Andrew Trick | 5d73448 | 2011-12-09 06:19:40 +0000 | [diff] [blame] | 1 | //===-- UnrollLoopRuntime.cpp - Runtime Loop unrolling utilities ----------===// |
| 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 file implements some loop unrolling utilities for loops with run-time |
| 11 | // trip counts. See LoopUnroll.cpp for unrolling loops with compile-time |
| 12 | // trip counts. |
| 13 | // |
| 14 | // The functions in this file are used to generate extra code when the |
| 15 | // run-time trip count modulo the unroll factor is not 0. When this is the |
| 16 | // case, we need to generate code to execute these 'left over' iterations. |
| 17 | // |
| 18 | // The current strategy generates an if-then-else sequence prior to the |
| 19 | // unrolled loop to execute the 'left over' iterations. Other strategies |
| 20 | // include generate a loop before or after the unrolled loop. |
| 21 | // |
| 22 | //===----------------------------------------------------------------------===// |
| 23 | |
| 24 | #define DEBUG_TYPE "loop-unroll" |
| 25 | #include "llvm/Transforms/Utils/UnrollLoop.h" |
| 26 | #include "llvm/BasicBlock.h" |
| 27 | #include "llvm/ADT/Statistic.h" |
| 28 | #include "llvm/Analysis/LoopIterator.h" |
| 29 | #include "llvm/Analysis/LoopPass.h" |
| 30 | #include "llvm/Analysis/ScalarEvolution.h" |
| 31 | #include "llvm/Analysis/ScalarEvolutionExpander.h" |
| 32 | #include "llvm/Support/Debug.h" |
| 33 | #include "llvm/Support/raw_ostream.h" |
| 34 | #include "llvm/Transforms/Utils/BasicBlockUtils.h" |
| 35 | #include "llvm/Transforms/Utils/Cloning.h" |
| 36 | #include <algorithm> |
| 37 | |
| 38 | using namespace llvm; |
| 39 | |
| 40 | STATISTIC(NumRuntimeUnrolled, |
| 41 | "Number of loops unrolled with run-time trip counts"); |
| 42 | |
| 43 | /// Connect the unrolling prolog code to the original loop. |
| 44 | /// The unrolling prolog code contains code to execute the |
| 45 | /// 'extra' iterations if the run-time trip count modulo the |
| 46 | /// unroll count is non-zero. |
| 47 | /// |
| 48 | /// This function performs the following: |
| 49 | /// - Create PHI nodes at prolog end block to combine values |
| 50 | /// that exit the prolog code and jump around the prolog. |
| 51 | /// - Add a PHI operand to a PHI node at the loop exit block |
| 52 | /// for values that exit the prolog and go around the loop. |
| 53 | /// - Branch around the original loop if the trip count is less |
| 54 | /// than the unroll factor. |
| 55 | /// |
| 56 | static void ConnectProlog(Loop *L, Value *TripCount, unsigned Count, |
| 57 | BasicBlock *LastPrologBB, BasicBlock *PrologEnd, |
| 58 | BasicBlock *OrigPH, BasicBlock *NewPH, |
| 59 | ValueToValueMapTy &LVMap, Pass *P) { |
| 60 | BasicBlock *Latch = L->getLoopLatch(); |
| 61 | assert(Latch != 0 && "Loop must have a latch"); |
| 62 | |
| 63 | // Create a PHI node for each outgoing value from the original loop |
| 64 | // (which means it is an outgoing value from the prolog code too). |
| 65 | // The new PHI node is inserted in the prolog end basic block. |
| 66 | // The new PHI name is added as an operand of a PHI node in either |
| 67 | // the loop header or the loop exit block. |
| 68 | for (succ_iterator SBI = succ_begin(Latch), SBE = succ_end(Latch); |
| 69 | SBI != SBE; ++SBI) { |
| 70 | for (BasicBlock::iterator BBI = (*SBI)->begin(); |
| 71 | PHINode *PN = dyn_cast<PHINode>(BBI); ++BBI) { |
| 72 | |
| 73 | // Add a new PHI node to the prolog end block and add the |
| 74 | // appropriate incoming values. |
