#define DEBUG_TYPE "jump-threading"
#include "llvm/Transforms/Scalar.h"
#include "llvm/IntrinsicInst.h"
+#include "llvm/LLVMContext.h"
#include "llvm/Pass.h"
#include "llvm/Analysis/ConstantFolding.h"
#include "llvm/Transforms/Utils/BasicBlockUtils.h"
#include "llvm/Support/Compiler.h"
#include "llvm/Support/Debug.h"
#include "llvm/Support/ValueHandle.h"
+#include "llvm/Support/raw_ostream.h"
using namespace llvm;
STATISTIC(NumThreads, "Number of jumps threaded");
JumpThreading() : FunctionPass(&ID) {}
virtual void getAnalysisUsage(AnalysisUsage &AU) const {
- AU.addRequired<TargetData>();
}
bool runOnFunction(Function &F);
bool ProcessBlock(BasicBlock *BB);
bool ThreadEdge(BasicBlock *BB, BasicBlock *PredBB, BasicBlock *SuccBB,
unsigned JumpThreadCost);
- BasicBlock *FactorCommonPHIPreds(PHINode *PN, Constant *CstVal);
+ BasicBlock *FactorCommonPHIPreds(PHINode *PN, Value *Val);
bool ProcessBranchOnDuplicateCond(BasicBlock *PredBB, BasicBlock *DestBB);
bool ProcessSwitchOnDuplicateCond(BasicBlock *PredBB, BasicBlock *DestBB);
/// runOnFunction - Top level algorithm.
///
bool JumpThreading::runOnFunction(Function &F) {
- DOUT << "Jump threading on function '" << F.getNameStart() << "'\n";
- TD = &getAnalysis<TargetData>();
+ DEBUG(errs() << "Jump threading on function '" << F.getName() << "'\n");
+ TD = getAnalysisIfAvailable<TargetData>();
FindLoopHeaders(F);
// 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();
+ DEBUG(errs() << " JT: Deleting dead block '" << BB->getName()
+ << "' with terminator: " << *BB->getTerminator());
LoopHeaders.erase(BB);
DeleteDeadBlock(BB);
Changed = true;
/// This is important for things like "phi i1 [true, true, false, true, x]"
/// where we only need to clone the block for the true blocks once.
///
-BasicBlock *JumpThreading::FactorCommonPHIPreds(PHINode *PN, Constant *CstVal) {
+BasicBlock *JumpThreading::FactorCommonPHIPreds(PHINode *PN, Value *Val) {
SmallVector<BasicBlock*, 16> CommonPreds;
for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
- if (PN->getIncomingValue(i) == CstVal)
+ if (PN->getIncomingValue(i) == Val)
CommonPreds.push_back(PN->getIncomingBlock(i));
if (CommonPreds.size() == 1)
return CommonPreds[0];
- DOUT << " Factoring out " << CommonPreds.size()
- << " common predecessors.\n";
+ DEBUG(errs() << " Factoring out " << CommonPreds.size()
+ << " common predecessors.\n");
return SplitBlockPredecessors(PN->getParent(),
&CommonPreds[0], CommonPreds.size(),
".thr_comm", this);
if (const CallInst *CI = dyn_cast<CallInst>(I)) {
if (!isa<IntrinsicInst>(CI))
Size += 3;
- else if (isa<VectorType>(CI->getType()))
+ else if (!isa<VectorType>(CI->getType()))
Size += 1;
}
}
// terminator to an unconditional branch. This can occur due to threading in
// other blocks.
if (isa<ConstantInt>(Condition)) {
- DOUT << " In block '" << BB->getNameStart()
- << "' folding terminator: " << *BB->getTerminator();
+ DEBUG(errs() << " In block '" << BB->getName()
+ << "' folding terminator: " << *BB->getTerminator());
++NumFolds;
ConstantFoldTerminator(BB);
return true;
BBTerm->getSuccessor(i)->removePredecessor(BB);
}
- DOUT << " In block '" << BB->getNameStart()
- << "' folding undef terminator: " << *BBTerm;
+ DEBUG(errs() << " In block '" << BB->getName()
+ << "' folding undef terminator: " << *BBTerm);
BranchInst::Create(BBTerm->getSuccessor(MinSucc), BBTerm);
BBTerm->eraseFromParent();
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;
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;
+ if (CmpInst *CondCmp = dyn_cast<CmpInst>(CondInst)) {
+ if (isa<PHINode>(CondCmp->getOperand(0))) {
+ // If we have "br (phi != 42)" and the phi node has any constant values
+ // as operands, we can thread through this block.
