bool partition::runOnLoop(Loop* L, LPPassManager &LPM) {
errs() << "*************************** Loop encountered: ************************" << '\n' << L->getHeader()->getName() << '\n';
if (function->getName() != "main")
return false;
IntegerType* int32Ty = Type::getInt32Ty(*context);
IntegerType* int64Ty = Type::getInt64Ty(*context);
PointerType* voidPtrTy = Type::getInt8PtrTy(*context);
FunctionType* funcTy = FunctionType::get(int32Ty, false);
Constant* func1_c;
Function* func1;
func1_c = module->getOrInsertFunction("func1", funcTy);
func1 = cast<Function>(func1_c);
Function* pro = module->getFunction("produce");
Function* con = module->getFunction("consume");
BasicBlock* func1EntryBlock = BasicBlock::Create(*context, "entry.func1", func1);
AllocaInst* i_var = new AllocaInst(int32Ty, NULL, 4, "i", func1EntryBlock);
Value* liveIn;
BasicBlock *forCond, *forBody, *forInc;
ValueToValueMapTy VMap;
std::map<BasicBlock *, BasicBlock *> BlockMap;
for (Loop::block_iterator BB = L->block_begin(), BBe = L->block_end(); BB != BBe; ++BB) {
BasicBlock* func1Block = CloneBasicBlock(*BB, VMap, ".func1", func1);
BlockMap[*BB] = func1Block;
if ((*BB)->getName() == "for.cond")
forCond = func1Block;
if ((*BB)->getName() == "for.body")
forBody = func1Block;
if ((*BB)->getName() == "for.inc")
forInc = func1Block;
for (BasicBlock::iterator it = func1Block->begin(), ite = func1Block->end(); it != ite; ++it) {
for (User::op_iterator oit = it->op_begin(), oite = it->op_end(); oit != oite; ++oit) {
if (VMap[*oit] != NULL) {
*oit = VMap[*oit];
} else {
Constant* cons = dyn_cast<Constant>(*oit);
BranchInst* br = dyn_cast<BranchInst>(it);
if (cons == NULL && br == NULL) {
liveIn = *oit;
*oit = i_var;
}
}
}
}
if ((*BB)->getName() == "for.body") {
Instruction* term = (*BB)->getTerminator();
term->removeFromParent();
for (int i = 0; i < 7; i++) {
(*BB)->back().eraseFromParent();
}
term->insertAfter(&(*BB)->back());
(*BB)->front().eraseFromParent();
LoadInst* load = new LoadInst(liveIn, "", false, 4, term);
std::vector<Value *> produce_args;
CastInst* cast = CastInst::CreateIntegerCast(load, int64Ty, true);
cast->insertAfter(load);
produce_args.push_back(cast);
ConstantInt* val = ConstantInt::get(int32Ty, (uint32_t) 3);
produce_args.push_back(val);
CallInst::Create(pro, ArrayRef<Value*>(produce_args), "", term);
produce_args.pop_back();
val = ConstantInt::get(int32Ty, (uint32_t) 2);
produce_args.push_back(val);
CallInst::Create(pro, ArrayRef<Value*>(produce_args), "", term);
}
}
// set branch instructions to restructure the CFG in created function
BasicBlock* func1EndBlock = BasicBlock::Create(*context, "if.end.func1", func1);
BasicBlock* garbageBB = BasicBlock::Create(*context, "garbage", func1);
ConstantInt* retVal_g = ConstantInt::get(int32Ty, (uint32_t) 0);
ReturnInst* ret_g = ReturnInst::Create(*context, retVal_g, garbageBB);
for (Function::iterator fit = func1->begin(), fite = func1->end(); fit != fite; ++fit) {
if (fit->getTerminator() == NULL || fit->getName() == "garbage")
continue;
BranchInst* br = dyn_cast<BranchInst>(fit->getTerminator());
int numSuccessors = br->getNumSuccessors();
for (int i = 0; i < numSuccessors; i++) {
BasicBlock* successor = br->getSuccessor(i);
if (BlockMap[successor] != NULL) {
br->setSuccessor(i, BlockMap[successor]);
}
else {
br->setSuccessor(i, func1EndBlock);
}
}
/*
if (fit->getName() == "for.body.func1") {
for (int i = 0; i < 4; i++) {
BasicBlock::iterator it = fit->begin();
it->moveBefore(ret_g);
}
}
*/
}
garbageBB->eraseFromParent();
BranchInst* br = dyn_cast<BranchInst>(forBody->getTerminator());
br->setSuccessor(0, forCond);
forInc->eraseFromParent();
// Create return instruction for func1EndBlock and set a branch from loop header to func1EndBlock
ConstantInt* retVal = ConstantInt::get(int32Ty, (uint32_t) 0);
ReturnInst* ret1 = ReturnInst::Create(*context, retVal, func1EndBlock);
BasicBlock* loopHeader = BlockMap.at(L->getHeader());
BranchInst* brInst = BranchInst::Create(loopHeader, func1EntryBlock);
// add produce function call
std::vector<Value *> produce_args;
ConstantInt* val = ConstantInt::get(int64Ty, (uint64_t) 0);
produce_args.push_back(val);
val = ConstantInt::get(int32Ty, (uint32_t) 5);
produce_args.push_back(val);
CallInst::Create(pro, ArrayRef<Value*>(produce_args), "", func1EndBlock->getTerminator());
// add consume function call
int q_id = 2;
for (Value::use_iterator uit = i_var->use_begin(), uite = i_var->use_end(); uit != uite; ++uit) {
std::vector<Value *> consume_args;
ConstantInt* val = ConstantInt::get(int32Ty, (uint32_t) q_id);
consume_args.push_back(val);
CallInst* call = CallInst::Create(con, ArrayRef<Value*>(consume_args));
Instruction* inst = dyn_cast<Instruction>(*uit);
call->insertAfter(inst);
CastInst* cast = CastInst::CreateIntegerCast(call, int32Ty, true);
cast->insertAfter(call);
(*uit)->replaceAllUsesWith(cast);
inst->eraseFromParent();
q_id++;
}
i_var->eraseFromParent();
// add produce and consume function calls to main thread
// transmit the function pointer to created function by a produce call
BasicBlock* loopPreheader = L->getLoopPreheader();
produce_args.clear();
CastInst* cast = CastInst::CreatePointerCast(func1, int64Ty);
cast->insertBefore(loopPreheader->getTerminator());
produce_args.push_back(cast);
val = ConstantInt::get(int32Ty, (uint32_t) 0);
produce_args.push_back(val);
CallInst::Create(pro, ArrayRef<Value*>(produce_args), "", loopPreheader->getTerminator());
// transmit induction variable to created function by a produce call
Instruction* load = &L->getHeader()->front();
produce_args.clear();
cast = CastInst::CreateIntegerCast(load, int64Ty, true);
cast->insertAfter(load);
produce_args.push_back(cast);
val = ConstantInt::get(int32Ty, (uint32_t) 4);
produce_args.push_back(val);
CallInst::Create(pro, ArrayRef<Value*>(produce_args))->insertAfter(cast);
return true;
}
bool llvm::UnrollRuntimeLoopRemainder(Loop *L, unsigned Count,
bool AllowExpensiveTripCount,
bool UseEpilogRemainder,
LoopInfo *LI, ScalarEvolution *SE,
DominatorTree *DT, bool PreserveLCSSA) {
// for now, only unroll loops that contain a single exit
if (!L->getExitingBlock())
return false;
// Make sure the loop is in canonical form, and there is a single
// exit block only.
if (!L->isLoopSimplifyForm())
return false;
BasicBlock *Exit = L->getUniqueExitBlock(); // successor out of loop
if (!Exit)
return false;
// Use Scalar Evolution to compute the trip count. This allows more loops to
// be unrolled than relying on induction var simplification.
if (!SE)
return false;
// 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 *BECountSC = SE->getBackedgeTakenCount(L);
if (isa<SCEVCouldNotCompute>(BECountSC) ||
!BECountSC->getType()->isIntegerTy())
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(BECountSC, SE->getConstant(BECountSC->getType(), 1));
if (isa<SCEVCouldNotCompute>(TripCountSC))
return false;
BasicBlock *Header = L->getHeader();
BasicBlock *PreHeader = L->getLoopPreheader();
BranchInst *PreHeaderBR = cast<BranchInst>(PreHeader->getTerminator());
const DataLayout &DL = Header->getModule()->getDataLayout();
SCEVExpander Expander(*SE, DL, "loop-unroll");
if (!AllowExpensiveTripCount &&
Expander.isHighCostExpansion(TripCountSC, L, PreHeaderBR))
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;
BasicBlock *Latch = L->getLoopLatch();
// Loop structure is the following:
//
// PreHeader
// Header
// ...
// Latch
// Exit
BasicBlock *NewPreHeader;
BasicBlock *NewExit = nullptr;
BasicBlock *PrologExit = nullptr;
BasicBlock *EpilogPreHeader = nullptr;
BasicBlock *PrologPreHeader = nullptr;
if (UseEpilogRemainder) {
// If epilog remainder
// Split PreHeader to insert a branch around loop for unrolling.
NewPreHeader = SplitBlock(PreHeader, PreHeader->getTerminator(), DT, LI);
NewPreHeader->setName(PreHeader->getName() + ".new");
// Split Exit to create phi nodes from branch above.
SmallVector<BasicBlock*, 4> Preds(predecessors(Exit));
NewExit = SplitBlockPredecessors(Exit, Preds, ".unr-lcssa",
DT, LI, PreserveLCSSA);
// Split NewExit to insert epilog remainder loop.
EpilogPreHeader = SplitBlock(NewExit, NewExit->getTerminator(), DT, LI);
EpilogPreHeader->setName(Header->getName() + ".epil.preheader");
} else {
// If prolog remainder
// Split the original preheader twice to insert prolog remainder loop
PrologPreHeader = SplitEdge(PreHeader, Header, DT, LI);
PrologPreHeader->setName(Header->getName() + ".prol.preheader");
PrologExit = SplitBlock(PrologPreHeader, PrologPreHeader->getTerminator(),
DT, LI);
PrologExit->setName(Header->getName() + ".prol.loopexit");
// Split PrologExit to get NewPreHeader.
NewPreHeader = SplitBlock(PrologExit, PrologExit->getTerminator(), DT, LI);
NewPreHeader->setName(PreHeader->getName() + ".new");
}
// Loop structure should be the following:
// Epilog Prolog
//
// PreHeader PreHeader
// *NewPreHeader *PrologPreHeader
// Header *PrologExit
// ... *NewPreHeader
// Latch Header
// *NewExit ...
// *EpilogPreHeader Latch
// Exit Exit
// Calculate conditions for branch around loop for unrolling
// in epilog case and around prolog remainder loop in prolog case.
// Compute the number of extra iterations required, which is:
// extra iterations = run-time trip count % loop unroll factor
PreHeaderBR = cast<BranchInst>(PreHeader->getTerminator());
Value *TripCount = Expander.expandCodeFor(TripCountSC, TripCountSC->getType(),
PreHeaderBR);
Value *BECount = Expander.expandCodeFor(BECountSC, BECountSC->getType(),
PreHeaderBR);
IRBuilder<> B(PreHeaderBR);
Value *ModVal;
// Calculate ModVal = (BECount + 1) % Count.
// Note that TripCount is BECount + 1.
if (isPowerOf2_32(Count)) {
// When Count is power of 2 we don't BECount for epilog case, however we'll
// need it for a branch around unrolling loop for prolog case.
ModVal = B.CreateAnd(TripCount, Count - 1, "xtraiter");
// 1. There are no iterations to be run in the prolog/epilog 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).
} else {
// As (BECount + 1) can potentially unsigned overflow we count
// (BECount % Count) + 1 which is overflow safe as BECount % Count < Count.
Value *ModValTmp = B.CreateURem(BECount,
ConstantInt::get(BECount->getType(),
Count));
Value *ModValAdd = B.CreateAdd(ModValTmp,
ConstantInt::get(ModValTmp->getType(), 1));
// At that point (BECount % Count) + 1 could be equal to Count.
// To handle this case we need to take mod by Count one more time.
ModVal = B.CreateURem(ModValAdd,
ConstantInt::get(BECount->getType(), Count),
"xtraiter");
}
Value *BranchVal =
UseEpilogRemainder ? B.CreateICmpULT(BECount,
ConstantInt::get(BECount->getType(),
Count - 1)) :
B.CreateIsNotNull(ModVal, "lcmp.mod");
BasicBlock *RemainderLoop = UseEpilogRemainder ? NewExit : PrologPreHeader;
BasicBlock *UnrollingLoop = UseEpilogRemainder ? NewPreHeader : PrologExit;
// Branch to either remainder (extra iterations) loop or unrolling loop.
B.CreateCondBr(BranchVal, RemainderLoop, UnrollingLoop);
PreHeaderBR->eraseFromParent();
Function *F = Header->getParent();
// Get an ordered list of blocks in the loop to help with the ordering of the
// cloned blocks in the prolog/epilog code
LoopBlocksDFS LoopBlocks(L);
LoopBlocks.perform(LI);
//
// For each extra loop iteration, create a copy of the loop's basic blocks
// and generate a condition that branches to the copy depending on the
// number of 'left over' iterations.
//
std::vector<BasicBlock *> NewBlocks;
ValueToValueMapTy VMap;
// For unroll factor 2 remainder loop will have 1 iterations.
// Do not create 1 iteration loop.
bool CreateRemainderLoop = (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
// iterations. This function adds the appropriate CFG connections.
BasicBlock *InsertBot = UseEpilogRemainder ? Exit : PrologExit;
BasicBlock *InsertTop = UseEpilogRemainder ? EpilogPreHeader : PrologPreHeader;
CloneLoopBlocks(L, ModVal, CreateRemainderLoop, UseEpilogRemainder, InsertTop,
InsertBot, NewPreHeader, NewBlocks, LoopBlocks, VMap, LI);
// Insert the cloned blocks into the function.
F->getBasicBlockList().splice(InsertBot->getIterator(),
F->getBasicBlockList(),
NewBlocks[0]->getIterator(),
F->end());
// Loop structure should be the following:
// Epilog Prolog
//
// PreHeader PreHeader
// NewPreHeader PrologPreHeader
// Header PrologHeader
// ... ...
// Latch PrologLatch
// NewExit PrologExit
// EpilogPreHeader NewPreHeader
// EpilogHeader Header
// ... ...
// EpilogLatch Latch
// Exit Exit
// Rewrite the cloned instruction operands to use the values created when the
// clone is created.
for (BasicBlock *BB : NewBlocks) {
for (Instruction &I : *BB) {
RemapInstruction(&I, VMap,
RF_NoModuleLevelChanges | RF_IgnoreMissingLocals);
}
}
if (UseEpilogRemainder) {
// Connect the epilog code to the original loop and update the
// PHI functions.
ConnectEpilog(L, ModVal, NewExit, Exit, PreHeader,
EpilogPreHeader, NewPreHeader, VMap, DT, LI,
PreserveLCSSA);
// Update counter in loop for unrolling.
// I should be multiply of Count.
IRBuilder<> B2(NewPreHeader->getTerminator());
Value *TestVal = B2.CreateSub(TripCount, ModVal, "unroll_iter");
BranchInst *LatchBR = cast<BranchInst>(Latch->getTerminator());
B2.SetInsertPoint(LatchBR);
PHINode *NewIdx = PHINode::Create(TestVal->getType(), 2, "niter",
Header->getFirstNonPHI());
Value *IdxSub =
B2.CreateSub(NewIdx, ConstantInt::get(NewIdx->getType(), 1),
NewIdx->getName() + ".nsub");
Value *IdxCmp;
if (LatchBR->getSuccessor(0) == Header)
IdxCmp = B2.CreateIsNotNull(IdxSub, NewIdx->getName() + ".ncmp");
else
IdxCmp = B2.CreateIsNull(IdxSub, NewIdx->getName() + ".ncmp");
NewIdx->addIncoming(TestVal, NewPreHeader);
NewIdx->addIncoming(IdxSub, Latch);
LatchBR->setCondition(IdxCmp);
} else {
// Connect the prolog code to the original loop and update the
// PHI functions.
ConnectProlog(L, BECount, Count, PrologExit, PreHeader, NewPreHeader,
VMap, DT, LI, PreserveLCSSA);
}
// If this loop is nested, then the loop unroller changes the code in the
// parent loop, so the Scalar Evolution pass needs to be run again.
if (Loop *ParentLoop = L->getParentLoop())
SE->forgetLoop(ParentLoop);
NumRuntimeUnrolled++;
return true;
}
/// HandleFloatingPointIV - If the loop has floating induction variable
/// then insert corresponding integer induction variable if possible.
/// For example,
/// for(double i = 0; i < 10000; ++i)
/// bar(i)
/// is converted into
/// for(int i = 0; i < 10000; ++i)
/// bar((double)i);
///
void IndVarSimplify::HandleFloatingPointIV(Loop *L, PHINode *PN) {
unsigned IncomingEdge = L->contains(PN->getIncomingBlock(0));
unsigned BackEdge = IncomingEdge^1;
// Check incoming value.
ConstantFP *InitValueVal =
dyn_cast<ConstantFP>(PN->getIncomingValue(IncomingEdge));
int64_t InitValue;
if (!InitValueVal || !ConvertToSInt(InitValueVal->getValueAPF(), InitValue))
return;
// Check IV increment. Reject this PN if increment operation is not
// an add or increment value can not be represented by an integer.
BinaryOperator *Incr =
dyn_cast<BinaryOperator>(PN->getIncomingValue(BackEdge));
if (Incr == 0 || Incr->getOpcode() != Instruction::FAdd) return;
// If this is not an add of the PHI with a constantfp, or if the constant fp
// is not an integer, bail out.
ConstantFP *IncValueVal = dyn_cast<ConstantFP>(Incr->getOperand(1));
int64_t IncValue;
if (IncValueVal == 0 || Incr->getOperand(0) != PN ||
!ConvertToSInt(IncValueVal->getValueAPF(), IncValue))
return;
// Check Incr uses. One user is PN and the other user is an exit condition
// used by the conditional terminator.
Value::use_iterator IncrUse = Incr->use_begin();
Instruction *U1 = cast<Instruction>(*IncrUse++);
if (IncrUse == Incr->use_end()) return;
Instruction *U2 = cast<Instruction>(*IncrUse++);
if (IncrUse != Incr->use_end()) return;
// Find exit condition, which is an fcmp. If it doesn't exist, or if it isn't
// only used by a branch, we can't transform it.
FCmpInst *Compare = dyn_cast<FCmpInst>(U1);
if (!Compare)
Compare = dyn_cast<FCmpInst>(U2);
if (Compare == 0 || !Compare->hasOneUse() ||
!isa<BranchInst>(Compare->use_back()))
return;
BranchInst *TheBr = cast<BranchInst>(Compare->use_back());
// We need to verify that the branch actually controls the iteration count
// of the loop. If not, the new IV can overflow and no one will notice.
// The branch block must be in the loop and one of the successors must be out
// of the loop.
assert(TheBr->isConditional() && "Can't use fcmp if not conditional");
if (!L->contains(TheBr->getParent()) ||
(L->contains(TheBr->getSuccessor(0)) &&
L->contains(TheBr->getSuccessor(1))))
return;
// If it isn't a comparison with an integer-as-fp (the exit value), we can't
// transform it.
ConstantFP *ExitValueVal = dyn_cast<ConstantFP>(Compare->getOperand(1));
int64_t ExitValue;
if (ExitValueVal == 0 ||
!ConvertToSInt(ExitValueVal->getValueAPF(), ExitValue))
return;
// Find new predicate for integer comparison.
CmpInst::Predicate NewPred = CmpInst::BAD_ICMP_PREDICATE;
switch (Compare->getPredicate()) {
default: return; // Unknown comparison.
case CmpInst::FCMP_OEQ:
case CmpInst::FCMP_UEQ: NewPred = CmpInst::ICMP_EQ; break;
case CmpInst::FCMP_ONE:
case CmpInst::FCMP_UNE: NewPred = CmpInst::ICMP_NE; break;
case CmpInst::FCMP_OGT:
case CmpInst::FCMP_UGT: NewPred = CmpInst::ICMP_SGT; break;
case CmpInst::FCMP_OGE:
case CmpInst::FCMP_UGE: NewPred = CmpInst::ICMP_SGE; break;
case CmpInst::FCMP_OLT:
case CmpInst::FCMP_ULT: NewPred = CmpInst::ICMP_SLT; break;
case CmpInst::FCMP_OLE:
case CmpInst::FCMP_ULE: NewPred = CmpInst::ICMP_SLE; break;
}
// We convert the floating point induction variable to a signed i32 value if
// we can. This is only safe if the comparison will not overflow in a way
// that won't be trapped by the integer equivalent operations. Check for this
// now.
// TODO: We could use i64 if it is native and the range requires it.
// The start/stride/exit values must all fit in signed i32.
if (!isInt<32>(InitValue) || !isInt<32>(IncValue) || !isInt<32>(ExitValue))
return;
// If not actually striding (add x, 0.0), avoid touching the code.
if (IncValue == 0)
return;
// Positive and negative strides have different safety conditions.
if (IncValue > 0) {
// If we have a positive stride, we require the init to be less than the
// exit value and an equality or less than comparison.
if (InitValue >= ExitValue ||
NewPred == CmpInst::ICMP_SGT || NewPred == CmpInst::ICMP_SGE)
return;
uint32_t Range = uint32_t(ExitValue-InitValue);
if (NewPred == CmpInst::ICMP_SLE) {
// Normalize SLE -> SLT, check for infinite loop.
if (++Range == 0) return; // Range overflows.
