branches, and BI is in one of the
3313 /// successor blocks of PBI - PBI branches to BI.
SimplifyCondBranchToCondBranch(BranchInst * PBI,BranchInst * BI,const DataLayout & DL,const TargetTransformInfo & TTI)3314 static bool SimplifyCondBranchToCondBranch(BranchInst *PBI, BranchInst *BI,
3315 const DataLayout &DL,
3316 const TargetTransformInfo &TTI) {
3317 assert(PBI->isConditional() && BI->isConditional());
3318 BasicBlock *BB = BI->getParent();
3319
3320 // If this block ends with a branch instruction, and if there is a
3321 // predecessor that ends on a branch of the same condition, make
3322 // this conditional branch redundant.
3323 if (PBI->getCondition() == BI->getCondition() &&
3324 PBI->getSuccessor(0) != PBI->getSuccessor(1)) {
3325 // Okay, the outcome of this conditional branch is statically
3326 // knowable. If this block had a single pred, handle specially.
3327 if (BB->getSinglePredecessor()) {
3328 // Turn this into a branch on constant.
3329 bool CondIsTrue = PBI->getSuccessor(0) == BB;
3330 BI->setCondition(
3331 ConstantInt::get(Type::getInt1Ty(BB->getContext()), CondIsTrue));
3332 return true; // Nuke the branch on constant.
3333 }
3334
3335 // Otherwise, if there are multiple predecessors, insert a PHI that merges
3336 // in the constant and simplify the block result. Subsequent passes of
3337 // simplifycfg will thread the block.
3338 if (BlockIsSimpleEnoughToThreadThrough(BB)) {
3339 pred_iterator PB = pred_begin(BB), PE = pred_end(BB);
3340 PHINode *NewPN = PHINode::Create(
3341 Type::getInt1Ty(BB->getContext()), std::distance(PB, PE),
3342 BI->getCondition()->getName() + ".pr", &BB->front());
3343 // Okay, we're going to insert the PHI node. Since PBI is not the only
3344 // predecessor, compute the PHI'd conditional value for all of the preds.
3345 // Any predecessor where the condition is not computable we keep symbolic.
3346 for (pred_iterator PI = PB; PI != PE; ++PI) {
3347 BasicBlock *P = *PI;
3348 if ((PBI = dyn_cast<BranchInst>(P->getTerminator())) && PBI != BI &&
3349 PBI->isConditional() && PBI->getCondition() == BI->getCondition() &&
3350 PBI->getSuccessor(0) != PBI->getSuccessor(1)) {
3351 bool CondIsTrue = PBI->getSuccessor(0) == BB;
3352 NewPN->addIncoming(
3353 ConstantInt::get(Type::getInt1Ty(BB->getContext()), CondIsTrue),
3354 P);
3355 } else {
3356 NewPN->addIncoming(BI->getCondition(), P);
3357 }
3358 }
3359
3360 BI->setCondition(NewPN);
3361 return true;
3362 }
3363 }
3364
3365 // If the previous block ended with a widenable branch, determine if reusing
3366 // the target block is profitable and legal. This will have the effect of
3367 // "widening" PBI, but doesn't require us to reason about hosting safety.
3368 if (tryWidenCondBranchToCondBranch(PBI, BI))
3369 return true;
3370
3371 if (auto *CE = dyn_cast<ConstantExpr>(BI->getCondition()))
3372 if (CE->canTrap())
3373 return false;
3374
3375 // If both branches are conditional and both contain stores to the same
3376 // address, remove the stores from the conditionals and create a conditional
3377 // merged store at the end.
3378 if (MergeCondStores && mergeConditionalStores(PBI, BI, DL, TTI))
3379 return true;
3380
3381 // If this is a conditional branch in an empty block, and if any
3382 // predecessors are a conditional branch to one of our destinations,
3383 // fold the conditions into logical ops and one cond br.
3384
3385 // Ignore dbg intrinsics.
3386 if (&*BB->instructionsWithoutDebug().begin() != BI)
3387 return false;
3388
3389 int PBIOp, BIOp;
3390 if (PBI->getSuccessor(0) == BI->getSuccessor(0)) {
3391 PBIOp = 0;
3392 BIOp = 0;
3393 } else if (PBI->getSuccessor(0) == BI->getSuccessor(1)) {
3394 PBIOp = 0;
3395 BIOp = 1;
3396 } else if (PBI->getSuccessor(1) == BI->getSuccessor(0)) {
3397 PBIOp = 1;
3398 BIOp = 0;
3399 } else if (PBI->getSuccessor(1) == BI->getSuccessor(1)) {
3400 PBIOp = 1;
3401 BIOp = 1;
3402 } else {
3403 return false;
3404 }
3405
3406 // Check to make sure that the other destination of this branch
3407 // isn't BB itself. If so, this is an infinite loop that will
3408 // keep getting unwound.
3409 if (PBI->getSuccessor(PBIOp) == BB)
3410 return false;
3411
3412 // Do not perform this transformation if it would require
3413 // insertion of a large number of select instructions. For targets
3414 // without predication/cmovs, this is a big pessimization.
3415
3416 // Also do not perform this transformation if any phi node in the common
3417 // destination block can trap when reached by BB or PBB (PR17073). In that
3418 // case, it would be unsafe to hoist the operation into a select instruction.
3419
3420 BasicBlock *CommonDest = PBI->getSuccessor(PBIOp);
3421 unsigned NumPhis = 0;
3422 for (BasicBlock::iterator II = CommonDest->begin(); isa<PHINode>(II);
3423 ++II, ++NumPhis) {
3424 if (NumPhis > 2) // Disable this xform.
3425 return false;
3426
3427 PHINode *PN = cast<PHINode>(II);
3428 Value *BIV = PN->getIncomingValueForBlock(BB);
3429 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(BIV))
3430 if (CE->canTrap())
3431 return false;
3432
3433 unsigned PBBIdx = PN->getBasicBlockIndex(PBI->getParent());
3434 Value *PBIV = PN->getIncomingValue(PBBIdx);
3435 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(PBIV))
3436 if (CE->canTrap())
3437 return false;
3438 }
3439
3440 // Finally, if everything is ok, fold the branches to logical ops.
3441 BasicBlock *OtherDest = BI->getSuccessor(BIOp ^ 1);
3442
3443 LLVM_DEBUG(dbgs() << "FOLDING BRs:" << *PBI->getParent()
3444 << "AND: " << *BI->getParent());
3445
3446 // If OtherDest *is* BB, then BB is a basic block with a single conditional
3447 // branch in it, where one edge (OtherDest) goes back to itself but the other
3448 // exits. We don't *know* that the program avoids the infinite loop
3449 // (even though that seems likely). If we do this xform naively, we'll end up
3450 // recursively unpeeling the loop. Since we know that (after the xform is
3451 // done) that the block *is* infinite if reached, we just make it an obviously
3452 // infinite loop with no cond branch.
3453 if (OtherDest == BB) {
3454 // Insert it at the end of the function, because it's either code,
3455 // or it won't matter if it's hot. :)
3456 BasicBlock *InfLoopBlock =
3457 BasicBlock::Create(BB->getContext(), "infloop", BB->getParent());
3458 BranchInst::Create(InfLoopBlock, InfLoopBlock);
3459 OtherDest = InfLoopBlock;
3460 }
3461
3462 LLVM_DEBUG(dbgs() << *PBI->getParent()->getParent());
3463
3464 // BI may have other predecessors. Because of this, we leave
3465 // it alone, but modify PBI.
3466
3467 // Make sure we get to CommonDest on True&True directions.
3468 Value *PBICond = PBI->getCondition();
3469 IRBuilder<NoFolder> Builder(PBI);
3470 if (PBIOp)
3471 PBICond = Builder.CreateNot(PBICond, PBICond->getName() + ".not");
3472
3473 Value *BICond = BI->getCondition();
3474 if (BIOp)
3475 BICond = Builder.CreateNot(BICond, BICond->getName() + ".not");
3476
3477 // Merge the conditions.
3478 Value *Cond = Builder.CreateOr(PBICond, BICond, "brmerge");
3479
3480 // Modify PBI to branch on the new condition to the new dests.
3481 PBI->setCondition(Cond);
3482 PBI->setSuccessor(0, CommonDest);
3483 PBI->setSuccessor(1, OtherDest);
3484
3485 // Update branch weight for PBI.
3486 uint64_t PredTrueWeight, PredFalseWeight, SuccTrueWeight, SuccFalseWeight;
3487 uint64_t PredCommon, PredOther, SuccCommon, SuccOther;
3488 bool HasWeights =
3489 extractPredSuccWeights(PBI, BI, PredTrueWeight, PredFalseWeight,
3490 SuccTrueWeight, SuccFalseWeight);
3491 if (HasWeights) {
3492 PredCommon = PBIOp ? PredFalseWeight : PredTrueWeight;
3493 PredOther = PBIOp ? PredTrueWeight : PredFalseWeight;
3494 SuccCommon = BIOp ? SuccFalseWeight : SuccTrueWeight;
3495 SuccOther = BIOp ? SuccTrueWeight : SuccFalseWeight;
3496 // The weight to CommonDest should be PredCommon * SuccTotal +
3497 // PredOther * SuccCommon.
3498 // The weight to OtherDest should be PredOther * SuccOther.
3499 uint64_t NewWeights[2] = {PredCommon * (SuccCommon + SuccOther) +
3500 PredOther * SuccCommon,
3501 PredOther * SuccOther};
3502 // Halve the weights if any of them cannot fit in an uint32_t
3503 FitWeights(NewWeights);
3504
3505 setBranchWeights(PBI, NewWeights[0], NewWeights[1]);
3506 }
3507
3508 // OtherDest may have phi nodes. If so, add an entry from PBI's
3509 // block that are identical to the entries for BI's block.
3510 AddPredecessorToBlock(OtherDest, PBI->getParent(), BB);
3511
3512 // We know that the CommonDest already had an edge from PBI to
3513 // it. If it has PHIs though, the PHIs may have different
3514 // entries for BB and PBI's BB. If so, insert a select to make
3515 // them agree.
3516 for (PHINode &PN : CommonDest->phis()) {
3517 Value *BIV = PN.getIncomingValueForBlock(BB);
3518 unsigned PBBIdx = PN.getBasicBlockIndex(PBI->getParent());
3519 Value *PBIV = PN.getIncomingValue(PBBIdx);
3520 if (BIV != PBIV) {
3521 // Insert a select in PBI to pick the right value.
3522 SelectInst *NV = cast<SelectInst>(
3523 Builder.CreateSelect(PBICond, PBIV, BIV, PBIV->getName() + ".mux"));
3524 PN.setIncomingValue(PBBIdx, NV);
3525 // Although the select has the same condition as PBI, the original branch
3526 // weights for PBI do not apply to the new select because the select's
3527 // 'logical' edges are incoming edges of the phi that is eliminated, not
3528 // the outgoing edges of PBI.
3529 if (HasWeights) {
3530 uint64_t PredCommon = PBIOp ? PredFalseWeight : PredTrueWeight;
3531 uint64_t PredOther = PBIOp ? PredTrueWeight : PredFalseWeight;
3532 uint64_t SuccCommon = BIOp ? SuccFalseWeight : SuccTrueWeight;
3533 uint64_t SuccOther = BIOp ? SuccTrueWeight : SuccFalseWeight;
3534 // The weight to PredCommonDest should be PredCommon * SuccTotal.
3535 // The weight to PredOtherDest should be PredOther * SuccCommon.
3536 uint64_t NewWeights[2] = {PredCommon * (SuccCommon + SuccOther),
3537 PredOther * SuccCommon};
3538
3539 FitWeights(NewWeights);
3540
3541 setBranchWeights(NV, NewWeights[0], NewWeights[1]);
3542 }
3543 }
3544 }
3545
3546 LLVM_DEBUG(dbgs() << "INTO: " << *PBI->getParent());
3547 LLVM_DEBUG(dbgs() << *PBI->getParent()->getParent());
3548
3549 // This basic block is probably dead. We know it has at least
3550 // one fewer predecessor.
3551 return true;
3552 }
3553
3554 // Simplifies a terminator by replacing it with a branch to TrueBB if Cond is
3555 // true or to FalseBB if Cond is false.
3556 // Takes care of updating the successors and removing the old terminator.
3557 // Also makes sure not to introduce new successors by assuming that edges to
3558 // non-successor TrueBBs and FalseBBs aren't reachable.
SimplifyTerminatorOnSelect(Instruction * OldTerm,Value * Cond,BasicBlock * TrueBB,BasicBlock * FalseBB,uint32_t TrueWeight,uint32_t FalseWeight)3559 bool SimplifyCFGOpt::SimplifyTerminatorOnSelect(Instruction *OldTerm,
3560 Value *Cond, BasicBlock *TrueBB,
3561 BasicBlock *FalseBB,
3562 uint32_t TrueWeight,
3563 uint32_t FalseWeight) {
3564 // Remove any superfluous successor edges from the CFG.
3565 // First, figure out which successors to preserve.
3566 // If TrueBB and FalseBB are equal, only try to preserve one copy of that
3567 // successor.
3568 BasicBlock *KeepEdge1 = TrueBB;
3569 BasicBlock *KeepEdge2 = TrueBB != FalseBB ? FalseBB : nullptr;
3570
3571 // Then remove the rest.
3572 for (BasicBlock *Succ : successors(OldTerm)) {
3573 // Make sure only to keep exactly one copy of each edge.
3574 if (Succ == KeepEdge1)
3575 KeepEdge1 = nullptr;
3576 else if (Succ == KeepEdge2)
3577 KeepEdge2 = nullptr;
3578 else
3579 Succ->removePredecessor(OldTerm->getParent(),
3580 /*KeepOneInputPHIs=*/true);
3581 }
3582
3583 IRBuilder<> Builder(OldTerm);
3584 Builder.SetCurrentDebugLocation(OldTerm->getDebugLoc());
3585
3586 // Insert an appropriate new terminator.
3587 if (!KeepEdge1 && !KeepEdge2) {
3588 if (TrueBB == FalseBB)
3589 // We were only looking for one successor, and it was present.
3590 // Create an unconditional branch to it.
3591 Builder.CreateBr(TrueBB);
3592 else {
3593 // We found both of the successors we were looking for.
3594 // Create a conditional branch sharing the condition of the select.
3595 BranchInst *NewBI = Builder.CreateCondBr(Cond, TrueBB, FalseBB);
3596 if (TrueWeight != FalseWeight)
3597 setBranchWeights(NewBI, TrueWeight, FalseWeight);
3598 }
3599 } else if (KeepEdge1 && (KeepEdge2 || TrueBB == FalseBB)) {
3600 // Neither of the selected blocks were successors, so this
3601 // terminator must be unreachable.
3602 new UnreachableInst(OldTerm->getContext(), OldTerm);
3603 } else {
3604 // One of the selected values was a successor, but the other wasn't.
3605 // Insert an unconditional branch to the one that was found;
3606 // the edge to the one that wasn't must be unreachable.
3607 if (!KeepEdge1)
3608 // Only TrueBB was found.
3609 Builder.CreateBr(TrueBB);
3610 else
3611 // Only FalseBB was found.
3612 Builder.CreateBr(FalseBB);
3613 }
3614
3615 EraseTerminatorAndDCECond(OldTerm);
3616 return true;
3617 }
3618
3619 // Replaces
3620 // (switch (select cond, X, Y)) on constant X, Y
3621 // with a branch - conditional if X and Y lead to distinct BBs,
3622 // unconditional otherwise.
SimplifySwitchOnSelect(SwitchInst * SI,SelectInst * Select)3623 bool SimplifyCFGOpt::SimplifySwitchOnSelect(SwitchInst *SI,
3624 SelectInst *Select) {
3625 // Check for constant integer values in the select.
3626 ConstantInt *TrueVal = dyn_cast<ConstantInt>(Select->getTrueValue());
3627 ConstantInt *FalseVal = dyn_cast<ConstantInt>(Select->getFalseValue());
3628 if (!TrueVal || !FalseVal)
3629 return false;
3630
3631 // Find the relevant condition and destinations.
3632 Value *Condition = Select->getCondition();
3633 BasicBlock *TrueBB = SI->findCaseValue(TrueVal)->getCaseSuccessor();
3634 BasicBlock *FalseBB = SI->findCaseValue(FalseVal)->getCaseSuccessor();
3635
3636 // Get weight for TrueBB and FalseBB.
3637 uint32_t TrueWeight = 0, FalseWeight = 0;
3638 SmallVector<uint64_t, 8> Weights;
3639 bool HasWeights = HasBranchWeights(SI);
3640 if (HasWeights) {
3641 GetBranchWeights(SI, Weights);
3642 if (Weights.size() == 1 + SI->getNumCases()) {
3643 TrueWeight =
3644 (uint32_t)Weights[SI->findCaseValue(TrueVal)->getSuccessorIndex()];
3645 FalseWeight =
3646 (uint32_t)Weights[SI->findCaseValue(FalseVal)->getSuccessorIndex()];
3647 }
3648 }
3649
3650 // Perform the actual simplification.
3651 return SimplifyTerminatorOnSelect(SI, Condition, TrueBB, FalseBB, TrueWeight,
3652 FalseWeight);
3653 }
3654
3655 // Replaces
3656 // (indirectbr (select cond, blockaddress(@fn, BlockA),
3657 // blockaddress(@fn, BlockB)))
3658 // with
3659 // (br cond, BlockA, BlockB).
