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/InstructionSimplify.h"
28 #include "llvm/Analysis/MemorySSA.h"
29 #include "llvm/Analysis/MemorySSAUpdater.h"
30 #include "llvm/Analysis/TargetTransformInfo.h"
31 #include "llvm/Analysis/ValueTracking.h"
32 #include "llvm/IR/Attributes.h"
33 #include "llvm/IR/BasicBlock.h"
34 #include "llvm/IR/CFG.h"
35 #include "llvm/IR/CallSite.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<bool> DupRet(
98 "simplifycfg-dup-ret", cl::Hidden, cl::init(false),
99 cl::desc("Duplicate return instructions into unconditional branches"));
100
101 static cl::opt<bool>
102 SinkCommon("simplifycfg-sink-common", cl::Hidden, cl::init(true),
103 cl::desc("Sink common instructions down to the end block"));
104
105 static cl::opt<bool> HoistCondStores(
106 "simplifycfg-hoist-cond-stores", cl::Hidden, cl::init(true),
107 cl::desc("Hoist conditional stores if an unconditional store precedes"));
108
109 static cl::opt<bool> MergeCondStores(
110 "simplifycfg-merge-cond-stores", cl::Hidden, cl::init(true),
111 cl::desc("Hoist conditional stores even if an unconditional store does not "
112 "precede - hoist multiple conditional stores into a single "
113 "predicated store"));
114
115 static cl::opt<bool> MergeCondStoresAggressively(
116 "simplifycfg-merge-cond-stores-aggressively", cl::Hidden, cl::init(false),
117 cl::desc("When merging conditional stores, do so even if the resultant "
118 "basic blocks are unlikely to be if-converted as a result"));
119
120 static cl::opt<bool> SpeculateOneExpensiveInst(
121 "speculate-one-expensive-inst", cl::Hidden, cl::init(true),
122 cl::desc("Allow exactly one expensive instruction to be speculatively "
123 "executed"));
124
125 static cl::opt<unsigned> MaxSpeculationDepth(
126 "max-speculation-depth", cl::Hidden, cl::init(10),
127 cl::desc("Limit maximum recursion depth when calculating costs of "
128 "speculatively executed instructions"));
129
130 STATISTIC(NumBitMaps, "Number of switch instructions turned into bitmaps");
131 STATISTIC(NumLinearMaps,
132 "Number of switch instructions turned into linear mapping");
133 STATISTIC(NumLookupTables,
134 "Number of switch instructions turned into lookup tables");
135 STATISTIC(
136 NumLookupTablesHoles,
137 "Number of switch instructions turned into lookup tables (holes checked)");
138 STATISTIC(NumTableCmpReuses, "Number of reused switch table lookup compares");
139 STATISTIC(NumSinkCommons,
140 "Number of common instructions sunk down to the end block");
141 STATISTIC(NumSpeculations, "Number of speculative executed instructions");
142
143 namespace {
144
145 // The first field contains the value that the switch produces when a certain
146 // case group is selected, and the second field is a vector containing the
147 // cases composing the case group.
148 using SwitchCaseResultVectorTy =
149 SmallVector<std::pair<Constant *, SmallVector<ConstantInt *, 4>>, 2>;
150
151 // The first field contains the phi node that generates a result of the switch
152 // and the second field contains the value generated for a certain case in the
153 // switch for that PHI.
154 using SwitchCaseResultsTy = SmallVector<std::pair<PHINode *, Constant *>, 4>;
155
156 /// ValueEqualityComparisonCase - Represents a case of a switch.
157 struct ValueEqualityComparisonCase {
158 ConstantInt *Value;
159 BasicBlock *Dest;
160
ValueEqualityComparisonCase__anon0ddf87310111::ValueEqualityComparisonCase161 ValueEqualityComparisonCase(ConstantInt *Value, BasicBlock *Dest)
162 : Value(Value), Dest(Dest) {}
163
operator <__anon0ddf87310111::ValueEqualityComparisonCase164 bool operator<(ValueEqualityComparisonCase RHS) const {
165 // Comparing pointers is ok as we only rely on the order for uniquing.
166 return Value < RHS.Value;
167 }
168
operator ==__anon0ddf87310111::ValueEqualityComparisonCase169 bool operator==(BasicBlock *RHSDest) const { return Dest == RHSDest; }
170 };
171
172 class SimplifyCFGOpt {
173 const TargetTransformInfo &TTI;
174 const DataLayout &DL;
175 SmallPtrSetImpl<BasicBlock *> *LoopHeaders;
176 const SimplifyCFGOptions &Options;
177 bool Resimplify;
178
179 Value *isValueEqualityComparison(Instruction *TI);
180 BasicBlock *GetValueEqualityComparisonCases(
181 Instruction *TI, std::vector<ValueEqualityComparisonCase> &Cases);
182 bool SimplifyEqualityComparisonWithOnlyPredecessor(Instruction *TI,
183 BasicBlock *Pred,
184 IRBuilder<> &Builder);
185 bool FoldValueComparisonIntoPredecessors(Instruction *TI,
186 IRBuilder<> &Builder);
187
188 bool SimplifyReturn(ReturnInst *RI, IRBuilder<> &Builder);
189 bool SimplifyResume(ResumeInst *RI, IRBuilder<> &Builder);
190 bool SimplifySingleResume(ResumeInst *RI);
191 bool SimplifyCommonResume(ResumeInst *RI);
192 bool SimplifyCleanupReturn(CleanupReturnInst *RI);
193 bool SimplifyUnreachable(UnreachableInst *UI);
194 bool SimplifySwitch(SwitchInst *SI, IRBuilder<> &Builder);
195 bool SimplifyIndirectBr(IndirectBrInst *IBI);
196 bool SimplifyUncondBranch(BranchInst *BI, IRBuilder<> &Builder);
197 bool SimplifyCondBranch(BranchInst *BI, IRBuilder<> &Builder);
198
199 bool tryToSimplifyUncondBranchWithICmpInIt(ICmpInst *ICI,
200 IRBuilder<> &Builder);
201
202 public:
SimplifyCFGOpt(const TargetTransformInfo & TTI,const DataLayout & DL,SmallPtrSetImpl<BasicBlock * > * LoopHeaders,const SimplifyCFGOptions & Opts)203 SimplifyCFGOpt(const TargetTransformInfo &TTI, const DataLayout &DL,
204 SmallPtrSetImpl<BasicBlock *> *LoopHeaders,
205 const SimplifyCFGOptions &Opts)
206 : TTI(TTI), DL(DL), LoopHeaders(LoopHeaders), Options(Opts) {}
207
208 bool run(BasicBlock *BB);
209 bool simplifyOnce(BasicBlock *BB);
210
211 // Helper to set Resimplify and return change indication.
requestResimplify()212 bool requestResimplify() {
213 Resimplify = true;
214 return true;
215 }
216 };
217
218 } // end anonymous namespace
219
220 /// Return true if it is safe to merge these two
221 /// terminator instructions together.
222 static bool
SafeToMergeTerminators(Instruction * SI1,Instruction * SI2,SmallSetVector<BasicBlock *,4> * FailBlocks=nullptr)223 SafeToMergeTerminators(Instruction *SI1, Instruction *SI2,
224 SmallSetVector<BasicBlock *, 4> *FailBlocks = nullptr) {
225 if (SI1 == SI2)
226 return false; // Can't merge with self!
227
228 // It is not safe to merge these two switch instructions if they have a common
229 // successor, and if that successor has a PHI node, and if *that* PHI node has
230 // conflicting incoming values from the two switch blocks.
231 BasicBlock *SI1BB = SI1->getParent();
232 BasicBlock *SI2BB = SI2->getParent();
233
234 SmallPtrSet<BasicBlock *, 16> SI1Succs(succ_begin(SI1BB), succ_end(SI1BB));
235 bool Fail = false;
236 for (BasicBlock *Succ : successors(SI2BB))
237 if (SI1Succs.count(Succ))
238 for (BasicBlock::iterator BBI = Succ->begin(); isa<PHINode>(BBI); ++BBI) {
239 PHINode *PN = cast<PHINode>(BBI);
240 if (PN->getIncomingValueForBlock(SI1BB) !=
241 PN->getIncomingValueForBlock(SI2BB)) {
242 if (FailBlocks)
243 FailBlocks->insert(Succ);
244 Fail = true;
245 }
246 }
247
248 return !Fail;
249 }
250
251 /// Return true if it is safe and profitable to merge these two terminator
252 /// instructions together, where SI1 is an unconditional branch. PhiNodes will
253 /// store all PHI nodes in common successors.
254 static bool
isProfitableToFoldUnconditional(BranchInst * SI1,BranchInst * SI2,Instruction * Cond,SmallVectorImpl<PHINode * > & PhiNodes)255 isProfitableToFoldUnconditional(BranchInst *SI1, BranchInst *SI2,
256 Instruction *Cond,
257 SmallVectorImpl<PHINode *> &PhiNodes) {
258 if (SI1 == SI2)
259 return false; // Can't merge with self!
260 assert(SI1->isUnconditional() && SI2->isConditional());
261
262 // We fold the unconditional branch if we can easily update all PHI nodes in
263 // common successors:
264 // 1> We have a constant incoming value for the conditional branch;
265 // 2> We have "Cond" as the incoming value for the unconditional branch;
266 // 3> SI2->getCondition() and Cond have same operands.
267 CmpInst *Ci2 = dyn_cast<CmpInst>(SI2->getCondition());
268 if (!Ci2)
269 return false;
270 if (!(Cond->getOperand(0) == Ci2->getOperand(0) &&
271 Cond->getOperand(1) == Ci2->getOperand(1)) &&
272 !(Cond->getOperand(0) == Ci2->getOperand(1) &&
273 Cond->getOperand(1) == Ci2->getOperand(0)))
274 return false;
275
276 BasicBlock *SI1BB = SI1->getParent();
277 BasicBlock *SI2BB = SI2->getParent();
278 SmallPtrSet<BasicBlock *, 16> SI1Succs(succ_begin(SI1BB), succ_end(SI1BB));
279 for (BasicBlock *Succ : successors(SI2BB))
280 if (SI1Succs.count(Succ))
281 for (BasicBlock::iterator BBI = Succ->begin(); isa<PHINode>(BBI); ++BBI) {
282 PHINode *PN = cast<PHINode>(BBI);
283 if (PN->getIncomingValueForBlock(SI1BB) != Cond ||
284 !isa<ConstantInt>(PN->getIncomingValueForBlock(SI2BB)))
285 return false;
286 PhiNodes.push_back(PN);
287 }
288 return true;
289 }
290
291 /// Update PHI nodes in Succ to indicate that there will now be entries in it
292 /// from the 'NewPred' block. The values that will be flowing into the PHI nodes
293 /// will be the same as those coming in from ExistPred, an existing predecessor
294 /// of Succ.
AddPredecessorToBlock(BasicBlock * Succ,BasicBlock * NewPred,BasicBlock * ExistPred,MemorySSAUpdater * MSSAU=nullptr)295 static void AddPredecessorToBlock(BasicBlock *Succ, BasicBlock *NewPred,
296 BasicBlock *ExistPred,
297 MemorySSAUpdater *MSSAU = nullptr) {
298 for (PHINode &PN : Succ->phis())
299 PN.addIncoming(PN.getIncomingValueForBlock(ExistPred), NewPred);
300 if (MSSAU)
301 if (auto *MPhi = MSSAU->getMemorySSA()->getMemoryAccess(Succ))
302 MPhi->addIncoming(MPhi->getIncomingValueForBlock(ExistPred), NewPred);
303 }
304
305 /// Compute an abstract "cost" of speculating the given instruction,
306 /// which is assumed to be safe to speculate. TCC_Free means cheap,
307 /// TCC_Basic means less cheap, and TCC_Expensive means prohibitively
308 /// expensive.
ComputeSpeculationCost(const User * I,const TargetTransformInfo & TTI)309 static unsigned ComputeSpeculationCost(const User *I,
310 const TargetTransformInfo &TTI) {
311 assert(isSafeToSpeculativelyExecute(I) &&
312 "Instruction is not safe to speculatively execute!");
313 return TTI.getUserCost(I);
314 }
315
316 /// If we have a merge point of an "if condition" as accepted above,
317 /// return true if the specified value dominates the block. We
318 /// don't handle the true generality of domination here, just a special case
319 /// which works well enough for us.
320 ///
321 /// If AggressiveInsts is non-null, and if V does not dominate BB, we check to
322 /// see if V (which must be an instruction) and its recursive operands
323 /// that do not dominate BB have a combined cost lower than CostRemaining and
324 /// are non-trapping. If both are true, the instruction is inserted into the
325 /// set and true is returned.
326 ///
327 /// The cost for most non-trapping instructions is defined as 1 except for
328 /// Select whose cost is 2.
329 ///
330 /// After this function returns, CostRemaining is decreased by the cost of
331 /// V plus its non-dominating operands. If that cost is greater than
332 /// CostRemaining, false is returned and CostRemaining is undefined.
DominatesMergePoint(Value * V,BasicBlock * BB,SmallPtrSetImpl<Instruction * > & AggressiveInsts,unsigned & CostRemaining,const TargetTransformInfo & TTI,unsigned Depth=0)333 static bool DominatesMergePoint(Value *V, BasicBlock *BB,
334 SmallPtrSetImpl<Instruction *> &AggressiveInsts,
335 unsigned &CostRemaining,
336 const TargetTransformInfo &TTI,
337 unsigned Depth = 0) {
338 // It is possible to hit a zero-cost cycle (phi/gep instructions for example),
339 // so limit the recursion depth.
340 // TODO: While this recursion limit does prevent pathological behavior, it
341 // would be better to track visited instructions to avoid cycles.
342 if (Depth == MaxSpeculationDepth)
343 return false;
344
345 Instruction *I = dyn_cast<Instruction>(V);
346 if (!I) {
347 // Non-instructions all dominate instructions, but not all constantexprs
348 // can be executed unconditionally.
349 if (ConstantExpr *C = dyn_cast<ConstantExpr>(V))
350 if (C->canTrap())
351 return false;
352 return true;
353 }
354 BasicBlock *PBB = I->getParent();
355
356 // We don't want to allow weird loops that might have the "if condition" in
357 // the bottom of this block.
358 if (PBB == BB)
359 return false;
360
361 // If this instruction is defined in a block that contains an unconditional
362 // branch to BB, then it must be in the 'conditional' part of the "if
363 // statement". If not, it definitely dominates the region.
364 BranchInst *BI = dyn_cast<BranchInst>(PBB->getTerminator());
365 if (!BI || BI->isConditional() || BI->getSuccessor(0) != BB)
366 return true;
367
368 // If we have seen this instruction before, don't count it again.
369 if (AggressiveInsts.count(I))
370 return true;
371
372 // Okay, it looks like the instruction IS in the "condition". Check to
373 // see if it's a cheap instruction to unconditionally compute, and if it
374 // only uses stuff defined outside of the condition. If so, hoist it out.
375 if (!isSafeToSpeculativelyExecute(I))
376 return false;
377
378 unsigned Cost = ComputeSpeculationCost(I, TTI);
379
380 // Allow exactly one instruction to be speculated regardless of its cost
381 // (as long as it is safe to do so).
382 // This is intended to flatten the CFG even if the instruction is a division
383 // or other expensive operation. The speculation of an expensive instruction
384 // is expected to be undone in CodeGenPrepare if the speculation has not
385 // enabled further IR optimizations.
386 if (Cost > CostRemaining &&
387 (!SpeculateOneExpensiveInst || !AggressiveInsts.empty() || Depth > 0))
388 return false;
389
390 // Avoid unsigned wrap.
391 CostRemaining = (Cost > CostRemaining) ? 0 : CostRemaining - Cost;
392
393 // Okay, we can only really hoist these out if their operands do
394 // not take us over the cost threshold.
395 for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i)
396 if (!DominatesMergePoint(*i, BB, AggressiveInsts, CostRemaining, TTI,
397 Depth + 1))
398 return false;
399 // Okay, it's safe to do this! Remember this instruction.
400 AggressiveInsts.insert(I);
401 return true;
402 }
403
404 /// Extract ConstantInt from value, looking through IntToPtr
405 /// and PointerNullValue. Return NULL if value is not a constant int.
GetConstantInt(Value * V,const DataLayout & DL)406 static ConstantInt *GetConstantInt(Value *V, const DataLayout &DL) {
407 // Normal constant int.
408 ConstantInt *CI = dyn_cast<ConstantInt>(V);
409 if (CI || !isa<Constant>(V) || !V->getType()->isPointerTy())
410 return CI;
411
412 // This is some kind of pointer constant. Turn it into a pointer-sized
413 // ConstantInt if possible.
414 IntegerType *PtrTy = cast<IntegerType>(DL.getIntPtrType(V->getType()));
415
416 // Null pointer means 0, see SelectionDAGBuilder::getValue(const Value*).
417 if (isa<ConstantPointerNull>(V))
418 return ConstantInt::get(PtrTy, 0);
419
420 // IntToPtr const int.
421 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
422 if (CE->getOpcode() == Instruction::IntToPtr)
423 if (ConstantInt *CI = dyn_cast<ConstantInt>(CE->getOperand(0))) {
424 // The constant is very likely to have the right type already.
425 if (CI->getType() == PtrTy)
426 return CI;
427 else
428 return cast<ConstantInt>(
429 ConstantExpr::getIntegerCast(CI, PtrTy, /*isSigned=*/false));
430 }
431 return nullptr;
432 }
433
434 namespace {
435
436 /// Given a chain of or (||) or and (&&) comparison of a value against a
437 /// constant, this will try to recover the information required for a switch
438 /// structure.
439 /// It will depth-first traverse the chain of comparison, seeking for patterns
440 /// like %a == 12 or %a < 4 and combine them to produce a set of integer
441 /// representing the different cases for the switch.
442 /// Note that if the chain is composed of '||' it will build the set of elements
443 /// that matches the comparisons (i.e. any of this value validate the chain)
444 /// while for a chain of '&&' it will build the set elements that make the test
445 /// fail.
446 struct ConstantComparesGatherer {
447 const DataLayout &DL;
448
449 /// Value found for the switch comparison
450 Value *CompValue = nullptr;
451
452 /// Extra clause to be checked before the switch
453 Value *Extra = nullptr;
454
455 /// Set of integers to match in switch
456 SmallVector<ConstantInt *, 8> Vals;
457
458 /// Number of comparisons matched in the and/or chain
459 unsigned UsedICmps = 0;
460
461 /// Construct and compute the result for the comparison instruction Cond
ConstantComparesGatherer__anon0ddf87310211::ConstantComparesGatherer462 ConstantComparesGatherer(Instruction *Cond, const DataLayout &DL) : DL(DL) {
463 gather(Cond);
464 }
465
466 ConstantComparesGatherer(const ConstantComparesGatherer &) = delete;
467 ConstantComparesGatherer &
468 operator=(const ConstantComparesGatherer &) = delete;
469
470 private:
471 /// Try to set the current value used for the comparison, it succeeds only if
472 /// it wasn't set before or if the new value is the same as the old one
setValueOnce__anon0ddf87310211::ConstantComparesGatherer473 bool setValueOnce(Value *NewVal) {
474 if (CompValue && CompValue != NewVal)
475 return false;
476 CompValue = NewVal;
477 return (CompValue != nullptr);
478 }
479
480 /// Try to match Instruction "I" as a comparison against a constant and
481 /// populates the array Vals with the set of values that match (or do not
482 /// match depending on isEQ).
483 /// Return false on failure. On success, the Value the comparison matched
484 /// against is placed in CompValue.
485 /// If CompValue is already set, the function is expected to fail if a match
486 /// is found but the value compared to is different.
matchInstruction__anon0ddf87310211::ConstantComparesGatherer487 bool matchInstruction(Instruction *I, bool isEQ) {
488 // If this is an icmp against a constant, handle this as one of the cases.
489 ICmpInst *ICI;
490 ConstantInt *C;
491 if (!((ICI = dyn_cast<ICmpInst>(I)) &&
492 (C = GetConstantInt(I->getOperand(1), DL)))) {
493 return false;
494 }
495
496 Value *RHSVal;
497 const APInt *RHSC;
498
499 // Pattern match a special case
500 // (x & ~2^z) == y --> x == y || x == y|2^z
501 // This undoes a transformation done by instcombine to fuse 2 compares.
502 if (ICI->getPredicate() == (isEQ ? ICmpInst::ICMP_EQ : ICmpInst::ICMP_NE)) {
503 // It's a little bit hard to see why the following transformations are
504 // correct. Here is a CVC3 program to verify them for 64-bit values:
505
506 /*
507 ONE : BITVECTOR(64) = BVZEROEXTEND(0bin1, 63);
508 x : BITVECTOR(64);
509 y : BITVECTOR(64);
510 z : BITVECTOR(64);
511 mask : BITVECTOR(64) = BVSHL(ONE, z);
512 QUERY( (y & ~mask = y) =>
513 ((x & ~mask = y) <=> (x = y OR x = (y | mask)))
514 );
515 QUERY( (y | mask = y) =>
516 ((x | mask = y) <=> (x = y OR x = (y & ~mask)))
517 );
518 */
519
520 // Please note that each pattern must be a dual implication (<--> or
521 // iff). One directional implication can create spurious matches. If the
522 // implication is only one-way, an unsatisfiable condition on the left
523 // side can imply a satisfiable condition on the right side. Dual
524 // implication ensures that satisfiable conditions are transformed to
525 // other satisfiable conditions and unsatisfiable conditions are
526 // transformed to other unsatisfiable conditions.
527
528 // Here is a concrete example of a unsatisfiable condition on the left
529 // implying a satisfiable condition on the right:
530 //
531 // mask = (1 << z)
532 // (x & ~mask) == y --> (x == y || x == (y | mask))
533 //
534 // Substituting y = 3, z = 0 yields:
535 // (x & -2) == 3 --> (x == 3 || x == 2)
536
537 // Pattern match a special case:
538 /*
539 QUERY( (y & ~mask = y) =>
540 ((x & ~mask = y) <=> (x = y OR x = (y | mask)))
541 );
542 */
543 if (match(ICI->getOperand(0),
544 m_And(m_Value(RHSVal), m_APInt(RHSC)))) {
545 APInt Mask = ~*RHSC;
546 if (Mask.isPowerOf2() && (C->getValue() & ~Mask) == C->getValue()) {
547 // If we already have a value for the switch, it has to match!
548 if (!setValueOnce(RHSVal))
549 return false;
550
551 Vals.push_back(C);
552 Vals.push_back(
553 ConstantInt::get(C->getContext(),
554 C->getValue() | Mask));
555 UsedICmps++;
556 return true;
557 }
558 }
559
560 // Pattern match a special case:
561 /*
562 QUERY( (y | mask = y) =>
563 ((x | mask = y) <=> (x = y OR x = (y & ~mask)))
564 );
565 */
566 if (match(ICI->getOperand(0),
567 m_Or(m_Value(RHSVal), m_APInt(RHSC)))) {
568 APInt Mask = *RHSC;
569 if (Mask.isPowerOf2() && (C->getValue() | Mask) == C->getValue()) {
570 // If we already have a value for the switch, it has to match!
571 if (!setValueOnce(RHSVal))
572 return false;
573
574 Vals.push_back(C);
575 Vals.push_back(ConstantInt::get(C->getContext(),
576 C->getValue() & ~Mask));
577 UsedICmps++;
578 return true;
579 }
580 }
581
582 // If we already have a value for the switch, it has to match!
583 if (!setValueOnce(ICI->getOperand(0)))
584 return false;
585
586 UsedICmps++;
587 Vals.push_back(C);
588 return ICI->getOperand(0);
589 }
590
591 // If we have "x ult 3", for example, then we can add 0,1,2 to the set.
592 ConstantRange Span = ConstantRange::makeAllowedICmpRegion(
593 ICI->getPredicate(), C->getValue());
594
595 // Shift the range if the compare is fed by an add. This is the range
596 // compare idiom as emitted by instcombine.
597 Value *CandidateVal = I->getOperand(0);
598 if (match(I->getOperand(0), m_Add(m_Value(RHSVal), m_APInt(RHSC)))) {
599 Span = Span.subtract(*RHSC);
600 CandidateVal = RHSVal;
601 }
602
603 // If this is an and/!= check, then we are looking to build the set of
604 // value that *don't* pass the and chain. I.e. to turn "x ugt 2" into
605 // x != 0 && x != 1.
606 if (!isEQ)
607 Span = Span.inverse();
608
609 // If there are a ton of values, we don't want to make a ginormous switch.
610 if (Span.isSizeLargerThan(8) || Span.isEmptySet()) {
611 return false;
612 }
613
614 // If we already have a value for the switch, it has to match!
615 if (!setValueOnce(CandidateVal))
616 return false;
617
618 // Add all values from the range to the set
619 for (APInt Tmp = Span.getLower(); Tmp != Span.getUpper(); ++Tmp)
620 Vals.push_back(ConstantInt::get(I->getContext(), Tmp));
621
622 UsedICmps++;
623 return true;
624 }
625
626 /// Given a potentially 'or'd or 'and'd together collection of icmp
627 /// eq/ne/lt/gt instructions that compare a value against a constant, extract
628 /// the value being compared, and stick the list constants into the Vals
629 /// vector.
630 /// One "Extra" case is allowed to differ from the other.
gather__anon0ddf87310211::ConstantComparesGatherer631 void gather(Value *V) {
632 Instruction *I = dyn_cast<Instruction>(V);
633 bool isEQ = (I->getOpcode() == Instruction::Or);
634
635 // Keep a stack (SmallVector for efficiency) for depth-first traversal
636 SmallVector<Value *, 8> DFT;
637 SmallPtrSet<Value *, 8> Visited;
638
639 // Initialize
640 Visited.insert(V);
641 DFT.push_back(V);
642
643 while (!DFT.empty()) {
644 V = DFT.pop_back_val();
645
646 if (Instruction *I = dyn_cast<Instruction>(V)) {
647 // If it is a || (or && depending on isEQ), process the operands.
648 if (I->getOpcode() == (isEQ ? Instruction::Or : Instruction::And)) {
649 if (Visited.insert(I->getOperand(1)).second)
650 DFT.push_back(I->getOperand(1));
651 if (Visited.insert(I->getOperand(0)).second)
652 DFT.push_back(I->getOperand(0));
653 continue;
654 }
655
656 // Try to match the current instruction
657 if (matchInstruction(I, isEQ))
658 // Match succeed, continue the loop
659 continue;
660 }
661
662 // One element of the sequence of || (or &&) could not be match as a
663 // comparison against the same value as the others.
664 // We allow only one "Extra" case to be checked before the switch
665 if (!Extra) {
666 Extra = V;
667 continue;
668 }
669 // Failed to parse a proper sequence, abort now
670 CompValue = nullptr;
671 break;
672 }
673 }
674 };
675
676 } // end anonymous namespace
677
EraseTerminatorAndDCECond(Instruction * TI,MemorySSAUpdater * MSSAU=nullptr)678 static void EraseTerminatorAndDCECond(Instruction *TI,
679 MemorySSAUpdater *MSSAU = nullptr) {
680 Instruction *Cond = nullptr;
681 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
682 Cond = dyn_cast<Instruction>(SI->getCondition());
683 } else if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
684 if (BI->isConditional())
685 Cond = dyn_cast<Instruction>(BI->getCondition());
686 } else if (IndirectBrInst *IBI = dyn_cast<IndirectBrInst>(TI)) {
687 Cond = dyn_cast<Instruction>(IBI->getAddress());
688 }
689
690 TI->eraseFromParent();
691 if (Cond)
692 RecursivelyDeleteTriviallyDeadInstructions(Cond, nullptr, MSSAU);
693 }
694
695 /// Return true if the specified terminator checks
696 /// to see if a value is equal to constant integer value.
isValueEqualityComparison(Instruction * TI)697 Value *SimplifyCFGOpt::isValueEqualityComparison(Instruction *TI) {
698 Value *CV = nullptr;
699 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
700 // Do not permit merging of large switch instructions into their
701 // predecessors unless there is only one predecessor.
702 if (!SI->getParent()->hasNPredecessorsOrMore(128 / SI->getNumSuccessors()))
703 CV = SI->getCondition();
704 } else if (BranchInst *BI = dyn_cast<BranchInst>(TI))
705 if (BI->isConditional() && BI->getCondition()->hasOneUse())
706 if (ICmpInst *ICI = dyn_cast<ICmpInst>(BI->getCondition())) {
707 if (ICI->isEquality() && GetConstantInt(ICI->getOperand(1), DL))
708 CV = ICI->getOperand(0);
709 }
710
711 // Unwrap any lossless ptrtoint cast.
712 if (CV) {
713 if (PtrToIntInst *PTII = dyn_cast<PtrToIntInst>(CV)) {
714 Value *Ptr = PTII->getPointerOperand();
715 if (PTII->getType() == DL.getIntPtrType(Ptr->getType()))
716 CV = Ptr;
717 }
718 }
719 return CV;
720 }
721
722 /// Given a value comparison instruction,
723 /// decode all of the 'cases' that it represents and return the 'default' block.
GetValueEqualityComparisonCases(Instruction * TI,std::vector<ValueEqualityComparisonCase> & Cases)724 BasicBlock *SimplifyCFGOpt::GetValueEqualityComparisonCases(
725 Instruction *TI, std::vector<ValueEqualityComparisonCase> &Cases) {
726 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
727 Cases.reserve(SI->getNumCases());
728 for (auto Case : SI->cases())
729 Cases.push_back(ValueEqualityComparisonCase(Case.getCaseValue(),
730 Case.getCaseSuccessor()));
731 return SI->getDefaultDest();
732 }
733
734 BranchInst *BI = cast<BranchInst>(TI);
735 ICmpInst *ICI = cast<ICmpInst>(BI->getCondition());
736 BasicBlock *Succ = BI->getSuccessor(ICI->getPredicate() == ICmpInst::ICMP_NE);
737 Cases.push_back(ValueEqualityComparisonCase(
738 GetConstantInt(ICI->getOperand(1), DL), Succ));
739 return BI->getSuccessor(ICI->getPredicate() == ICmpInst::ICMP_EQ);
740 }
741
742 /// Given a vector of bb/value pairs, remove any entries
743 /// in the list that match the specified block.
744 static void
EliminateBlockCases(BasicBlock * BB,std::vector<ValueEqualityComparisonCase> & Cases)745 EliminateBlockCases(BasicBlock *BB,
746 std::vector<ValueEqualityComparisonCase> &Cases) {
747 Cases.erase(std::remove(Cases.begin(), Cases.end(), BB), Cases.end());
748 }
749
750 /// Return true if there are any keys in C1 that exist in C2 as well.
ValuesOverlap(std::vector<ValueEqualityComparisonCase> & C1,std::vector<ValueEqualityComparisonCase> & C2)751 static bool ValuesOverlap(std::vector<ValueEqualityComparisonCase> &C1,
752 std::vector<ValueEqualityComparisonCase> &C2) {
753 std::vector<ValueEqualityComparisonCase> *V1 = &C1, *V2 = &C2;
754
755 // Make V1 be smaller than V2.
756 if (V1->size() > V2->size())
757 std::swap(V1, V2);
758
759 if (V1->empty())
760 return false;
761 if (V1->size() == 1) {
762 // Just scan V2.
763 ConstantInt *TheVal = (*V1)[0].Value;
764 for (unsigned i = 0, e = V2->size(); i != e; ++i)
765 if (TheVal == (*V2)[i].Value)
766 return true;
767 }
768
769 // Otherwise, just sort both lists and compare element by element.
770 array_pod_sort(V1->begin(), V1->end());
771 array_pod_sort(V2->begin(), V2->end());
772 unsigned i1 = 0, i2 = 0, e1 = V1->size(), e2 = V2->size();
773 while (i1 != e1 && i2 != e2) {
774 if ((*V1)[i1].Value == (*V2)[i2].Value)
775 return true;
776 if ((*V1)[i1].Value < (*V2)[i2].Value)
777 ++i1;
778 else
779 ++i2;
780 }
781 return false;
782 }
783
784 // Set branch weights on SwitchInst. This sets the metadata if there is at
785 // least one non-zero weight.
setBranchWeights(SwitchInst * SI,ArrayRef<uint32_t> Weights)786 static void setBranchWeights(SwitchInst *SI, ArrayRef<uint32_t> Weights) {
787 // Check that there is at least one non-zero weight. Otherwise, pass
788 // nullptr to setMetadata which will erase the existing metadata.
789 MDNode *N = nullptr;
790 if (llvm::any_of(Weights, [](uint32_t W) { return W != 0; }))
791 N = MDBuilder(SI->getParent()->getContext()).createBranchWeights(Weights);
792 SI->setMetadata(LLVMContext::MD_prof, N);
793 }
794
795 // Similar to the above, but for branch and select instructions that take
796 // exactly 2 weights.
setBranchWeights(Instruction * I,uint32_t TrueWeight,uint32_t FalseWeight)797 static void setBranchWeights(Instruction *I, uint32_t TrueWeight,
798 uint32_t FalseWeight) {
799 assert(isa<BranchInst>(I) || isa<SelectInst>(I));
800 // Check that there is at least one non-zero weight. Otherwise, pass
801 // nullptr to setMetadata which will erase the existing metadata.
802 MDNode *N = nullptr;
803 if (TrueWeight || FalseWeight)
804 N = MDBuilder(I->getParent()->getContext())
805 .createBranchWeights(TrueWeight, FalseWeight);
806 I->setMetadata(LLVMContext::MD_prof, N);
807 }
808
809 /// If TI is known to be a terminator instruction and its block is known to
810 /// only have a single predecessor block, check to see if that predecessor is
811 /// also a value comparison with the same value, and if that comparison
812 /// determines the outcome of this comparison. If so, simplify TI. This does a
813 /// very limited form of jump threading.
SimplifyEqualityComparisonWithOnlyPredecessor(Instruction * TI,BasicBlock * Pred,IRBuilder<> & Builder)814 bool SimplifyCFGOpt::SimplifyEqualityComparisonWithOnlyPredecessor(
815 Instruction *TI, BasicBlock *Pred, IRBuilder<> &Builder) {
816 Value *PredVal = isValueEqualityComparison(Pred->getTerminator());
817 if (!PredVal)
818 return false; // Not a value comparison in predecessor.
819
820 Value *ThisVal = isValueEqualityComparison(TI);
821 assert(ThisVal && "This isn't a value comparison!!");
822 if (ThisVal != PredVal)
823 return false; // Different predicates.
824
825 // TODO: Preserve branch weight metadata, similarly to how
826 // FoldValueComparisonIntoPredecessors preserves it.
827
828 // Find out information about when control will move from Pred to TI's block.
829 std::vector<ValueEqualityComparisonCase> PredCases;
830 BasicBlock *PredDef =
831 GetValueEqualityComparisonCases(Pred->getTerminator(), PredCases);
832 EliminateBlockCases(PredDef, PredCases); // Remove default from cases.
833
834 // Find information about how control leaves this block.
