#include "llvm/Pass.h"
#include "llvm/Type.h"
#include "llvm/Support/CFG.h"
+#include "llvm/Analysis/ConstantFolding.h"
#include "llvm/Transforms/Utils/Local.h"
#include "llvm/Support/CommandLine.h"
#include "llvm/Support/Compiler.h"
#include "llvm/Support/Debug.h"
+#include "llvm/Support/raw_ostream.h"
#include "llvm/ADT/Statistic.h"
#include "llvm/ADT/SmallPtrSet.h"
#include <map>
STATISTIC(NumEliminated, "Number of unconditional branches eliminated");
static cl::opt<unsigned>
-Threshold("taildup-threshold", cl::desc("Max block size to tail duplicate"),
- cl::init(6), cl::Hidden);
+TailDupThreshold("taildup-threshold",
+ cl::desc("Max block size to tail duplicate"),
+ cl::init(1), cl::Hidden);
namespace {
class VISIBILITY_HIDDEN TailDup : public FunctionPass {
bool runOnFunction(Function &F);
public:
static char ID; // Pass identification, replacement for typeid
- TailDup() : FunctionPass((intptr_t)&ID) {}
+ TailDup() : FunctionPass(&ID) {}
private:
- inline bool shouldEliminateUnconditionalBranch(TerminatorInst *TI);
+ inline bool shouldEliminateUnconditionalBranch(TerminatorInst *, unsigned);
inline void eliminateUnconditionalBranch(BranchInst *BI);
SmallPtrSet<BasicBlock*, 4> CycleDetector;
};
bool Changed = false;
CycleDetector.clear();
for (Function::iterator I = F.begin(), E = F.end(); I != E; ) {
- if (shouldEliminateUnconditionalBranch(I->getTerminator())) {
+ if (shouldEliminateUnconditionalBranch(I->getTerminator(),
+ TailDupThreshold)) {
eliminateUnconditionalBranch(cast<BranchInst>(I->getTerminator()));
Changed = true;
} else {
/// We don't count PHI nodes in the count since they will be removed when the
/// contents of the block are copied over.
///
-bool TailDup::shouldEliminateUnconditionalBranch(TerminatorInst *TI) {
+bool TailDup::shouldEliminateUnconditionalBranch(TerminatorInst *TI,
+ unsigned Threshold) {
BranchInst *BI = dyn_cast<BranchInst>(TI);
if (!BI || !BI->isUnconditional()) return false; // Not an uncond branch!
++PI;
if (PI == PE) return false; // Exactly one predecessor!
- BasicBlock::iterator I = Dest->begin();
- while (isa<PHINode>(*I)) ++I;
+ BasicBlock::iterator I = Dest->getFirstNonPHI();
for (unsigned Size = 0; I != Dest->end(); ++I) {
if (Size == Threshold) return false; // The block is too large.
// Don't tail duplicate call instructions. They are very large compared to
// other instructions.
if (isa<CallInst>(I) || isa<InvokeInst>(I)) return false;
+
+ // Allso alloca and malloc.
+ if (isa<AllocationInst>(I)) return false;
+
+ // Some vector instructions can expand into a number of instructions.
+ if (isa<ShuffleVectorInst>(I) || isa<ExtractElementInst>(I) ||
+ isa<InsertElementInst>(I)) return false;
// Only count instructions that are not debugger intrinsics.
if (!isa<DbgInfoIntrinsic>(I)) ++Size;
BasicBlock *DestBlock = Branch->getSuccessor(0);
assert(SourceBlock != DestBlock && "Our predicate is broken!");
- DOUT << "TailDuplication[" << SourceBlock->getParent()->getName()
- << "]: Eliminating branch: " << *Branch;
+ DEBUG(errs() << "TailDuplication[" << SourceBlock->getParent()->getName()
+ << "]: Eliminating branch: " << *Branch);
// See if we can avoid duplicating code by moving it up to a dominator of both
// blocks.
if (BasicBlock *DomBlock = FindObviousSharedDomOf(SourceBlock, DestBlock)) {
- DOUT << "Found shared dominator: " << DomBlock->getName() << "\n";
+ DEBUG(errs() << "Found shared dominator: " << DomBlock->getName() << "\n");
// If there are non-phi instructions in DestBlock that have no operands
// defined in DestBlock, and if the instruction has no side effects, we can
// move the instruction to DomBlock instead of duplicating it.
- BasicBlock::iterator BBI = DestBlock->begin();
- while (isa<PHINode>(BBI)) ++BBI;
+ BasicBlock::iterator BBI = DestBlock->getFirstNonPHI();
while (!isa<TerminatorInst>(BBI)) {
Instruction *I = BBI++;
- bool CanHoist = !I->isTrapping() && !I->mayWriteToMemory();
+ bool CanHoist = I->isSafeToSpeculativelyExecute() &&
+ !I->mayReadFromMemory();
if (CanHoist) {
for (unsigned op = 0, e = I->getNumOperands(); op != e; ++op)
if (Instruction *OpI = dyn_cast<Instruction>(I->getOperand(op)))
// keeping track of the mapping...
//
for (; BI != DestBlock->end(); ++BI) {
- Instruction *New = BI->clone();
+ Instruction *New = BI->clone(BI->getContext());
New->setName(BI->getName());
SourceBlock->getInstList().push_back(New);
ValueMapping[BI] = New;
//
BI = Branch; ++BI; // Get an iterator to the first new instruction
for (; BI != SourceBlock->end(); ++BI)
- for (unsigned i = 0, e = BI->getNumOperands(); i != e; ++i)
- if (Value *Remapped = ValueMapping[BI->getOperand(i)])
- BI->setOperand(i, Remapped);
+ for (unsigned i = 0, e = BI->getNumOperands(); i != e; ++i) {
+ std::map<Value*, Value*>::const_iterator I =
+ ValueMapping.find(BI->getOperand(i));
+ if (I != ValueMapping.end())
+ BI->setOperand(i, I->second);
+ }
// Next we check to see if any of the successors of DestBlock had PHI nodes.
// If so, we need to add entries to the PHI nodes for SourceBlock now.
Value *IV = PN->getIncomingValueForBlock(DestBlock);
// Remap the value if necessary...
- if (Value *MappedIV = ValueMapping[IV])
- IV = MappedIV;
+ std::map<Value*, Value*>::const_iterator I = ValueMapping.find(IV);
+ if (I != ValueMapping.end())
+ IV = I->second;
PN->addIncoming(IV, SourceBlock);
}
}
// instructions one last time, constant propagating and DCE'ing them, because
// they may not be needed anymore.
//
- if (HadPHINodes)
- while (BI != SourceBlock->end())
- if (!dceInstruction(BI) && !doConstantPropagation(BI))
- ++BI;
+ if (HadPHINodes) {
+ while (BI != SourceBlock->end()) {
+ Instruction *Inst = BI++;
+ if (isInstructionTriviallyDead(Inst))
+ Inst->eraseFromParent();
+ else if (Constant *C = ConstantFoldInstruction(Inst, BI->getContext())) {
+ Inst->replaceAllUsesWith(C);
+ Inst->eraseFromParent();
+ }
+ }
+ }
++NumEliminated; // We just killed a branch!
}
projects / oota-llvm.git / blobdiff