// actual pass or policy, but provides a single function to perform loop
// unrolling.
//
-// It works best when loops have been canonicalized by the -indvars pass,
-// allowing it to determine the trip counts of loops easily.
-//
// The process of unrolling can produce extraneous basic blocks linked with
// unconditional branches. This will be corrected in the future.
+//
//===----------------------------------------------------------------------===//
#define DEBUG_TYPE "loop-unroll"
#include "llvm/Transforms/Utils/UnrollLoop.h"
#include "llvm/BasicBlock.h"
#include "llvm/ADT/Statistic.h"
-#include "llvm/Analysis/ConstantFolding.h"
+#include "llvm/Analysis/InstructionSimplify.h"
+#include "llvm/Analysis/LoopIterator.h"
#include "llvm/Analysis/LoopPass.h"
#include "llvm/Analysis/ScalarEvolution.h"
#include "llvm/Support/Debug.h"
#include "llvm/Transforms/Utils/BasicBlockUtils.h"
#include "llvm/Transforms/Utils/Cloning.h"
#include "llvm/Transforms/Utils/Local.h"
-
+#include "llvm/Transforms/Utils/SimplifyIndVar.h"
using namespace llvm;
// TODO: Should these be here or in LoopUnroll?
STATISTIC(NumCompletelyUnrolled, "Number of loops completely unrolled");
-STATISTIC(NumUnrolled, "Number of loops unrolled (completely or otherwise)");
+STATISTIC(NumUnrolled, "Number of loops unrolled (completely or otherwise)");
/// RemapInstruction - Convert the instruction operands from referencing the
/// current values into those specified by VMap.
if (It != VMap.end())
I->setOperand(op, It->second);
}
+
+ if (PHINode *PN = dyn_cast<PHINode>(I)) {
+ for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
+ ValueToValueMapTy::iterator It = VMap.find(PN->getIncomingBlock(i));
+ if (It != VMap.end())
+ PN->setIncomingBlock(i, cast<BasicBlock>(It->second));
+ }
+ }
}
/// FoldBlockIntoPredecessor - Folds a basic block into its predecessor if it
/// only has one predecessor, and that predecessor only has one successor.
/// The LoopInfo Analysis that is passed will be kept consistent.
/// Returns the new combined block.
-static BasicBlock *FoldBlockIntoPredecessor(BasicBlock *BB, LoopInfo* LI) {
+static BasicBlock *FoldBlockIntoPredecessor(BasicBlock *BB, LoopInfo* LI,
+ LPPassManager *LPM) {
// Merge basic blocks into their predecessor if there is only one distinct
// pred, and if there is only one distinct successor of the predecessor, and
// if there are no PHI nodes.
// Delete the unconditional branch from the predecessor...
OnlyPred->getInstList().pop_back();
- // Move all definitions in the successor to the predecessor...
- OnlyPred->getInstList().splice(OnlyPred->end(), BB->getInstList());
-
// Make all PHI nodes that referred to BB now refer to Pred as their
// source...
BB->replaceAllUsesWith(OnlyPred);
+ // Move all definitions in the successor to the predecessor...
+ OnlyPred->getInstList().splice(OnlyPred->end(), BB->getInstList());
+
std::string OldName = BB->getName();
// Erase basic block from the function...
+
+ // ScalarEvolution holds references to loop exit blocks.
+ if (ScalarEvolution *SE = LPM->getAnalysisIfAvailable<ScalarEvolution>()) {
+ if (Loop *L = LI->getLoopFor(BB))
+ SE->forgetLoop(L);
+ }
LI->removeBlock(BB);
BB->eraseFromParent();
}
/// Unroll the given loop by Count. The loop must be in LCSSA form. Returns true
-/// if unrolling was succesful, or false if the loop was unmodified. Unrolling
+/// if unrolling was successful, or false if the loop was unmodified. Unrolling
/// can only fail when the loop's latch block is not terminated by a conditional
/// branch instruction. However, if the trip count (and multiple) are not known,
/// loop unrolling will mostly produce more code that is no faster.
///
+/// TripCount is generally defined as the number of times the loop header
+/// executes. UnrollLoop relaxes the definition to permit early exits: here
+/// TripCount is the iteration on which control exits LatchBlock if no early
+/// exits were taken. Note that UnrollLoop assumes that the loop counter test
+/// terminates LatchBlock in order to remove unnecesssary instances of the
+/// test. In other words, control may exit the loop prior to TripCount
+/// iterations via an early branch, but control may not exit the loop from the
+/// LatchBlock's terminator prior to TripCount iterations.
