1 //===- BreakCriticalEdges.cpp - Critical Edge Elimination Pass ------------===// 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 // BreakCriticalEdges pass - Break all of the critical edges in the CFG by 11 // inserting a dummy basic block. This pass may be "required" by passes that 12 // cannot deal with critical edges. For this usage, the structure type is 13 // forward declared. This pass obviously invalidates the CFG, but can update 14 // dominator trees. 15 // 16 //===----------------------------------------------------------------------===// 17 18 #define DEBUG_TYPE "break-crit-edges" 19 #include "llvm/Transforms/Scalar.h" 20 #include "llvm/ADT/SmallVector.h" 21 #include "llvm/ADT/Statistic.h" 22 #include "llvm/Analysis/CFG.h" 23 #include "llvm/Analysis/Dominators.h" 24 #include "llvm/Analysis/LoopInfo.h" 25 #include "llvm/Analysis/ProfileInfo.h" 26 #include "llvm/IR/Function.h" 27 #include "llvm/IR/Instructions.h" 28 #include "llvm/IR/Type.h" 29 #include "llvm/Support/CFG.h" 30 #include "llvm/Support/ErrorHandling.h" 31 #include "llvm/Transforms/Utils/BasicBlockUtils.h" 32 using namespace llvm; 33 34 STATISTIC(NumBroken, "Number of blocks inserted"); 35 36 namespace { 37 struct BreakCriticalEdges : public FunctionPass { 38 static char ID; // Pass identification, replacement for typeid 39 BreakCriticalEdges() : FunctionPass(ID) { 40 initializeBreakCriticalEdgesPass(*PassRegistry::getPassRegistry()); 41 } 42 43 virtual bool runOnFunction(Function &F); 44 45 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 46 AU.addPreserved<DominatorTree>(); 47 AU.addPreserved<LoopInfo>(); 48 AU.addPreserved<ProfileInfo>(); 49 50 // No loop canonicalization guarantees are broken by this pass. 51 AU.addPreservedID(LoopSimplifyID); 52 } 53 }; 54 } 55 56 char BreakCriticalEdges::ID = 0; 57 INITIALIZE_PASS(BreakCriticalEdges, "break-crit-edges", 58 "Break critical edges in CFG", false, false) 59 60 // Publicly exposed interface to pass... 61 char &llvm::BreakCriticalEdgesID = BreakCriticalEdges::ID; 62 FunctionPass *llvm::createBreakCriticalEdgesPass() { 63 return new BreakCriticalEdges(); 64 } 65 66 // runOnFunction - Loop over all of the edges in the CFG, breaking critical 67 // edges as they are found. 68 // 69 bool BreakCriticalEdges::runOnFunction(Function &F) { 70 bool Changed = false; 71 for (Function::iterator I = F.begin(), E = F.end(); I != E; ++I) { 72 TerminatorInst *TI = I->getTerminator(); 73 if (TI->getNumSuccessors() > 1 && !isa<IndirectBrInst>(TI)) 74 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i) 75 if (SplitCriticalEdge(TI, i, this)) { 76 ++NumBroken; 77 Changed = true; 78 } 79 } 80 81 return Changed; 82 } 83 84 //===----------------------------------------------------------------------===// 85 // Implementation of the external critical edge manipulation functions 86 //===----------------------------------------------------------------------===// 87 88 /// createPHIsForSplitLoopExit - When a loop exit edge is split, LCSSA form 89 /// may require new PHIs in the new exit block. This function inserts the 90 /// new PHIs, as needed. Preds is a list of preds inside the loop, SplitBB 91 /// is the new loop exit block, and DestBB is the old loop exit, now the 92 /// successor of SplitBB. 93 static void createPHIsForSplitLoopExit(ArrayRef<BasicBlock *> Preds, 94 BasicBlock *SplitBB, 95 BasicBlock *DestBB) { 96 // SplitBB shouldn't have anything non-trivial in it yet. 97 assert((SplitBB->getFirstNonPHI() == SplitBB->getTerminator() || 98 SplitBB->isLandingPad()) && "SplitBB has non-PHI nodes!"); 99 100 // For each PHI in the destination block. 