//===-- LICM.cpp - Loop Invariant Code Motion Pass ------------------------===//
//
-// This pass is a simple loop invariant code motion pass.
+// This pass is a simple loop invariant code motion pass. An interesting aspect
+// of this pass is that it uses alias analysis for two purposes:
//
-// Note that this pass does NOT require pre-headers to exist on loops in the
-// CFG, but if there is not distinct preheader for a loop, the hoisted code will
-// be *DUPLICATED* in every basic block, outside of the loop, that preceeds the
-// loop header. Additionally, any use of one of these hoisted expressions
-// cannot be loop invariant itself, because the expression hoisted gets a PHI
-// node that is loop variant.
-//
-// For these reasons, and many more, it makes sense to run a pass before this
-// that ensures that there are preheaders on all loops. That said, we don't
-// REQUIRE it. :)
+// 1. Moving loop invariant loads out of loops. If we can determine that a
+// load inside of a loop never aliases anything stored to, we can hoist it
+// like any other instruction.
+// 2. Scalar Promotion of Memory - If there is a store instruction inside of
+// the loop, we try to move the store to happen AFTER the loop instead of
+// inside of the loop. This can only happen if a few conditions are true:
+// A. The pointer stored through is loop invariant
+// B. There are no stores or loads in the loop which _may_ alias the
+// pointer. There are no calls in the loop which mod/ref the pointer.
+// If these conditions are true, we can promote the loads and stores in the
+// loop of the pointer to use a temporary alloca'd variable. We then use
+// the mem2reg functionality to construct the appropriate SSA form for the
+// variable.
//
//===----------------------------------------------------------------------===//
#include "llvm/Transforms/Scalar.h"
+#include "llvm/Transforms/Utils/PromoteMemToReg.h"
#include "llvm/Transforms/Utils/Local.h"
#include "llvm/Analysis/LoopInfo.h"
-#include "llvm/iOperators.h"
-#include "llvm/iPHINode.h"
+#include "llvm/Analysis/AliasAnalysis.h"
+#include "llvm/Analysis/AliasSetTracker.h"
+#include "llvm/Analysis/Dominators.h"
+#include "llvm/Instructions.h"
+#include "llvm/DerivedTypes.h"
+#include "llvm/Target/TargetData.h"
#include "llvm/Support/InstVisitor.h"
#include "llvm/Support/CFG.h"
-#include "Support/STLExtras.h"
-#include "Support/StatisticReporter.h"
+#include "Support/Statistic.h"
+#include "Support/CommandLine.h"
+#include "llvm/Assembly/Writer.h"
#include <algorithm>
-using std::string;
-
-static Statistic<> NumHoistedNPH("licm\t\t- Number of insts hoisted to multiple"
- " loop preds (bad, no loop pre-header)");
-static Statistic<> NumHoistedPH("licm\t\t- Number of insts hoisted to a loop "
- "pre-header");
namespace {
+ cl::opt<bool> DisablePromotion("disable-licm-promotion", cl::Hidden,
+ cl::desc("Disable memory promotion in LICM pass"));
+
+ Statistic<> NumHoisted("licm", "Number of instructions hoisted out of loop");
+ Statistic<> NumHoistedLoads("licm", "Number of load insts hoisted");
+ Statistic<> NumPromoted("licm", "Number of memory locations promoted to registers");
+
struct LICM : public FunctionPass, public InstVisitor<LICM> {
virtual bool runOnFunction(Function &F);
- // This transformation requires natural loop information...
+ /// This transformation requires natural loop information & requires that
+ /// loop preheaders be inserted into the CFG...
+ ///
virtual void getAnalysisUsage(AnalysisUsage &AU) const {
- AU.preservesCFG();
- AU.addRequired(LoopInfo::ID);
+ AU.setPreservesCFG();
+ AU.addRequiredID(LoopPreheadersID);
+ AU.addRequired<LoopInfo>();
+ AU.addRequired<DominatorTree>();
+ AU.addRequired<DominanceFrontier>(); // For scalar promotion (mem2reg)
+ AU.addRequired<AliasAnalysis>();
}
private:
- // List of predecessor blocks for the current loop - These blocks are where
- // we hoist loop invariants to for the current loop.
- //
- std::vector<BasicBlock*> LoopPreds, LoopBackEdges;
-
- Loop *CurLoop; // The current loop we are working on...
- bool Changed; // Set to true when we change anything.
-
- // visitLoop - Hoist expressions out of the specified loop...
