//
// 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.
//
//===----------------------------------------------------------------------===//
//
#include "llvm/DerivedTypes.h"
#include "llvm/Function.h"
#include "llvm/Instructions.h"
+#include "llvm/LLVMContext.h"
+#include "llvm/Analysis/ConstantFolding.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/PostOrderIterator.h"
-#include "llvm/Support/Debug.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");
-namespace {
- cl::opt<unsigned>
- Threshold("loop-unswitch-threshold", cl::desc("Max loop size to unswitch"),
- cl::init(10), cl::Hidden);
+static cl::opt<unsigned>
+Threshold("loop-unswitch-threshold", cl::desc("Max loop size to unswitch"),
+ cl::init(10), cl::Hidden);
- class LoopUnswitch : public FunctionPass {
+namespace {
+ class VISIBILITY_HIDDEN LoopUnswitch : public LoopPass {
LoopInfo *LI; // Loop information
+ LPPassManager *LPM;
- // LoopProcessWorklist - List of loops we need to process.
+ // LoopProcessWorklist - Used to check if second loop needs processing
+ // after RewriteLoopBodyWithConditionConstant rewrites first loop.
std::vector<Loop*> LoopProcessWorklist;
+ SmallPtrSet<Value *,8> UnswitchedVals;
+
+ 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
+ // loop, in that order.
+ std::vector<BasicBlock*> LoopBlocks;
+ // NewBlocks contained cloned copy of basic blocks from LoopBlocks.
+ std::vector<BasicBlock*> NewBlocks;
+
public:
- virtual bool runOnFunction(Function &F);
- bool visitLoop(Loop *L);
+ 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) {}
+
+ bool runOnLoop(Loop *L, LPPassManager &LPM);
+ bool processCurrentLoop();
/// This transformation requires natural loop information & requires that
/// loop preheaders be inserted into the CFG...
AU.addPreserved<LoopInfo>();
AU.addRequiredID(LCSSAID);
AU.addPreservedID(LCSSAID);
+ AU.addPreserved<DominatorTree>();
+ AU.addPreserved<DominanceFrontier>();
}
private:
+
/// RemoveLoopFromWorklist - If the specified loop is on the loop worklist,
/// remove it.
void RemoveLoopFromWorklist(Loop *L) {
if (I != LoopProcessWorklist.end())
LoopProcessWorklist.erase(I);
}
-
- bool UnswitchIfProfitable(Value *LoopCond, Constant *Val,Loop *L);
- unsigned getLoopUnswitchCost(Loop *L, Value *LIC);
+
+ void initLoopData() {
+ loopHeader = currentLoop->getHeader();
+ loopPreheader = currentLoop->getLoopPreheader();
+ }
+
+ /// Split all of the edges from inside the loop to their exit blocks.
+ /// Update the appropriate Phi nodes as we do so.
+ 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);
- BasicBlock *SplitEdge(BasicBlock *From, BasicBlock *To);
- BasicBlock *SplitBlock(BasicBlock *Old, Instruction *SplitPt);
void RewriteLoopBodyWithConditionConstant(Loop *L, Value *LIC,
Constant *Val, bool isEqual);
-
- void SimplifyCode(std::vector<Instruction*> &Worklist);
+
+ void EmitPreheaderBranchOnCondition(Value *LIC, Constant *Val,
+ BasicBlock *TrueDest,
+ BasicBlock *FalseDest,
+ Instruction *InsertPt);
+
+ void SimplifyCode(std::vector<Instruction*> &Worklist, Loop *L);
void RemoveBlockIfDead(BasicBlock *BB,
- std::vector<Instruction*> &Worklist);
+ std::vector<Instruction*> &Worklist, Loop *l);
void RemoveLoopFromHierarchy(Loop *L);
+ bool IsTrivialUnswitchCondition(Value *Cond, Constant **Val = 0,
+ BasicBlock **LoopExit = 0);
+
};
- RegisterPass<LoopUnswitch> X("loop-unswitch", "Unswitch loops");
}
+char LoopUnswitch::ID = 0;
+static RegisterPass<LoopUnswitch> X("loop-unswitch", "Unswitch loops");
-FunctionPass *llvm::createLoopUnswitchPass() { return new LoopUnswitch(); }
-
-bool LoopUnswitch::runOnFunction(Function &F) {
- bool Changed = false;
- LI = &getAnalysis<LoopInfo>();
-
- // Populate the worklist of loops to process in post-order.
