1 //===-- LCSSA.cpp - Convert loops into loop-closed SSA form ---------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This pass transforms loops by placing phi nodes at the end of the loops for
11 // all values that are live across the loop boundary. For example, it turns
12 // the left into the right code:
14 // for (...) for (...)
19 // X3 = phi(X1, X2) X3 = phi(X1, X2)
20 // ... = X3 + 4 X4 = phi(X3)
23 // This is still valid LLVM; the extra phi nodes are purely redundant, and will
24 // be trivially eliminated by InstCombine. The major benefit of this
25 // transformation is that it makes many other loop optimizations, such as
26 // LoopUnswitching, simpler.
28 //===----------------------------------------------------------------------===//
30 #include "llvm/Transforms/Utils/LCSSA.h"
31 #include "llvm/ADT/STLExtras.h"
32 #include "llvm/ADT/Statistic.h"
33 #include "llvm/Analysis/AliasAnalysis.h"
34 #include "llvm/Analysis/BasicAliasAnalysis.h"
35 #include "llvm/Analysis/GlobalsModRef.h"
36 #include "llvm/Analysis/LoopPass.h"
37 #include "llvm/Analysis/ScalarEvolution.h"
38 #include "llvm/Analysis/ScalarEvolutionAliasAnalysis.h"
39 #include "llvm/IR/Constants.h"
40 #include "llvm/IR/Dominators.h"
41 #include "llvm/IR/Function.h"
42 #include "llvm/IR/Instructions.h"
43 #include "llvm/IR/PredIteratorCache.h"
44 #include "llvm/Pass.h"
45 #include "llvm/Transforms/Scalar.h"
46 #include "llvm/Transforms/Utils/LoopUtils.h"
47 #include "llvm/Transforms/Utils/SSAUpdater.h"
50 #define DEBUG_TYPE "lcssa"
52 STATISTIC(NumLCSSA, "Number of live out of a loop variables");
54 #ifdef EXPENSIVE_CHECKS
55 static bool VerifyLoopLCSSA = true;
57 static bool VerifyLoopLCSSA = false;
59 static cl::opt<bool,true>
60 VerifyLoopLCSSAFlag("verify-loop-lcssa", cl::location(VerifyLoopLCSSA),
61 cl::desc("Verify loop lcssa form (time consuming)"));
63 /// Return true if the specified block is in the list.
64 static bool isExitBlock(BasicBlock *BB,
65 const SmallVectorImpl<BasicBlock *> &ExitBlocks) {
66 return is_contained(ExitBlocks, BB);
69 /// For every instruction from the worklist, check to see if it has any uses
70 /// that are outside the current loop. If so, insert LCSSA PHI nodes and
72 bool llvm::formLCSSAForInstructions(SmallVectorImpl<Instruction *> &Worklist,
73 DominatorTree &DT, LoopInfo &LI) {
74 SmallVector<Use *, 16> UsesToRewrite;
75 SmallSetVector<PHINode *, 16> PHIsToRemove;
76 PredIteratorCache PredCache;
79 // Cache the Loop ExitBlocks across this loop. We expect to get a lot of
80 // instructions within the same loops, computing the exit blocks is
81 // expensive, and we're not mutating the loop structure.
82 SmallDenseMap<Loop*, SmallVector<BasicBlock *,1>> LoopExitBlocks;
84 while (!Worklist.empty()) {
85 UsesToRewrite.clear();
87 Instruction *I = Worklist.pop_back_val();
88 BasicBlock *InstBB = I->getParent();
89 Loop *L = LI.getLoopFor(InstBB);
90 if (!LoopExitBlocks.count(L))
91 L->getExitBlocks(LoopExitBlocks[L]);
92 assert(LoopExitBlocks.count(L));
93 const SmallVectorImpl<BasicBlock *> &ExitBlocks = LoopExitBlocks[L];
95 if (ExitBlocks.empty())
98 // Tokens cannot be used in PHI nodes, so we skip over them.
