1 //===- SimplifyCFG.cpp - Code to perform CFG simplification ---------------===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 // Peephole optimize the CFG.
10 //
11 //===----------------------------------------------------------------------===//
12
13 #include "llvm/ADT/APInt.h"
14 #include "llvm/ADT/ArrayRef.h"
15 #include "llvm/ADT/DenseMap.h"
16 #include "llvm/ADT/Optional.h"
17 #include "llvm/ADT/STLExtras.h"
18 #include "llvm/ADT/SetOperations.h"
19 #include "llvm/ADT/SetVector.h"
20 #include "llvm/ADT/SmallPtrSet.h"
21 #include "llvm/ADT/SmallVector.h"
22 #include "llvm/ADT/Statistic.h"
23 #include "llvm/ADT/StringRef.h"
24 #include "llvm/Analysis/AssumptionCache.h"
25 #include "llvm/Analysis/ConstantFolding.h"
26 #include "llvm/Analysis/EHPersonalities.h"
27 #include "llvm/Analysis/GuardUtils.h"
28 #include "llvm/Analysis/InstructionSimplify.h"
29 #include "llvm/Analysis/MemorySSA.h"
30 #include "llvm/Analysis/MemorySSAUpdater.h"
31 #include "llvm/Analysis/TargetTransformInfo.h"
32 #include "llvm/Analysis/ValueTracking.h"
33 #include "llvm/IR/Attributes.h"
34 #include "llvm/IR/BasicBlock.h"
35 #include "llvm/IR/CFG.h"
36 #include "llvm/IR/CallSite.h"
37 #include "llvm/IR/Constant.h"
38 #include "llvm/IR/ConstantRange.h"
39 #include "llvm/IR/Constants.h"
40 #include "llvm/IR/DataLayout.h"
41 #include "llvm/IR/DerivedTypes.h"
42 #include "llvm/IR/Function.h"
43 #include "llvm/IR/GlobalValue.h"
44 #include "llvm/IR/GlobalVariable.h"
45 #include "llvm/IR/IRBuilder.h"
46 #include "llvm/IR/InstrTypes.h"
47 #include "llvm/IR/Instruction.h"
48 #include "llvm/IR/Instructions.h"
49 #include "llvm/IR/IntrinsicInst.h"
50 #include "llvm/IR/Intrinsics.h"
51 #include "llvm/IR/LLVMContext.h"
52 #include "llvm/IR/MDBuilder.h"
53 #include "llvm/IR/Metadata.h"
54 #include "llvm/IR/Module.h"
55 #include "llvm/IR/NoFolder.h"
56 #include "llvm/IR/Operator.h"
57 #include "llvm/IR/PatternMatch.h"
58 #include "llvm/IR/Type.h"
59 #include "llvm/IR/Use.h"
60 #include "llvm/IR/User.h"
61 #include "llvm/IR/Value.h"
62 #include "llvm/Support/Casting.h"
63 #include "llvm/Support/CommandLine.h"
64 #include "llvm/Support/Debug.h"
65 #include "llvm/Support/ErrorHandling.h"
66 #include "llvm/Support/KnownBits.h"
67 #include "llvm/Support/MathExtras.h"
68 #include "llvm/Support/raw_ostream.h"
69 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
70 #include "llvm/Transforms/Utils/Local.h"
71 #include "llvm/Transforms/Utils/ValueMapper.h"
72 #include <algorithm>
73 #include <cassert>
74 #include <climits>
75 #include <cstddef>
76 #include <cstdint>
77 #include <iterator>
78 #include <map>
79 #include <set>
80 #include <tuple>
81 #include <utility>
82 #include <vector>
83
84 using namespace llvm;
85 using namespace PatternMatch;
86
87 #define DEBUG_TYPE "simplifycfg"
88
89 // Chosen as 2 so as to be cheap, but still to have enough power to fold
90 // a select, so the "clamp" idiom (of a min followed by a max) will be caught.
91 // To catch this, we need to fold a compare and a select, hence '2' being the
92 // minimum reasonable default.
93 static cl::opt<unsigned> PHINodeFoldingThreshold(
94 "phi-node-folding-threshold", cl::Hidden, cl::init(2),
95 cl::desc(
96 "Control the amount of phi node folding to perform (default = 2)"));
97
98 static cl::opt<unsigned> TwoEntryPHINodeFoldingThreshold(
99 "two-entry-phi-node-folding-threshold", cl::Hidden, cl::init(4),
100 cl::desc("Control the maximal total instruction cost that we are willing "
101 "to speculatively execute to fold a 2-entry PHI node into a "
102 "select (default = 4)"));
103
104 static cl::opt<bool> DupRet(
105 "simplifycfg-dup-ret", cl::Hidden, cl::init(false),
106 cl::desc("Duplicate return instructions into unconditional branches"));
107
108 static cl::opt<bool>
109 SinkCommon("simplifycfg-sink-common", cl::Hidden, cl::init(true),
110 cl::desc("Sink common instructions down to the end block"));
111
112 static cl::opt<bool> HoistCondStores(
113 "simplifycfg-hoist-cond-stores", cl::Hidden, cl::init(true),
114 cl::desc("Hoist conditional stores if an unconditional store precedes"));
115
116 static cl::opt<bool> MergeCondStores(
117 "simplifycfg-merge-cond-stores", cl::Hidden, cl::init(true),
118 cl::desc("Hoist conditional stores even if an unconditional store does not "
119 "precede - hoist multiple conditional stores into a single "
120 "predicated store"));
121
122 static cl::opt<bool> MergeCondStoresAggressively(
123 "simplifycfg-merge-cond-stores-aggressively", cl::Hidden, cl::init(false),
124 cl::desc("When merging conditional stores, do so even if the resultant "
125 "basic blocks are unlikely to be if-converted as a result"));
126
127 static cl::opt<bool> SpeculateOneExpensiveInst(
128 "speculate-one-expensive-inst", cl::Hidden, cl::init(true),
129 cl::desc("Allow exactly one expensive instruction to be speculatively "
130 "executed"));
131
132 static cl::opt<unsigned> MaxSpeculationDepth(
133 "max-speculation-depth", cl::Hidden, cl::init(10),
134 cl::desc("Limit maximum recursion depth when calculating costs of "
135 "speculatively executed instructions"));
136
137 STATISTIC(NumBitMaps, "Number of switch instructions turned into bitmaps");
138 STATISTIC(NumLinearMaps,
139 "Number of switch instructions turned into linear mapping");
140 STATISTIC(NumLookupTables,
141 "Number of switch instructions turned into lookup tables");
142 STATISTIC(
143 NumLookupTablesHoles,
144 "Number of switch instructions turned into lookup tables (holes checked)");
145 STATISTIC(NumTableCmpReuses, "Number of reused switch table lookup compares");
146 STATISTIC(NumSinkCommons,
147 "Number of common instructions sunk down to the end block");
148 STATISTIC(NumSpeculations, "Number of speculative executed instructions");
149
150 namespace {
151
152 // The first field contains the value that the switch produces when a certain
153 // case group is selected, and the second field is a vector containing the
154 // cases composing the case group.
155 using SwitchCaseResultVectorTy =
156 SmallVector<std::pair<Constant *, SmallVector<ConstantInt *, 4>>, 2>;
157
158 // The first field contains the phi node that generates a result of the switch
159 // and the second field contains the value generated for a certain case in the
160 // switch for that PHI.
161 using SwitchCaseResultsTy = SmallVector<std::pair<PHINode *, Constant *>, 4>;
162
163 /// ValueEqualityComparisonCase - Represents a case of a switch.
164 struct ValueEqualityComparisonCase {
165 ConstantInt *Value;
166 BasicBlock *Dest;
167
ValueEqualityComparisonCase__anon82cdf8f10111::ValueEqualityComparisonCase168 ValueEqualityComparisonCase(ConstantInt *Value, BasicBlock *Dest)
169 : Value(Value), Dest(Dest) {}
170
operator <__anon82cdf8f10111::ValueEqualityComparisonCase171 bool operator<(ValueEqualityComparisonCase RHS) const {
172 // Comparing pointers is ok as we only rely on the order for uniquing.
173 return Value < RHS.Value;
174 }
175
operator ==__anon82cdf8f10111::ValueEqualityComparisonCase176 bool operator==(BasicBlock *RHSDest) const { return Dest == RHSDest; }
177 };
178
179 class SimplifyCFGOpt {
180 const TargetTransformInfo &TTI;
181 const DataLayout &DL;
182 SmallPtrSetImpl<BasicBlock *> *LoopHeaders;
183 const SimplifyCFGOptions &Options;
184 bool Resimplify;
185
186 Value *isValueEqualityComparison(Instruction *TI);
187 BasicBlock *GetValueEqualityComparisonCases(
188 Instruction *TI, std::vector<ValueEqualityComparisonCase> &Cases);
189 bool SimplifyEqualityComparisonWithOnlyPredecessor(Instruction *TI,
190 BasicBlock *Pred,
191 IRBuilder<> &Builder);
192 bool FoldValueComparisonIntoPredecessors(Instruction *TI,
193 IRBuilder<> &Builder);
194
195 bool SimplifyReturn(ReturnInst *RI, IRBuilder<> &Builder);
196 bool SimplifyResume(ResumeInst *RI, IRBuilder<> &Builder);
197 bool SimplifySingleResume(ResumeInst *RI);
198 bool SimplifyCommonResume(ResumeInst *RI);
199 bool SimplifyCleanupReturn(CleanupReturnInst *RI);
200 bool SimplifyUnreachable(UnreachableInst *UI);
201 bool SimplifySwitch(SwitchInst *SI, IRBuilder<> &Builder);
202 bool SimplifyIndirectBr(IndirectBrInst *IBI);
203 bool SimplifyUncondBranch(BranchInst *BI, IRBuilder<> &Builder);
204 bool SimplifyCondBranch(BranchInst *BI, IRBuilder<> &Builder);
205
206 bool tryToSimplifyUncondBranchWithICmpInIt(ICmpInst *ICI,
207 IRBuilder<> &Builder);
208
209 public:
SimplifyCFGOpt(const TargetTransformInfo & TTI,const DataLayout & DL,SmallPtrSetImpl<BasicBlock * > * LoopHeaders,const SimplifyCFGOptions & Opts)210 SimplifyCFGOpt(const TargetTransformInfo &TTI, const DataLayout &DL,
211 SmallPtrSetImpl<BasicBlock *> *LoopHeaders,
212 const SimplifyCFGOptions &Opts)
213 : TTI(TTI), DL(DL), LoopHeaders(LoopHeaders), Options(Opts) {}
214
215 bool run(BasicBlock *BB);
216 bool simplifyOnce(BasicBlock *BB);
217
218 // Helper to set Resimplify and return change indication.
requestResimplify()219 bool requestResimplify() {
220 Resimplify = true;
221 return true;
222 }
223 };
224
225 } // end anonymous namespace
226
227 /// Return true if it is safe to merge these two
228 /// terminator instructions together.
229 static bool
SafeToMergeTerminators(Instruction * SI1,Instruction * SI2,SmallSetVector<BasicBlock *,4> * FailBlocks=nullptr)230 SafeToMergeTerminators(Instruction *SI1, Instruction *SI2,
231 SmallSetVector<BasicBlock *, 4> *FailBlocks = nullptr) {
232 if (SI1 == SI2)
233 return false; // Can't merge with self!
234
235 // It is not safe to merge these two switch instructions if they have a common
236 // successor, and if that successor has a PHI node, and if *that* PHI node has
237 // conflicting incoming values from the two switch blocks.
238 BasicBlock *SI1BB = SI1->getParent();
239 BasicBlock *SI2BB = SI2->getParent();
240
241 SmallPtrSet<BasicBlock *, 16> SI1Succs(succ_begin(SI1BB), succ_end(SI1BB));
242 bool Fail = false;
243 for (BasicBlock *Succ : successors(SI2BB))
244 if (SI1Succs.count(Succ))
245 for (BasicBlock::iterator BBI = Succ->begin(); isa<PHINode>(BBI); ++BBI) {
246 PHINode *PN = cast<PHINode>(BBI);
247 if (PN->getIncomingValueForBlock(SI1BB) !=
248 PN->getIncomingValueForBlock(SI2BB)) {
249 if (FailBlocks)
250 FailBlocks->insert(Succ);
251 Fail = true;
252 }
253 }
254
255 return !Fail;
256 }
257
258 /// Return true if it is safe and profitable to merge these two terminator
259 /// instructions together, where SI1 is an unconditional branch. PhiNodes will
260 /// store all PHI nodes in common successors.
261 static bool
isProfitableToFoldUnconditional(BranchInst * SI1,BranchInst * SI2,Instruction * Cond,SmallVectorImpl<PHINode * > & PhiNodes)262 isProfitableToFoldUnconditional(BranchInst *SI1, BranchInst *SI2,
263 Instruction *Cond,
264 SmallVectorImpl<PHINode *> &PhiNodes) {
265 if (SI1 == SI2)
266 return false; // Can't merge with self!
267 assert(SI1->isUnconditional() && SI2->isConditional());
268
269 // We fold the unconditional branch if we can easily update all PHI nodes in
270 // common successors:
271 // 1> We have a constant incoming value for the conditional branch;
272 // 2> We have "Cond" as the incoming value for the unconditional branch;
273 // 3> SI2->getCondition() and Cond have same operands.
274 CmpInst *Ci2 = dyn_cast<CmpInst>(SI2->getCondition());
275 if (!Ci2)
276 return false;
277 if (!(Cond->getOperand(0) == Ci2->getOperand(0) &&
278 Cond->getOperand(1) == Ci2->getOperand(1)) &&
279 !(Cond->getOperand(0) == Ci2->getOperand(1) &&
280 Cond->getOperand(1) == Ci2->getOperand(0)))
281 return false;
282
283 BasicBlock *SI1BB = SI1->getParent();
284 BasicBlock *SI2BB = SI2->getParent();
285 SmallPtrSet<BasicBlock *, 16> SI1Succs(succ_begin(SI1BB), succ_end(SI1BB));
286 for (BasicBlock *Succ : successors(SI2BB))
287 if (SI1Succs.count(Succ))
288 for (BasicBlock::iterator BBI = Succ->begin(); isa<PHINode>(BBI); ++BBI) {
289 PHINode *PN = cast<PHINode>(BBI);
290 if (PN->getIncomingValueForBlock(SI1BB) != Cond ||
291 !isa<ConstantInt>(PN->getIncomingValueForBlock(SI2BB)))
292 return false;
293 PhiNodes.push_back(PN);
294 }
295 return true;
296 }
297
298 /// Update PHI nodes in Succ to indicate that there will now be entries in it
299 /// from the 'NewPred' block. The values that will be flowing into the PHI nodes
300 /// will be the same as those coming in from ExistPred, an existing predecessor
301 /// of Succ.
AddPredecessorToBlock(BasicBlock * Succ,BasicBlock * NewPred,BasicBlock * ExistPred,MemorySSAUpdater * MSSAU=nullptr)302 static void AddPredecessorToBlock(BasicBlock *Succ, BasicBlock *NewPred,
303 BasicBlock *ExistPred,
304 MemorySSAUpdater *MSSAU = nullptr) {
305 for (PHINode &PN : Succ->phis())
306 PN.addIncoming(PN.getIncomingValueForBlock(ExistPred), NewPred);
307 if (MSSAU)
308 if (auto *MPhi = MSSAU->getMemorySSA()->getMemoryAccess(Succ))
309 MPhi->addIncoming(MPhi->getIncomingValueForBlock(ExistPred), NewPred);
310 }
311
312 /// Compute an abstract "cost" of speculating the given instruction,
313 /// which is assumed to be safe to speculate. TCC_Free means cheap,
314 /// TCC_Basic means less cheap, and TCC_Expensive means prohibitively
315 /// expensive.
ComputeSpeculationCost(const User * I,const TargetTransformInfo & TTI)316 static unsigned ComputeSpeculationCost(const User *I,
317 const TargetTransformInfo &TTI) {
318 assert(isSafeToSpeculativelyExecute(I) &&
319 "Instruction is not safe to speculatively execute!");
320 return TTI.getUserCost(I);
321 }
322
323 /// If we have a merge point of an "if condition" as accepted above,
324 /// return true if the specified value dominates the block. We
325 /// don't handle the true generality of domination here, just a special case
326 /// which works well enough for us.
327 ///
328 /// If AggressiveInsts is non-null, and if V does not dominate BB, we check to
329 /// see if V (which must be an instruction) and its recursive operands
330 /// that do not dominate BB have a combined cost lower than CostRemaining and
331 /// are non-trapping. If both are true, the instruction is inserted into the
332 /// set and true is returned.
333 ///
334 /// The cost for most non-trapping instructions is defined as 1 except for
335 /// Select whose cost is 2.
336 ///
337 /// After this function returns, CostRemaining is decreased by the cost of
338 /// V plus its non-dominating operands. If that cost is greater than
339 /// CostRemaining, false is returned and CostRemaining is undefined.
DominatesMergePoint(Value * V,BasicBlock * BB,SmallPtrSetImpl<Instruction * > & AggressiveInsts,int & BudgetRemaining,const TargetTransformInfo & TTI,unsigned Depth=0)340 static bool DominatesMergePoint(Value *V, BasicBlock *BB,
341 SmallPtrSetImpl<Instruction *> &AggressiveInsts,
342 int &BudgetRemaining,
343 const TargetTransformInfo &TTI,
344 unsigned Depth = 0) {
345 // It is possible to hit a zero-cost cycle (phi/gep instructions for example),
346 // so limit the recursion depth.
347 // TODO: While this recursion limit does prevent pathological behavior, it
348 // would be better to track visited instructions to avoid cycles.
349 if (Depth == MaxSpeculationDepth)
350 return false;
351
352 Instruction *I = dyn_cast<Instruction>(V);
353 if (!I) {
354 // Non-instructions all dominate instructions, but not all constantexprs
355 // can be executed unconditionally.
356 if (ConstantExpr *C = dyn_cast<ConstantExpr>(V))
357 if (C->canTrap())
358 return false;
359 return true;
360 }
361 BasicBlock *PBB = I->getParent();
362
363 // We don't want to allow weird loops that might have the "if condition" in
364 // the bottom of this block.
365 if (PBB == BB)
366 return false;
367
368 // If this instruction is defined in a block that contains an unconditional
369 // branch to BB, then it must be in the 'conditional' part of the "if
370 // statement". If not, it definitely dominates the region.
371 BranchInst *BI = dyn_cast<BranchInst>(PBB->getTerminator());
372 if (!BI || BI->isConditional() || BI->getSuccessor(0) != BB)
373 return true;
374
375 // If we have seen this instruction before, don't count it again.
376 if (AggressiveInsts.count(I))
377 return true;
378
379 // Okay, it looks like the instruction IS in the "condition". Check to
380 // see if it's a cheap instruction to unconditionally compute, and if it
381 // only uses stuff defined outside of the condition. If so, hoist it out.
382 if (!isSafeToSpeculativelyExecute(I))
383 return false;
384
385 BudgetRemaining -= ComputeSpeculationCost(I, TTI);
386
387 // Allow exactly one instruction to be speculated regardless of its cost
388 // (as long as it is safe to do so).
389 // This is intended to flatten the CFG even if the instruction is a division
390 // or other expensive operation. The speculation of an expensive instruction
391 // is expected to be undone in CodeGenPrepare if the speculation has not
392 // enabled further IR optimizations.
393 if (BudgetRemaining < 0 &&
394 (!SpeculateOneExpensiveInst || !AggressiveInsts.empty() || Depth > 0))
395 return false;
396
397 // Okay, we can only really hoist these out if their operands do
398 // not take us over the cost threshold.
399 for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i)
400 if (!DominatesMergePoint(*i, BB, AggressiveInsts, BudgetRemaining, TTI,
401 Depth + 1))
402 return false;
403 // Okay, it's safe to do this! Remember this instruction.
404 AggressiveInsts.insert(I);
405 return true;
406 }
407
408 /// Extract ConstantInt from value, looking through IntToPtr
409 /// and PointerNullValue. Return NULL if value is not a constant int.
GetConstantInt(Value * V,const DataLayout & DL)410 static ConstantInt *GetConstantInt(Value *V, const DataLayout &DL) {
411 // Normal constant int.
412 ConstantInt *CI = dyn_cast<ConstantInt>(V);
413 if (CI || !isa<Constant>(V) || !V->getType()->isPointerTy())
414 return CI;
415
416 // This is some kind of pointer constant. Turn it into a pointer-sized
417 // ConstantInt if possible.
418 IntegerType *PtrTy = cast<IntegerType>(DL.getIntPtrType(V->getType()));
419
420 // Null pointer means 0, see SelectionDAGBuilder::getValue(const Value*).
421 if (isa<ConstantPointerNull>(V))
422 return ConstantInt::get(PtrTy, 0);
423
424 // IntToPtr const int.
425 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
426 if (CE->getOpcode() == Instruction::IntToPtr)
427 if (ConstantInt *CI = dyn_cast<ConstantInt>(CE->getOperand(0))) {
428 // The constant is very likely to have the right type already.
429 if (CI->getType() == PtrTy)
430 return CI;
431 else
432 return cast<ConstantInt>(
433 ConstantExpr::getIntegerCast(CI, PtrTy, /*isSigned=*/false));
434 }
435 return nullptr;
436 }
437
438 namespace {
439
440 /// Given a chain of or (||) or and (&&) comparison of a value against a
441 /// constant, this will try to recover the information required for a switch
442 /// structure.
443 /// It will depth-first traverse the chain of comparison, seeking for patterns
444 /// like %a == 12 or %a < 4 and combine them to produce a set of integer
445 /// representing the different cases for the switch.
446 /// Note that if the chain is composed of '||' it will build the set of elements
447 /// that matches the comparisons (i.e. any of this value validate the chain)
448 /// while for a chain of '&&' it will build the set elements that make the test
449 /// fail.
450 struct ConstantComparesGatherer {
451 const DataLayout &DL;
452
453 /// Value found for the switch comparison
454 Value *CompValue = nullptr;
455
456 /// Extra clause to be checked before the switch
457 Value *Extra = nullptr;
458
459 /// Set of integers to match in switch
460 SmallVector<ConstantInt *, 8> Vals;
461
462 /// Number of comparisons matched in the and/or chain
463 unsigned UsedICmps = 0;
464
465 /// Construct and compute the result for the comparison instruction Cond
ConstantComparesGatherer__anon82cdf8f10211::ConstantComparesGatherer466 ConstantComparesGatherer(Instruction *Cond, const DataLayout &DL) : DL(DL) {
467 gather(Cond);
468 }
469
470 ConstantComparesGatherer(const ConstantComparesGatherer &) = delete;
471 ConstantComparesGatherer &
472 operator=(const ConstantComparesGatherer &) = delete;
473
474 private:
475 /// Try to set the current value used for the comparison, it succeeds only if
476 /// it wasn't set before or if the new value is the same as the old one
setValueOnce__anon82cdf8f10211::ConstantComparesGatherer477 bool setValueOnce(Value *NewVal) {
478 if (CompValue && CompValue != NewVal)
479 return false;
480 CompValue = NewVal;
481 return (CompValue != nullptr);
482 }
483
484 /// Try to match Instruction "I" as a comparison against a constant and
485 /// populates the array Vals with the set of values that match (or do not
486 /// match depending on isEQ).
487 /// Return false on failure. On success, the Value the comparison matched
488 /// against is placed in CompValue.
489 /// If CompValue is already set, the function is expected to fail if a match
490 /// is found but the value compared to is different.
matchInstruction__anon82cdf8f10211::ConstantComparesGatherer491 bool matchInstruction(Instruction *I, bool isEQ) {
492 // If this is an icmp against a constant, handle this as one of the cases.
493 ICmpInst *ICI;
494 ConstantInt *C;
495 if (!((ICI = dyn_cast<ICmpInst>(I)) &&
496 (C = GetConstantInt(I->getOperand(1), DL)))) {
497 return false;
498 }
499
500 Value *RHSVal;
501 const APInt *RHSC;
502
503 // Pattern match a special case
504 // (x & ~2^z) == y --> x == y || x == y|2^z
505 // This undoes a transformation done by instcombine to fuse 2 compares.
506 if (ICI->getPredicate() == (isEQ ? ICmpInst::ICMP_EQ : ICmpInst::ICMP_NE)) {
507 // It's a little bit hard to see why the following transformations are
508 // correct. Here is a CVC3 program to verify them for 64-bit values:
509
510 /*
511 ONE : BITVECTOR(64) = BVZEROEXTEND(0bin1, 63);
512 x : BITVECTOR(64);
513 y : BITVECTOR(64);
514 z : BITVECTOR(64);
515 mask : BITVECTOR(64) = BVSHL(ONE, z);
516 QUERY( (y & ~mask = y) =>
517 ((x & ~mask = y) <=> (x = y OR x = (y | mask)))
518 );
519 QUERY( (y | mask = y) =>
520 ((x | mask = y) <=> (x = y OR x = (y & ~mask)))
521 );
522 */
523
524 // Please note that each pattern must be a dual implication (<--> or
525 // iff). One directional implication can create spurious matches. If the
526 // implication is only one-way, an unsatisfiable condition on the left
527 // side can imply a satisfiable condition on the right side. Dual
528 // implication ensures that satisfiable conditions are transformed to
529 // other satisfiable conditions and unsatisfiable conditions are
530 // transformed to other unsatisfiable conditions.
531
532 // Here is a concrete example of a unsatisfiable condition on the left
533 // implying a satisfiable condition on the right:
534 //
535 // mask = (1 << z)
536 // (x & ~mask) == y --> (x == y || x == (y | mask))
537 //
538 // Substituting y = 3, z = 0 yields:
539 // (x & -2) == 3 --> (x == 3 || x == 2)
540
541 // Pattern match a special case:
542 /*
543 QUERY( (y & ~mask = y) =>
544 ((x & ~mask = y) <=> (x = y OR x = (y | mask)))
545 );
546 */
547 if (match(ICI->getOperand(0),
548 m_And(m_Value(RHSVal), m_APInt(RHSC)))) {
549 APInt Mask = ~*RHSC;
550 if (Mask.isPowerOf2() && (C->getValue() & ~Mask) == C->getValue()) {
551 // If we already have a value for the switch, it has to match!
552 if (!setValueOnce(RHSVal))
553 return false;
554
555 Vals.push_back(C);
556 Vals.push_back(
557 ConstantInt::get(C->getContext(),
558 C->getValue() | Mask));
559 UsedICmps++;
560 return true;
561 }
562 }
563
564 // Pattern match a special case:
565 /*
566 QUERY( (y | mask = y) =>
567 ((x | mask = y) <=> (x = y OR x = (y & ~mask)))
568 );
569 */
570 if (match(ICI->getOperand(0),
571 m_Or(m_Value(RHSVal), m_APInt(RHSC)))) {
572 APInt Mask = *RHSC;
573 if (Mask.isPowerOf2() && (C->getValue() | Mask) == C->getValue()) {
574 // If we already have a value for the switch, it has to match!
575 if (!setValueOnce(RHSVal))
576 return false;
577
578 Vals.push_back(C);
579 Vals.push_back(ConstantInt::get(C->getContext(),
580 C->getValue() & ~Mask));
581 UsedICmps++;
582 return true;
583 }
584 }
585
586 // If we already have a value for the switch, it has to match!
587 if (!setValueOnce(ICI->getOperand(0)))
588 return false;
589
590 UsedICmps++;
591 Vals.push_back(C);
592 return ICI->getOperand(0);
593 }
594
595 // If we have "x ult 3", for example, then we can add 0,1,2 to the set.
596 ConstantRange Span = ConstantRange::makeAllowedICmpRegion(
597 ICI->getPredicate(), C->getValue());
598
599 // Shift the range if the compare is fed by an add. This is the range
600 // compare idiom as emitted by instcombine.
601 Value *CandidateVal = I->getOperand(0);
602 if (match(I->getOperand(0), m_Add(m_Value(RHSVal), m_APInt(RHSC)))) {
603 Span = Span.subtract(*RHSC);
604 CandidateVal = RHSVal;
605 }
606
607 // If this is an and/!= check, then we are looking to build the set of
608 // value that *don't* pass the and chain. I.e. to turn "x ugt 2" into
609 // x != 0 && x != 1.
610 if (!isEQ)
611 Span = Span.inverse();
612
613 // If there are a ton of values, we don't want to make a ginormous switch.
614 if (Span.isSizeLargerThan(8) || Span.isEmptySet()) {
615 return false;
616 }
617
618 // If we already have a value for the switch, it has to match!
619 if (!setValueOnce(CandidateVal))
620 return false;
621
622 // Add all values from the range to the set
623 for (APInt Tmp = Span.getLower(); Tmp != Span.getUpper(); ++Tmp)
624 Vals.push_back(ConstantInt::get(I->getContext(), Tmp));
625
626 UsedICmps++;
627 return true;
628 }
629
630 /// Given a potentially 'or'd or 'and'd together collection of icmp
631 /// eq/ne/lt/gt instructions that compare a value against a constant, extract
632 /// the value being compared, and stick the list constants into the Vals
633 /// vector.
634 /// One "Extra" case is allowed to differ from the other.
gather__anon82cdf8f10211::ConstantComparesGatherer635 void gather(Value *V) {
636 bool isEQ = (cast<Instruction>(V)->getOpcode() == Instruction::Or);
637
638 // Keep a stack (SmallVector for efficiency) for depth-first traversal
639 SmallVector<Value *, 8> DFT;
640 SmallPtrSet<Value *, 8> Visited;
641
642 // Initialize
643 Visited.insert(V);
644 DFT.push_back(V);
645
646 while (!DFT.empty()) {
647 V = DFT.pop_back_val();
648
649 if (Instruction *I = dyn_cast<Instruction>(V)) {
650 // If it is a || (or && depending on isEQ), process the operands.
651 if (I->getOpcode() == (isEQ ? Instruction::Or : Instruction::And)) {
652 if (Visited.insert(I->getOperand(1)).second)
653 DFT.push_back(I->getOperand(1));
654 if (Visited.insert(I->getOperand(0)).second)
655 DFT.push_back(I->getOperand(0));
656 continue;
657 }
658
659 // Try to match the current instruction
660 if (matchInstruction(I, isEQ))
661 // Match succeed, continue the loop
662 continue;
663 }
664
665 // One element of the sequence of || (or &&) could not be match as a
666 // comparison against the same value as the others.
667 // We allow only one "Extra" case to be checked before the switch
668 if (!Extra) {
669 Extra = V;
670 continue;
671 }
672 // Failed to parse a proper sequence, abort now
673 CompValue = nullptr;
674 break;
675 }
676 }
677 };
678
679 } // end anonymous namespace
680
EraseTerminatorAndDCECond(Instruction * TI,MemorySSAUpdater * MSSAU=nullptr)681 static void EraseTerminatorAndDCECond(Instruction *TI,
682 MemorySSAUpdater *MSSAU = nullptr) {
683 Instruction *Cond = nullptr;
684 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
685 Cond = dyn_cast<Instruction>(SI->getCondition());
686 } else if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
687 if (BI->isConditional())
688 Cond = dyn_cast<Instruction>(BI->getCondition());
689 } else if (IndirectBrInst *IBI = dyn_cast<IndirectBrInst>(TI)) {
690 Cond = dyn_cast<Instruction>(IBI->getAddress());
691 }
692
693 TI->eraseFromParent();
694 if (Cond)
695 RecursivelyDeleteTriviallyDeadInstructions(Cond, nullptr, MSSAU);
696 }
697
698 /// Return true if the specified terminator checks
699 /// to see if a value is equal to constant integer value.
isValueEqualityComparison(Instruction * TI)700 Value *SimplifyCFGOpt::isValueEqualityComparison(Instruction *TI) {
701 Value *CV = nullptr;
702 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
703 // Do not permit merging of large switch instructions into their
704 // predecessors unless there is only one predecessor.
705 if (!SI->getParent()->hasNPredecessorsOrMore(128 / SI->getNumSuccessors()))
706 CV = SI->getCondition();
707 } else if (BranchInst *BI = dyn_cast<BranchInst>(TI))
708 if (BI->isConditional() && BI->getCondition()->hasOneUse())
709 if (ICmpInst *ICI = dyn_cast<ICmpInst>(BI->getCondition())) {
710 if (ICI->isEquality() && GetConstantInt(ICI->getOperand(1), DL))
711 CV = ICI->getOperand(0);
712 }
713
714 // Unwrap any lossless ptrtoint cast.
715 if (CV) {
716 if (PtrToIntInst *PTII = dyn_cast<PtrToIntInst>(CV)) {
717 Value *Ptr = PTII->getPointerOperand();
718 if (PTII->getType() == DL.getIntPtrType(Ptr->getType()))
719 CV = Ptr;
720 }
721 }
722 return CV;
723 }
724
725 /// Given a value comparison instruction,
726 /// decode all of the 'cases' that it represents and return the 'default' block.
GetValueEqualityComparisonCases(Instruction * TI,std::vector<ValueEqualityComparisonCase> & Cases)727 BasicBlock *SimplifyCFGOpt::GetValueEqualityComparisonCases(
728 Instruction *TI, std::vector<ValueEqualityComparisonCase> &Cases) {
729 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
730 Cases.reserve(SI->getNumCases());
731 for (auto Case : SI->cases())
732 Cases.push_back(ValueEqualityComparisonCase(Case.getCaseValue(),
733 Case.getCaseSuccessor()));
734 return SI->getDefaultDest();
735 }
736
737 BranchInst *BI = cast<BranchInst>(TI);
738 ICmpInst *ICI = cast<ICmpInst>(BI->getCondition());
739 BasicBlock *Succ = BI->getSuccessor(ICI->getPredicate() == ICmpInst::ICMP_NE);
740 Cases.push_back(ValueEqualityComparisonCase(
741 GetConstantInt(ICI->getOperand(1), DL), Succ));
742 return BI->getSuccessor(ICI->getPredicate() == ICmpInst::ICMP_EQ);
743 }
744
745 /// Given a vector of bb/value pairs, remove any entries
746 /// in the list that match the specified block.
747 static void
EliminateBlockCases(BasicBlock * BB,std::vector<ValueEqualityComparisonCase> & Cases)748 EliminateBlockCases(BasicBlock *BB,
749 std::vector<ValueEqualityComparisonCase> &Cases) {
750 Cases.erase(std::remove(Cases.begin(), Cases.end(), BB), Cases.end());
751 }
752
753 /// Return true if there are any keys in C1 that exist in C2 as well.
ValuesOverlap(std::vector<ValueEqualityComparisonCase> & C1,std::vector<ValueEqualityComparisonCase> & C2)754 static bool ValuesOverlap(std::vector<ValueEqualityComparisonCase> &C1,
755 std::vector<ValueEqualityComparisonCase> &C2) {
756 std::vector<ValueEqualityComparisonCase> *V1 = &C1, *V2 = &C2;
757
758 // Make V1 be smaller than V2.
759 if (V1->size() > V2->size())
760 std::swap(V1, V2);
761
762 if (V1->empty())
763 return false;
764 if (V1->size() == 1) {
765 // Just scan V2.
766 ConstantInt *TheVal = (*V1)[0].Value;
767 for (unsigned i = 0, e = V2->size(); i != e; ++i)
768 if (TheVal == (*V2)[i].Value)
769 return true;
770 }
771
772 // Otherwise, just sort both lists and compare element by element.
773 array_pod_sort(V1->begin(), V1->end());
774 array_pod_sort(V2->begin(), V2->end());
775 unsigned i1 = 0, i2 = 0, e1 = V1->size(), e2 = V2->size();
776 while (i1 != e1 && i2 != e2) {
777 if ((*V1)[i1].Value == (*V2)[i2].Value)
778 return true;
779 if ((*V1)[i1].Value < (*V2)[i2].Value)
780 ++i1;
781 else
782 ++i2;
783 }
784 return false;
785 }
786
787 // Set branch weights on SwitchInst. This sets the metadata if there is at
788 // least one non-zero weight.
setBranchWeights(SwitchInst * SI,ArrayRef<uint32_t> Weights)789 static void setBranchWeights(SwitchInst *SI, ArrayRef<uint32_t> Weights) {
790 // Check that there is at least one non-zero weight. Otherwise, pass
791 // nullptr to setMetadata which will erase the existing metadata.
