1 //===- CorrelatedValuePropagation.cpp - Propagate CFG-derived info --------===//
2 //
3 // The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 // This file implements the Correlated Value Propagation pass.
11 //
12 //===----------------------------------------------------------------------===//
13
14 #include "llvm/Transforms/Scalar/CorrelatedValuePropagation.h"
15 #include "llvm/Transforms/Scalar.h"
16 #include "llvm/ADT/Statistic.h"
17 #include "llvm/Analysis/GlobalsModRef.h"
18 #include "llvm/Analysis/InstructionSimplify.h"
19 #include "llvm/Analysis/LazyValueInfo.h"
20 #include "llvm/IR/CFG.h"
21 #include "llvm/IR/Constants.h"
22 #include "llvm/IR/Function.h"
23 #include "llvm/IR/Instructions.h"
24 #include "llvm/IR/Module.h"
25 #include "llvm/Pass.h"
26 #include "llvm/Support/Debug.h"
27 #include "llvm/Support/raw_ostream.h"
28 #include "llvm/Transforms/Utils/Local.h"
29 using namespace llvm;
30
31 #define DEBUG_TYPE "correlated-value-propagation"
32
33 STATISTIC(NumPhis, "Number of phis propagated");
34 STATISTIC(NumSelects, "Number of selects propagated");
35 STATISTIC(NumMemAccess, "Number of memory access targets propagated");
36 STATISTIC(NumCmps, "Number of comparisons propagated");
37 STATISTIC(NumReturns, "Number of return values propagated");
38 STATISTIC(NumDeadCases, "Number of switch cases removed");
39 STATISTIC(NumSDivs, "Number of sdiv converted to udiv");
40 STATISTIC(NumSRems, "Number of srem converted to urem");
41
42 namespace {
43 class CorrelatedValuePropagation : public FunctionPass {
44 public:
45 static char ID;
CorrelatedValuePropagation()46 CorrelatedValuePropagation(): FunctionPass(ID) {
47 initializeCorrelatedValuePropagationPass(*PassRegistry::getPassRegistry());
48 }
49
50 bool runOnFunction(Function &F) override;
51
getAnalysisUsage(AnalysisUsage & AU) const52 void getAnalysisUsage(AnalysisUsage &AU) const override {
53 AU.addRequired<LazyValueInfoWrapperPass>();
54 AU.addPreserved<GlobalsAAWrapperPass>();
55 }
56 };
57 }
58
59 char CorrelatedValuePropagation::ID = 0;
60 INITIALIZE_PASS_BEGIN(CorrelatedValuePropagation, "correlated-propagation",
61 "Value Propagation", false, false)
INITIALIZE_PASS_DEPENDENCY(LazyValueInfoWrapperPass)62 INITIALIZE_PASS_DEPENDENCY(LazyValueInfoWrapperPass)
63 INITIALIZE_PASS_END(CorrelatedValuePropagation, "correlated-propagation",
64 "Value Propagation", false, false)
65
66 // Public interface to the Value Propagation pass
67 Pass *llvm::createCorrelatedValuePropagationPass() {
68 return new CorrelatedValuePropagation();
69 }
70
processSelect(SelectInst * S,LazyValueInfo * LVI)71 static bool processSelect(SelectInst *S, LazyValueInfo *LVI) {
72 if (S->getType()->isVectorTy()) return false;
73 if (isa<Constant>(S->getOperand(0))) return false;
74
75 Constant *C = LVI->getConstant(S->getOperand(0), S->getParent(), S);
76 if (!C) return false;
77
78 ConstantInt *CI = dyn_cast<ConstantInt>(C);
79 if (!CI) return false;
80
81 Value *ReplaceWith = S->getOperand(1);
82 Value *Other = S->getOperand(2);
83 if (!CI->isOne()) std::swap(ReplaceWith, Other);
84 if (ReplaceWith == S) ReplaceWith = UndefValue::get(S->getType());
85
86 S->replaceAllUsesWith(ReplaceWith);
87 S->eraseFromParent();
88
89 ++NumSelects;
90
91 return true;
92 }
93
processPHI(PHINode * P,LazyValueInfo * LVI)94 static bool processPHI(PHINode *P, LazyValueInfo *LVI) {
95 bool Changed = false;
96
97 BasicBlock *BB = P->getParent();
98 for (unsigned i = 0, e = P->getNumIncomingValues(); i < e; ++i) {
99 Value *Incoming = P->getIncomingValue(i);
100 if (isa<Constant>(Incoming)) continue;
101
102 Value *V = LVI->getConstantOnEdge(Incoming, P->getIncomingBlock(i), BB, P);
103
104 // Look if the incoming value is a select with a scalar condition for which
105 // LVI can tells us the value. In that case replace the incoming value with
106 // the appropriate value of the select. This often allows us to remove the
107 // select later.