}
unsigned Leftover = Range % uint32_t(IncValue);
// If this is an equality comparison, we require that the strided value
// exactly land on the exit value, otherwise the IV condition will wrap
// around and do things the fp IV wouldn't.
if ((NewPred == CmpInst::ICMP_EQ || NewPred == CmpInst::ICMP_NE) &&
Leftover != 0)
return;
// If the stride would wrap around the i32 before exiting, we can't
// transform the IV.
if (Leftover != 0 && int32_t(ExitValue+IncValue) < ExitValue)
return;
} else {
// If we have a negative stride, we require the init to be greater than the
// exit value and an equality or greater than comparison.
if (InitValue >= ExitValue ||
NewPred == CmpInst::ICMP_SLT || NewPred == CmpInst::ICMP_SLE)
return;
uint32_t Range = uint32_t(InitValue-ExitValue);
if (NewPred == CmpInst::ICMP_SGE) {
// Normalize SGE -> SGT, check for infinite loop.
if (++Range == 0) return; // Range overflows.
}
unsigned Leftover = Range % uint32_t(-IncValue);
// If this is an equality comparison, we require that the strided value
// exactly land on the exit value, otherwise the IV condition will wrap
// around and do things the fp IV wouldn't.
if ((NewPred == CmpInst::ICMP_EQ || NewPred == CmpInst::ICMP_NE) &&
Leftover != 0)
return;
// If the stride would wrap around the i32 before exiting, we can't
// transform the IV.
if (Leftover != 0 && int32_t(ExitValue+IncValue) > ExitValue)
return;
}
const IntegerType *Int32Ty = Type::getInt32Ty(PN->getContext());
// Insert new integer induction variable.
PHINode *NewPHI = PHINode::Create(Int32Ty, PN->getName()+".int", PN);
NewPHI->addIncoming(ConstantInt::get(Int32Ty, InitValue),
PN->getIncomingBlock(IncomingEdge));
Value *NewAdd =
BinaryOperator::CreateAdd(NewPHI, ConstantInt::get(Int32Ty, IncValue),
Incr->getName()+".int", Incr);
NewPHI->addIncoming(NewAdd, PN->getIncomingBlock(BackEdge));
ICmpInst *NewCompare = new ICmpInst(TheBr, NewPred, NewAdd,
ConstantInt::get(Int32Ty, ExitValue),
Compare->getName());
// In the following deletions, PN may become dead and may be deleted.
// Use a WeakVH to observe whether this happens.
WeakVH WeakPH = PN;
// Delete the old floating point exit comparison. The branch starts using the
// new comparison.
NewCompare->takeName(Compare);
Compare->replaceAllUsesWith(NewCompare);
RecursivelyDeleteTriviallyDeadInstructions(Compare);
// Delete the old floating point increment.
Incr->replaceAllUsesWith(UndefValue::get(Incr->getType()));
RecursivelyDeleteTriviallyDeadInstructions(Incr);
// If the FP induction variable still has uses, this is because something else
// in the loop uses its value. In order to canonicalize the induction
// variable, we chose to eliminate the IV and rewrite it in terms of an
// int->fp cast.
//
// We give preference to sitofp over uitofp because it is faster on most
// platforms.
if (WeakPH) {
Value *Conv = new SIToFPInst(NewPHI, PN->getType(), "indvar.conv",
PN->getParent()->getFirstNonPHI());
PN->replaceAllUsesWith(Conv);
RecursivelyDeleteTriviallyDeadInstructions(PN);
}
// Add a new IVUsers entry for the newly-created integer PHI.
IU->AddUsersIfInteresting(NewPHI);
}
/// Unroll the given loop by Count. The loop must be in LCSSA form. Returns true
/// 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.
///
/// If AllowRuntime is true then UnrollLoop will consider unrolling loops that
/// have a runtime (i.e. not compile time constant) trip count. Unrolling these
/// loops require a unroll "prologue" that runs "RuntimeTripCount % Count"
/// iterations before branching into the unrolled loop. UnrollLoop will not
/// runtime-unroll the loop if computing RuntimeTripCount will be expensive and
/// AllowExpensiveTripCount is false.
///
/// 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.
///
/// This utility preserves LoopInfo. If DominatorTree or ScalarEvolution are
/// available from the Pass it must also preserve those analyses.
bool llvm::UnrollLoop(Loop *L, unsigned Count, unsigned TripCount,
bool AllowRuntime, bool AllowExpensiveTripCount,
unsigned TripMultiple, LoopInfo *LI, Pass *PP,
LPPassManager *LPM, AssumptionCache *AC) {
BasicBlock *Preheader = L->getLoopPreheader();
if (!Preheader) {
DEBUG(dbgs() << " Can't unroll; loop preheader-insertion failed.\n");
return false;
}
BasicBlock *LatchBlock = L->getLoopLatch();
if (!LatchBlock) {
DEBUG(dbgs() << " Can't unroll; loop exit-block-insertion failed.\n");
return false;
}
// Loops with indirectbr cannot be cloned.
if (!L->isSafeToClone()) {
DEBUG(dbgs() << " Can't unroll; Loop body cannot be cloned.\n");
return false;
}
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() <<
" Can't unroll; loop not terminated by a conditional branch.\n");
return false;
}
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 (TripMultiple != 1)
DEBUG(dbgs() << " 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;
// 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, AllowExpensiveTripCount, LI, LPM))
return false;
// Notify ScalarEvolution that the loop will be substantially changed,
// if not outright eliminated.
auto *SEWP =
PP ? PP->getAnalysisIfAvailable<ScalarEvolutionWrapperPass>() : nullptr;
ScalarEvolution *SE = SEWP ? &SEWP->getSE() : nullptr;
if (SE)
SE->forgetLoop(L);
// 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);
}
// Report the unrolling decision.
DebugLoc LoopLoc = L->getStartLoc();
Function *F = Header->getParent();
LLVMContext &Ctx = F->getContext();
if (CompletelyUnroll) {
DEBUG(dbgs() << "COMPLETELY UNROLLING loop %" << Header->getName()
<< " with trip count " << TripCount << "!\n");
emitOptimizationRemark(Ctx, DEBUG_TYPE, *F, LoopLoc,
Twine("completely unrolled loop with ") +
Twine(TripCount) + " iterations");
} else {
auto EmitDiag = [&](const Twine &T) {
emitOptimizationRemark(Ctx, DEBUG_TYPE, *F, LoopLoc,
"unrolled loop by a factor of " + Twine(Count) +
T);
};
DEBUG(dbgs() << "UNROLLING loop %" << Header->getName()
<< " by " << Count);
if (TripMultiple == 0 || BreakoutTrip != TripMultiple) {
DEBUG(dbgs() << " with a breakout at trip " << BreakoutTrip);
EmitDiag(" with a breakout at trip " + Twine(BreakoutTrip));
} else if (TripMultiple != 1) {
DEBUG(dbgs() << " with " << TripMultiple << " trips per branch");
EmitDiag(" with " + Twine(TripMultiple) + " trips per branch");
} else if (RuntimeTripCount) {
DEBUG(dbgs() << " with run-time trip count");
EmitDiag(" with run-time trip count");
}
DEBUG(dbgs() << "!\n");
}
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.
ValueToValueMapTy LastValueMap;
std::vector<PHINode*> OrigPHINode;
for (BasicBlock::iterator I = Header->begin(); isa<PHINode>(I); ++I) {
OrigPHINode.push_back(cast<PHINode>(I));
}
std::vector<BasicBlock*> Headers;
std::vector<BasicBlock*> Latches;
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;
SmallDenseMap<const Loop *, Loop *, 4> NewLoops;
NewLoops[L] = L;
for (LoopBlocksDFS::RPOIterator BB = BlockBegin; BB != BlockEnd; ++BB) {
ValueToValueMapTy VMap;
BasicBlock *New = CloneBasicBlock(*BB, VMap, "." + Twine(It));
Header->getParent()->getBasicBlockList().push_back(New);
// Tell LI about New.
if (*BB == Header) {
assert(LI->getLoopFor(*BB) == L && "Header should not be in a sub-loop");
L->addBasicBlockToLoop(New, *LI);
} else {
// Figure out which loop New is in.
const Loop *OldLoop = LI->getLoopFor(*BB);
assert(OldLoop && "Should (at least) be in the loop being unrolled!");
Loop *&NewLoop = NewLoops[OldLoop];
if (!NewLoop) {
// Found a new sub-loop.
assert(*BB == OldLoop->getHeader() &&
"Header should be first in RPO");
Loop *NewLoopParent = NewLoops.lookup(OldLoop->getParentLoop());
assert(NewLoopParent &&
"Expected parent loop before sub-loop in RPO");
NewLoop = new Loop;
NewLoopParent->addChildLoop(NewLoop);
// Forget the old loop, since its inputs may have changed.
if (SE)
SE->forgetLoop(OldLoop);
}
NewLoop->addBasicBlockToLoop(New, *LI);
}
if (*BB == Header)
// Loop over all of the PHI nodes in the block, changing them to use
// the incoming values from the previous block.
for (unsigned i = 0, e = OrigPHINode.size(); i != e; ++i) {
PHINode *NewPHI = cast<PHINode>(VMap[OrigPHINode[i]]);
Value *InVal = NewPHI->getIncomingValueForBlock(LatchBlock);
if (Instruction *InValI = dyn_cast<Instruction>(InVal))
if (It > 1 && L->contains(InValI))
InVal = LastValueMap[InValI];
VMap[OrigPHINode[i]] = InVal;
New->getInstList().erase(NewPHI);
}
// Update our running map of newest clones
LastValueMap[*BB] = New;
for (ValueToValueMapTy::iterator VI = VMap.begin(), VE = VMap.end();
VI != VE; ++VI)
LastValueMap[VI->first] = VI->second;
// 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)
Latches.push_back(New);
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);
}
// 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 (Instruction *InValI = dyn_cast<Instruction>(InVal)) {
if (L->contains(InValI))
InVal = LastValueMap[InVal];
}
assert(Latches.back() == LastValueMap[LatchBlock] && "bad last latch");
PN->addIncoming(InVal, Latches.back());
}
}
// 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;
if (RuntimeTripCount && j != 0) {
NeedConditional = false;
}
// For a complete unroll, make the last iteration end with a branch
// to the exit block.
if (CompletelyUnroll) {
if (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 {
// 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.
SmallPtrSet<Loop *, 4> ForgottenLoops;
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,
ForgottenLoops))
std::replace(Latches.begin(), Latches.end(), Dest, Fold);
}
}
// FIXME: We could register any cloned assumptions instead of clearing the
// whole function's cache.
AC->clear();
DominatorTree *DT = nullptr;
if (PP) {
// FIXME: Reconstruct dom info, because it is not preserved properly.
// Incrementally updating domtree after loop unrolling would be easy.
if (DominatorTreeWrapperPass *DTWP =
PP->getAnalysisIfAvailable<DominatorTreeWrapperPass>()) {
DT = &DTWP->getDomTree();
DT->recalculate(*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.
const DataLayout &DL = Header->getModule()->getDataLayout();
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 (Value *V = SimplifyInstruction(Inst, DL))
if (LI->replacementPreservesLCSSAForm(Inst, V)) {
Inst->replaceAllUsesWith(V);
(*BB)->getInstList().erase(Inst);
}
}
NumCompletelyUnrolled += CompletelyUnroll;
++NumUnrolled;
Loop *OuterL = L->getParentLoop();
// Remove the loop from the LoopPassManager if it's completely removed.
if (CompletelyUnroll && LPM != nullptr)
LPM->deleteLoopFromQueue(L);
// If we have a pass and a DominatorTree we should re-simplify impacted loops
// to ensure subsequent analyses can rely on this form. We want to simplify
// at least one layer outside of the loop that was unrolled so that any
// changes to the parent loop exposed by the unrolling are considered.
if (PP && DT) {
if (!OuterL && !CompletelyUnroll)
OuterL = L;
if (OuterL) {
simplifyLoop(OuterL, DT, LI, PP, SE, AC);
// LCSSA must be performed on the outermost affected loop. The unrolled
// loop's last loop latch is guaranteed to be in the outermost loop after
// deleteLoopFromQueue updates LoopInfo.
Loop *LatchLoop = LI->getLoopFor(Latches.back());
if (!OuterL->contains(LatchLoop))
while (OuterL->getParentLoop() != LatchLoop)
OuterL = OuterL->getParentLoop();
formLCSSARecursively(*OuterL, *DT, LI, SE);
}
}
return true;
}
/// CloneAndPruneFunctionInto - This works exactly like CloneFunctionInto,
/// except that it does some simple constant prop and DCE on the fly. The
/// effect of this is to copy significantly less code in cases where (for
/// example) a function call with constant arguments is inlined, and those
/// constant arguments cause a significant amount of code in the callee to be
/// dead. Since this doesn't produce an exact copy of the input, it can't be
/// used for things like CloneFunction or CloneModule.
void llvm::CloneAndPruneFunctionInto(Function *NewFunc, const Function *OldFunc,
ValueToValueMapTy &VMap,
SmallVectorImpl<ReturnInst*> &Returns,
const char *NameSuffix,
ClonedCodeInfo *CodeInfo,
const TargetData *TD,
Instruction *TheCall) {
assert(NameSuffix && "NameSuffix cannot be null!");
#ifndef NDEBUG
for (Function::const_arg_iterator II = OldFunc->arg_begin(),
E = OldFunc->arg_end(); II != E; ++II)
assert(VMap.count(II) && "No mapping from source argument specified!");
#endif
PruningFunctionCloner PFC(NewFunc, OldFunc, VMap, Returns,
NameSuffix, CodeInfo, TD);
// Clone the entry block, and anything recursively reachable from it.
std::vector<const BasicBlock*> CloneWorklist;
CloneWorklist.push_back(&OldFunc->getEntryBlock());
while (!CloneWorklist.empty()) {
const BasicBlock *BB = CloneWorklist.back();
CloneWorklist.pop_back();
PFC.CloneBlock(BB, CloneWorklist);
}
// Loop over all of the basic blocks in the old function. If the block was
// reachable, we have cloned it and the old block is now in the value map:
// insert it into the new function in the right order. If not, ignore it.
//
// Defer PHI resolution until rest of function is resolved.
SmallVector<const PHINode*, 16> PHIToResolve;
for (Function::const_iterator BI = OldFunc->begin(), BE = OldFunc->end();
BI != BE; ++BI) {
BasicBlock *NewBB = cast_or_null<BasicBlock>(VMap[BI]);
if (NewBB == 0) continue; // Dead block.
// Add the new block to the new function.
NewFunc->getBasicBlockList().push_back(NewBB);
// Loop over all of the instructions in the block, fixing up operand
// references as we go. This uses VMap to do all the hard work.
//
BasicBlock::iterator I = NewBB->begin();
unsigned DbgKind = OldFunc->getContext().getMDKindID("dbg");
MDNode *TheCallMD = NULL;
if (TheCall && TheCall->hasMetadata())
TheCallMD = TheCall->getMetadata(DbgKind);
// Handle PHI nodes specially, as we have to remove references to dead
// blocks.
if (PHINode *PN = dyn_cast<PHINode>(I)) {
// Skip over all PHI nodes, remembering them for later.
BasicBlock::const_iterator OldI = BI->begin();
for (; (PN = dyn_cast<PHINode>(I)); ++I, ++OldI) {
if (I->hasMetadata()) {
if (TheCallMD) {
if (MDNode *IMD = I->getMetadata(DbgKind)) {
MDNode *NewMD = UpdateInlinedAtInfo(IMD, TheCallMD);
I->setMetadata(DbgKind, NewMD);
}
} else {
// The cloned instruction has dbg info but the call instruction
// does not have dbg info. Remove dbg info from cloned instruction.
I->setMetadata(DbgKind, 0);
}
}
PHIToResolve.push_back(cast<PHINode>(OldI));
}
}
// FIXME:
// FIXME:
// FIXME: Unclone all this metadata stuff.
// FIXME:
// FIXME:
// Otherwise, remap the rest of the instructions normally.
for (; I != NewBB->end(); ++I) {
if (I->hasMetadata()) {
if (TheCallMD) {
if (MDNode *IMD = I->getMetadata(DbgKind)) {
MDNode *NewMD = UpdateInlinedAtInfo(IMD, TheCallMD);
I->setMetadata(DbgKind, NewMD);
}
} else {
// The cloned instruction has dbg info but the call instruction
// does not have dbg info. Remove dbg info from cloned instruction.
I->setMetadata(DbgKind, 0);
}
}
RemapInstruction(I, VMap);
}
}
// Defer PHI resolution until rest of function is resolved, PHI resolution
// requires the CFG to be up-to-date.
for (unsigned phino = 0, e = PHIToResolve.size(); phino != e; ) {
const PHINode *OPN = PHIToResolve[phino];
unsigned NumPreds = OPN->getNumIncomingValues();
const BasicBlock *OldBB = OPN->getParent();
BasicBlock *NewBB = cast<BasicBlock>(VMap[OldBB]);
// Map operands for blocks that are live and remove operands for blocks
// that are dead.
for (; phino != PHIToResolve.size() &&
PHIToResolve[phino]->getParent() == OldBB; ++phino) {
OPN = PHIToResolve[phino];
PHINode *PN = cast<PHINode>(VMap[OPN]);
for (unsigned pred = 0, e = NumPreds; pred != e; ++pred) {
if (BasicBlock *MappedBlock =
cast_or_null<BasicBlock>(VMap[PN->getIncomingBlock(pred)])) {
Value *InVal = MapValue(PN->getIncomingValue(pred),
VMap);
assert(InVal && "Unknown input value?");
PN->setIncomingValue(pred, InVal);
PN->setIncomingBlock(pred, MappedBlock);
} else {
PN->removeIncomingValue(pred, false);
--pred, --e; // Revisit the next entry.
}
}
}
// The loop above has removed PHI entries for those blocks that are dead
// and has updated others. However, if a block is live (i.e. copied over)
// but its terminator has been changed to not go to this block, then our
// phi nodes will have invalid entries. Update the PHI nodes in this
// case.
PHINode *PN = cast<PHINode>(NewBB->begin());
NumPreds = std::distance(pred_begin(NewBB), pred_end(NewBB));
if (NumPreds != PN->getNumIncomingValues()) {
assert(NumPreds < PN->getNumIncomingValues());
// Count how many times each predecessor comes to this block.
std::map<BasicBlock*, unsigned> PredCount;
for (pred_iterator PI = pred_begin(NewBB), E = pred_end(NewBB);
PI != E; ++PI)
--PredCount[*PI];
// Figure out how many entries to remove from each PHI.
for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
++PredCount[PN->getIncomingBlock(i)];
// At this point, the excess predecessor entries are positive in the
// map. Loop over all of the PHIs and remove excess predecessor
// entries.
BasicBlock::iterator I = NewBB->begin();
for (; (PN = dyn_cast<PHINode>(I)); ++I) {
for (std::map<BasicBlock*, unsigned>::iterator PCI =PredCount.begin(),
E = PredCount.end(); PCI != E; ++PCI) {
BasicBlock *Pred = PCI->first;
for (unsigned NumToRemove = PCI->second; NumToRemove; --NumToRemove)
PN->removeIncomingValue(Pred, false);
}
}
}
// If the loops above have made these phi nodes have 0 or 1 operand,
// replace them with undef or the input value. We must do this for
// correctness, because 0-operand phis are not valid.
PN = cast<PHINode>(NewBB->begin());
if (PN->getNumIncomingValues() == 0) {
BasicBlock::iterator I = NewBB->begin();
BasicBlock::const_iterator OldI = OldBB->begin();
while ((PN = dyn_cast<PHINode>(I++))) {
Value *NV = UndefValue::get(PN->getType());
PN->replaceAllUsesWith(NV);
assert(VMap[OldI] == PN && "VMap mismatch");
VMap[OldI] = NV;
PN->eraseFromParent();
++OldI;
}
}
// NOTE: We cannot eliminate single entry phi nodes here, because of
// VMap. Single entry phi nodes can have multiple VMap entries
// pointing at them. Thus, deleting one would require scanning the VMap
// to update any entries in it that would require that. This would be
// really slow.
}
// Now that the inlined function body has been fully constructed, go through
// and zap unconditional fall-through branches. This happen all the time when
// specializing code: code specialization turns conditional branches into
// uncond branches, and this code folds them.
Function::iterator I = cast<BasicBlock>(VMap[&OldFunc->getEntryBlock()]);
while (I != NewFunc->end()) {
BranchInst *BI = dyn_cast<BranchInst>(I->getTerminator());
if (!BI || BI->isConditional()) { ++I; continue; }
// Note that we can't eliminate uncond branches if the destination has
// single-entry PHI nodes. Eliminating the single-entry phi nodes would
// require scanning the VMap to update any entries that point to the phi
// node.
BasicBlock *Dest = BI->getSuccessor(0);
if (!Dest->getSinglePredecessor() || isa<PHINode>(Dest->begin())) {
++I; continue;
}
// We know all single-entry PHI nodes in the inlined function have been
// removed, so we just need to splice the blocks.
BI->eraseFromParent();
// Move all the instructions in the succ to the pred.
I->getInstList().splice(I->end(), Dest->getInstList());
// Make all PHI nodes that referred to Dest now refer to I as their source.
Dest->replaceAllUsesWith(I);
// Remove the dest block.
Dest->eraseFromParent();
// Do not increment I, iteratively merge all things this block branches to.