- void visitLoop(Loop *L);
-
- // notInCurrentLoop - Little predicate that returns true if the specified
- // basic block is in a subloop of the current one, not the current one
- // itself.
- //
- bool notInCurrentLoop(BasicBlock *BB) {
+ LoopInfo *LI; // Current LoopInfo
+ AliasAnalysis *AA; // Current AliasAnalysis information
+ 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...
+
+ /// visitLoop - Hoist expressions out of the specified loop...
+ ///
+ void visitLoop(Loop *L, AliasSetTracker &AST);
+
+ /// 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 defintions before uses, allowing us to hoist a loop body in one
+ /// pass without iteration.
+ ///
+ void HoistRegion(DominatorTree::Node *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.
+ ///
+ bool inSubLoop(BasicBlock *BB) {
+ assert(CurLoop->contains(BB) && "Only valid if BB is IN the loop");
for (unsigned i = 0, e = CurLoop->getSubLoops().size(); i != e; ++i)
if (CurLoop->getSubLoops()[i]->contains(BB))
return true; // A subloop actually contains this block!
- return false;
+ return false;
}
- // hoist - When an instruction is found to only use loop invariant operands
- // that is safe to hoist, this instruction is called to do the dirty work.
- //
+ /// hoist - When an instruction is found to only use loop invariant operands
+ /// that is safe to hoist, this instruction is called to do the dirty work.
+ ///
void hoist(Instruction &I);
- // isLoopInvariant - Return true if the specified value is loop invariant
+ /// pointerInvalidatedByLoop - Return true if the body of this loop may
+ /// store into the memory location pointed to by V.
+ ///
+ bool pointerInvalidatedByLoop(Value *V) {
+ // Check to see if any of the basic blocks in CurLoop invalidate *V.
+ return CurAST->getAliasSetForPointer(V, 0).isMod();
+ }
+
+ /// isLoopInvariant - Return true if the specified value is loop invariant
+ ///
inline bool isLoopInvariant(Value *V) {
if (Instruction *I = dyn_cast<Instruction>(V))
return !CurLoop->contains(I->getParent());
return true; // All non-instructions are loop invariant
}
- // visitBasicBlock - Run LICM on a particular block.
- void visitBasicBlock(BasicBlock *BB);
+ /// PromoteValuesInLoop - Look at the stores in the loop and promote as many
+ /// to scalars as we can.
+ ///
+ void PromoteValuesInLoop();
+
+ /// findPromotableValuesInLoop - Check the current loop for stores to
+ /// definate 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...
+ ///
+ void findPromotableValuesInLoop(
+ std::vector<std::pair<AllocaInst*, Value*> > &PromotedValues,
+ std::map<Value*, AllocaInst*> &Val2AlMap);
+
- // Instruction visitation handlers... these basically control whether or not
- // the specified instruction types are hoisted.
- //
+ /// Instruction visitation handlers... these basically control whether or
+ /// not the specified instruction types are hoisted.
+ ///
friend class InstVisitor<LICM>;
- void visitUnaryOperator(Instruction &I) {
- if (isLoopInvariant(I.getOperand(0))) hoist(I);
- }
void visitBinaryOperator(Instruction &I) {
if (isLoopInvariant(I.getOperand(0)) && isLoopInvariant(I.getOperand(1)))
hoist(I);
}
-
- void visitCastInst(CastInst &I) { visitUnaryOperator((Instruction&)I); }
+ void visitCastInst(CastInst &CI) {
+ Instruction &I = (Instruction&)CI;
+ if (isLoopInvariant(I.getOperand(0))) hoist(I);
+ }
void visitShiftInst(ShiftInst &I) { visitBinaryOperator((Instruction&)I); }
+ void visitLoadInst(LoadInst &LI);
+
void visitGetElementPtrInst(GetElementPtrInst &GEPI) {
Instruction &I = (Instruction&)GEPI;
for (unsigned i = 0, e = I.getNumOperands(); i != e; ++i)
Pass *createLICMPass() { return new LICM(); }
+/// runOnFunction - For LICM, this simply traverses the loop structure of the
+/// function, hoisting expressions out of loops if possible.
+///
bool LICM::runOnFunction(Function &) {
- // get our loop information...
- const std::vector<Loop*> &TopLevelLoops =
- getAnalysis<LoopInfo>().getTopLevelLoops();
-
- // Traverse loops in postorder, hoisting expressions out of the deepest loops
- // first.