- for (LoopInfo::iterator I = LI->begin(), E = LI->end(); I != E; ++I)
- for (po_iterator<Loop*> LI = po_begin(*I), E = po_end(*I); LI != E; ++LI)
- LoopProcessWorklist.push_back(*LI);
-
- // Process the loops in worklist order, this is a post-order visitation of
- // the loops. We use a worklist of loops so that loops can be removed at any
- // time if they are deleted (e.g. the backedge of a loop is removed).
- while (!LoopProcessWorklist.empty()) {
- Loop *L = LoopProcessWorklist.back();
- LoopProcessWorklist.pop_back();
- Changed |= visitLoop(L);
- }
-
- return Changed;
+Pass *llvm::createLoopUnswitchPass(bool Os) {
+ return new LoopUnswitch(Os);
}
/// FindLIVLoopCondition - Cond is a condition that occurs in L. If it is
/// Otherwise, return null.
static Value *FindLIVLoopCondition(Value *Cond, Loop *L, bool &Changed) {
// Constants should be folded, not unswitched on!
- if (isa<Constant>(Cond)) return false;
-
+ 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) {
if (Value *RHS = FindLIVLoopCondition(BO->getOperand(1), L, Changed))
return RHS;
}
-
- return 0;
+
+ return 0;
}
-bool LoopUnswitch::visitLoop(Loop *L) {
- assert(L->isLCSSAForm());
-
+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());
+ redoLoop = false;
+ Changed |= processCurrentLoop();
+ } while(redoLoop);
+
+ if (Changed) {
+ // 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
+/// 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
// loop.
- for (Loop::block_iterator I = L->block_begin(), E = L->block_end();
+ 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 (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(), L, Changed);
- if (LoopCond && UnswitchIfProfitable(LoopCond, ConstantInt::getTrue(),
- L)) {
+ Value *LoopCond = FindLIVLoopCondition(BI->getCondition(),
+ currentLoop, Changed);
+ if (LoopCond && UnswitchIfProfitable(LoopCond,
+ ConstantInt::getTrue(Context))) {
++NumBranches;
return true;
}
}
} else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
- Value *LoopCond = FindLIVLoopCondition(SI->getCondition(), L, Changed);
+ Value *LoopCond = FindLIVLoopCondition(SI->getCondition(),
+ currentLoop, Changed);
if (LoopCond && SI->getNumCases() > 1) {
// Find a value to unswitch on:
// FIXME: this should chose the most expensive case!
Constant *UnswitchVal = SI->getCaseValue(1);
- if (UnswitchIfProfitable(LoopCond, UnswitchVal, L)) {
+ // Do not process same value again and again.
+ if (!UnswitchedVals.insert(UnswitchVal))
+ continue;
+
+ if (UnswitchIfProfitable(LoopCond, UnswitchVal)) {
++NumSwitches;
return true;
}
for (BasicBlock::iterator BBI = (*I)->begin(), E = (*I)->end();
BBI != E; ++BBI)
if (SelectInst *SI = dyn_cast<SelectInst>(BBI)) {
- Value *LoopCond = FindLIVLoopCondition(SI->getCondition(), L, Changed);
- if (LoopCond && UnswitchIfProfitable(LoopCond, ConstantInt::getTrue(),
- L)) {
+ Value *LoopCond = FindLIVLoopCondition(SI->getCondition(),
+ currentLoop, Changed);
+ if (LoopCond && UnswitchIfProfitable(LoopCond,
+ ConstantInt::getTrue(Context))) {
++NumSelects;
return true;
}
}
}
-
- assert(L->isLCSSAForm());
-
return Changed;
}
// 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;
/// exit. Finally, this sets LoopExit to the BB that the loop exits to when
/// Cond == Val.
///
-static bool IsTrivialUnswitchCondition(Loop *L, Value *Cond, Constant **Val = 0,
- BasicBlock **LoopExit = 0) {
- BasicBlock *Header = L->getHeader();
+bool LoopUnswitch::IsTrivialUnswitchCondition(Value *Cond, Constant **Val,
+ BasicBlock **LoopExit) {
+ BasicBlock *Header = currentLoop->getHeader();
TerminatorInst *HeaderTerm = Header->getTerminator();
+ LLVMContext &Context = Header->getContext();
BasicBlock *LoopExitBB = 0;
if (BranchInst *BI = dyn_cast<BranchInst>(HeaderTerm)) {
// 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.