}
}
bool PPCCTRLoops::convertToCTRLoop(Loop *L) {
bool MadeChange = false;
// Do not convert small short loops to CTR loop.
unsigned ConstTripCount = SE->getSmallConstantTripCount(L);
if (ConstTripCount && ConstTripCount < SmallCTRLoopThreshold) {
SmallPtrSet<const Value *, 32> EphValues;
auto AC = getAnalysis<AssumptionCacheTracker>().getAssumptionCache(
*L->getHeader()->getParent());
CodeMetrics::collectEphemeralValues(L, &AC, EphValues);
CodeMetrics Metrics;
for (BasicBlock *BB : L->blocks())
Metrics.analyzeBasicBlock(BB, *TTI, EphValues);
// 6 is an approximate latency for the mtctr instruction.
if (Metrics.NumInsts <= (6 * SchedModel.getIssueWidth()))
return false;
}
// Process nested loops first.
for (Loop::iterator I = L->begin(), E = L->end(); I != E; ++I) {
MadeChange |= convertToCTRLoop(*I);
DEBUG(dbgs() << "Nested loop converted\n");
}
// If a nested loop has been converted, then we can't convert this loop.
if (MadeChange)
return MadeChange;
#ifndef NDEBUG
// Stop trying after reaching the limit (if any).
int Limit = CTRLoopLimit;
if (Limit >= 0) {
if (Counter >= CTRLoopLimit)
return false;
Counter++;
}
#endif
// We don't want to spill/restore the counter register, and so we don't
// want to use the counter register if the loop contains calls.
for (Loop::block_iterator I = L->block_begin(), IE = L->block_end();
I != IE; ++I)
if (mightUseCTR(*I))
return MadeChange;
SmallVector<BasicBlock*, 4> ExitingBlocks;
L->getExitingBlocks(ExitingBlocks);
BasicBlock *CountedExitBlock = nullptr;
const SCEV *ExitCount = nullptr;
BranchInst *CountedExitBranch = nullptr;
for (SmallVectorImpl<BasicBlock *>::iterator I = ExitingBlocks.begin(),
IE = ExitingBlocks.end(); I != IE; ++I) {
const SCEV *EC = SE->getExitCount(L, *I);
DEBUG(dbgs() << "Exit Count for " << *L << " from block " <<
(*I)->getName() << ": " << *EC << "\n");
if (isa<SCEVCouldNotCompute>(EC))
continue;
if (const SCEVConstant *ConstEC = dyn_cast<SCEVConstant>(EC)) {
if (ConstEC->getValue()->isZero())
continue;
} else if (!SE->isLoopInvariant(EC, L))
continue;
if (SE->getTypeSizeInBits(EC->getType()) > (TM->isPPC64() ? 64 : 32))
continue;
// We now have a loop-invariant count of loop iterations (which is not the
// constant zero) for which we know that this loop will not exit via this
// exisiting block.
// We need to make sure that this block will run on every loop iteration.
// For this to be true, we must dominate all blocks with backedges. Such
// blocks are in-loop predecessors to the header block.
bool NotAlways = false;
for (pred_iterator PI = pred_begin(L->getHeader()),
PIE = pred_end(L->getHeader()); PI != PIE; ++PI) {
if (!L->contains(*PI))
continue;
if (!DT->dominates(*I, *PI)) {
NotAlways = true;
break;
}
}
if (NotAlways)
continue;
// Make sure this blocks ends with a conditional branch.
Instruction *TI = (*I)->getTerminator();
if (!TI)
continue;
if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
if (!BI->isConditional())
continue;
CountedExitBranch = BI;
} else
continue;
// Note that this block may not be the loop latch block, even if the loop
// has a latch block.
CountedExitBlock = *I;
ExitCount = EC;
break;
}
if (!CountedExitBlock)
return MadeChange;
BasicBlock *Preheader = L->getLoopPreheader();
// If we don't have a preheader, then insert one. If we already have a
// preheader, then we can use it (except if the preheader contains a use of
// the CTR register because some such uses might be reordered by the
// selection DAG after the mtctr instruction).
if (!Preheader || mightUseCTR(Preheader))
Preheader = InsertPreheaderForLoop(L, DT, LI, PreserveLCSSA);
if (!Preheader)
return MadeChange;
DEBUG(dbgs() << "Preheader for exit count: " << Preheader->getName() << "\n");
// Insert the count into the preheader and replace the condition used by the
// selected branch.
MadeChange = true;
SCEVExpander SCEVE(*SE, *DL, "loopcnt");
LLVMContext &C = SE->getContext();
Type *CountType = TM->isPPC64() ? Type::getInt64Ty(C) : Type::getInt32Ty(C);
if (!ExitCount->getType()->isPointerTy() &&
ExitCount->getType() != CountType)
ExitCount = SE->getZeroExtendExpr(ExitCount, CountType);
ExitCount = SE->getAddExpr(ExitCount, SE->getOne(CountType));
Value *ECValue =
SCEVE.expandCodeFor(ExitCount, CountType, Preheader->getTerminator());
IRBuilder<> CountBuilder(Preheader->getTerminator());
Module *M = Preheader->getParent()->getParent();
Value *MTCTRFunc = Intrinsic::getDeclaration(M, Intrinsic::ppc_mtctr,
CountType);
CountBuilder.CreateCall(MTCTRFunc, ECValue);
IRBuilder<> CondBuilder(CountedExitBranch);
Value *DecFunc =
Intrinsic::getDeclaration(M, Intrinsic::ppc_is_decremented_ctr_nonzero);
Value *NewCond = CondBuilder.CreateCall(DecFunc, {});
Value *OldCond = CountedExitBranch->getCondition();
CountedExitBranch->setCondition(NewCond);
// The false branch must exit the loop.
if (!L->contains(CountedExitBranch->getSuccessor(0)))
CountedExitBranch->swapSuccessors();
// The old condition may be dead now, and may have even created a dead PHI
// (the original induction variable).
RecursivelyDeleteTriviallyDeadInstructions(OldCond);
// Run through the basic blocks of the loop and see if any of them have dead
// PHIs that can be removed.
for (auto I : L->blocks())
DeleteDeadPHIs(I);
++NumCTRLoops;
return MadeChange;
}
bool LoopBogusCF::runOnLoop(Loop *loop, LPPassManager &LPM) {
if (disableLoopBcf)
return false;
++NumLoops;
DEBUG(errs() << "LoopBogusCF: Dumping loop info\n");
// DEBUG(loop->dump());
BasicBlock *header = loop->getHeader();
BranchInst *branch = dyn_cast<BranchInst>(header->getTerminator());
if (!branch || !branch->isConditional()) {
DEBUG(errs() << "\t Not trivial loop -- skipping\n");
return false;
}
if (!loop->isLoopSimplifyForm()) {
DEBUG(errs() << "\t Not simplified loop -- skipping\n");
return false;
}
BasicBlock *exitBlock = loop->getUniqueExitBlock();
if (!exitBlock) {
DEBUG(errs() << "\t No unique exit block -- skipping\n");
return false;
}
if (branch->getSuccessor(0) != exitBlock &&
branch->getSuccessor(1) != exitBlock) {
DEBUG(errs() << "\t Not trivial loop -- skipping\n");
return false;
}
++NumLoopsObf;
// DEBUG(header->getParent()->viewCFG());
DEBUG(errs() << "\tCreating dummy block\n");
LoopInfo &info = getAnalysis<LoopInfo>();
// Split header block
BasicBlock *dummy = header->splitBasicBlock(header->getTerminator());
loop->addBasicBlockToLoop(dummy, info.getBase());
BasicBlock *trueBlock, *falseBlock = exitBlock;
if (branch->getSuccessor(0) == exitBlock) {
trueBlock = branch->getSuccessor(1);
branch->setSuccessor(1, dummy);
} else {
trueBlock = branch->getSuccessor(0);
branch->setSuccessor(0, dummy);
}
branch->moveBefore(header->getTerminator());
header->getTerminator()->eraseFromParent();
OpaquePredicate::createStub(dummy, trueBlock, falseBlock,
OpaquePredicate::PredicateTrue, false);
// DEBUG(header->getParent()->viewCFG());
return true;
}
bool PPCCTRLoops::convertToCTRLoop(Loop *L) {
bool MadeChange = false;
Triple TT = Triple(L->getHeader()->getParent()->getParent()->
getTargetTriple());
if (!TT.isArch32Bit() && !TT.isArch64Bit())
return MadeChange; // Unknown arch. type.
// Process nested loops first.
for (Loop::iterator I = L->begin(), E = L->end(); I != E; ++I) {
MadeChange |= convertToCTRLoop(*I);
}
// If a nested loop has been converted, then we can't convert this loop.
if (MadeChange)
return MadeChange;
#ifndef NDEBUG
// Stop trying after reaching the limit (if any).
int Limit = CTRLoopLimit;
if (Limit >= 0) {
if (Counter >= CTRLoopLimit)
return false;
Counter++;
}
#endif
// We don't want to spill/restore the counter register, and so we don't
// want to use the counter register if the loop contains calls.
for (Loop::block_iterator I = L->block_begin(), IE = L->block_end();
I != IE; ++I)
if (mightUseCTR(TT, *I))
return MadeChange;
SmallVector<BasicBlock*, 4> ExitingBlocks;
L->getExitingBlocks(ExitingBlocks);
BasicBlock *CountedExitBlock = 0;
const SCEV *ExitCount = 0;
BranchInst *CountedExitBranch = 0;
for (SmallVector<BasicBlock*, 4>::iterator I = ExitingBlocks.begin(),
IE = ExitingBlocks.end(); I != IE; ++I) {
const SCEV *EC = SE->getExitCount(L, *I);
DEBUG(dbgs() << "Exit Count for " << *L << " from block " <<
(*I)->getName() << ": " << *EC << "\n");
if (isa<SCEVCouldNotCompute>(EC))
continue;
if (const SCEVConstant *ConstEC = dyn_cast<SCEVConstant>(EC)) {
if (ConstEC->getValue()->isZero())
continue;
} else if (!SE->isLoopInvariant(EC, L))
continue;
if (SE->getTypeSizeInBits(EC->getType()) > (TT.isArch64Bit() ? 64 : 32))
continue;
// We now have a loop-invariant count of loop iterations (which is not the
// constant zero) for which we know that this loop will not exit via this
// exisiting block.
// We need to make sure that this block will run on every loop iteration.
// For this to be true, we must dominate all blocks with backedges. Such
// blocks are in-loop predecessors to the header block.
bool NotAlways = false;
for (pred_iterator PI = pred_begin(L->getHeader()),
PIE = pred_end(L->getHeader()); PI != PIE; ++PI) {
if (!L->contains(*PI))
continue;
if (!DT->dominates(*I, *PI)) {
NotAlways = true;
break;
}
}
if (NotAlways)
continue;
// Make sure this blocks ends with a conditional branch.
Instruction *TI = (*I)->getTerminator();
if (!TI)
continue;
if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
if (!BI->isConditional())
continue;
CountedExitBranch = BI;
} else
continue;
// Note that this block may not be the loop latch block, even if the loop
// has a latch block.
CountedExitBlock = *I;
ExitCount = EC;
break;
}
if (!CountedExitBlock)
return MadeChange;
BasicBlock *Preheader = L->getLoopPreheader();
// If we don't have a preheader, then insert one. If we already have a
// preheader, then we can use it (except if the preheader contains a use of
// the CTR register because some such uses might be reordered by the
// selection DAG after the mtctr instruction).
if (!Preheader || mightUseCTR(TT, Preheader))
Preheader = InsertPreheaderForLoop(L, this);
if (!Preheader)
return MadeChange;
DEBUG(dbgs() << "Preheader for exit count: " << Preheader->getName() << "\n");
// Insert the count into the preheader and replace the condition used by the
// selected branch.
MadeChange = true;
SCEVExpander SCEVE(*SE, "loopcnt");
LLVMContext &C = SE->getContext();
Type *CountType = TT.isArch64Bit() ? Type::getInt64Ty(C) :
Type::getInt32Ty(C);
if (!ExitCount->getType()->isPointerTy() &&
ExitCount->getType() != CountType)
ExitCount = SE->getZeroExtendExpr(ExitCount, CountType);
ExitCount = SE->getAddExpr(ExitCount,
SE->getConstant(CountType, 1));
Value *ECValue = SCEVE.expandCodeFor(ExitCount, CountType,
Preheader->getTerminator());
IRBuilder<> CountBuilder(Preheader->getTerminator());
Module *M = Preheader->getParent()->getParent();
Value *MTCTRFunc = Intrinsic::getDeclaration(M, Intrinsic::ppc_mtctr,
CountType);
CountBuilder.CreateCall(MTCTRFunc, ECValue);
IRBuilder<> CondBuilder(CountedExitBranch);
Value *DecFunc =
Intrinsic::getDeclaration(M, Intrinsic::ppc_is_decremented_ctr_nonzero);
Value *NewCond = CondBuilder.CreateCall(DecFunc);
Value *OldCond = CountedExitBranch->getCondition();
CountedExitBranch->setCondition(NewCond);
// The false branch must exit the loop.
if (!L->contains(CountedExitBranch->getSuccessor(0)))
CountedExitBranch->swapSuccessors();
// The old condition may be dead now, and may have even created a dead PHI
// (the original induction variable).
RecursivelyDeleteTriviallyDeadInstructions(OldCond);
DeleteDeadPHIs(CountedExitBlock);
++NumCTRLoops;
return MadeChange;
}
/// If \param [in] BB has more than one predecessor that is a conditional
/// branch, attempt to use parallel and/or for the branch condition. \returns
/// true on success.
///
/// Before:
/// ......
/// %cmp10 = fcmp une float %tmp1, %tmp2
/// br i1 %cmp1, label %if.then, label %lor.rhs
///
/// lor.rhs:
/// ......
/// %cmp11 = fcmp une float %tmp3, %tmp4
/// br i1 %cmp11, label %if.then, label %ifend
///
/// if.end: // the merge block
/// ......
///
/// if.then: // has two predecessors, both of them contains conditional branch.
/// ......
/// br label %if.end;
///
/// After:
/// ......
/// %cmp10 = fcmp une float %tmp1, %tmp2
/// ......
/// %cmp11 = fcmp une float %tmp3, %tmp4
/// %cmp12 = or i1 %cmp10, %cmp11 // parallel-or mode.
/// br i1 %cmp12, label %if.then, label %ifend
///
/// if.end:
/// ......
///
/// if.then:
/// ......
/// br label %if.end;
///
/// Current implementation handles two cases.
/// Case 1: \param BB is on the else-path.
///
/// BB1
/// / |
/// BB2 |
/// / \ |
/// BB3 \ | where, BB1, BB2 contain conditional branches.
/// \ | / BB3 contains unconditional branch.
/// \ | / BB4 corresponds to \param BB which is also the merge.
/// BB => BB4
///
///
/// Corresponding source code:
///
/// if (a == b && c == d)
/// statement; // BB3
///
/// Case 2: \param BB BB is on the then-path.
///
/// BB1
/// / |
/// | BB2
/// \ / | where BB1, BB2 contain conditional branches.
/// BB => BB3 | BB3 contains unconditiona branch and corresponds
/// \ / to \param BB. BB4 is the merge.
/// BB4
///
/// Corresponding source code:
///
/// if (a == b || c == d)
/// statement; // BB3
///
/// In both cases, \param BB is the common successor of conditional branches.
/// In Case 1, \param BB (BB4) has an unconditional branch (BB3) as
/// its predecessor. In Case 2, \param BB (BB3) only has conditional branches
/// as its predecessors.
bool FlattenCFGOpt::FlattenParallelAndOr(BasicBlock *BB, IRBuilder<> &Builder) {
PHINode *PHI = dyn_cast<PHINode>(BB->begin());
if (PHI)
return false; // For simplicity, avoid cases containing PHI nodes.
BasicBlock *LastCondBlock = nullptr;
BasicBlock *FirstCondBlock = nullptr;
BasicBlock *UnCondBlock = nullptr;
int Idx = -1;
// Check predecessors of \param BB.
SmallPtrSet<BasicBlock *, 16> Preds(pred_begin(BB), pred_end(BB));
for (SmallPtrSetIterator<BasicBlock *> PI = Preds.begin(), PE = Preds.end();
PI != PE; ++PI) {
BasicBlock *Pred = *PI;
BranchInst *PBI = dyn_cast<BranchInst>(Pred->getTerminator());
// All predecessors should terminate with a branch.
if (!PBI)
return false;
BasicBlock *PP = Pred->getSinglePredecessor();
if (PBI->isUnconditional()) {
// Case 1: Pred (BB3) is an unconditional block, it should
// have a single predecessor (BB2) that is also a predecessor
// of \param BB (BB4) and should not have address-taken.
// There should exist only one such unconditional
// branch among the predecessors.
if (UnCondBlock || !PP || (Preds.count(PP) == 0) ||
Pred->hasAddressTaken())
return false;
UnCondBlock = Pred;
continue;
}
// Only conditional branches are allowed beyond this point.
assert(PBI->isConditional());
// Condition's unique use should be the branch instruction.
Value *PC = PBI->getCondition();
if (!PC || !PC->hasOneUse())
return false;
if (PP && Preds.count(PP)) {
// These are internal condition blocks to be merged from, e.g.,
// BB2 in both cases.
// Should not be address-taken.
if (Pred->hasAddressTaken())
return false;
// Instructions in the internal condition blocks should be safe
// to hoist up.
for (BasicBlock::iterator BI = Pred->begin(), BE = PBI->getIterator();
BI != BE;) {
Instruction *CI = &*BI++;
if (isa<PHINode>(CI) || !isSafeToSpeculativelyExecute(CI))
return false;
}
} else {
// This is the condition block to be merged into, e.g. BB1 in
// both cases.
if (FirstCondBlock)
return false;
FirstCondBlock = Pred;
}
// Find whether BB is uniformly on the true (or false) path
// for all of its predecessors.
BasicBlock *PS1 = PBI->getSuccessor(0);
BasicBlock *PS2 = PBI->getSuccessor(1);
BasicBlock *PS = (PS1 == BB) ? PS2 : PS1;
int CIdx = (PS1 == BB) ? 0 : 1;
if (Idx == -1)
Idx = CIdx;
else if (CIdx != Idx)
return false;
// PS is the successor which is not BB. Check successors to identify
// the last conditional branch.
if (Preds.count(PS) == 0) {
// Case 2.
LastCondBlock = Pred;
} else {
// Case 1
BranchInst *BPS = dyn_cast<BranchInst>(PS->getTerminator());
if (BPS && BPS->isUnconditional()) {
// Case 1: PS(BB3) should be an unconditional branch.
LastCondBlock = Pred;
}
}
}
if (!FirstCondBlock || !LastCondBlock || (FirstCondBlock == LastCondBlock))
return false;
TerminatorInst *TBB = LastCondBlock->getTerminator();
BasicBlock *PS1 = TBB->getSuccessor(0);
BasicBlock *PS2 = TBB->getSuccessor(1);
BranchInst *PBI1 = dyn_cast<BranchInst>(PS1->getTerminator());
BranchInst *PBI2 = dyn_cast<BranchInst>(PS2->getTerminator());
// If PS1 does not jump into PS2, but PS2 jumps into PS1,
// attempt branch inversion.
if (!PBI1 || !PBI1->isUnconditional() ||
(PS1->getTerminator()->getSuccessor(0) != PS2)) {
// Check whether PS2 jumps into PS1.
if (!PBI2 || !PBI2->isUnconditional() ||
(PS2->getTerminator()->getSuccessor(0) != PS1))
return false;
// Do branch inversion.
BasicBlock *CurrBlock = LastCondBlock;
bool EverChanged = false;
for (; CurrBlock != FirstCondBlock;
CurrBlock = CurrBlock->getSinglePredecessor()) {
BranchInst *BI = dyn_cast<BranchInst>(CurrBlock->getTerminator());
CmpInst *CI = dyn_cast<CmpInst>(BI->getCondition());
if (!CI)
continue;
CmpInst::Predicate Predicate = CI->getPredicate();
// Canonicalize icmp_ne -> icmp_eq, fcmp_one -> fcmp_oeq
if ((Predicate == CmpInst::ICMP_NE) || (Predicate == CmpInst::FCMP_ONE)) {
CI->setPredicate(ICmpInst::getInversePredicate(Predicate));
BI->swapSuccessors();
EverChanged = true;
}
}
return EverChanged;
}
// PS1 must have a conditional branch.
if (!PBI1 || !PBI1->isUnconditional())
return false;
// PS2 should not contain PHI node.
PHI = dyn_cast<PHINode>(PS2->begin());
if (PHI)
return false;
// Do the transformation.
BasicBlock *CB;
BranchInst *PBI = dyn_cast<BranchInst>(FirstCondBlock->getTerminator());
bool Iteration = true;
IRBuilder<>::InsertPointGuard Guard(Builder);
Value *PC = PBI->getCondition();
do {
CB = PBI->getSuccessor(1 - Idx);
// Delete the conditional branch.
FirstCondBlock->getInstList().pop_back();
FirstCondBlock->getInstList()
.splice(FirstCondBlock->end(), CB->getInstList());
PBI = cast<BranchInst>(FirstCondBlock->getTerminator());
Value *CC = PBI->getCondition();
// Merge conditions.
Builder.SetInsertPoint(PBI);
Value *NC;
if (Idx == 0)
// Case 2, use parallel or.
NC = Builder.CreateOr(PC, CC);
else
// Case 1, use parallel and.
NC = Builder.CreateAnd(PC, CC);
PBI->replaceUsesOfWith(CC, NC);
PC = NC;
if (CB == LastCondBlock)
Iteration = false;
// Remove internal conditional branches.
