LLVM API Documentation
00001 //===-- LCSSA.cpp - Convert loops into loop-closed SSA form ---------------===// 00002 // 00003 // The LLVM Compiler Infrastructure 00004 // 00005 // This file is distributed under the University of Illinois Open Source 00006 // License. See LICENSE.TXT for details. 00007 // 00008 //===----------------------------------------------------------------------===// 00009 // 00010 // This pass transforms loops by placing phi nodes at the end of the loops for 00011 // all values that are live across the loop boundary. For example, it turns 00012 // the left into the right code: 00013 // 00014 // for (...) for (...) 00015 // if (c) if (c) 00016 // X1 = ... X1 = ... 00017 // else else 00018 // X2 = ... X2 = ... 00019 // X3 = phi(X1, X2) X3 = phi(X1, X2) 00020 // ... = X3 + 4 X4 = phi(X3) 00021 // ... = X4 + 4 00022 // 00023 // This is still valid LLVM; the extra phi nodes are purely redundant, and will 00024 // be trivially eliminated by InstCombine. The major benefit of this 00025 // transformation is that it makes many other loop optimizations, such as 00026 // LoopUnswitching, simpler. 00027 // 00028 //===----------------------------------------------------------------------===// 00029 00030 #include "llvm/Transforms/Scalar.h" 00031 #include "llvm/ADT/STLExtras.h" 00032 #include "llvm/ADT/Statistic.h" 00033 #include "llvm/Analysis/AliasAnalysis.h" 00034 #include "llvm/Analysis/LoopPass.h" 00035 #include "llvm/Analysis/ScalarEvolution.h" 00036 #include "llvm/IR/Constants.h" 00037 #include "llvm/IR/Dominators.h" 00038 #include "llvm/IR/Function.h" 00039 #include "llvm/IR/Instructions.h" 00040 #include "llvm/IR/PredIteratorCache.h" 00041 #include "llvm/Pass.h" 00042 #include "llvm/Transforms/Utils/LoopUtils.h" 00043 #include "llvm/Transforms/Utils/SSAUpdater.h" 00044 using namespace llvm; 00045 00046 #define DEBUG_TYPE "lcssa" 00047 00048 STATISTIC(NumLCSSA, "Number of live out of a loop variables"); 00049 00050 /// Return true if the specified block is in the list. 00051 static bool isExitBlock(BasicBlock *BB, 00052 const SmallVectorImpl<BasicBlock *> &ExitBlocks) { 00053 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) 00054 if (ExitBlocks[i] == BB) 00055 return true; 00056 return false; 00057 } 00058 00059 /// Given an instruction in the loop, check to see if it has any uses that are 00060 /// outside the current loop. If so, insert LCSSA PHI nodes and rewrite the 00061 /// uses. 00062 static bool processInstruction(Loop &L, Instruction &Inst, DominatorTree &DT, 00063 const SmallVectorImpl<BasicBlock *> &ExitBlocks, 00064 PredIteratorCache &PredCache) { 00065 SmallVector<Use *, 16> UsesToRewrite; 00066 00067 BasicBlock *InstBB = Inst.getParent(); 00068 00069 for (Use &U : Inst.uses()) { 00070 Instruction *User = cast<Instruction>(U.getUser()); 00071 BasicBlock *UserBB = User->getParent(); 00072 if (PHINode *PN = dyn_cast<PHINode>(User)) 00073 UserBB = PN->getIncomingBlock(U); 00074 00075 if (InstBB != UserBB && !L.contains(UserBB)) 00076 UsesToRewrite.push_back(&U); 00077 } 00078 00079 // If there are no uses outside the loop, exit with no change. 00080 if (UsesToRewrite.empty()) 00081 return false; 00082 00083 ++NumLCSSA; // We are applying the transformation 00084 00085 // Invoke instructions are special in that their result value is not available 00086 // along their unwind edge. The code below tests to see whether DomBB 00087 // dominates 00088 // the value, so adjust DomBB to the normal destination block, which is 00089 // effectively where the value is first usable. 