- if ((LoopExitBB = isTrivialLoopExitBlock(L, BI->getSuccessor(0)))) {
- if (Val) *Val = ConstantInt::getTrue();
- } else if ((LoopExitBB = isTrivialLoopExitBlock(L, BI->getSuccessor(1)))) {
- if (Val) *Val = ConstantInt::getFalse();
+ if ((LoopExitBB = isTrivialLoopExitBlock(currentLoop,
+ BI->getSuccessor(0)))) {
+ if (Val) *Val = ConstantInt::getTrue(Context);
+ } 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.
// 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(L, SI->getSuccessor(i)))) {
+ if ((LoopExitBB = isTrivialLoopExitBlock(currentLoop,
+ SI->getSuccessor(i)))) {
// Okay, we found a trivial case, remember the value that is trivial.
if (Val) *Val = SI->getCaseValue(i);
break;
// 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 the specified loop on the specified value.
+/// we choose to unswitch current loop on the specified value.
///
-unsigned LoopUnswitch::getLoopUnswitchCost(Loop *L, Value *LIC) {
+unsigned LoopUnswitch::getLoopUnswitchCost(Value *LIC) {
// If the condition is trivial, always unswitch. There is no code growth for
// this case.
- if (IsTrivialUnswitchCondition(L, LIC))
+ 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 L->getBlocks().size();
+ return currentLoop->getBlocks().size();
unsigned Cost = 0;
// FIXME: this is brain dead. It should take into consideration code
// shrinkage.
- for (Loop::block_iterator I = L->block_begin(), E = L->block_end();
+ 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
return Cost;
}
-/// UnswitchIfProfitable - We have found that we can unswitch L when
+/// 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,Loop *L){
- // Check to see if it would be profitable to unswitch this loop.
- unsigned Cost = getLoopUnswitchCost(L, LoopCond);
+bool LoopUnswitch::UnswitchIfProfitable(Value *LoopCond, Constant *Val){
+
+ initLoopData();
+ Function *F = loopHeader->getParent();
+
+
+ // 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))
+ 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 %"
- << L->getHeader()->getName() << ", cost too high: "
- << L->getBlocks().size() << "\n";
+ DEBUG(errs() << "NOT unswitching loop %"
+ << currentLoop->getHeader()->getName() << ", cost too high: "
+ << currentLoop->getBlocks().size() << "\n");
return false;
}
-
- // If this is a trivial condition to unswitch (which results in no code
- // duplication), do it now.
+
Constant *CondVal;
BasicBlock *ExitBlock;
- if (IsTrivialUnswitchCondition(L, LoopCond, &CondVal, &ExitBlock)) {
- UnswitchTrivialCondition(L, LoopCond, CondVal, ExitBlock);
+ if (IsTrivialUnswitchCondition(LoopCond, &CondVal, &ExitBlock)) {
+ UnswitchTrivialCondition(currentLoop, LoopCond, CondVal, ExitBlock);
} else {
- UnswitchNontrivialCondition(LoopCond, Val, L);
+ UnswitchNontrivialCondition(LoopCond, Val, currentLoop);
}
-
- return true;
-}
-/// SplitBlock - Split the specified block at the specified instruction - every
-/// thing before SplitPt stays in Old and everything starting with SplitPt moves
-/// to a new block. The two blocks are joined by an unconditional branch and
-/// the loop info is updated.
-///
-BasicBlock *LoopUnswitch::SplitBlock(BasicBlock *Old, Instruction *SplitPt) {
- BasicBlock::iterator SplitIt = SplitPt;
- while (isa<PHINode>(SplitIt))
- ++SplitIt;
- BasicBlock *New = Old->splitBasicBlock(SplitIt, Old->getName()+".split");
-
- // The new block lives in whichever loop the old one did.
- if (Loop *L = LI->getLoopFor(Old))
- L->addBasicBlockToLoop(New, *LI);
-
- return New;
+ return true;
}
-
-BasicBlock *LoopUnswitch::SplitEdge(BasicBlock *BB, BasicBlock *Succ) {
- TerminatorInst *LatchTerm = BB->getTerminator();
- unsigned SuccNum = 0;
- for (unsigned i = 0, e = LatchTerm->getNumSuccessors(); ; ++i) {
- assert(i != e && "Didn't find edge?");
- if (LatchTerm->getSuccessor(i) == Succ) {
- SuccNum = i;
- break;
- }
- }
-
- // If this is a critical edge, let SplitCriticalEdge do it.
