#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/LoopPass.h"
+#include "llvm/LLVMContext.h"
#include "llvm/Analysis/AliasAnalysis.h"
#include "llvm/Analysis/AliasSetTracker.h"
+#include "llvm/Analysis/ConstantFolding.h"
+#include "llvm/Analysis/LoopInfo.h"
+#include "llvm/Analysis/LoopPass.h"
#include "llvm/Analysis/Dominators.h"
-#include "llvm/Analysis/ScalarEvolution.h"
+#include "llvm/Transforms/Utils/Local.h"
#include "llvm/Transforms/Utils/SSAUpdater.h"
#include "llvm/Support/CFG.h"
#include "llvm/Support/CommandLine.h"
namespace {
struct LICM : public LoopPass {
static char ID; // Pass identification, replacement for typeid
- LICM() : LoopPass(ID) {}
+ LICM() : LoopPass(ID) {
+ initializeLICMPass(*PassRegistry::getPassRegistry());
+ }
virtual bool runOnLoop(Loop *L, LPPassManager &LPM);
AU.addRequiredID(LoopSimplifyID);
AU.addRequired<AliasAnalysis>();
AU.addPreserved<AliasAnalysis>();
- AU.addPreserved<ScalarEvolution>();
+ AU.addPreserved("scalar-evolution");
AU.addPreservedID(LoopSimplifyID);
}
///
bool inSubLoop(BasicBlock *BB) {
assert(CurLoop->contains(BB) && "Only valid if BB is IN the loop");
- for (Loop::iterator I = CurLoop->begin(), E = CurLoop->end(); I != E; ++I)
- if ((*I)->contains(BB))
- return true; // A subloop actually contains this block!
- return false;
- }
-
- /// isExitBlockDominatedByBlockInLoop - This method checks to see if the
- /// specified exit block of the loop is dominated by the specified block
- /// that is in the body of the loop. We use these constraints to
- /// dramatically limit the amount of the dominator tree that needs to be
- /// searched.
- bool isExitBlockDominatedByBlockInLoop(BasicBlock *ExitBlock,
- BasicBlock *BlockInLoop) const {
- // If the block in the loop is the loop header, it must be dominated!
- BasicBlock *LoopHeader = CurLoop->getHeader();
- if (BlockInLoop == LoopHeader)
- return true;
-
- 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.
- 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;
-
- // Get next Immediate Dominator.
- IDom = IDom->getIDom();
- };
-
- return true;
+ return LI->getLoopFor(BB) != CurLoop;
}
/// sink - When an instruction is found to only be used outside of the loop,
/// pointerInvalidatedByLoop - Return true if the body of this loop may
/// store into the memory location pointed to by V.
///
- bool pointerInvalidatedByLoop(Value *V, unsigned Size) {
+ bool pointerInvalidatedByLoop(Value *V, uint64_t Size,
+ const MDNode *TBAAInfo) {
// Check to see if any of the basic blocks in CurLoop invalidate *V.
- return CurAST->getAliasSetForPointer(V, Size).isMod();
+ return CurAST->getAliasSetForPointer(V, Size, TBAAInfo).isMod();
}
bool canSinkOrHoistInst(Instruction &I);
- bool isLoopInvariantInst(Instruction &I);
bool isNotUsedInLoop(Instruction &I);
void PromoteAliasSet(AliasSet &AS);
}
char LICM::ID = 0;
-INITIALIZE_PASS(LICM, "licm", "Loop Invariant Code Motion", false, false);
+INITIALIZE_PASS_BEGIN(LICM, "licm", "Loop Invariant Code Motion", false, false)
+INITIALIZE_PASS_DEPENDENCY(DominatorTree)
+INITIALIZE_PASS_DEPENDENCY(LoopInfo)
+INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
+INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
+INITIALIZE_PASS_END(LICM, "licm", "Loop Invariant Code Motion", false, false)
Pass *llvm::createLICMPass() { return new LICM(); }
// If this subregion is not in the top level loop at all, exit.
if (!CurLoop->contains(BB)) return;
- // We are processing blocks in reverse dfo, so process children first...
+ // We are processing blocks in reverse dfo, so process children first.
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;
+
+ // If the instruction is dead, we would try to sink it because it isn't used
+ // in the loop, instead, just delete it.
