//===-- LICM.cpp - Loop Invariant Code Motion Pass ------------------------===//
-//
+//
// The LLVM Compiler Infrastructure
//
-// This file was developed by the LLVM research group and is distributed under
-// the University of Illinois Open Source License. See LICENSE.TXT for details.
-//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
//===----------------------------------------------------------------------===//
//
// This pass performs loop invariant code motion, attempting to remove as much
//
//===----------------------------------------------------------------------===//
+#define DEBUG_TYPE "licm"
#include "llvm/Transforms/Scalar.h"
+#include "llvm/Constants.h"
#include "llvm/DerivedTypes.h"
#include "llvm/Instructions.h"
+#include "llvm/LLVMContext.h"
#include "llvm/Target/TargetData.h"
#include "llvm/Analysis/LoopInfo.h"
+#include "llvm/Analysis/LoopPass.h"
#include "llvm/Analysis/AliasAnalysis.h"
#include "llvm/Analysis/AliasSetTracker.h"
#include "llvm/Analysis/Dominators.h"
-#include "llvm/Support/CFG.h"
+#include "llvm/Analysis/ScalarEvolution.h"
#include "llvm/Transforms/Utils/PromoteMemToReg.h"
-#include "llvm/Transforms/Utils/Local.h"
+#include "llvm/Support/CFG.h"
+#include "llvm/Support/Compiler.h"
#include "llvm/Support/CommandLine.h"
+#include "llvm/Support/raw_ostream.h"
#include "llvm/Support/Debug.h"
#include "llvm/ADT/Statistic.h"
#include <algorithm>
using namespace llvm;
-namespace {
- cl::opt<bool>
- DisablePromotion("disable-licm-promotion", cl::Hidden,
- cl::desc("Disable memory promotion in LICM pass"));
+STATISTIC(NumSunk , "Number of instructions sunk out of loop");
+STATISTIC(NumHoisted , "Number of instructions hoisted out of loop");
+STATISTIC(NumMovedLoads, "Number of load insts hoisted or sunk");
+STATISTIC(NumMovedCalls, "Number of call insts hoisted or sunk");
+STATISTIC(NumPromoted , "Number of memory locations promoted to registers");
+
+static cl::opt<bool>
+DisablePromotion("disable-licm-promotion", cl::Hidden,
+ cl::desc("Disable memory promotion in LICM pass"));
+
+// This feature is currently disabled by default because CodeGen is not yet
+// capable of rematerializing these constants in PIC mode, so it can lead to
+// degraded performance. Compile test/CodeGen/X86/remat-constant.ll with
+// -relocation-model=pic to see an example of this.
+static cl::opt<bool>
+EnableLICMConstantMotion("enable-licm-constant-variables", cl::Hidden,
+ cl::desc("Enable hoisting/sinking of constant "
+ "global variables"));
- Statistic<> NumSunk("licm", "Number of instructions sunk out of loop");
- Statistic<> NumHoisted("licm", "Number of instructions hoisted out of loop");
- Statistic<> NumMovedLoads("licm", "Number of load insts hoisted or sunk");
- Statistic<> NumMovedCalls("licm", "Number of call insts hoisted or sunk");
- Statistic<> NumPromoted("licm",
- "Number of memory locations promoted to registers");
+namespace {
+ struct VISIBILITY_HIDDEN LICM : public LoopPass {
+ static char ID; // Pass identification, replacement for typeid
+ LICM() : LoopPass(&ID) {}
- struct LICM : public FunctionPass {
- virtual bool runOnFunction(Function &F);
+ virtual bool runOnLoop(Loop *L, LPPassManager &LPM);
/// This transformation requires natural loop information & requires that
/// loop preheaders be inserted into the CFG...
AU.addRequired<DominatorTree>();
AU.addRequired<DominanceFrontier>(); // For scalar promotion (mem2reg)
AU.addRequired<AliasAnalysis>();
+ AU.addPreserved<ScalarEvolution>();
+ AU.addPreserved<DominanceFrontier>();
+ }
+
+ bool doFinalization() {
+ // Free the values stored in the map
+ for (std::map<Loop *, AliasSetTracker *>::iterator
+ I = LoopToAliasMap.begin(), E = LoopToAliasMap.end(); I != E; ++I)
+ delete I->second;
+
+ LoopToAliasMap.clear();
+ return false;
}
private:
BasicBlock *Preheader; // The preheader block of the current loop...
