//
// The LLVM Compiler Infrastructure
//
-// This file was developed by the LLVM research group and is distributed under
-// the University of Illinois Open Source License. See LICENSE.TXT for details.
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
//
//===----------------------------------------------------------------------===//
//
#include "llvm/Constants.h"
#include "llvm/Function.h"
#include "llvm/Instructions.h"
+#include "llvm/Analysis/ConstantFolding.h"
#include "llvm/Analysis/LoopInfo.h"
+#include "llvm/Analysis/LoopPass.h"
#include "llvm/Transforms/Utils/Cloning.h"
#include "llvm/Transforms/Utils/Local.h"
+#include "llvm/Support/CFG.h"
+#include "llvm/Support/Compiler.h"
#include "llvm/Support/CommandLine.h"
#include "llvm/Support/Debug.h"
+#include "llvm/Support/MathExtras.h"
#include "llvm/ADT/Statistic.h"
#include "llvm/ADT/STLExtras.h"
+#include "llvm/ADT/SmallPtrSet.h"
#include "llvm/IntrinsicInst.h"
-#include <cstdio>
-#include <set>
#include <algorithm>
-#include <iostream>
+#include <climits>
+#include <cstdio>
using namespace llvm;
+STATISTIC(NumCompletelyUnrolled, "Number of loops completely unrolled");
+STATISTIC(NumUnrolled, "Number of loops unrolled (completely or otherwise)");
+
namespace {
- Statistic<> NumUnrolled("loop-unroll", "Number of loops completely unrolled");
+ cl::opt<unsigned>
+ UnrollThreshold
+ ("unroll-threshold", cl::init(100), cl::Hidden,
+ cl::desc("The cut-off point for automatic loop unrolling"));
cl::opt<unsigned>
- UnrollThreshold("unroll-threshold", cl::init(100), cl::Hidden,
- cl::desc("The cut-off point for loop unrolling"));
+ UnrollCount
+ ("unroll-count", cl::init(0), cl::Hidden,
+ cl::desc("Use this unroll count for all loops, for testing purposes"));
- class LoopUnroll : public FunctionPass {
+ class VISIBILITY_HIDDEN LoopUnroll : public LoopPass {
LoopInfo *LI; // The current loop information
public:
- virtual bool runOnFunction(Function &F);
- bool visitLoop(Loop *L);
+ static char ID; // Pass ID, replacement for typeid
+ LoopUnroll() : LoopPass((intptr_t)&ID) {}
+
+ /// A magic value for use with the Threshold parameter to indicate
+ /// that the loop unroll should be performed regardless of how much
+ /// code expansion would result.
+ static const unsigned NoThreshold = UINT_MAX;
+
+ bool runOnLoop(Loop *L, LPPassManager &LPM);
+ bool unrollLoop(Loop *L, unsigned Count, unsigned Threshold);
+ BasicBlock *FoldBlockIntoPredecessor(BasicBlock *BB);
/// This transformation requires natural loop information & requires that
/// loop preheaders be inserted into the CFG...
AU.addPreserved<LoopInfo>();
}
};
- RegisterOpt<LoopUnroll> X("loop-unroll", "Unroll loops");
+ char LoopUnroll::ID = 0;
+ RegisterPass<LoopUnroll> X("loop-unroll", "Unroll loops");
}
-FunctionPass *llvm::createLoopUnrollPass() { return new LoopUnroll(); }
-
-bool LoopUnroll::runOnFunction(Function &F) {
- bool Changed = false;
- LI = &getAnalysis<LoopInfo>();
-
- // Transform all the top-level loops. Copy the loop list so that the child
- // can update the loop tree if it needs to delete the loop.
- std::vector<Loop*> SubLoops(LI->begin(), LI->end());
- for (unsigned i = 0, e = SubLoops.size(); i != e; ++i)
- Changed |= visitLoop(SubLoops[i]);
+LoopPass *llvm::createLoopUnrollPass() { return new LoopUnroll(); }
- return Changed;
-}
-
-/// ApproximateLoopSize - Approximate the size of the loop after it has been
-/// unrolled.
