//
// 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/Transforms/Scalar.h"
#include "llvm/Constants.h"
#include "llvm/DerivedTypes.h"
+#include "llvm/IntrinsicInst.h"
#include "llvm/Instructions.h"
#include "llvm/Target/TargetData.h"
#include "llvm/Analysis/LoopInfo.h"
#include "llvm/Analysis/ScalarEvolution.h"
#include "llvm/Transforms/Utils/PromoteMemToReg.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>
STATISTIC(NumMovedCalls, "Number of call insts hoisted or sunk");
STATISTIC(NumPromoted , "Number of memory locations promoted to registers");
-namespace {
- cl::opt<bool>
- DisablePromotion("disable-licm-promotion", cl::Hidden,
- cl::desc("Disable memory promotion in LICM pass"));
+static cl::opt<bool>
+DisablePromotion("disable-licm-promotion", cl::Hidden,
+ cl::desc("Disable memory promotion in LICM pass"));
- struct VISIBILITY_HIDDEN LICM : public LoopPass {
+namespace {
+ struct LICM : public LoopPass {
static char ID; // Pass identification, replacement for typeid
- LICM() : LoopPass((intptr_t)&ID) {}
+ LICM() : LoopPass(&ID) {}
virtual bool runOnLoop(Loop *L, LPPassManager &LPM);
AU.addRequired<AliasAnalysis>();
AU.addPreserved<ScalarEvolution>();
AU.addPreserved<DominanceFrontier>();
+ AU.addPreservedID(LoopSimplifyID);
}
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;
}
// 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();
-
+ IDom = IDom->getIDom();
+
+ while (IDom && IDom != BlockInLoopNode) {
// 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);
+ // Get next Immediate Dominator.
+ IDom = IDom->getIDom();
+ };
return true;
}
std::vector<std::pair<AllocaInst*, Value*> > &PromotedValues,
std::map<Value*, AllocaInst*> &Val2AlMap);
};
-
- char LICM::ID = 0;
- RegisterPass<LICM> X("licm", "Loop Invariant Code Motion");
}
-LoopPass *llvm::createLICMPass() { return new LICM(); }
+char LICM::ID = 0;
+static RegisterPass<LICM> X("licm", "Loop Invariant Code Motion");
+
+Pass *llvm::createLICMPass() { return new LICM(); }
-/// Hoist expressions out of the specified loop...
+/// 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::runOnLoop(Loop *L, LPPassManager &LPM) {
Changed = false;
// Get the preheader block to move instructions into...
Preheader = L->getLoopPreheader();
- assert(Preheader&&"Preheader insertion pass guarantees we have a preheader!");
// Loop over the body of this loop, looking for calls, invokes, and stores.
// 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...
- CurAST->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()));
- HoistRegion(DT->getNode(L->getHeader()));
+ if (L->hasDedicatedExits())
+ SinkRegion(DT->getNode(L->getHeader()));
+ if (Preheader)
+ HoistRegion(DT->getNode(L->getHeader()));
// Now that all loop invariants have been removed from the loop, promote any
// memory references to scalars that we can...
- if (!DisablePromotion)
+ if (!DisablePromotion && Preheader && L->hasDedicatedExits())
PromoteValuesInLoop();
// Clear out loops state information for the next iteration
}
}
-
/// HoistRegion - Walk the specified region of the CFG (defined by all blocks
/// dominated by the specified block, and that are in the current loop) in depth
/// first order w.r.t the DominatorTree. This allows us to visit definitions
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 (AA->pointsToConstantMemory(LI->getOperand(0)))
+ return true;
+
// Don't hoist loads which have may-aliased stores in loop.
unsigned Size = 0;
if (LI->getType()->isSized())
- Size = AA->getTargetData().getTypeSize(LI->getType());
+ 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()) {
- AliasAnalysis::ModRefBehavior Behavior =AA->getModRefBehavior(Callee, 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;
- }
+ 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
if (PN->getIncomingValue(i) == &I)
if (CurLoop->contains(PN->getIncomingBlock(i)))
return false;
- } else if (CurLoop->contains(User->getParent())) {
+ } else if (CurLoop->contains(User)) {
return false;
}
}
/// position, and may either delete it or move it to outside of the loop.
