#include "llvm/Pass.h"
#include "llvm/ADT/DenseMap.h"
#include "llvm/ADT/Statistic.h"
+#include "llvm/ADT/STLExtras.h"
+#include "llvm/Analysis/ConstantFolding.h"
#include "llvm/Transforms/Utils/BasicBlockUtils.h"
#include "llvm/Transforms/Utils/Local.h"
+#include "llvm/Target/TargetData.h"
#include "llvm/Support/CommandLine.h"
#include "llvm/Support/Compiler.h"
#include "llvm/Support/Debug.h"
+#include "llvm/ADT/SmallPtrSet.h"
using namespace llvm;
STATISTIC(NumThreads, "Number of jumps threaded");
/// revectored to the false side of the second if.
///
class VISIBILITY_HIDDEN JumpThreading : public FunctionPass {
+ TargetData *TD;
public:
static char ID; // Pass identification
- JumpThreading() : FunctionPass((intptr_t)&ID) {}
+ JumpThreading() : FunctionPass(&ID) {}
+
+ virtual void getAnalysisUsage(AnalysisUsage &AU) const {
+ AU.addRequired<TargetData>();
+ }
bool runOnFunction(Function &F);
- bool ThreadBlock(BasicBlock *BB);
+ bool ProcessBlock(BasicBlock *BB);
void ThreadEdge(BasicBlock *BB, BasicBlock *PredBB, BasicBlock *SuccBB);
BasicBlock *FactorCommonPHIPreds(PHINode *PN, Constant *CstVal);
+ bool ProcessBranchOnDuplicateCond(BasicBlock *PredBB, BasicBlock *DestBB);
+ bool ProcessSwitchOnDuplicateCond(BasicBlock *PredBB, BasicBlock *DestBB);
bool ProcessJumpOnPHI(PHINode *PN);
- bool ProcessJumpOnLogicalPHI(PHINode *PN, bool isAnd);
+ bool ProcessBranchOnLogical(Value *V, BasicBlock *BB, bool isAnd);
+ bool ProcessBranchOnCompare(CmpInst *Cmp, BasicBlock *BB);
+
+ bool SimplifyPartiallyRedundantLoad(LoadInst *LI);
};
- char JumpThreading::ID = 0;
- RegisterPass<JumpThreading> X("jump-threading", "Jump Threading");
}
+char JumpThreading::ID = 0;
+static RegisterPass<JumpThreading>
+X("jump-threading", "Jump Threading");
+
// Public interface to the Jump Threading pass
FunctionPass *llvm::createJumpThreadingPass() { return new JumpThreading(); }
///
bool JumpThreading::runOnFunction(Function &F) {
DOUT << "Jump threading on function '" << F.getNameStart() << "'\n";
+ TD = &getAnalysis<TargetData>();
bool AnotherIteration = true, EverChanged = false;
while (AnotherIteration) {
AnotherIteration = false;
bool Changed = false;
- for (Function::iterator I = F.begin(), E = F.end(); I != E; ++I)
- while (ThreadBlock(I))
+ for (Function::iterator I = F.begin(), E = F.end(); I != E;) {
+ BasicBlock *BB = I;
+ while (ProcessBlock(BB))
Changed = true;
+
+ ++I;
+
+ // If the block is trivially dead, zap it. This eliminates the successor
+ // edges which simplifies the CFG.
+ if (pred_begin(BB) == pred_end(BB) &&
+ BB != &BB->getParent()->getEntryBlock()) {
+ DOUT << " JT: Deleting dead block '" << BB->getNameStart()
+ << "' with terminator: " << *BB->getTerminator();
+ DeleteDeadBlock(BB);
+ Changed = true;
+ }
+ }
AnotherIteration = Changed;
EverChanged |= Changed;
}
/// getJumpThreadDuplicationCost - Return the cost of duplicating this block to
/// thread across it.
static unsigned getJumpThreadDuplicationCost(const BasicBlock *BB) {
- BasicBlock::const_iterator I = BB->begin();
/// Ignore PHI nodes, these will be flattened when duplication happens.