+/// ApproximateLoopSize - Approximate the size of the loop.
static unsigned ApproximateLoopSize(const Loop *L) {
unsigned Size = 0;
for (unsigned i = 0, e = L->getBlocks().size(); i != e; ++i) {
// Ignore PHI nodes in the header.
} else if (I->hasOneUse() && I->use_back() == Term) {
// Ignore instructions only used by the loop terminator.
- } else if (DbgInfoIntrinsic *DbgI = dyn_cast<DbgInfoIntrinsic>(I)) {
+ } else if (isa<DbgInfoIntrinsic>(I)) {
// Ignore debug instructions
} else {
++Size;
// current values into those specified by ValueMap.
//
static inline void RemapInstruction(Instruction *I,
- std::map<const Value *, Value*> &ValueMap) {
+ DenseMap<const Value *, Value*> &ValueMap) {
for (unsigned op = 0, E = I->getNumOperands(); op != E; ++op) {
Value *Op = I->getOperand(op);
- std::map<const Value *, Value*>::iterator It = ValueMap.find(Op);
+ DenseMap<const Value *, Value*>::iterator It = ValueMap.find(Op);
if (It != ValueMap.end()) Op = It->second;
I->setOperand(op, Op);
}
}
-bool LoopUnroll::visitLoop(Loop *L) {
- bool Changed = false;
+// FoldBlockIntoPredecessor - Folds a basic block into its predecessor if it
+// only has one predecessor, and that predecessor only has one successor.
+// Returns the new combined block.
+BasicBlock *LoopUnroll::FoldBlockIntoPredecessor(BasicBlock *BB) {
+ // 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.
+ //
+ BasicBlock *OnlyPred = BB->getSinglePredecessor();
+ if (!OnlyPred) return 0;
+
+ if (OnlyPred->getTerminator()->getNumSuccessors() != 1)
+ return 0;
+
+ DOUT << "Merging: " << *BB << "into: " << *OnlyPred;
+
+ // Resolve any PHI nodes at the start of the block. They are all
+ // guaranteed to have exactly one entry if they exist, unless there are
+ // multiple duplicate (but guaranteed to be equal) entries for the
+ // incoming edges. This occurs when there are multiple edges from
+ // OnlyPred to OnlySucc.
+ //
+ while (PHINode *PN = dyn_cast<PHINode>(&BB->front())) {
+ PN->replaceAllUsesWith(PN->getIncomingValue(0));
+ BB->getInstList().pop_front(); // Delete the phi node...
+ }
+
+ // 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);
+
+ std::string OldName = BB->getName();
+
+ // Erase basic block from the function...
+ LI->removeBlock(BB);
+ BB->eraseFromParent();
+
+ // Inherit predecessor's name if it exists...
+ if (!OldName.empty() && !OnlyPred->hasName())
+ OnlyPred->setName(OldName);
+
+ return OnlyPred;
+}
+
+bool LoopUnroll::runOnLoop(Loop *L, LPPassManager &LPM) {
+ LI = &getAnalysis<LoopInfo>();
- // Recurse through all subloops before we process this loop. Copy the loop
- // list so that the child can update the loop tree if it needs to delete the
- // loop.
- std::vector<Loop*> SubLoops(L->begin(), L->end());
- for (unsigned i = 0, e = SubLoops.size(); i != e; ++i)
- Changed |= visitLoop(SubLoops[i]);
+ // Unroll the loop.
+ if (!unrollLoop(L, UnrollCount, UnrollThreshold))
+ return false;
- BasicBlock* Header = L->getHeader();
- BasicBlock* LatchBlock = L->getLoopLatch();
+ // Update the loop information for this loop.
+ // If we completely unrolled the loop, remove it from the parent.
+ if (L->getNumBackEdges() == 0)
+ LPM.deleteLoopFromQueue(L);
+
+ return true;
+}
+/// Unroll the given loop by UnrollCount, or by a heuristically-determined
+/// value if Count is zero. If Threshold is not NoThreshold, it is a value
+/// to limit code size expansion. If the loop size would expand beyond the
+/// threshold value, unrolling is suppressed. The return value is true if
+/// any transformations are performed.