108 if (!V) {
109 SelectInst *SI = dyn_cast<SelectInst>(Incoming);
110 if (!SI) continue;
111
112 Value *Condition = SI->getCondition();
113 if (!Condition->getType()->isVectorTy()) {
114 if (Constant *C = LVI->getConstantOnEdge(
115 Condition, P->getIncomingBlock(i), BB, P)) {
116 if (C->isOneValue()) {
117 V = SI->getTrueValue();
118 } else if (C->isZeroValue()) {
119 V = SI->getFalseValue();
120 }
121 // Once LVI learns to handle vector types, we could also add support
122 // for vector type constants that are not all zeroes or all ones.
123 }
124 }
125
126 // Look if the select has a constant but LVI tells us that the incoming
127 // value can never be that constant. In that case replace the incoming
128 // value with the other value of the select. This often allows us to
129 // remove the select later.
130 if (!V) {
131 Constant *C = dyn_cast<Constant>(SI->getFalseValue());
132 if (!C) continue;
133
134 if (LVI->getPredicateOnEdge(ICmpInst::ICMP_EQ, SI, C,
135 P->getIncomingBlock(i), BB, P) !=
136 LazyValueInfo::False)
137 continue;
138 V = SI->getTrueValue();
139 }
140
141 DEBUG(dbgs() << "CVP: Threading PHI over " << *SI << '\n');
142 }
143
144 P->setIncomingValue(i, V);
145 Changed = true;
146 }
147
148 // FIXME: Provide TLI, DT, AT to SimplifyInstruction.
149 const DataLayout &DL = BB->getModule()->getDataLayout();
150 if (Value *V = SimplifyInstruction(P, DL)) {
151 P->replaceAllUsesWith(V);
152 P->eraseFromParent();
153 Changed = true;
154 }
155
156 if (Changed)
157 ++NumPhis;
158
159 return Changed;
160 }
161
processMemAccess(Instruction * I,LazyValueInfo * LVI)162 static bool processMemAccess(Instruction *I, LazyValueInfo *LVI) {
163 Value *Pointer = nullptr;
164 if (LoadInst *L = dyn_cast<LoadInst>(I))
165 Pointer = L->getPointerOperand();
166 else
167 Pointer = cast<StoreInst>(I)->getPointerOperand();
168
169 if (isa<Constant>(Pointer)) return false;
170
171 Constant *C = LVI->getConstant(Pointer, I->getParent(), I);
172 if (!C) return false;
173
174 ++NumMemAccess;
175 I->replaceUsesOfWith(Pointer, C);
176 return true;
177 }
178
179 /// See if LazyValueInfo's ability to exploit edge conditions or range
180 /// information is sufficient to prove this comparison. Even for local
181 /// conditions, this can sometimes prove conditions instcombine can't by
182 /// exploiting range information.
processCmp(CmpInst * C,LazyValueInfo * LVI)183 static bool processCmp(CmpInst *C, LazyValueInfo *LVI) {
184 Value *Op0 = C->getOperand(0);
185 Constant *Op1 = dyn_cast<Constant>(C->getOperand(1));
186 if (!Op1) return false;
187
188 // As a policy choice, we choose not to waste compile time on anything where
189 // the comparison is testing local values. While LVI can sometimes reason
190 // about such cases, it's not its primary purpose. We do make sure to do
191 // the block local query for uses from terminator instructions, but that's
192 // handled in the code for each terminator.
