#define DEBUG_TYPE "licm"
#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/LoopPass.h"
+#include "llvm/ADT/Statistic.h"
#include "llvm/Analysis/AliasAnalysis.h"
#include "llvm/Analysis/AliasSetTracker.h"
+#include "llvm/Analysis/ConstantFolding.h"
#include "llvm/Analysis/Dominators.h"
-#include "llvm/Analysis/ScalarEvolution.h"
-#include "llvm/Transforms/Utils/SSAUpdater.h"
+#include "llvm/Analysis/LoopInfo.h"
+#include "llvm/Analysis/LoopPass.h"
+#include "llvm/Analysis/ValueTracking.h"
+#include "llvm/IR/Constants.h"
+#include "llvm/IR/DataLayout.h"
+#include "llvm/IR/DerivedTypes.h"
+#include "llvm/IR/Instructions.h"
+#include "llvm/IR/IntrinsicInst.h"
+#include "llvm/IR/LLVMContext.h"
#include "llvm/Support/CFG.h"
#include "llvm/Support/CommandLine.h"
-#include "llvm/Support/raw_ostream.h"
#include "llvm/Support/Debug.h"
-#include "llvm/ADT/Statistic.h"
+#include "llvm/Support/raw_ostream.h"
+#include "llvm/Target/TargetLibraryInfo.h"
+#include "llvm/Transforms/Utils/Local.h"
+#include "llvm/Transforms/Utils/SSAUpdater.h"
#include <algorithm>
using namespace llvm;
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);
+ AU.addRequired<TargetLibraryInfo>();
}
+ using llvm::Pass::doFinalization;
+
bool doFinalization() {
assert(LoopToAliasSetMap.empty() && "Didn't free loop alias sets");
return false;
LoopInfo *LI; // Current LoopInfo
DominatorTree *DT; // Dominator Tree for the current Loop.
+ DataLayout *TD; // DataLayout for constant folding.
+ TargetLibraryInfo *TLI; // TargetLibraryInfo for constant folding.
+
// State that is updated as we process loops.
bool Changed; // Set to true when we change anything.
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...
+ bool MayThrow; // The current loop contains an instruction which
+ // may throw, thus preventing code motion of
+ // instructions with side effects.
DenseMap<Loop*, AliasSetTracker*> LoopToAliasSetMap;
/// cloneBasicBlockAnalysis - Simple Analysis hook. Clone alias set info.
///
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,
///
bool isSafeToExecuteUnconditionally(Instruction &I);
+ /// isGuaranteedToExecute - Check that the instruction is guaranteed to
+ /// execute.
+ ///
+ bool isGuaranteedToExecute(Instruction &I);
+
/// 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);
+ void PromoteAliasSet(AliasSet &AS,
+ SmallVectorImpl<BasicBlock*> &ExitBlocks,
+ SmallVectorImpl<Instruction*> &InsertPts);
};
}
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_PASS_DEPENDENCY(TargetLibraryInfo)
+INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
+INITIALIZE_PASS_END(LICM, "licm", "Loop Invariant Code Motion", false, false)
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
+/// 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) {
AA = &getAnalysis<AliasAnalysis>();
DT = &getAnalysis<DominatorTree>();
+ TD = getAnalysisIfAvailable<DataLayout>();
+ TLI = &getAnalysis<TargetLibraryInfo>();
+
CurAST = new AliasSetTracker(*AA);
// Collect Alias info from subloops.
for (Loop::iterator LoopItr = L->begin(), LoopItrE = L->end();
// What if InnerLoop was modified by other passes ?
CurAST->add(*InnerAST);
-
+
// Once we've incorporated the inner loop's AST into ours, we don't need the
// subloop's anymore.
delete InnerAST;
LoopToAliasSetMap.erase(InnerL);
}
-
+
CurLoop = L;
// Get the preheader block to move instructions into...
CurAST->add(*BB); // Incorporate the specified basic block
}
+ MayThrow = false;
+ // TODO: We've already searched for instructions which may throw in subloops.
+ // We may want to reuse this information.
