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/Scalar.h"
31 #include "llvm/ADT/STLExtras.h"
32 #include "llvm/ADT/Statistic.h"
33 #include "llvm/Analysis/AliasAnalysis.h"
34 #include "llvm/Analysis/GlobalsModRef.h"
35 #include "llvm/Analysis/LoopPass.h"
36 #include "llvm/Analysis/ScalarEvolution.h"
37 #include "llvm/Analysis/ScalarEvolutionAliasAnalysis.h"
38 #include "llvm/IR/Constants.h"
39 #include "llvm/IR/Dominators.h"
40 #include "llvm/IR/Function.h"
41 #include "llvm/IR/Instructions.h"
42 #include "llvm/IR/PredIteratorCache.h"
43 #include "llvm/Pass.h"
44 #include "llvm/Transforms/Utils/LoopUtils.h"
45 #include "llvm/Transforms/Utils/SSAUpdater.h"
48 #define DEBUG_TYPE "lcssa"
50 STATISTIC(NumLCSSA, "Number of live out of a loop variables");
52 /// Return true if the specified block is in the list.
53 static bool isExitBlock(BasicBlock *BB,
54 const SmallVectorImpl<BasicBlock *> &ExitBlocks) {
55 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i)
56 if (ExitBlocks[i] == BB)
61 /// Given an instruction in the loop, check to see if it has any uses that are
62 /// outside the current loop. If so, insert LCSSA PHI nodes and rewrite the
64 static bool processInstruction(Loop &L, Instruction &Inst, DominatorTree &DT,
65 const SmallVectorImpl<BasicBlock *> &ExitBlocks,
66 PredIteratorCache &PredCache, LoopInfo *LI) {
67 SmallVector<Use *, 16> UsesToRewrite;
69 BasicBlock *InstBB = Inst.getParent();
71 for (Use &U : Inst.uses()) {
72 Instruction *User = cast<Instruction>(U.getUser());
73 BasicBlock *UserBB = User->getParent();
74 if (PHINode *PN = dyn_cast<PHINode>(User))
75 UserBB = PN->getIncomingBlock(U);
77 if (InstBB != UserBB && !L.contains(UserBB))
78 UsesToRewrite.push_back(&U);
81 // If there are no uses outside the loop, exit with no change.
82 if (UsesToRewrite.empty())
85 ++NumLCSSA; // We are applying the transformation
87 // Invoke/CatchPad instructions are special in that their result value is not
88 // available along their unwind edge. The code below tests to see whether
89 // DomBB dominates the value, so adjust DomBB to the normal destination block,
90 // which is effectively where the value is first usable.
91 BasicBlock *DomBB = Inst.getParent();
92 if (InvokeInst *Inv = dyn_cast<InvokeInst>(&Inst))
93 DomBB = Inv->getNormalDest();
94 if (auto *CPI = dyn_cast<CatchPadInst>(&Inst))
95 DomBB = CPI->getNormalDest();
97 DomTreeNode *DomNode = DT.getNode(DomBB);
99 SmallVector<PHINode *, 16> AddedPHIs;
100 SmallVector<PHINode *, 8> PostProcessPHIs;
102 SSAUpdater SSAUpdate;
103 SSAUpdate.Initialize(Inst.getType(), Inst.getName());
105 // Insert the LCSSA phi's into all of the exit blocks dominated by the
106 // value, and add them to the Phi's map.
107 for (SmallVectorImpl<BasicBlock *>::const_iterator BBI = ExitBlocks.begin(),
108 BBE = ExitBlocks.end();
110 BasicBlock *ExitBB = *BBI;
111 if (!DT.dominates(DomNode, DT.getNode(ExitBB)))
114 // If we already inserted something for this BB, don't reprocess it.
115 if (SSAUpdate.HasValueForBlock(ExitBB))
118 PHINode *PN = PHINode::Create(Inst.getType(), PredCache.size(ExitBB),
119 Inst.getName() + ".lcssa", ExitBB->begin());
121 // Add inputs from inside the loop for this PHI.