| 75 | PHINode *NewPN = PHINode::Create(PN->getType(), 2, PN->getName()+".unr", |
| 76 | PrologEnd->getTerminator()); |
| 77 | // Adding a value to the new PHI node from the original loop preheader. |
| 78 | // This is the value that skips all the prolog code. |
| 79 | if (L->contains(PN)) { |
| 80 | NewPN->addIncoming(PN->getIncomingValueForBlock(NewPH), OrigPH); |
| 81 | } else { |
| 82 | NewPN->addIncoming(Constant::getNullValue(PN->getType()), OrigPH); |
| 83 | } |
| 84 | Value *OrigVal = PN->getIncomingValueForBlock(Latch); |
| 85 | Value *V = OrigVal; |
| 86 | if (Instruction *I = dyn_cast<Instruction>(V)) { |
| 87 | if (L->contains(I)) { |
| 88 | V = LVMap[I]; |
| 89 | } |
| 90 | } |
| 91 | // Adding a value to the new PHI node from the last prolog block |
| 92 | // that was created. |
| 93 | NewPN->addIncoming(V, LastPrologBB); |
| 94 | |
| 95 | // Update the existing PHI node operand with the value from the |
| 96 | // new PHI node. How this is done depends on if the existing |
| 97 | // PHI node is in the original loop block, or the exit block. |
| 98 | if (L->contains(PN)) { |
| 99 | PN->setIncomingValue(PN->getBasicBlockIndex(NewPH), NewPN); |
| 100 | } else { |
| 101 | PN->addIncoming(NewPN, PrologEnd); |
| 102 | } |
| 103 | } |
| 104 | } |
| 105 | |
| 106 | // Create a branch around the orignal loop, which is taken if the |
| 107 | // trip count is less than the unroll factor. |
| 108 | Instruction *InsertPt = PrologEnd->getTerminator(); |
| 109 | Instruction *BrLoopExit = |
| 110 | new ICmpInst(InsertPt, ICmpInst::ICMP_ULT, TripCount, |
| 111 | ConstantInt::get(TripCount->getType(), Count)); |
| 112 | BasicBlock *Exit = L->getUniqueExitBlock(); |
| 113 | assert(Exit != 0 && "Loop must have a single exit block only"); |
| 114 | // Split the exit to maintain loop canonicalization guarantees |
| 115 | SmallVector<BasicBlock*, 4> Preds(pred_begin(Exit), pred_end(Exit)); |
| 116 | if (!Exit->isLandingPad()) { |
Jakub Staszak | 2fac1d5 | 2011-12-09 21:19:53 +0000 | [diff] [blame^] | 117 | SplitBlockPredecessors(Exit, Preds, ".unr-lcssa", P); |
Andrew Trick | 5d73448 | 2011-12-09 06:19:40 +0000 | [diff] [blame] | 118 | } else { |
| 119 | SmallVector<BasicBlock*, 2> NewBBs; |
| 120 | SplitLandingPadPredecessors(Exit, Preds, ".unr1-lcssa", ".unr2-lcssa", |
| 121 | P, NewBBs); |
| 122 | } |
| 123 | // Add the branch to the exit block (around the unrolled loop) |
| 124 | BranchInst::Create(Exit, NewPH, BrLoopExit, InsertPt); |
| 125 | InsertPt->eraseFromParent(); |
| 126 | } |
| 127 | |
| 128 | /// Create a clone of the blocks in a loop and connect them together. |
| 129 | /// This function doesn't create a clone of the loop structure. |
| 130 | /// |
| 131 | /// There are two value maps that are defined and used. VMap is |
| 132 | /// for the values in the current loop instance. LVMap contains |
| 133 | /// the values from the last loop instance. We need the LVMap values |
| 134 | /// to update the inital values for the current loop instance. |
| 135 | /// |
| 136 | static void CloneLoopBlocks(Loop *L, |
| 137 | bool FirstCopy, |
| 138 | BasicBlock *InsertTop, |
| 139 | BasicBlock *InsertBot, |
| 140 | std::vector<BasicBlock *> &NewBlocks, |
| 141 | LoopBlocksDFS &LoopBlocks, |
| 142 | ValueToValueMapTy &VMap, |
| 143 | ValueToValueMapTy &LVMap, |
| 144 | LoopInfo *LI) { |
| 145 | |
| 146 | BasicBlock *Preheader = L->getLoopPreheader(); |
| 147 | BasicBlock *Header = L->getHeader(); |
| 148 | BasicBlock *Latch = L->getLoopLatch(); |
