// 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/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", cl::init(true));
+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;
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 {
bool mergeBlockIntoPredecessor(BasicBlock* BB);
Value* AttemptRedundancyElimination(Instruction* orig, unsigned valno);
void cleanupGlobalSets();
+ void verifyRemoved(const Instruction *I) const;
};
char GVN::ID = 0;
// 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());
if (BasicBlock *Pred = BB->getSinglePredecessor()) {
Value *ret = GetValueForBlock(Pred, orig, Phis);
DEBUG(cerr << "GVN removed: " << *PN);
MD->removeInstruction(PN);
PN->eraseFromParent();
+ DEBUG(verifyRemoved(PN));
Phis[BB] = v;
return v;
// 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())
+ 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
// Loading the allocation -> undef.
if (isa<AllocationInst>(DepInst)) {
ValuesPerBlock.push_back(std::make_pair(DepBB,
- UndefValue::get(LI->getType())));
+ Context->getUndef(LI->getType())));
continue;
}
Value* v = GetValueForBlock(LI->getParent(), LI, BlockReplValues, true);
LI->replaceAllUsesWith(v);
- if (!isa<GlobalValue>(v))
+ if (isa<PHINode>(v))
v->takeName(LI);
if (isa<PointerType>(v->getType()))
MD->invalidateCachedPointerInfo(v);
// 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
<< UnavailablePred->getName() << "': " << *LI);
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<GlobalValue>(v))
+ if (isa<PHINode>(v))
v->takeName(LI);
if (isa<PointerType>(v->getType()))
MD->invalidateCachedPointerInfo(v);
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())
// undef value. This can happen when loading for a fresh allocation with no
// intervening stores, for example.
if (isa<AllocationInst>(DepInst)) {
- L->replaceAllUsesWith(UndefValue::get(L->getType()));
+ L->replaceAllUsesWith(Context->getUndef(L->getType()));
toErase.push_back(L);
NumGVNLoad++;
return true;
DenseMap<BasicBlock*, ValueNumberScope*>::iterator LA =
localAvail.find(Current);
if (LA == localAvail.end()) return 0;
- DenseMap<unsigned, Value*>::iterator V = LA->second->table.find(valno);
+ DenseMap<uint32_t, Value*>::iterator V = LA->second->table.find(valno);
if (V != LA->second->table.end()) {
// Found an instance, record it.
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;
}
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));
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(BasicBlock* BB) {
- DomTreeNode* DTN = DT->getNode(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
// 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;
}
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
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