- while (isa<PHINode>(*I)) ++I;
+ BasicBlock::const_iterator I = BB->getFirstNonPHI();
// Sum up the cost of each instruction until we get to the terminator. Don't
// include the terminator because the copy won't include it.
return Size;
}
-
-/// ThreadBlock - If there are any predecessors whose control can be threaded
+/// ProcessBlock - If there are any predecessors whose control can be threaded
/// through to a successor, transform them now.
-bool JumpThreading::ThreadBlock(BasicBlock *BB) {
- // See if this block ends with a branch of switch. If so, see if the
+bool JumpThreading::ProcessBlock(BasicBlock *BB) {
+ // If this block has a single predecessor, and if that pred has a single
+ // successor, merge the blocks. This encourages recursive jump threading
+ // because now the condition in this block can be threaded through
+ // predecessors of our predecessor block.
+ if (BasicBlock *SinglePred = BB->getSinglePredecessor())
+ if (SinglePred->getTerminator()->getNumSuccessors() == 1 &&
+ SinglePred != BB) {
+ // 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(BB);
+
+ if (isEntry && BB != &BB->getParent()->getEntryBlock())
+ BB->moveBefore(&BB->getParent()->getEntryBlock());
+ return true;
+ }
+
+ // See if this block ends with a branch or switch. If so, see if the
// condition is a phi node. If so, and if an entry of the phi node is a
// constant, we can thread the block.
Value *Condition;
return true;
}
+ // If the terminator is branching on an undef, we can pick any of the
+ // successors to branch to. Since this is arbitrary, we pick the successor
+ // with the fewest predecessors. This should reduce the in-degree of the
+ // others.
+ if (isa<UndefValue>(Condition)) {
+ TerminatorInst *BBTerm = BB->getTerminator();
+ unsigned MinSucc = 0;
+ BasicBlock *TestBB = BBTerm->getSuccessor(MinSucc);
+ // Compute the successor with the minimum number of predecessors.
+ unsigned MinNumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB));
+ for (unsigned i = 1, e = BBTerm->getNumSuccessors(); i != e; ++i) {
+ TestBB = BBTerm->getSuccessor(i);
+ unsigned NumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB));
+ if (NumPreds < MinNumPreds)
+ MinSucc = i;
+ }
+
+ // Fold the branch/switch.
+ for (unsigned i = 0, e = BBTerm->getNumSuccessors(); i != e; ++i) {
+ if (i == MinSucc) continue;
+ BBTerm->getSuccessor(i)->removePredecessor(BB);
+ }
+
+ DOUT << " In block '" << BB->getNameStart()
+ << "' folding undef terminator: " << *BBTerm;
+ BranchInst::Create(BBTerm->getSuccessor(MinSucc), BBTerm);
+ BBTerm->eraseFromParent();
+ return true;
+ }
+
+ Instruction *CondInst = dyn_cast<Instruction>(Condition);
+
+ // If the condition is an instruction defined in another block, see if a
+ // predecessor has the same condition:
+ // br COND, BBX, BBY
+ // BBX:
+ // br COND, BBZ, BBW
+ if (!Condition->hasOneUse() && // Multiple uses.
+ (CondInst == 0 || CondInst->getParent() != BB)) { // Non-local definition.
+ pred_iterator PI = pred_begin(BB), E = pred_end(BB);
+ if (isa<BranchInst>(BB->getTerminator())) {
+ for (; PI != E; ++PI)
+ if (BranchInst *PBI = dyn_cast<BranchInst>((*PI)->getTerminator()))
+ if (PBI->isConditional() && PBI->getCondition() == Condition &&
+ ProcessBranchOnDuplicateCond(*PI, BB))
+ return true;
+ } else {
+ assert(isa<SwitchInst>(BB->getTerminator()) && "Unknown jump terminator");
+ for (; PI != E; ++PI)
+ if (SwitchInst *PSI = dyn_cast<SwitchInst>((*PI)->getTerminator()))
+ if (PSI->getCondition() == Condition &&
+ ProcessSwitchOnDuplicateCond(*PI, BB))
+ return true;
+ }
+ }
+
// If there is only a single predecessor of this block, nothing to fold.
if (BB->getSinglePredecessor())
return false;
-
+
+ // All the rest of our checks depend on the condition being an instruction.