- if (SplitCriticalEdge(BB->getTerminator(), SuccNum, this))
- return LatchTerm->getSuccessor(SuccNum);
-
- // If the edge isn't critical, then BB has a single successor or Succ has a
- // single pred. Split the block.
- BasicBlock::iterator SplitPoint;
- if (BasicBlock *SP = Succ->getSinglePredecessor()) {
- // If the successor only has a single pred, split the top of the successor
- // block.
- assert(SP == BB && "CFG broken");
- return SplitBlock(Succ, Succ->begin());
- } else {
- // Otherwise, if BB has a single successor, split it at the bottom of the
- // block.
- assert(BB->getTerminator()->getNumSuccessors() == 1 &&
- "Should have a single succ!");
- return SplitBlock(BB, BB->getTerminator());
- }
-}
-
-
-
// RemapInstruction - Convert the instruction operands from referencing the
// current values into those specified by ValueMap.
//
static inline void RemapInstruction(Instruction *I,
- std::map<const Value *, Value*> &ValueMap) {
+ DenseMap<const Value *, Value*> &ValueMap) {
for (unsigned op = 0, E = I->getNumOperands(); op != E; ++op) {
Value *Op = I->getOperand(op);
- std::map<const Value *, Value*>::iterator It = ValueMap.find(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, std::map<const Value*, Value*> &VM,
- LoopInfo *LI) {
+static Loop *CloneLoop(Loop *L, Loop *PL, DenseMap<const Value*, Value*> &VM,
+ LoopInfo *LI, LPPassManager *LPM) {
Loop *New = new Loop();
- if (PL)
- PL->addChildLoop(New);
- else
- LI->addTopLevelLoop(New);
+ LPM->insertLoop(New, PL);
// Add all of the blocks in L to the new loop.
for (Loop::block_iterator I = L->block_begin(), E = L->block_end();
I != E; ++I)
if (LI->getLoopFor(*I) == L)
- New->addBasicBlockToLoop(cast<BasicBlock>(VM[*I]), *LI);
+ New->addBasicBlockToLoop(cast<BasicBlock>(VM[*I]), LI->getBase());
// Add all of the subloops to the new loop.
for (Loop::iterator I = L->begin(), E = L->end(); I != E; ++I)
- CloneLoop(*I, New, VM, LI);
+ CloneLoop(*I, New, VM, LI, LPM);
return New;
}
/// EmitPreheaderBranchOnCondition - Emit a conditional branch on two values
/// if LIC == Val, branch to TrueDst, otherwise branch to FalseDest. Insert the
/// code immediately before InsertPt.
-static void EmitPreheaderBranchOnCondition(Value *LIC, Constant *Val,
- BasicBlock *TrueDest,
- BasicBlock *FalseDest,
- Instruction *InsertPt) {
+void LoopUnswitch::EmitPreheaderBranchOnCondition(Value *LIC, Constant *Val,
+ BasicBlock *TrueDest,
+ BasicBlock *FalseDest,
+ Instruction *InsertPt) {
// 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())
+ 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.
- new BranchInst(TrueDest, FalseDest, BranchVal, InsertPt);
+ BranchInst::Create(TrueDest, FalseDest, BranchVal, InsertPt);
}
-
/// 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
void LoopUnswitch::UnswitchTrivialCondition(Loop *L, Value *Cond,
Constant *Val,
BasicBlock *ExitBlock) {
- DOUT << "loop-unswitch: Trivial-Unswitch loop %"
- << L->getHeader()->getName() << " [" << L->getBlocks().size()
- << " blocks] in Function " << L->getHeader()->getParent()->getName()
- << " on cond: " << *Val << " == " << *Cond << "\n";
+ DEBUG(errs() << "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 'OrigPH' to have a
+ // to insert the conditional branch. We will change loopPreheader to have a
// conditional branch on Cond.
- BasicBlock *OrigPH = L->getLoopPreheader();
- BasicBlock *NewPH = SplitEdge(OrigPH, L->getHeader());
+ BasicBlock *NewPH = SplitEdge(loopPreheader, loopHeader, this);
// 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
// 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());
+ BasicBlock *NewExit = SplitBlock(ExitBlock, ExitBlock->begin(), this);
// 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,
- OrigPH->getTerminator());
- OrigPH->getTerminator()->eraseFromParent();
+ loopPreheader->getTerminator());
+ LPM->deleteSimpleAnalysisValue(loopPreheader->getTerminator(), L);
+ loopPreheader->getTerminator()->eraseFromParent();
// We need to reprocess this loop, it could be unswitched again.