Loop *CurLoop; // The current loop we are working on...
AliasSetTracker *CurAST; // AliasSet information for the current loop...
+ std::map<Loop *, AliasSetTracker *> LoopToAliasMap;
- /// visitLoop - Hoist expressions out of the specified loop...
- ///
- void visitLoop(Loop *L, AliasSetTracker &AST);
+ /// cloneBasicBlockAnalysis - Simple Analysis hook. Clone alias set info.
+ void cloneBasicBlockAnalysis(BasicBlock *From, BasicBlock *To, Loop *L);
+
+ /// deleteAnalysisValue - Simple Analysis hook. Delete value V from alias
+ /// set.
+ void deleteAnalysisValue(Value *V, Loop *L);
/// SinkRegion - Walk the specified region of the CFG (defined by all blocks
/// dominated by the specified block, and that are in the current loop) in
/// visit uses before definitions, allowing us to sink a loop body in one
/// pass without iteration.
///
- void SinkRegion(DominatorTree::Node *N);
+ void SinkRegion(DomTreeNode *N);
/// HoistRegion - Walk the specified region of the CFG (defined by all
/// blocks dominated by the specified block, and that are in the current
/// visit definitions before uses, allowing us to hoist a loop body in one
/// pass without iteration.
///
- void HoistRegion(DominatorTree::Node *N);
+ void HoistRegion(DomTreeNode *N);
/// inSubLoop - Little predicate that returns true if the specified basic
/// block is in a subloop of the current one, not the current one itself.
BasicBlock *LoopHeader = CurLoop->getHeader();
if (BlockInLoop == LoopHeader)
return true;
-
- DominatorTree::Node *BlockInLoopNode = DT->getNode(BlockInLoop);
- DominatorTree::Node *IDom = DT->getNode(ExitBlock);
-
+
+ DomTreeNode *BlockInLoopNode = DT->getNode(BlockInLoop);
+ DomTreeNode *IDom = DT->getNode(ExitBlock);
+
// Because the exit block is not in the loop, we know we have to get _at
// least_ its immediate dominator.
do {
// Get next Immediate Dominator.
IDom = IDom->getIDom();
-
+
// If we have got to the header of the loop, then the instructions block
// did not dominate the exit node, so we can't hoist it.
if (IDom->getBlock() == LoopHeader)
return false;
-
+
} while (IDom != BlockInLoopNode);
return true;
/// pointerInvalidatedByLoop - Return true if the body of this loop may
/// store into the memory location pointed to by V.
- ///
- bool pointerInvalidatedByLoop(Value *V) {
+ ///
+ bool pointerInvalidatedByLoop(Value *V, unsigned Size) {
// Check to see if any of the basic blocks in CurLoop invalidate *V.
- return CurAST->getAliasSetForPointer(V, 0).isMod();
+ return CurAST->getAliasSetForPointer(V, Size).isMod();
}
bool canSinkOrHoistInst(Instruction &I);
std::vector<std::pair<AllocaInst*, Value*> > &PromotedValues,
std::map<Value*, AllocaInst*> &Val2AlMap);
};
-
- RegisterOpt<LICM> X("licm", "Loop Invariant Code Motion");
}
-FunctionPass *llvm::createLICMPass() { return new LICM(); }
+char LICM::ID = 0;
+static RegisterPass<LICM> X("licm", "Loop Invariant Code Motion");
-/// runOnFunction - For LICM, this simply traverses the loop structure of the
-/// function, hoisting expressions out of loops if possible.
+Pass *llvm::createLICMPass() { return new LICM(); }
+
+/// Hoist expressions out of the specified loop. Note, alias info for inner
+/// loop is not preserved so it is not a good idea to run LICM multiple
+/// times on one loop.