CB->dropAllReferences();
// make CB unreachable and let downstream to delete the block.
new UnreachableInst(CB->getContext(), CB);
} while (Iteration);
LLVM_DEBUG(dbgs() << "Use parallel and/or in:\n" << *FirstCondBlock);
return true;
}
/*
This method performs Unroll and Jam. For a simple loop like:
for (i = ..)
Fore(i)
for (j = ..)
SubLoop(i, j)
Aft(i)
Instead of doing normal inner or outer unrolling, we do:
for (i = .., i+=2)
Fore(i)
Fore(i+1)
for (j = ..)
SubLoop(i, j)
SubLoop(i+1, j)
Aft(i)
Aft(i+1)
So the outer loop is essetially unrolled and then the inner loops are fused
("jammed") together into a single loop. This can increase speed when there
are loads in SubLoop that are invariant to i, as they become shared between
the now jammed inner loops.
We do this by spliting the blocks in the loop into Fore, Subloop and Aft.
Fore blocks are those before the inner loop, Aft are those after. Normal
Unroll code is used to copy each of these sets of blocks and the results are
combined together into the final form above.
isSafeToUnrollAndJam should be used prior to calling this to make sure the
unrolling will be valid. Checking profitablility is also advisable.
*/
LoopUnrollResult
llvm::UnrollAndJamLoop(Loop *L, unsigned Count, unsigned TripCount,
unsigned TripMultiple, bool UnrollRemainder,
LoopInfo *LI, ScalarEvolution *SE, DominatorTree *DT,
AssumptionCache *AC, OptimizationRemarkEmitter *ORE) {
// When we enter here we should have already checked that it is safe
BasicBlock *Header = L->getHeader();
assert(L->getSubLoops().size() == 1);
Loop *SubLoop = *L->begin();
// Don't enter the unroll code if there is nothing to do.
if (TripCount == 0 && Count < 2) {
LLVM_DEBUG(dbgs() << "Won't unroll; almost nothing to do\n");
return LoopUnrollResult::Unmodified;
}
assert(Count > 0);
assert(TripMultiple > 0);
assert(TripCount == 0 || TripCount % TripMultiple == 0);
// Are we eliminating the loop control altogether?
bool CompletelyUnroll = (Count == TripCount);
// We use the runtime remainder in cases where we don't know trip multiple
if (TripMultiple == 1 || TripMultiple % Count != 0) {
if (!UnrollRuntimeLoopRemainder(L, Count, /*AllowExpensiveTripCount*/ false,
/*UseEpilogRemainder*/ true,
UnrollRemainder, LI, SE, DT, AC, true)) {
LLVM_DEBUG(dbgs() << "Won't unroll-and-jam; remainder loop could not be "
"generated when assuming runtime trip count\n");
return LoopUnrollResult::Unmodified;
}
}
// Notify ScalarEvolution that the loop will be substantially changed,
// if not outright eliminated.
if (SE) {
SE->forgetLoop(L);
SE->forgetLoop(SubLoop);
}
using namespace ore;
// Report the unrolling decision.
if (CompletelyUnroll) {
LLVM_DEBUG(dbgs() << "COMPLETELY UNROLL AND JAMMING loop %"
<< Header->getName() << " with trip count " << TripCount
<< "!\n");
ORE->emit(OptimizationRemark(DEBUG_TYPE, "FullyUnrolled", L->getStartLoc(),
L->getHeader())
<< "completely unroll and jammed loop with "
<< NV("UnrollCount", TripCount) << " iterations");
} else {
auto DiagBuilder = [&]() {
OptimizationRemark Diag(DEBUG_TYPE, "PartialUnrolled", L->getStartLoc(),
L->getHeader());
return Diag << "unroll and jammed loop by a factor of "
<< NV("UnrollCount", Count);
};
LLVM_DEBUG(dbgs() << "UNROLL AND JAMMING loop %" << Header->getName()
<< " by " << Count);
if (TripMultiple != 1) {
LLVM_DEBUG(dbgs() << " with " << TripMultiple << " trips per branch");
ORE->emit([&]() {
return DiagBuilder() << " with " << NV("TripMultiple", TripMultiple)
<< " trips per branch";
});
} else {
LLVM_DEBUG(dbgs() << " with run-time trip count");
ORE->emit([&]() { return DiagBuilder() << " with run-time trip count"; });
}
LLVM_DEBUG(dbgs() << "!\n");
}
BasicBlock *Preheader = L->getLoopPreheader();
BasicBlock *LatchBlock = L->getLoopLatch();
BranchInst *BI = dyn_cast<BranchInst>(LatchBlock->getTerminator());
assert(Preheader && LatchBlock && Header);
assert(BI && !BI->isUnconditional());
bool ContinueOnTrue = L->contains(BI->getSuccessor(0));
BasicBlock *LoopExit = BI->getSuccessor(ContinueOnTrue);
bool SubLoopContinueOnTrue = SubLoop->contains(
SubLoop->getLoopLatch()->getTerminator()->getSuccessor(0));
// Partition blocks in an outer/inner loop pair into blocks before and after
// the loop
BasicBlockSet SubLoopBlocks;
BasicBlockSet ForeBlocks;
BasicBlockSet AftBlocks;
partitionOuterLoopBlocks(L, SubLoop, ForeBlocks, SubLoopBlocks, AftBlocks,
DT);
// We keep track of the entering/first and exiting/last block of each of
// Fore/SubLoop/Aft in each iteration. This helps make the stapling up of
// blocks easier.
std::vector<BasicBlock *> ForeBlocksFirst;
std::vector<BasicBlock *> ForeBlocksLast;
std::vector<BasicBlock *> SubLoopBlocksFirst;
std::vector<BasicBlock *> SubLoopBlocksLast;
std::vector<BasicBlock *> AftBlocksFirst;
std::vector<BasicBlock *> AftBlocksLast;
ForeBlocksFirst.push_back(Header);
ForeBlocksLast.push_back(SubLoop->getLoopPreheader());
SubLoopBlocksFirst.push_back(SubLoop->getHeader());
SubLoopBlocksLast.push_back(SubLoop->getExitingBlock());
AftBlocksFirst.push_back(SubLoop->getExitBlock());
AftBlocksLast.push_back(L->getExitingBlock());
// Maps Blocks[0] -> Blocks[It]
ValueToValueMapTy LastValueMap;
// Move any instructions from fore phi operands from AftBlocks into Fore.
moveHeaderPhiOperandsToForeBlocks(
Header, LatchBlock, SubLoop->getLoopPreheader()->getTerminator(),
AftBlocks);
// 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();
if (Header->getParent()->isDebugInfoForProfiling())
for (BasicBlock *BB : L->getBlocks())
for (Instruction &I : *BB)
if (!isa<DbgInfoIntrinsic>(&I))
if (const DILocation *DIL = I.getDebugLoc())
I.setDebugLoc(DIL->cloneWithDuplicationFactor(Count));
// Copy all blocks
for (unsigned It = 1; It != Count; ++It) {
std::vector<BasicBlock *> NewBlocks;
// Maps Blocks[It] -> Blocks[It-1]
DenseMap<Value *, Value *> PrevItValueMap;
for (LoopBlocksDFS::RPOIterator BB = BlockBegin; BB != BlockEnd; ++BB) {
ValueToValueMapTy VMap;
BasicBlock *New = CloneBasicBlock(*BB, VMap, "." + Twine(It));
Header->getParent()->getBasicBlockList().push_back(New);
if (ForeBlocks.count(*BB)) {
L->addBasicBlockToLoop(New, *LI);
if (*BB == ForeBlocksFirst[0])
ForeBlocksFirst.push_back(New);
if (*BB == ForeBlocksLast[0])
ForeBlocksLast.push_back(New);
} else if (SubLoopBlocks.count(*BB)) {
SubLoop->addBasicBlockToLoop(New, *LI);
if (*BB == SubLoopBlocksFirst[0])
SubLoopBlocksFirst.push_back(New);
if (*BB == SubLoopBlocksLast[0])
SubLoopBlocksLast.push_back(New);
} else if (AftBlocks.count(*BB)) {
L->addBasicBlockToLoop(New, *LI);
if (*BB == AftBlocksFirst[0])
AftBlocksFirst.push_back(New);
if (*BB == AftBlocksLast[0])
AftBlocksLast.push_back(New);
} else {
llvm_unreachable("BB being cloned should be in Fore/Sub/Aft");
}
// Update our running maps of newest clones
PrevItValueMap[New] = (It == 1 ? *BB : LastValueMap[*BB]);
LastValueMap[*BB] = New;
for (ValueToValueMapTy::iterator VI = VMap.begin(), VE = VMap.end();
VI != VE; ++VI) {
PrevItValueMap[VI->second] =
const_cast<Value *>(It == 1 ? VI->first : LastValueMap[VI->first]);
LastValueMap[VI->first] = VI->second;
}
NewBlocks.push_back(New);
// Update DomTree:
if (*BB == ForeBlocksFirst[0])
DT->addNewBlock(New, ForeBlocksLast[It - 1]);
else if (*BB == SubLoopBlocksFirst[0])
DT->addNewBlock(New, SubLoopBlocksLast[It - 1]);
else if (*BB == AftBlocksFirst[0])
DT->addNewBlock(New, AftBlocksLast[It - 1]);
else {
// Each set of blocks (Fore/Sub/Aft) will have the same internal domtree
// structure.
auto BBDomNode = DT->getNode(*BB);
auto BBIDom = BBDomNode->getIDom();
BasicBlock *OriginalBBIDom = BBIDom->getBlock();
assert(OriginalBBIDom);
assert(LastValueMap[cast<Value>(OriginalBBIDom)]);
DT->addNewBlock(
New, cast<BasicBlock>(LastValueMap[cast<Value>(OriginalBBIDom)]));
}
}
// Remap all instructions in the most recent iteration
for (BasicBlock *NewBlock : NewBlocks) {
for (Instruction &I : *NewBlock) {
::remapInstruction(&I, LastValueMap);
if (auto *II = dyn_cast<IntrinsicInst>(&I))
if (II->getIntrinsicID() == Intrinsic::assume)
AC->registerAssumption(II);
}
}
// Alter the ForeBlocks phi's, pointing them at the latest version of the
// value from the previous iteration's phis
for (PHINode &Phi : ForeBlocksFirst[It]->phis()) {
Value *OldValue = Phi.getIncomingValueForBlock(AftBlocksLast[It]);
assert(OldValue && "should have incoming edge from Aft[It]");
Value *NewValue = OldValue;
if (Value *PrevValue = PrevItValueMap[OldValue])
NewValue = PrevValue;
assert(Phi.getNumOperands() == 2);
Phi.setIncomingBlock(0, ForeBlocksLast[It - 1]);
Phi.setIncomingValue(0, NewValue);
Phi.removeIncomingValue(1);
}
}
// Now that all the basic blocks for the unrolled iterations are in place,
// finish up connecting the blocks and phi nodes. At this point LastValueMap
// is the last unrolled iterations values.
// Update Phis in BB from OldBB to point to NewBB
auto updatePHIBlocks = [](BasicBlock *BB, BasicBlock *OldBB,
BasicBlock *NewBB) {
for (PHINode &Phi : BB->phis()) {
int I = Phi.getBasicBlockIndex(OldBB);
Phi.setIncomingBlock(I, NewBB);
}
};
// Update Phis in BB from OldBB to point to NewBB and use the latest value
// from LastValueMap
auto updatePHIBlocksAndValues = [](BasicBlock *BB, BasicBlock *OldBB,
BasicBlock *NewBB,
ValueToValueMapTy &LastValueMap) {
for (PHINode &Phi : BB->phis()) {
for (unsigned b = 0; b < Phi.getNumIncomingValues(); ++b) {
if (Phi.getIncomingBlock(b) == OldBB) {
Value *OldValue = Phi.getIncomingValue(b);
if (Value *LastValue = LastValueMap[OldValue])
Phi.setIncomingValue(b, LastValue);
Phi.setIncomingBlock(b, NewBB);
break;
}
}
}
};
// Move all the phis from Src into Dest
auto movePHIs = [](BasicBlock *Src, BasicBlock *Dest) {
Instruction *insertPoint = Dest->getFirstNonPHI();
while (PHINode *Phi = dyn_cast<PHINode>(Src->begin()))
Phi->moveBefore(insertPoint);
};
// Update the PHI values outside the loop to point to the last block
updatePHIBlocksAndValues(LoopExit, AftBlocksLast[0], AftBlocksLast.back(),
LastValueMap);
// Update ForeBlocks successors and phi nodes
BranchInst *ForeTerm =
cast<BranchInst>(ForeBlocksLast.back()->getTerminator());
BasicBlock *Dest = SubLoopBlocksFirst[0];
ForeTerm->setSuccessor(0, Dest);
if (CompletelyUnroll) {
while (PHINode *Phi = dyn_cast<PHINode>(ForeBlocksFirst[0]->begin())) {
Phi->replaceAllUsesWith(Phi->getIncomingValueForBlock(Preheader));
Phi->getParent()->getInstList().erase(Phi);
}
} else {
// Update the PHI values to point to the last aft block
updatePHIBlocksAndValues(ForeBlocksFirst[0], AftBlocksLast[0],
AftBlocksLast.back(), LastValueMap);
}
for (unsigned It = 1; It != Count; It++) {
// Remap ForeBlock successors from previous iteration to this
BranchInst *ForeTerm =
cast<BranchInst>(ForeBlocksLast[It - 1]->getTerminator());
BasicBlock *Dest = ForeBlocksFirst[It];
ForeTerm->setSuccessor(0, Dest);
}
// Subloop successors and phis
BranchInst *SubTerm =
cast<BranchInst>(SubLoopBlocksLast.back()->getTerminator());
SubTerm->setSuccessor(!SubLoopContinueOnTrue, SubLoopBlocksFirst[0]);
SubTerm->setSuccessor(SubLoopContinueOnTrue, AftBlocksFirst[0]);
updatePHIBlocks(SubLoopBlocksFirst[0], ForeBlocksLast[0],
ForeBlocksLast.back());
updatePHIBlocks(SubLoopBlocksFirst[0], SubLoopBlocksLast[0],
SubLoopBlocksLast.back());
for (unsigned It = 1; It != Count; It++) {
// Replace the conditional branch of the previous iteration subloop with an
// unconditional one to this one
BranchInst *SubTerm =
cast<BranchInst>(SubLoopBlocksLast[It - 1]->getTerminator());
BranchInst::Create(SubLoopBlocksFirst[It], SubTerm);
SubTerm->eraseFromParent();
updatePHIBlocks(SubLoopBlocksFirst[It], ForeBlocksLast[It],
ForeBlocksLast.back());
updatePHIBlocks(SubLoopBlocksFirst[It], SubLoopBlocksLast[It],
SubLoopBlocksLast.back());
movePHIs(SubLoopBlocksFirst[It], SubLoopBlocksFirst[0]);
}
// Aft blocks successors and phis
BranchInst *Term = cast<BranchInst>(AftBlocksLast.back()->getTerminator());
if (CompletelyUnroll) {
BranchInst::Create(LoopExit, Term);
Term->eraseFromParent();
} else {
Term->setSuccessor(!ContinueOnTrue, ForeBlocksFirst[0]);
}
updatePHIBlocks(AftBlocksFirst[0], SubLoopBlocksLast[0],
SubLoopBlocksLast.back());
for (unsigned It = 1; It != Count; It++) {
// Replace the conditional branch of the previous iteration subloop with an
// unconditional one to this one
BranchInst *AftTerm =
cast<BranchInst>(AftBlocksLast[It - 1]->getTerminator());
BranchInst::Create(AftBlocksFirst[It], AftTerm);
AftTerm->eraseFromParent();
updatePHIBlocks(AftBlocksFirst[It], SubLoopBlocksLast[It],
SubLoopBlocksLast.back());
movePHIs(AftBlocksFirst[It], AftBlocksFirst[0]);
}
// Dominator Tree. Remove the old links between Fore, Sub and Aft, adding the
// new ones required.
if (Count != 1) {
SmallVector<DominatorTree::UpdateType, 4> DTUpdates;
DTUpdates.emplace_back(DominatorTree::UpdateKind::Delete, ForeBlocksLast[0],
SubLoopBlocksFirst[0]);
DTUpdates.emplace_back(DominatorTree::UpdateKind::Delete,
SubLoopBlocksLast[0], AftBlocksFirst[0]);
DTUpdates.emplace_back(DominatorTree::UpdateKind::Insert,
ForeBlocksLast.back(), SubLoopBlocksFirst[0]);
DTUpdates.emplace_back(DominatorTree::UpdateKind::Insert,
SubLoopBlocksLast.back(), AftBlocksFirst[0]);
DT->applyUpdates(DTUpdates);
}
// Merge adjacent basic blocks, if possible.
SmallPtrSet<BasicBlock *, 16> MergeBlocks;
MergeBlocks.insert(ForeBlocksLast.begin(), ForeBlocksLast.end());
MergeBlocks.insert(SubLoopBlocksLast.begin(), SubLoopBlocksLast.end());
MergeBlocks.insert(AftBlocksLast.begin(), AftBlocksLast.end());
while (!MergeBlocks.empty()) {
BasicBlock *BB = *MergeBlocks.begin();
BranchInst *Term = dyn_cast<BranchInst>(BB->getTerminator());
if (Term && Term->isUnconditional() && L->contains(Term->getSuccessor(0))) {
BasicBlock *Dest = Term->getSuccessor(0);
if (BasicBlock *Fold = foldBlockIntoPredecessor(Dest, LI, SE, DT)) {
// Don't remove BB and add Fold as they are the same BB
assert(Fold == BB);
(void)Fold;
MergeBlocks.erase(Dest);
} else
MergeBlocks.erase(BB);
} else
MergeBlocks.erase(BB);
}
// 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.
simplifyLoopAfterUnroll(SubLoop, true, LI, SE, DT, AC);
simplifyLoopAfterUnroll(L, !CompletelyUnroll && Count > 1, LI, SE, DT, AC);
NumCompletelyUnrolledAndJammed += CompletelyUnroll;
++NumUnrolledAndJammed;
#ifndef NDEBUG
// We shouldn't have done anything to break loop simplify form or LCSSA.
Loop *OuterL = L->getParentLoop();
Loop *OutestLoop = OuterL ? OuterL : (!CompletelyUnroll ? L : SubLoop);
assert(OutestLoop->isRecursivelyLCSSAForm(*DT, *LI));
if (!CompletelyUnroll)
assert(L->isLoopSimplifyForm());
assert(SubLoop->isLoopSimplifyForm());
assert(DT->verify());
#endif
// Update LoopInfo if the loop is completely removed.
if (CompletelyUnroll)
LI->erase(L);
return CompletelyUnroll ? LoopUnrollResult::FullyUnrolled
: LoopUnrollResult::PartiallyUnrolled;
}
bool llvm::UnrollRuntimeLoopRemainder(Loop *L, unsigned Count,
bool AllowExpensiveTripCount,
bool UseEpilogRemainder,
bool UnrollRemainder, LoopInfo *LI,
ScalarEvolution *SE, DominatorTree *DT,
AssumptionCache *AC, bool PreserveLCSSA,
Loop **ResultLoop) {
LLVM_DEBUG(dbgs() << "Trying runtime unrolling on Loop: \n");
LLVM_DEBUG(L->dump());
LLVM_DEBUG(UseEpilogRemainder ? dbgs() << "Using epilog remainder.\n"
: dbgs() << "Using prolog remainder.\n");
// Make sure the loop is in canonical form.
if (!L->isLoopSimplifyForm()) {
LLVM_DEBUG(dbgs() << "Not in simplify form!\n");
return false;
}
// Guaranteed by LoopSimplifyForm.
BasicBlock *Latch = L->getLoopLatch();
BasicBlock *Header = L->getHeader();
BranchInst *LatchBR = cast<BranchInst>(Latch->getTerminator());
if (!LatchBR || LatchBR->isUnconditional()) {
// The loop-rotate pass can be helpful to avoid this in many cases.
LLVM_DEBUG(
dbgs()
<< "Loop latch not terminated by a conditional branch.\n");
return false;
}
unsigned ExitIndex = LatchBR->getSuccessor(0) == Header ? 1 : 0;
BasicBlock *LatchExit = LatchBR->getSuccessor(ExitIndex);
if (L->contains(LatchExit)) {
// Cloning the loop basic blocks (`CloneLoopBlocks`) requires that one of the
// targets of the Latch be an exit block out of the loop.
LLVM_DEBUG(
dbgs()
<< "One of the loop latch successors must be the exit block.\n");
return false;
}
// These are exit blocks other than the target of the latch exiting block.
SmallVector<BasicBlock *, 4> OtherExits;
bool isMultiExitUnrollingEnabled =
canSafelyUnrollMultiExitLoop(L, OtherExits, LatchExit, PreserveLCSSA,
UseEpilogRemainder) &&
canProfitablyUnrollMultiExitLoop(L, OtherExits, LatchExit, PreserveLCSSA,
UseEpilogRemainder);
// Support only single exit and exiting block unless multi-exit loop unrolling is enabled.
if (!isMultiExitUnrollingEnabled &&
(!L->getExitingBlock() || OtherExits.size())) {
LLVM_DEBUG(
dbgs()
<< "Multiple exit/exiting blocks in loop and multi-exit unrolling not "
"enabled!\n");
return false;
}
// Use Scalar Evolution to compute the trip count. This allows more loops to
// be unrolled than relying on induction var simplification.
if (!SE)
return false;
// 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).
