X-Git-Url: http://plrg.eecs.uci.edu/git/?a=blobdiff_plain;f=lib%2FTransforms%2FScalar%2FLoopUnrollPass.cpp;h=e0f915b445fa03baeda40bf9f6aa7cd1f28fb053;hb=0b8c9a80f20772c3793201ab5b251d3520b9cea3;hp=dabb9481b4a55ad109982ea0ca7ffe48ef4941af;hpb=081c34b725980f995be9080eaec24cd3dfaaf065;p=oota-llvm.git diff --git a/lib/Transforms/Scalar/LoopUnrollPass.cpp b/lib/Transforms/Scalar/LoopUnrollPass.cpp index dabb9481b4a..e0f915b445f 100644 --- a/lib/Transforms/Scalar/LoopUnrollPass.cpp +++ b/lib/Transforms/Scalar/LoopUnrollPass.cpp @@ -13,11 +13,12 @@ //===----------------------------------------------------------------------===// #define DEBUG_TYPE "loop-unroll" -#include "llvm/IntrinsicInst.h" #include "llvm/Transforms/Scalar.h" +#include "llvm/Analysis/CodeMetrics.h" #include "llvm/Analysis/LoopPass.h" -#include "llvm/Analysis/InlineCost.h" #include "llvm/Analysis/ScalarEvolution.h" +#include "llvm/IR/DataLayout.h" +#include "llvm/IR/IntrinsicInst.h" #include "llvm/Support/CommandLine.h" #include "llvm/Support/Debug.h" #include "llvm/Support/raw_ostream.h" @@ -39,11 +40,21 @@ UnrollAllowPartial("unroll-allow-partial", cl::init(false), cl::Hidden, cl::desc("Allows loops to be partially unrolled until " "-unroll-threshold loop size is reached.")); +static cl::opt +UnrollRuntime("unroll-runtime", cl::ZeroOrMore, cl::init(false), cl::Hidden, + cl::desc("Unroll loops with run-time trip counts")); + namespace { class LoopUnroll : public LoopPass { public: static char ID; // Pass ID, replacement for typeid - LoopUnroll() : LoopPass(ID) { + LoopUnroll(int T = -1, int C = -1, int P = -1) : LoopPass(ID) { + CurrentThreshold = (T == -1) ? UnrollThreshold : unsigned(T); + CurrentCount = (C == -1) ? UnrollCount : unsigned(C); + CurrentAllowPartial = (P == -1) ? UnrollAllowPartial : (bool)P; + + UserThreshold = (T != -1) || (UnrollThreshold.getNumOccurrences() > 0); + initializeLoopUnrollPass(*PassRegistry::getPassRegistry()); } @@ -51,12 +62,19 @@ namespace { /// that the loop unroll should be performed regardless of how much /// code expansion would result. static const unsigned NoThreshold = UINT_MAX; - + // Threshold to use when optsize is specified (and there is no // explicit -unroll-threshold). static const unsigned OptSizeUnrollThreshold = 50; - + + // Default unroll count for loops with run-time trip count if + // -unroll-count is not set + static const unsigned UnrollRuntimeCount = 8; + + unsigned CurrentCount; unsigned CurrentThreshold; + bool CurrentAllowPartial; + bool UserThreshold; // CurrentThreshold is user-specified. bool runOnLoop(Loop *L, LPPassManager &LPM); @@ -70,6 +88,7 @@ namespace { AU.addPreservedID(LoopSimplifyID); AU.addRequiredID(LCSSAID); AU.addPreservedID(LCSSAID); + AU.addRequired(); AU.addPreserved(); // FIXME: Loop unroll requires LCSSA. And LCSSA requires dom info. // If loop unroll does not preserve dom info then LCSSA pass on next @@ -86,67 +105,70 @@ INITIALIZE_PASS_DEPENDENCY(LoopInfo) INITIALIZE_PASS_DEPENDENCY(LoopSimplify) INITIALIZE_PASS_DEPENDENCY(LCSSA) INITIALIZE_PASS_DEPENDENCY(ScalarEvolution) -INITIALIZE_PASS_DEPENDENCY(DominatorTree) INITIALIZE_PASS_END(LoopUnroll, "loop-unroll", "Unroll loops", false, false) -Pass *llvm::createLoopUnrollPass() { return new LoopUnroll(); } +Pass *llvm::createLoopUnrollPass(int Threshold, int Count, int AllowPartial) { + return new LoopUnroll(Threshold, Count, AllowPartial); +} /// ApproximateLoopSize - Approximate the size of the loop. -static unsigned ApproximateLoopSize(const Loop *L, unsigned &NumCalls) { +static unsigned ApproximateLoopSize(const Loop *L, unsigned &NumCalls, + bool &NotDuplicatable, const DataLayout *TD) { CodeMetrics Metrics; for (Loop::block_iterator I = L->block_begin(), E = L->block_end(); I != E; ++I) - Metrics.analyzeBasicBlock(*I); + Metrics.analyzeBasicBlock(*I, TD); NumCalls = Metrics.NumInlineCandidates; - + NotDuplicatable = Metrics.notDuplicatable; + unsigned LoopSize = Metrics.NumInsts; - - // If we can identify the induction variable, we know that it will become - // constant when we unroll the loop, so factor that into our loop size - // estimate. - // FIXME: We have to divide by InlineConstants::InstrCost because the - // measure returned by CountCodeReductionForConstant is not an instruction - // count, but rather a weight as defined by InlineConstants. It would - // probably be a good idea to standardize on a single weighting scheme by - // pushing more of the logic for weighting into CodeMetrics. - if (PHINode *IndVar = L->getCanonicalInductionVariable()) { - unsigned SizeDecrease = Metrics.CountCodeReductionForConstant(IndVar); - // NOTE: Because SizeDecrease is a fuzzy estimate, we don't want to allow - // it to totally negate the cost of unrolling a