1 //===-- LCSSA.cpp - Convert loops into loop-closed SSA form ---------------===// 2 // 3 // The LLVM Compiler Infrastructure 4 // 5 // This file is distributed under the University of Illinois Open Source 6 // License. See LICENSE.TXT for details. 7 // 8 //===----------------------------------------------------------------------===// 9 // 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: 13 // 14 // for (...) for (...) 15 // if (c) if (c) 16 // X1 = ... X1 = ... 17 // else else 18 // X2 = ... X2 = ... 19 // X3 = phi(X1, X2) X3 = phi(X1, X2) 20 // ... = X3 + 4 X4 = phi(X3) 21 // ... = X4 + 4 22 // 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. 27 // 28 //===----------------------------------------------------------------------===// 29 30 #define DEBUG_TYPE "lcssa" 31 #include "llvm/Transforms/Scalar.h" 32 #include "llvm/Constants.h" 33 #include "llvm/Pass.h" 34 #include "llvm/Function.h" 35 #include "llvm/Instructions.h" 36 #include "llvm/Analysis/Dominators.h" 37 #include "llvm/Analysis/LoopPass.h" 38 #include "llvm/Analysis/ScalarEvolution.h" 39 #include "llvm/Transforms/Utils/SSAUpdater.h" 40 #include "llvm/ADT/Statistic.h" 41 #include "llvm/ADT/STLExtras.h" 42 #include "llvm/Support/PredIteratorCache.h" 43 using namespace llvm; 44 45 STATISTIC(NumLCSSA, "Number of live out of a loop variables"); 46 47 namespace { 48 struct LCSSA : public LoopPass { 49 static char ID; // Pass identification, replacement for typeid 50 LCSSA() : LoopPass(ID) { 51 initializeLCSSAPass(*PassRegistry::getPassRegistry()); 52 } 53 54 // Cached analysis information for the current function. 55 DominatorTree *DT; 56 LoopInfo *LI; 57 ScalarEvolution *SE; 58 std::vector<BasicBlock*> LoopBlocks; 59 PredIteratorCache PredCache; 60 Loop *L; 61 62 virtual bool runOnLoop(Loop *L, LPPassManager &LPM); 63 64 /// This transformation requires natural loop information & requires that 65 /// loop preheaders be inserted into the CFG. It maintains both of these, 66 /// as well as the CFG. It also requires dominator information. 67 /// 68 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 69 AU.setPreservesCFG(); 70 71 AU.addRequired<DominatorTree>(); 72 AU.addRequired<LoopInfo>(); 73 AU.addPreservedID(LoopSimplifyID); 74 AU.addPreserved<ScalarEvolution>(); 75 } 76 private: 77 bool ProcessInstruction(Instruction *Inst, 78 const SmallVectorImpl<BasicBlock*> &ExitBlocks); 79 80 /// verifyAnalysis() - Verify loop nest. 81 virtual void verifyAnalysis() const { 82 // Check the special guarantees that LCSSA makes. 83 assert(L->isLCSSAForm(*DT) && "LCSSA form not preserved!"); 84 } 85 86 /// inLoop - returns true if the given block is within the current loop 87 bool inLoop(BasicBlock *B) const { 88 return std::binary_search(LoopBlocks.begin(), LoopBlocks.end(), B); 89 } 90 }; 91 } 92 93 char LCSSA::ID = 0; 94 INITIALIZE_PASS_BEGIN(LCSSA, "lcssa", "Loop-Closed SSA Form Pass", false, false) 95 INITIALIZE_PASS_DEPENDENCY(DominatorTree) 96 INITIALIZE_PASS_DEPENDENCY(LoopInfo) 97 INITIALIZE_PASS_END(LCSSA, "lcssa", "Loop-Closed SSA Form Pass", false, false) 98 99 Pass *llvm::createLCSSAPass() { return new LCSSA(); } 100 char &llvm::LCSSAID = LCSSA::ID; 101 102 103 /// BlockDominatesAnExit - Return true if the specified block dominates at least 104 /// one of the blocks in the specified list. 105 static bool BlockDominatesAnExit(BasicBlock *BB, 106 const SmallVectorImpl<BasicBlock*> &ExitBlocks, 107 DominatorTree *DT) { 108 DomTreeNode *DomNode = DT->getNode(BB); 109 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) 110 if (DT->dominates(DomNode, DT->getNode(ExitBlocks[i]))) 111 return true; 112 113 return false; 114 } 115 116 117 /// runOnFunction - Process all loops in the function, inner-most out. 118 bool LCSSA::runOnLoop(Loop *TheLoop, LPPassManager &LPM) { 119 L = TheLoop; 120 121 DT = &getAnalysis<DominatorTree>(); 122 LI = &getAnalysis<LoopInfo>(); 123 SE = getAnalysisIfAvailable<ScalarEvolution>(); 124 125 // Get the set of exiting blocks. 126 SmallVector<BasicBlock*, 8> ExitBlocks; 127 L->getExitBlocks(ExitBlocks); 128 129 if (ExitBlocks.empty()) 130 return false; 131 132 // Speed up queries by creating a sorted vector of blocks. 133 LoopBlocks.clear(); 134 LoopBlocks.insert(LoopBlocks.end(), L->block_begin(), L->block_end()); 135 array_pod_sort(LoopBlocks.begin(), LoopBlocks.end()); 136 137 // Look at all the instructions in the loop, checking to see if they have uses 138 // outside the loop. If so, rewrite those uses. 139 bool MadeChange = false; 140 141 for (Loop::block_iterator BBI = L->block_begin(), E = L->block_end(); 142 BBI != E; ++BBI) { 143 BasicBlock *BB = *BBI; 144 145 // For large loops, avoid use-scanning by using dominance information: In 146 // particular, if a block does not dominate any of the loop exits, then none 147 // of the values defined in the block could be used outside the loop. 148 if (!BlockDominatesAnExit(BB, ExitBlocks, DT)) 149 continue; 150 151 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); 152 I != E; ++I) { 153 // Reject two common cases fast: instructions with no uses (like stores) 154 // and instructions with one use that is in the same block as this. 155 if (I->use_empty() || 156 (I->hasOneUse() && I->use_back()->getParent() == BB && 157 !isa<PHINode>(I->use_back()))) 158 continue; 159 160 MadeChange |= ProcessInstruction(I, ExitBlocks); 161 } 162 } 163 164 assert(L->isLCSSAForm(*DT)); 165 PredCache.clear(); 166 167 return MadeChange; 168 } 169 170 /// isExitBlock - Return true if the specified block is in the list. 171 static bool isExitBlock(BasicBlock *BB, 172 const SmallVectorImpl<BasicBlock*> &ExitBlocks) { 173 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) 174 if (ExitBlocks[i] == BB) 175 return true; 176 return false; 177 } 178 179 /// ProcessInstruction - Given an instruction in the loop, check to see if it 180 /// has any uses that are outside the current loop. If so, insert LCSSA PHI 181 /// nodes and rewrite the uses. 182 bool LCSSA::ProcessInstruction(Instruction *Inst, 183 const SmallVectorImpl<BasicBlock*> &ExitBlocks) { 184 SmallVector<Use*, 16> UsesToRewrite; 185 186 BasicBlock *InstBB = Inst->getParent(); 187 188 for (Value::use_iterator UI = Inst->use_begin(), E = Inst->use_end(); 189 UI != E; ++UI) { 190 User *U = *UI; 191 BasicBlock *UserBB = cast<Instruction>(U)->getParent(); 192 if (PHINode *PN = dyn_cast<PHINode>(U)) 193 UserBB = PN->getIncomingBlock(UI); 194 195 if (InstBB != UserBB && !inLoop(UserBB)) 196 UsesToRewrite.push_back(&UI.getUse()); 197 } 198 199 // If there are no uses outside the loop, exit with no change. 