+///
+bool LoopUnroll::unrollLoop(Loop *L, unsigned Count, unsigned Threshold) {
+ assert(L->isLCSSAForm());
+
+ BasicBlock *Header = L->getHeader();
+ BasicBlock *LatchBlock = L->getLoopLatch();
BranchInst *BI = dyn_cast<BranchInst>(LatchBlock->getTerminator());
- if (BI == 0) return Changed; // Must end in a conditional branch
-
- ConstantInt *TripCountC = dyn_cast_or_null<ConstantInt>(L->getTripCount());
- if (!TripCountC) return Changed; // Must have constant trip count!
-
- uint64_t TripCountFull = TripCountC->getRawValue();
- if (TripCountFull != TripCountC->getRawValue() || TripCountFull == 0)
- return Changed; // More than 2^32 iterations???
-
- unsigned LoopSize = ApproximateLoopSize(L);
- DEBUG(std::cerr << "Loop Unroll: F[" << Header->getParent()->getName()
- << "] Loop %" << Header->getName() << " Loop Size = "
- << LoopSize << " Trip Count = " << TripCountFull << " - ");
- uint64_t Size = (uint64_t)LoopSize*TripCountFull;
- if (Size > UnrollThreshold) {
- DEBUG(std::cerr << "TOO LARGE: " << Size << ">" << UnrollThreshold << "\n");
- return Changed;
+
+ DOUT << "Loop Unroll: F[" << Header->getParent()->getName()
+ << "] Loop %" << Header->getName() << "\n";
+
+ if (!BI || BI->isUnconditional()) {
+ // The loop-rotate pass can be helpful to avoid this in many cases.
+ DOUT << " Can't unroll; loop not terminated by a conditional branch.\n";
+ return false;
}
- DEBUG(std::cerr << "UNROLLING!\n");
- std::vector<BasicBlock*> LoopBlocks = L->getBlocks();
+ // Determine the trip count and/or trip multiple. A TripCount value of zero
+ // is used to mean an unknown trip count. The TripMultiple value is the
+ // greatest known integer multiple of the trip count.
+ unsigned TripCount = 0;
+ unsigned TripMultiple = 1;
+ if (Value *TripCountValue = L->getTripCount()) {
+ if (ConstantInt *TripCountC = dyn_cast<ConstantInt>(TripCountValue)) {
+ // Guard against huge trip counts. This also guards against assertions in
+ // APInt from the use of getZExtValue, below.
+ if (TripCountC->getValue().getActiveBits() <= 32) {
+ TripCount = (unsigned)TripCountC->getZExtValue();
+ }
+ } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TripCountValue)) {
+ switch (BO->getOpcode()) {
+ case BinaryOperator::Mul:
+ if (ConstantInt *MultipleC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
+ if (MultipleC->getValue().getActiveBits() <= 32) {
+ TripMultiple = (unsigned)MultipleC->getZExtValue();
+ }
+ }
+ break;
+ default: break;
+ }
+ }
+ }
+ if (TripCount != 0)
+ DOUT << " Trip Count = " << TripCount << "\n";
+ if (TripMultiple != 1)
+ DOUT << " Trip Multiple = " << TripMultiple << "\n";
+
+ // 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
+ // don't unroll.
+ if (TripCount != 0) {
+ Count = TripCount;
+ } else {
+ return false;
+ }
+ }
- unsigned TripCount = (unsigned)TripCountFull;
+ // Effectively "DCE" unrolled iterations that are beyond the tripcount
+ // and will never be executed.
+ if (TripCount != 0 && Count > TripCount)
+ Count = TripCount;
+
+ assert(Count > 0);
+ assert(TripMultiple > 0);
+ assert(TripCount == 0 || TripCount % TripMultiple == 0);
+
+ // Enforce the threshold.
+ if (Threshold != NoThreshold) {
+ unsigned LoopSize = ApproximateLoopSize(L);
+ DOUT << " Loop Size = " << LoopSize << "\n";
+ uint64_t Size = (uint64_t)LoopSize*Count;
+ if (TripCount != 1 && Size > Threshold) {
+ DOUT << " TOO LARGE TO UNROLL: "
+ << Size << ">" << Threshold << "\n";
+ return false;
+ }
+ }
- BasicBlock *LoopExit = BI->getSuccessor(L->contains(BI->getSuccessor(0)));
+ // Are we eliminating the loop control altogether?