00090 BasicBlock *DomBB = Inst.getParent(); 00091 if (InvokeInst *Inv = dyn_cast<InvokeInst>(&Inst)) 00092 DomBB = Inv->getNormalDest(); 00093 00094 DomTreeNode *DomNode = DT.getNode(DomBB); 00095 00096 SmallVector<PHINode *, 16> AddedPHIs; 00097 00098 SSAUpdater SSAUpdate; 00099 SSAUpdate.Initialize(Inst.getType(), Inst.getName()); 00100 00101 // Insert the LCSSA phi's into all of the exit blocks dominated by the 00102 // value, and add them to the Phi's map. 00103 for (SmallVectorImpl<BasicBlock *>::const_iterator BBI = ExitBlocks.begin(), 00104 BBE = ExitBlocks.end(); 00105 BBI != BBE; ++BBI) { 00106 BasicBlock *ExitBB = *BBI; 00107 if (!DT.dominates(DomNode, DT.getNode(ExitBB))) 00108 continue; 00109 00110 // If we already inserted something for this BB, don't reprocess it. 00111 if (SSAUpdate.HasValueForBlock(ExitBB)) 00112 continue; 00113 00114 PHINode *PN = PHINode::Create(Inst.getType(), PredCache.GetNumPreds(ExitBB), 00115 Inst.getName() + ".lcssa", ExitBB->begin()); 00116 00117 // Add inputs from inside the loop for this PHI. 00118 for (BasicBlock **PI = PredCache.GetPreds(ExitBB); *PI; ++PI) { 00119 PN->addIncoming(&Inst, *PI); 00120 00121 // If the exit block has a predecessor not within the loop, arrange for 00122 // the incoming value use corresponding to that predecessor to be 00123 // rewritten in terms of a different LCSSA PHI. 00124 if (!L.contains(*PI)) 00125 UsesToRewrite.push_back( 00126 &PN->getOperandUse(PN->getOperandNumForIncomingValue( 00127 PN->getNumIncomingValues() - 1))); 00128 } 00129 00130 AddedPHIs.push_back(PN); 00131 00132 // Remember that this phi makes the value alive in this block. 00133 SSAUpdate.AddAvailableValue(ExitBB, PN); 00134 } 00135 00136 // Rewrite all uses outside the loop in terms of the new PHIs we just 00137 // inserted. 00138 for (unsigned i = 0, e = UsesToRewrite.size(); i != e; ++i) { 00139 // If this use is in an exit block, rewrite to use the newly inserted PHI. 00140 // This is required for correctness because SSAUpdate doesn't handle uses in 00141 // the same block. It assumes the PHI we inserted is at the end of the 00142 // block. 00143 Instruction *User = cast<Instruction>(UsesToRewrite[i]->getUser()); 00144 BasicBlock *UserBB = User->getParent(); 00145 if (PHINode *PN = dyn_cast<PHINode>(User)) 00146 UserBB = PN->getIncomingBlock(*UsesToRewrite[i]); 00147 00148 if (isa<PHINode>(UserBB->begin()) && isExitBlock(UserBB, ExitBlocks)) { 00149 // Tell the VHs that the uses changed. This updates SCEV's caches. 00150 if (UsesToRewrite[i]->get()->hasValueHandle()) 00151 ValueHandleBase::ValueIsRAUWd(*UsesToRewrite[i], UserBB->begin()); 00152 UsesToRewrite[i]->set(UserBB->begin()); 00153 continue; 00154 } 00155 00156 // Otherwise, do full PHI insertion. 00157 SSAUpdate.RewriteUse(*UsesToRewrite[i]); 00158 } 00159 00160 // Remove PHI nodes that did not have any uses rewritten. 00161 for (unsigned i = 0, e = AddedPHIs.size(); i != e; ++i) { 00162 if (AddedPHIs[i]->use_empty()) 00163 AddedPHIs[i]->eraseFromParent(); 00164 } 00165 00166 return true; 00167 } 00168 00169 /// Return true if the specified block dominates at least 00170 /// one of the blocks in the specified list. 00171 static bool 00172 blockDominatesAnExit(BasicBlock *BB, 00173 DominatorTree &DT, 00174 const SmallVectorImpl<BasicBlock *> &ExitBlocks) { 00175 DomTreeNode *DomNode = DT.getNode(BB); 00176 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) 00177 if (DT.dominates(DomNode, DT.getNode(ExitBlocks[i]))) 00178 return true; 00179 00180 return false; 00181 } 00182 00183 bool llvm::formLCSSA(Loop &L, DominatorTree &DT, ScalarEvolution *SE) { 00184 bool Changed = false; 00185 00186 // Get the set of exiting blocks. 00187 SmallVector<BasicBlock *, 8> ExitBlocks; 00188 L.getExitBlocks(ExitBlocks); 00189 00190 if (ExitBlocks.empty()) 00191 return false; 00192 00193 PredIteratorCache PredCache; 00194 00195 // Look at all the instructions in the loop, checking to see if they have uses 00196 // outside the loop. If so, rewrite those uses. 00197 for (Loop::block_iterator BBI = L.block_begin(), BBE = L.block_end(); 00198 BBI != BBE; ++BBI) { 00199 BasicBlock *BB = *BBI; 00200 00201 // For large loops, avoid use-scanning by using dominance information: In 00202 // particular, if a block does not dominate any of the loop exits, then none 00203 // of the values