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 assert(L && "Instruction belongs to a BB that's not part of a loop");
91 if (!LoopExitBlocks.count(L))
92 L->getExitBlocks(LoopExitBlocks[L]);
93 assert(LoopExitBlocks.count(L));
94 const SmallVectorImpl<BasicBlock *> &ExitBlocks = LoopExitBlocks[L];
96 if (ExitBlocks.empty())
99 // Tokens cannot be used in PHI nodes, so we skip over them.
100 // We can run into tokens which are live out of a loop with catchswitch
101 // instructions in Windows EH if the catchswitch has one catchpad which
102 // is inside the loop and another which is not.
103 if (I->getType()->isTokenTy())
106 for (Use &U : I->uses()) {
107 Instruction *User = cast<Instruction>(U.getUser());
108 BasicBlock *UserBB = User->getParent();
109 if (auto *PN = dyn_cast<PHINode>(User))
110 UserBB = PN->getIncomingBlock(U);
112 if (InstBB != UserBB && !L->contains(UserBB))
113 UsesToRewrite.push_back(&U);
116 // If there are no uses outside the loop, exit with no change.
117 if (UsesToRewrite.empty())
120 ++NumLCSSA; // We are applying the transformation
122 // Invoke instructions are special in that their result value is not
123 // available along their unwind edge. The code below tests to see whether
124 // DomBB dominates the value, so adjust DomBB to the normal destination
125 // block, which is effectively where the value is first usable.
126 BasicBlock *DomBB = InstBB;
127 if (auto *Inv = dyn_cast<InvokeInst>(I))
128 DomBB = Inv->getNormalDest();
130 DomTreeNode *DomNode = DT.getNode(DomBB);
132 SmallVector<PHINode *, 16> AddedPHIs;
133 SmallVector<PHINode *, 8> PostProcessPHIs;
135 SmallVector<PHINode *, 4> InsertedPHIs;
136 SSAUpdater SSAUpdate(&InsertedPHIs);
137 SSAUpdate.Initialize(I->getType(), I->getName());
139 // Insert the LCSSA phi's into all of the exit blocks dominated by the
140 // value, and add them to the Phi's map.
141 for (BasicBlock *ExitBB : ExitBlocks) {
142 if (!DT.dominates(DomNode, DT.getNode(ExitBB)))
145 // If we already inserted something for this BB, don't reprocess it.
146 if (SSAUpdate.HasValueForBlock(ExitBB))
149 PHINode *PN = PHINode::Create(I->getType(), PredCache.size(ExitBB),
150 I->getName() + ".lcssa", &ExitBB->front());
152 // Add inputs from inside the loop for this PHI.
153 for (BasicBlock *Pred : PredCache.get(ExitBB)) {
154 PN->addIncoming(I, Pred);
156 // If the exit block has a predecessor not within the loop, arrange for
157 // the incoming value use corresponding to that predecessor to be
158 // rewritten in terms of a different LCSSA PHI.
159 if (!L->contains(Pred))
160 UsesToRewrite.push_back(
161 &PN->getOperandUse(PN->getOperandNumForIncomingValue(
162 PN->getNumIncomingValues() - 1)));
165 AddedPHIs.push_back(PN);
167 // Remember that this phi makes the value alive in this block.
168 SSAUpdate.AddAvailableValue(ExitBB, PN);
170 // LoopSimplify might fail to simplify some loops (e.g. when indirect
171 // branches are involved). In such situations, it might happen that an
172 // exit for Loop L1 is the header of a disjoint Loop L2. Thus, when we
173 // create PHIs in such an exit block, we are also inserting PHIs into L2's
174 // header. This could break LCSSA form for L2 because these inserted PHIs
175 // can also have uses outside of L2. Remember all PHIs in such situation
176 // as to revisit than later on. FIXME: Remove this if indirectbr support
177 // into LoopSimplify gets improved.
178 if (auto *OtherLoop = LI.getLoopFor(ExitBB))
179 if (!L->contains(OtherLoop))
180 PostProcessPHIs.push_back(PN);
183 // Rewrite all uses outside the loop in terms of the new PHIs we just
185 for (Use *UseToRewrite : UsesToRewrite) {
186 // If this use is in an exit block, rewrite to use the newly inserted PHI.
187 // This is required for correctness because SSAUpdate doesn't handle uses
188 // in the same block. It assumes the PHI we inserted is at the end of the
190 Instruction *User = cast<Instruction>(UseToRewrite->getUser());
191 BasicBlock *UserBB = User->getParent();
192 if (auto *PN = dyn_cast<PHINode>(User))
193 UserBB = PN->getIncomingBlock(*UseToRewrite);
195 if (isa<PHINode>(UserBB->begin()) && isExitBlock(UserBB, ExitBlocks)) {
196 // Tell the VHs that the uses changed. This updates SCEV's caches.