193 auto *I = dyn_cast<Instruction>(Op0);
194 if (I && I->getParent() == C->getParent())
195 return false;
196
197 LazyValueInfo::Tristate Result =
198 LVI->getPredicateAt(C->getPredicate(), Op0, Op1, C);
199 if (Result == LazyValueInfo::Unknown) return false;
200
201 ++NumCmps;
202 if (Result == LazyValueInfo::True)
203 C->replaceAllUsesWith(ConstantInt::getTrue(C->getContext()));
204 else
205 C->replaceAllUsesWith(ConstantInt::getFalse(C->getContext()));
206 C->eraseFromParent();
207
208 return true;
209 }
210
211 /// Simplify a switch instruction by removing cases which can never fire. If the
212 /// uselessness of a case could be determined locally then constant propagation
213 /// would already have figured it out. Instead, walk the predecessors and
214 /// statically evaluate cases based on information available on that edge. Cases
215 /// that cannot fire no matter what the incoming edge can safely be removed. If
216 /// a case fires on every incoming edge then the entire switch can be removed
217 /// and replaced with a branch to the case destination.
processSwitch(SwitchInst * SI,LazyValueInfo * LVI)218 static bool processSwitch(SwitchInst *SI, LazyValueInfo *LVI) {
219 Value *Cond = SI->getCondition();
220 BasicBlock *BB = SI->getParent();
221
222 // If the condition was defined in same block as the switch then LazyValueInfo
223 // currently won't say anything useful about it, though in theory it could.
224 if (isa<Instruction>(Cond) && cast<Instruction>(Cond)->getParent() == BB)
225 return false;
226
227 // If the switch is unreachable then trying to improve it is a waste of time.
228 pred_iterator PB = pred_begin(BB), PE = pred_end(BB);
229 if (PB == PE) return false;
230
231 // Analyse each switch case in turn. This is done in reverse order so that
232 // removing a case doesn't cause trouble for the iteration.
233 bool Changed = false;
234 for (SwitchInst::CaseIt CI = SI->case_end(), CE = SI->case_begin(); CI-- != CE;
235 ) {
236 ConstantInt *Case = CI.getCaseValue();
237
238 // Check to see if the switch condition is equal to/not equal to the case
239 // value on every incoming edge, equal/not equal being the same each time.
240 LazyValueInfo::Tristate State = LazyValueInfo::Unknown;
241 for (pred_iterator PI = PB; PI != PE; ++PI) {
242 // Is the switch condition equal to the case value?
243 LazyValueInfo::Tristate Value = LVI->getPredicateOnEdge(CmpInst::ICMP_EQ,
244 Cond, Case, *PI,
245 BB, SI);
246 // Give up on this case if nothing is known.
247 if (Value == LazyValueInfo::Unknown) {
248 State = LazyValueInfo::Unknown;
249 break;
250 }
251
252 // If this was the first edge to be visited, record that all other edges
253 // need to give the same result.
254 if (PI == PB) {
255 State = Value;
256 continue;
257 }
258
259 // If this case is known to fire for some edges and known not to fire for
260 // others then there is nothing we can do - give up.
261 if (Value != State) {
262 State = LazyValueInfo::Unknown;
263 break;
264 }
265 }
266
267 if (State == LazyValueInfo::False) {
268 // This case never fires - remove it.
269 CI.getCaseSuccessor()->removePredecessor(BB);
270 SI->removeCase(CI); // Does not invalidate the iterator.
271
272 // The condition can be modified by removePredecessor's PHI simplification
273 // logic.
274 Cond = SI->getCondition();
275
276 ++NumDeadCases;
277 Changed = true;
278 } else if (State == LazyValueInfo::True) {
279 // This case always fires. Arrange for the switch to be turned into an
280 // unconditional branch by replacing the switch condition with the case
281 // value.
282 SI->setCondition(Case);
283 NumDeadCases += SI->getNumCases();
284 Changed = true;
285 break;
286 }
287 }
288
289 if (Changed)
290 // If the switch has been simplified to the point where it can be replaced
291 // by a branch then do so now.