defined in the block could be used outside the loop. 00204 if (!blockDominatesAnExit(BB, DT, ExitBlocks)) 00205 continue; 00206 00207 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) { 00208 // Reject two common cases fast: instructions with no uses (like stores) 00209 // and instructions with one use that is in the same block as this. 00210 if (I->use_empty() || 00211 (I->hasOneUse() && I->user_back()->getParent() == BB && 00212 !isa<PHINode>(I->user_back()))) 00213 continue; 00214 00215 Changed |= processInstruction(L, *I, DT, ExitBlocks, PredCache); 00216 } 00217 } 00218 00219 // If we modified the code, remove any caches about the loop from SCEV to 00220 // avoid dangling entries. 00221 // FIXME: This is a big hammer, can we clear the cache more selectively? 00222 if (SE && Changed) 00223 SE->forgetLoop(&L); 00224 00225 assert(L.isLCSSAForm(DT)); 00226 00227 return Changed; 00228 } 00229 00230 /// Process a loop nest depth first. 00231 bool llvm::formLCSSARecursively(Loop &L, DominatorTree &DT, 00232 ScalarEvolution *SE) { 00233 bool Changed = false; 00234 00235 // Recurse depth-first through inner loops. 00236 for (Loop::iterator LI = L.begin(), LE = L.end(); LI != LE; ++LI) 00237 Changed |= formLCSSARecursively(**LI, DT, SE); 00238 00239 Changed |= formLCSSA(L, DT, SE); 00240 return Changed; 00241 } 00242 00243 namespace { 00244 struct LCSSA : public FunctionPass { 00245 static char ID; // Pass identification, replacement for typeid 00246 LCSSA() : FunctionPass(ID) { 00247 initializeLCSSAPass(*PassRegistry::getPassRegistry()); 00248 } 00249 00250 // Cached analysis information for the current function. 00251 DominatorTree *DT; 00252 LoopInfo *LI; 00253 ScalarEvolution *SE; 00254 00255 bool runOnFunction(Function &F) override; 00256 00257 /// This transformation requires natural loop information & requires that 00258 /// loop preheaders be inserted into the CFG. It maintains both of these, 00259 /// as well as the CFG. It also requires dominator information. 00260 void getAnalysisUsage(AnalysisUsage &AU) const override { 00261 AU.setPreservesCFG(); 00262 00263 AU.addRequired<DominatorTreeWrapperPass>(); 00264 AU.addRequired<LoopInfo>(); 00265 AU.addPreservedID(LoopSimplifyID); 00266 AU.addPreserved<AliasAnalysis>(); 00267 AU.addPreserved<ScalarEvolution>(); 00268 } 00269 00270 private: 00271 void verifyAnalysis() const override; 00272 }; 00273 } 00274 00275 char LCSSA::ID = 0; 00276 INITIALIZE_PASS_BEGIN(LCSSA, "lcssa", "Loop-Closed SSA Form Pass", false, false) 00277 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass) 00278 INITIALIZE_PASS_DEPENDENCY(LoopInfo) 00279 INITIALIZE_PASS_END(LCSSA, "lcssa", "Loop-Closed SSA Form Pass", false, false) 00280 00281 Pass *llvm::createLCSSAPass() { return new LCSSA(); } 00282 char &llvm::LCSSAID = LCSSA::ID; 00283 00284 00285 /// Process all loops in the function, inner-most out. 00286 bool LCSSA::runOnFunction(Function &F) { 00287 bool Changed = false; 00288 LI = &getAnalysis<LoopInfo>(); 00289 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree(); 00290 SE = getAnalysisIfAvailable<ScalarEvolution>(); 00291 00292 // Simplify each loop nest in the function. 00293 for (LoopInfo::iterator I = LI->begin(), E = LI->end(); I != E; ++I) 00294 Changed |= formLCSSARecursively(**I, *DT, SE); 00295 00296 return Changed; 00297 } 00298 00299 static void verifyLoop(Loop &L, DominatorTree &DT) { 00300 // Recurse depth-first through inner loops. 00301 for (Loop::iterator LI = L.begin(), LE = L.end(); LI != LE; ++LI) 00302 verifyLoop(**LI, DT); 00303 00304 // Check the special guarantees that LCSSA makes. 00305 //assert(L.isLCSSAForm(DT) && "LCSSA form not preserved!"); 00306 } 00307 00308 void LCSSA::verifyAnalysis() const { 00309 // Verify each loop nest in the function, assuming LI still points at that 00310 // function's loop info. 00311 for (LoopInfo::iterator I = LI->begin(), E = LI->end(); I != E; ++I) 00312 verifyLoop(**I, *DT); 00313 }