SimplifyIndirectBrOnSelect(IndirectBrInst * IBI,SelectInst * SI)3660 bool SimplifyCFGOpt::SimplifyIndirectBrOnSelect(IndirectBrInst *IBI,
3661 SelectInst *SI) {
3662 // Check that both operands of the select are block addresses.
3663 BlockAddress *TBA = dyn_cast<BlockAddress>(SI->getTrueValue());
3664 BlockAddress *FBA = dyn_cast<BlockAddress>(SI->getFalseValue());
3665 if (!TBA || !FBA)
3666 return false;
3667
3668 // Extract the actual blocks.
3669 BasicBlock *TrueBB = TBA->getBasicBlock();
3670 BasicBlock *FalseBB = FBA->getBasicBlock();
3671
3672 // Perform the actual simplification.
3673 return SimplifyTerminatorOnSelect(IBI, SI->getCondition(), TrueBB, FalseBB, 0,
3674 0);
3675 }
3676
3677 /// This is called when we find an icmp instruction
3678 /// (a seteq/setne with a constant) as the only instruction in a
3679 /// block that ends with an uncond branch. We are looking for a very specific
3680 /// pattern that occurs when "A == 1 || A == 2 || A == 3" gets simplified. In
3681 /// this case, we merge the first two "or's of icmp" into a switch, but then the
3682 /// default value goes to an uncond block with a seteq in it, we get something
3683 /// like:
3684 ///
3685 /// switch i8 %A, label %DEFAULT [ i8 1, label %end i8 2, label %end ]
3686 /// DEFAULT:
3687 /// %tmp = icmp eq i8 %A, 92
3688 /// br label %end
3689 /// end:
3690 /// ... = phi i1 [ true, %entry ], [ %tmp, %DEFAULT ], [ true, %entry ]
3691 ///
3692 /// We prefer to split the edge to 'end' so that there is a true/false entry to
3693 /// the PHI, merging the third icmp into the switch.
tryToSimplifyUncondBranchWithICmpInIt(ICmpInst * ICI,IRBuilder<> & Builder)3694 bool SimplifyCFGOpt::tryToSimplifyUncondBranchWithICmpInIt(
3695 ICmpInst *ICI, IRBuilder<> &Builder) {
3696 BasicBlock *BB = ICI->getParent();
3697
3698 // If the block has any PHIs in it or the icmp has multiple uses, it is too
3699 // complex.
3700 if (isa<PHINode>(BB->begin()) || !ICI->hasOneUse())
3701 return false;
3702
3703 Value *V = ICI->getOperand(0);
3704 ConstantInt *Cst = cast<ConstantInt>(ICI->getOperand(1));
3705
3706 // The pattern we're looking for is where our only predecessor is a switch on
3707 // 'V' and this block is the default case for the switch. In this case we can
3708 // fold the compared value into the switch to simplify things.
3709 BasicBlock *Pred = BB->getSinglePredecessor();
3710 if (!Pred || !isa<SwitchInst>(Pred->getTerminator()))
3711 return false;
3712
3713 SwitchInst *SI = cast<SwitchInst>(Pred->getTerminator());
3714 if (SI->getCondition() != V)
3715 return false;
3716
3717 // If BB is reachable on a non-default case, then we simply know the value of
3718 // V in this block. Substitute it and constant fold the icmp instruction
3719 // away.
3720 if (SI->getDefaultDest() != BB) {
3721 ConstantInt *VVal = SI->findCaseDest(BB);
3722 assert(VVal && "Should have a unique destination value");
3723 ICI->setOperand(0, VVal);
3724
3725 if (Value *V = SimplifyInstruction(ICI, {DL, ICI})) {
3726 ICI->replaceAllUsesWith(V);
3727 ICI->eraseFromParent();
3728 }
3729 // BB is now empty, so it is likely to simplify away.
3730 return requestResimplify();
3731 }
3732
3733 // Ok, the block is reachable from the default dest. If the constant we're
3734 // comparing exists in one of the other edges, then we can constant fold ICI
3735 // and zap it.
3736 if (SI->findCaseValue(Cst) != SI->case_default()) {
3737 Value *V;
3738 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
3739 V = ConstantInt::getFalse(BB->getContext());
3740 else
3741 V = ConstantInt::getTrue(BB->getContext());
3742
3743 ICI->replaceAllUsesWith(V);
3744 ICI->eraseFromParent();
3745 // BB is now empty, so it is likely to simplify away.
3746 return requestResimplify();
3747 }
3748
3749 // The use of the icmp has to be in the 'end' block, by the only PHI node in
3750 // the block.
3751 BasicBlock *SuccBlock = BB->getTerminator()->getSuccessor(0);
3752 PHINode *PHIUse = dyn_cast<PHINode>(ICI->user_back());
3753 if (PHIUse == nullptr || PHIUse != &SuccBlock->front() ||
3754 isa<PHINode>(++BasicBlock::iterator(PHIUse)))
3755 return false;
3756
3757 // If the icmp is a SETEQ, then the default dest gets false, the new edge gets
3758 // true in the PHI.
3759 Constant *DefaultCst = ConstantInt::getTrue(BB->getContext());
3760 Constant *NewCst = ConstantInt::getFalse(BB->getContext());
3761
3762 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
3763 std::swap(DefaultCst, NewCst);
3764
3765 // Replace ICI (which is used by the PHI for the default value) with true or
3766 // false depending on if it is EQ or NE.
3767 ICI->replaceAllUsesWith(DefaultCst);
3768 ICI->eraseFromParent();
3769
3770 // Okay, the switch goes to this block on a default value. Add an edge from
3771 // the switch to the merge point on the compared value.
3772 BasicBlock *NewBB =
3773 BasicBlock::Create(BB->getContext(), "switch.edge", BB->getParent(), BB);
3774 {
3775 SwitchInstProfUpdateWrapper SIW(*SI);
3776 auto W0 = SIW.getSuccessorWeight(0);
3777 SwitchInstProfUpdateWrapper::CaseWeightOpt NewW;
3778 if (W0) {
3779 NewW = ((uint64_t(*W0) + 1) >> 1);
3780 SIW.setSuccessorWeight(0, *NewW);
3781 }
3782 SIW.addCase(Cst, NewBB, NewW);
3783 }
3784
3785 // NewBB branches to the phi block, add the uncond branch and the phi entry.
3786 Builder.SetInsertPoint(NewBB);
3787 Builder.SetCurrentDebugLocation(SI->getDebugLoc());
3788 Builder.CreateBr(SuccBlock);
3789 PHIUse->addIncoming(NewCst, NewBB);
3790 return true;
3791 }
3792
3793 /// The specified branch is a conditional branch.
3794 /// Check to see if it is branching on an or/and chain of icmp instructions, and
3795 /// fold it into a switch instruction if so.
SimplifyBranchOnICmpChain(BranchInst * BI,IRBuilder<> & Builder,const DataLayout & DL)3796 bool SimplifyCFGOpt::SimplifyBranchOnICmpChain(BranchInst *BI,
3797 IRBuilder<> &Builder,
3798 const DataLayout &DL) {
3799 Instruction *Cond = dyn_cast<Instruction>(BI->getCondition());
3800 if (!Cond)
3801 return false;
3802
3803 // Change br (X == 0 | X == 1), T, F into a switch instruction.
3804 // If this is a bunch of seteq's or'd together, or if it's a bunch of
3805 // 'setne's and'ed together, collect them.
3806
3807 // Try to gather values from a chain of and/or to be turned into a switch
3808 ConstantComparesGatherer ConstantCompare(Cond, DL);
3809 // Unpack the result
3810 SmallVectorImpl<ConstantInt *> &Values = ConstantCompare.Vals;
3811 Value *CompVal = ConstantCompare.CompValue;
3812 unsigned UsedICmps = ConstantCompare.UsedICmps;
3813 Value *ExtraCase = ConstantCompare.Extra;
3814
3815 // If we didn't have a multiply compared value, fail.
3816 if (!CompVal)
3817 return false;
3818
3819 // Avoid turning single icmps into a switch.
3820 if (UsedICmps <= 1)
3821 return false;
3822
3823 bool TrueWhenEqual = (Cond->getOpcode() == Instruction::Or);
3824
3825 // There might be duplicate constants in the list, which the switch
3826 // instruction can't handle, remove them now.
3827 array_pod_sort(Values.begin(), Values.end(), ConstantIntSortPredicate);
3828 Values.erase(std::unique(Values.begin(), Values.end()), Values.end());
3829
3830 // If Extra was used, we require at least two switch values to do the
3831 // transformation. A switch with one value is just a conditional branch.
3832 if (ExtraCase && Values.size() < 2)
3833 return false;
3834
3835 // TODO: Preserve branch weight metadata, similarly to how
3836 // FoldValueComparisonIntoPredecessors preserves it.
3837
3838 // Figure out which block is which destination.
3839 BasicBlock *DefaultBB = BI->getSuccessor(1);
3840 BasicBlock *EdgeBB = BI->getSuccessor(0);
3841 if (!TrueWhenEqual)
3842 std::swap(DefaultBB, EdgeBB);
3843
3844 BasicBlock *BB = BI->getParent();
3845
3846 // MSAN does not like undefs as branch condition which can be introduced
3847 // with "explicit branch".
3848 if (ExtraCase && BB->getParent()->hasFnAttribute(Attribute::SanitizeMemory))
3849 return false;
3850
3851 LLVM_DEBUG(dbgs() << "Converting 'icmp' chain with " << Values.size()
3852 << " cases into SWITCH. BB is:\n"
3853 << *BB);
3854
3855 // If there are any extra values that couldn't be folded into the switch
3856 // then we evaluate them with an explicit branch first. Split the block
3857 // right before the condbr to handle it.
3858 if (ExtraCase) {
3859 BasicBlock *NewBB =
3860 BB->splitBasicBlock(BI->getIterator(), "switch.early.test");
3861 // Remove the uncond branch added to the old block.
3862 Instruction *OldTI = BB->getTerminator();
3863 Builder.SetInsertPoint(OldTI);
3864
3865 if (TrueWhenEqual)
3866 Builder.CreateCondBr(ExtraCase, EdgeBB, NewBB);
3867 else
3868 Builder.CreateCondBr(ExtraCase, NewBB, EdgeBB);
3869
3870 OldTI->eraseFromParent();
3871
3872 // If there are PHI nodes in EdgeBB, then we need to add a new entry to them
3873 // for the edge we just added.
3874 AddPredecessorToBlock(EdgeBB, BB, NewBB);
3875
3876 LLVM_DEBUG(dbgs() << " ** 'icmp' chain unhandled condition: " << *ExtraCase
3877 << "\nEXTRABB = " << *BB);
3878 BB = NewBB;
3879 }
3880
3881 Builder.SetInsertPoint(BI);
3882 // Convert pointer to int before we switch.
3883 if (CompVal->getType()->isPointerTy()) {
3884 CompVal = Builder.CreatePtrToInt(
3885 CompVal, DL.getIntPtrType(CompVal->getType()), "magicptr");
3886 }
3887
3888 // Create the new switch instruction now.
3889 SwitchInst *New = Builder.CreateSwitch(CompVal, DefaultBB, Values.size());
3890
3891 // Add all of the 'cases' to the switch instruction.
3892 for (unsigned i = 0, e = Values.size(); i != e; ++i)
3893 New->addCase(Values[i], EdgeBB);
3894
3895 // We added edges from PI to the EdgeBB. As such, if there were any
3896 // PHI nodes in EdgeBB, they need entries to be added corresponding to
3897 // the number of edges added.
3898 for (BasicBlock::iterator BBI = EdgeBB->begin(); isa<PHINode>(BBI); ++BBI) {
3899 PHINode *PN = cast<PHINode>(BBI);
3900 Value *InVal = PN->getIncomingValueForBlock(BB);
3901 for (unsigned i = 0, e = Values.size() - 1; i != e; ++i)
3902 PN->addIncoming(InVal, BB);
3903 }
3904
3905 // Erase the old branch instruction.
3906 EraseTerminatorAndDCECond(BI);
3907
3908 LLVM_DEBUG(dbgs() << " ** 'icmp' chain result is:\n" << *BB << '\n');
3909 return true;
3910 }
3911
simplifyResume(ResumeInst * RI,IRBuilder<> & Builder)3912 bool SimplifyCFGOpt::simplifyResume(ResumeInst *RI, IRBuilder<> &Builder) {
3913 if (isa<PHINode>(RI->getValue()))
3914 return simplifyCommonResume(RI);
3915 else if (isa<LandingPadInst>(RI->getParent()->getFirstNonPHI()) &&
3916 RI->getValue() == RI->getParent()->getFirstNonPHI())
3917 // The resume must unwind the exception that caused control to branch here.
3918 return simplifySingleResume(RI);
3919
3920 return false;
3921 }
3922
3923 // Simplify resume that is shared by several landing pads (phi of landing pad).
simplifyCommonResume(ResumeInst * RI)3924 bool SimplifyCFGOpt::simplifyCommonResume(ResumeInst *RI) {
3925 BasicBlock *BB = RI->getParent();
3926
3927 // Check that there are no other instructions except for debug intrinsics
3928 // between the phi of landing pads (RI->getValue()) and resume instruction.
3929 BasicBlock::iterator I = cast<Instruction>(RI->getValue())->getIterator(),
3930 E = RI->getIterator();
3931 while (++I != E)
3932 if (!isa<DbgInfoIntrinsic>(I))
3933 return false;
3934
3935 SmallSetVector<BasicBlock *, 4> TrivialUnwindBlocks;
3936 auto *PhiLPInst = cast<PHINode>(RI->getValue());
3937
3938 // Check incoming blocks to see if any of them are trivial.
3939 for (unsigned Idx = 0, End = PhiLPInst->getNumIncomingValues(); Idx != End;
3940 Idx++) {
3941 auto *IncomingBB = PhiLPInst->getIncomingBlock(Idx);
3942 auto *IncomingValue = PhiLPInst->getIncomingValue(Idx);
3943
3944 // If the block has other successors, we can not delete it because
3945 // it has other dependents.
3946 if (IncomingBB->getUniqueSuccessor() != BB)
3947 continue;
3948
3949 auto *LandingPad = dyn_cast<LandingPadInst>(IncomingBB->getFirstNonPHI());
3950 // Not the landing pad that caused the control to branch here.
3951 if (IncomingValue != LandingPad)
3952 continue;
3953
3954 bool isTrivial = true;
3955
3956 I = IncomingBB->getFirstNonPHI()->getIterator();
3957 E = IncomingBB->getTerminator()->getIterator();
3958 while (++I != E)
3959 if (!isa<DbgInfoIntrinsic>(I)) {
3960 isTrivial = false;
3961 break;
3962 }
3963
3964 if (isTrivial)
3965 TrivialUnwindBlocks.insert(IncomingBB);
3966 }
3967
3968 // If no trivial unwind blocks, don't do any simplifications.
3969 if (TrivialUnwindBlocks.empty())
3970 return false;
3971
3972 // Turn all invokes that unwind here into calls.
3973 for (auto *TrivialBB : TrivialUnwindBlocks) {
3974 // Blocks that will be simplified should be removed from the phi node.
3975 // Note there could be multiple edges to the resume block, and we need
3976 // to remove them all.
3977 while (PhiLPInst->getBasicBlockIndex(TrivialBB) != -1)
3978 BB->removePredecessor(TrivialBB, true);
3979
3980 for (pred_iterator PI = pred_begin(TrivialBB), PE = pred_end(TrivialBB);
3981 PI != PE;) {
3982 BasicBlock *Pred = *PI++;
3983 removeUnwindEdge(Pred);
3984 }
3985
3986 // In each SimplifyCFG run, only the current processed block can be erased.
3987 // Otherwise, it will break the iteration of SimplifyCFG pass. So instead
3988 // of erasing TrivialBB, we only remove the branch to the common resume
3989 // block so that we can later erase the resume block since it has no
3990 // predecessors.
3991 TrivialBB->getTerminator()->eraseFromParent();
3992 new UnreachableInst(RI->getContext(), TrivialBB);
3993 }
3994
3995 // Delete the resume block if all its predecessors have been removed.
3996 if (pred_empty(BB))
3997 BB->eraseFromParent();
3998
3999 return !TrivialUnwindBlocks.empty();
4000 }
4001
4002 // Check if cleanup block is empty
isCleanupBlockEmpty(Instruction * Inst,Instruction * RI)4003 static bool isCleanupBlockEmpty(Instruction *Inst, Instruction *RI) {
4004 BasicBlock::iterator I = Inst->getIterator(), E = RI->getIterator();
4005 while (++I != E) {
4006 auto *II = dyn_cast<IntrinsicInst>(I);
4007 if (!II)
4008 return false;
4009
4010 Intrinsic::ID IntrinsicID = II->getIntrinsicID();
4011 switch (IntrinsicID) {
4012 case Intrinsic::dbg_declare:
4013 case Intrinsic::dbg_value:
4014 case Intrinsic::dbg_label:
4015 case Intrinsic::lifetime_end:
4016 break;
4017 default:
4018 return false;
4019 }
4020 }
4021 return true;
4022 }
4023
4024 // Simplify resume that is only used by a single (non-phi) landing pad.
simplifySingleResume(ResumeInst * RI)4025 bool SimplifyCFGOpt::simplifySingleResume(ResumeInst *RI) {
4026 BasicBlock *BB = RI->getParent();
4027 auto *LPInst = cast<LandingPadInst>(BB->getFirstNonPHI());
4028 assert(RI->getValue() == LPInst &&
4029 "Resume must unwind the exception that caused control to here");
4030
4031 // Check that there are no other instructions except for debug intrinsics.