292 ConstantFoldTerminator(BB);
293
294 return Changed;
295 }
296
297 /// Infer nonnull attributes for the arguments at the specified callsite.
processCallSite(CallSite CS,LazyValueInfo * LVI)298 static bool processCallSite(CallSite CS, LazyValueInfo *LVI) {
299 SmallVector<unsigned, 4> Indices;
300 unsigned ArgNo = 0;
301
302 for (Value *V : CS.args()) {
303 PointerType *Type = dyn_cast<PointerType>(V->getType());
304 // Try to mark pointer typed parameters as non-null. We skip the
305 // relatively expensive analysis for constants which are obviously either
306 // null or non-null to start with.
307 if (Type && !CS.paramHasAttr(ArgNo + 1, Attribute::NonNull) &&
308 !isa<Constant>(V) &&
309 LVI->getPredicateAt(ICmpInst::ICMP_EQ, V,
310 ConstantPointerNull::get(Type),
311 CS.getInstruction()) == LazyValueInfo::False)
312 Indices.push_back(ArgNo + 1);
313 ArgNo++;
314 }
315
316 assert(ArgNo == CS.arg_size() && "sanity check");
317
318 if (Indices.empty())
319 return false;
320
321 AttributeSet AS = CS.getAttributes();
322 LLVMContext &Ctx = CS.getInstruction()->getContext();
323 AS = AS.addAttribute(Ctx, Indices, Attribute::get(Ctx, Attribute::NonNull));
324 CS.setAttributes(AS);
325
326 return true;
327 }
328
329 // Helper function to rewrite srem and sdiv. As a policy choice, we choose not
330 // to waste compile time on anything where the operands are local defs. While
331 // LVI can sometimes reason about such cases, it's not its primary purpose.
hasLocalDefs(BinaryOperator * SDI)332 static bool hasLocalDefs(BinaryOperator *SDI) {
333 for (Value *O : SDI->operands()) {
334 auto *I = dyn_cast<Instruction>(O);
335 if (I && I->getParent() == SDI->getParent())
336 return true;
337 }
338 return false;
339 }
340
hasPositiveOperands(BinaryOperator * SDI,LazyValueInfo * LVI)341 static bool hasPositiveOperands(BinaryOperator *SDI, LazyValueInfo *LVI) {
342 Constant *Zero = ConstantInt::get(SDI->getType(), 0);
343 for (Value *O : SDI->operands()) {
344 auto Result = LVI->getPredicateAt(ICmpInst::ICMP_SGE, O, Zero, SDI);
345 if (Result != LazyValueInfo::True)
346 return false;
347 }
348 return true;
349 }
350
processSRem(BinaryOperator * SDI,LazyValueInfo * LVI)351 static bool processSRem(BinaryOperator *SDI, LazyValueInfo *LVI) {
352 if (SDI->getType()->isVectorTy() || hasLocalDefs(SDI) ||
353 !hasPositiveOperands(SDI, LVI))
354 return false;
355
356 ++NumSRems;
357 auto *BO = BinaryOperator::CreateURem(SDI->getOperand(0), SDI->getOperand(1),
358 SDI->getName(), SDI);
359 SDI->replaceAllUsesWith(BO);
360 SDI->eraseFromParent();
361 return true;
362 }
363
364 /// See if LazyValueInfo's ability to exploit edge conditions or range
365 /// information is sufficient to prove the both operands of this SDiv are
366 /// positive. If this is the case, replace the SDiv with a UDiv. Even for local
367 /// conditions, this can sometimes prove conditions instcombine can't by
368 /// exploiting range information.
processSDiv(BinaryOperator * SDI,LazyValueInfo * LVI)369 static bool processSDiv(BinaryOperator *SDI, LazyValueInfo *LVI) {
370 if (SDI->getType()->isVectorTy() || hasLocalDefs(SDI) ||
371 !hasPositiveOperands(SDI, LVI))
372 return false;
373
374 ++NumSDivs;
375 auto *BO = BinaryOperator::CreateUDiv(SDI->getOperand(0), SDI->getOperand(1),
376 SDI->getName(), SDI);
377 BO->setIsExact(SDI->isExact());
378 SDI->replaceAllUsesWith(BO);
379 SDI->eraseFromParent();
380
381 return true;
382 }
383
getConstantAt(Value * V,Instruction * At,LazyValueInfo * LVI)384 static Constant *getConstantAt(Value *V, Instruction *At, LazyValueInfo *LVI) {
385 if (Constant *C = LVI->getConstant(V, At->getParent(), At))
386 return C;
387
388 // TODO: The following really should be sunk inside LVI's core algorithm, or
389 // at least the outer shims around such.