+ if (CondInst == 0)
+ return false;
+
// See if this is a phi node in the current block.
- PHINode *PN = dyn_cast<PHINode>(Condition);
- if (PN && PN->getParent() == BB)
- return ProcessJumpOnPHI(PN);
+ if (PHINode *PN = dyn_cast<PHINode>(CondInst))
+ if (PN->getParent() == BB)
+ return ProcessJumpOnPHI(PN);
// If this is a conditional branch whose condition is and/or of a phi, try to
// simplify it.
- if (BinaryOperator *CondI = dyn_cast<BinaryOperator>(Condition)) {
- if ((CondI->getOpcode() == Instruction::And ||
- CondI->getOpcode() == Instruction::Or) &&
- isa<BranchInst>(BB->getTerminator())) {
- if (PHINode *PN = dyn_cast<PHINode>(CondI->getOperand(0)))
- if (PN->getParent() == BB &&
- ProcessJumpOnLogicalPHI(PN, CondI->getOpcode() == Instruction::And))
- return true;
- if (PHINode *PN = dyn_cast<PHINode>(CondI->getOperand(1)))
- if (PN->getParent() == BB &&
- ProcessJumpOnLogicalPHI(PN, CondI->getOpcode() == Instruction::And))
- return true;
+ if ((CondInst->getOpcode() == Instruction::And ||
+ CondInst->getOpcode() == Instruction::Or) &&
+ isa<BranchInst>(BB->getTerminator()) &&
+ ProcessBranchOnLogical(CondInst, BB,
+ CondInst->getOpcode() == Instruction::And))
+ return true;
+
+ // If we have "br (phi != 42)" and the phi node has any constant values as
+ // operands, we can thread through this block.
+ if (CmpInst *CondCmp = dyn_cast<CmpInst>(CondInst))
+ if (isa<PHINode>(CondCmp->getOperand(0)) &&
+ isa<Constant>(CondCmp->getOperand(1)) &&
+ ProcessBranchOnCompare(CondCmp, BB))
+ return true;
+
+ // Check for some cases that are worth simplifying. Right now we want to look
+ // for loads that are used by a switch or by the condition for the branch. If
+ // we see one, check to see if it's partially redundant. If so, insert a PHI
+ // which can then be used to thread the values.
+ //
+ // This is particularly important because reg2mem inserts loads and stores all
+ // over the place, and this blocks jump threading if we don't zap them.
+ Value *SimplifyValue = CondInst;
+ if (CmpInst *CondCmp = dyn_cast<CmpInst>(SimplifyValue))
+ if (isa<Constant>(CondCmp->getOperand(1)))
+ SimplifyValue = CondCmp->getOperand(0);
+
+ if (LoadInst *LI = dyn_cast<LoadInst>(SimplifyValue))
+ if (SimplifyPartiallyRedundantLoad(LI))
+ return true;
+
+ // TODO: If we have: "br (X > 0)" and we have a predecessor where we know
+ // "(X == 4)" thread through this block.
+
+ return false;
+}
+
+/// ProcessBranchOnDuplicateCond - We found a block and a predecessor of that
+/// block that jump on exactly the same condition. This means that we almost
+/// always know the direction of the edge in the DESTBB:
+/// PREDBB:
+/// br COND, DESTBB, BBY
+/// DESTBB:
+/// br COND, BBZ, BBW
+///
+/// If DESTBB has multiple predecessors, we can't just constant fold the branch
+/// in DESTBB, we have to thread over it.
+bool JumpThreading::ProcessBranchOnDuplicateCond(BasicBlock *PredBB,
+ BasicBlock *BB) {
+ BranchInst *PredBI = cast<BranchInst>(PredBB->getTerminator());
+
+ // If both successors of PredBB go to DESTBB, we don't know anything. We can
+ // fold the branch to an unconditional one, which allows other recursive
+ // simplifications.