101 for (BasicBlock::iterator I = DestBB->begin(); 102 PHINode *PN = dyn_cast<PHINode>(I); ++I) { 103 unsigned Idx = PN->getBasicBlockIndex(SplitBB); 104 Value *V = PN->getIncomingValue(Idx); 105 106 // If the input is a PHI which already satisfies LCSSA, don't create 107 // a new one. 108 if (const PHINode *VP = dyn_cast<PHINode>(V)) 109 if (VP->getParent() == SplitBB) 110 continue; 111 112 // Otherwise a new PHI is needed. Create one and populate it. 113 PHINode *NewPN = 114 PHINode::Create(PN->getType(), Preds.size(), "split", 115 SplitBB->isLandingPad() ? 116 SplitBB->begin() : SplitBB->getTerminator()); 117 for (unsigned i = 0, e = Preds.size(); i != e; ++i) 118 NewPN->addIncoming(V, Preds[i]); 119 120 // Update the original PHI. 121 PN->setIncomingValue(Idx, NewPN); 122 } 123 } 124 125 /// SplitCriticalEdge - If this edge is a critical edge, insert a new node to 126 /// split the critical edge. This will update DominatorTree information if it 127 /// is available, thus calling this pass will not invalidate either of them. 128 /// This returns the new block if the edge was split, null otherwise. 129 /// 130 /// If MergeIdenticalEdges is true (not the default), *all* edges from TI to the 131 /// specified successor will be merged into the same critical edge block. 132 /// This is most commonly interesting with switch instructions, which may 133 /// have many edges to any one destination. This ensures that all edges to that 134 /// dest go to one block instead of each going to a different block, but isn't 135 /// the standard definition of a "critical edge". 136 /// 137 /// It is invalid to call this function on a critical edge that starts at an 138 /// IndirectBrInst. Splitting these edges will almost always create an invalid 139 /// program because the address of the new block won't be the one that is jumped 140 /// to. 141 /// 142 BasicBlock *llvm::SplitCriticalEdge(TerminatorInst *TI, unsigned SuccNum, 143 Pass *P, bool MergeIdenticalEdges, 144 bool DontDeleteUselessPhis, 145 bool SplitLandingPads) { 146 if (!isCriticalEdge(TI, SuccNum, MergeIdenticalEdges)) return 0; 147 148 assert(!isa<IndirectBrInst>(TI) && 149 "Cannot split critical edge from IndirectBrInst"); 150 151 BasicBlock *TIBB = TI->getParent(); 152 BasicBlock *DestBB = TI->getSuccessor(SuccNum); 153 154 // Splitting the critical edge to a landing pad block is non-trivial. Don't do 155 // it in this generic function. 156 if (DestBB->isLandingPad()) return 0; 157 158 // Create a new basic block, linking it into the CFG. 159 BasicBlock *NewBB = BasicBlock::Create(TI->getContext(), 160 TIBB->getName() + "." + DestBB->getName() + "_crit_edge"); 161 // Create our unconditional branch. 162 BranchInst *NewBI = BranchInst::Create(DestBB, NewBB); 163 NewBI->setDebugLoc(TI->getDebugLoc()); 164 165 // Branch to the new block, breaking the edge. 166 TI->setSuccessor(SuccNum, NewBB); 167 168 // Insert the block into the function... right after the block TI lives in. 169 Function &F = *TIBB->getParent(); 170 Function::iterator FBBI = TIBB; 171 F.getBasicBlockList().insert(++FBBI, NewBB); 172 173 // If there are any PHI nodes in DestBB, we need to update them so that they 174 // merge incoming values from NewBB instead of from TIBB. 175 { 176 unsigned BBIdx = 0; 177 for (BasicBlock::iterator I = DestBB->begin(); isa<PHINode>(I); ++I) { 178 // We no longer enter through TIBB, now we come in through NewBB. 179 // Revector exactly one entry in the PHI node that used to come from 180 // TIBB to come from NewBB. 181 PHINode *PN = cast<PHINode>(I); 182 183 // Reuse the previous value of BBIdx if it lines up. In cases where we 184 // have multiple phi nodes with *lots* of predecessors, this is a speed 185 // win because we don't have to scan the PHI looking for TIBB. This 186 // happens because the BB list of PHI nodes are usually in the same 187 // order. 