-//===- GVN.cpp - Eliminate redundant values and loads ------------===//
+//===- GVN.cpp - Eliminate redundant values and loads ---------------------===//
//
// The LLVM Compiler Infrastructure
//
// This pass performs global value numbering to eliminate fully redundant
// instructions. It also performs simple dead load elimination.
//
-// Note that this pass does the value numbering itself, it does not use the
+// Note that this pass does the value numbering itself; it does not use the
// ValueNumbering analysis passes.
//
//===----------------------------------------------------------------------===//
#include "llvm/Constants.h"
#include "llvm/DerivedTypes.h"
#include "llvm/Function.h"
-#include "llvm/Instructions.h"
+#include "llvm/IntrinsicInst.h"
+#include "llvm/LLVMContext.h"
#include "llvm/Value.h"
#include "llvm/ADT/DenseMap.h"
#include "llvm/ADT/DepthFirstIterator.h"
+#include "llvm/ADT/PostOrderIterator.h"
#include "llvm/ADT/SmallPtrSet.h"
#include "llvm/ADT/SmallVector.h"
#include "llvm/ADT/Statistic.h"
#include "llvm/Support/CommandLine.h"
#include "llvm/Support/Compiler.h"
#include "llvm/Support/Debug.h"
+#include "llvm/Support/ErrorHandling.h"
#include "llvm/Transforms/Utils/BasicBlockUtils.h"
+#include "llvm/Transforms/Utils/Local.h"
#include <cstdio>
using namespace llvm;
-STATISTIC(NumGVNInstr, "Number of instructions deleted");
-STATISTIC(NumGVNLoad, "Number of loads deleted");
-STATISTIC(NumGVNPRE, "Number of instructions PRE'd");
+STATISTIC(NumGVNInstr, "Number of instructions deleted");
+STATISTIC(NumGVNLoad, "Number of loads deleted");
+STATISTIC(NumGVNPRE, "Number of instructions PRE'd");
STATISTIC(NumGVNBlocks, "Number of blocks merged");
-STATISTIC(NumPRELoad, "Number of loads PRE'd");
+STATISTIC(NumPRELoad, "Number of loads PRE'd");
static cl::opt<bool> EnablePRE("enable-pre",
cl::init(true), cl::Hidden);
-cl::opt<bool> EnableLoadPRE("enable-load-pre");
+static cl::opt<bool> EnableLoadPRE("enable-load-pre", cl::init(true));
//===----------------------------------------------------------------------===//
// ValueTable Class
/// two values.
namespace {
struct VISIBILITY_HIDDEN Expression {
- enum ExpressionOpcode { ADD, SUB, MUL, UDIV, SDIV, FDIV, UREM, SREM,
+ enum ExpressionOpcode { ADD, FADD, SUB, FSUB, MUL, FMUL,
+ UDIV, SDIV, FDIV, UREM, SREM,
FREM, SHL, LSHR, ASHR, AND, OR, XOR, ICMPEQ,
ICMPNE, ICMPUGT, ICMPUGE, ICMPULT, ICMPULE,
ICMPSGT, ICMPSGE, ICMPSLT, ICMPSLE, FCMPOEQ,
}
bool operator!=(const Expression &other) const {
- if (opcode != other.opcode)
- return true;
- else if (opcode == EMPTY || opcode == TOMBSTONE)
- return false;
- else if (type != other.type)
- return true;
- else if (function != other.function)
- return true;
- else if (firstVN != other.firstVN)
- return true;
- else if (secondVN != other.secondVN)
- return true;
- else if (thirdVN != other.thirdVN)
- return true;
- else {
- if (varargs.size() != other.varargs.size())
- return true;
-
- for (size_t i = 0; i < varargs.size(); ++i)
- if (varargs[i] != other.varargs[i])
- return true;
-
- return false;
- }
+ return !(*this == other);
}
};
void setMemDep(MemoryDependenceAnalysis* M) { MD = M; }
void setDomTree(DominatorTree* D) { DT = D; }
uint32_t getNextUnusedValueNumber() { return nextValueNumber; }
+ void verifyRemoved(const Value *) const;
};
}
Expression::ExpressionOpcode ValueTable::getOpcode(BinaryOperator* BO) {
switch(BO->getOpcode()) {
default: // THIS SHOULD NEVER HAPPEN
- assert(0 && "Binary operator with unknown opcode?");
+ llvm_unreachable("Binary operator with unknown opcode?");
case Instruction::Add: return Expression::ADD;
+ case Instruction::FAdd: return Expression::FADD;
case Instruction::Sub: return Expression::SUB;
+ case Instruction::FSub: return Expression::FSUB;
case Instruction::Mul: return Expression::MUL;
+ case Instruction::FMul: return Expression::FMUL;
case Instruction::UDiv: return Expression::UDIV;
case Instruction::SDiv: return Expression::SDIV;
case Instruction::FDiv: return Expression::FDIV;
}
Expression::ExpressionOpcode ValueTable::getOpcode(CmpInst* C) {
- if (isa<ICmpInst>(C) || isa<VICmpInst>(C)) {
+ if (isa<ICmpInst>(C)) {
switch (C->getPredicate()) {
default: // THIS SHOULD NEVER HAPPEN
- assert(0 && "Comparison with unknown predicate?");
+ llvm_unreachable("Comparison with unknown predicate?");
case ICmpInst::ICMP_EQ: return Expression::ICMPEQ;
case ICmpInst::ICMP_NE: return Expression::ICMPNE;
case ICmpInst::ICMP_UGT: return Expression::ICMPUGT;
case ICmpInst::ICMP_SLT: return Expression::ICMPSLT;
case ICmpInst::ICMP_SLE: return Expression::ICMPSLE;
