#define DEBUG_TYPE "jump-threading"
#include "llvm/Transforms/Scalar.h"
#include "llvm/IntrinsicInst.h"
+#include "llvm/LLVMContext.h"
#include "llvm/Pass.h"
-#include "llvm/ADT/DenseMap.h"
-#include "llvm/ADT/Statistic.h"
+#include "llvm/Analysis/InstructionSimplify.h"
+#include "llvm/Analysis/LazyValueInfo.h"
#include "llvm/Transforms/Utils/BasicBlockUtils.h"
#include "llvm/Transforms/Utils/Local.h"
+#include "llvm/Transforms/Utils/SSAUpdater.h"
+#include "llvm/Target/TargetData.h"
+#include "llvm/ADT/DenseMap.h"
+#include "llvm/ADT/Statistic.h"
+#include "llvm/ADT/STLExtras.h"
+#include "llvm/ADT/SmallPtrSet.h"
+#include "llvm/ADT/SmallSet.h"
#include "llvm/Support/CommandLine.h"
-#include "llvm/Support/Compiler.h"
#include "llvm/Support/Debug.h"
+#include "llvm/Support/raw_ostream.h"
using namespace llvm;
STATISTIC(NumThreads, "Number of jumps threaded");
STATISTIC(NumFolds, "Number of terminators folded");
+STATISTIC(NumDupes, "Number of branch blocks duplicated to eliminate phi");
static cl::opt<unsigned>
Threshold("jump-threading-threshold",
cl::desc("Max block size to duplicate for jump threading"),
cl::init(6), cl::Hidden);
+// Turn on use of LazyValueInfo.
+static cl::opt<bool>
+EnableLVI("enable-jump-threading-lvi", cl::ReallyHidden);
+
+
+
namespace {
/// This pass performs 'jump threading', which looks at blocks that have
/// multiple predecessors and multiple successors. If one or more of the
/// In this case, the unconditional branch at the end of the first if can be
/// revectored to the false side of the second if.
///
- class VISIBILITY_HIDDEN JumpThreading : public FunctionPass {
+ class JumpThreading : public FunctionPass {
+ TargetData *TD;
+ LazyValueInfo *LVI;
+#ifdef NDEBUG
+ SmallPtrSet<BasicBlock*, 16> LoopHeaders;
+#else
+ SmallSet<AssertingVH<BasicBlock>, 16> LoopHeaders;
+#endif
public:
static char ID; // Pass identification
- JumpThreading() : FunctionPass((intptr_t)&ID) {}
+ JumpThreading() : FunctionPass(&ID) {}
bool runOnFunction(Function &F);
- bool ThreadBlock(BasicBlock *BB);
- void ThreadEdge(BasicBlock *BB, BasicBlock *PredBB, BasicBlock *SuccBB);
- BasicBlock *FactorCommonPHIPreds(PHINode *PN, Constant *CstVal);
+
+ virtual void getAnalysisUsage(AnalysisUsage &AU) const {
+ if (EnableLVI)
+ AU.addRequired<LazyValueInfo>();
+ }
+
+ void FindLoopHeaders(Function &F);
+ bool ProcessBlock(BasicBlock *BB);
+ bool ThreadEdge(BasicBlock *BB, const SmallVectorImpl<BasicBlock*> &PredBBs,
+ BasicBlock *SuccBB);
+ bool DuplicateCondBranchOnPHIIntoPred(BasicBlock *BB,
+ BasicBlock *PredBB);
+
+ typedef SmallVectorImpl<std::pair<ConstantInt*,
+ BasicBlock*> > PredValueInfo;
+
+ bool ComputeValueKnownInPredecessors(Value *V, BasicBlock *BB,
+ PredValueInfo &Result);
+ bool ProcessThreadableEdges(Instruction *CondInst, BasicBlock *BB);
+
+
+ bool ProcessBranchOnDuplicateCond(BasicBlock *PredBB, BasicBlock *DestBB);
+ bool ProcessSwitchOnDuplicateCond(BasicBlock *PredBB, BasicBlock *DestBB);
bool ProcessJumpOnPHI(PHINode *PN);
- bool ProcessBranchOnLogical(Value *V, BasicBlock *BB, bool isAnd);
- bool ProcessBranchOnCompare(CmpInst *Cmp, BasicBlock *BB);
+
+ bool SimplifyPartiallyRedundantLoad(LoadInst *LI);
};
}
/// runOnFunction - Top level algorithm.
///
bool JumpThreading::runOnFunction(Function &F) {
- DOUT << "Jump threading on function '" << F.getNameStart() << "'\n";
+ DEBUG(errs() << "Jump threading on function '" << F.getName() << "'\n");
+ TD = getAnalysisIfAvailable<TargetData>();
+ LVI = EnableLVI ? &getAnalysis<LazyValueInfo>() : 0;
+
+ FindLoopHeaders(F);
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;
+ // Thread all of the branches we can over this block.
+ 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()) {
+ DEBUG(errs() << " JT: Deleting dead block '" << BB->getName()
+ << "' with terminator: " << *BB->getTerminator() << '\n');
+ LoopHeaders.erase(BB);
+ DeleteDeadBlock(BB);
+ Changed = true;
+ } else if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) {
+ // Can't thread an unconditional jump, but if the block is "almost
+ // empty", we can replace uses of it with uses of the successor and make
+ // this dead.
