#define DEBUG_TYPE "loop-unswitch"
#include "llvm/Transforms/Scalar.h"
-#include "llvm/Constants.h"
-#include "llvm/DerivedTypes.h"
-#include "llvm/Function.h"
-#include "llvm/Instructions.h"
-#include "llvm/LLVMContext.h"
-#include "llvm/Analysis/ConstantFolding.h"
+#include "llvm/ADT/STLExtras.h"
+#include "llvm/ADT/SmallPtrSet.h"
+#include "llvm/ADT/Statistic.h"
+#include "llvm/Analysis/CodeMetrics.h"
+#include "llvm/Analysis/Dominators.h"
+#include "llvm/Analysis/InstructionSimplify.h"
#include "llvm/Analysis/LoopInfo.h"
#include "llvm/Analysis/LoopPass.h"
-#include "llvm/Analysis/Dominators.h"
-#include "llvm/Transforms/Utils/Cloning.h"
-#include "llvm/Transforms/Utils/Local.h"
-#include "llvm/Transforms/Utils/BasicBlockUtils.h"
-#include "llvm/ADT/Statistic.h"
-#include "llvm/ADT/SmallPtrSet.h"
-#include "llvm/ADT/STLExtras.h"
+#include "llvm/Analysis/ScalarEvolution.h"
+#include "llvm/Analysis/TargetTransformInfo.h"
+#include "llvm/IR/Constants.h"
+#include "llvm/IR/DerivedTypes.h"
+#include "llvm/IR/Function.h"
+#include "llvm/IR/Instructions.h"
#include "llvm/Support/CommandLine.h"
#include "llvm/Support/Debug.h"
#include "llvm/Support/raw_ostream.h"
+#include "llvm/Transforms/Utils/BasicBlockUtils.h"
+#include "llvm/Transforms/Utils/Cloning.h"
+#include "llvm/Transforms/Utils/Local.h"
#include <algorithm>
+#include <map>
#include <set>
using namespace llvm;
STATISTIC(NumSelects , "Number of selects unswitched");
STATISTIC(NumTrivial , "Number of unswitches that are trivial");
STATISTIC(NumSimplify, "Number of simplifications of unswitched code");
+STATISTIC(TotalInsts, "Total number of instructions analyzed");
+// The specific value of 100 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(100), cl::Hidden);
+
namespace {
+
+ class LUAnalysisCache {
+
+ typedef DenseMap<const SwitchInst*, SmallPtrSet<const Value *, 8> >
+ UnswitchedValsMap;
+
+ typedef UnswitchedValsMap::iterator UnswitchedValsIt;
+
+ struct LoopProperties {
+ unsigned CanBeUnswitchedCount;
+ unsigned SizeEstimation;
+ UnswitchedValsMap UnswitchedVals;
+ };
+
+ // Here we use std::map instead of DenseMap, since we need to keep valid
+ // LoopProperties pointer for current loop for better performance.
+ typedef std::map<const Loop*, LoopProperties> LoopPropsMap;
+ typedef LoopPropsMap::iterator LoopPropsMapIt;
+
+ LoopPropsMap LoopsProperties;
+ UnswitchedValsMap* CurLoopInstructions;
+ LoopProperties* CurrentLoopProperties;
+
+ // Max size of code we can produce on remained iterations.
+ unsigned MaxSize;
+
+ public:
+
+ LUAnalysisCache() :
+ CurLoopInstructions(NULL), CurrentLoopProperties(NULL),
+ MaxSize(Threshold)
+ {}
+
+ // Analyze loop. Check its size, calculate is it possible to unswitch
+ // it. Returns true if we can unswitch this loop.
+ bool countLoop(const Loop* L, const TargetTransformInfo &TTI);
+
+ // Clean all data related to given loop.
+ void forgetLoop(const Loop* L);
+
+ // Mark case value as unswitched.
+ // Since SI instruction can be partly unswitched, in order to avoid
+ // extra unswitching in cloned loops keep track all unswitched values.
+ void setUnswitched(const SwitchInst* SI, const Value* V);
+
+ // Check was this case value unswitched before or not.
+ bool isUnswitched(const SwitchInst* SI, const Value* V);
+
+ // Clone all loop-unswitch related loop properties.
+ // Redistribute unswitching quotas.
+ // Note, that new loop data is stored inside the VMap.
+ void cloneData(const Loop* NewLoop, const Loop* OldLoop,
+ const ValueToValueMapTy& VMap);
+ };
+
class LoopUnswitch : public LoopPass {
LoopInfo *LI; // Loop information
LPPassManager *LPM;
// LoopProcessWorklist - Used to check if second loop needs processing
// after RewriteLoopBodyWithConditionConstant rewrites first loop.
std::vector<Loop*> LoopProcessWorklist;
- SmallPtrSet<Value *,8> UnswitchedVals;
-
+
+ LUAnalysisCache BranchesInfo;
+
bool OptimizeForSize;
bool redoLoop;
Loop *currentLoop;
- DominanceFrontier *DF;
DominatorTree *DT;
BasicBlock *loopHeader;
BasicBlock *loopPreheader;
-
+
// LoopBlocks contains all of the basic blocks of the loop, including the
- // preheader of the loop, the body of the loop, and the exit blocks of the
+ // preheader of the loop, the body of the loop, and the exit blocks of the
// loop, in that order.
std::vector<BasicBlock*> LoopBlocks;
// NewBlocks contained cloned copy of basic blocks from LoopBlocks.
public:
static char ID; // Pass ID, replacement for typeid
- explicit LoopUnswitch(bool Os = false) :
- LoopPass(&ID), OptimizeForSize(Os), redoLoop(false),
- currentLoop(NULL), DF(NULL), DT(NULL), loopHeader(NULL),
- loopPreheader(NULL) {}
+ explicit LoopUnswitch(bool Os = false) :
+ LoopPass(ID), OptimizeForSize(Os), redoLoop(false),
+ currentLoop(NULL), DT(NULL), loopHeader(NULL),
+ loopPreheader(NULL) {
+ initializeLoopUnswitchPass(*PassRegistry::getPassRegistry());
+ }
bool runOnLoop(Loop *L, LPPassManager &LPM);
bool processCurrentLoop();
/// This transformation requires natural loop information & requires that
- /// loop preheaders be inserted into the CFG...