+ for (Loop::block_iterator BB = L->block_begin(), BBE = L->block_end();
+ (BB != BBE) && !MayThrow ; ++BB)
+ for (BasicBlock::iterator I = (*BB)->begin(), E = (*BB)->end();
+ (I != E) && !MayThrow; ++I)
+ MayThrow |= I->mayThrow();
+
// 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
// their loop, into this loop, so there is no need to process the BODIES of
// Now that all loop invariants have been removed from the loop, promote any
// memory references to scalars that we can.
if (!DisablePromotion && Preheader && L->hasDedicatedExits()) {
+ SmallVector<BasicBlock *, 8> ExitBlocks;
+ SmallVector<Instruction *, 8> InsertPts;
+
// Loop over all of the alias sets in the tracker object.
for (AliasSetTracker::iterator I = CurAST->begin(), E = CurAST->end();
I != E; ++I)
- PromoteAliasSet(*I);
+ PromoteAliasSet(*I, ExitBlocks, InsertPts);
}
-
+
// Clear out loops state information for the next iteration
CurLoop = 0;
Preheader = 0;
// 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, TLI)) {
+ 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
// outside of the loop. In this case, it doesn't even matter if the
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, TD, TLI)) {
+ 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)
bool LICM::canSinkOrHoistInst(Instruction &I) {
// Loads have extra constraints we have to verify before we can hoist them.
if (LoadInst *LI = dyn_cast<LoadInst>(&I)) {
- if (LI->isVolatile())
- return false; // Don't hoist volatile loads!
+ if (!LI->isUnordered())
+ return false; // Don't hoist volatile/atomic 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;
-
+ if (LI->getMetadata("invariant.load"))
+ 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;
if (!FoundMod) return true;
}
- // FIXME: This should use mod/ref information to see if we can hoist or sink
- // the call.
+ // FIXME: This should use mod/ref information to see if we can hoist or
+ // sink the call.
return false;
}
- // 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);
+ // Only these instructions are hoistable/sinkable.
+ bool HoistableKind = (isa<BinaryOperator>(I) || isa<CastInst>(I) ||
+ isa<SelectInst>(I) || isa<GetElementPtrInst>(I) ||
+ isa<CmpInst>(I) || isa<InsertElementInst>(I) ||
+ isa<ExtractElementInst>(I) ||
+ isa<ShuffleVectorInst>(I));
+ if (!HoistableKind)
+ return false;
+
+ return isSafeToExecuteUnconditionally(I);
}
/// isNotUsedInLoop - Return true if the only users of this instruction are
}
-/// 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.
} else {
// Move the instruction to the start of the exit block, after any PHI
// nodes in it.
- I.removeFromParent();
- BasicBlock::iterator InsertPt = ExitBlocks[0]->getFirstNonPHI();
- ExitBlocks[0]->getInstList().insert(InsertPt, &I);
+ I.moveBefore(ExitBlocks[0]->getFirstInsertionPt());
// This instruction is no longer in the AST for the current loop, because
// we just sunk it out of the loop. If we just sunk it into an outer
}
return;
}
-
+
if (ExitBlocks.empty()) {
// The instruction is actually dead if there ARE NO exit blocks.
CurAST->deleteValue(&I);
I.eraseFromParent();
return;
}
-
+
// Otherwise, if we have multiple exits, use the SSAUpdater to do all of the
// hard work of inserting PHI nodes as necessary.
SmallVector<PHINode*, 8> NewPHIs;
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
// ExitBlocks list once.
BasicBlock *InstOrigBB = I.getParent();
unsigned NumInserted = 0;
-
+
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.
- BasicBlock::iterator InsertPt = ExitBlock->getFirstNonPHI();
-
+ BasicBlock::iterator InsertPt = ExitBlock->getFirstInsertionPt();
+
// If this is the first exit block processed, just move the original
// instruction, otherwise clone the original instruction and insert
// the copy.