835 std::vector<ValueEqualityComparisonCase> ThisCases;
836 BasicBlock *ThisDef = GetValueEqualityComparisonCases(TI, ThisCases);
837 EliminateBlockCases(ThisDef, ThisCases); // Remove default from cases.
838
839 // If TI's block is the default block from Pred's comparison, potentially
840 // simplify TI based on this knowledge.
841 if (PredDef == TI->getParent()) {
842 // If we are here, we know that the value is none of those cases listed in
843 // PredCases. If there are any cases in ThisCases that are in PredCases, we
844 // can simplify TI.
845 if (!ValuesOverlap(PredCases, ThisCases))
846 return false;
847
848 if (isa<BranchInst>(TI)) {
849 // Okay, one of the successors of this condbr is dead. Convert it to a
850 // uncond br.
851 assert(ThisCases.size() == 1 && "Branch can only have one case!");
852 // Insert the new branch.
853 Instruction *NI = Builder.CreateBr(ThisDef);
854 (void)NI;
855
856 // Remove PHI node entries for the dead edge.
857 ThisCases[0].Dest->removePredecessor(TI->getParent());
858
859 LLVM_DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator()
860 << "Through successor TI: " << *TI << "Leaving: " << *NI
861 << "\n");
862
863 EraseTerminatorAndDCECond(TI);
864 return true;
865 }
866
867 SwitchInstProfUpdateWrapper SI = *cast<SwitchInst>(TI);
868 // Okay, TI has cases that are statically dead, prune them away.
869 SmallPtrSet<Constant *, 16> DeadCases;
870 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
871 DeadCases.insert(PredCases[i].Value);
872
873 LLVM_DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator()
874 << "Through successor TI: " << *TI);
875
876 for (SwitchInst::CaseIt i = SI->case_end(), e = SI->case_begin(); i != e;) {
877 --i;
878 if (DeadCases.count(i->getCaseValue())) {
879 i->getCaseSuccessor()->removePredecessor(TI->getParent());
880 SI.removeCase(i);
881 }
882 }
883 LLVM_DEBUG(dbgs() << "Leaving: " << *TI << "\n");
884 return true;
885 }
886
887 // Otherwise, TI's block must correspond to some matched value. Find out
888 // which value (or set of values) this is.
889 ConstantInt *TIV = nullptr;
890 BasicBlock *TIBB = TI->getParent();
891 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
892 if (PredCases[i].Dest == TIBB) {
893 if (TIV)
894 return false; // Cannot handle multiple values coming to this block.
895 TIV = PredCases[i].Value;
896 }
897 assert(TIV && "No edge from pred to succ?");
898
899 // Okay, we found the one constant that our value can be if we get into TI's
900 // BB. Find out which successor will unconditionally be branched to.
901 BasicBlock *TheRealDest = nullptr;
902 for (unsigned i = 0, e = ThisCases.size(); i != e; ++i)
903 if (ThisCases[i].Value == TIV) {
904 TheRealDest = ThisCases[i].Dest;
905 break;
906 }
907
908 // If not handled by any explicit cases, it is handled by the default case.
909 if (!TheRealDest)
910 TheRealDest = ThisDef;
911
912 // Remove PHI node entries for dead edges.
913 BasicBlock *CheckEdge = TheRealDest;
914 for (BasicBlock *Succ : successors(TIBB))
915 if (Succ != CheckEdge)
916 Succ->removePredecessor(TIBB);
917 else
918 CheckEdge = nullptr;
919
920 // Insert the new branch.
921 Instruction *NI = Builder.CreateBr(TheRealDest);
922 (void)NI;
923
924 LLVM_DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator()
925 << "Through successor TI: " << *TI << "Leaving: " << *NI
926 << "\n");
927
928 EraseTerminatorAndDCECond(TI);
929 return true;
930 }
931
932 namespace {
933
934 /// This class implements a stable ordering of constant
935 /// integers that does not depend on their address. This is important for
936 /// applications that sort ConstantInt's to ensure uniqueness.
937 struct ConstantIntOrdering {
operator ()__anon0ddf87310411::ConstantIntOrdering938 bool operator()(const ConstantInt *LHS, const ConstantInt *RHS) const {
939 return LHS->getValue().ult(RHS->getValue());
940 }
941 };
942
943 } // end anonymous namespace
944
ConstantIntSortPredicate(ConstantInt * const * P1,ConstantInt * const * P2)945 static int ConstantIntSortPredicate(ConstantInt *const *P1,
946 ConstantInt *const *P2) {
947 const ConstantInt *LHS = *P1;
948 const ConstantInt *RHS = *P2;
949 if (LHS == RHS)
950 return 0;
951 return LHS->getValue().ult(RHS->getValue()) ? 1 : -1;
952 }
953
HasBranchWeights(const Instruction * I)954 static inline bool HasBranchWeights(const Instruction *I) {
955 MDNode *ProfMD = I->getMetadata(LLVMContext::MD_prof);
956 if (ProfMD && ProfMD->getOperand(0))
957 if (MDString *MDS = dyn_cast<MDString>(ProfMD->getOperand(0)))
958 return MDS->getString().equals("branch_weights");
959
960 return false;
961 }
962
963 /// Get Weights of a given terminator, the default weight is at the front
964 /// of the vector. If TI is a conditional eq, we need to swap the branch-weight
965 /// metadata.
GetBranchWeights(Instruction * TI,SmallVectorImpl<uint64_t> & Weights)966 static void GetBranchWeights(Instruction *TI,
967 SmallVectorImpl<uint64_t> &Weights) {
968 MDNode *MD = TI->getMetadata(LLVMContext::MD_prof);
969 assert(MD);
970 for (unsigned i = 1, e = MD->getNumOperands(); i < e; ++i) {
971 ConstantInt *CI = mdconst::extract<ConstantInt>(MD->getOperand(i));
972 Weights.push_back(CI->getValue().getZExtValue());
973 }
974
975 // If TI is a conditional eq, the default case is the false case,
976 // and the corresponding branch-weight data is at index 2. We swap the
977 // default weight to be the first entry.
978 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
979 assert(Weights.size() == 2);
980 ICmpInst *ICI = cast<ICmpInst>(BI->getCondition());
981 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
982 std::swap(Weights.front(), Weights.back());
983 }
984 }
985
986 /// Keep halving the weights until all can fit in uint32_t.
FitWeights(MutableArrayRef<uint64_t> Weights)987 static void FitWeights(MutableArrayRef<uint64_t> Weights) {
988 uint64_t Max = *std::max_element(Weights.begin(), Weights.end());
989 if (Max > UINT_MAX) {
990 unsigned Offset = 32 - countLeadingZeros(Max);
991 for (uint64_t &I : Weights)
992 I >>= Offset;
993 }
994 }
995
996 /// The specified terminator is a value equality comparison instruction
997 /// (either a switch or a branch on "X == c").
998 /// See if any of the predecessors of the terminator block are value comparisons
999 /// on the same value. If so, and if safe to do so, fold them together.
FoldValueComparisonIntoPredecessors(Instruction * TI,IRBuilder<> & Builder)1000 bool SimplifyCFGOpt::FoldValueComparisonIntoPredecessors(Instruction *TI,
1001 IRBuilder<> &Builder) {
1002 BasicBlock *BB = TI->getParent();
1003 Value *CV = isValueEqualityComparison(TI); // CondVal
1004 assert(CV && "Not a comparison?");
1005 bool Changed = false;
1006
1007 SmallVector<BasicBlock *, 16> Preds(pred_begin(BB), pred_end(BB));
1008 while (!Preds.empty()) {
1009 BasicBlock *Pred = Preds.pop_back_val();
1010
1011 // See if the predecessor is a comparison with the same value.
1012 Instruction *PTI = Pred->getTerminator();
1013 Value *PCV = isValueEqualityComparison(PTI); // PredCondVal
1014
1015 if (PCV == CV && TI != PTI) {
1016 SmallSetVector<BasicBlock*, 4> FailBlocks;
1017 if (!SafeToMergeTerminators(TI, PTI, &FailBlocks)) {
1018 for (auto *Succ : FailBlocks) {
1019 if (!SplitBlockPredecessors(Succ, TI->getParent(), ".fold.split"))
1020 return false;
1021 }
1022 }
1023
1024 // Figure out which 'cases' to copy from SI to PSI.
1025 std::vector<ValueEqualityComparisonCase> BBCases;
1026 BasicBlock *BBDefault = GetValueEqualityComparisonCases(TI, BBCases);
1027
1028 std::vector<ValueEqualityComparisonCase> PredCases;
1029 BasicBlock *PredDefault = GetValueEqualityComparisonCases(PTI, PredCases);
1030
1031 // Based on whether the default edge from PTI goes to BB or not, fill in
1032 // PredCases and PredDefault with the new switch cases we would like to
1033 // build.
1034 SmallVector<BasicBlock *, 8> NewSuccessors;
1035
1036 // Update the branch weight metadata along the way
1037 SmallVector<uint64_t, 8> Weights;
1038 bool PredHasWeights = HasBranchWeights(PTI);
1039 bool SuccHasWeights = HasBranchWeights(TI);
1040
1041 if (PredHasWeights) {
1042 GetBranchWeights(PTI, Weights);
1043 // branch-weight metadata is inconsistent here.
1044 if (Weights.size() != 1 + PredCases.size())
1045 PredHasWeights = SuccHasWeights = false;
1046 } else if (SuccHasWeights)
1047 // If there are no predecessor weights but there are successor weights,
1048 // populate Weights with 1, which will later be scaled to the sum of
1049 // successor's weights
1050 Weights.assign(1 + PredCases.size(), 1);
1051
1052 SmallVector<uint64_t, 8> SuccWeights;
1053 if (SuccHasWeights) {
1054 GetBranchWeights(TI, SuccWeights);
1055 // branch-weight metadata is inconsistent here.
1056 if (SuccWeights.size() != 1 + BBCases.size())
1057 PredHasWeights = SuccHasWeights = false;
1058 } else if (PredHasWeights)
1059 SuccWeights.assign(1 + BBCases.size(), 1);
1060
1061 if (PredDefault == BB) {
1062 // If this is the default destination from PTI, only the edges in TI
1063 // that don't occur in PTI, or that branch to BB will be activated.
1064 std::set<ConstantInt *, ConstantIntOrdering> PTIHandled;
1065 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
1066 if (PredCases[i].Dest != BB)
1067 PTIHandled.insert(PredCases[i].Value);
1068 else {
1069 // The default destination is BB, we don't need explicit targets.
1070 std::swap(PredCases[i], PredCases.back());
1071
1072 if (PredHasWeights || SuccHasWeights) {
1073 // Increase weight for the default case.
1074 Weights[0] += Weights[i + 1];
1075 std::swap(Weights[i + 1], Weights.back());
1076 Weights.pop_back();
1077 }
1078
1079 PredCases.pop_back();
1080 --i;
1081 --e;
1082 }
1083
1084 // Reconstruct the new switch statement we will be building.
1085 if (PredDefault != BBDefault) {
1086 PredDefault->removePredecessor(Pred);
1087 PredDefault = BBDefault;
1088 NewSuccessors.push_back(BBDefault);
1089 }
1090
1091 unsigned CasesFromPred = Weights.size();
1092 uint64_t ValidTotalSuccWeight = 0;
1093 for (unsigned i = 0, e = BBCases.size(); i != e; ++i)
1094 if (!PTIHandled.count(BBCases[i].Value) &&
1095 BBCases[i].Dest != BBDefault) {
1096 PredCases.push_back(BBCases[i]);
1097 NewSuccessors.push_back(BBCases[i].Dest);
1098 if (SuccHasWeights || PredHasWeights) {
1099 // The default weight is at index 0, so weight for the ith case
1100 // should be at index i+1. Scale the cases from successor by
1101 // PredDefaultWeight (Weights[0]).
1102 Weights.push_back(Weights[0] * SuccWeights[i + 1]);
1103 ValidTotalSuccWeight += SuccWeights[i + 1];
1104 }
1105 }
1106
1107 if (SuccHasWeights || PredHasWeights) {
1108 ValidTotalSuccWeight += SuccWeights[0];
1109 // Scale the cases from predecessor by ValidTotalSuccWeight.
1110 for (unsigned i = 1; i < CasesFromPred; ++i)
1111 Weights[i] *= ValidTotalSuccWeight;
1112 // Scale the default weight by SuccDefaultWeight (SuccWeights[0]).
1113 Weights[0] *= SuccWeights[0];
1114 }
1115 } else {
1116 // If this is not the default destination from PSI, only the edges
1117 // in SI that occur in PSI with a destination of BB will be
1118 // activated.
1119 std::set<ConstantInt *, ConstantIntOrdering> PTIHandled;
1120 std::map<ConstantInt *, uint64_t> WeightsForHandled;
1121 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
1122 if (PredCases[i].Dest == BB) {
1123 PTIHandled.insert(PredCases[i].Value);
1124
1125 if (PredHasWeights || SuccHasWeights) {
1126 WeightsForHandled[PredCases[i].Value] = Weights[i + 1];
1127 std::swap(Weights[i + 1], Weights.back());
1128 Weights.pop_back();
1129 }
1130
1131 std::swap(PredCases[i], PredCases.back());
1132 PredCases.pop_back();
1133 --i;
1134 --e;
1135 }
1136
1137 // Okay, now we know which constants were sent to BB from the
1138 // predecessor. Figure out where they will all go now.
1139 for (unsigned i = 0, e = BBCases.size(); i != e; ++i)
1140 if (PTIHandled.count(BBCases[i].Value)) {
1141 // If this is one we are capable of getting...
1142 if (PredHasWeights || SuccHasWeights)
1143 Weights.push_back(WeightsForHandled[BBCases[i].Value]);
1144 PredCases.push_back(BBCases[i]);
1145 NewSuccessors.push_back(BBCases[i].Dest);
1146 PTIHandled.erase(
1147 BBCases[i].Value); // This constant is taken care of
1148 }
1149
1150 // If there are any constants vectored to BB that TI doesn't handle,
1151 // they must go to the default destination of TI.
1152 for (ConstantInt *I : PTIHandled) {
1153 if (PredHasWeights || SuccHasWeights)
1154 Weights.push_back(WeightsForHandled[I]);
1155 PredCases.push_back(ValueEqualityComparisonCase(I, BBDefault));
1156 NewSuccessors.push_back(BBDefault);
1157 }
1158 }
1159
1160 // Okay, at this point, we know which new successor Pred will get. Make
1161 // sure we update the number of entries in the PHI nodes for these
1162 // successors.
1163 for (BasicBlock *NewSuccessor : NewSuccessors)
1164 AddPredecessorToBlock(NewSuccessor, Pred, BB);
1165
1166 Builder.SetInsertPoint(PTI);
1167 // Convert pointer to int before we switch.
1168 if (CV->getType()->isPointerTy()) {
1169 CV = Builder.CreatePtrToInt(CV, DL.getIntPtrType(CV->getType()),
1170 "magicptr");
1171 }
1172
1173 // Now that the successors are updated, create the new Switch instruction.
1174 SwitchInst *NewSI =
1175 Builder.CreateSwitch(CV, PredDefault, PredCases.size());
1176 NewSI->setDebugLoc(PTI->getDebugLoc());
1177 for (ValueEqualityComparisonCase &V : PredCases)
1178 NewSI->addCase(V.Value, V.Dest);
1179
1180 if (PredHasWeights || SuccHasWeights) {
1181 // Halve the weights if any of them cannot fit in an uint32_t
1182 FitWeights(Weights);
1183
1184 SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
1185
1186 setBranchWeights(NewSI, MDWeights);
1187 }
1188
1189 EraseTerminatorAndDCECond(PTI);
1190
1191 // Okay, last check. If BB is still a successor of PSI, then we must
1192 // have an infinite loop case. If so, add an infinitely looping block
1193 // to handle the case to preserve the behavior of the code.
1194 BasicBlock *InfLoopBlock = nullptr;
1195 for (unsigned i = 0, e = NewSI->getNumSuccessors(); i != e; ++i)
1196 if (NewSI->getSuccessor(i) == BB) {
1197 if (!InfLoopBlock) {
1198 // Insert it at the end of the function, because it's either code,
1199 // or it won't matter if it's hot. :)
1200 InfLoopBlock = BasicBlock::Create(BB->getContext(), "infloop",
1201 BB->getParent());
1202 BranchInst::Create(InfLoopBlock, InfLoopBlock);
1203 }
1204 NewSI->setSuccessor(i, InfLoopBlock);
1205 }
1206
1207 Changed = true;
1208 }
1209 }
1210 return Changed;
1211 }
1212
1213 // If we would need to insert a select that uses the value of this invoke
1214 // (comments in HoistThenElseCodeToIf explain why we would need to do this), we
1215 // 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)1216 static bool isSafeToHoistInvoke(BasicBlock *BB1, BasicBlock *BB2,
1217 Instruction *I1, Instruction *I2) {
1218 for (BasicBlock *Succ : successors(BB1)) {
1219 for (const PHINode &PN : Succ->phis()) {
1220 Value *BB1V = PN.getIncomingValueForBlock(BB1);
1221 Value *BB2V = PN.getIncomingValueForBlock(BB2);
1222 if (BB1V != BB2V && (BB1V == I1 || BB2V == I2)) {
1223 return false;
1224 }
1225 }
1226 }
1227 return true;
1228 }
1229
1230 static bool passingValueIsAlwaysUndefined(Value *V, Instruction *I);
1231
1232 /// Given a conditional branch that goes to BB1 and BB2, hoist any common code
1233 /// in the two blocks up into the branch block. The caller of this function
1234 /// guarantees that BI's block dominates BB1 and BB2.
HoistThenElseCodeToIf(BranchInst * BI,const TargetTransformInfo & TTI)1235 static bool HoistThenElseCodeToIf(BranchInst *BI,
1236 const TargetTransformInfo &TTI) {
1237 // This does very trivial matching, with limited scanning, to find identical
1238 // instructions in the two blocks. In particular, we don't want to get into
1239 // O(M*N) situations here where M and N are the sizes of BB1 and BB2. As
1240 // such, we currently just scan for obviously identical instructions in an
1241 // identical order.
1242 BasicBlock *BB1 = BI->getSuccessor(0); // The true destination.
1243 BasicBlock *BB2 = BI->getSuccessor(1); // The false destination
1244
1245 BasicBlock::iterator BB1_Itr = BB1->begin();
1246 BasicBlock::iterator BB2_Itr = BB2->begin();
1247
1248 Instruction *I1 = &*BB1_Itr++, *I2 = &*BB2_Itr++;
1249 // Skip debug info if it is not identical.
1250 DbgInfoIntrinsic *DBI1 = dyn_cast<DbgInfoIntrinsic>(I1);
1251 DbgInfoIntrinsic *DBI2 = dyn_cast<DbgInfoIntrinsic>(I2);
1252 if (!DBI1 || !DBI2 || !DBI1->isIdenticalToWhenDefined(DBI2)) {
1253 while (isa<DbgInfoIntrinsic>(I1))
1254 I1 = &*BB1_Itr++;
1255 while (isa<DbgInfoIntrinsic>(I2))
1256 I2 = &*BB2_Itr++;
1257 }
1258 // FIXME: Can we define a safety predicate for CallBr?
1259 if (isa<PHINode>(I1) || !I1->isIdenticalToWhenDefined(I2) ||
1260 (isa<InvokeInst>(I1) && !isSafeToHoistInvoke(BB1, BB2, I1, I2)) ||
1261 isa<CallBrInst>(I1))
1262 return false;
1263
1264 BasicBlock *BIParent = BI->getParent();
1265
1266 bool Changed = false;
1267 do {
1268 // If we are hoisting the terminator instruction, don't move one (making a
1269 // broken BB), instead clone it, and remove BI.
1270 if (I1->isTerminator())
1271 goto HoistTerminator;
1272
1273 // If we're going to hoist a call, make sure that the two instructions we're
1274 // commoning/hoisting are both marked with musttail, or neither of them is
1275 // marked as such. Otherwise, we might end up in a situation where we hoist
1276 // from a block where the terminator is a `ret` to a block where the terminator
1277 // is a `br`, and `musttail` calls expect to be followed by a return.
1278 auto *C1 = dyn_cast<CallInst>(I1);
1279 auto *C2 = dyn_cast<CallInst>(I2);
1280 if (C1 && C2)
1281 if (C1->isMustTailCall() != C2->isMustTailCall())
1282 return Changed;
1283
1284 if (!TTI.isProfitableToHoist(I1) || !TTI.isProfitableToHoist(I2))
1285 return Changed;
1286
1287 if (isa<DbgInfoIntrinsic>(I1) || isa<DbgInfoIntrinsic>(I2)) {
1288 assert (isa<DbgInfoIntrinsic>(I1) && isa<DbgInfoIntrinsic>(I2));
1289 // The debug location is an integral part of a debug info intrinsic
1290 // and can't be separated from it or replaced. Instead of attempting
1291 // to merge locations, simply hoist both copies of the intrinsic.
1292 BIParent->getInstList().splice(BI->getIterator(),
1293 BB1->getInstList(), I1);
1294 BIParent->getInstList().splice(BI->getIterator(),
1295 BB2->getInstList(), I2);
1296 Changed = true;
1297 } else {
1298 // For a normal instruction, we just move one to right before the branch,
1299 // then replace all uses of the other with the first. Finally, we remove
1300 // the now redundant second instruction.
1301 BIParent->getInstList().splice(BI->getIterator(),
1302 BB1->getInstList(), I1);
1303 if (!I2->use_empty())
1304 I2->replaceAllUsesWith(I1);
1305 I1->andIRFlags(I2);
1306 unsigned KnownIDs[] = {LLVMContext::MD_tbaa,
1307 LLVMContext::MD_range,
1308 LLVMContext::MD_fpmath,
1309 LLVMContext::MD_invariant_load,
1310 LLVMContext::MD_nonnull,
1311 LLVMContext::MD_invariant_group,
1312 LLVMContext::MD_align,
1313 LLVMContext::MD_dereferenceable,
1314 LLVMContext::MD_dereferenceable_or_null,
1315 LLVMContext::MD_mem_parallel_loop_access,
1316 LLVMContext::MD_access_group};
1317 combineMetadata(I1, I2, KnownIDs, true);
1318
1319 // I1 and I2 are being combined into a single instruction. Its debug
1320 // location is the merged locations of the original instructions.
1321 I1->applyMergedLocation(I1->getDebugLoc(), I2->getDebugLoc());
1322
1323 I2->eraseFromParent();
1324 Changed = true;
1325 }
1326
1327 I1 = &*BB1_Itr++;
1328 I2 = &*BB2_Itr++;
1329 // Skip debug info if it is not identical.
1330 DbgInfoIntrinsic *DBI1 = dyn_cast<DbgInfoIntrinsic>(I1);
1331 DbgInfoIntrinsic *DBI2 = dyn_cast<DbgInfoIntrinsic>(I2);
1332 if (!DBI1 || !DBI2 || !DBI1->isIdenticalToWhenDefined(DBI2)) {
1333 while (isa<DbgInfoIntrinsic>(I1))
1334 I1 = &*BB1_Itr++;
1335 while (isa<DbgInfoIntrinsic>(I2))
1336 I2 = &*BB2_Itr++;
1337 }
1338 } while (I1->isIdenticalToWhenDefined(I2));
1339
1340 return true;
1341
1342 HoistTerminator:
1343 // It may not be possible to hoist an invoke.
1344 // FIXME: Can we define a safety predicate for CallBr?
1345 if (isa<InvokeInst>(I1) && !isSafeToHoistInvoke(BB1, BB2, I1, I2))
1346 return Changed;
1347
1348 // TODO: callbr hoisting currently disabled pending further study.
1349 if (isa<CallBrInst>(I1))
1350 return Changed;
1351
1352 for (BasicBlock *Succ : successors(BB1)) {
1353 for (PHINode &PN : Succ->phis()) {
1354 Value *BB1V = PN.getIncomingValueForBlock(BB1);
1355 Value *BB2V = PN.getIncomingValueForBlock(BB2);
1356 if (BB1V == BB2V)
1357 continue;
1358
1359 // Check for passingValueIsAlwaysUndefined here because we would rather
1360 // eliminate undefined control flow then converting it to a select.
1361 if (passingValueIsAlwaysUndefined(BB1V, &PN) ||
1362 passingValueIsAlwaysUndefined(BB2V, &PN))
1363 return Changed;
1364
1365 if (isa<ConstantExpr>(BB1V) && !isSafeToSpeculativelyExecute(BB1V))
1366 return Changed;
1367 if (isa<ConstantExpr>(BB2V) && !isSafeToSpeculativelyExecute(BB2V))
1368 return Changed;
1369 }
1370 }
1371
1372 // Okay, it is safe to hoist the terminator.
1373 Instruction *NT = I1->clone();
1374 BIParent->getInstList().insert(BI->getIterator(), NT);
1375 if (!NT->getType()->isVoidTy()) {
1376 I1->replaceAllUsesWith(NT);
1377 I2->replaceAllUsesWith(NT);
1378 NT->takeName(I1);
1379 }
1380
1381 // Ensure terminator gets a debug location, even an unknown one, in case
1382 // it involves inlinable calls.
1383 NT->applyMergedLocation(I1->getDebugLoc(), I2->getDebugLoc());
1384
1385 // PHIs created below will adopt NT's merged DebugLoc.
1386 IRBuilder<NoFolder> Builder(NT);
1387
1388 // Hoisting one of the terminators from our successor is a great thing.
1389 // Unfortunately, the successors of the if/else blocks may have PHI nodes in
1390 // them. If they do, all PHI entries for BB1/BB2 must agree for all PHI
1391 // nodes, so we insert select instruction to compute the final result.
1392 std::map<std::pair<Value *, Value *>, SelectInst *> InsertedSelects;
1393 for (BasicBlock *Succ : successors(BB1)) {
1394 for (PHINode &PN : Succ->phis()) {
1395 Value *BB1V = PN.getIncomingValueForBlock(BB1);
1396 Value *BB2V = PN.getIncomingValueForBlock(BB2);
1397 if (BB1V == BB2V)
1398 continue;
1399
1400 // These values do not agree. Insert a select instruction before NT
1401 // that determines the right value.
1402 SelectInst *&SI = InsertedSelects[std::make_pair(BB1V, BB2V)];
1403 if (!SI)
1404 SI = cast<SelectInst>(
1405 Builder.CreateSelect(BI->getCondition(), BB1V, BB2V,
1406 BB1V->getName() + "." + BB2V->getName(), BI));
1407
1408 // Make the PHI node use the select for all incoming values for BB1/BB2
1409 for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i)
1410 if (PN.getIncomingBlock(i) == BB1 || PN.getIncomingBlock(i) == BB2)
1411 PN.setIncomingValue(i, SI);
1412 }
1413 }
1414
1415 // Update any PHI nodes in our new successors.
1416 for (BasicBlock *Succ : successors(BB1))
1417 AddPredecessorToBlock(Succ, BIParent, BB1);
1418
1419 EraseTerminatorAndDCECond(BI);
1420 return true;
1421 }
1422
1423 // All instructions in Insts belong to different blocks that all unconditionally
1424 // branch to a common successor. Analyze each instruction and return true if it
1425 // would be possible to sink them into their successor, creating one common
1426 // instruction instead. For every value that would be required to be provided by
1427 // PHI node (because an operand varies in each input block), add to PHIOperands.
canSinkInstructions(ArrayRef<Instruction * > Insts,DenseMap<Instruction *,SmallVector<Value *,4>> & PHIOperands)1428 static bool canSinkInstructions(
1429 ArrayRef<Instruction *> Insts,
1430 DenseMap<Instruction *, SmallVector<Value *, 4>> &PHIOperands) {
1431 // Prune out obviously bad instructions to move. Any non-store instruction
1432 // must have exactly one use, and we check later that use is by a single,
1433 // common PHI instruction in the successor.
1434 for (auto *I : Insts) {
1435 // These instructions may change or break semantics if moved.
1436 if (isa<PHINode>(I) || I->isEHPad() || isa<AllocaInst>(I) ||
1437 I->getType()->isTokenTy())
1438 return false;
1439
1440 // Conservatively return false if I is an inline-asm instruction. Sinking
1441 // and merging inline-asm instructions can potentially create arguments
1442 // that cannot satisfy the inline-asm constraints.
1443 if (const auto *C = dyn_cast<CallBase>(I))
1444 if (C->isInlineAsm())
1445 return false;
1446
1447 // Everything must have only one use too, apart from stores which
1448 // have no uses.
1449 if (!isa<StoreInst>(I) && !I->hasOneUse())
1450 return false;
1451 }
1452
1453 const Instruction *I0 = Insts.front();
1454 for (auto *I : Insts)
1455 if (!I->isSameOperationAs(I0))
1456 return false;
1457
1458 // All instructions in Insts are known to be the same opcode. If they aren't
1459 // stores, check the only user of each is a PHI or in the same block as the
1460 // instruction, because if a user is in the same block as an instruction
1461 // we're contemplating sinking, it must already be determined to be sinkable.
1462 if (!isa<StoreInst>(I0)) {
1463 auto *PNUse = dyn_cast<PHINode>(*I0->user_begin());
1464 auto *Succ = I0->getParent()->getTerminator()->getSuccessor(0);
1465 if (!all_of(Insts, [&PNUse,&Succ](const Instruction *I) -> bool {
1466 auto *U = cast<Instruction>(*I->user_begin());
1467 return (PNUse &&
1468 PNUse->getParent() == Succ &&
1469 PNUse->getIncomingValueForBlock(I->getParent()) == I) ||
1470 U->getParent() == I->getParent();
1471 }))
1472 return false;
1473 }
1474
1475 // Because SROA can't handle speculating stores of selects, try not
1476 // to sink loads or stores of allocas when we'd have to create a PHI for
1477 // the address operand. Also, because it is likely that loads or stores
1478 // of allocas will disappear when Mem2Reg/SROA is run, don't sink them.
1479 // This can cause code churn which can have unintended consequences down
1480 // the line - see https://llvm.org/bugs/show_bug.cgi?id=30244.
1481 // FIXME: This is a workaround for a deficiency in SROA - see
1482 // https://llvm.org/bugs/show_bug.cgi?id=30188
1483 if (isa<StoreInst>(I0) && any_of(Insts, [](const Instruction *I) {
1484 return isa<AllocaInst>(I->getOperand(1));
1485 }))
1486 return false;
1487 if (isa<LoadInst>(I0) && any_of(Insts, [](const Instruction *I) {
1488 return isa<AllocaInst>(I->getOperand(0));
1489 }))
1490 return false;
1491
1492 for (unsigned OI = 0, OE = I0->getNumOperands(); OI != OE; ++OI) {
1493 if (I0->getOperand(OI)->getType()->isTokenTy())
1494 // Don't touch any operand of token type.
1495 return false;
1496
1497 auto SameAsI0 = [&I0, OI](const Instruction *I) {
1498 assert(I->getNumOperands() == I0->getNumOperands());
1499 return I->getOperand(OI) == I0->getOperand(OI);
1500 };
1501 if (!all_of(Insts, SameAsI0)) {
1502 if (!canReplaceOperandWithVariable(I0, OI))
1503 // We can't create a PHI from this GEP.
1504 return false;
1505 // Don't create indirect calls! The called value is the final operand.
1506 if (isa<CallBase>(I0) && OI == OE - 1) {
1507 // FIXME: if the call was *already* indirect, we should do this.
1508 return false;
1509 }
1510 for (auto *I : Insts)
1511 PHIOperands[I].push_back(I->getOperand(OI));
1512 }
1513 }
1514 return true;
1515 }
1516
1517 // Assuming canSinkLastInstruction(Blocks) has returned true, sink the last
1518 // instruction of every block in Blocks to their common successor, commoning
1519 // into one instruction.
sinkLastInstruction(ArrayRef<BasicBlock * > Blocks)1520 static bool sinkLastInstruction(ArrayRef<BasicBlock*> Blocks) {
1521 auto *BBEnd = Blocks[0]->getTerminator()->getSuccessor(0);
1522
1523 // canSinkLastInstruction returning true guarantees that every block has at
1524 // least one non-terminator instruction.
1525 SmallVector<Instruction*,4> Insts;
1526 for (auto *BB : Blocks) {
1527 Instruction *I = BB->getTerminator();
1528 do {
1529 I = I->getPrevNode();
1530 } while (isa<DbgInfoIntrinsic>(I) && I != &BB->front());
1531 if (!isa<DbgInfoIntrinsic>(I))
1532 Insts.push_back(I);
1533 }
1534
1535 // The only checking we need to do now is that all users of all instructions
1536 // are the same PHI node. canSinkLastInstruction should have checked this but
1537 // it is slightly over-aggressive - it gets confused by commutative instructions
1538 // so double-check it here.
1539 Instruction *I0 = Insts.front();
1540 if (!isa<StoreInst>(I0)) {
1541 auto *PNUse = dyn_cast<PHINode>(*I0->user_begin());
1542 if (!all_of(Insts, [&PNUse](const Instruction *I) -> bool {
1543 auto *U = cast<Instruction>(*I->user_begin());
1544 return U == PNUse;
1545 }))
1546 return false;
1547 }
1548
1549 // We don't need to do any more checking here; canSinkLastInstruction should
1550 // have done it all for us.
1551 SmallVector<Value*, 4> NewOperands;
1552 for (unsigned O = 0, E = I0->getNumOperands(); O != E; ++O) {
1553 // This check is different to that in canSinkLastInstruction. There, we
1554 // cared about the global view once simplifycfg (and instcombine) have
1555 // completed - it takes into account PHIs that become trivially
1556 // simplifiable. However here we need a more local view; if an operand
1557 // differs we create a PHI and rely on instcombine to clean up the very
1558 // small mess we may make.
1559 bool NeedPHI = any_of(Insts, [&I0, O](const Instruction *I) {
1560 return I->getOperand(O) != I0->getOperand(O);
1561 });
1562 if (!NeedPHI) {
1563 NewOperands.push_back(I0->getOperand(O));
1564 continue;
1565 }
1566
1567 // Create a new PHI in the successor block and populate it.
1568 auto *Op = I0->getOperand(O);
1569 assert(!Op->getType()->isTokenTy() && "Can't PHI tokens!");
1570 auto *PN = PHINode::Create(Op->getType(), Insts.size(),
1571 Op->getName() + ".sink", &BBEnd->front());
1572 for (auto *I : Insts)
1573 PN->addIncoming(I->getOperand(O), I->getParent());
1574 NewOperands.push_back(PN);
1575 }
1576
1577 // Arbitrarily use I0 as the new "common" instruction; remap its operands
1578 // and move it to the start of the successor block.
1579 for (unsigned O = 0, E = I0->getNumOperands(); O != E; ++O)
1580 I0->getOperandUse(O).set(NewOperands[O]);
1581 I0->moveBefore(&*BBEnd->getFirstInsertionPt());
1582
1583 // Update metadata and IR flags, and merge debug locations.
1584 for (auto *I : Insts)
1585 if (I != I0) {
1586 // The debug location for the "common" instruction is the merged locations
1587 // of all the commoned instructions. We start with the original location
1588 // of the "common" instruction and iteratively merge each location in the
1589 // loop below.
1590 // This is an N-way merge, which will be inefficient if I0 is a CallInst.