+ /// loop preheaders be inserted into the CFG.
///
virtual void getAnalysisUsage(AnalysisUsage &AU) const {
AU.addRequiredID(LoopSimplifyID);
AU.addRequiredID(LCSSAID);
AU.addPreservedID(LCSSAID);
AU.addPreserved<DominatorTree>();
- AU.addPreserved<DominanceFrontier>();
+ AU.addPreserved<ScalarEvolution>();
+ AU.addRequired<TargetTransformInfo>();
}
private:
virtual void releaseMemory() {
- UnswitchedVals.clear();
+ BranchesInfo.forgetLoop(currentLoop);
}
/// RemoveLoopFromWorklist - If the specified loop is on the loop worklist,
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);
Constant *Val, bool isEqual);
void EmitPreheaderBranchOnCondition(Value *LIC, Constant *Val,
- BasicBlock *TrueDest,
+ BasicBlock *TrueDest,
BasicBlock *FalseDest,
Instruction *InsertPt);
};
}
-char LoopUnswitch::ID = 0;
-static RegisterPass<LoopUnswitch> X("loop-unswitch", "Unswitch loops");
-Pass *llvm::createLoopUnswitchPass(bool Os) {
- return new LoopUnswitch(Os);
+// Analyze loop. Check its size, calculate is it possible to unswitch
+// it. Returns true if we can unswitch this loop.
+bool LUAnalysisCache::countLoop(const Loop *L, const TargetTransformInfo &TTI) {
+
+ std::pair<LoopPropsMapIt, bool> InsertRes =
+ LoopsProperties.insert(std::make_pair(L, LoopProperties()));
+
+ LoopProperties& Props = InsertRes.first->second;
+
+ if (InsertRes.second) {
+ // New loop.
+
+ // 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.
+
+ // 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 = L->block_begin(),
+ E = L->block_end();
+ I != E; ++I)
+ Metrics.analyzeBasicBlock(*I, TTI);
+
+ Props.SizeEstimation = std::min(Metrics.NumInsts, Metrics.NumBlocks * 5);
+ Props.CanBeUnswitchedCount = MaxSize / (Props.SizeEstimation);
+ MaxSize -= Props.SizeEstimation * Props.CanBeUnswitchedCount;
+
+ if (Metrics.notDuplicatable) {
+ DEBUG(dbgs() << "NOT unswitching loop %"
+ << L->getHeader()->getName() << ", contents cannot be "
+ << "duplicated!\n");
+ return false;
+ }
+ }
+
+ if (!Props.CanBeUnswitchedCount) {
+ DEBUG(dbgs() << "NOT unswitching loop %"
+ << L->getHeader()->getName() << ", cost too high: "
+ << L->getBlocks().size() << "\n");
+
+ return false;
+ }
+
+ // Be careful. This links are good only before new loop addition.
+ CurrentLoopProperties = &Props;
+ CurLoopInstructions = &Props.UnswitchedVals;
+
+ return true;
+}
+
+// Clean all data related to given loop.
+void LUAnalysisCache::forgetLoop(const Loop* L) {
+
+ LoopPropsMapIt LIt = LoopsProperties.find(L);
+
+ if (LIt != LoopsProperties.end()) {
+ LoopProperties& Props = LIt->second;
+ MaxSize += Props.CanBeUnswitchedCount * Props.SizeEstimation;
+ LoopsProperties.erase(LIt);
+ }
+
+ CurrentLoopProperties = NULL;
+ CurLoopInstructions = NULL;
+}
+
+// Mark case value as unswitched.
+// Since SI instruction can be partly unswitched, in order to avoid
+// extra unswitching in cloned loops keep track all unswitched values.
+void LUAnalysisCache::setUnswitched(const SwitchInst* SI, const Value* V) {
+ (*CurLoopInstructions)[SI].insert(V);
+}
+
+// Check was this case value unswitched before or not.
+bool LUAnalysisCache::isUnswitched(const SwitchInst* SI, const Value* V) {
+ return (*CurLoopInstructions)[SI].count(V);
+}
+
+// Clone all loop-unswitch related loop properties.
+// Redistribute unswitching quotas.
+// Note, that new loop data is stored inside the VMap.
+void LUAnalysisCache::cloneData(const Loop* NewLoop, const Loop* OldLoop,
+ const ValueToValueMapTy& VMap) {
+
+ LoopProperties& NewLoopProps = LoopsProperties[NewLoop];
+ LoopProperties& OldLoopProps = *CurrentLoopProperties;
+ UnswitchedValsMap& Insts = OldLoopProps.UnswitchedVals;
+
+ // Reallocate "can-be-unswitched quota"
+
+ --OldLoopProps.CanBeUnswitchedCount;
+ unsigned Quota = OldLoopProps.CanBeUnswitchedCount;
+ NewLoopProps.CanBeUnswitchedCount = Quota / 2;
+ OldLoopProps.CanBeUnswitchedCount = Quota - Quota / 2;
+
+ NewLoopProps.SizeEstimation = OldLoopProps.SizeEstimation;
+
+ // Clone unswitched values info:
+ // for new loop switches we clone info about values that was
+ // already unswitched and has redundant successors.
+ for (UnswitchedValsIt I = Insts.begin(); I != Insts.end(); ++I) {
+ const SwitchInst* OldInst = I->first;
+ Value* NewI = VMap.lookup(OldInst);
+ const SwitchInst* NewInst = cast_or_null<SwitchInst>(NewI);
+ assert(NewInst && "All instructions that are in SrcBB must be in VMap.");
+
+ NewLoopProps.UnswitchedVals[NewInst] = OldLoopProps.UnswitchedVals[OldInst];
+ }
+}
+
+char LoopUnswitch::ID = 0;
+INITIALIZE_PASS_BEGIN(LoopUnswitch, "loop-unswitch", "Unswitch loops",
+ false, false)
+INITIALIZE_AG_DEPENDENCY(TargetTransformInfo)
+INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
+INITIALIZE_PASS_DEPENDENCY(LoopInfo)
+INITIALIZE_PASS_DEPENDENCY(LCSSA)
+INITIALIZE_PASS_END(LoopUnswitch, "loop-unswitch", "Unswitch loops",
+ false, false)
+
+Pass *llvm::createLoopUnswitchPass(bool Os) {
+ return new LoopUnswitch(Os);
}
/// FindLIVLoopCondition - Cond is a condition that occurs in L. If it is
/// invariant in the loop, or has an invariant piece, return the invariant.