+ bool BranchDir;
+ if (PredBI->getSuccessor(1) != BB)
+ BranchDir = true;
+ else if (PredBI->getSuccessor(0) != BB)
+ BranchDir = false;
+ else {
+ DOUT << " In block '" << PredBB->getNameStart()
+ << "' folding terminator: " << *PredBB->getTerminator();
+ ++NumFolds;
+ ConstantFoldTerminator(PredBB);
+ return true;
+ }
+
+ BranchInst *DestBI = cast<BranchInst>(BB->getTerminator());
+
+ // If the dest block has one predecessor, just fix the branch condition to a
+ // constant and fold it.
+ if (BB->getSinglePredecessor()) {
+ DOUT << " In block '" << BB->getNameStart()
+ << "' folding condition to '" << BranchDir << "': "
+ << *BB->getTerminator();
+ ++NumFolds;
+ DestBI->setCondition(ConstantInt::get(Type::Int1Ty, BranchDir));
+ ConstantFoldTerminator(BB);
+ return true;
+ }
+
+ // Otherwise we need to thread from PredBB to DestBB's successor which
+ // involves code duplication. Check to see if it is worth it.
+ unsigned JumpThreadCost = getJumpThreadDuplicationCost(BB);
+ if (JumpThreadCost > Threshold) {
+ DOUT << " Not threading BB '" << BB->getNameStart()
+ << "' - Cost is too high: " << JumpThreadCost << "\n";
+ return false;
+ }
+
+ // Next, figure out which successor we are threading to.
+ BasicBlock *SuccBB = DestBI->getSuccessor(!BranchDir);
+
+ // If threading to the same block as we come from, we would infinite loop.
+ if (SuccBB == BB) {
+ DOUT << " Not threading BB '" << BB->getNameStart()
+ << "' - would thread to self!\n";
+ return false;
+ }
+
+ // And finally, do it!
+ DOUT << " Threading edge from '" << PredBB->getNameStart() << "' to '"
+ << SuccBB->getNameStart() << "' with cost: " << JumpThreadCost
+ << ", across block:\n "
+ << *BB << "\n";
+
+ ThreadEdge(BB, PredBB, SuccBB);
+ ++NumThreads;
+ return true;
+}
+
+/// ProcessSwitchOnDuplicateCond - We found a block and a predecessor of that
+/// block that switch on exactly the same condition. This means that we almost
+/// always know the direction of the edge in the DESTBB:
+/// PREDBB:
+/// switch COND [... DESTBB, BBY ... ]
+/// DESTBB:
+/// switch COND [... BBZ, BBW ]
+///
+/// Optimizing switches like this is very important, because simplifycfg builds
+/// switches out of repeated 'if' conditions.
+bool JumpThreading::ProcessSwitchOnDuplicateCond(BasicBlock *PredBB,
+ BasicBlock *DestBB) {
+ SwitchInst *PredSI = cast<SwitchInst>(PredBB->getTerminator());
+ SwitchInst *DestSI = cast<SwitchInst>(DestBB->getTerminator());
+
+ // There are a variety of optimizations that we can potentially do on these
+ // blocks: we order them from most to least preferable.
+
+ // If DESTBB *just* contains the switch, then we can forward edges from PREDBB
+ // directly to their destination. This does not introduce *any* code size
+ // growth.
+
+ // FIXME: Thread if it just contains a PHI.
+ if (isa<SwitchInst>(DestBB->begin())) {
+ bool MadeChange = false;
+ // Ignore the default edge for now.
+ for (unsigned i = 1, e = DestSI->getNumSuccessors(); i != e; ++i) {
+ ConstantInt *DestVal = DestSI->getCaseValue(i);
+ BasicBlock *DestSucc = DestSI->getSuccessor(i);
+
+ // Okay, DestSI has a case for 'DestVal' that goes to 'DestSucc'. See if
+ // PredSI has an explicit case for it. If so, forward. If it is covered
+ // by the default case, we can't update PredSI.
+ unsigned PredCase = PredSI->findCaseValue(DestVal);
+ if (PredCase == 0) continue;
+
+ // If PredSI doesn't go to DestBB on this value, then it won't reach the
+ // case on this condition.
