/// updateLoopIterationSpace -- Update loop's iteration space if loop
/// body is executed for certain IV range only. For example,
///
/// for (i = 0; i < N; ++i) {
/// if ( i > A && i < B) {
/// ...
/// }
/// }
/// is transformed to iterators from A to B, if A > 0 and B < N.
///
bool LoopIndexSplit::updateLoopIterationSpace() {
SplitCondition = NULL;
if (ExitCondition->getPredicate() == ICmpInst::ICMP_NE
|| ExitCondition->getPredicate() == ICmpInst::ICMP_EQ)
return false;
BasicBlock *Latch = L->getLoopLatch();
BasicBlock *Header = L->getHeader();
BranchInst *BR = dyn_cast<BranchInst>(Header->getTerminator());
if (!BR) return false;
if (!isa<BranchInst>(Latch->getTerminator())) return false;
if (BR->isUnconditional()) return false;
BinaryOperator *AND = dyn_cast<BinaryOperator>(BR->getCondition());
if (!AND) return false;
if (AND->getOpcode() != Instruction::And) return false;
ICmpInst *Op0 = dyn_cast<ICmpInst>(AND->getOperand(0));
ICmpInst *Op1 = dyn_cast<ICmpInst>(AND->getOperand(1));
if (!Op0 || !Op1)
return false;
IVBasedValues.insert(AND);
IVBasedValues.insert(Op0);
IVBasedValues.insert(Op1);
if (!cleanBlock(Header)) return false;
BasicBlock *ExitingBlock = ExitCondition->getParent();
if (!cleanBlock(ExitingBlock)) return false;
// If the merge point for BR is not loop latch then skip this loop.
if (BR->getSuccessor(0) != Latch) {
DominanceFrontier::iterator DF0 = DF->find(BR->getSuccessor(0));
assert (DF0 != DF->end() && "Unable to find dominance frontier");
if (!DF0->second.count(Latch))
return false;
}
if (BR->getSuccessor(1) != Latch) {
DominanceFrontier::iterator DF1 = DF->find(BR->getSuccessor(1));
assert (DF1 != DF->end() && "Unable to find dominance frontier");
if (!DF1->second.count(Latch))
return false;
}
// Verify that loop exiting block has only two predecessor, where one pred
// is split condition block. The other predecessor will become exiting block's
// dominator after CFG is updated. TODO : Handle CFG's where exiting block has
// more then two predecessors. This requires extra work in updating dominator
// information.
BasicBlock *ExitingBBPred = NULL;
for (pred_iterator PI = pred_begin(ExitingBlock), PE = pred_end(ExitingBlock);
PI != PE; ++PI) {
BasicBlock *BB = *PI;
if (Header == BB)
continue;
if (ExitingBBPred)
return false;
else
ExitingBBPred = BB;
}
if (!restrictLoopBound(*Op0))
return false;
if (!restrictLoopBound(*Op1))
return false;
// Update CFG.
if (BR->getSuccessor(0) == ExitingBlock)
BR->setUnconditionalDest(BR->getSuccessor(1));
else
BR->setUnconditionalDest(BR->getSuccessor(0));
AND->eraseFromParent();
if (Op0->use_empty())
Op0->eraseFromParent();
if (Op1->use_empty())
Op1->eraseFromParent();
// Update domiantor info. Now, ExitingBlock has only one predecessor,
// ExitingBBPred, and it is ExitingBlock's immediate domiantor.
