//
// 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/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/PostOrderIterator.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);
+// 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(50), cl::Hidden);
- class VISIBILITY_HIDDEN LoopUnswitch : public LoopPass {
+namespace {
+ class LoopUnswitch : public LoopPass {
LoopInfo *LI; // Loop information
LPPassManager *LPM;
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:
static char ID; // Pass ID, replacement for typeid
explicit LoopUnswitch(bool Os = false) :
- LoopPass((intptr_t)&ID), OptimizeForSize(Os), redoLoop(false) {}
+ LoopPass(&ID), OptimizeForSize(Os), redoLoop(false),
+ currentLoop(NULL), DF(NULL), DT(NULL), loopHeader(NULL),
+ loopPreheader(NULL) {}
bool runOnLoop(Loop *L, LPPassManager &LPM);
- bool processLoop(Loop *L);
+ bool processCurrentLoop();
/// This transformation requires natural loop information & requires that
/// loop preheaders be inserted into the CFG...
private:
+ virtual void releaseMemory() {
+ UnswitchedVals.clear();
+ }
+
/// 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);
void UnswitchTrivialCondition(Loop *L, Value *Cond, Constant *Val,
BasicBlock *ExitBlock);
void UnswitchNontrivialCondition(Value *LIC, Constant *OnVal, Loop *L);
void RemoveBlockIfDead(BasicBlock *BB,
std::vector<Instruction*> &Worklist, Loop *l);
void RemoveLoopFromHierarchy(Loop *L);
+ bool IsTrivialUnswitchCondition(Value *Cond, Constant **Val = 0,
+ BasicBlock **LoopExit = 0);
+
};
- char LoopUnswitch::ID = 0;
- RegisterPass<LoopUnswitch> X("loop-unswitch", "Unswitch loops");
}
+char LoopUnswitch::ID = 0;
+static RegisterPass<LoopUnswitch> X("loop-unswitch", "Unswitch loops");
-LoopPass *llvm::createLoopUnswitchPass(bool Os) {
+Pass *llvm::createLoopUnswitchPass(bool Os) {
return new LoopUnswitch(Os);
}
/// 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::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 |= processLoop(L);
+ 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;
}
-/// processLoop - Do actual work and unswitch loop if possible and profitable.
-bool LoopUnswitch::processLoop(Loop *L) {
- assert(L->isLCSSAForm());
+/// 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!
if (!UnswitchedVals.insert(UnswitchVal))
continue;
- if (UnswitchIfProfitable(LoopCond, UnswitchVal, L)) {
+ 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.
-///
-unsigned LoopUnswitch::getLoopUnswitchCost(Loop *L, Value *LIC) {
- // If the condition is trivial, always unswitch. There is no code growth for
- // this case.
- if (IsTrivialUnswitchCondition(L, 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();
-
- 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();
- 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 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){
- // Do not do non-trivial unswitch while optimizing for size.
- if (Cost && OptimizeForSize)
- return false;
+ initLoopData();
- 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";
+ // If LoopSimplify was unable to form a preheader, don't do any unswitching.
+ if (!loopPreheader)
return false;
+
+ 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 (OptimizeForSize || F->hasFnAttr(Attribute::OptimizeForSize))
+ 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;
+ }
}
-
- // 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;
}
}
}
-// CloneDomInfo - NewBB is cloned from Orig basic block. Now clone Dominator Info.
-//
-// If Orig block's immediate dominator is mapped in VM then use corresponding
-// immediate dominator from the map. Otherwise Orig block's dominator is also
-// NewBB's dominator.
-//
-// OrigPreheader is loop pre-header before this pass started
-// updating CFG. NewPrehader is loops new pre-header. However, after CFG
-// manipulation, loop L may not exist. So rely on input parameter NewPreheader.
-void CloneDomInfo(BasicBlock *NewBB, BasicBlock *Orig,
- BasicBlock *NewPreheader, BasicBlock *OrigPreheader,
- BasicBlock *OrigHeader,
- DominatorTree *DT, DominanceFrontier *DF,
- DenseMap<const Value*, Value*> &VM) {
-
- // If NewBB alreay has found its place in domiantor tree then no need to do
- // anything.