188 if (PN->getIncomingBlock(BBIdx) != TIBB) 189 BBIdx = PN->getBasicBlockIndex(TIBB); 190 PN->setIncomingBlock(BBIdx, NewBB); 191 } 192 } 193 194 // If there are any other edges from TIBB to DestBB, update those to go 195 // through the split block, making those edges non-critical as well (and 196 // reducing the number of phi entries in the DestBB if relevant). 197 if (MergeIdenticalEdges) { 198 for (unsigned i = SuccNum+1, e = TI->getNumSuccessors(); i != e; ++i) { 199 if (TI->getSuccessor(i) != DestBB) continue; 200 201 // Remove an entry for TIBB from DestBB phi nodes. 202 DestBB->removePredecessor(TIBB, DontDeleteUselessPhis); 203 204 // We found another edge to DestBB, go to NewBB instead. 205 TI->setSuccessor(i, NewBB); 206 } 207 } 208 209 210 211 // If we don't have a pass object, we can't update anything... 212 if (P == 0) return NewBB; 213 214 DominatorTree *DT = P->getAnalysisIfAvailable<DominatorTree>(); 215 LoopInfo *LI = P->getAnalysisIfAvailable<LoopInfo>(); 216 ProfileInfo *PI = P->getAnalysisIfAvailable<ProfileInfo>(); 217 218 // If we have nothing to update, just return. 219 if (DT == 0 && LI == 0 && PI == 0) 220 return NewBB; 221 222 // Now update analysis information. Since the only predecessor of NewBB is 223 // the TIBB, TIBB clearly dominates NewBB. TIBB usually doesn't dominate 224 // anything, as there are other successors of DestBB. However, if all other 225 // predecessors of DestBB are already dominated by DestBB (e.g. DestBB is a 226 // loop header) then NewBB dominates DestBB. 227 SmallVector<BasicBlock*, 8> OtherPreds; 228 229 // If there is a PHI in the block, loop over predecessors with it, which is 230 // faster than iterating pred_begin/end. 231 if (PHINode *PN = dyn_cast<PHINode>(DestBB->begin())) { 232 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) 233 if (PN->getIncomingBlock(i) != NewBB) 234 OtherPreds.push_back(PN->getIncomingBlock(i)); 235 } else { 236 for (pred_iterator I = pred_begin(DestBB), E = pred_end(DestBB); 237 I != E; ++I) { 238 BasicBlock *P = *I; 239 if (P != NewBB) 240 OtherPreds.push_back(P); 241 } 242 } 243 244 bool NewBBDominatesDestBB = true; 245 246 // Should we update DominatorTree information? 247 if (DT) { 248 DomTreeNode *TINode = DT->getNode(TIBB); 249 250 // The new block is not the immediate dominator for any other nodes, but 251 // TINode is the immediate dominator for the new node. 252 // 253 if (TINode) { // Don't break unreachable code! 254 DomTreeNode *NewBBNode = DT->addNewBlock(NewBB, TIBB); 255 DomTreeNode *DestBBNode = 0; 256 257 // If NewBBDominatesDestBB hasn't been computed yet, do so with DT. 258 if (!OtherPreds.empty()) { 259 DestBBNode = DT->getNode(DestBB); 260 while (!OtherPreds.empty() && NewBBDominatesDestBB) { 261 if (DomTreeNode *OPNode = DT->getNode(OtherPreds.back())) 262 NewBBDominatesDestBB = DT->dominates(DestBBNode, OPNode); 263 OtherPreds.pop_back(); 264 } 265 OtherPreds.clear(); 266 } 267 268 // If NewBBDominatesDestBB, then NewBB dominates DestBB, otherwise it 269 // doesn't dominate anything. 270 if (NewBBDominatesDestBB) { 271 if (!DestBBNode) DestBBNode = DT->getNode(DestBB); 272 DT->changeImmediateDominator(DestBBNode, NewBBNode); 273 } 274 } 275 } 276 277 // Update LoopInfo if it is around. 278 if (LI) { 279 if (Loop *TIL = LI->getLoopFor(TIBB)) { 280 // If one or the other blocks were not in a loop, the new block is not 281 // either, and thus LI doesn't need to be updated. 