+///
+/// Similarly, TripMultiple divides the number of times that the LatchBlock may
+/// execute without exiting the loop.
+///
/// The LoopInfo Analysis that is passed will be kept consistent.
///
/// If a LoopPassManager is passed in, and the loop is fully removed, it will be
/// removed from the LoopPassManager as well. LPM can also be NULL.
-bool llvm::UnrollLoop(Loop *L, unsigned Count, LoopInfo* LI, LPPassManager* LPM) {
+///
+/// This utility preserves LoopInfo. If DominatorTree or ScalarEvolution are
+/// available it must also preserve those analyses.
+bool llvm::UnrollLoop(Loop *L, unsigned Count, unsigned TripCount,
+ bool AllowRuntime, unsigned TripMultiple,
+ LoopInfo *LI, LPPassManager *LPM) {
BasicBlock *Preheader = L->getLoopPreheader();
if (!Preheader) {
DEBUG(dbgs() << " Can't unroll; loop preheader-insertion failed.\n");
BasicBlock *Header = L->getHeader();
BranchInst *BI = dyn_cast<BranchInst>(LatchBlock->getTerminator());
-
+
if (!BI || BI->isUnconditional()) {
// The loop-rotate pass can be helpful to avoid this in many cases.
DEBUG(dbgs() <<
return false;
}
- // Notify ScalarEvolution that the loop will be substantially changed,
- // if not outright eliminated.
- if (ScalarEvolution *SE = LPM->getAnalysisIfAvailable<ScalarEvolution>())
- SE->forgetLoop(L);
-
- // Find trip count
- unsigned TripCount = L->getSmallConstantTripCount();
- // Find trip multiple if count is not available
- unsigned TripMultiple = 1;
- if (TripCount == 0)
- TripMultiple = L->getSmallConstantTripMultiple();
+ if (Header->hasAddressTaken()) {
+ // The loop-rotate pass can be helpful to avoid this in many cases.
+ DEBUG(dbgs() <<
+ " Won't unroll loop: address of header block is taken.\n");
+ return false;
+ }
if (TripCount != 0)
DEBUG(dbgs() << " Trip Count = " << TripCount << "\n");
if (TripCount != 0 && Count > TripCount)
Count = TripCount;
+ // Don't enter the unroll code if there is nothing to do. This way we don't
+ // need to support "partial unrolling by 1".
+ if (TripCount == 0 && Count < 2)
+ return false;
+
assert(Count > 0);
assert(TripMultiple > 0);
assert(TripCount == 0 || TripCount % TripMultiple == 0);
// Are we eliminating the loop control altogether?
bool CompletelyUnroll = Count == TripCount;
+ // We assume a run-time trip count if the compiler cannot
+ // figure out the loop trip count and the unroll-runtime
+ // flag is specified.
+ bool RuntimeTripCount = (TripCount == 0 && Count > 0 && AllowRuntime);
+
+ if (RuntimeTripCount && !UnrollRuntimeLoopProlog(L, Count, LI, LPM))
+ return false;
+
+ // Notify ScalarEvolution that the loop will be substantially changed,
+ // if not outright eliminated.
+ ScalarEvolution *SE = LPM->getAnalysisIfAvailable<ScalarEvolution>();
+ if (SE)
+ SE->forgetLoop(L);
+
// If we know the trip count, we know the multiple...
unsigned BreakoutTrip = 0;
if (TripCount != 0) {
DEBUG(dbgs() << " with a breakout at trip " << BreakoutTrip);
} else if (TripMultiple != 1) {
DEBUG(dbgs() << " with " << TripMultiple << " trips per branch");
+ } else if (RuntimeTripCount) {
+ DEBUG(dbgs() << " with run-time trip count");
}
DEBUG(dbgs() << "!\n");
}
ValueToValueMapTy LastValueMap;
std::vector<PHINode*> OrigPHINode;
for (BasicBlock::iterator I = Header->begin(); isa<PHINode>(I); ++I) {
- PHINode *PN = cast<PHINode>(I);
- OrigPHINode.push_back(PN);
- if (Instruction *I =
- dyn_cast<Instruction>(PN->getIncomingValueForBlock(LatchBlock)))
- if (L->contains(I))
- LastValueMap[I] = I;
+ OrigPHINode.push_back(cast<PHINode>(I));
}
std::vector<BasicBlock*> Headers;
Headers.push_back(Header);
Latches.push_back(LatchBlock);
+ // The current on-the-fly SSA update requires blocks to be processed in
+ // reverse postorder so that LastValueMap contains the correct value at each
+ // exit.