+ bool CompletelyUnroll = Count == TripCount;
+
+ // If we know the trip count, we know the multiple...
+ unsigned BreakoutTrip = 0;
+ if (TripCount != 0) {
+ BreakoutTrip = TripCount % Count;
+ TripMultiple = 0;
+ } else {
+ // Figure out what multiple to use.
+ BreakoutTrip = TripMultiple =
+ (unsigned)GreatestCommonDivisor64(Count, TripMultiple);
+ }
+
+ if (CompletelyUnroll) {
+ DOUT << "COMPLETELY UNROLLING loop %" << Header->getName()
+ << " with trip count " << TripCount << "!\n";
+ } else {
+ DOUT << "UNROLLING loop %" << Header->getName()
+ << " by " << Count;
+ if (TripMultiple == 0 || BreakoutTrip != TripMultiple) {
+ DOUT << " with a breakout at trip " << BreakoutTrip;
+ } else if (TripMultiple != 1) {
+ DOUT << " with " << TripMultiple << " trips per branch";
+ }
+ DOUT << "!\n";
+ }
+
+ std::vector<BasicBlock*> LoopBlocks = L->getBlocks();
+
+ bool ContinueOnTrue = L->contains(BI->getSuccessor(0));
+ BasicBlock *LoopExit = BI->getSuccessor(ContinueOnTrue);
// For the first iteration of the loop, we should use the precloned values for
// PHI nodes. Insert associations now.
- std::map<const Value*, Value*> LastValueMap;
+ typedef DenseMap<const Value*, Value*> ValueMapTy;
+ ValueMapTy LastValueMap;
std::vector<PHINode*> OrigPHINode;
for (BasicBlock::iterator I = Header->begin(); isa<PHINode>(I); ++I) {
PHINode *PN = cast<PHINode>(I);
LastValueMap[I] = I;
}
- // Remove the exit branch from the loop
- LatchBlock->getInstList().erase(BI);
-
std::vector<BasicBlock*> Headers;
std::vector<BasicBlock*> Latches;
Headers.push_back(Header);
Latches.push_back(LatchBlock);
- assert(TripCount != 0 && "Trip count of 0 is impossible!");
- for (unsigned It = 1; It != TripCount; ++It) {
+ for (unsigned It = 1; It != Count; ++It) {
char SuffixBuffer[100];
sprintf(SuffixBuffer, ".%d", It);
for (std::vector<BasicBlock*>::iterator BB = LoopBlocks.begin(),
E = LoopBlocks.end(); BB != E; ++BB) {
- std::map<const Value*, Value*> ValueMap;
+ ValueMapTy ValueMap;
BasicBlock *New = CloneBasicBlock(*BB, ValueMap, SuffixBuffer);
Header->getParent()->getBasicBlockList().push_back(New);
// Update our running map of newest clones
LastValueMap[*BB] = New;
- for (std::map<const Value*, Value*>::iterator VI = ValueMap.begin(),
- VE = ValueMap.end(); VI != VE; ++VI)
+ for (ValueMapTy::iterator VI = ValueMap.begin(), VE = ValueMap.end();
+ VI != VE; ++VI)
LastValueMap[VI->first] = VI->second;
- L->addBasicBlockToLoop(New, *LI);
+ 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; ++UI) {
- Instruction* UseInst = cast<Instruction>(*UI);
+ UI != UE;) {
+ Instruction *UseInst = cast<Instruction>(*UI);
+ ++UI;
if (isa<PHINode>(UseInst) && !L->contains(UseInst->getParent())) {
- PHINode* phi = cast<PHINode>(UseInst);
- Value* Incoming = phi->getIncomingValueForBlock(*BB);
- if (isa<Instruction>(Incoming))
- Incoming = LastValueMap[Incoming];
-
+ PHINode *phi = cast<PHINode>(UseInst);
+ Value *Incoming = phi->getIncomingValueForBlock(*BB);
phi->addIncoming(Incoming, New);
}
}
// 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);
}
E = NewBlocks[i]->end(); I != E; ++I)
RemapInstruction(I, LastValueMap);
}
-
- // Insert the branches that link the different iterations together
- for (unsigned i = 0; i < Latches.size()-1; ++i)
- new BranchInst(Headers[i+1], Latches[i]);
- // Finally, add an unconditional branch to the block to continue into the exit
- // block.