4032 if (!isCleanupBlockEmpty(LPInst, RI))
4033 return false;
4034
4035 // Turn all invokes that unwind here into calls and delete the basic block.
4036 for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE;) {
4037 BasicBlock *Pred = *PI++;
4038 removeUnwindEdge(Pred);
4039 }
4040
4041 // The landingpad is now unreachable. Zap it.
4042 if (LoopHeaders)
4043 LoopHeaders->erase(BB);
4044 BB->eraseFromParent();
4045 return true;
4046 }
4047
removeEmptyCleanup(CleanupReturnInst * RI)4048 static bool removeEmptyCleanup(CleanupReturnInst *RI) {
4049 // If this is a trivial cleanup pad that executes no instructions, it can be
4050 // eliminated. If the cleanup pad continues to the caller, any predecessor
4051 // that is an EH pad will be updated to continue to the caller and any
4052 // predecessor that terminates with an invoke instruction will have its invoke
4053 // instruction converted to a call instruction. If the cleanup pad being
4054 // simplified does not continue to the caller, each predecessor will be
4055 // updated to continue to the unwind destination of the cleanup pad being
4056 // simplified.
4057 BasicBlock *BB = RI->getParent();
4058 CleanupPadInst *CPInst = RI->getCleanupPad();
4059 if (CPInst->getParent() != BB)
4060 // This isn't an empty cleanup.
4061 return false;
4062
4063 // We cannot kill the pad if it has multiple uses. This typically arises
4064 // from unreachable basic blocks.
4065 if (!CPInst->hasOneUse())
4066 return false;
4067
4068 // Check that there are no other instructions except for benign intrinsics.
4069 if (!isCleanupBlockEmpty(CPInst, RI))
4070 return false;
4071
4072 // If the cleanup return we are simplifying unwinds to the caller, this will
4073 // set UnwindDest to nullptr.
4074 BasicBlock *UnwindDest = RI->getUnwindDest();
4075 Instruction *DestEHPad = UnwindDest ? UnwindDest->getFirstNonPHI() : nullptr;
4076
4077 // We're about to remove BB from the control flow. Before we do, sink any
4078 // PHINodes into the unwind destination. Doing this before changing the
4079 // control flow avoids some potentially slow checks, since we can currently
4080 // be certain that UnwindDest and BB have no common predecessors (since they
4081 // are both EH pads).
4082 if (UnwindDest) {
4083 // First, go through the PHI nodes in UnwindDest and update any nodes that
4084 // reference the block we are removing
4085 for (BasicBlock::iterator I = UnwindDest->begin(),
4086 IE = DestEHPad->getIterator();
4087 I != IE; ++I) {
4088 PHINode *DestPN = cast<PHINode>(I);
4089
4090 int Idx = DestPN->getBasicBlockIndex(BB);
4091 // Since BB unwinds to UnwindDest, it has to be in the PHI node.
4092 assert(Idx != -1);
4093 // This PHI node has an incoming value that corresponds to a control
4094 // path through the cleanup pad we are removing. If the incoming
4095 // value is in the cleanup pad, it must be a PHINode (because we
4096 // verified above that the block is otherwise empty). Otherwise, the
4097 // value is either a constant or a value that dominates the cleanup
4098 // pad being removed.
4099 //
4100 // Because BB and UnwindDest are both EH pads, all of their
4101 // predecessors must unwind to these blocks, and since no instruction
4102 // can have multiple unwind destinations, there will be no overlap in
4103 // incoming blocks between SrcPN and DestPN.
4104 Value *SrcVal = DestPN->getIncomingValue(Idx);
4105 PHINode *SrcPN = dyn_cast<PHINode>(SrcVal);
4106
4107 // Remove the entry for the block we are deleting.
4108 DestPN->removeIncomingValue(Idx, false);
4109
4110 if (SrcPN && SrcPN->getParent() == BB) {
4111 // If the incoming value was a PHI node in the cleanup pad we are
4112 // removing, we need to merge that PHI node's incoming values into
4113 // DestPN.
4114 for (unsigned SrcIdx = 0, SrcE = SrcPN->getNumIncomingValues();
4115 SrcIdx != SrcE; ++SrcIdx) {
4116 DestPN->addIncoming(SrcPN->getIncomingValue(SrcIdx),
4117 SrcPN->getIncomingBlock(SrcIdx));
4118 }
4119 } else {
4120 // Otherwise, the incoming value came from above BB and
4121 // so we can just reuse it. We must associate all of BB's
4122 // predecessors with this value.
4123 for (auto *pred : predecessors(BB)) {
4124 DestPN->addIncoming(SrcVal, pred);
4125 }
4126 }
4127 }
4128
4129 // Sink any remaining PHI nodes directly into UnwindDest.
4130 Instruction *InsertPt = DestEHPad;
4131 for (BasicBlock::iterator I = BB->begin(),
4132 IE = BB->getFirstNonPHI()->getIterator();
4133 I != IE;) {
4134 // The iterator must be incremented here because the instructions are
4135 // being moved to another block.
4136 PHINode *PN = cast<PHINode>(I++);
4137 if (PN->use_empty() || !PN->isUsedOutsideOfBlock(BB))
4138 // If the PHI node has no uses or all of its uses are in this basic
4139 // block (meaning they are debug or lifetime intrinsics), just leave
4140 // it. It will be erased when we erase BB below.
4141 continue;
4142
4143 // Otherwise, sink this PHI node into UnwindDest.
4144 // Any predecessors to UnwindDest which are not already represented
4145 // must be back edges which inherit the value from the path through
4146 // BB. In this case, the PHI value must reference itself.
4147 for (auto *pred : predecessors(UnwindDest))
4148 if (pred != BB)
4149 PN->addIncoming(PN, pred);
4150 PN->moveBefore(InsertPt);
4151 }
4152 }
4153
4154 for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE;) {
4155 // The iterator must be updated here because we are removing this pred.
4156 BasicBlock *PredBB = *PI++;
4157 if (UnwindDest == nullptr) {
4158 removeUnwindEdge(PredBB);
4159 } else {
4160 Instruction *TI = PredBB->getTerminator();
4161 TI->replaceUsesOfWith(BB, UnwindDest);
4162 }
4163 }
4164
4165 // The cleanup pad is now unreachable. Zap it.
4166 BB->eraseFromParent();
4167 return true;
4168 }
4169
4170 // Try to merge two cleanuppads together.
mergeCleanupPad(CleanupReturnInst * RI)4171 static bool mergeCleanupPad(CleanupReturnInst *RI) {
4172 // Skip any cleanuprets which unwind to caller, there is nothing to merge
4173 // with.
4174 BasicBlock *UnwindDest = RI->getUnwindDest();
4175 if (!UnwindDest)
4176 return false;
4177
4178 // This cleanupret isn't the only predecessor of this cleanuppad, it wouldn't
4179 // be safe to merge without code duplication.
4180 if (UnwindDest->getSinglePredecessor() != RI->getParent())
4181 return false;
4182
4183 // Verify that our cleanuppad's unwind destination is another cleanuppad.
4184 auto *SuccessorCleanupPad = dyn_cast<CleanupPadInst>(&UnwindDest->front());
4185 if (!SuccessorCleanupPad)
4186 return false;
4187
4188 CleanupPadInst *PredecessorCleanupPad = RI->getCleanupPad();
4189 // Replace any uses of the successor cleanupad with the predecessor pad
4190 // The only cleanuppad uses should be this cleanupret, it's cleanupret and
4191 // funclet bundle operands.
4192 SuccessorCleanupPad->replaceAllUsesWith(PredecessorCleanupPad);
4193 // Remove the old cleanuppad.
4194 SuccessorCleanupPad->eraseFromParent();
4195 // Now, we simply replace the cleanupret with a branch to the unwind
4196 // destination.
4197 BranchInst::Create(UnwindDest, RI->getParent());
4198 RI->eraseFromParent();
4199
4200 return true;
4201 }
4202
simplifyCleanupReturn(CleanupReturnInst * RI)4203 bool SimplifyCFGOpt::simplifyCleanupReturn(CleanupReturnInst *RI) {
4204 // It is possible to transiantly have an undef cleanuppad operand because we
4205 // have deleted some, but not all, dead blocks.
4206 // Eventually, this block will be deleted.
4207 if (isa<UndefValue>(RI->getOperand(0)))
4208 return false;
4209
4210 if (mergeCleanupPad(RI))
4211 return true;
4212
4213 if (removeEmptyCleanup(RI))
4214 return true;
4215
4216 return false;
4217 }
4218
simplifyReturn(ReturnInst * RI,IRBuilder<> & Builder)4219 bool SimplifyCFGOpt::simplifyReturn(ReturnInst *RI, IRBuilder<> &Builder) {
4220 BasicBlock *BB = RI->getParent();
4221 if (!BB->getFirstNonPHIOrDbg()->isTerminator())
4222 return false;
4223
4224 // Find predecessors that end with branches.
4225 SmallVector<BasicBlock *, 8> UncondBranchPreds;
4226 SmallVector<BranchInst *, 8> CondBranchPreds;
4227 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
4228 BasicBlock *P = *PI;
4229 Instruction *PTI = P->getTerminator();
4230 if (BranchInst *BI = dyn_cast<BranchInst>(PTI)) {
4231 if (BI->isUnconditional())
4232 UncondBranchPreds.push_back(P);
4233 else
4234 CondBranchPreds.push_back(BI);
4235 }
4236 }
4237
4238 // If we found some, do the transformation!
4239 if (!UncondBranchPreds.empty() && DupRet) {
4240 while (!UncondBranchPreds.empty()) {
4241 BasicBlock *Pred = UncondBranchPreds.pop_back_val();
4242 LLVM_DEBUG(dbgs() << "FOLDING: " << *BB
4243 << "INTO UNCOND BRANCH PRED: " << *Pred);
4244 (void)FoldReturnIntoUncondBranch(RI, BB, Pred);
4245 }
4246
4247 // If we eliminated all predecessors of the block, delete the block now.
4248 if (pred_empty(BB)) {
4249 // We know there are no successors, so just nuke the block.
4250 if (LoopHeaders)
4251 LoopHeaders->erase(BB);
4252 BB->eraseFromParent();
4253 }
4254
4255 return true;
4256 }
4257
4258 // Check out all of the conditional branches going to this return
4259 // instruction. If any of them just select between returns, change the
4260 // branch itself into a select/return pair.
4261 while (!CondBranchPreds.empty()) {
4262 BranchInst *BI = CondBranchPreds.pop_back_val();
4263
4264 // Check to see if the non-BB successor is also a return block.
4265 if (isa<ReturnInst>(BI->getSuccessor(0)->getTerminator()) &&
4266 isa<ReturnInst>(BI->getSuccessor(1)->getTerminator()) &&
4267 SimplifyCondBranchToTwoReturns(BI, Builder))
4268 return true;
4269 }
4270 return false;
4271 }
4272
simplifyUnreachable(UnreachableInst * UI)4273 bool SimplifyCFGOpt::simplifyUnreachable(UnreachableInst *UI) {
4274 BasicBlock *BB = UI->getParent();
4275
4276 bool Changed = false;
4277
4278 // If there are any instructions immediately before the unreachable that can
4279 // be removed, do so.
4280 while (UI->getIterator() != BB->begin()) {
4281 BasicBlock::iterator BBI = UI->getIterator();
4282 --BBI;
4283 // Do not delete instructions that can have side effects which might cause
4284 // the unreachable to not be reachable; specifically, calls and volatile
4285 // operations may have this effect.
4286 if (isa<CallInst>(BBI) && !isa<DbgInfoIntrinsic>(BBI))
4287 break;
4288
4289 if (BBI->mayHaveSideEffects()) {
4290 if (auto *SI = dyn_cast<StoreInst>(BBI)) {
4291 if (SI->isVolatile())
4292 break;
4293 } else if (auto *LI = dyn_cast<LoadInst>(BBI)) {
4294 if (LI->isVolatile())
4295 break;
4296 } else if (auto *RMWI = dyn_cast<AtomicRMWInst>(BBI)) {
4297 if (RMWI->isVolatile())
4298 break;
4299 } else if (auto *CXI = dyn_cast<AtomicCmpXchgInst>(BBI)) {
4300 if (CXI->isVolatile())
4301 break;
4302 } else if (isa<CatchPadInst>(BBI)) {
4303 // A catchpad may invoke exception object constructors and such, which
4304 // in some languages can be arbitrary code, so be conservative by
4305 // default.
4306 // For CoreCLR, it just involves a type test, so can be removed.
4307 if (classifyEHPersonality(BB->getParent()->getPersonalityFn()) !=
4308 EHPersonality::CoreCLR)
4309 break;
4310 } else if (!isa<FenceInst>(BBI) && !isa<VAArgInst>(BBI) &&
4311 !isa<LandingPadInst>(BBI)) {
4312 break;
4313 }
4314 // Note that deleting LandingPad's here is in fact okay, although it
4315 // involves a bit of subtle reasoning. If this inst is a LandingPad,
4316 // all the predecessors of this block will be the unwind edges of Invokes,
4317 // and we can therefore guarantee this block will be erased.
4318 }
4319
4320 // Delete this instruction (any uses are guaranteed to be dead)
4321 if (!BBI->use_empty())
4322 BBI->replaceAllUsesWith(UndefValue::get(BBI->getType()));
4323 BBI->eraseFromParent();
4324 Changed = true;
4325 }
4326
4327 // If the unreachable instruction is the first in the block, take a gander
4328 // at all of the predecessors of this instruction, and simplify them.
4329 if (&BB->front() != UI)
4330 return Changed;
4331
4332 SmallVector<BasicBlock *, 8> Preds(pred_begin(BB), pred_end(BB));
4333 for (unsigned i = 0, e = Preds.size(); i != e; ++i) {
4334 Instruction *TI = Preds[i]->getTerminator();
4335 IRBuilder<> Builder(TI);
4336 if (auto *BI = dyn_cast<BranchInst>(TI)) {
4337 if (BI->isUnconditional()) {
4338 assert(BI->getSuccessor(0) == BB && "Incorrect CFG");
4339 new UnreachableInst(TI->getContext(), TI);
4340 TI->eraseFromParent();
4341 Changed = true;
4342 } else {
4343 Value* Cond = BI->getCondition();
4344 if (BI->getSuccessor(0) == BB) {
4345 Builder.CreateAssumption(Builder.CreateNot(Cond));
4346 Builder.CreateBr(BI->getSuccessor(1));
4347 } else {
4348 assert(BI->getSuccessor(1) == BB && "Incorrect CFG");
4349 Builder.CreateAssumption(Cond);
4350 Builder.CreateBr(BI->getSuccessor(0));
4351 }
4352 EraseTerminatorAndDCECond(BI);
4353 Changed = true;
4354 }
4355 } else if (auto *SI = dyn_cast<SwitchInst>(TI)) {
4356 SwitchInstProfUpdateWrapper SU(*SI);
4357 for (auto i = SU->case_begin(), e = SU->case_end(); i != e;) {
4358 if (i->getCaseSuccessor() != BB) {
4359 ++i;
4360 continue;
4361 }
4362 BB->removePredecessor(SU->getParent());
4363 i = SU.removeCase(i);
4364 e = SU->case_end();
4365 Changed = true;
4366 }
4367 } else if (auto *II = dyn_cast<InvokeInst>(TI)) {
4368 if (II->getUnwindDest() == BB) {
4369 removeUnwindEdge(TI->getParent());
4370 Changed = true;
4371 }
4372 } else if (auto *CSI = dyn_cast<CatchSwitchInst>(TI)) {
4373 if (CSI->getUnwindDest() == BB) {
4374 removeUnwindEdge(TI->getParent());
4375 Changed = true;
4376 continue;
4377 }
4378
4379 for (CatchSwitchInst::handler_iterator I = CSI->handler_begin(),
4380 E = CSI->handler_end();
4381 I != E; ++I) {
4382 if (*I == BB) {
4383 CSI->removeHandler(I);
4384 --I;
4385 --E;
4386 Changed = true;
4387 }
4388 }
4389 if (CSI->getNumHandlers() == 0) {
4390 BasicBlock *CatchSwitchBB = CSI->getParent();
4391 if (CSI->hasUnwindDest()) {
4392 // Redirect preds to the unwind dest
4393 CatchSwitchBB->replaceAllUsesWith(CSI->getUnwindDest());
4394 } else {
4395 // Rewrite all preds to unwind to caller (or from invoke to call).
4396 SmallVector<BasicBlock *, 8> EHPreds(predecessors(CatchSwitchBB));
4397 for (BasicBlock *EHPred : EHPreds)
4398 removeUnwindEdge(EHPred);
4399 }
4400 // The catchswitch is no longer reachable.
4401 new UnreachableInst(CSI->getContext(), CSI);
4402 CSI->eraseFromParent();
4403 Changed = true;
4404 }
4405 } else if (isa<CleanupReturnInst>(TI)) {
4406 new UnreachableInst(TI->getContext(), TI);
4407 TI->eraseFromParent();
4408 Changed = true;
4409 }
4410 }
4411
4412 // If this block is now dead, remove it.
4413 if (pred_empty(BB) && BB != &BB->getParent()->getEntryBlock()) {
4414 // We know there are no successors, so just nuke the block.