+ if (isInstructionTriviallyDead(&I)) {
+ DEBUG(dbgs() << "LICM deleting dead inst: " << I << '\n');
+ ++II;
+ CurAST->deleteValue(&I);
+ I.eraseFromParent();
+ Changed = true;
+ continue;
+ }
// 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
for (BasicBlock::iterator II = BB->begin(), E = BB->end(); II != E; ) {
Instruction &I = *II++;
+ // Try constant folding this instruction. If all the operands are
+ // constants, it is technically hoistable, but it would be better to just
+ // fold it.
+ if (Constant *C = ConstantFoldInstruction(&I)) {
+ DEBUG(dbgs() << "LICM folding inst: " << I << " --> " << *C << '\n');
+ CurAST->copyValue(&I, C);
+ CurAST->deleteValue(&I);
+ I.replaceAllUsesWith(C);
+ I.eraseFromParent();
+ continue;
+ }
+
// 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 (CurLoop->hasLoopInvariantOperands(&I) && canSinkOrHoistInst(I) &&
isSafeToExecuteUnconditionally(I))
hoist(I);
- }
+ }
const std::vector<DomTreeNode*> &Children = N->getChildren();
for (unsigned i = 0, e = Children.size(); i != e; ++i)
return true;
// Don't hoist loads which have may-aliased stores in loop.
- unsigned Size = 0;
+ uint64_t Size = 0;
if (LI->getType()->isSized())
Size = AA->getTypeStoreSize(LI->getType());
- return !pointerInvalidatedByLoop(LI->getOperand(0), Size);
+ return !pointerInvalidatedByLoop(LI->getOperand(0), Size,
+ LI->getMetadata(LLVMContext::MD_tbaa));
} else if (CallInst *CI = dyn_cast<CallInst>(&I)) {
- // Handle obvious cases efficiently.
+ // Don't sink or hoist dbg info; it's legal, but not useful.
+ if (isa<DbgInfoIntrinsic>(I))
+ return false;
+
+ // Handle simple cases by querying alias analysis.
AliasAnalysis::ModRefBehavior Behavior = AA->getModRefBehavior(CI);
if (Behavior == AliasAnalysis::DoesNotAccessMemory)
return true;
- else if (Behavior == AliasAnalysis::OnlyReadsMemory) {
+ if (AliasAnalysis::onlyReadsMemory(Behavior)) {
// 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;
}
-/// isLoopInvariantInst - Return true if all operands of this instruction are
-/// loop invariant. We also filter out non-hoistable instructions here just for
-/// efficiency.
-///
-bool LICM::isLoopInvariantInst(Instruction &I) {
- // The instruction is loop invariant if all of its operands are loop-invariant
- for (unsigned i = 0, e = I.getNumOperands(); i != e; ++i)
- if (!CurLoop->isLoopInvariant(I.getOperand(i)))
- return false;
-
- // If we got this far, the instruction is loop invariant!
- return true;
-}
-
/// sink - When an instruction is found to only be used outside of the loop,
/// this function moves it to the exit blocks and patches up SSA form as needed.
/// This method is guaranteed to remove the original instruction from its
// 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 (ExitBlocks.size() == 1) {
- if (!isExitBlockDominatedByBlockInLoop(ExitBlocks[0], I.getParent())) {
+ if (!DT->dominates(I.getParent(), ExitBlocks[0])) {
// Instruction is not used, just delete it.
CurAST->deleteValue(&I);
// If I has users in unreachable blocks, eliminate.
SSAUpdater SSA(&NewPHIs);
if (!I.use_empty())
- SSA.Initialize(&I);
+ SSA.Initialize(I.getType(), I.getName());
// Insert a copy of the instruction in each exit block of the loop that is
// dominated by the instruction. Each exit block is known to only be in the
for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) {
BasicBlock *ExitBlock = ExitBlocks[i];
- if (!isExitBlockDominatedByBlockInLoop(ExitBlock, InstOrigBB))
+ if (!DT->dominates(InstOrigBB, ExitBlock))
continue;
// Insert the code after the last PHI node.
// Update CurAST for NewPHIs if I had pointer type.
if (I.getType()->isPointerTy())
for (unsigned i = 0, e = NewPHIs.size(); i != e; ++i)
- CurAST->copyValue(NewPHIs[i], &I);
+ CurAST->copyValue(&I, NewPHIs[i]);
// Finally, remove the instruction from CurAST. It is no longer in the loop.