- LoopProcessWorklist.push_back(L);
+ 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
++NumTrivial;
}
+/// 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)
+{
-/// VersionLoop - 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,
- Loop *L) {
- Function *F = L->getHeader()->getParent();
- DOUT << "loop-unswitch: Unswitching loop %"
- << L->getHeader()->getName() << " [" << L->getBlocks().size()
- << " blocks] in Function " << F->getName()
- << " when '" << *Val << "' == " << *LIC << "\n";
-
- // 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
- // loop, in that order.
- std::vector<BasicBlock*> LoopBlocks;
-
- // First step, split the preheader and exit blocks, and add these blocks to
- // the LoopBlocks list.
- BasicBlock *OrigPreheader = L->getLoopPreheader();
- LoopBlocks.push_back(SplitEdge(OrigPreheader, L->getHeader()));
-
- // We want the loop to come after the preheader, but before the exit blocks.
- LoopBlocks.insert(LoopBlocks.end(), L->block_begin(), L->block_end());
-
- std::vector<BasicBlock*> ExitBlocks;
- L->getUniqueExitBlocks(ExitBlocks);
-
- // Split all of the edges from inside the loop to their exit blocks. Update
- // the appropriate Phi nodes as we do so.
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) {
- assert(L->contains(Preds[j]) &&
- "All preds of loop exit blocks must be the same loop!");
- BasicBlock* MiddleBlock = SplitEdge(Preds[j], ExitBlock);
+ BasicBlock* NewExitBlock = SplitEdge(Preds[j], ExitBlock, this);
BasicBlock* StartBlock = Preds[j];
BasicBlock* EndBlock;
- if (MiddleBlock->getSinglePredecessor() == ExitBlock) {
- EndBlock = MiddleBlock;
- MiddleBlock = EndBlock->getSinglePredecessor();;
+ if (NewExitBlock->getSinglePredecessor() == ExitBlock) {
+ EndBlock = NewExitBlock;
+ NewExitBlock = EndBlock->getSinglePredecessor();
} else {
EndBlock = ExitBlock;
}
PHINode* OldLCSSA = 0;
for (BasicBlock::iterator I = EndBlock->begin();
(OldLCSSA = dyn_cast<PHINode>(I)); ++I) {
- Value* OldValue = OldLCSSA->getIncomingValueForBlock(MiddleBlock);
- PHINode* NewLCSSA = new PHINode(OldLCSSA->getType(),
- OldLCSSA->getName() + ".us-lcssa",
- MiddleBlock->getTerminator());
+ 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(MiddleBlock),
+ OldLCSSA->setIncomingValue(OldLCSSA->getBasicBlockIndex(NewExitBlock),
NewLCSSA);
InsertedPHIs.insert(NewLCSSA);
}
- BasicBlock::iterator InsertPt = EndBlock->begin();
- while (dyn_cast<PHINode>(InsertPt)) ++InsertPt;
- for (BasicBlock::iterator I = MiddleBlock->begin();
+ BasicBlock::iterator InsertPt = EndBlock->getFirstNonPHI();
+ for (BasicBlock::iterator I = NewExitBlock->begin();
(OldLCSSA = dyn_cast<PHINode>(I)) && InsertedPHIs.count(OldLCSSA) == 0;
++I) {
- PHINode *NewLCSSA = new PHINode(OldLCSSA->getType(),
- OldLCSSA->getName() + ".us-lcssa",
- InsertPt);
+ PHINode *NewLCSSA = PHINode::Create(OldLCSSA->getType(),
+ OldLCSSA->getName() + ".us-lcssa",
+ InsertPt);
OldLCSSA->replaceAllUsesWith(NewLCSSA);
- NewLCSSA->addIncoming(OldLCSSA, MiddleBlock);
+ NewLCSSA->addIncoming(OldLCSSA, NewExitBlock);
}
+
}
}
-
+
+}
+
+/// 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,
+ Loop *L) {
+ Function *F = loopHeader->getParent();
+ DEBUG(errs() << "loop-unswitch: Unswitching loop %"
+ << loopHeader->getName() << " [" << L->getBlocks().size()
+ << " blocks] in Function " << F->getName()
+ << " when '" << *Val << "' == " << *LIC << "\n");
+
+ LoopBlocks.clear();
+ NewBlocks.clear();
+
+ // First step, split the preheader and exit blocks, and add these blocks to
+ // the LoopBlocks list.