4415 if (LoopHeaders)
4416 LoopHeaders->erase(BB);
4417 BB->eraseFromParent();
4418 return true;
4419 }
4420
4421 return Changed;
4422 }
4423
CasesAreContiguous(SmallVectorImpl<ConstantInt * > & Cases)4424 static bool CasesAreContiguous(SmallVectorImpl<ConstantInt *> &Cases) {
4425 assert(Cases.size() >= 1);
4426
4427 array_pod_sort(Cases.begin(), Cases.end(), ConstantIntSortPredicate);
4428 for (size_t I = 1, E = Cases.size(); I != E; ++I) {
4429 if (Cases[I - 1]->getValue() != Cases[I]->getValue() + 1)
4430 return false;
4431 }
4432 return true;
4433 }
4434
createUnreachableSwitchDefault(SwitchInst * Switch)4435 static void createUnreachableSwitchDefault(SwitchInst *Switch) {
4436 LLVM_DEBUG(dbgs() << "SimplifyCFG: switch default is dead.\n");
4437 BasicBlock *NewDefaultBlock =
4438 SplitBlockPredecessors(Switch->getDefaultDest(), Switch->getParent(), "");
4439 Switch->setDefaultDest(&*NewDefaultBlock);
4440 SplitBlock(&*NewDefaultBlock, &NewDefaultBlock->front());
4441 auto *NewTerminator = NewDefaultBlock->getTerminator();
4442 new UnreachableInst(Switch->getContext(), NewTerminator);
4443 EraseTerminatorAndDCECond(NewTerminator);
4444 }
4445
4446 /// Turn a switch with two reachable destinations into an integer range
4447 /// comparison and branch.
TurnSwitchRangeIntoICmp(SwitchInst * SI,IRBuilder<> & Builder)4448 bool SimplifyCFGOpt::TurnSwitchRangeIntoICmp(SwitchInst *SI,
4449 IRBuilder<> &Builder) {
4450 assert(SI->getNumCases() > 1 && "Degenerate switch?");
4451
4452 bool HasDefault =
4453 !isa<UnreachableInst>(SI->getDefaultDest()->getFirstNonPHIOrDbg());
4454
4455 // Partition the cases into two sets with different destinations.
4456 BasicBlock *DestA = HasDefault ? SI->getDefaultDest() : nullptr;
4457 BasicBlock *DestB = nullptr;
4458 SmallVector<ConstantInt *, 16> CasesA;
4459 SmallVector<ConstantInt *, 16> CasesB;
4460
4461 for (auto Case : SI->cases()) {
4462 BasicBlock *Dest = Case.getCaseSuccessor();
4463 if (!DestA)
4464 DestA = Dest;
4465 if (Dest == DestA) {
4466 CasesA.push_back(Case.getCaseValue());
4467 continue;
4468 }
4469 if (!DestB)
4470 DestB = Dest;
4471 if (Dest == DestB) {
4472 CasesB.push_back(Case.getCaseValue());
4473 continue;
4474 }
4475 return false; // More than two destinations.
4476 }
4477
4478 assert(DestA && DestB &&
4479 "Single-destination switch should have been folded.");
4480 assert(DestA != DestB);
4481 assert(DestB != SI->getDefaultDest());
4482 assert(!CasesB.empty() && "There must be non-default cases.");
4483 assert(!CasesA.empty() || HasDefault);
4484
4485 // Figure out if one of the sets of cases form a contiguous range.
4486 SmallVectorImpl<ConstantInt *> *ContiguousCases = nullptr;
4487 BasicBlock *ContiguousDest = nullptr;
4488 BasicBlock *OtherDest = nullptr;
4489 if (!CasesA.empty() && CasesAreContiguous(CasesA)) {
4490 ContiguousCases = &CasesA;
4491 ContiguousDest = DestA;
4492 OtherDest = DestB;
4493 } else if (CasesAreContiguous(CasesB)) {
4494 ContiguousCases = &CasesB;
4495 ContiguousDest = DestB;
4496 OtherDest = DestA;
4497 } else
4498 return false;
4499
4500 // Start building the compare and branch.
4501
4502 Constant *Offset = ConstantExpr::getNeg(ContiguousCases->back());
4503 Constant *NumCases =
4504 ConstantInt::get(Offset->getType(), ContiguousCases->size());
4505
4506 Value *Sub = SI->getCondition();
4507 if (!Offset->isNullValue())
4508 Sub = Builder.CreateAdd(Sub, Offset, Sub->getName() + ".off");
4509
4510 Value *Cmp;
4511 // If NumCases overflowed, then all possible values jump to the successor.
4512 if (NumCases->isNullValue() && !ContiguousCases->empty())
4513 Cmp = ConstantInt::getTrue(SI->getContext());
4514 else
4515 Cmp = Builder.CreateICmpULT(Sub, NumCases, "switch");
4516 BranchInst *NewBI = Builder.CreateCondBr(Cmp, ContiguousDest, OtherDest);
4517
4518 // Update weight for the newly-created conditional branch.
4519 if (HasBranchWeights(SI)) {
4520 SmallVector<uint64_t, 8> Weights;
4521 GetBranchWeights(SI, Weights);
4522 if (Weights.size() == 1 + SI->getNumCases()) {
4523 uint64_t TrueWeight = 0;
4524 uint64_t FalseWeight = 0;
4525 for (size_t I = 0, E = Weights.size(); I != E; ++I) {
4526 if (SI->getSuccessor(I) == ContiguousDest)
4527 TrueWeight += Weights[I];
4528 else
4529 FalseWeight += Weights[I];
4530 }
4531 while (TrueWeight > UINT32_MAX || FalseWeight > UINT32_MAX) {
4532 TrueWeight /= 2;
4533 FalseWeight /= 2;
4534 }
4535 setBranchWeights(NewBI, TrueWeight, FalseWeight);
4536 }
4537 }
4538
4539 // Prune obsolete incoming values off the successors' PHI nodes.
4540 for (auto BBI = ContiguousDest->begin(); isa<PHINode>(BBI); ++BBI) {
4541 unsigned PreviousEdges = ContiguousCases->size();
4542 if (ContiguousDest == SI->getDefaultDest())
4543 ++PreviousEdges;
4544 for (unsigned I = 0, E = PreviousEdges - 1; I != E; ++I)
4545 cast<PHINode>(BBI)->removeIncomingValue(SI->getParent());
4546 }
4547 for (auto BBI = OtherDest->begin(); isa<PHINode>(BBI); ++BBI) {
4548 unsigned PreviousEdges = SI->getNumCases() - ContiguousCases->size();
4549 if (OtherDest == SI->getDefaultDest())
4550 ++PreviousEdges;
4551 for (unsigned I = 0, E = PreviousEdges - 1; I != E; ++I)
4552 cast<PHINode>(BBI)->removeIncomingValue(SI->getParent());
4553 }
4554
4555 // Clean up the default block - it may have phis or other instructions before
4556 // the unreachable terminator.
4557 if (!HasDefault)
4558 createUnreachableSwitchDefault(SI);
4559
4560 // Drop the switch.
4561 SI->eraseFromParent();
4562
4563 return true;
4564 }
4565
4566 /// Compute masked bits for the condition of a switch
4567 /// and use it to remove dead cases.
eliminateDeadSwitchCases(SwitchInst * SI,AssumptionCache * AC,const DataLayout & DL)4568 static bool eliminateDeadSwitchCases(SwitchInst *SI, AssumptionCache *AC,
4569 const DataLayout &DL) {
4570 Value *Cond = SI->getCondition();
4571 unsigned Bits = Cond->getType()->getIntegerBitWidth();
4572 KnownBits Known = computeKnownBits(Cond, DL, 0, AC, SI);
4573
4574 // We can also eliminate cases by determining that their values are outside of
4575 // the limited range of the condition based on how many significant (non-sign)
4576 // bits are in the condition value.
4577 unsigned ExtraSignBits = ComputeNumSignBits(Cond, DL, 0, AC, SI) - 1;
4578 unsigned MaxSignificantBitsInCond = Bits - ExtraSignBits;
4579
4580 // Gather dead cases.
4581 SmallVector<ConstantInt *, 8> DeadCases;
4582 for (auto &Case : SI->cases()) {
4583 const APInt &CaseVal = Case.getCaseValue()->getValue();
4584 if (Known.Zero.intersects(CaseVal) || !Known.One.isSubsetOf(CaseVal) ||
4585 (CaseVal.getMinSignedBits() > MaxSignificantBitsInCond)) {
4586 DeadCases.push_back(Case.getCaseValue());
4587 LLVM_DEBUG(dbgs() << "SimplifyCFG: switch case " << CaseVal
4588 << " is dead.\n");
4589 }
4590 }
4591
4592 // If we can prove that the cases must cover all possible values, the
4593 // default destination becomes dead and we can remove it. If we know some
4594 // of the bits in the value, we can use that to more precisely compute the
4595 // number of possible unique case values.
4596 bool HasDefault =
4597 !isa<UnreachableInst>(SI->getDefaultDest()->getFirstNonPHIOrDbg());
4598 const unsigned NumUnknownBits =
4599 Bits - (Known.Zero | Known.One).countPopulation();
4600 assert(NumUnknownBits <= Bits);
4601 if (HasDefault && DeadCases.empty() &&
4602 NumUnknownBits < 64 /* avoid overflow */ &&
4603 SI->getNumCases() == (1ULL << NumUnknownBits)) {
4604 createUnreachableSwitchDefault(SI);
4605 return true;
4606 }
4607
4608 if (DeadCases.empty())
4609 return false;
4610
4611 SwitchInstProfUpdateWrapper SIW(*SI);
4612 for (ConstantInt *DeadCase : DeadCases) {
4613 SwitchInst::CaseIt CaseI = SI->findCaseValue(DeadCase);
4614 assert(CaseI != SI->case_default() &&
4615 "Case was not found. Probably mistake in DeadCases forming.");
4616 // Prune unused values from PHI nodes.
4617 CaseI->getCaseSuccessor()->removePredecessor(SI->getParent());
4618 SIW.removeCase(CaseI);
4619 }
4620
4621 return true;
4622 }
4623
4624 /// If BB would be eligible for simplification by
4625 /// TryToSimplifyUncondBranchFromEmptyBlock (i.e. it is empty and terminated
4626 /// by an unconditional branch), look at the phi node for BB in the successor
4627 /// block and see if the incoming value is equal to CaseValue. If so, return
4628 /// the phi node, and set PhiIndex to BB's index in the phi node.
FindPHIForConditionForwarding(ConstantInt * CaseValue,BasicBlock * BB,int * PhiIndex)4629 static PHINode *FindPHIForConditionForwarding(ConstantInt *CaseValue,
4630 BasicBlock *BB, int *PhiIndex) {
4631 if (BB->getFirstNonPHIOrDbg() != BB->getTerminator())
4632 return nullptr; // BB must be empty to be a candidate for simplification.
4633 if (!BB->getSinglePredecessor())
4634 return nullptr; // BB must be dominated by the switch.
4635
4636 BranchInst *Branch = dyn_cast<BranchInst>(BB->getTerminator());
4637 if (!Branch || !Branch->isUnconditional())
4638 return nullptr; // Terminator must be unconditional branch.
4639
4640 BasicBlock *Succ = Branch->getSuccessor(0);
4641
4642 for (PHINode &PHI : Succ->phis()) {
4643 int Idx = PHI.getBasicBlockIndex(BB);
4644 assert(Idx >= 0 && "PHI has no entry for predecessor?");
4645
4646 Value *InValue = PHI.getIncomingValue(Idx);
4647 if (InValue != CaseValue)
4648 continue;
4649
4650 *PhiIndex = Idx;
4651 return &PHI;
4652 }
4653
4654 return nullptr;
4655 }
4656
4657 /// Try to forward the condition of a switch instruction to a phi node
4658 /// dominated by the switch, if that would mean that some of the destination
4659 /// blocks of the switch can be folded away. Return true if a change is made.
ForwardSwitchConditionToPHI(SwitchInst * SI)4660 static bool ForwardSwitchConditionToPHI(SwitchInst *SI) {
4661 using ForwardingNodesMap = DenseMap<PHINode *, SmallVector<int, 4>>;
4662
4663 ForwardingNodesMap ForwardingNodes;
4664 BasicBlock *SwitchBlock = SI->getParent();
4665 bool Changed = false;
4666 for (auto &Case : SI->cases()) {
4667 ConstantInt *CaseValue = Case.getCaseValue();
4668 BasicBlock *CaseDest = Case.getCaseSuccessor();
4669
4670 // Replace phi operands in successor blocks that are using the constant case
4671 // value rather than the switch condition variable:
4672 // switchbb:
4673 // switch i32 %x, label %default [
4674 // i32 17, label %succ
4675 // ...
4676 // succ:
4677 // %r = phi i32 ... [ 17, %switchbb ] ...
4678 // -->
4679 // %r = phi i32 ... [ %x, %switchbb ] ...
4680
4681 for (PHINode &Phi : CaseDest->phis()) {
4682 // This only works if there is exactly 1 incoming edge from the switch to
4683 // a phi. If there is >1, that means multiple cases of the switch map to 1
4684 // value in the phi, and that phi value is not the switch condition. Thus,
4685 // this transform would not make sense (the phi would be invalid because
4686 // a phi can't have different incoming values from the same block).
4687 int SwitchBBIdx = Phi.getBasicBlockIndex(SwitchBlock);
4688 if (Phi.getIncomingValue(SwitchBBIdx) == CaseValue &&
4689 count(Phi.blocks(), SwitchBlock) == 1) {
4690 Phi.setIncomingValue(SwitchBBIdx, SI->getCondition());
4691 Changed = true;
4692 }
4693 }
4694
4695 // Collect phi nodes that are indirectly using this switch's case constants.
4696 int PhiIdx;
4697 if (auto *Phi = FindPHIForConditionForwarding(CaseValue, CaseDest, &PhiIdx))
4698 ForwardingNodes[Phi].push_back(PhiIdx);
4699 }
4700
4701 for (auto &ForwardingNode : ForwardingNodes) {
4702 PHINode *Phi = ForwardingNode.first;
4703 SmallVectorImpl<int> &Indexes = ForwardingNode.second;
4704 if (Indexes.size() < 2)
4705 continue;
4706
4707 for (int Index : Indexes)
4708 Phi->setIncomingValue(Index, SI->getCondition());
4709 Changed = true;
4710 }
4711
4712 return Changed;
4713 }
4714
4715 /// Return true if the backend will be able to handle
4716 /// initializing an array of constants like C.
ValidLookupTableConstant(Constant * C,const TargetTransformInfo & TTI)4717 static bool ValidLookupTableConstant(Constant *C, const TargetTransformInfo &TTI) {
4718 if (C->isThreadDependent())
4719 return false;
4720 if (C->isDLLImportDependent())
4721 return false;
4722
4723 if (!isa<ConstantFP>(C) && !isa<ConstantInt>(C) &&
4724 !isa<ConstantPointerNull>(C) && !isa<GlobalValue>(C) &&
4725 !isa<UndefValue>(C) && !isa<ConstantExpr>(C))
4726 return false;
4727
4728 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) {
4729 if (!CE->isGEPWithNoNotionalOverIndexing())
4730 return false;
4731 if (!ValidLookupTableConstant(CE->getOperand(0), TTI))
4732 return false;
4733 }
4734
4735 if (!TTI.shouldBuildLookupTablesForConstant(C))
4736 return false;
4737
4738 return true;
4739 }
4740
4741 /// If V is a Constant, return it. Otherwise, try to look up
4742 /// its constant value in ConstantPool, returning 0 if it's not there.
4743 static Constant *
LookupConstant(Value * V,const SmallDenseMap<Value *,Constant * > & ConstantPool)4744 LookupConstant(Value *V,
4745 const SmallDenseMap<Value *, Constant *> &ConstantPool) {
4746 if (Constant *C = dyn_cast<Constant>(V))
4747 return C;
4748 return ConstantPool.lookup(V);
4749 }
4750
4751 /// Try to fold instruction I into a constant. This works for
4752 /// simple instructions such as binary operations where both operands are
4753 /// constant or can be replaced by constants from the ConstantPool. Returns the
4754 /// resulting constant on success, 0 otherwise.
4755 static Constant *
ConstantFold(Instruction * I,const DataLayout & DL,const SmallDenseMap<Value *,Constant * > & ConstantPool)4756 ConstantFold(Instruction *I, const DataLayout &DL,
4757 const SmallDenseMap<Value *, Constant *> &ConstantPool) {
4758 if (SelectInst *Select = dyn_cast<SelectInst>(I)) {
4759 Constant *A = LookupConstant(Select->getCondition(), ConstantPool);
4760 if (!A)
4761 return nullptr;
4762 if (A->isAllOnesValue())
4763 return LookupConstant(Select->getTrueValue(), ConstantPool);
4764 if (A->isNullValue())
4765 return LookupConstant(Select->getFalseValue(), ConstantPool);
4766 return nullptr;
4767 }
4768
4769 SmallVector<Constant *, 4> COps;
4770 for (unsigned N = 0, E = I->getNumOperands(); N != E; ++N) {
4771 if (Constant *A = LookupConstant(I->getOperand(N), ConstantPool))
4772 COps.push_back(A);
4773 else
4774 return nullptr;
4775 }
4776
4777 if (CmpInst *Cmp = dyn_cast<CmpInst>(I)) {
4778 return ConstantFoldCompareInstOperands(Cmp->getPredicate(), COps[0],
4779 COps[1], DL);
4780 }
4781
4782 return ConstantFoldInstOperands(I, COps, DL);
4783 }
4784
4785 /// Try to determine the resulting constant values in phi nodes
4786 /// at the common destination basic block, *CommonDest, for one of the case
4787 /// destionations CaseDest corresponding to value CaseVal (0 for the default
4788 /// case), of a switch instruction SI.
