#include "llvm/Function.h"
#include "llvm/Instructions.h"
#include "llvm/Analysis/ConstantFolding.h"
+#include "llvm/Analysis/InlineCost.h"
#include "llvm/Analysis/LoopInfo.h"
#include "llvm/Analysis/LoopPass.h"
#include "llvm/Analysis/Dominators.h"
#include "llvm/Transforms/Utils/BasicBlockUtils.h"
#include "llvm/ADT/Statistic.h"
#include "llvm/ADT/SmallPtrSet.h"
+#include "llvm/ADT/STLExtras.h"
#include "llvm/Support/CommandLine.h"
-#include "llvm/Support/Compiler.h"
#include "llvm/Support/Debug.h"
+#include "llvm/Support/raw_ostream.h"
#include <algorithm>
#include <set>
using namespace llvm;
STATISTIC(NumTrivial , "Number of unswitches that are trivial");
STATISTIC(NumSimplify, "Number of simplifications of unswitched code");
+// The specific value of 50 here was chosen based only on intuition and a
+// few specific examples.
static cl::opt<unsigned>
Threshold("loop-unswitch-threshold", cl::desc("Max loop size to unswitch"),
- cl::init(10), cl::Hidden);
+ cl::init(50), cl::Hidden);
namespace {
- class VISIBILITY_HIDDEN LoopUnswitch : public LoopPass {
+ class LoopUnswitch : public LoopPass {
LoopInfo *LI; // Loop information
LPPassManager *LPM;
private:
+ virtual void releaseMemory() {
+ UnswitchedVals.clear();
+ }
+
/// RemoveLoopFromWorklist - If the specified loop is on the loop worklist,
/// remove it.
void RemoveLoopFromWorklist(Loop *L) {
void SplitExitEdges(Loop *L, const SmallVector<BasicBlock *, 8> &ExitBlocks);
bool UnswitchIfProfitable(Value *LoopCond, Constant *Val);
- unsigned getLoopUnswitchCost(Value *LIC);
void UnswitchTrivialCondition(Loop *L, Value *Cond, Constant *Val,
BasicBlock *ExitBlock);
void UnswitchNontrivialCondition(Value *LIC, Constant *OnVal, Loop *L);
if (isa<Constant>(Cond)) return 0;
// TODO: Handle: br (VARIANT|INVARIANT).
- // TODO: Hoist simple expressions out of loops.
- if (L->isLoopInvariant(Cond)) return Cond;
-
+
+ // Hoist simple values out.
+ if (L->makeLoopInvariant(Cond, Changed))
+ return Cond;
+
if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Cond))
if (BO->getOpcode() == Instruction::And ||
BO->getOpcode() == Instruction::Or) {
bool LoopUnswitch::runOnLoop(Loop *L, LPPassManager &LPM_Ref) {
LI = &getAnalysis<LoopInfo>();
LPM = &LPM_Ref;
- DF = getAnalysisToUpdate<DominanceFrontier>();
- DT = getAnalysisToUpdate<DominatorTree>();
+ DF = getAnalysisIfAvailable<DominanceFrontier>();
+ DT = getAnalysisIfAvailable<DominatorTree>();
currentLoop = L;
Function *F = currentLoop->getHeader()->getParent();
bool Changed = false;
/// and profitable.
bool LoopUnswitch::processCurrentLoop() {
bool Changed = false;
+ LLVMContext &Context = currentLoop->getHeader()->getContext();
// Loop over all of the basic blocks in the loop. If we find an interior
// block that is branching on a loop-invariant condition, we can unswitch this
Value *LoopCond = FindLIVLoopCondition(BI->getCondition(),
currentLoop, Changed);
if (LoopCond && UnswitchIfProfitable(LoopCond,
- ConstantInt::getTrue())) {
+ ConstantInt::getTrue(Context))) {
++NumBranches;
return true;
}
Value *LoopCond = FindLIVLoopCondition(SI->getCondition(),
currentLoop, Changed);
if (LoopCond && UnswitchIfProfitable(LoopCond,
- ConstantInt::getTrue())) {
+ ConstantInt::getTrue(Context))) {
++NumSelects;
return true;
}
// Okay, everything after this looks good, check to make sure that this block
// doesn't include any side effects.
