// 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
+// ValueNumbering analysis passes.
+//
//===----------------------------------------------------------------------===//
#define DEBUG_TYPE "gvn"
-
#include "llvm/Transforms/Scalar.h"
#include "llvm/BasicBlock.h"
#include "llvm/Constants.h"
#include "llvm/DerivedTypes.h"
#include "llvm/Function.h"
-#include "llvm/IntrinsicInst.h"
#include "llvm/Instructions.h"
-#include "llvm/ParameterAttributes.h"
#include "llvm/Value.h"
-#include "llvm/ADT/BitVector.h"
#include "llvm/ADT/DenseMap.h"
#include "llvm/ADT/DepthFirstIterator.h"
#include "llvm/ADT/SmallPtrSet.h"
#include "llvm/Analysis/AliasAnalysis.h"
#include "llvm/Analysis/MemoryDependenceAnalysis.h"
#include "llvm/Support/CFG.h"
+#include "llvm/Support/CommandLine.h"
#include "llvm/Support/Compiler.h"
-#include "llvm/Target/TargetData.h"
+#include "llvm/Support/Debug.h"
+#include "llvm/Transforms/Utils/BasicBlockUtils.h"
using namespace llvm;
+STATISTIC(NumGVNInstr, "Number of instructions deleted");
+STATISTIC(NumGVNLoad, "Number of loads deleted");
+STATISTIC(NumGVNPRE, "Number of instructions PRE'd");
+
+static cl::opt<bool> EnablePRE("enable-pre",
+ cl::init(false), cl::Hidden);
+
//===----------------------------------------------------------------------===//
// ValueTable Class
//===----------------------------------------------------------------------===//
FCMPULT, FCMPULE, FCMPUNE, EXTRACT, INSERT,
SHUFFLE, SELECT, TRUNC, ZEXT, SEXT, FPTOUI,
FPTOSI, UITOFP, SITOFP, FPTRUNC, FPEXT,
- PTRTOINT, INTTOPTR, BITCAST, GEP, CALL, EMPTY,
- TOMBSTONE };
+ PTRTOINT, INTTOPTR, BITCAST, GEP, CALL, CONSTANT,
+ EMPTY, TOMBSTONE };
ExpressionOpcode opcode;
const Type* type;
DenseMap<Value*, uint32_t> valueNumbering;
DenseMap<Expression, uint32_t> expressionNumbering;
AliasAnalysis* AA;
+ MemoryDependenceAnalysis* MD;
+ DominatorTree* DT;
uint32_t nextValueNumber;
Expression create_expression(CastInst* C);
Expression create_expression(GetElementPtrInst* G);
Expression create_expression(CallInst* C);
+ Expression create_expression(Constant* C);
public:
ValueTable() : nextValueNumber(1) { }
uint32_t lookup_or_add(Value* V);
void erase(Value* v);
unsigned size();
void setAliasAnalysis(AliasAnalysis* A) { AA = A; }
- uint32_t hash_operand(Value* v);
+ void setMemDep(MemoryDependenceAnalysis* M) { MD = M; }
+ void setDomTree(DominatorTree* D) { DT = D; }
+ uint32_t getNextUnusedValueNumber() { return nextValueNumber; }
};
}
}
Expression::ExpressionOpcode ValueTable::getOpcode(CmpInst* C) {
- if (isa<ICmpInst>(C)) {
+ if (isa<ICmpInst>(C) || isa<VICmpInst>(C)) {
switch (C->getPredicate()) {
default: // THIS SHOULD NEVER HAPPEN
assert(0 && "Comparison with unknown predicate?");
case ICmpInst::ICMP_SLE: return Expression::ICMPSLE;
}
}
- assert(isa<FCmpInst>(C) && "Unknown compare");
+ assert((isa<FCmpInst>(C) || isa<VFCmpInst>(C)) && "Unknown compare");
switch (C->getPredicate()) {
default: // THIS SHOULD NEVER HAPPEN
assert(0 && "Comparison with unknown predicate?");
}
}
-uint32_t ValueTable::hash_operand(Value* v) {
- if (CallInst* CI = dyn_cast<CallInst>(v))
- if (!AA->doesNotAccessMemory(CI))
- return nextValueNumber++;
-
- return lookup_or_add(v);
-}
-
Expression ValueTable::create_expression(CallInst* C) {
Expression e;
for (CallInst::op_iterator I = C->op_begin()+1, E = C->op_end();
I != E; ++I)
- e.varargs.push_back(hash_operand(*I));
+ e.varargs.push_back(lookup_or_add(*I));
return e;
}
Expression ValueTable::create_expression(BinaryOperator* BO) {
Expression e;