4789 static bool
GetCaseResults(SwitchInst * SI,ConstantInt * CaseVal,BasicBlock * CaseDest,BasicBlock ** CommonDest,SmallVectorImpl<std::pair<PHINode *,Constant * >> & Res,const DataLayout & DL,const TargetTransformInfo & TTI)4790 GetCaseResults(SwitchInst *SI, ConstantInt *CaseVal, BasicBlock *CaseDest,
4791 BasicBlock **CommonDest,
4792 SmallVectorImpl<std::pair<PHINode *, Constant *>> &Res,
4793 const DataLayout &DL, const TargetTransformInfo &TTI) {
4794 // The block from which we enter the common destination.
4795 BasicBlock *Pred = SI->getParent();
4796
4797 // If CaseDest is empty except for some side-effect free instructions through
4798 // which we can constant-propagate the CaseVal, continue to its successor.
4799 SmallDenseMap<Value *, Constant *> ConstantPool;
4800 ConstantPool.insert(std::make_pair(SI->getCondition(), CaseVal));
4801 for (Instruction &I :CaseDest->instructionsWithoutDebug()) {
4802 if (I.isTerminator()) {
4803 // If the terminator is a simple branch, continue to the next block.
4804 if (I.getNumSuccessors() != 1 || I.isExceptionalTerminator())
4805 return false;
4806 Pred = CaseDest;
4807 CaseDest = I.getSuccessor(0);
4808 } else if (Constant *C = ConstantFold(&I, DL, ConstantPool)) {
4809 // Instruction is side-effect free and constant.
4810
4811 // If the instruction has uses outside this block or a phi node slot for
4812 // the block, it is not safe to bypass the instruction since it would then
4813 // no longer dominate all its uses.
4814 for (auto &Use : I.uses()) {
4815 User *User = Use.getUser();
4816 if (Instruction *I = dyn_cast<Instruction>(User))
4817 if (I->getParent() == CaseDest)
4818 continue;
4819 if (PHINode *Phi = dyn_cast<PHINode>(User))
4820 if (Phi->getIncomingBlock(Use) == CaseDest)
4821 continue;
4822 return false;
4823 }
4824
4825 ConstantPool.insert(std::make_pair(&I, C));
4826 } else {
4827 break;
4828 }
4829 }
4830
4831 // If we did not have a CommonDest before, use the current one.
4832 if (!*CommonDest)
4833 *CommonDest = CaseDest;
4834 // If the destination isn't the common one, abort.
4835 if (CaseDest != *CommonDest)
4836 return false;
4837
4838 // Get the values for this case from phi nodes in the destination block.
4839 for (PHINode &PHI : (*CommonDest)->phis()) {
4840 int Idx = PHI.getBasicBlockIndex(Pred);
4841 if (Idx == -1)
4842 continue;
4843
4844 Constant *ConstVal =
4845 LookupConstant(PHI.getIncomingValue(Idx), ConstantPool);
4846 if (!ConstVal)
4847 return false;
4848
4849 // Be conservative about which kinds of constants we support.
4850 if (!ValidLookupTableConstant(ConstVal, TTI))
4851 return false;
4852
4853 Res.push_back(std::make_pair(&PHI, ConstVal));
4854 }
4855
4856 return Res.size() > 0;
4857 }
4858
4859 // Helper function used to add CaseVal to the list of cases that generate
4860 // Result. Returns the updated number of cases that generate this result.
MapCaseToResult(ConstantInt * CaseVal,SwitchCaseResultVectorTy & UniqueResults,Constant * Result)4861 static uintptr_t MapCaseToResult(ConstantInt *CaseVal,
4862 SwitchCaseResultVectorTy &UniqueResults,
4863 Constant *Result) {
4864 for (auto &I : UniqueResults) {
4865 if (I.first == Result) {
4866 I.second.push_back(CaseVal);
4867 return I.second.size();
4868 }
4869 }
4870 UniqueResults.push_back(
4871 std::make_pair(Result, SmallVector<ConstantInt *, 4>(1, CaseVal)));
4872 return 1;
4873 }
4874
4875 // Helper function that initializes a map containing
4876 // results for the PHI node of the common destination block for a switch
4877 // instruction. Returns false if multiple PHI nodes have been found or if
4878 // there is not a common destination block for the switch.
4879 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)4880 InitializeUniqueCases(SwitchInst *SI, PHINode *&PHI, BasicBlock *&CommonDest,
4881 SwitchCaseResultVectorTy &UniqueResults,
4882 Constant *&DefaultResult, const DataLayout &DL,
4883 const TargetTransformInfo &TTI,
4884 uintptr_t MaxUniqueResults, uintptr_t MaxCasesPerResult) {
4885 for (auto &I : SI->cases()) {
4886 ConstantInt *CaseVal = I.getCaseValue();
4887
4888 // Resulting value at phi nodes for this case value.
4889 SwitchCaseResultsTy Results;
4890 if (!GetCaseResults(SI, CaseVal, I.getCaseSuccessor(), &CommonDest, Results,
4891 DL, TTI))
4892 return false;
4893
4894 // Only one value per case is permitted.
4895 if (Results.size() > 1)
4896 return false;
4897
4898 // Add the case->result mapping to UniqueResults.
4899 const uintptr_t NumCasesForResult =
4900 MapCaseToResult(CaseVal, UniqueResults, Results.begin()->second);
4901
4902 // Early out if there are too many cases for this result.
4903 if (NumCasesForResult > MaxCasesPerResult)
4904 return false;
4905
4906 // Early out if there are too many unique results.
4907 if (UniqueResults.size() > MaxUniqueResults)
4908 return false;
4909
4910 // Check the PHI consistency.
4911 if (!PHI)
4912 PHI = Results[0].first;
4913 else if (PHI != Results[0].first)
4914 return false;
4915 }
4916 // Find the default result value.
4917 SmallVector<std::pair<PHINode *, Constant *>, 1> DefaultResults;
4918 BasicBlock *DefaultDest = SI->getDefaultDest();
4919 GetCaseResults(SI, nullptr, SI->getDefaultDest(), &CommonDest, DefaultResults,
4920 DL, TTI);
4921 // If the default value is not found abort unless the default destination
4922 // is unreachable.
4923 DefaultResult =
4924 DefaultResults.size() == 1 ? DefaultResults.begin()->second : nullptr;
4925 if ((!DefaultResult &&
4926 !isa<UnreachableInst>(DefaultDest->getFirstNonPHIOrDbg())))
4927 return false;
4928
4929 return true;
4930 }
4931
4932 // Helper function that checks if it is possible to transform a switch with only
4933 // two cases (or two cases + default) that produces a result into a select.
4934 // Example:
4935 // switch (a) {
4936 // case 10: %0 = icmp eq i32 %a, 10
4937 // return 10; %1 = select i1 %0, i32 10, i32 4
4938 // case 20: ----> %2 = icmp eq i32 %a, 20
4939 // return 2; %3 = select i1 %2, i32 2, i32 %1
4940 // default:
4941 // return 4;
4942 // }
ConvertTwoCaseSwitch(const SwitchCaseResultVectorTy & ResultVector,Constant * DefaultResult,Value * Condition,IRBuilder<> & Builder)4943 static Value *ConvertTwoCaseSwitch(const SwitchCaseResultVectorTy &ResultVector,
4944 Constant *DefaultResult, Value *Condition,
4945 IRBuilder<> &Builder) {
4946 assert(ResultVector.size() == 2 &&
4947 "We should have exactly two unique results at this point");
4948 // If we are selecting between only two cases transform into a simple
4949 // select or a two-way select if default is possible.
4950 if (ResultVector[0].second.size() == 1 &&
4951 ResultVector[1].second.size() == 1) {
4952 ConstantInt *const FirstCase = ResultVector[0].second[0];
4953 ConstantInt *const SecondCase = ResultVector[1].second[0];
4954
4955 bool DefaultCanTrigger = DefaultResult;
4956 Value *SelectValue = ResultVector[1].first;
4957 if (DefaultCanTrigger) {
4958 Value *const ValueCompare =
4959 Builder.CreateICmpEQ(Condition, SecondCase, "switch.selectcmp");
4960 SelectValue = Builder.CreateSelect(ValueCompare, ResultVector[1].first,
4961 DefaultResult, "switch.select");
4962 }
4963 Value *const ValueCompare =
4964 Builder.CreateICmpEQ(Condition, FirstCase, "switch.selectcmp");
4965 return Builder.CreateSelect(ValueCompare, ResultVector[0].first,
4966 SelectValue, "switch.select");
4967 }
4968
4969 return nullptr;
4970 }
4971
4972 // Helper function to cleanup a switch instruction that has been converted into
4973 // a select, fixing up PHI nodes and basic blocks.
RemoveSwitchAfterSelectConversion(SwitchInst * SI,PHINode * PHI,Value * SelectValue,IRBuilder<> & Builder)4974 static void RemoveSwitchAfterSelectConversion(SwitchInst *SI, PHINode *PHI,
4975 Value *SelectValue,
4976 IRBuilder<> &Builder) {
4977 BasicBlock *SelectBB = SI->getParent();
4978 while (PHI->getBasicBlockIndex(SelectBB) >= 0)
4979 PHI->removeIncomingValue(SelectBB);
4980 PHI->addIncoming(SelectValue, SelectBB);
4981
4982 Builder.CreateBr(PHI->getParent());
4983
4984 // Remove the switch.
4985 for (unsigned i = 0, e = SI->getNumSuccessors(); i < e; ++i) {
4986 BasicBlock *Succ = SI->getSuccessor(i);
4987
4988 if (Succ == PHI->getParent())
4989 continue;
4990 Succ->removePredecessor(SelectBB);
4991 }
4992 SI->eraseFromParent();
4993 }
4994
4995 /// If the switch is only used to initialize one or more
4996 /// phi nodes in a common successor block with only two different
4997 /// constant values, replace the switch with select.
switchToSelect(SwitchInst * SI,IRBuilder<> & Builder,const DataLayout & DL,const TargetTransformInfo & TTI)4998 static bool switchToSelect(SwitchInst *SI, IRBuilder<> &Builder,
4999 const DataLayout &DL,
5000 const TargetTransformInfo &TTI) {
5001 Value *const Cond = SI->getCondition();
5002 PHINode *PHI = nullptr;
5003 BasicBlock *CommonDest = nullptr;
5004 Constant *DefaultResult;
5005 SwitchCaseResultVectorTy UniqueResults;
5006 // Collect all the cases that will deliver the same value from the switch.
5007 if (!InitializeUniqueCases(SI, PHI, CommonDest, UniqueResults, DefaultResult,
5008 DL, TTI, 2, 1))
5009 return false;
5010 // Selects choose between maximum two values.
5011 if (UniqueResults.size() != 2)
5012 return false;
5013 assert(PHI != nullptr && "PHI for value select not found");
5014
5015 Builder.SetInsertPoint(SI);
5016 Value *SelectValue =
5017 ConvertTwoCaseSwitch(UniqueResults, DefaultResult, Cond, Builder);
5018 if (SelectValue) {
5019 RemoveSwitchAfterSelectConversion(SI, PHI, SelectValue, Builder);
5020 return true;
5021 }
5022 // The switch couldn't be converted into a select.
5023 return false;
5024 }
5025
5026 namespace {
5027
5028 /// This class represents a lookup table that can be used to replace a switch.
5029 class SwitchLookupTable {
5030 public:
5031 /// Create a lookup table to use as a switch replacement with the contents
5032 /// of Values, using DefaultValue to fill any holes in the table.
5033 SwitchLookupTable(
5034 Module &M, uint64_t TableSize, ConstantInt *Offset,
5035 const SmallVectorImpl<std::pair<ConstantInt *, Constant *>> &Values,
5036 Constant *DefaultValue, const DataLayout &DL, const StringRef &FuncName);
5037
5038 /// Build instructions with Builder to retrieve the value at
5039 /// the position given by Index in the lookup table.
5040 Value *BuildLookup(Value *Index, IRBuilder<> &Builder);
5041
5042 /// Return true if a table with TableSize elements of
5043 /// type ElementType would fit in a target-legal register.
5044 static bool WouldFitInRegister(const DataLayout &DL, uint64_t TableSize,
5045 Type *ElementType);
5046
5047 private:
5048 // Depending on the contents of the table, it can be represented in
5049 // different ways.
5050 enum {
5051 // For tables where each element contains the same value, we just have to
5052 // store that single value and return it for each lookup.
5053 SingleValueKind,
5054
5055 // For tables where there is a linear relationship between table index
5056 // and values. We calculate the result with a simple multiplication
5057 // and addition instead of a table lookup.
5058 LinearMapKind,
5059
5060 // For small tables with integer elements, we can pack them into a bitmap
5061 // that fits into a target-legal register. Values are retrieved by
5062 // shift and mask operations.
5063 BitMapKind,
5064
5065 // The table is stored as an array of values. Values are retrieved by load
5066 // instructions from the table.
5067 ArrayKind
5068 } Kind;
5069
5070 // For SingleValueKind, this is the single value.
5071 Constant *SingleValue = nullptr;
5072
5073 // For BitMapKind, this is the bitmap.
5074 ConstantInt *BitMap = nullptr;
5075 IntegerType *BitMapElementTy = nullptr;
5076
5077 // For LinearMapKind, these are the constants used to derive the value.
5078 ConstantInt *LinearOffset = nullptr;
5079 ConstantInt *LinearMultiplier = nullptr;
5080
5081 // For ArrayKind, this is the array.
5082 GlobalVariable *Array = nullptr;
5083 };
5084
5085 } // end anonymous namespace
5086
SwitchLookupTable(Module & M,uint64_t TableSize,ConstantInt * Offset,const SmallVectorImpl<std::pair<ConstantInt *,Constant * >> & Values,Constant * DefaultValue,const DataLayout & DL,const StringRef & FuncName)5087 SwitchLookupTable::SwitchLookupTable(
5088 Module &M, uint64_t TableSize, ConstantInt *Offset,
5089 const SmallVectorImpl<std::pair<ConstantInt *, Constant *>> &Values,
5090 Constant *DefaultValue, const DataLayout &DL, const StringRef &FuncName) {
5091 assert(Values.size() && "Can't build lookup table without values!");
5092 assert(TableSize >= Values.size() && "Can't fit values in table!");
5093
5094 // If all values in the table are equal, this is that value.
5095 SingleValue = Values.begin()->second;
5096
5097 Type *ValueType = Values.begin()->second->getType();
5098
5099 // Build up the table contents.
5100 SmallVector<Constant *, 64> TableContents(TableSize);
5101 for (size_t I = 0, E = Values.size(); I != E; ++I) {
5102 ConstantInt *CaseVal = Values[I].first;
5103 Constant *CaseRes = Values[I].second;
5104 assert(CaseRes->getType() == ValueType);
5105
5106 uint64_t Idx = (CaseVal->getValue() - Offset->getValue()).getLimitedValue();
5107 TableContents[Idx] = CaseRes;
5108
5109 if (CaseRes != SingleValue)
5110 SingleValue = nullptr;
5111 }
5112
5113 // Fill in any holes in the table with the default result.
5114 if (Values.size() < TableSize) {
5115 assert(DefaultValue &&
5116 "Need a default value to fill the lookup table holes.");
5117 assert(DefaultValue->getType() == ValueType);
5118 for (uint64_t I = 0; I < TableSize; ++I) {
5119 if (!TableContents[I])
5120 TableContents[I] = DefaultValue;
5121 }
5122
5123 if (DefaultValue != SingleValue)
5124 SingleValue = nullptr;
5125 }
5126
5127 // If each element in the table contains the same value, we only need to store
5128 // that single value.
5129 if (SingleValue) {
5130 Kind = SingleValueKind;
5131 return;
5132 }
5133
5134 // Check if we can derive the value with a linear transformation from the
5135 // table index.
5136 if (isa<IntegerType>(ValueType)) {
5137 bool LinearMappingPossible = true;
5138 APInt PrevVal;
5139 APInt DistToPrev;
5140 assert(TableSize >= 2 && "Should be a SingleValue table.");
5141 // Check if there is the same distance between two consecutive values.
5142 for (uint64_t I = 0; I < TableSize; ++I) {
5143 ConstantInt *ConstVal = dyn_cast<ConstantInt>(TableContents[I]);
5144 if (!ConstVal) {
5145 // This is an undef. We could deal with it, but undefs in lookup tables
5146 // are very seldom. It's probably not worth the additional complexity.
5147 LinearMappingPossible = false;
5148 break;
5149 }
5150 const APInt &Val = ConstVal->getValue();
5151 if (I != 0) {
5152 APInt Dist = Val - PrevVal;
5153 if (I == 1) {
5154 DistToPrev = Dist;
5155 } else if (Dist != DistToPrev) {
5156 LinearMappingPossible = false;
5157 break;
5158 }
5159 }
5160 PrevVal = Val;
5161 }
5162 if (LinearMappingPossible) {
5163 LinearOffset = cast<ConstantInt>(TableContents[0]);
5164 LinearMultiplier = ConstantInt::get(M.getContext(), DistToPrev);
5165 Kind = LinearMapKind;
5166 ++NumLinearMaps;
5167 return;
5168 }
5169 }
5170
5171 // If the type is integer and the table fits in a register, build a bitmap.
