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 /// Return true if the specified block is in the list.
55 static bool isExitBlock(BasicBlock *BB,
56 const SmallVectorImpl<BasicBlock *> &ExitBlocks) {
57 return find(ExitBlocks, BB) != ExitBlocks.end();
60 /// For every instruction from the worklist, check to see if it has any uses
61 /// that are outside the current loop. If so, insert LCSSA PHI nodes and
63 bool llvm::formLCSSAForInstructions(SmallVectorImpl<Instruction *> &Worklist,
64 DominatorTree &DT, LoopInfo &LI) {
65 SmallVector<Use *, 16> UsesToRewrite;
66 SmallVector<BasicBlock *, 8> ExitBlocks;
67 SmallSetVector<PHINode *, 16> PHIsToRemove;
68 PredIteratorCache PredCache;
71 while (!Worklist.empty()) {
72 UsesToRewrite.clear();
75 Instruction *I = Worklist.pop_back_val();
76 BasicBlock *InstBB = I->getParent();
77 Loop *L = LI.getLoopFor(InstBB);
78 L->getExitBlocks(ExitBlocks);
80 if (ExitBlocks.empty())
83 // Tokens cannot be used in PHI nodes, so we skip over them.
84 // We can run into tokens which are live out of a loop with catchswitch
85 // instructions in Windows EH if the catchswitch has one catchpad which
86 // is inside the loop and another which is not.
87 if (I->getType()->isTokenTy())
90 for (Use &U : I->uses()) {
91 Instruction *User = cast<Instruction>(U.getUser());
92 BasicBlock *UserBB = User->getParent();
93 if (PHINode *PN = dyn_cast<PHINode>(User))
94 UserBB = PN->getIncomingBlock(U);
96 if (InstBB != UserBB && !L->contains(UserBB))
97 UsesToRewrite.push_back(&U);
100 // If there are no uses outside the loop, exit with no change.
101 if (UsesToRewrite.empty())
104 ++NumLCSSA; // We are applying the transformation
106 // Invoke instructions are special in that their result value is not
107 // available along their unwind edge. The code below tests to see whether
108 // DomBB dominates the value, so adjust DomBB to the normal destination
109 // block, which is effectively where the value is first usable.
110 BasicBlock *DomBB = InstBB;
111 if (InvokeInst *Inv = dyn_cast<InvokeInst>(I))
112 DomBB = Inv->getNormalDest();
114 DomTreeNode *DomNode = DT.getNode(DomBB);
116 SmallVector<PHINode *, 16> AddedPHIs;
117 SmallVector<PHINode *, 8> PostProcessPHIs;
119 SmallVector<PHINode *, 4> InsertedPHIs;
120 SSAUpdater SSAUpdate(&InsertedPHIs);
121 SSAUpdate.Initialize(I->getType(), I->getName());
123 // Insert the LCSSA phi's into all of the exit blocks dominated by the
124 // value, and add them to the Phi's map.
125 for (BasicBlock *ExitBB : ExitBlocks) {
126 if (!DT.dominates(DomNode, DT.getNode(ExitBB)))
129 // If we already inserted something for this BB, don't reprocess it.
130 if (SSAUpdate.HasValueForBlock(ExitBB))
133 PHINode *PN = PHINode::Create(I->getType(), PredCache.size(ExitBB),
134 I->getName() + ".lcssa", &ExitBB->front());
136 // Add inputs from inside the loop for this PHI.
137 for (BasicBlock *Pred : PredCache.get(ExitBB)) {
138 PN->addIncoming(I, Pred);
140 // If the exit block has a predecessor not within the loop, arrange for
141 // the incoming value use corresponding to that predecessor to be
142 // rewritten in terms of a different LCSSA PHI.
143 if (!L->contains(Pred))
144 UsesToRewrite.push_back(
145 &PN->getOperandUse(PN->getOperandNumForIncomingValue(
146 PN->getNumIncomingValues() - 1)));
149 AddedPHIs.push_back(PN);
151 // Remember that this phi makes the value alive in this block.