+ BasicBlock *NewPreheader = SplitEdge(loopPreheader, loopHeader, this);
+ LoopBlocks.push_back(NewPreheader);
+
+ // We want the loop to come after the preheader, but before the exit blocks.
+ LoopBlocks.insert(LoopBlocks.end(), L->block_begin(), L->block_end());
+
+ SmallVector<BasicBlock*, 8> ExitBlocks;
+ L->getUniqueExitBlocks(ExitBlocks);
+
+ // Split all of the edges from inside the loop to their exit blocks. Update
+ // the appropriate Phi nodes as we do so.
+ SplitExitEdges(L, ExitBlocks);
+
// The exit blocks may have been changed due to edge splitting, recompute.
ExitBlocks.clear();
L->getUniqueExitBlocks(ExitBlocks);
// Next step, clone all of the basic blocks that make up the loop (including
// the loop preheader and exit blocks), keeping track of the mapping between
// the instructions and blocks.
- std::vector<BasicBlock*> NewBlocks;
NewBlocks.reserve(LoopBlocks.size());
- std::map<const Value*, Value*> ValueMap;
+ DenseMap<const Value*, Value*> ValueMap;
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);
}
// Splice the newly inserted blocks into the function right before the
NewBlocks[0], F->end());
// Now we create the new Loop object for the versioned loop.
- Loop *NewLoop = CloneLoop(L, L->getParentLoop(), ValueMap, LI);
+ Loop *NewLoop = CloneLoop(L, L->getParentLoop(), ValueMap, LI, LPM);
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);
+ ParentLoop->addBasicBlockToLoop(NewBlocks[0], LI->getBase());
}
for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) {
BasicBlock *NewExit = cast<BasicBlock>(ValueMap[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);
+ 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) {
Value *V = PN->getIncomingValueForBlock(ExitBlocks[i]);
- std::map<const Value *, Value*>::iterator It = ValueMap.find(V);
+ DenseMap<const Value *, Value*>::iterator It = ValueMap.find(V);
if (It != ValueMap.end()) V = It->second;
PN->addIncoming(V, NewExit);
}
RemapInstruction(I, ValueMap);
// Rewrite the original preheader to select between versions of the loop.
- BranchInst *OldBR = cast<BranchInst>(OrigPreheader->getTerminator());
+ BranchInst *OldBR = cast<BranchInst>(loopPreheader->getTerminator());
assert(OldBR->isUnconditional() && OldBR->getSuccessor(0) == LoopBlocks[0] &&
"Preheader splitting did not work correctly!");
// Emit the new branch that selects between the two versions of this loop.
EmitPreheaderBranchOnCondition(LIC, Val, NewBlocks[0], LoopBlocks[0], OldBR);
+ LPM->deleteSimpleAnalysisValue(OldBR, L);
OldBR->eraseFromParent();
-
- LoopProcessWorklist.push_back(L);
+
LoopProcessWorklist.push_back(NewLoop);
+ redoLoop = true;
// Now we rewrite the original code to know that the condition is true and the
// new code to know that the condition is false.
// deleted. If so, don't simplify it.
if (!LoopProcessWorklist.empty() && LoopProcessWorklist.back() == NewLoop)
RewriteLoopBodyWithConditionConstant(NewLoop, LIC, Val, true);
+
}
/// RemoveFromWorklist - Remove all instances of I from the worklist vector
/// 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,
- std::vector<Instruction*> &Worklist) {
+ std::vector<Instruction*> &Worklist,
+ Loop *L, LPPassManager *LPM) {
DOUT << "Replace with '" << *V << "': " << *I;
// Add uses to the worklist, which may be dead now.
for (Value::use_iterator UI = I->use_begin(), E = I->use_end();
UI != E; ++UI)
Worklist.push_back(cast<Instruction>(*UI));
+ LPM->deleteSimpleAnalysisValue(I, L);
+ RemoveFromWorklist(I, Worklist);
I->replaceAllUsesWith(V);
I->eraseFromParent();
- RemoveFromWorklist(I, Worklist);
++NumSimplify;
}
/// information, and remove any dead successors it has.