///
-bool LICM::runOnFunction(Function &) {
+bool LICM::runOnLoop(Loop *L, LPPassManager &LPM) {
Changed = false;
// Get our Loop and Alias Analysis information...
DF = &getAnalysis<DominanceFrontier>();
DT = &getAnalysis<DominatorTree>();
- // Hoist expressions out of all of the top-level loops.
- for (LoopInfo::iterator I = LI->begin(), E = LI->end(); I != E; ++I) {
- AliasSetTracker AST(*AA);
- visitLoop(*I, AST);
- }
- return Changed;
-}
-
+ CurAST = new AliasSetTracker(*AA);
+ // Collect Alias info from subloops
+ for (Loop::iterator LoopItr = L->begin(), LoopItrE = L->end();
+ LoopItr != LoopItrE; ++LoopItr) {
+ Loop *InnerL = *LoopItr;
+ AliasSetTracker *InnerAST = LoopToAliasMap[InnerL];
+ assert (InnerAST && "Where is my AST?");
-/// visitLoop - Hoist expressions out of the specified loop...
-///
-void LICM::visitLoop(Loop *L, AliasSetTracker &AST) {
- // Recurse through all subloops before we process this loop...
- for (Loop::iterator I = L->begin(), E = L->end(); I != E; ++I) {
- AliasSetTracker SubAST(*AA);
- visitLoop(*I, SubAST);
-
- // Incorporate information about the subloops into this loop...
- AST.add(SubAST);
+ // What if InnerLoop was modified by other passes ?
+ CurAST->add(*InnerAST);
}
+
CurLoop = L;
- CurAST = &AST;
// Get the preheader block to move instructions into...
Preheader = L->getLoopPreheader();
// Because subloops have already been incorporated into AST, we skip blocks in
// subloops.
//
- for (std::vector<BasicBlock*>::const_iterator I = L->getBlocks().begin(),
- E = L->getBlocks().end(); I != E; ++I)
- if (LI->getLoopFor(*I) == L) // Ignore blocks in subloops...
- AST.add(**I); // Incorporate the specified basic block
+ for (Loop::block_iterator I = L->block_begin(), E = L->block_end();
+ I != E; ++I) {
+ BasicBlock *BB = *I;
+ if (LI->getLoopFor(BB) == L) // Ignore blocks in subloops...
+ CurAST->add(*BB); // Incorporate the specified basic block
+ }
// We want to visit all of the instructions in this loop... that are not parts
// of our subloops (they have already had their invariants hoisted out of
//
// Traverse the body of the loop in depth first order on the dominator tree so
// that we are guaranteed to see definitions before we see uses. This allows
- // us to sink instructions in one pass, without iteration. AFter sinking
+ // us to sink instructions in one pass, without iteration. After sinking
// instructions, we perform another pass to hoist them out of the loop.
//
SinkRegion(DT->getNode(L->getHeader()));
// Clear out loops state information for the next iteration
CurLoop = 0;
Preheader = 0;
+
+ LoopToAliasMap[L] = CurAST;
+ return Changed;
}
/// SinkRegion - Walk the specified region of the CFG (defined by all blocks
/// uses before definitions, allowing us to sink a loop body in one pass without
/// iteration.
///
-void LICM::SinkRegion(DominatorTree::Node *N) {
+void LICM::SinkRegion(DomTreeNode *N) {
assert(N != 0 && "Null dominator tree node?");
BasicBlock *BB = N->getBlock();
if (!CurLoop->contains(BB)) return;
// We are processing blocks in reverse dfo, so process children first...
- const std::vector<DominatorTree::Node*> &Children = N->getChildren();
+ const std::vector<DomTreeNode*> &Children = N->getChildren();
for (unsigned i = 0, e = Children.size(); i != e; ++i)
SinkRegion(Children[i]);
for (BasicBlock::iterator II = BB->end(); II != BB->begin(); ) {
Instruction &I = *--II;
-
+
// Check to see if we can sink this instruction to the exit blocks
// of the loop. We can do this if the all users of the instruction are
// outside of the loop. In this case, it doesn't even matter if the
// operands of the instruction are loop invariant.