/// Otherwise, return null.
static Value *FindLIVLoopCondition(Value *Cond, Loop *L, bool &Changed) {
+
+ // We started analyze new instruction, increment scanned instructions counter.
+ ++TotalInsts;
+
+ // We can never unswitch on vector conditions.
+ if (Cond->getType()->isVectorTy())
+ return 0;
+
// Constants should be folded, not unswitched on!
if (isa<Constant>(Cond)) return 0;
if (Value *RHS = FindLIVLoopCondition(BO->getOperand(1), L, Changed))
return RHS;
}
-
+
return 0;
}
bool LoopUnswitch::runOnLoop(Loop *L, LPPassManager &LPM_Ref) {
LI = &getAnalysis<LoopInfo>();
LPM = &LPM_Ref;
- DF = getAnalysisIfAvailable<DominanceFrontier>();
DT = getAnalysisIfAvailable<DominatorTree>();
currentLoop = L;
Function *F = currentLoop->getHeader()->getParent();
bool Changed = false;
do {
- assert(currentLoop->isLCSSAForm());
+ assert(currentLoop->isLCSSAForm(*DT));
redoLoop = false;
Changed |= processCurrentLoop();
} while(redoLoop);
// FIXME: Reconstruct dom info, because it is not preserved properly.
if (DT)
DT->runOnFunction(*F);
- if (DF)
- DF->runOnFunction(*F);
}
return Changed;
}
-/// processCurrentLoop - Do actual work and unswitch loop if possible
+/// processCurrentLoop - Do actual work and unswitch loop if possible
/// and profitable.
bool LoopUnswitch::processCurrentLoop() {
bool Changed = false;
- LLVMContext &Context = currentLoop->getHeader()->getContext();
+
+ initLoopData();
+
+ // If LoopSimplify was unable to form a preheader, don't do any unswitching.
+ if (!loopPreheader)
+ return false;
+
+ // Loops with indirectbr cannot be cloned.
+ if (!currentLoop->isSafeToClone())
+ return false;
+
+ // Without dedicated exits, splitting the exit edge may fail.
+ if (!currentLoop->hasDedicatedExits())
+ return false;
+
+ LLVMContext &Context = loopHeader->getContext();
+
+ // Probably we reach the quota of branches for this loop. If so
+ // stop unswitching.
+ if (!BranchesInfo.countLoop(currentLoop, getAnalysis<TargetTransformInfo>()))
+ return false;
// 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
// loop.
- for (Loop::block_iterator I = currentLoop->block_begin(),
- E = currentLoop->block_end();
- I != E; ++I) {
+ for (Loop::block_iterator I = currentLoop->block_begin(),
+ E = currentLoop->block_end(); I != E; ++I) {
TerminatorInst *TI = (*I)->getTerminator();
if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
// If this isn't branching on an invariant condition, we can't unswitch
if (BI->isConditional()) {
// See if this, or some part of it, is loop invariant. If so, we can
// unswitch on it if we desire.
- Value *LoopCond = FindLIVLoopCondition(BI->getCondition(),
+ Value *LoopCond = FindLIVLoopCondition(BI->getCondition(),
currentLoop, Changed);
- if (LoopCond && UnswitchIfProfitable(LoopCond,
+ if (LoopCond && UnswitchIfProfitable(LoopCond,
ConstantInt::getTrue(Context))) {
++NumBranches;
return true;
}
- }
+ }
} else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
- Value *LoopCond = FindLIVLoopCondition(SI->getCondition(),
+ Value *LoopCond = FindLIVLoopCondition(SI->getCondition(),
currentLoop, Changed);
- if (LoopCond && SI->getNumCases() > 1) {
+ unsigned NumCases = SI->getNumCases();
+ if (LoopCond && NumCases) {
// Find a value to unswitch on:
// FIXME: this should chose the most expensive case!
- Constant *UnswitchVal = SI->getCaseValue(1);
+ // FIXME: scan for a case with a non-critical edge?
+ Constant *UnswitchVal = NULL;
+
// Do not process same value again and again.
- if (!UnswitchedVals.insert(UnswitchVal))
+ // At this point we have some cases already unswitched and
+ // some not yet unswitched. Let's find the first not yet unswitched one.
+ for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end();
+ i != e; ++i) {
+ Constant* UnswitchValCandidate = i.getCaseValue();
+ if (!BranchesInfo.isUnswitched(SI, UnswitchValCandidate)) {
+ UnswitchVal = UnswitchValCandidate;
+ break;
+ }
+ }
+
+ if (!UnswitchVal)
continue;
if (UnswitchIfProfitable(LoopCond, UnswitchVal)) {
}
}
}
-
+
// Scan the instructions to check for unswitchable values.
- for (BasicBlock::iterator BBI = (*I)->begin(), E = (*I)->end();
+ for (BasicBlock::iterator BBI = (*I)->begin(), E = (*I)->end();
BBI != E; ++BBI)
if (SelectInst *SI = dyn_cast<SelectInst>(BBI)) {
- Value *LoopCond = FindLIVLoopCondition(SI->getCondition(),
+ Value *LoopCond = FindLIVLoopCondition(SI->getCondition(),
currentLoop, Changed);
- if (LoopCond && UnswitchIfProfitable(LoopCond,
+ if (LoopCond && UnswitchIfProfitable(LoopCond,
ConstantInt::getTrue(Context))) {
++NumSelects;
return true;
return Changed;
}
-/// isTrivialLoopExitBlock - Check to see if all paths from BB either:
-/// 1. Exit the loop with no side effects.
-/// 2. Branch to the latch block with no side-effects.
+/// isTrivialLoopExitBlock - Check to see if all paths from BB exit the
+/// loop with no side effects (including infinite loops).
///
-/// If these conditions are true, we return true and set ExitBB to the block we
+/// If true, we return true and set ExitBB to the block we
/// exit through.
///
static bool isTrivialLoopExitBlockHelper(Loop *L, BasicBlock *BB,
BasicBlock *&ExitBB,
std::set<BasicBlock*> &Visited) {
if (!Visited.insert(BB).second) {
- // Already visited and Ok, end of recursion.
