#include "llvm/Transforms/Utils/UnrollLoop.h"
#include "llvm/ADT/Statistic.h"
+#include "llvm/Analysis/AliasAnalysis.h"
#include "llvm/Analysis/LoopIterator.h"
#include "llvm/Analysis/LoopPass.h"
#include "llvm/Analysis/ScalarEvolution.h"
#include "llvm/Analysis/ScalarEvolutionExpander.h"
#include "llvm/IR/BasicBlock.h"
+#include "llvm/IR/Dominators.h"
#include "llvm/IR/Metadata.h"
+#include "llvm/IR/Module.h"
#include "llvm/Support/Debug.h"
#include "llvm/Support/raw_ostream.h"
+#include "llvm/Transforms/Scalar.h"
#include "llvm/Transforms/Utils/BasicBlockUtils.h"
#include "llvm/Transforms/Utils/Cloning.h"
#include <algorithm>
/// - Branch around the original loop if the trip count is less
/// than the unroll factor.
///
-static void ConnectProlog(Loop *L, Value *TripCount, unsigned Count,
+static void ConnectProlog(Loop *L, Value *BECount, unsigned Count,
BasicBlock *LastPrologBB, BasicBlock *PrologEnd,
BasicBlock *OrigPH, BasicBlock *NewPH,
- ValueToValueMapTy &VMap, Pass *P) {
+ ValueToValueMapTy &VMap, DominatorTree *DT,
+ LoopInfo *LI, Pass *P) {
BasicBlock *Latch = L->getLoopLatch();
assert(Latch && "Loop must have a latch");
if (L->contains(PN)) {
NewPN->addIncoming(PN->getIncomingValueForBlock(NewPH), OrigPH);
} else {
- NewPN->addIncoming(Constant::getNullValue(PN->getType()), OrigPH);
+ NewPN->addIncoming(UndefValue::get(PN->getType()), OrigPH);
}
Value *V = PN->getIncomingValueForBlock(Latch);
}
}
- // Create a branch around the orignal loop, which is taken if the
- // trip count is less than the unroll factor.
+ // Create a branch around the orignal loop, which is taken if there are no
+ // iterations remaining to be executed after running the prologue.
Instruction *InsertPt = PrologEnd->getTerminator();
- Instruction *BrLoopExit =
- new ICmpInst(InsertPt, ICmpInst::ICMP_ULT, TripCount,
- ConstantInt::get(TripCount->getType(), Count));
+ IRBuilder<> B(InsertPt);
+
+ assert(Count != 0 && "nonsensical Count!");
+
+ // If BECount <u (Count - 1) then (BECount + 1) & (Count - 1) == (BECount + 1)
+ // (since Count is a power of 2). This means %xtraiter is (BECount + 1) and
+ // and all of the iterations of this loop were executed by the prologue. Note
+ // that if BECount <u (Count - 1) then (BECount + 1) cannot unsigned-overflow.
+ Value *BrLoopExit =
+ B.CreateICmpULT(BECount, ConstantInt::get(BECount->getType(), Count - 1));
BasicBlock *Exit = L->getUniqueExitBlock();
assert(Exit && "Loop must have a single exit block only");
// Split the exit to maintain loop canonicalization guarantees
SmallVector<BasicBlock*, 4> Preds(pred_begin(Exit), pred_end(Exit));
- if (!Exit->isLandingPad()) {
- SplitBlockPredecessors(Exit, Preds, ".unr-lcssa", P);
- } else {
- SmallVector<BasicBlock*, 2> NewBBs;
- SplitLandingPadPredecessors(Exit, Preds, ".unr1-lcssa", ".unr2-lcssa",
- P, NewBBs);
- }
+ SplitBlockPredecessors(Exit, Preds, ".unr-lcssa", DT, LI,
+ P->mustPreserveAnalysisID(LCSSAID));
// Add the branch to the exit block (around the unrolled loop)
- BranchInst::Create(Exit, NewPH, BrLoopExit, InsertPt);
+ B.CreateCondBr(BrLoopExit, Exit, NewPH);
InsertPt->eraseFromParent();
}
Function *F = Header->getParent();
LoopBlocksDFS::RPOIterator BlockBegin = LoopBlocks.beginRPO();
LoopBlocksDFS::RPOIterator BlockEnd = LoopBlocks.endRPO();
- Loop *NewLoop = 0;
+ Loop *NewLoop = nullptr;
Loop *ParentLoop = L->getParentLoop();
if (!UnrollProlog) {
NewLoop = new Loop();
VMap.erase((*BB)->getTerminator());
BasicBlock *FirstLoopBB = cast<BasicBlock>(VMap[Header]);
BranchInst *LatchBR = cast<BranchInst>(NewBB->getTerminator());
+ IRBuilder<> Builder(LatchBR);
if (UnrollProlog) {
- LatchBR->eraseFromParent();
- BranchInst::Create(InsertBot, NewBB);
+ Builder.CreateBr(InsertBot);
} else {
PHINode *NewIdx = PHINode::Create(NewIter->getType(), 2, "prol.iter",
FirstLoopBB->getFirstNonPHI());
- IRBuilder<> Builder(LatchBR);
Value *IdxSub =
Builder.CreateSub(NewIdx, ConstantInt::get(NewIdx->getType(), 1),
NewIdx->getName() + ".sub");
Value *IdxCmp =
Builder.CreateIsNotNull(IdxSub, NewIdx->getName() + ".cmp");
- BranchInst::Create(FirstLoopBB, InsertBot, IdxCmp, NewBB);
+ Builder.CreateCondBr(IdxCmp, FirstLoopBB, InsertBot);
NewIdx->addIncoming(NewIter, InsertTop);
NewIdx->addIncoming(IdxSub, NewBB);
- LatchBR->eraseFromParent();
}
+ LatchBR->eraseFromParent();
}
}
// Change the incoming values to the ones defined in the preheader or
// cloned loop.