+ if (BI->isUnconditional() &&
+ BB != &BB->getParent()->getEntryBlock()) {
+ BasicBlock::iterator BBI = BB->getFirstNonPHI();
+ // Ignore dbg intrinsics.
+ while (isa<DbgInfoIntrinsic>(BBI))
+ ++BBI;
+ // If the terminator is the only non-phi instruction, try to nuke it.
+ if (BBI->isTerminator()) {
+ // Since TryToSimplifyUncondBranchFromEmptyBlock may delete the
+ // block, we have to make sure it isn't in the LoopHeaders set. We
+ // reinsert afterward in the rare case when the block isn't deleted.
+ bool ErasedFromLoopHeaders = LoopHeaders.erase(BB);
+
+ if (TryToSimplifyUncondBranchFromEmptyBlock(BB))
+ Changed = true;
+ else if (ErasedFromLoopHeaders)
+ LoopHeaders.insert(BB);
+ }
+ }
+ }
+ }
AnotherIteration = Changed;
EverChanged |= Changed;
}
+
+ LoopHeaders.clear();
return EverChanged;
}
-/// FactorCommonPHIPreds - If there are multiple preds with the same incoming
-/// value for the PHI, factor them together so we get one block to thread for
-/// the whole group.
-/// 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) {
- SmallVector<BasicBlock*, 16> CommonPreds;
- for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
- if (PN->getIncomingValue(i) == CstVal)
- CommonPreds.push_back(PN->getIncomingBlock(i));
-
- if (CommonPreds.size() == 1)
- return CommonPreds[0];
-
- DOUT << " Factoring out " << CommonPreds.size()
- << " common predecessors.\n";
- return SplitBlockPredecessors(PN->getParent(),
- &CommonPreds[0], CommonPreds.size(),
- ".thr_comm", this);
-}
-
-
/// getJumpThreadDuplicationCost - Return the cost of duplicating this block to
/// thread across it.
static unsigned getJumpThreadDuplicationCost(const BasicBlock *BB) {
/// Ignore PHI nodes, these will be flattened when duplication happens.
BasicBlock::const_iterator I = BB->getFirstNonPHI();
-
+
+ // FIXME: THREADING will delete values that are just used to compute the
+ // branch, so they shouldn't count against the duplication cost.
+
+
// 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.
unsigned Size = 0;
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;
}
}
return Size;
}
+/// FindLoopHeaders - We do not want jump threading to turn proper loop
+/// structures into irreducible loops. Doing this breaks up the loop nesting
+/// hierarchy and pessimizes later transformations. To prevent this from
+/// happening, we first have to find the loop headers. Here we approximate this
+/// by finding targets of backedges in the CFG.
+///
+/// Note that there definitely are cases when we want to allow threading of
+/// edges across a loop header. For example, threading a jump from outside the
+/// loop (the preheader) to an exit block of the loop is definitely profitable.
+/// It is also almost always profitable to thread backedges from within the loop
+/// to exit blocks, and is often profitable to thread backedges to other blocks
+/// within the loop (forming a nested loop). This simple analysis is not rich
+/// enough to track all of these properties and keep it up-to-date as the CFG
+/// mutates, so we don't allow any of these transformations.
+///
+void JumpThreading::FindLoopHeaders(Function &F) {
+ SmallVector<std::pair<const BasicBlock*,const BasicBlock*>, 32> Edges;
+ FindFunctionBackedges(F, Edges);
+
+ for (unsigned i = 0, e = Edges.size(); i != e; ++i)
+ LoopHeaders.insert(const_cast<BasicBlock*>(Edges[i].second));
+}
+
+/// ComputeValueKnownInPredecessors - Given a basic block BB and a value V, see
+/// if we can infer that the value is a known ConstantInt in any of our
+/// predecessors. If so, return the known list of value and pred BB in the
+/// result vector. If a value is known to be undef, it is returned as null.
+///
+/// This returns true if there were any known values.
+///
+bool JumpThreading::
+ComputeValueKnownInPredecessors(Value *V, BasicBlock *BB,PredValueInfo &Result){
+ // If V is a constantint, then it is known in all predecessors.
+ if (isa<ConstantInt>(V) || isa<UndefValue>(V)) {
+ ConstantInt *CI = dyn_cast<ConstantInt>(V);
+
+ for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
+ Result.push_back(std::make_pair(CI, *PI));
+ return true;
+ }
+
+ // If V is a non-instruction value, or an instruction in a different block,
+ // then it can't be derived from a PHI.
+ Instruction *I = dyn_cast<Instruction>(V);
+ if (I == 0 || I->getParent() != BB) {
+
+ // Okay, if this is a live-in value, see if it has a known value at the end
+ // of any of our predecessors.
+ //
+ // FIXME: This should be an edge property, not a block end property.
+ /// TODO: Per PR2563, we could infer value range information about a
+ /// predecessor based on its terminator.
+ //
+ if (LVI) {
+ for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
+ // If the value is known by LazyValueInfo to be a constant in a
+ // predecessor, use that information to try to thread this block.
