//===-- LICM.cpp - Loop Invariant Code Motion Pass ------------------------===//
-//
+//
// 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 pass performs loop invariant code motion, attempting to remove as much
// code from the body of a loop as possible. It does this by either hoisting
// code into the preheader block, or by sinking code to the exit blocks if it is
// safe. This pass also promotes must-aliased memory locations in the loop to
-// live in registers.
+// live in registers, thus hoisting and sinking "invariant" loads and stores.
//
// This pass uses alias analysis for two purposes:
//
-// 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
-// or sink it like any other instruction.
+// 1. Moving loop invariant loads and calls out of loops. If we can determine
+// that a load or call inside of a loop never aliases anything stored to,
+// we can hoist it or sink 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:
//
//===----------------------------------------------------------------------===//
+#define DEBUG_TYPE "licm"
#include "llvm/Transforms/Scalar.h"
-#include "llvm/Transforms/Utils/PromoteMemToReg.h"
-#include "llvm/Transforms/Utils/Local.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/Instructions.h"
-#include "llvm/DerivedTypes.h"
-#include "llvm/Target/TargetData.h"
+#include "llvm/Transforms/Utils/PromoteMemToReg.h"
#include "llvm/Support/CFG.h"
-#include "Support/CommandLine.h"
-#include "Support/Debug.h"
-#include "Support/Statistic.h"
-#include "llvm/Assembly/Writer.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<> 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...
+ std::map<Loop *, AliasSetTracker *> LoopToAliasMap;
- /// visitLoop - Hoist expressions out of the specified loop...
+ /// 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 ETForest. This allows us to
+ /// visit uses before definitions, allowing us to sink a loop body in one
+ /// pass without iteration.
///
- void visitLoop(Loop *L, AliasSetTracker &AST);
+ 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.
///
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))
+ 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;
}
BasicBlock *LoopHeader = CurLoop->getHeader();
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_ it's immediate dominator.
+ // 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;
}
/// 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
+ bool pointerInvalidatedByLoop(Value *V, unsigned Size) {
+ // Check to see if any of the basic blocks in CurLoop invalidate *V.
+ return CurAST->getAliasSetForPointer(V, Size).isMod();
}
bool canSinkOrHoistInst(Instruction &I);
///
void PromoteValuesInLoop();
- /// findPromotableValuesInLoop - Check the current loop for stores to
+ /// 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...
///
- void findPromotableValuesInLoop(
+ void FindPromotableValuesInLoop(
std::vector<std::pair<AllocaInst*, Value*> > &PromotedValues,
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.
- 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);
- 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 (std::vector<Loop*>::const_iterator I = L->getSubLoops().begin(),
- E = L->getSubLoops().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
//
// 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.
+ // us to sink instructions in one pass, without iteration. AFter sinking
+ // instructions, we perform another pass to hoist them out of the loop.
//
- 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 ETForest. This allows us to visit
+/// uses before definitions, allowing us to sink a loop body in one pass without
+/// iteration.
+///
+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...
+ std::vector<BasicBlock*> Children;
+ ET->getChildren(BB, Children);
+ for (unsigned i = 0, e = Children.size(); i != e; ++i)
+ SinkRegion(Children[i]);
+
+ // Only need to process the contents of this block if it is not part of a
+ // subloop (which would already have been processed).
+ if (inSubLoop(BB)) return;
+
+ for (BasicBlock::iterator II = BB->end(); II != BB->begin(); ) {
+ Instruction &I = *--II;
+
+ // 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
+ // operands of the instruction are loop invariant.
+ //
+ if (isNotUsedInLoop(I) && canSinkOrHoistInst(I)) {
+ ++II;
+ sink(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;
for (BasicBlock::iterator II = BB->begin(), E = BB->end(); II != E; ) {
Instruction &I = *II++;
- // We can only handle simple expressions and loads with this code.
- if (canSinkOrHoistInst(I)) {
- // First check to see if we can sink this instruction to the exit blocks
- // of the loop. We can do this if the only users of the instruction are
- // outside of the loop. In this case, it doesn't even matter if the
- // operands of the instruction are loop invariant.
- //
- if (isNotUsedInLoop(I))
- sink(I);
-
- // If we can't sink the instruction, try hoisting it 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.
- //
- else if (isLoopInvariantInst(I) && isSafeToExecuteUnconditionally(I))
- hoist(I);
+ // 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) &&
+ isSafeToExecuteUnconditionally(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; // Don't hoist volatile loads!