792 MDNode *N = nullptr;
793 if (llvm::any_of(Weights, [](uint32_t W) { return W != 0; }))
794 N = MDBuilder(SI->getParent()->getContext()).createBranchWeights(Weights);
795 SI->setMetadata(LLVMContext::MD_prof, N);
796 }
797
798 // Similar to the above, but for branch and select instructions that take
799 // exactly 2 weights.
setBranchWeights(Instruction * I,uint32_t TrueWeight,uint32_t FalseWeight)800 static void setBranchWeights(Instruction *I, uint32_t TrueWeight,
801 uint32_t FalseWeight) {
802 assert(isa<BranchInst>(I) || isa<SelectInst>(I));
803 // Check that there is at least one non-zero weight. Otherwise, pass
804 // nullptr to setMetadata which will erase the existing metadata.
805 MDNode *N = nullptr;
806 if (TrueWeight || FalseWeight)
807 N = MDBuilder(I->getParent()->getContext())
808 .createBranchWeights(TrueWeight, FalseWeight);
809 I->setMetadata(LLVMContext::MD_prof, N);
810 }
811
812 /// If TI is known to be a terminator instruction and its block is known to
813 /// only have a single predecessor block, check to see if that predecessor is
814 /// also a value comparison with the same value, and if that comparison
815 /// determines the outcome of this comparison. If so, simplify TI. This does a
816 /// very limited form of jump threading.
SimplifyEqualityComparisonWithOnlyPredecessor(Instruction * TI,BasicBlock * Pred,IRBuilder<> & Builder)817 bool SimplifyCFGOpt::SimplifyEqualityComparisonWithOnlyPredecessor(
818 Instruction *TI, BasicBlock *Pred, IRBuilder<> &Builder) {
819 Value *PredVal = isValueEqualityComparison(Pred->getTerminator());
820 if (!PredVal)
821 return false; // Not a value comparison in predecessor.
822
823 Value *ThisVal = isValueEqualityComparison(TI);
824 assert(ThisVal && "This isn't a value comparison!!");
825 if (ThisVal != PredVal)
826 return false; // Different predicates.
827
828 // TODO: Preserve branch weight metadata, similarly to how
829 // FoldValueComparisonIntoPredecessors preserves it.
830
831 // Find out information about when control will move from Pred to TI's block.
832 std::vector<ValueEqualityComparisonCase> PredCases;
833 BasicBlock *PredDef =
834 GetValueEqualityComparisonCases(Pred->getTerminator(), PredCases);
835 EliminateBlockCases(PredDef, PredCases); // Remove default from cases.
836
837 // Find information about how control leaves this block.
838 std::vector<ValueEqualityComparisonCase> ThisCases;
839 BasicBlock *ThisDef = GetValueEqualityComparisonCases(TI, ThisCases);
840 EliminateBlockCases(ThisDef, ThisCases); // Remove default from cases.
841
842 // If TI's block is the default block from Pred's comparison, potentially
843 // simplify TI based on this knowledge.
844 if (PredDef == TI->getParent()) {
845 // If we are here, we know that the value is none of those cases listed in
846 // PredCases. If there are any cases in ThisCases that are in PredCases, we
847 // can simplify TI.
848 if (!ValuesOverlap(PredCases, ThisCases))
849 return false;
850
851 if (isa<BranchInst>(TI)) {
852 // Okay, one of the successors of this condbr is dead. Convert it to a
853 // uncond br.
854 assert(ThisCases.size() == 1 && "Branch can only have one case!");
855 // Insert the new branch.
856 Instruction *NI = Builder.CreateBr(ThisDef);
857 (void)NI;
858
859 // Remove PHI node entries for the dead edge.
860 ThisCases[0].Dest->removePredecessor(TI->getParent());
861
862 LLVM_DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator()
863 << "Through successor TI: " << *TI << "Leaving: " << *NI
864 << "\n");
865
866 EraseTerminatorAndDCECond(TI);
867 return true;
868 }
869
870 SwitchInstProfUpdateWrapper SI = *cast<SwitchInst>(TI);
871 // Okay, TI has cases that are statically dead, prune them away.
872 SmallPtrSet<Constant *, 16> DeadCases;
873 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
874 DeadCases.insert(PredCases[i].Value);
875
876 LLVM_DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator()
877 << "Through successor TI: " << *TI);
878
879 for (SwitchInst::CaseIt i = SI->case_end(), e = SI->case_begin(); i != e;) {
880 --i;
881 if (DeadCases.count(i->getCaseValue())) {
882 i->getCaseSuccessor()->removePredecessor(TI->getParent());
883 SI.removeCase(i);
884 }
885 }
886 LLVM_DEBUG(dbgs() << "Leaving: " << *TI << "\n");
887 return true;
888 }
889
890 // Otherwise, TI's block must correspond to some matched value. Find out
891 // which value (or set of values) this is.
892 ConstantInt *TIV = nullptr;
893 BasicBlock *TIBB = TI->getParent();
894 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
895 if (PredCases[i].Dest == TIBB) {
896 if (TIV)
897 return false; // Cannot handle multiple values coming to this block.
898 TIV = PredCases[i].Value;
899 }
900 assert(TIV && "No edge from pred to succ?");
901
902 // Okay, we found the one constant that our value can be if we get into TI's
903 // BB. Find out which successor will unconditionally be branched to.
904 BasicBlock *TheRealDest = nullptr;
905 for (unsigned i = 0, e = ThisCases.size(); i != e; ++i)
906 if (ThisCases[i].Value == TIV) {
907 TheRealDest = ThisCases[i].Dest;
908 break;
909 }
910
911 // If not handled by any explicit cases, it is handled by the default case.
912 if (!TheRealDest)
913 TheRealDest = ThisDef;
914
915 // Remove PHI node entries for dead edges.
916 BasicBlock *CheckEdge = TheRealDest;
917 for (BasicBlock *Succ : successors(TIBB))
918 if (Succ != CheckEdge)
919 Succ->removePredecessor(TIBB);
920 else
921 CheckEdge = nullptr;
922
923 // Insert the new branch.
924 Instruction *NI = Builder.CreateBr(TheRealDest);
925 (void)NI;
926
927 LLVM_DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator()
928 << "Through successor TI: " << *TI << "Leaving: " << *NI
929 << "\n");
930
931 EraseTerminatorAndDCECond(TI);
932 return true;
933 }
934
935 namespace {
936
937 /// This class implements a stable ordering of constant
938 /// integers that does not depend on their address. This is important for
939 /// applications that sort ConstantInt's to ensure uniqueness.
940 struct ConstantIntOrdering {
operator ()__anon82cdf8f10411::ConstantIntOrdering941 bool operator()(const ConstantInt *LHS, const ConstantInt *RHS) const {
942 return LHS->getValue().ult(RHS->getValue());
943 }
944 };
945
946 } // end anonymous namespace
947
ConstantIntSortPredicate(ConstantInt * const * P1,ConstantInt * const * P2)948 static int ConstantIntSortPredicate(ConstantInt *const *P1,
949 ConstantInt *const *P2) {
950 const ConstantInt *LHS = *P1;
951 const ConstantInt *RHS = *P2;
952 if (LHS == RHS)
953 return 0;
954 return LHS->getValue().ult(RHS->getValue()) ? 1 : -1;
955 }
956
HasBranchWeights(const Instruction * I)957 static inline bool HasBranchWeights(const Instruction *I) {
958 MDNode *ProfMD = I->getMetadata(LLVMContext::MD_prof);
959 if (ProfMD && ProfMD->getOperand(0))
960 if (MDString *MDS = dyn_cast<MDString>(ProfMD->getOperand(0)))
961 return MDS->getString().equals("branch_weights");
962
963 return false;
964 }
965
966 /// Get Weights of a given terminator, the default weight is at the front
967 /// of the vector. If TI is a conditional eq, we need to swap the branch-weight
968 /// metadata.
GetBranchWeights(Instruction * TI,SmallVectorImpl<uint64_t> & Weights)969 static void GetBranchWeights(Instruction *TI,
970 SmallVectorImpl<uint64_t> &Weights) {
971 MDNode *MD = TI->getMetadata(LLVMContext::MD_prof);
972 assert(MD);
973 for (unsigned i = 1, e = MD->getNumOperands(); i < e; ++i) {
974 ConstantInt *CI = mdconst::extract<ConstantInt>(MD->getOperand(i));
975 Weights.push_back(CI->getValue().getZExtValue());
976 }
977
978 // If TI is a conditional eq, the default case is the false case,
979 // and the corresponding branch-weight data is at index 2. We swap the
980 // default weight to be the first entry.
981 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
982 assert(Weights.size() == 2);
983 ICmpInst *ICI = cast<ICmpInst>(BI->getCondition());
984 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
985 std::swap(Weights.front(), Weights.back());
986 }
987 }
988
989 /// Keep halving the weights until all can fit in uint32_t.
FitWeights(MutableArrayRef<uint64_t> Weights)990 static void FitWeights(MutableArrayRef<uint64_t> Weights) {
991 uint64_t Max = *std::max_element(Weights.begin(), Weights.end());
992 if (Max > UINT_MAX) {
993 unsigned Offset = 32 - countLeadingZeros(Max);
994 for (uint64_t &I : Weights)
995 I >>= Offset;
996 }
997 }
998
999 /// The specified terminator is a value equality comparison instruction
1000 /// (either a switch or a branch on "X == c").
1001 /// See if any of the predecessors of the terminator block are value comparisons
1002 /// on the same value. If so, and if safe to do so, fold them together.
FoldValueComparisonIntoPredecessors(Instruction * TI,IRBuilder<> & Builder)1003 bool SimplifyCFGOpt::FoldValueComparisonIntoPredecessors(Instruction *TI,
1004 IRBuilder<> &Builder) {
1005 BasicBlock *BB = TI->getParent();
1006 Value *CV = isValueEqualityComparison(TI); // CondVal
1007 assert(CV && "Not a comparison?");
1008 bool Changed = false;
1009
1010 SmallVector<BasicBlock *, 16> Preds(pred_begin(BB), pred_end(BB));
1011 while (!Preds.empty()) {
1012 BasicBlock *Pred = Preds.pop_back_val();
1013
1014 // See if the predecessor is a comparison with the same value.
1015 Instruction *PTI = Pred->getTerminator();
1016 Value *PCV = isValueEqualityComparison(PTI); // PredCondVal
1017
1018 if (PCV == CV && TI != PTI) {
1019 SmallSetVector<BasicBlock*, 4> FailBlocks;
1020 if (!SafeToMergeTerminators(TI, PTI, &FailBlocks)) {
1021 for (auto *Succ : FailBlocks) {
1022 if (!SplitBlockPredecessors(Succ, TI->getParent(), ".fold.split"))
1023 return false;
1024 }
1025 }
1026
1027 // Figure out which 'cases' to copy from SI to PSI.
1028 std::vector<ValueEqualityComparisonCase> BBCases;
1029 BasicBlock *BBDefault = GetValueEqualityComparisonCases(TI, BBCases);
1030
1031 std::vector<ValueEqualityComparisonCase> PredCases;
1032 BasicBlock *PredDefault = GetValueEqualityComparisonCases(PTI, PredCases);
1033
1034 // Based on whether the default edge from PTI goes to BB or not, fill in
1035 // PredCases and PredDefault with the new switch cases we would like to
1036 // build.
1037 SmallVector<BasicBlock *, 8> NewSuccessors;
1038
1039 // Update the branch weight metadata along the way
1040 SmallVector<uint64_t, 8> Weights;
1041 bool PredHasWeights = HasBranchWeights(PTI);
1042 bool SuccHasWeights = HasBranchWeights(TI);
1043
1044 if (PredHasWeights) {
1045 GetBranchWeights(PTI, Weights);
1046 // branch-weight metadata is inconsistent here.
1047 if (Weights.size() != 1 + PredCases.size())
1048 PredHasWeights = SuccHasWeights = false;
1049 } else if (SuccHasWeights)
1050 // If there are no predecessor weights but there are successor weights,
1051 // populate Weights with 1, which will later be scaled to the sum of
1052 // successor's weights
1053 Weights.assign(1 + PredCases.size(), 1);
1054
1055 SmallVector<uint64_t, 8> SuccWeights;
1056 if (SuccHasWeights) {
1057 GetBranchWeights(TI, SuccWeights);
1058 // branch-weight metadata is inconsistent here.
1059 if (SuccWeights.size() != 1 + BBCases.size())
1060 PredHasWeights = SuccHasWeights = false;
1061 } else if (PredHasWeights)
1062 SuccWeights.assign(1 + BBCases.size(), 1);
1063
1064 if (PredDefault == BB) {
1065 // If this is the default destination from PTI, only the edges in TI
1066 // that don't occur in PTI, or that branch to BB will be activated.
1067 std::set<ConstantInt *, ConstantIntOrdering> PTIHandled;
1068 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
1069 if (PredCases[i].Dest != BB)
1070 PTIHandled.insert(PredCases[i].Value);
1071 else {
1072 // The default destination is BB, we don't need explicit targets.
1073 std::swap(PredCases[i], PredCases.back());
1074
1075 if (PredHasWeights || SuccHasWeights) {
1076 // Increase weight for the default case.
1077 Weights[0] += Weights[i + 1];
1078 std::swap(Weights[i + 1], Weights.back());
1079 Weights.pop_back();
1080 }
1081
1082 PredCases.pop_back();
1083 --i;
1084 --e;
1085 }
1086
1087 // Reconstruct the new switch statement we will be building.
1088 if (PredDefault != BBDefault) {
1089 PredDefault->removePredecessor(Pred);
1090 PredDefault = BBDefault;
1091 NewSuccessors.push_back(BBDefault);
1092 }
1093
1094 unsigned CasesFromPred = Weights.size();
1095 uint64_t ValidTotalSuccWeight = 0;
1096 for (unsigned i = 0, e = BBCases.size(); i != e; ++i)
1097 if (!PTIHandled.count(BBCases[i].Value) &&
1098 BBCases[i].Dest != BBDefault) {
1099 PredCases.push_back(BBCases[i]);
1100 NewSuccessors.push_back(BBCases[i].Dest);
1101 if (SuccHasWeights || PredHasWeights) {
1102 // The default weight is at index 0, so weight for the ith case
1103 // should be at index i+1. Scale the cases from successor by
1104 // PredDefaultWeight (Weights[0]).
1105 Weights.push_back(Weights[0] * SuccWeights[i + 1]);
1106 ValidTotalSuccWeight += SuccWeights[i + 1];
1107 }
1108 }
1109
1110 if (SuccHasWeights || PredHasWeights) {
1111 ValidTotalSuccWeight += SuccWeights[0];
1112 // Scale the cases from predecessor by ValidTotalSuccWeight.
1113 for (unsigned i = 1; i < CasesFromPred; ++i)
1114 Weights[i] *= ValidTotalSuccWeight;
1115 // Scale the default weight by SuccDefaultWeight (SuccWeights[0]).
1116 Weights[0] *= SuccWeights[0];
1117 }
1118 } else {
1119 // If this is not the default destination from PSI, only the edges
1120 // in SI that occur in PSI with a destination of BB will be
1121 // activated.
1122 std::set<ConstantInt *, ConstantIntOrdering> PTIHandled;
1123 std::map<ConstantInt *, uint64_t> WeightsForHandled;
1124 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
1125 if (PredCases[i].Dest == BB) {
1126 PTIHandled.insert(PredCases[i].Value);
1127
1128 if (PredHasWeights || SuccHasWeights) {
1129 WeightsForHandled[PredCases[i].Value] = Weights[i + 1];
1130 std::swap(Weights[i + 1], Weights.back());
1131 Weights.pop_back();
1132 }
1133
1134 std::swap(PredCases[i], PredCases.back());
1135 PredCases.pop_back();
1136 --i;
1137 --e;
1138 }
1139
1140 // Okay, now we know which constants were sent to BB from the
1141 // predecessor. Figure out where they will all go now.
1142 for (unsigned i = 0, e = BBCases.size(); i != e; ++i)
1143 if (PTIHandled.count(BBCases[i].Value)) {
1144 // If this is one we are capable of getting...
1145 if (PredHasWeights || SuccHasWeights)
1146 Weights.push_back(WeightsForHandled[BBCases[i].Value]);
1147 PredCases.push_back(BBCases[i]);
1148 NewSuccessors.push_back(BBCases[i].Dest);
1149 PTIHandled.erase(
1150 BBCases[i].Value); // This constant is taken care of
1151 }
1152
1153 // If there are any constants vectored to BB that TI doesn't handle,
1154 // they must go to the default destination of TI.
1155 for (ConstantInt *I : PTIHandled) {
1156 if (PredHasWeights || SuccHasWeights)
1157 Weights.push_back(WeightsForHandled[I]);
1158 PredCases.push_back(ValueEqualityComparisonCase(I, BBDefault));
1159 NewSuccessors.push_back(BBDefault);
1160 }
1161 }
1162
1163 // Okay, at this point, we know which new successor Pred will get. Make
1164 // sure we update the number of entries in the PHI nodes for these
1165 // successors.
1166 for (BasicBlock *NewSuccessor : NewSuccessors)
1167 AddPredecessorToBlock(NewSuccessor, Pred, BB);
1168
1169 Builder.SetInsertPoint(PTI);
1170 // Convert pointer to int before we switch.
1171 if (CV->getType()->isPointerTy()) {
1172 CV = Builder.CreatePtrToInt(CV, DL.getIntPtrType(CV->getType()),
1173 "magicptr");
1174 }
1175
1176 // Now that the successors are updated, create the new Switch instruction.
1177 SwitchInst *NewSI =
1178 Builder.CreateSwitch(CV, PredDefault, PredCases.size());
1179 NewSI->setDebugLoc(PTI->getDebugLoc());
1180 for (ValueEqualityComparisonCase &V : PredCases)
1181 NewSI->addCase(V.Value, V.Dest);
1182
1183 if (PredHasWeights || SuccHasWeights) {
1184 // Halve the weights if any of them cannot fit in an uint32_t
1185 FitWeights(Weights);
1186
1187 SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
1188
1189 setBranchWeights(NewSI, MDWeights);
1190 }
1191
1192 EraseTerminatorAndDCECond(PTI);
1193
1194 // Okay, last check. If BB is still a successor of PSI, then we must
1195 // have an infinite loop case. If so, add an infinitely looping block
1196 // to handle the case to preserve the behavior of the code.
1197 BasicBlock *InfLoopBlock = nullptr;
1198 for (unsigned i = 0, e = NewSI->getNumSuccessors(); i != e; ++i)
1199 if (NewSI->getSuccessor(i) == BB) {
1200 if (!InfLoopBlock) {
1201 // Insert it at the end of the function, because it's either code,
1202 // or it won't matter if it's hot. :)
1203 InfLoopBlock = BasicBlock::Create(BB->getContext(), "infloop",
1204 BB->getParent());
1205 BranchInst::Create(InfLoopBlock, InfLoopBlock);
1206 }
1207 NewSI->setSuccessor(i, InfLoopBlock);
1208 }
1209
1210 Changed = true;
1211 }
1212 }
1213 return Changed;
1214 }
1215
1216 // If we would need to insert a select that uses the value of this invoke
1217 // (comments in HoistThenElseCodeToIf explain why we would need to do this), we
1218 // 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)1219 static bool isSafeToHoistInvoke(BasicBlock *BB1, BasicBlock *BB2,
1220 Instruction *I1, Instruction *I2) {
1221 for (BasicBlock *Succ : successors(BB1)) {
1222 for (const PHINode &PN : Succ->phis()) {
1223 Value *BB1V = PN.getIncomingValueForBlock(BB1);
1224 Value *BB2V = PN.getIncomingValueForBlock(BB2);
1225 if (BB1V != BB2V && (BB1V == I1 || BB2V == I2)) {
1226 return false;
1227 }
1228 }
1229 }
1230 return true;
1231 }
1232
1233 static bool passingValueIsAlwaysUndefined(Value *V, Instruction *I);
1234
1235 /// Given a conditional branch that goes to BB1 and BB2, hoist any common code
1236 /// in the two blocks up into the branch block. The caller of this function
1237 /// guarantees that BI's block dominates BB1 and BB2.
HoistThenElseCodeToIf(BranchInst * BI,const TargetTransformInfo & TTI)1238 static bool HoistThenElseCodeToIf(BranchInst *BI,
1239 const TargetTransformInfo &TTI) {
1240 // This does very trivial matching, with limited scanning, to find identical
1241 // instructions in the two blocks. In particular, we don't want to get into
1242 // O(M*N) situations here where M and N are the sizes of BB1 and BB2. As
1243 // such, we currently just scan for obviously identical instructions in an
1244 // identical order.
1245 BasicBlock *BB1 = BI->getSuccessor(0); // The true destination.
1246 BasicBlock *BB2 = BI->getSuccessor(1); // The false destination
1247
1248 BasicBlock::iterator BB1_Itr = BB1->begin();
1249 BasicBlock::iterator BB2_Itr = BB2->begin();
1250
1251 Instruction *I1 = &*BB1_Itr++, *I2 = &*BB2_Itr++;
1252 // Skip debug info if it is not identical.
1253 DbgInfoIntrinsic *DBI1 = dyn_cast<DbgInfoIntrinsic>(I1);
1254 DbgInfoIntrinsic *DBI2 = dyn_cast<DbgInfoIntrinsic>(I2);
1255 if (!DBI1 || !DBI2 || !DBI1->isIdenticalToWhenDefined(DBI2)) {
1256 while (isa<DbgInfoIntrinsic>(I1))
1257 I1 = &*BB1_Itr++;
1258 while (isa<DbgInfoIntrinsic>(I2))
1259 I2 = &*BB2_Itr++;
1260 }
1261 // FIXME: Can we define a safety predicate for CallBr?
1262 if (isa<PHINode>(I1) || !I1->isIdenticalToWhenDefined(I2) ||
1263 (isa<InvokeInst>(I1) && !isSafeToHoistInvoke(BB1, BB2, I1, I2)) ||
1264 isa<CallBrInst>(I1))
1265 return false;
1266
1267 BasicBlock *BIParent = BI->getParent();
1268
1269 bool Changed = false;
1270 do {
1271 // If we are hoisting the terminator instruction, don't move one (making a
1272 // broken BB), instead clone it, and remove BI.
1273 if (I1->isTerminator())
1274 goto HoistTerminator;
1275
1276 // If we're going to hoist a call, make sure that the two instructions we're
1277 // commoning/hoisting are both marked with musttail, or neither of them is
1278 // marked as such. Otherwise, we might end up in a situation where we hoist
1279 // from a block where the terminator is a `ret` to a block where the terminator
1280 // is a `br`, and `musttail` calls expect to be followed by a return.
1281 auto *C1 = dyn_cast<CallInst>(I1);
1282 auto *C2 = dyn_cast<CallInst>(I2);
1283 if (C1 && C2)
1284 if (C1->isMustTailCall() != C2->isMustTailCall())
1285 return Changed;
1286
1287 if (!TTI.isProfitableToHoist(I1) || !TTI.isProfitableToHoist(I2))
1288 return Changed;
1289
1290 if (isa<DbgInfoIntrinsic>(I1) || isa<DbgInfoIntrinsic>(I2)) {
1291 assert (isa<DbgInfoIntrinsic>(I1) && isa<DbgInfoIntrinsic>(I2));
1292 // The debug location is an integral part of a debug info intrinsic
1293 // and can't be separated from it or replaced. Instead of attempting
1294 // to merge locations, simply hoist both copies of the intrinsic.
1295 BIParent->getInstList().splice(BI->getIterator(),
1296 BB1->getInstList(), I1);
1297 BIParent->getInstList().splice(BI->getIterator(),
1298 BB2->getInstList(), I2);
1299 Changed = true;
1300 } else {
1301 // For a normal instruction, we just move one to right before the branch,
1302 // then replace all uses of the other with the first. Finally, we remove
1303 // the now redundant second instruction.
1304 BIParent->getInstList().splice(BI->getIterator(),
1305 BB1->getInstList(), I1);
1306 if (!I2->use_empty())
1307 I2->replaceAllUsesWith(I1);
1308 I1->andIRFlags(I2);
1309 unsigned KnownIDs[] = {LLVMContext::MD_tbaa,
1310 LLVMContext::MD_range,
1311 LLVMContext::MD_fpmath,
1312 LLVMContext::MD_invariant_load,
1313 LLVMContext::MD_nonnull,
1314 LLVMContext::MD_invariant_group,
1315 LLVMContext::MD_align,
1316 LLVMContext::MD_dereferenceable,
1317 LLVMContext::MD_dereferenceable_or_null,
1318 LLVMContext::MD_mem_parallel_loop_access,
1319 LLVMContext::MD_access_group,
1320 LLVMContext::MD_preserve_access_index};
1321 combineMetadata(I1, I2, KnownIDs, true);
1322
1323 // I1 and I2 are being combined into a single instruction. Its debug
1324 // location is the merged locations of the original instructions.
1325 I1->applyMergedLocation(I1->getDebugLoc(), I2->getDebugLoc());
1326
1327 I2->eraseFromParent();
1328 Changed = true;
1329 }
1330
1331 I1 = &*BB1_Itr++;
1332 I2 = &*BB2_Itr++;
1333 // Skip debug info if it is not identical.
1334 DbgInfoIntrinsic *DBI1 = dyn_cast<DbgInfoIntrinsic>(I1);
1335 DbgInfoIntrinsic *DBI2 = dyn_cast<DbgInfoIntrinsic>(I2);
1336 if (!DBI1 || !DBI2 || !DBI1->isIdenticalToWhenDefined(DBI2)) {
1337 while (isa<DbgInfoIntrinsic>(I1))
1338 I1 = &*BB1_Itr++;
1339 while (isa<DbgInfoIntrinsic>(I2))
1340 I2 = &*BB2_Itr++;
1341 }
1342 } while (I1->isIdenticalToWhenDefined(I2));
1343
1344 return true;
1345
1346 HoistTerminator:
1347 // It may not be possible to hoist an invoke.
1348 // FIXME: Can we define a safety predicate for CallBr?
1349 if (isa<InvokeInst>(I1) && !isSafeToHoistInvoke(BB1, BB2, I1, I2))
1350 return Changed;
1351
1352 // TODO: callbr hoisting currently disabled pending further study.
1353 if (isa<CallBrInst>(I1))
1354 return Changed;
1355
1356 for (BasicBlock *Succ : successors(BB1)) {
1357 for (PHINode &PN : Succ->phis()) {
1358 Value *BB1V = PN.getIncomingValueForBlock(BB1);
1359 Value *BB2V = PN.getIncomingValueForBlock(BB2);
1360 if (BB1V == BB2V)
1361 continue;
1362
1363 // Check for passingValueIsAlwaysUndefined here because we would rather
1364 // eliminate undefined control flow then converting it to a select.
1365 if (passingValueIsAlwaysUndefined(BB1V, &PN) ||
1366 passingValueIsAlwaysUndefined(BB2V, &PN))
1367 return Changed;
1368
1369 if (isa<ConstantExpr>(BB1V) && !isSafeToSpeculativelyExecute(BB1V))
1370 return Changed;
1371 if (isa<ConstantExpr>(BB2V) && !isSafeToSpeculativelyExecute(BB2V))
1372 return Changed;
1373 }
1374 }
1375
1376 // Okay, it is safe to hoist the terminator.
1377 Instruction *NT = I1->clone();
1378 BIParent->getInstList().insert(BI->getIterator(), NT);
1379 if (!NT->getType()->isVoidTy()) {
1380 I1->replaceAllUsesWith(NT);
1381 I2->replaceAllUsesWith(NT);
1382 NT->takeName(I1);
1383 }
1384
1385 // Ensure terminator gets a debug location, even an unknown one, in case
1386 // it involves inlinable calls.
1387 NT->applyMergedLocation(I1->getDebugLoc(), I2->getDebugLoc());
1388
1389 // PHIs created below will adopt NT's merged DebugLoc.
1390 IRBuilder<NoFolder> Builder(NT);
1391
1392 // Hoisting one of the terminators from our successor is a great thing.
1393 // Unfortunately, the successors of the if/else blocks may have PHI nodes in
1394 // them. If they do, all PHI entries for BB1/BB2 must agree for all PHI
1395 // nodes, so we insert select instruction to compute the final result.
1396 std::map<std::pair<Value *, Value *>, SelectInst *> InsertedSelects;
1397 for (BasicBlock *Succ : successors(BB1)) {
1398 for (PHINode &PN : Succ->phis()) {
1399 Value *BB1V = PN.getIncomingValueForBlock(BB1);
1400 Value *BB2V = PN.getIncomingValueForBlock(BB2);
1401 if (BB1V == BB2V)
1402 continue;
1403
1404 // These values do not agree. Insert a select instruction before NT
1405 // that determines the right value.
1406 SelectInst *&SI = InsertedSelects[std::make_pair(BB1V, BB2V)];
1407 if (!SI) {
1408 // Propagate fast-math-flags from phi node to its replacement select.
1409 IRBuilder<>::FastMathFlagGuard FMFGuard(Builder);
1410 if (isa<FPMathOperator>(PN))
1411 Builder.setFastMathFlags(PN.getFastMathFlags());
1412
1413 SI = cast<SelectInst>(
1414 Builder.CreateSelect(BI->getCondition(), BB1V, BB2V,
1415 BB1V->getName() + "." + BB2V->getName(), BI));
1416 }
1417
1418 // Make the PHI node use the select for all incoming values for BB1/BB2
1419 for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i)
1420 if (PN.getIncomingBlock(i) == BB1 || PN.getIncomingBlock(i) == BB2)
1421 PN.setIncomingValue(i, SI);
1422 }
1423 }
1424
1425 // Update any PHI nodes in our new successors.
1426 for (BasicBlock *Succ : successors(BB1))
1427 AddPredecessorToBlock(Succ, BIParent, BB1);
1428
1429 EraseTerminatorAndDCECond(BI);
1430 return true;
1431 }
1432
1433 // Check lifetime markers.
isLifeTimeMarker(const Instruction * I)1434 static bool isLifeTimeMarker(const Instruction *I) {
1435 if (auto II = dyn_cast<IntrinsicInst>(I)) {
1436 switch (II->getIntrinsicID()) {
1437 default:
1438 break;
1439 case Intrinsic::lifetime_start:
1440 case Intrinsic::lifetime_end:
1441 return true;
1442 }
1443 }
1444 return false;
1445 }
1446
1447 // All instructions in Insts belong to different blocks that all unconditionally
1448 // branch to a common successor. Analyze each instruction and return true if it
1449 // would be possible to sink them into their successor, creating one common
1450 // instruction instead. For every value that would be required to be provided by
1451 // PHI node (because an operand varies in each input block), add to PHIOperands.
canSinkInstructions(ArrayRef<Instruction * > Insts,DenseMap<Instruction *,SmallVector<Value *,4>> & PHIOperands)1452 static bool canSinkInstructions(
1453 ArrayRef<Instruction *> Insts,
1454 DenseMap<Instruction *, SmallVector<Value *, 4>> &PHIOperands) {
1455 // Prune out obviously bad instructions to move. Any non-store instruction
1456 // must have exactly one use, and we check later that use is by a single,
1457 // common PHI instruction in the successor.
1458 for (auto *I : Insts) {
1459 // These instructions may change or break semantics if moved.
1460 if (isa<PHINode>(I) || I->isEHPad() || isa<AllocaInst>(I) ||
1461 I->getType()->isTokenTy())
1462 return false;
1463
1464 // Conservatively return false if I is an inline-asm instruction. Sinking
1465 // and merging inline-asm instructions can potentially create arguments
1466 // that cannot satisfy the inline-asm constraints.
1467 if (const auto *C = dyn_cast<CallBase>(I))
1468 if (C->isInlineAsm())
1469 return false;
1470
1471 // Everything must have only one use too, apart from stores which
1472 // have no uses.
1473 if (!isa<StoreInst>(I) && !I->hasOneUse())
1474 return false;
1475 }
1476
1477 const Instruction *I0 = Insts.front();
1478 for (auto *I : Insts)
1479 if (!I->isSameOperationAs(I0))
1480 return false;
1481
1482 // All instructions in Insts are known to be the same opcode. If they aren't
1483 // stores, check the only user of each is a PHI or in the same block as the
1484 // instruction, because if a user is in the same block as an instruction
1485 // we're contemplating sinking, it must already be determined to be sinkable.
1486 if (!isa<StoreInst>(I0)) {
1487 auto *PNUse = dyn_cast<PHINode>(*I0->user_begin());
1488 auto *Succ = I0->getParent()->getTerminator()->getSuccessor(0);
1489 if (!all_of(Insts, [&PNUse,&Succ](const Instruction *I) -> bool {
1490 auto *U = cast<Instruction>(*I->user_begin());
1491 return (PNUse &&
1492 PNUse->getParent() == Succ &&
1493 PNUse->getIncomingValueForBlock(I->getParent()) == I) ||
1494 U->getParent() == I->getParent();
1495 }))
1496 return false;
1497 }
1498
1499 // Because SROA can't handle speculating stores of selects, try not to sink
1500 // loads, stores or lifetime markers of allocas when we'd have to create a
1501 // PHI for the address operand. Also, because it is likely that loads or
1502 // stores of allocas will disappear when Mem2Reg/SROA is run, don't sink
1503 // them.
1504 // This can cause code churn which can have unintended consequences down
1505 // the line - see https://llvm.org/bugs/show_bug.cgi?id=30244.
1506 // FIXME: This is a workaround for a deficiency in SROA - see
1507 // https://llvm.org/bugs/show_bug.cgi?id=30188
1508 if (isa<StoreInst>(I0) && any_of(Insts, [](const Instruction *I) {
1509 return isa<AllocaInst>(I->getOperand(1)->stripPointerCasts());
1510 }))
1511 return false;
1512 if (isa<LoadInst>(I0) && any_of(Insts, [](const Instruction *I) {
1513 return isa<AllocaInst>(I->getOperand(0)->stripPointerCasts());
1514 }))
1515 return false;
1516 if (isLifeTimeMarker(I0) && any_of(Insts, [](const Instruction *I) {
1517 return isa<AllocaInst>(I->getOperand(1)->stripPointerCasts());
1518 }))
1519 return false;
1520
1521 for (unsigned OI = 0, OE = I0->getNumOperands(); OI != OE; ++OI) {
1522 if (I0->getOperand(OI)->getType()->isTokenTy())
1523 // Don't touch any operand of token type.
1524 return false;
1525
1526 auto SameAsI0 = [&I0, OI](const Instruction *I) {
1527 assert(I->getNumOperands() == I0->getNumOperands());
1528 return I->getOperand(OI) == I0->getOperand(OI);
1529 };
1530 if (!all_of(Insts, SameAsI0)) {
1531 if (!canReplaceOperandWithVariable(I0, OI))
1532 // We can't create a PHI from this GEP.
1533 return false;
1534 // Don't create indirect calls! The called value is the final operand.
1535 if (isa<CallBase>(I0) && OI == OE - 1) {
1536 // FIXME: if the call was *already* indirect, we should do this.
1537 return false;
1538 }
1539 for (auto *I : Insts)
1540 PHIOperands[I].push_back(I->getOperand(OI));
1541 }
1542 }
1543 return true;
1544 }
1545
1546 // Assuming canSinkLastInstruction(Blocks) has returned true, sink the last
1547 // instruction of every block in Blocks to their common successor, commoning
1548 // into one instruction.
sinkLastInstruction(ArrayRef<BasicBlock * > Blocks)1549 static bool sinkLastInstruction(ArrayRef<BasicBlock*> Blocks) {
1550 auto *BBEnd = Blocks[0]->getTerminator()->getSuccessor(0);
1551
1552 // canSinkLastInstruction returning true guarantees that every block has at
1553 // least one non-terminator instruction.
1554 SmallVector<Instruction*,4> Insts;
1555 for (auto *BB : Blocks) {
1556 Instruction *I = BB->getTerminator();
1557 do {
1558 I = I->getPrevNode();
1559 } while (isa<DbgInfoIntrinsic>(I) && I != &BB->front());
1560 if (!isa<DbgInfoIntrinsic>(I))
1561 Insts.push_back(I);
1562 }
1563
1564 // The only checking we need to do now is that all users of all instructions
1565 // are the same PHI node. canSinkLastInstruction should have checked this but
1566 // it is slightly over-aggressive - it gets confused by commutative instructions
1567 // so double-check it here.