1591 // However, as N-way merge for CallInst is rare, so we use simplified API
1592 // instead of using complex API for N-way merge.
1593 I0->applyMergedLocation(I0->getDebugLoc(), I->getDebugLoc());
1594 combineMetadataForCSE(I0, I, true);
1595 I0->andIRFlags(I);
1596 }
1597
1598 if (!isa<StoreInst>(I0)) {
1599 // canSinkLastInstruction checked that all instructions were used by
1600 // one and only one PHI node. Find that now, RAUW it to our common
1601 // instruction and nuke it.
1602 assert(I0->hasOneUse());
1603 auto *PN = cast<PHINode>(*I0->user_begin());
1604 PN->replaceAllUsesWith(I0);
1605 PN->eraseFromParent();
1606 }
1607
1608 // Finally nuke all instructions apart from the common instruction.
1609 for (auto *I : Insts)
1610 if (I != I0)
1611 I->eraseFromParent();
1612
1613 return true;
1614 }
1615
1616 namespace {
1617
1618 // LockstepReverseIterator - Iterates through instructions
1619 // in a set of blocks in reverse order from the first non-terminator.
1620 // For example (assume all blocks have size n):
1621 // LockstepReverseIterator I([B1, B2, B3]);
1622 // *I-- = [B1[n], B2[n], B3[n]];
1623 // *I-- = [B1[n-1], B2[n-1], B3[n-1]];
1624 // *I-- = [B1[n-2], B2[n-2], B3[n-2]];
1625 // ...
1626 class LockstepReverseIterator {
1627 ArrayRef<BasicBlock*> Blocks;
1628 SmallVector<Instruction*,4> Insts;
1629 bool Fail;
1630
1631 public:
LockstepReverseIterator(ArrayRef<BasicBlock * > Blocks)1632 LockstepReverseIterator(ArrayRef<BasicBlock*> Blocks) : Blocks(Blocks) {
1633 reset();
1634 }
1635
reset()1636 void reset() {
1637 Fail = false;
1638 Insts.clear();
1639 for (auto *BB : Blocks) {
1640 Instruction *Inst = BB->getTerminator();
1641 for (Inst = Inst->getPrevNode(); Inst && isa<DbgInfoIntrinsic>(Inst);)
1642 Inst = Inst->getPrevNode();
1643 if (!Inst) {
1644 // Block wasn't big enough.
1645 Fail = true;
1646 return;
1647 }
1648 Insts.push_back(Inst);
1649 }
1650 }
1651
isValid() const1652 bool isValid() const {
1653 return !Fail;
1654 }
1655
operator --()1656 void operator--() {
1657 if (Fail)
1658 return;
1659 for (auto *&Inst : Insts) {
1660 for (Inst = Inst->getPrevNode(); Inst && isa<DbgInfoIntrinsic>(Inst);)
1661 Inst = Inst->getPrevNode();
1662 // Already at beginning of block.
1663 if (!Inst) {
1664 Fail = true;
1665 return;
1666 }
1667 }
1668 }
1669
operator *() const1670 ArrayRef<Instruction*> operator * () const {
1671 return Insts;
1672 }
1673 };
1674
1675 } // end anonymous namespace
1676
1677 /// Check whether BB's predecessors end with unconditional branches. If it is
1678 /// true, sink any common code from the predecessors to BB.
1679 /// We also allow one predecessor to end with conditional branch (but no more
1680 /// than one).
SinkCommonCodeFromPredecessors(BasicBlock * BB)1681 static bool SinkCommonCodeFromPredecessors(BasicBlock *BB) {
1682 // We support two situations:
1683 // (1) all incoming arcs are unconditional
1684 // (2) one incoming arc is conditional
1685 //
1686 // (2) is very common in switch defaults and
1687 // else-if patterns;
1688 //
1689 // if (a) f(1);
1690 // else if (b) f(2);
1691 //
1692 // produces:
1693 //
1694 // [if]
1695 // / \
1696 // [f(1)] [if]
1697 // | | \
1698 // | | |
1699 // | [f(2)]|
1700 // \ | /
1701 // [ end ]
1702 //
1703 // [end] has two unconditional predecessor arcs and one conditional. The
1704 // conditional refers to the implicit empty 'else' arc. This conditional
1705 // arc can also be caused by an empty default block in a switch.
1706 //
1707 // In this case, we attempt to sink code from all *unconditional* arcs.
1708 // If we can sink instructions from these arcs (determined during the scan
1709 // phase below) we insert a common successor for all unconditional arcs and
1710 // connect that to [end], to enable sinking:
1711 //
1712 // [if]
1713 // / \
1714 // [x(1)] [if]
1715 // | | \
1716 // | | \
1717 // | [x(2)] |
1718 // \ / |
1719 // [sink.split] |
1720 // \ /
1721 // [ end ]
1722 //
1723 SmallVector<BasicBlock*,4> UnconditionalPreds;
1724 Instruction *Cond = nullptr;
1725 for (auto *B : predecessors(BB)) {
1726 auto *T = B->getTerminator();
1727 if (isa<BranchInst>(T) && cast<BranchInst>(T)->isUnconditional())
1728 UnconditionalPreds.push_back(B);
1729 else if ((isa<BranchInst>(T) || isa<SwitchInst>(T)) && !Cond)
1730 Cond = T;
1731 else
1732 return false;
1733 }
1734 if (UnconditionalPreds.size() < 2)
1735 return false;
1736
1737 bool Changed = false;
1738 // We take a two-step approach to tail sinking. First we scan from the end of
1739 // each block upwards in lockstep. If the n'th instruction from the end of each
1740 // block can be sunk, those instructions are added to ValuesToSink and we
1741 // carry on. If we can sink an instruction but need to PHI-merge some operands
1742 // (because they're not identical in each instruction) we add these to
1743 // PHIOperands.
1744 unsigned ScanIdx = 0;
1745 SmallPtrSet<Value*,4> InstructionsToSink;
1746 DenseMap<Instruction*, SmallVector<Value*,4>> PHIOperands;
1747 LockstepReverseIterator LRI(UnconditionalPreds);
1748 while (LRI.isValid() &&
1749 canSinkInstructions(*LRI, PHIOperands)) {
1750 LLVM_DEBUG(dbgs() << "SINK: instruction can be sunk: " << *(*LRI)[0]
1751 << "\n");
1752 InstructionsToSink.insert((*LRI).begin(), (*LRI).end());
1753 ++ScanIdx;
1754 --LRI;
1755 }
1756
1757 auto ProfitableToSinkInstruction = [&](LockstepReverseIterator &LRI) {
1758 unsigned NumPHIdValues = 0;
1759 for (auto *I : *LRI)
1760 for (auto *V : PHIOperands[I])
1761 if (InstructionsToSink.count(V) == 0)
1762 ++NumPHIdValues;
1763 LLVM_DEBUG(dbgs() << "SINK: #phid values: " << NumPHIdValues << "\n");
1764 unsigned NumPHIInsts = NumPHIdValues / UnconditionalPreds.size();
1765 if ((NumPHIdValues % UnconditionalPreds.size()) != 0)
1766 NumPHIInsts++;
1767
1768 return NumPHIInsts <= 1;
1769 };
1770
1771 if (ScanIdx > 0 && Cond) {
1772 // Check if we would actually sink anything first! This mutates the CFG and
1773 // adds an extra block. The goal in doing this is to allow instructions that
1774 // couldn't be sunk before to be sunk - obviously, speculatable instructions
1775 // (such as trunc, add) can be sunk and predicated already. So we check that
1776 // we're going to sink at least one non-speculatable instruction.
1777 LRI.reset();
1778 unsigned Idx = 0;
1779 bool Profitable = false;
1780 while (ProfitableToSinkInstruction(LRI) && Idx < ScanIdx) {
1781 if (!isSafeToSpeculativelyExecute((*LRI)[0])) {
1782 Profitable = true;
1783 break;
1784 }
1785 --LRI;
1786 ++Idx;
1787 }
1788 if (!Profitable)
1789 return false;
1790
1791 LLVM_DEBUG(dbgs() << "SINK: Splitting edge\n");
1792 // We have a conditional edge and we're going to sink some instructions.
1793 // Insert a new block postdominating all blocks we're going to sink from.
1794 if (!SplitBlockPredecessors(BB, UnconditionalPreds, ".sink.split"))
1795 // Edges couldn't be split.
1796 return false;
1797 Changed = true;
1798 }
1799
1800 // Now that we've analyzed all potential sinking candidates, perform the
1801 // actual sink. We iteratively sink the last non-terminator of the source
1802 // blocks into their common successor unless doing so would require too
1803 // many PHI instructions to be generated (currently only one PHI is allowed
1804 // per sunk instruction).
1805 //
1806 // We can use InstructionsToSink to discount values needing PHI-merging that will
1807 // actually be sunk in a later iteration. This allows us to be more
1808 // aggressive in what we sink. This does allow a false positive where we
1809 // sink presuming a later value will also be sunk, but stop half way through
1810 // and never actually sink it which means we produce more PHIs than intended.
1811 // This is unlikely in practice though.
1812 for (unsigned SinkIdx = 0; SinkIdx != ScanIdx; ++SinkIdx) {
1813 LLVM_DEBUG(dbgs() << "SINK: Sink: "
1814 << *UnconditionalPreds[0]->getTerminator()->getPrevNode()
1815 << "\n");
1816
1817 // Because we've sunk every instruction in turn, the current instruction to
1818 // sink is always at index 0.
1819 LRI.reset();
1820 if (!ProfitableToSinkInstruction(LRI)) {
1821 // Too many PHIs would be created.
1822 LLVM_DEBUG(
1823 dbgs() << "SINK: stopping here, too many PHIs would be created!\n");
1824 break;
1825 }
1826
1827 if (!sinkLastInstruction(UnconditionalPreds))
1828 return Changed;
1829 NumSinkCommons++;
1830 Changed = true;
1831 }
1832 return Changed;
1833 }
1834
1835 /// Determine if we can hoist sink a sole store instruction out of a
1836 /// conditional block.
1837 ///
1838 /// We are looking for code like the following:
1839 /// BrBB:
1840 /// store i32 %add, i32* %arrayidx2
1841 /// ... // No other stores or function calls (we could be calling a memory
1842 /// ... // function).
1843 /// %cmp = icmp ult %x, %y
1844 /// br i1 %cmp, label %EndBB, label %ThenBB
1845 /// ThenBB:
1846 /// store i32 %add5, i32* %arrayidx2
1847 /// br label EndBB
1848 /// EndBB:
1849 /// ...
1850 /// We are going to transform this into:
1851 /// BrBB:
1852 /// store i32 %add, i32* %arrayidx2
1853 /// ... //
1854 /// %cmp = icmp ult %x, %y
1855 /// %add.add5 = select i1 %cmp, i32 %add, %add5
1856 /// store i32 %add.add5, i32* %arrayidx2
1857 /// ...
1858 ///
1859 /// \return The pointer to the value of the previous store if the store can be
1860 /// hoisted into the predecessor block. 0 otherwise.
isSafeToSpeculateStore(Instruction * I,BasicBlock * BrBB,BasicBlock * StoreBB,BasicBlock * EndBB)1861 static Value *isSafeToSpeculateStore(Instruction *I, BasicBlock *BrBB,
1862 BasicBlock *StoreBB, BasicBlock *EndBB) {
1863 StoreInst *StoreToHoist = dyn_cast<StoreInst>(I);
1864 if (!StoreToHoist)
1865 return nullptr;
1866
1867 // Volatile or atomic.
1868 if (!StoreToHoist->isSimple())
1869 return nullptr;
1870
1871 Value *StorePtr = StoreToHoist->getPointerOperand();
1872
1873 // Look for a store to the same pointer in BrBB.
1874 unsigned MaxNumInstToLookAt = 9;
1875 for (Instruction &CurI : reverse(BrBB->instructionsWithoutDebug())) {
1876 if (!MaxNumInstToLookAt)
1877 break;
1878 --MaxNumInstToLookAt;
1879
1880 // Could be calling an instruction that affects memory like free().
1881 if (CurI.mayHaveSideEffects() && !isa<StoreInst>(CurI))
1882 return nullptr;
1883
1884 if (auto *SI = dyn_cast<StoreInst>(&CurI)) {
1885 // Found the previous store make sure it stores to the same location.
1886 if (SI->getPointerOperand() == StorePtr)
1887 // Found the previous store, return its value operand.
1888 return SI->getValueOperand();
1889 return nullptr; // Unknown store.
1890 }
1891 }
1892
1893 return nullptr;
1894 }
1895
1896 /// Speculate a conditional basic block flattening the CFG.
1897 ///
1898 /// Note that this is a very risky transform currently. Speculating
1899 /// instructions like this is most often not desirable. Instead, there is an MI
1900 /// pass which can do it with full awareness of the resource constraints.
1901 /// However, some cases are "obvious" and we should do directly. An example of
1902 /// this is speculating a single, reasonably cheap instruction.
1903 ///
1904 /// There is only one distinct advantage to flattening the CFG at the IR level:
1905 /// it makes very common but simplistic optimizations such as are common in
1906 /// instcombine and the DAG combiner more powerful by removing CFG edges and
1907 /// modeling their effects with easier to reason about SSA value graphs.
1908 ///
1909 ///
1910 /// An illustration of this transform is turning this IR:
1911 /// \code
1912 /// BB:
1913 /// %cmp = icmp ult %x, %y
1914 /// br i1 %cmp, label %EndBB, label %ThenBB
1915 /// ThenBB:
1916 /// %sub = sub %x, %y
1917 /// br label BB2
1918 /// EndBB:
1919 /// %phi = phi [ %sub, %ThenBB ], [ 0, %EndBB ]
1920 /// ...
1921 /// \endcode
1922 ///
1923 /// Into this IR:
1924 /// \code
1925 /// BB:
1926 /// %cmp = icmp ult %x, %y
1927 /// %sub = sub %x, %y
1928 /// %cond = select i1 %cmp, 0, %sub
1929 /// ...
1930 /// \endcode
1931 ///
1932 /// \returns true if the conditional block is removed.
SpeculativelyExecuteBB(BranchInst * BI,BasicBlock * ThenBB,const TargetTransformInfo & TTI)1933 static bool SpeculativelyExecuteBB(BranchInst *BI, BasicBlock *ThenBB,
1934 const TargetTransformInfo &TTI) {
1935 // Be conservative for now. FP select instruction can often be expensive.
1936 Value *BrCond = BI->getCondition();
1937 if (isa<FCmpInst>(BrCond))
1938 return false;
1939
1940 BasicBlock *BB = BI->getParent();
1941 BasicBlock *EndBB = ThenBB->getTerminator()->getSuccessor(0);
1942
1943 // If ThenBB is actually on the false edge of the conditional branch, remember
1944 // to swap the select operands later.
1945 bool Invert = false;
1946 if (ThenBB != BI->getSuccessor(0)) {
1947 assert(ThenBB == BI->getSuccessor(1) && "No edge from 'if' block?");
1948 Invert = true;
1949 }
1950 assert(EndBB == BI->getSuccessor(!Invert) && "No edge from to end block");
1951
1952 // Keep a count of how many times instructions are used within ThenBB when
1953 // they are candidates for sinking into ThenBB. Specifically:
1954 // - They are defined in BB, and
1955 // - They have no side effects, and
1956 // - All of their uses are in ThenBB.
1957 SmallDenseMap<Instruction *, unsigned, 4> SinkCandidateUseCounts;
1958
1959 SmallVector<Instruction *, 4> SpeculatedDbgIntrinsics;
1960
1961 unsigned SpeculationCost = 0;
1962 Value *SpeculatedStoreValue = nullptr;
1963 StoreInst *SpeculatedStore = nullptr;
1964 for (BasicBlock::iterator BBI = ThenBB->begin(),
1965 BBE = std::prev(ThenBB->end());
1966 BBI != BBE; ++BBI) {
1967 Instruction *I = &*BBI;
1968 // Skip debug info.
1969 if (isa<DbgInfoIntrinsic>(I)) {
1970 SpeculatedDbgIntrinsics.push_back(I);
1971 continue;
1972 }
1973
1974 // Only speculatively execute a single instruction (not counting the
1975 // terminator) for now.
1976 ++SpeculationCost;
1977 if (SpeculationCost > 1)
1978 return false;
1979
1980 // Don't hoist the instruction if it's unsafe or expensive.
1981 if (!isSafeToSpeculativelyExecute(I) &&
1982 !(HoistCondStores && (SpeculatedStoreValue = isSafeToSpeculateStore(
1983 I, BB, ThenBB, EndBB))))
1984 return false;
1985 if (!SpeculatedStoreValue &&
1986 ComputeSpeculationCost(I, TTI) >
1987 PHINodeFoldingThreshold * TargetTransformInfo::TCC_Basic)
1988 return false;
1989
1990 // Store the store speculation candidate.
1991 if (SpeculatedStoreValue)
1992 SpeculatedStore = cast<StoreInst>(I);
1993
1994 // Do not hoist the instruction if any of its operands are defined but not
1995 // used in BB. The transformation will prevent the operand from
1996 // being sunk into the use block.
1997 for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i) {
1998 Instruction *OpI = dyn_cast<Instruction>(*i);
1999 if (!OpI || OpI->getParent() != BB || OpI->mayHaveSideEffects())
2000 continue; // Not a candidate for sinking.
2001
2002 ++SinkCandidateUseCounts[OpI];
2003 }
2004 }
2005
2006 // Consider any sink candidates which are only used in ThenBB as costs for
2007 // speculation. Note, while we iterate over a DenseMap here, we are summing
2008 // and so iteration order isn't significant.
2009 for (SmallDenseMap<Instruction *, unsigned, 4>::iterator
2010 I = SinkCandidateUseCounts.begin(),
2011 E = SinkCandidateUseCounts.end();
2012 I != E; ++I)
2013 if (I->first->hasNUses(I->second)) {
2014 ++SpeculationCost;
2015 if (SpeculationCost > 1)
2016 return false;
2017 }
2018
2019 // Check that the PHI nodes can be converted to selects.
2020 bool HaveRewritablePHIs = false;
2021 for (PHINode &PN : EndBB->phis()) {
2022 Value *OrigV = PN.getIncomingValueForBlock(BB);
2023 Value *ThenV = PN.getIncomingValueForBlock(ThenBB);
2024
2025 // FIXME: Try to remove some of the duplication with HoistThenElseCodeToIf.
2026 // Skip PHIs which are trivial.
2027 if (ThenV == OrigV)
2028 continue;
2029
2030 // Don't convert to selects if we could remove undefined behavior instead.
2031 if (passingValueIsAlwaysUndefined(OrigV, &PN) ||
2032 passingValueIsAlwaysUndefined(ThenV, &PN))
2033 return false;
2034
2035 HaveRewritablePHIs = true;
2036 ConstantExpr *OrigCE = dyn_cast<ConstantExpr>(OrigV);
2037 ConstantExpr *ThenCE = dyn_cast<ConstantExpr>(ThenV);
2038 if (!OrigCE && !ThenCE)
2039 continue; // Known safe and cheap.
2040
2041 if ((ThenCE && !isSafeToSpeculativelyExecute(ThenCE)) ||
2042 (OrigCE && !isSafeToSpeculativelyExecute(OrigCE)))
2043 return false;
2044 unsigned OrigCost = OrigCE ? ComputeSpeculationCost(OrigCE, TTI) : 0;
2045 unsigned ThenCost = ThenCE ? ComputeSpeculationCost(ThenCE, TTI) : 0;
2046 unsigned MaxCost =
2047 2 * PHINodeFoldingThreshold * TargetTransformInfo::TCC_Basic;
2048 if (OrigCost + ThenCost > MaxCost)
2049 return false;
2050
2051 // Account for the cost of an unfolded ConstantExpr which could end up
2052 // getting expanded into Instructions.
2053 // FIXME: This doesn't account for how many operations are combined in the
2054 // constant expression.
2055 ++SpeculationCost;
2056 if (SpeculationCost > 1)
2057 return false;
2058 }
2059
2060 // If there are no PHIs to process, bail early. This helps ensure idempotence
2061 // as well.
2062 if (!HaveRewritablePHIs && !(HoistCondStores && SpeculatedStoreValue))
2063 return false;
2064
2065 // If we get here, we can hoist the instruction and if-convert.
2066 LLVM_DEBUG(dbgs() << "SPECULATIVELY EXECUTING BB" << *ThenBB << "\n";);
2067
2068 // Insert a select of the value of the speculated store.
2069 if (SpeculatedStoreValue) {
2070 IRBuilder<NoFolder> Builder(BI);
2071 Value *TrueV = SpeculatedStore->getValueOperand();
2072 Value *FalseV = SpeculatedStoreValue;
2073 if (Invert)
2074 std::swap(TrueV, FalseV);
2075 Value *S = Builder.CreateSelect(
2076 BrCond, TrueV, FalseV, "spec.store.select", BI);
2077 SpeculatedStore->setOperand(0, S);
2078 SpeculatedStore->applyMergedLocation(BI->getDebugLoc(),
2079 SpeculatedStore->getDebugLoc());
2080 }
2081
2082 // Metadata can be dependent on the condition we are hoisting above.
2083 // Conservatively strip all metadata on the instruction.
2084 for (auto &I : *ThenBB)
2085 I.dropUnknownNonDebugMetadata();
2086
2087 // Hoist the instructions.
2088 BB->getInstList().splice(BI->getIterator(), ThenBB->getInstList(),
2089 ThenBB->begin(), std::prev(ThenBB->end()));
2090
2091 // Insert selects and rewrite the PHI operands.
2092 IRBuilder<NoFolder> Builder(BI);
2093 for (PHINode &PN : EndBB->phis()) {
2094 unsigned OrigI = PN.getBasicBlockIndex(BB);
2095 unsigned ThenI = PN.getBasicBlockIndex(ThenBB);
2096 Value *OrigV = PN.getIncomingValue(OrigI);
2097 Value *ThenV = PN.getIncomingValue(ThenI);
2098
2099 // Skip PHIs which are trivial.
2100 if (OrigV == ThenV)
2101 continue;
2102
2103 // Create a select whose true value is the speculatively executed value and
2104 // false value is the preexisting value. Swap them if the branch
2105 // destinations were inverted.
2106 Value *TrueV = ThenV, *FalseV = OrigV;
2107 if (Invert)
2108 std::swap(TrueV, FalseV);
2109 Value *V = Builder.CreateSelect(
2110 BrCond, TrueV, FalseV, "spec.select", BI);
2111 PN.setIncomingValue(OrigI, V);
2112 PN.setIncomingValue(ThenI, V);
2113 }
2114
2115 // Remove speculated dbg intrinsics.
2116 // FIXME: Is it possible to do this in a more elegant way? Moving/merging the
2117 // dbg value for the different flows and inserting it after the select.
2118 for (Instruction *I : SpeculatedDbgIntrinsics)
2119 I->eraseFromParent();
2120
2121 ++NumSpeculations;
2122 return true;
2123 }
2124
2125 /// Return true if we can thread a branch across this block.
BlockIsSimpleEnoughToThreadThrough(BasicBlock * BB)2126 static bool BlockIsSimpleEnoughToThreadThrough(BasicBlock *BB) {
2127 unsigned Size = 0;
2128
2129 for (Instruction &I : BB->instructionsWithoutDebug()) {
2130 if (Size > 10)
2131 return false; // Don't clone large BB's.
2132 ++Size;
2133
2134 // We can only support instructions that do not define values that are
2135 // live outside of the current basic block.
2136 for (User *U : I.users()) {
2137 Instruction *UI = cast<Instruction>(U);
2138 if (UI->getParent() != BB || isa<PHINode>(UI))
2139 return false;
2140 }
2141
2142 // Looks ok, continue checking.
2143 }
2144
2145 return true;
2146 }
2147
2148 /// If we have a conditional branch on a PHI node value that is defined in the
2149 /// same block as the branch and if any PHI entries are constants, thread edges
2150 /// corresponding to that entry to be branches to their ultimate destination.
FoldCondBranchOnPHI(BranchInst * BI,const DataLayout & DL,AssumptionCache * AC)2151 static bool FoldCondBranchOnPHI(BranchInst *BI, const DataLayout &DL,
2152 AssumptionCache *AC) {
2153 BasicBlock *BB = BI->getParent();
2154 PHINode *PN = dyn_cast<PHINode>(BI->getCondition());
2155 // NOTE: we currently cannot transform this case if the PHI node is used
2156 // outside of the block.
2157 if (!PN || PN->getParent() != BB || !PN->hasOneUse())
2158 return false;
2159
2160 // Degenerate case of a single entry PHI.
2161 if (PN->getNumIncomingValues() == 1) {
2162 FoldSingleEntryPHINodes(PN->getParent());
2163 return true;
2164 }
2165
2166 // Now we know that this block has multiple preds and two succs.
2167 if (!BlockIsSimpleEnoughToThreadThrough(BB))
2168 return false;
2169
2170 // Can't fold blocks that contain noduplicate or convergent calls.
2171 if (any_of(*BB, [](const Instruction &I) {
2172 const CallInst *CI = dyn_cast<CallInst>(&I);
2173 return CI && (CI->cannotDuplicate() || CI->isConvergent());
2174 }))
2175 return false;
2176
2177 // Okay, this is a simple enough basic block. See if any phi values are
2178 // constants.
2179 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
2180 ConstantInt *CB = dyn_cast<ConstantInt>(PN->getIncomingValue(i));
2181 if (!CB || !CB->getType()->isIntegerTy(1))
2182 continue;
2183
2184 // Okay, we now know that all edges from PredBB should be revectored to
2185 // branch to RealDest.
2186 BasicBlock *PredBB = PN->getIncomingBlock(i);
2187 BasicBlock *RealDest = BI->getSuccessor(!CB->getZExtValue());
2188
2189 if (RealDest == BB)
2190 continue; // Skip self loops.
2191 // Skip if the predecessor's terminator is an indirect branch.
2192 if (isa<IndirectBrInst>(PredBB->getTerminator()))
2193 continue;
2194
2195 // The dest block might have PHI nodes, other predecessors and other
2196 // difficult cases. Instead of being smart about this, just insert a new
2197 // block that jumps to the destination block, effectively splitting
2198 // the edge we are about to create.
2199 BasicBlock *EdgeBB =
2200 BasicBlock::Create(BB->getContext(), RealDest->getName() + ".critedge",
2201 RealDest->getParent(), RealDest);
2202 BranchInst *CritEdgeBranch = BranchInst::Create(RealDest, EdgeBB);
2203 CritEdgeBranch->setDebugLoc(BI->getDebugLoc());
2204
2205 // Update PHI nodes.
2206 AddPredecessorToBlock(RealDest, EdgeBB, BB);
2207
2208 // BB may have instructions that are being threaded over. Clone these
2209 // instructions into EdgeBB. We know that there will be no uses of the
2210 // cloned instructions outside of EdgeBB.
2211 BasicBlock::iterator InsertPt = EdgeBB->begin();
2212 DenseMap<Value *, Value *> TranslateMap; // Track translated values.
2213 for (BasicBlock::iterator BBI = BB->begin(); &*BBI != BI; ++BBI) {
2214 if (PHINode *PN = dyn_cast<PHINode>(BBI)) {
2215 TranslateMap[PN] = PN->getIncomingValueForBlock(PredBB);
2216 continue;
2217 }
2218 // Clone the instruction.
2219 Instruction *N = BBI->clone();
2220 if (BBI->hasName())
2221 N->setName(BBI->getName() + ".c");
2222
2223 // Update operands due to translation.
2224 for (User::op_iterator i = N->op_begin(), e = N->op_end(); i != e; ++i) {
2225 DenseMap<Value *, Value *>::iterator PI = TranslateMap.find(*i);
2226 if (PI != TranslateMap.end())
2227 *i = PI->second;
2228 }
2229
2230 // Check for trivial simplification.
2231 if (Value *V = SimplifyInstruction(N, {DL, nullptr, nullptr, AC})) {
2232 if (!BBI->use_empty())
2233 TranslateMap[&*BBI] = V;
2234 if (!N->mayHaveSideEffects()) {
2235 N->deleteValue(); // Instruction folded away, don't need actual inst
2236 N = nullptr;
2237 }
2238 } else {
2239 if (!BBI->use_empty())
2240 TranslateMap[&*BBI] = N;
2241 }
2242 // Insert the new instruction into its new home.
2243 if (N)
2244 EdgeBB->getInstList().insert(InsertPt, N);
2245
2246 // Register the new instruction with the assumption cache if necessary.
2247 if (auto *II = dyn_cast_or_null<IntrinsicInst>(N))
2248 if (II->getIntrinsicID() == Intrinsic::assume)
2249 AC->registerAssumption(II);
2250 }
2251
2252 // Loop over all of the edges from PredBB to BB, changing them to branch
2253 // to EdgeBB instead.
2254 Instruction *PredBBTI = PredBB->getTerminator();
2255 for (unsigned i = 0, e = PredBBTI->getNumSuccessors(); i != e; ++i)
2256 if (PredBBTI->getSuccessor(i) == BB) {
2257 BB->removePredecessor(PredBB);
2258 PredBBTI->setSuccessor(i, EdgeBB);
2259 }
2260
2261 // Recurse, simplifying any other constants.
2262 return FoldCondBranchOnPHI(BI, DL, AC) || true;
2263 }
2264
2265 return false;
2266 }
2267
2268 /// Given a BB that starts with the specified two-entry PHI node,
2269 /// see if we can eliminate it.
FoldTwoEntryPHINode(PHINode * PN,const TargetTransformInfo & TTI,const DataLayout & DL)2270 static bool FoldTwoEntryPHINode(PHINode *PN, const TargetTransformInfo &TTI,
2271 const DataLayout &DL) {
2272 // Ok, this is a two entry PHI node. Check to see if this is a simple "if
2273 // statement", which has a very simple dominance structure. Basically, we
2274 // are trying to find the condition that is being branched on, which
2275 // subsequently causes this merge to happen. We really want control
2276 // dependence information for this check, but simplifycfg can't keep it up
2277 // to date, and this catches most of the cases we care about anyway.
2278 BasicBlock *BB = PN->getParent();
2279 const Function *Fn = BB->getParent();
2280 if (Fn && Fn->hasFnAttribute(Attribute::OptForFuzzing))
2281 return false;
2282
2283 BasicBlock *IfTrue, *IfFalse;
2284 Value *IfCond = GetIfCondition(BB, IfTrue, IfFalse);
2285 if (!IfCond ||
2286 // Don't bother if the branch will be constant folded trivially.
2287 isa<ConstantInt>(IfCond))
2288 return false;
2289
2290 // Okay, we found that we can merge this two-entry phi node into a select.
2291 // Doing so would require us to fold *all* two entry phi nodes in this block.
2292 // At some point this becomes non-profitable (particularly if the target
2293 // doesn't support cmov's). Only do this transformation if there are two or
2294 // fewer PHI nodes in this block.
2295 unsigned NumPhis = 0;
2296 for (BasicBlock::iterator I = BB->begin(); isa<PHINode>(I); ++NumPhis, ++I)
2297 if (NumPhis > 2)
2298 return false;
2299
2300 // Loop over the PHI's seeing if we can promote them all to select
2301 // instructions. While we are at it, keep track of the instructions
2302 // that need to be moved to the dominating block.
2303 SmallPtrSet<Instruction *, 4> AggressiveInsts;
2304 unsigned MaxCostVal0 = PHINodeFoldingThreshold,
2305 MaxCostVal1 = PHINodeFoldingThreshold;
2306 MaxCostVal0 *= TargetTransformInfo::TCC_Basic;
2307 MaxCostVal1 *= TargetTransformInfo::TCC_Basic;
2308
2309 for (BasicBlock::iterator II = BB->begin(); isa<PHINode>(II);) {
2310 PHINode *PN = cast<PHINode>(II++);
2311 if (Value *V = SimplifyInstruction(PN, {DL, PN})) {
2312 PN->replaceAllUsesWith(V);
2313 PN->eraseFromParent();
2314 continue;
2315 }
2316
2317 if (!DominatesMergePoint(PN->getIncomingValue(0), BB, AggressiveInsts,
2318 MaxCostVal0, TTI) ||
2319 !DominatesMergePoint(PN->getIncomingValue(1), BB, AggressiveInsts,
2320 MaxCostVal1, TTI))
2321 return false;
2322 }
2323
2324 // If we folded the first phi, PN dangles at this point. Refresh it. If
2325 // we ran out of PHIs then we simplified them all.
2326 PN = dyn_cast<PHINode>(BB->begin());
2327 if (!PN)
2328 return true;
2329
2330 // Don't fold i1 branches on PHIs which contain binary operators. These can
2331 // often be turned into switches and other things.
2332 if (PN->getType()->isIntegerTy(1) &&
2333 (isa<BinaryOperator>(PN->getIncomingValue(0)) ||
2334 isa<BinaryOperator>(PN->getIncomingValue(1)) ||
2335 isa<BinaryOperator>(IfCond)))
2336 return false;
2337
2338 // If all PHI nodes are promotable, check to make sure that all instructions
2339 // in the predecessor blocks can be promoted as well. If not, we won't be able
2340 // to get rid of the control flow, so it's not worth promoting to select
2341 // instructions.
2342 BasicBlock *DomBlock = nullptr;
2343 BasicBlock *IfBlock1 = PN->getIncomingBlock(0);
2344 BasicBlock *IfBlock2 = PN->getIncomingBlock(1);
2345 if (cast<BranchInst>(IfBlock1->getTerminator())->isConditional()) {
2346 IfBlock1 = nullptr;
2347 } else {
2348 DomBlock = *pred_begin(IfBlock1);
2349 for (BasicBlock::iterator I = IfBlock1->begin(); !I->isTerminator(); ++I)
2350 if (!AggressiveInsts.count(&*I) && !isa<DbgInfoIntrinsic>(I)) {
2351 // This is not an aggressive instruction that we can promote.
2352 // Because of this, we won't be able to get rid of the control flow, so
2353 // the xform is not worth it.
2354 return false;
2355 }
2356 }
2357
2358 if (cast<BranchInst>(IfBlock2->getTerminator())->isConditional()) {
2359 IfBlock2 = nullptr;
2360 } else {
2361 DomBlock = *pred_begin(IfBlock2);
2362 for (BasicBlock::iterator I = IfBlock2->begin(); !I->isTerminator(); ++I)
2363 if (!AggressiveInsts.count(&*I) && !isa<DbgInfoIntrinsic>(I)) {
2364 // This is not an aggressive instruction that we can promote.
2365 // Because of this, we won't be able to get rid of the control flow, so
2366 // the xform is not worth it.
2367 return false;
2368 }
2369 }
2370
2371 LLVM_DEBUG(dbgs() << "FOUND IF CONDITION! " << *IfCond
2372 << " T: " << IfTrue->getName()
2373 << " F: " << IfFalse->getName() << "\n");
2374
2375 // If we can still promote the PHI nodes after this gauntlet of tests,
2376 // do all of the PHI's now.
2377 Instruction *InsertPt = DomBlock->getTerminator();
2378 IRBuilder<NoFolder> Builder(InsertPt);
2379
2380 // Move all 'aggressive' instructions, which are defined in the
2381 // conditional parts of the if's up to the dominating block.