- //
Changed = false;
- std::for_each(TopLevelLoops.begin(), TopLevelLoops.end(),
- bind_obj(this, &LICM::visitLoop));
- return Changed;
-}
-void LICM::visitLoop(Loop *L) {
- // Recurse through all subloops before we process this loop...
- std::for_each(L->getSubLoops().begin(), L->getSubLoops().end(),
- bind_obj(this, &LICM::visitLoop));
- CurLoop = L;
+ // Get our Loop and Alias Analysis information...
+ LI = &getAnalysis<LoopInfo>();
+ AA = &getAnalysis<AliasAnalysis>();
- // Calculate the set of predecessors for this loop. The predecessors for this
- // loop are equal to the predecessors for the header node of the loop that are
- // not themselves in the loop.
- //
- BasicBlock *Header = L->getHeader();
+ // Hoist expressions out of all of the top-level loops.
+ const std::vector<Loop*> &TopLevelLoops = LI->getTopLevelLoops();
+ for (std::vector<Loop*>::const_iterator I = TopLevelLoops.begin(),
+ E = TopLevelLoops.end(); I != E; ++I) {
+ AliasSetTracker AST(*AA);
+ LICM::visitLoop(*I, AST);
+ }
+ return Changed;
+}
- // Calculate the sets of predecessors and backedges of the loop...
- LoopBackEdges.insert(LoopBackEdges.end(),pred_begin(Header),pred_end(Header));
- std::vector<BasicBlock*>::iterator LPI =
- std::partition(LoopBackEdges.begin(), LoopBackEdges.end(),
- bind_obj(CurLoop, &Loop::contains));
+/// 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 (std::vector<Loop*>::const_iterator I = L->getSubLoops().begin(),
+ E = L->getSubLoops().end(); I != E; ++I) {
+ AliasSetTracker SubAST(*AA);
+ LICM::visitLoop(*I, SubAST);
- // Move all predecessors to the LoopPreds vector...
- LoopPreds.insert(LoopPreds.end(), LPI, LoopBackEdges.end());
+ // Incorporate information about the subloops into this loop...
+ AST.add(SubAST);
+ }
+ CurLoop = L;
+ CurAST = &AST;
- // Remove predecessors from backedges list...
- LoopBackEdges.erase(LPI, LoopBackEdges.end());
-
+ // Get the preheader block to move instructions into...
+ Preheader = L->getLoopPreheader();
+ assert(Preheader&&"Preheader insertion pass guarantees we have a preheader!");
- // The only way that there could be no predecessors to a loop is if the loop
- // is not reachable. Since we don't care about optimizing dead loops,
- // summarily ignore them.
+ // 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.
//
- if (LoopPreds.empty()) return;
-
+ const std::vector<BasicBlock*> &LoopBBs = L->getBlocks();
+ for (std::vector<BasicBlock*>::const_iterator I = LoopBBs.begin(),
+ E = LoopBBs.end(); I != E; ++I)
+ if (LI->getLoopFor(*I) == L) // Ignore blocks in subloops...
+ AST.add(**I); // 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
// their loop, into this loop, so there is no need to process the BODIES of
// the subloops).
//
- std::vector<BasicBlock*> BBs(L->getBlocks().begin(), L->getBlocks().end());
-
- // Remove blocks that are actually in subloops...
- BBs.erase(std::remove_if(BBs.begin(), BBs.end(),
- bind_obj(this, &LICM::notInCurrentLoop)), BBs.end());
-
- // Visit all of the basic blocks we have chosen, hoisting out the instructions
- // as neccesary. This leaves dead copies of the instruction in the loop
- // unfortunately...
+ // 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 perform the LICM transformation in one pass, without iteration.
//
- for_each(BBs.begin(), BBs.end(), bind_obj(this, &LICM::visitBasicBlock));
+ HoistRegion(getAnalysis<DominatorTree>()[L->getHeader()]);
+
+ // Now that all loop invariants have been removed from the loop, promote any
+ // memory references to scalars that we can...
+ if (!DisablePromotion)
+ PromoteValuesInLoop();
// Clear out loops state information for the next iteration
CurLoop = 0;
- LoopPreds.clear();
- LoopBackEdges.clear();
+ Preheader = 0;
}
-void LICM::visitBasicBlock(BasicBlock *BB) {
- for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ) {
- visit(*I);
-
- if (dceInstruction(I))
- Changed = true;
- else
- ++I;
- }
+/// 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 defintions
+/// before uses, allowing us to hoist a loop body in one pass without iteration.