1568 Instruction *I0 = Insts.front();
1569 if (!isa<StoreInst>(I0)) {
1570 auto *PNUse = dyn_cast<PHINode>(*I0->user_begin());
1571 if (!all_of(Insts, [&PNUse](const Instruction *I) -> bool {
1572 auto *U = cast<Instruction>(*I->user_begin());
1573 return U == PNUse;
1574 }))
1575 return false;
1576 }
1577
1578 // We don't need to do any more checking here; canSinkLastInstruction should
1579 // have done it all for us.
1580 SmallVector<Value*, 4> NewOperands;
1581 for (unsigned O = 0, E = I0->getNumOperands(); O != E; ++O) {
1582 // This check is different to that in canSinkLastInstruction. There, we
1583 // cared about the global view once simplifycfg (and instcombine) have
1584 // completed - it takes into account PHIs that become trivially
1585 // simplifiable. However here we need a more local view; if an operand
1586 // differs we create a PHI and rely on instcombine to clean up the very
1587 // small mess we may make.
1588 bool NeedPHI = any_of(Insts, [&I0, O](const Instruction *I) {
1589 return I->getOperand(O) != I0->getOperand(O);
1590 });
1591 if (!NeedPHI) {
1592 NewOperands.push_back(I0->getOperand(O));
1593 continue;
1594 }
1595
1596 // Create a new PHI in the successor block and populate it.
1597 auto *Op = I0->getOperand(O);
1598 assert(!Op->getType()->isTokenTy() && "Can't PHI tokens!");
1599 auto *PN = PHINode::Create(Op->getType(), Insts.size(),
1600 Op->getName() + ".sink", &BBEnd->front());
1601 for (auto *I : Insts)
1602 PN->addIncoming(I->getOperand(O), I->getParent());
1603 NewOperands.push_back(PN);
1604 }
1605
1606 // Arbitrarily use I0 as the new "common" instruction; remap its operands
1607 // and move it to the start of the successor block.
1608 for (unsigned O = 0, E = I0->getNumOperands(); O != E; ++O)
1609 I0->getOperandUse(O).set(NewOperands[O]);
1610 I0->moveBefore(&*BBEnd->getFirstInsertionPt());
1611
1612 // Update metadata and IR flags, and merge debug locations.
1613 for (auto *I : Insts)
1614 if (I != I0) {
1615 // The debug location for the "common" instruction is the merged locations
1616 // of all the commoned instructions. We start with the original location
1617 // of the "common" instruction and iteratively merge each location in the
1618 // loop below.
1619 // This is an N-way merge, which will be inefficient if I0 is a CallInst.
1620 // However, as N-way merge for CallInst is rare, so we use simplified API
1621 // instead of using complex API for N-way merge.
1622 I0->applyMergedLocation(I0->getDebugLoc(), I->getDebugLoc());
1623 combineMetadataForCSE(I0, I, true);
1624 I0->andIRFlags(I);
1625 }
1626
1627 if (!isa<StoreInst>(I0)) {
1628 // canSinkLastInstruction checked that all instructions were used by
1629 // one and only one PHI node. Find that now, RAUW it to our common
1630 // instruction and nuke it.
1631 assert(I0->hasOneUse());
1632 auto *PN = cast<PHINode>(*I0->user_begin());
1633 PN->replaceAllUsesWith(I0);
1634 PN->eraseFromParent();
1635 }
1636
1637 // Finally nuke all instructions apart from the common instruction.
1638 for (auto *I : Insts)
1639 if (I != I0)
1640 I->eraseFromParent();
1641
1642 return true;
1643 }
1644
1645 namespace {
1646
1647 // LockstepReverseIterator - Iterates through instructions
1648 // in a set of blocks in reverse order from the first non-terminator.
1649 // For example (assume all blocks have size n):
1650 // LockstepReverseIterator I([B1, B2, B3]);
1651 // *I-- = [B1[n], B2[n], B3[n]];
1652 // *I-- = [B1[n-1], B2[n-1], B3[n-1]];
1653 // *I-- = [B1[n-2], B2[n-2], B3[n-2]];
1654 // ...
1655 class LockstepReverseIterator {
1656 ArrayRef<BasicBlock*> Blocks;
1657 SmallVector<Instruction*,4> Insts;
1658 bool Fail;
1659
1660 public:
LockstepReverseIterator(ArrayRef<BasicBlock * > Blocks)1661 LockstepReverseIterator(ArrayRef<BasicBlock*> Blocks) : Blocks(Blocks) {
1662 reset();
1663 }
1664
reset()1665 void reset() {
1666 Fail = false;
1667 Insts.clear();
1668 for (auto *BB : Blocks) {
1669 Instruction *Inst = BB->getTerminator();
1670 for (Inst = Inst->getPrevNode(); Inst && isa<DbgInfoIntrinsic>(Inst);)
1671 Inst = Inst->getPrevNode();
1672 if (!Inst) {
1673 // Block wasn't big enough.
1674 Fail = true;
1675 return;
1676 }
1677 Insts.push_back(Inst);
1678 }
1679 }
1680
isValid() const1681 bool isValid() const {
1682 return !Fail;
1683 }
1684
operator --()1685 void operator--() {
1686 if (Fail)
1687 return;
1688 for (auto *&Inst : Insts) {
1689 for (Inst = Inst->getPrevNode(); Inst && isa<DbgInfoIntrinsic>(Inst);)
1690 Inst = Inst->getPrevNode();
1691 // Already at beginning of block.
1692 if (!Inst) {
1693 Fail = true;
1694 return;
1695 }
1696 }
1697 }
1698
operator *() const1699 ArrayRef<Instruction*> operator * () const {
1700 return Insts;
1701 }
1702 };
1703
1704 } // end anonymous namespace
1705
1706 /// Check whether BB's predecessors end with unconditional branches. If it is
1707 /// true, sink any common code from the predecessors to BB.
1708 /// We also allow one predecessor to end with conditional branch (but no more
1709 /// than one).
SinkCommonCodeFromPredecessors(BasicBlock * BB)1710 static bool SinkCommonCodeFromPredecessors(BasicBlock *BB) {
1711 // We support two situations:
1712 // (1) all incoming arcs are unconditional
1713 // (2) one incoming arc is conditional
1714 //
1715 // (2) is very common in switch defaults and
1716 // else-if patterns;
1717 //
1718 // if (a) f(1);
1719 // else if (b) f(2);
1720 //
1721 // produces:
1722 //
1723 // [if]
1724 // / \
1725 // [f(1)] [if]
1726 // | | \
1727 // | | |
1728 // | [f(2)]|
1729 // \ | /
1730 // [ end ]
1731 //
1732 // [end] has two unconditional predecessor arcs and one conditional. The
1733 // conditional refers to the implicit empty 'else' arc. This conditional
1734 // arc can also be caused by an empty default block in a switch.
1735 //
1736 // In this case, we attempt to sink code from all *unconditional* arcs.
1737 // If we can sink instructions from these arcs (determined during the scan
1738 // phase below) we insert a common successor for all unconditional arcs and
1739 // connect that to [end], to enable sinking:
1740 //
1741 // [if]
1742 // / \
1743 // [x(1)] [if]
1744 // | | \
1745 // | | \
1746 // | [x(2)] |
1747 // \ / |
1748 // [sink.split] |
1749 // \ /
1750 // [ end ]
1751 //
1752 SmallVector<BasicBlock*,4> UnconditionalPreds;
1753 Instruction *Cond = nullptr;
1754 for (auto *B : predecessors(BB)) {
1755 auto *T = B->getTerminator();
1756 if (isa<BranchInst>(T) && cast<BranchInst>(T)->isUnconditional())
1757 UnconditionalPreds.push_back(B);
1758 else if ((isa<BranchInst>(T) || isa<SwitchInst>(T)) && !Cond)
1759 Cond = T;
1760 else
1761 return false;
1762 }
1763 if (UnconditionalPreds.size() < 2)
1764 return false;
1765
1766 bool Changed = false;
1767 // We take a two-step approach to tail sinking. First we scan from the end of
1768 // each block upwards in lockstep. If the n'th instruction from the end of each
1769 // block can be sunk, those instructions are added to ValuesToSink and we
1770 // carry on. If we can sink an instruction but need to PHI-merge some operands
1771 // (because they're not identical in each instruction) we add these to
1772 // PHIOperands.
1773 unsigned ScanIdx = 0;
1774 SmallPtrSet<Value*,4> InstructionsToSink;
1775 DenseMap<Instruction*, SmallVector<Value*,4>> PHIOperands;
1776 LockstepReverseIterator LRI(UnconditionalPreds);
1777 while (LRI.isValid() &&
1778 canSinkInstructions(*LRI, PHIOperands)) {
1779 LLVM_DEBUG(dbgs() << "SINK: instruction can be sunk: " << *(*LRI)[0]
1780 << "\n");
1781 InstructionsToSink.insert((*LRI).begin(), (*LRI).end());
1782 ++ScanIdx;
1783 --LRI;
1784 }
1785
1786 auto ProfitableToSinkInstruction = [&](LockstepReverseIterator &LRI) {
1787 unsigned NumPHIdValues = 0;
1788 for (auto *I : *LRI)
1789 for (auto *V : PHIOperands[I])
1790 if (InstructionsToSink.count(V) == 0)
1791 ++NumPHIdValues;
1792 LLVM_DEBUG(dbgs() << "SINK: #phid values: " << NumPHIdValues << "\n");
1793 unsigned NumPHIInsts = NumPHIdValues / UnconditionalPreds.size();
1794 if ((NumPHIdValues % UnconditionalPreds.size()) != 0)
1795 NumPHIInsts++;
1796
1797 return NumPHIInsts <= 1;
1798 };
1799
1800 if (ScanIdx > 0 && Cond) {
1801 // Check if we would actually sink anything first! This mutates the CFG and
1802 // adds an extra block. The goal in doing this is to allow instructions that
1803 // couldn't be sunk before to be sunk - obviously, speculatable instructions
1804 // (such as trunc, add) can be sunk and predicated already. So we check that
1805 // we're going to sink at least one non-speculatable instruction.
1806 LRI.reset();
1807 unsigned Idx = 0;
1808 bool Profitable = false;
1809 while (ProfitableToSinkInstruction(LRI) && Idx < ScanIdx) {
1810 if (!isSafeToSpeculativelyExecute((*LRI)[0])) {
1811 Profitable = true;
1812 break;
1813 }
1814 --LRI;
1815 ++Idx;
1816 }
1817 if (!Profitable)
1818 return false;
1819
1820 LLVM_DEBUG(dbgs() << "SINK: Splitting edge\n");
1821 // We have a conditional edge and we're going to sink some instructions.
1822 // Insert a new block postdominating all blocks we're going to sink from.
1823 if (!SplitBlockPredecessors(BB, UnconditionalPreds, ".sink.split"))
1824 // Edges couldn't be split.
1825 return false;
1826 Changed = true;
1827 }
1828
1829 // Now that we've analyzed all potential sinking candidates, perform the
1830 // actual sink. We iteratively sink the last non-terminator of the source
1831 // blocks into their common successor unless doing so would require too
1832 // many PHI instructions to be generated (currently only one PHI is allowed
1833 // per sunk instruction).
1834 //
1835 // We can use InstructionsToSink to discount values needing PHI-merging that will
1836 // actually be sunk in a later iteration. This allows us to be more
1837 // aggressive in what we sink. This does allow a false positive where we
1838 // sink presuming a later value will also be sunk, but stop half way through
1839 // and never actually sink it which means we produce more PHIs than intended.
1840 // This is unlikely in practice though.
1841 for (unsigned SinkIdx = 0; SinkIdx != ScanIdx; ++SinkIdx) {
1842 LLVM_DEBUG(dbgs() << "SINK: Sink: "
1843 << *UnconditionalPreds[0]->getTerminator()->getPrevNode()
1844 << "\n");
1845
1846 // Because we've sunk every instruction in turn, the current instruction to
1847 // sink is always at index 0.
1848 LRI.reset();
1849 if (!ProfitableToSinkInstruction(LRI)) {
1850 // Too many PHIs would be created.
1851 LLVM_DEBUG(
1852 dbgs() << "SINK: stopping here, too many PHIs would be created!\n");
1853 break;
1854 }
1855
1856 if (!sinkLastInstruction(UnconditionalPreds))
1857 return Changed;
1858 NumSinkCommons++;
1859 Changed = true;
1860 }
1861 return Changed;
1862 }
1863
1864 /// Determine if we can hoist sink a sole store instruction out of a
1865 /// conditional block.
1866 ///
1867 /// We are looking for code like the following:
1868 /// BrBB:
1869 /// store i32 %add, i32* %arrayidx2
1870 /// ... // No other stores or function calls (we could be calling a memory
1871 /// ... // function).
1872 /// %cmp = icmp ult %x, %y
1873 /// br i1 %cmp, label %EndBB, label %ThenBB
1874 /// ThenBB:
1875 /// store i32 %add5, i32* %arrayidx2
1876 /// br label EndBB
1877 /// EndBB:
1878 /// ...
1879 /// We are going to transform this into:
1880 /// BrBB:
1881 /// store i32 %add, i32* %arrayidx2
1882 /// ... //
1883 /// %cmp = icmp ult %x, %y
1884 /// %add.add5 = select i1 %cmp, i32 %add, %add5
1885 /// store i32 %add.add5, i32* %arrayidx2
1886 /// ...
1887 ///
1888 /// \return The pointer to the value of the previous store if the store can be
1889 /// hoisted into the predecessor block. 0 otherwise.
isSafeToSpeculateStore(Instruction * I,BasicBlock * BrBB,BasicBlock * StoreBB,BasicBlock * EndBB)1890 static Value *isSafeToSpeculateStore(Instruction *I, BasicBlock *BrBB,
1891 BasicBlock *StoreBB, BasicBlock *EndBB) {
1892 StoreInst *StoreToHoist = dyn_cast<StoreInst>(I);
1893 if (!StoreToHoist)
1894 return nullptr;
1895
1896 // Volatile or atomic.
1897 if (!StoreToHoist->isSimple())
1898 return nullptr;
1899
1900 Value *StorePtr = StoreToHoist->getPointerOperand();
1901
1902 // Look for a store to the same pointer in BrBB.
1903 unsigned MaxNumInstToLookAt = 9;
1904 for (Instruction &CurI : reverse(BrBB->instructionsWithoutDebug())) {
1905 if (!MaxNumInstToLookAt)
1906 break;
1907 --MaxNumInstToLookAt;
1908
1909 // Could be calling an instruction that affects memory like free().
1910 if (CurI.mayHaveSideEffects() && !isa<StoreInst>(CurI))
1911 return nullptr;
1912
1913 if (auto *SI = dyn_cast<StoreInst>(&CurI)) {
1914 // Found the previous store make sure it stores to the same location.
1915 if (SI->getPointerOperand() == StorePtr)
1916 // Found the previous store, return its value operand.
1917 return SI->getValueOperand();
1918 return nullptr; // Unknown store.
1919 }
1920 }
1921
1922 return nullptr;
1923 }
1924
1925 /// Speculate a conditional basic block flattening the CFG.
1926 ///
1927 /// Note that this is a very risky transform currently. Speculating
1928 /// instructions like this is most often not desirable. Instead, there is an MI
1929 /// pass which can do it with full awareness of the resource constraints.
1930 /// However, some cases are "obvious" and we should do directly. An example of
1931 /// this is speculating a single, reasonably cheap instruction.
1932 ///
1933 /// There is only one distinct advantage to flattening the CFG at the IR level:
1934 /// it makes very common but simplistic optimizations such as are common in
1935 /// instcombine and the DAG combiner more powerful by removing CFG edges and
1936 /// modeling their effects with easier to reason about SSA value graphs.
1937 ///
1938 ///
1939 /// An illustration of this transform is turning this IR:
1940 /// \code
1941 /// BB:
1942 /// %cmp = icmp ult %x, %y
1943 /// br i1 %cmp, label %EndBB, label %ThenBB
1944 /// ThenBB:
1945 /// %sub = sub %x, %y
1946 /// br label BB2
1947 /// EndBB:
1948 /// %phi = phi [ %sub, %ThenBB ], [ 0, %EndBB ]
1949 /// ...
1950 /// \endcode
1951 ///
1952 /// Into this IR:
1953 /// \code
1954 /// BB:
1955 /// %cmp = icmp ult %x, %y
1956 /// %sub = sub %x, %y
1957 /// %cond = select i1 %cmp, 0, %sub
1958 /// ...
1959 /// \endcode
1960 ///
1961 /// \returns true if the conditional block is removed.
SpeculativelyExecuteBB(BranchInst * BI,BasicBlock * ThenBB,const TargetTransformInfo & TTI)1962 static bool SpeculativelyExecuteBB(BranchInst *BI, BasicBlock *ThenBB,
1963 const TargetTransformInfo &TTI) {
1964 // Be conservative for now. FP select instruction can often be expensive.
1965 Value *BrCond = BI->getCondition();
1966 if (isa<FCmpInst>(BrCond))
1967 return false;
1968
1969 BasicBlock *BB = BI->getParent();
1970 BasicBlock *EndBB = ThenBB->getTerminator()->getSuccessor(0);
1971
1972 // If ThenBB is actually on the false edge of the conditional branch, remember
1973 // to swap the select operands later.
1974 bool Invert = false;
1975 if (ThenBB != BI->getSuccessor(0)) {
1976 assert(ThenBB == BI->getSuccessor(1) && "No edge from 'if' block?");
1977 Invert = true;
1978 }
1979 assert(EndBB == BI->getSuccessor(!Invert) && "No edge from to end block");
1980
1981 // Keep a count of how many times instructions are used within ThenBB when
1982 // they are candidates for sinking into ThenBB. Specifically:
1983 // - They are defined in BB, and
1984 // - They have no side effects, and
1985 // - All of their uses are in ThenBB.
1986 SmallDenseMap<Instruction *, unsigned, 4> SinkCandidateUseCounts;
1987
1988 SmallVector<Instruction *, 4> SpeculatedDbgIntrinsics;
1989
1990 unsigned SpeculatedInstructions = 0;
1991 Value *SpeculatedStoreValue = nullptr;
1992 StoreInst *SpeculatedStore = nullptr;
1993 for (BasicBlock::iterator BBI = ThenBB->begin(),
1994 BBE = std::prev(ThenBB->end());
1995 BBI != BBE; ++BBI) {
1996 Instruction *I = &*BBI;
1997 // Skip debug info.
1998 if (isa<DbgInfoIntrinsic>(I)) {
1999 SpeculatedDbgIntrinsics.push_back(I);
2000 continue;
2001 }
2002
2003 // Only speculatively execute a single instruction (not counting the
2004 // terminator) for now.
2005 ++SpeculatedInstructions;
2006 if (SpeculatedInstructions > 1)
2007 return false;
2008
2009 // Don't hoist the instruction if it's unsafe or expensive.
2010 if (!isSafeToSpeculativelyExecute(I) &&
2011 !(HoistCondStores && (SpeculatedStoreValue = isSafeToSpeculateStore(
2012 I, BB, ThenBB, EndBB))))
2013 return false;
2014 if (!SpeculatedStoreValue &&
2015 ComputeSpeculationCost(I, TTI) >
2016 PHINodeFoldingThreshold * TargetTransformInfo::TCC_Basic)
2017 return false;
2018
2019 // Store the store speculation candidate.
2020 if (SpeculatedStoreValue)
2021 SpeculatedStore = cast<StoreInst>(I);
2022
2023 // Do not hoist the instruction if any of its operands are defined but not
2024 // used in BB. The transformation will prevent the operand from
2025 // being sunk into the use block.
2026 for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i) {
2027 Instruction *OpI = dyn_cast<Instruction>(*i);
2028 if (!OpI || OpI->getParent() != BB || OpI->mayHaveSideEffects())
2029 continue; // Not a candidate for sinking.
2030
2031 ++SinkCandidateUseCounts[OpI];
2032 }
2033 }
2034
2035 // Consider any sink candidates which are only used in ThenBB as costs for
2036 // speculation. Note, while we iterate over a DenseMap here, we are summing
2037 // and so iteration order isn't significant.
2038 for (SmallDenseMap<Instruction *, unsigned, 4>::iterator
2039 I = SinkCandidateUseCounts.begin(),
2040 E = SinkCandidateUseCounts.end();
2041 I != E; ++I)
2042 if (I->first->hasNUses(I->second)) {
2043 ++SpeculatedInstructions;
2044 if (SpeculatedInstructions > 1)
2045 return false;
2046 }
2047
2048 // Check that the PHI nodes can be converted to selects.
2049 bool HaveRewritablePHIs = false;
2050 for (PHINode &PN : EndBB->phis()) {
2051 Value *OrigV = PN.getIncomingValueForBlock(BB);
2052 Value *ThenV = PN.getIncomingValueForBlock(ThenBB);
2053
2054 // FIXME: Try to remove some of the duplication with HoistThenElseCodeToIf.
2055 // Skip PHIs which are trivial.
2056 if (ThenV == OrigV)
2057 continue;
2058
2059 // Don't convert to selects if we could remove undefined behavior instead.
2060 if (passingValueIsAlwaysUndefined(OrigV, &PN) ||
2061 passingValueIsAlwaysUndefined(ThenV, &PN))
2062 return false;
2063
2064 HaveRewritablePHIs = true;
2065 ConstantExpr *OrigCE = dyn_cast<ConstantExpr>(OrigV);
2066 ConstantExpr *ThenCE = dyn_cast<ConstantExpr>(ThenV);
2067 if (!OrigCE && !ThenCE)
2068 continue; // Known safe and cheap.
2069
2070 if ((ThenCE && !isSafeToSpeculativelyExecute(ThenCE)) ||
2071 (OrigCE && !isSafeToSpeculativelyExecute(OrigCE)))
2072 return false;
2073 unsigned OrigCost = OrigCE ? ComputeSpeculationCost(OrigCE, TTI) : 0;
2074 unsigned ThenCost = ThenCE ? ComputeSpeculationCost(ThenCE, TTI) : 0;
2075 unsigned MaxCost =
2076 2 * PHINodeFoldingThreshold * TargetTransformInfo::TCC_Basic;
2077 if (OrigCost + ThenCost > MaxCost)
2078 return false;
2079
2080 // Account for the cost of an unfolded ConstantExpr which could end up
2081 // getting expanded into Instructions.
2082 // FIXME: This doesn't account for how many operations are combined in the
2083 // constant expression.
2084 ++SpeculatedInstructions;
2085 if (SpeculatedInstructions > 1)
2086 return false;
2087 }
2088
2089 // If there are no PHIs to process, bail early. This helps ensure idempotence
2090 // as well.
2091 if (!HaveRewritablePHIs && !(HoistCondStores && SpeculatedStoreValue))
2092 return false;
2093
2094 // If we get here, we can hoist the instruction and if-convert.
2095 LLVM_DEBUG(dbgs() << "SPECULATIVELY EXECUTING BB" << *ThenBB << "\n";);
2096
2097 // Insert a select of the value of the speculated store.
2098 if (SpeculatedStoreValue) {
2099 IRBuilder<NoFolder> Builder(BI);
2100 Value *TrueV = SpeculatedStore->getValueOperand();
2101 Value *FalseV = SpeculatedStoreValue;
2102 if (Invert)
2103 std::swap(TrueV, FalseV);
2104 Value *S = Builder.CreateSelect(
2105 BrCond, TrueV, FalseV, "spec.store.select", BI);
2106 SpeculatedStore->setOperand(0, S);
2107 SpeculatedStore->applyMergedLocation(BI->getDebugLoc(),
2108 SpeculatedStore->getDebugLoc());
2109 }
2110
2111 // Metadata can be dependent on the condition we are hoisting above.
2112 // Conservatively strip all metadata on the instruction.
2113 for (auto &I : *ThenBB)
2114 I.dropUnknownNonDebugMetadata();
2115
2116 // Hoist the instructions.
2117 BB->getInstList().splice(BI->getIterator(), ThenBB->getInstList(),
2118 ThenBB->begin(), std::prev(ThenBB->end()));
2119
2120 // Insert selects and rewrite the PHI operands.
2121 IRBuilder<NoFolder> Builder(BI);
2122 for (PHINode &PN : EndBB->phis()) {
2123 unsigned OrigI = PN.getBasicBlockIndex(BB);
2124 unsigned ThenI = PN.getBasicBlockIndex(ThenBB);
2125 Value *OrigV = PN.getIncomingValue(OrigI);
2126 Value *ThenV = PN.getIncomingValue(ThenI);
2127
2128 // Skip PHIs which are trivial.
2129 if (OrigV == ThenV)
2130 continue;
2131
2132 // Create a select whose true value is the speculatively executed value and
2133 // false value is the preexisting value. Swap them if the branch
2134 // destinations were inverted.
2135 Value *TrueV = ThenV, *FalseV = OrigV;
2136 if (Invert)
2137 std::swap(TrueV, FalseV);
2138 Value *V = Builder.CreateSelect(
2139 BrCond, TrueV, FalseV, "spec.select", BI);
2140 PN.setIncomingValue(OrigI, V);
2141 PN.setIncomingValue(ThenI, V);
2142 }
2143
2144 // Remove speculated dbg intrinsics.
2145 // FIXME: Is it possible to do this in a more elegant way? Moving/merging the
2146 // dbg value for the different flows and inserting it after the select.
2147 for (Instruction *I : SpeculatedDbgIntrinsics)
2148 I->eraseFromParent();
2149
2150 ++NumSpeculations;
2151 return true;
2152 }
2153
2154 /// Return true if we can thread a branch across this block.
BlockIsSimpleEnoughToThreadThrough(BasicBlock * BB)2155 static bool BlockIsSimpleEnoughToThreadThrough(BasicBlock *BB) {
2156 unsigned Size = 0;
2157
2158 for (Instruction &I : BB->instructionsWithoutDebug()) {
2159 if (Size > 10)
2160 return false; // Don't clone large BB's.
2161 ++Size;
2162
2163 // We can only support instructions that do not define values that are
2164 // live outside of the current basic block.
2165 for (User *U : I.users()) {
2166 Instruction *UI = cast<Instruction>(U);
2167 if (UI->getParent() != BB || isa<PHINode>(UI))
2168 return false;
2169 }
2170
2171 // Looks ok, continue checking.
2172 }
2173
2174 return true;
2175 }
2176
2177 /// If we have a conditional branch on a PHI node value that is defined in the
2178 /// same block as the branch and if any PHI entries are constants, thread edges
2179 /// corresponding to that entry to be branches to their ultimate destination.
FoldCondBranchOnPHI(BranchInst * BI,const DataLayout & DL,AssumptionCache * AC)2180 static bool FoldCondBranchOnPHI(BranchInst *BI, const DataLayout &DL,
2181 AssumptionCache *AC) {
2182 BasicBlock *BB = BI->getParent();
2183 PHINode *PN = dyn_cast<PHINode>(BI->getCondition());
2184 // NOTE: we currently cannot transform this case if the PHI node is used
2185 // outside of the block.
2186 if (!PN || PN->getParent() != BB || !PN->hasOneUse())
2187 return false;
2188
2189 // Degenerate case of a single entry PHI.
2190 if (PN->getNumIncomingValues() == 1) {
2191 FoldSingleEntryPHINodes(PN->getParent());
2192 return true;
2193 }
2194
2195 // Now we know that this block has multiple preds and two succs.
2196 if (!BlockIsSimpleEnoughToThreadThrough(BB))
2197 return false;
2198
2199 // Can't fold blocks that contain noduplicate or convergent calls.
2200 if (any_of(*BB, [](const Instruction &I) {
2201 const CallInst *CI = dyn_cast<CallInst>(&I);
2202 return CI && (CI->cannotDuplicate() || CI->isConvergent());
2203 }))
2204 return false;
2205
2206 // Okay, this is a simple enough basic block. See if any phi values are
2207 // constants.
2208 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
2209 ConstantInt *CB = dyn_cast<ConstantInt>(PN->getIncomingValue(i));
2210 if (!CB || !CB->getType()->isIntegerTy(1))
2211 continue;
2212
2213 // Okay, we now know that all edges from PredBB should be revectored to
2214 // branch to RealDest.
2215 BasicBlock *PredBB = PN->getIncomingBlock(i);
2216 BasicBlock *RealDest = BI->getSuccessor(!CB->getZExtValue());
2217
2218 if (RealDest == BB)
2219 continue; // Skip self loops.
2220 // Skip if the predecessor's terminator is an indirect branch.
2221 if (isa<IndirectBrInst>(PredBB->getTerminator()))
2222 continue;
2223
2224 // The dest block might have PHI nodes, other predecessors and other
2225 // difficult cases. Instead of being smart about this, just insert a new
2226 // block that jumps to the destination block, effectively splitting
2227 // the edge we are about to create.
2228 BasicBlock *EdgeBB =
2229 BasicBlock::Create(BB->getContext(), RealDest->getName() + ".critedge",
2230 RealDest->getParent(), RealDest);
2231 BranchInst *CritEdgeBranch = BranchInst::Create(RealDest, EdgeBB);
2232 CritEdgeBranch->setDebugLoc(BI->getDebugLoc());
2233
2234 // Update PHI nodes.
2235 AddPredecessorToBlock(RealDest, EdgeBB, BB);
2236
2237 // BB may have instructions that are being threaded over. Clone these
2238 // instructions into EdgeBB. We know that there will be no uses of the
2239 // cloned instructions outside of EdgeBB.
2240 BasicBlock::iterator InsertPt = EdgeBB->begin();
2241 DenseMap<Value *, Value *> TranslateMap; // Track translated values.
2242 for (BasicBlock::iterator BBI = BB->begin(); &*BBI != BI; ++BBI) {
2243 if (PHINode *PN = dyn_cast<PHINode>(BBI)) {
2244 TranslateMap[PN] = PN->getIncomingValueForBlock(PredBB);
2245 continue;
2246 }
2247 // Clone the instruction.
2248 Instruction *N = BBI->clone();
2249 if (BBI->hasName())
2250 N->setName(BBI->getName() + ".c");
2251
2252 // Update operands due to translation.
2253 for (User::op_iterator i = N->op_begin(), e = N->op_end(); i != e; ++i) {
2254 DenseMap<Value *, Value *>::iterator PI = TranslateMap.find(*i);
2255 if (PI != TranslateMap.end())
2256 *i = PI->second;
2257 }
2258
2259 // Check for trivial simplification.
2260 if (Value *V = SimplifyInstruction(N, {DL, nullptr, nullptr, AC})) {
2261 if (!BBI->use_empty())
2262 TranslateMap[&*BBI] = V;
2263 if (!N->mayHaveSideEffects()) {
2264 N->deleteValue(); // Instruction folded away, don't need actual inst
2265 N = nullptr;
2266 }
2267 } else {
2268 if (!BBI->use_empty())
2269 TranslateMap[&*BBI] = N;
2270 }
2271 if (N) {
2272 // Insert the new instruction into its new home.
2273 EdgeBB->getInstList().insert(InsertPt, N);
2274
2275 // Register the new instruction with the assumption cache if necessary.
2276 if (AC && match(N, m_Intrinsic<Intrinsic::assume>()))
2277 AC->registerAssumption(cast<IntrinsicInst>(N));
2278 }
2279 }
2280
2281 // Loop over all of the edges from PredBB to BB, changing them to branch
2282 // to EdgeBB instead.
2283 Instruction *PredBBTI = PredBB->getTerminator();
2284 for (unsigned i = 0, e = PredBBTI->getNumSuccessors(); i != e; ++i)
2285 if (PredBBTI->getSuccessor(i) == BB) {
2286 BB->removePredecessor(PredBB);
2287 PredBBTI->setSuccessor(i, EdgeBB);
2288 }
2289
2290 // Recurse, simplifying any other constants.
2291 return FoldCondBranchOnPHI(BI, DL, AC) || true;
2292 }
2293
2294 return false;
2295 }
2296
2297 /// Given a BB that starts with the specified two-entry PHI node,
2298 /// see if we can eliminate it.
FoldTwoEntryPHINode(PHINode * PN,const TargetTransformInfo & TTI,const DataLayout & DL)2299 static bool FoldTwoEntryPHINode(PHINode *PN, const TargetTransformInfo &TTI,
2300 const DataLayout &DL) {
2301 // Ok, this is a two entry PHI node. Check to see if this is a simple "if
2302 // statement", which has a very simple dominance structure. Basically, we
2303 // are trying to find the condition that is being branched on, which
2304 // subsequently causes this merge to happen. We really want control
2305 // dependence information for this check, but simplifycfg can't keep it up
2306 // to date, and this catches most of the cases we care about anyway.
2307 BasicBlock *BB = PN->getParent();
2308 const Function *Fn = BB->getParent();
2309 if (Fn && Fn->hasFnAttribute(Attribute::OptForFuzzing))
2310 return false;
2311
2312 BasicBlock *IfTrue, *IfFalse;
2313 Value *IfCond = GetIfCondition(BB, IfTrue, IfFalse);
2314 if (!IfCond ||
2315 // Don't bother if the branch will be constant folded trivially.
2316 isa<ConstantInt>(IfCond))
2317 return false;
2318
2319 // Okay, we found that we can merge this two-entry phi node into a select.
2320 // Doing so would require us to fold *all* two entry phi nodes in this block.
2321 // At some point this becomes non-profitable (particularly if the target
2322 // doesn't support cmov's). Only do this transformation if there are two or
2323 // fewer PHI nodes in this block.
2324 unsigned NumPhis = 0;
2325 for (BasicBlock::iterator I = BB->begin(); isa<PHINode>(I); ++NumPhis, ++I)
2326 if (NumPhis > 2)
2327 return false;
2328
2329 // Loop over the PHI's seeing if we can promote them all to select
2330 // instructions. While we are at it, keep track of the instructions
2331 // that need to be moved to the dominating block.
2332 SmallPtrSet<Instruction *, 4> AggressiveInsts;
2333 int BudgetRemaining =
2334 TwoEntryPHINodeFoldingThreshold * TargetTransformInfo::TCC_Basic;
2335
2336 for (BasicBlock::iterator II = BB->begin(); isa<PHINode>(II);) {
2337 PHINode *PN = cast<PHINode>(II++);
2338 if (Value *V = SimplifyInstruction(PN, {DL, PN})) {
2339 PN->replaceAllUsesWith(V);
2340 PN->eraseFromParent();
2341 continue;
2342 }
2343
2344 if (!DominatesMergePoint(PN->getIncomingValue(0), BB, AggressiveInsts,
2345 BudgetRemaining, TTI) ||
2346 !DominatesMergePoint(PN->getIncomingValue(1), BB, AggressiveInsts,
2347 BudgetRemaining, TTI))
2348 return false;
2349 }
2350
2351 // If we folded the first phi, PN dangles at this point. Refresh it. If
2352 // we ran out of PHIs then we simplified them all.
2353 PN = dyn_cast<PHINode>(BB->begin());
2354 if (!PN)
2355 return true;
2356
2357 // Return true if at least one of these is a 'not', and another is either
2358 // a 'not' too, or a constant.
2359 auto CanHoistNotFromBothValues = [](Value *V0, Value *V1) {
2360 if (!match(V0, m_Not(m_Value())))
2361 std::swap(V0, V1);
2362 auto Invertible = m_CombineOr(m_Not(m_Value()), m_AnyIntegralConstant());
2363 return match(V0, m_Not(m_Value())) && match(V1, Invertible);
2364 };
2365
2366 // Don't fold i1 branches on PHIs which contain binary operators, unless one
2367 // of the incoming values is an 'not' and another one is freely invertible.
2368 // These can often be turned into switches and other things.
2369 if (PN->getType()->isIntegerTy(1) &&
2370 (isa<BinaryOperator>(PN->getIncomingValue(0)) ||
2371 isa<BinaryOperator>(PN->getIncomingValue(1)) ||
2372 isa<BinaryOperator>(IfCond)) &&
2373 !CanHoistNotFromBothValues(PN->getIncomingValue(0),
2374 PN->getIncomingValue(1)))
2375 return false;
2376
2377 // If all PHI nodes are promotable, check to make sure that all instructions
2378 // in the predecessor blocks can be promoted as well. If not, we won't be able
2379 // to get rid of the control flow, so it's not worth promoting to select
2380 // instructions.