}
- }
- assert((isa<FCmpInst>(C) || isa<VFCmpInst>(C)) && "Unknown compare");
- switch (C->getPredicate()) {
- default: // THIS SHOULD NEVER HAPPEN
- assert(0 && "Comparison with unknown predicate?");
- case FCmpInst::FCMP_OEQ: return Expression::FCMPOEQ;
- case FCmpInst::FCMP_OGT: return Expression::FCMPOGT;
- case FCmpInst::FCMP_OGE: return Expression::FCMPOGE;
- case FCmpInst::FCMP_OLT: return Expression::FCMPOLT;
- case FCmpInst::FCMP_OLE: return Expression::FCMPOLE;
- case FCmpInst::FCMP_ONE: return Expression::FCMPONE;
- case FCmpInst::FCMP_ORD: return Expression::FCMPORD;
- case FCmpInst::FCMP_UNO: return Expression::FCMPUNO;
- case FCmpInst::FCMP_UEQ: return Expression::FCMPUEQ;
- case FCmpInst::FCMP_UGT: return Expression::FCMPUGT;
- case FCmpInst::FCMP_UGE: return Expression::FCMPUGE;
- case FCmpInst::FCMP_ULT: return Expression::FCMPULT;
- case FCmpInst::FCMP_ULE: return Expression::FCMPULE;
- case FCmpInst::FCMP_UNE: return Expression::FCMPUNE;
+ } else {
+ switch (C->getPredicate()) {
+ default: // THIS SHOULD NEVER HAPPEN
+ llvm_unreachable("Comparison with unknown predicate?");
+ case FCmpInst::FCMP_OEQ: return Expression::FCMPOEQ;
+ case FCmpInst::FCMP_OGT: return Expression::FCMPOGT;
+ case FCmpInst::FCMP_OGE: return Expression::FCMPOGE;
+ case FCmpInst::FCMP_OLT: return Expression::FCMPOLT;
+ case FCmpInst::FCMP_OLE: return Expression::FCMPOLE;
+ case FCmpInst::FCMP_ONE: return Expression::FCMPONE;
+ case FCmpInst::FCMP_ORD: return Expression::FCMPORD;
+ case FCmpInst::FCMP_UNO: return Expression::FCMPUNO;
+ case FCmpInst::FCMP_UEQ: return Expression::FCMPUEQ;
+ case FCmpInst::FCMP_UGT: return Expression::FCMPUGT;
+ case FCmpInst::FCMP_UGE: return Expression::FCMPUGE;
+ case FCmpInst::FCMP_ULT: return Expression::FCMPULT;
+ case FCmpInst::FCMP_ULE: return Expression::FCMPULE;
+ case FCmpInst::FCMP_UNE: return Expression::FCMPUNE;
+ }
}
}
Expression::ExpressionOpcode ValueTable::getOpcode(CastInst* C) {
switch(C->getOpcode()) {
default: // THIS SHOULD NEVER HAPPEN
- assert(0 && "Cast operator with unknown opcode?");
+ llvm_unreachable("Cast operator with unknown opcode?");
case Instruction::Trunc: return Expression::TRUNC;
case Instruction::ZExt: return Expression::ZEXT;
case Instruction::SExt: return Expression::SEXT;
MemDepResult local_dep = MD->getDependency(C);
- if (local_dep.isNone()) {
+ if (!local_dep.isDef() && !local_dep.isNonLocal()) {
valueNumbering.insert(std::make_pair(V, nextValueNumber));
return nextValueNumber++;
}
-
- if (Instruction *LocalDepInst = local_dep.getInst()) {
- if (!isa<CallInst>(LocalDepInst)) {
- valueNumbering.insert(std::make_pair(V, nextValueNumber));
- return nextValueNumber++;
- }
-
- CallInst* local_cdep = cast<CallInst>(LocalDepInst);
-
- if (local_cdep->getCalledFunction() != C->getCalledFunction() ||
- local_cdep->getNumOperands() != C->getNumOperands()) {
- valueNumbering.insert(std::make_pair(V, nextValueNumber));
- return nextValueNumber++;
- }
+
+ if (local_dep.isDef()) {
+ CallInst* local_cdep = cast<CallInst>(local_dep.getInst());
- if (!C->getCalledFunction()) {
+ if (local_cdep->getNumOperands() != C->getNumOperands()) {
valueNumbering.insert(std::make_pair(V, nextValueNumber));
return nextValueNumber++;
}
-
+
for (unsigned i = 1; i < C->getNumOperands(); ++i) {
uint32_t c_vn = lookup_or_add(C->getOperand(i));
uint32_t cd_vn = lookup_or_add(local_cdep->getOperand(i));
valueNumbering.insert(std::make_pair(V, v));
return v;
}
-
+ // Non-local case.
const MemoryDependenceAnalysis::NonLocalDepInfo &deps =
- MD->getNonLocalDependency(C);
+ MD->getNonLocalCallDependency(CallSite(C));
+ // FIXME: call/call dependencies for readonly calls should return def, not
+ // clobber! Move the checking logic to MemDep!
CallInst* cdep = 0;
// Check to see if we have a single dominating call instruction that is
// We don't handle non-depedencies. If we already have a call, reject
// instruction dependencies.
- if (I->second.isNone() || cdep != 0) {
+ if (I->second.isClobber() || cdep != 0) {
cdep = 0;
break;
}
return nextValueNumber++;
}
- if (cdep->getCalledFunction() != C->getCalledFunction() ||
- cdep->getNumOperands() != C->getNumOperands()) {
- valueNumbering.insert(std::make_pair(V, nextValueNumber));
- return nextValueNumber++;
- }
- if (!C->getCalledFunction()) {
+ if (cdep->getNumOperands() != C->getNumOperands()) {
valueNumbering.insert(std::make_pair(V, nextValueNumber));
return nextValueNumber++;
}
valueNumbering.erase(V);
}
+/// verifyRemoved - Verify that the value is removed from all internal data
+/// structures.