+///
+void LICM::HoistRegion(DominatorTree::Node *N) {
+ assert(N != 0 && "Null dominator tree node?");
+
+ // If this subregion is not in the top level loop at all, exit.
+ if (!CurLoop->contains(N->getNode())) return;
+
+ // Only need to hoist the contents of this block if it is not part of a
+ // subloop (which would already have been hoisted)
+ if (!inSubLoop(N->getNode()))
+ visit(*N->getNode());
+
+ const std::vector<DominatorTree::Node*> &Children = N->getChildren();
+ for (unsigned i = 0, e = Children.size(); i != e; ++i)
+ HoistRegion(Children[i]);
}
+/// hoist - When an instruction is found to only use loop invariant operands
+/// that is safe to hoist, this instruction is called to do the dirty work.
+///
void LICM::hoist(Instruction &Inst) {
- if (Inst.use_empty()) return; // Don't (re) hoist dead instructions!
- //cerr << "Hoisting " << Inst;
-
- BasicBlock *Header = CurLoop->getHeader();
-
- // Old instruction will be removed, so take it's name...
- string InstName = Inst.getName();
- Inst.setName("");
-
- // The common case is that we have a pre-header. Generate special case code
- // that is faster if that is the case.
- //
- if (LoopPreds.size() == 1) {
- BasicBlock *Pred = LoopPreds[0];
-
- // Create a new copy of the instruction, for insertion into Pred.
- Instruction *New = Inst.clone();
- New->setName(InstName);
+ DEBUG(std::cerr << "LICM hoisting to";
+ WriteAsOperand(std::cerr, Preheader, false);
+ std::cerr << ": " << Inst);
- // Insert the new node in Pred, before the terminator.
- Pred->getInstList().insert(--Pred->end(), New);
+ // Remove the instruction from its current basic block... but don't delete the
+ // instruction.
+ Inst.getParent()->getInstList().remove(&Inst);
- // Kill the old instruction...
- Inst.replaceAllUsesWith(New);
- ++NumHoistedPH;
-
- } else {
- // No loop pre-header, insert a PHI node into header to capture all of the
- // incoming versions of the value.
- //
- PHINode *LoopVal = new PHINode(Inst.getType(), InstName+".phi");
+ // Insert the new node in Preheader, before the terminator.
+ Preheader->getInstList().insert(Preheader->getTerminator(), &Inst);
+
+ ++NumHoisted;
+ Changed = true;
+}
- // Insert the new PHI node into the loop header...
- Header->getInstList().push_front(LoopVal);
- // Insert cloned versions of the instruction into all of the loop preds.
- for (unsigned i = 0, e = LoopPreds.size(); i != e; ++i) {
- BasicBlock *Pred = LoopPreds[i];
-
- // Create a new copy of the instruction, for insertion into Pred.
- Instruction *New = Inst.clone();
- New->setName(InstName);
+void LICM::visitLoadInst(LoadInst &LI) {
+ if (isLoopInvariant(LI.getOperand(0)) &&
+ !pointerInvalidatedByLoop(LI.getOperand(0))) {
+ hoist(LI);
+ ++NumHoistedLoads;
+ }
+}
- // Insert the new node in Pred, before the terminator.
- Pred->getInstList().insert(--Pred->end(), New);
+/// PromoteValuesInLoop - 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. We promote these memory locations to use allocas
+/// instead. These allocas can easily be raised to register values by the
+/// PromoteMem2Reg functionality.
+///
+void LICM::PromoteValuesInLoop() {
+ // PromotedValues - List of values that are promoted out of the loop. Each
+ // value has an alloca instruction for it, and a cannonical version of the
+ // pointer.
+ std::vector<std::pair<AllocaInst*, Value*> > PromotedValues;
+ std::map<Value*, AllocaInst*> ValueToAllocaMap; // Map of ptr to alloca
+
+ findPromotableValuesInLoop(PromotedValues, ValueToAllocaMap);
+ if (ValueToAllocaMap.empty()) return; // If there are values to promote...
- // Add the incoming value to the PHI node.
- LoopVal->addIncoming(New, Pred);
+ Changed = true;
+ NumPromoted += PromotedValues.size();
+
+ // Emit a copy from the value into the alloca'd value in the loop preheader
+ TerminatorInst *LoopPredInst = Preheader->getTerminator();
+ for (unsigned i = 0, e = PromotedValues.size(); i != e; ++i) {
+ // Load from the memory we are promoting...
