#define DEBUG_TYPE "licm"
#include "llvm/Transforms/Scalar.h"
+#include "llvm/Constants.h"
#include "llvm/DerivedTypes.h"
#include "llvm/Instructions.h"
#include "llvm/Target/TargetData.h"
#include "llvm/Analysis/LoopInfo.h"
+#include "llvm/Analysis/LoopPass.h"
#include "llvm/Analysis/AliasAnalysis.h"
#include "llvm/Analysis/AliasSetTracker.h"
#include "llvm/Analysis/Dominators.h"
-#include "llvm/Support/CFG.h"
#include "llvm/Transforms/Utils/PromoteMemToReg.h"
-#include "llvm/Transforms/Utils/Local.h"
+#include "llvm/Support/CFG.h"
+#include "llvm/Support/Compiler.h"
#include "llvm/Support/CommandLine.h"
#include "llvm/Support/Debug.h"
#include "llvm/ADT/Statistic.h"
#include <algorithm>
using namespace llvm;
+STATISTIC(NumSunk , "Number of instructions sunk out of loop");
+STATISTIC(NumHoisted , "Number of instructions hoisted out of loop");
+STATISTIC(NumMovedLoads, "Number of load insts hoisted or sunk");
+STATISTIC(NumMovedCalls, "Number of call insts hoisted or sunk");
+STATISTIC(NumPromoted , "Number of memory locations promoted to registers");
+
namespace {
cl::opt<bool>
DisablePromotion("disable-licm-promotion", cl::Hidden,
cl::desc("Disable memory promotion in LICM pass"));
- Statistic<> NumSunk("licm", "Number of instructions sunk out of loop");
- Statistic<> NumHoisted("licm", "Number of instructions hoisted out of loop");
- Statistic<> NumMovedLoads("licm", "Number of load insts hoisted or sunk");
- Statistic<> NumMovedCalls("licm", "Number of call insts hoisted or sunk");
- Statistic<> NumPromoted("licm",
- "Number of memory locations promoted to registers");
-
- struct LICM : public FunctionPass {
- virtual bool runOnFunction(Function &F);
+ struct VISIBILITY_HIDDEN LICM : public LoopPass {
+ virtual bool runOnLoop(Loop *L, LPPassManager &LPM);
/// This transformation requires natural loop information & requires that
/// loop preheaders be inserted into the CFG...
AU.setPreservesCFG();
AU.addRequiredID(LoopSimplifyID);
AU.addRequired<LoopInfo>();
- AU.addRequired<DominatorTree>();
+ AU.addRequired<ETForest>();
AU.addRequired<DominanceFrontier>(); // For scalar promotion (mem2reg)
AU.addRequired<AliasAnalysis>();
}
+ bool doFinalization() {
+ LoopToAliasMap.clear();
+ return false;
+ }
+
private:
// Various analyses that we use...
AliasAnalysis *AA; // Current AliasAnalysis information
LoopInfo *LI; // Current LoopInfo
- DominatorTree *DT; // Dominator Tree for the current Loop...
+ ETForest *ET; // ETForest for the current Loop...
DominanceFrontier *DF; // Current Dominance Frontier
// State that is updated as we process loops
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);
+ std::map<Loop *, AliasSetTracker *> LoopToAliasMap;
/// SinkRegion - Walk the specified region of the CFG (defined by all blocks
/// dominated by the specified block, and that are in the current loop) in
- /// reverse depth first order w.r.t the DominatorTree. This allows us to
+ /// reverse depth first order w.r.t the ETForest. This allows us to
/// visit uses before definitions, allowing us to sink a loop body in one
/// pass without iteration.
///
- void SinkRegion(DominatorTree::Node *N);
+ void SinkRegion(BasicBlock *BB);
/// 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
+ /// loop) in depth first order w.r.t the ETForest. This allows us to
/// visit definitions before uses, allowing us to hoist a loop body in one
/// pass without iteration.
///
- void HoistRegion(DominatorTree::Node *N);
+ void HoistRegion(BasicBlock *BB);
/// inSubLoop - Little predicate that returns true if the specified basic
/// block is in a subloop of the current one, not the current one itself.
if (BlockInLoop == LoopHeader)
return true;
- DominatorTree::Node *BlockInLoopNode = DT->getNode(BlockInLoop);
- DominatorTree::Node *IDom = DT->getNode(ExitBlock);
+ BasicBlock *IDom = ExitBlock;
// Because the exit block is not in the loop, we know we have to get _at
// least_ its immediate dominator.
do {
// Get next Immediate Dominator.