+ Constant *PredCst = LVI->getConstant(V, *PI);
+ if (PredCst == 0 ||
+ (!isa<ConstantInt>(PredCst) && !isa<UndefValue>(PredCst)))
+ continue;
+
+ Result.push_back(std::make_pair(dyn_cast<ConstantInt>(PredCst), *PI));
+ }
+
+ return !Result.empty();
+ }
+
+ return false;
+ }
+
+ /// If I is a PHI node, then we know the incoming values for any constants.
+ if (PHINode *PN = dyn_cast<PHINode>(I)) {
+ for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
+ Value *InVal = PN->getIncomingValue(i);
+ if (isa<ConstantInt>(InVal) || isa<UndefValue>(InVal)) {
+ ConstantInt *CI = dyn_cast<ConstantInt>(InVal);
+ Result.push_back(std::make_pair(CI, PN->getIncomingBlock(i)));
+ }
+ }
+ return !Result.empty();
+ }
+
+ SmallVector<std::pair<ConstantInt*, BasicBlock*>, 8> LHSVals, RHSVals;
+
+ // Handle some boolean conditions.
+ if (I->getType()->getPrimitiveSizeInBits() == 1) {
+ // X | true -> true
+ // X & false -> false
+ if (I->getOpcode() == Instruction::Or ||
+ I->getOpcode() == Instruction::And) {
+ ComputeValueKnownInPredecessors(I->getOperand(0), BB, LHSVals);
+ ComputeValueKnownInPredecessors(I->getOperand(1), BB, RHSVals);
+
+ if (LHSVals.empty() && RHSVals.empty())
+ return false;
+
+ ConstantInt *InterestingVal;
+ if (I->getOpcode() == Instruction::Or)
+ InterestingVal = ConstantInt::getTrue(I->getContext());
+ else
+ InterestingVal = ConstantInt::getFalse(I->getContext());
+
+ // Scan for the sentinel.
+ for (unsigned i = 0, e = LHSVals.size(); i != e; ++i)
+ if (LHSVals[i].first == InterestingVal || LHSVals[i].first == 0)
+ Result.push_back(LHSVals[i]);
+ for (unsigned i = 0, e = RHSVals.size(); i != e; ++i)
+ if (RHSVals[i].first == InterestingVal || RHSVals[i].first == 0)
+ Result.push_back(RHSVals[i]);
+ return !Result.empty();
+ }
+
+ // Handle the NOT form of XOR.
+ if (I->getOpcode() == Instruction::Xor &&
+ isa<ConstantInt>(I->getOperand(1)) &&
+ cast<ConstantInt>(I->getOperand(1))->isOne()) {
+ ComputeValueKnownInPredecessors(I->getOperand(0), BB, Result);
+ if (Result.empty())
+ return false;
+
+ // Invert the known values.
+ for (unsigned i = 0, e = Result.size(); i != e; ++i)
+ Result[i].first =
+ cast<ConstantInt>(ConstantExpr::getNot(Result[i].first));
+ return true;
+ }
+ }
+
+ // Handle compare with phi operand, where the PHI is defined in this block.
+ if (CmpInst *Cmp = dyn_cast<CmpInst>(I)) {
+ PHINode *PN = dyn_cast<PHINode>(Cmp->getOperand(0));
+ if (PN && PN->getParent() == BB) {
+ // We can do this simplification if any comparisons fold to true or false.
+ // See if any do.
+ for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
+ BasicBlock *PredBB = PN->getIncomingBlock(i);
+ Value *LHS = PN->getIncomingValue(i);
+ Value *RHS = Cmp->getOperand(1)->DoPHITranslation(BB, PredBB);
+
+ Value *Res = SimplifyCmpInst(Cmp->getPredicate(), LHS, RHS);
+ if (Res == 0) continue;
+
+ if (isa<UndefValue>(Res))
+ Result.push_back(std::make_pair((ConstantInt*)0, PredBB));
+ else if (ConstantInt *CI = dyn_cast<ConstantInt>(Res))
+ Result.push_back(std::make_pair(CI, PredBB));
+ }
+
+ return !Result.empty();
+ }
+
+ // TODO: We could also recurse to see if we can determine constants another
+ // way.
+ }
+ return false;
+}
+
+
+
+/// GetBestDestForBranchOnUndef - If we determine that the specified block ends
+/// in an undefined jump, decide which block is best to revector to.
+///
+/// Since we can pick an arbitrary destination, we pick the successor with the
+/// fewest predecessors. This should reduce the in-degree of the others.
+///
+static unsigned GetBestDestForJumpOnUndef(BasicBlock *BB) {
+ 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;
+ }
+
+ return MinSucc;
+}
-/// 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 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.
+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) {
+ // If SinglePred was a loop header, BB becomes one.
+ if (LoopHeaders.erase(SinglePred))
+ LoopHeaders.insert(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;
+ }
+ }
+
+ // Look to see if the terminator is a branch of switch, if not we can't thread
+ // it.
Value *Condition;
if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) {
// Can't thread an unconditional jump.
// 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() << '\n');
++NumFolds;
ConstantFoldTerminator(BB);
return true;
}
- // If there is only a single predecessor of this block, nothing to fold.
- if (BB->getSinglePredecessor())
- return false;
+ // If the terminator is branching on an undef, we can pick any of the
+ // successors to branch to. Let GetBestDestForJumpOnUndef decide.
+ if (isa<UndefValue>(Condition)) {
+ unsigned BestSucc = GetBestDestForJumpOnUndef(BB);
+
+ // Fold the branch/switch.