// We calculate the backedge count by using getExitCount on the Latch block,
// which is proven to be the only exiting block in this loop. This is same as
// calculating getBackedgeTakenCount on the loop (which computes SCEV for all
// exiting blocks).
const SCEV *BECountSC = SE->getExitCount(L, Latch);
if (isa<SCEVCouldNotCompute>(BECountSC) ||
!BECountSC->getType()->isIntegerTy()) {
LLVM_DEBUG(dbgs() << "Could not compute exit block SCEV\n");
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(BECountSC, SE->getConstant(BECountSC->getType(), 1));
if (isa<SCEVCouldNotCompute>(TripCountSC)) {
LLVM_DEBUG(dbgs() << "Could not compute trip count SCEV.\n");
return false;
}
BasicBlock *PreHeader = L->getLoopPreheader();
BranchInst *PreHeaderBR = cast<BranchInst>(PreHeader->getTerminator());
const DataLayout &DL = Header->getModule()->getDataLayout();
SCEVExpander Expander(*SE, DL, "loop-unroll");
if (!AllowExpensiveTripCount &&
Expander.isHighCostExpansion(TripCountSC, L, PreHeaderBR)) {
LLVM_DEBUG(dbgs() << "High cost for expanding trip count scev!\n");
return false;
}
// This constraint lets us deal with an overflowing trip count easily; see the
// comment on ModVal below.
if (Log2_32(Count) > BEWidth) {
LLVM_DEBUG(
dbgs()
<< "Count failed constraint on overflow trip count calculation.\n");
return false;
}
// Loop structure is the following:
//
// PreHeader
// Header
// ...
// Latch
// LatchExit
BasicBlock *NewPreHeader;
BasicBlock *NewExit = nullptr;
BasicBlock *PrologExit = nullptr;
BasicBlock *EpilogPreHeader = nullptr;
BasicBlock *PrologPreHeader = nullptr;
if (UseEpilogRemainder) {
// If epilog remainder
// Split PreHeader to insert a branch around loop for unrolling.
NewPreHeader = SplitBlock(PreHeader, PreHeader->getTerminator(), DT, LI);
NewPreHeader->setName(PreHeader->getName() + ".new");
// Split LatchExit to create phi nodes from branch above.
SmallVector<BasicBlock*, 4> Preds(predecessors(LatchExit));
NewExit = SplitBlockPredecessors(LatchExit, Preds, ".unr-lcssa", DT, LI,
nullptr, PreserveLCSSA);
// NewExit gets its DebugLoc from LatchExit, which is not part of the
// original Loop.
// Fix this by setting Loop's DebugLoc to NewExit.
auto *NewExitTerminator = NewExit->getTerminator();
NewExitTerminator->setDebugLoc(Header->getTerminator()->getDebugLoc());
// Split NewExit to insert epilog remainder loop.
EpilogPreHeader = SplitBlock(NewExit, NewExitTerminator, DT, LI);
EpilogPreHeader->setName(Header->getName() + ".epil.preheader");
} else {
// If prolog remainder
// Split the original preheader twice to insert prolog remainder loop
PrologPreHeader = SplitEdge(PreHeader, Header, DT, LI);
PrologPreHeader->setName(Header->getName() + ".prol.preheader");
PrologExit = SplitBlock(PrologPreHeader, PrologPreHeader->getTerminator(),
DT, LI);
PrologExit->setName(Header->getName() + ".prol.loopexit");
// Split PrologExit to get NewPreHeader.
NewPreHeader = SplitBlock(PrologExit, PrologExit->getTerminator(), DT, LI);
NewPreHeader->setName(PreHeader->getName() + ".new");
}
// Loop structure should be the following:
// Epilog Prolog
//
// PreHeader PreHeader
// *NewPreHeader *PrologPreHeader
// Header *PrologExit
// ... *NewPreHeader
// Latch Header
// *NewExit ...
// *EpilogPreHeader Latch
// LatchExit LatchExit
// Calculate conditions for branch around loop for unrolling
// in epilog case and around prolog remainder loop in prolog case.
// Compute the number of extra iterations required, which is:
// extra iterations = run-time trip count % loop unroll factor
PreHeaderBR = cast<BranchInst>(PreHeader->getTerminator());
Value *TripCount = Expander.expandCodeFor(TripCountSC, TripCountSC->getType(),
PreHeaderBR);
Value *BECount = Expander.expandCodeFor(BECountSC, BECountSC->getType(),
PreHeaderBR);
IRBuilder<> B(PreHeaderBR);
Value *ModVal;
// Calculate ModVal = (BECount + 1) % Count.
// Note that TripCount is BECount + 1.
if (isPowerOf2_32(Count)) {
// When Count is power of 2 we don't BECount for epilog case, however we'll
// need it for a branch around unrolling loop for prolog case.
ModVal = B.CreateAnd(TripCount, Count - 1, "xtraiter");
// 1. There are no iterations to be run in the prolog/epilog 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).
} else {
// As (BECount + 1) can potentially unsigned overflow we count
// (BECount % Count) + 1 which is overflow safe as BECount % Count < Count.
Value *ModValTmp = B.CreateURem(BECount,
ConstantInt::get(BECount->getType(),
Count));
Value *ModValAdd = B.CreateAdd(ModValTmp,
ConstantInt::get(ModValTmp->getType(), 1));
// At that point (BECount % Count) + 1 could be equal to Count.
// To handle this case we need to take mod by Count one more time.
ModVal = B.CreateURem(ModValAdd,
ConstantInt::get(BECount->getType(), Count),
"xtraiter");
}
Value *BranchVal =
UseEpilogRemainder ? B.CreateICmpULT(BECount,
ConstantInt::get(BECount->getType(),
Count - 1)) :
B.CreateIsNotNull(ModVal, "lcmp.mod");
BasicBlock *RemainderLoop = UseEpilogRemainder ? NewExit : PrologPreHeader;
BasicBlock *UnrollingLoop = UseEpilogRemainder ? NewPreHeader : PrologExit;
// Branch to either remainder (extra iterations) loop or unrolling loop.
B.CreateCondBr(BranchVal, RemainderLoop, UnrollingLoop);
PreHeaderBR->eraseFromParent();
if (DT) {
if (UseEpilogRemainder)
DT->changeImmediateDominator(NewExit, PreHeader);
else
DT->changeImmediateDominator(PrologExit, PreHeader);
}
Function *F = Header->getParent();
// Get an ordered list of blocks in the loop to help with the ordering of the
// cloned blocks in the prolog/epilog code
LoopBlocksDFS LoopBlocks(L);
LoopBlocks.perform(LI);
//
// For each extra loop iteration, create a copy of the loop's basic blocks
// and generate a condition that branches to the copy depending on the
// number of 'left over' iterations.
//
std::vector<BasicBlock *> NewBlocks;
ValueToValueMapTy VMap;
// For unroll factor 2 remainder loop will have 1 iterations.
// Do not create 1 iteration loop.
bool CreateRemainderLoop = (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
// iterations. This function adds the appropriate CFG connections.
BasicBlock *InsertBot = UseEpilogRemainder ? LatchExit : PrologExit;
BasicBlock *InsertTop = UseEpilogRemainder ? EpilogPreHeader : PrologPreHeader;
Loop *remainderLoop = CloneLoopBlocks(
L, ModVal, CreateRemainderLoop, UseEpilogRemainder, UnrollRemainder,
InsertTop, InsertBot,
NewPreHeader, NewBlocks, LoopBlocks, VMap, DT, LI);
// Insert the cloned blocks into the function.
F->getBasicBlockList().splice(InsertBot->getIterator(),
F->getBasicBlockList(),
NewBlocks[0]->getIterator(),
F->end());
// Now the loop blocks are cloned and the other exiting blocks from the
// remainder are connected to the original Loop's exit blocks. The remaining
// work is to update the phi nodes in the original loop, and take in the
// values from the cloned region.
for (auto *BB : OtherExits) {
for (auto &II : *BB) {
// Given we preserve LCSSA form, we know that the values used outside the
// loop will be used through these phi nodes at the exit blocks that are
// transformed below.
if (!isa<PHINode>(II))
break;
PHINode *Phi = cast<PHINode>(&II);
unsigned oldNumOperands = Phi->getNumIncomingValues();
// Add the incoming values from the remainder code to the end of the phi
// node.
for (unsigned i =0; i < oldNumOperands; i++){
Value *newVal = VMap.lookup(Phi->getIncomingValue(i));
// newVal can be a constant or derived from values outside the loop, and
// hence need not have a VMap value. Also, since lookup already generated
// a default "null" VMap entry for this value, we need to populate that
// VMap entry correctly, with the mapped entry being itself.
if (!newVal) {
newVal = Phi->getIncomingValue(i);
VMap[Phi->getIncomingValue(i)] = Phi->getIncomingValue(i);
}
Phi->addIncoming(newVal,
cast<BasicBlock>(VMap[Phi->getIncomingBlock(i)]));
}
}
#if defined(EXPENSIVE_CHECKS) && !defined(NDEBUG)
for (BasicBlock *SuccBB : successors(BB)) {
assert(!(any_of(OtherExits,
[SuccBB](BasicBlock *EB) { return EB == SuccBB; }) ||
SuccBB == LatchExit) &&
"Breaks the definition of dedicated exits!");
}
#endif
}
// Update the immediate dominator of the exit blocks and blocks that are
// reachable from the exit blocks. This is needed because we now have paths
// from both the original loop and the remainder code reaching the exit
// blocks. While the IDom of these exit blocks were from the original loop,
// now the IDom is the preheader (which decides whether the original loop or
// remainder code should run).
if (DT && !L->getExitingBlock()) {
SmallVector<BasicBlock *, 16> ChildrenToUpdate;
// NB! We have to examine the dom children of all loop blocks, not just
// those which are the IDom of the exit blocks. This is because blocks
// reachable from the exit blocks can have their IDom as the nearest common
// dominator of the exit blocks.
for (auto *BB : L->blocks()) {
auto *DomNodeBB = DT->getNode(BB);
for (auto *DomChild : DomNodeBB->getChildren()) {
auto *DomChildBB = DomChild->getBlock();
if (!L->contains(LI->getLoopFor(DomChildBB)))
ChildrenToUpdate.push_back(DomChildBB);
}
}
for (auto *BB : ChildrenToUpdate)
DT->changeImmediateDominator(BB, PreHeader);
}
// Loop structure should be the following:
// Epilog Prolog
//
// PreHeader PreHeader
// NewPreHeader PrologPreHeader
// Header PrologHeader
// ... ...
// Latch PrologLatch
// NewExit PrologExit
// EpilogPreHeader NewPreHeader
// EpilogHeader Header
// ... ...
// EpilogLatch Latch
// LatchExit LatchExit
// Rewrite the cloned instruction operands to use the values created when the
// clone is created.
for (BasicBlock *BB : NewBlocks) {
for (Instruction &I : *BB) {
RemapInstruction(&I, VMap,
RF_NoModuleLevelChanges | RF_IgnoreMissingLocals);
}
}
if (UseEpilogRemainder) {
// Connect the epilog code to the original loop and update the
// PHI functions.
ConnectEpilog(L, ModVal, NewExit, LatchExit, PreHeader,
EpilogPreHeader, NewPreHeader, VMap, DT, LI,
PreserveLCSSA);
// Update counter in loop for unrolling.
// I should be multiply of Count.
IRBuilder<> B2(NewPreHeader->getTerminator());
Value *TestVal = B2.CreateSub(TripCount, ModVal, "unroll_iter");
BranchInst *LatchBR = cast<BranchInst>(Latch->getTerminator());
B2.SetInsertPoint(LatchBR);
PHINode *NewIdx = PHINode::Create(TestVal->getType(), 2, "niter",
Header->getFirstNonPHI());
Value *IdxSub =
B2.CreateSub(NewIdx, ConstantInt::get(NewIdx->getType(), 1),
NewIdx->getName() + ".nsub");
Value *IdxCmp;
if (LatchBR->getSuccessor(0) == Header)
IdxCmp = B2.CreateIsNotNull(IdxSub, NewIdx->getName() + ".ncmp");
else
IdxCmp = B2.CreateIsNull(IdxSub, NewIdx->getName() + ".ncmp");
NewIdx->addIncoming(TestVal, NewPreHeader);
NewIdx->addIncoming(IdxSub, Latch);
LatchBR->setCondition(IdxCmp);
} else {
// Connect the prolog code to the original loop and update the
// PHI functions.
ConnectProlog(L, BECount, Count, PrologExit, LatchExit, PreHeader,
NewPreHeader, VMap, DT, LI, PreserveLCSSA);
}
// If this loop is nested, then the loop unroller changes the code in the any
// of its parent loops, so the Scalar Evolution pass needs to be run again.
SE->forgetTopmostLoop(L);
// Verify that the Dom Tree is correct.
#if defined(EXPENSIVE_CHECKS) && !defined(NDEBUG)
if (DT)
assert(DT->verify(DominatorTree::VerificationLevel::Full));
#endif
// Canonicalize to LoopSimplifyForm both original and remainder loops. We
// cannot rely on the LoopUnrollPass to do this because it only does
// canonicalization for parent/subloops and not the sibling loops.
if (OtherExits.size() > 0) {
// Generate dedicated exit blocks for the original loop, to preserve
// LoopSimplifyForm.
formDedicatedExitBlocks(L, DT, LI, nullptr, PreserveLCSSA);
// Generate dedicated exit blocks for the remainder loop if one exists, to
// preserve LoopSimplifyForm.
if (remainderLoop)
formDedicatedExitBlocks(remainderLoop, DT, LI, nullptr, PreserveLCSSA);
}
auto UnrollResult = LoopUnrollResult::Unmodified;
if (remainderLoop && UnrollRemainder) {
LLVM_DEBUG(dbgs() << "Unrolling remainder loop\n");
UnrollResult =
UnrollLoop(remainderLoop, /*Count*/ Count - 1, /*TripCount*/ Count - 1,
/*Force*/ false, /*AllowRuntime*/ false,
/*AllowExpensiveTripCount*/ false, /*PreserveCondBr*/ true,
/*PreserveOnlyFirst*/ false, /*TripMultiple*/ 1,
/*PeelCount*/ 0, /*UnrollRemainder*/ false, LI, SE, DT, AC,
/*ORE*/ nullptr, PreserveLCSSA);
}
if (ResultLoop && UnrollResult != LoopUnrollResult::FullyUnrolled)
*ResultLoop = remainderLoop;
NumRuntimeUnrolled++;
return true;
}
bool FunctionStaticSlicer::computeBC() {
bool changed = false;
#ifdef DEBUG_BC
errs() << " ====== BEG BC Computation ======\n";
#endif
if (!forward) {
PostDominanceFrontier &PDF = MP->getAnalysis<PostDominanceFrontier>(fun);
for (inst_iterator I = inst_begin(fun), E = inst_end(fun); I != E; I++) {
Instruction *i = &*I;
const InsInfo *ii = getInsInfo(i);
if (ii->isSliced())
continue;
BasicBlock *BB = i->getParent();
#ifdef DEBUG_BC
errs() << " ";
i->print(errs());
errs() << " -> bb=" << BB->getName() << '\n';
#endif
PostDominanceFrontier::const_iterator frontier = PDF.find(BB);
if (frontier == PDF.end())
continue;
changed |= updateRCSC(frontier->second.begin(), frontier->second.end());
}
}
else {
for (inst_iterator I = inst_begin(fun), E = inst_end(fun); I != E; I++) {
Instruction *i = &*I;
if (!isa<BranchInst>(i))
continue;
BranchInst * bi = dyn_cast<BranchInst>(i);
if (bi->isUnconditional()) // skip unconditional inst
continue;
const InsInfo *ii = getInsInfo(i);
if (ii->isSliced())
continue;
unsigned succs = bi->getNumSuccessors();
for (unsigned si = 0; si < succs; si++) {
BasicBlock * BB = bi->getSuccessor(si);
BasicBlock * Pred = BB->getUniquePredecessor();
if (Pred == NULL)
continue;
for (BasicBlock::iterator BI = BB->begin(), BE = BB->end(); BI != BE; BI++) {
InsInfo *bii = getInsInfo(&*BI);
bii->deslice();
for (ValSet::const_iterator VI = bii->DEF_begin(), VE = bii->DEF_end();
VI != VE; VI++) {
if (bii->addRC(*VI)) {
changed = true;
#ifdef DEBUG_RC
errs() << " added ";
printVal(*VI);
errs() << "\n";
#endif
}
}
}
}
}
}
#ifdef DEBUG_BC
errs() << " ====== END BC Computation ======\n";
#endif
return changed;
}
void Executor::executeBranch(Instruction *i, bool cond) {
assert(i && "Expecting an instruction!");
LastExecutedBB = i->getParent();
if (CollectBlockTraces) {
ExecutedBlocks.push_back(LastExecutedBB);
}
BranchInst *br = dyn_cast<BranchInst>(i);
if (br->isConditional()) {
// Loops: check if the control-flow
// jump back to a loop header.
std::string bbname = PathLoopInfo->getLoopLatchName(br);
BasicBlock *nextBB1 = br->getSuccessor(0);
BasicBlock *nextBB2 = br->getSuccessor(1);
if ((cond && bbname==nextBB1->getName().str())
|| (!cond && bbname==nextBB2->getName().str())) {
unsigned line = 0;
if (MDNode *N = br->getMetadata("dbg")) {
DILocation Loc(N);
line = Loc.getLineNumber();
}
//std::cout << "Current: " << LastExecutedBB->getName().str() << std::endl;
//std::cout << "LoopLatch: " << bbname << std::endl;
std::cout << "Violated property: unwinding assertion";
std::cout << " (line " << line << ")\n";
//exit(1);
DisabledSymbolicExeCurRun = true;
return;
}
// Next block to be executed
if (cond) {
NextBB = nextBB1;
} else {
NextBB = nextBB2;
}
} else {
// Next block to be executed
NextBB = br->getSuccessor(0);
}
if (!DisabledSymbolicExeCurRun) {
PathContext->propagatePointers(LastExecutedBB, NextBB);
}
// For program coverage analysis
//Profile->covered(br, cond);
//Profile->covered(br->getParent());
if(br->isConditional()) {
// Update the number of time we jump in the IR
GotoCount++;
if (GotoCount>MaxDepth) {
std::cout << "warning: maximum goto-count (";
std::cout << GotoCount << ") reached.\n";
DisabledSymbolicExeCurRun = true;
return;
}
if (DisabledSymbolicExeCurRun) {
return;
}
Value *vcond = br->getCondition();
// vcond in domain(S)
if (SMAP->contains(vcond)) {
SymbolPtr Svcond = SMAP->get(vcond);
// Report(S(vcond), branchTaken)
Path->addBranchCell(Svcond, cond, i);
}
}
}
/// [LIBUNWIND] Find the (possibly absent) call to @llvm.eh.selector
/// in the given landing pad. In principle, llvm.eh.exception is
/// required to be in the landing pad; in practice, SplitCriticalEdge
/// can break that invariant, and then inlining can break it further.
/// There's a real need for a reliable solution here, but until that
/// happens, we have some fragile workarounds here.
static EHSelectorInst *findSelectorForLandingPad(BasicBlock *lpad) {
// Look for an exception call in the actual landing pad.
EHExceptionInst *exn = findExceptionInBlock(lpad);
if (exn) return findSelectorForException(exn);
// Okay, if that failed, look for one in an obvious successor. If
// we find one, we'll fix the IR by moving things back to the
// landing pad.
bool dominates = true; // does the lpad dominate the exn call
BasicBlock *nonDominated = 0; // if not, the first non-dominated block
BasicBlock *lastDominated = 0; // and the block which branched to it
BasicBlock *exnBlock = lpad;
// We need to protect against lpads that lead into infinite loops.
SmallPtrSet<BasicBlock*,4> visited;
visited.insert(exnBlock);
do {
// We're not going to apply this hack to anything more complicated
// than a series of unconditional branches, so if the block
// doesn't terminate in an unconditional branch, just fail. More
// complicated cases can arise when, say, sinking a call into a
// split unwind edge and then inlining it; but that can do almost
// *anything* to the CFG, including leaving the selector
// completely unreachable. The only way to fix that properly is
// to (1) prohibit transforms which move the exception or selector
// values away from the landing pad, e.g. by producing them with
// instructions that are pinned to an edge like a phi, or
// producing them with not-really-instructions, and (2) making
// transforms which split edges deal with that.
BranchInst *branch = dyn_cast<BranchInst>(&exnBlock->back());
if (!branch || branch->isConditional()) return 0;
BasicBlock *successor = branch->getSuccessor(0);
// Fail if we found an infinite loop.
if (!visited.insert(successor)) return 0;
// If the successor isn't dominated by exnBlock:
if (!successor->getSinglePredecessor()) {
// We don't want to have to deal with threading the exception
// through multiple levels of phi, so give up if we've already
// followed a non-dominating edge.
if (!dominates) return 0;
// Otherwise, remember this as a non-dominating edge.
dominates = false;
nonDominated = successor;
lastDominated = exnBlock;
}
exnBlock = successor;
// Can we stop here?
exn = findExceptionInBlock(exnBlock);
} while (!exn);
// Look for a selector call for the exception we found.
EHSelectorInst *selector = findSelectorForException(exn);
if (!selector) return 0;
// The easy case is when the landing pad still dominates the
// exception call, in which case we can just move both calls back to
// the landing pad.
if (dominates) {
selector->moveBefore(lpad->getFirstNonPHI());
exn->moveBefore(selector);
return selector;
}
// Otherwise, we have to split at the first non-dominating block.
// The CFG looks basically like this:
// lpad:
// phis_0
// insnsAndBranches_1
// br label %nonDominated
// nonDominated:
// phis_2
// insns_3
// %exn = call i8* @llvm.eh.exception()
// insnsAndBranches_4
// %selector = call @llvm.eh.selector(i8* %exn, ...