- IDom = IDom->getIDom();
+ IDom = ET->getIDom(IDom);
// 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)
+ if (IDom == LoopHeader)
return false;
- } while (IDom != BlockInLoopNode);
+ } while (IDom != BlockInLoop);
return true;
}
std::map<Value*, AllocaInst*> &Val2AlMap);
};
- RegisterOpt<LICM> X("licm", "Loop Invariant Code Motion");
+ RegisterPass<LICM> X("licm", "Loop Invariant Code Motion");
}
-FunctionPass *llvm::createLICMPass() { return new LICM(); }
+LoopPass *llvm::createLICMPass() { return new LICM(); }
-/// runOnFunction - For LICM, this simply traverses the loop structure of the
-/// function, hoisting expressions out of loops if possible.
+/// Hoist expressions out of the specified loop...
///
-bool LICM::runOnFunction(Function &) {
+bool LICM::runOnLoop(Loop *L, LPPassManager &LPM) {
Changed = false;
// Get our Loop and Alias Analysis information...
LI = &getAnalysis<LoopInfo>();
AA = &getAnalysis<AliasAnalysis>();
DF = &getAnalysis<DominanceFrontier>();
- DT = &getAnalysis<DominatorTree>();
-
- // Hoist expressions out of all of the top-level loops.
- for (LoopInfo::iterator I = LI->begin(), E = LI->end(); I != E; ++I) {
- AliasSetTracker AST(*AA);
- visitLoop(*I, AST);
- }
- return Changed;
-}
-
-
-/// visitLoop - Hoist expressions out of the specified loop...
-///
-void LICM::visitLoop(Loop *L, AliasSetTracker &AST) {
- // Recurse through all subloops before we process this loop...
- for (Loop::iterator I = L->begin(), E = L->end(); I != E; ++I) {
- AliasSetTracker SubAST(*AA);
- visitLoop(*I, SubAST);
-
- // Incorporate information about the subloops into this loop...
- AST.add(SubAST);
+ ET = &getAnalysis<ETForest>();
+
+ CurAST = new AliasSetTracker(*AA);
+ // Collect Alias info frmo subloops
+ for (Loop::iterator LoopItr = L->begin(), LoopItrE = L->end();
+ LoopItr != LoopItrE; ++LoopItr) {
+ Loop *InnerL = *LoopItr;
+ AliasSetTracker *InnerAST = LoopToAliasMap[InnerL];
+ assert (InnerAST && "Where is my AST?");
+
+ // What if InnerLoop was modified by other passes ?
+ CurAST->add(*InnerAST);
}
+
CurLoop = L;
- CurAST = &AST;
// Get the preheader block to move instructions into...
Preheader = L->getLoopPreheader();
for (std::vector<BasicBlock*>::const_iterator I = L->getBlocks().begin(),
E = L->getBlocks().end(); I != E; ++I)
if (LI->getLoopFor(*I) == L) // Ignore blocks in subloops...
- AST.add(**I); // Incorporate the specified basic block
+ CurAST->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
// 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()));
+ SinkRegion(L->getHeader());
+ HoistRegion(L->getHeader());
// Now that all loop invariants have been removed from the loop, promote any
// memory references to scalars that we can...
// Clear out loops state information for the next iteration
CurLoop = 0;
Preheader = 0;
+
+ LoopToAliasMap[L] = CurAST;
+ return Changed;
}
/// SinkRegion - Walk the specified region of the CFG (defined by all blocks
/// dominated by the specified block, and that are in the current loop) in
-/// reverse depth first order w.r.t the DominatorTree. This allows us to visit
+/// reverse depth first order w.r.t the ETForest. This allows us to visit
/// uses before definitions, allowing us to sink a loop body in one pass without
/// iteration.
///
-void LICM::SinkRegion(DominatorTree::Node *N) {
- assert(N != 0 && "Null dominator tree node?");
- BasicBlock *BB = N->getBlock();
+void LICM::SinkRegion(BasicBlock *BB) {
+ assert(BB != 0 && "Null sink block?");
// 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...
- const std::vector<DominatorTree::Node*> &Children = N->getChildren();
+ std::vector<BasicBlock*> Children;
+ ET->getChildren(BB, Children);
for (unsigned i = 0, e = Children.size(); i != e; ++i)
SinkRegion(Children[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 definitions
+/// first order w.r.t the ETForest. This allows us to visit definitions
/// before uses, allowing us to hoist a loop body in one pass without iteration.
///
-void LICM::HoistRegion(DominatorTree::Node *N) {
- assert(N != 0 && "Null dominator tree node?");
- BasicBlock *BB = N->getBlock();
+void LICM::HoistRegion(BasicBlock *BB) {
+ assert(BB != 0 && "Null hoist block?");
// If this subregion is not in the top level loop at all, exit.
if (!CurLoop->contains(BB)) return;
//
if (isLoopInvariantInst(I) && canSinkOrHoistInst(I) &&
isSafeToExecuteUnconditionally(I))
- hoist(I);
+ hoist(I);
}
- const std::vector<DominatorTree::Node*> &Children = N->getChildren();
+ std::vector<BasicBlock*> Children;
+ ET->getChildren(BB, Children);
for (unsigned i = 0, e = Children.size(); i != e; ++i)
HoistRegion(Children[i]);
}
return false;
}
- return isa<BinaryOperator>(I) || isa<ShiftInst>(I) || isa<CastInst>(I) ||
- isa<SelectInst>(I) ||
- isa<GetElementPtrInst>(I) || isa<VAArgInst>(I);
+ // Otherwise these instructions are hoistable/sinkable
+ return isa<BinaryOperator>(I) || isa<CastInst>(I) ||
+ isa<SelectInst>(I) || isa<GetElementPtrInst>(I) || isa<CmpInst>(I);
}
/// isNotUsedInLoop - Return true if the only users of this instruction are
/// position, and may either delete it or move it to outside of the loop.