- if (DT->getNode(NewBB))
- return;
-
- // If Orig does not have any immediate domiantor then its clone, NewBB, does
- // not need any immediate dominator.
- DomTreeNode *OrigNode = DT->getNode(Orig);
- if (!OrigNode)
- return;
- DomTreeNode *OrigIDomNode = OrigNode->getIDom();
- if (!OrigIDomNode)
- return;
-
- BasicBlock *OrigIDom = NULL;
-
- // If Orig is original loop header then its immediate dominator is
- // NewPreheader.
- if (Orig == OrigHeader)
- OrigIDom = NewPreheader;
-
- // If Orig is new pre-header then its immediate dominator is
- // original pre-header.
- else if (Orig == NewPreheader)
- OrigIDom = OrigPreheader;
-
- // Other as DT to find Orig's immediate dominator.
- else
- OrigIDom = OrigIDomNode->getBlock();
-
- // Initially use Orig's immediate dominator as NewBB's immediate dominator.
- BasicBlock *NewIDom = OrigIDom;
- DenseMap<const Value*, Value*>::iterator I = VM.find(OrigIDom);
- if (I != VM.end()) {
- NewIDom = cast<BasicBlock>(I->second);
-
- // If NewIDom does not have corresponding dominatore tree node then
- // get one.
- if (!DT->getNode(NewIDom))
- CloneDomInfo(NewIDom, OrigIDom, NewPreheader, OrigPreheader,
- OrigHeader, DT, DF, VM);
- }
-
- DT->addNewBlock(NewBB, NewIDom);
-
- // Copy cloned dominance frontiner set
- DominanceFrontier::DomSetType NewDFSet;
- if (DF) {
- DominanceFrontier::iterator DFI = DF->find(Orig);
- if ( DFI != DF->end()) {
- DominanceFrontier::DomSetType S = DFI->second;
- for (DominanceFrontier::DomSetType::iterator I = S.begin(), E = S.end();
- I != E; ++I) {
- BasicBlock *BB = *I;
- DenseMap<const Value*, Value*>::iterator IDM = VM.find(BB);
- if (IDM != VM.end())
- NewDFSet.insert(cast<BasicBlock>(IDM->second));
- else
- NewDFSet.insert(BB);
- }
- }
- DF->addBasicBlock(NewBB, NewDFSet);
- }
-}
-
/// 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,
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)
// 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.
- new BranchInst(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
/// 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(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 '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(), this);
+ 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
// 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();
- LPM->deleteSimpleAnalysisValue(OrigPH->getTerminator(), L);
+ loopPreheader->getTerminator());
+ LPM->deleteSimpleAnalysisValue(loopPreheader->getTerminator(), L);
+ loopPreheader->getTerminator()->eraseFromParent();
// We need to reprocess this loop, it could be unswitched again.
redoLoop = true;
++NumTrivial;
}
-/// 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.
+/// 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)
+{
+
+ 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);
+ }
+}
+
+/// 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 = L->getHeader()->getParent();
- DOUT << "loop-unswitch: Unswitching loop %"
- << L->getHeader()->getName() << " [" << L->getBlocks().size()
- << " blocks] in Function " << F->getName()
- << " when '" << *Val << "' == " << *LIC << "\n";
+ Function *F = loopHeader->getParent();
+ DEBUG(dbgs() << "loop-unswitch: Unswitching loop %"
+ << loopHeader->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;
+ LoopBlocks.clear();
+ NewBlocks.clear();
// First step, split the preheader and exit blocks, and add these blocks to
// the LoopBlocks list.
- BasicBlock *OrigHeader = L->getHeader();
- BasicBlock *OrigPreheader = L->getLoopPreheader();
- BasicBlock *NewPreheader = SplitEdge(OrigPreheader, L->getHeader(), this);
+ 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());
- std::vector<BasicBlock*> ExitBlocks;
+ 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.