CurAST->deleteValue(&I);
SmallVector<BasicBlock*, 8> ExitBlocks;
CurLoop->getExitBlocks(ExitBlocks);
- // For each exit block, get the DT node and walk up the DT until the
- // instruction's basic block is found or we exit the loop.
+ // Verify that the block dominates each of the exit blocks of the loop.
for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i)
- if (!isExitBlockDominatedByBlockInLoop(ExitBlocks[i], Inst.getParent()))
+ if (!DT->dominates(Inst.getParent(), ExitBlocks[i]))
return false;
return true;
}
+namespace {
+ class LoopPromoter : public LoadAndStorePromoter {
+ Value *SomePtr; // Designated pointer to store to.
+ SmallPtrSet<Value*, 4> &PointerMustAliases;
+ SmallVectorImpl<BasicBlock*> &LoopExitBlocks;
+ AliasSetTracker &AST;
+ DebugLoc DL;
+ int Alignment;
+ public:
+ LoopPromoter(Value *SP,
+ const SmallVectorImpl<Instruction*> &Insts, SSAUpdater &S,
+ SmallPtrSet<Value*, 4> &PMA,
+ SmallVectorImpl<BasicBlock*> &LEB, AliasSetTracker &ast,
+ DebugLoc dl, int alignment)
+ : LoadAndStorePromoter(Insts, S, 0, 0), SomePtr(SP),
+ PointerMustAliases(PMA), LoopExitBlocks(LEB), AST(ast), DL(dl),
+ Alignment(alignment) {}
+
+ virtual bool isInstInList(Instruction *I,
+ const SmallVectorImpl<Instruction*> &) const {
+ Value *Ptr;
+ if (LoadInst *LI = dyn_cast<LoadInst>(I))
+ Ptr = LI->getOperand(0);
+ else
+ Ptr = cast<StoreInst>(I)->getPointerOperand();
+ return PointerMustAliases.count(Ptr);
+ }
+
+ virtual void doExtraRewritesBeforeFinalDeletion() const {
+ // Insert stores after in the loop exit blocks. Each exit block gets a
+ // store of the live-out values that feed them. Since we've already told
+ // the SSA updater about the defs in the loop and the preheader
+ // definition, it is all set and we can start using it.
+ for (unsigned i = 0, e = LoopExitBlocks.size(); i != e; ++i) {
+ BasicBlock *ExitBlock = LoopExitBlocks[i];
+ Value *LiveInValue = SSA.GetValueInMiddleOfBlock(ExitBlock);
+ Instruction *InsertPos = ExitBlock->getFirstNonPHI();
+ StoreInst *NewSI = new StoreInst(LiveInValue, SomePtr, InsertPos);
+ NewSI->setAlignment(Alignment);
+ NewSI->setDebugLoc(DL);
+ }
+ }
+
+ virtual void replaceLoadWithValue(LoadInst *LI, Value *V) const {
+ // Update alias analysis.
+ AST.copyValue(LI, V);
+ }
+ virtual void instructionDeleted(Instruction *I) const {
+ AST.deleteValue(I);
+ }
+ };
+} // end anon namespace
+
/// PromoteAliasSet - Try to promote memory values to scalars by sinking
/// stores out of the loop and moving loads to before the loop. We do this by
/// looping over the stores in the loop, looking for stores to Must pointers
// 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;
-
+
SmallVector<Instruction*, 64> LoopUses;
SmallPtrSet<Value*, 4> PointerMustAliases;
+ // We start with an alignment of one and try to find instructions that allow
+ // us to prove better alignment.
+ unsigned Alignment = 1;
+
// 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 there is an non-load/store instruction in the loop, we can't promote
// it.
- if (isa<LoadInst>(Use))
+ unsigned InstAlignment;
+ if (LoadInst *load = dyn_cast<LoadInst>(Use)) {
assert(!cast<LoadInst>(Use)->isVolatile() && "AST broken");
- else if (isa<StoreInst>(Use))
- assert(!cast<StoreInst>(Use)->isVolatile() &&
- Use->getOperand(0) != ASIV && "AST broken");
- else
+ InstAlignment = load->getAlignment();
+ } else if (StoreInst *store = dyn_cast<StoreInst>(Use)) {
+ // Stores *of* the pointer are not interesting, only stores *to* the
+ // pointer.
+ if (Use->getOperand(1) != ASIV)
+ continue;
+ InstAlignment = store->getAlignment();
+ assert(!cast<StoreInst>(Use)->isVolatile() && "AST broken");
+ } else
return; // Not a load or store.