+void ValueTable::verifyRemoved(const Value *V) const {
+ for (DenseMap<Value*, uint32_t>::iterator
+ I = valueNumbering.begin(), E = valueNumbering.end(); I != E; ++I) {
+ assert(I->first != V && "Inst still occurs in value numbering map!");
+ }
+}
+
//===----------------------------------------------------------------------===//
-// GVN Pass
+// GVN Pass
//===----------------------------------------------------------------------===//
namespace {
// Helper fuctions
// FIXME: eliminate or document these better
bool processLoad(LoadInst* L,
- DenseMap<Value*, LoadInst*> &lastLoad,
SmallVectorImpl<Instruction*> &toErase);
bool processInstruction(Instruction* I,
- DenseMap<Value*, LoadInst*>& lastSeenLoad,
SmallVectorImpl<Instruction*> &toErase);
bool processNonLocalLoad(LoadInst* L,
SmallVectorImpl<Instruction*> &toErase);
- bool processBlock(DomTreeNode* DTN);
- Value *GetValueForBlock(BasicBlock *BB, LoadInst* orig,
+ bool processBlock(BasicBlock* BB);
+ Value *GetValueForBlock(BasicBlock *BB, Instruction* orig,
DenseMap<BasicBlock*, Value*> &Phis,
bool top_level = false);
void dump(DenseMap<uint32_t, Value*>& d);
bool performPRE(Function& F);
Value* lookupNumber(BasicBlock* BB, uint32_t num);
bool mergeBlockIntoPredecessor(BasicBlock* BB);
+ Value* AttemptRedundancyElimination(Instruction* orig, unsigned valno);
void cleanupGlobalSets();
+ void verifyRemoved(const Instruction *I) const;
};
char GVN::ID = 0;
/// GetValueForBlock - Get the value to use within the specified basic block.
/// available values are in Phis.
-Value *GVN::GetValueForBlock(BasicBlock *BB, LoadInst* orig,
+Value *GVN::GetValueForBlock(BasicBlock *BB, Instruction* orig,
DenseMap<BasicBlock*, Value*> &Phis,
bool top_level) {
// If the block is unreachable, just return undef, since this path
// can't actually occur at runtime.
if (!DT->isReachableFromEntry(BB))
- return Phis[BB] = UndefValue::get(orig->getType());
+ return Phis[BB] = Context->getUndef(orig->getType());
- BasicBlock* singlePred = BB->getSinglePredecessor();
- if (singlePred) {
- Value *ret = GetValueForBlock(singlePred, orig, Phis);
+ if (BasicBlock *Pred = BB->getSinglePredecessor()) {
+ Value *ret = GetValueForBlock(Pred, orig, Phis);
Phis[BB] = ret;
return ret;
}
+
+ // Get the number of predecessors of this block so we can reserve space later.
+ // If there is already a PHI in it, use the #preds from it, otherwise count.
+ // Getting it from the PHI is constant time.
+ unsigned NumPreds;
+ if (PHINode *ExistingPN = dyn_cast<PHINode>(BB->begin()))
+ NumPreds = ExistingPN->getNumIncomingValues();
+ else
+ NumPreds = std::distance(pred_begin(BB), pred_end(BB));
// Otherwise, the idom is the loop, so we need to insert a PHI node. Do so
// now, then get values to fill in the incoming values for the PHI.
PHINode *PN = PHINode::Create(orig->getType(), orig->getName()+".rle",
BB->begin());
- PN->reserveOperandSpace(std::distance(pred_begin(BB), pred_end(BB)));
+ PN->reserveOperandSpace(NumPreds);
- if (Phis.count(BB) == 0)
- Phis.insert(std::make_pair(BB, PN));
+ Phis.insert(std::make_pair(BB, PN));
// Fill in the incoming values for the block.
for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
Value* v = CollapsePhi(PN);
if (!v) {
// Cache our phi construction results
- phiMap[orig->getPointerOperand()].insert(PN);
+ if (LoadInst* L = dyn_cast<LoadInst>(orig))
+ phiMap[L->getPointerOperand()].insert(PN);
+ else
+ phiMap[orig].insert(PN);
+
return PN;
}
PN->replaceAllUsesWith(v);
+ if (isa<PointerType>(v->getType()))
+ MD->invalidateCachedPointerInfo(v);
for (DenseMap<BasicBlock*, Value*>::iterator I = Phis.begin(),
E = Phis.end(); I != E; ++I)
DEBUG(cerr << "GVN removed: " << *PN);
MD->removeInstruction(PN);
PN->eraseFromParent();
+ DEBUG(verifyRemoved(PN));
Phis[BB] = v;
return v;
/// IsValueFullyAvailableInBlock - Return true if we can prove that the value
/// we're analyzing is fully available in the specified block. As we go, keep
-/// track of which blocks we know it is fully alive or not in
-/// FullyAvailableBlocks.
+/// track of which blocks we know are fully alive in FullyAvailableBlocks. This
+/// map is actually a tri-state map with the following values:
+/// 0) we know the block *is not* fully available.
+/// 1) we know the block *is* fully available.
+/// 2) we do not know whether the block is fully available or not, but we are
+/// currently speculating that it will be.
+/// 3) we are speculating for this block and have used that to speculate for
+/// other blocks.
static bool IsValueFullyAvailableInBlock(BasicBlock *BB,
- DenseMap<BasicBlock*, bool> &FullyAvailableBlocks) {
+ DenseMap<BasicBlock*, char> &FullyAvailableBlocks) {
// Optimistically assume that the block is fully available and check to see
// if we already know about this block in one lookup.
- std::pair<DenseMap<BasicBlock*, bool>::iterator, bool> IV =
- FullyAvailableBlocks.insert(std::make_pair(BB, true));
+ std::pair<DenseMap<BasicBlock*, char>::iterator, char> IV =
+ FullyAvailableBlocks.insert(std::make_pair(BB, 2));
// If the entry already existed for this block, return the precomputed value.
- if (!IV.second)
- return IV.first->second;
+ if (!IV.second) {
+ // If this is a speculative "available" value, mark it as being used for
+ // speculation of other blocks.