2381 BasicBlock *DomBlock = nullptr;
2382 BasicBlock *IfBlock1 = PN->getIncomingBlock(0);
2383 BasicBlock *IfBlock2 = PN->getIncomingBlock(1);
2384 if (cast<BranchInst>(IfBlock1->getTerminator())->isConditional()) {
2385 IfBlock1 = nullptr;
2386 } else {
2387 DomBlock = *pred_begin(IfBlock1);
2388 for (BasicBlock::iterator I = IfBlock1->begin(); !I->isTerminator(); ++I)
2389 if (!AggressiveInsts.count(&*I) && !isa<DbgInfoIntrinsic>(I)) {
2390 // This is not an aggressive instruction that we can promote.
2391 // Because of this, we won't be able to get rid of the control flow, so
2392 // the xform is not worth it.
2393 return false;
2394 }
2395 }
2396
2397 if (cast<BranchInst>(IfBlock2->getTerminator())->isConditional()) {
2398 IfBlock2 = nullptr;
2399 } else {
2400 DomBlock = *pred_begin(IfBlock2);
2401 for (BasicBlock::iterator I = IfBlock2->begin(); !I->isTerminator(); ++I)
2402 if (!AggressiveInsts.count(&*I) && !isa<DbgInfoIntrinsic>(I)) {
2403 // This is not an aggressive instruction that we can promote.
2404 // Because of this, we won't be able to get rid of the control flow, so
2405 // the xform is not worth it.
2406 return false;
2407 }
2408 }
2409 assert(DomBlock && "Failed to find root DomBlock");
2410
2411 LLVM_DEBUG(dbgs() << "FOUND IF CONDITION! " << *IfCond
2412 << " T: " << IfTrue->getName()
2413 << " F: " << IfFalse->getName() << "\n");
2414
2415 // If we can still promote the PHI nodes after this gauntlet of tests,
2416 // do all of the PHI's now.
2417 Instruction *InsertPt = DomBlock->getTerminator();
2418 IRBuilder<NoFolder> Builder(InsertPt);
2419
2420 // Move all 'aggressive' instructions, which are defined in the
2421 // conditional parts of the if's up to the dominating block.
2422 if (IfBlock1)
2423 hoistAllInstructionsInto(DomBlock, InsertPt, IfBlock1);
2424 if (IfBlock2)
2425 hoistAllInstructionsInto(DomBlock, InsertPt, IfBlock2);
2426
2427 // Propagate fast-math-flags from phi nodes to replacement selects.
2428 IRBuilder<>::FastMathFlagGuard FMFGuard(Builder);
2429 while (PHINode *PN = dyn_cast<PHINode>(BB->begin())) {
2430 if (isa<FPMathOperator>(PN))
2431 Builder.setFastMathFlags(PN->getFastMathFlags());
2432
2433 // Change the PHI node into a select instruction.
2434 Value *TrueVal = PN->getIncomingValue(PN->getIncomingBlock(0) == IfFalse);
2435 Value *FalseVal = PN->getIncomingValue(PN->getIncomingBlock(0) == IfTrue);
2436
2437 Value *Sel = Builder.CreateSelect(IfCond, TrueVal, FalseVal, "", InsertPt);
2438 PN->replaceAllUsesWith(Sel);
2439 Sel->takeName(PN);
2440 PN->eraseFromParent();
2441 }
2442
2443 // At this point, IfBlock1 and IfBlock2 are both empty, so our if statement
2444 // has been flattened. Change DomBlock to jump directly to our new block to
2445 // avoid other simplifycfg's kicking in on the diamond.
2446 Instruction *OldTI = DomBlock->getTerminator();
2447 Builder.SetInsertPoint(OldTI);
2448 Builder.CreateBr(BB);
2449 OldTI->eraseFromParent();
2450 return true;
2451 }
2452
2453 /// If we found a conditional branch that goes to two returning blocks,
2454 /// try to merge them together into one return,
2455 /// introducing a select if the return values disagree.
SimplifyCondBranchToTwoReturns(BranchInst * BI,IRBuilder<> & Builder)2456 static bool SimplifyCondBranchToTwoReturns(BranchInst *BI,
2457 IRBuilder<> &Builder) {
2458 assert(BI->isConditional() && "Must be a conditional branch");
2459 BasicBlock *TrueSucc = BI->getSuccessor(0);
2460 BasicBlock *FalseSucc = BI->getSuccessor(1);
2461 ReturnInst *TrueRet = cast<ReturnInst>(TrueSucc->getTerminator());
2462 ReturnInst *FalseRet = cast<ReturnInst>(FalseSucc->getTerminator());
2463
2464 // Check to ensure both blocks are empty (just a return) or optionally empty
2465 // with PHI nodes. If there are other instructions, merging would cause extra
2466 // computation on one path or the other.
2467 if (!TrueSucc->getFirstNonPHIOrDbg()->isTerminator())
2468 return false;
2469 if (!FalseSucc->getFirstNonPHIOrDbg()->isTerminator())
2470 return false;
2471
2472 Builder.SetInsertPoint(BI);
2473 // Okay, we found a branch that is going to two return nodes. If
2474 // there is no return value for this function, just change the
2475 // branch into a return.
2476 if (FalseRet->getNumOperands() == 0) {
2477 TrueSucc->removePredecessor(BI->getParent());
2478 FalseSucc->removePredecessor(BI->getParent());
2479 Builder.CreateRetVoid();
2480 EraseTerminatorAndDCECond(BI);
2481 return true;
2482 }
2483
2484 // Otherwise, figure out what the true and false return values are
2485 // so we can insert a new select instruction.
2486 Value *TrueValue = TrueRet->getReturnValue();
2487 Value *FalseValue = FalseRet->getReturnValue();
2488
2489 // Unwrap any PHI nodes in the return blocks.
2490 if (PHINode *TVPN = dyn_cast_or_null<PHINode>(TrueValue))
2491 if (TVPN->getParent() == TrueSucc)
2492 TrueValue = TVPN->getIncomingValueForBlock(BI->getParent());
2493 if (PHINode *FVPN = dyn_cast_or_null<PHINode>(FalseValue))
2494 if (FVPN->getParent() == FalseSucc)
2495 FalseValue = FVPN->getIncomingValueForBlock(BI->getParent());
2496
2497 // In order for this transformation to be safe, we must be able to
2498 // unconditionally execute both operands to the return. This is
2499 // normally the case, but we could have a potentially-trapping
2500 // constant expression that prevents this transformation from being
2501 // safe.
2502 if (ConstantExpr *TCV = dyn_cast_or_null<ConstantExpr>(TrueValue))
2503 if (TCV->canTrap())
2504 return false;
2505 if (ConstantExpr *FCV = dyn_cast_or_null<ConstantExpr>(FalseValue))
2506 if (FCV->canTrap())
2507 return false;
2508
2509 // Okay, we collected all the mapped values and checked them for sanity, and
2510 // defined to really do this transformation. First, update the CFG.
2511 TrueSucc->removePredecessor(BI->getParent());
2512 FalseSucc->removePredecessor(BI->getParent());
2513
2514 // Insert select instructions where needed.
2515 Value *BrCond = BI->getCondition();
2516 if (TrueValue) {
2517 // Insert a select if the results differ.
2518 if (TrueValue == FalseValue || isa<UndefValue>(FalseValue)) {
2519 } else if (isa<UndefValue>(TrueValue)) {
2520 TrueValue = FalseValue;
2521 } else {
2522 TrueValue =
2523 Builder.CreateSelect(BrCond, TrueValue, FalseValue, "retval", BI);
2524 }
2525 }
2526
2527 Value *RI =
2528 !TrueValue ? Builder.CreateRetVoid() : Builder.CreateRet(TrueValue);
2529
2530 (void)RI;
2531
2532 LLVM_DEBUG(dbgs() << "\nCHANGING BRANCH TO TWO RETURNS INTO SELECT:"
2533 << "\n " << *BI << "NewRet = " << *RI << "TRUEBLOCK: "
2534 << *TrueSucc << "FALSEBLOCK: " << *FalseSucc);
2535
2536 EraseTerminatorAndDCECond(BI);
2537
2538 return true;
2539 }
2540
2541 /// Return true if the given instruction is available
2542 /// in its predecessor block. If yes, the instruction will be removed.
tryCSEWithPredecessor(Instruction * Inst,BasicBlock * PB)2543 static bool tryCSEWithPredecessor(Instruction *Inst, BasicBlock *PB) {
2544 if (!isa<BinaryOperator>(Inst) && !isa<CmpInst>(Inst))
2545 return false;
2546 for (Instruction &I : *PB) {
2547 Instruction *PBI = &I;
2548 // Check whether Inst and PBI generate the same value.
2549 if (Inst->isIdenticalTo(PBI)) {
2550 Inst->replaceAllUsesWith(PBI);
2551 Inst->eraseFromParent();
2552 return true;
2553 }
2554 }
2555 return false;
2556 }
2557
2558 /// Return true if either PBI or BI has branch weight available, and store
2559 /// the weights in {Pred|Succ}{True|False}Weight. If one of PBI and BI does
2560 /// 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)2561 static bool extractPredSuccWeights(BranchInst *PBI, BranchInst *BI,
2562 uint64_t &PredTrueWeight,
2563 uint64_t &PredFalseWeight,
2564 uint64_t &SuccTrueWeight,
2565 uint64_t &SuccFalseWeight) {
2566 bool PredHasWeights =
2567 PBI->extractProfMetadata(PredTrueWeight, PredFalseWeight);
2568 bool SuccHasWeights =
2569 BI->extractProfMetadata(SuccTrueWeight, SuccFalseWeight);
2570 if (PredHasWeights || SuccHasWeights) {
2571 if (!PredHasWeights)
2572 PredTrueWeight = PredFalseWeight = 1;
2573 if (!SuccHasWeights)
2574 SuccTrueWeight = SuccFalseWeight = 1;
2575 return true;
2576 } else {
2577 return false;
2578 }
2579 }
2580
2581 /// If this basic block is simple enough, and if a predecessor branches to us
2582 /// and one of our successors, fold the block into the predecessor and use
2583 /// logical operations to pick the right destination.
FoldBranchToCommonDest(BranchInst * BI,MemorySSAUpdater * MSSAU,unsigned BonusInstThreshold)2584 bool llvm::FoldBranchToCommonDest(BranchInst *BI, MemorySSAUpdater *MSSAU,
2585 unsigned BonusInstThreshold) {
2586 BasicBlock *BB = BI->getParent();
2587
2588 const unsigned PredCount = pred_size(BB);
2589
2590 Instruction *Cond = nullptr;
2591 if (BI->isConditional())
2592 Cond = dyn_cast<Instruction>(BI->getCondition());
2593 else {
2594 // For unconditional branch, check for a simple CFG pattern, where
2595 // BB has a single predecessor and BB's successor is also its predecessor's
2596 // successor. If such pattern exists, check for CSE between BB and its
2597 // predecessor.
2598 if (BasicBlock *PB = BB->getSinglePredecessor())
2599 if (BranchInst *PBI = dyn_cast<BranchInst>(PB->getTerminator()))
2600 if (PBI->isConditional() &&
2601 (BI->getSuccessor(0) == PBI->getSuccessor(0) ||
2602 BI->getSuccessor(0) == PBI->getSuccessor(1))) {
2603 for (auto I = BB->instructionsWithoutDebug().begin(),
2604 E = BB->instructionsWithoutDebug().end();
2605 I != E;) {
2606 Instruction *Curr = &*I++;
2607 if (isa<CmpInst>(Curr)) {
2608 Cond = Curr;
2609 break;
2610 }
2611 // Quit if we can't remove this instruction.
2612 if (!tryCSEWithPredecessor(Curr, PB))
2613 return false;
2614 }
2615 }
2616
2617 if (!Cond)
2618 return false;
2619 }
2620
2621 if (!Cond || (!isa<CmpInst>(Cond) && !isa<BinaryOperator>(Cond)) ||
2622 Cond->getParent() != BB || !Cond->hasOneUse())
2623 return false;
2624
2625 // Make sure the instruction after the condition is the cond branch.
2626 BasicBlock::iterator CondIt = ++Cond->getIterator();
2627
2628 // Ignore dbg intrinsics.
2629 while (isa<DbgInfoIntrinsic>(CondIt))
2630 ++CondIt;
2631
2632 if (&*CondIt != BI)
2633 return false;
2634
2635 // Only allow this transformation if computing the condition doesn't involve
2636 // too many instructions and these involved instructions can be executed
2637 // unconditionally. We denote all involved instructions except the condition
2638 // as "bonus instructions", and only allow this transformation when the
2639 // number of the bonus instructions we'll need to create when cloning into
2640 // each predecessor does not exceed a certain threshold.
2641 unsigned NumBonusInsts = 0;
2642 for (auto I = BB->begin(); Cond != &*I; ++I) {
2643 // Ignore dbg intrinsics.
2644 if (isa<DbgInfoIntrinsic>(I))
2645 continue;
2646 if (!I->hasOneUse() || !isSafeToSpeculativelyExecute(&*I))
2647 return false;
2648 // I has only one use and can be executed unconditionally.
2649 Instruction *User = dyn_cast<Instruction>(I->user_back());
2650 if (User == nullptr || User->getParent() != BB)
2651 return false;
2652 // I is used in the same BB. Since BI uses Cond and doesn't have more slots
2653 // to use any other instruction, User must be an instruction between next(I)
2654 // and Cond.
2655
2656 // Account for the cost of duplicating this instruction into each
2657 // predecessor.
2658 NumBonusInsts += PredCount;
2659 // Early exits once we reach the limit.
2660 if (NumBonusInsts > BonusInstThreshold)
2661 return false;
2662 }
2663
2664 // Cond is known to be a compare or binary operator. Check to make sure that
2665 // neither operand is a potentially-trapping constant expression.
2666 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(0)))
2667 if (CE->canTrap())
2668 return false;
2669 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(1)))
2670 if (CE->canTrap())
2671 return false;
2672
2673 // Finally, don't infinitely unroll conditional loops.
2674 BasicBlock *TrueDest = BI->getSuccessor(0);
2675 BasicBlock *FalseDest = (BI->isConditional()) ? BI->getSuccessor(1) : nullptr;
2676 if (TrueDest == BB || FalseDest == BB)
2677 return false;
2678
2679 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
2680 BasicBlock *PredBlock = *PI;
2681 BranchInst *PBI = dyn_cast<BranchInst>(PredBlock->getTerminator());
2682
2683 // Check that we have two conditional branches. If there is a PHI node in
2684 // the common successor, verify that the same value flows in from both
2685 // blocks.
2686 SmallVector<PHINode *, 4> PHIs;
2687 if (!PBI || PBI->isUnconditional() ||
2688 (BI->isConditional() && !SafeToMergeTerminators(BI, PBI)) ||
2689 (!BI->isConditional() &&
2690 !isProfitableToFoldUnconditional(BI, PBI, Cond, PHIs)))
2691 continue;
2692
2693 // Determine if the two branches share a common destination.
2694 Instruction::BinaryOps Opc = Instruction::BinaryOpsEnd;
2695 bool InvertPredCond = false;
2696
2697 if (BI->isConditional()) {
2698 if (PBI->getSuccessor(0) == TrueDest) {
2699 Opc = Instruction::Or;
2700 } else if (PBI->getSuccessor(1) == FalseDest) {
2701 Opc = Instruction::And;
2702 } else if (PBI->getSuccessor(0) == FalseDest) {
2703 Opc = Instruction::And;
2704 InvertPredCond = true;
2705 } else if (PBI->getSuccessor(1) == TrueDest) {
2706 Opc = Instruction::Or;
2707 InvertPredCond = true;
2708 } else {
2709 continue;
2710 }
2711 } else {
2712 if (PBI->getSuccessor(0) != TrueDest && PBI->getSuccessor(1) != TrueDest)
2713 continue;
2714 }
2715
2716 LLVM_DEBUG(dbgs() << "FOLDING BRANCH TO COMMON DEST:\n" << *PBI << *BB);
2717 IRBuilder<> Builder(PBI);
2718
2719 // If we need to invert the condition in the pred block to match, do so now.
2720 if (InvertPredCond) {
2721 Value *NewCond = PBI->getCondition();
2722
2723 if (NewCond->hasOneUse() && isa<CmpInst>(NewCond)) {
2724 CmpInst *CI = cast<CmpInst>(NewCond);
2725 CI->setPredicate(CI->getInversePredicate());
2726 } else {
2727 NewCond =
2728 Builder.CreateNot(NewCond, PBI->getCondition()->getName() + ".not");
2729 }
2730
2731 PBI->setCondition(NewCond);
2732 PBI->swapSuccessors();
2733 }
2734
2735 // If we have bonus instructions, clone them into the predecessor block.
2736 // Note that there may be multiple predecessor blocks, so we cannot move
2737 // bonus instructions to a predecessor block.
2738 ValueToValueMapTy VMap; // maps original values to cloned values
2739 // We already make sure Cond is the last instruction before BI. Therefore,
2740 // all instructions before Cond other than DbgInfoIntrinsic are bonus
2741 // instructions.
2742 for (auto BonusInst = BB->begin(); Cond != &*BonusInst; ++BonusInst) {
2743 if (isa<DbgInfoIntrinsic>(BonusInst))
2744 continue;
2745 Instruction *NewBonusInst = BonusInst->clone();
2746 RemapInstruction(NewBonusInst, VMap,
2747 RF_NoModuleLevelChanges | RF_IgnoreMissingLocals);
2748 VMap[&*BonusInst] = NewBonusInst;
2749
2750 // If we moved a load, we cannot any longer claim any knowledge about
2751 // its potential value. The previous information might have been valid
2752 // only given the branch precondition.
2753 // For an analogous reason, we must also drop all the metadata whose
2754 // semantics we don't understand.
2755 NewBonusInst->dropUnknownNonDebugMetadata();
2756
2757 PredBlock->getInstList().insert(PBI->getIterator(), NewBonusInst);
2758 NewBonusInst->takeName(&*BonusInst);
2759 BonusInst->setName(BonusInst->getName() + ".old");
2760 }
2761
2762 // Clone Cond into the predecessor basic block, and or/and the
2763 // two conditions together.
2764 Instruction *CondInPred = Cond->clone();
2765 RemapInstruction(CondInPred, VMap,
2766 RF_NoModuleLevelChanges | RF_IgnoreMissingLocals);
2767 PredBlock->getInstList().insert(PBI->getIterator(), CondInPred);
2768 CondInPred->takeName(Cond);
2769 Cond->setName(CondInPred->getName() + ".old");
2770
2771 if (BI->isConditional()) {
2772 Instruction *NewCond = cast<Instruction>(
2773 Builder.CreateBinOp(Opc, PBI->getCondition(), CondInPred, "or.cond"));
2774 PBI->setCondition(NewCond);
2775
2776 uint64_t PredTrueWeight, PredFalseWeight, SuccTrueWeight, SuccFalseWeight;
2777 bool HasWeights =
2778 extractPredSuccWeights(PBI, BI, PredTrueWeight, PredFalseWeight,
2779 SuccTrueWeight, SuccFalseWeight);
2780 SmallVector<uint64_t, 8> NewWeights;
2781
2782 if (PBI->getSuccessor(0) == BB) {
2783 if (HasWeights) {
2784 // PBI: br i1 %x, BB, FalseDest
2785 // BI: br i1 %y, TrueDest, FalseDest
2786 // TrueWeight is TrueWeight for PBI * TrueWeight for BI.
2787 NewWeights.push_back(PredTrueWeight * SuccTrueWeight);
2788 // FalseWeight is FalseWeight for PBI * TotalWeight for BI +
2789 // TrueWeight for PBI * FalseWeight for BI.
2790 // We assume that total weights of a BranchInst can fit into 32 bits.
2791 // Therefore, we will not have overflow using 64-bit arithmetic.
2792 NewWeights.push_back(PredFalseWeight *
2793 (SuccFalseWeight + SuccTrueWeight) +
2794 PredTrueWeight * SuccFalseWeight);
2795 }
2796 AddPredecessorToBlock(TrueDest, PredBlock, BB, MSSAU);
2797 PBI->setSuccessor(0, TrueDest);
2798 }
2799 if (PBI->getSuccessor(1) == BB) {
2800 if (HasWeights) {
2801 // PBI: br i1 %x, TrueDest, BB
2802 // BI: br i1 %y, TrueDest, FalseDest
2803 // TrueWeight is TrueWeight for PBI * TotalWeight for BI +
2804 // FalseWeight for PBI * TrueWeight for BI.
2805 NewWeights.push_back(PredTrueWeight *
2806 (SuccFalseWeight + SuccTrueWeight) +
2807 PredFalseWeight * SuccTrueWeight);
2808 // FalseWeight is FalseWeight for PBI * FalseWeight for BI.
2809 NewWeights.push_back(PredFalseWeight * SuccFalseWeight);
2810 }
2811 AddPredecessorToBlock(FalseDest, PredBlock, BB, MSSAU);
2812 PBI->setSuccessor(1, FalseDest);
2813 }
2814 if (NewWeights.size() == 2) {
2815 // Halve the weights if any of them cannot fit in an uint32_t
2816 FitWeights(NewWeights);
2817
2818 SmallVector<uint32_t, 8> MDWeights(NewWeights.begin(),
2819 NewWeights.end());
2820 setBranchWeights(PBI, MDWeights[0], MDWeights[1]);
2821 } else
2822 PBI->setMetadata(LLVMContext::MD_prof, nullptr);
2823 } else {
2824 // Update PHI nodes in the common successors.
2825 for (unsigned i = 0, e = PHIs.size(); i != e; ++i) {
2826 ConstantInt *PBI_C = cast<ConstantInt>(
2827 PHIs[i]->getIncomingValueForBlock(PBI->getParent()));
2828 assert(PBI_C->getType()->isIntegerTy(1));
2829 Instruction *MergedCond = nullptr;
2830 if (PBI->getSuccessor(0) == TrueDest) {
2831 // Create (PBI_Cond and PBI_C) or (!PBI_Cond and BI_Value)
2832 // PBI_C is true: PBI_Cond or (!PBI_Cond and BI_Value)
2833 // is false: !PBI_Cond and BI_Value
2834 Instruction *NotCond = cast<Instruction>(
2835 Builder.CreateNot(PBI->getCondition(), "not.cond"));
2836 MergedCond = cast<Instruction>(
2837 Builder.CreateBinOp(Instruction::And, NotCond, CondInPred,
2838 "and.cond"));
2839 if (PBI_C->isOne())
2840 MergedCond = cast<Instruction>(Builder.CreateBinOp(
2841 Instruction::Or, PBI->getCondition(), MergedCond, "or.cond"));
2842 } else {
2843 // Create (PBI_Cond and BI_Value) or (!PBI_Cond and PBI_C)
2844 // PBI_C is true: (PBI_Cond and BI_Value) or (!PBI_Cond)
2845 // is false: PBI_Cond and BI_Value
2846 MergedCond = cast<Instruction>(Builder.CreateBinOp(
2847 Instruction::And, PBI->getCondition(), CondInPred, "and.cond"));
2848 if (PBI_C->isOne()) {
2849 Instruction *NotCond = cast<Instruction>(
2850 Builder.CreateNot(PBI->getCondition(), "not.cond"));
2851 MergedCond = cast<Instruction>(Builder.CreateBinOp(
2852 Instruction::Or, NotCond, MergedCond, "or.cond"));
2853 }
2854 }
2855 // Update PHI Node.
2856 PHIs[i]->setIncomingValueForBlock(PBI->getParent(), MergedCond);
2857 }
2858
2859 // PBI is changed to branch to TrueDest below. Remove itself from
2860 // potential phis from all other successors.
2861 if (MSSAU)
2862 MSSAU->changeCondBranchToUnconditionalTo(PBI, TrueDest);
2863
2864 // Change PBI from Conditional to Unconditional.
2865 BranchInst *New_PBI = BranchInst::Create(TrueDest, PBI);
2866 EraseTerminatorAndDCECond(PBI, MSSAU);
2867 PBI = New_PBI;
2868 }
2869
2870 // If BI was a loop latch, it may have had associated loop metadata.
2871 // We need to copy it to the new latch, that is, PBI.
2872 if (MDNode *LoopMD = BI->getMetadata(LLVMContext::MD_loop))
2873 PBI->setMetadata(LLVMContext::MD_loop, LoopMD);
2874
2875 // TODO: If BB is reachable from all paths through PredBlock, then we
2876 // could replace PBI's branch probabilities with BI's.
2877
2878 // Copy any debug value intrinsics into the end of PredBlock.
2879 for (Instruction &I : *BB)
2880 if (isa<DbgInfoIntrinsic>(I))
2881 I.clone()->insertBefore(PBI);
2882
2883 return true;
2884 }
2885 return false;
2886 }
2887
2888 // If there is only one store in BB1 and BB2, return it, otherwise return
2889 // nullptr.
findUniqueStoreInBlocks(BasicBlock * BB1,BasicBlock * BB2)2890 static StoreInst *findUniqueStoreInBlocks(BasicBlock *BB1, BasicBlock *BB2) {
2891 StoreInst *S = nullptr;
2892 for (auto *BB : {BB1, BB2}) {
2893 if (!BB)
2894 continue;
2895 for (auto &I : *BB)
2896 if (auto *SI = dyn_cast<StoreInst>(&I)) {
2897 if (S)
2898 // Multiple stores seen.
2899 return nullptr;
2900 else
2901 S = SI;
2902 }
2903 }
2904 return S;
2905 }
2906
ensureValueAvailableInSuccessor(Value * V,BasicBlock * BB,Value * AlternativeV=nullptr)2907 static Value *ensureValueAvailableInSuccessor(Value *V, BasicBlock *BB,
2908 Value *AlternativeV = nullptr) {
2909 // PHI is going to be a PHI node that allows the value V that is defined in
2910 // BB to be referenced in BB's only successor.
2911 //
2912 // If AlternativeV is nullptr, the only value we care about in PHI is V. It
2913 // doesn't matter to us what the other operand is (it'll never get used). We
2914 // could just create a new PHI with an undef incoming value, but that could
2915 // increase register pressure if EarlyCSE/InstCombine can't fold it with some
2916 // other PHI. So here we directly look for some PHI in BB's successor with V
2917 // as an incoming operand. If we find one, we use it, else we create a new
2918 // one.
2919 //
2920 // If AlternativeV is not nullptr, we care about both incoming values in PHI.
2921 // PHI must be exactly: phi <ty> [ %BB, %V ], [ %OtherBB, %AlternativeV]
2922 // where OtherBB is the single other predecessor of BB's only successor.
2923 PHINode *PHI = nullptr;
2924 BasicBlock *Succ = BB->getSingleSuccessor();
2925
2926 for (auto I = Succ->begin(); isa<PHINode>(I); ++I)
2927 if (cast<PHINode>(I)->getIncomingValueForBlock(BB) == V) {
2928 PHI = cast<PHINode>(I);
2929 if (!AlternativeV)
2930 break;
2931
2932 assert(Succ->hasNPredecessors(2));
2933 auto PredI = pred_begin(Succ);
2934 BasicBlock *OtherPredBB = *PredI == BB ? *++PredI : *PredI;
2935 if (PHI->getIncomingValueForBlock(OtherPredBB) == AlternativeV)
2936 break;
2937 PHI = nullptr;
2938 }
2939 if (PHI)
2940 return PHI;
2941
2942 // If V is not an instruction defined in BB, just return it.
2943 if (!AlternativeV &&
2944 (!isa<Instruction>(V) || cast<Instruction>(V)->getParent() != BB))
2945 return V;
2946
2947 PHI = PHINode::Create(V->getType(), 2, "simplifycfg.merge", &Succ->front());
2948 PHI->addIncoming(V, BB);
2949 for (BasicBlock *PredBB : predecessors(Succ))
2950 if (PredBB != BB)
2951 PHI->addIncoming(
2952 AlternativeV ? AlternativeV : UndefValue::get(V->getType()), PredBB);
2953 return PHI;
2954 }
2955
mergeConditionalStoreToAddress(BasicBlock * PTB,BasicBlock * PFB,BasicBlock * QTB,BasicBlock * QFB,BasicBlock * PostBB,Value * Address,bool InvertPCond,bool InvertQCond,const DataLayout & DL,const TargetTransformInfo & TTI)2956 static bool mergeConditionalStoreToAddress(BasicBlock *PTB, BasicBlock *PFB,
2957 BasicBlock *QTB, BasicBlock *QFB,
2958 BasicBlock *PostBB, Value *Address,
2959 bool InvertPCond, bool InvertQCond,
2960 const DataLayout &DL,
2961 const TargetTransformInfo &TTI) {
2962 // For every pointer, there must be exactly two stores, one coming from
2963 // PTB or PFB, and the other from QTB or QFB. We don't support more than one
2964 // store (to any address) in PTB,PFB or QTB,QFB.
2965 // FIXME: We could relax this restriction with a bit more work and performance
2966 // testing.
2967 StoreInst *PStore = findUniqueStoreInBlocks(PTB, PFB);
2968 StoreInst *QStore = findUniqueStoreInBlocks(QTB, QFB);
2969 if (!PStore || !QStore)
2970 return false;
2971
2972 // Now check the stores are compatible.
2973 if (!QStore->isUnordered() || !PStore->isUnordered())
2974 return false;
2975
2976 // Check that sinking the store won't cause program behavior changes. Sinking
2977 // the store out of the Q blocks won't change any behavior as we're sinking
2978 // from a block to its unconditional successor. But we're moving a store from
2979 // the P blocks down through the middle block (QBI) and past both QFB and QTB.
2980 // So we need to check that there are no aliasing loads or stores in
2981 // QBI, QTB and QFB. We also need to check there are no conflicting memory
2982 // operations between PStore and the end of its parent block.
2983 //
2984 // The ideal way to do this is to query AliasAnalysis, but we don't
2985 // preserve AA currently so that is dangerous. Be super safe and just
2986 // check there are no other memory operations at all.
2987 for (auto &I : *QFB->getSinglePredecessor())
2988 if (I.mayReadOrWriteMemory())
2989 return false;
2990 for (auto &I : *QFB)
2991 if (&I != QStore && I.mayReadOrWriteMemory())
2992 return false;
2993 if (QTB)
2994 for (auto &I : *QTB)
2995 if (&I != QStore && I.mayReadOrWriteMemory())
2996 return false;
2997 for (auto I = BasicBlock::iterator(PStore), E = PStore->getParent()->end();
2998 I != E; ++I)
2999 if (&*I != PStore && I->mayReadOrWriteMemory())
3000 return false;
3001
3002 // If we're not in aggressive mode, we only optimize if we have some
3003 // confidence that by optimizing we'll allow P and/or Q to be if-converted.
3004 auto IsWorthwhile = [&](BasicBlock *BB, ArrayRef<StoreInst *> FreeStores) {
3005 if (!BB)
3006 return true;
3007 // Heuristic: if the block can be if-converted/phi-folded and the
3008 // instructions inside are all cheap (arithmetic/GEPs), it's worthwhile to
3009 // thread this store.
3010 int BudgetRemaining =
3011 PHINodeFoldingThreshold * TargetTransformInfo::TCC_Basic;
3012 for (auto &I : BB->instructionsWithoutDebug()) {
3013 // Consider terminator instruction to be free.
3014 if (I.isTerminator())
3015 continue;
3016 // If this is one the stores that we want to speculate out of this BB,
3017 // then don't count it's cost, consider it to be free.
3018 if (auto *S = dyn_cast<StoreInst>(&I))
3019 if (llvm::find(FreeStores, S))
3020 continue;
3021 // Else, we have a white-list of instructions that we are ak speculating.
3022 if (!isa<BinaryOperator>(I) && !isa<GetElementPtrInst>(I))
3023 return false; // Not in white-list - not worthwhile folding.
3024 // And finally, if this is a non-free instruction that we are okay
3025 // speculating, ensure that we consider the speculation budget.
3026 BudgetRemaining -= TTI.getUserCost(&I);
3027 if (BudgetRemaining < 0)
3028 return false; // Eagerly refuse to fold as soon as we're out of budget.
3029 }
3030 assert(BudgetRemaining >= 0 &&
3031 "When we run out of budget we will eagerly return from within the "
3032 "per-instruction loop.");
3033 return true;
3034 };
3035
3036 const SmallVector<StoreInst *, 2> FreeStores = {PStore, QStore};
3037 if (!MergeCondStoresAggressively &&
3038 (!IsWorthwhile(PTB, FreeStores) || !IsWorthwhile(PFB, FreeStores) ||
3039 !IsWorthwhile(QTB, FreeStores) || !IsWorthwhile(QFB, FreeStores)))
3040 return false;
3041
3042 // If PostBB has more than two predecessors, we need to split it so we can
3043 // sink the store.
3044 if (std::next(pred_begin(PostBB), 2) != pred_end(PostBB)) {
3045 // We know that QFB's only successor is PostBB. And QFB has a single
3046 // predecessor. If QTB exists, then its only successor is also PostBB.
3047 // If QTB does not exist, then QFB's only predecessor has a conditional
3048 // branch to QFB and PostBB.
3049 BasicBlock *TruePred = QTB ? QTB : QFB->getSinglePredecessor();
3050 BasicBlock *NewBB = SplitBlockPredecessors(PostBB, { QFB, TruePred},
3051 "condstore.split");
3052 if (!NewBB)
3053 return false;
3054 PostBB = NewBB;
3055 }
3056
3057 // OK, we're going to sink the stores to PostBB. The store has to be
3058 // conditional though, so first create the predicate.
3059 Value *PCond = cast<BranchInst>(PFB->getSinglePredecessor()->getTerminator())
3060 ->getCondition();
3061 Value *QCond = cast<BranchInst>(QFB->getSinglePredecessor()->getTerminator())
3062 ->getCondition();
3063
3064 Value *PPHI = ensureValueAvailableInSuccessor(PStore->getValueOperand(),
3065 PStore->getParent());
3066 Value *QPHI = ensureValueAvailableInSuccessor(QStore->getValueOperand(),
3067 QStore->getParent(), PPHI);
3068
3069 IRBuilder<> QB(&*PostBB->getFirstInsertionPt());
3070
3071 Value *PPred = PStore->getParent() == PTB ? PCond : QB.CreateNot(PCond);
3072 Value *QPred = QStore->getParent() == QTB ? QCond : QB.CreateNot(QCond);
3073
3074 if (InvertPCond)
3075 PPred = QB.CreateNot(PPred);
3076 if (InvertQCond)
3077 QPred = QB.CreateNot(QPred);
3078 Value *CombinedPred = QB.CreateOr(PPred, QPred);
3079
3080 auto *T =
3081 SplitBlockAndInsertIfThen(CombinedPred, &*QB.GetInsertPoint(), false);
3082 QB.SetInsertPoint(T);
3083 StoreInst *SI = cast<StoreInst>(QB.CreateStore(QPHI, Address));
3084 AAMDNodes AAMD;
3085 PStore->getAAMetadata(AAMD, /*Merge=*/false);
3086 PStore->getAAMetadata(AAMD, /*Merge=*/true);
3087 SI->setAAMetadata(AAMD);
3088 unsigned PAlignment = PStore->getAlignment();
3089 unsigned QAlignment = QStore->getAlignment();
3090 unsigned TypeAlignment =
3091 DL.getABITypeAlignment(SI->getValueOperand()->getType());
3092 unsigned MinAlignment;
3093 unsigned MaxAlignment;
3094 std::tie(MinAlignment, MaxAlignment) = std::minmax(PAlignment, QAlignment);
3095 // Choose the minimum alignment. If we could prove both stores execute, we
3096 // could use biggest one. In this case, though, we only know that one of the
3097 // stores executes. And we don't know it's safe to take the alignment from a
3098 // store that doesn't execute.
3099 if (MinAlignment != 0) {
3100 // Choose the minimum of all non-zero alignments.
3101 SI->setAlignment(Align(MinAlignment));
3102 } else if (MaxAlignment != 0) {
3103 // Choose the minimal alignment between the non-zero alignment and the ABI
3104 // default alignment for the type of the stored value.
3105 SI->setAlignment(Align(std::min(MaxAlignment, TypeAlignment)));
3106 } else {
3107 // If both alignments are zero, use ABI default alignment for the type of
3108 // the stored value.