+ LoadInst *LI = new LoadInst(PromotedValues[i].second,
+ PromotedValues[i].second->getName()+".promoted",
+ LoopPredInst);
+ // 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. If we find a branch that exits the
+ // loop, make sure to put reload code into all of the successors of the
+ // loop.
+ //
+ const std::vector<BasicBlock*> &LoopBBs = CurLoop->getBlocks();
+ for (std::vector<BasicBlock*>::const_iterator I = LoopBBs.begin(),
+ E = LoopBBs.end(); I != E; ++I) {
+ // Rewrite all loads and stores in the block of the pointer...
+ for (BasicBlock::iterator II = (*I)->begin(), E = (*I)->end();
+ II != E; ++II) {
+ if (LoadInst *L = dyn_cast<LoadInst>(II)) {
+ std::map<Value*, AllocaInst*>::iterator
+ I = ValueToAllocaMap.find(L->getOperand(0));
+ if (I != ValueToAllocaMap.end())
+ L->setOperand(0, I->second); // Rewrite load instruction...
+ } else if (StoreInst *S = dyn_cast<StoreInst>(II)) {
+ std::map<Value*, AllocaInst*>::iterator
+ I = ValueToAllocaMap.find(S->getOperand(1));
+ if (I != ValueToAllocaMap.end())
+ S->setOperand(1, I->second); // Rewrite store instruction...
+ }
}
- // Add incoming values to the PHI node for all backedges in the loop...
- for (unsigned i = 0, e = LoopBackEdges.size(); i != e; ++i)
- LoopVal->addIncoming(LoopVal, LoopBackEdges[i]);
-
- // Replace all uses of the old version of the instruction in the loop with
- // the new version that is out of the loop. We know that this is ok,
- // because the new definition is in the loop header, which dominates the
- // entire loop body. The old definition was defined _inside_ of the loop,
- // so the scope cannot extend outside of the loop, so we're ok.
+ // Check to see if any successors of this block are outside of the loop.
+ // If so, we need to copy the value from the alloca back into the memory
+ // location...
//
- Inst.replaceAllUsesWith(LoopVal);
- ++NumHoistedNPH;
+ for (succ_iterator SI = succ_begin(*I), SE = succ_end(*I); SI != SE; ++SI)
+ if (!CurLoop->contains(*SI)) {
+ // Copy all of the allocas into their memory locations...
+ BasicBlock::iterator BI = (*SI)->begin();
+ while (isa<PHINode>(*BI))
+ ++BI; // Skip over all of the phi nodes in the block...
+ Instruction *InsertPos = BI;
+ for (unsigned i = 0, e = PromotedValues.size(); i != e; ++i) {
+ // Load from the alloca...
+ LoadInst *LI = new LoadInst(PromotedValues[i].first, "", InsertPos);
+ // Store into the memory we promoted...
+ new StoreInst(LI, PromotedValues[i].second, InsertPos);
+ }
+ }
}
- Changed = true;
+ // 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...
+ //
+ std::vector<AllocaInst*> PromotedAllocas;
+ PromotedAllocas.reserve(PromotedValues.size());
+ for (unsigned i = 0, e = PromotedValues.size(); i != e; ++i)
+ PromotedAllocas.push_back(PromotedValues[i].first);
+ PromoteMemToReg(PromotedAllocas, getAnalysis<DominanceFrontier>(),
+ AA->getTargetData());
}
+/// findPromotableValuesInLoop - Check the current loop for stores to definate
+/// 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...
+///
+void LICM::findPromotableValuesInLoop(
+ std::vector<std::pair<AllocaInst*, Value*> > &PromotedValues,
+ std::map<Value*, AllocaInst*> &ValueToAllocaMap) {
+ Instruction *FnStart = CurLoop->getHeader()->getParent()->begin()->begin();
+
+ // Loop over all of the alias sets in the tracker object...
+ for (AliasSetTracker::iterator I = CurAST->begin(), E = CurAST->end();
+ I != E; ++I) {
+ AliasSet &AS = *I;
+ // We can promote this alias set if it has a store, if it is a "Must" alias
+ // set, and if the pointer is loop invariant.
+ if (!AS.isForwardingAliasSet() && AS.isMod() && AS.isMustAlias() &&
+ 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.
+ bool PointerOk = true;
+ for (AliasSet::iterator I = AS.begin(), E = AS.end(); I != E; ++I)
+ if (V->getType() != I->first->getType()) {
+ PointerOk = false;
+ break;
+ }
+
+ 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));
+
+ 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");
+ }
+ }
+ }
+}