5172 if (WouldFitInRegister(DL, TableSize, ValueType)) {
5173 IntegerType *IT = cast<IntegerType>(ValueType);
5174 APInt TableInt(TableSize * IT->getBitWidth(), 0);
5175 for (uint64_t I = TableSize; I > 0; --I) {
5176 TableInt <<= IT->getBitWidth();
5177 // Insert values into the bitmap. Undef values are set to zero.
5178 if (!isa<UndefValue>(TableContents[I - 1])) {
5179 ConstantInt *Val = cast<ConstantInt>(TableContents[I - 1]);
5180 TableInt |= Val->getValue().zext(TableInt.getBitWidth());
5181 }
5182 }
5183 BitMap = ConstantInt::get(M.getContext(), TableInt);
5184 BitMapElementTy = IT;
5185 Kind = BitMapKind;
5186 ++NumBitMaps;
5187 return;
5188 }
5189
5190 // Store the table in an array.
5191 ArrayType *ArrayTy = ArrayType::get(ValueType, TableSize);
5192 Constant *Initializer = ConstantArray::get(ArrayTy, TableContents);
5193
5194 Array = new GlobalVariable(M, ArrayTy, /*isConstant=*/true,
5195 GlobalVariable::PrivateLinkage, Initializer,
5196 "switch.table." + FuncName);
5197 Array->setUnnamedAddr(GlobalValue::UnnamedAddr::Global);
5198 // Set the alignment to that of an array items. We will be only loading one
5199 // value out of it.
5200 Array->setAlignment(Align(DL.getPrefTypeAlignment(ValueType)));
5201 Kind = ArrayKind;
5202 }
5203
BuildLookup(Value * Index,IRBuilder<> & Builder)5204 Value *SwitchLookupTable::BuildLookup(Value *Index, IRBuilder<> &Builder) {
5205 switch (Kind) {
5206 case SingleValueKind:
5207 return SingleValue;
5208 case LinearMapKind: {
5209 // Derive the result value from the input value.
5210 Value *Result = Builder.CreateIntCast(Index, LinearMultiplier->getType(),
5211 false, "switch.idx.cast");
5212 if (!LinearMultiplier->isOne())
5213 Result = Builder.CreateMul(Result, LinearMultiplier, "switch.idx.mult");
5214 if (!LinearOffset->isZero())
5215 Result = Builder.CreateAdd(Result, LinearOffset, "switch.offset");
5216 return Result;
5217 }
5218 case BitMapKind: {
5219 // Type of the bitmap (e.g. i59).
5220 IntegerType *MapTy = BitMap->getType();
5221
5222 // Cast Index to the same type as the bitmap.
5223 // Note: The Index is <= the number of elements in the table, so
5224 // truncating it to the width of the bitmask is safe.
5225 Value *ShiftAmt = Builder.CreateZExtOrTrunc(Index, MapTy, "switch.cast");
5226
5227 // Multiply the shift amount by the element width.
5228 ShiftAmt = Builder.CreateMul(
5229 ShiftAmt, ConstantInt::get(MapTy, BitMapElementTy->getBitWidth()),
5230 "switch.shiftamt");
5231
5232 // Shift down.
5233 Value *DownShifted =
5234 Builder.CreateLShr(BitMap, ShiftAmt, "switch.downshift");
5235 // Mask off.
5236 return Builder.CreateTrunc(DownShifted, BitMapElementTy, "switch.masked");
5237 }
5238 case ArrayKind: {
5239 // Make sure the table index will not overflow when treated as signed.
5240 IntegerType *IT = cast<IntegerType>(Index->getType());
5241 uint64_t TableSize =
5242 Array->getInitializer()->getType()->getArrayNumElements();
5243 if (TableSize > (1ULL << (IT->getBitWidth() - 1)))
5244 Index = Builder.CreateZExt(
5245 Index, IntegerType::get(IT->getContext(), IT->getBitWidth() + 1),
5246 "switch.tableidx.zext");
5247
5248 Value *GEPIndices[] = {Builder.getInt32(0), Index};
5249 Value *GEP = Builder.CreateInBoundsGEP(Array->getValueType(), Array,
5250 GEPIndices, "switch.gep");
5251 return Builder.CreateLoad(
5252 cast<ArrayType>(Array->getValueType())->getElementType(), GEP,
5253 "switch.load");
5254 }
5255 }
5256 llvm_unreachable("Unknown lookup table kind!");
5257 }
5258
WouldFitInRegister(const DataLayout & DL,uint64_t TableSize,Type * ElementType)5259 bool SwitchLookupTable::WouldFitInRegister(const DataLayout &DL,
5260 uint64_t TableSize,
5261 Type *ElementType) {
5262 auto *IT = dyn_cast<IntegerType>(ElementType);
5263 if (!IT)
5264 return false;
5265 // FIXME: If the type is wider than it needs to be, e.g. i8 but all values
5266 // are <= 15, we could try to narrow the type.
5267
5268 // Avoid overflow, fitsInLegalInteger uses unsigned int for the width.
5269 if (TableSize >= UINT_MAX / IT->getBitWidth())
5270 return false;
5271 return DL.fitsInLegalInteger(TableSize * IT->getBitWidth());
5272 }
5273
5274 /// Determine whether a lookup table should be built for this switch, based on
5275 /// the number of cases, size of the table, and the types of the results.
5276 static bool
ShouldBuildLookupTable(SwitchInst * SI,uint64_t TableSize,const TargetTransformInfo & TTI,const DataLayout & DL,const SmallDenseMap<PHINode *,Type * > & ResultTypes)5277 ShouldBuildLookupTable(SwitchInst *SI, uint64_t TableSize,
5278 const TargetTransformInfo &TTI, const DataLayout &DL,
5279 const SmallDenseMap<PHINode *, Type *> &ResultTypes) {
5280 if (SI->getNumCases() > TableSize || TableSize >= UINT64_MAX / 10)
5281 return false; // TableSize overflowed, or mul below might overflow.
5282
5283 bool AllTablesFitInRegister = true;
5284 bool HasIllegalType = false;
5285 for (const auto &I : ResultTypes) {
5286 Type *Ty = I.second;
5287
5288 // Saturate this flag to true.
5289 HasIllegalType = HasIllegalType || !TTI.isTypeLegal(Ty);
5290
5291 // Saturate this flag to false.
5292 AllTablesFitInRegister =
5293 AllTablesFitInRegister &&
5294 SwitchLookupTable::WouldFitInRegister(DL, TableSize, Ty);
5295
5296 // If both flags saturate, we're done. NOTE: This *only* works with
5297 // saturating flags, and all flags have to saturate first due to the
5298 // non-deterministic behavior of iterating over a dense map.
5299 if (HasIllegalType && !AllTablesFitInRegister)
5300 break;
5301 }
5302
5303 // If each table would fit in a register, we should build it anyway.
5304 if (AllTablesFitInRegister)
5305 return true;
5306
5307 // Don't build a table that doesn't fit in-register if it has illegal types.
5308 if (HasIllegalType)
5309 return false;
5310
5311 // The table density should be at least 40%. This is the same criterion as for
5312 // jump tables, see SelectionDAGBuilder::handleJTSwitchCase.
5313 // FIXME: Find the best cut-off.
5314 return SI->getNumCases() * 10 >= TableSize * 4;
5315 }
5316
5317 /// Try to reuse the switch table index compare. Following pattern:
5318 /// \code
5319 /// if (idx < tablesize)
5320 /// r = table[idx]; // table does not contain default_value
5321 /// else
5322 /// r = default_value;
5323 /// if (r != default_value)
5324 /// ...
5325 /// \endcode
5326 /// Is optimized to:
5327 /// \code
5328 /// cond = idx < tablesize;
5329 /// if (cond)
5330 /// r = table[idx];
5331 /// else
5332 /// r = default_value;
5333 /// if (cond)
5334 /// ...
5335 /// \endcode
5336 /// 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)5337 static void reuseTableCompare(
5338 User *PhiUser, BasicBlock *PhiBlock, BranchInst *RangeCheckBranch,
5339 Constant *DefaultValue,
5340 const SmallVectorImpl<std::pair<ConstantInt *, Constant *>> &Values) {
5341 ICmpInst *CmpInst = dyn_cast<ICmpInst>(PhiUser);
5342 if (!CmpInst)
5343 return;
5344
5345 // We require that the compare is in the same block as the phi so that jump
5346 // threading can do its work afterwards.
5347 if (CmpInst->getParent() != PhiBlock)
5348 return;
5349
5350 Constant *CmpOp1 = dyn_cast<Constant>(CmpInst->getOperand(1));
5351 if (!CmpOp1)
5352 return;
5353
5354 Value *RangeCmp = RangeCheckBranch->getCondition();
5355 Constant *TrueConst = ConstantInt::getTrue(RangeCmp->getType());
5356 Constant *FalseConst = ConstantInt::getFalse(RangeCmp->getType());
5357
5358 // Check if the compare with the default value is constant true or false.
5359 Constant *DefaultConst = ConstantExpr::getICmp(CmpInst->getPredicate(),
5360 DefaultValue, CmpOp1, true);
5361 if (DefaultConst != TrueConst && DefaultConst != FalseConst)
5362 return;
5363
5364 // Check if the compare with the case values is distinct from the default
5365 // compare result.
5366 for (auto ValuePair : Values) {
5367 Constant *CaseConst = ConstantExpr::getICmp(CmpInst->getPredicate(),
5368 ValuePair.second, CmpOp1, true);
5369 if (!CaseConst || CaseConst == DefaultConst || isa<UndefValue>(CaseConst))
5370 return;
5371 assert((CaseConst == TrueConst || CaseConst == FalseConst) &&
5372 "Expect true or false as compare result.");
5373 }
5374
5375 // Check if the branch instruction dominates the phi node. It's a simple
5376 // dominance check, but sufficient for our needs.
5377 // Although this check is invariant in the calling loops, it's better to do it
5378 // at this late stage. Practically we do it at most once for a switch.
5379 BasicBlock *BranchBlock = RangeCheckBranch->getParent();
5380 for (auto PI = pred_begin(PhiBlock), E = pred_end(PhiBlock); PI != E; ++PI) {
5381 BasicBlock *Pred = *PI;
5382 if (Pred != BranchBlock && Pred->getUniquePredecessor() != BranchBlock)
5383 return;
5384 }
5385
5386 if (DefaultConst == FalseConst) {
5387 // The compare yields the same result. We can replace it.
5388 CmpInst->replaceAllUsesWith(RangeCmp);
5389 ++NumTableCmpReuses;
5390 } else {
5391 // The compare yields the same result, just inverted. We can replace it.
5392 Value *InvertedTableCmp = BinaryOperator::CreateXor(
5393 RangeCmp, ConstantInt::get(RangeCmp->getType(), 1), "inverted.cmp",
5394 RangeCheckBranch);
5395 CmpInst->replaceAllUsesWith(InvertedTableCmp);
5396 ++NumTableCmpReuses;
5397 }
5398 }
5399
5400 /// If the switch is only used to initialize one or more phi nodes in a common
5401 /// successor block with different constant values, replace the switch with
5402 /// lookup tables.
SwitchToLookupTable(SwitchInst * SI,IRBuilder<> & Builder,const DataLayout & DL,const TargetTransformInfo & TTI)5403 static bool SwitchToLookupTable(SwitchInst *SI, IRBuilder<> &Builder,
5404 const DataLayout &DL,
5405 const TargetTransformInfo &TTI) {
5406 assert(SI->getNumCases() > 1 && "Degenerate switch?");
5407
5408 Function *Fn = SI->getParent()->getParent();
5409 // Only build lookup table when we have a target that supports it or the
5410 // attribute is not set.
5411 if (!TTI.shouldBuildLookupTables() ||
5412 (Fn->getFnAttribute("no-jump-tables").getValueAsString() == "true"))
5413 return false;
5414
5415 // FIXME: If the switch is too sparse for a lookup table, perhaps we could
5416 // split off a dense part and build a lookup table for that.
5417
5418 // FIXME: This creates arrays of GEPs to constant strings, which means each
5419 // GEP needs a runtime relocation in PIC code. We should just build one big
5420 // string and lookup indices into that.
5421
5422 // Ignore switches with less than three cases. Lookup tables will not make
5423 // them faster, so we don't analyze them.
5424 if (SI->getNumCases() < 3)
5425 return false;
5426
5427 // Figure out the corresponding result for each case value and phi node in the
5428 // common destination, as well as the min and max case values.
5429 assert(!SI->cases().empty());
5430 SwitchInst::CaseIt CI = SI->case_begin();
5431 ConstantInt *MinCaseVal = CI->getCaseValue();
5432 ConstantInt *MaxCaseVal = CI->getCaseValue();
5433
5434 BasicBlock *CommonDest = nullptr;
5435
5436 using ResultListTy = SmallVector<std::pair<ConstantInt *, Constant *>, 4>;
5437 SmallDenseMap<PHINode *, ResultListTy> ResultLists;
5438
5439 SmallDenseMap<PHINode *, Constant *> DefaultResults;
5440 SmallDenseMap<PHINode *, Type *> ResultTypes;
5441 SmallVector<PHINode *, 4> PHIs;
5442
5443 for (SwitchInst::CaseIt E = SI->case_end(); CI != E; ++CI) {
5444 ConstantInt *CaseVal = CI->getCaseValue();
5445 if (CaseVal->getValue().slt(MinCaseVal->getValue()))
5446 MinCaseVal = CaseVal;
5447 if (CaseVal->getValue().sgt(MaxCaseVal->getValue()))
5448 MaxCaseVal = CaseVal;
5449
5450 // Resulting value at phi nodes for this case value.
5451 using ResultsTy = SmallVector<std::pair<PHINode *, Constant *>, 4>;
5452 ResultsTy Results;
5453 if (!GetCaseResults(SI, CaseVal, CI->getCaseSuccessor(), &CommonDest,
5454 Results, DL, TTI))
5455 return false;
5456
5457 // Append the result from this case to the list for each phi.
5458 for (const auto &I : Results) {
5459 PHINode *PHI = I.first;
5460 Constant *Value = I.second;
5461 if (!ResultLists.count(PHI))
5462 PHIs.push_back(PHI);
5463 ResultLists[PHI].push_back(std::make_pair(CaseVal, Value));
5464 }
5465 }
5466
5467 // Keep track of the result types.
5468 for (PHINode *PHI : PHIs) {
5469 ResultTypes[PHI] = ResultLists[PHI][0].second->getType();
5470 }
5471
5472 uint64_t NumResults = ResultLists[PHIs[0]].size();
5473 APInt RangeSpread = MaxCaseVal->getValue() - MinCaseVal->getValue();
5474 uint64_t TableSize = RangeSpread.getLimitedValue() + 1;
5475 bool TableHasHoles = (NumResults < TableSize);
5476
5477 // If the table has holes, we need a constant result for the default case
5478 // or a bitmask that fits in a register.
5479 SmallVector<std::pair<PHINode *, Constant *>, 4> DefaultResultsList;
5480 bool HasDefaultResults =
5481 GetCaseResults(SI, nullptr, SI->getDefaultDest(), &CommonDest,
5482 DefaultResultsList, DL, TTI);
5483
5484 bool NeedMask = (TableHasHoles && !HasDefaultResults);
5485 if (NeedMask) {
5486 // As an extra penalty for the validity test we require more cases.
5487 if (SI->getNumCases() < 4) // FIXME: Find best threshold value (benchmark).
5488 return false;
5489 if (!DL.fitsInLegalInteger(TableSize))
5490 return false;
5491 }
5492
5493 for (const auto &I : DefaultResultsList) {
5494 PHINode *PHI = I.first;
5495 Constant *Result = I.second;
5496 DefaultResults[PHI] = Result;
5497 }
5498
5499 if (!ShouldBuildLookupTable(SI, TableSize, TTI, DL, ResultTypes))
5500 return false;
5501
5502 // Create the BB that does the lookups.
5503 Module &Mod = *CommonDest->getParent()->getParent();
5504 BasicBlock *LookupBB = BasicBlock::Create(
5505 Mod.getContext(), "switch.lookup", CommonDest->getParent(), CommonDest);
5506
5507 // Compute the table index value.
5508 Builder.SetInsertPoint(SI);
5509 Value *TableIndex;
5510 if (MinCaseVal->isNullValue())
5511 TableIndex = SI->getCondition();
5512 else
5513 TableIndex = Builder.CreateSub(SI->getCondition(), MinCaseVal,
5514 "switch.tableidx");
5515
5516 // Compute the maximum table size representable by the integer type we are
5517 // switching upon.
5518 unsigned CaseSize = MinCaseVal->getType()->getPrimitiveSizeInBits();
5519 uint64_t MaxTableSize = CaseSize > 63 ? UINT64_MAX : 1ULL << CaseSize;
5520 assert(MaxTableSize >= TableSize &&
5521 "It is impossible for a switch to have more entries than the max "
5522 "representable value of its input integer type's size.");
5523
5524 // If the default destination is unreachable, or if the lookup table covers
5525 // all values of the conditional variable, branch directly to the lookup table
5526 // BB. Otherwise, check that the condition is within the case range.
5527 const bool DefaultIsReachable =
5528 !isa<UnreachableInst>(SI->getDefaultDest()->getFirstNonPHIOrDbg());
5529 const bool GeneratingCoveredLookupTable = (MaxTableSize == TableSize);
5530 BranchInst *RangeCheckBranch = nullptr;
5531
5532 if (!DefaultIsReachable || GeneratingCoveredLookupTable) {
5533 Builder.CreateBr(LookupBB);
5534 // Note: We call removeProdecessor later since we need to be able to get the
5535 // PHI value for the default case in case we're using a bit mask.
