//===- TailDuplication.cpp - Simplify CFG through tail duplication --------===//
-//
+//
// The LLVM Compiler Infrastructure
//
-// This file was developed by the LLVM research group and is distributed under
-// the University of Illinois Open Source License. See LICENSE.TXT for details.
-//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
//===----------------------------------------------------------------------===//
//
// This pass performs a limited form of tail duplication, intended to simplify
//
//===----------------------------------------------------------------------===//
+#define DEBUG_TYPE "tailduplicate"
#include "llvm/Transforms/Scalar.h"
#include "llvm/Constant.h"
#include "llvm/Function.h"
#include "llvm/Support/CFG.h"
#include "llvm/Transforms/Utils/Local.h"
#include "llvm/Support/CommandLine.h"
+#include "llvm/Support/Compiler.h"
#include "llvm/Support/Debug.h"
#include "llvm/ADT/Statistic.h"
+#include "llvm/ADT/SmallPtrSet.h"
+#include <map>
using namespace llvm;
+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);
+
namespace {
- cl::opt<unsigned>
- Threshold("taildup-threshold", cl::desc("Max block size to tail duplicate"),
- cl::init(6), cl::Hidden);
- Statistic<> NumEliminated("tailduplicate",
- "Number of unconditional branches eliminated");
- Statistic<> NumPHINodes("tailduplicate", "Number of phi nodes inserted");
-
- class TailDup : public FunctionPass {
+ class VISIBILITY_HIDDEN TailDup : public FunctionPass {
bool runOnFunction(Function &F);
+ public:
+ static char ID; // Pass identification, replacement for typeid
+ TailDup() : FunctionPass((intptr_t)&ID) {}
+
private:
inline bool shouldEliminateUnconditionalBranch(TerminatorInst *TI);
inline void eliminateUnconditionalBranch(BranchInst *BI);
+ SmallPtrSet<BasicBlock*, 4> CycleDetector;
};
- RegisterOpt<TailDup> X("tailduplicate", "Tail Duplication");
}
+char TailDup::ID = 0;
+static RegisterPass<TailDup> X("tailduplicate", "Tail Duplication");
+
// Public interface to the Tail Duplication pass
FunctionPass *llvm::createTailDuplicationPass() { return new TailDup(); }
/// runOnFunction - Top level algorithm - Loop over each unconditional branch in
-/// the function, eliminating it if it looks attractive enough.
-///
+/// the function, eliminating it if it looks attractive enough. CycleDetector
+/// prevents infinite loops by checking that we aren't redirecting a branch to
+/// a place it already pointed to earlier; see PR 2323.
bool TailDup::runOnFunction(Function &F) {
bool Changed = false;
- for (Function::iterator I = F.begin(), E = F.end(); I != E; )
+ CycleDetector.clear();
+ for (Function::iterator I = F.begin(), E = F.end(); I != E; ) {
if (shouldEliminateUnconditionalBranch(I->getTerminator())) {
eliminateUnconditionalBranch(cast<BranchInst>(I->getTerminator()));
Changed = true;
} else {
++I;
+ CycleDetector.clear();
}
+ }
return Changed;
}
if (!DTI->use_empty())
return false;
- // Do not bother working on dead blocks...
- pred_iterator PI = pred_begin(Dest), PE = pred_end(Dest);
- if (PI == PE && Dest != Dest->getParent()->begin())
- return false; // It's just a dead block, ignore it...
-
- // Also, do not bother with blocks with only a single predecessor: simplify
+ // Do not bother with blocks with only a single predecessor: simplify
// CFG will fold these two blocks together!
+ pred_iterator PI = pred_begin(Dest), PE = pred_end(Dest);
++PI;
if (PI == PE) return false; // Exactly one predecessor!
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;
+
// Only count instructions that are not debugger intrinsics.
if (!isa<DbgInfoIntrinsic>(I)) ++Size;
}
for (; PI != PE; ++PI)
if (TooMany-- == 0) return false;
}
+
+ // If this unconditional branch is a fall-through, be careful about
+ // tail duplicating it. In particular, we don't want to taildup it if the
+ // original block will still be there after taildup is completed: doing so
+ // would eliminate the fall-through, requiring unconditional branches.