// We need to turn this into:
// lpad:
// phis_0
// %exn0 = call i8* @llvm.eh.exception()
// %selector0 = call @llvm.eh.selector(i8* %exn0, ...
// insnsAndBranches_1
// br label %split // from lastDominated
// nonDominated:
// phis_2 (without edge from lastDominated)
// %exn1 = call i8* @llvm.eh.exception()
// %selector1 = call i8* @llvm.eh.selector(i8* %exn1, ...
// br label %split
// split:
// phis_2 (edge from lastDominated, edge from split)
// %exn = phi ...
// %selector = phi ...
// insns_3
// insnsAndBranches_4
assert(nonDominated);
assert(lastDominated);
// First, make clones of the intrinsics to go in lpad.
EHExceptionInst *lpadExn = cast<EHExceptionInst>(exn->clone());
EHSelectorInst *lpadSelector = cast<EHSelectorInst>(selector->clone());
lpadSelector->setArgOperand(0, lpadExn);
lpadSelector->insertBefore(lpad->getFirstNonPHI());
lpadExn->insertBefore(lpadSelector);
// Split the non-dominated block.
BasicBlock *split =
nonDominated->splitBasicBlock(nonDominated->getFirstNonPHI(),
nonDominated->getName() + ".lpad-fix");
// Redirect the last dominated branch there.
cast<BranchInst>(lastDominated->back()).setSuccessor(0, split);
// Move the existing intrinsics to the end of the old block.
selector->moveBefore(&nonDominated->back());
exn->moveBefore(selector);
Instruction *splitIP = &split->front();
// For all the phis in nonDominated, make a new phi in split to join
// that phi with the edge from lastDominated.
for (BasicBlock::iterator
i = nonDominated->begin(), e = nonDominated->end(); i != e; ++i) {
PHINode *phi = dyn_cast<PHINode>(i);
if (!phi) break;
PHINode *splitPhi = PHINode::Create(phi->getType(), 2, phi->getName(),
splitIP);
phi->replaceAllUsesWith(splitPhi);
splitPhi->addIncoming(phi, nonDominated);
splitPhi->addIncoming(phi->removeIncomingValue(lastDominated),
lastDominated);
}
// Make new phis for the exception and selector.
PHINode *exnPhi = PHINode::Create(exn->getType(), 2, "", splitIP);
exn->replaceAllUsesWith(exnPhi);
selector->setArgOperand(0, exn); // except for this use
exnPhi->addIncoming(exn, nonDominated);
exnPhi->addIncoming(lpadExn, lastDominated);
PHINode *selectorPhi = PHINode::Create(selector->getType(), 2, "", splitIP);
selector->replaceAllUsesWith(selectorPhi);
selectorPhi->addIncoming(selector, nonDominated);
selectorPhi->addIncoming(lpadSelector, lastDominated);
return lpadSelector;
}
/// Unroll the given loop by Count. The loop must be in LCSSA form. Returns true
/// 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.
///
/// 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");
return false;
}
BasicBlock *LatchBlock = L->getLoopLatch();
if (!LatchBlock) {
DEBUG(dbgs() << " Can't unroll; loop exit-block-insertion failed.\n");
return false;
}
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() <<
" Can't unroll; loop not terminated by a conditional branch.\n");
return false;
}
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 (TripMultiple != 1)
DEBUG(dbgs() << " 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;
// 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) {
BreakoutTrip = TripCount % Count;
TripMultiple = 0;
} else {
// Figure out what multiple to use.
BreakoutTrip = TripMultiple =
(unsigned)GreatestCommonDivisor64(Count, TripMultiple);
}
if (CompletelyUnroll) {
DEBUG(dbgs() << "COMPLETELY UNROLLING loop %" << Header->getName()
<< " with trip count " << TripCount << "!\n");
} else {
DEBUG(dbgs() << "UNROLLING loop %" << Header->getName()
<< " by " << Count);
if (TripMultiple == 0 || BreakoutTrip != TripMultiple) {
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");
}
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.
ValueToValueMapTy LastValueMap;
std::vector<PHINode*> OrigPHINode;
for (BasicBlock::iterator I = Header->begin(); isa<PHINode>(I); ++I) {
OrigPHINode.push_back(cast<PHINode>(I));
}
std::vector<BasicBlock*> Headers;
std::vector<BasicBlock*> Latches;
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 (LoopBlocksDFS::RPOIterator BB = BlockBegin; BB != BlockEnd; ++BB) {
ValueToValueMapTy VMap;
BasicBlock *New = CloneBasicBlock(*BB, VMap, "." + Twine(It));
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>(VMap[OrigPHINode[i]]);
Value *InVal = NewPHI->getIncomingValueForBlock(LatchBlock);
if (Instruction *InValI = dyn_cast<Instruction>(InVal))
if (It > 1 && L->contains(InValI))
InVal = LastValueMap[InValI];
VMap[OrigPHINode[i]] = InVal;
New->getInstList().erase(NewPHI);
}
// Update our running map of newest clones
LastValueMap[*BB] = New;
for (ValueToValueMapTy::iterator VI = VMap.begin(), VE = VMap.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.
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)
Latches.push_back(New);
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);
}
// 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 (Instruction *InValI = dyn_cast<Instruction>(InVal)) {
if (L->contains(InValI))
InVal = LastValueMap[InVal];
}
assert(Latches.back() == LastValueMap[LatchBlock] && "bad last latch");
PN->addIncoming(InVal, Latches.back());
}
}
// 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;
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) {
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 {
// 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.
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 (Value *V = SimplifyInstruction(Inst))
if (LI->replacementPreservesLCSSAForm(Inst, V)) {
Inst->replaceAllUsesWith(V);
(*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);
return true;
}
/// EliminateMostlyEmptyBlock - Eliminate a basic block that have only phi's and
/// an unconditional branch in it.
void CodeGenPrepare::EliminateMostlyEmptyBlock(BasicBlock *BB) {
BranchInst *BI = cast<BranchInst>(BB->getTerminator());
BasicBlock *DestBB = BI->getSuccessor(0);
DEBUG(dbgs() << "MERGING MOSTLY EMPTY BLOCKS - BEFORE:\n" << *BB << *DestBB);
// If the destination block has a single pred, then this is a trivial edge,
// just collapse it.
if (BasicBlock *SinglePred = DestBB->getSinglePredecessor()) {
if (SinglePred != DestBB) {
// Remember if SinglePred was the entry block of the function. If so, we
// will need to move BB back to the entry position.
bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock();
MergeBasicBlockIntoOnlyPred(DestBB, this);
if (isEntry && BB != &BB->getParent()->getEntryBlock())
BB->moveBefore(&BB->getParent()->getEntryBlock());
DEBUG(dbgs() << "AFTER:\n" << *DestBB << "\n\n\n");
return;
}
}
// Otherwise, we have multiple predecessors of BB. Update the PHIs in DestBB
// to handle the new incoming edges it is about to have.
PHINode *PN;
for (BasicBlock::iterator BBI = DestBB->begin();
(PN = dyn_cast<PHINode>(BBI)); ++BBI) {
// Remove the incoming value for BB, and remember it.
Value *InVal = PN->removeIncomingValue(BB, false);
// Two options: either the InVal is a phi node defined in BB or it is some
// value that dominates BB.
PHINode *InValPhi = dyn_cast<PHINode>(InVal);
if (InValPhi && InValPhi->getParent() == BB) {
// Add all of the input values of the input PHI as inputs of this phi.
for (unsigned i = 0, e = InValPhi->getNumIncomingValues(); i != e; ++i)
PN->addIncoming(InValPhi->getIncomingValue(i),
InValPhi->getIncomingBlock(i));
} else {
// Otherwise, add one instance of the dominating value for each edge that
// we will be adding.
if (PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) {
for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i)
PN->addIncoming(InVal, BBPN->getIncomingBlock(i));
} else {
for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
PN->addIncoming(InVal, *PI);
}
}
}
// The PHIs are now updated, change everything that refers to BB to use
// DestBB and remove BB.
BB->replaceAllUsesWith(DestBB);
if (DT) {
BasicBlock *BBIDom = DT->getNode(BB)->getIDom()->getBlock();
BasicBlock *DestBBIDom = DT->getNode(DestBB)->getIDom()->getBlock();
BasicBlock *NewIDom = DT->findNearestCommonDominator(BBIDom, DestBBIDom);
DT->changeImmediateDominator(DestBB, NewIDom);
DT->eraseNode(BB);
}
if (PFI) {
PFI->replaceAllUses(BB, DestBB);
PFI->removeEdge(ProfileInfo::getEdge(BB, DestBB));
}
BB->eraseFromParent();
++NumBlocksElim;
DEBUG(dbgs() << "AFTER:\n" << *DestBB << "\n\n\n");
}
/// UnswitchNontrivialCondition - We determined that the loop is profitable
/// to unswitch when LIC equal Val. Split it into loop versions and test the
/// condition outside of either loop. Return the loops created as Out1/Out2.
void LoopUnswitch::UnswitchNontrivialCondition(Value *LIC, Constant *Val,
Loop *L) {
Function *F = loopHeader->getParent();
DEBUG(dbgs() << "loop-unswitch: Unswitching loop %"
<< loopHeader->getName() << " [" << L->getBlocks().size()
<< " blocks] in Function " << F->getName()
<< " when '" << *Val << "' == " << *LIC << "\n");
if (ScalarEvolution *SE = getAnalysisIfAvailable<ScalarEvolution>())
SE->forgetLoop(L);
LoopBlocks.clear();
NewBlocks.clear();
// First step, split the preheader and exit blocks, and add these blocks to
// the LoopBlocks list.
BasicBlock *NewPreheader = SplitEdge(loopPreheader, loopHeader, this);
LoopBlocks.push_back(NewPreheader);
// We want the loop to come after the preheader, but before the exit blocks.
LoopBlocks.insert(LoopBlocks.end(), L->block_begin(), L->block_end());
SmallVector<BasicBlock*, 8> ExitBlocks;
L->getUniqueExitBlocks(ExitBlocks);
// Split all of the edges from inside the loop to their exit blocks. Update
// the appropriate Phi nodes as we do so.
SplitExitEdges(L, ExitBlocks);
// The exit blocks may have been changed due to edge splitting, recompute.
ExitBlocks.clear();
L->getUniqueExitBlocks(ExitBlocks);
// Add exit blocks to the loop blocks.
LoopBlocks.insert(LoopBlocks.end(), ExitBlocks.begin(), ExitBlocks.end());
// Next step, clone all of the basic blocks that make up the loop (including
// the loop preheader and exit blocks), keeping track of the mapping between
// the instructions and blocks.
NewBlocks.reserve(LoopBlocks.size());
ValueToValueMapTy VMap;
for (unsigned i = 0, e = LoopBlocks.size(); i != e; ++i) {
BasicBlock *NewBB = CloneBasicBlock(LoopBlocks[i], VMap, ".us", F);
NewBlocks.push_back(NewBB);
VMap[LoopBlocks[i]] = NewBB; // Keep the BB mapping.
LPM->cloneBasicBlockSimpleAnalysis(LoopBlocks[i], NewBB, L);
}
// Splice the newly inserted blocks into the function right before the
// original preheader.
F->getBasicBlockList().splice(NewPreheader, F->getBasicBlockList(),
NewBlocks[0], F->end());
// Now we create the new Loop object for the versioned loop.
Loop *NewLoop = CloneLoop(L, L->getParentLoop(), VMap, LI, LPM);
Loop *ParentLoop = L->getParentLoop();
if (ParentLoop) {
// Make sure to add the cloned preheader and exit blocks to the parent loop
// as well.
ParentLoop->addBasicBlockToLoop(NewBlocks[0], LI->getBase());
}
for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) {
BasicBlock *NewExit = cast<BasicBlock>(VMap[ExitBlocks[i]]);
// The new exit block should be in the same loop as the old one.
if (Loop *ExitBBLoop = LI->getLoopFor(ExitBlocks[i]))
ExitBBLoop->addBasicBlockToLoop(NewExit, LI->getBase());
assert(NewExit->getTerminator()->getNumSuccessors() == 1 &&
"Exit block should have been split to have one successor!");
BasicBlock *ExitSucc = NewExit->getTerminator()->getSuccessor(0);
// If the successor of the exit block had PHI nodes, add an entry for
// NewExit.
PHINode *PN;
for (BasicBlock::iterator I = ExitSucc->begin(); isa<PHINode>(I); ++I) {
PN = cast<PHINode>(I);
Value *V = PN->getIncomingValueForBlock(ExitBlocks[i]);
ValueToValueMapTy::iterator It = VMap.find(V);
if (It != VMap.end()) V = It->second;
PN->addIncoming(V, NewExit);
}
}
// Rewrite the code to refer to itself.
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, VMap,RF_NoModuleLevelChanges|RF_IgnoreMissingEntries);
// Rewrite the original preheader to select between versions of the loop.
BranchInst *OldBR = cast<BranchInst>(loopPreheader->getTerminator());
assert(OldBR->isUnconditional() && OldBR->getSuccessor(0) == LoopBlocks[0] &&
"Preheader splitting did not work correctly!");
// Emit the new branch that selects between the two versions of this loop.
EmitPreheaderBranchOnCondition(LIC, Val, NewBlocks[0], LoopBlocks[0], OldBR);
LPM->deleteSimpleAnalysisValue(OldBR, L);
OldBR->eraseFromParent();
LoopProcessWorklist.push_back(NewLoop);
redoLoop = true;
// Keep a WeakVH holding onto LIC. If the first call to RewriteLoopBody
// deletes the instruction (for example by simplifying a PHI that feeds into
// the condition that we're unswitching on), we don't rewrite the second
// iteration.
WeakVH LICHandle(LIC);
// Now we rewrite the original code to know that the condition is true and the
// new code to know that the condition is false.
RewriteLoopBodyWithConditionConstant(L, LIC, Val, false);
// It's possible that simplifying one loop could cause the other to be
// changed to another value or a constant. If its a constant, don't simplify
// it.
if (!LoopProcessWorklist.empty() && LoopProcessWorklist.back() == NewLoop &&
LICHandle && !isa<Constant>(LICHandle))
RewriteLoopBodyWithConditionConstant(NewLoop, LICHandle, Val, true);
}
/// SimplifyStoreAtEndOfBlock - Turn things like:
/// if () { *P = v1; } else { *P = v2 }
/// into a phi node with a store in the successor.
///
/// Simplify things like:
/// *P = v1; if () { *P = v2; }
/// into a phi node with a store in the successor.
///
bool InstCombiner::SimplifyStoreAtEndOfBlock(StoreInst &SI) {
BasicBlock *StoreBB = SI.getParent();
// Check to see if the successor block has exactly two incoming edges. If
// so, see if the other predecessor contains a store to the same location.
// if so, insert a PHI node (if needed) and move the stores down.
BasicBlock *DestBB = StoreBB->getTerminator()->getSuccessor(0);
// Determine whether Dest has exactly two predecessors and, if so, compute
// the other predecessor.
pred_iterator PI = pred_begin(DestBB);
BasicBlock *P = *PI;
BasicBlock *OtherBB = 0;
if (P != StoreBB)
OtherBB = P;
if (++PI == pred_end(DestBB))
return false;
P = *PI;
if (P != StoreBB) {
if (OtherBB)
return false;
OtherBB = P;
}
if (++PI != pred_end(DestBB))
return false;
// Bail out if all the relevant blocks aren't distinct (this can happen,
// for example, if SI is in an infinite loop)
if (StoreBB == DestBB || OtherBB == DestBB)
return false;
// Verify that the other block ends in a branch and is not otherwise empty.
BasicBlock::iterator BBI = OtherBB->getTerminator();
BranchInst *OtherBr = dyn_cast<BranchInst>(BBI);
if (!OtherBr || BBI == OtherBB->begin())
return false;
// If the other block ends in an unconditional branch, check for the 'if then
// else' case. there is an instruction before the branch.
StoreInst *OtherStore = 0;
if (OtherBr->isUnconditional()) {
--BBI;
// Skip over debugging info.
while (isa<DbgInfoIntrinsic>(BBI) ||
(isa<BitCastInst>(BBI) && BBI->getType()->isPointerTy())) {
if (BBI==OtherBB->begin())
return false;
--BBI;
}
// If this isn't a store, isn't a store to the same location, or is not the
// right kind of store, bail out.
OtherStore = dyn_cast<StoreInst>(BBI);
if (!OtherStore || OtherStore->getOperand(1) != SI.getOperand(1) ||
!SI.isSameOperationAs(OtherStore))
return false;
} else {
// Otherwise, the other block ended with a conditional branch. If one of the
// destinations is StoreBB, then we have the if/then case.
if (OtherBr->getSuccessor(0) != StoreBB &&
OtherBr->getSuccessor(1) != StoreBB)
return false;
// Okay, we know that OtherBr now goes to Dest and StoreBB, so this is an
// if/then triangle. See if there is a store to the same ptr as SI that
// lives in OtherBB.
for (;; --BBI) {
// Check to see if we find the matching store.
if ((OtherStore = dyn_cast<StoreInst>(BBI))) {
if (OtherStore->getOperand(1) != SI.getOperand(1) ||
!SI.isSameOperationAs(OtherStore))
return false;
break;
}
// If we find something that may be using or overwriting the stored
// value, or if we run out of instructions, we can't do the xform.
if (BBI->mayReadFromMemory() || BBI->mayWriteToMemory() ||
BBI == OtherBB->begin())
return false;
}
// In order to eliminate the store in OtherBr, we have to
// make sure nothing reads or overwrites the stored value in
// StoreBB.
for (BasicBlock::iterator I = StoreBB->begin(); &*I != &SI; ++I) {
// FIXME: This should really be AA driven.
if (I->mayReadFromMemory() || I->mayWriteToMemory())
return false;
}
}
// Insert a PHI node now if we need it.
Value *MergedVal = OtherStore->getOperand(0);
if (MergedVal != SI.getOperand(0)) {
PHINode *PN = PHINode::Create(MergedVal->getType(), 2, "storemerge");
PN->addIncoming(SI.getOperand(0), SI.getParent());
PN->addIncoming(OtherStore->getOperand(0), OtherBB);
MergedVal = InsertNewInstBefore(PN, DestBB->front());
}
// Advance to a place where it is safe to insert the new store and
// insert it.
BBI = DestBB->getFirstInsertionPt();
StoreInst *NewSI = new StoreInst(MergedVal, SI.getOperand(1),
SI.isVolatile(),
SI.getAlignment(),
SI.getOrdering(),
SI.getSynchScope());
InsertNewInstBefore(NewSI, *BBI);
NewSI->setDebugLoc(OtherStore->getDebugLoc());
// If the two stores had the same TBAA tag, preserve it.
if (MDNode *TBAATag = SI.getMetadata(LLVMContext::MD_tbaa))
if ((TBAATag = MDNode::getMostGenericTBAA(TBAATag,
OtherStore->getMetadata(LLVMContext::MD_tbaa))))
NewSI->setMetadata(LLVMContext::MD_tbaa, TBAATag);
// Nuke the old stores.
EraseInstFromFunction(SI);
EraseInstFromFunction(*OtherStore);
return true;
}
/// Insert code in the prolog code when unrolling a loop with a
/// run-time trip-count.
///
/// This method assumes that the loop unroll factor is total number
/// of loop bodes in the loop after unrolling. (Some folks refer
/// to the unroll factor as the number of *extra* copies added).
/// We assume also that the loop unroll factor is a power-of-two. So, after
/// unrolling the loop, the number of loop bodies executed is 2,
/// 4, 8, etc. Note - LLVM converts the if-then-sequence to a switch
/// instruction in SimplifyCFG.cpp. Then, the backend decides how code for
/// the switch instruction is generated.
///
/// extraiters = tripcount % loopfactor
/// if (extraiters == 0) jump Loop:
/// else jump Prol
/// Prol: LoopBody;
/// extraiters -= 1 // Omitted if unroll factor is 2.
/// if (extraiters != 0) jump Prol: // Omitted if unroll factor is 2.
/// if (tripcount < loopfactor) jump End
/// Loop:
/// ...
/// End:
///
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())
return false;
// Make sure the loop is in canonical form, and there is a single
// exit block only.
if (!L->isLoopSimplifyForm() || !L->getUniqueExitBlock())
return false;
// Use Scalar Evolution to compute the trip count. This allows more
// loops to be unrolled than relying on induction var simplification
if (!LPM)
return false;
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 *BECountSC = SE.getBackedgeTakenCount(L);
if (isa<SCEVCouldNotCompute>(BECountSC) ||
!BECountSC->getType()->isIntegerTy())
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(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 (!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);
// Grab analyses that we preserve.
auto *DTWP = LPM->getAnalysisIfAvailable<DominatorTreeWrapperPass>();
auto *DT = DTWP ? &DTWP->getDomTree() : nullptr;
BasicBlock *PH = L->getLoopPreheader();
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, 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)
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");
// 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
B.CreateCondBr(BranchVal, PEnd, PEnd);
assert(PreHeaderBR->isUnconditional() &&
PreHeaderBR->getSuccessor(0) == PEnd &&
"CFG edges in Preheader are not correct");
PreHeaderBR->eraseFromParent();
Function *F = Header->getParent();
// Get an ordered list of blocks in the loop to help with the ordering of the
// cloned blocks in the prolog code
LoopBlocksDFS LoopBlocks(L);
LoopBlocks.perform(LI);
//
// For each extra loop iteration, create a copy of the loop's basic blocks
// and generate a condition that branches to the copy depending on the
// number of 'left over' iterations.