5536 } else {
5537 Value *Cmp = Builder.CreateICmpULT(
5538 TableIndex, ConstantInt::get(MinCaseVal->getType(), TableSize));
5539 RangeCheckBranch =
5540 Builder.CreateCondBr(Cmp, LookupBB, SI->getDefaultDest());
5541 }
5542
5543 // Populate the BB that does the lookups.
5544 Builder.SetInsertPoint(LookupBB);
5545
5546 if (NeedMask) {
5547 // Before doing the lookup, we do the hole check. The LookupBB is therefore
5548 // re-purposed to do the hole check, and we create a new LookupBB.
5549 BasicBlock *MaskBB = LookupBB;
5550 MaskBB->setName("switch.hole_check");
5551 LookupBB = BasicBlock::Create(Mod.getContext(), "switch.lookup",
5552 CommonDest->getParent(), CommonDest);
5553
5554 // Make the mask's bitwidth at least 8-bit and a power-of-2 to avoid
5555 // unnecessary illegal types.
5556 uint64_t TableSizePowOf2 = NextPowerOf2(std::max(7ULL, TableSize - 1ULL));
5557 APInt MaskInt(TableSizePowOf2, 0);
5558 APInt One(TableSizePowOf2, 1);
5559 // Build bitmask; fill in a 1 bit for every case.
5560 const ResultListTy &ResultList = ResultLists[PHIs[0]];
5561 for (size_t I = 0, E = ResultList.size(); I != E; ++I) {
5562 uint64_t Idx = (ResultList[I].first->getValue() - MinCaseVal->getValue())
5563 .getLimitedValue();
5564 MaskInt |= One << Idx;
5565 }
5566 ConstantInt *TableMask = ConstantInt::get(Mod.getContext(), MaskInt);
5567
5568 // Get the TableIndex'th bit of the bitmask.
5569 // If this bit is 0 (meaning hole) jump to the default destination,
5570 // else continue with table lookup.
5571 IntegerType *MapTy = TableMask->getType();
5572 Value *MaskIndex =
5573 Builder.CreateZExtOrTrunc(TableIndex, MapTy, "switch.maskindex");
5574 Value *Shifted = Builder.CreateLShr(TableMask, MaskIndex, "switch.shifted");
5575 Value *LoBit = Builder.CreateTrunc(
5576 Shifted, Type::getInt1Ty(Mod.getContext()), "switch.lobit");
5577 Builder.CreateCondBr(LoBit, LookupBB, SI->getDefaultDest());
5578
5579 Builder.SetInsertPoint(LookupBB);
5580 AddPredecessorToBlock(SI->getDefaultDest(), MaskBB, SI->getParent());
5581 }
5582
5583 if (!DefaultIsReachable || GeneratingCoveredLookupTable) {
5584 // We cached PHINodes in PHIs. To avoid accessing deleted PHINodes later,
5585 // do not delete PHINodes here.
5586 SI->getDefaultDest()->removePredecessor(SI->getParent(),
5587 /*KeepOneInputPHIs=*/true);
5588 }
5589
5590 bool ReturnedEarly = false;
5591 for (PHINode *PHI : PHIs) {
5592 const ResultListTy &ResultList = ResultLists[PHI];
5593
5594 // If using a bitmask, use any value to fill the lookup table holes.
5595 Constant *DV = NeedMask ? ResultLists[PHI][0].second : DefaultResults[PHI];
5596 StringRef FuncName = Fn->getName();
5597 SwitchLookupTable Table(Mod, TableSize, MinCaseVal, ResultList, DV, DL,
5598 FuncName);
5599
5600 Value *Result = Table.BuildLookup(TableIndex, Builder);
5601
5602 // If the result is used to return immediately from the function, we want to
5603 // do that right here.
5604 if (PHI->hasOneUse() && isa<ReturnInst>(*PHI->user_begin()) &&
5605 PHI->user_back() == CommonDest->getFirstNonPHIOrDbg()) {
5606 Builder.CreateRet(Result);
5607 ReturnedEarly = true;
5608 break;
5609 }
5610
5611 // Do a small peephole optimization: re-use the switch table compare if
5612 // possible.
5613 if (!TableHasHoles && HasDefaultResults && RangeCheckBranch) {
5614 BasicBlock *PhiBlock = PHI->getParent();
5615 // Search for compare instructions which use the phi.
5616 for (auto *User : PHI->users()) {
5617 reuseTableCompare(User, PhiBlock, RangeCheckBranch, DV, ResultList);
5618 }
5619 }
5620
5621 PHI->addIncoming(Result, LookupBB);
5622 }
5623
5624 if (!ReturnedEarly)
5625 Builder.CreateBr(CommonDest);
5626
5627 // Remove the switch.
5628 for (unsigned i = 0, e = SI->getNumSuccessors(); i < e; ++i) {
5629 BasicBlock *Succ = SI->getSuccessor(i);
5630
5631 if (Succ == SI->getDefaultDest())
5632 continue;
5633 Succ->removePredecessor(SI->getParent());
5634 }
5635 SI->eraseFromParent();
5636
5637 ++NumLookupTables;
5638 if (NeedMask)
5639 ++NumLookupTablesHoles;
5640 return true;
5641 }
5642
isSwitchDense(ArrayRef<int64_t> Values)5643 static bool isSwitchDense(ArrayRef<int64_t> Values) {
5644 // See also SelectionDAGBuilder::isDense(), which this function was based on.
5645 uint64_t Diff = (uint64_t)Values.back() - (uint64_t)Values.front();
5646 uint64_t Range = Diff + 1;
5647 uint64_t NumCases = Values.size();
5648 // 40% is the default density for building a jump table in optsize/minsize mode.
5649 uint64_t MinDensity = 40;
5650
5651 return NumCases * 100 >= Range * MinDensity;
5652 }
5653
5654 /// Try to transform a switch that has "holes" in it to a contiguous sequence
5655 /// of cases.
5656 ///
5657 /// A switch such as: switch(i) {case 5: case 9: case 13: case 17:} can be
5658 /// range-reduced to: switch ((i-5) / 4) {case 0: case 1: case 2: case 3:}.
5659 ///
5660 /// This converts a sparse switch into a dense switch which allows better
5661 /// lowering and could also allow transforming into a lookup table.
ReduceSwitchRange(SwitchInst * SI,IRBuilder<> & Builder,const DataLayout & DL,const TargetTransformInfo & TTI)5662 static bool ReduceSwitchRange(SwitchInst *SI, IRBuilder<> &Builder,
5663 const DataLayout &DL,
5664 const TargetTransformInfo &TTI) {
5665 auto *CondTy = cast<IntegerType>(SI->getCondition()->getType());
5666 if (CondTy->getIntegerBitWidth() > 64 ||
5667 !DL.fitsInLegalInteger(CondTy->getIntegerBitWidth()))
5668 return false;
5669 // Only bother with this optimization if there are more than 3 switch cases;
5670 // SDAG will only bother creating jump tables for 4 or more cases.
5671 if (SI->getNumCases() < 4)
5672 return false;
5673
5674 // This transform is agnostic to the signedness of the input or case values. We
5675 // can treat the case values as signed or unsigned. We can optimize more common
5676 // cases such as a sequence crossing zero {-4,0,4,8} if we interpret case values
5677 // as signed.
5678 SmallVector<int64_t,4> Values;
5679 for (auto &C : SI->cases())
5680 Values.push_back(C.getCaseValue()->getValue().getSExtValue());
5681 llvm::sort(Values);
5682
5683 // If the switch is already dense, there's nothing useful to do here.
5684 if (isSwitchDense(Values))
5685 return false;
5686
5687 // First, transform the values such that they start at zero and ascend.
5688 int64_t Base = Values[0];
5689 for (auto &V : Values)
5690 V -= (uint64_t)(Base);
5691
5692 // Now we have signed numbers that have been shifted so that, given enough
5693 // precision, there are no negative values. Since the rest of the transform
5694 // is bitwise only, we switch now to an unsigned representation.
5695
5696 // This transform can be done speculatively because it is so cheap - it
5697 // results in a single rotate operation being inserted.
5698 // FIXME: It's possible that optimizing a switch on powers of two might also
5699 // be beneficial - flag values are often powers of two and we could use a CLZ
5700 // as the key function.
5701
5702 // countTrailingZeros(0) returns 64. As Values is guaranteed to have more than
5703 // one element and LLVM disallows duplicate cases, Shift is guaranteed to be
5704 // less than 64.
5705 unsigned Shift = 64;
5706 for (auto &V : Values)
5707 Shift = std::min(Shift, countTrailingZeros((uint64_t)V));
5708 assert(Shift < 64);
5709 if (Shift > 0)
5710 for (auto &V : Values)
5711 V = (int64_t)((uint64_t)V >> Shift);
5712
5713 if (!isSwitchDense(Values))
5714 // Transform didn't create a dense switch.
5715 return false;
5716
5717 // The obvious transform is to shift the switch condition right and emit a
5718 // check that the condition actually cleanly divided by GCD, i.e.
5719 // C & (1 << Shift - 1) == 0
5720 // inserting a new CFG edge to handle the case where it didn't divide cleanly.
5721 //
5722 // A cheaper way of doing this is a simple ROTR(C, Shift). This performs the
5723 // shift and puts the shifted-off bits in the uppermost bits. If any of these
5724 // are nonzero then the switch condition will be very large and will hit the
5725 // default case.
5726
5727 auto *Ty = cast<IntegerType>(SI->getCondition()->getType());
5728 Builder.SetInsertPoint(SI);
5729 auto *ShiftC = ConstantInt::get(Ty, Shift);
5730 auto *Sub = Builder.CreateSub(SI->getCondition(), ConstantInt::get(Ty, Base));
5731 auto *LShr = Builder.CreateLShr(Sub, ShiftC);
5732 auto *Shl = Builder.CreateShl(Sub, Ty->getBitWidth() - Shift);
5733 auto *Rot = Builder.CreateOr(LShr, Shl);
5734 SI->replaceUsesOfWith(SI->getCondition(), Rot);
5735
5736 for (auto Case : SI->cases()) {
5737 auto *Orig = Case.getCaseValue();
5738 auto Sub = Orig->getValue() - APInt(Ty->getBitWidth(), Base);
5739 Case.setValue(
5740 cast<ConstantInt>(ConstantInt::get(Ty, Sub.lshr(ShiftC->getValue()))));
5741 }
5742 return true;
5743 }
5744
simplifySwitch(SwitchInst * SI,IRBuilder<> & Builder)5745 bool SimplifyCFGOpt::simplifySwitch(SwitchInst *SI, IRBuilder<> &Builder) {
5746 BasicBlock *BB = SI->getParent();
5747
5748 if (isValueEqualityComparison(SI)) {
5749 // If we only have one predecessor, and if it is a branch on this value,
5750 // see if that predecessor totally determines the outcome of this switch.
5751 if (BasicBlock *OnlyPred = BB->getSinglePredecessor())
5752 if (SimplifyEqualityComparisonWithOnlyPredecessor(SI, OnlyPred, Builder))
5753 return requestResimplify();
5754
5755 Value *Cond = SI->getCondition();
5756 if (SelectInst *Select = dyn_cast<SelectInst>(Cond))
5757 if (SimplifySwitchOnSelect(SI, Select))
5758 return requestResimplify();
5759
5760 // If the block only contains the switch, see if we can fold the block
5761 // away into any preds.
5762 if (SI == &*BB->instructionsWithoutDebug().begin())
5763 if (FoldValueComparisonIntoPredecessors(SI, Builder))
5764 return requestResimplify();
5765 }
5766
5767 // Try to transform the switch into an icmp and a branch.
5768 if (TurnSwitchRangeIntoICmp(SI, Builder))
5769 return requestResimplify();
5770
5771 // Remove unreachable cases.
5772 if (eliminateDeadSwitchCases(SI, Options.AC, DL))
5773 return requestResimplify();
5774
5775 if (switchToSelect(SI, Builder, DL, TTI))
5776 return requestResimplify();
5777
5778 if (Options.ForwardSwitchCondToPhi && ForwardSwitchConditionToPHI(SI))
5779 return requestResimplify();
5780
5781 // The conversion from switch to lookup tables results in difficult-to-analyze
5782 // code and makes pruning branches much harder. This is a problem if the
5783 // switch expression itself can still be restricted as a result of inlining or
5784 // CVP. Therefore, only apply this transformation during late stages of the
5785 // optimisation pipeline.
5786 if (Options.ConvertSwitchToLookupTable &&
5787 SwitchToLookupTable(SI, Builder, DL, TTI))
5788 return requestResimplify();
5789
5790 if (ReduceSwitchRange(SI, Builder, DL, TTI))
5791 return requestResimplify();
5792
5793 return false;
5794 }
5795
simplifyIndirectBr(IndirectBrInst * IBI)5796 bool SimplifyCFGOpt::simplifyIndirectBr(IndirectBrInst *IBI) {
5797 BasicBlock *BB = IBI->getParent();
5798 bool Changed = false;
5799
5800 // Eliminate redundant destinations.
5801 SmallPtrSet<Value *, 8> Succs;
5802 for (unsigned i = 0, e = IBI->getNumDestinations(); i != e; ++i) {
5803 BasicBlock *Dest = IBI->getDestination(i);
5804 if (!Dest->hasAddressTaken() || !Succs.insert(Dest).second) {
5805 Dest->removePredecessor(BB);
5806 IBI->removeDestination(i);
5807 --i;
5808 --e;
5809 Changed = true;
5810 }
5811 }
5812
5813 if (IBI->getNumDestinations() == 0) {
5814 // If the indirectbr has no successors, change it to unreachable.
5815 new UnreachableInst(IBI->getContext(), IBI);
5816 EraseTerminatorAndDCECond(IBI);
5817 return true;
5818 }
5819
5820 if (IBI->getNumDestinations() == 1) {
5821 // If the indirectbr has one successor, change it to a direct branch.
5822 BranchInst::Create(IBI->getDestination(0), IBI);
5823 EraseTerminatorAndDCECond(IBI);
5824 return true;
5825 }
5826
5827 if (SelectInst *SI = dyn_cast<SelectInst>(IBI->getAddress())) {
5828 if (SimplifyIndirectBrOnSelect(IBI, SI))
5829 return requestResimplify();
5830 }
5831 return Changed;
5832 }
5833
5834 /// Given an block with only a single landing pad and a unconditional branch
5835 /// try to find another basic block which this one can be merged with. This
5836 /// handles cases where we have multiple invokes with unique landing pads, but
5837 /// a shared handler.
5838 ///
5839 /// We specifically choose to not worry about merging non-empty blocks
5840 /// here. That is a PRE/scheduling problem and is best solved elsewhere. In
5841 /// practice, the optimizer produces empty landing pad blocks quite frequently
5842 /// when dealing with exception dense code. (see: instcombine, gvn, if-else
5843 /// sinking in this file)
5844 ///
5845 /// This is primarily a code size optimization. We need to avoid performing
5846 /// any transform which might inhibit optimization (such as our ability to
5847 /// specialize a particular handler via tail commoning). We do this by not
5848 /// merging any blocks which require us to introduce a phi. Since the same
5849 /// values are flowing through both blocks, we don't lose any ability to
5850 /// specialize. If anything, we make such specialization more likely.
5851 ///
5852 /// TODO - This transformation could remove entries from a phi in the target
5853 /// block when the inputs in the phi are the same for the two blocks being
5854 /// merged. In some cases, this could result in removal of the PHI entirely.
TryToMergeLandingPad(LandingPadInst * LPad,BranchInst * BI,BasicBlock * BB)5855 static bool TryToMergeLandingPad(LandingPadInst *LPad, BranchInst *BI,
5856 BasicBlock *BB) {
5857 auto Succ = BB->getUniqueSuccessor();
5858 assert(Succ);
5859 // If there's a phi in the successor block, we'd likely have to introduce
5860 // a phi into the merged landing pad block.
5861 if (isa<PHINode>(*Succ->begin()))
5862 return false;
5863
5864 for (BasicBlock *OtherPred : predecessors(Succ)) {
5865 if (BB == OtherPred)
5866 continue;
5867 BasicBlock::iterator I = OtherPred->begin();
5868 LandingPadInst *LPad2 = dyn_cast<LandingPadInst>(I);
5869 if (!LPad2 || !LPad2->isIdenticalTo(LPad))
5870 continue;
5871 for (++I; isa<DbgInfoIntrinsic>(I); ++I)
5872 ;
5873 BranchInst *BI2 = dyn_cast<BranchInst>(I);
5874 if (!BI2 || !BI2->isIdenticalTo(BI))
5875 continue;
5876
5877 // We've found an identical block. Update our predecessors to take that
5878 // path instead and make ourselves dead.
5879 SmallPtrSet<BasicBlock *, 16> Preds;
5880 Preds.insert(pred_begin(BB), pred_end(BB));
5881 for (BasicBlock *Pred : Preds) {
5882 InvokeInst *II = cast<InvokeInst>(Pred->getTerminator());
5883 assert(II->getNormalDest() != BB && II->getUnwindDest() == BB &&
5884 "unexpected successor");
5885 II->setUnwindDest(OtherPred);
5886 }
5887
5888 // The debug info in OtherPred doesn't cover the merged control flow that
5889 // used to go through BB. We need to delete it or update it.