197 if (UseToRewrite->get()->hasValueHandle())
198 ValueHandleBase::ValueIsRAUWd(*UseToRewrite, &UserBB->front());
199 UseToRewrite->set(&UserBB->front());
203 // Otherwise, do full PHI insertion.
204 SSAUpdate.RewriteUse(*UseToRewrite);
207 // SSAUpdater might have inserted phi-nodes inside other loops. We'll need
208 // to post-process them to keep LCSSA form.
209 for (PHINode *InsertedPN : InsertedPHIs) {
210 if (auto *OtherLoop = LI.getLoopFor(InsertedPN->getParent()))
211 if (!L->contains(OtherLoop))
212 PostProcessPHIs.push_back(InsertedPN);
215 // Post process PHI instructions that were inserted into another disjoint
216 // loop and update their exits properly.
217 for (auto *PostProcessPN : PostProcessPHIs) {
218 if (PostProcessPN->use_empty())
221 // Reprocess each PHI instruction.
222 Worklist.push_back(PostProcessPN);
225 // Keep track of PHI nodes that we want to remove because they did not have
226 // any uses rewritten.
227 for (PHINode *PN : AddedPHIs)
229 PHIsToRemove.insert(PN);
233 // Remove PHI nodes that did not have any uses rewritten.
234 for (PHINode *PN : PHIsToRemove) {
235 assert (PN->use_empty() && "Trying to remove a phi with uses.");
236 PN->eraseFromParent();
241 // Compute the set of BasicBlocks in the loop `L` dominating at least one exit.
242 static void computeBlocksDominatingExits(
243 Loop &L, DominatorTree &DT, SmallVector<BasicBlock *, 8> &ExitBlocks,
244 SmallPtrSet<BasicBlock *, 8> &BlocksDominatingExits) {
245 SmallVector<BasicBlock *, 8> BBWorklist;
247 // We start from the exit blocks, as every block trivially dominates itself
249 for (BasicBlock *BB : ExitBlocks)
250 BBWorklist.push_back(BB);
252 while (!BBWorklist.empty()) {
253 BasicBlock *BB = BBWorklist.pop_back_val();
255 // Check if this is a loop header. If this is the case, we're done.
256 if (L.getHeader() == BB)
259 // Otherwise, add its immediate predecessor in the dominator tree to the
260 // worklist, unless we visited it already.
261 BasicBlock *IDomBB = DT.getNode(BB)->getIDom()->getBlock();
263 // Exit blocks can have an immediate dominator not beloinging to the
264 // loop. For an exit block to be immediately dominated by another block
265 // outside the loop, it implies not all paths from that dominator, to the
266 // exit block, go through the loop.
277 // C is the exit block of the loop and it's immediately dominated by A,
278 // which doesn't belong to the loop.
279 if (!L.contains(IDomBB))
282 if (BlocksDominatingExits.insert(IDomBB).second)
283 BBWorklist.push_back(IDomBB);
287 bool llvm::formLCSSA(Loop &L, DominatorTree &DT, LoopInfo *LI,
288 ScalarEvolution *SE) {
289 bool Changed = false;
291 SmallVector<BasicBlock *, 8> ExitBlocks;
292 L.getExitBlocks(ExitBlocks);
293 if (ExitBlocks.empty())
296 SmallPtrSet<BasicBlock *, 8> BlocksDominatingExits;
298 // We want to avoid use-scanning leveraging dominance informations.
299 // If a block doesn't dominate any of the loop exits, the none of the values
300 // defined in the loop can be used outside.
301 // We compute the set of blocks fullfilling the conditions in advance
302 // walking the dominator tree upwards until we hit a loop header.
303 computeBlocksDominatingExits(L, DT, ExitBlocks, BlocksDominatingExits);
305 SmallVector<Instruction *, 8> Worklist;
307 // Look at all the instructions in the loop, checking to see if they have uses
308 // outside the loop. If so, put them into the worklist to rewrite those uses.
309 for (BasicBlock *BB : BlocksDominatingExits) {
310 for (Instruction &I : *BB) {
311 // Reject two common cases fast: instructions with no uses (like stores)
312 // and instructions with one use that is in the same block as this.
314 (I.hasOneUse() && I.user_back()->getParent() == BB &&
315 !isa<PHINode>(I.user_back())))
318 Worklist.push_back(&I);
321 Changed = formLCSSAForInstructions(Worklist, DT, *LI);
323 // If we modified the code, remove any caches about the loop from SCEV to
324 // avoid dangling entries.