for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
- if (I->mayWriteToMemory())
+ if (I->mayHaveSideEffects())
return false;
return true;
BasicBlock **LoopExit) {
BasicBlock *Header = currentLoop->getHeader();
TerminatorInst *HeaderTerm = Header->getTerminator();
+ LLVMContext &Context = Header->getContext();
BasicBlock *LoopExitBB = 0;
if (BranchInst *BI = dyn_cast<BranchInst>(HeaderTerm)) {
// this.
if ((LoopExitBB = isTrivialLoopExitBlock(currentLoop,
BI->getSuccessor(0)))) {
- if (Val) *Val = ConstantInt::getTrue();
+ if (Val) *Val = ConstantInt::getTrue(Context);
} else if ((LoopExitBB = isTrivialLoopExitBlock(currentLoop,
BI->getSuccessor(1)))) {
- if (Val) *Val = ConstantInt::getFalse();
+ if (Val) *Val = ConstantInt::getFalse(Context);
}
} else if (SwitchInst *SI = dyn_cast<SwitchInst>(HeaderTerm)) {
// If this isn't a switch on Cond, we can't handle it.
// part of the loop that the code *would* execute. We already checked the
// tail, check the header now.
for (BasicBlock::iterator I = Header->begin(), E = Header->end(); I != E; ++I)
- if (I->mayWriteToMemory())
+ if (I->mayHaveSideEffects())
return false;
return true;
}
-/// getLoopUnswitchCost - Return the cost (code size growth) that will happen if
-/// we choose to unswitch current loop on the specified value.
-///
-unsigned LoopUnswitch::getLoopUnswitchCost(Value *LIC) {
- // If the condition is trivial, always unswitch. There is no code growth for
- // this case.
- if (IsTrivialUnswitchCondition(LIC))
- return 0;
-
- // FIXME: This is really overly conservative. However, more liberal
- // estimations have thus far resulted in excessive unswitching, which is bad
- // both in compile time and in code size. This should be replaced once
- // someone figures out how a good estimation.
- return currentLoop->getBlocks().size();
-
- unsigned Cost = 0;
- // FIXME: this is brain dead. It should take into consideration code
- // shrinkage.
- for (Loop::block_iterator I = currentLoop->block_begin(),
- E = currentLoop->block_end();
- I != E; ++I) {
- BasicBlock *BB = *I;
- // Do not include empty blocks in the cost calculation. This happen due to
- // loop canonicalization and will be removed.
- if (BB->begin() == BasicBlock::iterator(BB->getTerminator()))
- continue;
-
- // Count basic blocks.
- ++Cost;
- }
-
- return Cost;
-}
-
/// UnswitchIfProfitable - We have found that we can unswitch currentLoop when
/// LoopCond == Val to simplify the loop. If we decide that this is profitable,
/// unswitch the loop, reprocess the pieces, then return true.
bool LoopUnswitch::UnswitchIfProfitable(Value *LoopCond, Constant *Val){
initLoopData();
- Function *F = loopHeader->getParent();
- // Do not unswitch if the function is optimized for size.
- if (!F->isDeclaration() && F->hasFnAttr(Attribute::OptimizeForSize))
+ // If LoopSimplify was unable to form a preheader, don't do any unswitching.
+ if (!loopPreheader)
return false;
- // Check to see if it would be profitable to unswitch current loop.
- unsigned Cost = getLoopUnswitchCost(LoopCond);
+ Function *F = loopHeader->getParent();
+
+ // If the condition is trivial, always unswitch. There is no code growth for
+ // this case.
+ if (!IsTrivialUnswitchCondition(LoopCond)) {
+ // Check to see if it would be profitable to unswitch current loop.
- // Do not do non-trivial unswitch while optimizing for size.