New->setName(I.getName()+".le");
ExitBlock->getInstList().insert(InsertPt, New);
}
-
+
// Now that we have inserted the instruction, inform SSAUpdater.
if (!I.use_empty())
SSA.AddAvailableValue(ExitBlock, New);
}
-
+
// If the instruction doesn't dominate any exit blocks, it must be dead.
if (NumInserted == 0) {
CurAST->deleteValue(&I);
I.eraseFromParent();
return;
}
-
+
// Next, rewrite uses of the instruction, inserting PHI nodes as needed.
for (Value::use_iterator UI = I.use_begin(), UE = I.use_end(); UI != UE; ) {
// Grab the use before incrementing the iterator.
++UI;
SSA.RewriteUseAfterInsertions(U);
}
-
+
// 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);
}
DEBUG(dbgs() << "LICM hoisting to " << Preheader->getName() << ": "
<< I << "\n");
- // Remove the instruction from its current basic block... but don't delete the
- // instruction.
- I.removeFromParent();
-
- // Insert the new node in Preheader, before the terminator.
- Preheader->getInstList().insert(Preheader->getTerminator(), &I);
+ // Move the new node to the Preheader, before its terminator.
+ I.moveBefore(Preheader->getTerminator());
if (isa<LoadInst>(I)) ++NumMovedLoads;
else if (isa<CallInst>(I)) ++NumMovedCalls;
///
bool LICM::isSafeToExecuteUnconditionally(Instruction &Inst) {
// If it is not a trapping instruction, it is always safe to hoist.
- if (Inst.isSafeToSpeculativelyExecute())
+ if (isSafeToSpeculativelyExecute(&Inst))
return true;
+ return isGuaranteedToExecute(Inst);
+}
+
+bool LICM::isGuaranteedToExecute(Instruction &Inst) {
+
+ // Somewhere in this loop there is an instruction which may throw and make us
+ // exit the loop.
+ if (MayThrow)
+ return false;
+
// 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.
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;
+ // As a degenerate case, if the loop is statically infinite then we haven't
+ // proven anything since there are no exit blocks.
+ if (ExitBlocks.empty())
+ return false;
+
return true;
}
+namespace {
+ class LoopPromoter : public LoadAndStorePromoter {
+ Value *SomePtr; // Designated pointer to store to.
+ SmallPtrSet<Value*, 4> &PointerMustAliases;
+ SmallVectorImpl<BasicBlock*> &LoopExitBlocks;
+ SmallVectorImpl<Instruction*> &LoopInsertPts;
+ AliasSetTracker &AST;
+ DebugLoc DL;
+ int Alignment;
+ MDNode *TBAATag;
+ public:
+ LoopPromoter(Value *SP,
+ const SmallVectorImpl<Instruction*> &Insts, SSAUpdater &S,
+ SmallPtrSet<Value*, 4> &PMA,
+ SmallVectorImpl<BasicBlock*> &LEB,
+ SmallVectorImpl<Instruction*> &LIP,
+ AliasSetTracker &ast, DebugLoc dl, int alignment,
+ MDNode *TBAATag)
+ : LoadAndStorePromoter(Insts, S), SomePtr(SP),
+ PointerMustAliases(PMA), LoopExitBlocks(LEB), LoopInsertPts(LIP),
+ AST(ast), DL(dl), Alignment(alignment), TBAATag(TBAATag) {}
+
+ 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 = LoopInsertPts[i];
+ StoreInst *NewSI = new StoreInst(LiveInValue, SomePtr, InsertPos);
+ NewSI->setAlignment(Alignment);
+ NewSI->setDebugLoc(DL);
+ if (TBAATag) NewSI->setMetadata(LLVMContext::MD_tbaa, TBAATag);
+ }
+ }
+
+ 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
/// which are loop invariant.
///
-void LICM::PromoteAliasSet(AliasSet &AS) {
+void LICM::PromoteAliasSet(AliasSet &AS,
+ SmallVectorImpl<BasicBlock*> &ExitBlocks,
+ SmallVectorImpl<Instruction*> &InsertPts) {
// 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()->getValue()))
return;
-
+
assert(!AS.empty() &&
"Must alias set should have at least one pointer element in it!");
Value *SomePtr = AS.begin()->getValue();
// 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;
-
+
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;
+ MDNode *TBAATag = 0;
+
// 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.