+ Function::iterator DestI = Dest;
+ if (&*--DestI == BI->getParent()) {
+ // The uncond branch is a fall-through. Tail duplication of the block is
+ // will eliminate the fall-through-ness and end up cloning the terminator
+ // at the end of the Dest block. Since the original Dest block will
+ // continue to exist, this means that one or the other will not be able to
+ // fall through. One typical example that this helps with is code like:
+ // if (a)
+ // foo();
+ // if (b)
+ // foo();
+ // Cloning the 'if b' block into the end of the first foo block is messy.
+
+ // The messy case is when the fall-through block falls through to other
+ // blocks. This is what we would be preventing if we cloned the block.
+ DestI = Dest;
+ if (++DestI != Dest->getParent()->end()) {
+ BasicBlock *DestSucc = DestI;
+ // If any of Dest's successors are fall-throughs, don't do this xform.
+ for (succ_iterator SI = succ_begin(Dest), SE = succ_end(Dest);
+ SI != SE; ++SI)
+ if (*SI == DestSucc)
+ return false;
+ }
+ }
- return true;
+ // Finally, check that we haven't redirected to this target block earlier;
+ // there are cases where we loop forever if we don't check this (PR 2323).
+ if (!CycleDetector.insert(Dest))
+ return false;
+
+ return true;
}
/// FindObviousSharedDomOf - We know there is a branch from SrcBlock to
if (PI == PE || ++PI != PE) return 0;
BasicBlock *SrcPred = *pred_begin(SrcBlock);
-
+
// Look at the predecessors of DstBlock. One of them will be SrcBlock. If
// there is only one other pred, get it, otherwise we can't handle it.
PI = pred_begin(DstBlock); PE = pred_end(DstBlock);
BasicBlock *DestBlock = Branch->getSuccessor(0);
assert(SourceBlock != DestBlock && "Our predicate is broken!");
- DEBUG(std::cerr << "TailDuplication[" << SourceBlock->getParent()->getName()
- << "]: Eliminating branch: " << *Branch);
+ DOUT << "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)) {
- DEBUG(std::cerr << "Found shared dominator: " << DomBlock->getName()
- << "\n");
+ DOUT << "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
while (isa<PHINode>(BBI)) ++BBI;
while (!isa<TerminatorInst>(BBI)) {
Instruction *I = BBI++;
-
+
bool CanHoist = !I->isTrapping() && !I->mayWriteToMemory();
if (CanHoist) {
for (unsigned op = 0, e = I->getNumOperands(); op != e; ++op)
// Remove from DestBlock, move right before the term in DomBlock.
DestBlock->getInstList().remove(I);
DomBlock->getInstList().insert(DomBlock->getTerminator(), I);
- DEBUG(std::cerr << "Hoisted: " << *I);
+ DOUT << "Hoisted: " << *I;
}
}
}
// this reason, we spill all values that are used outside of the tail to the
// stack.
for (BasicBlock::iterator I = DestBlock->begin(); I != DestBlock->end(); ++I)
- for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E;
- ++UI) {
- bool ShouldDemote = false;
- if (cast<Instruction>(*UI)->getParent() != DestBlock) {
- // We must allow our successors to use tail values in their PHI nodes
- // (if the incoming value corresponds to the tail block).
- if (PHINode *PN = dyn_cast<PHINode>(*UI)) {
- for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
- if (PN->getIncomingValue(i) == I &&
- PN->getIncomingBlock(i) != DestBlock) {
- ShouldDemote = true;
- break;
- }
-
- } else {
- ShouldDemote = true;
- }
- } else if (PHINode *PN = dyn_cast<PHINode>(cast<Instruction>(*UI))) {
- // If the user of this instruction is a PHI node in the current block,
- // which has an entry from another block using the value, spill it.
- for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
- if (PN->getIncomingValue(i) == I &&
- PN->getIncomingBlock(i) != DestBlock) {
- ShouldDemote = true;
- break;
- }
- }
-
- if (ShouldDemote) {
- // We found a use outside of the tail. Create a new stack slot to
- // break this inter-block usage pattern.
- DemoteRegToStack(*I);
- break;
- }
+ if (I->isUsedOutsideOfBlock(DestBlock)) {
+ // We found a use outside of the tail. 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 block B to the new
// Ok, we have a PHI node. Figure out what the incoming value was for the
// DestBlock.
Value *IV = PN->getIncomingValueForBlock(DestBlock);
-
+
// Remap the value if necessary...
if (Value *MappedIV = ValueMapping[IV])
IV = MappedIV;