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 #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/LoopPass.h" 35 #include "llvm/Analysis/ScalarEvolution.h" 36 #include "llvm/IR/Constants.h" 37 #include "llvm/IR/Dominators.h" 38 #include "llvm/IR/Function.h" 39 #include "llvm/IR/Instructions.h" 40 #include "llvm/IR/PredIteratorCache.h" 41 #include "llvm/Pass.h" 42 #include "llvm/Transforms/Utils/LoopUtils.h" 43 #include "llvm/Transforms/Utils/SSAUpdater.h" 44 using namespace llvm; 45 46 #define DEBUG_TYPE "lcssa" 47 48 STATISTIC(NumLCSSA, "Number of live out of a loop variables"); 49 50 /// Return true if the specified block is in the list. 51 static bool isExitBlock(BasicBlock *BB, 52 const SmallVectorImpl<BasicBlock *> &ExitBlocks) { 53 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) 54 if (ExitBlocks[i] == BB) 55 return true; 56 return false; 57 } 58 59 /// Given an instruction in the loop, check to see if it has any uses that are 60 /// outside the current loop. If so, insert LCSSA PHI nodes and rewrite the 61 /// uses. 62 static bool processInstruction(Loop &L, Instruction &Inst, DominatorTree &DT, 63 const SmallVectorImpl<BasicBlock *> &ExitBlocks, 64 PredIteratorCache &PredCache, LoopInfo *LI) { 65 SmallVector<Use *, 16> UsesToRewrite; 66 67 BasicBlock *InstBB = Inst.getParent(); 68 69 for (Use &U : Inst.uses()) { 70 Instruction *User = cast<Instruction>(U.getUser()); 71 BasicBlock *UserBB = User->getParent(); 72 if (PHINode *PN = dyn_cast<PHINode>(User)) 73 UserBB = PN->getIncomingBlock(U); 74 75 if (InstBB != UserBB && !L.contains(UserBB)) 76 UsesToRewrite.push_back(&U); 77 } 78 79 // If there are no uses outside the loop, exit with no change. 80 if (UsesToRewrite.empty()) 81 return false; 82 83 ++NumLCSSA; // We are applying the transformation 84 85 // Invoke/CatchPad instructions are special in that their result value is not 86 // available along their unwind edge. The code below tests to see whether 87 // DomBB dominates the value, so adjust DomBB to the normal destination block, 88 // which is effectively where the value is first usable. 89 BasicBlock *DomBB = Inst.getParent(); 90 if (InvokeInst *Inv = dyn_cast<InvokeInst>(&Inst)) 91 DomBB = Inv->getNormalDest(); 92 if (auto *CPI = dyn_cast<CatchPadInst>(&Inst)) 93 DomBB = CPI->getNormalDest(); 94 95 DomTreeNode *DomNode = DT.getNode(DomBB); 96 97 SmallVector<PHINode *, 16> AddedPHIs; 98 SmallVector<PHINode *, 8> PostProcessPHIs; 99 100 SSAUpdater SSAUpdate; 101 SSAUpdate.Initialize(Inst.getType(), Inst.getName()); 102 103 // Insert the LCSSA phi's into all of the exit blocks dominated by the 104 // value, and add them to the Phi's map. 105 for (SmallVectorImpl<BasicBlock *>::const_iterator BBI = ExitBlocks.begin(), 106 BBE = ExitBlocks.end(); 107 BBI != BBE; ++BBI) { 108 BasicBlock *ExitBB = *BBI; 109 if (!DT.dominates(DomNode, DT.getNode(ExitBB))) 110 continue; 111 112 // If we already inserted something for this BB, don't reprocess it. 113 if (SSAUpdate.HasValueForBlock(ExitBB)) 114 continue; 115 116 PHINode *PN = PHINode::Create(Inst.getType(), PredCache.size(ExitBB), 117 Inst.getName() + ".lcssa", ExitBB->begin()); 118 119 // Add inputs from inside the loop for this PHI. 120 for (BasicBlock *Pred : PredCache.get(ExitBB)) { 121 PN->addIncoming(&Inst, Pred); 122 123 // If the exit block has a predecessor not within the loop, arrange for 124 // the incoming value use corresponding to that predecessor to be 125 // rewritten in terms of a different LCSSA PHI. 126 if (!L.contains(Pred)) 127 UsesToRewrite.push_back( 128 &PN->getOperandUse(PN->getOperandNumForIncomingValue( 129 PN->getNumIncomingValues() - 1))); 130 } 131 132 AddedPHIs.push_back(PN); 133 134 // Remember that this phi makes the value alive in this block. 