3109 SI->setAlignment(Align(TypeAlignment));
3110 }
3111
3112 QStore->eraseFromParent();
3113 PStore->eraseFromParent();
3114
3115 return true;
3116 }
3117
mergeConditionalStores(BranchInst * PBI,BranchInst * QBI,const DataLayout & DL,const TargetTransformInfo & TTI)3118 static bool mergeConditionalStores(BranchInst *PBI, BranchInst *QBI,
3119 const DataLayout &DL,
3120 const TargetTransformInfo &TTI) {
3121 // The intention here is to find diamonds or triangles (see below) where each
3122 // conditional block contains a store to the same address. Both of these
3123 // stores are conditional, so they can't be unconditionally sunk. But it may
3124 // be profitable to speculatively sink the stores into one merged store at the
3125 // end, and predicate the merged store on the union of the two conditions of
3126 // PBI and QBI.
3127 //
3128 // This can reduce the number of stores executed if both of the conditions are
3129 // true, and can allow the blocks to become small enough to be if-converted.
3130 // This optimization will also chain, so that ladders of test-and-set
3131 // sequences can be if-converted away.
3132 //
3133 // We only deal with simple diamonds or triangles:
3134 //
3135 // PBI or PBI or a combination of the two
3136 // / \ | \
3137 // PTB PFB | PFB
3138 // \ / | /
3139 // QBI QBI
3140 // / \ | \
3141 // QTB QFB | QFB
3142 // \ / | /
3143 // PostBB PostBB
3144 //
3145 // We model triangles as a type of diamond with a nullptr "true" block.
3146 // Triangles are canonicalized so that the fallthrough edge is represented by
3147 // a true condition, as in the diagram above.
3148 BasicBlock *PTB = PBI->getSuccessor(0);
3149 BasicBlock *PFB = PBI->getSuccessor(1);
3150 BasicBlock *QTB = QBI->getSuccessor(0);
3151 BasicBlock *QFB = QBI->getSuccessor(1);
3152 BasicBlock *PostBB = QFB->getSingleSuccessor();
3153
3154 // Make sure we have a good guess for PostBB. If QTB's only successor is
3155 // QFB, then QFB is a better PostBB.
3156 if (QTB->getSingleSuccessor() == QFB)
3157 PostBB = QFB;
3158
3159 // If we couldn't find a good PostBB, stop.
3160 if (!PostBB)
3161 return false;
3162
3163 bool InvertPCond = false, InvertQCond = false;
3164 // Canonicalize fallthroughs to the true branches.
3165 if (PFB == QBI->getParent()) {
3166 std::swap(PFB, PTB);
3167 InvertPCond = true;
3168 }
3169 if (QFB == PostBB) {
3170 std::swap(QFB, QTB);
3171 InvertQCond = true;
3172 }
3173
3174 // From this point on we can assume PTB or QTB may be fallthroughs but PFB
3175 // and QFB may not. Model fallthroughs as a nullptr block.
3176 if (PTB == QBI->getParent())
3177 PTB = nullptr;
3178 if (QTB == PostBB)
3179 QTB = nullptr;
3180
3181 // Legality bailouts. We must have at least the non-fallthrough blocks and
3182 // the post-dominating block, and the non-fallthroughs must only have one
3183 // predecessor.
3184 auto HasOnePredAndOneSucc = [](BasicBlock *BB, BasicBlock *P, BasicBlock *S) {
3185 return BB->getSinglePredecessor() == P && BB->getSingleSuccessor() == S;
3186 };
3187 if (!HasOnePredAndOneSucc(PFB, PBI->getParent(), QBI->getParent()) ||
3188 !HasOnePredAndOneSucc(QFB, QBI->getParent(), PostBB))
3189 return false;
3190 if ((PTB && !HasOnePredAndOneSucc(PTB, PBI->getParent(), QBI->getParent())) ||
3191 (QTB && !HasOnePredAndOneSucc(QTB, QBI->getParent(), PostBB)))
3192 return false;
3193 if (!QBI->getParent()->hasNUses(2))
3194 return false;
3195
3196 // OK, this is a sequence of two diamonds or triangles.
3197 // Check if there are stores in PTB or PFB that are repeated in QTB or QFB.
3198 SmallPtrSet<Value *, 4> PStoreAddresses, QStoreAddresses;
3199 for (auto *BB : {PTB, PFB}) {
3200 if (!BB)
3201 continue;
3202 for (auto &I : *BB)
3203 if (StoreInst *SI = dyn_cast<StoreInst>(&I))
3204 PStoreAddresses.insert(SI->getPointerOperand());
3205 }
3206 for (auto *BB : {QTB, QFB}) {
3207 if (!BB)
3208 continue;
3209 for (auto &I : *BB)
3210 if (StoreInst *SI = dyn_cast<StoreInst>(&I))
3211 QStoreAddresses.insert(SI->getPointerOperand());
3212 }
3213
3214 set_intersect(PStoreAddresses, QStoreAddresses);
3215 // set_intersect mutates PStoreAddresses in place. Rename it here to make it
3216 // clear what it contains.
3217 auto &CommonAddresses = PStoreAddresses;
3218
3219 bool Changed = false;
3220 for (auto *Address : CommonAddresses)
3221 Changed |= mergeConditionalStoreToAddress(
3222 PTB, PFB, QTB, QFB, PostBB, Address, InvertPCond, InvertQCond, DL, TTI);
3223 return Changed;
3224 }
3225
3226
3227 /// If the previous block ended with a widenable branch, determine if reusing
3228 /// the target block is profitable and legal. This will have the effect of
3229 /// "widening" PBI, but doesn't require us to reason about hosting safety.
tryWidenCondBranchToCondBranch(BranchInst * PBI,BranchInst * BI)3230 static bool tryWidenCondBranchToCondBranch(BranchInst *PBI, BranchInst *BI) {
3231 // TODO: This can be generalized in two important ways:
3232 // 1) We can allow phi nodes in IfFalseBB and simply reuse all the input
3233 // values from the PBI edge.
3234 // 2) We can sink side effecting instructions into BI's fallthrough
3235 // successor provided they doesn't contribute to computation of
3236 // BI's condition.
3237 Value *CondWB, *WC;
3238 BasicBlock *IfTrueBB, *IfFalseBB;
3239 if (!parseWidenableBranch(PBI, CondWB, WC, IfTrueBB, IfFalseBB) ||
3240 IfTrueBB != BI->getParent() || !BI->getParent()->getSinglePredecessor())
3241 return false;
3242 if (!IfFalseBB->phis().empty())
3243 return false; // TODO
3244 // Use lambda to lazily compute expensive condition after cheap ones.
3245 auto NoSideEffects = [](BasicBlock &BB) {
3246 return !llvm::any_of(BB, [](const Instruction &I) {
3247 return I.mayWriteToMemory() || I.mayHaveSideEffects();
3248 });
3249 };
3250 if (BI->getSuccessor(1) != IfFalseBB && // no inf looping
3251 BI->getSuccessor(1)->getTerminatingDeoptimizeCall() && // profitability
3252 NoSideEffects(*BI->getParent())) {
3253 BI->getSuccessor(1)->removePredecessor(BI->getParent());
3254 BI->setSuccessor(1, IfFalseBB);
3255 return true;
3256 }
3257 if (BI->getSuccessor(0) != IfFalseBB && // no inf looping
3258 BI->getSuccessor(0)->getTerminatingDeoptimizeCall() && // profitability
3259 NoSideEffects(*BI->getParent())) {
3260 BI->getSuccessor(0)->removePredecessor(BI->getParent());
3261 BI->setSuccessor(0, IfFalseBB);
3262 return true;
3263 }
3264 return false;
3265 }
3266
3267 /// If we have a conditional branch as a predecessor of another block,
3268 /// this function tries to simplify it. We know
3269 /// that PBI and BI are both conditional branches, and BI is in one of the
3270 /// successor blocks of PBI - PBI branches to BI.
SimplifyCondBranchToCondBranch(BranchInst * PBI,BranchInst * BI,const DataLayout & DL,const TargetTransformInfo & TTI)3271 static bool SimplifyCondBranchToCondBranch(BranchInst *PBI, BranchInst *BI,
3272 const DataLayout &DL,
3273 const TargetTransformInfo &TTI) {
3274 assert(PBI->isConditional() && BI->isConditional());
3275 BasicBlock *BB = BI->getParent();
3276
3277 // If this block ends with a branch instruction, and if there is a
3278 // predecessor that ends on a branch of the same condition, make
3279 // this conditional branch redundant.
3280 if (PBI->getCondition() == BI->getCondition() &&
3281 PBI->getSuccessor(0) != PBI->getSuccessor(1)) {
3282 // Okay, the outcome of this conditional branch is statically
3283 // knowable. If this block had a single pred, handle specially.
3284 if (BB->getSinglePredecessor()) {
3285 // Turn this into a branch on constant.
3286 bool CondIsTrue = PBI->getSuccessor(0) == BB;
3287 BI->setCondition(
3288 ConstantInt::get(Type::getInt1Ty(BB->getContext()), CondIsTrue));
3289 return true; // Nuke the branch on constant.
3290 }
3291
3292 // Otherwise, if there are multiple predecessors, insert a PHI that merges
3293 // in the constant and simplify the block result. Subsequent passes of
3294 // simplifycfg will thread the block.
3295 if (BlockIsSimpleEnoughToThreadThrough(BB)) {
3296 pred_iterator PB = pred_begin(BB), PE = pred_end(BB);
3297 PHINode *NewPN = PHINode::Create(
3298 Type::getInt1Ty(BB->getContext()), std::distance(PB, PE),
3299 BI->getCondition()->getName() + ".pr", &BB->front());
3300 // Okay, we're going to insert the PHI node. Since PBI is not the only
3301 // predecessor, compute the PHI'd conditional value for all of the preds.
3302 // Any predecessor where the condition is not computable we keep symbolic.
3303 for (pred_iterator PI = PB; PI != PE; ++PI) {
3304 BasicBlock *P = *PI;
3305 if ((PBI = dyn_cast<BranchInst>(P->getTerminator())) && PBI != BI &&
3306 PBI->isConditional() && PBI->getCondition() == BI->getCondition() &&
3307 PBI->getSuccessor(0) != PBI->getSuccessor(1)) {
3308 bool CondIsTrue = PBI->getSuccessor(0) == BB;
3309 NewPN->addIncoming(
3310 ConstantInt::get(Type::getInt1Ty(BB->getContext()), CondIsTrue),
3311 P);
3312 } else {
3313 NewPN->addIncoming(BI->getCondition(), P);
3314 }
3315 }
3316
3317 BI->setCondition(NewPN);
3318 return true;
3319 }
3320 }
3321
3322 // If the previous block ended with a widenable branch, determine if reusing
3323 // the target block is profitable and legal. This will have the effect of
3324 // "widening" PBI, but doesn't require us to reason about hosting safety.
3325 if (tryWidenCondBranchToCondBranch(PBI, BI))
3326 return true;
3327
3328 if (auto *CE = dyn_cast<ConstantExpr>(BI->getCondition()))
3329 if (CE->canTrap())
3330 return false;
3331
3332 // If both branches are conditional and both contain stores to the same
3333 // address, remove the stores from the conditionals and create a conditional
3334 // merged store at the end.
3335 if (MergeCondStores && mergeConditionalStores(PBI, BI, DL, TTI))
3336 return true;
3337
3338 // If this is a conditional branch in an empty block, and if any
3339 // predecessors are a conditional branch to one of our destinations,
3340 // fold the conditions into logical ops and one cond br.
3341
3342 // Ignore dbg intrinsics.
3343 if (&*BB->instructionsWithoutDebug().begin() != BI)
3344 return false;
3345
3346 int PBIOp, BIOp;
3347 if (PBI->getSuccessor(0) == BI->getSuccessor(0)) {
3348 PBIOp = 0;
3349 BIOp = 0;
3350 } else if (PBI->getSuccessor(0) == BI->getSuccessor(1)) {
3351 PBIOp = 0;
3352 BIOp = 1;
3353 } else if (PBI->getSuccessor(1) == BI->getSuccessor(0)) {
3354 PBIOp = 1;
3355 BIOp = 0;
3356 } else if (PBI->getSuccessor(1) == BI->getSuccessor(1)) {
3357 PBIOp = 1;
3358 BIOp = 1;
3359 } else {
3360 return false;
3361 }
3362
3363 // Check to make sure that the other destination of this branch
3364 // isn't BB itself. If so, this is an infinite loop that will
3365 // keep getting unwound.
3366 if (PBI->getSuccessor(PBIOp) == BB)
3367 return false;
3368
3369 // Do not perform this transformation if it would require
3370 // insertion of a large number of select instructions. For targets
3371 // without predication/cmovs, this is a big pessimization.
3372
3373 // Also do not perform this transformation if any phi node in the common
3374 // destination block can trap when reached by BB or PBB (PR17073). In that
3375 // case, it would be unsafe to hoist the operation into a select instruction.
3376
3377 BasicBlock *CommonDest = PBI->getSuccessor(PBIOp);
3378 unsigned NumPhis = 0;
3379 for (BasicBlock::iterator II = CommonDest->begin(); isa<PHINode>(II);
3380 ++II, ++NumPhis) {
3381 if (NumPhis > 2) // Disable this xform.
3382 return false;
3383
3384 PHINode *PN = cast<PHINode>(II);
3385 Value *BIV = PN->getIncomingValueForBlock(BB);
3386 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(BIV))
3387 if (CE->canTrap())
3388 return false;
3389
3390 unsigned PBBIdx = PN->getBasicBlockIndex(PBI->getParent());
3391 Value *PBIV = PN->getIncomingValue(PBBIdx);
3392 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(PBIV))
3393 if (CE->canTrap())
3394 return false;
3395 }
3396
3397 // Finally, if everything is ok, fold the branches to logical ops.
3398 BasicBlock *OtherDest = BI->getSuccessor(BIOp ^ 1);
3399
3400 LLVM_DEBUG(dbgs() << "FOLDING BRs:" << *PBI->getParent()
3401 << "AND: " << *BI->getParent());
3402
3403 // If OtherDest *is* BB, then BB is a basic block with a single conditional
3404 // branch in it, where one edge (OtherDest) goes back to itself but the other
3405 // exits. We don't *know* that the program avoids the infinite loop
3406 // (even though that seems likely). If we do this xform naively, we'll end up
3407 // recursively unpeeling the loop. Since we know that (after the xform is
3408 // done) that the block *is* infinite if reached, we just make it an obviously
3409 // infinite loop with no cond branch.
3410 if (OtherDest == BB) {
3411 // Insert it at the end of the function, because it's either code,
3412 // or it won't matter if it's hot. :)
3413 BasicBlock *InfLoopBlock =
3414 BasicBlock::Create(BB->getContext(), "infloop", BB->getParent());
3415 BranchInst::Create(InfLoopBlock, InfLoopBlock);
3416 OtherDest = InfLoopBlock;
3417 }
3418
3419 LLVM_DEBUG(dbgs() << *PBI->getParent()->getParent());
3420
3421 // BI may have other predecessors. Because of this, we leave
3422 // it alone, but modify PBI.
3423
3424 // Make sure we get to CommonDest on True&True directions.
3425 Value *PBICond = PBI->getCondition();
3426 IRBuilder<NoFolder> Builder(PBI);
3427 if (PBIOp)
3428 PBICond = Builder.CreateNot(PBICond, PBICond->getName() + ".not");
3429
3430 Value *BICond = BI->getCondition();
3431 if (BIOp)
3432 BICond = Builder.CreateNot(BICond, BICond->getName() + ".not");
3433
3434 // Merge the conditions.
3435 Value *Cond = Builder.CreateOr(PBICond, BICond, "brmerge");
3436
3437 // Modify PBI to branch on the new condition to the new dests.
3438 PBI->setCondition(Cond);
3439 PBI->setSuccessor(0, CommonDest);
3440 PBI->setSuccessor(1, OtherDest);
3441
3442 // Update branch weight for PBI.
3443 uint64_t PredTrueWeight, PredFalseWeight, SuccTrueWeight, SuccFalseWeight;
3444 uint64_t PredCommon, PredOther, SuccCommon, SuccOther;
3445 bool HasWeights =
3446 extractPredSuccWeights(PBI, BI, PredTrueWeight, PredFalseWeight,
3447 SuccTrueWeight, SuccFalseWeight);
3448 if (HasWeights) {
3449 PredCommon = PBIOp ? PredFalseWeight : PredTrueWeight;
3450 PredOther = PBIOp ? PredTrueWeight : PredFalseWeight;
3451 SuccCommon = BIOp ? SuccFalseWeight : SuccTrueWeight;
3452 SuccOther = BIOp ? SuccTrueWeight : SuccFalseWeight;
3453 // The weight to CommonDest should be PredCommon * SuccTotal +
3454 // PredOther * SuccCommon.
3455 // The weight to OtherDest should be PredOther * SuccOther.
3456 uint64_t NewWeights[2] = {PredCommon * (SuccCommon + SuccOther) +
3457 PredOther * SuccCommon,
3458 PredOther * SuccOther};
3459 // Halve the weights if any of them cannot fit in an uint32_t
3460 FitWeights(NewWeights);
3461
3462 setBranchWeights(PBI, NewWeights[0], NewWeights[1]);
3463 }
3464
3465 // OtherDest may have phi nodes. If so, add an entry from PBI's
3466 // block that are identical to the entries for BI's block.
3467 AddPredecessorToBlock(OtherDest, PBI->getParent(), BB);
3468
3469 // We know that the CommonDest already had an edge from PBI to
3470 // it. If it has PHIs though, the PHIs may have different
3471 // entries for BB and PBI's BB. If so, insert a select to make
3472 // them agree.
3473 for (PHINode &PN : CommonDest->phis()) {
3474 Value *BIV = PN.getIncomingValueForBlock(BB);
3475 unsigned PBBIdx = PN.getBasicBlockIndex(PBI->getParent());
3476 Value *PBIV = PN.getIncomingValue(PBBIdx);
3477 if (BIV != PBIV) {
3478 // Insert a select in PBI to pick the right value.
3479 SelectInst *NV = cast<SelectInst>(
3480 Builder.CreateSelect(PBICond, PBIV, BIV, PBIV->getName() + ".mux"));
3481 PN.setIncomingValue(PBBIdx, NV);
3482 // Although the select has the same condition as PBI, the original branch
3483 // weights for PBI do not apply to the new select because the select's
3484 // 'logical' edges are incoming edges of the phi that is eliminated, not
3485 // the outgoing edges of PBI.
3486 if (HasWeights) {
3487 uint64_t PredCommon = PBIOp ? PredFalseWeight : PredTrueWeight;
3488 uint64_t PredOther = PBIOp ? PredTrueWeight : PredFalseWeight;
3489 uint64_t SuccCommon = BIOp ? SuccFalseWeight : SuccTrueWeight;
3490 uint64_t SuccOther = BIOp ? SuccTrueWeight : SuccFalseWeight;
3491 // The weight to PredCommonDest should be PredCommon * SuccTotal.
3492 // The weight to PredOtherDest should be PredOther * SuccCommon.
3493 uint64_t NewWeights[2] = {PredCommon * (SuccCommon + SuccOther),
3494 PredOther * SuccCommon};
3495
3496 FitWeights(NewWeights);
3497
3498 setBranchWeights(NV, NewWeights[0], NewWeights[1]);
3499 }
3500 }
3501 }
3502
3503 LLVM_DEBUG(dbgs() << "INTO: " << *PBI->getParent());
3504 LLVM_DEBUG(dbgs() << *PBI->getParent()->getParent());
3505
3506 // This basic block is probably dead. We know it has at least
3507 // one fewer predecessor.
3508 return true;
3509 }
3510
3511 // Simplifies a terminator by replacing it with a branch to TrueBB if Cond is
3512 // true or to FalseBB if Cond is false.
3513 // Takes care of updating the successors and removing the old terminator.
3514 // Also makes sure not to introduce new successors by assuming that edges to
3515 // non-successor TrueBBs and FalseBBs aren't reachable.
SimplifyTerminatorOnSelect(Instruction * OldTerm,Value * Cond,BasicBlock * TrueBB,BasicBlock * FalseBB,uint32_t TrueWeight,uint32_t FalseWeight)3516 static bool SimplifyTerminatorOnSelect(Instruction *OldTerm, Value *Cond,
3517 BasicBlock *TrueBB, BasicBlock *FalseBB,
3518 uint32_t TrueWeight,
3519 uint32_t FalseWeight) {
3520 // Remove any superfluous successor edges from the CFG.
3521 // First, figure out which successors to preserve.
3522 // If TrueBB and FalseBB are equal, only try to preserve one copy of that
3523 // successor.
3524 BasicBlock *KeepEdge1 = TrueBB;
3525 BasicBlock *KeepEdge2 = TrueBB != FalseBB ? FalseBB : nullptr;
3526
3527 // Then remove the rest.
3528 for (BasicBlock *Succ : successors(OldTerm)) {
3529 // Make sure only to keep exactly one copy of each edge.
3530 if (Succ == KeepEdge1)
3531 KeepEdge1 = nullptr;
3532 else if (Succ == KeepEdge2)
3533 KeepEdge2 = nullptr;
3534 else
3535 Succ->removePredecessor(OldTerm->getParent(),
3536 /*KeepOneInputPHIs=*/true);
3537 }
3538
3539 IRBuilder<> Builder(OldTerm);
3540 Builder.SetCurrentDebugLocation(OldTerm->getDebugLoc());
3541
3542 // Insert an appropriate new terminator.
3543 if (!KeepEdge1 && !KeepEdge2) {
3544 if (TrueBB == FalseBB)
3545 // We were only looking for one successor, and it was present.
3546 // Create an unconditional branch to it.
3547 Builder.CreateBr(TrueBB);
3548 else {
3549 // We found both of the successors we were looking for.
3550 // Create a conditional branch sharing the condition of the select.
3551 BranchInst *NewBI = Builder.CreateCondBr(Cond, TrueBB, FalseBB);
3552 if (TrueWeight != FalseWeight)
3553 setBranchWeights(NewBI, TrueWeight, FalseWeight);
3554 }
3555 } else if (KeepEdge1 && (KeepEdge2 || TrueBB == FalseBB)) {
3556 // Neither of the selected blocks were successors, so this
3557 // terminator must be unreachable.
3558 new UnreachableInst(OldTerm->getContext(), OldTerm);
3559 } else {
3560 // One of the selected values was a successor, but the other wasn't.
3561 // Insert an unconditional branch to the one that was found;
3562 // the edge to the one that wasn't must be unreachable.
3563 if (!KeepEdge1)
3564 // Only TrueBB was found.
3565 Builder.CreateBr(TrueBB);
3566 else
3567 // Only FalseBB was found.
3568 Builder.CreateBr(FalseBB);
3569 }
3570
3571 EraseTerminatorAndDCECond(OldTerm);
3572 return true;
3573 }
3574
3575 // Replaces
3576 // (switch (select cond, X, Y)) on constant X, Y
3577 // with a branch - conditional if X and Y lead to distinct BBs,
3578 // unconditional otherwise.
SimplifySwitchOnSelect(SwitchInst * SI,SelectInst * Select)3579 static bool SimplifySwitchOnSelect(SwitchInst *SI, SelectInst *Select) {
3580 // Check for constant integer values in the select.
3581 ConstantInt *TrueVal = dyn_cast<ConstantInt>(Select->getTrueValue());
3582 ConstantInt *FalseVal = dyn_cast<ConstantInt>(Select->getFalseValue());
3583 if (!TrueVal || !FalseVal)
3584 return false;
3585
3586 // Find the relevant condition and destinations.
3587 Value *Condition = Select->getCondition();
3588 BasicBlock *TrueBB = SI->findCaseValue(TrueVal)->getCaseSuccessor();
3589 BasicBlock *FalseBB = SI->findCaseValue(FalseVal)->getCaseSuccessor();
3590
3591 // Get weight for TrueBB and FalseBB.
3592 uint32_t TrueWeight = 0, FalseWeight = 0;
3593 SmallVector<uint64_t, 8> Weights;
3594 bool HasWeights = HasBranchWeights(SI);
3595 if (HasWeights) {
3596 GetBranchWeights(SI, Weights);
3597 if (Weights.size() == 1 + SI->getNumCases()) {
3598 TrueWeight =
3599 (uint32_t)Weights[SI->findCaseValue(TrueVal)->getSuccessorIndex()];
3600 FalseWeight =
3601 (uint32_t)Weights[SI->findCaseValue(FalseVal)->getSuccessorIndex()];
3602 }
3603 }
3604
3605 // Perform the actual simplification.
3606 return SimplifyTerminatorOnSelect(SI, Condition, TrueBB, FalseBB, TrueWeight,
3607 FalseWeight);
3608 }
3609
3610 // Replaces
3611 // (indirectbr (select cond, blockaddress(@fn, BlockA),
3612 // blockaddress(@fn, BlockB)))
3613 // with
3614 // (br cond, BlockA, BlockB).
SimplifyIndirectBrOnSelect(IndirectBrInst * IBI,SelectInst * SI)3615 static bool SimplifyIndirectBrOnSelect(IndirectBrInst *IBI, SelectInst *SI) {
3616 // Check that both operands of the select are block addresses.
3617 BlockAddress *TBA = dyn_cast<BlockAddress>(SI->getTrueValue());
3618 BlockAddress *FBA = dyn_cast<BlockAddress>(SI->getFalseValue());
3619 if (!TBA || !FBA)
3620 return false;
3621
3622 // Extract the actual blocks.
3623 BasicBlock *TrueBB = TBA->getBasicBlock();
3624 BasicBlock *FalseBB = FBA->getBasicBlock();
3625
3626 // Perform the actual simplification.
3627 return SimplifyTerminatorOnSelect(IBI, SI->getCondition(), TrueBB, FalseBB, 0,
3628 0);
3629 }
3630
3631 /// This is called when we find an icmp instruction
3632 /// (a seteq/setne with a constant) as the only instruction in a
3633 /// block that ends with an uncond branch. We are looking for a very specific
3634 /// pattern that occurs when "A == 1 || A == 2 || A == 3" gets simplified. In
3635 /// this case, we merge the first two "or's of icmp" into a switch, but then the
3636 /// default value goes to an uncond block with a seteq in it, we get something
3637 /// like:
3638 ///
3639 /// switch i8 %A, label %DEFAULT [ i8 1, label %end i8 2, label %end ]
3640 /// DEFAULT:
3641 /// %tmp = icmp eq i8 %A, 92
3642 /// br label %end
3643 /// end:
3644 /// ... = phi i1 [ true, %entry ], [ %tmp, %DEFAULT ], [ true, %entry ]
3645 ///
3646 /// We prefer to split the edge to 'end' so that there is a true/false entry to
3647 /// the PHI, merging the third icmp into the switch.
tryToSimplifyUncondBranchWithICmpInIt(ICmpInst * ICI,IRBuilder<> & Builder)3648 bool SimplifyCFGOpt::tryToSimplifyUncondBranchWithICmpInIt(
3649 ICmpInst *ICI, IRBuilder<> &Builder) {
3650 BasicBlock *BB = ICI->getParent();
3651
3652 // If the block has any PHIs in it or the icmp has multiple uses, it is too
3653 // complex.
3654 if (isa<PHINode>(BB->begin()) || !ICI->hasOneUse())
3655 return false;
3656
3657 Value *V = ICI->getOperand(0);
3658 ConstantInt *Cst = cast<ConstantInt>(ICI->getOperand(1));
3659
3660 // The pattern we're looking for is where our only predecessor is a switch on
3661 // 'V' and this block is the default case for the switch. In this case we can
3662 // fold the compared value into the switch to simplify things.
3663 BasicBlock *Pred = BB->getSinglePredecessor();
3664 if (!Pred || !isa<SwitchInst>(Pred->getTerminator()))
3665 return false;
3666
3667 SwitchInst *SI = cast<SwitchInst>(Pred->getTerminator());
3668 if (SI->getCondition() != V)
3669 return false;
3670
3671 // If BB is reachable on a non-default case, then we simply know the value of
3672 // V in this block. Substitute it and constant fold the icmp instruction
3673 // away.
3674 if (SI->getDefaultDest() != BB) {
3675 ConstantInt *VVal = SI->findCaseDest(BB);
3676 assert(VVal && "Should have a unique destination value");
3677 ICI->setOperand(0, VVal);
3678
3679 if (Value *V = SimplifyInstruction(ICI, {DL, ICI})) {
3680 ICI->replaceAllUsesWith(V);
3681 ICI->eraseFromParent();
3682 }
3683 // BB is now empty, so it is likely to simplify away.
3684 return requestResimplify();
3685 }
3686
3687 // Ok, the block is reachable from the default dest. If the constant we're
3688 // comparing exists in one of the other edges, then we can constant fold ICI
3689 // and zap it.
3690 if (SI->findCaseValue(Cst) != SI->case_default()) {
3691 Value *V;
3692 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
3693 V = ConstantInt::getFalse(BB->getContext());
3694 else
3695 V = ConstantInt::getTrue(BB->getContext());
3696
3697 ICI->replaceAllUsesWith(V);
3698 ICI->eraseFromParent();
3699 // BB is now empty, so it is likely to simplify away.
3700 return requestResimplify();
3701 }
3702
3703 // The use of the icmp has to be in the 'end' block, by the only PHI node in
3704 // the block.
3705 BasicBlock *SuccBlock = BB->getTerminator()->getSuccessor(0);
3706 PHINode *PHIUse = dyn_cast<PHINode>(ICI->user_back());
3707 if (PHIUse == nullptr || PHIUse != &SuccBlock->front() ||
3708 isa<PHINode>(++BasicBlock::iterator(PHIUse)))
3709 return false;
3710
3711 // If the icmp is a SETEQ, then the default dest gets false, the new edge gets
3712 // true in the PHI.
3713 Constant *DefaultCst = ConstantInt::getTrue(BB->getContext());
3714 Constant *NewCst = ConstantInt::getFalse(BB->getContext());
3715
3716 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
3717 std::swap(DefaultCst, NewCst);
3718
3719 // Replace ICI (which is used by the PHI for the default value) with true or
3720 // false depending on if it is EQ or NE.
3721 ICI->replaceAllUsesWith(DefaultCst);
3722 ICI->eraseFromParent();
3723
3724 // Okay, the switch goes to this block on a default value. Add an edge from
3725 // the switch to the merge point on the compared value.
3726 BasicBlock *NewBB =
3727 BasicBlock::Create(BB->getContext(), "switch.edge", BB->getParent(), BB);
3728 {
3729 SwitchInstProfUpdateWrapper SIW(*SI);
3730 auto W0 = SIW.getSuccessorWeight(0);
3731 SwitchInstProfUpdateWrapper::CaseWeightOpt NewW;
3732 if (W0) {
3733 NewW = ((uint64_t(*W0) + 1) >> 1);
3734 SIW.setSuccessorWeight(0, *NewW);
3735 }
3736 SIW.addCase(Cst, NewBB, NewW);
3737 }
3738
3739 // NewBB branches to the phi block, add the uncond branch and the phi entry.
3740 Builder.SetInsertPoint(NewBB);
3741 Builder.SetCurrentDebugLocation(SI->getDebugLoc());
3742 Builder.CreateBr(SuccBlock);
3743 PHIUse->addIncoming(NewCst, NewBB);
3744 return true;
3745 }
3746
3747 /// The specified branch is a conditional branch.
3748 /// Check to see if it is branching on an or/and chain of icmp instructions, and
3749 /// fold it into a switch instruction if so.
SimplifyBranchOnICmpChain(BranchInst * BI,IRBuilder<> & Builder,const DataLayout & DL)3750 static bool SimplifyBranchOnICmpChain(BranchInst *BI, IRBuilder<> &Builder,
3751 const DataLayout &DL) {
3752 Instruction *Cond = dyn_cast<Instruction>(BI->getCondition());
3753 if (!Cond)
3754 return false;
3755
3756 // Change br (X == 0 | X == 1), T, F into a switch instruction.
3757 // If this is a bunch of seteq's or'd together, or if it's a bunch of
3758 // 'setne's and'ed together, collect them.
3759
3760 // Try to gather values from a chain of and/or to be turned into a switch
3761 ConstantComparesGatherer ConstantCompare(Cond, DL);
3762 // Unpack the result
3763 SmallVectorImpl<ConstantInt *> &Values = ConstantCompare.Vals;
3764 Value *CompVal = ConstantCompare.CompValue;
3765 unsigned UsedICmps = ConstantCompare.UsedICmps;
3766 Value *ExtraCase = ConstantCompare.Extra;
3767
3768 // If we didn't have a multiply compared value, fail.
3769 if (!CompVal)
3770 return false;
3771
3772 // Avoid turning single icmps into a switch.
3773 if (UsedICmps <= 1)
3774 return false;
3775
3776 bool TrueWhenEqual = (Cond->getOpcode() == Instruction::Or);
3777
3778 // There might be duplicate constants in the list, which the switch
3779 // instruction can't handle, remove them now.
3780 array_pod_sort(Values.begin(), Values.end(), ConstantIntSortPredicate);
3781 Values.erase(std::unique(Values.begin(), Values.end()), Values.end());
3782
3783 // If Extra was used, we require at least two switch values to do the
3784 // transformation. A switch with one value is just a conditional branch.
3785 if (ExtraCase && Values.size() < 2)
3786 return false;
3787
3788 // TODO: Preserve branch weight metadata, similarly to how
3789 // FoldValueComparisonIntoPredecessors preserves it.
3790
3791 // Figure out which block is which destination.
3792 BasicBlock *DefaultBB = BI->getSuccessor(1);
3793 BasicBlock *EdgeBB = BI->getSuccessor(0);
3794 if (!TrueWhenEqual)
3795 std::swap(DefaultBB, EdgeBB);
3796
3797 BasicBlock *BB = BI->getParent();
3798
3799 // MSAN does not like undefs as branch condition which can be introduced
3800 // with "explicit branch".
3801 if (ExtraCase && BB->getParent()->hasFnAttribute(Attribute::SanitizeMemory))
3802 return false;
3803
3804 LLVM_DEBUG(dbgs() << "Converting 'icmp' chain with " << Values.size()
3805 << " cases into SWITCH. BB is:\n"
3806 << *BB);
3807
3808 // If there are any extra values that couldn't be folded into the switch
3809 // then we evaluate them with an explicit branch first. Split the block
3810 // right before the condbr to handle it.
3811 if (ExtraCase) {
3812 BasicBlock *NewBB =
3813 BB->splitBasicBlock(BI->getIterator(), "switch.early.test");
3814 // Remove the uncond branch added to the old block.
3815 Instruction *OldTI = BB->getTerminator();
3816 Builder.SetInsertPoint(OldTI);
3817
3818 if (TrueWhenEqual)
3819 Builder.CreateCondBr(ExtraCase, EdgeBB, NewBB);
3820 else
3821 Builder.CreateCondBr(ExtraCase, NewBB, EdgeBB);
3822
3823 OldTI->eraseFromParent();
3824
3825 // If there are PHI nodes in EdgeBB, then we need to add a new entry to them
3826 // for the edge we just added.
3827 AddPredecessorToBlock(EdgeBB, BB, NewBB);
3828
3829 LLVM_DEBUG(dbgs() << " ** 'icmp' chain unhandled condition: " << *ExtraCase
3830 << "\nEXTRABB = " << *BB);
3831 BB = NewBB;
3832 }
3833
3834 Builder.SetInsertPoint(BI);
3835 // Convert pointer to int before we switch.
3836 if (CompVal->getType()->isPointerTy()) {
3837 CompVal = Builder.CreatePtrToInt(
3838 CompVal, DL.getIntPtrType(CompVal->getType()), "magicptr");
3839 }
3840
3841 // Create the new switch instruction now.
3842 SwitchInst *New = Builder.CreateSwitch(CompVal, DefaultBB, Values.size());
3843
3844 // Add all of the 'cases' to the switch instruction.
3845 for (unsigned i = 0, e = Values.size(); i != e; ++i)
3846 New->addCase(Values[i], EdgeBB);
3847
3848 // We added edges from PI to the EdgeBB. As such, if there were any
3849 // PHI nodes in EdgeBB, they need entries to be added corresponding to
3850 // the number of edges added.
3851 for (BasicBlock::iterator BBI = EdgeBB->begin(); isa<PHINode>(BBI); ++BBI) {
3852 PHINode *PN = cast<PHINode>(BBI);
3853 Value *InVal = PN->getIncomingValueForBlock(BB);
3854 for (unsigned i = 0, e = Values.size() - 1; i != e; ++i)
3855 PN->addIncoming(InVal, BB);
3856 }
3857
3858 // Erase the old branch instruction.
