BasicBlock *FactorCommonPHIPreds(PHINode *PN, Constant *CstVal);
bool ProcessJumpOnPHI(PHINode *PN);
- bool ProcessJumpOnLogicalPHI(PHINode *PN, bool isAnd);
+ bool ProcessBranchOnLogical(Value *V, BasicBlock *BB, bool isAnd);
+ bool ProcessBranchOnCompare(CmpInst *Cmp, BasicBlock *BB);
};
- 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(); }
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;
- }
+ isa<BranchInst>(BB->getTerminator()) &&
+ ProcessBranchOnLogical(CondI, BB,
+ CondI->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>(Condition))
+ if (isa<PHINode>(CondCmp->getOperand(0)) &&
+ isa<Constant>(CondCmp->getOperand(1)) &&
+ ProcessBranchOnCompare(CondCmp, BB))
+ return true;
+
return false;
}
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.
+ 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.
+ if (!isa<StructType>(I->getType())) {
DemoteRegToStack(*I);
+ continue;
}
+
+ // Alternatively, I must be a call or invoke that returns multiple retvals.
+ // We can't use 'DemoteRegToStack' because that will create loads and
+ // stores of aggregates which is not valid yet. If I is a call, we can just
+ // pull all the getresult instructions up to this block. If I is an invoke,
+ // we are out of luck.
+ BasicBlock::iterator IP = I; ++IP;
+ for (Value::use_iterator UI = I->use_begin(), E = I->use_end();
+ UI != E; ++UI)
+ cast<GetResultInst>(UI)->moveBefore(IP);
+ }
// 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