// 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/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 <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"/*, cl::init(true)*/);
+cl::opt<bool> EnableLoadPRE("enable-load-pre", cl::init(true));
//===----------------------------------------------------------------------===//
// ValueTable Class
}
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;
};
}
// 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;
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 {
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) {
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;
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;
+
+ // 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())
return false;
-
- BasicBlock *EntryBlock = &LI->getParent()->getParent()->getEntryBlock();
// 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;
-
- 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);
- continue;
- }
+ for (unsigned i = 0, e = Deps.size(); i != e; ++i) {
+ BasicBlock *DepBB = Deps[i].first;
+ MemDepResult DepInfo = Deps[i].second;
if (DepInfo.isClobber()) {
UnavailableBlocks.push_back(DepBB);
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
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 (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;
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);
- v->takeName(LI);
+ if (isa<PHINode>(v))
+ v->takeName(LI);
+ if (isa<PointerType>(v->getType()))
+ MD->invalidateCachedPointerInfo(v);
toErase.push_back(LI);
NumPRELoad++;
return true;
MemDepResult dep = MD->getDependency(L);
// If the value isn't available, don't do anything!
- if (dep.isClobber())
+ 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())
// 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;
// Remove it!
L->replaceAllUsesWith(DepLI);
+ if (isa<PointerType>(DepLI->getType()))
+ MD->invalidateCachedPointerInfo(DepLI);
toErase.push_back(L);
NumGVNLoad++;
return true;
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,
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] = ConstantInt::getTrue();
+ if (falseSucc->getSinglePredecessor())
+ localAvail[falseSucc]->table[condVN] = ConstantInt::getFalse();
+
+ 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;
}
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>();
}
-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;
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, toErase);
DEBUG(cerr << "GVN removed: " << **I);
MD->removeInstruction(*I);
(*I)->eraseFromParent();
+ DEBUG(verifyRemoved(*I));
}
toErase.clear();
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
// 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;
}
}
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;
+ }
+ }
+}
projects / oota-llvm.git / blobdiff