- if (Cost && OptimizeForSize)
- return false;
+ // Do not do non-trivial unswitch while optimizing for size.
+ if (OptimizeForSize || F->hasFnAttr(Attribute::OptimizeForSize))
+ return false;
- if (Cost > Threshold) {
- // FIXME: this should estimate growth by the amount of code shared by the
- // resultant unswitched loops.
- //
- DOUT << "NOT unswitching loop %"
- << currentLoop->getHeader()->getName() << ", cost too high: "
- << currentLoop->getBlocks().size() << "\n";
- return false;
+ // FIXME: This is overly conservative because it does not take into
+ // consideration code simplification opportunities and code that can
+ // be shared by the resultant unswitched loops.
+ CodeMetrics Metrics;
+ for (Loop::block_iterator I = currentLoop->block_begin(),
+ E = currentLoop->block_end();
+ I != E; ++I)
+ Metrics.analyzeBasicBlock(*I);
+
+ // Limit the number of instructions to avoid causing significant code
+ // expansion, and the number of basic blocks, to avoid loops with
+ // large numbers of branches which cause loop unswitching to go crazy.
+ // This is a very ad-hoc heuristic.
+ if (Metrics.NumInsts > Threshold ||
+ Metrics.NumBlocks * 5 > Threshold ||
+ Metrics.NeverInline) {
+ DEBUG(dbgs() << "NOT unswitching loop %"
+ << currentLoop->getHeader()->getName() << ", cost too high: "
+ << currentLoop->getBlocks().size() << "\n");
+ return false;
+ }
}
Constant *CondVal;
// Insert a conditional branch on LIC to the two preheaders. The original
// code is the true version and the new code is the false version.
Value *BranchVal = LIC;
- if (!isa<ConstantInt>(Val) || Val->getType() != Type::Int1Ty)
- BranchVal = new ICmpInst(ICmpInst::ICMP_EQ, LIC, Val, "tmp", InsertPt);
- else if (Val != ConstantInt::getTrue())
+ if (!isa<ConstantInt>(Val) ||
+ Val->getType() != Type::getInt1Ty(LIC->getContext()))
+ BranchVal = new ICmpInst(InsertPt, ICmpInst::ICMP_EQ, LIC, Val, "tmp");
+ else if (Val != ConstantInt::getTrue(Val->getContext()))
// We want to enter the new loop when the condition is true.
std::swap(TrueDest, FalseDest);
// Insert the new branch.
- BranchInst::Create(TrueDest, FalseDest, BranchVal, InsertPt);
+ BranchInst *BI = BranchInst::Create(TrueDest, FalseDest, BranchVal, InsertPt);
+
+ // If either edge is critical, split it. This helps preserve LoopSimplify
+ // form for enclosing loops.
+ SplitCriticalEdge(BI, 0, this);
+ SplitCriticalEdge(BI, 1, this);
}
/// UnswitchTrivialCondition - Given a loop that has a trivial unswitchable
void LoopUnswitch::UnswitchTrivialCondition(Loop *L, Value *Cond,
Constant *Val,
BasicBlock *ExitBlock) {
- DOUT << "loop-unswitch: Trivial-Unswitch loop %"
- << loopHeader->getName() << " [" << L->getBlocks().size()
- << " blocks] in Function " << L->getHeader()->getParent()->getName()
- << " on cond: " << *Val << " == " << *Cond << "\n";
+ DEBUG(dbgs() << "loop-unswitch: Trivial-Unswitch loop %"
+ << loopHeader->getName() << " [" << L->getBlocks().size()
+ << " blocks] in Function " << L->getHeader()->getParent()->getName()
+ << " on cond: " << *Val << " == " << *Cond << "\n");
// First step, split the preheader, so that we know that there is a safe place
// to insert the conditional branch. We will change loopPreheader to have a