+ if (PredSI->getSuccessor(PredCase) != DestBB &&
+ DestSI->getSuccessor(i) != DestBB)
+ continue;
+
+ // Otherwise, we're safe to make the change. Make sure that the edge from
+ // DestSI to DestSucc is not critical and has no PHI nodes.
+ DOUT << "FORWARDING EDGE " << *DestVal << " FROM: " << *PredSI;
+ DOUT << "THROUGH: " << *DestSI;
+
+ // If the destination has PHI nodes, just split the edge for updating
+ // simplicity.
+ if (isa<PHINode>(DestSucc->begin()) && !DestSucc->getSinglePredecessor()){
+ SplitCriticalEdge(DestSI, i, this);
+ DestSucc = DestSI->getSuccessor(i);
+ }
+ FoldSingleEntryPHINodes(DestSucc);
+ PredSI->setSuccessor(PredCase, DestSucc);
+ MadeChange = true;
}
+
+ if (MadeChange)
+ return true;
}
return false;
}
+
+/// SimplifyPartiallyRedundantLoad - If LI is an obviously partially redundant
+/// load instruction, eliminate it by replacing it with a PHI node. This is an
+/// important optimization that encourages jump threading, and needs to be run
+/// interlaced with other jump threading tasks.
+bool JumpThreading::SimplifyPartiallyRedundantLoad(LoadInst *LI) {
+ // Don't hack volatile loads.
+ if (LI->isVolatile()) return false;
+
+ // If the load is defined in a block with exactly one predecessor, it can't be
+ // partially redundant.
+ BasicBlock *LoadBB = LI->getParent();
+ if (LoadBB->getSinglePredecessor())
+ return false;
+
+ Value *LoadedPtr = LI->getOperand(0);
+
+ // If the loaded operand is defined in the LoadBB, it can't be available.
+ // FIXME: Could do PHI translation, that would be fun :)
+ if (Instruction *PtrOp = dyn_cast<Instruction>(LoadedPtr))
+ if (PtrOp->getParent() == LoadBB)
+ return false;
+
+ // Scan a few instructions up from the load, to see if it is obviously live at
+ // the entry to its block.
+ BasicBlock::iterator BBIt = LI;
+
+ if (Value *AvailableVal = FindAvailableLoadedValue(LoadedPtr, LoadBB,
+ BBIt, 6)) {
+ // If the value if the load is locally available within the block, just use
+ // it. This frequently occurs for reg2mem'd allocas.
+ //cerr << "LOAD ELIMINATED:\n" << *BBIt << *LI << "\n";
+ LI->replaceAllUsesWith(AvailableVal);
+ LI->eraseFromParent();
+ return true;
+ }
+
+ // Otherwise, if we scanned the whole block and got to the top of the block,
+ // we know the block is locally transparent to the load. If not, something
+ // might clobber its value.
+ if (BBIt != LoadBB->begin())
+ return false;
+
+
+ SmallPtrSet<BasicBlock*, 8> PredsScanned;
+ typedef SmallVector<std::pair<BasicBlock*, Value*>, 8> AvailablePredsTy;
+ AvailablePredsTy AvailablePreds;
+ BasicBlock *OneUnavailablePred = 0;
+
+ // If we got here, the loaded value is transparent through to the start of the
+ // block. Check to see if it is available in any of the predecessor blocks.
+ for (pred_iterator PI = pred_begin(LoadBB), PE = pred_end(LoadBB);
+ PI != PE; ++PI) {
+ BasicBlock *PredBB = *PI;
+
+ // If we already scanned this predecessor, skip it.
+ if (!PredsScanned.insert(PredBB))
+ continue;
+
+ // Scan the predecessor to see if the value is available in the pred.
+ BBIt = PredBB->end();
+ Value *PredAvailable = FindAvailableLoadedValue(LoadedPtr, PredBB, BBIt, 6);
+ if (!PredAvailable) {
+ OneUnavailablePred = PredBB;
+ continue;
+ }
+
+ // If so, this load is partially redundant. Remember this info so that we
+ // can create a PHI node.