152 SSAUpdate.AddAvailableValue(ExitBB, PN);
154 // LoopSimplify might fail to simplify some loops (e.g. when indirect
155 // branches are involved). In such situations, it might happen that an
156 // exit for Loop L1 is the header of a disjoint Loop L2. Thus, when we
157 // create PHIs in such an exit block, we are also inserting PHIs into L2's
158 // header. This could break LCSSA form for L2 because these inserted PHIs
159 // can also have uses outside of L2. Remember all PHIs in such situation
160 // as to revisit than later on. FIXME: Remove this if indirectbr support
161 // into LoopSimplify gets improved.
162 if (auto *OtherLoop = LI.getLoopFor(ExitBB))
163 if (!L->contains(OtherLoop))
164 PostProcessPHIs.push_back(PN);
167 // Rewrite all uses outside the loop in terms of the new PHIs we just
169 for (Use *UseToRewrite : UsesToRewrite) {
170 // If this use is in an exit block, rewrite to use the newly inserted PHI.
171 // This is required for correctness because SSAUpdate doesn't handle uses
172 // in the same block. It assumes the PHI we inserted is at the end of the
174 Instruction *User = cast<Instruction>(UseToRewrite->getUser());
175 BasicBlock *UserBB = User->getParent();
176 if (PHINode *PN = dyn_cast<PHINode>(User))
177 UserBB = PN->getIncomingBlock(*UseToRewrite);
179 if (isa<PHINode>(UserBB->begin()) && isExitBlock(UserBB, ExitBlocks)) {
180 // Tell the VHs that the uses changed. This updates SCEV's caches.
181 if (UseToRewrite->get()->hasValueHandle())
182 ValueHandleBase::ValueIsRAUWd(*UseToRewrite, &UserBB->front());
183 UseToRewrite->set(&UserBB->front());
187 // Otherwise, do full PHI insertion.
188 SSAUpdate.RewriteUse(*UseToRewrite);
190 // SSAUpdater might have inserted phi-nodes inside other loops. We'll need
191 // to post-process them to keep LCSSA form.
192 for (PHINode *InsertedPN : InsertedPHIs) {
193 if (auto *OtherLoop = LI.getLoopFor(InsertedPN->getParent()))
194 if (!L->contains(OtherLoop))
195 PostProcessPHIs.push_back(InsertedPN);
199 // Post process PHI instructions that were inserted into another disjoint
200 // loop and update their exits properly.
201 for (auto *PostProcessPN : PostProcessPHIs) {
202 if (PostProcessPN->use_empty())
205 // Reprocess each PHI instruction.
206 Worklist.push_back(PostProcessPN);
209 // Keep track of PHI nodes that we want to remove because they did not have
210 // any uses rewritten.
211 for (PHINode *PN : AddedPHIs)
213 PHIsToRemove.insert(PN);
217 // Remove PHI nodes that did not have any uses rewritten.
218 for (PHINode *PN : PHIsToRemove) {
219 assert (PN->use_empty() && "Trying to remove a phi with uses.");
220 PN->eraseFromParent();
225 /// Return true if the specified block dominates at least
226 /// one of the blocks in the specified list.
228 blockDominatesAnExit(BasicBlock *BB,
230 const SmallVectorImpl<BasicBlock *> &ExitBlocks) {
231 DomTreeNode *DomNode = DT.getNode(BB);
232 return llvm::any_of(ExitBlocks, [&](BasicBlock * EB) {
233 return DT.dominates(DomNode, DT.getNode(EB));
237 bool llvm::formLCSSA(Loop &L, DominatorTree &DT, LoopInfo *LI,
238 ScalarEvolution *SE) {
239 bool Changed = false;
241 // Get the set of exiting blocks.
242 SmallVector<BasicBlock *, 8> ExitBlocks;
243 L.getExitBlocks(ExitBlocks);
245 if (ExitBlocks.empty())
248 SmallVector<Instruction *, 8> Worklist;
250 // Look at all the instructions in the loop, checking to see if they have uses
251 // outside the loop. If so, put them into the worklist to rewrite those uses.
252 for (BasicBlock *BB : L.blocks()) {
253 // For large loops, avoid use-scanning by using dominance information: In
254 // particular, if a block does not dominate any of the loop exits, then none
255 // of the values defined in the block could be used outside the loop.