loop. - SizeDecrease = SizeDecrease > LoopSize / 2 ? LoopSize / 2 : SizeDecrease; - } - + // Don't allow an estimate of size zero. This would allows unrolling of loops // with huge iteration counts, which is a compile time problem even if it's // not a problem for code quality. if (LoopSize == 0) LoopSize = 1; - + return LoopSize; } bool LoopUnroll::runOnLoop(Loop *L, LPPassManager &LPM) { - LoopInfo *LI = &getAnalysis(); + ScalarEvolution *SE = &getAnalysis(); BasicBlock *Header = L->getHeader(); DEBUG(dbgs() << "Loop Unroll: F[" << Header->getParent()->getName() << "] Loop %" << Header->getName() << "\n"); (void)Header; - + // Determine the current unrolling threshold. While this is normally set // from UnrollThreshold, it is overridden to a smaller value if the current // function is marked as optimize-for-size, and the unroll threshold was // not user specified. - CurrentThreshold = UnrollThreshold; - if (Header->getParent()->hasFnAttr(Attribute::OptimizeForSize) && - UnrollThreshold.getNumOccurrences() == 0) - CurrentThreshold = OptSizeUnrollThreshold; - - // Find trip count - unsigned TripCount = L->getSmallConstantTripCount(); - unsigned Count = UnrollCount; + unsigned Threshold = CurrentThreshold; + if (!UserThreshold && + Header->getParent()->getAttributes(). + hasAttribute(AttributeSet::FunctionIndex, + Attribute::OptimizeForSize)) + Threshold = OptSizeUnrollThreshold; + + // Find trip count and trip multiple if count is not available + unsigned TripCount = 0; + unsigned TripMultiple = 1; + // Find "latch trip count". UnrollLoop assumes that control cannot exit + // via the loop latch on any iteration prior to TripCount. The loop may exit + // early via an earlier branch. + BasicBlock *LatchBlock = L->getLoopLatch(); + if (LatchBlock) { + TripCount = SE->getSmallConstantTripCount(L, LatchBlock); + TripMultiple = SE->getSmallConstantTripMultiple(L, LatchBlock); + } + // Use a default unroll-count if the user doesn't specify a value + // and the trip count is a run-time value. The default is different + // for run-time or compile-time trip count loops. + unsigned Count = CurrentCount; + if (UnrollRuntime && CurrentCount == 0 && TripCount == 0) + Count = UnrollRuntimeCount; - // Automatically select an unroll count. if (Count == 0) { // Conservative heuristic: if we know the trip count, see if we can // completely unroll (subject to the threshold, checked below); otherwise @@ -158,27 +180,43 @@ bool LoopUnroll::runOnLoop(Loop *L, LPPassManager &LPM) { } // Enforce the threshold. - if (CurrentThreshold != NoThreshold) { + if (Threshold != NoThreshold) { + const DataLayout *TD = getAnalysisIfAvailable(); unsigned NumInlineCandidates; - unsigned LoopSize = ApproximateLoopSize(L, NumInlineCandidates); + bool notDuplicatable; + unsigned LoopSize = ApproximateLoopSize(L, NumInlineCandidates, + notDuplicatable, TD); DEBUG(dbgs() << " Loop Size = " << LoopSize << "\n"); + if (notDuplicatable) { + DEBUG(dbgs() << " Not unrolling loop which contains non duplicatable" + << " instructions.\n"); + return false; + } if (NumInlineCandidates != 0) { DEBUG(dbgs() << " Not unrolling loop with inlinable calls.\n"); return false; } uint64_t Size = (uint64_t)LoopSize*Count; - if (TripCount != 1 && Size > CurrentThreshold) { + if (TripCount != 1 && Size > Threshold) { DEBUG(dbgs() << " Too large to fully unroll with count: " << Count - << " because size: " << Size << ">" << CurrentThreshold << "\n"); - if (!UnrollAllowPartial) { + << " because size: " << Size << ">" << Threshold << "\n"); + if (!CurrentAllowPartial && !(UnrollRuntime && TripCount == 0)) { DEBUG(dbgs() << " will not try to unroll partially because " << "-unroll-allow-partial not given\n"); return false; } - // Reduce unroll count to be modulo of TripCount for partial unrolling - Count = CurrentThreshold / LoopSize; - while (Count != 0 && TripCount%Count != 0) { - Count--; + if (TripCount) { + // Reduce unroll count to be modulo of TripCount for partial unrolling + Count = Threshold / LoopSize; + while (Count != 0 && TripCount%Count != 0) + Count--; + } + else if (UnrollRuntime) { + // Reduce unroll count to be a lower power-of-two value + while (Count != 0 && Size > Threshold) { + Count >>= 1; + Size = LoopSize*Count; + } } if (Count < 2) { DEBUG(dbgs() << " could not unroll partially\n"); @@ -189,12 +227,8 @@ bool LoopUnroll::runOnLoop(Loop *L, LPPassManager &LPM) { } // Unroll the loop. - Function *F = L->getHeader()->getParent(); - if (!UnrollLoop(L, Count, LI, &LPM)) + if (!UnrollLoop(L, Count, TripCount, UnrollRuntime, TripMultiple, LI, &LPM)) return false; - // FIXME: Reconstruct dom info, because it is not preserved properly. - if (DominatorTree *DT = getAnalysisIfAvailable()) - DT->runOnFunction(*F); return true; }