DT->changeImmediateDominator(ExitingBlock, ExitingBBPred);
BasicBlock *ExitBlock = ExitingBlock->getTerminator()->getSuccessor(1);
if (L->contains(ExitBlock))
ExitBlock = ExitingBlock->getTerminator()->getSuccessor(0);
// If ExitingBlock is a member of the loop basic blocks' DF list then
// replace ExitingBlock with header and exit block in the DF list
DominanceFrontier::iterator ExitingBlockDF = DF->find(ExitingBlock);
for (Loop::block_iterator I = L->block_begin(), E = L->block_end();
I != E; ++I) {
BasicBlock *BB = *I;
if (BB == Header || BB == ExitingBlock)
continue;
DominanceFrontier::iterator BBDF = DF->find(BB);
DominanceFrontier::DomSetType::iterator DomSetI = BBDF->second.begin();
DominanceFrontier::DomSetType::iterator DomSetE = BBDF->second.end();
while (DomSetI != DomSetE) {
DominanceFrontier::DomSetType::iterator CurrentItr = DomSetI;
++DomSetI;
BasicBlock *DFBB = *CurrentItr;
if (DFBB == ExitingBlock) {
BBDF->second.erase(DFBB);
for (DominanceFrontier::DomSetType::iterator
EBI = ExitingBlockDF->second.begin(),
EBE = ExitingBlockDF->second.end(); EBI != EBE; ++EBI)
BBDF->second.insert(*EBI);
}
}
}
NumRestrictBounds++;
return true;
}
/// 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());
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) {
DEBUG(errs() << "COMPLETELY UNROLLING loop %" << Header->getName()
<< " with trip count " << TripCount << "!\n");
} else {
DEBUG(errs() << "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.
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);
OrigPHINode.push_back(PN);
if (Instruction *I =
dyn_cast<Instruction>(PN->getIncomingValueForBlock(LatchBlock)))
if (L->contains(I->getParent()))
LastValueMap[I] = I;
}
std::vector<BasicBlock*> Headers;
std::vector<BasicBlock*> Latches;
Headers.push_back(Header);
Latches.push_back(LatchBlock);
for (unsigned It = 1; It != Count; ++It) {
char SuffixBuffer[100];
sprintf(SuffixBuffer, ".%d", It);
std::vector<BasicBlock*> NewBlocks;
for (std::vector<BasicBlock*>::iterator BB = LoopBlocks.begin(),
E = LoopBlocks.end(); BB != E; ++BB) {
ValueMapTy ValueMap;
BasicBlock *New = CloneBasicBlock(*BB, ValueMap, SuffixBuffer);
Header->getParent()->getBasicBlockList().push_back(New);
// Loop over all of the PHI nodes in the block, changing them to use the
// incoming values from the previous block.
if (*BB == Header)
for (unsigned i = 0, e = OrigPHINode.size(); i != e; ++i) {
PHINode *NewPHI = cast<PHINode>(ValueMap[OrigPHINode[i]]);
Value *InVal = NewPHI->getIncomingValueForBlock(LatchBlock);
if (Instruction *InValI = dyn_cast<Instruction>(InVal))
if (It > 1 && L->contains(InValI->getParent()))
InVal = LastValueMap[InValI];
ValueMap[OrigPHINode[i]] = InVal;
New->getInstList().erase(NewPHI);
}
// 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;
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->getParent())) {
PHINode *phi = cast<PHINode>(UseInst);
Value *Incoming = phi->getIncomingValueForBlock(*BB);
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) {
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;
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 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, LI)) {
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(),
BBE = NewLoopBlocks.end(); BB != BBE; ++BB)
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,
Header->getContext())) {
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;
}
/// processOneIterationLoop -- Eliminate loop if loop body is executed
/// only once. For example,
/// for (i = 0; i < N; ++i) {
/// if ( i == X) {
/// ...
/// }
/// }
///
bool LoopIndexSplit::processOneIterationLoop() {
SplitCondition = NULL;
BasicBlock *Latch = L->getLoopLatch();
BasicBlock *Header = L->getHeader();
BranchInst *BR = dyn_cast<BranchInst>(Header->getTerminator());
if (!BR) return false;
if (!isa<BranchInst>(Latch->getTerminator())) return false;
if (BR->isUnconditional()) return false;
SplitCondition = dyn_cast<ICmpInst>(BR->getCondition());
if (!SplitCondition) return false;
if (SplitCondition == ExitCondition) return false;
if (SplitCondition->getPredicate() != ICmpInst::ICMP_EQ) return false;
if (BR->getOperand(1) != Latch) return false;
if (!IVBasedValues.count(SplitCondition->getOperand(0))
&& !IVBasedValues.count(SplitCondition->getOperand(1)))
return false;
// If IV is used outside the loop then this loop traversal is required.