//
- if (canSinkOrHoistInst(I) && isNotUsedInLoop(I)) {
+ if (isNotUsedInLoop(I) && canSinkOrHoistInst(I)) {
++II;
sink(I);
}
/// first order w.r.t the DominatorTree. This allows us to visit definitions
/// before uses, allowing us to hoist a loop body in one pass without iteration.
///
-void LICM::HoistRegion(DominatorTree::Node *N) {
+void LICM::HoistRegion(DomTreeNode *N) {
assert(N != 0 && "Null dominator tree node?");
BasicBlock *BB = N->getBlock();
if (!inSubLoop(BB))
for (BasicBlock::iterator II = BB->begin(), E = BB->end(); II != E; ) {
Instruction &I = *II++;
-
+
// Try hoisting the instruction out to the preheader. We can only do this
// if all of the operands of the instruction are loop invariant and if it
// is safe to hoist the instruction.
//
- if (isLoopInvariantInst(I) && canSinkOrHoistInst(I) &&
+ if (isLoopInvariantInst(I) && canSinkOrHoistInst(I) &&
isSafeToExecuteUnconditionally(I))
- hoist(I);
+ hoist(I);
}
- const std::vector<DominatorTree::Node*> &Children = N->getChildren();
+ const std::vector<DomTreeNode*> &Children = N->getChildren();
for (unsigned i = 0, e = Children.size(); i != e; ++i)
HoistRegion(Children[i]);
}
if (LI->isVolatile())
return false; // Don't hoist volatile loads!
+ // Loads from constant memory are always safe to move, even if they end up
+ // in the same alias set as something that ends up being modified.
+ if (EnableLICMConstantMotion &&
+ AA->pointsToConstantMemory(LI->getOperand(0)))
+ return true;
+
// Don't hoist loads which have may-aliased stores in loop.
- return !pointerInvalidatedByLoop(LI->getOperand(0));
+ unsigned Size = 0;
+ if (LI->getType()->isSized())
+ Size = AA->getTypeStoreSize(LI->getType());
+ return !pointerInvalidatedByLoop(LI->getOperand(0), Size);
} else if (CallInst *CI = dyn_cast<CallInst>(&I)) {
// Handle obvious cases efficiently.
- if (Function *Callee = CI->getCalledFunction()) {
- if (AA->doesNotAccessMemory(Callee))
- return true;
- else if (AA->onlyReadsMemory(Callee)) {
- // If this call only reads from memory and there are no writes to memory
- // in the loop, we can hoist or sink the call as appropriate.
- bool FoundMod = false;
- for (AliasSetTracker::iterator I = CurAST->begin(), E = CurAST->end();
- I != E; ++I) {
- AliasSet &AS = *I;
- if (!AS.isForwardingAliasSet() && AS.isMod()) {
- FoundMod = true;
- break;
- }
+ AliasAnalysis::ModRefBehavior Behavior = AA->getModRefBehavior(CI);
+ if (Behavior == AliasAnalysis::DoesNotAccessMemory)
+ return true;
+ else if (Behavior == AliasAnalysis::OnlyReadsMemory) {
+ // If this call only reads from memory and there are no writes to memory
+ // in the loop, we can hoist or sink the call as appropriate.
+ bool FoundMod = false;
+ for (AliasSetTracker::iterator I = CurAST->begin(), E = CurAST->end();
+ I != E; ++I) {
+ AliasSet &AS = *I;
+ if (!AS.isForwardingAliasSet() && AS.isMod()) {
+ FoundMod = true;
+ break;
}
- if (!FoundMod) return true;
}
+ if (!FoundMod) return true;
}
// FIXME: This should use mod/ref information to see if we can hoist or sink
// the call.
-
+
return false;
}
- return isa<BinaryOperator>(I) || isa<ShiftInst>(I) || isa<CastInst>(I) ||
- isa<SelectInst>(I) ||
- isa<GetElementPtrInst>(I) || isa<VANextInst>(I) || isa<VAArgInst>(I);
+ // Otherwise these instructions are hoistable/sinkable
+ return isa<BinaryOperator>(I) || isa<CastInst>(I) ||
+ isa<SelectInst>(I) || isa<GetElementPtrInst>(I) || isa<CmpInst>(I) ||
+ isa<InsertElementInst>(I) || isa<ExtractElementInst>(I) ||
+ isa<ShuffleVectorInst>(I);
}
/// isNotUsedInLoop - Return true if the only users of this instruction are
/// position, and may either delete it or move it to outside of the loop.