//
std::vector<BasicBlock *> NewBlocks;
ValueToValueMapTy VMap;
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
// iterations. This function adds the appropriate CFG connections.
CloneLoopBlocks(L, ModVal, UnrollPrologue, PH, PEnd, NewBlocks, LoopBlocks,
VMap, LI);
// Insert the cloned blocks into function just before the original loop
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 (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, 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, BECount, Count, LastLoopBB, PEnd, PH, NewPH, VMap, DT, LI,
LPM->getAsPass());
NumRuntimeUnrolled++;
return true;
}
/// This works like CloneAndPruneFunctionInto, except that it does not clone the
/// entire function. Instead it starts at an instruction provided by the caller
/// and copies (and prunes) only the code reachable from that instruction.
void llvm::CloneAndPruneIntoFromInst(Function *NewFunc, const Function *OldFunc,
const Instruction *StartingInst,
ValueToValueMapTy &VMap,
bool ModuleLevelChanges,
SmallVectorImpl<ReturnInst *> &Returns,
const char *NameSuffix,
ClonedCodeInfo *CodeInfo,
CloningDirector *Director) {
assert(NameSuffix && "NameSuffix cannot be null!");
ValueMapTypeRemapper *TypeMapper = nullptr;
ValueMaterializer *Materializer = nullptr;
if (Director) {
TypeMapper = Director->getTypeRemapper();
Materializer = Director->getValueMaterializer();
}
#ifndef NDEBUG
// If the cloning starts at the beginning of the function, verify that
// the function arguments are mapped.
if (!StartingInst)
for (const Argument &II : OldFunc->args())
assert(VMap.count(&II) && "No mapping from source argument specified!");
#endif
PruningFunctionCloner PFC(NewFunc, OldFunc, VMap, ModuleLevelChanges,
NameSuffix, CodeInfo, Director);
const BasicBlock *StartingBB;
if (StartingInst)
StartingBB = StartingInst->getParent();
else {
StartingBB = &OldFunc->getEntryBlock();
StartingInst = &StartingBB->front();
}
// Clone the entry block, and anything recursively reachable from it.
std::vector<const BasicBlock*> CloneWorklist;
PFC.CloneBlock(StartingBB, StartingInst->getIterator(), CloneWorklist);
while (!CloneWorklist.empty()) {
const BasicBlock *BB = CloneWorklist.back();
CloneWorklist.pop_back();
PFC.CloneBlock(BB, BB->begin(), CloneWorklist);
}
// Loop over all of the basic blocks in the old function. If the block was
// reachable, we have cloned it and the old block is now in the value map:
// insert it into the new function in the right order. If not, ignore it.
//
// Defer PHI resolution until rest of function is resolved.
SmallVector<const PHINode*, 16> PHIToResolve;
for (const BasicBlock &BI : *OldFunc) {
Value *V = VMap[&BI];
BasicBlock *NewBB = cast_or_null<BasicBlock>(V);
if (!NewBB) continue; // Dead block.
// Add the new block to the new function.
NewFunc->getBasicBlockList().push_back(NewBB);
// Handle PHI nodes specially, as we have to remove references to dead
// blocks.
for (BasicBlock::const_iterator I = BI.begin(), E = BI.end(); I != E; ++I) {
// PHI nodes may have been remapped to non-PHI nodes by the caller or
// during the cloning process.
if (const PHINode *PN = dyn_cast<PHINode>(I)) {
if (isa<PHINode>(VMap[PN]))
PHIToResolve.push_back(PN);
else
break;
} else {
break;
}
}
// Finally, remap the terminator instructions, as those can't be remapped
// until all BBs are mapped.
RemapInstruction(NewBB->getTerminator(), VMap,
ModuleLevelChanges ? RF_None : RF_NoModuleLevelChanges,
TypeMapper, Materializer);
}
// Defer PHI resolution until rest of function is resolved, PHI resolution
// requires the CFG to be up-to-date.
for (unsigned phino = 0, e = PHIToResolve.size(); phino != e; ) {
const PHINode *OPN = PHIToResolve[phino];
unsigned NumPreds = OPN->getNumIncomingValues();
const BasicBlock *OldBB = OPN->getParent();
BasicBlock *NewBB = cast<BasicBlock>(VMap[OldBB]);
// Map operands for blocks that are live and remove operands for blocks
// that are dead.
for (; phino != PHIToResolve.size() &&
PHIToResolve[phino]->getParent() == OldBB; ++phino) {
OPN = PHIToResolve[phino];
PHINode *PN = cast<PHINode>(VMap[OPN]);
for (unsigned pred = 0, e = NumPreds; pred != e; ++pred) {
Value *V = VMap[PN->getIncomingBlock(pred)];
if (BasicBlock *MappedBlock = cast_or_null<BasicBlock>(V)) {
Value *InVal = MapValue(PN->getIncomingValue(pred),
VMap,
ModuleLevelChanges ? RF_None : RF_NoModuleLevelChanges);
assert(InVal && "Unknown input value?");
PN->setIncomingValue(pred, InVal);
PN->setIncomingBlock(pred, MappedBlock);
} else {
PN->removeIncomingValue(pred, false);
--pred, --e; // Revisit the next entry.
}
}
}
// The loop above has removed PHI entries for those blocks that are dead
// and has updated others. However, if a block is live (i.e. copied over)
// but its terminator has been changed to not go to this block, then our
// phi nodes will have invalid entries. Update the PHI nodes in this
// case.
PHINode *PN = cast<PHINode>(NewBB->begin());
NumPreds = std::distance(pred_begin(NewBB), pred_end(NewBB));
if (NumPreds != PN->getNumIncomingValues()) {
assert(NumPreds < PN->getNumIncomingValues());
// Count how many times each predecessor comes to this block.
std::map<BasicBlock*, unsigned> PredCount;
for (pred_iterator PI = pred_begin(NewBB), E = pred_end(NewBB);
PI != E; ++PI)
--PredCount[*PI];
// Figure out how many entries to remove from each PHI.
for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
++PredCount[PN->getIncomingBlock(i)];
// At this point, the excess predecessor entries are positive in the
// map. Loop over all of the PHIs and remove excess predecessor
// entries.
BasicBlock::iterator I = NewBB->begin();
for (; (PN = dyn_cast<PHINode>(I)); ++I) {
for (std::map<BasicBlock*, unsigned>::iterator PCI =PredCount.begin(),
E = PredCount.end(); PCI != E; ++PCI) {
BasicBlock *Pred = PCI->first;
for (unsigned NumToRemove = PCI->second; NumToRemove; --NumToRemove)
PN->removeIncomingValue(Pred, false);
}
}
}
// If the loops above have made these phi nodes have 0 or 1 operand,
// replace them with undef or the input value. We must do this for
// correctness, because 0-operand phis are not valid.
PN = cast<PHINode>(NewBB->begin());
if (PN->getNumIncomingValues() == 0) {
BasicBlock::iterator I = NewBB->begin();
BasicBlock::const_iterator OldI = OldBB->begin();
while ((PN = dyn_cast<PHINode>(I++))) {
Value *NV = UndefValue::get(PN->getType());
PN->replaceAllUsesWith(NV);
assert(VMap[&*OldI] == PN && "VMap mismatch");
VMap[&*OldI] = NV;
PN->eraseFromParent();
++OldI;
}
}
}
// Make a second pass over the PHINodes now that all of them have been
// remapped into the new function, simplifying the PHINode and performing any
// recursive simplifications exposed. This will transparently update the
// WeakVH in the VMap. Notably, we rely on that so that if we coalesce
// two PHINodes, the iteration over the old PHIs remains valid, and the
// mapping will just map us to the new node (which may not even be a PHI
// node).
for (unsigned Idx = 0, Size = PHIToResolve.size(); Idx != Size; ++Idx)
if (PHINode *PN = dyn_cast<PHINode>(VMap[PHIToResolve[Idx]]))
recursivelySimplifyInstruction(PN);
// Now that the inlined function body has been fully constructed, go through
// and zap unconditional fall-through branches. This happens all the time when
// specializing code: code specialization turns conditional branches into
// uncond branches, and this code folds them.
Function::iterator Begin = cast<BasicBlock>(VMap[StartingBB])->getIterator();
Function::iterator I = Begin;
while (I != NewFunc->end()) {
// Check if this block has become dead during inlining or other
// simplifications. Note that the first block will appear dead, as it has
// not yet been wired up properly.
if (I != Begin && (pred_begin(&*I) == pred_end(&*I) ||
I->getSinglePredecessor() == &*I)) {
BasicBlock *DeadBB = &*I++;
DeleteDeadBlock(DeadBB);
continue;
}
// We need to simplify conditional branches and switches with a constant
// operand. We try to prune these out when cloning, but if the
// simplification required looking through PHI nodes, those are only
// available after forming the full basic block. That may leave some here,
// and we still want to prune the dead code as early as possible.
ConstantFoldTerminator(&*I);
BranchInst *BI = dyn_cast<BranchInst>(I->getTerminator());
if (!BI || BI->isConditional()) { ++I; continue; }
BasicBlock *Dest = BI->getSuccessor(0);
if (!Dest->getSinglePredecessor()) {
++I; continue;
}
// We shouldn't be able to get single-entry PHI nodes here, as instsimplify
// above should have zapped all of them..
assert(!isa<PHINode>(Dest->begin()));
// We know all single-entry PHI nodes in the inlined function have been
// removed, so we just need to splice the blocks.
BI->eraseFromParent();
// Make all PHI nodes that referred to Dest now refer to I as their source.
Dest->replaceAllUsesWith(&*I);
// Move all the instructions in the succ to the pred.
I->getInstList().splice(I->end(), Dest->getInstList());
// Remove the dest block.
Dest->eraseFromParent();
// Do not increment I, iteratively merge all things this block branches to.
}
// Make a final pass over the basic blocks from the old function to gather
// any return instructions which survived folding. We have to do this here
// because we can iteratively remove and merge returns above.
for (Function::iterator I = cast<BasicBlock>(VMap[StartingBB])->getIterator(),
E = NewFunc->end();
I != E; ++I)
if (ReturnInst *RI = dyn_cast<ReturnInst>(I->getTerminator()))
Returns.push_back(RI);
}
void TempScopInfo::buildCondition(BasicBlock *BB, Region &R) {
BasicBlock *RegionEntry = R.getEntry();
BBCond Cond;
DomTreeNode *BBNode = DT->getNode(BB), *EntryNode = DT->getNode(RegionEntry);
assert(BBNode && EntryNode && "Get null node while building condition!");
// Walk up the dominance tree until reaching the entry node. Collect all
// branching blocks on the path to BB except if BB postdominates the block
// containing the condition.
SmallVector<BasicBlock *, 4> DominatorBrBlocks;
while (BBNode != EntryNode) {
BasicBlock *CurBB = BBNode->getBlock();
BBNode = BBNode->getIDom();
assert(BBNode && "BBNode should not reach the root node!");
if (PDT->dominates(CurBB, BBNode->getBlock()))
continue;
BranchInst *Br = dyn_cast<BranchInst>(BBNode->getBlock()->getTerminator());
assert(Br && "A Valid Scop should only contain branch instruction");
if (Br->isUnconditional())
continue;
DominatorBrBlocks.push_back(BBNode->getBlock());
}
RegionInfo *RI = R.getRegionInfo();
// Iterate in reverse order over the dominating blocks. Until a non-affine
// branch was encountered add all conditions collected. If a non-affine branch
// was encountered, stop as we overapproximate from here on anyway.
for (auto BIt = DominatorBrBlocks.rbegin(), BEnd = DominatorBrBlocks.rend();
BIt != BEnd; BIt++) {
BasicBlock *BBNode = *BIt;
BranchInst *Br = dyn_cast<BranchInst>(BBNode->getTerminator());
assert(Br && "A Valid Scop should only contain branch instruction");
assert(Br->isConditional() && "Assumed a conditional branch");
if (SD->isNonAffineSubRegion(RI->getRegionFor(BBNode), &R))
break;
BasicBlock *TrueBB = Br->getSuccessor(0), *FalseBB = Br->getSuccessor(1);
// Is BB on the ELSE side of the branch?
bool inverted = DT->dominates(FalseBB, BB);
// If both TrueBB and FalseBB dominate BB, one of them must be the target of
// a back-edge, i.e. a loop header.
if (inverted && DT->dominates(TrueBB, BB)) {
assert(
(DT->dominates(TrueBB, FalseBB) || DT->dominates(FalseBB, TrueBB)) &&
"One of the successors should be the loop header and dominate the"
"other!");
// It is not an invert if the FalseBB is the header.
if (DT->dominates(FalseBB, TrueBB))
inverted = false;
}
Comparison *Cmp;
buildAffineCondition(*(Br->getCondition()), inverted, &Cmp);
Cond.push_back(*Cmp);
}
if (!Cond.empty())
BBConds[BB] = Cond;
}
bool LoopUnroll::visitLoop(Loop *L) {
bool Changed = false;
// 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]);
// We only handle single basic block loops right now.
if (L->getBlocks().size() != 1)
return Changed;
BasicBlock *BB = L->getHeader();
BranchInst *BI = dyn_cast<BranchInst>(BB->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!
unsigned TripCount = TripCountC->getRawValue();
if (TripCount != TripCountC->getRawValue() || TripCount == 0)
return Changed; // More than 2^32 iterations???
unsigned LoopSize = ApproximateLoopSize(L);
DEBUG(std::cerr << "Loop Unroll: F[" << BB->getParent()->getName()
<< "] Loop %" << BB->getName() << " Loop Size = " << LoopSize
<< " Trip Count = " << TripCount << " - ");
uint64_t Size = (uint64_t)LoopSize*(uint64_t)TripCount;
if (Size > UnrollThreshold) {
DEBUG(std::cerr << "TOO LARGE: " << Size << ">" << UnrollThreshold << "\n");
return Changed;
}
DEBUG(std::cerr << "UNROLLING!\n");
BasicBlock *LoopExit = BI->getSuccessor(L->contains(BI->getSuccessor(0)));
// Create a new basic block to temporarily hold all of the cloned code.
BasicBlock *NewBlock = new BasicBlock();
// 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;
std::vector<PHINode*> OrigPHINode;
for (BasicBlock::iterator I = BB->begin();
PHINode *PN = dyn_cast<PHINode>(I); ++I) {
OrigPHINode.push_back(PN);
if (Instruction *I =dyn_cast<Instruction>(PN->getIncomingValueForBlock(BB)))
if (I->getParent() == BB)
LastValueMap[I] = I;
}
// Remove the exit branch from the loop
BB->getInstList().erase(BI);
assert(TripCount != 0 && "Trip count of 0 is impossible!");
for (unsigned It = 1; It != TripCount; ++It) {
char SuffixBuffer[100];
sprintf(SuffixBuffer, ".%d", It);
std::map<const Value*, Value*> ValueMap;
BasicBlock *New = CloneBasicBlock(BB, ValueMap, SuffixBuffer);
// Loop over all of the PHI nodes in the block, changing them to use the
// incoming values from the previous block.
for (unsigned i = 0, e = OrigPHINode.size(); i != e; ++i) {
PHINode *NewPHI = cast<PHINode>(ValueMap[OrigPHINode[i]]);
Value *InVal = NewPHI->getIncomingValueForBlock(BB);
if (Instruction *InValI = dyn_cast<Instruction>(InVal))
if (InValI->getParent() == BB)
InVal = LastValueMap[InValI];
ValueMap[OrigPHINode[i]] = InVal;
New->getInstList().erase(NewPHI);
}
for (BasicBlock::iterator I = New->begin(), E = New->end(); I != E; ++I)
RemapInstruction(I, ValueMap);
// Now that all of the instructions are remapped, splice them into the end
// of the NewBlock.
NewBlock->getInstList().splice(NewBlock->end(), New->getInstList());
delete New;
// LastValue map now contains values from this iteration.
std::swap(LastValueMap, ValueMap);
}
// If there was more than one iteration, replace any uses of values computed
// in the loop with values computed during the last iteration of the loop.
if (TripCount != 1) {
std::set<User*> Users;
for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
Users.insert(I->use_begin(), I->use_end());
// We don't want to reprocess entries with PHI nodes in them. For this
// reason, we look at each operand of each user exactly once, performing the
// stubstitution exactly once.
for (std::set<User*>::iterator UI = Users.begin(), E = Users.end(); UI != E;
++UI) {
Instruction *I = cast<Instruction>(*UI);
if (I->getParent() != BB && I->getParent() != NewBlock)
RemapInstruction(I, LastValueMap);
}
}
// Now that we cloned the block as many times as we needed, stitch the new
// code into the original block and delete the temporary block.
BB->getInstList().splice(BB->end(), NewBlock->getInstList());
delete NewBlock;
// 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));
BB->getInstList().erase(PN);
}
// Finally, add an unconditional branch to the block to continue into the exit
// block.
new BranchInst(LoopExit, BB);
// 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.
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);
}
}
// Update the loop information for this loop.
Loop *Parent = L->getParentLoop();
// Move all of the basic blocks in the loop into the parent loop.
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));
// FIXME: Should update dominator analyses
// Now that everything is up-to-date that will be, we fold the loop block into
// the preheader and exit block, updating our analyses as we go.
LoopExit->getInstList().splice(LoopExit->begin(), BB->getInstList(),
BB->getInstList().begin(),
prior(BB->getInstList().end()));
LoopExit->getInstList().splice(LoopExit->begin(), Preheader->getInstList(),
Preheader->getInstList().begin(),
prior(Preheader->getInstList().end()));
// Make all other blocks in the program branch to LoopExit now instead of
// Preheader.
Preheader->replaceAllUsesWith(LoopExit);
// Remove BB and LoopExit from our analyses.
LI->removeBlock(Preheader);
LI->removeBlock(BB);
// If the preheader was the entry block of this function, move the exit block
// to be the new entry of the loop.
Function *F = LoopExit->getParent();
if (Preheader == &F->front())
F->getBasicBlockList().splice(F->begin(), F->getBasicBlockList(), LoopExit);
// Actually delete the blocks now.
F->getBasicBlockList().erase(Preheader);
F->getBasicBlockList().erase(BB);
++NumUnrolled;
return true;
}
/// Insert code in the prolog code when unrolling a loop with a
/// run-time trip-count.
///
/// This method assumes that the loop unroll factor is total number
/// of loop bodes in the loop after unrolling. (Some folks refer
/// to the unroll factor as the number of *extra* copies added).
/// We assume also that the loop unroll factor is a power-of-two. So, after
/// unrolling the loop, the number of loop bodies executed is 2,
/// 4, 8, etc. Note - LLVM converts the if-then-sequence to a switch
/// instruction in SimplifyCFG.cpp. Then, the backend decides how code for
/// the switch instruction is generated.
///
/// extraiters = tripcount % loopfactor
/// if (extraiters == 0) jump Loop:
/// if (extraiters == loopfactor) jump L1
/// if (extraiters == loopfactor-1) jump L2
/// ...
/// L1: LoopBody;
/// L2: LoopBody;
/// ...
/// if tripcount < loopfactor jump End
/// Loop:
/// ...
/// End:
///
bool llvm::UnrollRuntimeLoopProlog(Loop *L, unsigned Count, LoopInfo *LI,
LPPassManager *LPM) {
// for now, only unroll loops that contain a single exit
if (!L->getExitingBlock())
return false;
// Make sure the loop is in canonical form, and there is a single
// exit block only.
if (!L->isLoopSimplifyForm() || L->getUniqueExitBlock() == 0)
return false;
// Use Scalar Evolution to compute the trip count. This allows more
// loops to be unrolled than relying on induction var simplification
if (!LPM)
return false;
ScalarEvolution *SE = LPM->getAnalysisIfAvailable<ScalarEvolution>();
if (SE == 0)
return false;
// 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())
return false;
// Add 1 since the backedge count doesn't include the first loop iteration
const SCEV *TripCountSC =
SE->getAddExpr(BECount, SE->getConstant(BECount->getType(), 1));
if (isa<SCEVCouldNotCompute>(TripCountSC))
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)
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);
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());
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);
Type *CountTy = TripCount->getType();
BinaryOperator *ModVal =
BinaryOperator::CreateURem(TripCount,
ConstantInt::get(CountTy, Count),
"xtraiter");
ModVal->insertBefore(PreHeaderBR);
// Check if for no extra iterations, then jump to unrolled loop
Value *BranchVal = new ICmpInst(PreHeaderBR,
ICmpInst::ICMP_NE, ModVal,
ConstantInt::get(CountTy, 0), "lcmp");
// Branch to either the extra iterations or the unrolled loop
// We will fix up the true branch label when adding loop body copies
BranchInst::Create(PEnd, PEnd, BranchVal, PreHeaderBR);
assert(PreHeaderBR->isUnconditional() &&
PreHeaderBR->getSuccessor(0) == PEnd &&
"CFG edges in Preheader are not correct");
PreHeaderBR->eraseFromParent();
ValueToValueMapTy LVMap;
Function *F = Header->getParent();
// These variables are used to update the CFG links in each iteration
BasicBlock *CompareBB = 0;
BasicBlock *LastLoopBB = PH;
// Get an ordered list of blocks in the loop to help with the ordering of the
// cloned blocks in the prolog code
LoopBlocksDFS LoopBlocks(L);
LoopBlocks.perform(LI);
//
// For each extra loop iteration, create a copy of the loop's basic blocks
// and generate a condition that branches to the copy depending on the
// number of 'left over' iterations.
//
for (unsigned leftOverIters = Count-1; leftOverIters > 0; --leftOverIters) {
std::vector<BasicBlock*> NewBlocks;
ValueToValueMapTy VMap;
// Clone all the basic blocks in the loop, but we don't clone the loop
// This function adds the appropriate CFG connections.