// Don't hoist loads which have may-aliased stores in loop.
- return !pointerInvalidatedByLoop(LI->getOperand(0));
+ unsigned Size = 0;
+ if (LI->getType()->isSized())
+ Size = AA->getTargetData().getTypeSize(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;
+ }
+ }
+ if (!FoundMod) return true;
+ }
+ }
+
+ // FIXME: This should use mod/ref information to see if we can hoist or sink
+ // the call.
+
+ return false;
}
- return isa<BinaryOperator>(I) || isa<ShiftInst>(I) || isa<CastInst>(I) ||
- isa<GetElementPtrInst>(I) || isa<VANextInst>(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
/// exit blocks of the loop.
///
bool LICM::isNotUsedInLoop(Instruction &I) {
- for (Value::use_iterator UI = I.use_begin(), E = I.use_end(); UI != E; ++UI)
- if (CurLoop->contains(cast<Instruction>(*UI)->getParent()))
+ for (Value::use_iterator UI = I.use_begin(), E = I.use_end(); UI != E; ++UI) {
+ Instruction *User = cast<Instruction>(*UI);
+ if (PHINode *PN = dyn_cast<PHINode>(User)) {
+ // PHI node uses occur in predecessor blocks!
+ for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
+ if (PN->getIncomingValue(i) == &I)
+ if (CurLoop->contains(PN->getIncomingBlock(i)))
+ return false;
+ } else if (CurLoop->contains(User->getParent())) {
return false;
+ }
+ }
return true;
}
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 (!isLoopInvariant(I.getOperand(i)))
+ if (!CurLoop->isLoopInvariant(I.getOperand(i)))
return false;
// If we got this far, the instruction is loop invariant!
}
/// 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 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
+/// 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;
- const std::vector<BasicBlock*> &ExitBlocks = CurLoop->getExitBlocks();
- std::vector<Value*> Operands(I.op_begin(), I.op_end());
+ std::vector<BasicBlock*> ExitBlocks;
+ CurLoop->getExitBlocks(ExitBlocks);
if (isa<LoadInst>(I)) ++NumMovedLoads;
+ else if (isa<CallInst>(I)) ++NumMovedCalls;
++NumSunk;
Changed = true;
if (ExitBlocks.size() == 1) {
if (!isExitBlockDominatedByBlockInLoop(ExitBlocks[0], I.getParent())) {
// Instruction is not used, just delete it.
- I.getParent()->getInstList().erase(&I);
+ CurAST->deleteValue(&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;
ExitBlocks[0]->getInstList().insert(InsertPt, &I);
}
} else if (ExitBlocks.size() == 0) {
// The instruction is actually dead if there ARE NO exit blocks.
- I.getParent()->getInstList().erase(&I);
+ CurAST->deleteValue(&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
// the value into a stack object to get it to do this.
// Firstly, we create a stack object to hold the value...
- AllocaInst *AI = new AllocaInst(I.getType(), 0, I.getName(),
- I.getParent()->getParent()->front().begin());
+ AllocaInst *AI = 0;
+
+ if (I.getType() != Type::VoidTy)
+ AI = new AllocaInst(I.getType(), 0, I.getName(),
+ I.getParent()->getParent()->getEntryBlock().begin());
// Secondly, insert load instructions for each use of the instruction
// outside of the loop.
// Insert the code after the last PHI node...
BasicBlock::iterator InsertPt = ExitBlock->begin();
while (isa<PHINode>(InsertPt)) ++InsertPt;
-
+
// If this is the first exit block processed, just move the original
// instruction, otherwise clone the original instruction and insert
// the copy.
Instruction *New;
- if (InsertedBlocks.empty()) {
- I.getParent()->getInstList().remove(&I);
+ if (InsertedBlocks.size() == 1) {
+ I.removeFromParent();
ExitBlock->getInstList().insert(InsertPt, &I);
New = &I;
} else {
New = I.clone();
- New->setName(I.getName()+".le");
+ CurAST->copyValue(&I, New);
+ if (!I.getName().empty())
+ New->setName(I.getName()+".le");
ExitBlock->getInstList().insert(InsertPt, New);
}
-
+
// Now that we have inserted the instruction, store it into the alloca
- new StoreInst(New, AI, InsertPt);
+ if (AI) new StoreInst(New, AI, InsertPt);
}
}
}
-
+
+ // If the instruction doesn't dominate any exit blocks, it must be dead.