for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) {
BasicBlock *ExitBlock = ExitBlocks[i];
- std::vector<BasicBlock*> Preds(pred_begin(ExitBlock), pred_end(ExitBlock));
-
- for (unsigned j = 0, e = Preds.size(); j != e; ++j) {
- BasicBlock* NewExitBlock = SplitEdge(Preds[j], ExitBlock, this);
- BasicBlock* StartBlock = Preds[j];
- BasicBlock* EndBlock;
- if (NewExitBlock->getSinglePredecessor() == ExitBlock) {
- EndBlock = NewExitBlock;
- NewExitBlock = EndBlock->getSinglePredecessor();;
- } else {
- EndBlock = ExitBlock;
- }
-
- std::set<PHINode*> InsertedPHIs;
- PHINode* OldLCSSA = 0;
- for (BasicBlock::iterator I = EndBlock->begin();
- (OldLCSSA = dyn_cast<PHINode>(I)); ++I) {
- Value* OldValue = OldLCSSA->getIncomingValueForBlock(NewExitBlock);
- PHINode* NewLCSSA = PHINode::Create(OldLCSSA->getType(),
- OldLCSSA->getName() + ".us-lcssa",
- NewExitBlock->getTerminator());
- NewLCSSA->addIncoming(OldValue, StartBlock);
- OldLCSSA->setIncomingValue(OldLCSSA->getBasicBlockIndex(NewExitBlock),
- NewLCSSA);
- InsertedPHIs.insert(NewLCSSA);
- }
-
- BasicBlock::iterator InsertPt = EndBlock->getFirstNonPHI();
- for (BasicBlock::iterator I = NewExitBlock->begin();
- (OldLCSSA = dyn_cast<PHINode>(I)) && InsertedPHIs.count(OldLCSSA) == 0;
- ++I) {
- PHINode *NewLCSSA = PHINode::Create(OldLCSSA->getType(),
- OldLCSSA->getName() + ".us-lcssa",
- InsertPt);
- OldLCSSA->replaceAllUsesWith(NewLCSSA);
- NewLCSSA->addIncoming(OldLCSSA, NewExitBlock);
- }
-
- }
+ SmallVector<BasicBlock *, 4> Preds(pred_begin(ExitBlock),
+ pred_end(ExitBlock));
+ SplitBlockPredecessors(ExitBlock, Preds.data(), Preds.size(),
+ ".us-lcssa", this);
}
-
}
/// UnswitchNontrivialCondition - We determined that the loop is profitable
void LoopUnswitch::UnswitchNontrivialCondition(Value *LIC, Constant *Val,
Loop *L) {
Function *F = loopHeader->getParent();
- DOUT << "loop-unswitch: Unswitching loop %"
- << loopHeader->getName() << " [" << L->getBlocks().size()
- << " blocks] in Function " << F->getName()
- << " when '" << *Val << "' == " << *LIC << "\n";
+ DEBUG(dbgs() << "loop-unswitch: Unswitching loop %"
+ << loopHeader->getName() << " [" << L->getBlocks().size()
+ << " blocks] in Function " << F->getName()
+ << " when '" << *Val << "' == " << *LIC << "\n");
LoopBlocks.clear();
NewBlocks.clear();
static void ReplaceUsesOfWith(Instruction *I, Value *V,
std::vector<Instruction*> &Worklist,
Loop *L, LPPassManager *LPM) {
- DOUT << "Replace with '" << *V << "': " << *I;
+ DEBUG(dbgs() << "Replace with '" << *V << "': " << *I);
// Add uses to the worklist, which may be dead now.
for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
// dominates the latch).
LPM->deleteSimpleAnalysisValue(Pred->getTerminator(), L);
Pred->getTerminator()->eraseFromParent();
- new UnreachableInst(Pred);
+ new UnreachableInst(BB->getContext(), Pred);
// The loop is now broken, remove it from LI.
RemoveLoopFromHierarchy(L);
return;
}
- DOUT << "Nuking dead block: " << *BB;
+ DEBUG(dbgs() << "Nuking dead block: " << *BB);
// Remove the instructions in the basic block from the worklist.
for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
// Anything that uses the instructions in this basic block should have their
// uses replaced with undefs.