| 149 | Function *F = Header->getParent(); |
| 150 | LoopBlocksDFS::RPOIterator BlockBegin = LoopBlocks.beginRPO(); |
| 151 | LoopBlocksDFS::RPOIterator BlockEnd = LoopBlocks.endRPO(); |
| 152 | // For each block in the original loop, create a new copy, |
| 153 | // and update the value map with the newly created values. |
| 154 | for (LoopBlocksDFS::RPOIterator BB = BlockBegin; BB != BlockEnd; ++BB) { |
| 155 | BasicBlock *NewBB = CloneBasicBlock(*BB, VMap, ".unr", F); |
| 156 | NewBlocks.push_back(NewBB); |
| 157 | |
| 158 | if (Loop *ParentLoop = L->getParentLoop()) |
| 159 | ParentLoop->addBasicBlockToLoop(NewBB, LI->getBase()); |
| 160 | |
| 161 | VMap[*BB] = NewBB; |
| 162 | if (Header == *BB) { |
| 163 | // For the first block, add a CFG connection to this newly |
| 164 | // created block |
| 165 | InsertTop->getTerminator()->setSuccessor(0, NewBB); |
| 166 | |
| 167 | // Change the incoming values to the ones defined in the |
| 168 | // previously cloned loop. |
| 169 | for (BasicBlock::iterator I = Header->begin(); isa<PHINode>(I); ++I) { |
| 170 | PHINode *NewPHI = cast<PHINode>(VMap[I]); |
| 171 | if (FirstCopy) { |
| 172 | // We replace the first phi node with the value from the preheader |
| 173 | VMap[I] = NewPHI->getIncomingValueForBlock(Preheader); |
| 174 | NewBB->getInstList().erase(NewPHI); |
| 175 | } else { |
| 176 | // Update VMap with values from the previous block |
| 177 | unsigned idx = NewPHI->getBasicBlockIndex(Latch); |
| 178 | Value *InVal = NewPHI->getIncomingValue(idx); |
| 179 | if (Instruction *I = dyn_cast<Instruction>(InVal)) |
| 180 | if (L->contains(I)) |
| 181 | InVal = LVMap[InVal]; |
| 182 | NewPHI->setIncomingValue(idx, InVal); |
| 183 | NewPHI->setIncomingBlock(idx, InsertTop); |
| 184 | } |
| 185 | } |
| 186 | } |
| 187 | |
| 188 | if (Latch == *BB) { |
| 189 | VMap.erase((*BB)->getTerminator()); |
| 190 | NewBB->getTerminator()->eraseFromParent(); |
| 191 | BranchInst::Create(InsertBot, NewBB); |
| 192 | } |
| 193 | } |
| 194 | // LastValueMap is updated with the values for the current loop |
| 195 | // which are used the next time this function is called. |
| 196 | for (ValueToValueMapTy::iterator VI = VMap.begin(), VE = VMap.end(); |
| 197 | VI != VE; ++VI) { |
| 198 | LVMap[VI->first] = VI->second; |
| 199 | } |
| 200 | } |
| 201 | |
| 202 | /// Insert code in the prolog code when unrolling a loop with a |
| 203 | /// run-time trip-count. |
| 204 | /// |
| 205 | /// This method assumes that the loop unroll factor is total number |
| 206 | /// of loop bodes in the loop after unrolling. (Some folks refer |
| 207 | /// to the unroll factor as the number of *extra* copies added). |
| 208 | /// We assume also that the loop unroll factor is a power-of-two. So, after |
| 209 | /// unrolling the loop, the number of loop bodies executed is 2, |
| 210 | /// 4, 8, etc. Note - LLVM converts the if-then-sequence to a switch |
| 211 | /// instruction in SimplifyCFG.cpp. Then, the backend decides how code for |
| 212 | /// the switch instruction is generated. |
| 213 | /// |
| 214 | /// extraiters = tripcount % loopfactor |
| 215 | /// if (extraiters == 0) jump Loop: |
| 216 | /// if (extraiters == loopfactor) jump L1 |
| 217 | /// if (extraiters == loopfactor-1) jump L2 |
| 218 | /// ... |
| 219 | /// L1: LoopBody; |
| 220 | /// L2: LoopBody; |
| 221 | /// ... |
| 222 | /// if tripcount < loopfactor jump End |
| 223 | /// Loop: |
| 224 | /// ... |
| 225 | /// End: |
| 226 | /// |