2382 if (IfBlock1)
2383 hoistAllInstructionsInto(DomBlock, InsertPt, IfBlock1);
2384 if (IfBlock2)
2385 hoistAllInstructionsInto(DomBlock, InsertPt, IfBlock2);
2386
2387 while (PHINode *PN = dyn_cast<PHINode>(BB->begin())) {
2388 // Change the PHI node into a select instruction.
2389 Value *TrueVal = PN->getIncomingValue(PN->getIncomingBlock(0) == IfFalse);
2390 Value *FalseVal = PN->getIncomingValue(PN->getIncomingBlock(0) == IfTrue);
2391
2392 Value *Sel = Builder.CreateSelect(IfCond, TrueVal, FalseVal, "", InsertPt);
2393 PN->replaceAllUsesWith(Sel);
2394 Sel->takeName(PN);
2395 PN->eraseFromParent();
2396 }
2397
2398 // At this point, IfBlock1 and IfBlock2 are both empty, so our if statement
2399 // has been flattened. Change DomBlock to jump directly to our new block to
2400 // avoid other simplifycfg's kicking in on the diamond.
2401 Instruction *OldTI = DomBlock->getTerminator();
2402 Builder.SetInsertPoint(OldTI);
2403 Builder.CreateBr(BB);
2404 OldTI->eraseFromParent();
2405 return true;
2406 }
2407
2408 /// If we found a conditional branch that goes to two returning blocks,
2409 /// try to merge them together into one return,
2410 /// introducing a select if the return values disagree.
SimplifyCondBranchToTwoReturns(BranchInst * BI,IRBuilder<> & Builder)2411 static bool SimplifyCondBranchToTwoReturns(BranchInst *BI,
2412 IRBuilder<> &Builder) {
2413 assert(BI->isConditional() && "Must be a conditional branch");
2414 BasicBlock *TrueSucc = BI->getSuccessor(0);
2415 BasicBlock *FalseSucc = BI->getSuccessor(1);
2416 ReturnInst *TrueRet = cast<ReturnInst>(TrueSucc->getTerminator());
2417 ReturnInst *FalseRet = cast<ReturnInst>(FalseSucc->getTerminator());
2418
2419 // Check to ensure both blocks are empty (just a return) or optionally empty
2420 // with PHI nodes. If there are other instructions, merging would cause extra
2421 // computation on one path or the other.
2422 if (!TrueSucc->getFirstNonPHIOrDbg()->isTerminator())
2423 return false;
2424 if (!FalseSucc->getFirstNonPHIOrDbg()->isTerminator())
2425 return false;
2426
2427 Builder.SetInsertPoint(BI);
2428 // Okay, we found a branch that is going to two return nodes. If
2429 // there is no return value for this function, just change the
2430 // branch into a return.
2431 if (FalseRet->getNumOperands() == 0) {
2432 TrueSucc->removePredecessor(BI->getParent());
2433 FalseSucc->removePredecessor(BI->getParent());
2434 Builder.CreateRetVoid();
2435 EraseTerminatorAndDCECond(BI);
2436 return true;
2437 }
2438
2439 // Otherwise, figure out what the true and false return values are
2440 // so we can insert a new select instruction.
2441 Value *TrueValue = TrueRet->getReturnValue();
2442 Value *FalseValue = FalseRet->getReturnValue();
2443
2444 // Unwrap any PHI nodes in the return blocks.
2445 if (PHINode *TVPN = dyn_cast_or_null<PHINode>(TrueValue))
2446 if (TVPN->getParent() == TrueSucc)
2447 TrueValue = TVPN->getIncomingValueForBlock(BI->getParent());
2448 if (PHINode *FVPN = dyn_cast_or_null<PHINode>(FalseValue))
2449 if (FVPN->getParent() == FalseSucc)
2450 FalseValue = FVPN->getIncomingValueForBlock(BI->getParent());
2451
2452 // In order for this transformation to be safe, we must be able to
2453 // unconditionally execute both operands to the return. This is
2454 // normally the case, but we could have a potentially-trapping
2455 // constant expression that prevents this transformation from being
2456 // safe.
2457 if (ConstantExpr *TCV = dyn_cast_or_null<ConstantExpr>(TrueValue))
2458 if (TCV->canTrap())
2459 return false;
2460 if (ConstantExpr *FCV = dyn_cast_or_null<ConstantExpr>(FalseValue))
2461 if (FCV->canTrap())
2462 return false;
2463
2464 // Okay, we collected all the mapped values and checked them for sanity, and
2465 // defined to really do this transformation. First, update the CFG.
2466 TrueSucc->removePredecessor(BI->getParent());
2467 FalseSucc->removePredecessor(BI->getParent());
2468
2469 // Insert select instructions where needed.
2470 Value *BrCond = BI->getCondition();
2471 if (TrueValue) {
2472 // Insert a select if the results differ.
2473 if (TrueValue == FalseValue || isa<UndefValue>(FalseValue)) {
2474 } else if (isa<UndefValue>(TrueValue)) {
2475 TrueValue = FalseValue;
2476 } else {
2477 TrueValue =
2478 Builder.CreateSelect(BrCond, TrueValue, FalseValue, "retval", BI);
2479 }
2480 }
2481
2482 Value *RI =
2483 !TrueValue ? Builder.CreateRetVoid() : Builder.CreateRet(TrueValue);
2484
2485 (void)RI;
2486
2487 LLVM_DEBUG(dbgs() << "\nCHANGING BRANCH TO TWO RETURNS INTO SELECT:"
2488 << "\n " << *BI << "NewRet = " << *RI << "TRUEBLOCK: "
2489 << *TrueSucc << "FALSEBLOCK: " << *FalseSucc);
2490
2491 EraseTerminatorAndDCECond(BI);
2492
2493 return true;
2494 }
2495
2496 /// Return true if the given instruction is available
2497 /// in its predecessor block. If yes, the instruction will be removed.
tryCSEWithPredecessor(Instruction * Inst,BasicBlock * PB)2498 static bool tryCSEWithPredecessor(Instruction *Inst, BasicBlock *PB) {
2499 if (!isa<BinaryOperator>(Inst) && !isa<CmpInst>(Inst))
2500 return false;
2501 for (Instruction &I : *PB) {
2502 Instruction *PBI = &I;
2503 // Check whether Inst and PBI generate the same value.
2504 if (Inst->isIdenticalTo(PBI)) {
2505 Inst->replaceAllUsesWith(PBI);
2506 Inst->eraseFromParent();
2507 return true;
2508 }
2509 }
2510 return false;
2511 }
2512
2513 /// Return true if either PBI or BI has branch weight available, and store
2514 /// the weights in {Pred|Succ}{True|False}Weight. If one of PBI and BI does
2515 /// 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)2516 static bool extractPredSuccWeights(BranchInst *PBI, BranchInst *BI,
2517 uint64_t &PredTrueWeight,
2518 uint64_t &PredFalseWeight,
2519 uint64_t &SuccTrueWeight,
2520 uint64_t &SuccFalseWeight) {
2521 bool PredHasWeights =
2522 PBI->extractProfMetadata(PredTrueWeight, PredFalseWeight);
2523 bool SuccHasWeights =
2524 BI->extractProfMetadata(SuccTrueWeight, SuccFalseWeight);
2525 if (PredHasWeights || SuccHasWeights) {
2526 if (!PredHasWeights)
2527 PredTrueWeight = PredFalseWeight = 1;
2528 if (!SuccHasWeights)
2529 SuccTrueWeight = SuccFalseWeight = 1;
2530 return true;
2531 } else {
2532 return false;
2533 }
2534 }
2535
2536 /// If this basic block is simple enough, and if a predecessor branches to us
2537 /// and one of our successors, fold the block into the predecessor and use
2538 /// logical operations to pick the right destination.
FoldBranchToCommonDest(BranchInst * BI,MemorySSAUpdater * MSSAU,unsigned BonusInstThreshold)2539 bool llvm::FoldBranchToCommonDest(BranchInst *BI, MemorySSAUpdater *MSSAU,
2540 unsigned BonusInstThreshold) {
2541 BasicBlock *BB = BI->getParent();
2542
2543 const unsigned PredCount = pred_size(BB);
2544
2545 Instruction *Cond = nullptr;
2546 if (BI->isConditional())
2547 Cond = dyn_cast<Instruction>(BI->getCondition());
2548 else {
2549 // For unconditional branch, check for a simple CFG pattern, where
2550 // BB has a single predecessor and BB's successor is also its predecessor's
2551 // successor. If such pattern exists, check for CSE between BB and its
2552 // predecessor.
2553 if (BasicBlock *PB = BB->getSinglePredecessor())
2554 if (BranchInst *PBI = dyn_cast<BranchInst>(PB->getTerminator()))
2555 if (PBI->isConditional() &&
2556 (BI->getSuccessor(0) == PBI->getSuccessor(0) ||
2557 BI->getSuccessor(0) == PBI->getSuccessor(1))) {
2558 for (auto I = BB->instructionsWithoutDebug().begin(),
2559 E = BB->instructionsWithoutDebug().end();
2560 I != E;) {
2561 Instruction *Curr = &*I++;
2562 if (isa<CmpInst>(Curr)) {
2563 Cond = Curr;
2564 break;
2565 }
2566 // Quit if we can't remove this instruction.
2567 if (!tryCSEWithPredecessor(Curr, PB))
2568 return false;
2569 }
2570 }
2571
2572 if (!Cond)
2573 return false;
2574 }
2575
2576 if (!Cond || (!isa<CmpInst>(Cond) && !isa<BinaryOperator>(Cond)) ||
2577 Cond->getParent() != BB || !Cond->hasOneUse())
2578 return false;
2579
2580 // Make sure the instruction after the condition is the cond branch.
2581 BasicBlock::iterator CondIt = ++Cond->getIterator();
2582
2583 // Ignore dbg intrinsics.
2584 while (isa<DbgInfoIntrinsic>(CondIt))
2585 ++CondIt;
2586
2587 if (&*CondIt != BI)
2588 return false;
2589
2590 // Only allow this transformation if computing the condition doesn't involve
2591 // too many instructions and these involved instructions can be executed
2592 // unconditionally. We denote all involved instructions except the condition
2593 // as "bonus instructions", and only allow this transformation when the
2594 // number of the bonus instructions we'll need to create when cloning into
2595 // each predecessor does not exceed a certain threshold.
2596 unsigned NumBonusInsts = 0;
2597 for (auto I = BB->begin(); Cond != &*I; ++I) {
2598 // Ignore dbg intrinsics.
2599 if (isa<DbgInfoIntrinsic>(I))
2600 continue;
2601 if (!I->hasOneUse() || !isSafeToSpeculativelyExecute(&*I))
2602 return false;
2603 // I has only one use and can be executed unconditionally.
2604 Instruction *User = dyn_cast<Instruction>(I->user_back());
2605 if (User == nullptr || User->getParent() != BB)
2606 return false;
2607 // I is used in the same BB. Since BI uses Cond and doesn't have more slots
2608 // to use any other instruction, User must be an instruction between next(I)
2609 // and Cond.
2610
2611 // Account for the cost of duplicating this instruction into each
2612 // predecessor.
2613 NumBonusInsts += PredCount;
2614 // Early exits once we reach the limit.
2615 if (NumBonusInsts > BonusInstThreshold)
2616 return false;
2617 }
2618
2619 // Cond is known to be a compare or binary operator. Check to make sure that
2620 // neither operand is a potentially-trapping constant expression.
2621 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(0)))
2622 if (CE->canTrap())
2623 return false;
2624 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(1)))
2625 if (CE->canTrap())
2626 return false;
2627
2628 // Finally, don't infinitely unroll conditional loops.
2629 BasicBlock *TrueDest = BI->getSuccessor(0);
2630 BasicBlock *FalseDest = (BI->isConditional()) ? BI->getSuccessor(1) : nullptr;
2631 if (TrueDest == BB || FalseDest == BB)
2632 return false;
2633
2634 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
2635 BasicBlock *PredBlock = *PI;
2636 BranchInst *PBI = dyn_cast<BranchInst>(PredBlock->getTerminator());
2637
2638 // Check that we have two conditional branches. If there is a PHI node in
2639 // the common successor, verify that the same value flows in from both
2640 // blocks.
2641 SmallVector<PHINode *, 4> PHIs;
2642 if (!PBI || PBI->isUnconditional() ||
2643 (BI->isConditional() && !SafeToMergeTerminators(BI, PBI)) ||
2644 (!BI->isConditional() &&
2645 !isProfitableToFoldUnconditional(BI, PBI, Cond, PHIs)))
2646 continue;
2647
2648 // Determine if the two branches share a common destination.
2649 Instruction::BinaryOps Opc = Instruction::BinaryOpsEnd;
2650 bool InvertPredCond = false;
2651
2652 if (BI->isConditional()) {
2653 if (PBI->getSuccessor(0) == TrueDest) {
2654 Opc = Instruction::Or;
2655 } else if (PBI->getSuccessor(1) == FalseDest) {
2656 Opc = Instruction::And;
2657 } else if (PBI->getSuccessor(0) == FalseDest) {
2658 Opc = Instruction::And;
2659 InvertPredCond = true;
2660 } else if (PBI->getSuccessor(1) == TrueDest) {
2661 Opc = Instruction::Or;
2662 InvertPredCond = true;
2663 } else {
2664 continue;
2665 }
2666 } else {
2667 if (PBI->getSuccessor(0) != TrueDest && PBI->getSuccessor(1) != TrueDest)
2668 continue;
2669 }
2670
2671 LLVM_DEBUG(dbgs() << "FOLDING BRANCH TO COMMON DEST:\n" << *PBI << *BB);
2672 IRBuilder<> Builder(PBI);
2673
2674 // If we need to invert the condition in the pred block to match, do so now.
2675 if (InvertPredCond) {
2676 Value *NewCond = PBI->getCondition();
2677
2678 if (NewCond->hasOneUse() && isa<CmpInst>(NewCond)) {
2679 CmpInst *CI = cast<CmpInst>(NewCond);
2680 CI->setPredicate(CI->getInversePredicate());
2681 } else {
2682 NewCond =
2683 Builder.CreateNot(NewCond, PBI->getCondition()->getName() + ".not");
2684 }
2685
2686 PBI->setCondition(NewCond);
2687 PBI->swapSuccessors();
2688 }
2689
2690 // If we have bonus instructions, clone them into the predecessor block.
2691 // Note that there may be multiple predecessor blocks, so we cannot move
2692 // bonus instructions to a predecessor block.
2693 ValueToValueMapTy VMap; // maps original values to cloned values
2694 // We already make sure Cond is the last instruction before BI. Therefore,
2695 // all instructions before Cond other than DbgInfoIntrinsic are bonus
2696 // instructions.
2697 for (auto BonusInst = BB->begin(); Cond != &*BonusInst; ++BonusInst) {
2698 if (isa<DbgInfoIntrinsic>(BonusInst))
2699 continue;
2700 Instruction *NewBonusInst = BonusInst->clone();
2701 RemapInstruction(NewBonusInst, VMap,
2702 RF_NoModuleLevelChanges | RF_IgnoreMissingLocals);
2703 VMap[&*BonusInst] = NewBonusInst;
2704
2705 // If we moved a load, we cannot any longer claim any knowledge about
2706 // its potential value. The previous information might have been valid
2707 // only given the branch precondition.
2708 // For an analogous reason, we must also drop all the metadata whose
2709 // semantics we don't understand.
2710 NewBonusInst->dropUnknownNonDebugMetadata();
2711
2712 PredBlock->getInstList().insert(PBI->getIterator(), NewBonusInst);
2713 NewBonusInst->takeName(&*BonusInst);
2714 BonusInst->setName(BonusInst->getName() + ".old");
2715 }
2716
2717 // Clone Cond into the predecessor basic block, and or/and the
2718 // two conditions together.
2719 Instruction *CondInPred = Cond->clone();
2720 RemapInstruction(CondInPred, VMap,
2721 RF_NoModuleLevelChanges | RF_IgnoreMissingLocals);
2722 PredBlock->getInstList().insert(PBI->getIterator(), CondInPred);
2723 CondInPred->takeName(Cond);
2724 Cond->setName(CondInPred->getName() + ".old");
2725
2726 if (BI->isConditional()) {
2727 Instruction *NewCond = cast<Instruction>(
2728 Builder.CreateBinOp(Opc, PBI->getCondition(), CondInPred, "or.cond"));
2729 PBI->setCondition(NewCond);
2730
2731 uint64_t PredTrueWeight, PredFalseWeight, SuccTrueWeight, SuccFalseWeight;
2732 bool HasWeights =
2733 extractPredSuccWeights(PBI, BI, PredTrueWeight, PredFalseWeight,
2734 SuccTrueWeight, SuccFalseWeight);
2735 SmallVector<uint64_t, 8> NewWeights;
2736
2737 if (PBI->getSuccessor(0) == BB) {
2738 if (HasWeights) {
2739 // PBI: br i1 %x, BB, FalseDest
2740 // BI: br i1 %y, TrueDest, FalseDest
2741 // TrueWeight is TrueWeight for PBI * TrueWeight for BI.
2742 NewWeights.push_back(PredTrueWeight * SuccTrueWeight);
2743 // FalseWeight is FalseWeight for PBI * TotalWeight for BI +
2744 // TrueWeight for PBI * FalseWeight for BI.
2745 // We assume that total weights of a BranchInst can fit into 32 bits.
2746 // Therefore, we will not have overflow using 64-bit arithmetic.
2747 NewWeights.push_back(PredFalseWeight *
2748 (SuccFalseWeight + SuccTrueWeight) +
2749 PredTrueWeight * SuccFalseWeight);
2750 }
2751 AddPredecessorToBlock(TrueDest, PredBlock, BB, MSSAU);
2752 PBI->setSuccessor(0, TrueDest);
2753 }
2754 if (PBI->getSuccessor(1) == BB) {
2755 if (HasWeights) {
2756 // PBI: br i1 %x, TrueDest, BB
2757 // BI: br i1 %y, TrueDest, FalseDest
2758 // TrueWeight is TrueWeight for PBI * TotalWeight for BI +
2759 // FalseWeight for PBI * TrueWeight for BI.
2760 NewWeights.push_back(PredTrueWeight *
2761 (SuccFalseWeight + SuccTrueWeight) +
2762 PredFalseWeight * SuccTrueWeight);
2763 // FalseWeight is FalseWeight for PBI * FalseWeight for BI.
2764 NewWeights.push_back(PredFalseWeight * SuccFalseWeight);
2765 }
2766 AddPredecessorToBlock(FalseDest, PredBlock, BB, MSSAU);
2767 PBI->setSuccessor(1, FalseDest);
2768 }
2769 if (NewWeights.size() == 2) {
2770 // Halve the weights if any of them cannot fit in an uint32_t
2771 FitWeights(NewWeights);
2772
2773 SmallVector<uint32_t, 8> MDWeights(NewWeights.begin(),
2774 NewWeights.end());
2775 setBranchWeights(PBI, MDWeights[0], MDWeights[1]);
2776 } else
2777 PBI->setMetadata(LLVMContext::MD_prof, nullptr);
2778 } else {
2779 // Update PHI nodes in the common successors.
2780 for (unsigned i = 0, e = PHIs.size(); i != e; ++i) {
2781 ConstantInt *PBI_C = cast<ConstantInt>(
2782 PHIs[i]->getIncomingValueForBlock(PBI->getParent()));
2783 assert(PBI_C->getType()->isIntegerTy(1));
2784 Instruction *MergedCond = nullptr;
2785 if (PBI->getSuccessor(0) == TrueDest) {
2786 // Create (PBI_Cond and PBI_C) or (!PBI_Cond and BI_Value)
2787 // PBI_C is true: PBI_Cond or (!PBI_Cond and BI_Value)
2788 // is false: !PBI_Cond and BI_Value
2789 Instruction *NotCond = cast<Instruction>(
2790 Builder.CreateNot(PBI->getCondition(), "not.cond"));
2791 MergedCond = cast<Instruction>(
2792 Builder.CreateBinOp(Instruction::And, NotCond, CondInPred,
2793 "and.cond"));
2794 if (PBI_C->isOne())
2795 MergedCond = cast<Instruction>(Builder.CreateBinOp(
2796 Instruction::Or, PBI->getCondition(), MergedCond, "or.cond"));
2797 } else {
2798 // Create (PBI_Cond and BI_Value) or (!PBI_Cond and PBI_C)
2799 // PBI_C is true: (PBI_Cond and BI_Value) or (!PBI_Cond)
2800 // is false: PBI_Cond and BI_Value
2801 MergedCond = cast<Instruction>(Builder.CreateBinOp(
2802 Instruction::And, PBI->getCondition(), CondInPred, "and.cond"));
2803 if (PBI_C->isOne()) {
2804 Instruction *NotCond = cast<Instruction>(
2805 Builder.CreateNot(PBI->getCondition(), "not.cond"));
2806 MergedCond = cast<Instruction>(Builder.CreateBinOp(
2807 Instruction::Or, NotCond, MergedCond, "or.cond"));
2808 }
2809 }
2810 // Update PHI Node.
2811 PHIs[i]->setIncomingValueForBlock(PBI->getParent(), MergedCond);
2812 }
2813
2814 // PBI is changed to branch to TrueDest below. Remove itself from
2815 // potential phis from all other successors.
2816 if (MSSAU)
2817 MSSAU->changeCondBranchToUnconditionalTo(PBI, TrueDest);
2818
2819 // Change PBI from Conditional to Unconditional.
2820 BranchInst *New_PBI = BranchInst::Create(TrueDest, PBI);
2821 EraseTerminatorAndDCECond(PBI, MSSAU);
2822 PBI = New_PBI;
2823 }
2824
2825 // If BI was a loop latch, it may have had associated loop metadata.
2826 // We need to copy it to the new latch, that is, PBI.
2827 if (MDNode *LoopMD = BI->getMetadata(LLVMContext::MD_loop))
2828 PBI->setMetadata(LLVMContext::MD_loop, LoopMD);
2829
2830 // TODO: If BB is reachable from all paths through PredBlock, then we
2831 // could replace PBI's branch probabilities with BI's.
2832
2833 // Copy any debug value intrinsics into the end of PredBlock.
2834 for (Instruction &I : *BB)
2835 if (isa<DbgInfoIntrinsic>(I))
2836 I.clone()->insertBefore(PBI);
2837
2838 return true;
2839 }
2840 return false;
2841 }
2842
2843 // If there is only one store in BB1 and BB2, return it, otherwise return
2844 // nullptr.
findUniqueStoreInBlocks(BasicBlock * BB1,BasicBlock * BB2)2845 static StoreInst *findUniqueStoreInBlocks(BasicBlock *BB1, BasicBlock *BB2) {
2846 StoreInst *S = nullptr;
2847 for (auto *BB : {BB1, BB2}) {
2848 if (!BB)
2849 continue;
2850 for (auto &I : *BB)
2851 if (auto *SI = dyn_cast<StoreInst>(&I)) {
2852 if (S)
2853 // Multiple stores seen.
2854 return nullptr;
2855 else
2856 S = SI;
2857 }
2858 }
2859 return S;
2860 }
2861
ensureValueAvailableInSuccessor(Value * V,BasicBlock * BB,Value * AlternativeV=nullptr)2862 static Value *ensureValueAvailableInSuccessor(Value *V, BasicBlock *BB,
2863 Value *AlternativeV = nullptr) {
2864 // PHI is going to be a PHI node that allows the value V that is defined in
2865 // BB to be referenced in BB's only successor.
2866 //
2867 // If AlternativeV is nullptr, the only value we care about in PHI is V. It
2868 // doesn't matter to us what the other operand is (it'll never get used). We
2869 // could just create a new PHI with an undef incoming value, but that could
2870 // increase register pressure if EarlyCSE/InstCombine can't fold it with some
2871 // other PHI. So here we directly look for some PHI in BB's successor with V
2872 // as an incoming operand. If we find one, we use it, else we create a new
2873 // one.
2874 //
2875 // If AlternativeV is not nullptr, we care about both incoming values in PHI.
2876 // PHI must be exactly: phi <ty> [ %BB, %V ], [ %OtherBB, %AlternativeV]
2877 // where OtherBB is the single other predecessor of BB's only successor.
2878 PHINode *PHI = nullptr;
2879 BasicBlock *Succ = BB->getSingleSuccessor();
2880
2881 for (auto I = Succ->begin(); isa<PHINode>(I); ++I)
2882 if (cast<PHINode>(I)->getIncomingValueForBlock(BB) == V) {
2883 PHI = cast<PHINode>(I);
2884 if (!AlternativeV)
2885 break;
2886
2887 assert(Succ->hasNPredecessors(2));
2888 auto PredI = pred_begin(Succ);
2889 BasicBlock *OtherPredBB = *PredI == BB ? *++PredI : *PredI;
2890 if (PHI->getIncomingValueForBlock(OtherPredBB) == AlternativeV)
2891 break;
2892 PHI = nullptr;
2893 }
2894 if (PHI)
2895 return PHI;
2896
2897 // If V is not an instruction defined in BB, just return it.
2898 if (!AlternativeV &&
2899 (!isa<Instruction>(V) || cast<Instruction>(V)->getParent() != BB))
2900 return V;
2901
2902 PHI = PHINode::Create(V->getType(), 2, "simplifycfg.merge", &Succ->front());
2903 PHI->addIncoming(V, BB);
2904 for (BasicBlock *PredBB : predecessors(Succ))
2905 if (PredBB != BB)
2906 PHI->addIncoming(
2907 AlternativeV ? AlternativeV : UndefValue::get(V->getType()), PredBB);
2908 return PHI;
2909 }
2910
mergeConditionalStoreToAddress(BasicBlock * PTB,BasicBlock * PFB,BasicBlock * QTB,BasicBlock * QFB,BasicBlock * PostBB,Value * Address,bool InvertPCond,bool InvertQCond,const DataLayout & DL)2911 static bool mergeConditionalStoreToAddress(BasicBlock *PTB, BasicBlock *PFB,
2912 BasicBlock *QTB, BasicBlock *QFB,
2913 BasicBlock *PostBB, Value *Address,
2914 bool InvertPCond, bool InvertQCond,
2915 const DataLayout &DL) {
2916 auto IsaBitcastOfPointerType = [](const Instruction &I) {
2917 return Operator::getOpcode(&I) == Instruction::BitCast &&
2918 I.getType()->isPointerTy();
2919 };
2920
2921 // If we're not in aggressive mode, we only optimize if we have some
2922 // confidence that by optimizing we'll allow P and/or Q to be if-converted.
2923 auto IsWorthwhile = [&](BasicBlock *BB) {
2924 if (!BB)
2925 return true;
2926 // Heuristic: if the block can be if-converted/phi-folded and the
2927 // instructions inside are all cheap (arithmetic/GEPs), it's worthwhile to
2928 // thread this store.
2929 unsigned N = 0;
2930 for (auto &I : BB->instructionsWithoutDebug()) {
2931 // Cheap instructions viable for folding.
2932 if (isa<BinaryOperator>(I) || isa<GetElementPtrInst>(I) ||
2933 isa<StoreInst>(I))
2934 ++N;
2935 // Free instructions.
2936 else if (I.isTerminator() || IsaBitcastOfPointerType(I))
2937 continue;
2938 else
2939 return false;
2940 }
2941 // The store we want to merge is counted in N, so add 1 to make sure
2942 // we're counting the instructions that would be left.
2943 return N <= (PHINodeFoldingThreshold + 1);
2944 };
2945
2946 if (!MergeCondStoresAggressively &&
2947 (!IsWorthwhile(PTB) || !IsWorthwhile(PFB) || !IsWorthwhile(QTB) ||
2948 !IsWorthwhile(QFB)))
2949 return false;
2950
2951 // For every pointer, there must be exactly two stores, one coming from
2952 // PTB or PFB, and the other from QTB or QFB. We don't support more than one
2953 // store (to any address) in PTB,PFB or QTB,QFB.
2954 // FIXME: We could relax this restriction with a bit more work and performance
2955 // testing.
2956 StoreInst *PStore = findUniqueStoreInBlocks(PTB, PFB);
2957 StoreInst *QStore = findUniqueStoreInBlocks(QTB, QFB);
2958 if (!PStore || !QStore)
2959 return false;
2960
2961 // Now check the stores are compatible.
2962 if (!QStore->isUnordered() || !PStore->isUnordered())
2963 return false;
2964
2965 // Check that sinking the store won't cause program behavior changes. Sinking
2966 // the store out of the Q blocks won't change any behavior as we're sinking
2967 // from a block to its unconditional successor. But we're moving a store from
2968 // the P blocks down through the middle block (QBI) and past both QFB and QTB.
2969 // So we need to check that there are no aliasing loads or stores in
2970 // QBI, QTB and QFB. We also need to check there are no conflicting memory
2971 // operations between PStore and the end of its parent block.
2972 //
2973 // The ideal way to do this is to query AliasAnalysis, but we don't
2974 // preserve AA currently so that is dangerous. Be super safe and just
2975 // check there are no other memory operations at all.
2976 for (auto &I : *QFB->getSinglePredecessor())
2977 if (I.mayReadOrWriteMemory())
2978 return false;
2979 for (auto &I : *QFB)
2980 if (&I != QStore && I.mayReadOrWriteMemory())
2981 return false;
2982 if (QTB)
2983 for (auto &I : *QTB)
2984 if (&I != QStore && I.mayReadOrWriteMemory())
2985 return false;
2986 for (auto I = BasicBlock::iterator(PStore), E = PStore->getParent()->end();
2987 I != E; ++I)
2988 if (&*I != PStore && I->mayReadOrWriteMemory())
2989 return false;
2990
2991 // If PostBB has more than two predecessors, we need to split it so we can
2992 // sink the store.
2993 if (std::next(pred_begin(PostBB), 2) != pred_end(PostBB)) {
2994 // We know that QFB's only successor is PostBB. And QFB has a single
2995 // predecessor. If QTB exists, then its only successor is also PostBB.
2996 // If QTB does not exist, then QFB's only predecessor has a conditional
2997 // branch to QFB and PostBB.
2998 BasicBlock *TruePred = QTB ? QTB : QFB->getSinglePredecessor();
2999 BasicBlock *NewBB = SplitBlockPredecessors(PostBB, { QFB, TruePred},
3000 "condstore.split");
3001 if (!NewBB)
3002 return false;
3003 PostBB = NewBB;
3004 }
3005
3006 // OK, we're going to sink the stores to PostBB. The store has to be
3007 // conditional though, so first create the predicate.
3008 Value *PCond = cast<BranchInst>(PFB->getSinglePredecessor()->getTerminator())
3009 ->getCondition();
3010 Value *QCond = cast<BranchInst>(QFB->getSinglePredecessor()->getTerminator())
3011 ->getCondition();
3012
3013 Value *PPHI = ensureValueAvailableInSuccessor(PStore->getValueOperand(),
3014 PStore->getParent());
3015 Value *QPHI = ensureValueAvailableInSuccessor(QStore->getValueOperand(),
3016 QStore->getParent(), PPHI);
3017
3018 IRBuilder<> QB(&*PostBB->getFirstInsertionPt());
3019
3020 Value *PPred = PStore->getParent() == PTB ? PCond : QB.CreateNot(PCond);
3021 Value *QPred = QStore->getParent() == QTB ? QCond : QB.CreateNot(QCond);
3022
3023 if (InvertPCond)
3024 PPred = QB.CreateNot(PPred);
3025 if (InvertQCond)
3026 QPred = QB.CreateNot(QPred);
3027 Value *CombinedPred = QB.CreateOr(PPred, QPred);
3028
3029 auto *T =
3030 SplitBlockAndInsertIfThen(CombinedPred, &*QB.GetInsertPoint(), false);
3031 QB.SetInsertPoint(T);
3032 StoreInst *SI = cast<StoreInst>(QB.CreateStore(QPHI, Address));
3033 AAMDNodes AAMD;
3034 PStore->getAAMetadata(AAMD, /*Merge=*/false);
3035 PStore->getAAMetadata(AAMD, /*Merge=*/true);
3036 SI->setAAMetadata(AAMD);
3037 unsigned PAlignment = PStore->getAlignment();
3038 unsigned QAlignment = QStore->getAlignment();
3039 unsigned TypeAlignment =
3040 DL.getABITypeAlignment(SI->getValueOperand()->getType());
3041 unsigned MinAlignment;
3042 unsigned MaxAlignment;
3043 std::tie(MinAlignment, MaxAlignment) = std::minmax(PAlignment, QAlignment);
3044 // Choose the minimum alignment. If we could prove both stores execute, we
3045 // could use biggest one. In this case, though, we only know that one of the
3046 // stores executes. And we don't know it's safe to take the alignment from a
3047 // store that doesn't execute.
3048 if (MinAlignment != 0) {
3049 // Choose the minimum of all non-zero alignments.
3050 SI->setAlignment(MinAlignment);
3051 } else if (MaxAlignment != 0) {
3052 // Choose the minimal alignment between the non-zero alignment and the ABI
3053 // default alignment for the type of the stored value.
3054 SI->setAlignment(std::min(MaxAlignment, TypeAlignment));
3055 } else {
3056 // If both alignments are zero, use ABI default alignment for the type of
3057 // the stored value.
3058 SI->setAlignment(TypeAlignment);
3059 }
3060
3061 QStore->eraseFromParent();
3062 PStore->eraseFromParent();
3063
3064 return true;
3065 }
3066
mergeConditionalStores(BranchInst * PBI,BranchInst * QBI,const DataLayout & DL)3067 static bool mergeConditionalStores(BranchInst *PBI, BranchInst *QBI,
3068 const DataLayout &DL) {
3069 // The intention here is to find diamonds or triangles (see below) where each
3070 // conditional block contains a store to the same address. Both of these
3071 // stores are conditional, so they can't be unconditionally sunk. But it may
3072 // be profitable to speculatively sink the stores into one merged store at the
3073 // end, and predicate the merged store on the union of the two conditions of
3074 // PBI and QBI.
3075 //
3076 // This can reduce the number of stores executed if both of the conditions are
3077 // true, and can allow the blocks to become small enough to be if-converted.
3078 // This optimization will also chain, so that ladders of test-and-set
3079 // sequences can be if-converted away.
3080 //
3081 // We only deal with simple diamonds or triangles:
3082 //
3083 // PBI or PBI or a combination of the two
3084 // / \ | \
3085 // PTB PFB | PFB
3086 // \ / | /
3087 // QBI QBI
3088 // / \ | \
3089 // QTB QFB | QFB
3090 // \ / | /
3091 // PostBB PostBB
3092 //
3093 // We model triangles as a type of diamond with a nullptr "true" block.
3094 // Triangles are canonicalized so that the fallthrough edge is represented by
3095 // a true condition, as in the diagram above.
3096 BasicBlock *PTB = PBI->getSuccessor(0);
3097 BasicBlock *PFB = PBI->getSuccessor(1);
3098 BasicBlock *QTB = QBI->getSuccessor(0);
3099 BasicBlock *QFB = QBI->getSuccessor(1);
3100 BasicBlock *PostBB = QFB->getSingleSuccessor();
3101
3102 // Make sure we have a good guess for PostBB. If QTB's only successor is
3103 // QFB, then QFB is a better PostBB.