///
void LICM::sink(Instruction &I) {
- DOUT << "LICM sinking instruction: " << I;
+ DEBUG(dbgs() << "LICM sinking instruction: " << I);
- std::vector<BasicBlock*> ExitBlocks;
+ SmallVector<BasicBlock*, 8> ExitBlocks;
CurLoop->getExitBlocks(ExitBlocks);
if (isa<LoadInst>(I)) ++NumMovedLoads;
if (!isExitBlockDominatedByBlockInLoop(ExitBlocks[0], I.getParent())) {
// Instruction is not used, just delete it.
CurAST->deleteValue(&I);
- if (!I.use_empty()) // If I has users in unreachable blocks, eliminate.
+ // If I has users in unreachable blocks, eliminate.
+ // 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()));
I.eraseFromParent();
} else {
// Move the instruction to the start of the exit block, after any PHI
// nodes in it.
I.removeFromParent();
-
- BasicBlock::iterator InsertPt = ExitBlocks[0]->begin();
- while (isa<PHINode>(InsertPt)) ++InsertPt;
+ 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);
- if (!I.use_empty()) // If I has users in unreachable blocks, eliminate.
+ // If I has users in unreachable blocks, eliminate.
+ // 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()));
I.eraseFromParent();
} else {
// Firstly, we create a stack object to hold the value...
AllocaInst *AI = 0;
- if (I.getType() != Type::VoidTy) {
+ if (!I.getType()->isVoidTy()) {
AI = new AllocaInst(I.getType(), 0, I.getName(),
I.getParent()->getParent()->getEntryBlock().begin());
CurAST->add(AI);
// 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
/// that is safe to hoist, this instruction is called to do the dirty work.
///
void LICM::hoist(Instruction &I) {
- DOUT << "LICM hoisting to " << Preheader->getName() << ": " << I;
+ DEBUG(dbgs() << "LICM hoisting to " << Preheader->getName() << ": "
+ << I << "\n");
// Remove the instruction from its current basic block... but don't delete the
// instruction.
///
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
if (Inst.getParent() == CurLoop->getHeader())
return true;
- // It's always safe to load from a global or alloca.
- if (isa<LoadInst>(Inst))
- if (isa<AllocationInst>(Inst.getOperand(0)) ||
- isa<GlobalVariable>(Inst.getOperand(0)))
- 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
// If we are promoting a pointer value, update alias information for the
// inserted load.
Value *LoadValue = 0;
- if (isa<PointerType>(cast<PointerType>(Ptr->getType())->getElementType())) {
+ if (cast<PointerType>(Ptr->getType())->getElementType()->isPointerTy()) {
// Locate a load or store through the pointer, and assign the same value
// to LI as we are loading or storing. Since we know that the value is
// stored in this loop, this will always succeed.
// 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 (LI->getType()->isPointerTy())
+ 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.
}
/// 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));
+ // 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))
+ continue;
+
+ if (!isa<LoadInst>(Use) && !isa<StoreInst>(Use)) {
+ InvalidInst = true;
+ break;
+ }
+
+ if (!GuaranteedToExecute)
+ GuaranteedToExecute = isSafeToExecuteUnconditionally(*Use);
+ }
- // Update the AST and alias analysis.
- CurAST->copyValue(V, AI);
+ // 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));
- for (AliasSet::iterator I = AS.begin(), E = AS.end(); I != E; ++I)
- ValueToAllocaMap.insert(std::make_pair(I->first, AI));
+ // Update the AST and alias analysis.
+ CurAST->copyValue(V, AI);
- DOUT << "LICM: Promoting value: " << *V << "\n";
- }
- }
+ for (AliasSet::iterator I = AS.begin(), E = AS.end(); I != E; ++I)
+ ValueToAllocaMap.insert(std::make_pair(I->getValue(), AI));
+
+ DEBUG(dbgs() << "LICM: Promoting value: " << *V << "\n");
}
}
projects / oota-llvm.git / blobdiff