122 for (BasicBlock *Pred : PredCache.get(ExitBB)) {
123 PN->addIncoming(&Inst, Pred);
125 // If the exit block has a predecessor not within the loop, arrange for
126 // the incoming value use corresponding to that predecessor to be
127 // rewritten in terms of a different LCSSA PHI.
128 if (!L.contains(Pred))
129 UsesToRewrite.push_back(
130 &PN->getOperandUse(PN->getOperandNumForIncomingValue(
131 PN->getNumIncomingValues() - 1)));
134 AddedPHIs.push_back(PN);
136 // Remember that this phi makes the value alive in this block.
137 SSAUpdate.AddAvailableValue(ExitBB, PN);
139 // LoopSimplify might fail to simplify some loops (e.g. when indirect
140 // branches are involved). In such situations, it might happen that an exit
141 // for Loop L1 is the header of a disjoint Loop L2. Thus, when we create
142 // PHIs in such an exit block, we are also inserting PHIs into L2's header.
143 // This could break LCSSA form for L2 because these inserted PHIs can also
144 // have uses outside of L2. Remember all PHIs in such situation as to
145 // revisit than later on. FIXME: Remove this if indirectbr support into
146 // LoopSimplify gets improved.
147 if (auto *OtherLoop = LI->getLoopFor(ExitBB))
148 if (!L.contains(OtherLoop))
149 PostProcessPHIs.push_back(PN);
152 // Rewrite all uses outside the loop in terms of the new PHIs we just
154 for (unsigned i = 0, e = UsesToRewrite.size(); i != e; ++i) {
155 // If this use is in an exit block, rewrite to use the newly inserted PHI.
156 // This is required for correctness because SSAUpdate doesn't handle uses in
157 // the same block. It assumes the PHI we inserted is at the end of the
159 Instruction *User = cast<Instruction>(UsesToRewrite[i]->getUser());
160 BasicBlock *UserBB = User->getParent();
161 if (PHINode *PN = dyn_cast<PHINode>(User))
162 UserBB = PN->getIncomingBlock(*UsesToRewrite[i]);
164 if (isa<PHINode>(UserBB->begin()) && isExitBlock(UserBB, ExitBlocks)) {
165 // Tell the VHs that the uses changed. This updates SCEV's caches.
166 if (UsesToRewrite[i]->get()->hasValueHandle())
167 ValueHandleBase::ValueIsRAUWd(*UsesToRewrite[i], UserBB->begin());
168 UsesToRewrite[i]->set(UserBB->begin());
172 // Otherwise, do full PHI insertion.
173 SSAUpdate.RewriteUse(*UsesToRewrite[i]);
176 // Post process PHI instructions that were inserted into another disjoint loop
177 // and update their exits properly.
178 for (auto *I : PostProcessPHIs) {
182 BasicBlock *PHIBB = I->getParent();
183 Loop *OtherLoop = LI->getLoopFor(PHIBB);
184 SmallVector<BasicBlock *, 8> EBs;
185 OtherLoop->getExitBlocks(EBs);
189 // Recurse and re-process each PHI instruction. FIXME: we should really
190 // convert this entire thing to a worklist approach where we process a
191 // vector of instructions...
192 processInstruction(*OtherLoop, *I, DT, EBs, PredCache, LI);
195 // Remove PHI nodes that did not have any uses rewritten.
196 for (unsigned i = 0, e = AddedPHIs.size(); i != e; ++i) {
197 if (AddedPHIs[i]->use_empty())
198 AddedPHIs[i]->eraseFromParent();
204 /// Return true if the specified block dominates at least
205 /// one of the blocks in the specified list.
207 blockDominatesAnExit(BasicBlock *BB,
209 const SmallVectorImpl<BasicBlock *> &ExitBlocks) {
210 DomTreeNode *DomNode = DT.getNode(BB);
211 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i)
212 if (DT.dominates(DomNode, DT.getNode(ExitBlocks[i])))
218 bool llvm::formLCSSA(Loop &L, DominatorTree &DT, LoopInfo *LI,
219 ScalarEvolution *SE) {
220 bool Changed = false;
222 // Get the set of exiting blocks.