+ if (IV.first->second == 2)
+ IV.first->second = 3;
+ return IV.first->second != 0;
+ }
// Otherwise, see if it is fully available in all predecessors.
pred_iterator PI = pred_begin(BB), PE = pred_end(BB);
// If this block has no predecessors, it isn't live-in here.
if (PI == PE)
- return FullyAvailableBlocks[BB] = false;
+ goto SpeculationFailure;
for (; PI != PE; ++PI)
// If the value isn't fully available in one of our predecessors, then it
// isn't fully available in this block either. Undo our previous
// optimistic assumption and bail out.
if (!IsValueFullyAvailableInBlock(*PI, FullyAvailableBlocks))
- return FullyAvailableBlocks[BB] = false;
+ goto SpeculationFailure;
return true;
+
+// SpeculationFailure - If we get here, we found out that this is not, after
+// all, a fully-available block. We have a problem if we speculated on this and
+// used the speculation to mark other blocks as available.
+SpeculationFailure:
+ char &BBVal = FullyAvailableBlocks[BB];
+
+ // If we didn't speculate on this, just return with it set to false.
+ if (BBVal == 2) {
+ BBVal = 0;
+ return false;
+ }
+
+ // If we did speculate on this value, we could have blocks set to 1 that are
+ // incorrect. Walk the (transitive) successors of this block and mark them as
+ // 0 if set to one.
+ SmallVector<BasicBlock*, 32> BBWorklist;
+ BBWorklist.push_back(BB);
+
+ while (!BBWorklist.empty()) {
+ BasicBlock *Entry = BBWorklist.pop_back_val();
+ // Note that this sets blocks to 0 (unavailable) if they happen to not
+ // already be in FullyAvailableBlocks. This is safe.
+ char &EntryVal = FullyAvailableBlocks[Entry];
+ if (EntryVal == 0) continue; // Already unavailable.
+
+ // Mark as unavailable.
+ EntryVal = 0;
+
+ for (succ_iterator I = succ_begin(Entry), E = succ_end(Entry); I != E; ++I)
+ BBWorklist.push_back(*I);
+ }
+
+ return false;
}
/// processNonLocalLoad - Attempt to eliminate a load whose dependencies are
bool GVN::processNonLocalLoad(LoadInst *LI,
SmallVectorImpl<Instruction*> &toErase) {
// Find the non-local dependencies of the load.
- const MemoryDependenceAnalysis::NonLocalDepInfo &deps =
- MD->getNonLocalDependency(LI);
- //DEBUG(cerr << "INVESTIGATING NONLOCAL LOAD: " << deps.size() << *LI);
+ SmallVector<MemoryDependenceAnalysis::NonLocalDepEntry, 64> Deps;
+ MD->getNonLocalPointerDependency(LI->getOperand(0), true, LI->getParent(),
+ Deps);
+ //DEBUG(cerr << "INVESTIGATING NONLOCAL LOAD: " << Deps.size() << *LI);
// If we had to process more than one hundred blocks to find the
// dependencies, this load isn't worth worrying about. Optimizing
// it will be too expensive.
- if (deps.size() > 100)
+ if (Deps.size() > 100)
return false;
-
- BasicBlock *EntryBlock = &LI->getParent()->getParent()->getEntryBlock();
+
+ // If we had a phi translation failure, we'll have a single entry which is a
+ // clobber in the current block. Reject this early.
+ if (Deps.size() == 1 && Deps[0].second.isClobber()) {
+ DEBUG(
+ DOUT << "GVN: non-local load ";
+ WriteAsOperand(*DOUT.stream(), LI);
+ DOUT << " is clobbered by " << *Deps[0].second.getInst();
+ );
+ return false;
+ }
// Filter out useless results (non-locals, etc). Keep track of the blocks
// where we have a value available in repl, also keep track of whether we see
SmallVector<std::pair<BasicBlock*, Value*>, 16> ValuesPerBlock;
SmallVector<BasicBlock*, 16> UnavailableBlocks;
- for (unsigned i = 0, e = deps.size(); i != e; ++i) {
- BasicBlock *DepBB = deps[i].first;
- MemDepResult DepInfo = deps[i].second;
+ for (unsigned i = 0, e = Deps.size(); i != e; ++i) {
+ BasicBlock *DepBB = Deps[i].first;
+ MemDepResult DepInfo = Deps[i].second;
- if (DepInfo.isNonLocal()) {
- // If this is a non-local dependency in the entry block, then we depend on
- // the value live-in at the start of the function. We could insert a load
- // in the entry block to get this, but for now we'll just bail out.
- if (DepBB == EntryBlock)
- UnavailableBlocks.push_back(DepBB);
+ if (DepInfo.isClobber()) {
+ UnavailableBlocks.push_back(DepBB);
continue;
}
- if (DepInfo.isNone()) {
+ Instruction *DepInst = DepInfo.getInst();
+
+ // Loading the allocation -> undef.
+ if (isa<AllocationInst>(DepInst)) {
ValuesPerBlock.push_back(std::make_pair(DepBB,
- UndefValue::get(LI->getType())));
+ Context->getUndef(LI->getType())));
continue;
}
- if (StoreInst* S = dyn_cast<StoreInst>(DepInfo.getInst())) {
+ if (StoreInst* S = dyn_cast<StoreInst>(DepInst)) {
// Reject loads and stores that are to the same address but are of
// different types.