- return true;
+ // Already visited. Without more analysis, this could indicate an infinite
+ // loop.
+ return false;
} else if (!L->contains(BB)) {
// Otherwise, this is a loop exit, this is fine so long as this is the
// first exit.
ExitBB = BB;
return true;
}
-
+
// Otherwise, this is an unvisited intra-loop node. Check all successors.
for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI) {
// Check to see if the successor is a trivial loop exit.
for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
if (I->mayHaveSideEffects())
return false;
-
+
return true;
}
/// isTrivialLoopExitBlock - Return true if the specified block unconditionally
-/// leads to an exit from the specified loop, and has no side-effects in the
+/// leads to an exit from the specified loop, and has no side-effects in the
/// process. If so, return the block that is exited to, otherwise return null.
static BasicBlock *isTrivialLoopExitBlock(Loop *L, BasicBlock *BB) {
std::set<BasicBlock*> Visited;
- Visited.insert(L->getHeader()); // Branches to header are ok.
+ Visited.insert(L->getHeader()); // Branches to header make infinite loops.
BasicBlock *ExitBB = 0;
if (isTrivialLoopExitBlockHelper(L, BB, ExitBB, Visited))
return ExitBB;
BasicBlock *Header = currentLoop->getHeader();
TerminatorInst *HeaderTerm = Header->getTerminator();
LLVMContext &Context = Header->getContext();
-
+
BasicBlock *LoopExitBB = 0;
if (BranchInst *BI = dyn_cast<BranchInst>(HeaderTerm)) {
// If the header block doesn't end with a conditional branch on Cond, we
// can't handle it.
if (!BI->isConditional() || BI->getCondition() != Cond)
return false;
-
- // Check to see if a successor of the branch is guaranteed to go to the
- // latch block or exit through a one exit block without having any
+
+ // Check to see if a successor of the branch is guaranteed to
+ // exit through a unique exit block without having any
// side-effects. If so, determine the value of Cond that causes it to do
// this.
- if ((LoopExitBB = isTrivialLoopExitBlock(currentLoop,
+ if ((LoopExitBB = isTrivialLoopExitBlock(currentLoop,
BI->getSuccessor(0)))) {
if (Val) *Val = ConstantInt::getTrue(Context);
- } else if ((LoopExitBB = isTrivialLoopExitBlock(currentLoop,
+ } else if ((LoopExitBB = isTrivialLoopExitBlock(currentLoop,
BI->getSuccessor(1)))) {
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.
if (SI->getCondition() != Cond) return false;
-
+
// Check to see if a successor of the switch is guaranteed to go to the
- // latch block or exit through a one exit block without having any
+ // latch block or exit through a one exit block without having any
// side-effects. If so, determine the value of Cond that causes it to do
- // this. Note that we can't trivially unswitch on the default case.
- for (unsigned i = 1, e = SI->getNumSuccessors(); i != e; ++i)
- if ((LoopExitBB = isTrivialLoopExitBlock(currentLoop,
- SI->getSuccessor(i)))) {
+ // this.
+ // Note that we can't trivially unswitch on the default case or
+ // on already unswitched cases.
+ for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end();
+ i != e; ++i) {
+ BasicBlock* LoopExitCandidate;
+ if ((LoopExitCandidate = isTrivialLoopExitBlock(currentLoop,
+ i.getCaseSuccessor()))) {
// Okay, we found a trivial case, remember the value that is trivial.
- if (Val) *Val = SI->getCaseValue(i);
+ ConstantInt* CaseVal = i.getCaseValue();
+
+ // Check that it was not unswitched before, since already unswitched
+ // trivial vals are looks trivial too.
+ if (BranchesInfo.isUnswitched(SI, CaseVal))
+ continue;
+ LoopExitBB = LoopExitCandidate;
+ if (Val) *Val = CaseVal;
break;
}
+ }
}
// If we didn't find a single unique LoopExit block, or if the loop exit block
// contains phi nodes, this isn't trivial.
if (!LoopExitBB || isa<PHINode>(LoopExitBB->begin()))
return false; // Can't handle this.
-
+
if (LoopExit) *LoopExit = LoopExitBB;
-
+
// We already know that nothing uses any scalar values defined inside of this
// loop. As such, we just have to check to see if this loop will execute any
// side-effecting instructions (e.g. stores, calls, volatile loads) in the
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();
+bool LoopUnswitch::UnswitchIfProfitable(Value *LoopCond, Constant *Val) {
Function *F = loopHeader->getParent();
+ Constant *CondVal = 0;
+ BasicBlock *ExitBlock = 0;
+ if (IsTrivialUnswitchCondition(LoopCond, &CondVal, &ExitBlock)) {
+ // If the condition is trivial, always unswitch. There is no code growth
+ // for this case.
+ UnswitchTrivialCondition(currentLoop, LoopCond, CondVal, ExitBlock);
+ return true;
+ }
// Check to see if it would be profitable to unswitch current loop.
- unsigned Cost = getLoopUnswitchCost(LoopCond);
// Do not do non-trivial unswitch while optimizing for size.
- if (Cost && OptimizeForSize)
- return false;
- if (Cost && !F->isDeclaration() && F->hasFnAttr(Attribute::OptimizeForSize))
+ if (OptimizeForSize ||
+ F->getAttributes().hasAttribute(AttributeSet::FunctionIndex,
+ Attribute::OptimizeForSize))
return false;
- if (Cost > Threshold) {
- // FIXME: this should estimate growth by the amount of code shared by the
- // resultant unswitched loops.
- //
- DEBUG(errs() << "NOT unswitching loop %"
- << currentLoop->getHeader()->getName() << ", cost too high: "
- << currentLoop->getBlocks().size() << "\n");
- return false;
- }
-
- Constant *CondVal;
- BasicBlock *ExitBlock;
- if (IsTrivialUnswitchCondition(LoopCond, &CondVal, &ExitBlock)) {
- UnswitchTrivialCondition(currentLoop, LoopCond, CondVal, ExitBlock);
- } else {
- UnswitchNontrivialCondition(LoopCond, Val, currentLoop);
- }
-
+ UnswitchNontrivialCondition(LoopCond, Val, currentLoop);
return true;
}
-// RemapInstruction - Convert the instruction operands from referencing the
-// current values into those specified by ValueMap.