+ //
+ // If we have "br (cmp phi, x)" and the phi node contains x such that the
+ // comparison uniquely identifies the branch target, we can thread
+ // through this block.
+
+ if (ProcessBranchOnCompare(CondCmp, BB))
+ return true;
+ }
+
+ // If we have a comparison, loop over the predecessors to see if there is
+ // a condition with the same value.
+ pred_iterator PI = pred_begin(BB), E = pred_end(BB);
+ for (; PI != E; ++PI)
+ if (BranchInst *PBI = dyn_cast<BranchInst>((*PI)->getTerminator()))
+ if (PBI->isConditional() && *PI != BB) {
+ if (CmpInst *CI = dyn_cast<CmpInst>(PBI->getCondition())) {
+ if (CI->getOperand(0) == CondCmp->getOperand(0) &&
+ CI->getOperand(1) == CondCmp->getOperand(1) &&
+ CI->getPredicate() == CondCmp->getPredicate()) {
+ // TODO: Could handle things like (x != 4) --> (x == 17)
+ if (ProcessBranchOnDuplicateCond(*PI, 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
else if (PredBI->getSuccessor(0) != BB)
BranchDir = false;
else {
- DOUT << " In block '" << PredBB->getNameStart()
- << "' folding terminator: " << *PredBB->getTerminator();
+ DEBUG(errs() << " In block '" << PredBB->getName()
+ << "' folding terminator: " << *PredBB->getTerminator());
++NumFolds;
ConstantFoldTerminator(PredBB);
return true;
// 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();
+ DEBUG(errs() << " In block '" << BB->getName()
+ << "' folding condition to '" << BranchDir << "': "
+ << *BB->getTerminator());
++NumFolds;
DestBI->setCondition(ConstantInt::get(Type::Int1Ty, BranchDir));
ConstantFoldTerminator(BB);
// 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";
+ DEBUG(errs() << " Not threading BB '" << BB->getName()
+ << "' - Cost is too high: " << JumpThreadCost << "\n");
return false;
}
// 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;
+ DEBUG(errs() << "FORWARDING EDGE " << *DestVal << " FROM: " << *PredSI);
+ DEBUG(errs() << "THROUGH: " << *DestSI);
// If the destination has PHI nodes, just split the edge for updating
// simplicity.
// If the returned value is the load itself, replace with an undef. This can
// only happen in dead loops.
- if (AvailableVal == LI) AvailableVal = UndefValue::get(LI->getType());
+ if (AvailableVal == LI) AvailableVal =
+ AvailableVal->getContext().getUndef(LI->getType());
LI->replaceAllUsesWith(AvailableVal);
LI->eraseFromParent();
return true;
BasicBlock *BB = PN->getParent();
unsigned JumpThreadCost = getJumpThreadDuplicationCost(BB);
if (JumpThreadCost > Threshold) {
- DOUT << " Not threading BB '" << BB->getNameStart()
- << "' - Cost is too high: " << JumpThreadCost << "\n";
+ DEBUG(errs() << " Not threading BB '" << BB->getName()
+ << "' - Cost is too high: " << JumpThreadCost << "\n");
return false;
}
// Next, figure out which successor we are threading to.
BasicBlock *SuccBB;
if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator()))
- SuccBB = BI->getSuccessor(PredCst == ConstantInt::getFalse());
+ SuccBB = BI->getSuccessor(PredCst == PredBB->getContext().getFalse());
else {
SwitchInst *SI = cast<SwitchInst>(BB->getTerminator());
SuccBB = SI->getSuccessor(SI->findCaseValue(PredCst));
// 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";
+ DEBUG(errs() << " Not threading BB '" << BB->getName()
+ << "' - Cost is too high: " << JumpThreadCost << "\n");
return false;
}
return ThreadEdge(BB, PredBB, SuccBB, JumpThreadCost);
}
+/// GetResultOfComparison - Given an icmp/fcmp predicate and the left and right
+/// hand sides of the compare instruction, try to determine the result. If the
+/// result can not be determined, a null pointer is returned.
+static Constant *GetResultOfComparison(CmpInst::Predicate pred,
+ Value *LHS, Value *RHS,
+ LLVMContext &Context) {
+ if (Constant *CLHS = dyn_cast<Constant>(LHS))
+ if (Constant *CRHS = dyn_cast<Constant>(RHS))
+ return Context.getConstantExprCompare(pred, CLHS, CRHS);
+
+ if (LHS == RHS)
+ if (isa<IntegerType>(LHS->getType()) || isa<PointerType>(LHS->getType()))
+ return ICmpInst::isTrueWhenEqual(pred) ?