- SmallVector<BasicBlock *,8> MiddleBlocks;
- 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* MiddleBlock = SplitEdge(Preds[j], ExitBlock, this);
- MiddleBlocks.push_back(MiddleBlock);
- BasicBlock* StartBlock = Preds[j];
- BasicBlock* EndBlock;
- if (MiddleBlock->getSinglePredecessor() == ExitBlock) {
- EndBlock = MiddleBlock;
- MiddleBlock = 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(MiddleBlock);
- PHINode* NewLCSSA = new PHINode(OldLCSSA->getType(),
- OldLCSSA->getName() + ".us-lcssa",
- MiddleBlock->getTerminator());
- NewLCSSA->addIncoming(OldValue, StartBlock);
- OldLCSSA->setIncomingValue(OldLCSSA->getBasicBlockIndex(MiddleBlock),
- NewLCSSA);
- InsertedPHIs.insert(NewLCSSA);
- }
+ SplitExitEdges(L, ExitBlocks);
- BasicBlock::iterator InsertPt = EndBlock->begin();
- while (dyn_cast<PHINode>(InsertPt)) ++InsertPt;
- for (BasicBlock::iterator I = MiddleBlock->begin();
- (OldLCSSA = dyn_cast<PHINode>(I)) && InsertedPHIs.count(OldLCSSA) == 0;
- ++I) {
- PHINode *NewLCSSA = new PHINode(OldLCSSA->getType(),
- OldLCSSA->getName() + ".us-lcssa",
- InsertPt);
- OldLCSSA->replaceAllUsesWith(NewLCSSA);
- NewLCSSA->addIncoming(OldLCSSA, MiddleBlock);
- }
- }
- }
-
// The exit blocks may have been changed due to edge splitting, recompute.
ExitBlocks.clear();
L->getUniqueExitBlocks(ExitBlocks);
// Add exit blocks to the loop blocks.
LoopBlocks.insert(LoopBlocks.end(), ExitBlocks.begin(), ExitBlocks.end());
- DominanceFrontier *DF = getAnalysisToUpdate<DominanceFrontier>();
- DominatorTree *DT = getAnalysisToUpdate<DominatorTree>();
-
// 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());
DenseMap<const Value*, Value*> ValueMap;
for (unsigned i = 0, e = LoopBlocks.size(); i != e; ++i) {
LPM->cloneBasicBlockSimpleAnalysis(LoopBlocks[i], New, L);
}
- // OutSiders are basic block that are dominated by original header and
- // at the same time they are not part of loop.
- SmallPtrSet<BasicBlock *, 8> OutSiders;
- if (DT) {
- DomTreeNode *OrigHeaderNode = DT->getNode(OrigHeader);
- for(std::vector<DomTreeNode*>::iterator DI = OrigHeaderNode->begin(),
- DE = OrigHeaderNode->end(); DI != DE; ++DI) {
- BasicBlock *B = (*DI)->getBlock();
-
- DenseMap<const Value*, Value*>::iterator VI = ValueMap.find(B);
- if (VI == ValueMap.end())
- OutSiders.insert(B);
- }
- }
-
// Splice the newly inserted blocks into the function right before the
// original preheader.
F->getBasicBlockList().splice(LoopBlocks[0], F->getBasicBlockList(),
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;
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);
- OldBR->eraseFromParent();
LPM->deleteSimpleAnalysisValue(OldBR, L);
-
- // Update dominator info
- if (DF && DT) {
-
- // Clone dominator info for all cloned basic block.
- for (unsigned i = 0, e = LoopBlocks.size(); i != e; ++i) {
- BasicBlock *LBB = LoopBlocks[i];
- BasicBlock *NBB = NewBlocks[i];
- CloneDomInfo(NBB, LBB, NewPreheader, OrigPreheader,
- OrigHeader, DT, DF, ValueMap);
-
- // Remove any OutSiders from LBB and NBB's dominance frontier.
- DominanceFrontier::iterator LBBI = DF->find(LBB);
- if (LBBI != DF->end()) {
- DominanceFrontier::DomSetType &LBSet = LBBI->second;
- for (DominanceFrontier::DomSetType::iterator LI = LBSet.begin(),
- LE = LBSet.end(); LI != LE; ++LI) {
- BasicBlock *B = *LI;
- if (OutSiders.count(B))
- DF->removeFromFrontier(LBBI, B);
- }
- }
-
- // Remove any OutSiders from LBB and NBB's dominance frontier.