- if (!I->use_empty())
+ // If I is not void type then replaceAllUsesWith undef.
+ // This allows ValueHandlers and custom metadata to adjust itself.
+ if (!I->getType()->isVoidTy())
I->replaceAllUsesWith(UndefValue::get(I->getType()));
}
// Remove phi node entries in successors for this block.
TerminatorInst *TI = BB->getTerminator();
- std::vector<BasicBlock*> Succs;
+ SmallVector<BasicBlock*, 4> Succs;
for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i) {
Succs.push_back(TI->getSuccessor(i));
TI->getSuccessor(i)->removePredecessor(BB);
}
// Unique the successors, remove anything with multiple uses.
- std::sort(Succs.begin(), Succs.end());
+ array_pod_sort(Succs.begin(), Succs.end());
Succs.erase(std::unique(Succs.begin(), Succs.end()), Succs.end());
// Remove the basic block, including all of the instructions contained in it.
// selects, switches.
std::vector<User*> Users(LIC->use_begin(), LIC->use_end());
std::vector<Instruction*> Worklist;
+ LLVMContext &Context = Val->getContext();
+
// If we know that LIC == Val, or that LIC == NotVal, just replace uses of LIC
// in the loop with the appropriate one directly.
- if (IsEqual || (isa<ConstantInt>(Val) && Val->getType() == Type::Int1Ty)) {
+ if (IsEqual || (isa<ConstantInt>(Val) &&
+ Val->getType()->isInteger(1))) {
Value *Replacement;
if (IsEqual)
Replacement = Val;
else
- Replacement = ConstantInt::get(Type::Int1Ty,
+ Replacement = ConstantInt::get(Type::getInt1Ty(Val->getContext()),
!cast<ConstantInt>(Val)->getZExtValue());
for (unsigned i = 0, e = Users.size(); i != e; ++i)
if (Instruction *U = cast<Instruction>(Users[i])) {
- if (!L->contains(U->getParent()))
+ if (!L->contains(U))
continue;
U->replaceUsesOfWith(LIC, Replacement);
Worklist.push_back(U);
// can. This case occurs when we unswitch switch statements.
for (unsigned i = 0, e = Users.size(); i != e; ++i)
if (Instruction *U = cast<Instruction>(Users[i])) {
- if (!L->contains(U->getParent()))
+ if (!L->contains(U))
continue;
Worklist.push_back(U);
// FIXME: This is a hack. We need to keep the successor around
// and hooked up so as to preserve the loop structure, because
// trying to update it is complicated. So instead we preserve the
- // loop structure and put the block on an dead code path.
-
- BasicBlock *SISucc = SI->getSuccessor(i);
- BasicBlock* Old = SI->getParent();
- BasicBlock* Split = SplitBlock(Old, SI, this);
-
- Instruction* OldTerm = Old->getTerminator();
- BranchInst::Create(Split, SISucc,
- ConstantInt::getTrue(), OldTerm);
-
- LPM->deleteSimpleAnalysisValue(Old->getTerminator(), L);
- Old->getTerminator()->eraseFromParent();
-
- PHINode *PN;
- for (BasicBlock::iterator II = SISucc->begin();
- (PN = dyn_cast<PHINode>(II)); ++II) {
- Value *InVal = PN->removeIncomingValue(Split, false);
- PN->addIncoming(InVal, Old);
- }
-
- SI->removeCase(i);
+ // loop structure and put the block on a dead code path.
+ BasicBlock *Switch = SI->getParent();
+ SplitEdge(Switch, SI->getSuccessor(i), this);
+ // Compute the successors instead of relying on the return value
+ // of SplitEdge, since it may have split the switch successor
+ // after PHI nodes.
+ BasicBlock *NewSISucc = SI->getSuccessor(i);
+ BasicBlock *OldSISucc = *succ_begin(NewSISucc);
+ // Create an "unreachable" destination.