+ Context.getTrue() : Context.getFalse();
+
+ return 0;
+}
+
/// 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.
+/// node and a value. If we can identify when the comparison is true between
+/// the phi inputs and the value, 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));
+ Value *RHS = Cmp->getOperand(1);
// If the phi isn't in the current block, an incoming edge to this block
// doesn't control the destination.
// We can do this simplification if any comparisons fold to true or false.
// See if any do.
- Constant *PredCst = 0;
+ Value *PredVal = 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;
+ PredVal = PN->getIncomingValue(i);
+
+ Constant *Res = GetResultOfComparison(Cmp->getPredicate(), PredVal,
+ RHS, Cmp->getContext());
+ if (!Res) {
+ PredVal = 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();
}
// Otherwise, we can't fold this input.
- PredCst = 0;
+ PredVal = 0;
}
// If no match, bail out.
- if (PredCst == 0)
+ if (PredVal == 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";
+ DEBUG(errs() << " Not threading BB '" << BB->getName()
+ << "' - 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);
+ BasicBlock *PredBB = FactorCommonPHIPreds(PN, PredVal);
// 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 across BB '" << BB->getNameStart()
- << "' - would thread to self!\n";
+ DEBUG(errs() << " Not threading across BB '" << BB->getName()
+ << "' - would thread to self!\n");
return false;
}
// If threading this would thread across a loop header, don't thread the edge.
// See the comments above FindLoopHeaders for justifications and caveats.
if (LoopHeaders.count(BB)) {
- DOUT << " Not threading from '" << PredBB->getNameStart()
- << "' across loop header BB '" << BB->getNameStart()
- << "' to dest BB '" << SuccBB->getNameStart()
- << "' - it might create an irreducible loop!\n";
+ DEBUG(errs() << " Not threading from '" << PredBB->getName()
+ << "' across loop header BB '" << BB->getName()
+ << "' to dest BB '" << SuccBB->getName()
+ << "' - it might create an irreducible loop!\n");
return false;
}
// And finally, do it!
- DOUT << " Threading edge from '" << PredBB->getNameStart() << "' to '"
- << SuccBB->getNameStart() << "' with cost: " << JumpThreadCost
- << ", across block:\n "
- << *BB << "\n";
+ DEBUG(errs() << " Threading edge from '" << PredBB->getName() << "' to '"
+ << SuccBB->getName() << "' with cost: " << JumpThreadCost
+ << ", across block:\n "
+ << *BB << "\n");
// Jump Threading can not update SSA properties correctly if the values
// defined in the duplicated block are used outside of the block itself. For
// Clone the non-phi instructions of BB into NewBB, keeping track of the
// mapping and using it to remap operands in the cloned instructions.
for (; !isa<TerminatorInst>(BI); ++BI) {
- Instruction *New = BI->clone();
- New->setName(BI->getNameStart());
+ Instruction *New = BI->clone(BI->getContext());
+ New->setName(BI->getName());
NewBB->getInstList().push_back(New);
ValueMapping[BI] = New;
// Remap operands to patch up intra-block references.
for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i)
- if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i)))
- if (Value *Remapped = ValueMapping[Inst])
- New->setOperand(i, Remapped);
+ if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) {
+ DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst);
+ if (I != ValueMapping.end())
+ New->setOperand(i, I->second);
+ }
}
// We didn't copy the terminator from BB over to NewBB, because there is now
Value *IV = PN->getIncomingValueForBlock(BB);
// Remap the value if necessary.
- if (Instruction *Inst = dyn_cast<Instruction>(IV))
- if (Value *MappedIV = ValueMapping[Inst])
- IV = MappedIV;
+ if (Instruction *Inst = dyn_cast<Instruction>(IV)) {
+ DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst);
+ if (I != ValueMapping.end())
+ IV = I->second;
+ }
PN->addIncoming(IV, NewBB);
}
BI = NewBB->begin();
for (BasicBlock::iterator E = NewBB->end(); BI != E; ) {
Instruction *Inst = BI++;
- if (Constant *C = ConstantFoldInstruction(Inst, TD)) {
+ if (Constant *C = ConstantFoldInstruction(Inst, BB->getContext(), TD)) {
Inst->replaceAllUsesWith(C);
Inst->eraseFromParent();
continue;