99 // We can run into tokens which are live out of a loop with catchswitch
100 // instructions in Windows EH if the catchswitch has one catchpad which
101 // is inside the loop and another which is not.
102 if (I->getType()->isTokenTy())
105 for (Use &U : I->uses()) {
106 Instruction *User = cast<Instruction>(U.getUser());
107 BasicBlock *UserBB = User->getParent();
108 if (PHINode *PN = dyn_cast<PHINode>(User))
109 UserBB = PN->getIncomingBlock(U);
111 if (InstBB != UserBB && !L->contains(UserBB))
112 UsesToRewrite.push_back(&U);
115 // If there are no uses outside the loop, exit with no change.
116 if (UsesToRewrite.empty())
119 ++NumLCSSA; // We are applying the transformation
121 // Invoke instructions are special in that their result value is not
122 // available along their unwind edge. The code below tests to see whether
123 // DomBB dominates the value, so adjust DomBB to the normal destination
124 // block, which is effectively where the value is first usable.
125 BasicBlock *DomBB = InstBB;
126 if (InvokeInst *Inv = dyn_cast<InvokeInst>(I))
127 DomBB = Inv->getNormalDest();
129 DomTreeNode *DomNode = DT.getNode(DomBB);
131 SmallVector<PHINode *, 16> AddedPHIs;
132 SmallVector<PHINode *, 8> PostProcessPHIs;
134 SmallVector<PHINode *, 4> InsertedPHIs;
135 SSAUpdater SSAUpdate(&InsertedPHIs);
136 SSAUpdate.Initialize(I->getType(), I->getName());
138 // Insert the LCSSA phi's into all of the exit blocks dominated by the
139 // value, and add them to the Phi's map.
140 for (BasicBlock *ExitBB : ExitBlocks) {
141 if (!DT.dominates(DomNode, DT.getNode(ExitBB)))
144 // If we already inserted something for this BB, don't reprocess it.
145 if (SSAUpdate.HasValueForBlock(ExitBB))
148 PHINode *PN = PHINode::Create(I->getType(), PredCache.size(ExitBB),
149 I->getName() + ".lcssa", &ExitBB->front());
151 // Add inputs from inside the loop for this PHI.
152 for (BasicBlock *Pred : PredCache.get(ExitBB)) {
153 PN->addIncoming(I, Pred);
155 // If the exit block has a predecessor not within the loop, arrange for
156 // the incoming value use corresponding to that predecessor to be
157 // rewritten in terms of a different LCSSA PHI.
158 if (!L->contains(Pred))
159 UsesToRewrite.push_back(
160 &PN->getOperandUse(PN->getOperandNumForIncomingValue(
161 PN->getNumIncomingValues() - 1)));
164 AddedPHIs.push_back(PN);
166 // Remember that this phi makes the value alive in this block.
167 SSAUpdate.AddAvailableValue(ExitBB, PN);
169 // LoopSimplify might fail to simplify some loops (e.g. when indirect
170 // branches are involved). In such situations, it might happen that an
171 // exit for Loop L1 is the header of a disjoint Loop L2. Thus, when we
172 // create PHIs in such an exit block, we are also inserting PHIs into L2's
173 // header. This could break LCSSA form for L2 because these inserted PHIs
174 // can also have uses outside of L2. Remember all PHIs in such situation
175 // as to revisit than later on. FIXME: Remove this if indirectbr support
176 // into LoopSimplify gets improved.
177 if (auto *OtherLoop = LI.getLoopFor(ExitBB))
178 if (!L->contains(OtherLoop))
179 PostProcessPHIs.push_back(PN);
182 // Rewrite all uses outside the loop in terms of the new PHIs we just
184 for (Use *UseToRewrite : UsesToRewrite) {
185 // If this use is in an exit block, rewrite to use the newly inserted PHI.
