llvm/lib/Transforms/Utils/UnrollLoop.cpp

389 lines
15 KiB
C++
Raw Normal View History

//===-- UnrollLoop.cpp - Loop unrolling utilities -------------------------===//
//
// The LLVM Compiler Infrastructure
//
// This file is distributed under the University of Illinois Open Source
// License. See LICENSE.TXT for details.
//
//===----------------------------------------------------------------------===//
//
// This file implements some loop unrolling utilities. It does not define any
// 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/ADT/STLExtras.h"
#include "llvm/Analysis/ConstantFolding.h"
#include "llvm/Analysis/LoopPass.h"
#include "llvm/Support/Debug.h"
#include "llvm/Transforms/Utils/Cloning.h"
#include "llvm/Transforms/Utils/Local.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)");
/// RemapInstruction - Convert the instruction operands from referencing the
/// current values into those specified by ValueMap.
static inline void RemapInstruction(Instruction *I,
DenseMap<const Value *, Value*> &ValueMap) {
for (unsigned op = 0, E = I->getNumOperands(); op != E; ++op) {
Value *Op = I->getOperand(op);
DenseMap<const Value *, Value*>::iterator It = ValueMap.find(Op);
if (It != ValueMap.end()) Op = It->second;
I->setOperand(op, Op);
}
}
/// 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) {
// 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;
}
/// 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
/// 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.
///
/// 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) {
assert(L->isLCSSAForm());
BasicBlock *Header = L->getHeader();
BasicBlock *LatchBlock = L->getLoopLatch();
BranchInst *BI = dyn_cast<BranchInst>(LatchBlock->getTerminator());
Function *Func = Header->getParent();
Function::iterator BBInsertPt = next(Function::iterator(LatchBlock));
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;
}
// 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 (TripCount != 0)
DOUT << " Trip Count = " << TripCount << "\n";
if (TripMultiple != 1)
DOUT << " Trip Multiple = " << TripMultiple << "\n";
// 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);
// 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";
}
// Make a copy of the original LoopBlocks list so we can keep referring
// to it while hacking on the loop.
std::vector<BasicBlock*> LoopBlocks = L->getBlocks();
bool ContinueOnTrue = BI->getSuccessor(0) == Header;
BasicBlock *LoopExit = BI->getSuccessor(ContinueOnTrue);
// For the first iteration of the loop, we should use the precloned values for
// PHI nodes. Insert associations now.
typedef DenseMap<const Value*, Value*> ValueMapTy;
ValueMapTy LastValueMap;
for (BasicBlock::iterator I = Header->begin(); isa<PHINode>(I); ++I) {
PHINode *PN = cast<PHINode>(I);
if (Instruction *I =
dyn_cast<Instruction>(PN->getIncomingValueForBlock(LatchBlock)))
if (L->contains(I->getParent()))
LastValueMap[I] = I;
}
// Keep track of all the headers and latches that we create. These are
// needed by the logic that inserts the branches to connect all the
// new blocks.
std::vector<BasicBlock*> Headers;
std::vector<BasicBlock*> Latches;
Headers.reserve(Count);
Latches.reserve(Count);
Headers.push_back(Header);
Latches.push_back(LatchBlock);
// Iterate through all but the first iterations, cloning blocks from
// the first iteration to populate the subsequent iterations.
for (unsigned It = 1; It != Count; ++It) {
char SuffixBuffer[100];
sprintf(SuffixBuffer, ".%d", It);
std::vector<BasicBlock*> NewBlocks;
NewBlocks.reserve(LoopBlocks.size());
// Iterate through all the blocks in the original loop.
for (std::vector<BasicBlock*>::const_iterator BBI = LoopBlocks.begin(),
E = LoopBlocks.end(); BBI != E; ++BBI) {
bool SuppressExitEdges = false;
BasicBlock *BB = *BBI;
ValueMapTy ValueMap;
BasicBlock *New = CloneBasicBlock(BB, ValueMap, SuffixBuffer);
NewBlocks.push_back(New);
Func->getBasicBlockList().insert(BBInsertPt, New);
L->addBasicBlockToLoop(New, LI->getBase());
// Special handling for the loop header block.
if (BB == Header) {
// Keep track of new headers as we create them, so that we can insert
// the proper branches later.
Headers[It] = New;
// Loop over all of the PHI nodes in the block, changing them to use
// the incoming values from the previous block.
for (BasicBlock::iterator I = Header->begin(); isa<PHINode>(I); ++I) {
PHINode *NewPHI = cast<PHINode>(ValueMap[I]);
Value *InVal = NewPHI->getIncomingValueForBlock(LatchBlock);
if (Instruction *InValI = dyn_cast<Instruction>(InVal))
if (It > 1 && L->contains(InValI->getParent()))
InVal = LastValueMap[InValI];
ValueMap[I] = InVal;
New->getInstList().erase(NewPHI);
}
}
// Special handling for the loop latch block.
if (BB == LatchBlock) {
// Keep track of new latches as we create them, so that we can insert
// the proper branches later.