-//
-static inline void RemapInstruction(Instruction *I,
- DenseMap<const Value *, Value*> &ValueMap) {
- for (unsigned op = 0, E = I->getNumOperands(); op != E; ++op) {
- Value *Op = I->getOperand(op);
- DenseMap<const Value *, Value*>::iterator It = ValueMap.find(Op);
- if (It != ValueMap.end()) Op = It->second;
- I->setOperand(op, Op);
- }
-}
-
/// CloneLoop - Recursively clone the specified loop and all of its children,
/// mapping the blocks with the specified map.
-static Loop *CloneLoop(Loop *L, Loop *PL, DenseMap<const Value*, Value*> &VM,
+static Loop *CloneLoop(Loop *L, Loop *PL, ValueToValueMapTy &VM,
LoopInfo *LI, LPPassManager *LPM) {
Loop *New = new Loop();
-
LPM->insertLoop(New, PL);
// Add all of the blocks in L to the new loop.
Value *BranchVal = LIC;
if (!isa<ConstantInt>(Val) ||
Val->getType() != Type::getInt1Ty(LIC->getContext()))
- BranchVal = new ICmpInst(InsertPt, ICmpInst::ICMP_EQ, LIC, Val, "tmp");
+ BranchVal = new ICmpInst(InsertPt, ICmpInst::ICMP_EQ, LIC, Val);
else if (Val != ConstantInt::getTrue(Val->getContext()))
// We want to enter the new loop when the condition is true.
std::swap(TrueDest, FalseDest);
// If either edge is critical, split it. This helps preserve LoopSimplify
// form for enclosing loops.
- SplitCriticalEdge(BI, 0, this);
- SplitCriticalEdge(BI, 1, this);
+ SplitCriticalEdge(BI, 0, this, false, false, true);
+ SplitCriticalEdge(BI, 1, this, false, false, true);
}
/// UnswitchTrivialCondition - Given a loop that has a trivial unswitchable
/// condition in it (a cond branch from its header block to its latch block,
-/// where the path through the loop that doesn't execute its body has no
+/// where the path through the loop that doesn't execute its body has no
/// side-effects), unswitch it. This doesn't involve any code duplication, just
/// moving the conditional branch outside of the loop and updating loop info.
-void LoopUnswitch::UnswitchTrivialCondition(Loop *L, Value *Cond,
- Constant *Val,
+void LoopUnswitch::UnswitchTrivialCondition(Loop *L, Value *Cond,
+ Constant *Val,
BasicBlock *ExitBlock) {
- DEBUG(errs() << "loop-unswitch: Trivial-Unswitch loop %"
+ 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
// conditional branch on Cond.
// Now that we have a place to insert the conditional branch, create a place
// to branch to: this is the exit block out of the loop that we should
// short-circuit to.
-
+
// Split this block now, so that the loop maintains its exit block, and so
// that the jump from the preheader can execute the contents of the exit block
// without actually branching to it (the exit block should be dominated by the
// loop header, not the preheader).
assert(!L->contains(ExitBlock) && "Exit block is in the loop?");
BasicBlock *NewExit = SplitBlock(ExitBlock, ExitBlock->begin(), this);
-
- // Okay, now we have a position to branch from and a position to branch to,
+
+ // Okay, now we have a position to branch from and a position to branch to,
// insert the new conditional branch.
- EmitPreheaderBranchOnCondition(Cond, Val, NewExit, NewPH,
+ EmitPreheaderBranchOnCondition(Cond, Val, NewExit, NewPH,
loopPreheader->getTerminator());
LPM->deleteSimpleAnalysisValue(loopPreheader->getTerminator(), L);
loopPreheader->getTerminator()->eraseFromParent();
// We need to reprocess this loop, it could be unswitched again.
redoLoop = true;
-
+
// Now that we know that the loop is never entered when this condition is a
// particular value, rewrite the loop with this info. We know that this will
// at least eliminate the old branch.
/// SplitExitEdges - Split all of the edges from inside the loop to their exit
/// blocks. Update the appropriate Phi nodes as we do so.
-void LoopUnswitch::SplitExitEdges(Loop *L,
- const SmallVector<BasicBlock *, 8> &ExitBlocks)
-{
+void LoopUnswitch::SplitExitEdges(Loop *L,
+ const SmallVector<BasicBlock *, 8> &ExitBlocks){
for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) {
BasicBlock *ExitBlock = ExitBlocks[i];
SmallVector<BasicBlock *, 4> Preds(pred_begin(ExitBlock),
pred_end(ExitBlock));
- SplitBlockPredecessors(ExitBlock, Preds.data(), Preds.size(),
- ".us-lcssa", this);
+
+ // Although SplitBlockPredecessors doesn't preserve loop-simplify in
+ // general, if we call it on all predecessors of all exits then it does.
+ if (!ExitBlock->isLandingPad()) {
+ SplitBlockPredecessors(ExitBlock, Preds, ".us-lcssa", this);
+ } else {
+ SmallVector<BasicBlock*, 2> NewBBs;
+ SplitLandingPadPredecessors(ExitBlock, Preds, ".us-lcssa", ".us-lcssa",
+ this, NewBBs);
+ }
}
}
-/// UnswitchNontrivialCondition - We determined that the loop is profitable
-/// to unswitch when LIC equal Val. Split it into loop versions and test the
+/// UnswitchNontrivialCondition - We determined that the loop is profitable
+/// to unswitch when LIC equal Val. Split it into loop versions and test the
/// condition outside of either loop. Return the loops created as Out1/Out2.
-void LoopUnswitch::UnswitchNontrivialCondition(Value *LIC, Constant *Val,
+void LoopUnswitch::UnswitchNontrivialCondition(Value *LIC, Constant *Val,
Loop *L) {
Function *F = loopHeader->getParent();
- DEBUG(errs() << "loop-unswitch: Unswitching loop %"
+ DEBUG(dbgs() << "loop-unswitch: Unswitching loop %"
<< loopHeader->getName() << " [" << L->getBlocks().size()
<< " blocks] in Function " << F->getName()
<< " when '" << *Val << "' == " << *LIC << "\n");
+ if (ScalarEvolution *SE = getAnalysisIfAvailable<ScalarEvolution>())
+ SE->forgetLoop(L);
+
LoopBlocks.clear();
NewBlocks.clear();
// the loop preheader and exit blocks), keeping track of the mapping between
// the instructions and blocks.