for (BasicBlock::iterator I = Header->begin(); isa<PHINode>(I); ++I) {
- PHINode *NewPHI = cast<PHINode>(VMap[I]);
+ PHINode *NewPHI = cast<PHINode>(VMap[&*I]);
if (UnrollProlog) {
- VMap[I] = NewPHI->getIncomingValueForBlock(Preheader);
+ VMap[&*I] = NewPHI->getIncomingValueForBlock(Preheader);
cast<BasicBlock>(VMap[Header])->getInstList().erase(NewPHI);
} else {
unsigned idx = NewPHI->getBasicBlockIndex(Preheader);
/// ...
/// End:
///
-bool llvm::UnrollRuntimeLoopProlog(Loop *L, unsigned Count, LoopInfo *LI,
+bool llvm::UnrollRuntimeLoopProlog(Loop *L, unsigned Count,
+ bool AllowExpensiveTripCount, LoopInfo *LI,
LPPassManager *LPM) {
// for now, only unroll loops that contain a single exit
if (!L->getExitingBlock())
// loops to be unrolled than relying on induction var simplification
if (!LPM)
return false;
- ScalarEvolution *SE = LPM->getAnalysisIfAvailable<ScalarEvolution>();
- if (!SE)
+ auto *SEWP = LPM->getAnalysisIfAvailable<ScalarEvolutionWrapperPass>();
+ if (!SEWP)
return false;
+ ScalarEvolution &SE = SEWP->getSE();
// Only unroll loops with a computable trip count and the trip count needs
// to be an int value (allowing a pointer type is a TODO item)
- const SCEV *BECount = SE->getBackedgeTakenCount(L);
- if (isa<SCEVCouldNotCompute>(BECount) || !BECount->getType()->isIntegerTy())
+ const SCEV *BECountSC = SE.getBackedgeTakenCount(L);
+ if (isa<SCEVCouldNotCompute>(BECountSC) ||
+ !BECountSC->getType()->isIntegerTy())
return false;
- // If BECount is INT_MAX, we can't compute trip-count without overflow.
- if (BECount->isAllOnesValue())
- return false;
+ unsigned BEWidth = cast<IntegerType>(BECountSC->getType())->getBitWidth();
// Add 1 since the backedge count doesn't include the first loop iteration
const SCEV *TripCountSC =
- SE->getAddExpr(BECount, SE->getConstant(BECount->getType(), 1));
+ SE.getAddExpr(BECountSC, SE.getConstant(BECountSC->getType(), 1));
if (isa<SCEVCouldNotCompute>(TripCountSC))
return false;
+ BasicBlock *Header = L->getHeader();
+ const DataLayout &DL = Header->getModule()->getDataLayout();
+ SCEVExpander Expander(SE, DL, "loop-unroll");
+ if (!AllowExpensiveTripCount && Expander.isHighCostExpansion(TripCountSC, L))
+ return false;
+
// We only handle cases when the unroll factor is a power of 2.
// Count is the loop unroll factor, the number of extra copies added + 1.
- if ((Count & (Count-1)) != 0)
+ if (!isPowerOf2_32(Count))
+ return false;
+
+ // This constraint lets us deal with an overflowing trip count easily; see the
+ // comment on ModVal below.
+ if (Log2_32(Count) > BEWidth)
return false;
// If this loop is nested, then the loop unroller changes the code in
// parent loop, so the Scalar Evolution pass needs to be run again
if (Loop *ParentLoop = L->getParentLoop())
- SE->forgetLoop(ParentLoop);
+ SE.forgetLoop(ParentLoop);
+
+ // Grab analyses that we preserve.