+ // different sizes. While we are at it, collect alignment and TBAA info.
for (AliasSet::iterator ASI = AS.begin(), E = AS.end(); ASI != E; ++ASI) {
Value *ASIV = ASI->getValue();
PointerMustAliases.insert(ASIV);
-
+
// 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 (SomePtr->getType() != ASIV->getType())
return;
-
+
for (Value::use_iterator UI = ASIV->use_begin(), UE = ASIV->use_end();
UI != UE; ++UI) {
// Ignore instructions that are outside the loop.
Instruction *Use = dyn_cast<Instruction>(*UI);
if (!Use || !CurLoop->contains(Use))
continue;
-
+
// If there is an non-load/store instruction in the loop, we can't promote
// it.
- if (isa<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
+ if (LoadInst *load = dyn_cast<LoadInst>(Use)) {
+ assert(!load->isVolatile() && "AST broken");
+ if (!load->isSimple())
+ return;
+ } 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;
+ assert(!store->isVolatile() && "AST broken");
+ if (!store->isSimple())
+ return;
+
+ // Note that we only check GuaranteedToExecute inside the store case
+ // so that we do not introduce stores where they did not exist before
+ // (which would break the LLVM concurrency model).
+
+ // 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.
+ unsigned InstAlignment = store->getAlignment();
+ if ((InstAlignment > Alignment || InstAlignment == 0) && Alignment != 0)
+ if (isGuaranteedToExecute(*Use)) {
+ GuaranteedToExecute = true;
+ Alignment = InstAlignment;
+ }
+
+ if (!GuaranteedToExecute)
+ GuaranteedToExecute = isGuaranteedToExecute(*Use);
+
+ } else
return; // Not a load or store.
-
- if (!GuaranteedToExecute)
- GuaranteedToExecute = isSafeToExecuteUnconditionally(*Use);
-
+
+ // Merge the TBAA tags.
+ if (LoopUses.empty()) {
+ // On the first load/store, just take its TBAA tag.
+ TBAATag = Use->getMetadata(LLVMContext::MD_tbaa);
+ } else if (TBAATag && TBAATag != Use->getMetadata(LLVMContext::MD_tbaa)) {
+ TBAATag = 0;
+ }
+
LoopUses.push_back(Use);
}
}
-
+
// If there isn't a guaranteed-to-execute instruction, we can't promote.
if (!GuaranteedToExecute)
return;
-
+
// Otherwise, this is safe to promote, lets do it!
- DEBUG(dbgs() << "LICM: Promoting value stored to in loop: " <<*SomePtr<<'\n');
+ DEBUG(dbgs() << "LICM: Promoting value stored to in loop: " <<*SomePtr<<'\n');
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();
+
+ // Figure out the loop exits and their insertion points, if this is the
+ // first promotion.
+ if (ExitBlocks.empty()) {
+ CurLoop->getUniqueExitBlocks(ExitBlocks);
+ InsertPts.resize(ExitBlocks.size());
+ for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i)
+ InsertPts[i] = ExitBlocks[i]->getFirstInsertionPt();
+ }
+
// We use the SSAUpdater interface to insert phi nodes as required.
SmallVector<PHINode*, 16> NewPHIs;
SSAUpdater SSA(&NewPHIs);
-
- // 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());
+ LoopPromoter Promoter(SomePtr, LoopUses, SSA, PointerMustAliases, ExitBlocks,
+ InsertPts, *CurAST, DL, Alignment, TBAATag);
+
+ // 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);
+ if (TBAATag) PreheaderLoad->setMetadata(LLVMContext::MD_tbaa, TBAATag);
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);
- }
-
- // 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);
+ // Rewrite all the loads in the loop and remember all the definitions from
+ // stores in the loop.
+ Promoter.run(LoopUses);
- 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();
}
projects / oota-llvm.git / blobdiff