///
void LoopUnswitch::RemoveBlockIfDead(BasicBlock *BB,
- std::vector<Instruction*> &Worklist) {
+ std::vector<Instruction*> &Worklist,
+ Loop *L) {
if (pred_begin(BB) != pred_end(BB)) {
// This block isn't dead, since an edge to BB was just removed, see if there
// are any easy simplifications we can do now.
while (isa<PHINode>(BB->begin()))
ReplaceUsesOfWith(BB->begin(),
cast<PHINode>(BB->begin())->getIncomingValue(0),
- Worklist);
+ 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))
- if (L->getHeader() == BB && L->contains(Pred)) {
+ if (loopHeader == BB && L->contains(Pred)) {
// Remove the branch from the latch to the header block, this makes
// the header dead, which will make the latch dead (because the header
// dominates the latch).
+ LPM->deleteSimpleAnalysisValue(Pred->getTerminator(), L);
Pred->getTerminator()->eraseFromParent();
new UnreachableInst(Pred);
RemoveLoopFromHierarchy(L);
// Reprocess the header, which now IS dead.
- RemoveBlockIfDead(BB, Worklist);
+ RemoveBlockIfDead(BB, Worklist, L);
return;
}
// 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.
+ 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,
// then getting removed before we revisit them, which is badness.
}
for (unsigned i = 0, e = Succs.size(); i != e; ++i)
- RemoveBlockIfDead(Succs[i], Worklist);
+ RemoveBlockIfDead(Succs[i], Worklist, L);
}
/// RemoveLoopFromHierarchy - We have discovered that the specified loop has
/// so they just reparent loops. If the loops are actually dead, they will be
/// removed later.
void LoopUnswitch::RemoveLoopFromHierarchy(Loop *L) {
- if (Loop *ParentLoop = L->getParentLoop()) { // Not a top-level loop.
- // Reparent all of the blocks in this loop. Since BBLoop had a parent,
- // they are now all in it.
- for (Loop::block_iterator I = L->block_begin(), E = L->block_end();
- I != E; ++I)
- if (LI->getLoopFor(*I) == L) // Don't change blocks in subloops.
- LI->changeLoopFor(*I, ParentLoop);
-
- // Remove the loop from its parent loop.
- for (Loop::iterator I = ParentLoop->begin(), E = ParentLoop->end();;
- ++I) {
- assert(I != E && "Couldn't find loop");
- if (*I == L) {
- ParentLoop->removeChildLoop(I);
- break;
- }
- }
-
- // Move all subloops into the parent loop.
- while (L->begin() != L->end())
- ParentLoop->addChildLoop(L->removeChildLoop(L->end()-1));
- } else {
- // Reparent all of the blocks in this loop. Since BBLoop had no parent,
- // they no longer in a loop at all.
-
- for (unsigned i = 0; i != L->getBlocks().size(); ++i) {
- // Don't change blocks in subloops.
- if (LI->getLoopFor(L->getBlocks()[i]) == L) {
- LI->removeBlock(L->getBlocks()[i]);
- --i;
- }
- }
-
- // Remove the loop from the top-level LoopInfo object.
- for (LoopInfo::iterator I = LI->begin(), E = LI->end();; ++I) {
- assert(I != E && "Couldn't find loop");
- if (*I == L) {
- LI->removeLoop(I);
- break;
- }
- }
-
- // Move all of the subloops to the top-level.
- while (L->begin() != L->end())
- LI->addTopLevelLoop(L->removeChildLoop(L->end()-1));
- }
-
- delete L;
+ LPM->deleteLoopFromQueue(L);
RemoveLoopFromWorklist(L);
}
-
-
// RewriteLoopBodyWithConditionConstant - We know either that the value LIC has
// the value specified by Val in the specified loop, or we know it does NOT have
// that value. Rewrite any uses of LIC or of properties correlated to 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.
// 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);
+ BasicBlock* Split = SplitBlock(Old, SI, this);
Instruction* OldTerm = Old->getTerminator();
- new BranchInst(Split, SI->getSuccessor(i),
- ConstantInt::getTrue(), OldTerm);
-
+ BranchInst::Create(Split, SISucc,
+ ConstantInt::getTrue(Context), OldTerm);
+
+ LPM->deleteSimpleAnalysisValue(Old->getTerminator(), L);
Old->getTerminator()->eraseFromParent();
-
PHINode *PN;
- for (BasicBlock::iterator II = SI->getSuccessor(i)->begin();
+ for (BasicBlock::iterator II = SISucc->begin();
(PN = dyn_cast<PHINode>(II)); ++II) {
Value *InVal = PN->removeIncomingValue(Split, false);
PN->addIncoming(InVal, Old);
}
}
- SimplifyCode(Worklist);
+ SimplifyCode(Worklist, L);
}
/// SimplifyCode - Okay, now that we have simplified some instructions in the
/// FIXME: When the loop optimizer is more mature, separate this out to a new
/// pass.