200 if (UsesToRewrite.empty()) return false; 201 202 ++NumLCSSA; // We are applying the transformation 203 204 // Invoke instructions are special in that their result value is not available 205 // along their unwind edge. The code below tests to see whether DomBB dominates 206 // the value, so adjust DomBB to the normal destination block, which is 207 // effectively where the value is first usable. 208 BasicBlock *DomBB = Inst->getParent(); 209 if (InvokeInst *Inv = dyn_cast<InvokeInst>(Inst)) 210 DomBB = Inv->getNormalDest(); 211 212 DomTreeNode *DomNode = DT->getNode(DomBB); 213 214 SmallVector<PHINode*, 16> AddedPHIs; 215 216 SSAUpdater SSAUpdate; 217 SSAUpdate.Initialize(Inst->getType(), Inst->getName()); 218 219 // Insert the LCSSA phi's into all of the exit blocks dominated by the 220 // value, and add them to the Phi's map. 221 for (SmallVectorImpl<BasicBlock*>::const_iterator BBI = ExitBlocks.begin(), 222 BBE = ExitBlocks.end(); BBI != BBE; ++BBI) { 223 BasicBlock *ExitBB = *BBI; 224 if (!DT->dominates(DomNode, DT->getNode(ExitBB))) continue; 225 226 // If we already inserted something for this BB, don't reprocess it. 227 if (SSAUpdate.HasValueForBlock(ExitBB)) continue; 228 229 PHINode *PN = PHINode::Create(Inst->getType(), 230 PredCache.GetNumPreds(ExitBB), 231 Inst->getName()+".lcssa", 232 ExitBB->begin()); 233 234 // Add inputs from inside the loop for this PHI. 235 for (BasicBlock **PI = PredCache.GetPreds(ExitBB); *PI; ++PI) { 236 PN->addIncoming(Inst, *PI); 237 238 // If the exit block has a predecessor not within the loop, arrange for 239 // the incoming value use corresponding to that predecessor to be 240 // rewritten in terms of a different LCSSA PHI. 241 if (!inLoop(*PI)) 242 UsesToRewrite.push_back( 243 &PN->getOperandUse( 244 PN->getOperandNumForIncomingValue(PN->getNumIncomingValues()-1))); 245 } 246 247 AddedPHIs.push_back(PN); 248 249 // Remember that this phi makes the value alive in this block. 250 SSAUpdate.AddAvailableValue(ExitBB, PN); 251 252 // If the exiting block is part of a loop inserting a PHI may change its 253 // SCEV analysis. Conservatively drop any caches from it. 254 if (SE) 255 if (Loop *L = LI->getLoopFor(ExitBB)) 256 SE->forgetLoop(L); 257 } 258 259 // Rewrite all uses outside the loop in terms of the new PHIs we just 260 // inserted. 261 for (unsigned i = 0, e = UsesToRewrite.size(); i != e; ++i) { 262 // If this use is in an exit block, rewrite to use the newly inserted PHI. 263 // This is required for correctness because SSAUpdate doesn't handle uses in 264 // the same block. It assumes the PHI we inserted is at the end of the 265 // block. 266 Instruction *User = cast<Instruction>(UsesToRewrite[i]->getUser()); 267 BasicBlock *UserBB = User->getParent(); 268 if (PHINode *PN = dyn_cast<PHINode>(User)) 269 UserBB = PN->getIncomingBlock(*UsesToRewrite[i]); 270 271 // Tell SCEV to reanalyze the value that's about to change. 272 if (SE) 273 SE->forgetValue(*UsesToRewrite[i]); 274 275 if (isa<PHINode>(UserBB->begin()) && 276 isExitBlock(UserBB, ExitBlocks)) { 277 UsesToRewrite[i]->set(UserBB->begin()); 278 continue; 279 } 280 281 // Otherwise, do full PHI insertion. 282 SSAUpdate.RewriteUse(*UsesToRewrite[i]); 283 } 284 285 // Remove PHI nodes that did not have any uses rewritten. 286 for (unsigned i = 0, e = AddedPHIs.size(); i != e; ++i) { 287 if (AddedPHIs[i]->use_empty()) 288 AddedPHIs[i]->eraseFromParent(); 289 } 290 291 return true; 292 } 293 294