| 227 | bool llvm::UnrollRuntimeLoopProlog(Loop *L, unsigned Count, LoopInfo *LI, |
| 228 | LPPassManager *LPM) { |
| 229 | // for now, only unroll loops that contain a single exit |
| 230 | SmallVector<BasicBlock*, 4> ExitingBlocks; |
| 231 | L->getExitingBlocks(ExitingBlocks); |
| 232 | if (ExitingBlocks.size() > 1) |
| 233 | return false; |
| 234 | |
| 235 | // Make sure the loop is in canonical form, and there is a single |
| 236 | // exit block only. |
| 237 | if (!L->isLoopSimplifyForm() || L->getUniqueExitBlock() == 0) |
| 238 | return false; |
| 239 | |
| 240 | // Use Scalar Evolution to compute the trip count. This allows more |
| 241 | // loops to be unrolled than relying on induction var simplification |
| 242 | ScalarEvolution *SE = LPM->getAnalysisIfAvailable<ScalarEvolution>(); |
| 243 | if (SE == 0) |
| 244 | return false; |
| 245 | |
| 246 | // Only unroll loops with a computable trip count and the trip count needs |
| 247 | // to be an int value (allowing a pointer type is a TODO item) |
| 248 | const SCEV *BECount = SE->getBackedgeTakenCount(L); |
| 249 | if (isa<SCEVCouldNotCompute>(BECount) || !BECount->getType()->isIntegerTy()) |
| 250 | return false; |
| 251 | |
| 252 | // Add 1 since the backedge count doesn't include the first loop iteration |
| 253 | const SCEV *TripCountSC = |
| 254 | SE->getAddExpr(BECount, SE->getConstant(BECount->getType(), 1)); |
| 255 | if (isa<SCEVCouldNotCompute>(TripCountSC)) |
| 256 | return false; |
| 257 | |
| 258 | // We only handle cases when the unroll factor is a power of 2. |
| 259 | // Count is the loop unroll factor, the number of extra copies added + 1. |
| 260 | if ((Count & (Count-1)) != 0) |
| 261 | return false; |
| 262 | |
| 263 | // If this loop is nested, then the loop unroller changes the code in |
| 264 | // parent loop, so the Scalar Evolution pass needs to be run again |
| 265 | if (Loop *ParentLoop = L->getParentLoop()) |
| 266 | SE->forgetLoop(ParentLoop); |
| 267 | |
| 268 | BasicBlock *PH = L->getLoopPreheader(); |
| 269 | BasicBlock *Header = L->getHeader(); |
| 270 | BasicBlock *Latch = L->getLoopLatch(); |
| 271 | // It helps to splits the original preheader twice, one for the end of the |
| 272 | // prolog code and one for a new loop preheader |
| 273 | BasicBlock *PEnd = SplitEdge(PH, Header, LPM->getAsPass()); |
| 274 | BasicBlock *NewPH = SplitBlock(PEnd, PEnd->getTerminator(), LPM->getAsPass()); |
| 275 | BranchInst *PreHeaderBR = cast<BranchInst>(PH->getTerminator()); |
| 276 | |
| 277 | // Compute the number of extra iterations required, which is: |
| 278 | // extra iterations = run-time trip count % (loop unroll factor + 1) |
| 279 | SCEVExpander Expander(*SE, "loop-unroll"); |
| 280 | Value *TripCount = Expander.expandCodeFor(TripCountSC, TripCountSC->getType(), |
| 281 | PreHeaderBR); |
| 282 | Type *CountTy = TripCount->getType(); |
| 283 | BinaryOperator *ModVal = |
| 284 | BinaryOperator::CreateURem(TripCount, |
| 285 | ConstantInt::get(CountTy, Count), |
| 286 | "xtraiter"); |
| 287 | ModVal->insertBefore(PreHeaderBR); |
| 288 | |
| 289 | // Check if for no extra iterations, then jump to unrolled loop |
| 290 | Value *BranchVal = new ICmpInst(PreHeaderBR, |
| 291 | ICmpInst::ICMP_NE, ModVal, |
| 292 | ConstantInt::get(CountTy, 0), "lcmp"); |
| 293 | // Branch to either the extra iterations or the unrolled loop |
| 294 | // We will fix up the true branch label when adding loop body copies |
| 295 | BranchInst::Create(PEnd, PEnd, BranchVal, PreHeaderBR); |
| 296 | assert(PreHeaderBR->isUnconditional() && |