+ AvailablePreds.push_back(std::make_pair(PredBB, PredAvailable));
+ }
+
+ // If the loaded value isn't available in any predecessor, it isn't partially
+ // redundant.
+ if (AvailablePreds.empty()) return false;
+
+ // Okay, the loaded value is available in at least one (and maybe all!)
+ // predecessors. If the value is unavailable in more than one unique
+ // predecessor, we want to insert a merge block for those common predecessors.
+ // This ensures that we only have to insert one reload, thus not increasing
+ // code size.
+ BasicBlock *UnavailablePred = 0;
+
+ // If there is exactly one predecessor where the value is unavailable, the
+ // already computed 'OneUnavailablePred' block is it. If it ends in an
+ // unconditional branch, we know that it isn't a critical edge.
+ if (PredsScanned.size() == AvailablePreds.size()+1 &&
+ OneUnavailablePred->getTerminator()->getNumSuccessors() == 1) {
+ UnavailablePred = OneUnavailablePred;
+ } else if (PredsScanned.size() != AvailablePreds.size()) {
+ // Otherwise, we had multiple unavailable predecessors or we had a critical
+ // edge from the one.
+ SmallVector<BasicBlock*, 8> PredsToSplit;
+ SmallPtrSet<BasicBlock*, 8> AvailablePredSet;
+
+ for (unsigned i = 0, e = AvailablePreds.size(); i != e; ++i)
+ AvailablePredSet.insert(AvailablePreds[i].first);
+
+ // Add all the unavailable predecessors to the PredsToSplit list.
+ for (pred_iterator PI = pred_begin(LoadBB), PE = pred_end(LoadBB);
+ PI != PE; ++PI)
+ if (!AvailablePredSet.count(*PI))
+ PredsToSplit.push_back(*PI);
+
+ // Split them out to their own block.
+ UnavailablePred =
+ SplitBlockPredecessors(LoadBB, &PredsToSplit[0], PredsToSplit.size(),
+ "thread-split", this);
+ }
+
+ // If the value isn't available in all predecessors, then there will be
+ // exactly one where it isn't available. Insert a load on that edge and add
+ // it to the AvailablePreds list.
+ if (UnavailablePred) {
+ assert(UnavailablePred->getTerminator()->getNumSuccessors() == 1 &&
+ "Can't handle critical edge here!");
+ Value *NewVal = new LoadInst(LoadedPtr, LI->getName()+".pr",
+ UnavailablePred->getTerminator());
+ AvailablePreds.push_back(std::make_pair(UnavailablePred, NewVal));
+ }
+
+ // Now we know that each predecessor of this block has a value in
+ // AvailablePreds, sort them for efficient access as we're walking the preds.
+ array_pod_sort(AvailablePreds.begin(), AvailablePreds.end());
+
+ // Create a PHI node at the start of the block for the PRE'd load value.
+ PHINode *PN = PHINode::Create(LI->getType(), "", LoadBB->begin());
+ PN->takeName(LI);
+
+ // Insert new entries into the PHI for each predecessor. A single block may
+ // have multiple entries here.
+ for (pred_iterator PI = pred_begin(LoadBB), E = pred_end(LoadBB); PI != E;
+ ++PI) {
+ AvailablePredsTy::iterator I =
+ std::lower_bound(AvailablePreds.begin(), AvailablePreds.end(),
+ std::make_pair(*PI, (Value*)0));
+
+ assert(I != AvailablePreds.end() && I->first == *PI &&
+ "Didn't find entry for predecessor!");
+
+ PN->addIncoming(I->second, I->first);
+ }
+
+ //cerr << "PRE: " << *LI << *PN << "\n";
+
+ LI->replaceAllUsesWith(PN);
+ LI->eraseFromParent();
+
+ return true;
+}
+
+
/// ProcessJumpOnPHI - We have a conditional branch of switch on a PHI node in
/// the current block. See if there are any simplifications we can do based on
/// inputs to the phi node.
bool JumpThreading::ProcessJumpOnPHI(PHINode *PN) {
// See if the phi node has any constant values. If so, we can determine where
// the corresponding predecessor will branch.