-
+
+ // If the alignment of this instruction allows us to specify a more
+ // restrictive (and performant) alignment and if we are sure this
+ // instruction will be executed, update the alignment.
+ // Larger is better, with the exception of 0 being the best alignment.
+ if ((InstAlignment > Alignment || InstAlignment == 0)
+ && (Alignment != 0))
+ if (isSafeToExecuteUnconditionally(*Use)) {
+ GuaranteedToExecute = true;
+ Alignment = InstAlignment;
+ }
+
if (!GuaranteedToExecute)
GuaranteedToExecute = isSafeToExecuteUnconditionally(*Use);
LoopUses.push_back(Use);
}
}
-
+
// If there isn't a guaranteed-to-execute instruction, we can't promote.
if (!GuaranteedToExecute)
return;
Changed = true;
++NumPromoted;
+ // Grab a debug location for the inserted loads/stores; given that the
+ // inserted loads/stores have little relation to the original loads/stores,
+ // this code just arbitrarily picks a location from one, since any debug
+ // location is better than none.
+ DebugLoc DL = LoopUses[0]->getDebugLoc();
+
+ SmallVector<BasicBlock*, 8> ExitBlocks;
+ CurLoop->getUniqueExitBlocks(ExitBlocks);
+
// We use the SSAUpdater interface to insert phi nodes as required.
SmallVector<PHINode*, 16> NewPHIs;
SSAUpdater SSA(&NewPHIs);
+ LoopPromoter Promoter(SomePtr, LoopUses, SSA, PointerMustAliases, ExitBlocks,
+ *CurAST, DL, Alignment);
- // It wants to know some value of the same type as what we'll be inserting.
- Value *SomeValue;
- if (isa<LoadInst>(LoopUses[0]))
- SomeValue = LoopUses[0];
- else
- SomeValue = cast<StoreInst>(LoopUses[0])->getOperand(0);
- SSA.Initialize(SomeValue);
-
- // First step: bucket up uses of the pointers by the block they occur in.
- // This is important because we have to handle multiple defs/uses in a block
- // ourselves: SSAUpdater is purely for cross-block references.
- // FIXME: Want a TinyVector<Instruction*> since there is usually 0/1 element.
- DenseMap<BasicBlock*, std::vector<Instruction*> > UsesByBlock;
- for (unsigned i = 0, e = LoopUses.size(); i != e; ++i) {
- Instruction *User = LoopUses[i];
- UsesByBlock[User->getParent()].push_back(User);
- }
-
- // Okay, now we can iterate over all the blocks in the loop with uses,
- // processing them. Keep track of which loads are loading a live-in value.
- SmallVector<LoadInst*, 32> LiveInLoads;
-
- for (unsigned LoopUse = 0, e = LoopUses.size(); LoopUse != e; ++LoopUse) {
- Instruction *User = LoopUses[LoopUse];
- std::vector<Instruction*> &BlockUses = UsesByBlock[User->getParent()];
-
- // If this block has already been processed, ignore this repeat use.
- if (BlockUses.empty()) continue;
-
- // Okay, this is the first use in the block. If this block just has a
- // single user in it, we can rewrite it trivially.
- if (BlockUses.size() == 1) {
- // If it is a store, it is a trivial def of the value in the block.
- if (isa<StoreInst>(User)) {
- SSA.AddAvailableValue(User->getParent(),
- cast<StoreInst>(User)->getOperand(0));
- } else {
- // Otherwise it is a load, queue it to rewrite as a live-in load.
- LiveInLoads.push_back(cast<LoadInst>(User));
- }
- BlockUses.clear();
- continue;
- }
-
- // Otherwise, check to see if this block is all loads. If so, we can queue
- // them all as live in loads.
- bool HasStore = false;
- for (unsigned i = 0, e = BlockUses.size(); i != e; ++i) {
- if (isa<StoreInst>(BlockUses[i])) {
- HasStore = true;
- break;
- }
- }
-
- if (!HasStore) {
- for (unsigned i = 0, e = BlockUses.size(); i != e; ++i)
- LiveInLoads.push_back(cast<LoadInst>(BlockUses[i]));
- BlockUses.clear();
- continue;
- }
-
- // Otherwise, we have mixed loads and stores (or just a bunch of stores).