///
void LICM::sink(Instruction &I) {
- DEBUG(std::cerr << "LICM sinking instruction: " << I);
+ DOUT << "LICM sinking instruction: " << I;
- std::vector<BasicBlock*> ExitBlocks;
+ SmallVector<BasicBlock*, 8> ExitBlocks;
CurLoop->getExitBlocks(ExitBlocks);
if (isa<LoadInst>(I)) ++NumMovedLoads;
++NumSunk;
Changed = true;
+ LLVMContext &Context = I.getContext();
+
// The case where there is only a single exit node of this loop is common
// enough that we handle it as a special (more efficient) case. It is more
// efficient to handle because there are no PHI nodes that need to be placed.
if (!isExitBlockDominatedByBlockInLoop(ExitBlocks[0], I.getParent())) {
// Instruction is not used, just delete it.
CurAST->deleteValue(&I);
- I.getParent()->getInstList().erase(&I);
+ if (!I.use_empty()) // If I has users in unreachable blocks, eliminate.
+ I.replaceAllUsesWith(UndefValue::get(I.getType()));
+ I.eraseFromParent();
} else {
// Move the instruction to the start of the exit block, after any PHI
// nodes in it.
- I.getParent()->getInstList().remove(&I);
-
- BasicBlock::iterator InsertPt = ExitBlocks[0]->begin();
- while (isa<PHINode>(InsertPt)) ++InsertPt;
+ I.removeFromParent();
+
+ BasicBlock::iterator InsertPt = ExitBlocks[0]->getFirstNonPHI();
ExitBlocks[0]->getInstList().insert(InsertPt, &I);
}
- } else if (ExitBlocks.size() == 0) {
+ } else if (ExitBlocks.empty()) {
// The instruction is actually dead if there ARE NO exit blocks.
CurAST->deleteValue(&I);
- I.getParent()->getInstList().erase(&I);
+ if (!I.use_empty()) // If I has users in unreachable blocks, eliminate.
+ I.replaceAllUsesWith(UndefValue::get(I.getType()));
+ I.eraseFromParent();
} else {
// Otherwise, if we have multiple exits, use the PromoteMem2Reg function to
// do all of the hard work of inserting PHI nodes as necessary. We convert
// Firstly, we create a stack object to hold the value...
AllocaInst *AI = 0;
- if (I.getType() != Type::VoidTy)
+ if (I.getType() != Type::VoidTy) {
AI = new AllocaInst(I.getType(), 0, I.getName(),
- I.getParent()->getParent()->front().begin());
-
+ I.getParent()->getParent()->getEntryBlock().begin());
+ CurAST->add(AI);
+ }
+
// Secondly, insert load instructions for each use of the instruction
// outside of the loop.
while (!I.use_empty()) {
// Insert a new load instruction right before the terminator in
// the predecessor block.
PredVal = new LoadInst(AI, "", Pred->getTerminator());
+ CurAST->add(cast<LoadInst>(PredVal));
}
UPN->setIncomingValue(i, PredVal);
} else {
LoadInst *L = new LoadInst(AI, "", U);
U->replaceUsesOfWith(&I, L);
+ CurAST->add(L);
}
}
// If we haven't already processed this exit block, do so now.
if (InsertedBlocks.insert(ExitBlock).second) {
// Insert the code after the last PHI node...
- BasicBlock::iterator InsertPt = ExitBlock->begin();
- while (isa<PHINode>(InsertPt)) ++InsertPt;
-
+ BasicBlock::iterator InsertPt = ExitBlock->getFirstNonPHI();
+
// If this is the first exit block processed, just move the original
// instruction, otherwise clone the original instruction and insert
// the copy.