3104 if (QTB->getSingleSuccessor() == QFB)
3105 PostBB = QFB;
3106
3107 // If we couldn't find a good PostBB, stop.
3108 if (!PostBB)
3109 return false;
3110
3111 bool InvertPCond = false, InvertQCond = false;
3112 // Canonicalize fallthroughs to the true branches.
3113 if (PFB == QBI->getParent()) {
3114 std::swap(PFB, PTB);
3115 InvertPCond = true;
3116 }
3117 if (QFB == PostBB) {
3118 std::swap(QFB, QTB);
3119 InvertQCond = true;
3120 }
3121
3122 // From this point on we can assume PTB or QTB may be fallthroughs but PFB
3123 // and QFB may not. Model fallthroughs as a nullptr block.
3124 if (PTB == QBI->getParent())
3125 PTB = nullptr;
3126 if (QTB == PostBB)
3127 QTB = nullptr;
3128
3129 // Legality bailouts. We must have at least the non-fallthrough blocks and
3130 // the post-dominating block, and the non-fallthroughs must only have one
3131 // predecessor.
3132 auto HasOnePredAndOneSucc = [](BasicBlock *BB, BasicBlock *P, BasicBlock *S) {
3133 return BB->getSinglePredecessor() == P && BB->getSingleSuccessor() == S;
3134 };
3135 if (!HasOnePredAndOneSucc(PFB, PBI->getParent(), QBI->getParent()) ||
3136 !HasOnePredAndOneSucc(QFB, QBI->getParent(), PostBB))
3137 return false;
3138 if ((PTB && !HasOnePredAndOneSucc(PTB, PBI->getParent(), QBI->getParent())) ||
3139 (QTB && !HasOnePredAndOneSucc(QTB, QBI->getParent(), PostBB)))
3140 return false;
3141 if (!QBI->getParent()->hasNUses(2))
3142 return false;
3143
3144 // OK, this is a sequence of two diamonds or triangles.
3145 // Check if there are stores in PTB or PFB that are repeated in QTB or QFB.
3146 SmallPtrSet<Value *, 4> PStoreAddresses, QStoreAddresses;
3147 for (auto *BB : {PTB, PFB}) {
3148 if (!BB)
3149 continue;
3150 for (auto &I : *BB)
3151 if (StoreInst *SI = dyn_cast<StoreInst>(&I))
3152 PStoreAddresses.insert(SI->getPointerOperand());
3153 }
3154 for (auto *BB : {QTB, QFB}) {
3155 if (!BB)
3156 continue;
3157 for (auto &I : *BB)
3158 if (StoreInst *SI = dyn_cast<StoreInst>(&I))
3159 QStoreAddresses.insert(SI->getPointerOperand());
3160 }
3161
3162 set_intersect(PStoreAddresses, QStoreAddresses);
3163 // set_intersect mutates PStoreAddresses in place. Rename it here to make it
3164 // clear what it contains.
3165 auto &CommonAddresses = PStoreAddresses;
3166
3167 bool Changed = false;
3168 for (auto *Address : CommonAddresses)
3169 Changed |= mergeConditionalStoreToAddress(
3170 PTB, PFB, QTB, QFB, PostBB, Address, InvertPCond, InvertQCond, DL);
3171 return Changed;
3172 }
3173
3174 /// If we have a conditional branch as a predecessor of another block,
3175 /// this function tries to simplify it. We know
3176 /// that PBI and BI are both conditional branches, and BI is in one of the
3177 /// successor blocks of PBI - PBI branches to BI.
SimplifyCondBranchToCondBranch(BranchInst * PBI,BranchInst * BI,const DataLayout & DL)3178 static bool SimplifyCondBranchToCondBranch(BranchInst *PBI, BranchInst *BI,
3179 const DataLayout &DL) {
3180 assert(PBI->isConditional() && BI->isConditional());
3181 BasicBlock *BB = BI->getParent();
3182
3183 // If this block ends with a branch instruction, and if there is a
3184 // predecessor that ends on a branch of the same condition, make
3185 // this conditional branch redundant.
3186 if (PBI->getCondition() == BI->getCondition() &&
3187 PBI->getSuccessor(0) != PBI->getSuccessor(1)) {
3188 // Okay, the outcome of this conditional branch is statically
3189 // knowable. If this block had a single pred, handle specially.
3190 if (BB->getSinglePredecessor()) {
3191 // Turn this into a branch on constant.
3192 bool CondIsTrue = PBI->getSuccessor(0) == BB;
3193 BI->setCondition(
3194 ConstantInt::get(Type::getInt1Ty(BB->getContext()), CondIsTrue));
3195 return true; // Nuke the branch on constant.
3196 }
3197
3198 // Otherwise, if there are multiple predecessors, insert a PHI that merges
3199 // in the constant and simplify the block result. Subsequent passes of
3200 // simplifycfg will thread the block.
3201 if (BlockIsSimpleEnoughToThreadThrough(BB)) {
3202 pred_iterator PB = pred_begin(BB), PE = pred_end(BB);
3203 PHINode *NewPN = PHINode::Create(
3204 Type::getInt1Ty(BB->getContext()), std::distance(PB, PE),
3205 BI->getCondition()->getName() + ".pr", &BB->front());
3206 // Okay, we're going to insert the PHI node. Since PBI is not the only
3207 // predecessor, compute the PHI'd conditional value for all of the preds.
3208 // Any predecessor where the condition is not computable we keep symbolic.
3209 for (pred_iterator PI = PB; PI != PE; ++PI) {
3210 BasicBlock *P = *PI;
3211 if ((PBI = dyn_cast<BranchInst>(P->getTerminator())) && PBI != BI &&
3212 PBI->isConditional() && PBI->getCondition() == BI->getCondition() &&
3213 PBI->getSuccessor(0) != PBI->getSuccessor(1)) {
3214 bool CondIsTrue = PBI->getSuccessor(0) == BB;
3215 NewPN->addIncoming(
3216 ConstantInt::get(Type::getInt1Ty(BB->getContext()), CondIsTrue),
3217 P);
3218 } else {
3219 NewPN->addIncoming(BI->getCondition(), P);
3220 }
3221 }
3222
3223 BI->setCondition(NewPN);
3224 return true;
3225 }
3226 }
3227
3228 if (auto *CE = dyn_cast<ConstantExpr>(BI->getCondition()))
3229 if (CE->canTrap())
3230 return false;
3231
3232 // If both branches are conditional and both contain stores to the same
3233 // address, remove the stores from the conditionals and create a conditional
3234 // merged store at the end.
3235 if (MergeCondStores && mergeConditionalStores(PBI, BI, DL))
3236 return true;
3237
3238 // If this is a conditional branch in an empty block, and if any
3239 // predecessors are a conditional branch to one of our destinations,
3240 // fold the conditions into logical ops and one cond br.
3241
3242 // Ignore dbg intrinsics.
3243 if (&*BB->instructionsWithoutDebug().begin() != BI)
3244 return false;
3245
3246 int PBIOp, BIOp;
3247 if (PBI->getSuccessor(0) == BI->getSuccessor(0)) {
3248 PBIOp = 0;
3249 BIOp = 0;
3250 } else if (PBI->getSuccessor(0) == BI->getSuccessor(1)) {
3251 PBIOp = 0;
3252 BIOp = 1;
3253 } else if (PBI->getSuccessor(1) == BI->getSuccessor(0)) {
3254 PBIOp = 1;
3255 BIOp = 0;
3256 } else if (PBI->getSuccessor(1) == BI->getSuccessor(1)) {
3257 PBIOp = 1;
3258 BIOp = 1;
3259 } else {
3260 return false;
3261 }
3262
3263 // Check to make sure that the other destination of this branch
3264 // isn't BB itself. If so, this is an infinite loop that will
3265 // keep getting unwound.
3266 if (PBI->getSuccessor(PBIOp) == BB)
3267 return false;
3268
3269 // Do not perform this transformation if it would require
3270 // insertion of a large number of select instructions. For targets
3271 // without predication/cmovs, this is a big pessimization.
3272
3273 // Also do not perform this transformation if any phi node in the common
3274 // destination block can trap when reached by BB or PBB (PR17073). In that
3275 // case, it would be unsafe to hoist the operation into a select instruction.
3276
3277 BasicBlock *CommonDest = PBI->getSuccessor(PBIOp);
3278 unsigned NumPhis = 0;
3279 for (BasicBlock::iterator II = CommonDest->begin(); isa<PHINode>(II);
3280 ++II, ++NumPhis) {
3281 if (NumPhis > 2) // Disable this xform.
3282 return false;
3283
3284 PHINode *PN = cast<PHINode>(II);
3285 Value *BIV = PN->getIncomingValueForBlock(BB);
3286 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(BIV))
3287 if (CE->canTrap())
3288 return false;
3289
3290 unsigned PBBIdx = PN->getBasicBlockIndex(PBI->getParent());
3291 Value *PBIV = PN->getIncomingValue(PBBIdx);
3292 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(PBIV))
3293 if (CE->canTrap())
3294 return false;
3295 }
3296
3297 // Finally, if everything is ok, fold the branches to logical ops.
3298 BasicBlock *OtherDest = BI->getSuccessor(BIOp ^ 1);
3299
3300 LLVM_DEBUG(dbgs() << "FOLDING BRs:" << *PBI->getParent()
3301 << "AND: " << *BI->getParent());
3302
3303 // If OtherDest *is* BB, then BB is a basic block with a single conditional
3304 // branch in it, where one edge (OtherDest) goes back to itself but the other
3305 // exits. We don't *know* that the program avoids the infinite loop
3306 // (even though that seems likely). If we do this xform naively, we'll end up
3307 // recursively unpeeling the loop. Since we know that (after the xform is
3308 // done) that the block *is* infinite if reached, we just make it an obviously
3309 // infinite loop with no cond branch.
3310 if (OtherDest == BB) {
3311 // Insert it at the end of the function, because it's either code,
3312 // or it won't matter if it's hot. :)
3313 BasicBlock *InfLoopBlock =
3314 BasicBlock::Create(BB->getContext(), "infloop", BB->getParent());
3315 BranchInst::Create(InfLoopBlock, InfLoopBlock);
3316 OtherDest = InfLoopBlock;
3317 }
3318
3319 LLVM_DEBUG(dbgs() << *PBI->getParent()->getParent());
3320
3321 // BI may have other predecessors. Because of this, we leave
3322 // it alone, but modify PBI.
3323
3324 // Make sure we get to CommonDest on True&True directions.
3325 Value *PBICond = PBI->getCondition();
3326 IRBuilder<NoFolder> Builder(PBI);
3327 if (PBIOp)
3328 PBICond = Builder.CreateNot(PBICond, PBICond->getName() + ".not");
3329
3330 Value *BICond = BI->getCondition();
3331 if (BIOp)
3332 BICond = Builder.CreateNot(BICond, BICond->getName() + ".not");
3333
3334 // Merge the conditions.
3335 Value *Cond = Builder.CreateOr(PBICond, BICond, "brmerge");
3336
3337 // Modify PBI to branch on the new condition to the new dests.
3338 PBI->setCondition(Cond);
3339 PBI->setSuccessor(0, CommonDest);
3340 PBI->setSuccessor(1, OtherDest);
3341
3342 // Update branch weight for PBI.
3343 uint64_t PredTrueWeight, PredFalseWeight, SuccTrueWeight, SuccFalseWeight;
3344 uint64_t PredCommon, PredOther, SuccCommon, SuccOther;
3345 bool HasWeights =
3346 extractPredSuccWeights(PBI, BI, PredTrueWeight, PredFalseWeight,
3347 SuccTrueWeight, SuccFalseWeight);
3348 if (HasWeights) {
3349 PredCommon = PBIOp ? PredFalseWeight : PredTrueWeight;
3350 PredOther = PBIOp ? PredTrueWeight : PredFalseWeight;
3351 SuccCommon = BIOp ? SuccFalseWeight : SuccTrueWeight;
3352 SuccOther = BIOp ? SuccTrueWeight : SuccFalseWeight;
3353 // The weight to CommonDest should be PredCommon * SuccTotal +
3354 // PredOther * SuccCommon.
3355 // The weight to OtherDest should be PredOther * SuccOther.
3356 uint64_t NewWeights[2] = {PredCommon * (SuccCommon + SuccOther) +
3357 PredOther * SuccCommon,
3358 PredOther * SuccOther};
3359 // Halve the weights if any of them cannot fit in an uint32_t
3360 FitWeights(NewWeights);
3361
3362 setBranchWeights(PBI, NewWeights[0], NewWeights[1]);
3363 }
3364
3365 // OtherDest may have phi nodes. If so, add an entry from PBI's
3366 // block that are identical to the entries for BI's block.
3367 AddPredecessorToBlock(OtherDest, PBI->getParent(), BB);
3368
3369 // We know that the CommonDest already had an edge from PBI to
3370 // it. If it has PHIs though, the PHIs may have different
3371 // entries for BB and PBI's BB. If so, insert a select to make
3372 // them agree.
3373 for (PHINode &PN : CommonDest->phis()) {
3374 Value *BIV = PN.getIncomingValueForBlock(BB);
3375 unsigned PBBIdx = PN.getBasicBlockIndex(PBI->getParent());
3376 Value *PBIV = PN.getIncomingValue(PBBIdx);
3377 if (BIV != PBIV) {
3378 // Insert a select in PBI to pick the right value.
3379 SelectInst *NV = cast<SelectInst>(
3380 Builder.CreateSelect(PBICond, PBIV, BIV, PBIV->getName() + ".mux"));
3381 PN.setIncomingValue(PBBIdx, NV);
3382 // Although the select has the same condition as PBI, the original branch
3383 // weights for PBI do not apply to the new select because the select's
3384 // 'logical' edges are incoming edges of the phi that is eliminated, not
3385 // the outgoing edges of PBI.
3386 if (HasWeights) {
3387 uint64_t PredCommon = PBIOp ? PredFalseWeight : PredTrueWeight;
3388 uint64_t PredOther = PBIOp ? PredTrueWeight : PredFalseWeight;
3389 uint64_t SuccCommon = BIOp ? SuccFalseWeight : SuccTrueWeight;
3390 uint64_t SuccOther = BIOp ? SuccTrueWeight : SuccFalseWeight;
3391 // The weight to PredCommonDest should be PredCommon * SuccTotal.
3392 // The weight to PredOtherDest should be PredOther * SuccCommon.
3393 uint64_t NewWeights[2] = {PredCommon * (SuccCommon + SuccOther),
3394 PredOther * SuccCommon};
3395
3396 FitWeights(NewWeights);
3397
3398 setBranchWeights(NV, NewWeights[0], NewWeights[1]);
3399 }
3400 }
3401 }
3402
3403 LLVM_DEBUG(dbgs() << "INTO: " << *PBI->getParent());
3404 LLVM_DEBUG(dbgs() << *PBI->getParent()->getParent());
3405
3406 // This basic block is probably dead. We know it has at least
3407 // one fewer predecessor.
3408 return true;
3409 }
3410
3411 // Simplifies a terminator by replacing it with a branch to TrueBB if Cond is
3412 // true or to FalseBB if Cond is false.
3413 // Takes care of updating the successors and removing the old terminator.
3414 // Also makes sure not to introduce new successors by assuming that edges to
3415 // non-successor TrueBBs and FalseBBs aren't reachable.
SimplifyTerminatorOnSelect(Instruction * OldTerm,Value * Cond,BasicBlock * TrueBB,BasicBlock * FalseBB,uint32_t TrueWeight,uint32_t FalseWeight)3416 static bool SimplifyTerminatorOnSelect(Instruction *OldTerm, Value *Cond,
3417 BasicBlock *TrueBB, BasicBlock *FalseBB,
3418 uint32_t TrueWeight,
3419 uint32_t FalseWeight) {
3420 // Remove any superfluous successor edges from the CFG.
3421 // First, figure out which successors to preserve.
3422 // If TrueBB and FalseBB are equal, only try to preserve one copy of that
3423 // successor.
3424 BasicBlock *KeepEdge1 = TrueBB;
3425 BasicBlock *KeepEdge2 = TrueBB != FalseBB ? FalseBB : nullptr;
3426
3427 // Then remove the rest.
3428 for (BasicBlock *Succ : successors(OldTerm)) {
3429 // Make sure only to keep exactly one copy of each edge.
3430 if (Succ == KeepEdge1)
3431 KeepEdge1 = nullptr;
3432 else if (Succ == KeepEdge2)
3433 KeepEdge2 = nullptr;
3434 else
3435 Succ->removePredecessor(OldTerm->getParent(),
3436 /*KeepOneInputPHIs=*/true);
3437 }
3438
3439 IRBuilder<> Builder(OldTerm);
3440 Builder.SetCurrentDebugLocation(OldTerm->getDebugLoc());
3441
3442 // Insert an appropriate new terminator.
3443 if (!KeepEdge1 && !KeepEdge2) {
3444 if (TrueBB == FalseBB)
3445 // We were only looking for one successor, and it was present.
3446 // Create an unconditional branch to it.
3447 Builder.CreateBr(TrueBB);
3448 else {
3449 // We found both of the successors we were looking for.
3450 // Create a conditional branch sharing the condition of the select.
3451 BranchInst *NewBI = Builder.CreateCondBr(Cond, TrueBB, FalseBB);
3452 if (TrueWeight != FalseWeight)
3453 setBranchWeights(NewBI, TrueWeight, FalseWeight);
3454 }
3455 } else if (KeepEdge1 && (KeepEdge2 || TrueBB == FalseBB)) {
3456 // Neither of the selected blocks were successors, so this
3457 // terminator must be unreachable.
3458 new UnreachableInst(OldTerm->getContext(), OldTerm);
3459 } else {
3460 // One of the selected values was a successor, but the other wasn't.
3461 // Insert an unconditional branch to the one that was found;
3462 // the edge to the one that wasn't must be unreachable.
3463 if (!KeepEdge1)
3464 // Only TrueBB was found.
3465 Builder.CreateBr(TrueBB);
3466 else
3467 // Only FalseBB was found.
3468 Builder.CreateBr(FalseBB);
3469 }
3470
3471 EraseTerminatorAndDCECond(OldTerm);
3472 return true;
3473 }
3474
3475 // Replaces
3476 // (switch (select cond, X, Y)) on constant X, Y
3477 // with a branch - conditional if X and Y lead to distinct BBs,
3478 // unconditional otherwise.
SimplifySwitchOnSelect(SwitchInst * SI,SelectInst * Select)3479 static bool SimplifySwitchOnSelect(SwitchInst *SI, SelectInst *Select) {
3480 // Check for constant integer values in the select.
3481 ConstantInt *TrueVal = dyn_cast<ConstantInt>(Select->getTrueValue());
3482 ConstantInt *FalseVal = dyn_cast<ConstantInt>(Select->getFalseValue());
3483 if (!TrueVal || !FalseVal)
3484 return false;
3485
3486 // Find the relevant condition and destinations.
3487 Value *Condition = Select->getCondition();
3488 BasicBlock *TrueBB = SI->findCaseValue(TrueVal)->getCaseSuccessor();
3489 BasicBlock *FalseBB = SI->findCaseValue(FalseVal)->getCaseSuccessor();
3490
3491 // Get weight for TrueBB and FalseBB.
3492 uint32_t TrueWeight = 0, FalseWeight = 0;
3493 SmallVector<uint64_t, 8> Weights;
3494 bool HasWeights = HasBranchWeights(SI);
3495 if (HasWeights) {
3496 GetBranchWeights(SI, Weights);
3497 if (Weights.size() == 1 + SI->getNumCases()) {
3498 TrueWeight =
3499 (uint32_t)Weights[SI->findCaseValue(TrueVal)->getSuccessorIndex()];
3500 FalseWeight =
3501 (uint32_t)Weights[SI->findCaseValue(FalseVal)->getSuccessorIndex()];
3502 }
3503 }
3504
3505 // Perform the actual simplification.
3506 return SimplifyTerminatorOnSelect(SI, Condition, TrueBB, FalseBB, TrueWeight,
3507 FalseWeight);
3508 }
3509
3510 // Replaces
3511 // (indirectbr (select cond, blockaddress(@fn, BlockA),
3512 // blockaddress(@fn, BlockB)))
3513 // with
3514 // (br cond, BlockA, BlockB).
SimplifyIndirectBrOnSelect(IndirectBrInst * IBI,SelectInst * SI)3515 static bool SimplifyIndirectBrOnSelect(IndirectBrInst *IBI, SelectInst *SI) {
3516 // Check that both operands of the select are block addresses.
3517 BlockAddress *TBA = dyn_cast<BlockAddress>(SI->getTrueValue());
3518 BlockAddress *FBA = dyn_cast<BlockAddress>(SI->getFalseValue());
3519 if (!TBA || !FBA)
3520 return false;
3521
3522 // Extract the actual blocks.
3523 BasicBlock *TrueBB = TBA->getBasicBlock();
3524 BasicBlock *FalseBB = FBA->getBasicBlock();
3525
3526 // Perform the actual simplification.
3527 return SimplifyTerminatorOnSelect(IBI, SI->getCondition(), TrueBB, FalseBB, 0,
3528 0);
3529 }
3530
3531 /// This is called when we find an icmp instruction
3532 /// (a seteq/setne with a constant) as the only instruction in a
3533 /// block that ends with an uncond branch. We are looking for a very specific
3534 /// pattern that occurs when "A == 1 || A == 2 || A == 3" gets simplified. In
3535 /// this case, we merge the first two "or's of icmp" into a switch, but then the
3536 /// default value goes to an uncond block with a seteq in it, we get something
3537 /// like:
3538 ///
3539 /// switch i8 %A, label %DEFAULT [ i8 1, label %end i8 2, label %end ]
3540 /// DEFAULT:
3541 /// %tmp = icmp eq i8 %A, 92
3542 /// br label %end
3543 /// end:
3544 /// ... = phi i1 [ true, %entry ], [ %tmp, %DEFAULT ], [ true, %entry ]
3545 ///
3546 /// We prefer to split the edge to 'end' so that there is a true/false entry to
3547 /// the PHI, merging the third icmp into the switch.
tryToSimplifyUncondBranchWithICmpInIt(ICmpInst * ICI,IRBuilder<> & Builder)3548 bool SimplifyCFGOpt::tryToSimplifyUncondBranchWithICmpInIt(
3549 ICmpInst *ICI, IRBuilder<> &Builder) {
3550 BasicBlock *BB = ICI->getParent();
3551
3552 // If the block has any PHIs in it or the icmp has multiple uses, it is too
3553 // complex.
3554 if (isa<PHINode>(BB->begin()) || !ICI->hasOneUse())
3555 return false;
3556
3557 Value *V = ICI->getOperand(0);
3558 ConstantInt *Cst = cast<ConstantInt>(ICI->getOperand(1));
3559
3560 // The pattern we're looking for is where our only predecessor is a switch on
3561 // 'V' and this block is the default case for the switch. In this case we can
3562 // fold the compared value into the switch to simplify things.
3563 BasicBlock *Pred = BB->getSinglePredecessor();
3564 if (!Pred || !isa<SwitchInst>(Pred->getTerminator()))
3565 return false;
3566
3567 SwitchInst *SI = cast<SwitchInst>(Pred->getTerminator());
3568 if (SI->getCondition() != V)
3569 return false;
3570
3571 // If BB is reachable on a non-default case, then we simply know the value of
3572 // V in this block. Substitute it and constant fold the icmp instruction
3573 // away.
3574 if (SI->getDefaultDest() != BB) {
3575 ConstantInt *VVal = SI->findCaseDest(BB);
3576 assert(VVal && "Should have a unique destination value");
3577 ICI->setOperand(0, VVal);
3578
3579 if (Value *V = SimplifyInstruction(ICI, {DL, ICI})) {
3580 ICI->replaceAllUsesWith(V);
3581 ICI->eraseFromParent();
3582 }
3583 // BB is now empty, so it is likely to simplify away.
3584 return requestResimplify();
3585 }
3586
3587 // Ok, the block is reachable from the default dest. If the constant we're
3588 // comparing exists in one of the other edges, then we can constant fold ICI
3589 // and zap it.
3590 if (SI->findCaseValue(Cst) != SI->case_default()) {
3591 Value *V;
3592 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
3593 V = ConstantInt::getFalse(BB->getContext());
3594 else
3595 V = ConstantInt::getTrue(BB->getContext());
3596
3597 ICI->replaceAllUsesWith(V);
3598 ICI->eraseFromParent();
3599 // BB is now empty, so it is likely to simplify away.
3600 return requestResimplify();
3601 }
3602
3603 // The use of the icmp has to be in the 'end' block, by the only PHI node in
3604 // the block.
3605 BasicBlock *SuccBlock = BB->getTerminator()->getSuccessor(0);
3606 PHINode *PHIUse = dyn_cast<PHINode>(ICI->user_back());
3607 if (PHIUse == nullptr || PHIUse != &SuccBlock->front() ||
3608 isa<PHINode>(++BasicBlock::iterator(PHIUse)))
3609 return false;
3610
3611 // If the icmp is a SETEQ, then the default dest gets false, the new edge gets
3612 // true in the PHI.
3613 Constant *DefaultCst = ConstantInt::getTrue(BB->getContext());
3614 Constant *NewCst = ConstantInt::getFalse(BB->getContext());
3615
3616 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
3617 std::swap(DefaultCst, NewCst);
3618
3619 // Replace ICI (which is used by the PHI for the default value) with true or
3620 // false depending on if it is EQ or NE.
3621 ICI->replaceAllUsesWith(DefaultCst);
3622 ICI->eraseFromParent();
3623
3624 // Okay, the switch goes to this block on a default value. Add an edge from
3625 // the switch to the merge point on the compared value.
3626 BasicBlock *NewBB =
3627 BasicBlock::Create(BB->getContext(), "switch.edge", BB->getParent(), BB);
3628 {
3629 SwitchInstProfUpdateWrapper SIW(*SI);
3630 auto W0 = SIW.getSuccessorWeight(0);
3631 SwitchInstProfUpdateWrapper::CaseWeightOpt NewW;
3632 if (W0) {
3633 NewW = ((uint64_t(*W0) + 1) >> 1);
3634 SIW.setSuccessorWeight(0, *NewW);
3635 }
3636 SIW.addCase(Cst, NewBB, NewW);
3637 }
3638
3639 // NewBB branches to the phi block, add the uncond branch and the phi entry.
3640 Builder.SetInsertPoint(NewBB);
3641 Builder.SetCurrentDebugLocation(SI->getDebugLoc());
3642 Builder.CreateBr(SuccBlock);
3643 PHIUse->addIncoming(NewCst, NewBB);
3644 return true;
3645 }
3646
3647 /// The specified branch is a conditional branch.
3648 /// Check to see if it is branching on an or/and chain of icmp instructions, and
3649 /// fold it into a switch instruction if so.
SimplifyBranchOnICmpChain(BranchInst * BI,IRBuilder<> & Builder,const DataLayout & DL)3650 static bool SimplifyBranchOnICmpChain(BranchInst *BI, IRBuilder<> &Builder,
3651 const DataLayout &DL) {
3652 Instruction *Cond = dyn_cast<Instruction>(BI->getCondition());
3653 if (!Cond)
3654 return false;
3655
3656 // Change br (X == 0 | X == 1), T, F into a switch instruction.
3657 // If this is a bunch of seteq's or'd together, or if it's a bunch of
3658 // 'setne's and'ed together, collect them.
3659
3660 // Try to gather values from a chain of and/or to be turned into a switch
3661 ConstantComparesGatherer ConstantCompare(Cond, DL);
3662 // Unpack the result
3663 SmallVectorImpl<ConstantInt *> &Values = ConstantCompare.Vals;
3664 Value *CompVal = ConstantCompare.CompValue;
3665 unsigned UsedICmps = ConstantCompare.UsedICmps;
3666 Value *ExtraCase = ConstantCompare.Extra;
3667
3668 // If we didn't have a multiply compared value, fail.
3669 if (!CompVal)
3670 return false;
3671
3672 // Avoid turning single icmps into a switch.
3673 if (UsedICmps <= 1)
3674 return false;
3675
3676 bool TrueWhenEqual = (Cond->getOpcode() == Instruction::Or);
3677
3678 // There might be duplicate constants in the list, which the switch
3679 // instruction can't handle, remove them now.
3680 array_pod_sort(Values.begin(), Values.end(), ConstantIntSortPredicate);
3681 Values.erase(std::unique(Values.begin(), Values.end()), Values.end());
3682
3683 // If Extra was used, we require at least two switch values to do the
3684 // transformation. A switch with one value is just a conditional branch.
3685 if (ExtraCase && Values.size() < 2)
3686 return false;
3687
3688 // TODO: Preserve branch weight metadata, similarly to how
3689 // FoldValueComparisonIntoPredecessors preserves it.
3690
3691 // Figure out which block is which destination.
3692 BasicBlock *DefaultBB = BI->getSuccessor(1);
3693 BasicBlock *EdgeBB = BI->getSuccessor(0);
3694 if (!TrueWhenEqual)
3695 std::swap(DefaultBB, EdgeBB);
3696
3697 BasicBlock *BB = BI->getParent();
3698
3699 LLVM_DEBUG(dbgs() << "Converting 'icmp' chain with " << Values.size()
3700 << " cases into SWITCH. BB is:\n"
3701 << *BB);
3702
3703 // If there are any extra values that couldn't be folded into the switch
3704 // then we evaluate them with an explicit branch first. Split the block
3705 // right before the condbr to handle it.
3706 if (ExtraCase) {
3707 BasicBlock *NewBB =
3708 BB->splitBasicBlock(BI->getIterator(), "switch.early.test");
3709 // Remove the uncond branch added to the old block.
3710 Instruction *OldTI = BB->getTerminator();
3711 Builder.SetInsertPoint(OldTI);
3712
3713 if (TrueWhenEqual)
3714 Builder.CreateCondBr(ExtraCase, EdgeBB, NewBB);
3715 else
3716 Builder.CreateCondBr(ExtraCase, NewBB, EdgeBB);
3717
3718 OldTI->eraseFromParent();
3719
3720 // If there are PHI nodes in EdgeBB, then we need to add a new entry to them
3721 // for the edge we just added.
3722 AddPredecessorToBlock(EdgeBB, BB, NewBB);
3723
3724 LLVM_DEBUG(dbgs() << " ** 'icmp' chain unhandled condition: " << *ExtraCase
3725 << "\nEXTRABB = " << *BB);
3726 BB = NewBB;
3727 }
3728
3729 Builder.SetInsertPoint(BI);
3730 // Convert pointer to int before we switch.
3731 if (CompVal->getType()->isPointerTy()) {
3732 CompVal = Builder.CreatePtrToInt(
3733 CompVal, DL.getIntPtrType(CompVal->getType()), "magicptr");
3734 }
3735
3736 // Create the new switch instruction now.
3737 SwitchInst *New = Builder.CreateSwitch(CompVal, DefaultBB, Values.size());
3738
3739 // Add all of the 'cases' to the switch instruction.
3740 for (unsigned i = 0, e = Values.size(); i != e; ++i)
3741 New->addCase(Values[i], EdgeBB);
3742
3743 // We added edges from PI to the EdgeBB. As such, if there were any
3744 // PHI nodes in EdgeBB, they need entries to be added corresponding to
3745 // the number of edges added.
3746 for (BasicBlock::iterator BBI = EdgeBB->begin(); isa<PHINode>(BBI); ++BBI) {
3747 PHINode *PN = cast<PHINode>(BBI);
3748 Value *InVal = PN->getIncomingValueForBlock(BB);
3749 for (unsigned i = 0, e = Values.size() - 1; i != e; ++i)
3750 PN->addIncoming(InVal, BB);
3751 }
3752
3753 // Erase the old branch instruction.
3754 EraseTerminatorAndDCECond(BI);
3755
3756 LLVM_DEBUG(dbgs() << " ** 'icmp' chain result is:\n" << *BB << '\n');
3757 return true;
3758 }
3759
SimplifyResume(ResumeInst * RI,IRBuilder<> & Builder)3760 bool SimplifyCFGOpt::SimplifyResume(ResumeInst *RI, IRBuilder<> &Builder) {
3761 if (isa<PHINode>(RI->getValue()))
3762 return SimplifyCommonResume(RI);
3763 else if (isa<LandingPadInst>(RI->getParent()->getFirstNonPHI()) &&
3764 RI->getValue() == RI->getParent()->getFirstNonPHI())
3765 // The resume must unwind the exception that caused control to branch here.
3766 return SimplifySingleResume(RI);
3767
3768 return false;
3769 }
3770
3771 // Simplify resume that is shared by several landing pads (phi of landing pad).
SimplifyCommonResume(ResumeInst * RI)3772 bool SimplifyCFGOpt::SimplifyCommonResume(ResumeInst *RI) {
3773 BasicBlock *BB = RI->getParent();
3774
3775 // Check that there are no other instructions except for debug intrinsics
3776 // between the phi of landing pads (RI->getValue()) and resume instruction.
3777 BasicBlock::iterator I = cast<Instruction>(RI->getValue())->getIterator(),
3778 E = RI->getIterator();
3779 while (++I != E)
3780 if (!isa<DbgInfoIntrinsic>(I))
3781 return false;
3782
3783 SmallSetVector<BasicBlock *, 4> TrivialUnwindBlocks;
3784 auto *PhiLPInst = cast<PHINode>(RI->getValue());
3785
3786 // Check incoming blocks to see if any of them are trivial.
3787 for (unsigned Idx = 0, End = PhiLPInst->getNumIncomingValues(); Idx != End;
3788 Idx++) {
3789 auto *IncomingBB = PhiLPInst->getIncomingBlock(Idx);
3790 auto *IncomingValue = PhiLPInst->getIncomingValue(Idx);
3791
3792 // If the block has other successors, we can not delete it because
3793 // it has other dependents.
3794 if (IncomingBB->getUniqueSuccessor() != BB)
3795 continue;
3796
3797 auto *LandingPad = dyn_cast<LandingPadInst>(IncomingBB->getFirstNonPHI());
3798 // Not the landing pad that caused the control to branch here.
3799 if (IncomingValue != LandingPad)
3800 continue;
3801
3802 bool isTrivial = true;
3803
3804 I = IncomingBB->getFirstNonPHI()->getIterator();
3805 E = IncomingBB->getTerminator()->getIterator();
3806 while (++I != E)
3807 if (!isa<DbgInfoIntrinsic>(I)) {
3808 isTrivial = false;
3809 break;
3810 }
3811
3812 if (isTrivial)
3813 TrivialUnwindBlocks.insert(IncomingBB);
3814 }
3815
3816 // If no trivial unwind blocks, don't do any simplifications.
3817 if (TrivialUnwindBlocks.empty())
3818 return false;
3819
3820 // Turn all invokes that unwind here into calls.
3821 for (auto *TrivialBB : TrivialUnwindBlocks) {
3822 // Blocks that will be simplified should be removed from the phi node.
3823 // Note there could be multiple edges to the resume block, and we need
3824 // to remove them all.