+ TerminatorInst *BBTerm = BB->getTerminator();
+ for (unsigned i = 0, e = BBTerm->getNumSuccessors(); i != e; ++i) {
+ if (i == BestSucc) continue;
+ RemovePredecessorAndSimplify(BBTerm->getSuccessor(i), BB, TD);
+ }
+
+ DEBUG(errs() << " In block '" << BB->getName()
+ << "' folding undef terminator: " << *BBTerm << '\n');
+ BranchInst::Create(BBTerm->getSuccessor(BestSucc), 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;
+ }
+ }
+
+ // 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 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()) &&
- ProcessBranchOnLogical(CondI, BB,
- CondI->getOpcode() == Instruction::And))
- return true;
+ if (PHINode *PN = dyn_cast<PHINode>(CondInst))
+ if (PN->getParent() == BB)
+ return ProcessJumpOnPHI(PN);
+
+ if (CmpInst *CondCmp = dyn_cast<CmpInst>(CondInst)) {
+ if (!isa<PHINode>(CondCmp->getOperand(0)) ||
+ cast<PHINode>(CondCmp->getOperand(0))->getParent() != BB) {
+ // If we have a comparison, loop over the predecessors to see if there is
+ // a condition with a lexically identical 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
+ // 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 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))
+ if (LoadInst *LI = dyn_cast<LoadInst>(SimplifyValue))
+ if (SimplifyPartiallyRedundantLoad(LI))
return true;
+
+ // Handle a variety of cases where we are branching on something derived from
+ // a PHI node in the current block. If we can prove that any predecessors
+ // compute a predictable value based on a PHI node, thread those predecessors.
+ //
+ if (ProcessThreadableEdges(CondInst, BB))
+ 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;
}
-/// 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.
- ConstantInt *PredCst = 0;
- 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 (PredCst == 0)
- 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());
- // 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()
- << "' - Cost is too high: " << JumpThreadCost << "\n";
- return false;
+ // 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 {
+ DEBUG(errs() << " In block '" << PredBB->getName()
+ << "' folding terminator: " << *PredBB->getTerminator() << '\n');
+ ++NumFolds;
+ ConstantFoldTerminator(PredBB);
+ return true;
}
-
- // If so, we can actually do this threading. Merge any common predecessors
- // that will act the same.
- BasicBlock *PredBB = FactorCommonPHIPreds(PN, PredCst);
+
+ 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()) {
+ DEBUG(errs() << " In block '" << BB->getName()
+ << "' folding condition to '" << BranchDir << "': "
+ << *BB->getTerminator() << '\n');
+ ++NumFolds;
+ Value *OldCond = DestBI->getCondition();
+ DestBI->setCondition(ConstantInt::get(Type::getInt1Ty(BB->getContext()),
+ BranchDir));
+ ConstantFoldTerminator(BB);
+ RecursivelyDeleteTriviallyDeadInstructions(OldCond);
+ return true;
+ }
+
// 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());
- else {
- SwitchInst *SI = cast<SwitchInst>(BB->getTerminator());
- SuccBB = SI->getSuccessor(SI->findCaseValue(PredCst));
- }
+ 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;
- }
+ SmallVector<BasicBlock*, 2> Preds;
+ Preds.push_back(PredBB);
- // 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;
+ // Ok, try to thread it!
+ return ThreadEdge(BB, Preds, SuccBB);
}
-/// ProcessJumpOnLogicalPHI - PN's basic block contains a conditional branch
-/// whose condition is an AND/OR where one side is PN. If PN has constant
-/// operands that permit us to evaluate the condition for some operand, thread
-/// through the block. For example with:
-/// br (and X, phi(Y, Z, false))
-/// the predecessor corresponding to the 'false' will always jump to the false
-/// destination of the branch.
+/// 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 ]
///
-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;
- }
+/// Optimizing switches like this is very important, because simplifycfg builds
+/// switches out of repeated 'if' conditions.
+bool JumpThreading::ProcessSwitchOnDuplicateCond(BasicBlock *PredBB,
+ BasicBlock *DestBB) {
+ // Can't thread edge to self.
+ if (PredBB == DestBB)
+ return false;
+
+ 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. Skip debug info first.
+ BasicBlock::iterator BBI = DestBB->begin();
+ while (isa<DbgInfoIntrinsic>(BBI))
+ BBI++;
+
+ // FIXME: Thread if it just contains a PHI.
+ if (isa<SwitchInst>(BBI)) {
+ 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);
- // 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;
- ConstantInt *PredCst = ConstantInt::get(Type::Int1Ty, !isAnd);
- for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
- if (PN->getIncomingValue(i) == PredCst) {
- PredNo = i;
- break;
+ // 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.
+ 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 (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;
}
- // If no match, bail out.
- if (PredNo == ~0U)
+ 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;
- // 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";
+ 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";
+
+ // 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());
+ 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;
- // If so, we can actually do this threading. Merge any common predecessors
- // that will act the same.
- BasicBlock *PredBB = FactorCommonPHIPreds(PN, PredCst);
+ 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);
+ }
- // Next, figure out which successor we are threading to. If this was an AND,
- // the constant must be FALSE, and we must be targeting the 'false' block.