Instruction *New;
if (InsertedBlocks.size() == 1) {
- I.getParent()->getInstList().remove(&I);
+ I.removeFromParent();
ExitBlock->getInstList().insert(InsertPt, &I);
New = &I;
} else {
- New = I.clone();
+ New = I.clone(Context);
+ CurAST->copyValue(&I, New);
if (!I.getName().empty())
New->setName(I.getName()+".le");
ExitBlock->getInstList().insert(InsertPt, New);
}
-
+
// Now that we have inserted the instruction, store it into the alloca
if (AI) new StoreInst(New, AI, InsertPt);
}
// If the instruction doesn't dominate any exit blocks, it must be dead.
if (InsertedBlocks.empty()) {
CurAST->deleteValue(&I);
- I.getParent()->getInstList().erase(&I);
+ I.eraseFromParent();
}
-
+
// Finally, promote the fine value to SSA form.
if (AI) {
std::vector<AllocaInst*> Allocas;
Allocas.push_back(AI);
- PromoteMemToReg(Allocas, *DT, *DF, AA->getTargetData(), CurAST);
+ PromoteMemToReg(Allocas, *DT, *DF, Context, CurAST);
}
}
}
/// that is safe to hoist, this instruction is called to do the dirty work.
///
void LICM::hoist(Instruction &I) {
- DEBUG(std::cerr << "LICM hoisting to " << Preheader->getName()
- << ": " << I);
+ DEBUG(errs() << "LICM hoisting to " << Preheader->getName() << ": " << I);
// Remove the instruction from its current basic block... but don't delete the
// instruction.
- I.getParent()->getInstList().remove(&I);
+ I.removeFromParent();
// Insert the new node in Preheader, before the terminator.
Preheader->getInstList().insert(Preheader->getTerminator(), &I);
-
+
if (isa<LoadInst>(I)) ++NumMovedLoads;
else if (isa<CallInst>(I)) ++NumMovedCalls;
++NumHoisted;
///
bool LICM::isSafeToExecuteUnconditionally(Instruction &Inst) {
// If it is not a trapping instruction, it is always safe to hoist.
- if (!Inst.isTrapping()) return true;
-
+ if (Inst.isSafeToSpeculativelyExecute())
+ return true;
+
// Otherwise we have to check to make sure that the instruction dominates all
// of the exit blocks. If it doesn't, then there is a path out of the loop
// which does not execute this instruction, so we can't hoist it.
return true;
// Get the exit blocks for the current loop.
- std::vector<BasicBlock*> ExitBlocks;
+ SmallVector<BasicBlock*, 8> ExitBlocks;
CurLoop->getExitBlocks(ExitBlocks);
// For each exit block, get the DT node and walk up the DT until the
for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i)
if (!isExitBlockDominatedByBlockInLoop(ExitBlocks[i], Inst.getParent()))
return false;
-
+
return true;
}
LoadValue = LI;
break;
} else if (StoreInst *SI = dyn_cast<StoreInst>(*UI)) {
- if (SI->getOperand(1) == LI) {
+ if (SI->getOperand(1) == Ptr) {
LoadValue = SI->getOperand(0);
break;
}
// Store into the temporary alloca.
new StoreInst(LI, PromotedValues[i].first, LoopPredInst);
}
-
+
// Scan the basic blocks in the loop, replacing uses of our pointers with
// uses of the allocas in question.
//
- const std::vector<BasicBlock*> &LoopBBs = CurLoop->getBlocks();
- for (std::vector<BasicBlock*>::const_iterator I = LoopBBs.begin(),
- E = LoopBBs.end(); I != E; ++I) {
+ for (Loop::block_iterator I = CurLoop->block_begin(),
+ E = CurLoop->block_end(); I != E; ++I) {
+ BasicBlock *BB = *I;
// Rewrite all loads and stores in the block of the pointer...
- for (BasicBlock::iterator II = (*I)->begin(), E = (*I)->end();
- II != E; ++II) {
+ for (BasicBlock::iterator II = BB->begin(), E = BB->end(); II != E; ++II) {
if (LoadInst *L = dyn_cast<LoadInst>(II)) {
std::map<Value*, AllocaInst*>::iterator
I = ValueToAllocaMap.find(L->getOperand(0));
// want to insert one copy of the code in each exit block, though the loop may
// exit to the same block more than once.