3825 while (PhiLPInst->getBasicBlockIndex(TrivialBB) != -1)
3826 BB->removePredecessor(TrivialBB, true);
3827
3828 for (pred_iterator PI = pred_begin(TrivialBB), PE = pred_end(TrivialBB);
3829 PI != PE;) {
3830 BasicBlock *Pred = *PI++;
3831 removeUnwindEdge(Pred);
3832 }
3833
3834 // In each SimplifyCFG run, only the current processed block can be erased.
3835 // Otherwise, it will break the iteration of SimplifyCFG pass. So instead
3836 // of erasing TrivialBB, we only remove the branch to the common resume
3837 // block so that we can later erase the resume block since it has no
3838 // predecessors.
3839 TrivialBB->getTerminator()->eraseFromParent();
3840 new UnreachableInst(RI->getContext(), TrivialBB);
3841 }
3842
3843 // Delete the resume block if all its predecessors have been removed.
3844 if (pred_empty(BB))
3845 BB->eraseFromParent();
3846
3847 return !TrivialUnwindBlocks.empty();
3848 }
3849
3850 // Simplify resume that is only used by a single (non-phi) landing pad.
SimplifySingleResume(ResumeInst * RI)3851 bool SimplifyCFGOpt::SimplifySingleResume(ResumeInst *RI) {
3852 BasicBlock *BB = RI->getParent();
3853 LandingPadInst *LPInst = dyn_cast<LandingPadInst>(BB->getFirstNonPHI());
3854 assert(RI->getValue() == LPInst &&
3855 "Resume must unwind the exception that caused control to here");
3856
3857 // Check that there are no other instructions except for debug intrinsics.
3858 BasicBlock::iterator I = LPInst->getIterator(), E = RI->getIterator();
3859 while (++I != E)
3860 if (!isa<DbgInfoIntrinsic>(I))
3861 return false;
3862
3863 // Turn all invokes that unwind here into calls and delete the basic block.
3864 for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE;) {
3865 BasicBlock *Pred = *PI++;
3866 removeUnwindEdge(Pred);
3867 }
3868
3869 // The landingpad is now unreachable. Zap it.
3870 if (LoopHeaders)
3871 LoopHeaders->erase(BB);
3872 BB->eraseFromParent();
3873 return true;
3874 }
3875
removeEmptyCleanup(CleanupReturnInst * RI)3876 static bool removeEmptyCleanup(CleanupReturnInst *RI) {
3877 // If this is a trivial cleanup pad that executes no instructions, it can be
3878 // eliminated. If the cleanup pad continues to the caller, any predecessor
3879 // that is an EH pad will be updated to continue to the caller and any
3880 // predecessor that terminates with an invoke instruction will have its invoke
3881 // instruction converted to a call instruction. If the cleanup pad being
3882 // simplified does not continue to the caller, each predecessor will be
3883 // updated to continue to the unwind destination of the cleanup pad being
3884 // simplified.
3885 BasicBlock *BB = RI->getParent();
3886 CleanupPadInst *CPInst = RI->getCleanupPad();
3887 if (CPInst->getParent() != BB)
3888 // This isn't an empty cleanup.
3889 return false;
3890
3891 // We cannot kill the pad if it has multiple uses. This typically arises
3892 // from unreachable basic blocks.
3893 if (!CPInst->hasOneUse())
3894 return false;
3895
3896 // Check that there are no other instructions except for benign intrinsics.
3897 BasicBlock::iterator I = CPInst->getIterator(), E = RI->getIterator();
3898 while (++I != E) {
3899 auto *II = dyn_cast<IntrinsicInst>(I);
3900 if (!II)
3901 return false;
3902
3903 Intrinsic::ID IntrinsicID = II->getIntrinsicID();
3904 switch (IntrinsicID) {
3905 case Intrinsic::dbg_declare:
3906 case Intrinsic::dbg_value:
3907 case Intrinsic::dbg_label:
3908 case Intrinsic::lifetime_end:
3909 break;
3910 default:
3911 return false;
3912 }
3913 }
3914
3915 // If the cleanup return we are simplifying unwinds to the caller, this will
3916 // set UnwindDest to nullptr.
3917 BasicBlock *UnwindDest = RI->getUnwindDest();
3918 Instruction *DestEHPad = UnwindDest ? UnwindDest->getFirstNonPHI() : nullptr;
3919
3920 // We're about to remove BB from the control flow. Before we do, sink any
3921 // PHINodes into the unwind destination. Doing this before changing the
3922 // control flow avoids some potentially slow checks, since we can currently
3923 // be certain that UnwindDest and BB have no common predecessors (since they
3924 // are both EH pads).
3925 if (UnwindDest) {
3926 // First, go through the PHI nodes in UnwindDest and update any nodes that
3927 // reference the block we are removing
3928 for (BasicBlock::iterator I = UnwindDest->begin(),
3929 IE = DestEHPad->getIterator();
3930 I != IE; ++I) {
3931 PHINode *DestPN = cast<PHINode>(I);
3932
3933 int Idx = DestPN->getBasicBlockIndex(BB);
3934 // Since BB unwinds to UnwindDest, it has to be in the PHI node.
3935 assert(Idx != -1);
3936 // This PHI node has an incoming value that corresponds to a control
3937 // path through the cleanup pad we are removing. If the incoming
3938 // value is in the cleanup pad, it must be a PHINode (because we
3939 // verified above that the block is otherwise empty). Otherwise, the
3940 // value is either a constant or a value that dominates the cleanup
3941 // pad being removed.
3942 //
3943 // Because BB and UnwindDest are both EH pads, all of their
3944 // predecessors must unwind to these blocks, and since no instruction
3945 // can have multiple unwind destinations, there will be no overlap in
3946 // incoming blocks between SrcPN and DestPN.
3947 Value *SrcVal = DestPN->getIncomingValue(Idx);
3948 PHINode *SrcPN = dyn_cast<PHINode>(SrcVal);
3949
3950 // Remove the entry for the block we are deleting.
3951 DestPN->removeIncomingValue(Idx, false);
3952
3953 if (SrcPN && SrcPN->getParent() == BB) {
3954 // If the incoming value was a PHI node in the cleanup pad we are
3955 // removing, we need to merge that PHI node's incoming values into
3956 // DestPN.
3957 for (unsigned SrcIdx = 0, SrcE = SrcPN->getNumIncomingValues();
3958 SrcIdx != SrcE; ++SrcIdx) {
3959 DestPN->addIncoming(SrcPN->getIncomingValue(SrcIdx),
3960 SrcPN->getIncomingBlock(SrcIdx));
3961 }
3962 } else {
3963 // Otherwise, the incoming value came from above BB and
3964 // so we can just reuse it. We must associate all of BB's
3965 // predecessors with this value.
3966 for (auto *pred : predecessors(BB)) {
3967 DestPN->addIncoming(SrcVal, pred);
3968 }
3969 }
3970 }
3971
3972 // Sink any remaining PHI nodes directly into UnwindDest.
3973 Instruction *InsertPt = DestEHPad;
3974 for (BasicBlock::iterator I = BB->begin(),
3975 IE = BB->getFirstNonPHI()->getIterator();
3976 I != IE;) {
3977 // The iterator must be incremented here because the instructions are
3978 // being moved to another block.
3979 PHINode *PN = cast<PHINode>(I++);
3980 if (PN->use_empty())
3981 // If the PHI node has no uses, just leave it. It will be erased
3982 // when we erase BB below.
3983 continue;
3984
3985 // Otherwise, sink this PHI node into UnwindDest.
3986 // Any predecessors to UnwindDest which are not already represented
3987 // must be back edges which inherit the value from the path through
3988 // BB. In this case, the PHI value must reference itself.
3989 for (auto *pred : predecessors(UnwindDest))
3990 if (pred != BB)
3991 PN->addIncoming(PN, pred);
3992 PN->moveBefore(InsertPt);
3993 }
3994 }
3995
3996 for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE;) {
3997 // The iterator must be updated here because we are removing this pred.
3998 BasicBlock *PredBB = *PI++;
3999 if (UnwindDest == nullptr) {
4000 removeUnwindEdge(PredBB);
4001 } else {
4002 Instruction *TI = PredBB->getTerminator();
4003 TI->replaceUsesOfWith(BB, UnwindDest);
4004 }
4005 }
4006
4007 // The cleanup pad is now unreachable. Zap it.
4008 BB->eraseFromParent();
4009 return true;
4010 }
4011
4012 // Try to merge two cleanuppads together.
mergeCleanupPad(CleanupReturnInst * RI)4013 static bool mergeCleanupPad(CleanupReturnInst *RI) {
4014 // Skip any cleanuprets which unwind to caller, there is nothing to merge
4015 // with.
4016 BasicBlock *UnwindDest = RI->getUnwindDest();
4017 if (!UnwindDest)
4018 return false;
4019
4020 // This cleanupret isn't the only predecessor of this cleanuppad, it wouldn't
4021 // be safe to merge without code duplication.
4022 if (UnwindDest->getSinglePredecessor() != RI->getParent())
4023 return false;
4024
4025 // Verify that our cleanuppad's unwind destination is another cleanuppad.
4026 auto *SuccessorCleanupPad = dyn_cast<CleanupPadInst>(&UnwindDest->front());
4027 if (!SuccessorCleanupPad)
4028 return false;
4029
4030 CleanupPadInst *PredecessorCleanupPad = RI->getCleanupPad();
4031 // Replace any uses of the successor cleanupad with the predecessor pad
4032 // The only cleanuppad uses should be this cleanupret, it's cleanupret and
4033 // funclet bundle operands.
4034 SuccessorCleanupPad->replaceAllUsesWith(PredecessorCleanupPad);
4035 // Remove the old cleanuppad.
4036 SuccessorCleanupPad->eraseFromParent();
4037 // Now, we simply replace the cleanupret with a branch to the unwind
4038 // destination.
4039 BranchInst::Create(UnwindDest, RI->getParent());
4040 RI->eraseFromParent();
4041
4042 return true;
4043 }
4044
SimplifyCleanupReturn(CleanupReturnInst * RI)4045 bool SimplifyCFGOpt::SimplifyCleanupReturn(CleanupReturnInst *RI) {
4046 // It is possible to transiantly have an undef cleanuppad operand because we
4047 // have deleted some, but not all, dead blocks.
4048 // Eventually, this block will be deleted.
4049 if (isa<UndefValue>(RI->getOperand(0)))
4050 return false;
4051
4052 if (mergeCleanupPad(RI))
4053 return true;
4054
4055 if (removeEmptyCleanup(RI))
4056 return true;
4057
4058 return false;
4059 }
4060
SimplifyReturn(ReturnInst * RI,IRBuilder<> & Builder)4061 bool SimplifyCFGOpt::SimplifyReturn(ReturnInst *RI, IRBuilder<> &Builder) {
4062 BasicBlock *BB = RI->getParent();
4063 if (!BB->getFirstNonPHIOrDbg()->isTerminator())
4064 return false;
4065
4066 // Find predecessors that end with branches.
4067 SmallVector<BasicBlock *, 8> UncondBranchPreds;
4068 SmallVector<BranchInst *, 8> CondBranchPreds;
4069 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
4070 BasicBlock *P = *PI;
4071 Instruction *PTI = P->getTerminator();
4072 if (BranchInst *BI = dyn_cast<BranchInst>(PTI)) {
4073 if (BI->isUnconditional())
4074 UncondBranchPreds.push_back(P);
4075 else
4076 CondBranchPreds.push_back(BI);
4077 }
4078 }
4079
4080 // If we found some, do the transformation!
4081 if (!UncondBranchPreds.empty() && DupRet) {
4082 while (!UncondBranchPreds.empty()) {
4083 BasicBlock *Pred = UncondBranchPreds.pop_back_val();
4084 LLVM_DEBUG(dbgs() << "FOLDING: " << *BB
4085 << "INTO UNCOND BRANCH PRED: " << *Pred);
4086 (void)FoldReturnIntoUncondBranch(RI, BB, Pred);
4087 }
4088
4089 // If we eliminated all predecessors of the block, delete the block now.
4090 if (pred_empty(BB)) {
4091 // We know there are no successors, so just nuke the block.
4092 if (LoopHeaders)
4093 LoopHeaders->erase(BB);
4094 BB->eraseFromParent();
4095 }
4096
4097 return true;
4098 }
4099
4100 // Check out all of the conditional branches going to this return
4101 // instruction. If any of them just select between returns, change the
4102 // branch itself into a select/return pair.
4103 while (!CondBranchPreds.empty()) {
4104 BranchInst *BI = CondBranchPreds.pop_back_val();
4105
4106 // Check to see if the non-BB successor is also a return block.
4107 if (isa<ReturnInst>(BI->getSuccessor(0)->getTerminator()) &&
4108 isa<ReturnInst>(BI->getSuccessor(1)->getTerminator()) &&
4109 SimplifyCondBranchToTwoReturns(BI, Builder))
4110 return true;
4111 }
4112 return false;
4113 }
4114
SimplifyUnreachable(UnreachableInst * UI)4115 bool SimplifyCFGOpt::SimplifyUnreachable(UnreachableInst *UI) {
4116 BasicBlock *BB = UI->getParent();
4117
4118 bool Changed = false;
4119
4120 // If there are any instructions immediately before the unreachable that can
4121 // be removed, do so.
4122 while (UI->getIterator() != BB->begin()) {
4123 BasicBlock::iterator BBI = UI->getIterator();
4124 --BBI;
4125 // Do not delete instructions that can have side effects which might cause
4126 // the unreachable to not be reachable; specifically, calls and volatile
4127 // operations may have this effect.
4128 if (isa<CallInst>(BBI) && !isa<DbgInfoIntrinsic>(BBI))
4129 break;
4130
4131 if (BBI->mayHaveSideEffects()) {
4132 if (auto *SI = dyn_cast<StoreInst>(BBI)) {
4133 if (SI->isVolatile())
4134 break;
4135 } else if (auto *LI = dyn_cast<LoadInst>(BBI)) {
4136 if (LI->isVolatile())
4137 break;
4138 } else if (auto *RMWI = dyn_cast<AtomicRMWInst>(BBI)) {
4139 if (RMWI->isVolatile())
4140 break;
4141 } else if (auto *CXI = dyn_cast<AtomicCmpXchgInst>(BBI)) {
4142 if (CXI->isVolatile())
4143 break;
4144 } else if (isa<CatchPadInst>(BBI)) {
4145 // A catchpad may invoke exception object constructors and such, which
4146 // in some languages can be arbitrary code, so be conservative by
4147 // default.
4148 // For CoreCLR, it just involves a type test, so can be removed.
4149 if (classifyEHPersonality(BB->getParent()->getPersonalityFn()) !=
4150 EHPersonality::CoreCLR)
4151 break;
4152 } else if (!isa<FenceInst>(BBI) && !isa<VAArgInst>(BBI) &&
4153 !isa<LandingPadInst>(BBI)) {
4154 break;
4155 }
4156 // Note that deleting LandingPad's here is in fact okay, although it
4157 // involves a bit of subtle reasoning. If this inst is a LandingPad,
4158 // all the predecessors of this block will be the unwind edges of Invokes,
4159 // and we can therefore guarantee this block will be erased.
4160 }
4161
4162 // Delete this instruction (any uses are guaranteed to be dead)
4163 if (!BBI->use_empty())
4164 BBI->replaceAllUsesWith(UndefValue::get(BBI->getType()));
4165 BBI->eraseFromParent();
4166 Changed = true;
4167 }
4168
4169 // If the unreachable instruction is the first in the block, take a gander
4170 // at all of the predecessors of this instruction, and simplify them.
4171 if (&BB->front() != UI)
4172 return Changed;
4173
4174 SmallVector<BasicBlock *, 8> Preds(pred_begin(BB), pred_end(BB));
4175 for (unsigned i = 0, e = Preds.size(); i != e; ++i) {
4176 Instruction *TI = Preds[i]->getTerminator();
4177 IRBuilder<> Builder(TI);
4178 if (auto *BI = dyn_cast<BranchInst>(TI)) {
4179 if (BI->isUnconditional()) {
4180 if (BI->getSuccessor(0) == BB) {
4181 new UnreachableInst(TI->getContext(), TI);
4182 TI->eraseFromParent();
4183 Changed = true;
4184 }
4185 } else {
4186 Value* Cond = BI->getCondition();
4187 if (BI->getSuccessor(0) == BB) {
4188 Builder.CreateAssumption(Builder.CreateNot(Cond));
4189 Builder.CreateBr(BI->getSuccessor(1));
4190 EraseTerminatorAndDCECond(BI);
4191 } else if (BI->getSuccessor(1) == BB) {
4192 Builder.CreateAssumption(Cond);
4193 Builder.CreateBr(BI->getSuccessor(0));
4194 EraseTerminatorAndDCECond(BI);
4195 Changed = true;
4196 }
4197 }
4198 } else if (auto *SI = dyn_cast<SwitchInst>(TI)) {
4199 SwitchInstProfUpdateWrapper SU(*SI);
4200 for (auto i = SU->case_begin(), e = SU->case_end(); i != e;) {
4201 if (i->getCaseSuccessor() != BB) {
4202 ++i;
4203 continue;
4204 }
4205 BB->removePredecessor(SU->getParent());
4206 i = SU.removeCase(i);
4207 e = SU->case_end();
4208 Changed = true;
4209 }
4210 } else if (auto *II = dyn_cast<InvokeInst>(TI)) {
4211 if (II->getUnwindDest() == BB) {
4212 removeUnwindEdge(TI->getParent());
4213 Changed = true;
4214 }
4215 } else if (auto *CSI = dyn_cast<CatchSwitchInst>(TI)) {
4216 if (CSI->getUnwindDest() == BB) {
4217 removeUnwindEdge(TI->getParent());
4218 Changed = true;
4219 continue;
4220 }
4221
4222 for (CatchSwitchInst::handler_iterator I = CSI->handler_begin(),
4223 E = CSI->handler_end();
4224 I != E; ++I) {
4225 if (*I == BB) {
4226 CSI->removeHandler(I);
4227 --I;
4228 --E;
4229 Changed = true;
4230 }
4231 }
4232 if (CSI->getNumHandlers() == 0) {
4233 BasicBlock *CatchSwitchBB = CSI->getParent();
4234 if (CSI->hasUnwindDest()) {
4235 // Redirect preds to the unwind dest
4236 CatchSwitchBB->replaceAllUsesWith(CSI->getUnwindDest());
4237 } else {
4238 // Rewrite all preds to unwind to caller (or from invoke to call).
4239 SmallVector<BasicBlock *, 8> EHPreds(predecessors(CatchSwitchBB));
4240 for (BasicBlock *EHPred : EHPreds)
4241 removeUnwindEdge(EHPred);
4242 }
4243 // The catchswitch is no longer reachable.
4244 new UnreachableInst(CSI->getContext(), CSI);
4245 CSI->eraseFromParent();
4246 Changed = true;
4247 }
4248 } else if (isa<CleanupReturnInst>(TI)) {
4249 new UnreachableInst(TI->getContext(), TI);
4250 TI->eraseFromParent();
4251 Changed = true;
4252 }
4253 }
4254
4255 // If this block is now dead, remove it.
4256 if (pred_empty(BB) && BB != &BB->getParent()->getEntryBlock()) {
4257 // We know there are no successors, so just nuke the block.
4258 if (LoopHeaders)
4259 LoopHeaders->erase(BB);
4260 BB->eraseFromParent();
4261 return true;
4262 }
4263
4264 return Changed;
4265 }
4266
CasesAreContiguous(SmallVectorImpl<ConstantInt * > & Cases)4267 static bool CasesAreContiguous(SmallVectorImpl<ConstantInt *> &Cases) {
4268 assert(Cases.size() >= 1);
4269
4270 array_pod_sort(Cases.begin(), Cases.end(), ConstantIntSortPredicate);
4271 for (size_t I = 1, E = Cases.size(); I != E; ++I) {
4272 if (Cases[I - 1]->getValue() != Cases[I]->getValue() + 1)
4273 return false;
4274 }
4275 return true;
4276 }
4277
4278 /// Turn a switch with two reachable destinations into an integer range
4279 /// comparison and branch.
TurnSwitchRangeIntoICmp(SwitchInst * SI,IRBuilder<> & Builder)4280 static bool TurnSwitchRangeIntoICmp(SwitchInst *SI, IRBuilder<> &Builder) {
4281 assert(SI->getNumCases() > 1 && "Degenerate switch?");
4282
4283 bool HasDefault =
4284 !isa<UnreachableInst>(SI->getDefaultDest()->getFirstNonPHIOrDbg());
4285
4286 // Partition the cases into two sets with different destinations.
4287 BasicBlock *DestA = HasDefault ? SI->getDefaultDest() : nullptr;
4288 BasicBlock *DestB = nullptr;
4289 SmallVector<ConstantInt *, 16> CasesA;
4290 SmallVector<ConstantInt *, 16> CasesB;
4291
4292 for (auto Case : SI->cases()) {
4293 BasicBlock *Dest = Case.getCaseSuccessor();
4294 if (!DestA)
4295 DestA = Dest;
4296 if (Dest == DestA) {
4297 CasesA.push_back(Case.getCaseValue());
4298 continue;
4299 }
4300 if (!DestB)
4301 DestB = Dest;
4302 if (Dest == DestB) {
4303 CasesB.push_back(Case.getCaseValue());
4304 continue;
4305 }
4306 return false; // More than two destinations.
4307 }
4308
4309 assert(DestA && DestB &&
4310 "Single-destination switch should have been folded.");
4311 assert(DestA != DestB);
4312 assert(DestB != SI->getDefaultDest());
4313 assert(!CasesB.empty() && "There must be non-default cases.");
4314 assert(!CasesA.empty() || HasDefault);
4315
4316 // Figure out if one of the sets of cases form a contiguous range.
4317 SmallVectorImpl<ConstantInt *> *ContiguousCases = nullptr;
4318 BasicBlock *ContiguousDest = nullptr;
4319 BasicBlock *OtherDest = nullptr;
4320 if (!CasesA.empty() && CasesAreContiguous(CasesA)) {
4321 ContiguousCases = &CasesA;
4322 ContiguousDest = DestA;
4323 OtherDest = DestB;
4324 } else if (CasesAreContiguous(CasesB)) {
4325 ContiguousCases = &CasesB;
4326 ContiguousDest = DestB;
4327 OtherDest = DestA;
4328 } else
4329 return false;
4330
4331 // Start building the compare and branch.
4332
4333 Constant *Offset = ConstantExpr::getNeg(ContiguousCases->back());
4334 Constant *NumCases =
4335 ConstantInt::get(Offset->getType(), ContiguousCases->size());
4336
4337 Value *Sub = SI->getCondition();
4338 if (!Offset->isNullValue())
4339 Sub = Builder.CreateAdd(Sub, Offset, Sub->getName() + ".off");
4340
4341 Value *Cmp;
4342 // If NumCases overflowed, then all possible values jump to the successor.
4343 if (NumCases->isNullValue() && !ContiguousCases->empty())
4344 Cmp = ConstantInt::getTrue(SI->getContext());
4345 else
4346 Cmp = Builder.CreateICmpULT(Sub, NumCases, "switch");
4347 BranchInst *NewBI = Builder.CreateCondBr(Cmp, ContiguousDest, OtherDest);
4348
4349 // Update weight for the newly-created conditional branch.
4350 if (HasBranchWeights(SI)) {
4351 SmallVector<uint64_t, 8> Weights;
4352 GetBranchWeights(SI, Weights);
4353 if (Weights.size() == 1 + SI->getNumCases()) {
4354 uint64_t TrueWeight = 0;
4355 uint64_t FalseWeight = 0;
4356 for (size_t I = 0, E = Weights.size(); I != E; ++I) {
4357 if (SI->getSuccessor(I) == ContiguousDest)
4358 TrueWeight += Weights[I];
4359 else
4360 FalseWeight += Weights[I];
4361 }
4362 while (TrueWeight > UINT32_MAX || FalseWeight > UINT32_MAX) {
4363 TrueWeight /= 2;
4364 FalseWeight /= 2;
4365 }
4366 setBranchWeights(NewBI, TrueWeight, FalseWeight);
4367 }
4368 }
4369
4370 // Prune obsolete incoming values off the successors' PHI nodes.
4371 for (auto BBI = ContiguousDest->begin(); isa<PHINode>(BBI); ++BBI) {
4372 unsigned PreviousEdges = ContiguousCases->size();
4373 if (ContiguousDest == SI->getDefaultDest())
4374 ++PreviousEdges;
4375 for (unsigned I = 0, E = PreviousEdges - 1; I != E; ++I)
4376 cast<PHINode>(BBI)->removeIncomingValue(SI->getParent());
4377 }
4378 for (auto BBI = OtherDest->begin(); isa<PHINode>(BBI); ++BBI) {
4379 unsigned PreviousEdges = SI->getNumCases() - ContiguousCases->size();
4380 if (OtherDest == SI->getDefaultDest())
4381 ++PreviousEdges;
4382 for (unsigned I = 0, E = PreviousEdges - 1; I != E; ++I)
4383 cast<PHINode>(BBI)->removeIncomingValue(SI->getParent());
4384 }
4385
4386 // Drop the switch.
4387 SI->eraseFromParent();
4388
4389 return true;
4390 }
4391
4392 /// Compute masked bits for the condition of a switch
4393 /// and use it to remove dead cases.
eliminateDeadSwitchCases(SwitchInst * SI,AssumptionCache * AC,const DataLayout & DL)4394 static bool eliminateDeadSwitchCases(SwitchInst *SI, AssumptionCache *AC,
4395 const DataLayout &DL) {
4396 Value *Cond = SI->getCondition();
4397 unsigned Bits = Cond->getType()->getIntegerBitWidth();
4398 KnownBits Known = computeKnownBits(Cond, DL, 0, AC, SI);
4399
4400 // We can also eliminate cases by determining that their values are outside of
4401 // the limited range of the condition based on how many significant (non-sign)
4402 // bits are in the condition value.
4403 unsigned ExtraSignBits = ComputeNumSignBits(Cond, DL, 0, AC, SI) - 1;
4404 unsigned MaxSignificantBitsInCond = Bits - ExtraSignBits;
4405
4406 // Gather dead cases.
4407 SmallVector<ConstantInt *, 8> DeadCases;
4408 for (auto &Case : SI->cases()) {
4409 const APInt &CaseVal = Case.getCaseValue()->getValue();
4410 if (Known.Zero.intersects(CaseVal) || !Known.One.isSubsetOf(CaseVal) ||
4411 (CaseVal.getMinSignedBits() > MaxSignificantBitsInCond)) {
4412 DeadCases.push_back(Case.getCaseValue());
4413 LLVM_DEBUG(dbgs() << "SimplifyCFG: switch case " << CaseVal
4414 << " is dead.\n");
4415 }
4416 }
4417
4418 // If we can prove that the cases must cover all possible values, the
4419 // default destination becomes dead and we can remove it. If we know some
4420 // of the bits in the value, we can use that to more precisely compute the
4421 // number of possible unique case values.
4422 bool HasDefault =
4423 !isa<UnreachableInst>(SI->getDefaultDest()->getFirstNonPHIOrDbg());
4424 const unsigned NumUnknownBits =
4425 Bits - (Known.Zero | Known.One).countPopulation();
4426 assert(NumUnknownBits <= Bits);
4427 if (HasDefault && DeadCases.empty() &&
4428 NumUnknownBits < 64 /* avoid overflow */ &&
4429 SI->getNumCases() == (1ULL << NumUnknownBits)) {
4430 LLVM_DEBUG(dbgs() << "SimplifyCFG: switch default is dead.\n");
4431 BasicBlock *NewDefault =
4432 SplitBlockPredecessors(SI->getDefaultDest(), SI->getParent(), "");
4433 SI->setDefaultDest(&*NewDefault);
4434 SplitBlock(&*NewDefault, &NewDefault->front());
4435 auto *OldTI = NewDefault->getTerminator();
4436 new UnreachableInst(SI->getContext(), OldTI);
4437 EraseTerminatorAndDCECond(OldTI);
4438 return true;
4439 }
4440
4441 if (DeadCases.empty())
4442 return false;
4443
4444 SwitchInstProfUpdateWrapper SIW(*SI);
4445 for (ConstantInt *DeadCase : DeadCases) {
4446 SwitchInst::CaseIt CaseI = SI->findCaseValue(DeadCase);
4447 assert(CaseI != SI->case_default() &&
4448 "Case was not found. Probably mistake in DeadCases forming.");
4449 // Prune unused values from PHI nodes.
4450 CaseI->getCaseSuccessor()->removePredecessor(SI->getParent());
4451 SIW.removeCase(CaseI);
4452 }
4453
4454 return true;
4455 }
4456
4457 /// If BB would be eligible for simplification by
4458 /// TryToSimplifyUncondBranchFromEmptyBlock (i.e. it is empty and terminated
4459 /// by an unconditional branch), look at the phi node for BB in the successor
4460 /// block and see if the incoming value is equal to CaseValue. If so, return
4461 /// the phi node, and set PhiIndex to BB's index in the phi node.
FindPHIForConditionForwarding(ConstantInt * CaseValue,BasicBlock * BB,int * PhiIndex)4462 static PHINode *FindPHIForConditionForwarding(ConstantInt *CaseValue,
4463 BasicBlock *BB, int *PhiIndex) {
4464 if (BB->getFirstNonPHIOrDbg() != BB->getTerminator())
4465 return nullptr; // BB must be empty to be a candidate for simplification.
4466 if (!BB->getSinglePredecessor())
4467 return nullptr; // BB must be dominated by the switch.
4468
4469 BranchInst *Branch = dyn_cast<BranchInst>(BB->getTerminator());
4470 if (!Branch || !Branch->isUnconditional())
4471 return nullptr; // Terminator must be unconditional branch.
4472
4473 BasicBlock *Succ = Branch->getSuccessor(0);
4474
4475 for (PHINode &PHI : Succ->phis()) {
4476 int Idx = PHI.getBasicBlockIndex(BB);
4477 assert(Idx >= 0 && "PHI has no entry for predecessor?");
4478
4479 Value *InValue = PHI.getIncomingValue(Idx);
4480 if (InValue != CaseValue)
4481 continue;
4482
4483 *PhiIndex = Idx;
4484 return &PHI;
4485 }
4486
4487 return nullptr;
4488 }
4489
4490 /// Try to forward the condition of a switch instruction to a phi node
4491 /// dominated by the switch, if that would mean that some of the destination
4492 /// blocks of the switch can be folded away. Return true if a change is made.
ForwardSwitchConditionToPHI(SwitchInst * SI)4493 static bool ForwardSwitchConditionToPHI(SwitchInst *SI) {
4494 using ForwardingNodesMap = DenseMap<PHINode *, SmallVector<int, 4>>;
4495
4496 ForwardingNodesMap ForwardingNodes;
4497 BasicBlock *SwitchBlock = SI->getParent();
4498 bool Changed = false;
4499 for (auto &Case : SI->cases()) {
4500 ConstantInt *CaseValue = Case.getCaseValue();
4501 BasicBlock *CaseDest = Case.getCaseSuccessor();
4502
4503 // Replace phi operands in successor blocks that are using the constant case
4504 // value rather than the switch condition variable:
4505 // switchbb:
4506 // switch i32 %x, label %default [
4507 // i32 17, label %succ
4508 // ...
4509 // succ:
4510 // %r = phi i32 ... [ 17, %switchbb ] ...
4511 // -->
4512 // %r = phi i32 ... [ %x, %switchbb ] ...
4513
4514 for (PHINode &Phi : CaseDest->phis()) {
4515 // This only works if there is exactly 1 incoming edge from the switch to
4516 // a phi. If there is >1, that means multiple cases of the switch map to 1
4517 // value in the phi, and that phi value is not the switch condition. Thus,
4518 // this transform would not make sense (the phi would be invalid because
4519 // a phi can't have different incoming values from the same block).
4520 int SwitchBBIdx = Phi.getBasicBlockIndex(SwitchBlock);
4521 if (Phi.getIncomingValue(SwitchBBIdx) == CaseValue &&
4522 count(Phi.blocks(), SwitchBlock) == 1) {
4523 Phi.setIncomingValue(SwitchBBIdx, SI->getCondition());
4524 Changed = true;
4525 }
4526 }
4527
4528 // Collect phi nodes that are indirectly using this switch's case constants.
4529 int PhiIdx;
4530 if (auto *Phi = FindPHIForConditionForwarding(CaseValue, CaseDest, &PhiIdx))
4531 ForwardingNodes[Phi].push_back(PhiIdx);
4532 }
4533
4534 for (auto &ForwardingNode : ForwardingNodes) {
4535 PHINode *Phi = ForwardingNode.first;
4536 SmallVectorImpl<int> &Indexes = ForwardingNode.second;
4537 if (Indexes.size() < 2)
4538 continue;
4539
4540 for (int Index : Indexes)
4541 Phi->setIncomingValue(Index, SI->getCondition());
4542 Changed = true;
4543 }
4544
4545 return Changed;
4546 }
4547
4548 /// Return true if the backend will be able to handle
4549 /// initializing an array of constants like C.
ValidLookupTableConstant(Constant * C,const TargetTransformInfo & TTI)4550 static bool ValidLookupTableConstant(Constant *C, const TargetTransformInfo &TTI) {
4551 if (C->isThreadDependent())
4552 return false;
4553 if (C->isDLLImportDependent())
4554 return false;
4555
4556 if (!isa<ConstantFP>(C) && !isa<ConstantInt>(C) &&
4557 !isa<ConstantPointerNull>(C) && !isa<GlobalValue>(C) &&
4558 !isa<UndefValue>(C) && !isa<ConstantExpr>(C))
4559 return false;
4560
4561 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) {
4562 if (!CE->isGEPWithNoNotionalOverIndexing())
4563 return false;
4564 if (!ValidLookupTableConstant(CE->getOperand(0), TTI))
4565 return false;
4566 }
4567
4568 if (!TTI.shouldBuildLookupTablesForConstant(C))
4569 return false;
4570
4571 return true;
4572 }
4573
4574 /// If V is a Constant, return it. Otherwise, try to look up
4575 /// its constant value in ConstantPool, returning 0 if it's not there.
4576 static Constant *
LookupConstant(Value * V,const SmallDenseMap<Value *,Constant * > & ConstantPool)4577 LookupConstant(Value *V,
4578 const SmallDenseMap<Value *, Constant *> &ConstantPool) {
4579 if (Constant *C = dyn_cast<Constant>(V))
4580 return C;
4581 return ConstantPool.lookup(V);
4582 }
4583
4584 /// Try to fold instruction I into a constant. This works for
4585 /// simple instructions such as binary operations where both operands are
4586 /// constant or can be replaced by constants from the ConstantPool. Returns the
4587 /// resulting constant on success, 0 otherwise.