- // If this is an OR, the constant must be TRUE, and we must be targeting the
- // 'true' block.
- BasicBlock *SuccBB = BB->getTerminator()->getSuccessor(isAnd);
+ // 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));
+ }
- // 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;
+ // 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);
}
- // And finally, do it!
- DOUT << " Threading edge through bool from '" << PredBB->getNameStart()
- << "' to '" << SuccBB->getNameStart() << "' with cost: "
- << JumpThreadCost << ", across block:\n "
- << *BB << "\n";
+ //cerr << "PRE: " << *LI << *PN << "\n";
+
+ LI->replaceAllUsesWith(PN);
+ LI->eraseFromParent();
- 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));
+/// FindMostPopularDest - The specified list contains multiple possible
+/// threadable destinations. Pick the one that occurs the most frequently in
+/// the list.
+static BasicBlock *
+FindMostPopularDest(BasicBlock *BB,
+ const SmallVectorImpl<std::pair<BasicBlock*,
+ BasicBlock*> > &PredToDestList) {
+ assert(!PredToDestList.empty());
- // 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;
+ // Determine popularity. If there are multiple possible destinations, we
+ // explicitly choose to ignore 'undef' destinations. We prefer to thread
+ // blocks with known and real destinations to threading undef. We'll handle
+ // them later if interesting.
+ DenseMap<BasicBlock*, unsigned> DestPopularity;
+ for (unsigned i = 0, e = PredToDestList.size(); i != e; ++i)
+ if (PredToDestList[i].second)
+ DestPopularity[PredToDestList[i].second]++;
- // 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;
+ // Find the most popular dest.
+ DenseMap<BasicBlock*, unsigned>::iterator DPI = DestPopularity.begin();
+ BasicBlock *MostPopularDest = DPI->first;
+ unsigned Popularity = DPI->second;
+ SmallVector<BasicBlock*, 4> SamePopularity;
+
+ for (++DPI; DPI != DestPopularity.end(); ++DPI) {
+ // If the popularity of this entry isn't higher than the popularity we've
+ // seen so far, ignore it.
+ if (DPI->second < Popularity)
+ ; // ignore.
+ else if (DPI->second == Popularity) {
+ // If it is the same as what we've seen so far, keep track of it.
+ SamePopularity.push_back(DPI->first);
+ } else {
+ // If it is more popular, remember it.
+ SamePopularity.clear();
+ MostPopularDest = DPI->first;
+ Popularity = DPI->second;
+ }
+ }
+
+ // Okay, now we know the most popular destination. If there is more than
+ // destination, we need to determine one. This is arbitrary, but we need
+ // to make a deterministic decision. Pick the first one that appears in the
+ // successor list.
+ if (!SamePopularity.empty()) {
+ SamePopularity.push_back(MostPopularDest);
+ TerminatorInst *TI = BB->getTerminator();
+ for (unsigned i = 0; ; ++i) {
+ assert(i != TI->getNumSuccessors() && "Didn't find any successor!");
+
+ if (std::find(SamePopularity.begin(), SamePopularity.end(),
+ TI->getSuccessor(i)) == SamePopularity.end())
+ continue;
+
+ MostPopularDest = TI->getSuccessor(i);
break;
}
-
- // Otherwise, we can't fold this input.
- PredCst = 0;
}
- // If no match, bail out.
- if (PredCst == 0)
+ // Okay, we have finally picked the most popular destination.
+ return MostPopularDest;
+}
+
+bool JumpThreading::ProcessThreadableEdges(Instruction *CondInst,
+ BasicBlock *BB) {
+ // If threading this would thread across a loop header, don't even try to
+ // thread the edge.
+ if (LoopHeaders.count(BB))
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";
+ SmallVector<std::pair<ConstantInt*, BasicBlock*>, 8> PredValues;
+ if (!ComputeValueKnownInPredecessors(CondInst, BB, PredValues))
return false;
- }
+ assert(!PredValues.empty() &&
+ "ComputeValueKnownInPredecessors returned true with no values");
+
+ DEBUG(errs() << "IN BB: " << *BB;
+ for (unsigned i = 0, e = PredValues.size(); i != e; ++i) {
+ errs() << " BB '" << BB->getName() << "': FOUND condition = ";
+ if (PredValues[i].first)
+ errs() << *PredValues[i].first;
+ else
+ errs() << "UNDEF";
+ errs() << " for pred '" << PredValues[i].second->getName()
+ << "'.\n";
+ });
- // If so, we can actually do this threading. Merge any common predecessors
- // that will act the same.
- BasicBlock *PredBB = FactorCommonPHIPreds(PN, PredCst);
+ // Decide what we want to thread through. Convert our list of known values to
+ // a list of known destinations for each pred. This also discards duplicate
+ // predecessors and keeps track of the undefined inputs (which are represented
+ // as a null dest in the PredToDestList).
+ SmallPtrSet<BasicBlock*, 16> SeenPreds;
+ SmallVector<std::pair<BasicBlock*, BasicBlock*>, 16> PredToDestList;
- // Next, get our successor.
- BasicBlock *SuccBB = BB->getTerminator()->getSuccessor(!TrueDirection);
+ BasicBlock *OnlyDest = 0;
+ BasicBlock *MultipleDestSentinel = (BasicBlock*)(intptr_t)~0ULL;
- // 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;
+ for (unsigned i = 0, e = PredValues.size(); i != e; ++i) {
+ BasicBlock *Pred = PredValues[i].second;
+ if (!SeenPreds.insert(Pred))
+ continue; // Duplicate predecessor entry.