// NOTE: 403.gcc does have this case (e.g. in readonly_fields_p) because
continue;
}
- if (S->getPointerOperand() != LI->getPointerOperand() &&
- VN.getAliasAnalysis()->alias(S->getPointerOperand(), 1,
- LI->getPointerOperand(), 1)
- != AliasAnalysis::MustAlias) {
- UnavailableBlocks.push_back(DepBB);
- continue;
- }
ValuesPerBlock.push_back(std::make_pair(DepBB, S->getOperand(0)));
- } else if (LoadInst* LD = dyn_cast<LoadInst>(DepInfo.getInst())) {
+ } else if (LoadInst* LD = dyn_cast<LoadInst>(DepInst)) {
if (LD->getType() != LI->getType()) {
UnavailableBlocks.push_back(DepBB);
continue;
}
-
- if (LD->getPointerOperand() != LI->getPointerOperand() &&
- VN.getAliasAnalysis()->alias(LD->getPointerOperand(), 1,
- LI->getPointerOperand(), 1)
- != AliasAnalysis::MustAlias) {
- UnavailableBlocks.push_back(DepBB);
- continue;
- }
ValuesPerBlock.push_back(std::make_pair(DepBB, LD));
} else {
UnavailableBlocks.push_back(DepBB);
if (UnavailableBlocks.empty()) {
// Use cached PHI construction information from previous runs
SmallPtrSet<Instruction*, 4> &p = phiMap[LI->getPointerOperand()];
+ // FIXME: What does phiMap do? Are we positive it isn't getting invalidated?
for (SmallPtrSet<Instruction*, 4>::iterator I = p.begin(), E = p.end();
I != E; ++I) {
if ((*I)->getParent() == LI->getParent()) {
DEBUG(cerr << "GVN REMOVING NONLOCAL LOAD #1: " << *LI);
LI->replaceAllUsesWith(*I);
+ if (isa<PointerType>((*I)->getType()))
+ MD->invalidateCachedPointerInfo(*I);
toErase.push_back(LI);
NumGVNLoad++;
return true;
// Perform PHI construction.
Value* v = GetValueForBlock(LI->getParent(), LI, BlockReplValues, true);
LI->replaceAllUsesWith(v);
+
+ if (isa<PHINode>(v))
+ v->takeName(LI);
+ if (isa<PointerType>(v->getType()))
+ MD->invalidateCachedPointerInfo(v);
toErase.push_back(LI);
NumGVNLoad++;
return true;
// that we only have to insert *one* load (which means we're basically moving
// the load, not inserting a new one).
- // Everything we do here is based on local predecessors of LI's block. If it
- // only has one predecessor, bail now.
+ SmallPtrSet<BasicBlock *, 4> Blockers;
+ for (unsigned i = 0, e = UnavailableBlocks.size(); i != e; ++i)
+ Blockers.insert(UnavailableBlocks[i]);
+
+ // Lets find first basic block with more than one predecessor. Walk backwards
+ // through predecessors if needed.
BasicBlock *LoadBB = LI->getParent();
- if (LoadBB->getSinglePredecessor())
- return false;
+ BasicBlock *TmpBB = LoadBB;
+
+ bool isSinglePred = false;
+ bool allSingleSucc = true;
+ while (TmpBB->getSinglePredecessor()) {
+ isSinglePred = true;
+ TmpBB = TmpBB->getSinglePredecessor();
+ if (!TmpBB) // If haven't found any, bail now.
+ return false;
+ if (TmpBB == LoadBB) // Infinite (unreachable) loop.
+ return false;
+ if (Blockers.count(TmpBB))
+ return false;
+ if (TmpBB->getTerminator()->getNumSuccessors() != 1)
+ allSingleSucc = false;
+ }
+
+ assert(TmpBB);
+ LoadBB = TmpBB;
// If we have a repl set with LI itself in it, this means we have a loop where
// at least one of the values is LI. Since this means that we won't be able
if (ValuesPerBlock[i].second == LI)
return false;
+ if (isSinglePred) {
+ bool isHot = false;
+ for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i)
+ if (Instruction *I = dyn_cast<Instruction>(ValuesPerBlock[i].second))
+ // "Hot" Instruction is in some loop (because it dominates its dep.
+ // instruction).
+ if (DT->dominates(LI, I)) {
+ isHot = true;
+ break;
+ }
+
+ // We are interested only in "hot" instructions. We don't want to do any
+ // mis-optimizations here.
+ if (!isHot)
+ return false;
+ }
+
// Okay, we have some hope :). Check to see if the loaded value is fully
// available in all but one predecessor.
// FIXME: If we could restructure the CFG, we could make a common pred with
// that one block.
BasicBlock *UnavailablePred = 0;
- DenseMap<BasicBlock*, bool> FullyAvailableBlocks;
+ DenseMap<BasicBlock*, char> FullyAvailableBlocks;
for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i)
FullyAvailableBlocks[ValuesPerBlock[i].first] = true;
for (unsigned i = 0, e = UnavailableBlocks.size(); i != e; ++i)
return false;
}
+ // Make sure it is valid to move this load here. We have to watch out for:
+ // @1 = getelementptr (i8* p, ...
+ // test p and branch if == 0
+ // load @1
+ // It is valid to have the getelementptr before the test, even if p can be 0,
+ // as getelementptr only does address arithmetic.
+ // If we are not pushing the value through any multiple-successor blocks
+ // we do not have this case. Otherwise, check that the load is safe to
+ // put anywhere; this can be improved, but should be conservatively safe.
+ if (!allSingleSucc &&
+ !isSafeToLoadUnconditionally(LoadPtr, UnavailablePred->getTerminator()))
+ return false;
+
// Okay, we can eliminate this load by inserting a reload in the predecessor
// and using PHI construction to get the value in the other predecessors, do
// it.
LI->getAlignment(),
UnavailablePred->getTerminator());
+ SmallPtrSet<Instruction*, 4> &p = phiMap[LI->getPointerOperand()];
+ for (SmallPtrSet<Instruction*, 4>::iterator I = p.begin(), E = p.end();
+ I != E; ++I)
+ ValuesPerBlock.push_back(std::make_pair((*I)->getParent(), *I));
+
DenseMap<BasicBlock*, Value*> BlockReplValues;
BlockReplValues.insert(ValuesPerBlock.begin(), ValuesPerBlock.end());
BlockReplValues[UnavailablePred] = NewLoad;
// Perform PHI construction.
Value* v = GetValueForBlock(LI->getParent(), LI, BlockReplValues, true);
LI->replaceAllUsesWith(v);
+ if (isa<PHINode>(v))
+ v->takeName(LI);
+ if (isa<PointerType>(v->getType()))
+ MD->invalidateCachedPointerInfo(v);
toErase.push_back(LI);
NumPRELoad++;
return true;
/// processLoad - Attempt to eliminate a load, first by eliminating it
/// locally, and then attempting non-local elimination if that fails.