390 auto *C = dyn_cast<CmpInst>(V);
391 if (!C) return nullptr;
392
393 Value *Op0 = C->getOperand(0);
394 Constant *Op1 = dyn_cast<Constant>(C->getOperand(1));
395 if (!Op1) return nullptr;
396
397 LazyValueInfo::Tristate Result =
398 LVI->getPredicateAt(C->getPredicate(), Op0, Op1, At);
399 if (Result == LazyValueInfo::Unknown)
400 return nullptr;
401
402 return (Result == LazyValueInfo::True) ?
403 ConstantInt::getTrue(C->getContext()) :
404 ConstantInt::getFalse(C->getContext());
405 }
406
runImpl(Function & F,LazyValueInfo * LVI)407 static bool runImpl(Function &F, LazyValueInfo *LVI) {
408 bool FnChanged = false;
409
410 for (BasicBlock &BB : F) {
411 bool BBChanged = false;
412 for (BasicBlock::iterator BI = BB.begin(), BE = BB.end(); BI != BE;) {
413 Instruction *II = &*BI++;
414 switch (II->getOpcode()) {
415 case Instruction::Select:
416 BBChanged |= processSelect(cast<SelectInst>(II), LVI);
417 break;
418 case Instruction::PHI:
419 BBChanged |= processPHI(cast<PHINode>(II), LVI);
420 break;
421 case Instruction::ICmp:
422 case Instruction::FCmp:
423 BBChanged |= processCmp(cast<CmpInst>(II), LVI);
424 break;
425 case Instruction::Load:
426 case Instruction::Store:
427 BBChanged |= processMemAccess(II, LVI);
428 break;
429 case Instruction::Call:
430 case Instruction::Invoke:
431 BBChanged |= processCallSite(CallSite(II), LVI);
432 break;
433 case Instruction::SRem:
434 BBChanged |= processSRem(cast<BinaryOperator>(II), LVI);
435 break;
436 case Instruction::SDiv:
437 BBChanged |= processSDiv(cast<BinaryOperator>(II), LVI);
438 break;
439 }
440 }
441
442 Instruction *Term = BB.getTerminator();
443 switch (Term->getOpcode()) {
444 case Instruction::Switch:
445 BBChanged |= processSwitch(cast<SwitchInst>(Term), LVI);
446 break;
447 case Instruction::Ret: {
448 auto *RI = cast<ReturnInst>(Term);
449 // Try to determine the return value if we can. This is mainly here to
450 // simplify the writing of unit tests, but also helps to enable IPO by
451 // constant folding the return values of callees.
452 auto *RetVal = RI->getReturnValue();
453 if (!RetVal) break; // handle "ret void"
454 if (isa<Constant>(RetVal)) break; // nothing to do
455 if (auto *C = getConstantAt(RetVal, RI, LVI)) {
456 ++NumReturns;
457 RI->replaceUsesOfWith(RetVal, C);
458 BBChanged = true;
459 }
460 }
461 };
462
463 FnChanged |= BBChanged;
464 }
465
466 return FnChanged;
467 }
468
runOnFunction(Function & F)469 bool CorrelatedValuePropagation::runOnFunction(Function &F) {
470 if (skipFunction(F))
471 return false;
472
473 LazyValueInfo *LVI = &getAnalysis<LazyValueInfoWrapperPass>().getLVI();
474 return runImpl(F, LVI);
475 }
476
477 PreservedAnalyses
run(Function & F,FunctionAnalysisManager & AM)478 CorrelatedValuePropagationPass::run(Function &F, FunctionAnalysisManager &AM) {
479
480 LazyValueInfo *LVI = &AM.getResult<LazyValueAnalysis>(F);
481 bool Changed = runImpl(F, LVI);
482
483 // FIXME: We need to invalidate LVI to avoid PR28400. Is there a better
484 // solution?
485 AM.invalidate<LazyValueAnalysis>(F);
486
487 if (!Changed)
488 return PreservedAnalyses::all();
489 PreservedAnalyses PA;
490 PA.preserve<GlobalsAA>();
491 return PA;
492 }
493