+
+ // If the predecessor ends with an indirect goto, we can't change its
+ // destination.
+ if (isa<IndirectBrInst>(Pred->getTerminator()))
+ continue;
+
+ ConstantInt *Val = PredValues[i].first;
+
+ BasicBlock *DestBB;
+ if (Val == 0) // Undef.
+ DestBB = 0;
+ else if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator()))
+ DestBB = BI->getSuccessor(Val->isZero());
+ else {
+ SwitchInst *SI = cast<SwitchInst>(BB->getTerminator());
+ DestBB = SI->getSuccessor(SI->findCaseValue(Val));
+ }
+
+ // If we have exactly one destination, remember it for efficiency below.
+ if (i == 0)
+ OnlyDest = DestBB;
+ else if (OnlyDest != DestBB)
+ OnlyDest = MultipleDestSentinel;
+
+ PredToDestList.push_back(std::make_pair(Pred, DestBB));
}
+ // If all edges were unthreadable, we fail.
+ if (PredToDestList.empty())
+ 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";
+ // Determine which is the most common successor. If we have many inputs and
+ // this block is a switch, we want to start by threading the batch that goes
+ // to the most popular destination first. If we only know about one
+ // threadable destination (the common case) we can avoid this.
+ BasicBlock *MostPopularDest = OnlyDest;
- ThreadEdge(BB, PredBB, SuccBB);
- ++NumThreads;
- return true;
+ if (MostPopularDest == MultipleDestSentinel)
+ MostPopularDest = FindMostPopularDest(BB, PredToDestList);
+
+ // Now that we know what the most popular destination is, factor all
+ // predecessors that will jump to it into a single predecessor.
+ SmallVector<BasicBlock*, 16> PredsToFactor;
+ for (unsigned i = 0, e = PredToDestList.size(); i != e; ++i)
+ if (PredToDestList[i].second == MostPopularDest) {
+ BasicBlock *Pred = PredToDestList[i].first;
+
+ // This predecessor may be a switch or something else that has multiple
+ // edges to the block. Factor each of these edges by listing them
+ // according to # occurrences in PredsToFactor.
+ TerminatorInst *PredTI = Pred->getTerminator();
+ for (unsigned i = 0, e = PredTI->getNumSuccessors(); i != e; ++i)
+ if (PredTI->getSuccessor(i) == BB)
+ PredsToFactor.push_back(Pred);
+ }
+
+ // If the threadable edges are branching on an undefined value, we get to pick
+ // the destination that these predecessors should get to.
+ if (MostPopularDest == 0)
+ MostPopularDest = BB->getTerminator()->
+ getSuccessor(GetBestDestForJumpOnUndef(BB));
+
+ // Ok, try to thread it!
+ return ThreadEdge(BB, PredsToFactor, MostPopularDest);
}
+/// ProcessJumpOnPHI - We have a conditional branch or 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) {
+ BasicBlock *BB = PN->getParent();
+
+ // If any of the predecessor blocks end in an unconditional branch, we can
+ // *duplicate* the jump into that block in order to further encourage jump
+ // threading and to eliminate cases where we have branch on a phi of an icmp
+ // (branch on icmp is much better).
-/// ThreadEdge - We have decided that it is safe and profitable to thread an
-/// edge from PredBB to SuccBB across BB. Transform the IR to reflect this
-/// change.
-void JumpThreading::ThreadEdge(BasicBlock *BB, BasicBlock *PredBB,
- BasicBlock *SuccBB) {
+ // We don't want to do this tranformation for switches, because we don't
+ // really want to duplicate a switch.
+ if (isa<SwitchInst>(BB->getTerminator()))
+ return false;
+
+ // Look for unconditional branch predecessors.
+ for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
+ BasicBlock *PredBB = PN->getIncomingBlock(i);
+ if (BranchInst *PredBr = dyn_cast<BranchInst>(PredBB->getTerminator()))
+ if (PredBr->isUnconditional() &&
+ // Try to duplicate BB into PredBB.
+ DuplicateCondBranchOnPHIIntoPred(BB, PredBB))
+ return true;
+ }
- // 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))
- continue;
+ return false;
+}
+
+
+/// AddPHINodeEntriesForMappedBlock - We're adding 'NewPred' as a new
+/// predecessor to the PHIBB block. If it has PHI nodes, add entries for
+/// NewPred using the entries from OldPred (suitably mapped).
+static void AddPHINodeEntriesForMappedBlock(BasicBlock *PHIBB,
+ BasicBlock *OldPred,
+ BasicBlock *NewPred,
+ DenseMap<Instruction*, Value*> &ValueMap) {
+ for (BasicBlock::iterator PNI = PHIBB->begin();
+ PHINode *PN = dyn_cast<PHINode>(PNI); ++PNI) {
+ // Ok, we have a PHI node. Figure out what the incoming value was for the
+ // DestBlock.
+ Value *IV = PN->getIncomingValueForBlock(OldPred);
- // 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;
+ // Remap the value if necessary.