| 297 | PreHeaderBR->getSuccessor(0) == PEnd && |
| 298 | "CFG edges in Preheader are not correct"); |
| 299 | PreHeaderBR->eraseFromParent(); |
| 300 | |
| 301 | ValueToValueMapTy LVMap; |
| 302 | Function *F = Header->getParent(); |
| 303 | // These variables are used to update the CFG links in each iteration |
| 304 | BasicBlock *CompareBB = 0; |
| 305 | BasicBlock *LastLoopBB = PH; |
| 306 | // Get an ordered list of blocks in the loop to help with the ordering of the |
| 307 | // cloned blocks in the prolog code |
| 308 | LoopBlocksDFS LoopBlocks(L); |
| 309 | LoopBlocks.perform(LI); |
| 310 | |
| 311 | // |
| 312 | // For each extra loop iteration, create a copy of the loop's basic blocks |
| 313 | // and generate a condition that branches to the copy depending on the |
| 314 | // number of 'left over' iterations. |
| 315 | // |
| 316 | for (unsigned leftOverIters = Count-1; leftOverIters > 0; --leftOverIters) { |
| 317 | std::vector<BasicBlock*> NewBlocks; |
| 318 | ValueToValueMapTy VMap; |
| 319 | |
| 320 | // Clone all the basic blocks in the loop, but we don't clone the loop |
| 321 | // This function adds the appropriate CFG connections. |
| 322 | CloneLoopBlocks(L, (leftOverIters == Count-1), LastLoopBB, PEnd, NewBlocks, |
| 323 | LoopBlocks, VMap, LVMap, LI); |
| 324 | LastLoopBB = cast<BasicBlock>(VMap[Latch]); |
| 325 | |
| 326 | // Insert the cloned blocks into function just before the original loop |
| 327 | F->getBasicBlockList().splice(PEnd, F->getBasicBlockList(), |
| 328 | NewBlocks[0], F->end()); |
| 329 | |
| 330 | // Generate the code for the comparison which determines if the loop |
| 331 | // prolog code needs to be executed. |
| 332 | if (leftOverIters == Count-1) { |
| 333 | // There is no compare block for the fall-thru case when for the last |
| 334 | // left over iteration |
| 335 | CompareBB = NewBlocks[0]; |
| 336 | } else { |
| 337 | // Create a new block for the comparison |
| 338 | BasicBlock *NewBB = BasicBlock::Create(CompareBB->getContext(), "unr.cmp", |
| 339 | F, CompareBB); |
| 340 | if (Loop *ParentLoop = L->getParentLoop()) { |
| 341 | // Add the new block to the parent loop, if needed |
| 342 | ParentLoop->addBasicBlockToLoop(NewBB, LI->getBase()); |
| 343 | } |
| 344 | |
| 345 | // The comparison w/ the extra iteration value and branch |
| 346 | Value *BranchVal = new ICmpInst(*NewBB, ICmpInst::ICMP_EQ, ModVal, |
| 347 | ConstantInt::get(CountTy, leftOverIters), |
| 348 | "un.tmp"); |
| 349 | // Branch to either the extra iterations or the unrolled loop |
| 350 | BranchInst::Create(NewBlocks[0], CompareBB, |
| 351 | BranchVal, NewBB); |
| 352 | CompareBB = NewBB; |
| 353 | PH->getTerminator()->setSuccessor(0, NewBB); |
| 354 | VMap[NewPH] = CompareBB; |
| 355 | } |
| 356 | |
| 357 | // Rewrite the cloned instruction operands to use the values |
| 358 | // created when the clone is created. |
| 359 | for (unsigned i = 0, e = NewBlocks.size(); i != e; ++i) { |
| 360 | for (BasicBlock::iterator I = NewBlocks[i]->begin(), |
| 361 | E = NewBlocks[i]->end(); I != E; ++I) { |
| 362 | RemapInstruction(I, VMap, |
| 363 | RF_NoModuleLevelChanges|RF_IgnoreMissingEntries); |
| 364 | } |
| 365 | } |
| 366 | } |
| 367 | |
| 368 | // Connect the prolog code to the original loop and update the |
| 369 | // PHI functions. |
| 370 | ConnectProlog(L, TripCount, Count, LastLoopBB, PEnd, PH, NewPH, LVMap, |
| 371 | LPM->getAsPass()); |
| 372 | NumRuntimeUnrolled++; |
| 373 | return true; |
| 374 | } |