186 // This is required for correctness because SSAUpdate doesn't handle uses
187 // in the same block. It assumes the PHI we inserted is at the end of the
189 Instruction *User = cast<Instruction>(UseToRewrite->getUser());
190 BasicBlock *UserBB = User->getParent();
191 if (PHINode *PN = dyn_cast<PHINode>(User))
192 UserBB = PN->getIncomingBlock(*UseToRewrite);
194 if (isa<PHINode>(UserBB->begin()) && isExitBlock(UserBB, ExitBlocks)) {
195 // Tell the VHs that the uses changed. This updates SCEV's caches.
196 if (UseToRewrite->get()->hasValueHandle())
197 ValueHandleBase::ValueIsRAUWd(*UseToRewrite, &UserBB->front());
198 UseToRewrite->set(&UserBB->front());
202 // Otherwise, do full PHI insertion.
203 SSAUpdate.RewriteUse(*UseToRewrite);
206 // SSAUpdater might have inserted phi-nodes inside other loops. We'll need
207 // to post-process them to keep LCSSA form.
208 for (PHINode *InsertedPN : InsertedPHIs) {
209 if (auto *OtherLoop = LI.getLoopFor(InsertedPN->getParent()))
210 if (!L->contains(OtherLoop))
211 PostProcessPHIs.push_back(InsertedPN);
214 // Post process PHI instructions that were inserted into another disjoint
215 // loop and update their exits properly.
216 for (auto *PostProcessPN : PostProcessPHIs) {
217 if (PostProcessPN->use_empty())
220 // Reprocess each PHI instruction.
221 Worklist.push_back(PostProcessPN);
224 // Keep track of PHI nodes that we want to remove because they did not have
225 // any uses rewritten.
226 for (PHINode *PN : AddedPHIs)
228 PHIsToRemove.insert(PN);
232 // Remove PHI nodes that did not have any uses rewritten.
233 for (PHINode *PN : PHIsToRemove) {
234 assert (PN->use_empty() && "Trying to remove a phi with uses.");
235 PN->eraseFromParent();
240 /// Return true if the specified block dominates at least
241 /// one of the blocks in the specified list.
243 blockDominatesAnExit(BasicBlock *BB,
245 const SmallVectorImpl<BasicBlock *> &ExitBlocks) {
246 DomTreeNode *DomNode = DT.getNode(BB);
247 return any_of(ExitBlocks, [&](BasicBlock *EB) {
248 return DT.dominates(DomNode, DT.getNode(EB));
252 bool llvm::formLCSSA(Loop &L, DominatorTree &DT, LoopInfo *LI,
253 ScalarEvolution *SE) {
254 bool Changed = false;
256 // Get the set of exiting blocks.
257 SmallVector<BasicBlock *, 8> ExitBlocks;
258 L.getExitBlocks(ExitBlocks);
260 if (ExitBlocks.empty())
263 SmallVector<Instruction *, 8> Worklist;
265 // Look at all the instructions in the loop, checking to see if they have uses
266 // outside the loop. If so, put them into the worklist to rewrite those uses.
267 for (BasicBlock *BB : L.blocks()) {
268 // For large loops, avoid use-scanning by using dominance information: In
269 // particular, if a block does not dominate any of the loop exits, then none
270 // of the values defined in the block could be used outside the loop.
271 if (!blockDominatesAnExit(BB, DT, ExitBlocks))
274 for (Instruction &I : *BB) {
275 // Reject two common cases fast: instructions with no uses (like stores)
276 // and instructions with one use that is in the same block as this.
278 (I.hasOneUse() && I.user_back()->getParent() == BB &&
279 !isa<PHINode>(I.user_back())))
282 Worklist.push_back(&I);
285 Changed = formLCSSAForInstructions(Worklist, DT, *LI);
287 // If we modified the code, remove any caches about the loop from SCEV to
288 // avoid dangling entries.
289 // FIXME: This is a big hammer, can we clear the cache more selectively?
293 assert(L.isLCSSAForm(DT));
298 /// Process a loop nest depth first.
299 bool llvm::formLCSSARecursively(Loop &L, DominatorTree &DT, LoopInfo *LI,
300 ScalarEvolution *SE) {
301 bool Changed = false;
303 // Recurse depth-first through inner loops.