-bool GVN::processLoad(LoadInst *L, DenseMap<Value*, LoadInst*> &lastLoad,
- SmallVectorImpl<Instruction*> &toErase) {
- if (L->isVolatile()) {
- lastLoad[L->getPointerOperand()] = L;
+bool GVN::processLoad(LoadInst *L, SmallVectorImpl<Instruction*> &toErase) {
+ if (L->isVolatile())
return false;
- }
Value* pointer = L->getPointerOperand();
- LoadInst*& last = lastLoad[pointer];
-
+
// ... to a pointer that has been loaded from before...
- bool removedNonLocal = false;
MemDepResult dep = MD->getDependency(L);
- if (dep.isNonLocal() &&
- L->getParent() != &L->getParent()->getParent()->getEntryBlock()) {
- removedNonLocal = processNonLocalLoad(L, toErase);
+
+ // If the value isn't available, don't do anything!
+ if (dep.isClobber()) {
+ DEBUG(
+ // fast print dep, using operator<< on instruction would be too slow
+ DOUT << "GVN: load ";
+ WriteAsOperand(*DOUT.stream(), L);
+ Instruction *I = dep.getInst();
+ DOUT << " is clobbered by " << *I;
+ );
+ return false;
+ }
+
+ // If it is defined in another block, try harder.
+ if (dep.isNonLocal())
+ return processNonLocalLoad(L, toErase);
+
+ Instruction *DepInst = dep.getInst();
+ if (StoreInst *DepSI = dyn_cast<StoreInst>(DepInst)) {
+ // Only forward substitute stores to loads of the same type.
+ // FIXME: Could do better!
+ if (DepSI->getPointerOperand()->getType() != pointer->getType())
+ return false;
- if (!removedNonLocal)
- last = L;
+ // Remove it!
+ L->replaceAllUsesWith(DepSI->getOperand(0));
+ if (isa<PointerType>(DepSI->getOperand(0)->getType()))
+ MD->invalidateCachedPointerInfo(DepSI->getOperand(0));
+ toErase.push_back(L);
+ NumGVNLoad++;
+ return true;
+ }
+
+ if (LoadInst *DepLI = dyn_cast<LoadInst>(DepInst)) {
+ // Only forward substitute stores to loads of the same type.
+ // FIXME: Could do better! load i32 -> load i8 -> truncate on little endian.
+ if (DepLI->getType() != L->getType())
+ return false;
- return removedNonLocal;
+ // Remove it!
+ L->replaceAllUsesWith(DepLI);
+ if (isa<PointerType>(DepLI->getType()))
+ MD->invalidateCachedPointerInfo(DepLI);
+ toErase.push_back(L);
+ NumGVNLoad++;
+ return true;
}
-
- bool deletedLoad = false;
-
- // Walk up the dependency chain until we either find
- // a dependency we can use, or we can't walk any further
- while (Instruction *DepInst = dep.getInst()) {
- // ... that depends on a store ...
- if (StoreInst* S = dyn_cast<StoreInst>(DepInst)) {
- if (S->getPointerOperand() == pointer) {
- // Remove it!
- L->replaceAllUsesWith(S->getOperand(0));
- toErase.push_back(L);
- deletedLoad = true;
- NumGVNLoad++;
- }
-
- // Whether we removed it or not, we can't
- // go any further
- break;
- } else if (!isa<LoadInst>(DepInst)) {
- // Only want to handle loads below.
- break;
- } else if (!last) {
- // If we don't depend on a store, and we haven't
- // been loaded before, bail.
- break;
- } else if (DepInst == last) {
- // Remove it!
- L->replaceAllUsesWith(last);
- toErase.push_back(L);
- deletedLoad = true;
- NumGVNLoad++;
- break;
- } else {
- dep = MD->getDependencyFrom(L, DepInst, DepInst->getParent());
- }
- }
-
// If this load really doesn't depend on anything, then we must be loading an
// undef value. This can happen when loading for a fresh allocation with no
// intervening stores, for example.
- if (dep.isNone()) {
- // If this load depends directly on an allocation, there isn't
- // anything stored there; therefore, we can optimize this load
- // to undef.
- L->replaceAllUsesWith(UndefValue::get(L->getType()));
+ if (isa<AllocationInst>(DepInst)) {
+ L->replaceAllUsesWith(Context->getUndef(L->getType()));
toErase.push_back(L);
- deletedLoad = true;
NumGVNLoad++;
+ return true;
}
- if (!deletedLoad)
- last = L;
-
- return deletedLoad;
+ return false;
}
Value* GVN::lookupNumber(BasicBlock* BB, uint32_t num) {
return 0;
}
+/// AttemptRedundancyElimination - If the "fast path" of redundancy elimination
+/// by inheritance from the dominator fails, see if we can perform phi
+/// construction to eliminate the redundancy.
+Value* GVN::AttemptRedundancyElimination(Instruction* orig, unsigned valno) {
+ BasicBlock* BaseBlock = orig->getParent();
+
+ SmallPtrSet<BasicBlock*, 4> Visited;
+ SmallVector<BasicBlock*, 8> Stack;
+ Stack.push_back(BaseBlock);
+
+ DenseMap<BasicBlock*, Value*> Results;
+
+ // Walk backwards through our predecessors, looking for instances of the
+ // value number we're looking for. Instances are recorded in the Results
+ // map, which is then used to perform phi construction.
+ while (!Stack.empty()) {
+ BasicBlock* Current = Stack.back();
+ Stack.pop_back();
+
+ // If we've walked all the way to a proper dominator, then give up. Cases
+ // where the instance is in the dominator will have been caught by the fast
+ // path, and any cases that require phi construction further than this are
+ // probably not worth it anyways. Note that this is a SIGNIFICANT compile
+ // time improvement.