///
void LICM::sink(Instruction &I) {
- DEBUG(std::cerr << "LICM sinking instruction: " << I);
+ DOUT << "LICM sinking instruction: " << I;
std::vector<BasicBlock*> ExitBlocks;
CurLoop->getExitBlocks(ExitBlocks);
if (!isExitBlockDominatedByBlockInLoop(ExitBlocks[0], I.getParent())) {
// Instruction is not used, just delete it.
CurAST->deleteValue(&I);
- I.getParent()->getInstList().erase(&I);
+ if (!I.use_empty()) // If I has users in unreachable blocks, eliminate.
+ I.replaceAllUsesWith(UndefValue::get(I.getType()));
+ I.eraseFromParent();
} else {
// Move the instruction to the start of the exit block, after any PHI
// nodes in it.
- I.getParent()->getInstList().remove(&I);
+ I.removeFromParent();
BasicBlock::iterator InsertPt = ExitBlocks[0]->begin();
while (isa<PHINode>(InsertPt)) ++InsertPt;
} else if (ExitBlocks.size() == 0) {
// The instruction is actually dead if there ARE NO exit blocks.
CurAST->deleteValue(&I);
- I.getParent()->getInstList().erase(&I);
+ if (!I.use_empty()) // If I has users in unreachable blocks, eliminate.
+ I.replaceAllUsesWith(UndefValue::get(I.getType()));
+ I.eraseFromParent();
} else {
// Otherwise, if we have multiple exits, use the PromoteMem2Reg function to
// do all of the hard work of inserting PHI nodes as necessary. We convert
if (I.getType() != Type::VoidTy)
AI = new AllocaInst(I.getType(), 0, I.getName(),
- I.getParent()->getParent()->front().begin());
+ I.getParent()->getParent()->getEntryBlock().begin());
// Secondly, insert load instructions for each use of the instruction
// outside of the loop.
// the copy.
Instruction *New;
if (InsertedBlocks.size() == 1) {
- I.getParent()->getInstList().remove(&I);
+ I.removeFromParent();
ExitBlock->getInstList().insert(InsertPt, &I);
New = &I;
} else {
// If the instruction doesn't dominate any exit blocks, it must be dead.
if (InsertedBlocks.empty()) {
CurAST->deleteValue(&I);
- I.getParent()->getInstList().erase(&I);
+ I.eraseFromParent();
}
// Finally, promote the fine value to SSA form.
if (AI) {
std::vector<AllocaInst*> Allocas;
Allocas.push_back(AI);
- PromoteMemToReg(Allocas, *DT, *DF, AA->getTargetData(), CurAST);
+ PromoteMemToReg(Allocas, *ET, *DF, AA->getTargetData(), CurAST);
}
}
}
/// that is safe to hoist, this instruction is called to do the dirty work.
///
void LICM::hoist(Instruction &I) {
- DEBUG(std::cerr << "LICM hoisting to " << Preheader->getName()
- << ": " << I);
+ DOUT << "LICM hoisting to " << Preheader->getName() << ": " << I;
// Remove the instruction from its current basic block... but don't delete the
// instruction.
- I.getParent()->getInstList().remove(&I);
+ I.removeFromParent();
// Insert the new node in Preheader, before the terminator.
Preheader->getInstList().insert(Preheader->getTerminator(), &I);
std::vector<BasicBlock*> ExitBlocks;
CurLoop->getExitBlocks(ExitBlocks);
- // For each exit block, get the DT node and walk up the DT until the
+ // For each exit block, walk up the ET until the
// instruction's basic block is found or we exit the loop.
for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i)
if (!isExitBlockDominatedByBlockInLoop(ExitBlocks[i], Inst.getParent()))
PromotedAllocas.reserve(PromotedValues.size());
for (unsigned i = 0, e = PromotedValues.size(); i != e; ++i)
PromotedAllocas.push_back(PromotedValues[i].first);
- PromoteMemToReg(PromotedAllocas, *DT, *DF, AA->getTargetData(), CurAST);
+ PromoteMemToReg(PromotedAllocas, *ET, *DF, AA->getTargetData(), CurAST);
}
/// FindPromotableValuesInLoop - Check the current loop for stores to definite
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");
+ DOUT << "LICM: Promoting value: " << *V << "\n";
}
}
}