CloneLoopBlocks(L, (leftOverIters == Count-1), LastLoopBB, PEnd, NewBlocks,
LoopBlocks, VMap, LVMap, LI);
LastLoopBB = cast<BasicBlock>(VMap[Latch]);
// Insert the cloned blocks into function just before the original loop
F->getBasicBlockList().splice(PEnd, F->getBasicBlockList(),
NewBlocks[0], F->end());
// Generate the code for the comparison which determines if the loop
// prolog code needs to be executed.
if (leftOverIters == Count-1) {
// There is no compare block for the fall-thru case when for the last
// left over iteration
CompareBB = NewBlocks[0];
} else {
// Create a new block for the comparison
BasicBlock *NewBB = BasicBlock::Create(CompareBB->getContext(), "unr.cmp",
F, CompareBB);
if (Loop *ParentLoop = L->getParentLoop()) {
// Add the new block to the parent loop, if needed
ParentLoop->addBasicBlockToLoop(NewBB, LI->getBase());
}
// The comparison w/ the extra iteration value and branch
Value *BranchVal = new ICmpInst(*NewBB, ICmpInst::ICMP_EQ, ModVal,
ConstantInt::get(CountTy, leftOverIters),
"un.tmp");
// Branch to either the extra iterations or the unrolled loop
BranchInst::Create(NewBlocks[0], CompareBB,
BranchVal, NewBB);
CompareBB = NewBB;
PH->getTerminator()->setSuccessor(0, NewBB);
VMap[NewPH] = CompareBB;
}
// Rewrite the cloned instruction operands to use the values
// created when the clone is created.
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, VMap,
RF_NoModuleLevelChanges|RF_IgnoreMissingEntries);
}
}
}
// Connect the prolog code to the original loop and update the
// PHI functions.
ConnectProlog(L, TripCount, Count, LastLoopBB, PEnd, PH, NewPH, LVMap,
LPM->getAsPass());
NumRuntimeUnrolled++;
return true;
}
/// ProcessLoop - Walk the loop structure in depth first order, ensuring that
/// all loops have preheaders.
///
bool LoopSimplify::ProcessLoop(Loop *L, LPPassManager &LPM) {
bool Changed = false;
ReprocessLoop:
// Check to see that no blocks (other than the header) in this loop have
// predecessors that are not in the loop. This is not valid for natural
// loops, but can occur if the blocks are unreachable. Since they are
// unreachable we can just shamelessly delete those CFG edges!
for (Loop::block_iterator BB = L->block_begin(), E = L->block_end();
BB != E; ++BB) {
if (*BB == L->getHeader()) continue;
SmallPtrSet<BasicBlock*, 4> BadPreds;
for (pred_iterator PI = pred_begin(*BB),
PE = pred_end(*BB); PI != PE; ++PI) {
BasicBlock *P = *PI;
if (!L->contains(P))
BadPreds.insert(P);
}
// Delete each unique out-of-loop (and thus dead) predecessor.
for (SmallPtrSet<BasicBlock*, 4>::iterator I = BadPreds.begin(),
E = BadPreds.end(); I != E; ++I) {
DEBUG(dbgs() << "LoopSimplify: Deleting edge from dead predecessor "
<< (*I)->getName() << "\n");
// Inform each successor of each dead pred.
for (succ_iterator SI = succ_begin(*I), SE = succ_end(*I); SI != SE; ++SI)
(*SI)->removePredecessor(*I);
// Zap the dead pred's terminator and replace it with unreachable.
TerminatorInst *TI = (*I)->getTerminator();
TI->replaceAllUsesWith(UndefValue::get(TI->getType()));
(*I)->getTerminator()->eraseFromParent();
new UnreachableInst((*I)->getContext(), *I);
Changed = true;
}
}
// If there are exiting blocks with branches on undef, resolve the undef in
// the direction which will exit the loop. This will help simplify loop
// trip count computations.
SmallVector<BasicBlock*, 8> ExitingBlocks;
L->getExitingBlocks(ExitingBlocks);
for (SmallVectorImpl<BasicBlock *>::iterator I = ExitingBlocks.begin(),
E = ExitingBlocks.end(); I != E; ++I)
if (BranchInst *BI = dyn_cast<BranchInst>((*I)->getTerminator()))
if (BI->isConditional()) {
if (UndefValue *Cond = dyn_cast<UndefValue>(BI->getCondition())) {
DEBUG(dbgs() << "LoopSimplify: Resolving \"br i1 undef\" to exit in "
<< (*I)->getName() << "\n");
BI->setCondition(ConstantInt::get(Cond->getType(),
!L->contains(BI->getSuccessor(0))));
// This may make the loop analyzable, force SCEV recomputation.
if (SE)
SE->forgetLoop(L);
Changed = true;
}
}
// Does the loop already have a preheader? If so, don't insert one.
BasicBlock *Preheader = L->getLoopPreheader();
if (!Preheader) {
Preheader = InsertPreheaderForLoop(L, this);
if (Preheader) {
++NumInserted;
Changed = true;
}
}
// Next, check to make sure that all exit nodes of the loop only have
// predecessors that are inside of the loop. This check guarantees that the
// loop preheader/header will dominate the exit blocks. If the exit block has
// predecessors from outside of the loop, split the edge now.
SmallVector<BasicBlock*, 8> ExitBlocks;
L->getExitBlocks(ExitBlocks);
SmallSetVector<BasicBlock *, 8> ExitBlockSet(ExitBlocks.begin(),
ExitBlocks.end());
for (SmallSetVector<BasicBlock *, 8>::iterator I = ExitBlockSet.begin(),
E = ExitBlockSet.end(); I != E; ++I) {
BasicBlock *ExitBlock = *I;
for (pred_iterator PI = pred_begin(ExitBlock), PE = pred_end(ExitBlock);
PI != PE; ++PI)
// Must be exactly this loop: no subloops, parent loops, or non-loop preds
// allowed.
if (!L->contains(*PI)) {
if (RewriteLoopExitBlock(L, ExitBlock)) {
++NumInserted;
Changed = true;
}
break;
}
}
// If the header has more than two predecessors at this point (from the
// preheader and from multiple backedges), we must adjust the loop.
BasicBlock *LoopLatch = L->getLoopLatch();
if (!LoopLatch) {
// If this is really a nested loop, rip it out into a child loop. Don't do
// this for loops with a giant number of backedges, just factor them into a
// common backedge instead.
if (L->getNumBackEdges() < 8) {
if (SeparateNestedLoop(L, LPM, Preheader)) {
++NumNested;
// This is a big restructuring change, reprocess the whole loop.
Changed = true;
// GCC doesn't tail recursion eliminate this.
goto ReprocessLoop;
}
}
// If we either couldn't, or didn't want to, identify nesting of the loops,
// insert a new block that all backedges target, then make it jump to the
// loop header.
LoopLatch = InsertUniqueBackedgeBlock(L, Preheader);
if (LoopLatch) {
++NumInserted;
Changed = true;
}
}
// Scan over the PHI nodes in the loop header. Since they now have only two
// incoming values (the loop is canonicalized), we may have simplified the PHI
// down to 'X = phi [X, Y]', which should be replaced with 'Y'.
PHINode *PN;
for (BasicBlock::iterator I = L->getHeader()->begin();
(PN = dyn_cast<PHINode>(I++)); )
if (Value *V = SimplifyInstruction(PN, 0, 0, DT)) {
if (AA) AA->deleteValue(PN);
if (SE) SE->forgetValue(PN);
PN->replaceAllUsesWith(V);
PN->eraseFromParent();
}
// If this loop has multiple exits and the exits all go to the same
// block, attempt to merge the exits. This helps several passes, such
// as LoopRotation, which do not support loops with multiple exits.
// SimplifyCFG also does this (and this code uses the same utility
// function), however this code is loop-aware, where SimplifyCFG is
// not. That gives it the advantage of being able to hoist
// loop-invariant instructions out of the way to open up more
// opportunities, and the disadvantage of having the responsibility
// to preserve dominator information.
bool UniqueExit = true;
if (!ExitBlocks.empty())
for (unsigned i = 1, e = ExitBlocks.size(); i != e; ++i)
if (ExitBlocks[i] != ExitBlocks[0]) {
UniqueExit = false;
break;
}
if (UniqueExit) {
for (unsigned i = 0, e = ExitingBlocks.size(); i != e; ++i) {
BasicBlock *ExitingBlock = ExitingBlocks[i];
if (!ExitingBlock->getSinglePredecessor()) continue;
BranchInst *BI = dyn_cast<BranchInst>(ExitingBlock->getTerminator());
if (!BI || !BI->isConditional()) continue;
CmpInst *CI = dyn_cast<CmpInst>(BI->getCondition());
if (!CI || CI->getParent() != ExitingBlock) continue;
// Attempt to hoist out all instructions except for the
// comparison and the branch.
bool AllInvariant = true;
for (BasicBlock::iterator I = ExitingBlock->begin(); &*I != BI; ) {
Instruction *Inst = I++;
// Skip debug info intrinsics.
if (isa<DbgInfoIntrinsic>(Inst))
continue;
if (Inst == CI)
continue;
if (!L->makeLoopInvariant(Inst, Changed,
Preheader ? Preheader->getTerminator() : 0)) {
AllInvariant = false;
break;
}
}
if (!AllInvariant) continue;
// The block has now been cleared of all instructions except for
// a comparison and a conditional branch. SimplifyCFG may be able
// to fold it now.
if (!FoldBranchToCommonDest(BI)) continue;
// Success. The block is now dead, so remove it from the loop,
// update the dominator tree and delete it.
DEBUG(dbgs() << "LoopSimplify: Eliminating exiting block "
<< ExitingBlock->getName() << "\n");
// If any reachable control flow within this loop has changed, notify
// ScalarEvolution. Currently assume the parent loop doesn't change
// (spliting edges doesn't count). If blocks, CFG edges, or other values
// in the parent loop change, then we need call to forgetLoop() for the
// parent instead.
if (SE)
SE->forgetLoop(L);
assert(pred_begin(ExitingBlock) == pred_end(ExitingBlock));
Changed = true;
LI->removeBlock(ExitingBlock);
DomTreeNode *Node = DT->getNode(ExitingBlock);
const std::vector<DomTreeNodeBase<BasicBlock> *> &Children =
Node->getChildren();
while (!Children.empty()) {
DomTreeNode *Child = Children.front();
DT->changeImmediateDominator(Child, Node->getIDom());
}
DT->eraseNode(ExitingBlock);
BI->getSuccessor(0)->removePredecessor(ExitingBlock);
BI->getSuccessor(1)->removePredecessor(ExitingBlock);
ExitingBlock->eraseFromParent();
}
}
return Changed;
}
/// DupRetToEnableTailCallOpts - Look for opportunities to duplicate return
/// instructions to the predecessor to enable tail call optimizations. The
/// case it is currently looking for is:
/// bb0:
/// %tmp0 = tail call i32 @f0()
/// br label %return
/// bb1:
/// %tmp1 = tail call i32 @f1()
/// br label %return
/// bb2:
/// %tmp2 = tail call i32 @f2()
/// br label %return
/// return:
/// %retval = phi i32 [ %tmp0, %bb0 ], [ %tmp1, %bb1 ], [ %tmp2, %bb2 ]
/// ret i32 %retval
///
/// =>
///
/// bb0:
/// %tmp0 = tail call i32 @f0()
/// ret i32 %tmp0
/// bb1:
/// %tmp1 = tail call i32 @f1()
/// ret i32 %tmp1
/// bb2:
/// %tmp2 = tail call i32 @f2()
/// ret i32 %tmp2
///
bool CodeGenPrepare::DupRetToEnableTailCallOpts(ReturnInst *RI) {
if (!TLI)
return false;
Value *V = RI->getReturnValue();
PHINode *PN = V ? dyn_cast<PHINode>(V) : NULL;
if (V && !PN)
return false;
BasicBlock *BB = RI->getParent();
if (PN && PN->getParent() != BB)
return false;
// It's not safe to eliminate the sign / zero extension of the return value.
// See llvm::isInTailCallPosition().
const Function *F = BB->getParent();
Attributes CallerRetAttr = F->getAttributes().getRetAttributes();
if ((CallerRetAttr & Attribute::ZExt) || (CallerRetAttr & Attribute::SExt))
return false;
// Make sure there are no instructions between the PHI and return, or that the
// return is the first instruction in the block.
if (PN) {
BasicBlock::iterator BI = BB->begin();
do {
++BI;
}
while (isa<DbgInfoIntrinsic>(BI));
if (&*BI != RI)
return false;
} else {
BasicBlock::iterator BI = BB->begin();
while (isa<DbgInfoIntrinsic>(BI)) ++BI;
if (&*BI != RI)
return false;
}
/// Only dup the ReturnInst if the CallInst is likely to be emitted as a tail
/// call.
SmallVector<CallInst*, 4> TailCalls;
if (PN) {
for (unsigned I = 0, E = PN->getNumIncomingValues(); I != E; ++I) {
CallInst *CI = dyn_cast<CallInst>(PN->getIncomingValue(I));
// Make sure the phi value is indeed produced by the tail call.
if (CI && CI->hasOneUse() && CI->getParent() == PN->getIncomingBlock(I) &&
TLI->mayBeEmittedAsTailCall(CI))
TailCalls.push_back(CI);
}
} else {
SmallPtrSet<BasicBlock*, 4> VisitedBBs;
for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE; ++PI) {
if (!VisitedBBs.insert(*PI))
continue;
BasicBlock::InstListType &InstList = (*PI)->getInstList();
BasicBlock::InstListType::reverse_iterator RI = InstList.rbegin();
BasicBlock::InstListType::reverse_iterator RE = InstList.rend();
do {
++RI;
}
while (RI != RE && isa<DbgInfoIntrinsic>(&*RI));
if (RI == RE)
continue;
CallInst *CI = dyn_cast<CallInst>(&*RI);
if (CI && CI->use_empty() && TLI->mayBeEmittedAsTailCall(CI))
TailCalls.push_back(CI);
}
}
bool Changed = false;
for (unsigned i = 0, e = TailCalls.size(); i != e; ++i) {
CallInst *CI = TailCalls[i];
CallSite CS(CI);
// Conservatively require the attributes of the call to match those of the
// return. Ignore noalias because it doesn't affect the call sequence.
Attributes CalleeRetAttr = CS.getAttributes().getRetAttributes();
if ((CalleeRetAttr ^ CallerRetAttr) & ~Attribute::NoAlias)
continue;
// Make sure the call instruction is followed by an unconditional branch to
// the return block.
BasicBlock *CallBB = CI->getParent();
BranchInst *BI = dyn_cast<BranchInst>(CallBB->getTerminator());
if (!BI || !BI->isUnconditional() || BI->getSuccessor(0) != BB)
continue;
// Duplicate the return into CallBB.
(void)FoldReturnIntoUncondBranch(RI, BB, CallBB);
ModifiedDT = Changed = true;
++NumRetsDup;
}
// If we eliminated all predecessors of the block, delete the block now.
if (Changed && pred_begin(BB) == pred_end(BB))
BB->eraseFromParent();
return Changed;
}
/// \brief Analyze the predecessors of each block and build up predicates
void StructurizeCFG::gatherPredicates(RegionNode *N) {
RegionInfo *RI = ParentRegion->getRegionInfo();
BasicBlock *BB = N->getEntry();
BBPredicates &Pred = Predicates[BB];
BBPredicates &LPred = LoopPreds[BB];
for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB);
PI != PE; ++PI) {
// Ignore it if it's a branch from outside into our region entry
if (!ParentRegion->contains(*PI))
continue;
Region *R = RI->getRegionFor(*PI);
if (R == ParentRegion) {
// It's a top level block in our region
BranchInst *Term = cast<BranchInst>((*PI)->getTerminator());
for (unsigned i = 0, e = Term->getNumSuccessors(); i != e; ++i) {
BasicBlock *Succ = Term->getSuccessor(i);
if (Succ != BB)
continue;
if (Visited.count(*PI)) {
// Normal forward edge
if (Term->isConditional()) {
// Try to treat it like an ELSE block
BasicBlock *Other = Term->getSuccessor(!i);
if (Visited.count(Other) && !Loops.count(Other) &&
!Pred.count(Other) && !Pred.count(*PI)) {
Pred[Other] = BoolFalse;
Pred[*PI] = BoolTrue;
continue;
}
}
Pred[*PI] = buildCondition(Term, i, false);
} else {
// Back edge
LPred[*PI] = buildCondition(Term, i, true);
}
}
} else {
// It's an exit from a sub region
while(R->getParent() != ParentRegion)
R = R->getParent();
// Edge from inside a subregion to its entry, ignore it
if (R == N)
continue;
BasicBlock *Entry = R->getEntry();
if (Visited.count(Entry))
Pred[Entry] = BoolTrue;
else
LPred[Entry] = BoolFalse;
}
}
}
bool LoopInterchangeTransform::adjustLoopBranches() {
DEBUG(dbgs() << "adjustLoopBranches called\n");
// Adjust the loop preheader
BasicBlock *InnerLoopHeader = InnerLoop->getHeader();
BasicBlock *OuterLoopHeader = OuterLoop->getHeader();
BasicBlock *InnerLoopLatch = InnerLoop->getLoopLatch();
BasicBlock *OuterLoopLatch = OuterLoop->getLoopLatch();
BasicBlock *OuterLoopPreHeader = OuterLoop->getLoopPreheader();
BasicBlock *InnerLoopPreHeader = InnerLoop->getLoopPreheader();
BasicBlock *OuterLoopPredecessor = OuterLoopPreHeader->getUniquePredecessor();
BasicBlock *InnerLoopLatchPredecessor =
InnerLoopLatch->getUniquePredecessor();
BasicBlock *InnerLoopLatchSuccessor;
BasicBlock *OuterLoopLatchSuccessor;
BranchInst *OuterLoopLatchBI =
dyn_cast<BranchInst>(OuterLoopLatch->getTerminator());
BranchInst *InnerLoopLatchBI =
dyn_cast<BranchInst>(InnerLoopLatch->getTerminator());
BranchInst *OuterLoopHeaderBI =
dyn_cast<BranchInst>(OuterLoopHeader->getTerminator());
BranchInst *InnerLoopHeaderBI =
dyn_cast<BranchInst>(InnerLoopHeader->getTerminator());
if (!OuterLoopPredecessor || !InnerLoopLatchPredecessor ||
!OuterLoopLatchBI || !InnerLoopLatchBI || !OuterLoopHeaderBI ||
!InnerLoopHeaderBI)
return false;
BranchInst *InnerLoopLatchPredecessorBI =
dyn_cast<BranchInst>(InnerLoopLatchPredecessor->getTerminator());
BranchInst *OuterLoopPredecessorBI =
dyn_cast<BranchInst>(OuterLoopPredecessor->getTerminator());
if (!OuterLoopPredecessorBI || !InnerLoopLatchPredecessorBI)
return false;
BasicBlock *InnerLoopHeaderSucessor = InnerLoopHeader->getUniqueSuccessor();
if (!InnerLoopHeaderSucessor)
return false;
// Adjust Loop Preheader and headers
unsigned NumSucc = OuterLoopPredecessorBI->getNumSuccessors();
for (unsigned i = 0; i < NumSucc; ++i) {
if (OuterLoopPredecessorBI->getSuccessor(i) == OuterLoopPreHeader)
OuterLoopPredecessorBI->setSuccessor(i, InnerLoopPreHeader);
}
NumSucc = OuterLoopHeaderBI->getNumSuccessors();
for (unsigned i = 0; i < NumSucc; ++i) {
if (OuterLoopHeaderBI->getSuccessor(i) == OuterLoopLatch)
OuterLoopHeaderBI->setSuccessor(i, LoopExit);
else if (OuterLoopHeaderBI->getSuccessor(i) == InnerLoopPreHeader)
OuterLoopHeaderBI->setSuccessor(i, InnerLoopHeaderSucessor);
}
BranchInst::Create(OuterLoopPreHeader, InnerLoopHeaderBI);
InnerLoopHeaderBI->eraseFromParent();
// -------------Adjust loop latches-----------
if (InnerLoopLatchBI->getSuccessor(0) == InnerLoopHeader)
InnerLoopLatchSuccessor = InnerLoopLatchBI->getSuccessor(1);
else
InnerLoopLatchSuccessor = InnerLoopLatchBI->getSuccessor(0);
NumSucc = InnerLoopLatchPredecessorBI->getNumSuccessors();
for (unsigned i = 0; i < NumSucc; ++i) {
if (InnerLoopLatchPredecessorBI->getSuccessor(i) == InnerLoopLatch)
InnerLoopLatchPredecessorBI->setSuccessor(i, InnerLoopLatchSuccessor);
}
if (OuterLoopLatchBI->getSuccessor(0) == OuterLoopHeader)
OuterLoopLatchSuccessor = OuterLoopLatchBI->getSuccessor(1);
else
OuterLoopLatchSuccessor = OuterLoopLatchBI->getSuccessor(0);
if (InnerLoopLatchBI->getSuccessor(1) == InnerLoopLatchSuccessor)
InnerLoopLatchBI->setSuccessor(1, OuterLoopLatchSuccessor);
else
InnerLoopLatchBI->setSuccessor(0, OuterLoopLatchSuccessor);
if (OuterLoopLatchBI->getSuccessor(0) == OuterLoopLatchSuccessor) {
OuterLoopLatchBI->setSuccessor(0, InnerLoopLatch);
} else {
OuterLoopLatchBI->setSuccessor(1, InnerLoopLatch);
}
return true;
}