4588 static Constant *
ConstantFold(Instruction * I,const DataLayout & DL,const SmallDenseMap<Value *,Constant * > & ConstantPool)4589 ConstantFold(Instruction *I, const DataLayout &DL,
4590 const SmallDenseMap<Value *, Constant *> &ConstantPool) {
4591 if (SelectInst *Select = dyn_cast<SelectInst>(I)) {
4592 Constant *A = LookupConstant(Select->getCondition(), ConstantPool);
4593 if (!A)
4594 return nullptr;
4595 if (A->isAllOnesValue())
4596 return LookupConstant(Select->getTrueValue(), ConstantPool);
4597 if (A->isNullValue())
4598 return LookupConstant(Select->getFalseValue(), ConstantPool);
4599 return nullptr;
4600 }
4601
4602 SmallVector<Constant *, 4> COps;
4603 for (unsigned N = 0, E = I->getNumOperands(); N != E; ++N) {
4604 if (Constant *A = LookupConstant(I->getOperand(N), ConstantPool))
4605 COps.push_back(A);
4606 else
4607 return nullptr;
4608 }
4609
4610 if (CmpInst *Cmp = dyn_cast<CmpInst>(I)) {
4611 return ConstantFoldCompareInstOperands(Cmp->getPredicate(), COps[0],
4612 COps[1], DL);
4613 }
4614
4615 return ConstantFoldInstOperands(I, COps, DL);
4616 }
4617
4618 /// Try to determine the resulting constant values in phi nodes
4619 /// at the common destination basic block, *CommonDest, for one of the case
4620 /// destionations CaseDest corresponding to value CaseVal (0 for the default
4621 /// case), of a switch instruction SI.
4622 static bool
GetCaseResults(SwitchInst * SI,ConstantInt * CaseVal,BasicBlock * CaseDest,BasicBlock ** CommonDest,SmallVectorImpl<std::pair<PHINode *,Constant * >> & Res,const DataLayout & DL,const TargetTransformInfo & TTI)4623 GetCaseResults(SwitchInst *SI, ConstantInt *CaseVal, BasicBlock *CaseDest,
4624 BasicBlock **CommonDest,
4625 SmallVectorImpl<std::pair<PHINode *, Constant *>> &Res,
4626 const DataLayout &DL, const TargetTransformInfo &TTI) {
4627 // The block from which we enter the common destination.
4628 BasicBlock *Pred = SI->getParent();
4629
4630 // If CaseDest is empty except for some side-effect free instructions through
4631 // which we can constant-propagate the CaseVal, continue to its successor.
4632 SmallDenseMap<Value *, Constant *> ConstantPool;
4633 ConstantPool.insert(std::make_pair(SI->getCondition(), CaseVal));
4634 for (Instruction &I :CaseDest->instructionsWithoutDebug()) {
4635 if (I.isTerminator()) {
4636 // If the terminator is a simple branch, continue to the next block.
4637 if (I.getNumSuccessors() != 1 || I.isExceptionalTerminator())
4638 return false;
4639 Pred = CaseDest;
4640 CaseDest = I.getSuccessor(0);
4641 } else if (Constant *C = ConstantFold(&I, DL, ConstantPool)) {
4642 // Instruction is side-effect free and constant.
4643
4644 // If the instruction has uses outside this block or a phi node slot for
4645 // the block, it is not safe to bypass the instruction since it would then
4646 // no longer dominate all its uses.
4647 for (auto &Use : I.uses()) {
4648 User *User = Use.getUser();
4649 if (Instruction *I = dyn_cast<Instruction>(User))
4650 if (I->getParent() == CaseDest)
4651 continue;
4652 if (PHINode *Phi = dyn_cast<PHINode>(User))
4653 if (Phi->getIncomingBlock(Use) == CaseDest)
4654 continue;
4655 return false;
4656 }
4657
4658 ConstantPool.insert(std::make_pair(&I, C));
4659 } else {
4660 break;
4661 }
4662 }
4663
4664 // If we did not have a CommonDest before, use the current one.
4665 if (!*CommonDest)
4666 *CommonDest = CaseDest;
4667 // If the destination isn't the common one, abort.
4668 if (CaseDest != *CommonDest)
4669 return false;
4670
4671 // Get the values for this case from phi nodes in the destination block.
4672 for (PHINode &PHI : (*CommonDest)->phis()) {
4673 int Idx = PHI.getBasicBlockIndex(Pred);
4674 if (Idx == -1)
4675 continue;
4676
4677 Constant *ConstVal =
4678 LookupConstant(PHI.getIncomingValue(Idx), ConstantPool);
4679 if (!ConstVal)
4680 return false;
4681
4682 // Be conservative about which kinds of constants we support.
4683 if (!ValidLookupTableConstant(ConstVal, TTI))
4684 return false;
4685
4686 Res.push_back(std::make_pair(&PHI, ConstVal));
4687 }
4688
4689 return Res.size() > 0;
4690 }
4691
4692 // Helper function used to add CaseVal to the list of cases that generate
4693 // Result. Returns the updated number of cases that generate this result.
MapCaseToResult(ConstantInt * CaseVal,SwitchCaseResultVectorTy & UniqueResults,Constant * Result)4694 static uintptr_t MapCaseToResult(ConstantInt *CaseVal,
4695 SwitchCaseResultVectorTy &UniqueResults,
4696 Constant *Result) {
4697 for (auto &I : UniqueResults) {
4698 if (I.first == Result) {
4699 I.second.push_back(CaseVal);
4700 return I.second.size();
4701 }
4702 }
4703 UniqueResults.push_back(
4704 std::make_pair(Result, SmallVector<ConstantInt *, 4>(1, CaseVal)));
4705 return 1;
4706 }
4707
4708 // Helper function that initializes a map containing
4709 // results for the PHI node of the common destination block for a switch
4710 // instruction. Returns false if multiple PHI nodes have been found or if
4711 // there is not a common destination block for the switch.
4712 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)4713 InitializeUniqueCases(SwitchInst *SI, PHINode *&PHI, BasicBlock *&CommonDest,
4714 SwitchCaseResultVectorTy &UniqueResults,
4715 Constant *&DefaultResult, const DataLayout &DL,
4716 const TargetTransformInfo &TTI,
4717 uintptr_t MaxUniqueResults, uintptr_t MaxCasesPerResult) {
4718 for (auto &I : SI->cases()) {
4719 ConstantInt *CaseVal = I.getCaseValue();
4720
4721 // Resulting value at phi nodes for this case value.
4722 SwitchCaseResultsTy Results;
4723 if (!GetCaseResults(SI, CaseVal, I.getCaseSuccessor(), &CommonDest, Results,
4724 DL, TTI))
4725 return false;
4726
4727 // Only one value per case is permitted.
4728 if (Results.size() > 1)
4729 return false;
4730
4731 // Add the case->result mapping to UniqueResults.
4732 const uintptr_t NumCasesForResult =
4733 MapCaseToResult(CaseVal, UniqueResults, Results.begin()->second);
4734
4735 // Early out if there are too many cases for this result.
4736 if (NumCasesForResult > MaxCasesPerResult)
4737 return false;
4738
4739 // Early out if there are too many unique results.
4740 if (UniqueResults.size() > MaxUniqueResults)
4741 return false;
4742
4743 // Check the PHI consistency.
4744 if (!PHI)
4745 PHI = Results[0].first;
4746 else if (PHI != Results[0].first)
4747 return false;
4748 }
4749 // Find the default result value.
4750 SmallVector<std::pair<PHINode *, Constant *>, 1> DefaultResults;
4751 BasicBlock *DefaultDest = SI->getDefaultDest();
4752 GetCaseResults(SI, nullptr, SI->getDefaultDest(), &CommonDest, DefaultResults,
4753 DL, TTI);
4754 // If the default value is not found abort unless the default destination
4755 // is unreachable.
4756 DefaultResult =
4757 DefaultResults.size() == 1 ? DefaultResults.begin()->second : nullptr;
4758 if ((!DefaultResult &&
4759 !isa<UnreachableInst>(DefaultDest->getFirstNonPHIOrDbg())))
4760 return false;
4761
4762 return true;
4763 }
4764
4765 // Helper function that checks if it is possible to transform a switch with only
4766 // two cases (or two cases + default) that produces a result into a select.
4767 // Example:
4768 // switch (a) {
4769 // case 10: %0 = icmp eq i32 %a, 10
4770 // return 10; %1 = select i1 %0, i32 10, i32 4
4771 // case 20: ----> %2 = icmp eq i32 %a, 20
4772 // return 2; %3 = select i1 %2, i32 2, i32 %1
4773 // default:
4774 // return 4;
4775 // }
ConvertTwoCaseSwitch(const SwitchCaseResultVectorTy & ResultVector,Constant * DefaultResult,Value * Condition,IRBuilder<> & Builder)4776 static Value *ConvertTwoCaseSwitch(const SwitchCaseResultVectorTy &ResultVector,
4777 Constant *DefaultResult, Value *Condition,
4778 IRBuilder<> &Builder) {
4779 assert(ResultVector.size() == 2 &&
4780 "We should have exactly two unique results at this point");
4781 // If we are selecting between only two cases transform into a simple
4782 // select or a two-way select if default is possible.
4783 if (ResultVector[0].second.size() == 1 &&
4784 ResultVector[1].second.size() == 1) {
4785 ConstantInt *const FirstCase = ResultVector[0].second[0];
4786 ConstantInt *const SecondCase = ResultVector[1].second[0];
4787
4788 bool DefaultCanTrigger = DefaultResult;
4789 Value *SelectValue = ResultVector[1].first;
4790 if (DefaultCanTrigger) {
4791 Value *const ValueCompare =
4792 Builder.CreateICmpEQ(Condition, SecondCase, "switch.selectcmp");
4793 SelectValue = Builder.CreateSelect(ValueCompare, ResultVector[1].first,
4794 DefaultResult, "switch.select");
4795 }
4796 Value *const ValueCompare =
4797 Builder.CreateICmpEQ(Condition, FirstCase, "switch.selectcmp");
4798 return Builder.CreateSelect(ValueCompare, ResultVector[0].first,
4799 SelectValue, "switch.select");
4800 }
4801
4802 return nullptr;
4803 }
4804
4805 // Helper function to cleanup a switch instruction that has been converted into
4806 // a select, fixing up PHI nodes and basic blocks.
RemoveSwitchAfterSelectConversion(SwitchInst * SI,PHINode * PHI,Value * SelectValue,IRBuilder<> & Builder)4807 static void RemoveSwitchAfterSelectConversion(SwitchInst *SI, PHINode *PHI,
4808 Value *SelectValue,
4809 IRBuilder<> &Builder) {
4810 BasicBlock *SelectBB = SI->getParent();
4811 while (PHI->getBasicBlockIndex(SelectBB) >= 0)
4812 PHI->removeIncomingValue(SelectBB);
4813 PHI->addIncoming(SelectValue, SelectBB);
4814
4815 Builder.CreateBr(PHI->getParent());
4816
4817 // Remove the switch.
4818 for (unsigned i = 0, e = SI->getNumSuccessors(); i < e; ++i) {
4819 BasicBlock *Succ = SI->getSuccessor(i);
4820
4821 if (Succ == PHI->getParent())
4822 continue;
4823 Succ->removePredecessor(SelectBB);
4824 }
4825 SI->eraseFromParent();
4826 }
4827
4828 /// If the switch is only used to initialize one or more
4829 /// phi nodes in a common successor block with only two different
4830 /// constant values, replace the switch with select.
switchToSelect(SwitchInst * SI,IRBuilder<> & Builder,const DataLayout & DL,const TargetTransformInfo & TTI)4831 static bool switchToSelect(SwitchInst *SI, IRBuilder<> &Builder,
4832 const DataLayout &DL,
4833 const TargetTransformInfo &TTI) {
4834 Value *const Cond = SI->getCondition();
4835 PHINode *PHI = nullptr;
4836 BasicBlock *CommonDest = nullptr;
4837 Constant *DefaultResult;
4838 SwitchCaseResultVectorTy UniqueResults;
4839 // Collect all the cases that will deliver the same value from the switch.
4840 if (!InitializeUniqueCases(SI, PHI, CommonDest, UniqueResults, DefaultResult,
4841 DL, TTI, 2, 1))
4842 return false;
4843 // Selects choose between maximum two values.
4844 if (UniqueResults.size() != 2)
4845 return false;
4846 assert(PHI != nullptr && "PHI for value select not found");
4847
4848 Builder.SetInsertPoint(SI);
4849 Value *SelectValue =
4850 ConvertTwoCaseSwitch(UniqueResults, DefaultResult, Cond, Builder);
4851 if (SelectValue) {
4852 RemoveSwitchAfterSelectConversion(SI, PHI, SelectValue, Builder);
4853 return true;
4854 }
4855 // The switch couldn't be converted into a select.
4856 return false;
4857 }
4858
4859 namespace {
4860
4861 /// This class represents a lookup table that can be used to replace a switch.
4862 class SwitchLookupTable {
4863 public:
4864 /// Create a lookup table to use as a switch replacement with the contents
4865 /// of Values, using DefaultValue to fill any holes in the table.
4866 SwitchLookupTable(
4867 Module &M, uint64_t TableSize, ConstantInt *Offset,
4868 const SmallVectorImpl<std::pair<ConstantInt *, Constant *>> &Values,
4869 Constant *DefaultValue, const DataLayout &DL, const StringRef &FuncName);
4870
4871 /// Build instructions with Builder to retrieve the value at
4872 /// the position given by Index in the lookup table.
4873 Value *BuildLookup(Value *Index, IRBuilder<> &Builder);
4874
4875 /// Return true if a table with TableSize elements of
4876 /// type ElementType would fit in a target-legal register.
4877 static bool WouldFitInRegister(const DataLayout &DL, uint64_t TableSize,
4878 Type *ElementType);
4879
4880 private:
4881 // Depending on the contents of the table, it can be represented in
4882 // different ways.
4883 enum {
4884 // For tables where each element contains the same value, we just have to
4885 // store that single value and return it for each lookup.
4886 SingleValueKind,
4887
4888 // For tables where there is a linear relationship between table index
4889 // and values. We calculate the result with a simple multiplication
4890 // and addition instead of a table lookup.
4891 LinearMapKind,
4892
4893 // For small tables with integer elements, we can pack them into a bitmap
4894 // that fits into a target-legal register. Values are retrieved by
4895 // shift and mask operations.
4896 BitMapKind,
4897
4898 // The table is stored as an array of values. Values are retrieved by load
4899 // instructions from the table.
4900 ArrayKind
4901 } Kind;
4902
4903 // For SingleValueKind, this is the single value.
4904 Constant *SingleValue = nullptr;
4905
4906 // For BitMapKind, this is the bitmap.
4907 ConstantInt *BitMap = nullptr;
4908 IntegerType *BitMapElementTy = nullptr;
4909
4910 // For LinearMapKind, these are the constants used to derive the value.
4911 ConstantInt *LinearOffset = nullptr;
4912 ConstantInt *LinearMultiplier = nullptr;
4913
4914 // For ArrayKind, this is the array.
4915 GlobalVariable *Array = nullptr;
4916 };
4917
4918 } // end anonymous namespace
4919
SwitchLookupTable(Module & M,uint64_t TableSize,ConstantInt * Offset,const SmallVectorImpl<std::pair<ConstantInt *,Constant * >> & Values,Constant * DefaultValue,const DataLayout & DL,const StringRef & FuncName)4920 SwitchLookupTable::SwitchLookupTable(
4921 Module &M, uint64_t TableSize, ConstantInt *Offset,
4922 const SmallVectorImpl<std::pair<ConstantInt *, Constant *>> &Values,
4923 Constant *DefaultValue, const DataLayout &DL, const StringRef &FuncName) {
4924 assert(Values.size() && "Can't build lookup table without values!");
4925 assert(TableSize >= Values.size() && "Can't fit values in table!");
4926
4927 // If all values in the table are equal, this is that value.
4928 SingleValue = Values.begin()->second;
4929
4930 Type *ValueType = Values.begin()->second->getType();
4931
4932 // Build up the table contents.
4933 SmallVector<Constant *, 64> TableContents(TableSize);
4934 for (size_t I = 0, E = Values.size(); I != E; ++I) {
4935 ConstantInt *CaseVal = Values[I].first;
4936 Constant *CaseRes = Values[I].second;
4937 assert(CaseRes->getType() == ValueType);
4938
4939 uint64_t Idx = (CaseVal->getValue() - Offset->getValue()).getLimitedValue();
4940 TableContents[Idx] = CaseRes;
4941
4942 if (CaseRes != SingleValue)
4943 SingleValue = nullptr;
4944 }
4945
4946 // Fill in any holes in the table with the default result.
4947 if (Values.size() < TableSize) {
4948 assert(DefaultValue &&
4949 "Need a default value to fill the lookup table holes.");
4950 assert(DefaultValue->getType() == ValueType);
4951 for (uint64_t I = 0; I < TableSize; ++I) {
4952 if (!TableContents[I])
4953 TableContents[I] = DefaultValue;
4954 }
4955
4956 if (DefaultValue != SingleValue)
4957 SingleValue = nullptr;
4958 }
4959
4960 // If each element in the table contains the same value, we only need to store
4961 // that single value.
4962 if (SingleValue) {
4963 Kind = SingleValueKind;
4964 return;
4965 }
4966
4967 // Check if we can derive the value with a linear transformation from the
4968 // table index.
4969 if (isa<IntegerType>(ValueType)) {
4970 bool LinearMappingPossible = true;
4971 APInt PrevVal;
4972 APInt DistToPrev;
4973 assert(TableSize >= 2 && "Should be a SingleValue table.");
4974 // Check if there is the same distance between two consecutive values.
4975 for (uint64_t I = 0; I < TableSize; ++I) {
4976 ConstantInt *ConstVal = dyn_cast<ConstantInt>(TableContents[I]);
4977 if (!ConstVal) {
4978 // This is an undef. We could deal with it, but undefs in lookup tables
4979 // are very seldom. It's probably not worth the additional complexity.
4980 LinearMappingPossible = false;
4981 break;
4982 }
4983 const APInt &Val = ConstVal->getValue();
4984 if (I != 0) {
4985 APInt Dist = Val - PrevVal;
4986 if (I == 1) {
4987 DistToPrev = Dist;
4988 } else if (Dist != DistToPrev) {
4989 LinearMappingPossible = false;
4990 break;
4991 }
4992 }
4993 PrevVal = Val;
4994 }
4995 if (LinearMappingPossible) {
4996 LinearOffset = cast<ConstantInt>(TableContents[0]);
4997 LinearMultiplier = ConstantInt::get(M.getContext(), DistToPrev);
4998 Kind = LinearMapKind;
4999 ++NumLinearMaps;
5000 return;
5001 }
5002 }
5003
5004 // If the type is integer and the table fits in a register, build a bitmap.
5005 if (WouldFitInRegister(DL, TableSize, ValueType)) {
5006 IntegerType *IT = cast<IntegerType>(ValueType);
5007 APInt TableInt(TableSize * IT->getBitWidth(), 0);
5008 for (uint64_t I = TableSize; I > 0; --I) {
5009 TableInt <<= IT->getBitWidth();
5010 // Insert values into the bitmap. Undef values are set to zero.
5011 if (!isa<UndefValue>(TableContents[I - 1])) {
5012 ConstantInt *Val = cast<ConstantInt>(TableContents[I - 1]);
5013 TableInt |= Val->getValue().zext(TableInt.getBitWidth());
5014 }
5015 }
5016 BitMap = ConstantInt::get(M.getContext(), TableInt);
5017 BitMapElementTy = IT;
5018 Kind = BitMapKind;
5019 ++NumBitMaps;
5020 return;
5021 }
5022
5023 // Store the table in an array.
5024 ArrayType *ArrayTy = ArrayType::get(ValueType, TableSize);
5025 Constant *Initializer = ConstantArray::get(ArrayTy, TableContents);
5026
5027 Array = new GlobalVariable(M, ArrayTy, /*isConstant=*/true,
5028 GlobalVariable::PrivateLinkage, Initializer,
5029 "switch.table." + FuncName);
5030 Array->setUnnamedAddr(GlobalValue::UnnamedAddr::Global);
5031 // Set the alignment to that of an array items. We will be only loading one
5032 // value out of it.
5033 Array->setAlignment(DL.getPrefTypeAlignment(ValueType));
5034 Kind = ArrayKind;
5035 }
5036
BuildLookup(Value * Index,IRBuilder<> & Builder)5037 Value *SwitchLookupTable::BuildLookup(Value *Index, IRBuilder<> &Builder) {
5038 switch (Kind) {
5039 case SingleValueKind:
5040 return SingleValue;
5041 case LinearMapKind: {
5042 // Derive the result value from the input value.
5043 Value *Result = Builder.CreateIntCast(Index, LinearMultiplier->getType(),
5044 false, "switch.idx.cast");
5045 if (!LinearMultiplier->isOne())
5046 Result = Builder.CreateMul(Result, LinearMultiplier, "switch.idx.mult");
5047 if (!LinearOffset->isZero())
5048 Result = Builder.CreateAdd(Result, LinearOffset, "switch.offset");
5049 return Result;
5050 }
5051 case BitMapKind: {
5052 // Type of the bitmap (e.g. i59).
5053 IntegerType *MapTy = BitMap->getType();
5054
5055 // Cast Index to the same type as the bitmap.
5056 // Note: The Index is <= the number of elements in the table, so
5057 // truncating it to the width of the bitmask is safe.
5058 Value *ShiftAmt = Builder.CreateZExtOrTrunc(Index, MapTy, "switch.cast");
5059
5060 // Multiply the shift amount by the element width.
5061 ShiftAmt = Builder.CreateMul(
5062 ShiftAmt, ConstantInt::get(MapTy, BitMapElementTy->getBitWidth()),
5063 "switch.shiftamt");
5064
5065 // Shift down.
5066 Value *DownShifted =
5067 Builder.CreateLShr(BitMap, ShiftAmt, "switch.downshift");
5068 // Mask off.
5069 return Builder.CreateTrunc(DownShifted, BitMapElementTy, "switch.masked");
5070 }
5071 case ArrayKind: {
5072 // Make sure the table index will not overflow when treated as signed.
5073 IntegerType *IT = cast<IntegerType>(Index->getType());
5074 uint64_t TableSize =
5075 Array->getInitializer()->getType()->getArrayNumElements();
5076 if (TableSize > (1ULL << (IT->getBitWidth() - 1)))
5077 Index = Builder.CreateZExt(
5078 Index, IntegerType::get(IT->getContext(), IT->getBitWidth() + 1),
5079 "switch.tableidx.zext");
5080
5081 Value *GEPIndices[] = {Builder.getInt32(0), Index};
5082 Value *GEP = Builder.CreateInBoundsGEP(Array->getValueType(), Array,
5083 GEPIndices, "switch.gep");
5084 return Builder.CreateLoad(
5085 cast<ArrayType>(Array->getValueType())->getElementType(), GEP,
5086 "switch.load");
5087 }
5088 }
5089 llvm_unreachable("Unknown lookup table kind!");
5090 }
5091
WouldFitInRegister(const DataLayout & DL,uint64_t TableSize,Type * ElementType)5092 bool SwitchLookupTable::WouldFitInRegister(const DataLayout &DL,
5093 uint64_t TableSize,
5094 Type *ElementType) {
5095 auto *IT = dyn_cast<IntegerType>(ElementType);
5096 if (!IT)
5097 return false;
5098 // FIXME: If the type is wider than it needs to be, e.g. i8 but all values
5099 // are <= 15, we could try to narrow the type.
5100
5101 // Avoid overflow, fitsInLegalInteger uses unsigned int for the width.
5102 if (TableSize >= UINT_MAX / IT->getBitWidth())
5103 return false;
5104 return DL.fitsInLegalInteger(TableSize * IT->getBitWidth());
5105 }
5106
5107 /// Determine whether a lookup table should be built for this switch, based on
5108 /// the number of cases, size of the table, and the types of the results.
5109 static bool
ShouldBuildLookupTable(SwitchInst * SI,uint64_t TableSize,const TargetTransformInfo & TTI,const DataLayout & DL,const SmallDenseMap<PHINode *,Type * > & ResultTypes)5110 ShouldBuildLookupTable(SwitchInst *SI, uint64_t TableSize,
5111 const TargetTransformInfo &TTI, const DataLayout &DL,
5112 const SmallDenseMap<PHINode *, Type *> &ResultTypes) {
5113 if (SI->getNumCases() > TableSize || TableSize >= UINT64_MAX / 10)
5114 return false; // TableSize overflowed, or mul below might overflow.
5115
5116 bool AllTablesFitInRegister = true;
5117 bool HasIllegalType = false;
5118 for (const auto &I : ResultTypes) {
5119 Type *Ty = I.second;
5120
5121 // Saturate this flag to true.
5122 HasIllegalType = HasIllegalType || !TTI.isTypeLegal(Ty);
5123
5124 // Saturate this flag to false.
5125 AllTablesFitInRegister =
5126 AllTablesFitInRegister &&
5127 SwitchLookupTable::WouldFitInRegister(DL, TableSize, Ty);
5128
5129 // If both flags saturate, we're done. NOTE: This *only* works with
5130 // saturating flags, and all flags have to saturate first due to the
5131 // non-deterministic behavior of iterating over a dense map.
5132 if (HasIllegalType && !AllTablesFitInRegister)
5133 break;
5134 }
5135
5136 // If each table would fit in a register, we should build it anyway.
5137 if (AllTablesFitInRegister)
5138 return true;
5139
5140 // Don't build a table that doesn't fit in-register if it has illegal types.
5141 if (HasIllegalType)
5142 return false;
5143
5144 // The table density should be at least 40%. This is the same criterion as for
5145 // jump tables, see SelectionDAGBuilder::handleJTSwitchCase.
5146 // FIXME: Find the best cut-off.
5147 return SI->getNumCases() * 10 >= TableSize * 4;
5148 }
5149
5150 /// Try to reuse the switch table index compare. Following pattern:
5151 /// \code
5152 /// if (idx < tablesize)
5153 /// r = table[idx]; // table does not contain default_value
5154 /// else
5155 /// r = default_value;
5156 /// if (r != default_value)
5157 /// ...
5158 /// \endcode
5159 /// Is optimized to:
5160 /// \code
5161 /// cond = idx < tablesize;
5162 /// if (cond)
5163 /// r = table[idx];
5164 /// else
5165 /// r = default_value;
5166 /// if (cond)
5167 /// ...
5168 /// \endcode
5169 /// 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)5170 static void reuseTableCompare(
5171 User *PhiUser, BasicBlock *PhiBlock, BranchInst *RangeCheckBranch,
5172 Constant *DefaultValue,
5173 const SmallVectorImpl<std::pair<ConstantInt *, Constant *>> &Values) {
5174 ICmpInst *CmpInst = dyn_cast<ICmpInst>(PhiUser);
5175 if (!CmpInst)
5176 return;
5177
5178 // We require that the compare is in the same block as the phi so that jump
5179 // threading can do its work afterwards.
5180 if (CmpInst->getParent() != PhiBlock)
5181 return;
5182
5183 Constant *CmpOp1 = dyn_cast<Constant>(CmpInst->getOperand(1));
5184 if (!CmpOp1)
5185 return;
5186
5187 Value *RangeCmp = RangeCheckBranch->getCondition();
5188 Constant *TrueConst = ConstantInt::getTrue(RangeCmp->getType());
5189 Constant *FalseConst = ConstantInt::getFalse(RangeCmp->getType());
5190
5191 // Check if the compare with the default value is constant true or false.
5192 Constant *DefaultConst = ConstantExpr::getICmp(CmpInst->getPredicate(),
5193 DefaultValue, CmpOp1, true);
5194 if (DefaultConst != TrueConst && DefaultConst != FalseConst)
5195 return;
5196
5197 // Check if the compare with the case values is distinct from the default
5198 // compare result.
5199 for (auto ValuePair : Values) {
5200 Constant *CaseConst = ConstantExpr::getICmp(CmpInst->getPredicate(),
5201 ValuePair.second, CmpOp1, true);
5202 if (!CaseConst || CaseConst == DefaultConst || isa<UndefValue>(CaseConst))
5203 return;
5204 assert((CaseConst == TrueConst || CaseConst == FalseConst) &&
5205 "Expect true or false as compare result.");
5206 }
5207
5208 // Check if the branch instruction dominates the phi node. It's a simple
5209 // dominance check, but sufficient for our needs.
5210 // Although this check is invariant in the calling loops, it's better to do it
5211 // at this late stage. Practically we do it at most once for a switch.
5212 BasicBlock *BranchBlock = RangeCheckBranch->getParent();
5213 for (auto PI = pred_begin(PhiBlock), E = pred_end(PhiBlock); PI != E; ++PI) {
5214 BasicBlock *Pred = *PI;
5215 if (Pred != BranchBlock && Pred->getUniquePredecessor() != BranchBlock)
5216 return;
5217 }
5218
5219 if (DefaultConst == FalseConst) {
5220 // The compare yields the same result. We can replace it.
5221 CmpInst->replaceAllUsesWith(RangeCmp);
5222 ++NumTableCmpReuses;
5223 } else {
5224 // The compare yields the same result, just inverted. We can replace it.
5225 Value *InvertedTableCmp = BinaryOperator::CreateXor(
5226 RangeCmp, ConstantInt::get(RangeCmp->getType(), 1), "inverted.cmp",
5227 RangeCheckBranch);
5228 CmpInst->replaceAllUsesWith(InvertedTableCmp);
5229 ++NumTableCmpReuses;
5230 }
5231 }
5232
5233 /// If the switch is only used to initialize one or more phi nodes in a common
5234 /// successor block with different constant values, replace the switch with
5235 /// lookup tables.
SwitchToLookupTable(SwitchInst * SI,IRBuilder<> & Builder,const DataLayout & DL,const TargetTransformInfo & TTI)5236 static bool SwitchToLookupTable(SwitchInst *SI, IRBuilder<> &Builder,
5237 const DataLayout &DL,
5238 const TargetTransformInfo &TTI) {
5239 assert(SI->getNumCases() > 1 && "Degenerate switch?");
5240
5241 Function *Fn = SI->getParent()->getParent();
5242 // Only build lookup table when we have a target that supports it or the
5243 // attribute is not set.
5244 if (!TTI.shouldBuildLookupTables() ||
5245 (Fn->getFnAttribute("no-jump-tables").getValueAsString() == "true"))
5246 return false;
5247
5248 // FIXME: If the switch is too sparse for a lookup table, perhaps we could
5249 // split off a dense part and build a lookup table for that.
5250
5251 // FIXME: This creates arrays of GEPs to constant strings, which means each
5252 // GEP needs a runtime relocation in PIC code. We should just build one big
5253 // string and lookup indices into that.
5254
5255 // Ignore switches with less than three cases. Lookup tables will not make
5256 // them faster, so we don't analyze them.
5257 if (SI->getNumCases() < 3)
5258 return false;
5259
5260 // Figure out the corresponding result for each case value and phi node in the
5261 // common destination, as well as the min and max case values.
5262 assert(!empty(SI->cases()));
5263 SwitchInst::CaseIt CI = SI->case_begin();
5264 ConstantInt *MinCaseVal = CI->getCaseValue();
5265 ConstantInt *MaxCaseVal = CI->getCaseValue();
5266
5267 BasicBlock *CommonDest = nullptr;
5268
5269 using ResultListTy = SmallVector<std::pair<ConstantInt *, Constant *>, 4>;
5270 SmallDenseMap<PHINode *, ResultListTy> ResultLists;
5271
5272 SmallDenseMap<PHINode *, Constant *> DefaultResults;
5273 SmallDenseMap<PHINode *, Type *> ResultTypes;
5274 SmallVector<PHINode *, 4> PHIs;
5275
5276 for (SwitchInst::CaseIt E = SI->case_end(); CI != E; ++CI) {
5277 ConstantInt *CaseVal = CI->getCaseValue();
5278 if (CaseVal->getValue().slt(MinCaseVal->getValue()))
5279 MinCaseVal = CaseVal;
5280 if (CaseVal->getValue().sgt(MaxCaseVal->getValue()))
5281 MaxCaseVal = CaseVal;
5282
5283 // Resulting value at phi nodes for this case value.
5284 using ResultsTy = SmallVector<std::pair<PHINode *, Constant *>, 4>;
5285 ResultsTy Results;
5286 if (!GetCaseResults(SI, CaseVal, CI->getCaseSuccessor(), &CommonDest,
5287 Results, DL, TTI))
5288 return false;
5289
5290 // Append the result from this case to the list for each phi.
5291 for (const auto &I : Results) {
5292 PHINode *PHI = I.first;
5293 Constant *Value = I.second;
5294 if (!ResultLists.count(PHI))
5295 PHIs.push_back(PHI);
5296 ResultLists[PHI].push_back(std::make_pair(CaseVal, Value));
5297 }
5298 }
5299
5300 // Keep track of the result types.
5301 for (PHINode *PHI : PHIs) {
5302 ResultTypes[PHI] = ResultLists[PHI][0].second->getType();
5303 }
5304
5305 uint64_t NumResults = ResultLists[PHIs[0]].size();
5306 APInt RangeSpread = MaxCaseVal->getValue() - MinCaseVal->getValue();
5307 uint64_t TableSize = RangeSpread.getLimitedValue() + 1;
5308 bool TableHasHoles = (NumResults < TableSize);
5309
5310 // If the table has holes, we need a constant result for the default case
5311 // or a bitmask that fits in a register.
5312 SmallVector<std::pair<PHINode *, Constant *>, 4> DefaultResultsList;
5313 bool HasDefaultResults =
5314 GetCaseResults(SI, nullptr, SI->getDefaultDest(), &CommonDest,
5315 DefaultResultsList, DL, TTI);
5316
5317 bool NeedMask = (TableHasHoles && !HasDefaultResults);
5318 if (NeedMask) {
5319 // As an extra penalty for the validity test we require more cases.
5320 if (SI->getNumCases() < 4) // FIXME: Find best threshold value (benchmark).
5321 return false;
5322 if (!DL.fitsInLegalInteger(TableSize))
5323 return false;
5324 }
5325
5326 for (const auto &I : DefaultResultsList) {
5327 PHINode *PHI = I.first;
5328 Constant *Result = I.second;
5329 DefaultResults[PHI] = Result;
5330 }
5331
5332 if (!ShouldBuildLookupTable(SI, TableSize, TTI, DL, ResultTypes))
5333 return false;
5334
5335 // Create the BB that does the lookups.
5336 Module &Mod = *CommonDest->getParent()->getParent();
5337 BasicBlock *LookupBB = BasicBlock::Create(
5338 Mod.getContext(), "switch.lookup", CommonDest->getParent(), CommonDest);
5339
5340 // Compute the table index value.
5341 Builder.SetInsertPoint(SI);
5342 Value *TableIndex;
5343 if (MinCaseVal->isNullValue())
5344 TableIndex = SI->getCondition();
5345 else
5346 TableIndex = Builder.CreateSub(SI->getCondition(), MinCaseVal,
5347 "switch.tableidx");
5348
5349 // Compute the maximum table size representable by the integer type we are
5350 // switching upon.
5351 unsigned CaseSize = MinCaseVal->getType()->getPrimitiveSizeInBits();
5352 uint64_t MaxTableSize = CaseSize > 63 ? UINT64_MAX : 1ULL << CaseSize;
5353 assert(MaxTableSize >= TableSize &&
5354 "It is impossible for a switch to have more entries than the max "
5355 "representable value of its input integer type's size.");
5356
5357 // If the default destination is unreachable, or if the lookup table covers
5358 // all values of the conditional variable, branch directly to the lookup table
5359 // BB. Otherwise, check that the condition is within the case range.