///
-void LoopUnswitch::SimplifyCode(std::vector<Instruction*> &Worklist) {
+void LoopUnswitch::SimplifyCode(std::vector<Instruction*> &Worklist, Loop *L) {
while (!Worklist.empty()) {
Instruction *I = Worklist.back();
Worklist.pop_back();
// Simple constant folding.
- if (Constant *C = ConstantFoldInstruction(I)) {
- ReplaceUsesOfWith(I, C, Worklist);
+ if (Constant *C = ConstantFoldInstruction(I, I->getContext())) {
+ ReplaceUsesOfWith(I, C, Worklist, L, LPM);
continue;
}
for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
if (Instruction *Use = dyn_cast<Instruction>(I->getOperand(i)))
Worklist.push_back(Use);
- I->eraseFromParent();
+ LPM->deleteSimpleAnalysisValue(I, L);
RemoveFromWorklist(I, Worklist);
+ I->eraseFromParent();
++NumSimplify;
continue;
}
switch (I->getOpcode()) {
case Instruction::Select:
if (ConstantInt *CB = dyn_cast<ConstantInt>(I->getOperand(0))) {
- ReplaceUsesOfWith(I, I->getOperand(!CB->getZExtValue()+1), Worklist);
+ ReplaceUsesOfWith(I, I->getOperand(!CB->getZExtValue()+1), Worklist, L,
+ LPM);
continue;
}
break;
cast<BinaryOperator>(I)->swapOperands();
if (ConstantInt *CB = dyn_cast<ConstantInt>(I->getOperand(1)))
if (CB->getType() == Type::Int1Ty) {
- if (CB->getZExtValue()) // X & 1 -> X
- ReplaceUsesOfWith(I, I->getOperand(0), Worklist);
+ if (CB->isOne()) // X & 1 -> X
+ ReplaceUsesOfWith(I, I->getOperand(0), Worklist, L, LPM);
else // X & 0 -> 0
- ReplaceUsesOfWith(I, I->getOperand(1), Worklist);
+ ReplaceUsesOfWith(I, I->getOperand(1), Worklist, L, LPM);
continue;
}
break;
cast<BinaryOperator>(I)->swapOperands();
if (ConstantInt *CB = dyn_cast<ConstantInt>(I->getOperand(1)))
if (CB->getType() == Type::Int1Ty) {
- if (CB->getZExtValue()) // X | 1 -> 1
- ReplaceUsesOfWith(I, I->getOperand(1), Worklist);
+ if (CB->isOne()) // X | 1 -> 1
+ ReplaceUsesOfWith(I, I->getOperand(1), Worklist, L, LPM);
else // X | 0 -> X
- ReplaceUsesOfWith(I, I->getOperand(0), Worklist);
+ ReplaceUsesOfWith(I, I->getOperand(0), Worklist, L, LPM);
continue;
}
break;
if (!SinglePred) continue; // Nothing to do.
assert(SinglePred == Pred && "CFG broken");
- DOUT << "Merging blocks: " << Pred->getName() << " <- "
- << Succ->getName() << "\n";
+ DEBUG(errs() << "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);
+ ReplaceUsesOfWith(PN, PN->getIncomingValue(0), Worklist, L, LPM);
// 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);
// 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())){
BasicBlock *DeadSucc = BI->getSuccessor(CB->getZExtValue());
BasicBlock *LiveSucc = BI->getSuccessor(!CB->getZExtValue());
DeadSucc->removePredecessor(BI->getParent(), true);
- Worklist.push_back(new BranchInst(LiveSucc, BI));
+ Worklist.push_back(BranchInst::Create(LiveSucc, BI));
+ LPM->deleteSimpleAnalysisValue(BI, L);
BI->eraseFromParent();
RemoveFromWorklist(BI, Worklist);
++NumSimplify;
- RemoveBlockIfDead(DeadSucc, Worklist);
+ RemoveBlockIfDead(DeadSucc, Worklist, L);
}
break;
}