//
- std::set<BasicBlock*> ProcessedBlocks;
+ SmallPtrSet<BasicBlock*, 16> ProcessedBlocks;
- std::vector<BasicBlock*> ExitBlocks;
+ SmallVector<BasicBlock*, 8> ExitBlocks;
CurLoop->getExitBlocks(ExitBlocks);
- for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i)
- if (ProcessedBlocks.insert(ExitBlocks[i]).second) {
- // Copy all of the allocas into their memory locations.
- BasicBlock::iterator BI = ExitBlocks[i]->begin();
- while (isa<PHINode>(*BI))
- ++BI; // Skip over all of the phi nodes in the block.
- Instruction *InsertPos = BI;
- unsigned PVN = 0;
- for (unsigned i = 0, e = PromotedValues.size(); i != e; ++i) {
- // Load from the alloca.
- LoadInst *LI = new LoadInst(PromotedValues[i].first, "", InsertPos);
-
- // If this is a pointer type, update alias info appropriately.
- if (isa<PointerType>(LI->getType()))
- CurAST->copyValue(PointerValueNumbers[PVN++], LI);
-
- // Store into the memory we promoted.
- new StoreInst(LI, PromotedValues[i].second, InsertPos);
- }
+ for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) {
+ if (!ProcessedBlocks.insert(ExitBlocks[i]))
+ continue;
+
+ // Copy all of the allocas into their memory locations.
+ BasicBlock::iterator BI = ExitBlocks[i]->getFirstNonPHI();
+ Instruction *InsertPos = BI;
+ unsigned PVN = 0;
+ for (unsigned i = 0, e = PromotedValues.size(); i != e; ++i) {
+ // Load from the alloca.
+ LoadInst *LI = new LoadInst(PromotedValues[i].first, "", InsertPos);
+
+ // If this is a pointer type, update alias info appropriately.
+ if (isa<PointerType>(LI->getType()))
+ CurAST->copyValue(PointerValueNumbers[PVN++], LI);
+
+ // Store into the memory we promoted.
+ new StoreInst(LI, PromotedValues[i].second, InsertPos);
}
+ }
// Now that we have done the deed, use the mem2reg functionality to promote
// all of the new allocas we just created into real SSA registers.
PromotedAllocas.reserve(PromotedValues.size());
for (unsigned i = 0, e = PromotedValues.size(); i != e; ++i)
PromotedAllocas.push_back(PromotedValues[i].first);
- PromoteMemToReg(PromotedAllocas, *DT, *DF, AA->getTargetData(), CurAST);
+ PromoteMemToReg(PromotedAllocas, *DT, *DF, Preheader->getContext(), CurAST);
}
/// FindPromotableValuesInLoop - Check the current loop for stores to definite
-/// pointers, which are not loaded and stored through may aliases. If these are
-/// found, create an alloca for the value, add it to the PromotedValues list,
-/// and keep track of the mapping from value to alloca.
-///
+/// pointers, which are not loaded and stored through may aliases and are safe
+/// for promotion. If these are found, create an alloca for the value, add it
+/// to the PromotedValues list, and keep track of the mapping from value to
+/// alloca.
void LICM::FindPromotableValuesInLoop(
std::vector<std::pair<AllocaInst*, Value*> > &PromotedValues,
std::map<Value*, AllocaInst*> &ValueToAllocaMap) {
// We can promote this alias set if it has a store, if it is a "Must" alias
// set, if the pointer is loop invariant, and if we are not eliminating any
// volatile loads or stores.
- if (!AS.isForwardingAliasSet() && AS.isMod() && AS.isMustAlias() &&
- !AS.isVolatile() && CurLoop->isLoopInvariant(AS.begin()->first)) {
- assert(AS.begin() != AS.end() &&
- "Must alias set should have at least one pointer element in it!");
- Value *V = AS.begin()->first;
-
- // Check that all of the pointers in the alias set have the same type. We
- // cannot (yet) promote a memory location that is loaded and stored in
- // different sizes.