+ if (InsertedBlocks.empty()) {
+ CurAST->deleteValue(&I);
+ I.eraseFromParent();
+ }
+
// Finally, promote the fine value to SSA form.
- std::vector<AllocaInst*> Allocas;
- Allocas.push_back(AI);
- PromoteMemToReg(Allocas, *DT, *DF, AA->getTargetData());
+ if (AI) {
+ std::vector<AllocaInst*> Allocas;
+ Allocas.push_back(AI);
+ PromoteMemToReg(Allocas, *ET, *DF, AA->getTargetData(), CurAST);
+ }
}
-
- // Since we just sunk an instruction, check to see if any other instructions
- // used by this instruction are now sinkable. If so, sink them too.
- for (unsigned i = 0, e = Operands.size(); i != e; ++i)
- if (Instruction *OpI = dyn_cast<Instruction>(Operands[i]))
- if (CurLoop->contains(OpI->getParent()) && canSinkOrHoistInst(*OpI) &&
- isNotUsedInLoop(*OpI) && isSafeToExecuteUnconditionally(*OpI))
- sink(*OpI);
}
/// 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 &I) {
- DEBUG(std::cerr << "LICM hoisting to";
- WriteAsOperand(std::cerr, Preheader, false);
- std::cerr << ": " << 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);
-
+
if (isa<LoadInst>(I)) ++NumMovedLoads;
+ else if (isa<CallInst>(I)) ++NumMovedCalls;
++NumHoisted;
Changed = true;
}
bool LICM::isSafeToExecuteUnconditionally(Instruction &Inst) {
// If it is not a trapping instruction, it is always safe to hoist.
if (!Inst.isTrapping()) 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
// which does not execute this instruction, so we can't hoist it.
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.
- const std::vector<BasicBlock*> &ExitBlocks = CurLoop->getExitBlocks();
+ 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()))
return false;
-
+
return true;
}
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...
+ FindPromotableValuesInLoop(PromotedValues, ValueToAllocaMap);
+ if (ValueToAllocaMap.empty()) return; // If there are values to promote.
Changed = true;
NumPromoted += PromotedValues.size();
+ std::vector<Value*> PointerValueNumbers;
+
// 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...
+ Value *Ptr = PromotedValues[i].second;
+
+ // 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())) {
+ // 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.
+ for (Value::use_iterator UI = Ptr->use_begin(), E = Ptr->use_end();
+ UI != E; ++UI)
+ if (LoadInst *LI = dyn_cast<LoadInst>(*UI)) {
+ LoadValue = LI;
+ break;
+ } else if (StoreInst *SI = dyn_cast<StoreInst>(*UI)) {
+ if (SI->getOperand(1) == Ptr) {
+ LoadValue = SI->getOperand(0);
+ break;
+ }
+ }
+ assert(LoadValue && "No store through the pointer found!");
+ PointerValueNumbers.push_back(LoadValue); // Remember this for later.
+ }
+
+ // Load from the memory we are promoting.
+ LoadInst *LI = new LoadInst(Ptr, Ptr->getName()+".promoted", LoopPredInst);
+
+ if (LoadValue) CurAST->copyValue(LoadValue, LI);
+
+ // 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.
//
//
std::set<BasicBlock*> ProcessedBlocks;
- const std::vector<BasicBlock*> &ExitBlocks = CurLoop->getExitBlocks();
+ std::vector<BasicBlock*> 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...
+ // 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...
+ ++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...
+ // Load from the alloca.
LoadInst *LI = new LoadInst(PromotedValues[i].first, "", InsertPos);
- // Store into the memory we promoted...
+
+ // 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);
}
}
// 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...
+ // 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, *DT, *DF, AA->getTargetData());
+ PromoteMemToReg(PromotedAllocas, *ET, *DF, AA->getTargetData(), CurAST);
}
-/// findPromotableValuesInLoop - Check the current loop for stores to definite
+/// 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...
+/// and keep track of the mapping from value to alloca.
///
-void LICM::findPromotableValuesInLoop(
+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...
+ // 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.
+ // 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() &&
- isLoopInvariant(AS.begin()->first)) {
+ !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;
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));
-
+
+ // Update the AST and alias analysis.
+ CurAST->copyValue(V, AI);
+
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";
}
}
}