- DominanceFrontier::iterator NBBI = DF->find(NBB);
- if (NBBI != DF->end()) {
- DominanceFrontier::DomSetType NBSet = NBBI->second;
- for (DominanceFrontier::DomSetType::iterator NI = NBSet.begin(),
- NE = NBSet.end(); NI != NE; ++NI) {
- BasicBlock *B = *NI;
- if (OutSiders.count(B))
- DF->removeFromFrontier(NBBI, B);
- }
- }
- }
-
- // MiddleBlocks are dominated by original pre header. SplitEdge updated
- // MiddleBlocks' dominance frontier appropriately.
- for (unsigned i = 0, e = MiddleBlocks.size(); i != e; ++i) {
- BasicBlock *MBB = MiddleBlocks[i];
- if (!MBB->getSinglePredecessor())
- DT->changeImmediateDominator(MBB, OrigPreheader);
- }
-
- // All Outsiders are now dominated by original pre header.
- for (SmallPtrSet<BasicBlock *, 8>::iterator OI = OutSiders.begin(),
- OE = OutSiders.end(); OI != OE; ++OI) {
- BasicBlock *OB = *OI;
- DT->changeImmediateDominator(OB, OrigPreheader);
- }
-
- // New loop headers are dominated by original preheader
- DT->changeImmediateDominator(NewBlocks[0], OrigPreheader);
- DT->changeImmediateDominator(LoopBlocks[0], OrigPreheader);
- }
+ OldBR->eraseFromParent();
LoopProcessWorklist.push_back(NewLoop);
redoLoop = true;
// 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
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)
for (Value::use_iterator UI = I->use_begin(), E = I->use_end();
UI != E; ++UI)
Worklist.push_back(cast<Instruction>(*UI));
- I->replaceAllUsesWith(V);
- I->eraseFromParent();
LPM->deleteSimpleAnalysisValue(I, L);
RemoveFromWorklist(I, Worklist);
+ I->replaceAllUsesWith(V);
+ I->eraseFromParent();
++NumSimplify;
}
// 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).
- Pred->getTerminator()->eraseFromParent();
LPM->deleteSimpleAnalysisValue(Pred->getTerminator(), L);
- new UnreachableInst(Pred);
+ Pred->getTerminator()->eraseFromParent();
+ 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.
- BB->eraseFromParent();
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.
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.
- 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* Old = SI->getParent();
- BasicBlock* Split = SplitBlock(Old, SI, this);
-
- Instruction* OldTerm = Old->getTerminator();
- new BranchInst(Split, SI->getSuccessor(i),
- ConstantInt::getTrue(), OldTerm);
-
- Old->getTerminator()->eraseFromParent();
-
-
- PHINode *PN;
- for (BasicBlock::iterator II = SI->getSuccessor(i)->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)
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, L, LPM);
+ ReplaceUsesOfWith(I, I->getOperand(!CB->getZExtValue()+1), Worklist, L,
+ LPM);
continue;
}
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()))
// Move all of the successor contents from Succ to Pred.
Pred->getInstList().splice(BI, Succ->getInstList(), Succ->begin(),
Succ->end());
- BI->eraseFromParent();
LPM->deleteSimpleAnalysisValue(BI, L);
+ BI->eraseFromParent();
RemoveFromWorklist(BI, Worklist);
// If Succ has any successors with PHI nodes, update them to have
// Remove Succ from the loop tree.
LI->removeBlock(Succ);
- Succ->eraseFromParent();
LPM->deleteSimpleAnalysisValue(Succ, L);
+ Succ->eraseFromParent();
++NumSimplify;
} else if (ConstantInt *CB = dyn_cast<ConstantInt>(BI->getCondition())){
// Conditional branch. Turn it into an unconditional branch, then
// 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);
- Worklist.push_back(new BranchInst(LiveSucc, BI));
- BI->eraseFromParent();
+ Worklist.push_back(BranchInst::Create(LiveSucc, BI));
LPM->deleteSimpleAnalysisValue(BI, L);
+ BI->eraseFromParent();
RemoveFromWorklist(BI, Worklist);
++NumSimplify;