+ if (DT->properlyDominates(Current, orig->getParent())) return 0;
+
+ DenseMap<BasicBlock*, ValueNumberScope*>::iterator LA =
+ localAvail.find(Current);
+ if (LA == localAvail.end()) return 0;
+ DenseMap<uint32_t, Value*>::iterator V = LA->second->table.find(valno);
+
+ if (V != LA->second->table.end()) {
+ // Found an instance, record it.
+ Results.insert(std::make_pair(Current, V->second));
+ continue;
+ }
+
+ // If we reach the beginning of the function, then give up.
+ if (pred_begin(Current) == pred_end(Current))
+ return 0;
+
+ for (pred_iterator PI = pred_begin(Current), PE = pred_end(Current);
+ PI != PE; ++PI)
+ if (Visited.insert(*PI))
+ Stack.push_back(*PI);
+ }
+
+ // If we didn't find instances, give up. Otherwise, perform phi construction.
+ if (Results.size() == 0)
+ return 0;
+ else
+ return GetValueForBlock(BaseBlock, orig, Results, true);
+}
+
/// processInstruction - When calculating availability, handle an instruction
/// by inserting it into the appropriate sets
bool GVN::processInstruction(Instruction *I,
- DenseMap<Value*, LoadInst*> &lastSeenLoad,
SmallVectorImpl<Instruction*> &toErase) {
if (LoadInst* L = dyn_cast<LoadInst>(I)) {
- bool changed = processLoad(L, lastSeenLoad, toErase);
+ bool changed = processLoad(L, toErase);
if (!changed) {
unsigned num = VN.lookup_or_add(L);
uint32_t nextNum = VN.getNextUnusedValueNumber();
unsigned num = VN.lookup_or_add(I);
+ if (BranchInst* BI = dyn_cast<BranchInst>(I)) {
+ localAvail[I->getParent()]->table.insert(std::make_pair(num, I));
+
+ if (!BI->isConditional() || isa<Constant>(BI->getCondition()))
+ return false;
+
+ Value* branchCond = BI->getCondition();
+ uint32_t condVN = VN.lookup_or_add(branchCond);
+
+ BasicBlock* trueSucc = BI->getSuccessor(0);
+ BasicBlock* falseSucc = BI->getSuccessor(1);
+
+ if (trueSucc->getSinglePredecessor())
+ localAvail[trueSucc]->table[condVN] = Context->getConstantIntTrue();
+ if (falseSucc->getSinglePredecessor())
+ localAvail[falseSucc]->table[condVN] = Context->getConstantIntFalse();
+
+ return false;
+
// Allocations are always uniquely numbered, so we can save time and memory
- // by fast failing them.
- if (isa<AllocationInst>(I) || isa<TerminatorInst>(I)) {
+ // by fast failing them.
+ } else if (isa<AllocationInst>(I) || isa<TerminatorInst>(I)) {
localAvail[I->getParent()]->table.insert(std::make_pair(num, I));
return false;
}
if (constVal) {
for (PhiMapType::iterator PI = phiMap.begin(), PE = phiMap.end();
PI != PE; ++PI)
- if (PI->second.count(p))
- PI->second.erase(p);
+ PI->second.erase(p);
p->replaceAllUsesWith(constVal);
+ if (isa<PointerType>(constVal->getType()))
+ MD->invalidateCachedPointerInfo(constVal);
+ VN.erase(p);
+
toErase.push_back(p);
} else {
localAvail[I->getParent()]->table.insert(std::make_pair(num, I));
} else if (num == nextNum) {
localAvail[I->getParent()]->table.insert(std::make_pair(num, I));
- // Perform value-number based elimination
+ // Perform fast-path value-number based elimination of values inherited from
+ // dominators.
} else if (Value* repl = lookupNumber(I->getParent(), num)) {
// Remove it!
VN.erase(I);
I->replaceAllUsesWith(repl);
+ if (isa<PointerType>(repl->getType()))
+ MD->invalidateCachedPointerInfo(repl);
toErase.push_back(I);
return true;
+
+#if 0
+ // Perform slow-pathvalue-number based elimination with phi construction.
+ } else if (Value* repl = AttemptRedundancyElimination(I, num)) {
+ // Remove it!
+ VN.erase(I);
+ I->replaceAllUsesWith(repl);
+ if (isa<PointerType>(repl->getType()))
+ MD->invalidateCachedPointerInfo(repl);
+ toErase.push_back(I);
+ return true;
+#endif
} else {
localAvail[I->getParent()]->table.insert(std::make_pair(num, I));
}
return false;
}
-// GVN::runOnFunction - This is the main transformation entry point for a
-// function.
-//
+/// runOnFunction - This is the main transformation entry point for a function.
bool GVN::runOnFunction(Function& F) {
MD = &getAnalysis<MemoryDependenceAnalysis>();
DT = &getAnalysis<DominatorTree>();
changed |= removedBlock;
}
+ unsigned Iteration = 0;
+
while (shouldContinue) {
+ DEBUG(cerr << "GVN iteration: " << Iteration << "\n");
shouldContinue = iterateOnFunction(F);
changed |= shouldContinue;
+ ++Iteration;
}
if (EnablePRE) {
changed |= PREChanged;
}
}
+ // FIXME: Should perform GVN again after PRE does something. PRE can move
+ // computations into blocks where they become fully redundant. Note that
+ // we can't do this until PRE's critical edge splitting updates memdep.
+ // Actually, when this happens, we should just fully integrate PRE into GVN.
cleanupGlobalSets();
}
-bool GVN::processBlock(DomTreeNode* DTN) {
- BasicBlock* BB = DTN->getBlock();
+bool GVN::processBlock(BasicBlock* BB) {
+ // FIXME: Kill off toErase by doing erasing eagerly in a helper function (and
+ // incrementing BI before processing an instruction).