NewBlocks.reserve(LoopBlocks.size());
- DenseMap<const Value*, Value*> ValueMap;
+ ValueToValueMapTy VMap;
for (unsigned i = 0, e = LoopBlocks.size(); i != e; ++i) {
- BasicBlock *New = CloneBasicBlock(LoopBlocks[i], ValueMap, ".us", F);
- NewBlocks.push_back(New);
- ValueMap[LoopBlocks[i]] = New; // Keep the BB mapping.
- LPM->cloneBasicBlockSimpleAnalysis(LoopBlocks[i], New, L);
+ BasicBlock *NewBB = CloneBasicBlock(LoopBlocks[i], VMap, ".us", F);
+
+ NewBlocks.push_back(NewBB);
+ VMap[LoopBlocks[i]] = NewBB; // Keep the BB mapping.
+ LPM->cloneBasicBlockSimpleAnalysis(LoopBlocks[i], NewBB, L);
}
// Splice the newly inserted blocks into the function right before the
// original preheader.
- F->getBasicBlockList().splice(LoopBlocks[0], F->getBasicBlockList(),
+ F->getBasicBlockList().splice(NewPreheader, F->getBasicBlockList(),
NewBlocks[0], F->end());
// Now we create the new Loop object for the versioned loop.
- Loop *NewLoop = CloneLoop(L, L->getParentLoop(), ValueMap, LI, LPM);
+ Loop *NewLoop = CloneLoop(L, L->getParentLoop(), VMap, LI, LPM);
+
+ // Recalculate unswitching quota, inherit simplified switches info for NewBB,
+ // Probably clone more loop-unswitch related loop properties.
+ BranchesInfo.cloneData(NewLoop, L, VMap);
+
Loop *ParentLoop = L->getParentLoop();
if (ParentLoop) {
// Make sure to add the cloned preheader and exit blocks to the parent loop
// as well.
ParentLoop->addBasicBlockToLoop(NewBlocks[0], LI->getBase());
}
-
+
for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) {
- BasicBlock *NewExit = cast<BasicBlock>(ValueMap[ExitBlocks[i]]);
+ BasicBlock *NewExit = cast<BasicBlock>(VMap[ExitBlocks[i]]);
// The new exit block should be in the same loop as the old one.
if (Loop *ExitBBLoop = LI->getLoopFor(ExitBlocks[i]))
ExitBBLoop->addBasicBlockToLoop(NewExit, LI->getBase());
-
+
assert(NewExit->getTerminator()->getNumSuccessors() == 1 &&
"Exit block should have been split to have one successor!");
BasicBlock *ExitSucc = NewExit->getTerminator()->getSuccessor(0);
// If the successor of the exit block had PHI nodes, add an entry for
// NewExit.
PHINode *PN;
- for (BasicBlock::iterator I = ExitSucc->begin();
- (PN = dyn_cast<PHINode>(I)); ++I) {
+ for (BasicBlock::iterator I = ExitSucc->begin(); isa<PHINode>(I); ++I) {
+ PN = cast<PHINode>(I);
Value *V = PN->getIncomingValueForBlock(ExitBlocks[i]);
- DenseMap<const Value *, Value*>::iterator It = ValueMap.find(V);
- if (It != ValueMap.end()) V = It->second;
+ ValueToValueMapTy::iterator It = VMap.find(V);
+ if (It != VMap.end()) V = It->second;
PN->addIncoming(V, NewExit);
}
+
+ if (LandingPadInst *LPad = NewExit->getLandingPadInst()) {
+ PN = PHINode::Create(LPad->getType(), 0, "",
+ ExitSucc->getFirstInsertionPt());
+
+ for (pred_iterator I = pred_begin(ExitSucc), E = pred_end(ExitSucc);
+ I != E; ++I) {
+ BasicBlock *BB = *I;
+ LandingPadInst *LPI = BB->getLandingPadInst();
+ LPI->replaceAllUsesWith(PN);
+ PN->addIncoming(LPI, BB);
+ }
+ }
}
// Rewrite the code to refer to itself.
for (unsigned i = 0, e = NewBlocks.size(); i != e; ++i)
for (BasicBlock::iterator I = NewBlocks[i]->begin(),
E = NewBlocks[i]->end(); I != E; ++I)
- RemapInstruction(I, ValueMap);
-
+ RemapInstruction(I, VMap,RF_NoModuleLevelChanges|RF_IgnoreMissingEntries);
+
// Rewrite the original preheader to select between versions of the loop.
BranchInst *OldBR = cast<BranchInst>(loopPreheader->getTerminator());
assert(OldBR->isUnconditional() && OldBR->getSuccessor(0) == LoopBlocks[0] &&
LoopProcessWorklist.push_back(NewLoop);
redoLoop = true;
+ // Keep a WeakVH holding onto LIC. If the first call to RewriteLoopBody
+ // deletes the instruction (for example by simplifying a PHI that feeds into
+ // the condition that we're unswitching on), we don't rewrite the second
+ // iteration.
+ WeakVH LICHandle(LIC);
+
// Now we rewrite the original code to know that the condition is true and the
// new code to know that the condition is false.
- RewriteLoopBodyWithConditionConstant(L , LIC, Val, false);
-
- // It's possible that simplifying one loop could cause the other to be
- // deleted. If so, don't simplify it.
- if (!LoopProcessWorklist.empty() && LoopProcessWorklist.back() == NewLoop)
- RewriteLoopBodyWithConditionConstant(NewLoop, LIC, Val, true);
+ RewriteLoopBodyWithConditionConstant(L, LIC, Val, false);
+ // It's possible that simplifying one loop could cause the other to be
+ // changed to another value or a constant. If its a constant, don't simplify
+ // it.
+ if (!LoopProcessWorklist.empty() && LoopProcessWorklist.back() == NewLoop &&
+ LICHandle && !isa<Constant>(LICHandle))
+ RewriteLoopBodyWithConditionConstant(NewLoop, LICHandle, Val, true);
}
/// RemoveFromWorklist - Remove all instances of I from the worklist vector
/// specified.