+ BasicBlock *Abort = BasicBlock::Create(Context, "us-unreachable",
+ Switch->getParent(),
+ OldSISucc);
+ new UnreachableInst(Context, Abort);
+ // Force the new case destination to branch to the "unreachable"
+ // block while maintaining a (dead) CFG edge to the old block.
+ NewSISucc->getTerminator()->eraseFromParent();
+ BranchInst::Create(Abort, OldSISucc,
+ ConstantInt::getTrue(Context), NewSISucc);
+ // Release the PHI operands for this edge.
+ for (BasicBlock::iterator II = NewSISucc->begin();
+ PHINode *PN = dyn_cast<PHINode>(II); ++II)
+ PN->setIncomingValue(PN->getBasicBlockIndex(Switch),
+ UndefValue::get(PN->getType()));
+ // Tell the domtree about the new block. We don't fully update the
+ // domtree here -- instead we force it to do a full recomputation
+ // after the pass is complete -- but we do need to inform it of
+ // new blocks.
+ if (DT)
+ DT->addNewBlock(Abort, NewSISucc);
break;
}
}
SimplifyCode(Worklist, L);
}
-/// SimplifyCode - Okay, now that we have simplified some instructions in the
+/// SimplifyCode - Okay, now that we have simplified some instructions in the
/// loop, walk over it and constant prop, dce, and fold control flow where
/// possible. Note that this is effectively a very simple loop-structure-aware
/// optimizer. During processing of this loop, L could very well be deleted, so
// Simple DCE.
if (isInstructionTriviallyDead(I)) {
- DOUT << "Remove dead instruction '" << *I;
+ DEBUG(dbgs() << "Remove dead instruction '" << *I);
// Add uses to the worklist, which may be dead now.
for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
break;
case Instruction::And:
if (isa<ConstantInt>(I->getOperand(0)) &&
- I->getOperand(0)->getType() == Type::Int1Ty) // constant -> RHS
+ // constant -> RHS
+ I->getOperand(0)->getType()->isInteger(1))
cast<BinaryOperator>(I)->swapOperands();
if (ConstantInt *CB = dyn_cast<ConstantInt>(I->getOperand(1)))
- if (CB->getType() == Type::Int1Ty) {
+ if (CB->getType()->isInteger(1)) {
if (CB->isOne()) // X & 1 -> X
ReplaceUsesOfWith(I, I->getOperand(0), Worklist, L, LPM);
else // X & 0 -> 0
break;
case Instruction::Or:
if (isa<ConstantInt>(I->getOperand(0)) &&
- I->getOperand(0)->getType() == Type::Int1Ty) // constant -> RHS
+ // constant -> RHS
+ I->getOperand(0)->getType()->isInteger(1))
cast<BinaryOperator>(I)->swapOperands();
if (ConstantInt *CB = dyn_cast<ConstantInt>(I->getOperand(1)))
- if (CB->getType() == Type::Int1Ty) {
+ if (CB->getType()->isInteger(1)) {
if (CB->isOne()) // X | 1 -> 1
ReplaceUsesOfWith(I, I->getOperand(1), Worklist, L, LPM);
else // X | 0 -> X
if (!SinglePred) continue; // Nothing to do.
assert(SinglePred == Pred && "CFG broken");
- DOUT << "Merging blocks: " << Pred->getName() << " <- "
- << Succ->getName() << "\n";
+ DEBUG(dbgs() << "Merging blocks: " << Pred->getName() << " <- "
+ << Succ->getName() << "\n");
// Resolve any single entry PHI nodes in Succ.
while (PHINode *PN = dyn_cast<PHINode>(Succ->begin()))
// remove dead blocks.
break; // FIXME: Enable.
- DOUT << "Folded branch: " << *BI;
+ DEBUG(dbgs() << "Folded branch: " << *BI);
BasicBlock *DeadSucc = BI->getSuccessor(CB->getZExtValue());
BasicBlock *LiveSucc = BI->getSuccessor(!CB->getZExtValue());
DeadSucc->removePredecessor(BI->getParent(), true);
projects / oota-llvm.git / blobdiff