135 SSAUpdate.AddAvailableValue(ExitBB, PN); 136 137 // LoopSimplify might fail to simplify some loops (e.g. when indirect 138 // branches are involved). In such situations, it might happen that an exit 139 // for Loop L1 is the header of a disjoint Loop L2. Thus, when we create 140 // PHIs in such an exit block, we are also inserting PHIs into L2's header. 141 // This could break LCSSA form for L2 because these inserted PHIs can also 142 // have uses outside of L2. Remember all PHIs in such situation as to 143 // revisit than later on. FIXME: Remove this if indirectbr support into 144 // LoopSimplify gets improved. 145 if (auto *OtherLoop = LI->getLoopFor(ExitBB)) 146 if (!L.contains(OtherLoop)) 147 PostProcessPHIs.push_back(PN); 148 } 149 150 // Rewrite all uses outside the loop in terms of the new PHIs we just 151 // inserted. 152 for (unsigned i = 0, e = UsesToRewrite.size(); i != e; ++i) { 153 // If this use is in an exit block, rewrite to use the newly inserted PHI. 154 // This is required for correctness because SSAUpdate doesn't handle uses in 155 // the same block. It assumes the PHI we inserted is at the end of the 156 // block. 157 Instruction *User = cast<Instruction>(UsesToRewrite[i]->getUser()); 158 BasicBlock *UserBB = User->getParent(); 159 if (PHINode *PN = dyn_cast<PHINode>(User)) 160 UserBB = PN->getIncomingBlock(*UsesToRewrite[i]); 161 162 if (isa<PHINode>(UserBB->begin()) && isExitBlock(UserBB, ExitBlocks)) { 163 // Tell the VHs that the uses changed. This updates SCEV's caches. 164 if (UsesToRewrite[i]->get()->hasValueHandle()) 165 ValueHandleBase::ValueIsRAUWd(*UsesToRewrite[i], UserBB->begin()); 166 UsesToRewrite[i]->set(UserBB->begin()); 167 continue; 168 } 169 170 // Otherwise, do full PHI insertion. 171 SSAUpdate.RewriteUse(*UsesToRewrite[i]); 172 } 173 174 // Post process PHI instructions that were inserted into another disjoint loop 175 // and update their exits properly. 176 for (auto *I : PostProcessPHIs) { 177 if (I->use_empty()) 178 continue; 179 180 BasicBlock *PHIBB = I->getParent(); 181 Loop *OtherLoop = LI->getLoopFor(PHIBB); 182 SmallVector<BasicBlock *, 8> EBs; 183 OtherLoop->getExitBlocks(EBs); 184 if (EBs.empty()) 185 continue; 186 187 // Recurse and re-process each PHI instruction. FIXME: we should really 188 // convert this entire thing to a worklist approach where we process a 189 // vector of instructions... 190 processInstruction(*OtherLoop, *I, DT, EBs, PredCache, LI); 191 } 192 193 // Remove PHI nodes that did not have any uses rewritten. 194 for (unsigned i = 0, e = AddedPHIs.size(); i != e; ++i) { 195 if (AddedPHIs[i]->use_empty()) 196 AddedPHIs[i]->eraseFromParent(); 197 } 198 199 return true; 200 } 201 202 /// Return true if the specified block dominates at least 203 /// one of the blocks in the specified list. 204 static bool 205 blockDominatesAnExit(BasicBlock *BB, 206 DominatorTree &DT, 207 const SmallVectorImpl<BasicBlock *> &ExitBlocks) { 208 DomTreeNode *DomNode = DT.getNode(BB); 209 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) 210 if (DT.dominates(DomNode, DT.getNode(ExitBlocks[i]))) 211 return true; 212 213 return false; 214 } 215 216 bool llvm::formLCSSA(Loop &L, DominatorTree &DT, LoopInfo *LI, 217 ScalarEvolution *SE) { 218 bool Changed = false; 219 220 // Get the set of exiting blocks. 