304 for (Loop *SubLoop : L.getSubLoops())
305 Changed |= formLCSSARecursively(*SubLoop, DT, LI, SE);
307 Changed |= formLCSSA(L, DT, LI, SE);
311 /// Process all loops in the function, inner-most out.
312 static bool formLCSSAOnAllLoops(LoopInfo *LI, DominatorTree &DT,
313 ScalarEvolution *SE) {
314 bool Changed = false;
316 Changed |= formLCSSARecursively(*L, DT, LI, SE);
321 struct LCSSAWrapperPass : public FunctionPass {
322 static char ID; // Pass identification, replacement for typeid
323 LCSSAWrapperPass() : FunctionPass(ID) {
324 initializeLCSSAWrapperPassPass(*PassRegistry::getPassRegistry());
327 // Cached analysis information for the current function.
332 bool runOnFunction(Function &F) override;
333 void verifyAnalysis() const override {
334 // This check is very expensive. On the loop intensive compiles it may cause
335 // up to 10x slowdown. Currently it's disabled by default. LPPassManager
336 // always does limited form of the LCSSA verification. Similar reasoning
337 // was used for the LoopInfo verifier.
338 if (VerifyLoopLCSSA) {
341 return L->isRecursivelyLCSSAForm(*DT, *LI);
343 "LCSSA form is broken!");
347 /// This transformation requires natural loop information & requires that
348 /// loop preheaders be inserted into the CFG. It maintains both of these,
349 /// as well as the CFG. It also requires dominator information.
350 void getAnalysisUsage(AnalysisUsage &AU) const override {
351 AU.setPreservesCFG();
353 AU.addRequired<DominatorTreeWrapperPass>();
354 AU.addRequired<LoopInfoWrapperPass>();
355 AU.addPreservedID(LoopSimplifyID);
356 AU.addPreserved<AAResultsWrapperPass>();
357 AU.addPreserved<BasicAAWrapperPass>();
358 AU.addPreserved<GlobalsAAWrapperPass>();
359 AU.addPreserved<ScalarEvolutionWrapperPass>();
360 AU.addPreserved<SCEVAAWrapperPass>();
362 // This is needed to perform LCSSA verification inside LPPassManager
363 AU.addRequired<LCSSAVerificationPass>();
364 AU.addPreserved<LCSSAVerificationPass>();
369 char LCSSAWrapperPass::ID = 0;
370 INITIALIZE_PASS_BEGIN(LCSSAWrapperPass, "lcssa", "Loop-Closed SSA Form Pass",
372 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
373 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
374 INITIALIZE_PASS_DEPENDENCY(LCSSAVerificationPass)
375 INITIALIZE_PASS_END(LCSSAWrapperPass, "lcssa", "Loop-Closed SSA Form Pass",
378 Pass *llvm::createLCSSAPass() { return new LCSSAWrapperPass(); }
379 char &llvm::LCSSAID = LCSSAWrapperPass::ID;
381 /// Transform \p F into loop-closed SSA form.
382 bool LCSSAWrapperPass::runOnFunction(Function &F) {
383 LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
384 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
385 auto *SEWP = getAnalysisIfAvailable<ScalarEvolutionWrapperPass>();
386 SE = SEWP ? &SEWP->getSE() : nullptr;
388 return formLCSSAOnAllLoops(LI, *DT, SE);
391 PreservedAnalyses LCSSAPass::run(Function &F, FunctionAnalysisManager &AM) {
392 auto &LI = AM.getResult<LoopAnalysis>(F);
393 auto &DT = AM.getResult<DominatorTreeAnalysis>(F);
394 auto *SE = AM.getCachedResult<ScalarEvolutionAnalysis>(F);
395 if (!formLCSSAOnAllLoops(&LI, DT, SE))
396 return PreservedAnalyses::all();
398 // FIXME: This should also 'preserve the CFG'.
399 PreservedAnalyses PA;
400 PA.preserve<BasicAA>();
401 PA.preserve<GlobalsAA>();
402 PA.preserve<SCEVAA>();
403 PA.preserve<ScalarEvolutionAnalysis>();