// FIXME: Calculate and use last IV value.
if (isUsedOutsideLoop(IVIncrement, L))
return false;
// If BR operands are not IV or not loop invariants then skip this loop.
Value *OPV = SplitCondition->getOperand(0);
Value *SplitValue = SplitCondition->getOperand(1);
if (!L->isLoopInvariant(SplitValue))
std::swap(OPV, SplitValue);
if (!L->isLoopInvariant(SplitValue))
return false;
Instruction *OPI = dyn_cast<Instruction>(OPV);
if (!OPI)
return false;
if (OPI->getParent() != Header || isUsedOutsideLoop(OPI, L))
return false;
Value *StartValue = IVStartValue;
Value *ExitValue = IVExitValue;;
if (OPV != IndVar) {
// If BR operand is IV based then use this operand to calculate
// effective conditions for loop body.
BinaryOperator *BOPV = dyn_cast<BinaryOperator>(OPV);
if (!BOPV)
return false;
if (BOPV->getOpcode() != Instruction::Add)
return false;
StartValue = BinaryOperator::CreateAdd(OPV, StartValue, "" , BR);
ExitValue = BinaryOperator::CreateAdd(OPV, ExitValue, "" , BR);
}
if (!cleanBlock(Header))
return false;
if (!cleanBlock(Latch))
return false;
// If the merge point for BR is not loop latch then skip this loop.
if (BR->getSuccessor(0) != Latch) {
DominanceFrontier::iterator DF0 = DF->find(BR->getSuccessor(0));
assert (DF0 != DF->end() && "Unable to find dominance frontier");
if (!DF0->second.count(Latch))
return false;
}
if (BR->getSuccessor(1) != Latch) {
DominanceFrontier::iterator DF1 = DF->find(BR->getSuccessor(1));
assert (DF1 != DF->end() && "Unable to find dominance frontier");
if (!DF1->second.count(Latch))
return false;
}
// Now, Current loop L contains compare instruction
// that compares induction variable, IndVar, against loop invariant. And
// entire (i.e. meaningful) loop body is dominated by this compare
// instruction. In such case eliminate
// loop structure surrounding this loop body. For example,
// for (int i = start; i < end; ++i) {
// if ( i == somevalue) {
// loop_body
// }
// }
// can be transformed into
// if (somevalue >= start && somevalue < end) {
// i = somevalue;
// loop_body
// }
// Replace index variable with split value in loop body. Loop body is executed
// only when index variable is equal to split value.
IndVar->replaceAllUsesWith(SplitValue);
// Replace split condition in header.
// Transform
// SplitCondition : icmp eq i32 IndVar, SplitValue
// into
// c1 = icmp uge i32 SplitValue, StartValue
// c2 = icmp ult i32 SplitValue, ExitValue
// and i32 c1, c2
Instruction *C1 = new ICmpInst(BR, ExitCondition->isSigned() ?
ICmpInst::ICMP_SGE : ICmpInst::ICMP_UGE,
SplitValue, StartValue, "lisplit");
CmpInst::Predicate C2P = ExitCondition->getPredicate();
BranchInst *LatchBR = cast<BranchInst>(Latch->getTerminator());
if (LatchBR->getOperand(1) != Header)
C2P = CmpInst::getInversePredicate(C2P);
Instruction *C2 = new ICmpInst(BR, C2P, SplitValue, ExitValue, "lisplit");
Instruction *NSplitCond = BinaryOperator::CreateAnd(C1, C2, "lisplit", BR);
SplitCondition->replaceAllUsesWith(NSplitCond);
SplitCondition->eraseFromParent();
// Remove Latch to Header edge.
BasicBlock *LatchSucc = NULL;
Header->removePredecessor(Latch);
for (succ_iterator SI = succ_begin(Latch), E = succ_end(Latch);
SI != E; ++SI) {
if (Header != *SI)
LatchSucc = *SI;
}
// Clean up latch block.