3859 EraseTerminatorAndDCECond(BI);
3860
3861 LLVM_DEBUG(dbgs() << " ** 'icmp' chain result is:\n" << *BB << '\n');
3862 return true;
3863 }
3864
SimplifyResume(ResumeInst * RI,IRBuilder<> & Builder)3865 bool SimplifyCFGOpt::SimplifyResume(ResumeInst *RI, IRBuilder<> &Builder) {
3866 if (isa<PHINode>(RI->getValue()))
3867 return SimplifyCommonResume(RI);
3868 else if (isa<LandingPadInst>(RI->getParent()->getFirstNonPHI()) &&
3869 RI->getValue() == RI->getParent()->getFirstNonPHI())
3870 // The resume must unwind the exception that caused control to branch here.
3871 return SimplifySingleResume(RI);
3872
3873 return false;
3874 }
3875
3876 // Simplify resume that is shared by several landing pads (phi of landing pad).
SimplifyCommonResume(ResumeInst * RI)3877 bool SimplifyCFGOpt::SimplifyCommonResume(ResumeInst *RI) {
3878 BasicBlock *BB = RI->getParent();
3879
3880 // Check that there are no other instructions except for debug intrinsics
3881 // between the phi of landing pads (RI->getValue()) and resume instruction.
3882 BasicBlock::iterator I = cast<Instruction>(RI->getValue())->getIterator(),
3883 E = RI->getIterator();
3884 while (++I != E)
3885 if (!isa<DbgInfoIntrinsic>(I))
3886 return false;
3887
3888 SmallSetVector<BasicBlock *, 4> TrivialUnwindBlocks;
3889 auto *PhiLPInst = cast<PHINode>(RI->getValue());
3890
3891 // Check incoming blocks to see if any of them are trivial.
3892 for (unsigned Idx = 0, End = PhiLPInst->getNumIncomingValues(); Idx != End;
3893 Idx++) {
3894 auto *IncomingBB = PhiLPInst->getIncomingBlock(Idx);
3895 auto *IncomingValue = PhiLPInst->getIncomingValue(Idx);
3896
3897 // If the block has other successors, we can not delete it because
3898 // it has other dependents.
3899 if (IncomingBB->getUniqueSuccessor() != BB)
3900 continue;
3901
3902 auto *LandingPad = dyn_cast<LandingPadInst>(IncomingBB->getFirstNonPHI());
3903 // Not the landing pad that caused the control to branch here.
3904 if (IncomingValue != LandingPad)
3905 continue;
3906
3907 bool isTrivial = true;
3908
3909 I = IncomingBB->getFirstNonPHI()->getIterator();
3910 E = IncomingBB->getTerminator()->getIterator();
3911 while (++I != E)
3912 if (!isa<DbgInfoIntrinsic>(I)) {
3913 isTrivial = false;
3914 break;
3915 }
3916
3917 if (isTrivial)
3918 TrivialUnwindBlocks.insert(IncomingBB);
3919 }
3920
3921 // If no trivial unwind blocks, don't do any simplifications.
3922 if (TrivialUnwindBlocks.empty())
3923 return false;
3924
3925 // Turn all invokes that unwind here into calls.
3926 for (auto *TrivialBB : TrivialUnwindBlocks) {
3927 // Blocks that will be simplified should be removed from the phi node.
3928 // Note there could be multiple edges to the resume block, and we need
3929 // to remove them all.
3930 while (PhiLPInst->getBasicBlockIndex(TrivialBB) != -1)
3931 BB->removePredecessor(TrivialBB, true);
3932
3933 for (pred_iterator PI = pred_begin(TrivialBB), PE = pred_end(TrivialBB);
3934 PI != PE;) {
3935 BasicBlock *Pred = *PI++;
3936 removeUnwindEdge(Pred);
3937 }
3938
3939 // In each SimplifyCFG run, only the current processed block can be erased.
3940 // Otherwise, it will break the iteration of SimplifyCFG pass. So instead
3941 // of erasing TrivialBB, we only remove the branch to the common resume
3942 // block so that we can later erase the resume block since it has no
3943 // predecessors.
3944 TrivialBB->getTerminator()->eraseFromParent();
3945 new UnreachableInst(RI->getContext(), TrivialBB);
3946 }
3947
3948 // Delete the resume block if all its predecessors have been removed.
3949 if (pred_empty(BB))
3950 BB->eraseFromParent();
3951
3952 return !TrivialUnwindBlocks.empty();
3953 }
3954
3955 // Simplify resume that is only used by a single (non-phi) landing pad.
SimplifySingleResume(ResumeInst * RI)3956 bool SimplifyCFGOpt::SimplifySingleResume(ResumeInst *RI) {
3957 BasicBlock *BB = RI->getParent();
3958 auto *LPInst = cast<LandingPadInst>(BB->getFirstNonPHI());
3959 assert(RI->getValue() == LPInst &&
3960 "Resume must unwind the exception that caused control to here");
3961
3962 // Check that there are no other instructions except for debug intrinsics.
3963 BasicBlock::iterator I = LPInst->getIterator(), E = RI->getIterator();
3964 while (++I != E)
3965 if (!isa<DbgInfoIntrinsic>(I))
3966 return false;
3967
3968 // Turn all invokes that unwind here into calls and delete the basic block.
3969 for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE;) {
3970 BasicBlock *Pred = *PI++;
3971 removeUnwindEdge(Pred);
3972 }
3973
3974 // The landingpad is now unreachable. Zap it.
3975 if (LoopHeaders)
3976 LoopHeaders->erase(BB);
3977 BB->eraseFromParent();
3978 return true;
3979 }
3980
removeEmptyCleanup(CleanupReturnInst * RI)3981 static bool removeEmptyCleanup(CleanupReturnInst *RI) {
3982 // If this is a trivial cleanup pad that executes no instructions, it can be
3983 // eliminated. If the cleanup pad continues to the caller, any predecessor
3984 // that is an EH pad will be updated to continue to the caller and any
3985 // predecessor that terminates with an invoke instruction will have its invoke
3986 // instruction converted to a call instruction. If the cleanup pad being
3987 // simplified does not continue to the caller, each predecessor will be
3988 // updated to continue to the unwind destination of the cleanup pad being
3989 // simplified.
3990 BasicBlock *BB = RI->getParent();
3991 CleanupPadInst *CPInst = RI->getCleanupPad();
3992 if (CPInst->getParent() != BB)
3993 // This isn't an empty cleanup.
3994 return false;
3995
3996 // We cannot kill the pad if it has multiple uses. This typically arises
3997 // from unreachable basic blocks.
3998 if (!CPInst->hasOneUse())
3999 return false;
4000
4001 // Check that there are no other instructions except for benign intrinsics.
4002 BasicBlock::iterator I = CPInst->getIterator(), E = RI->getIterator();
4003 while (++I != E) {
4004 auto *II = dyn_cast<IntrinsicInst>(I);
4005 if (!II)
4006 return false;
4007
4008 Intrinsic::ID IntrinsicID = II->getIntrinsicID();
4009 switch (IntrinsicID) {
4010 case Intrinsic::dbg_declare:
4011 case Intrinsic::dbg_value:
4012 case Intrinsic::dbg_label:
4013 case Intrinsic::lifetime_end:
4014 break;
4015 default:
4016 return false;
4017 }
4018 }
4019
4020 // If the cleanup return we are simplifying unwinds to the caller, this will
4021 // set UnwindDest to nullptr.
4022 BasicBlock *UnwindDest = RI->getUnwindDest();
4023 Instruction *DestEHPad = UnwindDest ? UnwindDest->getFirstNonPHI() : nullptr;
4024
4025 // We're about to remove BB from the control flow. Before we do, sink any
4026 // PHINodes into the unwind destination. Doing this before changing the
4027 // control flow avoids some potentially slow checks, since we can currently
4028 // be certain that UnwindDest and BB have no common predecessors (since they
4029 // are both EH pads).
4030 if (UnwindDest) {
4031 // First, go through the PHI nodes in UnwindDest and update any nodes that
4032 // reference the block we are removing
4033 for (BasicBlock::iterator I = UnwindDest->begin(),
4034 IE = DestEHPad->getIterator();
4035 I != IE; ++I) {
4036 PHINode *DestPN = cast<PHINode>(I);
4037
4038 int Idx = DestPN->getBasicBlockIndex(BB);
4039 // Since BB unwinds to UnwindDest, it has to be in the PHI node.
4040 assert(Idx != -1);
4041 // This PHI node has an incoming value that corresponds to a control
4042 // path through the cleanup pad we are removing. If the incoming
4043 // value is in the cleanup pad, it must be a PHINode (because we
4044 // verified above that the block is otherwise empty). Otherwise, the
4045 // value is either a constant or a value that dominates the cleanup
4046 // pad being removed.
4047 //
4048 // Because BB and UnwindDest are both EH pads, all of their
4049 // predecessors must unwind to these blocks, and since no instruction
4050 // can have multiple unwind destinations, there will be no overlap in
4051 // incoming blocks between SrcPN and DestPN.
4052 Value *SrcVal = DestPN->getIncomingValue(Idx);
4053 PHINode *SrcPN = dyn_cast<PHINode>(SrcVal);
4054
4055 // Remove the entry for the block we are deleting.
4056 DestPN->removeIncomingValue(Idx, false);
4057
4058 if (SrcPN && SrcPN->getParent() == BB) {
4059 // If the incoming value was a PHI node in the cleanup pad we are
4060 // removing, we need to merge that PHI node's incoming values into
4061 // DestPN.
4062 for (unsigned SrcIdx = 0, SrcE = SrcPN->getNumIncomingValues();
4063 SrcIdx != SrcE; ++SrcIdx) {
4064 DestPN->addIncoming(SrcPN->getIncomingValue(SrcIdx),
4065 SrcPN->getIncomingBlock(SrcIdx));
4066 }
4067 } else {
4068 // Otherwise, the incoming value came from above BB and
4069 // so we can just reuse it. We must associate all of BB's
4070 // predecessors with this value.
4071 for (auto *pred : predecessors(BB)) {
4072 DestPN->addIncoming(SrcVal, pred);
4073 }
4074 }
4075 }
4076
4077 // Sink any remaining PHI nodes directly into UnwindDest.
4078 Instruction *InsertPt = DestEHPad;
4079 for (BasicBlock::iterator I = BB->begin(),
4080 IE = BB->getFirstNonPHI()->getIterator();
4081 I != IE;) {
4082 // The iterator must be incremented here because the instructions are
4083 // being moved to another block.
4084 PHINode *PN = cast<PHINode>(I++);
4085 if (PN->use_empty())
4086 // If the PHI node has no uses, just leave it. It will be erased
4087 // when we erase BB below.
4088 continue;
4089
4090 // Otherwise, sink this PHI node into UnwindDest.
4091 // Any predecessors to UnwindDest which are not already represented
4092 // must be back edges which inherit the value from the path through
4093 // BB. In this case, the PHI value must reference itself.
4094 for (auto *pred : predecessors(UnwindDest))
4095 if (pred != BB)
4096 PN->addIncoming(PN, pred);
4097 PN->moveBefore(InsertPt);
4098 }
4099 }
4100
4101 for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE;) {
4102 // The iterator must be updated here because we are removing this pred.
4103 BasicBlock *PredBB = *PI++;
4104 if (UnwindDest == nullptr) {
4105 removeUnwindEdge(PredBB);
4106 } else {
4107 Instruction *TI = PredBB->getTerminator();
4108 TI->replaceUsesOfWith(BB, UnwindDest);
4109 }
4110 }
4111
4112 // The cleanup pad is now unreachable. Zap it.
4113 BB->eraseFromParent();
4114 return true;
4115 }
4116
4117 // Try to merge two cleanuppads together.
mergeCleanupPad(CleanupReturnInst * RI)4118 static bool mergeCleanupPad(CleanupReturnInst *RI) {
4119 // Skip any cleanuprets which unwind to caller, there is nothing to merge
4120 // with.
4121 BasicBlock *UnwindDest = RI->getUnwindDest();
4122 if (!UnwindDest)
4123 return false;
4124
4125 // This cleanupret isn't the only predecessor of this cleanuppad, it wouldn't
4126 // be safe to merge without code duplication.
4127 if (UnwindDest->getSinglePredecessor() != RI->getParent())
4128 return false;
4129
4130 // Verify that our cleanuppad's unwind destination is another cleanuppad.
4131 auto *SuccessorCleanupPad = dyn_cast<CleanupPadInst>(&UnwindDest->front());
4132 if (!SuccessorCleanupPad)
4133 return false;
4134
4135 CleanupPadInst *PredecessorCleanupPad = RI->getCleanupPad();
4136 // Replace any uses of the successor cleanupad with the predecessor pad
4137 // The only cleanuppad uses should be this cleanupret, it's cleanupret and
4138 // funclet bundle operands.
4139 SuccessorCleanupPad->replaceAllUsesWith(PredecessorCleanupPad);
4140 // Remove the old cleanuppad.
4141 SuccessorCleanupPad->eraseFromParent();
4142 // Now, we simply replace the cleanupret with a branch to the unwind
4143 // destination.
4144 BranchInst::Create(UnwindDest, RI->getParent());
4145 RI->eraseFromParent();
4146
4147 return true;
4148 }
4149
SimplifyCleanupReturn(CleanupReturnInst * RI)4150 bool SimplifyCFGOpt::SimplifyCleanupReturn(CleanupReturnInst *RI) {
4151 // It is possible to transiantly have an undef cleanuppad operand because we
4152 // have deleted some, but not all, dead blocks.
4153 // Eventually, this block will be deleted.
4154 if (isa<UndefValue>(RI->getOperand(0)))
4155 return false;
4156
4157 if (mergeCleanupPad(RI))
4158 return true;
4159
4160 if (removeEmptyCleanup(RI))
4161 return true;
4162
4163 return false;
4164 }
4165
SimplifyReturn(ReturnInst * RI,IRBuilder<> & Builder)4166 bool SimplifyCFGOpt::SimplifyReturn(ReturnInst *RI, IRBuilder<> &Builder) {
4167 BasicBlock *BB = RI->getParent();
4168 if (!BB->getFirstNonPHIOrDbg()->isTerminator())
4169 return false;
4170
4171 // Find predecessors that end with branches.
4172 SmallVector<BasicBlock *, 8> UncondBranchPreds;
4173 SmallVector<BranchInst *, 8> CondBranchPreds;
4174 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
4175 BasicBlock *P = *PI;
4176 Instruction *PTI = P->getTerminator();
4177 if (BranchInst *BI = dyn_cast<BranchInst>(PTI)) {
4178 if (BI->isUnconditional())
4179 UncondBranchPreds.push_back(P);
4180 else
4181 CondBranchPreds.push_back(BI);
4182 }
4183 }
4184
4185 // If we found some, do the transformation!
4186 if (!UncondBranchPreds.empty() && DupRet) {
4187 while (!UncondBranchPreds.empty()) {
4188 BasicBlock *Pred = UncondBranchPreds.pop_back_val();
4189 LLVM_DEBUG(dbgs() << "FOLDING: " << *BB
4190 << "INTO UNCOND BRANCH PRED: " << *Pred);
4191 (void)FoldReturnIntoUncondBranch(RI, BB, Pred);
4192 }
4193
4194 // If we eliminated all predecessors of the block, delete the block now.
4195 if (pred_empty(BB)) {
4196 // We know there are no successors, so just nuke the block.
4197 if (LoopHeaders)
4198 LoopHeaders->erase(BB);
4199 BB->eraseFromParent();
4200 }
4201
4202 return true;
4203 }
4204
4205 // Check out all of the conditional branches going to this return
4206 // instruction. If any of them just select between returns, change the
4207 // branch itself into a select/return pair.
4208 while (!CondBranchPreds.empty()) {
4209 BranchInst *BI = CondBranchPreds.pop_back_val();
4210
4211 // Check to see if the non-BB successor is also a return block.
4212 if (isa<ReturnInst>(BI->getSuccessor(0)->getTerminator()) &&
4213 isa<ReturnInst>(BI->getSuccessor(1)->getTerminator()) &&
4214 SimplifyCondBranchToTwoReturns(BI, Builder))
4215 return true;
4216 }
4217 return false;
4218 }
4219
SimplifyUnreachable(UnreachableInst * UI)4220 bool SimplifyCFGOpt::SimplifyUnreachable(UnreachableInst *UI) {
4221 BasicBlock *BB = UI->getParent();
4222
4223 bool Changed = false;
4224
4225 // If there are any instructions immediately before the unreachable that can
4226 // be removed, do so.
4227 while (UI->getIterator() != BB->begin()) {
4228 BasicBlock::iterator BBI = UI->getIterator();
4229 --BBI;
4230 // Do not delete instructions that can have side effects which might cause
4231 // the unreachable to not be reachable; specifically, calls and volatile
4232 // operations may have this effect.
4233 if (isa<CallInst>(BBI) && !isa<DbgInfoIntrinsic>(BBI))
4234 break;
4235
4236 if (BBI->mayHaveSideEffects()) {
4237 if (auto *SI = dyn_cast<StoreInst>(BBI)) {
4238 if (SI->isVolatile())
4239 break;
4240 } else if (auto *LI = dyn_cast<LoadInst>(BBI)) {
4241 if (LI->isVolatile())
4242 break;
4243 } else if (auto *RMWI = dyn_cast<AtomicRMWInst>(BBI)) {
4244 if (RMWI->isVolatile())
4245 break;
4246 } else if (auto *CXI = dyn_cast<AtomicCmpXchgInst>(BBI)) {
4247 if (CXI->isVolatile())
4248 break;
4249 } else if (isa<CatchPadInst>(BBI)) {
4250 // A catchpad may invoke exception object constructors and such, which
4251 // in some languages can be arbitrary code, so be conservative by
4252 // default.
4253 // For CoreCLR, it just involves a type test, so can be removed.
4254 if (classifyEHPersonality(BB->getParent()->getPersonalityFn()) !=
4255 EHPersonality::CoreCLR)
4256 break;
4257 } else if (!isa<FenceInst>(BBI) && !isa<VAArgInst>(BBI) &&
4258 !isa<LandingPadInst>(BBI)) {
4259 break;
4260 }
4261 // Note that deleting LandingPad's here is in fact okay, although it
4262 // involves a bit of subtle reasoning. If this inst is a LandingPad,
4263 // all the predecessors of this block will be the unwind edges of Invokes,
4264 // and we can therefore guarantee this block will be erased.
4265 }
4266
4267 // Delete this instruction (any uses are guaranteed to be dead)
4268 if (!BBI->use_empty())
4269 BBI->replaceAllUsesWith(UndefValue::get(BBI->getType()));
4270 BBI->eraseFromParent();
4271 Changed = true;
4272 }
4273
4274 // If the unreachable instruction is the first in the block, take a gander
4275 // at all of the predecessors of this instruction, and simplify them.
4276 if (&BB->front() != UI)
4277 return Changed;
4278
4279 SmallVector<BasicBlock *, 8> Preds(pred_begin(BB), pred_end(BB));
4280 for (unsigned i = 0, e = Preds.size(); i != e; ++i) {
4281 Instruction *TI = Preds[i]->getTerminator();
4282 IRBuilder<> Builder(TI);
4283 if (auto *BI = dyn_cast<BranchInst>(TI)) {
4284 if (BI->isUnconditional()) {
4285 assert(BI->getSuccessor(0) == BB && "Incorrect CFG");
4286 new UnreachableInst(TI->getContext(), TI);
4287 TI->eraseFromParent();
4288 Changed = true;
4289 } else {
4290 Value* Cond = BI->getCondition();
4291 if (BI->getSuccessor(0) == BB) {
4292 Builder.CreateAssumption(Builder.CreateNot(Cond));
4293 Builder.CreateBr(BI->getSuccessor(1));
4294 } else {
4295 assert(BI->getSuccessor(1) == BB && "Incorrect CFG");
4296 Builder.CreateAssumption(Cond);
4297 Builder.CreateBr(BI->getSuccessor(0));
4298 }
4299 EraseTerminatorAndDCECond(BI);
4300 Changed = true;
4301 }
4302 } else if (auto *SI = dyn_cast<SwitchInst>(TI)) {
4303 SwitchInstProfUpdateWrapper SU(*SI);
4304 for (auto i = SU->case_begin(), e = SU->case_end(); i != e;) {
4305 if (i->getCaseSuccessor() != BB) {
4306 ++i;
4307 continue;
4308 }
4309 BB->removePredecessor(SU->getParent());
4310 i = SU.removeCase(i);
4311 e = SU->case_end();
4312 Changed = true;
4313 }
4314 } else if (auto *II = dyn_cast<InvokeInst>(TI)) {
4315 if (II->getUnwindDest() == BB) {
4316 removeUnwindEdge(TI->getParent());
4317 Changed = true;
4318 }
4319 } else if (auto *CSI = dyn_cast<CatchSwitchInst>(TI)) {
4320 if (CSI->getUnwindDest() == BB) {
4321 removeUnwindEdge(TI->getParent());
4322 Changed = true;
4323 continue;
4324 }
4325
4326 for (CatchSwitchInst::handler_iterator I = CSI->handler_begin(),
4327 E = CSI->handler_end();
4328 I != E; ++I) {
4329 if (*I == BB) {
4330 CSI->removeHandler(I);
4331 --I;
4332 --E;
4333 Changed = true;
4334 }
4335 }
4336 if (CSI->getNumHandlers() == 0) {
4337 BasicBlock *CatchSwitchBB = CSI->getParent();
4338 if (CSI->hasUnwindDest()) {
4339 // Redirect preds to the unwind dest
4340 CatchSwitchBB->replaceAllUsesWith(CSI->getUnwindDest());
4341 } else {
4342 // Rewrite all preds to unwind to caller (or from invoke to call).
4343 SmallVector<BasicBlock *, 8> EHPreds(predecessors(CatchSwitchBB));
4344 for (BasicBlock *EHPred : EHPreds)
4345 removeUnwindEdge(EHPred);
4346 }
4347 // The catchswitch is no longer reachable.
4348 new UnreachableInst(CSI->getContext(), CSI);
4349 CSI->eraseFromParent();
4350 Changed = true;
4351 }
4352 } else if (isa<CleanupReturnInst>(TI)) {
4353 new UnreachableInst(TI->getContext(), TI);
4354 TI->eraseFromParent();
4355 Changed = true;
4356 }
4357 }
4358
4359 // If this block is now dead, remove it.
4360 if (pred_empty(BB) && BB != &BB->getParent()->getEntryBlock()) {
4361 // We know there are no successors, so just nuke the block.
4362 if (LoopHeaders)
4363 LoopHeaders->erase(BB);
4364 BB->eraseFromParent();
4365 return true;
4366 }
4367
4368 return Changed;
4369 }
4370
CasesAreContiguous(SmallVectorImpl<ConstantInt * > & Cases)4371 static bool CasesAreContiguous(SmallVectorImpl<ConstantInt *> &Cases) {
4372 assert(Cases.size() >= 1);
4373
4374 array_pod_sort(Cases.begin(), Cases.end(), ConstantIntSortPredicate);
4375 for (size_t I = 1, E = Cases.size(); I != E; ++I) {
4376 if (Cases[I - 1]->getValue() != Cases[I]->getValue() + 1)
4377 return false;
4378 }
4379 return true;
4380 }
4381
createUnreachableSwitchDefault(SwitchInst * Switch)4382 static void createUnreachableSwitchDefault(SwitchInst *Switch) {
4383 LLVM_DEBUG(dbgs() << "SimplifyCFG: switch default is dead.\n");
4384 BasicBlock *NewDefaultBlock =
4385 SplitBlockPredecessors(Switch->getDefaultDest(), Switch->getParent(), "");
4386 Switch->setDefaultDest(&*NewDefaultBlock);
4387 SplitBlock(&*NewDefaultBlock, &NewDefaultBlock->front());
4388 auto *NewTerminator = NewDefaultBlock->getTerminator();
4389 new UnreachableInst(Switch->getContext(), NewTerminator);
4390 EraseTerminatorAndDCECond(NewTerminator);
4391 }
4392
4393 /// Turn a switch with two reachable destinations into an integer range
4394 /// comparison and branch.
TurnSwitchRangeIntoICmp(SwitchInst * SI,IRBuilder<> & Builder)4395 static bool TurnSwitchRangeIntoICmp(SwitchInst *SI, IRBuilder<> &Builder) {
4396 assert(SI->getNumCases() > 1 && "Degenerate switch?");
4397
4398 bool HasDefault =
4399 !isa<UnreachableInst>(SI->getDefaultDest()->getFirstNonPHIOrDbg());
4400
4401 // Partition the cases into two sets with different destinations.
4402 BasicBlock *DestA = HasDefault ? SI->getDefaultDest() : nullptr;
4403 BasicBlock *DestB = nullptr;
4404 SmallVector<ConstantInt *, 16> CasesA;
4405 SmallVector<ConstantInt *, 16> CasesB;
4406
4407 for (auto Case : SI->cases()) {
4408 BasicBlock *Dest = Case.getCaseSuccessor();
4409 if (!DestA)
4410 DestA = Dest;
4411 if (Dest == DestA) {
4412 CasesA.push_back(Case.getCaseValue());
4413 continue;
4414 }
4415 if (!DestB)
4416 DestB = Dest;
4417 if (Dest == DestB) {
4418 CasesB.push_back(Case.getCaseValue());
4419 continue;
4420 }
4421 return false; // More than two destinations.
4422 }
4423
4424 assert(DestA && DestB &&
4425 "Single-destination switch should have been folded.");
4426 assert(DestA != DestB);
4427 assert(DestB != SI->getDefaultDest());
4428 assert(!CasesB.empty() && "There must be non-default cases.");
4429 assert(!CasesA.empty() || HasDefault);
4430
4431 // Figure out if one of the sets of cases form a contiguous range.
4432 SmallVectorImpl<ConstantInt *> *ContiguousCases = nullptr;
4433 BasicBlock *ContiguousDest = nullptr;
4434 BasicBlock *OtherDest = nullptr;
4435 if (!CasesA.empty() && CasesAreContiguous(CasesA)) {
4436 ContiguousCases = &CasesA;
4437 ContiguousDest = DestA;
4438 OtherDest = DestB;
4439 } else if (CasesAreContiguous(CasesB)) {
4440 ContiguousCases = &CasesB;
4441 ContiguousDest = DestB;
4442 OtherDest = DestA;
4443 } else
4444 return false;
4445
4446 // Start building the compare and branch.
4447
4448 Constant *Offset = ConstantExpr::getNeg(ContiguousCases->back());
4449 Constant *NumCases =
4450 ConstantInt::get(Offset->getType(), ContiguousCases->size());
4451
4452 Value *Sub = SI->getCondition();
4453 if (!Offset->isNullValue())
4454 Sub = Builder.CreateAdd(Sub, Offset, Sub->getName() + ".off");
4455
4456 Value *Cmp;
4457 // If NumCases overflowed, then all possible values jump to the successor.
4458 if (NumCases->isNullValue() && !ContiguousCases->empty())
4459 Cmp = ConstantInt::getTrue(SI->getContext());
4460 else
4461 Cmp = Builder.CreateICmpULT(Sub, NumCases, "switch");
4462 BranchInst *NewBI = Builder.CreateCondBr(Cmp, ContiguousDest, OtherDest);
4463
4464 // Update weight for the newly-created conditional branch.
4465 if (HasBranchWeights(SI)) {
4466 SmallVector<uint64_t, 8> Weights;
4467 GetBranchWeights(SI, Weights);
4468 if (Weights.size() == 1 + SI->getNumCases()) {
4469 uint64_t TrueWeight = 0;
4470 uint64_t FalseWeight = 0;
4471 for (size_t I = 0, E = Weights.size(); I != E; ++I) {
4472 if (SI->getSuccessor(I) == ContiguousDest)
4473 TrueWeight += Weights[I];
4474 else
4475 FalseWeight += Weights[I];
4476 }
4477 while (TrueWeight > UINT32_MAX || FalseWeight > UINT32_MAX) {
4478 TrueWeight /= 2;
4479 FalseWeight /= 2;
4480 }
4481 setBranchWeights(NewBI, TrueWeight, FalseWeight);
4482 }
4483 }
4484
4485 // Prune obsolete incoming values off the successors' PHI nodes.
4486 for (auto BBI = ContiguousDest->begin(); isa<PHINode>(BBI); ++BBI) {
4487 unsigned PreviousEdges = ContiguousCases->size();
4488 if (ContiguousDest == SI->getDefaultDest())
4489 ++PreviousEdges;
4490 for (unsigned I = 0, E = PreviousEdges - 1; I != E; ++I)
4491 cast<PHINode>(BBI)->removeIncomingValue(SI->getParent());
4492 }
4493 for (auto BBI = OtherDest->begin(); isa<PHINode>(BBI); ++BBI) {
4494 unsigned PreviousEdges = SI->getNumCases() - ContiguousCases->size();
4495 if (OtherDest == SI->getDefaultDest())
4496 ++PreviousEdges;
4497 for (unsigned I = 0, E = PreviousEdges - 1; I != E; ++I)
4498 cast<PHINode>(BBI)->removeIncomingValue(SI->getParent());
4499 }
4500
4501 // Clean up the default block - it may have phis or other instructions before
4502 // the unreachable terminator.
4503 if (!HasDefault)
4504 createUnreachableSwitchDefault(SI);
4505
4506 // Drop the switch.
4507 SI->eraseFromParent();
4508
4509 return true;
4510 }
4511
4512 /// Compute masked bits for the condition of a switch
4513 /// and use it to remove dead cases.
eliminateDeadSwitchCases(SwitchInst * SI,AssumptionCache * AC,const DataLayout & DL)4514 static bool eliminateDeadSwitchCases(SwitchInst *SI, AssumptionCache *AC,
4515 const DataLayout &DL) {
4516 Value *Cond = SI->getCondition();
4517 unsigned Bits = Cond->getType()->getIntegerBitWidth();
4518 KnownBits Known = computeKnownBits(Cond, DL, 0, AC, SI);
4519
4520 // We can also eliminate cases by determining that their values are outside of
4521 // the limited range of the condition based on how many significant (non-sign)
4522 // bits are in the condition value.
4523 unsigned ExtraSignBits = ComputeNumSignBits(Cond, DL, 0, AC, SI) - 1;
4524 unsigned MaxSignificantBitsInCond = Bits - ExtraSignBits;
4525
4526 // Gather dead cases.
4527 SmallVector<ConstantInt *, 8> DeadCases;
4528 for (auto &Case : SI->cases()) {
4529 const APInt &CaseVal = Case.getCaseValue()->getValue();
4530 if (Known.Zero.intersects(CaseVal) || !Known.One.isSubsetOf(CaseVal) ||
4531 (CaseVal.getMinSignedBits() > MaxSignificantBitsInCond)) {
4532 DeadCases.push_back(Case.getCaseValue());
4533 LLVM_DEBUG(dbgs() << "SimplifyCFG: switch case " << CaseVal
4534 << " is dead.\n");
4535 }
4536 }
4537
4538 // If we can prove that the cases must cover all possible values, the
4539 // default destination becomes dead and we can remove it. If we know some
4540 // of the bits in the value, we can use that to more precisely compute the
4541 // number of possible unique case values.
4542 bool HasDefault =
4543 !isa<UnreachableInst>(SI->getDefaultDest()->getFirstNonPHIOrDbg());
4544 const unsigned NumUnknownBits =
4545 Bits - (Known.Zero | Known.One).countPopulation();
4546 assert(NumUnknownBits <= Bits);
4547 if (HasDefault && DeadCases.empty() &&
4548 NumUnknownBits < 64 /* avoid overflow */ &&
4549 SI->getNumCases() == (1ULL << NumUnknownBits)) {
4550 createUnreachableSwitchDefault(SI);
4551 return true;
4552 }
4553
4554 if (DeadCases.empty())
4555 return false;
4556
4557 SwitchInstProfUpdateWrapper SIW(*SI);
4558 for (ConstantInt *DeadCase : DeadCases) {
4559 SwitchInst::CaseIt CaseI = SI->findCaseValue(DeadCase);
4560 assert(CaseI != SI->case_default() &&
4561 "Case was not found. Probably mistake in DeadCases forming.");
4562 // Prune unused values from PHI nodes.
4563 CaseI->getCaseSuccessor()->removePredecessor(SI->getParent());
4564 SIW.removeCase(CaseI);
4565 }
4566
4567 return true;
4568 }
4569
4570 /// If BB would be eligible for simplification by
4571 /// TryToSimplifyUncondBranchFromEmptyBlock (i.e. it is empty and terminated
4572 /// by an unconditional branch), look at the phi node for BB in the successor
4573 /// block and see if the incoming value is equal to CaseValue. If so, return
4574 /// the phi node, and set PhiIndex to BB's index in the phi node.
FindPHIForConditionForwarding(ConstantInt * CaseValue,BasicBlock * BB,int * PhiIndex)4575 static PHINode *FindPHIForConditionForwarding(ConstantInt *CaseValue,
4576 BasicBlock *BB, int *PhiIndex) {
4577 if (BB->getFirstNonPHIOrDbg() != BB->getTerminator())
4578 return nullptr; // BB must be empty to be a candidate for simplification.
4579 if (!BB->getSinglePredecessor())
4580 return nullptr; // BB must be dominated by the switch.
4581
4582 BranchInst *Branch = dyn_cast<BranchInst>(BB->getTerminator());
4583 if (!Branch || !Branch->isUnconditional())
4584 return nullptr; // Terminator must be unconditional branch.
4585
4586 BasicBlock *Succ = Branch->getSuccessor(0);
4587
4588 for (PHINode &PHI : Succ->phis()) {
4589 int Idx = PHI.getBasicBlockIndex(BB);
4590 assert(Idx >= 0 && "PHI has no entry for predecessor?");
4591
4592 Value *InValue = PHI.getIncomingValue(Idx);
4593 if (InValue != CaseValue)
4594 continue;
4595
4596 *PhiIndex = Idx;
4597 return &PHI;
4598 }
4599
4600 return nullptr;
4601 }
4602
4603 /// Try to forward the condition of a switch instruction to a phi node
4604 /// dominated by the switch, if that would mean that some of the destination
4605 /// blocks of the switch can be folded away. Return true if a change is made.
ForwardSwitchConditionToPHI(SwitchInst * SI)4606 static bool ForwardSwitchConditionToPHI(SwitchInst *SI) {
4607 using ForwardingNodesMap = DenseMap<PHINode *, SmallVector<int, 4>>;
4608
4609 ForwardingNodesMap ForwardingNodes;
4610 BasicBlock *SwitchBlock = SI->getParent();
4611 bool Changed = false;
4612 for (auto &Case : SI->cases()) {
4613 ConstantInt *CaseValue = Case.getCaseValue();
4614 BasicBlock *CaseDest = Case.getCaseSuccessor();
4615
4616 // Replace phi operands in successor blocks that are using the constant case
4617 // value rather than the switch condition variable:
4618 // switchbb:
4619 // switch i32 %x, label %default [
4620 // i32 17, label %succ
4621 // ...
4622 // succ:
4623 // %r = phi i32 ... [ 17, %switchbb ] ...
4624 // -->
4625 // %r = phi i32 ... [ %x, %switchbb ] ...
4626
4627 for (PHINode &Phi : CaseDest->phis()) {
4628 // This only works if there is exactly 1 incoming edge from the switch to
4629 // a phi. If there is >1, that means multiple cases of the switch map to 1
4630 // value in the phi, and that phi value is not the switch condition. Thus,
4631 // this transform would not make sense (the phi would be invalid because
4632 // a phi can't have different incoming values from the same block).
4633 int SwitchBBIdx = Phi.getBasicBlockIndex(SwitchBlock);
4634 if (Phi.getIncomingValue(SwitchBBIdx) == CaseValue &&
4635 count(Phi.blocks(), SwitchBlock) == 1) {
4636 Phi.setIncomingValue(SwitchBBIdx, SI->getCondition());
4637 Changed = true;
4638 }
4639 }
4640
4641 // Collect phi nodes that are indirectly using this switch's case constants.