SmallVector<Instruction*, 8> toErase;
- DenseMap<Value*, LoadInst*> lastSeenLoad;
bool changed_function = false;
- if (DTN->getIDom())
- localAvail[BB] =
- new ValueNumberScope(localAvail[DTN->getIDom()->getBlock()]);
- else
- localAvail[BB] = new ValueNumberScope(0);
-
for (BasicBlock::iterator BI = BB->begin(), BE = BB->end();
BI != BE;) {
- changed_function |= processInstruction(BI, lastSeenLoad, toErase);
+ changed_function |= processInstruction(BI, toErase);
if (toErase.empty()) {
++BI;
continue;
DEBUG(cerr << "GVN removed: " << **I);
MD->removeInstruction(*I);
(*I)->eraseFromParent();
+ DEBUG(verifyRemoved(*I));
}
+ toErase.clear();
if (AtStart)
BI = BB->begin();
else
++BI;
-
- toErase.clear();
}
return changed_function;
for (BasicBlock::iterator BI = CurrentBlock->begin(),
BE = CurrentBlock->end(); BI != BE; ) {
Instruction *CurInst = BI++;
-
+
if (isa<AllocationInst>(CurInst) || isa<TerminatorInst>(CurInst) ||
- isa<PHINode>(CurInst) || CurInst->mayReadFromMemory() ||
- CurInst->mayWriteToMemory())
+ isa<PHINode>(CurInst) || (CurInst->getType() == Type::VoidTy) ||
+ CurInst->mayReadFromMemory() || CurInst->mayHaveSideEffects() ||
+ isa<DbgInfoIntrinsic>(CurInst))
continue;
-
+
uint32_t valno = VN.lookup(CurInst);
// Look for the predecessors for PRE opportunities. We're
if (isCriticalEdge(PREPred->getTerminator(), succNum)) {
toSplit.push_back(std::make_pair(PREPred->getTerminator(), succNum));
- Changed = true;
continue;
}
// will be available in the predecessor by the time we need them. Any
// that weren't original present will have been instantiated earlier
// in this loop.
- Instruction* PREInstr = CurInst->clone();
+ Instruction* PREInstr = CurInst->clone(*Context);
bool success = true;
for (unsigned i = 0, e = CurInst->getNumOperands(); i != e; ++i) {
Value *Op = PREInstr->getOperand(i);
// are not value numbered precisely.
if (!success) {
delete PREInstr;
+ DEBUG(verifyRemoved(PREInstr));
continue;
}
localAvail[CurrentBlock]->table[valno] = Phi;
CurInst->replaceAllUsesWith(Phi);
+ if (isa<PointerType>(Phi->getType()))
+ MD->invalidateCachedPointerInfo(Phi);
VN.erase(CurInst);
DEBUG(cerr << "GVN PRE removed: " << *CurInst);
MD->removeInstruction(CurInst);
CurInst->eraseFromParent();
+ DEBUG(verifyRemoved(CurInst));
Changed = true;
}
}
for (SmallVector<std::pair<TerminatorInst*, unsigned>, 4>::iterator
- I = toSplit.begin(), E = toSplit.end(); I != E; ++I) {
+ I = toSplit.begin(), E = toSplit.end(); I != E; ++I)
SplitCriticalEdge(I->first, I->second, this);
- BasicBlock* NewBlock = I->first->getSuccessor(I->second);
- localAvail[NewBlock] =
- new ValueNumberScope(localAvail[I->first->getParent()]);
- }
- return Changed;
+ return Changed || toSplit.size();
}
-// iterateOnFunction - Executes one iteration of GVN
+/// iterateOnFunction - Executes one iteration of GVN
bool GVN::iterateOnFunction(Function &F) {
cleanupGlobalSets();
+ for (df_iterator<DomTreeNode*> DI = df_begin(DT->getRootNode()),
+ DE = df_end(DT->getRootNode()); DI != DE; ++DI) {
+ if (DI->getIDom())
+ localAvail[DI->getBlock()] =
+ new ValueNumberScope(localAvail[DI->getIDom()->getBlock()]);
+ else
+ localAvail[DI->getBlock()] = new ValueNumberScope(0);
+ }
+
// Top-down walk of the dominator tree
bool changed = false;
+#if 0
+ // Needed for value numbering with phi construction to work.
+ ReversePostOrderTraversal<Function*> RPOT(&F);
+ for (ReversePostOrderTraversal<Function*>::rpo_iterator RI = RPOT.begin(),
+ RE = RPOT.end(); RI != RE; ++RI)
+ changed |= processBlock(*RI);
+#else
for (df_iterator<DomTreeNode*> DI = df_begin(DT->getRootNode()),
DE = df_end(DT->getRootNode()); DI != DE; ++DI)
- changed |= processBlock(*DI);
-
+ changed |= processBlock(DI->getBlock());
+#endif
+
return changed;
}
delete I->second;
localAvail.clear();
}
+
+/// verifyRemoved - Verify that the specified instruction does not occur in our
+/// internal data structures.
+void GVN::verifyRemoved(const Instruction *Inst) const {
+ VN.verifyRemoved(Inst);
+
+ // Walk through the PHI map to make sure the instruction isn't hiding in there
+ // somewhere.
+ for (PhiMapType::iterator
+ I = phiMap.begin(), E = phiMap.end(); I != E; ++I) {
+ assert(I->first != Inst && "Inst is still a key in PHI map!");
+
+ for (SmallPtrSet<Instruction*, 4>::iterator
+ II = I->second.begin(), IE = I->second.end(); II != IE; ++II) {
+ assert(*II != Inst && "Inst is still a value in PHI map!");
+ }
+ }
+
+ // Walk through the value number scope to make sure the instruction isn't
+ // ferreted away in it.
+ for (DenseMap<BasicBlock*, ValueNumberScope*>::iterator
+ I = localAvail.begin(), E = localAvail.end(); I != E; ++I) {
+ const ValueNumberScope *VNS = I->second;
+
+ while (VNS) {
+ for (DenseMap<uint32_t, Value*>::iterator
+ II = VNS->table.begin(), IE = VNS->table.end(); II != IE; ++II) {
+ assert(II->second != Inst && "Inst still in value numbering scope!");
+ }
+
+ VNS = VNS->parent;
+ }
+ }
+}