282 if (Loop *DestLoop = LI->getLoopFor(DestBB)) { 283 if (TIL == DestLoop) { 284 // Both in the same loop, the NewBB joins loop. 285 DestLoop->addBasicBlockToLoop(NewBB, LI->getBase()); 286 } else if (TIL->contains(DestLoop)) { 287 // Edge from an outer loop to an inner loop. Add to the outer loop. 288 TIL->addBasicBlockToLoop(NewBB, LI->getBase()); 289 } else if (DestLoop->contains(TIL)) { 290 // Edge from an inner loop to an outer loop. Add to the outer loop. 291 DestLoop->addBasicBlockToLoop(NewBB, LI->getBase()); 292 } else { 293 // Edge from two loops with no containment relation. Because these 294 // are natural loops, we know that the destination block must be the 295 // header of its loop (adding a branch into a loop elsewhere would 296 // create an irreducible loop). 297 assert(DestLoop->getHeader() == DestBB && 298 "Should not create irreducible loops!"); 299 if (Loop *P = DestLoop->getParentLoop()) 300 P->addBasicBlockToLoop(NewBB, LI->getBase()); 301 } 302 } 303 // If TIBB is in a loop and DestBB is outside of that loop, split the 304 // other exit blocks of the loop that also have predecessors outside 305 // the loop, to maintain a LoopSimplify guarantee. 306 if (!TIL->contains(DestBB) && 307 P->mustPreserveAnalysisID(LoopSimplifyID)) { 308 assert(!TIL->contains(NewBB) && 309 "Split point for loop exit is contained in loop!"); 310 311 // Update LCSSA form in the newly created exit block. 312 if (P->mustPreserveAnalysisID(LCSSAID)) 313 createPHIsForSplitLoopExit(TIBB, NewBB, DestBB); 314 315 // For each unique exit block... 316 // FIXME: This code is functionally equivalent to the corresponding 317 // loop in LoopSimplify. 318 SmallVector<BasicBlock *, 4> ExitBlocks; 319 TIL->getExitBlocks(ExitBlocks); 320 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) { 321 // Collect all the preds that are inside the loop, and note 322 // whether there are any preds outside the loop. 323 SmallVector<BasicBlock *, 4> Preds; 324 bool HasPredOutsideOfLoop = false; 325 BasicBlock *Exit = ExitBlocks[i]; 326 for (pred_iterator I = pred_begin(Exit), E = pred_end(Exit); 327 I != E; ++I) { 328 BasicBlock *P = *I; 329 if (TIL->contains(P)) { 330 if (isa<IndirectBrInst>(P->getTerminator())) { 331 Preds.clear(); 332 break; 333 } 334 Preds.push_back(P); 335 } else { 336 HasPredOutsideOfLoop = true; 337 } 338 } 339 // If there are any preds not in the loop, we'll need to split 340 // the edges. The Preds.empty() check is needed because a block 341 // may appear multiple times in the list. We can't use 342 // getUniqueExitBlocks above because that depends on LoopSimplify 343 // form, which we're in the process of restoring! 344 if (!Preds.empty() && HasPredOutsideOfLoop) { 345 if (!Exit->isLandingPad()) { 346 BasicBlock *NewExitBB = 347 SplitBlockPredecessors(Exit, Preds, "split", P); 348 if (P->mustPreserveAnalysisID(LCSSAID)) 349 createPHIsForSplitLoopExit(Preds, NewExitBB, Exit); 350 } else if (SplitLandingPads) { 351 SmallVector<BasicBlock*, 8> NewBBs; 352 SplitLandingPadPredecessors(Exit, Preds, 353 ".split1", ".split2", 354 P, NewBBs); 355 if (P->mustPreserveAnalysisID(LCSSAID)) 356 createPHIsForSplitLoopExit(Preds, NewBBs[0], Exit); 357 } 358 } 359 } 360 } 361 // LCSSA form was updated above for the case where LoopSimplify is 362 // available, which means that all predecessors of loop exit blocks 363 // are within the loop. Without LoopSimplify form, it would be 364 // necessary to insert a new phi. 365 assert((!P->mustPreserveAnalysisID(LCSSAID) || 366 P->mustPreserveAnalysisID(LoopSimplifyID)) && 367 "SplitCriticalEdge doesn't know how to update LCCSA form " 368 "without LoopSimplify!"); 369 } 370 } 371 372 // Update ProfileInfo if it is around. 373 if (PI) 374 PI->splitEdge(TIBB, DestBB, NewBB, MergeIdenticalEdges); 375 376 return NewBB; 377 } 378