+ if (AS.isForwardingAliasSet() || !AS.isMod() || !AS.isMustAlias() ||
+ AS.isVolatile() || !CurLoop->isLoopInvariant(AS.begin()->getValue()))
+ continue;
+
+ assert(!AS.empty() &&
+ "Must alias set should have at least one pointer element in it!");
+ Value *V = AS.begin()->getValue();
+
+ // Check that all of the pointers in the alias set have the same type. We
+ // cannot (yet) promote a memory location that is loaded and stored in
+ // different sizes.
+ {
bool PointerOk = true;
for (AliasSet::iterator I = AS.begin(), E = AS.end(); I != E; ++I)
- if (V->getType() != I->first->getType()) {
+ if (V->getType() != I->getValue()->getType()) {
PointerOk = false;
break;
}
+ if (!PointerOk)
+ continue;
+ }
- if (PointerOk) {
- const Type *Ty = cast<PointerType>(V->getType())->getElementType();
- AllocaInst *AI = new AllocaInst(Ty, 0, V->getName()+".tmp", FnStart);
- PromotedValues.push_back(std::make_pair(AI, V));
-
- // Update the AST and alias analysis.
- CurAST->copyValue(V, AI);
-
- for (AliasSet::iterator I = AS.begin(), E = AS.end(); I != E; ++I)
- ValueToAllocaMap.insert(std::make_pair(I->first, AI));
-
- DEBUG(std::cerr << "LICM: Promoting value: " << *V << "\n");
+ // It isn't safe to promote a load/store from the loop if the load/store is
+ // conditional. For example, turning:
+ //
+ // for () { if (c) *P += 1; }
+ //
+ // into:
+ //
+ // tmp = *P; for () { if (c) tmp +=1; } *P = tmp;
+ //
+ // is not safe, because *P may only be valid to access if 'c' is true.
+ //
+ // It is safe to promote P if all uses are direct load/stores and if at
+ // least one is guaranteed to be executed.
+ bool GuaranteedToExecute = false;
+ bool InvalidInst = false;
+ for (Value::use_iterator UI = V->use_begin(), UE = V->use_end();
+ UI != UE; ++UI) {
+ // Ignore instructions not in this loop.
+ Instruction *Use = dyn_cast<Instruction>(*UI);
+ if (!Use || !CurLoop->contains(Use->getParent()))
+ continue;
+
+ if (!isa<LoadInst>(Use) && !isa<StoreInst>(Use)) {
+ InvalidInst = true;
+ break;
}
+
+ if (!GuaranteedToExecute)
+ GuaranteedToExecute = isSafeToExecuteUnconditionally(*Use);
}
+
+ // If there is an non-load/store instruction in the loop, we can't promote
+ // it. If there isn't a guaranteed-to-execute instruction, we can't
+ // promote.
+ if (InvalidInst || !GuaranteedToExecute)
+ continue;
+
+ const Type *Ty = cast<PointerType>(V->getType())->getElementType();
+ AllocaInst *AI = new AllocaInst(Ty, 0, V->getName()+".tmp", FnStart);
+ PromotedValues.push_back(std::make_pair(AI, V));
+
+ // Update the AST and alias analysis.
+ CurAST->copyValue(V, AI);
+
+ for (AliasSet::iterator I = AS.begin(), E = AS.end(); I != E; ++I)
+ ValueToAllocaMap.insert(std::make_pair(I->getValue(), AI));
+
+ DOUT << "LICM: Promoting value: " << *V << "\n";
}
}
+
+/// cloneBasicBlockAnalysis - Simple Analysis hook. Clone alias set info.
+void LICM::cloneBasicBlockAnalysis(BasicBlock *From, BasicBlock *To, Loop *L) {
+ AliasSetTracker *AST = LoopToAliasMap[L];
+ if (!AST)
+ return;
+
+ AST->copyValue(From, To);
+}
+
+/// deleteAnalysisValue - Simple Analysis hook. Delete value V from alias
+/// set.
+void LICM::deleteAnalysisValue(Value *V, Loop *L) {
+ AliasSetTracker *AST = LoopToAliasMap[L];
+ if (!AST)
+ return;
+
+ AST->deleteValue(V);
+}