- unsigned PredNo = ~0U;
ConstantInt *PredCst = 0;
- for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
- if ((PredCst = dyn_cast<ConstantInt>(PN->getIncomingValue(i)))) {
- PredNo = i;
+ for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
+ if ((PredCst = dyn_cast<ConstantInt>(PN->getIncomingValue(i))))
break;
- }
- }
// If no incoming value has a constant, we don't know the destination of any
// predecessors.
- if (PredNo == ~0U)
+ if (PredCst == 0)
return false;
// See if the cost of duplicating this block is low enough.
SuccBB = SI->getSuccessor(SI->findCaseValue(PredCst));
}
+ // If threading to the same block as we come from, we would infinite loop.
+ if (SuccBB == BB) {
+ DOUT << " Not threading BB '" << BB->getNameStart()
+ << "' - would thread to self!\n";
+ return false;
+ }
+
// And finally, do it!
DOUT << " Threading edge from '" << PredBB->getNameStart() << "' to '"
<< SuccBB->getNameStart() << "' with cost: " << JumpThreadCost
/// the predecessor corresponding to the 'false' will always jump to the false
/// destination of the branch.
///
-bool JumpThreading::ProcessJumpOnLogicalPHI(PHINode *PN, bool isAnd) {
-
+bool JumpThreading::ProcessBranchOnLogical(Value *V, BasicBlock *BB,
+ bool isAnd) {
+ // If this is a binary operator tree of the same AND/OR opcode, check the
+ // LHS/RHS.
+ if (BinaryOperator *BO = dyn_cast<BinaryOperator>(V))
+ if ((isAnd && BO->getOpcode() == Instruction::And) ||
+ (!isAnd && BO->getOpcode() == Instruction::Or)) {
+ if (ProcessBranchOnLogical(BO->getOperand(0), BB, isAnd))
+ return true;
+ if (ProcessBranchOnLogical(BO->getOperand(1), BB, isAnd))
+ return true;
+ }
+
+ // If this isn't a PHI node, we can't handle it.
+ PHINode *PN = dyn_cast<PHINode>(V);
+ if (!PN || PN->getParent() != BB) return false;
+
// We can only do the simplification for phi nodes of 'false' with AND or
// 'true' with OR. See if we have any entries in the phi for this.
unsigned PredNo = ~0U;
return false;
// See if the cost of duplicating this block is low enough.
- BasicBlock *BB = PN->getParent();
unsigned JumpThreadCost = getJumpThreadDuplicationCost(BB);
if (JumpThreadCost > Threshold) {
DOUT << " Not threading BB '" << BB->getNameStart()
// 'true' block.
BasicBlock *SuccBB = BB->getTerminator()->getSuccessor(isAnd);
+ // If threading to the same block as we come from, we would infinite loop.
+ if (SuccBB == BB) {
+ DOUT << " Not threading BB '" << BB->getNameStart()
+ << "' - would thread to self!\n";
+ return false;
+ }
+
+ // And finally, do it!
+ DOUT << " Threading edge through bool from '" << PredBB->getNameStart()
+ << "' to '" << SuccBB->getNameStart() << "' with cost: "
+ << JumpThreadCost << ", across block:\n "
+ << *BB << "\n";
+
+ ThreadEdge(BB, PredBB, SuccBB);
+ ++NumThreads;
+ return true;
+}
+
+/// ProcessBranchOnCompare - We found a branch on a comparison between a phi
+/// node and a constant. If the PHI node contains any constants as inputs, we
+/// can fold the compare for that edge and thread through it.
+bool JumpThreading::ProcessBranchOnCompare(CmpInst *Cmp, BasicBlock *BB) {
+ PHINode *PN = cast<PHINode>(Cmp->getOperand(0));
+ Constant *RHS = cast<Constant>(Cmp->getOperand(1));
+
+ // If the phi isn't in the current block, an incoming edge to this block
+ // doesn't control the destination.
+ if (PN->getParent() != BB)
+ return false;
+
+ // We can do this simplification if any comparisons fold to true or false.
+ // See if any do.