221 SmallVector<BasicBlock *, 8> ExitBlocks; 222 L.getExitBlocks(ExitBlocks); 223 224 if (ExitBlocks.empty()) 225 return false; 226 227 PredIteratorCache PredCache; 228 229 // Look at all the instructions in the loop, checking to see if they have uses 230 // outside the loop. If so, rewrite those uses. 231 for (Loop::block_iterator BBI = L.block_begin(), BBE = L.block_end(); 232 BBI != BBE; ++BBI) { 233 BasicBlock *BB = *BBI; 234 235 // For large loops, avoid use-scanning by using dominance information: In 236 // particular, if a block does not dominate any of the loop exits, then none 237 // of the values defined in the block could be used outside the loop. 238 if (!blockDominatesAnExit(BB, DT, ExitBlocks)) 239 continue; 240 241 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) { 242 // Reject two common cases fast: instructions with no uses (like stores) 243 // and instructions with one use that is in the same block as this. 244 if (I->use_empty() || 245 (I->hasOneUse() && I->user_back()->getParent() == BB && 246 !isa<PHINode>(I->user_back()))) 247 continue; 248 249 Changed |= processInstruction(L, *I, DT, ExitBlocks, PredCache, LI); 250 } 251 } 252 253 // If we modified the code, remove any caches about the loop from SCEV to 254 // avoid dangling entries. 255 // FIXME: This is a big hammer, can we clear the cache more selectively? 256 if (SE && Changed) 257 SE->forgetLoop(&L); 258 259 assert(L.isLCSSAForm(DT)); 260 261 return Changed; 262 } 263 264 /// Process a loop nest depth first. 265 bool llvm::formLCSSARecursively(Loop &L, DominatorTree &DT, LoopInfo *LI, 266 ScalarEvolution *SE) { 267 bool Changed = false; 268 269 // Recurse depth-first through inner loops. 270 for (Loop::iterator I = L.begin(), E = L.end(); I != E; ++I) 271 Changed |= formLCSSARecursively(**I, DT, LI, SE); 272 273 Changed |= formLCSSA(L, DT, LI, SE); 274 return Changed; 275 } 276 277 namespace { 278 struct LCSSA : public FunctionPass { 279 static char ID; // Pass identification, replacement for typeid 280 LCSSA() : FunctionPass(ID) { 281 initializeLCSSAPass(*PassRegistry::getPassRegistry()); 282 } 283 284 // Cached analysis information for the current function. 285 DominatorTree *DT; 286 LoopInfo *LI; 287 ScalarEvolution *SE; 288 289 bool runOnFunction(Function &F) override; 290 291 /// This transformation requires natural loop information & requires that 292 /// loop preheaders be inserted into the CFG. It maintains both of these, 293 /// as well as the CFG. It also requires dominator information. 294 void getAnalysisUsage(AnalysisUsage &AU) const override { 295 AU.setPreservesCFG(); 296 297 AU.addRequired<DominatorTreeWrapperPass>(); 298 AU.addRequired<LoopInfoWrapperPass>(); 299 AU.addPreservedID(LoopSimplifyID); 300 AU.addPreserved<AliasAnalysis>(); 301 AU.addPreserved<ScalarEvolutionWrapperPass>(); 302 } 303 }; 304 } 305 306 char LCSSA::ID = 0; 307 INITIALIZE_PASS_BEGIN(LCSSA, "lcssa", "Loop-Closed SSA Form Pass", false, false) 308 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass) 309 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass) 310 INITIALIZE_PASS_END(LCSSA, "lcssa", "Loop-Closed SSA Form Pass", false, false) 311 312 Pass *llvm::createLCSSAPass() { return new LCSSA(); } 313 char &llvm::LCSSAID = LCSSA::ID; 314 315 316 /// Process all loops in the function, inner-most out. 317 bool LCSSA::runOnFunction(Function &F) { 318 bool Changed = false; 319 LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo(); 320 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree(); 321 auto *SEWP = getAnalysisIfAvailable<ScalarEvolutionWrapperPass>(); 322 SE = SEWP ? &SEWP->getSE() : nullptr; 323 324 // Simplify each loop nest in the function. 325 for (LoopInfo::iterator I = LI->begin(), E = LI->end(); I != E; ++I) 326 Changed |= formLCSSARecursively(**I, *DT, LI, SE); 327 328 return Changed; 329 } 330 331