Value *LatchBRCond = LatchBR->getCondition();
LatchBR->setUnconditionalDest(LatchSucc);
RecursivelyDeleteTriviallyDeadInstructions(LatchBRCond);
LPM->deleteLoopFromQueue(L);
// Update Dominator Info.
// Only CFG change done is to remove Latch to Header edge. This
// does not change dominator tree because Latch did not dominate
// Header.
if (DF) {
DominanceFrontier::iterator HeaderDF = DF->find(Header);
if (HeaderDF != DF->end())
DF->removeFromFrontier(HeaderDF, Header);
DominanceFrontier::iterator LatchDF = DF->find(Latch);
if (LatchDF != DF->end())
DF->removeFromFrontier(LatchDF, Header);
}
++NumIndexSplitRemoved;
return true;
}
//insert a basic block with appropriate code
//along a given edge
void insertBB(Edge ed,
getEdgeCode *edgeCode,
Instruction *rInst,
Value *countInst,
int numPaths, int Methno, Value *threshold){
BasicBlock* BB1=ed.getFirst()->getElement();
BasicBlock* BB2=ed.getSecond()->getElement();
#ifdef DEBUG_PATH_PROFILES
//debugging info
cerr<<"Edges with codes ######################\n";
cerr<<BB1->getName()<<"->"<<BB2->getName()<<"\n";
cerr<<"########################\n";
#endif
//We need to insert a BB between BB1 and BB2
TerminatorInst *TI=BB1->getTerminator();
BasicBlock *newBB=new BasicBlock("counter", BB1->getParent());
//get code for the new BB
vector<Value *> retVec;
edgeCode->getCode(rInst, countInst, BB1->getParent(), newBB, retVec);
BranchInst *BI = cast<BranchInst>(TI);
//Is terminator a branch instruction?
//then we need to change branch destinations to include new BB
if(BI->isUnconditional()){
BI->setUnconditionalDest(newBB);
}
else{
if(BI->getSuccessor(0)==BB2)
BI->setSuccessor(0, newBB);
if(BI->getSuccessor(1)==BB2)
BI->setSuccessor(1, newBB);
}
BasicBlock *triggerBB = NULL;
if(retVec.size()>0){
triggerBB = new BasicBlock("trigger", BB1->getParent());
getTriggerCode(BB1->getParent()->getParent(), triggerBB, Methno,
retVec[1], countInst, rInst);//retVec[0]);
//Instruction *castInst = new CastInst(retVec[0], Type::IntTy, "");
Instruction *etr = new LoadInst(threshold, "threshold");
//std::cerr<<"type1: "<<etr->getType()<<" type2: "<<retVec[0]->getType()<<"\n";
Instruction *cmpInst = new SetCondInst(Instruction::SetLE, etr,
retVec[0], "");
Instruction *newBI2 = new BranchInst(triggerBB, BB2, cmpInst);
//newBB->getInstList().push_back(castInst);
newBB->getInstList().push_back(etr);
newBB->getInstList().push_back(cmpInst);
newBB->getInstList().push_back(newBI2);
//triggerBB->getInstList().push_back(triggerInst);
new BranchInst(BB2, 0, 0, triggerBB);
}
else{
new BranchInst(BB2, 0, 0, newBB);
}
//now iterate over BB2, and set its Phi nodes right
for(BasicBlock::iterator BB2Inst = BB2->begin(), BBend = BB2->end();
BB2Inst != BBend; ++BB2Inst){
if(PHINode *phiInst=dyn_cast<PHINode>(BB2Inst)){
int bbIndex=phiInst->getBasicBlockIndex(BB1);
assert(bbIndex>=0);
phiInst->setIncomingBlock(bbIndex, newBB);
///check if trigger!=null, then add value corresponding to it too!
if(retVec.size()>0){
assert(triggerBB && "BasicBlock with trigger should not be null!");
Value *vl = phiInst->getIncomingValue((unsigned int)bbIndex);
phiInst->addIncoming(vl, triggerBB);
}
}
}
}