256 if (!blockDominatesAnExit(BB, DT, ExitBlocks))
259 for (Instruction &I : *BB) {
260 // Reject two common cases fast: instructions with no uses (like stores)
261 // and instructions with one use that is in the same block as this.
263 (I.hasOneUse() && I.user_back()->getParent() == BB &&
264 !isa<PHINode>(I.user_back())))
267 Worklist.push_back(&I);
270 Changed = formLCSSAForInstructions(Worklist, DT, *LI);
272 // If we modified the code, remove any caches about the loop from SCEV to
273 // avoid dangling entries.
274 // FIXME: This is a big hammer, can we clear the cache more selectively?
278 assert(L.isLCSSAForm(DT));
283 /// Process a loop nest depth first.
284 bool llvm::formLCSSARecursively(Loop &L, DominatorTree &DT, LoopInfo *LI,
285 ScalarEvolution *SE) {
286 bool Changed = false;
288 // Recurse depth-first through inner loops.
289 for (Loop *SubLoop : L.getSubLoops())
290 Changed |= formLCSSARecursively(*SubLoop, DT, LI, SE);
292 Changed |= formLCSSA(L, DT, LI, SE);
296 /// Process all loops in the function, inner-most out.
297 static bool formLCSSAOnAllLoops(LoopInfo *LI, DominatorTree &DT,
298 ScalarEvolution *SE) {
299 bool Changed = false;
301 Changed |= formLCSSARecursively(*L, DT, LI, SE);
306 struct LCSSAWrapperPass : public FunctionPass {
307 static char ID; // Pass identification, replacement for typeid
308 LCSSAWrapperPass() : FunctionPass(ID) {
309 initializeLCSSAWrapperPassPass(*PassRegistry::getPassRegistry());
312 // Cached analysis information for the current function.
317 bool runOnFunction(Function &F) override;
319 /// This transformation requires natural loop information & requires that
320 /// loop preheaders be inserted into the CFG. It maintains both of these,
321 /// as well as the CFG. It also requires dominator information.
322 void getAnalysisUsage(AnalysisUsage &AU) const override {
323 AU.setPreservesCFG();
325 AU.addRequired<DominatorTreeWrapperPass>();
326 AU.addRequired<LoopInfoWrapperPass>();
327 AU.addPreservedID(LoopSimplifyID);
328 AU.addPreserved<AAResultsWrapperPass>();
329 AU.addPreserved<BasicAAWrapperPass>();
330 AU.addPreserved<GlobalsAAWrapperPass>();
331 AU.addPreserved<ScalarEvolutionWrapperPass>();
332 AU.addPreserved<SCEVAAWrapperPass>();
337 char LCSSAWrapperPass::ID = 0;
338 INITIALIZE_PASS_BEGIN(LCSSAWrapperPass, "lcssa", "Loop-Closed SSA Form Pass",
340 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
341 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
342 INITIALIZE_PASS_END(LCSSAWrapperPass, "lcssa", "Loop-Closed SSA Form Pass",
345 Pass *llvm::createLCSSAPass() { return new LCSSAWrapperPass(); }
346 char &llvm::LCSSAID = LCSSAWrapperPass::ID;
348 /// Transform \p F into loop-closed SSA form.
349 bool LCSSAWrapperPass::runOnFunction(Function &F) {
350 LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
351 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
352 auto *SEWP = getAnalysisIfAvailable<ScalarEvolutionWrapperPass>();
353 SE = SEWP ? &SEWP->getSE() : nullptr;
355 return formLCSSAOnAllLoops(LI, *DT, SE);
358 PreservedAnalyses LCSSAPass::run(Function &F, AnalysisManager<Function> &AM) {
359 auto &LI = AM.getResult<LoopAnalysis>(F);
360 auto &DT = AM.getResult<DominatorTreeAnalysis>(F);
361 auto *SE = AM.getCachedResult<ScalarEvolutionAnalysis>(F);
362 if (!formLCSSAOnAllLoops(&LI, DT, SE))
363 return PreservedAnalyses::all();
365 // FIXME: This should also 'preserve the CFG'.
366 PreservedAnalyses PA;
367 PA.preserve<BasicAA>();
368 PA.preserve<GlobalsAA>();
369 PA.preserve<SCEVAA>();
370 PA.preserve<ScalarEvolutionAnalysis>();