-static void RemoveFromWorklist(Instruction *I,
+static void RemoveFromWorklist(Instruction *I,
std::vector<Instruction*> &Worklist) {
- std::vector<Instruction*>::iterator WI = std::find(Worklist.begin(),
- Worklist.end(), I);
- while (WI != Worklist.end()) {
- unsigned Offset = WI-Worklist.begin();
- Worklist.erase(WI);
- WI = std::find(Worklist.begin()+Offset, Worklist.end(), I);
- }
+
+ Worklist.erase(std::remove(Worklist.begin(), Worklist.end(), I),
+ Worklist.end());
}
/// ReplaceUsesOfWith - When we find that I really equals V, remove I from the
/// program, replacing all uses with V and update the worklist.
-static void ReplaceUsesOfWith(Instruction *I, Value *V,
+static void ReplaceUsesOfWith(Instruction *I, Value *V,
std::vector<Instruction*> &Worklist,
Loop *L, LPPassManager *LPM) {
- DEBUG(errs() << "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)
if (BasicBlock *Pred = BB->getSinglePredecessor()) {
// If it has one pred, fold phi nodes in BB.
while (isa<PHINode>(BB->begin()))
- ReplaceUsesOfWith(BB->begin(),
- cast<PHINode>(BB->begin())->getIncomingValue(0),
+ ReplaceUsesOfWith(BB->begin(),
+ cast<PHINode>(BB->begin())->getIncomingValue(0),
Worklist, L, LPM);
-
+
// If this is the header of a loop and the only pred is the latch, we now
// have an unreachable loop.
if (Loop *L = LI->getLoopFor(BB))
LPM->deleteSimpleAnalysisValue(Pred->getTerminator(), L);
Pred->getTerminator()->eraseFromParent();
new UnreachableInst(BB->getContext(), Pred);
-
+
// The loop is now broken, remove it from LI.
RemoveLoopFromHierarchy(L);
-
+
// Reprocess the header, which now IS dead.
RemoveBlockIfDead(BB, Worklist, L);
return;
}
-
+
// If pred ends in a uncond branch, add uncond branch to worklist so that
// the two blocks will get merged.
if (BranchInst *BI = dyn_cast<BranchInst>(Pred->getTerminator()))
return;
}
- DEBUG(errs() << "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) {
RemoveFromWorklist(I, Worklist);
-
+
// 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()));
}
-
+
// If this is the edge to the header block for a loop, remove the loop and
// promote all subloops.
if (Loop *BBLoop = LI->getLoopFor(BB)) {
- if (BBLoop->getLoopLatch() == BB)
+ if (BBLoop->getLoopLatch() == BB) {
RemoveLoopFromHierarchy(BBLoop);
+ if (currentLoop == BBLoop) {
+ currentLoop = 0;
+ redoLoop = false;
+ }
+ }
}
// Remove the block from the loop info, which removes it from any loops it
// was in.
LI->removeBlock(BB);
-
-
+
+
// Remove phi node entries in successors for this block.
TerminatorInst *TI = BB->getTerminator();
SmallVector<BasicBlock*, 4> Succs;
Succs.push_back(TI->getSuccessor(i));
TI->getSuccessor(i)->removePredecessor(BB);
}
-
+
// Unique the successors, remove anything with multiple uses.
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.
- LPM->deleteSimpleAnalysisValue(BB, L);
+ LPM->deleteSimpleAnalysisValue(BB, L);
BB->eraseFromParent();
// Remove successor blocks here that are not dead, so that we know we only
// have dead blocks in this list. Nondead blocks have a way of becoming dead,
--i;
}
}
-
+
for (unsigned i = 0, e = Succs.size(); i != e; ++i)
RemoveBlockIfDead(Succs[i], Worklist, L);
}
Constant *Val,
bool IsEqual) {
assert(!isa<Constant>(LIC) && "Why are we unswitching on a constant?");
-
+
// FIXME: Support correlated properties, like:
// for (...)
// if (li1 < li2)
// ...
// if (li1 > li2)
// ...
-
+
// FOLD boolean conditions (X|LIC), (X&LIC). Fold conditional branches,
// 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::getInt1Ty(Val->getContext()))) {
+ Val->getType()->isIntegerTy(1))) {
Value *Replacement;
if (IsEqual)
Replacement = Val;
else
- Replacement = ConstantInt::get(Type::getInt1Ty(Val->getContext()),
+ 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()))
- continue;
- U->replaceUsesOfWith(LIC, Replacement);
- Worklist.push_back(U);
- }
- } else {
- // Otherwise, we don't know the precise value of LIC, but we do know that it
- // is certainly NOT "Val". As such, simplify any uses in the loop that we
- // 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()))
- continue;
- Worklist.push_back(U);
-
- // If we know that LIC is not Val, use this info to simplify code.
- if (SwitchInst *SI = dyn_cast<SwitchInst>(U)) {
- for (unsigned i = 1, e = SI->getNumCases(); i != e; ++i) {
- if (SI->getCaseValue(i) == Val) {
- // Found a dead case value. Don't remove PHI nodes in the
- // successor if they become single-entry, those PHI nodes may
- // be in the Users list.
-
- // 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 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;
- }
- }
- }
-
- // TODO: We could do other simplifications, for example, turning
- // LIC == Val -> false.
- }
+ for (Value::use_iterator UI = LIC->use_begin(), E = LIC->use_end();
+ UI != E; ++UI) {
+ Instruction *U = dyn_cast<Instruction>(*UI);
+ if (!U || !L->contains(U))
+ continue;
+ Worklist.push_back(U);
+ }
+
+ for (std::vector<Instruction*>::iterator UI = Worklist.begin();
+ UI != Worklist.end(); ++UI)
+ (*UI)->replaceUsesOfWith(LIC, Replacement);
+
+ SimplifyCode(Worklist, L);
+ return;
+ }
+
+ // Otherwise, we don't know the precise value of LIC, but we do know that it
+ // is certainly NOT "Val". As such, simplify any uses in the loop that we
+ // can. This case occurs when we unswitch switch statements.
+ for (Value::use_iterator UI = LIC->use_begin(), E = LIC->use_end();
+ UI != E; ++UI) {
+ Instruction *U = dyn_cast<Instruction>(*UI);
+ if (!U || !L->contains(U))
+ continue;
+
+ Worklist.push_back(U);
+
+ // TODO: We could do other simplifications, for example, turning
+ // 'icmp eq LIC, Val' -> false.