223 SmallVector<BasicBlock *, 8> ExitBlocks;
224 L.getExitBlocks(ExitBlocks);
226 if (ExitBlocks.empty())
229 PredIteratorCache PredCache;
231 // Look at all the instructions in the loop, checking to see if they have uses
232 // outside the loop. If so, rewrite those uses.
233 for (Loop::block_iterator BBI = L.block_begin(), BBE = L.block_end();
235 BasicBlock *BB = *BBI;
237 // For large loops, avoid use-scanning by using dominance information: In
238 // particular, if a block does not dominate any of the loop exits, then none
239 // of the values defined in the block could be used outside the loop.
240 if (!blockDominatesAnExit(BB, DT, ExitBlocks))
243 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
244 // Reject two common cases fast: instructions with no uses (like stores)
245 // and instructions with one use that is in the same block as this.
246 if (I->use_empty() ||
247 (I->hasOneUse() && I->user_back()->getParent() == BB &&
248 !isa<PHINode>(I->user_back())))
251 Changed |= processInstruction(L, *I, DT, ExitBlocks, PredCache, LI);
255 // If we modified the code, remove any caches about the loop from SCEV to
256 // avoid dangling entries.
257 // FIXME: This is a big hammer, can we clear the cache more selectively?
261 assert(L.isLCSSAForm(DT));
266 /// Process a loop nest depth first.
267 bool llvm::formLCSSARecursively(Loop &L, DominatorTree &DT, LoopInfo *LI,
268 ScalarEvolution *SE) {
269 bool Changed = false;
271 // Recurse depth-first through inner loops.
272 for (Loop::iterator I = L.begin(), E = L.end(); I != E; ++I)
273 Changed |= formLCSSARecursively(**I, DT, LI, SE);
275 Changed |= formLCSSA(L, DT, LI, SE);
280 struct LCSSA : public FunctionPass {
281 static char ID; // Pass identification, replacement for typeid
282 LCSSA() : FunctionPass(ID) {
283 initializeLCSSAPass(*PassRegistry::getPassRegistry());
286 // Cached analysis information for the current function.
291 bool runOnFunction(Function &F) override;
293 /// This transformation requires natural loop information & requires that
294 /// loop preheaders be inserted into the CFG. It maintains both of these,
295 /// as well as the CFG. It also requires dominator information.
296 void getAnalysisUsage(AnalysisUsage &AU) const override {
297 AU.setPreservesCFG();
299 AU.addRequired<DominatorTreeWrapperPass>();
300 AU.addRequired<LoopInfoWrapperPass>();
301 AU.addPreservedID(LoopSimplifyID);
302 AU.addPreserved<AAResultsWrapperPass>();
303 AU.addPreserved<GlobalsAAWrapperPass>();
304 AU.addPreserved<ScalarEvolutionWrapperPass>();
305 AU.addPreserved<SCEVAAWrapperPass>();
311 INITIALIZE_PASS_BEGIN(LCSSA, "lcssa", "Loop-Closed SSA Form Pass", false, false)
312 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
313 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
314 INITIALIZE_PASS_DEPENDENCY(GlobalsAAWrapperPass)
315 INITIALIZE_PASS_DEPENDENCY(SCEVAAWrapperPass)
316 INITIALIZE_PASS_END(LCSSA, "lcssa", "Loop-Closed SSA Form Pass", false, false)
318 Pass *llvm::createLCSSAPass() { return new LCSSA(); }
319 char &llvm::LCSSAID = LCSSA::ID;
322 /// Process all loops in the function, inner-most out.
323 bool LCSSA::runOnFunction(Function &F) {
324 bool Changed = false;
325 LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
326 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
327 auto *SEWP = getAnalysisIfAvailable<ScalarEvolutionWrapperPass>();
328 SE = SEWP ? &SEWP->getSE() : nullptr;
330 // Simplify each loop nest in the function.
331 for (LoopInfo::iterator I = LI->begin(), E = LI->end(); I != E; ++I)
332 Changed |= formLCSSARecursively(**I, *DT, LI, SE);