+ auto *DTWP = LPM->getAnalysisIfAvailable<DominatorTreeWrapperPass>();
+ auto *DT = DTWP ? &DTWP->getDomTree() : nullptr;
BasicBlock *PH = L->getLoopPreheader();
- BasicBlock *Header = L->getHeader();
BasicBlock *Latch = L->getLoopLatch();
// It helps to splits the original preheader twice, one for the end of the
// prolog code and one for a new loop preheader
- BasicBlock *PEnd = SplitEdge(PH, Header, LPM->getAsPass());
- BasicBlock *NewPH = SplitBlock(PEnd, PEnd->getTerminator(), LPM->getAsPass());
+ BasicBlock *PEnd = SplitEdge(PH, Header, DT, LI);
+ BasicBlock *NewPH = SplitBlock(PEnd, PEnd->getTerminator(), DT, LI);
BranchInst *PreHeaderBR = cast<BranchInst>(PH->getTerminator());
// Compute the number of extra iterations required, which is:
// extra iterations = run-time trip count % (loop unroll factor + 1)
- SCEVExpander Expander(*SE, "loop-unroll");
Value *TripCount = Expander.expandCodeFor(TripCountSC, TripCountSC->getType(),
PreHeaderBR);
+ Value *BECount = Expander.expandCodeFor(BECountSC, BECountSC->getType(),
+ PreHeaderBR);
IRBuilder<> B(PreHeaderBR);
Value *ModVal = B.CreateAnd(TripCount, Count - 1, "xtraiter");
- // Check if for no extra iterations, then jump to cloned/unrolled loop.
- // We have to check that the trip count computation didn't overflow when
- // adding one to the backedge taken count.
- Value *LCmp = B.CreateIsNotNull(ModVal, "lcmp.mod");
- Value *OverflowCheck = B.CreateIsNull(TripCount, "lcmp.overflow");
- Value *BranchVal = B.CreateOr(OverflowCheck, LCmp, "lcmp.or");
+ // If ModVal is zero, we know that either
+ // 1. there are no iteration to be run in the prologue loop
+ // OR
+ // 2. the addition computing TripCount overflowed
+ //
+ // If (2) is true, we know that TripCount really is (1 << BEWidth) and so the
+ // number of iterations that remain to be run in the original loop is a
+ // multiple Count == (1 << Log2(Count)) because Log2(Count) <= BEWidth (we
+ // explicitly check this above).
+
+ Value *BranchVal = B.CreateIsNotNull(ModVal, "lcmp.mod");
// Branch to either the extra iterations or the cloned/unrolled loop
// We will fix up the true branch label when adding loop body copies
- BranchInst::Create(PEnd, PEnd, BranchVal, PreHeaderBR);
+ B.CreateCondBr(BranchVal, PEnd, PEnd);
assert(PreHeaderBR->isUnconditional() &&
PreHeaderBR->getSuccessor(0) == PEnd &&
"CFG edges in Preheader are not correct");
std::vector<BasicBlock *> NewBlocks;
ValueToValueMapTy VMap;
- // If unroll count is 2 and we can't overflow in tripcount computation (which
- // is BECount + 1), then we don't need a loop for prologue, and we can unroll
- // it. We can be sure that we don't overflow only if tripcount is a constant.
- bool UnrollPrologue = (Count == 2 && isa<ConstantInt>(TripCount));
+ bool UnrollPrologue = Count == 2;
// Clone all the basic blocks in the loop. If Count is 2, we don't clone
// the loop, otherwise we create a cloned loop to execute the extra
VMap, LI);
// Insert the cloned blocks into function just before the original loop
- F->getBasicBlockList().splice(PEnd, F->getBasicBlockList(), NewBlocks[0],
- F->end());
+ F->getBasicBlockList().splice(PEnd->getIterator(), F->getBasicBlockList(),
+ NewBlocks[0]->getIterator(), F->end());
// Rewrite the cloned instruction operands to use the values
// created when the clone is created.
for (BasicBlock::iterator I = NewBlocks[i]->begin(),
E = NewBlocks[i]->end();
I != E; ++I) {
- RemapInstruction(I, VMap,
+ RemapInstruction(&*I, VMap,
RF_NoModuleLevelChanges | RF_IgnoreMissingEntries);
}
}
// Connect the prolog code to the original loop and update the
// PHI functions.
BasicBlock *LastLoopBB = cast<BasicBlock>(VMap[Latch]);
- ConnectProlog(L, TripCount, Count, LastLoopBB, PEnd, PH, NewPH, VMap,
+ ConnectProlog(L, BECount, Count, LastLoopBB, PEnd, PH, NewPH, VMap, DT, LI,
LPM->getAsPass());
NumRuntimeUnrolled++;
return true;
projects / oota-llvm.git / blobdiff