4642 int PhiIdx;
4643 if (auto *Phi = FindPHIForConditionForwarding(CaseValue, CaseDest, &PhiIdx))
4644 ForwardingNodes[Phi].push_back(PhiIdx);
4645 }
4646
4647 for (auto &ForwardingNode : ForwardingNodes) {
4648 PHINode *Phi = ForwardingNode.first;
4649 SmallVectorImpl<int> &Indexes = ForwardingNode.second;
4650 if (Indexes.size() < 2)
4651 continue;
4652
4653 for (int Index : Indexes)
4654 Phi->setIncomingValue(Index, SI->getCondition());
4655 Changed = true;
4656 }
4657
4658 return Changed;
4659 }
4660
4661 /// Return true if the backend will be able to handle
4662 /// initializing an array of constants like C.
ValidLookupTableConstant(Constant * C,const TargetTransformInfo & TTI)4663 static bool ValidLookupTableConstant(Constant *C, const TargetTransformInfo &TTI) {
4664 if (C->isThreadDependent())
4665 return false;
4666 if (C->isDLLImportDependent())
4667 return false;
4668
4669 if (!isa<ConstantFP>(C) && !isa<ConstantInt>(C) &&
4670 !isa<ConstantPointerNull>(C) && !isa<GlobalValue>(C) &&
4671 !isa<UndefValue>(C) && !isa<ConstantExpr>(C))
4672 return false;
4673
4674 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) {
4675 if (!CE->isGEPWithNoNotionalOverIndexing())
4676 return false;
4677 if (!ValidLookupTableConstant(CE->getOperand(0), TTI))
4678 return false;
4679 }
4680
4681 if (!TTI.shouldBuildLookupTablesForConstant(C))
4682 return false;
4683
4684 return true;
4685 }
4686
4687 /// If V is a Constant, return it. Otherwise, try to look up
4688 /// its constant value in ConstantPool, returning 0 if it's not there.
4689 static Constant *
LookupConstant(Value * V,const SmallDenseMap<Value *,Constant * > & ConstantPool)4690 LookupConstant(Value *V,
4691 const SmallDenseMap<Value *, Constant *> &ConstantPool) {
4692 if (Constant *C = dyn_cast<Constant>(V))
4693 return C;
4694 return ConstantPool.lookup(V);
4695 }
4696
4697 /// Try to fold instruction I into a constant. This works for
4698 /// simple instructions such as binary operations where both operands are
4699 /// constant or can be replaced by constants from the ConstantPool. Returns the
4700 /// resulting constant on success, 0 otherwise.
4701 static Constant *
ConstantFold(Instruction * I,const DataLayout & DL,const SmallDenseMap<Value *,Constant * > & ConstantPool)4702 ConstantFold(Instruction *I, const DataLayout &DL,
4703 const SmallDenseMap<Value *, Constant *> &ConstantPool) {
4704 if (SelectInst *Select = dyn_cast<SelectInst>(I)) {
4705 Constant *A = LookupConstant(Select->getCondition(), ConstantPool);
4706 if (!A)
4707 return nullptr;
4708 if (A->isAllOnesValue())
4709 return LookupConstant(Select->getTrueValue(), ConstantPool);
4710 if (A->isNullValue())
4711 return LookupConstant(Select->getFalseValue(), ConstantPool);
4712 return nullptr;
4713 }
4714
4715 SmallVector<Constant *, 4> COps;
4716 for (unsigned N = 0, E = I->getNumOperands(); N != E; ++N) {
4717 if (Constant *A = LookupConstant(I->getOperand(N), ConstantPool))
4718 COps.push_back(A);
4719 else
4720 return nullptr;
4721 }
4722
4723 if (CmpInst *Cmp = dyn_cast<CmpInst>(I)) {
4724 return ConstantFoldCompareInstOperands(Cmp->getPredicate(), COps[0],
4725 COps[1], DL);
4726 }
4727
4728 return ConstantFoldInstOperands(I, COps, DL);
4729 }
4730
4731 /// Try to determine the resulting constant values in phi nodes
4732 /// at the common destination basic block, *CommonDest, for one of the case
4733 /// destionations CaseDest corresponding to value CaseVal (0 for the default
4734 /// case), of a switch instruction SI.
4735 static bool
GetCaseResults(SwitchInst * SI,ConstantInt * CaseVal,BasicBlock * CaseDest,BasicBlock ** CommonDest,SmallVectorImpl<std::pair<PHINode *,Constant * >> & Res,const DataLayout & DL,const TargetTransformInfo & TTI)4736 GetCaseResults(SwitchInst *SI, ConstantInt *CaseVal, BasicBlock *CaseDest,
4737 BasicBlock **CommonDest,
4738 SmallVectorImpl<std::pair<PHINode *, Constant *>> &Res,
4739 const DataLayout &DL, const TargetTransformInfo &TTI) {
4740 // The block from which we enter the common destination.
4741 BasicBlock *Pred = SI->getParent();
4742
4743 // If CaseDest is empty except for some side-effect free instructions through
4744 // which we can constant-propagate the CaseVal, continue to its successor.
4745 SmallDenseMap<Value *, Constant *> ConstantPool;
4746 ConstantPool.insert(std::make_pair(SI->getCondition(), CaseVal));
4747 for (Instruction &I :CaseDest->instructionsWithoutDebug()) {
4748 if (I.isTerminator()) {
4749 // If the terminator is a simple branch, continue to the next block.
4750 if (I.getNumSuccessors() != 1 || I.isExceptionalTerminator())
4751 return false;
4752 Pred = CaseDest;
4753 CaseDest = I.getSuccessor(0);
4754 } else if (Constant *C = ConstantFold(&I, DL, ConstantPool)) {
4755 // Instruction is side-effect free and constant.
4756
4757 // If the instruction has uses outside this block or a phi node slot for
4758 // the block, it is not safe to bypass the instruction since it would then
4759 // no longer dominate all its uses.
4760 for (auto &Use : I.uses()) {
4761 User *User = Use.getUser();
4762 if (Instruction *I = dyn_cast<Instruction>(User))
4763 if (I->getParent() == CaseDest)
4764 continue;
4765 if (PHINode *Phi = dyn_cast<PHINode>(User))
4766 if (Phi->getIncomingBlock(Use) == CaseDest)
4767 continue;
4768 return false;
4769 }
4770
4771 ConstantPool.insert(std::make_pair(&I, C));
4772 } else {
4773 break;
4774 }
4775 }
4776
4777 // If we did not have a CommonDest before, use the current one.
4778 if (!*CommonDest)
4779 *CommonDest = CaseDest;
4780 // If the destination isn't the common one, abort.
4781 if (CaseDest != *CommonDest)
4782 return false;
4783
4784 // Get the values for this case from phi nodes in the destination block.
4785 for (PHINode &PHI : (*CommonDest)->phis()) {
4786 int Idx = PHI.getBasicBlockIndex(Pred);
4787 if (Idx == -1)
4788 continue;
4789
4790 Constant *ConstVal =
4791 LookupConstant(PHI.getIncomingValue(Idx), ConstantPool);
4792 if (!ConstVal)
4793 return false;
4794
4795 // Be conservative about which kinds of constants we support.
4796 if (!ValidLookupTableConstant(ConstVal, TTI))
4797 return false;
4798
4799 Res.push_back(std::make_pair(&PHI, ConstVal));
4800 }
4801
4802 return Res.size() > 0;
4803 }
4804
4805 // Helper function used to add CaseVal to the list of cases that generate
4806 // Result. Returns the updated number of cases that generate this result.
MapCaseToResult(ConstantInt * CaseVal,SwitchCaseResultVectorTy & UniqueResults,Constant * Result)4807 static uintptr_t MapCaseToResult(ConstantInt *CaseVal,
4808 SwitchCaseResultVectorTy &UniqueResults,
4809 Constant *Result) {
4810 for (auto &I : UniqueResults) {
4811 if (I.first == Result) {
4812 I.second.push_back(CaseVal);
4813 return I.second.size();
4814 }
4815 }
4816 UniqueResults.push_back(
4817 std::make_pair(Result, SmallVector<ConstantInt *, 4>(1, CaseVal)));
4818 return 1;
4819 }
4820
4821 // Helper function that initializes a map containing
4822 // results for the PHI node of the common destination block for a switch
4823 // instruction. Returns false if multiple PHI nodes have been found or if
4824 // there is not a common destination block for the switch.
4825 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)4826 InitializeUniqueCases(SwitchInst *SI, PHINode *&PHI, BasicBlock *&CommonDest,
4827 SwitchCaseResultVectorTy &UniqueResults,
4828 Constant *&DefaultResult, const DataLayout &DL,
4829 const TargetTransformInfo &TTI,
4830 uintptr_t MaxUniqueResults, uintptr_t MaxCasesPerResult) {
4831 for (auto &I : SI->cases()) {
4832 ConstantInt *CaseVal = I.getCaseValue();
4833
4834 // Resulting value at phi nodes for this case value.
4835 SwitchCaseResultsTy Results;
4836 if (!GetCaseResults(SI, CaseVal, I.getCaseSuccessor(), &CommonDest, Results,
4837 DL, TTI))
4838 return false;
4839
4840 // Only one value per case is permitted.
4841 if (Results.size() > 1)
4842 return false;
4843
4844 // Add the case->result mapping to UniqueResults.
4845 const uintptr_t NumCasesForResult =
4846 MapCaseToResult(CaseVal, UniqueResults, Results.begin()->second);
4847
4848 // Early out if there are too many cases for this result.
4849 if (NumCasesForResult > MaxCasesPerResult)
4850 return false;
4851
4852 // Early out if there are too many unique results.
4853 if (UniqueResults.size() > MaxUniqueResults)
4854 return false;
4855
4856 // Check the PHI consistency.
4857 if (!PHI)
4858 PHI = Results[0].first;
4859 else if (PHI != Results[0].first)
4860 return false;
4861 }
4862 // Find the default result value.
4863 SmallVector<std::pair<PHINode *, Constant *>, 1> DefaultResults;
4864 BasicBlock *DefaultDest = SI->getDefaultDest();
4865 GetCaseResults(SI, nullptr, SI->getDefaultDest(), &CommonDest, DefaultResults,
4866 DL, TTI);
4867 // If the default value is not found abort unless the default destination
4868 // is unreachable.
4869 DefaultResult =
4870 DefaultResults.size() == 1 ? DefaultResults.begin()->second : nullptr;
4871 if ((!DefaultResult &&
4872 !isa<UnreachableInst>(DefaultDest->getFirstNonPHIOrDbg())))
4873 return false;
4874
4875 return true;
4876 }
4877
4878 // Helper function that checks if it is possible to transform a switch with only
4879 // two cases (or two cases + default) that produces a result into a select.
4880 // Example:
4881 // switch (a) {
4882 // case 10: %0 = icmp eq i32 %a, 10
4883 // return 10; %1 = select i1 %0, i32 10, i32 4
4884 // case 20: ----> %2 = icmp eq i32 %a, 20
4885 // return 2; %3 = select i1 %2, i32 2, i32 %1
4886 // default:
4887 // return 4;
4888 // }
ConvertTwoCaseSwitch(const SwitchCaseResultVectorTy & ResultVector,Constant * DefaultResult,Value * Condition,IRBuilder<> & Builder)4889 static Value *ConvertTwoCaseSwitch(const SwitchCaseResultVectorTy &ResultVector,
4890 Constant *DefaultResult, Value *Condition,
4891 IRBuilder<> &Builder) {
4892 assert(ResultVector.size() == 2 &&
4893 "We should have exactly two unique results at this point");
4894 // If we are selecting between only two cases transform into a simple
4895 // select or a two-way select if default is possible.
4896 if (ResultVector[0].second.size() == 1 &&
4897 ResultVector[1].second.size() == 1) {
4898 ConstantInt *const FirstCase = ResultVector[0].second[0];
4899 ConstantInt *const SecondCase = ResultVector[1].second[0];
4900
4901 bool DefaultCanTrigger = DefaultResult;
4902 Value *SelectValue = ResultVector[1].first;
4903 if (DefaultCanTrigger) {
4904 Value *const ValueCompare =
4905 Builder.CreateICmpEQ(Condition, SecondCase, "switch.selectcmp");
4906 SelectValue = Builder.CreateSelect(ValueCompare, ResultVector[1].first,
4907 DefaultResult, "switch.select");
4908 }
4909 Value *const ValueCompare =
4910 Builder.CreateICmpEQ(Condition, FirstCase, "switch.selectcmp");
4911 return Builder.CreateSelect(ValueCompare, ResultVector[0].first,
4912 SelectValue, "switch.select");
4913 }
4914
4915 return nullptr;
4916 }
4917
4918 // Helper function to cleanup a switch instruction that has been converted into
4919 // a select, fixing up PHI nodes and basic blocks.
RemoveSwitchAfterSelectConversion(SwitchInst * SI,PHINode * PHI,Value * SelectValue,IRBuilder<> & Builder)4920 static void RemoveSwitchAfterSelectConversion(SwitchInst *SI, PHINode *PHI,
4921 Value *SelectValue,
4922 IRBuilder<> &Builder) {
4923 BasicBlock *SelectBB = SI->getParent();
4924 while (PHI->getBasicBlockIndex(SelectBB) >= 0)
4925 PHI->removeIncomingValue(SelectBB);
4926 PHI->addIncoming(SelectValue, SelectBB);
4927
4928 Builder.CreateBr(PHI->getParent());
4929
4930 // Remove the switch.
4931 for (unsigned i = 0, e = SI->getNumSuccessors(); i < e; ++i) {
4932 BasicBlock *Succ = SI->getSuccessor(i);
4933
4934 if (Succ == PHI->getParent())
4935 continue;
4936 Succ->removePredecessor(SelectBB);
4937 }
4938 SI->eraseFromParent();
4939 }
4940
4941 /// If the switch is only used to initialize one or more
4942 /// phi nodes in a common successor block with only two different
4943 /// constant values, replace the switch with select.
switchToSelect(SwitchInst * SI,IRBuilder<> & Builder,const DataLayout & DL,const TargetTransformInfo & TTI)4944 static bool switchToSelect(SwitchInst *SI, IRBuilder<> &Builder,
4945 const DataLayout &DL,
4946 const TargetTransformInfo &TTI) {
4947 Value *const Cond = SI->getCondition();
4948 PHINode *PHI = nullptr;
4949 BasicBlock *CommonDest = nullptr;
4950 Constant *DefaultResult;
4951 SwitchCaseResultVectorTy UniqueResults;
4952 // Collect all the cases that will deliver the same value from the switch.
4953 if (!InitializeUniqueCases(SI, PHI, CommonDest, UniqueResults, DefaultResult,
4954 DL, TTI, 2, 1))
4955 return false;
4956 // Selects choose between maximum two values.
4957 if (UniqueResults.size() != 2)
4958 return false;
4959 assert(PHI != nullptr && "PHI for value select not found");
4960
4961 Builder.SetInsertPoint(SI);
4962 Value *SelectValue =
4963 ConvertTwoCaseSwitch(UniqueResults, DefaultResult, Cond, Builder);
4964 if (SelectValue) {
4965 RemoveSwitchAfterSelectConversion(SI, PHI, SelectValue, Builder);
4966 return true;
4967 }
4968 // The switch couldn't be converted into a select.
4969 return false;
4970 }
4971
4972 namespace {
4973
4974 /// This class represents a lookup table that can be used to replace a switch.
4975 class SwitchLookupTable {
4976 public:
4977 /// Create a lookup table to use as a switch replacement with the contents
4978 /// of Values, using DefaultValue to fill any holes in the table.
4979 SwitchLookupTable(
4980 Module &M, uint64_t TableSize, ConstantInt *Offset,
4981 const SmallVectorImpl<std::pair<ConstantInt *, Constant *>> &Values,
4982 Constant *DefaultValue, const DataLayout &DL, const StringRef &FuncName);
4983
4984 /// Build instructions with Builder to retrieve the value at
4985 /// the position given by Index in the lookup table.
4986 Value *BuildLookup(Value *Index, IRBuilder<> &Builder);
4987
4988 /// Return true if a table with TableSize elements of
4989 /// type ElementType would fit in a target-legal register.
4990 static bool WouldFitInRegister(const DataLayout &DL, uint64_t TableSize,
4991 Type *ElementType);
4992
4993 private:
4994 // Depending on the contents of the table, it can be represented in
4995 // different ways.
4996 enum {
4997 // For tables where each element contains the same value, we just have to
4998 // store that single value and return it for each lookup.
4999 SingleValueKind,
5000
5001 // For tables where there is a linear relationship between table index
5002 // and values. We calculate the result with a simple multiplication
5003 // and addition instead of a table lookup.
5004 LinearMapKind,
5005
5006 // For small tables with integer elements, we can pack them into a bitmap
5007 // that fits into a target-legal register. Values are retrieved by
5008 // shift and mask operations.
5009 BitMapKind,
5010
5011 // The table is stored as an array of values. Values are retrieved by load
5012 // instructions from the table.
5013 ArrayKind
5014 } Kind;
5015
5016 // For SingleValueKind, this is the single value.
5017 Constant *SingleValue = nullptr;
5018
5019 // For BitMapKind, this is the bitmap.
5020 ConstantInt *BitMap = nullptr;
5021 IntegerType *BitMapElementTy = nullptr;
5022
5023 // For LinearMapKind, these are the constants used to derive the value.
5024 ConstantInt *LinearOffset = nullptr;
5025 ConstantInt *LinearMultiplier = nullptr;
5026
5027 // For ArrayKind, this is the array.
5028 GlobalVariable *Array = nullptr;
5029 };
5030
5031 } // end anonymous namespace
5032
SwitchLookupTable(Module & M,uint64_t TableSize,ConstantInt * Offset,const SmallVectorImpl<std::pair<ConstantInt *,Constant * >> & Values,Constant * DefaultValue,const DataLayout & DL,const StringRef & FuncName)5033 SwitchLookupTable::SwitchLookupTable(
5034 Module &M, uint64_t TableSize, ConstantInt *Offset,
5035 const SmallVectorImpl<std::pair<ConstantInt *, Constant *>> &Values,
5036 Constant *DefaultValue, const DataLayout &DL, const StringRef &FuncName) {
5037 assert(Values.size() && "Can't build lookup table without values!");
5038 assert(TableSize >= Values.size() && "Can't fit values in table!");
5039
5040 // If all values in the table are equal, this is that value.
5041 SingleValue = Values.begin()->second;
5042
5043 Type *ValueType = Values.begin()->second->getType();
5044
5045 // Build up the table contents.
5046 SmallVector<Constant *, 64> TableContents(TableSize);
5047 for (size_t I = 0, E = Values.size(); I != E; ++I) {
5048 ConstantInt *CaseVal = Values[I].first;
5049 Constant *CaseRes = Values[I].second;
5050 assert(CaseRes->getType() == ValueType);
5051
5052 uint64_t Idx = (CaseVal->getValue() - Offset->getValue()).getLimitedValue();
5053 TableContents[Idx] = CaseRes;
5054
5055 if (CaseRes != SingleValue)
5056 SingleValue = nullptr;
5057 }
5058
5059 // Fill in any holes in the table with the default result.
5060 if (Values.size() < TableSize) {
5061 assert(DefaultValue &&
5062 "Need a default value to fill the lookup table holes.");
5063 assert(DefaultValue->getType() == ValueType);
5064 for (uint64_t I = 0; I < TableSize; ++I) {
5065 if (!TableContents[I])
5066 TableContents[I] = DefaultValue;
5067 }
5068
5069 if (DefaultValue != SingleValue)
5070 SingleValue = nullptr;
5071 }
5072
5073 // If each element in the table contains the same value, we only need to store
5074 // that single value.
5075 if (SingleValue) {
5076 Kind = SingleValueKind;
5077 return;
5078 }
5079
5080 // Check if we can derive the value with a linear transformation from the
5081 // table index.
5082 if (isa<IntegerType>(ValueType)) {
5083 bool LinearMappingPossible = true;
5084 APInt PrevVal;
5085 APInt DistToPrev;
5086 assert(TableSize >= 2 && "Should be a SingleValue table.");
5087 // Check if there is the same distance between two consecutive values.
5088 for (uint64_t I = 0; I < TableSize; ++I) {
5089 ConstantInt *ConstVal = dyn_cast<ConstantInt>(TableContents[I]);
5090 if (!ConstVal) {
5091 // This is an undef. We could deal with it, but undefs in lookup tables
5092 // are very seldom. It's probably not worth the additional complexity.
5093 LinearMappingPossible = false;
5094 break;
5095 }
5096 const APInt &Val = ConstVal->getValue();
5097 if (I != 0) {
5098 APInt Dist = Val - PrevVal;
5099 if (I == 1) {
5100 DistToPrev = Dist;
5101 } else if (Dist != DistToPrev) {
5102 LinearMappingPossible = false;
5103 break;
5104 }
5105 }
5106 PrevVal = Val;
5107 }
5108 if (LinearMappingPossible) {
5109 LinearOffset = cast<ConstantInt>(TableContents[0]);
5110 LinearMultiplier = ConstantInt::get(M.getContext(), DistToPrev);
5111 Kind = LinearMapKind;
5112 ++NumLinearMaps;
5113 return;
5114 }
5115 }
5116
5117 // If the type is integer and the table fits in a register, build a bitmap.
5118 if (WouldFitInRegister(DL, TableSize, ValueType)) {
5119 IntegerType *IT = cast<IntegerType>(ValueType);
5120 APInt TableInt(TableSize * IT->getBitWidth(), 0);
5121 for (uint64_t I = TableSize; I > 0; --I) {
5122 TableInt <<= IT->getBitWidth();
5123 // Insert values into the bitmap. Undef values are set to zero.
5124 if (!isa<UndefValue>(TableContents[I - 1])) {
5125 ConstantInt *Val = cast<ConstantInt>(TableContents[I - 1]);
5126 TableInt |= Val->getValue().zext(TableInt.getBitWidth());
5127 }
5128 }
5129 BitMap = ConstantInt::get(M.getContext(), TableInt);
5130 BitMapElementTy = IT;
5131 Kind = BitMapKind;
5132 ++NumBitMaps;
5133 return;
5134 }
5135
5136 // Store the table in an array.
5137 ArrayType *ArrayTy = ArrayType::get(ValueType, TableSize);
5138 Constant *Initializer = ConstantArray::get(ArrayTy, TableContents);
5139
5140 Array = new GlobalVariable(M, ArrayTy, /*isConstant=*/true,
5141 GlobalVariable::PrivateLinkage, Initializer,
5142 "switch.table." + FuncName);
5143 Array->setUnnamedAddr(GlobalValue::UnnamedAddr::Global);
5144 // Set the alignment to that of an array items. We will be only loading one
5145 // value out of it.
5146 Array->setAlignment(Align(DL.getPrefTypeAlignment(ValueType)));
5147 Kind = ArrayKind;
5148 }
5149
BuildLookup(Value * Index,IRBuilder<> & Builder)5150 Value *SwitchLookupTable::BuildLookup(Value *Index, IRBuilder<> &Builder) {
5151 switch (Kind) {
5152 case SingleValueKind:
5153 return SingleValue;
5154 case LinearMapKind: {
5155 // Derive the result value from the input value.
5156 Value *Result = Builder.CreateIntCast(Index, LinearMultiplier->getType(),
5157 false, "switch.idx.cast");
5158 if (!LinearMultiplier->isOne())
5159 Result = Builder.CreateMul(Result, LinearMultiplier, "switch.idx.mult");
5160 if (!LinearOffset->isZero())
5161 Result = Builder.CreateAdd(Result, LinearOffset, "switch.offset");
5162 return Result;
5163 }
5164 case BitMapKind: {
5165 // Type of the bitmap (e.g. i59).
5166 IntegerType *MapTy = BitMap->getType();
5167
5168 // Cast Index to the same type as the bitmap.
5169 // Note: The Index is <= the number of elements in the table, so
5170 // truncating it to the width of the bitmask is safe.
5171 Value *ShiftAmt = Builder.CreateZExtOrTrunc(Index, MapTy, "switch.cast");
5172
5173 // Multiply the shift amount by the element width.
5174 ShiftAmt = Builder.CreateMul(
5175 ShiftAmt, ConstantInt::get(MapTy, BitMapElementTy->getBitWidth()),
5176 "switch.shiftamt");
5177
5178 // Shift down.
5179 Value *DownShifted =
5180 Builder.CreateLShr(BitMap, ShiftAmt, "switch.downshift");
5181 // Mask off.
5182 return Builder.CreateTrunc(DownShifted, BitMapElementTy, "switch.masked");
5183 }
5184 case ArrayKind: {
5185 // Make sure the table index will not overflow when treated as signed.
5186 IntegerType *IT = cast<IntegerType>(Index->getType());
5187 uint64_t TableSize =
5188 Array->getInitializer()->getType()->getArrayNumElements();
5189 if (TableSize > (1ULL << (IT->getBitWidth() - 1)))
5190 Index = Builder.CreateZExt(
5191 Index, IntegerType::get(IT->getContext(), IT->getBitWidth() + 1),
5192 "switch.tableidx.zext");
5193
5194 Value *GEPIndices[] = {Builder.getInt32(0), Index};
5195 Value *GEP = Builder.CreateInBoundsGEP(Array->getValueType(), Array,
5196 GEPIndices, "switch.gep");
5197 return Builder.CreateLoad(
5198 cast<ArrayType>(Array->getValueType())->getElementType(), GEP,
5199 "switch.load");
5200 }
5201 }
5202 llvm_unreachable("Unknown lookup table kind!");
5203 }
5204
WouldFitInRegister(const DataLayout & DL,uint64_t TableSize,Type * ElementType)5205 bool SwitchLookupTable::WouldFitInRegister(const DataLayout &DL,
5206 uint64_t TableSize,
5207 Type *ElementType) {
5208 auto *IT = dyn_cast<IntegerType>(ElementType);
5209 if (!IT)
5210 return false;
5211 // FIXME: If the type is wider than it needs to be, e.g. i8 but all values
5212 // are <= 15, we could try to narrow the type.
5213
5214 // Avoid overflow, fitsInLegalInteger uses unsigned int for the width.
5215 if (TableSize >= UINT_MAX / IT->getBitWidth())
5216 return false;
5217 return DL.fitsInLegalInteger(TableSize * IT->getBitWidth());
5218 }
5219
5220 /// Determine whether a lookup table should be built for this switch, based on
5221 /// the number of cases, size of the table, and the types of the results.
5222 static bool
ShouldBuildLookupTable(SwitchInst * SI,uint64_t TableSize,const TargetTransformInfo & TTI,const DataLayout & DL,const SmallDenseMap<PHINode *,Type * > & ResultTypes)5223 ShouldBuildLookupTable(SwitchInst *SI, uint64_t TableSize,
5224 const TargetTransformInfo &TTI, const DataLayout &DL,
5225 const SmallDenseMap<PHINode *, Type *> &ResultTypes) {
5226 if (SI->getNumCases() > TableSize || TableSize >= UINT64_MAX / 10)
5227 return false; // TableSize overflowed, or mul below might overflow.
5228
5229 bool AllTablesFitInRegister = true;
5230 bool HasIllegalType = false;
5231 for (const auto &I : ResultTypes) {
5232 Type *Ty = I.second;
5233
5234 // Saturate this flag to true.
5235 HasIllegalType = HasIllegalType || !TTI.isTypeLegal(Ty);
5236
5237 // Saturate this flag to false.
5238 AllTablesFitInRegister =
5239 AllTablesFitInRegister &&
5240 SwitchLookupTable::WouldFitInRegister(DL, TableSize, Ty);
5241
5242 // If both flags saturate, we're done. NOTE: This *only* works with
5243 // saturating flags, and all flags have to saturate first due to the
5244 // non-deterministic behavior of iterating over a dense map.
5245 if (HasIllegalType && !AllTablesFitInRegister)
5246 break;
5247 }
5248
5249 // If each table would fit in a register, we should build it anyway.
5250 if (AllTablesFitInRegister)
5251 return true;
5252
5253 // Don't build a table that doesn't fit in-register if it has illegal types.
5254 if (HasIllegalType)
5255 return false;
5256
5257 // The table density should be at least 40%. This is the same criterion as for
5258 // jump tables, see SelectionDAGBuilder::handleJTSwitchCase.
5259 // FIXME: Find the best cut-off.
5260 return SI->getNumCases() * 10 >= TableSize * 4;
5261 }
5262
5263 /// Try to reuse the switch table index compare. Following pattern:
5264 /// \code
5265 /// if (idx < tablesize)
5266 /// r = table[idx]; // table does not contain default_value
5267 /// else
5268 /// r = default_value;
5269 /// if (r != default_value)
5270 /// ...
5271 /// \endcode
5272 /// Is optimized to:
5273 /// \code
5274 /// cond = idx < tablesize;
5275 /// if (cond)
5276 /// r = table[idx];
5277 /// else
5278 /// r = default_value;
5279 /// if (cond)
5280 /// ...
5281 /// \endcode
5282 /// 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)5283 static void reuseTableCompare(
5284 User *PhiUser, BasicBlock *PhiBlock, BranchInst *RangeCheckBranch,
5285 Constant *DefaultValue,
5286 const SmallVectorImpl<std::pair<ConstantInt *, Constant *>> &Values) {
5287 ICmpInst *CmpInst = dyn_cast<ICmpInst>(PhiUser);
5288 if (!CmpInst)
5289 return;
5290
5291 // We require that the compare is in the same block as the phi so that jump
5292 // threading can do its work afterwards.
5293 if (CmpInst->getParent() != PhiBlock)
5294 return;
5295
5296 Constant *CmpOp1 = dyn_cast<Constant>(CmpInst->getOperand(1));
5297 if (!CmpOp1)
5298 return;
5299
5300 Value *RangeCmp = RangeCheckBranch->getCondition();
5301 Constant *TrueConst = ConstantInt::getTrue(RangeCmp->getType());
5302 Constant *FalseConst = ConstantInt::getFalse(RangeCmp->getType());
5303
5304 // Check if the compare with the default value is constant true or false.
5305 Constant *DefaultConst = ConstantExpr::getICmp(CmpInst->getPredicate(),
5306 DefaultValue, CmpOp1, true);
5307 if (DefaultConst != TrueConst && DefaultConst != FalseConst)
5308 return;
5309
5310 // Check if the compare with the case values is distinct from the default
5311 // compare result.
5312 for (auto ValuePair : Values) {
5313 Constant *CaseConst = ConstantExpr::getICmp(CmpInst->getPredicate(),
5314 ValuePair.second, CmpOp1, true);
5315 if (!CaseConst || CaseConst == DefaultConst || isa<UndefValue>(CaseConst))
5316 return;
5317 assert((CaseConst == TrueConst || CaseConst == FalseConst) &&
5318 "Expect true or false as compare result.");
5319 }
5320
5321 // Check if the branch instruction dominates the phi node. It's a simple
5322 // dominance check, but sufficient for our needs.
5323 // Although this check is invariant in the calling loops, it's better to do it
5324 // at this late stage. Practically we do it at most once for a switch.
5325 BasicBlock *BranchBlock = RangeCheckBranch->getParent();
5326 for (auto PI = pred_begin(PhiBlock), E = pred_end(PhiBlock); PI != E; ++PI) {
5327 BasicBlock *Pred = *PI;
5328 if (Pred != BranchBlock && Pred->getUniquePredecessor() != BranchBlock)
5329 return;
5330 }
5331
5332 if (DefaultConst == FalseConst) {
5333 // The compare yields the same result. We can replace it.
5334 CmpInst->replaceAllUsesWith(RangeCmp);
5335 ++NumTableCmpReuses;
5336 } else {
5337 // The compare yields the same result, just inverted. We can replace it.
5338 Value *InvertedTableCmp = BinaryOperator::CreateXor(
5339 RangeCmp, ConstantInt::get(RangeCmp->getType(), 1), "inverted.cmp",
5340 RangeCheckBranch);
5341 CmpInst->replaceAllUsesWith(InvertedTableCmp);
5342 ++NumTableCmpReuses;
5343 }
5344 }
5345
5346 /// If the switch is only used to initialize one or more phi nodes in a common
5347 /// successor block with different constant values, replace the switch with
5348 /// lookup tables.
SwitchToLookupTable(SwitchInst * SI,IRBuilder<> & Builder,const DataLayout & DL,const TargetTransformInfo & TTI)5349 static bool SwitchToLookupTable(SwitchInst *SI, IRBuilder<> &Builder,
5350 const DataLayout &DL,
5351 const TargetTransformInfo &TTI) {
5352 assert(SI->getNumCases() > 1 && "Degenerate switch?");
5353
5354 Function *Fn = SI->getParent()->getParent();
5355 // Only build lookup table when we have a target that supports it or the
5356 // attribute is not set.
5357 if (!TTI.shouldBuildLookupTables() ||
5358 (Fn->getFnAttribute("no-jump-tables").getValueAsString() == "true"))
5359 return false;
5360
5361 // FIXME: If the switch is too sparse for a lookup table, perhaps we could
5362 // split off a dense part and build a lookup table for that.
5363
5364 // FIXME: This creates arrays of GEPs to constant strings, which means each
5365 // GEP needs a runtime relocation in PIC code. We should just build one big
5366 // string and lookup indices into that.
5367
5368 // Ignore switches with less than three cases. Lookup tables will not make
5369 // them faster, so we don't analyze them.
5370 if (SI->getNumCases() < 3)
5371 return false;
5372
5373 // Figure out the corresponding result for each case value and phi node in the
5374 // common destination, as well as the min and max case values.
5375 assert(!SI->cases().empty());
5376 SwitchInst::CaseIt CI = SI->case_begin();
5377 ConstantInt *MinCaseVal = CI->getCaseValue();
5378 ConstantInt *MaxCaseVal = CI->getCaseValue();
5379
5380 BasicBlock *CommonDest = nullptr;
5381
5382 using ResultListTy = SmallVector<std::pair<ConstantInt *, Constant *>, 4>;
5383 SmallDenseMap<PHINode *, ResultListTy> ResultLists;
5384
5385 SmallDenseMap<PHINode *, Constant *> DefaultResults;
5386 SmallDenseMap<PHINode *, Type *> ResultTypes;
5387 SmallVector<PHINode *, 4> PHIs;
5388
5389 for (SwitchInst::CaseIt E = SI->case_end(); CI != E; ++CI) {
5390 ConstantInt *CaseVal = CI->getCaseValue();
5391 if (CaseVal->getValue().slt(MinCaseVal->getValue()))
5392 MinCaseVal = CaseVal;
5393 if (CaseVal->getValue().sgt(MaxCaseVal->getValue()))
5394 MaxCaseVal = CaseVal;
5395
5396 // Resulting value at phi nodes for this case value.
5397 using ResultsTy = SmallVector<std::pair<PHINode *, Constant *>, 4>;
5398 ResultsTy Results;
5399 if (!GetCaseResults(SI, CaseVal, CI->getCaseSuccessor(), &CommonDest,
5400 Results, DL, TTI))
5401 return false;
5402
5403 // Append the result from this case to the list for each phi.
5404 for (const auto &I : Results) {
5405 PHINode *PHI = I.first;
5406 Constant *Value = I.second;
5407 if (!ResultLists.count(PHI))
5408 PHIs.push_back(PHI);
5409 ResultLists[PHI].push_back(std::make_pair(CaseVal, Value));
5410 }
5411 }
5412
5413 // Keep track of the result types.
5414 for (PHINode *PHI : PHIs) {
5415 ResultTypes[PHI] = ResultLists[PHI][0].second->getType();
5416 }
5417
5418 uint64_t NumResults = ResultLists[PHIs[0]].size();
5419 APInt RangeSpread = MaxCaseVal->getValue() - MinCaseVal->getValue();
5420 uint64_t TableSize = RangeSpread.getLimitedValue() + 1;
5421 bool TableHasHoles = (NumResults < TableSize);
5422
5423 // If the table has holes, we need a constant result for the default case
5424 // or a bitmask that fits in a register.
5425 SmallVector<std::pair<PHINode *, Constant *>, 4> DefaultResultsList;
5426 bool HasDefaultResults =
5427 GetCaseResults(SI, nullptr, SI->getDefaultDest(), &CommonDest,
5428 DefaultResultsList, DL, TTI);
5429
5430 bool NeedMask = (TableHasHoles && !HasDefaultResults);
5431 if (NeedMask) {
5432 // As an extra penalty for the validity test we require more cases.
5433 if (SI->getNumCases() < 4) // FIXME: Find best threshold value (benchmark).
5434 return false;
5435 if (!DL.fitsInLegalInteger(TableSize))
5436 return false;
5437 }
5438
5439 for (const auto &I : DefaultResultsList) {
5440 PHINode *PHI = I.first;
5441 Constant *Result = I.second;
5442 DefaultResults[PHI] = Result;
5443 }
5444
5445 if (!ShouldBuildLookupTable(SI, TableSize, TTI, DL, ResultTypes))
5446 return false;
5447
5448 // Create the BB that does the lookups.