+ LoopBlocksDFS DFS(L);
+ DFS.perform(LI);
+
+ // Stash the DFS iterators before adding blocks to the loop.
+ LoopBlocksDFS::RPOIterator BlockBegin = DFS.beginRPO();
+ LoopBlocksDFS::RPOIterator BlockEnd = DFS.endRPO();
+
for (unsigned It = 1; It != Count; ++It) {
std::vector<BasicBlock*> NewBlocks;
-
- for (std::vector<BasicBlock*>::iterator BB = LoopBlocks.begin(),
- E = LoopBlocks.end(); BB != E; ++BB) {
+
+ for (LoopBlocksDFS::RPOIterator BB = BlockBegin; BB != BlockEnd; ++BB) {
ValueToValueMapTy VMap;
BasicBlock *New = CloneBasicBlock(*BB, VMap, "." + Twine(It));
Header->getParent()->getBasicBlockList().push_back(New);
L->addBasicBlockToLoop(New, LI->getBase());
- // Add phi entries for newly created values to all exit blocks except
- // the successor of the latch block. The successor of the exit block will
- // be updated specially after unrolling all the way.
- if (*BB != LatchBlock)
- for (Value::use_iterator UI = (*BB)->use_begin(), UE = (*BB)->use_end();
- UI != UE;) {
- Instruction *UseInst = cast<Instruction>(*UI);
- ++UI;
- if (isa<PHINode>(UseInst) && !L->contains(UseInst)) {
- PHINode *phi = cast<PHINode>(UseInst);
- Value *Incoming = phi->getIncomingValueForBlock(*BB);
- phi->addIncoming(Incoming, New);
- }
+ // Add phi entries for newly created values to all exit blocks.
+ for (succ_iterator SI = succ_begin(*BB), SE = succ_end(*BB);
+ SI != SE; ++SI) {
+ if (L->contains(*SI))
+ continue;
+ for (BasicBlock::iterator BBI = (*SI)->begin();
+ PHINode *phi = dyn_cast<PHINode>(BBI); ++BBI) {
+ Value *Incoming = phi->getIncomingValueForBlock(*BB);
+ ValueToValueMapTy::iterator It = LastValueMap.find(Incoming);
+ if (It != LastValueMap.end())
+ Incoming = It->second;
+ phi->addIncoming(Incoming, New);
}
-
+ }
// Keep track of new headers and latches as we create them, so that
// we can insert the proper branches later.
if (*BB == Header)
Headers.push_back(New);
- if (*BB == LatchBlock) {
+ if (*BB == LatchBlock)
Latches.push_back(New);
- // Also, clear out the new latch's back edge so that it doesn't look
- // like a new loop, so that it's amenable to being merged with adjacent
- // blocks later on.
- TerminatorInst *Term = New->getTerminator();
- assert(L->contains(Term->getSuccessor(!ContinueOnTrue)));
- assert(Term->getSuccessor(ContinueOnTrue) == LoopExit);
- Term->setSuccessor(!ContinueOnTrue, NULL);
- }
-
NewBlocks.push_back(New);
}
-
+
// Remap all instructions in the most recent iteration
for (unsigned i = 0; i < NewBlocks.size(); ++i)
for (BasicBlock::iterator I = NewBlocks[i]->begin(),
E = NewBlocks[i]->end(); I != E; ++I)
::RemapInstruction(I, LastValueMap);
}
-
- // The latch block exits the loop. If there are any PHI nodes in the
- // successor blocks, update them to use the appropriate values computed as the
- // last iteration of the loop.
- if (Count != 1) {
- SmallPtrSet<PHINode*, 8> Users;
- for (Value::use_iterator UI = LatchBlock->use_begin(),
- UE = LatchBlock->use_end(); UI != UE; ++UI)
- if (PHINode *phi = dyn_cast<PHINode>(*UI))
- Users.insert(phi);
-
- BasicBlock *LastIterationBB = cast<BasicBlock>(LastValueMap[LatchBlock]);
- for (SmallPtrSet<PHINode*,8>::iterator SI = Users.begin(), SE = Users.end();
- SI != SE; ++SI) {
- PHINode *PN = *SI;
+
+ // Loop over the PHI nodes in the original block, setting incoming values.