- new BranchInst(LoopExit, Latches[Latches.size()-1]);
-
- // Update PHI nodes that reference the final latch block
- if (TripCount > 1) {
- std::set<PHINode*> Users;
+ // 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))
+ if (PHINode *phi = dyn_cast<PHINode>(*UI))
Users.insert(phi);
-
- for (std::set<PHINode*>::iterator SI = Users.begin(), SE = Users.end();
+
+ BasicBlock *LastIterationBB = cast<BasicBlock>(LastValueMap[LatchBlock]);
+ for (SmallPtrSet<PHINode*,8>::iterator SI = Users.begin(), SE = Users.end();
SI != SE; ++SI) {
- Value* InVal = (*SI)->getIncomingValueForBlock(LatchBlock);
- if (isa<Instruction>(InVal))
- InVal = LastValueMap[InVal];
- (*SI)->removeIncomingValue(LatchBlock, false);
- (*SI)->addIncoming(InVal, cast<BasicBlock>(LastValueMap[LatchBlock]));
+ PHINode *PN = *SI;
+ 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 (Instruction *InValI = dyn_cast<Instruction>(InVal)) {
+ if (L->contains(InValI->getParent()))
+ 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);
}
}
- // Now loop over the PHI nodes in the original block, setting them to their
- // incoming values.
- 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);
- }
+ // Now that all the basic blocks for the unrolled iterations are in place,
+ // set up the branches to connect them.
+ for (unsigned i = 0, e = Latches.size(); i != e; ++i) {
+ // The original branch was replicated in each unrolled iteration.
+ BranchInst *Term = cast<BranchInst>(Latches[i]->getTerminator());
+
+ // The branch destination.
+ unsigned j = (i + 1) % e;
+ BasicBlock *Dest = Headers[j];
+ bool NeedConditional = true;
+
+ // For a complete unroll, make the last iteration end with a branch
+ // to the exit block.
+ if (CompletelyUnroll && j == 0) {
+ Dest = LoopExit;
+ NeedConditional = false;
+ }
+
+ // If we know the trip count or a multiple of it, we can safely use an
+ // unconditional branch for some iterations.
+ if (j != BreakoutTrip && (TripMultiple == 0 || j % TripMultiple != 0)) {
+ NeedConditional = false;
+ }
+ if (NeedConditional) {
+ // Update the conditional branch's successor for the following
+ // iteration.
+ Term->setSuccessor(!ContinueOnTrue, Dest);
+ } else {
+ Term->setUnconditionalDest(Dest);
+ // Merge adjacent basic blocks, if possible.
+ if (BasicBlock *Fold = FoldBlockIntoPredecessor(Dest)) {
+ std::replace(Latches.begin(), Latches.end(), Dest, Fold);
+ std::replace(Headers.begin(), Headers.end(), Dest, Fold);
+ }
+ }
+ }
+
// 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.
const std::vector<BasicBlock*> &NewLoopBlocks = L->getBlocks();
for (std::vector<BasicBlock*>::const_iterator BB = NewLoopBlocks.begin(),
- E = NewLoopBlocks.end(); BB != E; ++BB)
+ BBE = NewLoopBlocks.end(); BB != BBE; ++BB)
for (BasicBlock::iterator I = (*BB)->begin(), E = (*BB)->end(); I != E; ) {
Instruction *Inst = I++;
}
}
- // Update the loop information for this loop.
- Loop *Parent = L->getParentLoop();
-
- // Move all of the basic blocks in the loop into the parent loop.
- for (std::vector<BasicBlock*>::const_iterator BB = NewLoopBlocks.begin(),
- E = NewLoopBlocks.end(); BB != E; ++BB)
- LI->changeLoopFor(*BB, Parent);
-
- // Remove the loop from the parent.
- if (Parent)
- delete Parent->removeChildLoop(std::find(Parent->begin(), Parent->end(),L));
- else
- delete LI->removeLoop(std::find(LI->begin(), LI->end(), L));
-
+ NumCompletelyUnrolled += CompletelyUnroll;
++NumUnrolled;
return true;
}