//===- 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/Instructions.h"
+#include "llvm/IntrinsicInst.h"
#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>
using namespace llvm;
+STATISTIC(NumEliminated, "Number of unconditional branches eliminated");
+
+static cl::opt<unsigned>
+TailDupThreshold("taildup-threshold",
+ cl::desc("Max block size to tail duplicate"),
+ cl::init(1), 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(&ID) {}
+
private:
- inline bool shouldEliminateUnconditionalBranch(TerminatorInst *TI);
+ inline bool shouldEliminateUnconditionalBranch(TerminatorInst *, unsigned);
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; )
- if (shouldEliminateUnconditionalBranch(I->getTerminator())) {
+ CycleDetector.clear();
+ for (Function::iterator I = F.begin(), E = F.end(); I != E; ) {
+ if (shouldEliminateUnconditionalBranch(I->getTerminator(),
+ TailDupThreshold)) {
eliminateUnconditionalBranch(cast<BranchInst>(I->getTerminator()));
Changed = true;
} else {
++I;
+ CycleDetector.clear();
}
+ }
return Changed;
}
/// 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!
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!
- 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;
- for (unsigned Size = 0; I != Dest->end(); ++Size, ++I)
- if (Size == Threshold) return false; // The block is too large...
+ // 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;
+ }
// Do not tail duplicate a block that has thousands of successors into a block
// with a single successor if the block has many other predecessors. This can
// cause an N^2 explosion in CFG edges (and PHI node entries), as seen in
// cases that have a large number of indirect gotos.
- if (DTI->getNumSuccessors() > 8)
- if (std::distance(PI, PE) * DTI->getNumSuccessors() > 128)
- return false;
+ unsigned NumSuccs = DTI->getNumSuccessors();
+ if (NumSuccs > 8) {
+ unsigned TooMany = 128;
+ if (NumSuccs >= TooMany) return false;
+ TooMany = TooMany/NumSuccs;
+ 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;
+ }
+ }
+
+ // 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;
+ return true;
+}
+
+/// FindObviousSharedDomOf - We know there is a branch from SrcBlock to
+/// DestBlock, and that SrcBlock is not the only predecessor of DstBlock. If we
+/// can find a predecessor of SrcBlock that is a dominator of both SrcBlock and
+/// DstBlock, return it.
+static BasicBlock *FindObviousSharedDomOf(BasicBlock *SrcBlock,
+ BasicBlock *DstBlock) {
+ // SrcBlock must have a single predecessor.
+ pred_iterator PI = pred_begin(SrcBlock), PE = pred_end(SrcBlock);
+ 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 *DstOtherPred = 0;
+ if (*PI == SrcBlock) {
+ if (++PI == PE) return 0;
+ DstOtherPred = *PI;
+ if (++PI != PE) return 0;
+ } else {
+ DstOtherPred = *PI;
+ if (++PI == PE || *PI != SrcBlock || ++PI != PE) return 0;
+ }
+
+ // We can handle two situations here: "if then" and "if then else" blocks. An
+ // 'if then' situation is just where DstOtherPred == SrcPred.
+ if (DstOtherPred == SrcPred)
+ return SrcPred;
+
+ // Check to see if we have an "if then else" situation, which means that
+ // DstOtherPred will have a single predecessor and it will be SrcPred.
+ PI = pred_begin(DstOtherPred); PE = pred_end(DstOtherPred);
+ if (PI != PE && *PI == SrcPred) {
+ if (++PI != PE) return 0; // Not a single pred.
+ return SrcPred; // Otherwise, it's an "if then" situation. Return the if.
+ }
+
+ // Otherwise, this is something we can't handle.
+ return 0;
}
BasicBlock *DestBlock = Branch->getSuccessor(0);
assert(SourceBlock != DestBlock && "Our predicate is broken!");
- DEBUG(std::cerr << "TailDuplication[" << SourceBlock->getParent()->getName()
- << "]: Eliminating branch: " << *Branch);
+ DEBUG(errs() << "TailDuplication[" << SourceBlock->getParent()->getName()
+ << "]: Eliminating branch: " << *Branch);
- // Tail duplication can not update SSA properties correctly if the values
- // defined in the duplicated tail are used outside of the tail itself. For
- // 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;
+ // See if we can avoid duplicating code by moving it up to a dominator of both
+ // blocks.
+ if (BasicBlock *DomBlock = FindObviousSharedDomOf(SourceBlock, DestBlock)) {
+ 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->getFirstNonPHI();
+ while (!isa<TerminatorInst>(BBI)) {
+ Instruction *I = BBI++;
+
+ 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)))
+ if (OpI->getParent() == DestBlock ||
+ (isa<InvokeInst>(OpI) && OpI->getParent() == DomBlock)) {
+ CanHoist = false;
break;
}
-
- } else {
- ShouldDemote = true;
+ if (CanHoist) {
+ // Remove from DestBlock, move right before the term in DomBlock.
+ DestBlock->getInstList().remove(I);
+ DomBlock->getInstList().insert(DomBlock->getTerminator(), I);
+ DOUT << "Hoisted: " << *I;
}
- } 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;
- }
+ // Tail duplication can not update SSA properties correctly if the values
+ // defined in the duplicated tail are used outside of the tail itself. For
+ // 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)
+ 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
// 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.
// 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;
+ 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!
}