+ Constant *PredCst = 0;
+ bool TrueDirection = false;
+ for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
+ PredCst = dyn_cast<Constant>(PN->getIncomingValue(i));
+ if (PredCst == 0) continue;
+
+ Constant *Res;
+ if (ICmpInst *ICI = dyn_cast<ICmpInst>(Cmp))
+ Res = ConstantExpr::getICmp(ICI->getPredicate(), PredCst, RHS);
+ else
+ Res = ConstantExpr::getFCmp(cast<FCmpInst>(Cmp)->getPredicate(),
+ PredCst, RHS);
+ // If this folded to a constant expr, we can't do anything.
+ if (ConstantInt *ResC = dyn_cast<ConstantInt>(Res)) {
+ TrueDirection = ResC->getZExtValue();
+ break;
+ }
+ // If this folded to undef, just go the false way.
+ if (isa<UndefValue>(Res)) {
+ TrueDirection = false;
+ break;
+ }
+
+ // Otherwise, we can't fold this input.
+ PredCst = 0;
+ }
+
+ // If no match, bail out.
+ if (PredCst == 0)
+ return false;
+
+ // See if the cost of duplicating this block is low enough.
+ unsigned JumpThreadCost = getJumpThreadDuplicationCost(BB);
+ if (JumpThreadCost > Threshold) {
+ DOUT << " Not threading BB '" << BB->getNameStart()
+ << "' - Cost is too high: " << JumpThreadCost << "\n";
+ return false;
+ }
+
+ // If so, we can actually do this threading. Merge any common predecessors
+ // that will act the same.
+ BasicBlock *PredBB = FactorCommonPHIPreds(PN, PredCst);
+
+ // Next, get our successor.
+ BasicBlock *SuccBB = BB->getTerminator()->getSuccessor(!TrueDirection);
+
+ // If threading to the same block as we come from, we would infinite loop.
+ if (SuccBB == BB) {
+ DOUT << " Not threading BB '" << BB->getNameStart()
+ << "' - would thread to self!\n";
+ return false;
+ }
+
+
// And finally, do it!
DOUT << " Threading edge through bool from '" << PredBB->getNameStart()
<< "' to '" << SuccBB->getNameStart() << "' with cost: "
// Jump Threading can not update SSA properties correctly if the values
// defined in the duplicated block are used outside of the block itself. For
// this reason, we spill all values that are used outside of BB to the stack.
- for (BasicBlock::iterator I = BB->begin(); I != BB->end(); ++I)
- if (I->isUsedOutsideOfBlock(BB)) {
- // We found a use of I outside of BB. Create a new stack slot to
- // break this inter-block usage pattern.
- DemoteRegToStack(*I);
- }
+ for (BasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) {
+ if (!I->isUsedOutsideOfBlock(BB))
+ continue;
+
+ // We found a use of I outside of BB. Create a new stack slot to
+ // break this inter-block usage pattern.
+ DemoteRegToStack(*I);
+ }
// We are going to have to map operands from the original BB block to the new
// copy of the block 'NewBB'. If there are PHI nodes in BB, evaluate them to
PN->addIncoming(IV, NewBB);
}
- // Finally, NewBB is good to go. Update the terminator of PredBB to jump to
+ // Ok, NewBB is good to go. Update the terminator of PredBB to jump to
// NewBB instead of BB. This eliminates predecessors from BB, which requires
// us to simplify any PHI nodes in BB.
TerminatorInst *PredTerm = PredBB->getTerminator();
BB->removePredecessor(PredBB);
PredTerm->setSuccessor(i, NewBB);
}
+
+ // At this point, the IR is fully up to date and consistent. Do a quick scan
+ // over the new instructions and zap any that are constants or dead. This
+ // frequently happens because of phi translation.
+ BI = NewBB->begin();
+ for (BasicBlock::iterator E = NewBB->end(); BI != E; ) {
+ Instruction *Inst = BI++;
+ if (Constant *C = ConstantFoldInstruction(Inst, TD)) {
+ Inst->replaceAllUsesWith(C);
+ Inst->eraseFromParent();
+ continue;
+ }
+
+ RecursivelyDeleteTriviallyDeadInstructions(Inst);
+ }
}
projects / oota-llvm.git / blobdiff