+ for (unsigned i = 0, e = OrigPHINode.size(); i != e; ++i) {
+ PHINode *PN = OrigPHINode[i];
+ if (CompletelyUnroll) {
+ PN->replaceAllUsesWith(PN->getIncomingValueForBlock(Preheader));
+ Header->getInstList().erase(PN);
+ }
+ else if (Count > 1) {
Value *InVal = PN->removeIncomingValue(LatchBlock, false);
// If this value was defined in the loop, take the value defined by the
// last iteration of the loop.
if (L->contains(InValI))
InVal = LastValueMap[InVal];
}
- PN->addIncoming(InVal, LastIterationBB);
- }
- }
-
- // Now, if we're doing complete unrolling, loop over the PHI nodes in the
- // original block, setting them to their incoming values.
- if (CompletelyUnroll) {
- BasicBlock *Preheader = L->getLoopPreheader();
- for (unsigned i = 0, e = OrigPHINode.size(); i != e; ++i) {
- PHINode *PN = OrigPHINode[i];
- PN->replaceAllUsesWith(PN->getIncomingValueForBlock(Preheader));
- Header->getInstList().erase(PN);
+ assert(Latches.back() == LastValueMap[LatchBlock] && "bad last latch");
+ PN->addIncoming(InVal, Latches.back());
}
}
BasicBlock *Dest = Headers[j];
bool NeedConditional = true;
+ if (RuntimeTripCount && j != 0) {
+ NeedConditional = false;
+ }
+
// For a complete unroll, make the last iteration end with a branch
// to the exit block.
if (CompletelyUnroll && j == 0) {
// iteration.
Term->setSuccessor(!ContinueOnTrue, Dest);
} else {
- Term->setUnconditionalDest(Dest);
- // Merge adjacent basic blocks, if possible.
- if (BasicBlock *Fold = FoldBlockIntoPredecessor(Dest, LI)) {
- std::replace(Latches.begin(), Latches.end(), Dest, Fold);
- std::replace(Headers.begin(), Headers.end(), Dest, Fold);
+ // Remove phi operands at this loop exit
+ if (Dest != LoopExit) {
+ BasicBlock *BB = Latches[i];
+ for (succ_iterator SI = succ_begin(BB), SE = succ_end(BB);
+ SI != SE; ++SI) {
+ if (*SI == Headers[i])
+ continue;
+ for (BasicBlock::iterator BBI = (*SI)->begin();
+ PHINode *Phi = dyn_cast<PHINode>(BBI); ++BBI) {
+ Phi->removeIncomingValue(BB, false);
+ }
+ }
}
+ // Replace the conditional branch with an unconditional one.
+ BranchInst::Create(Dest, Term);
+ Term->eraseFromParent();
+ }
+ }
+
+ // Merge adjacent basic blocks, if possible.
+ for (unsigned i = 0, e = Latches.size(); i != e; ++i) {
+ BranchInst *Term = cast<BranchInst>(Latches[i]->getTerminator());
+ if (Term->isUnconditional()) {
+ BasicBlock *Dest = Term->getSuccessor(0);
+ if (BasicBlock *Fold = FoldBlockIntoPredecessor(Dest, LI, LPM))
+ std::replace(Latches.begin(), Latches.end(), Dest, Fold);
}
}
-
+
+ // FIXME: Reconstruct dom info, because it is not preserved properly.
+ // Incrementally updating domtree after loop unrolling would be easy.
+ if (DominatorTree *DT = LPM->getAnalysisIfAvailable<DominatorTree>())
+ DT->runOnFunction(*L->getHeader()->getParent());
+
+ // Simplify any new induction variables in the partially unrolled loop.
+ if (SE && !CompletelyUnroll) {
+ SmallVector<WeakVH, 16> DeadInsts;
+ simplifyLoopIVs(L, SE, LPM, DeadInsts);
+
+ // Aggressively clean up dead instructions that simplifyLoopIVs already
+ // identified. Any remaining should be cleaned up below.
+ while (!DeadInsts.empty())
+ if (Instruction *Inst =
+ dyn_cast_or_null<Instruction>(&*DeadInsts.pop_back_val()))
+ RecursivelyDeleteTriviallyDeadInstructions(Inst);
+ }
+
// At this point, the code is well formed. We now do a quick sweep over the
// inserted code, doing constant propagation and dead code elimination as we
// go.
if (isInstructionTriviallyDead(Inst))
(*BB)->getInstList().erase(Inst);
- else if (Constant *C = ConstantFoldInstruction(Inst)) {
- Inst->replaceAllUsesWith(C);
- (*BB)->getInstList().erase(Inst);
- }
+ else if (Value *V = SimplifyInstruction(Inst))
+ if (LI->replacementPreservesLCSSAForm(Inst, V)) {
+ Inst->replaceAllUsesWith(V);
+ (*BB)->getInstList().erase(Inst);
+ }
}
NumCompletelyUnrolled += CompletelyUnroll;