5360 const bool DefaultIsReachable =
5361 !isa<UnreachableInst>(SI->getDefaultDest()->getFirstNonPHIOrDbg());
5362 const bool GeneratingCoveredLookupTable = (MaxTableSize == TableSize);
5363 BranchInst *RangeCheckBranch = nullptr;
5364
5365 if (!DefaultIsReachable || GeneratingCoveredLookupTable) {
5366 Builder.CreateBr(LookupBB);
5367 // Note: We call removeProdecessor later since we need to be able to get the
5368 // PHI value for the default case in case we're using a bit mask.
5369 } else {
5370 Value *Cmp = Builder.CreateICmpULT(
5371 TableIndex, ConstantInt::get(MinCaseVal->getType(), TableSize));
5372 RangeCheckBranch =
5373 Builder.CreateCondBr(Cmp, LookupBB, SI->getDefaultDest());
5374 }
5375
5376 // Populate the BB that does the lookups.
5377 Builder.SetInsertPoint(LookupBB);
5378
5379 if (NeedMask) {
5380 // Before doing the lookup, we do the hole check. The LookupBB is therefore
5381 // re-purposed to do the hole check, and we create a new LookupBB.
5382 BasicBlock *MaskBB = LookupBB;
5383 MaskBB->setName("switch.hole_check");
5384 LookupBB = BasicBlock::Create(Mod.getContext(), "switch.lookup",
5385 CommonDest->getParent(), CommonDest);
5386
5387 // Make the mask's bitwidth at least 8-bit and a power-of-2 to avoid
5388 // unnecessary illegal types.
5389 uint64_t TableSizePowOf2 = NextPowerOf2(std::max(7ULL, TableSize - 1ULL));
5390 APInt MaskInt(TableSizePowOf2, 0);
5391 APInt One(TableSizePowOf2, 1);
5392 // Build bitmask; fill in a 1 bit for every case.
5393 const ResultListTy &ResultList = ResultLists[PHIs[0]];
5394 for (size_t I = 0, E = ResultList.size(); I != E; ++I) {
5395 uint64_t Idx = (ResultList[I].first->getValue() - MinCaseVal->getValue())
5396 .getLimitedValue();
5397 MaskInt |= One << Idx;
5398 }
5399 ConstantInt *TableMask = ConstantInt::get(Mod.getContext(), MaskInt);
5400
5401 // Get the TableIndex'th bit of the bitmask.
5402 // If this bit is 0 (meaning hole) jump to the default destination,
5403 // else continue with table lookup.
5404 IntegerType *MapTy = TableMask->getType();
5405 Value *MaskIndex =
5406 Builder.CreateZExtOrTrunc(TableIndex, MapTy, "switch.maskindex");
5407 Value *Shifted = Builder.CreateLShr(TableMask, MaskIndex, "switch.shifted");
5408 Value *LoBit = Builder.CreateTrunc(
5409 Shifted, Type::getInt1Ty(Mod.getContext()), "switch.lobit");
5410 Builder.CreateCondBr(LoBit, LookupBB, SI->getDefaultDest());
5411
5412 Builder.SetInsertPoint(LookupBB);
5413 AddPredecessorToBlock(SI->getDefaultDest(), MaskBB, SI->getParent());
5414 }
5415
5416 if (!DefaultIsReachable || GeneratingCoveredLookupTable) {
5417 // We cached PHINodes in PHIs. To avoid accessing deleted PHINodes later,
5418 // do not delete PHINodes here.
5419 SI->getDefaultDest()->removePredecessor(SI->getParent(),
5420 /*KeepOneInputPHIs=*/true);
5421 }
5422
5423 bool ReturnedEarly = false;
5424 for (PHINode *PHI : PHIs) {
5425 const ResultListTy &ResultList = ResultLists[PHI];
5426
5427 // If using a bitmask, use any value to fill the lookup table holes.
5428 Constant *DV = NeedMask ? ResultLists[PHI][0].second : DefaultResults[PHI];
5429 StringRef FuncName = Fn->getName();
5430 SwitchLookupTable Table(Mod, TableSize, MinCaseVal, ResultList, DV, DL,
5431 FuncName);
5432
5433 Value *Result = Table.BuildLookup(TableIndex, Builder);
5434
5435 // If the result is used to return immediately from the function, we want to
5436 // do that right here.
5437 if (PHI->hasOneUse() && isa<ReturnInst>(*PHI->user_begin()) &&
5438 PHI->user_back() == CommonDest->getFirstNonPHIOrDbg()) {
5439 Builder.CreateRet(Result);
5440 ReturnedEarly = true;
5441 break;
5442 }
5443
5444 // Do a small peephole optimization: re-use the switch table compare if
5445 // possible.
5446 if (!TableHasHoles && HasDefaultResults && RangeCheckBranch) {
5447 BasicBlock *PhiBlock = PHI->getParent();
5448 // Search for compare instructions which use the phi.
5449 for (auto *User : PHI->users()) {
5450 reuseTableCompare(User, PhiBlock, RangeCheckBranch, DV, ResultList);
5451 }
5452 }
5453
5454 PHI->addIncoming(Result, LookupBB);
5455 }
5456
5457 if (!ReturnedEarly)
5458 Builder.CreateBr(CommonDest);
5459
5460 // Remove the switch.
5461 for (unsigned i = 0, e = SI->getNumSuccessors(); i < e; ++i) {
5462 BasicBlock *Succ = SI->getSuccessor(i);
5463
5464 if (Succ == SI->getDefaultDest())
5465 continue;
5466 Succ->removePredecessor(SI->getParent());
5467 }
5468 SI->eraseFromParent();
5469
5470 ++NumLookupTables;
5471 if (NeedMask)
5472 ++NumLookupTablesHoles;
5473 return true;
5474 }
5475
isSwitchDense(ArrayRef<int64_t> Values)5476 static bool isSwitchDense(ArrayRef<int64_t> Values) {
5477 // See also SelectionDAGBuilder::isDense(), which this function was based on.
5478 uint64_t Diff = (uint64_t)Values.back() - (uint64_t)Values.front();
5479 uint64_t Range = Diff + 1;
5480 uint64_t NumCases = Values.size();
5481 // 40% is the default density for building a jump table in optsize/minsize mode.
5482 uint64_t MinDensity = 40;
5483
5484 return NumCases * 100 >= Range * MinDensity;
5485 }
5486
5487 /// Try to transform a switch that has "holes" in it to a contiguous sequence
5488 /// of cases.
5489 ///
5490 /// A switch such as: switch(i) {case 5: case 9: case 13: case 17:} can be
5491 /// range-reduced to: switch ((i-5) / 4) {case 0: case 1: case 2: case 3:}.
5492 ///
5493 /// This converts a sparse switch into a dense switch which allows better
5494 /// lowering and could also allow transforming into a lookup table.
ReduceSwitchRange(SwitchInst * SI,IRBuilder<> & Builder,const DataLayout & DL,const TargetTransformInfo & TTI)5495 static bool ReduceSwitchRange(SwitchInst *SI, IRBuilder<> &Builder,
5496 const DataLayout &DL,
5497 const TargetTransformInfo &TTI) {
5498 auto *CondTy = cast<IntegerType>(SI->getCondition()->getType());
5499 if (CondTy->getIntegerBitWidth() > 64 ||
5500 !DL.fitsInLegalInteger(CondTy->getIntegerBitWidth()))
5501 return false;
5502 // Only bother with this optimization if there are more than 3 switch cases;
5503 // SDAG will only bother creating jump tables for 4 or more cases.
5504 if (SI->getNumCases() < 4)
5505 return false;
5506
5507 // This transform is agnostic to the signedness of the input or case values. We
5508 // can treat the case values as signed or unsigned. We can optimize more common
5509 // cases such as a sequence crossing zero {-4,0,4,8} if we interpret case values
5510 // as signed.
5511 SmallVector<int64_t,4> Values;
5512 for (auto &C : SI->cases())
5513 Values.push_back(C.getCaseValue()->getValue().getSExtValue());
5514 llvm::sort(Values);
5515
5516 // If the switch is already dense, there's nothing useful to do here.
5517 if (isSwitchDense(Values))
5518 return false;
5519
5520 // First, transform the values such that they start at zero and ascend.
5521 int64_t Base = Values[0];
5522 for (auto &V : Values)
5523 V -= (uint64_t)(Base);
5524
5525 // Now we have signed numbers that have been shifted so that, given enough
5526 // precision, there are no negative values. Since the rest of the transform
5527 // is bitwise only, we switch now to an unsigned representation.
5528
5529 // This transform can be done speculatively because it is so cheap - it
5530 // results in a single rotate operation being inserted.
5531 // FIXME: It's possible that optimizing a switch on powers of two might also
5532 // be beneficial - flag values are often powers of two and we could use a CLZ
5533 // as the key function.
5534
5535 // countTrailingZeros(0) returns 64. As Values is guaranteed to have more than
5536 // one element and LLVM disallows duplicate cases, Shift is guaranteed to be
5537 // less than 64.
5538 unsigned Shift = 64;
5539 for (auto &V : Values)
5540 Shift = std::min(Shift, countTrailingZeros((uint64_t)V));
5541 assert(Shift < 64);
5542 if (Shift > 0)
5543 for (auto &V : Values)
5544 V = (int64_t)((uint64_t)V >> Shift);
5545
5546 if (!isSwitchDense(Values))
5547 // Transform didn't create a dense switch.
5548 return false;
5549
5550 // The obvious transform is to shift the switch condition right and emit a
5551 // check that the condition actually cleanly divided by GCD, i.e.
5552 // C & (1 << Shift - 1) == 0
5553 // inserting a new CFG edge to handle the case where it didn't divide cleanly.
5554 //
5555 // A cheaper way of doing this is a simple ROTR(C, Shift). This performs the
5556 // shift and puts the shifted-off bits in the uppermost bits. If any of these
5557 // are nonzero then the switch condition will be very large and will hit the
5558 // default case.
5559
5560 auto *Ty = cast<IntegerType>(SI->getCondition()->getType());
5561 Builder.SetInsertPoint(SI);
5562 auto *ShiftC = ConstantInt::get(Ty, Shift);
5563 auto *Sub = Builder.CreateSub(SI->getCondition(), ConstantInt::get(Ty, Base));
5564 auto *LShr = Builder.CreateLShr(Sub, ShiftC);
5565 auto *Shl = Builder.CreateShl(Sub, Ty->getBitWidth() - Shift);
5566 auto *Rot = Builder.CreateOr(LShr, Shl);
5567 SI->replaceUsesOfWith(SI->getCondition(), Rot);
5568
5569 for (auto Case : SI->cases()) {
5570 auto *Orig = Case.getCaseValue();
5571 auto Sub = Orig->getValue() - APInt(Ty->getBitWidth(), Base);
5572 Case.setValue(
5573 cast<ConstantInt>(ConstantInt::get(Ty, Sub.lshr(ShiftC->getValue()))));
5574 }
5575 return true;
5576 }
5577
SimplifySwitch(SwitchInst * SI,IRBuilder<> & Builder)5578 bool SimplifyCFGOpt::SimplifySwitch(SwitchInst *SI, IRBuilder<> &Builder) {
5579 BasicBlock *BB = SI->getParent();
5580
5581 if (isValueEqualityComparison(SI)) {
5582 // If we only have one predecessor, and if it is a branch on this value,
5583 // see if that predecessor totally determines the outcome of this switch.
5584 if (BasicBlock *OnlyPred = BB->getSinglePredecessor())
5585 if (SimplifyEqualityComparisonWithOnlyPredecessor(SI, OnlyPred, Builder))
5586 return requestResimplify();
5587
5588 Value *Cond = SI->getCondition();
5589 if (SelectInst *Select = dyn_cast<SelectInst>(Cond))
5590 if (SimplifySwitchOnSelect(SI, Select))
5591 return requestResimplify();
5592
5593 // If the block only contains the switch, see if we can fold the block
5594 // away into any preds.
5595 if (SI == &*BB->instructionsWithoutDebug().begin())
5596 if (FoldValueComparisonIntoPredecessors(SI, Builder))
5597 return requestResimplify();
5598 }
5599
5600 // Try to transform the switch into an icmp and a branch.
5601 if (TurnSwitchRangeIntoICmp(SI, Builder))
5602 return requestResimplify();
5603
5604 // Remove unreachable cases.
5605 if (eliminateDeadSwitchCases(SI, Options.AC, DL))
5606 return requestResimplify();
5607
5608 if (switchToSelect(SI, Builder, DL, TTI))
5609 return requestResimplify();
5610
5611 if (Options.ForwardSwitchCondToPhi && ForwardSwitchConditionToPHI(SI))
5612 return requestResimplify();
5613
5614 // The conversion from switch to lookup tables results in difficult-to-analyze
5615 // code and makes pruning branches much harder. This is a problem if the
5616 // switch expression itself can still be restricted as a result of inlining or
5617 // CVP. Therefore, only apply this transformation during late stages of the
5618 // optimisation pipeline.
5619 if (Options.ConvertSwitchToLookupTable &&
5620 SwitchToLookupTable(SI, Builder, DL, TTI))
5621 return requestResimplify();
5622
5623 if (ReduceSwitchRange(SI, Builder, DL, TTI))
5624 return requestResimplify();
5625
5626 return false;
5627 }
5628
SimplifyIndirectBr(IndirectBrInst * IBI)5629 bool SimplifyCFGOpt::SimplifyIndirectBr(IndirectBrInst *IBI) {
5630 BasicBlock *BB = IBI->getParent();
5631 bool Changed = false;
5632
5633 // Eliminate redundant destinations.
5634 SmallPtrSet<Value *, 8> Succs;
5635 for (unsigned i = 0, e = IBI->getNumDestinations(); i != e; ++i) {
5636 BasicBlock *Dest = IBI->getDestination(i);
5637 if (!Dest->hasAddressTaken() || !Succs.insert(Dest).second) {
5638 Dest->removePredecessor(BB);
5639 IBI->removeDestination(i);
5640 --i;
5641 --e;
5642 Changed = true;
5643 }
5644 }
5645
5646 if (IBI->getNumDestinations() == 0) {
5647 // If the indirectbr has no successors, change it to unreachable.
5648 new UnreachableInst(IBI->getContext(), IBI);
5649 EraseTerminatorAndDCECond(IBI);
5650 return true;
5651 }
5652
5653 if (IBI->getNumDestinations() == 1) {
5654 // If the indirectbr has one successor, change it to a direct branch.
5655 BranchInst::Create(IBI->getDestination(0), IBI);
5656 EraseTerminatorAndDCECond(IBI);
5657 return true;
5658 }
5659
5660 if (SelectInst *SI = dyn_cast<SelectInst>(IBI->getAddress())) {
5661 if (SimplifyIndirectBrOnSelect(IBI, SI))
5662 return requestResimplify();
5663 }
5664 return Changed;
5665 }
5666
5667 /// Given an block with only a single landing pad and a unconditional branch
5668 /// try to find another basic block which this one can be merged with. This
5669 /// handles cases where we have multiple invokes with unique landing pads, but
5670 /// a shared handler.
5671 ///
5672 /// We specifically choose to not worry about merging non-empty blocks
5673 /// here. That is a PRE/scheduling problem and is best solved elsewhere. In
5674 /// practice, the optimizer produces empty landing pad blocks quite frequently
5675 /// when dealing with exception dense code. (see: instcombine, gvn, if-else
5676 /// sinking in this file)
5677 ///
5678 /// This is primarily a code size optimization. We need to avoid performing
5679 /// any transform which might inhibit optimization (such as our ability to
5680 /// specialize a particular handler via tail commoning). We do this by not
5681 /// merging any blocks which require us to introduce a phi. Since the same
5682 /// values are flowing through both blocks, we don't lose any ability to
5683 /// specialize. If anything, we make such specialization more likely.
5684 ///
5685 /// TODO - This transformation could remove entries from a phi in the target
5686 /// block when the inputs in the phi are the same for the two blocks being
5687 /// merged. In some cases, this could result in removal of the PHI entirely.
TryToMergeLandingPad(LandingPadInst * LPad,BranchInst * BI,BasicBlock * BB)5688 static bool TryToMergeLandingPad(LandingPadInst *LPad, BranchInst *BI,
5689 BasicBlock *BB) {
5690 auto Succ = BB->getUniqueSuccessor();
5691 assert(Succ);
5692 // If there's a phi in the successor block, we'd likely have to introduce
5693 // a phi into the merged landing pad block.
5694 if (isa<PHINode>(*Succ->begin()))
5695 return false;
5696
5697 for (BasicBlock *OtherPred : predecessors(Succ)) {
5698 if (BB == OtherPred)
5699 continue;
5700 BasicBlock::iterator I = OtherPred->begin();
5701 LandingPadInst *LPad2 = dyn_cast<LandingPadInst>(I);
5702 if (!LPad2 || !LPad2->isIdenticalTo(LPad))
5703 continue;
5704 for (++I; isa<DbgInfoIntrinsic>(I); ++I)
5705 ;
5706 BranchInst *BI2 = dyn_cast<BranchInst>(I);
5707 if (!BI2 || !BI2->isIdenticalTo(BI))
5708 continue;
5709
5710 // We've found an identical block. Update our predecessors to take that
5711 // path instead and make ourselves dead.
5712 SmallPtrSet<BasicBlock *, 16> Preds;
5713 Preds.insert(pred_begin(BB), pred_end(BB));
5714 for (BasicBlock *Pred : Preds) {
5715 InvokeInst *II = cast<InvokeInst>(Pred->getTerminator());
5716 assert(II->getNormalDest() != BB && II->getUnwindDest() == BB &&
5717 "unexpected successor");
5718 II->setUnwindDest(OtherPred);
5719 }
5720
5721 // The debug info in OtherPred doesn't cover the merged control flow that
5722 // used to go through BB. We need to delete it or update it.
5723 for (auto I = OtherPred->begin(), E = OtherPred->end(); I != E;) {
5724 Instruction &Inst = *I;
5725 I++;
5726 if (isa<DbgInfoIntrinsic>(Inst))
5727 Inst.eraseFromParent();
5728 }
5729
5730 SmallPtrSet<BasicBlock *, 16> Succs;
5731 Succs.insert(succ_begin(BB), succ_end(BB));
5732 for (BasicBlock *Succ : Succs) {
5733 Succ->removePredecessor(BB);
5734 }
5735
5736 IRBuilder<> Builder(BI);
5737 Builder.CreateUnreachable();
5738 BI->eraseFromParent();
5739 return true;
5740 }
5741 return false;
5742 }
5743
SimplifyUncondBranch(BranchInst * BI,IRBuilder<> & Builder)5744 bool SimplifyCFGOpt::SimplifyUncondBranch(BranchInst *BI,
5745 IRBuilder<> &Builder) {
5746 BasicBlock *BB = BI->getParent();
5747 BasicBlock *Succ = BI->getSuccessor(0);
5748
5749 // If the Terminator is the only non-phi instruction, simplify the block.
5750 // If LoopHeader is provided, check if the block or its successor is a loop
5751 // header. (This is for early invocations before loop simplify and
5752 // vectorization to keep canonical loop forms for nested loops. These blocks
5753 // can be eliminated when the pass is invoked later in the back-end.)
5754 // Note that if BB has only one predecessor then we do not introduce new
5755 // backedge, so we can eliminate BB.
5756 bool NeedCanonicalLoop =
5757 Options.NeedCanonicalLoop &&
5758 (LoopHeaders && BB->hasNPredecessorsOrMore(2) &&
5759 (LoopHeaders->count(BB) || LoopHeaders->count(Succ)));
5760 BasicBlock::iterator I = BB->getFirstNonPHIOrDbg()->getIterator();
5761 if (I->isTerminator() && BB != &BB->getParent()->getEntryBlock() &&
5762 !NeedCanonicalLoop && TryToSimplifyUncondBranchFromEmptyBlock(BB))
5763 return true;
5764
5765 // If the only instruction in the block is a seteq/setne comparison against a
5766 // constant, try to simplify the block.
5767 if (ICmpInst *ICI = dyn_cast<ICmpInst>(I))
5768 if (ICI->isEquality() && isa<ConstantInt>(ICI->getOperand(1))) {
5769 for (++I; isa<DbgInfoIntrinsic>(I); ++I)
5770 ;
5771 if (I->isTerminator() &&
5772 tryToSimplifyUncondBranchWithICmpInIt(ICI, Builder))
5773 return true;
5774 }
5775
5776 // See if we can merge an empty landing pad block with another which is
5777 // equivalent.
5778 if (LandingPadInst *LPad = dyn_cast<LandingPadInst>(I)) {
5779 for (++I; isa<DbgInfoIntrinsic>(I); ++I)
5780 ;
5781 if (I->isTerminator() && TryToMergeLandingPad(LPad, BI, BB))
5782 return true;
5783 }
5784
5785 // If this basic block is ONLY a compare and a branch, and if a predecessor
5786 // branches to us and our successor, fold the comparison into the
5787 // predecessor and use logical operations to update the incoming value
5788 // for PHI nodes in common successor.
5789 if (FoldBranchToCommonDest(BI, nullptr, Options.BonusInstThreshold))
5790 return requestResimplify();
5791 return false;
5792 }
5793
allPredecessorsComeFromSameSource(BasicBlock * BB)5794 static BasicBlock *allPredecessorsComeFromSameSource(BasicBlock *BB) {
5795 BasicBlock *PredPred = nullptr;
5796 for (auto *P : predecessors(BB)) {
5797 BasicBlock *PPred = P->getSinglePredecessor();
5798 if (!PPred || (PredPred && PredPred != PPred))
5799 return nullptr;
5800 PredPred = PPred;
5801 }
5802 return PredPred;
5803 }
5804
SimplifyCondBranch(BranchInst * BI,IRBuilder<> & Builder)5805 bool SimplifyCFGOpt::SimplifyCondBranch(BranchInst *BI, IRBuilder<> &Builder) {
5806 BasicBlock *BB = BI->getParent();
5807 const Function *Fn = BB->getParent();
5808 if (Fn && Fn->hasFnAttribute(Attribute::OptForFuzzing))
5809 return false;
5810
5811 // Conditional branch
5812 if (isValueEqualityComparison(BI)) {
5813 // If we only have one predecessor, and if it is a branch on this value,
5814 // see if that predecessor totally determines the outcome of this
5815 // switch.
5816 if (BasicBlock *OnlyPred = BB->getSinglePredecessor())
5817 if (SimplifyEqualityComparisonWithOnlyPredecessor(BI, OnlyPred, Builder))
5818 return requestResimplify();
5819
5820 // This block must be empty, except for the setcond inst, if it exists.
5821 // Ignore dbg intrinsics.
5822 auto I = BB->instructionsWithoutDebug().begin();
5823 if (&*I == BI) {
5824 if (FoldValueComparisonIntoPredecessors(BI, Builder))
5825 return requestResimplify();
5826 } else if (&*I == cast<Instruction>(BI->getCondition())) {
5827 ++I;
5828 if (&*I == BI && FoldValueComparisonIntoPredecessors(BI, Builder))
5829 return requestResimplify();
5830 }
5831 }
5832
5833 // Try to turn "br (X == 0 | X == 1), T, F" into a switch instruction.
5834 if (SimplifyBranchOnICmpChain(BI, Builder, DL))
5835 return true;
5836
5837 // If this basic block has dominating predecessor blocks and the dominating
5838 // blocks' conditions imply BI's condition, we know the direction of BI.
5839 Optional<bool> Imp = isImpliedByDomCondition(BI->getCondition(), BI, DL);
5840 if (Imp) {
5841 // Turn this into a branch on constant.
5842 auto *OldCond = BI->getCondition();
5843 ConstantInt *TorF = *Imp ? ConstantInt::getTrue(BB->getContext())
5844 : ConstantInt::getFalse(BB->getContext());
5845 BI->setCondition(TorF);
5846 RecursivelyDeleteTriviallyDeadInstructions(OldCond);
5847 return requestResimplify();
5848 }
5849
5850 // If this basic block is ONLY a compare and a branch, and if a predecessor
5851 // branches to us and one of our successors, fold the comparison into the
5852 // predecessor and use logical operations to pick the right destination.
5853 if (FoldBranchToCommonDest(BI, nullptr, Options.BonusInstThreshold))
5854 return requestResimplify();
5855
5856 // We have a conditional branch to two blocks that are only reachable
5857 // from BI. We know that the condbr dominates the two blocks, so see if
5858 // there is any identical code in the "then" and "else" blocks. If so, we
5859 // can hoist it up to the branching block.
5860 if (BI->getSuccessor(0)->getSinglePredecessor()) {
5861 if (BI->getSuccessor(1)->getSinglePredecessor()) {
5862 if (HoistThenElseCodeToIf(BI, TTI))
5863 return requestResimplify();
5864 } else {
5865 // If Successor #1 has multiple preds, we may be able to conditionally
5866 // execute Successor #0 if it branches to Successor #1.
5867 Instruction *Succ0TI = BI->getSuccessor(0)->getTerminator();
5868 if (Succ0TI->getNumSuccessors() == 1 &&
5869 Succ0TI->getSuccessor(0) == BI->getSuccessor(1))
5870 if (SpeculativelyExecuteBB(BI, BI->getSuccessor(0), TTI))
5871 return requestResimplify();
5872 }
5873 } else if (BI->getSuccessor(1)->getSinglePredecessor()) {
5874 // If Successor #0 has multiple preds, we may be able to conditionally
5875 // execute Successor #1 if it branches to Successor #0.
5876 Instruction *Succ1TI = BI->getSuccessor(1)->getTerminator();
5877 if (Succ1TI->getNumSuccessors() == 1 &&
5878 Succ1TI->getSuccessor(0) == BI->getSuccessor(0))
5879 if (SpeculativelyExecuteBB(BI, BI->getSuccessor(1), TTI))
5880 return requestResimplify();
5881 }
5882
5883 // If this is a branch on a phi node in the current block, thread control
5884 // through this block if any PHI node entries are constants.
5885 if (PHINode *PN = dyn_cast<PHINode>(BI->getCondition()))
5886 if (PN->getParent() == BI->getParent())
5887 if (FoldCondBranchOnPHI(BI, DL, Options.AC))
5888 return requestResimplify();
5889
5890 // Scan predecessor blocks for conditional branches.
5891 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
5892 if (BranchInst *PBI = dyn_cast<BranchInst>((*PI)->getTerminator()))
5893 if (PBI != BI && PBI->isConditional())
5894 if (SimplifyCondBranchToCondBranch(PBI, BI, DL))
5895 return requestResimplify();
5896
5897 // Look for diamond patterns.
5898 if (MergeCondStores)
5899 if (BasicBlock *PrevBB = allPredecessorsComeFromSameSource(BB))
5900 if (BranchInst *PBI = dyn_cast<BranchInst>(PrevBB->getTerminator()))
5901 if (PBI != BI && PBI->isConditional())
5902 if (mergeConditionalStores(PBI, BI, DL))
5903 return requestResimplify();
5904
5905 return false;
5906 }
5907
5908 /// Check if passing a value to an instruction will cause undefined behavior.
passingValueIsAlwaysUndefined(Value * V,Instruction * I)5909 static bool passingValueIsAlwaysUndefined(Value *V, Instruction *I) {
5910 Constant *C = dyn_cast<Constant>(V);
5911 if (!C)
5912 return false;
5913
5914 if (I->use_empty())
5915 return false;
5916
5917 if (C->isNullValue() || isa<UndefValue>(C)) {
5918 // Only look at the first use, avoid hurting compile time with long uselists
5919 User *Use = *I->user_begin();
5920
5921 // Now make sure that there are no instructions in between that can alter
5922 // control flow (eg. calls)
5923 for (BasicBlock::iterator
5924 i = ++BasicBlock::iterator(I),
5925 UI = BasicBlock::iterator(dyn_cast<Instruction>(Use));
5926 i != UI; ++i)
5927 if (i == I->getParent()->end() || i->mayHaveSideEffects())
5928 return false;
5929
5930 // Look through GEPs. A load from a GEP derived from NULL is still undefined
5931 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Use))
5932 if (GEP->getPointerOperand() == I)
5933 return passingValueIsAlwaysUndefined(V, GEP);
5934
5935 // Look through bitcasts.
5936 if (BitCastInst *BC = dyn_cast<BitCastInst>(Use))
5937 return passingValueIsAlwaysUndefined(V, BC);
5938
5939 // Load from null is undefined.
5940 if (LoadInst *LI = dyn_cast<LoadInst>(Use))
5941 if (!LI->isVolatile())
5942 return !NullPointerIsDefined(LI->getFunction(),
5943 LI->getPointerAddressSpace());
5944
5945 // Store to null is undefined.
5946 if (StoreInst *SI = dyn_cast<StoreInst>(Use))
5947 if (!SI->isVolatile())
5948 return (!NullPointerIsDefined(SI->getFunction(),
5949 SI->getPointerAddressSpace())) &&
5950 SI->getPointerOperand() == I;
5951
5952 // A call to null is undefined.
5953 if (auto CS = CallSite(Use))
5954 return !NullPointerIsDefined(CS->getFunction()) &&
5955 CS.getCalledValue() == I;
5956 }
5957 return false;
5958 }
5959
5960 /// If BB has an incoming value that will always trigger undefined behavior
5961 /// (eg. null pointer dereference), remove the branch leading here.
removeUndefIntroducingPredecessor(BasicBlock * BB)5962 static bool removeUndefIntroducingPredecessor(BasicBlock *BB) {
5963 for (PHINode &PHI : BB->phis())
5964 for (unsigned i = 0, e = PHI.getNumIncomingValues(); i != e; ++i)
5965 if (passingValueIsAlwaysUndefined(PHI.getIncomingValue(i), &PHI)) {
5966 Instruction *T = PHI.getIncomingBlock(i)->getTerminator();
5967 IRBuilder<> Builder(T);
5968 if (BranchInst *BI = dyn_cast<BranchInst>(T)) {
5969 BB->removePredecessor(PHI.getIncomingBlock(i));
5970 // Turn uncoditional branches into unreachables and remove the dead
5971 // destination from conditional branches.
5972 if (BI->isUnconditional())
5973 Builder.CreateUnreachable();
5974 else
5975 Builder.CreateBr(BI->getSuccessor(0) == BB ? BI->getSuccessor(1)
5976 : BI->getSuccessor(0));
5977 BI->eraseFromParent();
5978 return true;
5979 }
5980 // TODO: SwitchInst.
5981 }
5982
5983 return false;
5984 }
5985
simplifyOnce(BasicBlock * BB)5986 bool SimplifyCFGOpt::simplifyOnce(BasicBlock *BB) {
5987 bool Changed = false;
5988
5989 assert(BB && BB->getParent() && "Block not embedded in function!");
5990 assert(BB->getTerminator() && "Degenerate basic block encountered!");
5991
5992 // Remove basic blocks that have no predecessors (except the entry block)...
5993 // or that just have themself as a predecessor. These are unreachable.
5994 if ((pred_empty(BB) && BB != &BB->getParent()->getEntryBlock()) ||
5995 BB->getSinglePredecessor() == BB) {
5996 LLVM_DEBUG(dbgs() << "Removing BB: \n" << *BB);
5997 DeleteDeadBlock(BB);
5998 return true;
5999 }
6000
6001 // Check to see if we can constant propagate this terminator instruction
6002 // away...
6003 Changed |= ConstantFoldTerminator(BB, true);
6004
6005 // Check for and eliminate duplicate PHI nodes in this block.
6006 Changed |= EliminateDuplicatePHINodes(BB);
6007
6008 // Check for and remove branches that will always cause undefined behavior.
6009 Changed |= removeUndefIntroducingPredecessor(BB);
6010
6011 // Merge basic blocks into their predecessor if there is only one distinct
6012 // pred, and if there is only one distinct successor of the predecessor, and
6013 // if there are no PHI nodes.
6014 if (MergeBlockIntoPredecessor(BB))
6015 return true;
6016
6017 if (SinkCommon && Options.SinkCommonInsts)
6018 Changed |= SinkCommonCodeFromPredecessors(BB);
6019
6020 IRBuilder<> Builder(BB);
6021
6022 // If there is a trivial two-entry PHI node in this basic block, and we can
6023 // eliminate it, do so now.
6024 if (auto *PN = dyn_cast<PHINode>(BB->begin()))
6025 if (PN->getNumIncomingValues() == 2)
6026 Changed |= FoldTwoEntryPHINode(PN, TTI, DL);
6027
6028 Builder.SetInsertPoint(BB->getTerminator());
6029 if (auto *BI = dyn_cast<BranchInst>(BB->getTerminator())) {
6030 if (BI->isUnconditional()) {
6031 if (SimplifyUncondBranch(BI, Builder))
6032 return true;
6033 } else {
6034 if (SimplifyCondBranch(BI, Builder))
6035 return true;
6036 }
6037 } else if (auto *RI = dyn_cast<ReturnInst>(BB->getTerminator())) {
6038 if (SimplifyReturn(RI, Builder))
6039 return true;
6040 } else if (auto *RI = dyn_cast<ResumeInst>(BB->getTerminator())) {
6041 if (SimplifyResume(RI, Builder))
6042 return true;
6043 } else if (auto *RI = dyn_cast<CleanupReturnInst>(BB->getTerminator())) {
6044 if (SimplifyCleanupReturn(RI))
6045 return true;
6046 } else if (auto *SI = dyn_cast<SwitchInst>(BB->getTerminator())) {
6047 if (SimplifySwitch(SI, Builder))
6048 return true;
6049 } else if (auto *UI = dyn_cast<UnreachableInst>(BB->getTerminator())) {
6050 if (SimplifyUnreachable(UI))
6051 return true;
6052 } else if (auto *IBI = dyn_cast<IndirectBrInst>(BB->getTerminator())) {
6053 if (SimplifyIndirectBr(IBI))
6054 return true;
6055 }
6056
6057 return Changed;
6058 }
6059
run(BasicBlock * BB)6060 bool SimplifyCFGOpt::run(BasicBlock *BB) {
6061 bool Changed = false;
6062
6063 // Repeated simplify BB as long as resimplification is requested.
6064 do {
6065 Resimplify = false;
6066
6067 // Perform one round of simplifcation. Resimplify flag will be set if
6068 // another iteration is requested.
6069 Changed |= simplifyOnce(BB);
6070 } while (Resimplify);
6071
6072 return Changed;
6073 }
6074
simplifyCFG(BasicBlock * BB,const TargetTransformInfo & TTI,const SimplifyCFGOptions & Options,SmallPtrSetImpl<BasicBlock * > * LoopHeaders)6075 bool llvm::simplifyCFG(BasicBlock *BB, const TargetTransformInfo &TTI,
6076 const SimplifyCFGOptions &Options,
6077 SmallPtrSetImpl<BasicBlock *> *LoopHeaders) {
6078 return SimplifyCFGOpt(TTI, BB->getModule()->getDataLayout(), LoopHeaders,
6079 Options)
6080 .run(BB);
6081 }
6082