+ if (Instruction *Inst = dyn_cast<Instruction>(IV)) {
+ DenseMap<Instruction*, Value*>::iterator I = ValueMap.find(Inst);
+ if (I != ValueMap.end())
+ IV = I->second;
}
- // 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);
+ PN->addIncoming(IV, NewPred);
}
-
+}
+
+/// ThreadEdge - We have decided that it is safe and profitable to factor the
+/// blocks in PredBBs to one predecessor, then thread an edge from it to SuccBB
+/// across BB. Transform the IR to reflect this change.
+bool JumpThreading::ThreadEdge(BasicBlock *BB,
+ const SmallVectorImpl<BasicBlock*> &PredBBs,
+ BasicBlock *SuccBB) {
+ // If threading to the same block as we come from, we would infinite loop.
+ if (SuccBB == BB) {
+ 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)) {
+ DEBUG(errs() << " Not threading across loop header BB '" << BB->getName()
+ << "' to dest BB '" << SuccBB->getName()
+ << "' - it might create an irreducible loop!\n");
+ return false;
+ }
+
+ unsigned JumpThreadCost = getJumpThreadDuplicationCost(BB);
+ if (JumpThreadCost > Threshold) {
+ DEBUG(errs() << " Not threading BB '" << BB->getName()
+ << "' - Cost is too high: " << JumpThreadCost << "\n");
+ return false;
+ }
+
+ // And finally, do it! Start by factoring the predecessors is needed.
+ BasicBlock *PredBB;
+ if (PredBBs.size() == 1)
+ PredBB = PredBBs[0];
+ else {
+ DEBUG(errs() << " Factoring out " << PredBBs.size()
+ << " common predecessors.\n");
+ PredBB = SplitBlockPredecessors(BB, &PredBBs[0], PredBBs.size(),
+ ".thr_comm", this);
+ }
+
+ // And finally, do it!
+ DEBUG(errs() << " Threading edge from '" << PredBB->getName() << "' to '"
+ << SuccBB->getName() << "' with cost: " << JumpThreadCost
+ << ", across block:\n "
+ << *BB << "\n");
+
// 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
// account for entry from PredBB.
DenseMap<Instruction*, Value*> ValueMapping;
- BasicBlock *NewBB =
- BasicBlock::Create(BB->getName()+".thread", BB->getParent(), BB);
+ BasicBlock *NewBB = BasicBlock::Create(BB->getContext(),
+ BB->getName()+".thread",
+ BB->getParent(), BB);
NewBB->moveAfter(PredBB);
BasicBlock::iterator BI = BB->begin();
// mapping and using it to remap operands in the cloned instructions.
for (; !isa<TerminatorInst>(BI); ++BI) {
Instruction *New = BI->clone();
- New->setName(BI->getNameStart());
+ 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
// Check to see if SuccBB has PHI nodes. If so, we need to add entries to the
// PHI nodes for NewBB now.
- for (BasicBlock::iterator PNI = SuccBB->begin(); isa<PHINode>(PNI); ++PNI) {
- PHINode *PN = cast<PHINode>(PNI);
- // Ok, we have a PHI node. Figure out what the incoming value was for the
- // DestBlock.
- Value *IV = PN->getIncomingValueForBlock(BB);
+ AddPHINodeEntriesForMappedBlock(SuccBB, BB, NewBB, ValueMapping);
+
+ // If there were values defined in BB that are used outside the block, then we
+ // now have to update all uses of the value to use either the original value,
+ // the cloned value, or some PHI derived value. This can require arbitrary
+ // PHI insertion, of which we are prepared to do, clean these up now.
+ SSAUpdater SSAUpdate;
+ SmallVector<Use*, 16> UsesToRename;
+ for (BasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) {
+ // Scan all uses of this instruction to see if it is used outside of its
+ // block, and if so, record them in UsesToRename.
+ for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E;
+ ++UI) {
+ Instruction *User = cast<Instruction>(*UI);
+ if (PHINode *UserPN = dyn_cast<PHINode>(User)) {
+ if (UserPN->getIncomingBlock(UI) == BB)
+ continue;
+ } else if (User->getParent() == BB)
+ continue;
+
+ UsesToRename.push_back(&UI.getUse());
+ }
- // Remap the value if necessary.
- if (Instruction *Inst = dyn_cast<Instruction>(IV))
- if (Value *MappedIV = ValueMapping[Inst])
- IV = MappedIV;
- PN->addIncoming(IV, NewBB);
+ // If there are no uses outside the block, we're done with this instruction.
+ if (UsesToRename.empty())
+ continue;
+
+ DEBUG(errs() << "JT: Renaming non-local uses of: " << *I << "\n");
+
+ // We found a use of I outside of BB. Rename all uses of I that are outside
+ // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks
+ // with the two values we know.
+ SSAUpdate.Initialize(I);
+ SSAUpdate.AddAvailableValue(BB, I);
+ SSAUpdate.AddAvailableValue(NewBB, ValueMapping[I]);
+
+ while (!UsesToRename.empty())
+ SSAUpdate.RewriteUse(*UsesToRename.pop_back_val());
+ DEBUG(errs() << "\n");
}
- // 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();
for (unsigned i = 0, e = PredTerm->getNumSuccessors(); i != e; ++i)
if (PredTerm->getSuccessor(i) == BB) {
- BB->removePredecessor(PredBB);
+ RemovePredecessorAndSimplify(BB, PredBB, TD);
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 (Value *V = SimplifyInstruction(Inst, TD)) {
+ WeakVH BIHandle(BI);
+ ReplaceAndSimplifyAllUses(Inst, V, TD);
+ if (BIHandle == 0)
+ BI = NewBB->begin();
+ continue;
+ }
+
+ RecursivelyDeleteTriviallyDeadInstructions(Inst);
+ }
+
+ // Threaded an edge!