325 // FIXME: This is a big hammer, can we clear the cache more selectively?
329 assert(L.isLCSSAForm(DT));
334 /// Process a loop nest depth first.
335 bool llvm::formLCSSARecursively(Loop &L, DominatorTree &DT, LoopInfo *LI,
336 ScalarEvolution *SE) {
337 bool Changed = false;
339 // Recurse depth-first through inner loops.
340 for (Loop *SubLoop : L.getSubLoops())
341 Changed |= formLCSSARecursively(*SubLoop, DT, LI, SE);
343 Changed |= formLCSSA(L, DT, LI, SE);
347 /// Process all loops in the function, inner-most out.
348 static bool formLCSSAOnAllLoops(LoopInfo *LI, DominatorTree &DT,
349 ScalarEvolution *SE) {
350 bool Changed = false;
352 Changed |= formLCSSARecursively(*L, DT, LI, SE);
357 struct LCSSAWrapperPass : public FunctionPass {
358 static char ID; // Pass identification, replacement for typeid
359 LCSSAWrapperPass() : FunctionPass(ID) {
360 initializeLCSSAWrapperPassPass(*PassRegistry::getPassRegistry());
363 // Cached analysis information for the current function.
368 bool runOnFunction(Function &F) override;
369 void verifyAnalysis() const override {
370 // This check is very expensive. On the loop intensive compiles it may cause
371 // up to 10x slowdown. Currently it's disabled by default. LPPassManager
372 // always does limited form of the LCSSA verification. Similar reasoning
373 // was used for the LoopInfo verifier.
374 if (VerifyLoopLCSSA) {
377 return L->isRecursivelyLCSSAForm(*DT, *LI);
379 "LCSSA form is broken!");
383 /// This transformation requires natural loop information & requires that
384 /// loop preheaders be inserted into the CFG. It maintains both of these,
385 /// as well as the CFG. It also requires dominator information.
386 void getAnalysisUsage(AnalysisUsage &AU) const override {
387 AU.setPreservesCFG();
389 AU.addRequired<DominatorTreeWrapperPass>();
390 AU.addRequired<LoopInfoWrapperPass>();
391 AU.addPreservedID(LoopSimplifyID);
392 AU.addPreserved<AAResultsWrapperPass>();
393 AU.addPreserved<BasicAAWrapperPass>();
394 AU.addPreserved<GlobalsAAWrapperPass>();
395 AU.addPreserved<ScalarEvolutionWrapperPass>();
396 AU.addPreserved<SCEVAAWrapperPass>();
398 // This is needed to perform LCSSA verification inside LPPassManager
399 AU.addRequired<LCSSAVerificationPass>();
400 AU.addPreserved<LCSSAVerificationPass>();
405 char LCSSAWrapperPass::ID = 0;
406 INITIALIZE_PASS_BEGIN(LCSSAWrapperPass, "lcssa", "Loop-Closed SSA Form Pass",
408 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
409 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
410 INITIALIZE_PASS_DEPENDENCY(LCSSAVerificationPass)
411 INITIALIZE_PASS_END(LCSSAWrapperPass, "lcssa", "Loop-Closed SSA Form Pass",
414 Pass *llvm::createLCSSAPass() { return new LCSSAWrapperPass(); }
415 char &llvm::LCSSAID = LCSSAWrapperPass::ID;
417 /// Transform \p F into loop-closed SSA form.
418 bool LCSSAWrapperPass::runOnFunction(Function &F) {
419 LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
420 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
421 auto *SEWP = getAnalysisIfAvailable<ScalarEvolutionWrapperPass>();
422 SE = SEWP ? &SEWP->getSE() : nullptr;
424 return formLCSSAOnAllLoops(LI, *DT, SE);
427 PreservedAnalyses LCSSAPass::run(Function &F, FunctionAnalysisManager &AM) {
428 auto &LI = AM.getResult<LoopAnalysis>(F);
429 auto &DT = AM.getResult<DominatorTreeAnalysis>(F);
430 auto *SE = AM.getCachedResult<ScalarEvolutionAnalysis>(F);
431 if (!formLCSSAOnAllLoops(&LI, DT, SE))
432 return PreservedAnalyses::all();
434 PreservedAnalyses PA;
435 PA.preserveSet<CFGAnalyses>();
436 PA.preserve<BasicAA>();
437 PA.preserve<GlobalsAA>();
438 PA.preserve<SCEVAA>();
439 PA.preserve<ScalarEvolutionAnalysis>();