5890 for (auto I = OtherPred->begin(), E = OtherPred->end(); I != E;) {
5891 Instruction &Inst = *I;
5892 I++;
5893 if (isa<DbgInfoIntrinsic>(Inst))
5894 Inst.eraseFromParent();
5895 }
5896
5897 SmallPtrSet<BasicBlock *, 16> Succs;
5898 Succs.insert(succ_begin(BB), succ_end(BB));
5899 for (BasicBlock *Succ : Succs) {
5900 Succ->removePredecessor(BB);
5901 }
5902
5903 IRBuilder<> Builder(BI);
5904 Builder.CreateUnreachable();
5905 BI->eraseFromParent();
5906 return true;
5907 }
5908 return false;
5909 }
5910
simplifyBranch(BranchInst * Branch,IRBuilder<> & Builder)5911 bool SimplifyCFGOpt::simplifyBranch(BranchInst *Branch, IRBuilder<> &Builder) {
5912 return Branch->isUnconditional() ? simplifyUncondBranch(Branch, Builder)
5913 : simplifyCondBranch(Branch, Builder);
5914 }
5915
simplifyUncondBranch(BranchInst * BI,IRBuilder<> & Builder)5916 bool SimplifyCFGOpt::simplifyUncondBranch(BranchInst *BI,
5917 IRBuilder<> &Builder) {
5918 BasicBlock *BB = BI->getParent();
5919 BasicBlock *Succ = BI->getSuccessor(0);
5920
5921 // If the Terminator is the only non-phi instruction, simplify the block.
5922 // If LoopHeader is provided, check if the block or its successor is a loop
5923 // header. (This is for early invocations before loop simplify and
5924 // vectorization to keep canonical loop forms for nested loops. These blocks
5925 // can be eliminated when the pass is invoked later in the back-end.)
5926 // Note that if BB has only one predecessor then we do not introduce new
5927 // backedge, so we can eliminate BB.
5928 bool NeedCanonicalLoop =
5929 Options.NeedCanonicalLoop &&
5930 (LoopHeaders && BB->hasNPredecessorsOrMore(2) &&
5931 (LoopHeaders->count(BB) || LoopHeaders->count(Succ)));
5932 BasicBlock::iterator I = BB->getFirstNonPHIOrDbg()->getIterator();
5933 if (I->isTerminator() && BB != &BB->getParent()->getEntryBlock() &&
5934 !NeedCanonicalLoop && TryToSimplifyUncondBranchFromEmptyBlock(BB))
5935 return true;
5936
5937 // If the only instruction in the block is a seteq/setne comparison against a
5938 // constant, try to simplify the block.
5939 if (ICmpInst *ICI = dyn_cast<ICmpInst>(I))
5940 if (ICI->isEquality() && isa<ConstantInt>(ICI->getOperand(1))) {
5941 for (++I; isa<DbgInfoIntrinsic>(I); ++I)
5942 ;
5943 if (I->isTerminator() &&
5944 tryToSimplifyUncondBranchWithICmpInIt(ICI, Builder))
5945 return true;
5946 }
5947
5948 // See if we can merge an empty landing pad block with another which is
5949 // equivalent.
5950 if (LandingPadInst *LPad = dyn_cast<LandingPadInst>(I)) {
5951 for (++I; isa<DbgInfoIntrinsic>(I); ++I)
5952 ;
5953 if (I->isTerminator() && TryToMergeLandingPad(LPad, BI, BB))
5954 return true;
5955 }
5956
5957 // If this basic block is ONLY a compare and a branch, and if a predecessor
5958 // branches to us and our successor, fold the comparison into the
5959 // predecessor and use logical operations to update the incoming value
5960 // for PHI nodes in common successor.
5961 if (FoldBranchToCommonDest(BI, nullptr, Options.BonusInstThreshold))
5962 return requestResimplify();
5963 return false;
5964 }
5965
allPredecessorsComeFromSameSource(BasicBlock * BB)5966 static BasicBlock *allPredecessorsComeFromSameSource(BasicBlock *BB) {
5967 BasicBlock *PredPred = nullptr;
5968 for (auto *P : predecessors(BB)) {
5969 BasicBlock *PPred = P->getSinglePredecessor();
5970 if (!PPred || (PredPred && PredPred != PPred))
5971 return nullptr;
5972 PredPred = PPred;
5973 }
5974 return PredPred;
5975 }
5976
simplifyCondBranch(BranchInst * BI,IRBuilder<> & Builder)5977 bool SimplifyCFGOpt::simplifyCondBranch(BranchInst *BI, IRBuilder<> &Builder) {
5978 BasicBlock *BB = BI->getParent();
5979 if (!Options.SimplifyCondBranch)
5980 return false;
5981
5982 // Conditional branch
5983 if (isValueEqualityComparison(BI)) {
5984 // If we only have one predecessor, and if it is a branch on this value,
5985 // see if that predecessor totally determines the outcome of this
5986 // switch.
5987 if (BasicBlock *OnlyPred = BB->getSinglePredecessor())
5988 if (SimplifyEqualityComparisonWithOnlyPredecessor(BI, OnlyPred, Builder))
5989 return requestResimplify();
5990
5991 // This block must be empty, except for the setcond inst, if it exists.
5992 // Ignore dbg intrinsics.
5993 auto I = BB->instructionsWithoutDebug().begin();
5994 if (&*I == BI) {
5995 if (FoldValueComparisonIntoPredecessors(BI, Builder))
5996 return requestResimplify();
5997 } else if (&*I == cast<Instruction>(BI->getCondition())) {
5998 ++I;
5999 if (&*I == BI && FoldValueComparisonIntoPredecessors(BI, Builder))
6000 return requestResimplify();
6001 }
6002 }
6003
6004 // Try to turn "br (X == 0 | X == 1), T, F" into a switch instruction.
6005 if (SimplifyBranchOnICmpChain(BI, Builder, DL))
6006 return true;
6007
6008 // If this basic block has dominating predecessor blocks and the dominating
6009 // blocks' conditions imply BI's condition, we know the direction of BI.
6010 Optional<bool> Imp = isImpliedByDomCondition(BI->getCondition(), BI, DL);
6011 if (Imp) {
6012 // Turn this into a branch on constant.
6013 auto *OldCond = BI->getCondition();
6014 ConstantInt *TorF = *Imp ? ConstantInt::getTrue(BB->getContext())
6015 : ConstantInt::getFalse(BB->getContext());
6016 BI->setCondition(TorF);
6017 RecursivelyDeleteTriviallyDeadInstructions(OldCond);
6018 return requestResimplify();
6019 }
6020
6021 // If this basic block is ONLY a compare and a branch, and if a predecessor
6022 // branches to us and one of our successors, fold the comparison into the
6023 // predecessor and use logical operations to pick the right destination.
6024 if (FoldBranchToCommonDest(BI, nullptr, Options.BonusInstThreshold))
6025 return requestResimplify();
6026
6027 // We have a conditional branch to two blocks that are only reachable
6028 // from BI. We know that the condbr dominates the two blocks, so see if
6029 // there is any identical code in the "then" and "else" blocks. If so, we
6030 // can hoist it up to the branching block.
6031 if (BI->getSuccessor(0)->getSinglePredecessor()) {
6032 if (BI->getSuccessor(1)->getSinglePredecessor()) {
6033 if (HoistThenElseCodeToIf(BI, TTI))
6034 return requestResimplify();
6035 } else {
6036 // If Successor #1 has multiple preds, we may be able to conditionally
6037 // execute Successor #0 if it branches to Successor #1.
6038 Instruction *Succ0TI = BI->getSuccessor(0)->getTerminator();
6039 if (Succ0TI->getNumSuccessors() == 1 &&
6040 Succ0TI->getSuccessor(0) == BI->getSuccessor(1))
6041 if (SpeculativelyExecuteBB(BI, BI->getSuccessor(0), TTI))
6042 return requestResimplify();
6043 }
6044 } else if (BI->getSuccessor(1)->getSinglePredecessor()) {
6045 // If Successor #0 has multiple preds, we may be able to conditionally
6046 // execute Successor #1 if it branches to Successor #0.
6047 Instruction *Succ1TI = BI->getSuccessor(1)->getTerminator();
6048 if (Succ1TI->getNumSuccessors() == 1 &&
6049 Succ1TI->getSuccessor(0) == BI->getSuccessor(0))
6050 if (SpeculativelyExecuteBB(BI, BI->getSuccessor(1), TTI))
6051 return requestResimplify();
6052 }
6053
6054 // If this is a branch on a phi node in the current block, thread control
6055 // through this block if any PHI node entries are constants.
6056 if (PHINode *PN = dyn_cast<PHINode>(BI->getCondition()))
6057 if (PN->getParent() == BI->getParent())
6058 if (FoldCondBranchOnPHI(BI, DL, Options.AC))
6059 return requestResimplify();
6060
6061 // Scan predecessor blocks for conditional branches.
6062 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
6063 if (BranchInst *PBI = dyn_cast<BranchInst>((*PI)->getTerminator()))
6064 if (PBI != BI && PBI->isConditional())
6065 if (SimplifyCondBranchToCondBranch(PBI, BI, DL, TTI))
6066 return requestResimplify();
6067
6068 // Look for diamond patterns.
6069 if (MergeCondStores)
6070 if (BasicBlock *PrevBB = allPredecessorsComeFromSameSource(BB))
6071 if (BranchInst *PBI = dyn_cast<BranchInst>(PrevBB->getTerminator()))
6072 if (PBI != BI && PBI->isConditional())
6073 if (mergeConditionalStores(PBI, BI, DL, TTI))
6074 return requestResimplify();
6075
6076 return false;
6077 }
6078
6079 /// Check if passing a value to an instruction will cause undefined behavior.
passingValueIsAlwaysUndefined(Value * V,Instruction * I)6080 static bool passingValueIsAlwaysUndefined(Value *V, Instruction *I) {
6081 Constant *C = dyn_cast<Constant>(V);
6082 if (!C)
6083 return false;
6084
6085 if (I->use_empty())
6086 return false;
6087
6088 if (C->isNullValue() || isa<UndefValue>(C)) {
6089 // Only look at the first use, avoid hurting compile time with long uselists
6090 User *Use = *I->user_begin();
6091
6092 // Now make sure that there are no instructions in between that can alter
6093 // control flow (eg. calls)
6094 for (BasicBlock::iterator
6095 i = ++BasicBlock::iterator(I),
6096 UI = BasicBlock::iterator(dyn_cast<Instruction>(Use));
6097 i != UI; ++i)
6098 if (i == I->getParent()->end() || i->mayHaveSideEffects())
6099 return false;
6100
6101 // Look through GEPs. A load from a GEP derived from NULL is still undefined
6102 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Use))
6103 if (GEP->getPointerOperand() == I)
6104 return passingValueIsAlwaysUndefined(V, GEP);
6105
6106 // Look through bitcasts.
6107 if (BitCastInst *BC = dyn_cast<BitCastInst>(Use))
6108 return passingValueIsAlwaysUndefined(V, BC);
6109
6110 // Load from null is undefined.
6111 if (LoadInst *LI = dyn_cast<LoadInst>(Use))
6112 if (!LI->isVolatile())
6113 return !NullPointerIsDefined(LI->getFunction(),
6114 LI->getPointerAddressSpace());
6115
6116 // Store to null is undefined.
6117 if (StoreInst *SI = dyn_cast<StoreInst>(Use))
6118 if (!SI->isVolatile())
6119 return (!NullPointerIsDefined(SI->getFunction(),
6120 SI->getPointerAddressSpace())) &&
6121 SI->getPointerOperand() == I;
6122
6123 // A call to null is undefined.
6124 if (auto *CB = dyn_cast<CallBase>(Use))
6125 return !NullPointerIsDefined(CB->getFunction()) &&
6126 CB->getCalledOperand() == I;
6127 }
6128 return false;
6129 }
6130
6131 /// If BB has an incoming value that will always trigger undefined behavior
6132 /// (eg. null pointer dereference), remove the branch leading here.
removeUndefIntroducingPredecessor(BasicBlock * BB)6133 static bool removeUndefIntroducingPredecessor(BasicBlock *BB) {
6134 for (PHINode &PHI : BB->phis())
6135 for (unsigned i = 0, e = PHI.getNumIncomingValues(); i != e; ++i)
6136 if (passingValueIsAlwaysUndefined(PHI.getIncomingValue(i), &PHI)) {
6137 Instruction *T = PHI.getIncomingBlock(i)->getTerminator();
6138 IRBuilder<> Builder(T);
6139 if (BranchInst *BI = dyn_cast<BranchInst>(T)) {
6140 BB->removePredecessor(PHI.getIncomingBlock(i));
6141 // Turn uncoditional branches into unreachables and remove the dead
6142 // destination from conditional branches.
6143 if (BI->isUnconditional())
6144 Builder.CreateUnreachable();
6145 else
6146 Builder.CreateBr(BI->getSuccessor(0) == BB ? BI->getSuccessor(1)
6147 : BI->getSuccessor(0));
6148 BI->eraseFromParent();
6149 return true;
6150 }
6151 // TODO: SwitchInst.
6152 }
6153
6154 return false;
6155 }
6156
simplifyOnce(BasicBlock * BB)6157 bool SimplifyCFGOpt::simplifyOnce(BasicBlock *BB) {
6158 bool Changed = false;
6159
6160 assert(BB && BB->getParent() && "Block not embedded in function!");
6161 assert(BB->getTerminator() && "Degenerate basic block encountered!");
6162
6163 // Remove basic blocks that have no predecessors (except the entry block)...
6164 // or that just have themself as a predecessor. These are unreachable.
6165 if ((pred_empty(BB) && BB != &BB->getParent()->getEntryBlock()) ||
6166 BB->getSinglePredecessor() == BB) {
6167 LLVM_DEBUG(dbgs() << "Removing BB: \n" << *BB);
6168 DeleteDeadBlock(BB);
6169 return true;
6170 }
6171
6172 // Check to see if we can constant propagate this terminator instruction
6173 // away...
6174 Changed |= ConstantFoldTerminator(BB, true);
6175
6176 // Check for and eliminate duplicate PHI nodes in this block.
6177 Changed |= EliminateDuplicatePHINodes(BB);
6178
6179 // Check for and remove branches that will always cause undefined behavior.
6180 Changed |= removeUndefIntroducingPredecessor(BB);
6181
6182 // Merge basic blocks into their predecessor if there is only one distinct
6183 // pred, and if there is only one distinct successor of the predecessor, and
6184 // if there are no PHI nodes.
6185 if (MergeBlockIntoPredecessor(BB))
6186 return true;
6187
6188 if (SinkCommon && Options.SinkCommonInsts)
6189 Changed |= SinkCommonCodeFromPredecessors(BB);
6190
6191 IRBuilder<> Builder(BB);
6192
6193 if (Options.FoldTwoEntryPHINode) {
6194 // If there is a trivial two-entry PHI node in this basic block, and we can
6195 // eliminate it, do so now.
6196 if (auto *PN = dyn_cast<PHINode>(BB->begin()))
6197 if (PN->getNumIncomingValues() == 2)
6198 Changed |= FoldTwoEntryPHINode(PN, TTI, DL);
6199 }
6200
6201 Instruction *Terminator = BB->getTerminator();
6202 Builder.SetInsertPoint(Terminator);
6203 switch (Terminator->getOpcode()) {
6204 case Instruction::Br:
6205 Changed |= simplifyBranch(cast<BranchInst>(Terminator), Builder);
6206 break;
6207 case Instruction::Ret:
6208 Changed |= simplifyReturn(cast<ReturnInst>(Terminator), Builder);
6209 break;
6210 case Instruction::Resume:
6211 Changed |= simplifyResume(cast<ResumeInst>(Terminator), Builder);
6212 break;
6213 case Instruction::CleanupRet:
6214 Changed |= simplifyCleanupReturn(cast<CleanupReturnInst>(Terminator));
6215 break;
6216 case Instruction::Switch:
6217 Changed |= simplifySwitch(cast<SwitchInst>(Terminator), Builder);
6218 break;
6219 case Instruction::Unreachable:
6220 Changed |= simplifyUnreachable(cast<UnreachableInst>(Terminator));
6221 break;
6222 case Instruction::IndirectBr:
6223 Changed |= simplifyIndirectBr(cast<IndirectBrInst>(Terminator));
6224 break;
6225 }
6226
6227 return Changed;
6228 }
6229
run(BasicBlock * BB)6230 bool SimplifyCFGOpt::run(BasicBlock *BB) {
6231 bool Changed = false;
6232
6233 // Repeated simplify BB as long as resimplification is requested.
6234 do {
6235 Resimplify = false;
6236
6237 // Perform one round of simplifcation. Resimplify flag will be set if
6238 // another iteration is requested.
6239 Changed |= simplifyOnce(BB);
6240 } while (Resimplify);
6241
6242 return Changed;
6243 }
6244
simplifyCFG(BasicBlock * BB,const TargetTransformInfo & TTI,const SimplifyCFGOptions & Options,SmallPtrSetImpl<BasicBlock * > * LoopHeaders)6245 bool llvm::simplifyCFG(BasicBlock *BB, const TargetTransformInfo &TTI,
6246 const SimplifyCFGOptions &Options,
6247 SmallPtrSetImpl<BasicBlock *> *LoopHeaders) {
6248 return SimplifyCFGOpt(TTI, BB->getModule()->getDataLayout(), LoopHeaders,
6249 Options)
6250 .run(BB);
6251 }
6252