+ ++NumThreads;
+ return true;
+}
+
+/// DuplicateCondBranchOnPHIIntoPred - PredBB contains an unconditional branch
+/// to BB which contains an i1 PHI node and a conditional branch on that PHI.
+/// If we can duplicate the contents of BB up into PredBB do so now, this
+/// improves the odds that the branch will be on an analyzable instruction like
+/// a compare.
+bool JumpThreading::DuplicateCondBranchOnPHIIntoPred(BasicBlock *BB,
+ BasicBlock *PredBB) {
+ // If BB is a loop header, then duplicating this block outside the loop would
+ // cause us to transform this into an irreducible loop, don't do this.
+ // See the comments above FindLoopHeaders for justifications and caveats.
+ if (LoopHeaders.count(BB)) {
+ DEBUG(errs() << " Not duplicating loop header '" << BB->getName()
+ << "' into predecessor block '" << PredBB->getName()
+ << "' - it might create an irreducible loop!\n");
+ return false;
+ }
+
+ unsigned DuplicationCost = getJumpThreadDuplicationCost(BB);
+ if (DuplicationCost > Threshold) {
+ DEBUG(errs() << " Not duplicating BB '" << BB->getName()
+ << "' - Cost is too high: " << DuplicationCost << "\n");
+ return false;
+ }
+
+ // Okay, we decided to do this! Clone all the instructions in BB onto the end
+ // of PredBB.
+ DEBUG(errs() << " Duplicating block '" << BB->getName() << "' into end of '"
+ << PredBB->getName() << "' to eliminate branch on phi. Cost: "
+ << DuplicationCost << " block is:" << *BB << "\n");
+
+ // We are going to have to map operands from the original BB block into the
+ // PredBB block. Evaluate PHI nodes in BB.
+ DenseMap<Instruction*, Value*> ValueMapping;
+
+ BasicBlock::iterator BI = BB->begin();
+ for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI)
+ ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB);
+
+ BranchInst *OldPredBranch = cast<BranchInst>(PredBB->getTerminator());
+
+ // Clone the non-phi instructions of BB into PredBB, keeping track of the
+ // mapping and using it to remap operands in the cloned instructions.
+ for (; BI != BB->end(); ++BI) {
+ Instruction *New = BI->clone();
+ New->setName(BI->getName());
+ PredBB->getInstList().insert(OldPredBranch, 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))) {
+ DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst);
+ if (I != ValueMapping.end())
+ New->setOperand(i, I->second);
+ }
+ }
+
+ // Check to see if the targets of the branch had PHI nodes. If so, we need to
+ // add entries to the PHI nodes for branch from PredBB now.
+ BranchInst *BBBranch = cast<BranchInst>(BB->getTerminator());
+ AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(0), BB, PredBB,
+ ValueMapping);
+ AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(1), BB, PredBB,
+ ValueMapping);
+
+ // If there were values defined in BB that are used outside the block, then we
+ // now have to update all uses of the value to use either the original value,
+ // the cloned value, or some PHI derived value. This can require arbitrary
+ // PHI insertion, of which we are prepared to do, clean these up now.
+ SSAUpdater SSAUpdate;
+ SmallVector<Use*, 16> UsesToRename;
+ for (BasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) {
+ // Scan all uses of this instruction to see if it is used outside of its
+ // block, and if so, record them in UsesToRename.
+ for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E;
+ ++UI) {
+ Instruction *User = cast<Instruction>(*UI);
+ if (PHINode *UserPN = dyn_cast<PHINode>(User)) {
+ if (UserPN->getIncomingBlock(UI) == BB)
+ continue;
+ } else if (User->getParent() == BB)
+ continue;
+
+ UsesToRename.push_back(&UI.getUse());
+ }
+
+ // If there are no uses outside the block, we're done with this instruction.
+ if (UsesToRename.empty())
+ continue;
+
+ DEBUG(errs() << "JT: Renaming non-local uses of: " << *I << "\n");
+
+ // We found a use of I outside of BB. Rename all uses of I that are outside
+ // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks
+ // with the two values we know.
+ SSAUpdate.Initialize(I);
+ SSAUpdate.AddAvailableValue(BB, I);
+ SSAUpdate.AddAvailableValue(PredBB, ValueMapping[I]);
+
+ while (!UsesToRename.empty())
+ SSAUpdate.RewriteUse(*UsesToRename.pop_back_val());
+ DEBUG(errs() << "\n");
+ }
+
+ // PredBB no longer jumps to BB, remove entries in the PHI node for the edge
+ // that we nuked.
+ RemovePredecessorAndSimplify(BB, PredBB, TD);
+
+ // Remove the unconditional branch at the end of the PredBB block.
+ OldPredBranch->eraseFromParent();
+
+ ++NumDupes;
+ return true;
}
+
+
projects / oota-llvm.git / blobdiff