Latches[It] = New;
// If knowledge of the trip count and/or multiple will allow us
// to emit unconditional branches in some of the new latch blocks,
// those blocks shouldn't be referenced by PHIs that reference
// the original latch.
unsigned NextIt = (It + 1) % Count;
SuppressExitEdges =
NextIt != BreakoutTrip &&
(TripMultiple == 0 || NextIt % TripMultiple != 0);
}
// Update our running map of newest clones
LastValueMap[BB] = New;
for (ValueMapTy::iterator VI = ValueMap.begin(), VE = ValueMap.end();
VI != VE; ++VI)
LastValueMap[VI->first] = VI->second;
// Add incoming values to phi nodes that reference this block. The last
// latch block may need to be referenced by the first header, and any
// block with an exit edge may be referenced from outside the loop.
for (Value::use_iterator UI = BB->use_begin(), UE = BB->use_end();
UI != UE; ) {
PHINode *PN = dyn_cast<PHINode>(*UI++);
if (PN &&
((BB == LatchBlock && It == Count - 1 && !CompletelyUnroll) ||
(!SuppressExitEdges && !L->contains(PN->getParent())))) {
Value *InVal = PN->getIncomingValueForBlock(BB);
// If this value was defined in the loop, take the value defined
// by the last iteration of the loop.
ValueMapTy::iterator VI = LastValueMap.find(InVal);
if (VI != LastValueMap.end())
InVal = VI->second;
PN->addIncoming(InVal, New);
}
}
}
// Remap all instructions in the most recent iteration
for (unsigned i = 0, e = NewBlocks.size(); i != e; ++i)
for (BasicBlock::iterator I = NewBlocks[i]->begin(),
E = NewBlocks[i]->end(); I != E; ++I)
RemapInstruction(I, LastValueMap);
}
// Now that all the basic blocks for the unrolled iterations are in place,
// set up the branches to connect them.
for (unsigned It = 0; It != Count; ++It) {
// The original branch was replicated in each unrolled iteration.
BranchInst *Term = cast<BranchInst>(Latches[It]->getTerminator());
// The branch destination.
unsigned NextIt = (It + 1) % Count;
BasicBlock *Dest = Headers[NextIt];
bool NeedConditional = true;
bool HasExit = true;
// For a complete unroll, make the last iteration end with an
// unconditional branch to the exit block.
if (CompletelyUnroll && NextIt == 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 (NextIt != BreakoutTrip &&
(TripMultiple == 0 || NextIt % TripMultiple != 0)) {
NeedConditional = false;
HasExit = 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, LI)) {
std::replace(Latches.begin(), Latches.end(), Dest, Fold);
std::replace(Headers.begin(), Headers.end(), Dest, Fold);
}
}
// Special handling for the first iteration. If the first latch is
// now unconditionally branching to the second header, then it is
// no longer an exit node. Delete PHI references to it both from
// the first header and from outsie the loop.
if (It == 0)
for (Value::use_iterator UI = LatchBlock->use_begin(),
UE = LatchBlock->use_end(); UI != UE; ) {
PHINode *PN = dyn_cast<PHINode>(*UI++);
if (PN && (PN->getParent() == Header ? Count > 1 : !HasExit))
PN->removeIncomingValue(LatchBlock);
}
}
// At this point, unrolling is complete and the code is well formed.
// Now, do some simplifications.
// 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 (BasicBlock::iterator I = Header->begin(); isa<PHINode>(I); ) {
PHINode *PN = cast<PHINode>(I++);
PN->replaceAllUsesWith(PN->getIncomingValueForBlock(Preheader));
Header->getInstList().erase(PN);
}
}
// We now do a quick sweep over the inserted code, doing constant
// propagation and dead code elimination as we go.
for (Loop::block_iterator BI = L->block_begin(), BBE = L->block_end();
BI != BBE; ++BI) {
BasicBlock *BB = *BI;
for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ) {
Instruction *Inst = I++;
if (isInstructionTriviallyDead(Inst))
BB->getInstList().erase(Inst);
else if (Constant *C = ConstantFoldInstruction(Inst)) {
Inst->replaceAllUsesWith(C);
BB->getInstList().erase(Inst);
}
}
}
NumCompletelyUnrolled += CompletelyUnroll;
++NumUnrolled;
// Remove the loop from the LoopPassManager if it's completely removed.
if (CompletelyUnroll && LPM != NULL)
LPM->deleteLoopFromQueue(L);
// If we didn't completely unroll the loop, it should still be in LCSSA form.
if (!CompletelyUnroll)
assert(L->isLCSSAForm());
return true;
}