+
+ // If we know that LIC is not Val, use this info to simplify code.
+ SwitchInst *SI = dyn_cast<SwitchInst>(U);
+ if (SI == 0 || !isa<ConstantInt>(Val)) continue;
+
+ SwitchInst::CaseIt DeadCase = SI->findCaseValue(cast<ConstantInt>(Val));
+ // Default case is live for multiple values.
+ if (DeadCase == SI->case_default()) continue;
+
+ // Found a dead case value. Don't remove PHI nodes in the
+ // successor if they become single-entry, those PHI nodes may
+ // be in the Users list.
+
+ BasicBlock *Switch = SI->getParent();
+ BasicBlock *SISucc = DeadCase.getCaseSuccessor();
+ BasicBlock *Latch = L->getLoopLatch();
+
+ BranchesInfo.setUnswitched(SI, Val);
+
+ if (!SI->findCaseDest(SISucc)) continue; // Edge is critical.
+ // If the DeadCase successor dominates the loop latch, then the
+ // transformation isn't safe since it will delete the sole predecessor edge
+ // to the latch.
+ if (Latch && DT->dominates(SISucc, Latch))
+ continue;
+
+ // 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 a dead code path.
+ SplitEdge(Switch, SISucc, 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 = DeadCase.getCaseSuccessor();
+ 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);
}
-
+
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
while (!Worklist.empty()) {
Instruction *I = Worklist.back();
Worklist.pop_back();
-
- // Simple constant folding.
- if (Constant *C = ConstantFoldInstruction(I, I->getContext())) {
- ReplaceUsesOfWith(I, C, Worklist, L, LPM);
- continue;
- }
-
+
// Simple DCE.
if (isInstructionTriviallyDead(I)) {
- DEBUG(errs() << "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)
if (Instruction *Use = dyn_cast<Instruction>(I->getOperand(i)))
++NumSimplify;
continue;
}
-
- // Special case hacks that appear commonly in unswitched code.
- switch (I->getOpcode()) {
- case Instruction::Select:
- if (ConstantInt *CB = dyn_cast<ConstantInt>(I->getOperand(0))) {
- ReplaceUsesOfWith(I, I->getOperand(!CB->getZExtValue()+1), Worklist, L,
- LPM);
+
+ // See if instruction simplification can hack this up. This is common for
+ // things like "select false, X, Y" after unswitching made the condition be
+ // 'false'. TODO: update the domtree properly so we can pass it here.
+ if (Value *V = SimplifyInstruction(I))
+ if (LI->replacementPreservesLCSSAForm(I, V)) {
+ ReplaceUsesOfWith(I, V, Worklist, L, LPM);
continue;
}
- break;
- case Instruction::And:
- if (isa<ConstantInt>(I->getOperand(0)) &&
- // constant -> RHS
- I->getOperand(0)->getType() == Type::getInt1Ty(I->getContext()))
- cast<BinaryOperator>(I)->swapOperands();
- if (ConstantInt *CB = dyn_cast<ConstantInt>(I->getOperand(1)))
- if (CB->getType() == Type::getInt1Ty(I->getContext())) {
- if (CB->isOne()) // X & 1 -> X
- ReplaceUsesOfWith(I, I->getOperand(0), Worklist, L, LPM);
- else // X & 0 -> 0
- ReplaceUsesOfWith(I, I->getOperand(1), Worklist, L, LPM);
- continue;
- }
- break;
- case Instruction::Or:
- if (isa<ConstantInt>(I->getOperand(0)) &&
- // constant -> RHS
- I->getOperand(0)->getType() == Type::getInt1Ty(I->getContext()))
- cast<BinaryOperator>(I)->swapOperands();
- if (ConstantInt *CB = dyn_cast<ConstantInt>(I->getOperand(1)))
- if (CB->getType() == Type::getInt1Ty(I->getContext())) {
- if (CB->isOne()) // X | 1 -> 1
- ReplaceUsesOfWith(I, I->getOperand(1), Worklist, L, LPM);
- else // X | 0 -> X
- ReplaceUsesOfWith(I, I->getOperand(0), Worklist, L, LPM);
- continue;
- }
- break;
- case Instruction::Br: {
- BranchInst *BI = cast<BranchInst>(I);
+
+ // Special case hacks that appear commonly in unswitched code.
+ if (BranchInst *BI = dyn_cast<BranchInst>(I)) {
if (BI->isUnconditional()) {
// If BI's parent is the only pred of the successor, fold the two blocks
// together.
if (!SinglePred) continue; // Nothing to do.
assert(SinglePred == Pred && "CFG broken");
- DEBUG(errs() << "Merging blocks: " << Pred->getName() << " <- "
+ 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()))
ReplaceUsesOfWith(PN, PN->getIncomingValue(0), Worklist, L, LPM);
-
+
+ // If Succ has any successors with PHI nodes, update them to have
+ // entries coming from Pred instead of Succ.
+ Succ->replaceAllUsesWith(Pred);
+
// Move all of the successor contents from Succ to Pred.
Pred->getInstList().splice(BI, Succ->getInstList(), Succ->begin(),
Succ->end());
LPM->deleteSimpleAnalysisValue(BI, L);
BI->eraseFromParent();
RemoveFromWorklist(BI, Worklist);
-
- // If Succ has any successors with PHI nodes, update them to have
- // entries coming from Pred instead of Succ.
- Succ->replaceAllUsesWith(Pred);
-
+
// Remove Succ from the loop tree.
LI->removeBlock(Succ);
LPM->deleteSimpleAnalysisValue(Succ, L);
Succ->eraseFromParent();
++NumSimplify;
- } else if (ConstantInt *CB = dyn_cast<ConstantInt>(BI->getCondition())){
+ continue;
+ }
+
+ if (ConstantInt *CB = dyn_cast<ConstantInt>(BI->getCondition())){
// Conditional branch. Turn it into an unconditional branch, then
// remove dead blocks.
- break; // FIXME: Enable.
+ continue; // FIXME: Enable.
- DEBUG(errs() << "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);
RemoveBlockIfDead(DeadSucc, Worklist, L);
}
- break;
- }
+ continue;
}
}
}