- // Since SSAUpdater is purely for cross-block values, we need to determine
- // the order of these instructions in the block. If the first use in the
- // block is a load, then it uses the live in value. The last store defines
- // the live out value. We handle this by doing a linear scan of the block.
- BasicBlock *BB = User->getParent();
- Value *StoredValue = 0;
- for (BasicBlock::iterator II = BB->begin(), E = BB->end(); II != E; ++II) {
- if (LoadInst *L = dyn_cast<LoadInst>(II)) {
- // If this is a load to an unrelated pointer, ignore it.
- if (!PointerMustAliases.count(L->getOperand(0))) continue;
-
- // If we haven't seen a store yet, this is a live in use, otherwise
- // use the stored value.
- if (StoredValue)
- L->replaceAllUsesWith(StoredValue);
- else
- LiveInLoads.push_back(L);
- continue;
- }
-
- if (StoreInst *S = dyn_cast<StoreInst>(II)) {
- // If this is a load to an unrelated pointer, ignore it.
- if (!PointerMustAliases.count(S->getOperand(1))) continue;
-
- // Remember that this is the active value in the block.
- StoredValue = S->getOperand(0);
- }
- }
-
- // The last stored value that happened is the live-out for the block.
- assert(StoredValue && "Already checked that there is a store in block");
- SSA.AddAvailableValue(BB, StoredValue);
- BlockUses.clear();
- }
-
- // Now that all the intra-loop values are classified, set up the preheader.
- // It gets a load of the pointer we're promoting, and it is the live-out value
- // from the preheader.
- LoadInst *PreheaderLoad = new LoadInst(SomePtr,SomePtr->getName()+".promoted",
- Preheader->getTerminator());
+ // Set up the preheader to have a definition of the value. It is the live-out
+ // value from the preheader that uses in the loop will use.
+ LoadInst *PreheaderLoad =
+ new LoadInst(SomePtr, SomePtr->getName()+".promoted",
+ Preheader->getTerminator());
+ PreheaderLoad->setAlignment(Alignment);
+ PreheaderLoad->setDebugLoc(DL);
SSA.AddAvailableValue(Preheader, PreheaderLoad);
- // Now that the preheader is good to go, set up the exit blocks. Each exit
- // block gets a store of the live-out values that feed them. Since we've
- // already told the SSA updater about the defs in the loop and the preheader
- // definition, it is all set and we can start using it.
- SmallVector<BasicBlock*, 8> ExitBlocks;
- CurLoop->getUniqueExitBlocks(ExitBlocks);
- for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) {
- BasicBlock *ExitBlock = ExitBlocks[i];
- Value *LiveInValue = SSA.GetValueInMiddleOfBlock(ExitBlock);
- Instruction *InsertPos = ExitBlock->getFirstNonPHI();
- new StoreInst(LiveInValue, SomePtr, InsertPos);
- }
+ // Rewrite all the loads in the loop and remember all the definitions from
+ // stores in the loop.
+ Promoter.run(LoopUses);
- // Okay, now we rewrite all loads that use live-in values in the loop,
- // inserting PHI nodes as necessary.
- for (unsigned i = 0, e = LiveInLoads.size(); i != e; ++i) {
- LoadInst *ALoad = LiveInLoads[i];
- ALoad->replaceAllUsesWith(SSA.GetValueInMiddleOfBlock(ALoad->getParent()));
- }
-
- // Now that everything is rewritten, delete the old instructions from the body
- // of the loop. They should all be dead now.
- for (unsigned i = 0, e = LoopUses.size(); i != e; ++i) {
- Instruction *User = LoopUses[i];
- CurAST->deleteValue(User);
- User->eraseFromParent();
- }
-
- // If the preheader load is itself a pointer, we need to tell alias analysis
- // about the new pointer we created in the preheader block and about any PHI
- // nodes that just got inserted.
- if (PreheaderLoad->getType()->isPointerTy()) {
- // Copy any value stored to or loaded from a must-alias of the pointer.
- CurAST->copyValue(SomeValue, PreheaderLoad);
-
- for (unsigned i = 0, e = NewPHIs.size(); i != e; ++i)
- CurAST->copyValue(SomeValue, NewPHIs[i]);
- }
-
- // fwew, we're done!
+ // If the SSAUpdater didn't use the load in the preheader, just zap it now.
+ if (PreheaderLoad->use_empty())
+ PreheaderLoad->eraseFromParent();
}