5449 Module &Mod = *CommonDest->getParent()->getParent();
5450 BasicBlock *LookupBB = BasicBlock::Create(
5451 Mod.getContext(), "switch.lookup", CommonDest->getParent(), CommonDest);
5452
5453 // Compute the table index value.
5454 Builder.SetInsertPoint(SI);
5455 Value *TableIndex;
5456 if (MinCaseVal->isNullValue())
5457 TableIndex = SI->getCondition();
5458 else
5459 TableIndex = Builder.CreateSub(SI->getCondition(), MinCaseVal,
5460 "switch.tableidx");
5461
5462 // Compute the maximum table size representable by the integer type we are
5463 // switching upon.
5464 unsigned CaseSize = MinCaseVal->getType()->getPrimitiveSizeInBits();
5465 uint64_t MaxTableSize = CaseSize > 63 ? UINT64_MAX : 1ULL << CaseSize;
5466 assert(MaxTableSize >= TableSize &&
5467 "It is impossible for a switch to have more entries than the max "
5468 "representable value of its input integer type's size.");
5469
5470 // If the default destination is unreachable, or if the lookup table covers
5471 // all values of the conditional variable, branch directly to the lookup table
5472 // BB. Otherwise, check that the condition is within the case range.
5473 const bool DefaultIsReachable =
5474 !isa<UnreachableInst>(SI->getDefaultDest()->getFirstNonPHIOrDbg());
5475 const bool GeneratingCoveredLookupTable = (MaxTableSize == TableSize);
5476 BranchInst *RangeCheckBranch = nullptr;
5477
5478 if (!DefaultIsReachable || GeneratingCoveredLookupTable) {
5479 Builder.CreateBr(LookupBB);
5480 // Note: We call removeProdecessor later since we need to be able to get the
5481 // PHI value for the default case in case we're using a bit mask.
5482 } else {
5483 Value *Cmp = Builder.CreateICmpULT(
5484 TableIndex, ConstantInt::get(MinCaseVal->getType(), TableSize));
5485 RangeCheckBranch =
5486 Builder.CreateCondBr(Cmp, LookupBB, SI->getDefaultDest());
5487 }
5488
5489 // Populate the BB that does the lookups.
5490 Builder.SetInsertPoint(LookupBB);
5491
5492 if (NeedMask) {
5493 // Before doing the lookup, we do the hole check. The LookupBB is therefore
5494 // re-purposed to do the hole check, and we create a new LookupBB.
5495 BasicBlock *MaskBB = LookupBB;
5496 MaskBB->setName("switch.hole_check");
5497 LookupBB = BasicBlock::Create(Mod.getContext(), "switch.lookup",
5498 CommonDest->getParent(), CommonDest);
5499
5500 // Make the mask's bitwidth at least 8-bit and a power-of-2 to avoid
5501 // unnecessary illegal types.
5502 uint64_t TableSizePowOf2 = NextPowerOf2(std::max(7ULL, TableSize - 1ULL));
5503 APInt MaskInt(TableSizePowOf2, 0);
5504 APInt One(TableSizePowOf2, 1);
5505 // Build bitmask; fill in a 1 bit for every case.
5506 const ResultListTy &ResultList = ResultLists[PHIs[0]];
5507 for (size_t I = 0, E = ResultList.size(); I != E; ++I) {
5508 uint64_t Idx = (ResultList[I].first->getValue() - MinCaseVal->getValue())
5509 .getLimitedValue();
5510 MaskInt |= One << Idx;
5511 }
5512 ConstantInt *TableMask = ConstantInt::get(Mod.getContext(), MaskInt);
5513
5514 // Get the TableIndex'th bit of the bitmask.
5515 // If this bit is 0 (meaning hole) jump to the default destination,
5516 // else continue with table lookup.
5517 IntegerType *MapTy = TableMask->getType();
5518 Value *MaskIndex =
5519 Builder.CreateZExtOrTrunc(TableIndex, MapTy, "switch.maskindex");
5520 Value *Shifted = Builder.CreateLShr(TableMask, MaskIndex, "switch.shifted");
5521 Value *LoBit = Builder.CreateTrunc(
5522 Shifted, Type::getInt1Ty(Mod.getContext()), "switch.lobit");
5523 Builder.CreateCondBr(LoBit, LookupBB, SI->getDefaultDest());
5524
5525 Builder.SetInsertPoint(LookupBB);
5526 AddPredecessorToBlock(SI->getDefaultDest(), MaskBB, SI->getParent());
5527 }
5528
5529 if (!DefaultIsReachable || GeneratingCoveredLookupTable) {
5530 // We cached PHINodes in PHIs. To avoid accessing deleted PHINodes later,
5531 // do not delete PHINodes here.
5532 SI->getDefaultDest()->removePredecessor(SI->getParent(),
5533 /*KeepOneInputPHIs=*/true);
5534 }
5535
5536 bool ReturnedEarly = false;
5537 for (PHINode *PHI : PHIs) {
5538 const ResultListTy &ResultList = ResultLists[PHI];
5539
5540 // If using a bitmask, use any value to fill the lookup table holes.
5541 Constant *DV = NeedMask ? ResultLists[PHI][0].second : DefaultResults[PHI];
5542 StringRef FuncName = Fn->getName();
5543 SwitchLookupTable Table(Mod, TableSize, MinCaseVal, ResultList, DV, DL,
5544 FuncName);
5545
5546 Value *Result = Table.BuildLookup(TableIndex, Builder);
5547
5548 // If the result is used to return immediately from the function, we want to
5549 // do that right here.
5550 if (PHI->hasOneUse() && isa<ReturnInst>(*PHI->user_begin()) &&
5551 PHI->user_back() == CommonDest->getFirstNonPHIOrDbg()) {
5552 Builder.CreateRet(Result);
5553 ReturnedEarly = true;
5554 break;
5555 }
5556
5557 // Do a small peephole optimization: re-use the switch table compare if
5558 // possible.
5559 if (!TableHasHoles && HasDefaultResults && RangeCheckBranch) {
5560 BasicBlock *PhiBlock = PHI->getParent();
5561 // Search for compare instructions which use the phi.
5562 for (auto *User : PHI->users()) {
5563 reuseTableCompare(User, PhiBlock, RangeCheckBranch, DV, ResultList);
5564 }
5565 }
5566
5567 PHI->addIncoming(Result, LookupBB);
5568 }
5569
5570 if (!ReturnedEarly)
5571 Builder.CreateBr(CommonDest);
5572
5573 // Remove the switch.
5574 for (unsigned i = 0, e = SI->getNumSuccessors(); i < e; ++i) {
5575 BasicBlock *Succ = SI->getSuccessor(i);
5576
5577 if (Succ == SI->getDefaultDest())
5578 continue;
5579 Succ->removePredecessor(SI->getParent());
5580 }
5581 SI->eraseFromParent();
5582
5583 ++NumLookupTables;
5584 if (NeedMask)
5585 ++NumLookupTablesHoles;
5586 return true;
5587 }
5588
isSwitchDense(ArrayRef<int64_t> Values)5589 static bool isSwitchDense(ArrayRef<int64_t> Values) {
5590 // See also SelectionDAGBuilder::isDense(), which this function was based on.
5591 uint64_t Diff = (uint64_t)Values.back() - (uint64_t)Values.front();
5592 uint64_t Range = Diff + 1;
5593 uint64_t NumCases = Values.size();
5594 // 40% is the default density for building a jump table in optsize/minsize mode.
5595 uint64_t MinDensity = 40;
5596
5597 return NumCases * 100 >= Range * MinDensity;
5598 }
5599
5600 /// Try to transform a switch that has "holes" in it to a contiguous sequence
5601 /// of cases.
5602 ///
5603 /// A switch such as: switch(i) {case 5: case 9: case 13: case 17:} can be
5604 /// range-reduced to: switch ((i-5) / 4) {case 0: case 1: case 2: case 3:}.
5605 ///
5606 /// This converts a sparse switch into a dense switch which allows better
5607 /// lowering and could also allow transforming into a lookup table.
ReduceSwitchRange(SwitchInst * SI,IRBuilder<> & Builder,const DataLayout & DL,const TargetTransformInfo & TTI)5608 static bool ReduceSwitchRange(SwitchInst *SI, IRBuilder<> &Builder,
5609 const DataLayout &DL,
5610 const TargetTransformInfo &TTI) {
5611 auto *CondTy = cast<IntegerType>(SI->getCondition()->getType());
5612 if (CondTy->getIntegerBitWidth() > 64 ||
5613 !DL.fitsInLegalInteger(CondTy->getIntegerBitWidth()))
5614 return false;
5615 // Only bother with this optimization if there are more than 3 switch cases;
5616 // SDAG will only bother creating jump tables for 4 or more cases.
5617 if (SI->getNumCases() < 4)
5618 return false;
5619
5620 // This transform is agnostic to the signedness of the input or case values. We
5621 // can treat the case values as signed or unsigned. We can optimize more common
5622 // cases such as a sequence crossing zero {-4,0,4,8} if we interpret case values
5623 // as signed.
5624 SmallVector<int64_t,4> Values;
5625 for (auto &C : SI->cases())
5626 Values.push_back(C.getCaseValue()->getValue().getSExtValue());
5627 llvm::sort(Values);
5628
5629 // If the switch is already dense, there's nothing useful to do here.
5630 if (isSwitchDense(Values))
5631 return false;
5632
5633 // First, transform the values such that they start at zero and ascend.
5634 int64_t Base = Values[0];
5635 for (auto &V : Values)
5636 V -= (uint64_t)(Base);
5637
5638 // Now we have signed numbers that have been shifted so that, given enough
5639 // precision, there are no negative values. Since the rest of the transform
5640 // is bitwise only, we switch now to an unsigned representation.
5641
5642 // This transform can be done speculatively because it is so cheap - it
5643 // results in a single rotate operation being inserted.
5644 // FIXME: It's possible that optimizing a switch on powers of two might also
5645 // be beneficial - flag values are often powers of two and we could use a CLZ
5646 // as the key function.
5647
5648 // countTrailingZeros(0) returns 64. As Values is guaranteed to have more than
5649 // one element and LLVM disallows duplicate cases, Shift is guaranteed to be
5650 // less than 64.
5651 unsigned Shift = 64;
5652 for (auto &V : Values)
5653 Shift = std::min(Shift, countTrailingZeros((uint64_t)V));
5654 assert(Shift < 64);
5655 if (Shift > 0)
5656 for (auto &V : Values)
5657 V = (int64_t)((uint64_t)V >> Shift);
5658
5659 if (!isSwitchDense(Values))
5660 // Transform didn't create a dense switch.
5661 return false;
5662
5663 // The obvious transform is to shift the switch condition right and emit a
5664 // check that the condition actually cleanly divided by GCD, i.e.
5665 // C & (1 << Shift - 1) == 0
5666 // inserting a new CFG edge to handle the case where it didn't divide cleanly.
5667 //
5668 // A cheaper way of doing this is a simple ROTR(C, Shift). This performs the
5669 // shift and puts the shifted-off bits in the uppermost bits. If any of these
5670 // are nonzero then the switch condition will be very large and will hit the
5671 // default case.
5672
5673 auto *Ty = cast<IntegerType>(SI->getCondition()->getType());
5674 Builder.SetInsertPoint(SI);
5675 auto *ShiftC = ConstantInt::get(Ty, Shift);
5676 auto *Sub = Builder.CreateSub(SI->getCondition(), ConstantInt::get(Ty, Base));
5677 auto *LShr = Builder.CreateLShr(Sub, ShiftC);
5678 auto *Shl = Builder.CreateShl(Sub, Ty->getBitWidth() - Shift);
5679 auto *Rot = Builder.CreateOr(LShr, Shl);
5680 SI->replaceUsesOfWith(SI->getCondition(), Rot);
5681
5682 for (auto Case : SI->cases()) {
5683 auto *Orig = Case.getCaseValue();
5684 auto Sub = Orig->getValue() - APInt(Ty->getBitWidth(), Base);
5685 Case.setValue(
5686 cast<ConstantInt>(ConstantInt::get(Ty, Sub.lshr(ShiftC->getValue()))));
5687 }
5688 return true;
5689 }
5690
SimplifySwitch(SwitchInst * SI,IRBuilder<> & Builder)5691 bool SimplifyCFGOpt::SimplifySwitch(SwitchInst *SI, IRBuilder<> &Builder) {
5692 BasicBlock *BB = SI->getParent();
5693
5694 if (isValueEqualityComparison(SI)) {
5695 // If we only have one predecessor, and if it is a branch on this value,
5696 // see if that predecessor totally determines the outcome of this switch.
5697 if (BasicBlock *OnlyPred = BB->getSinglePredecessor())
5698 if (SimplifyEqualityComparisonWithOnlyPredecessor(SI, OnlyPred, Builder))
5699 return requestResimplify();
5700
5701 Value *Cond = SI->getCondition();
5702 if (SelectInst *Select = dyn_cast<SelectInst>(Cond))
5703 if (SimplifySwitchOnSelect(SI, Select))
5704 return requestResimplify();
5705
5706 // If the block only contains the switch, see if we can fold the block
5707 // away into any preds.
5708 if (SI == &*BB->instructionsWithoutDebug().begin())
5709 if (FoldValueComparisonIntoPredecessors(SI, Builder))
5710 return requestResimplify();
5711 }
5712
5713 // Try to transform the switch into an icmp and a branch.
5714 if (TurnSwitchRangeIntoICmp(SI, Builder))
5715 return requestResimplify();
5716
5717 // Remove unreachable cases.
5718 if (eliminateDeadSwitchCases(SI, Options.AC, DL))
5719 return requestResimplify();
5720
5721 if (switchToSelect(SI, Builder, DL, TTI))
5722 return requestResimplify();
5723
5724 if (Options.ForwardSwitchCondToPhi && ForwardSwitchConditionToPHI(SI))
5725 return requestResimplify();
5726
5727 // The conversion from switch to lookup tables results in difficult-to-analyze
5728 // code and makes pruning branches much harder. This is a problem if the
5729 // switch expression itself can still be restricted as a result of inlining or
5730 // CVP. Therefore, only apply this transformation during late stages of the
5731 // optimisation pipeline.
5732 if (Options.ConvertSwitchToLookupTable &&
5733 SwitchToLookupTable(SI, Builder, DL, TTI))
5734 return requestResimplify();
5735
5736 if (ReduceSwitchRange(SI, Builder, DL, TTI))
5737 return requestResimplify();
5738
5739 return false;
5740 }
5741
SimplifyIndirectBr(IndirectBrInst * IBI)5742 bool SimplifyCFGOpt::SimplifyIndirectBr(IndirectBrInst *IBI) {
5743 BasicBlock *BB = IBI->getParent();
5744 bool Changed = false;
5745
5746 // Eliminate redundant destinations.
5747 SmallPtrSet<Value *, 8> Succs;
5748 for (unsigned i = 0, e = IBI->getNumDestinations(); i != e; ++i) {
5749 BasicBlock *Dest = IBI->getDestination(i);
5750 if (!Dest->hasAddressTaken() || !Succs.insert(Dest).second) {
5751 Dest->removePredecessor(BB);
5752 IBI->removeDestination(i);
5753 --i;
5754 --e;
5755 Changed = true;
5756 }
5757 }
5758
5759 if (IBI->getNumDestinations() == 0) {
5760 // If the indirectbr has no successors, change it to unreachable.
5761 new UnreachableInst(IBI->getContext(), IBI);
5762 EraseTerminatorAndDCECond(IBI);
5763 return true;
5764 }
5765
5766 if (IBI->getNumDestinations() == 1) {
5767 // If the indirectbr has one successor, change it to a direct branch.
5768 BranchInst::Create(IBI->getDestination(0), IBI);
5769 EraseTerminatorAndDCECond(IBI);
5770 return true;
5771 }
5772
5773 if (SelectInst *SI = dyn_cast<SelectInst>(IBI->getAddress())) {
5774 if (SimplifyIndirectBrOnSelect(IBI, SI))
5775 return requestResimplify();
5776 }
5777 return Changed;
5778 }
5779
5780 /// Given an block with only a single landing pad and a unconditional branch
5781 /// try to find another basic block which this one can be merged with. This
5782 /// handles cases where we have multiple invokes with unique landing pads, but
5783 /// a shared handler.
5784 ///
5785 /// We specifically choose to not worry about merging non-empty blocks
5786 /// here. That is a PRE/scheduling problem and is best solved elsewhere. In
5787 /// practice, the optimizer produces empty landing pad blocks quite frequently
5788 /// when dealing with exception dense code. (see: instcombine, gvn, if-else
5789 /// sinking in this file)
5790 ///
5791 /// This is primarily a code size optimization. We need to avoid performing
5792 /// any transform which might inhibit optimization (such as our ability to
5793 /// specialize a particular handler via tail commoning). We do this by not
5794 /// merging any blocks which require us to introduce a phi. Since the same
5795 /// values are flowing through both blocks, we don't lose any ability to
5796 /// specialize. If anything, we make such specialization more likely.
5797 ///
5798 /// TODO - This transformation could remove entries from a phi in the target
5799 /// block when the inputs in the phi are the same for the two blocks being
5800 /// merged. In some cases, this could result in removal of the PHI entirely.
TryToMergeLandingPad(LandingPadInst * LPad,BranchInst * BI,BasicBlock * BB)5801 static bool TryToMergeLandingPad(LandingPadInst *LPad, BranchInst *BI,
5802 BasicBlock *BB) {
5803 auto Succ = BB->getUniqueSuccessor();
5804 assert(Succ);
5805 // If there's a phi in the successor block, we'd likely have to introduce
5806 // a phi into the merged landing pad block.
5807 if (isa<PHINode>(*Succ->begin()))
5808 return false;
5809
5810 for (BasicBlock *OtherPred : predecessors(Succ)) {
5811 if (BB == OtherPred)
5812 continue;
5813 BasicBlock::iterator I = OtherPred->begin();
5814 LandingPadInst *LPad2 = dyn_cast<LandingPadInst>(I);
5815 if (!LPad2 || !LPad2->isIdenticalTo(LPad))
5816 continue;
5817 for (++I; isa<DbgInfoIntrinsic>(I); ++I)
5818 ;
5819 BranchInst *BI2 = dyn_cast<BranchInst>(I);
5820 if (!BI2 || !BI2->isIdenticalTo(BI))
5821 continue;
5822
5823 // We've found an identical block. Update our predecessors to take that
5824 // path instead and make ourselves dead.
5825 SmallPtrSet<BasicBlock *, 16> Preds;
5826 Preds.insert(pred_begin(BB), pred_end(BB));
5827 for (BasicBlock *Pred : Preds) {
5828 InvokeInst *II = cast<InvokeInst>(Pred->getTerminator());
5829 assert(II->getNormalDest() != BB && II->getUnwindDest() == BB &&
5830 "unexpected successor");
5831 II->setUnwindDest(OtherPred);
5832 }
5833
5834 // The debug info in OtherPred doesn't cover the merged control flow that
5835 // used to go through BB. We need to delete it or update it.
5836 for (auto I = OtherPred->begin(), E = OtherPred->end(); I != E;) {
5837 Instruction &Inst = *I;
5838 I++;
5839 if (isa<DbgInfoIntrinsic>(Inst))
5840 Inst.eraseFromParent();
5841 }
5842
5843 SmallPtrSet<BasicBlock *, 16> Succs;
5844 Succs.insert(succ_begin(BB), succ_end(BB));
5845 for (BasicBlock *Succ : Succs) {
5846 Succ->removePredecessor(BB);
5847 }
5848
5849 IRBuilder<> Builder(BI);
5850 Builder.CreateUnreachable();
5851 BI->eraseFromParent();
5852 return true;
5853 }
5854 return false;
5855 }
5856
SimplifyUncondBranch(BranchInst * BI,IRBuilder<> & Builder)5857 bool SimplifyCFGOpt::SimplifyUncondBranch(BranchInst *BI,
5858 IRBuilder<> &Builder) {
5859 BasicBlock *BB = BI->getParent();
5860 BasicBlock *Succ = BI->getSuccessor(0);
5861
5862 // If the Terminator is the only non-phi instruction, simplify the block.
5863 // If LoopHeader is provided, check if the block or its successor is a loop
5864 // header. (This is for early invocations before loop simplify and
5865 // vectorization to keep canonical loop forms for nested loops. These blocks
5866 // can be eliminated when the pass is invoked later in the back-end.)
5867 // Note that if BB has only one predecessor then we do not introduce new
5868 // backedge, so we can eliminate BB.
5869 bool NeedCanonicalLoop =
5870 Options.NeedCanonicalLoop &&
5871 (LoopHeaders && BB->hasNPredecessorsOrMore(2) &&
5872 (LoopHeaders->count(BB) || LoopHeaders->count(Succ)));
5873 BasicBlock::iterator I = BB->getFirstNonPHIOrDbg()->getIterator();
5874 if (I->isTerminator() && BB != &BB->getParent()->getEntryBlock() &&
5875 !NeedCanonicalLoop && TryToSimplifyUncondBranchFromEmptyBlock(BB))
5876 return true;
5877
5878 // If the only instruction in the block is a seteq/setne comparison against a
5879 // constant, try to simplify the block.
5880 if (ICmpInst *ICI = dyn_cast<ICmpInst>(I))
5881 if (ICI->isEquality() && isa<ConstantInt>(ICI->getOperand(1))) {
5882 for (++I; isa<DbgInfoIntrinsic>(I); ++I)
5883 ;
5884 if (I->isTerminator() &&
5885 tryToSimplifyUncondBranchWithICmpInIt(ICI, Builder))
5886 return true;
5887 }
5888
5889 // See if we can merge an empty landing pad block with another which is
5890 // equivalent.
5891 if (LandingPadInst *LPad = dyn_cast<LandingPadInst>(I)) {
5892 for (++I; isa<DbgInfoIntrinsic>(I); ++I)
5893 ;
5894 if (I->isTerminator() && TryToMergeLandingPad(LPad, BI, BB))
5895 return true;
5896 }
5897
5898 // If this basic block is ONLY a compare and a branch, and if a predecessor
5899 // branches to us and our successor, fold the comparison into the
5900 // predecessor and use logical operations to update the incoming value
5901 // for PHI nodes in common successor.
5902 if (FoldBranchToCommonDest(BI, nullptr, Options.BonusInstThreshold))
5903 return requestResimplify();
5904 return false;
5905 }
5906
allPredecessorsComeFromSameSource(BasicBlock * BB)5907 static BasicBlock *allPredecessorsComeFromSameSource(BasicBlock *BB) {
5908 BasicBlock *PredPred = nullptr;
5909 for (auto *P : predecessors(BB)) {
5910 BasicBlock *PPred = P->getSinglePredecessor();
5911 if (!PPred || (PredPred && PredPred != PPred))
5912 return nullptr;
5913 PredPred = PPred;
5914 }
5915 return PredPred;
5916 }
5917
SimplifyCondBranch(BranchInst * BI,IRBuilder<> & Builder)5918 bool SimplifyCFGOpt::SimplifyCondBranch(BranchInst *BI, IRBuilder<> &Builder) {
5919 BasicBlock *BB = BI->getParent();
5920 const Function *Fn = BB->getParent();
5921 if (Fn && Fn->hasFnAttribute(Attribute::OptForFuzzing))
5922 return false;
5923
5924 // Conditional branch
5925 if (isValueEqualityComparison(BI)) {
5926 // If we only have one predecessor, and if it is a branch on this value,
5927 // see if that predecessor totally determines the outcome of this
5928 // switch.
5929 if (BasicBlock *OnlyPred = BB->getSinglePredecessor())
5930 if (SimplifyEqualityComparisonWithOnlyPredecessor(BI, OnlyPred, Builder))
5931 return requestResimplify();
5932
5933 // This block must be empty, except for the setcond inst, if it exists.
5934 // Ignore dbg intrinsics.
5935 auto I = BB->instructionsWithoutDebug().begin();
5936 if (&*I == BI) {
5937 if (FoldValueComparisonIntoPredecessors(BI, Builder))
5938 return requestResimplify();
5939 } else if (&*I == cast<Instruction>(BI->getCondition())) {
5940 ++I;
5941 if (&*I == BI && FoldValueComparisonIntoPredecessors(BI, Builder))
5942 return requestResimplify();
5943 }
5944 }
5945
5946 // Try to turn "br (X == 0 | X == 1), T, F" into a switch instruction.
5947 if (SimplifyBranchOnICmpChain(BI, Builder, DL))
5948 return true;
5949
5950 // If this basic block has dominating predecessor blocks and the dominating
5951 // blocks' conditions imply BI's condition, we know the direction of BI.
5952 Optional<bool> Imp = isImpliedByDomCondition(BI->getCondition(), BI, DL);
5953 if (Imp) {
5954 // Turn this into a branch on constant.
5955 auto *OldCond = BI->getCondition();
5956 ConstantInt *TorF = *Imp ? ConstantInt::getTrue(BB->getContext())
5957 : ConstantInt::getFalse(BB->getContext());
5958 BI->setCondition(TorF);
5959 RecursivelyDeleteTriviallyDeadInstructions(OldCond);
5960 return requestResimplify();
5961 }
5962
5963 // If this basic block is ONLY a compare and a branch, and if a predecessor
5964 // branches to us and one of our successors, fold the comparison into the
5965 // predecessor and use logical operations to pick the right destination.
5966 if (FoldBranchToCommonDest(BI, nullptr, Options.BonusInstThreshold))
5967 return requestResimplify();
5968
5969 // We have a conditional branch to two blocks that are only reachable
5970 // from BI. We know that the condbr dominates the two blocks, so see if
5971 // there is any identical code in the "then" and "else" blocks. If so, we
5972 // can hoist it up to the branching block.
5973 if (BI->getSuccessor(0)->getSinglePredecessor()) {
5974 if (BI->getSuccessor(1)->getSinglePredecessor()) {
5975 if (HoistThenElseCodeToIf(BI, TTI))
5976 return requestResimplify();
5977 } else {
5978 // If Successor #1 has multiple preds, we may be able to conditionally
5979 // execute Successor #0 if it branches to Successor #1.
5980 Instruction *Succ0TI = BI->getSuccessor(0)->getTerminator();
5981 if (Succ0TI->getNumSuccessors() == 1 &&
5982 Succ0TI->getSuccessor(0) == BI->getSuccessor(1))
5983 if (SpeculativelyExecuteBB(BI, BI->getSuccessor(0), TTI))
5984 return requestResimplify();
5985 }
5986 } else if (BI->getSuccessor(1)->getSinglePredecessor()) {
5987 // If Successor #0 has multiple preds, we may be able to conditionally
5988 // execute Successor #1 if it branches to Successor #0.
5989 Instruction *Succ1TI = BI->getSuccessor(1)->getTerminator();
5990 if (Succ1TI->getNumSuccessors() == 1 &&
5991 Succ1TI->getSuccessor(0) == BI->getSuccessor(0))
5992 if (SpeculativelyExecuteBB(BI, BI->getSuccessor(1), TTI))
5993 return requestResimplify();
5994 }
5995
5996 // If this is a branch on a phi node in the current block, thread control
5997 // through this block if any PHI node entries are constants.
5998 if (PHINode *PN = dyn_cast<PHINode>(BI->getCondition()))
5999 if (PN->getParent() == BI->getParent())
6000 if (FoldCondBranchOnPHI(BI, DL, Options.AC))
6001 return requestResimplify();
6002
6003 // Scan predecessor blocks for conditional branches.
6004 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
6005 if (BranchInst *PBI = dyn_cast<BranchInst>((*PI)->getTerminator()))
6006 if (PBI != BI && PBI->isConditional())
6007 if (SimplifyCondBranchToCondBranch(PBI, BI, DL, TTI))
6008 return requestResimplify();
6009
6010 // Look for diamond patterns.
6011 if (MergeCondStores)
6012 if (BasicBlock *PrevBB = allPredecessorsComeFromSameSource(BB))
6013 if (BranchInst *PBI = dyn_cast<BranchInst>(PrevBB->getTerminator()))
6014 if (PBI != BI && PBI->isConditional())
6015 if (mergeConditionalStores(PBI, BI, DL, TTI))
6016 return requestResimplify();
6017
6018 return false;
6019 }
6020
6021 /// Check if passing a value to an instruction will cause undefined behavior.
passingValueIsAlwaysUndefined(Value * V,Instruction * I)6022 static bool passingValueIsAlwaysUndefined(Value *V, Instruction *I) {
6023 Constant *C = dyn_cast<Constant>(V);
6024 if (!C)
6025 return false;
6026
6027 if (I->use_empty())
6028 return false;
6029
6030 if (C->isNullValue() || isa<UndefValue>(C)) {
6031 // Only look at the first use, avoid hurting compile time with long uselists
6032 User *Use = *I->user_begin();
6033
6034 // Now make sure that there are no instructions in between that can alter
6035 // control flow (eg. calls)
6036 for (BasicBlock::iterator
6037 i = ++BasicBlock::iterator(I),
6038 UI = BasicBlock::iterator(dyn_cast<Instruction>(Use));
6039 i != UI; ++i)
6040 if (i == I->getParent()->end() || i->mayHaveSideEffects())
6041 return false;
6042
6043 // Look through GEPs. A load from a GEP derived from NULL is still undefined
6044 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Use))
6045 if (GEP->getPointerOperand() == I)
6046 return passingValueIsAlwaysUndefined(V, GEP);
6047
6048 // Look through bitcasts.
6049 if (BitCastInst *BC = dyn_cast<BitCastInst>(Use))
6050 return passingValueIsAlwaysUndefined(V, BC);
6051
6052 // Load from null is undefined.
6053 if (LoadInst *LI = dyn_cast<LoadInst>(Use))
6054 if (!LI->isVolatile())
6055 return !NullPointerIsDefined(LI->getFunction(),
6056 LI->getPointerAddressSpace());
6057
6058 // Store to null is undefined.
6059 if (StoreInst *SI = dyn_cast<StoreInst>(Use))
6060 if (!SI->isVolatile())
6061 return (!NullPointerIsDefined(SI->getFunction(),
6062 SI->getPointerAddressSpace())) &&
6063 SI->getPointerOperand() == I;
6064
6065 // A call to null is undefined.
6066 if (auto CS = CallSite(Use))
6067 return !NullPointerIsDefined(CS->getFunction()) &&
6068 CS.getCalledValue() == I;
6069 }
6070 return false;
6071 }
6072
6073 /// If BB has an incoming value that will always trigger undefined behavior
6074 /// (eg. null pointer dereference), remove the branch leading here.
removeUndefIntroducingPredecessor(BasicBlock * BB)6075 static bool removeUndefIntroducingPredecessor(BasicBlock *BB) {
6076 for (PHINode &PHI : BB->phis())
6077 for (unsigned i = 0, e = PHI.getNumIncomingValues(); i != e; ++i)
6078 if (passingValueIsAlwaysUndefined(PHI.getIncomingValue(i), &PHI)) {
6079 Instruction *T = PHI.getIncomingBlock(i)->getTerminator();
6080 IRBuilder<> Builder(T);
6081 if (BranchInst *BI = dyn_cast<BranchInst>(T)) {
6082 BB->removePredecessor(PHI.getIncomingBlock(i));
6083 // Turn uncoditional branches into unreachables and remove the dead
6084 // destination from conditional branches.
6085 if (BI->isUnconditional())
6086 Builder.CreateUnreachable();
6087 else
6088 Builder.CreateBr(BI->getSuccessor(0) == BB ? BI->getSuccessor(1)
6089 : BI->getSuccessor(0));
6090 BI->eraseFromParent();
6091 return true;
6092 }
6093 // TODO: SwitchInst.
6094 }
6095
6096 return false;
6097 }
6098
simplifyOnce(BasicBlock * BB)6099 bool SimplifyCFGOpt::simplifyOnce(BasicBlock *BB) {
6100 bool Changed = false;
6101
6102 assert(BB && BB->getParent() && "Block not embedded in function!");
6103 assert(BB->getTerminator() && "Degenerate basic block encountered!");
6104
6105 // Remove basic blocks that have no predecessors (except the entry block)...
6106 // or that just have themself as a predecessor. These are unreachable.
6107 if ((pred_empty(BB) && BB != &BB->getParent()->getEntryBlock()) ||
6108 BB->getSinglePredecessor() == BB) {
6109 LLVM_DEBUG(dbgs() << "Removing BB: \n" << *BB);
6110 DeleteDeadBlock(BB);
6111 return true;
6112 }
6113
6114 // Check to see if we can constant propagate this terminator instruction
6115 // away...
6116 Changed |= ConstantFoldTerminator(BB, true);
6117
6118 // Check for and eliminate duplicate PHI nodes in this block.
6119 Changed |= EliminateDuplicatePHINodes(BB);
6120
6121 // Check for and remove branches that will always cause undefined behavior.
6122 Changed |= removeUndefIntroducingPredecessor(BB);
6123
6124 // Merge basic blocks into their predecessor if there is only one distinct
6125 // pred, and if there is only one distinct successor of the predecessor, and
6126 // if there are no PHI nodes.
6127 if (MergeBlockIntoPredecessor(BB))
6128 return true;
6129
6130 if (SinkCommon && Options.SinkCommonInsts)
6131 Changed |= SinkCommonCodeFromPredecessors(BB);
6132
6133 IRBuilder<> Builder(BB);
6134
6135 // If there is a trivial two-entry PHI node in this basic block, and we can
6136 // eliminate it, do so now.
6137 if (auto *PN = dyn_cast<PHINode>(BB->begin()))
6138 if (PN->getNumIncomingValues() == 2)
6139 Changed |= FoldTwoEntryPHINode(PN, TTI, DL);
6140
6141 Builder.SetInsertPoint(BB->getTerminator());
6142 if (auto *BI = dyn_cast<BranchInst>(BB->getTerminator())) {
6143 if (BI->isUnconditional()) {
6144 if (SimplifyUncondBranch(BI, Builder))
6145 return true;
6146 } else {
6147 if (SimplifyCondBranch(BI, Builder))
6148 return true;
6149 }
6150 } else if (auto *RI = dyn_cast<ReturnInst>(BB->getTerminator())) {
6151 if (SimplifyReturn(RI, Builder))
6152 return true;
6153 } else if (auto *RI = dyn_cast<ResumeInst>(BB->getTerminator())) {
6154 if (SimplifyResume(RI, Builder))
6155 return true;
6156 } else if (auto *RI = dyn_cast<CleanupReturnInst>(BB->getTerminator())) {
6157 if (SimplifyCleanupReturn(RI))
6158 return true;
6159 } else if (auto *SI = dyn_cast<SwitchInst>(BB->getTerminator())) {
6160 if (SimplifySwitch(SI, Builder))
6161 return true;
6162 } else if (auto *UI = dyn_cast<UnreachableInst>(BB->getTerminator())) {
6163 if (SimplifyUnreachable(UI))
6164 return true;
6165 } else if (auto *IBI = dyn_cast<IndirectBrInst>(BB->getTerminator())) {
6166 if (SimplifyIndirectBr(IBI))
6167 return true;
6168 }
6169
6170 return Changed;
6171 }
6172
run(BasicBlock * BB)6173 bool SimplifyCFGOpt::run(BasicBlock *BB) {
6174 bool Changed = false;
6175
6176 // Repeated simplify BB as long as resimplification is requested.
6177 do {
6178 Resimplify = false;
6179
6180 // Perform one round of simplifcation. Resimplify flag will be set if
6181 // another iteration is requested.
6182 Changed |= simplifyOnce(BB);
6183 } while (Resimplify);
6184
6185 return Changed;
6186 }
6187
simplifyCFG(BasicBlock * BB,const TargetTransformInfo & TTI,const SimplifyCFGOptions & Options,SmallPtrSetImpl<BasicBlock * > * LoopHeaders)6188 bool llvm::simplifyCFG(BasicBlock *BB, const TargetTransformInfo &TTI,
6189 const SimplifyCFGOptions &Options,
6190 SmallPtrSetImpl<BasicBlock *> *LoopHeaders) {
6191 return SimplifyCFGOpt(TTI, BB->getModule()->getDataLayout(), LoopHeaders,
6192 Options)
6193 .run(BB);
6194 }
6195