- e.firstVN = hash_operand(BO->getOperand(0));
- e.secondVN = hash_operand(BO->getOperand(1));
+ e.firstVN = lookup_or_add(BO->getOperand(0));
+ e.secondVN = lookup_or_add(BO->getOperand(1));
e.thirdVN = 0;
e.function = 0;
e.type = BO->getType();
Expression ValueTable::create_expression(CmpInst* C) {
Expression e;
- e.firstVN = hash_operand(C->getOperand(0));
- e.secondVN = hash_operand(C->getOperand(1));
+ e.firstVN = lookup_or_add(C->getOperand(0));
+ e.secondVN = lookup_or_add(C->getOperand(1));
e.thirdVN = 0;
e.function = 0;
e.type = C->getType();
Expression ValueTable::create_expression(CastInst* C) {
Expression e;
- e.firstVN = hash_operand(C->getOperand(0));
+ e.firstVN = lookup_or_add(C->getOperand(0));
e.secondVN = 0;
e.thirdVN = 0;
e.function = 0;
Expression ValueTable::create_expression(ShuffleVectorInst* S) {
Expression e;
- e.firstVN = hash_operand(S->getOperand(0));
- e.secondVN = hash_operand(S->getOperand(1));
- e.thirdVN = hash_operand(S->getOperand(2));
+ e.firstVN = lookup_or_add(S->getOperand(0));
+ e.secondVN = lookup_or_add(S->getOperand(1));
+ e.thirdVN = lookup_or_add(S->getOperand(2));
e.function = 0;
e.type = S->getType();
e.opcode = Expression::SHUFFLE;
Expression ValueTable::create_expression(ExtractElementInst* E) {
Expression e;
- e.firstVN = hash_operand(E->getOperand(0));
- e.secondVN = hash_operand(E->getOperand(1));
+ e.firstVN = lookup_or_add(E->getOperand(0));
+ e.secondVN = lookup_or_add(E->getOperand(1));
e.thirdVN = 0;
e.function = 0;
e.type = E->getType();
Expression ValueTable::create_expression(InsertElementInst* I) {
Expression e;
- e.firstVN = hash_operand(I->getOperand(0));
- e.secondVN = hash_operand(I->getOperand(1));
- e.thirdVN = hash_operand(I->getOperand(2));
+ e.firstVN = lookup_or_add(I->getOperand(0));
+ e.secondVN = lookup_or_add(I->getOperand(1));
+ e.thirdVN = lookup_or_add(I->getOperand(2));
e.function = 0;
e.type = I->getType();
e.opcode = Expression::INSERT;
Expression ValueTable::create_expression(SelectInst* I) {
Expression e;
- e.firstVN = hash_operand(I->getCondition());
- e.secondVN = hash_operand(I->getTrueValue());
- e.thirdVN = hash_operand(I->getFalseValue());
+ e.firstVN = lookup_or_add(I->getCondition());
+ e.secondVN = lookup_or_add(I->getTrueValue());
+ e.thirdVN = lookup_or_add(I->getFalseValue());
e.function = 0;
e.type = I->getType();
e.opcode = Expression::SELECT;
Expression ValueTable::create_expression(GetElementPtrInst* G) {
Expression e;
-
- e.firstVN = hash_operand(G->getPointerOperand());
+
+ e.firstVN = lookup_or_add(G->getPointerOperand());
e.secondVN = 0;
e.thirdVN = 0;
e.function = 0;
for (GetElementPtrInst::op_iterator I = G->idx_begin(), E = G->idx_end();
I != E; ++I)
- e.varargs.push_back(hash_operand(*I));
+ e.varargs.push_back(lookup_or_add(*I));
return e;
}
// ValueTable External Functions
//===----------------------------------------------------------------------===//
+/// add - Insert a value into the table with a specified value number.
+void ValueTable::add(Value* V, uint32_t num) {
+ valueNumbering.insert(std::make_pair(V, num));
+}
+
/// lookup_or_add - Returns the value number for the specified value, assigning
/// it a new number if it did not have one before.
uint32_t ValueTable::lookup_or_add(Value* V) {
return VI->second;
if (CallInst* C = dyn_cast<CallInst>(V)) {
- if (AA->onlyReadsMemory(C)) { // includes doesNotAccessMemory
+ if (AA->doesNotAccessMemory(C)) {
Expression e = create_expression(C);
DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e);
return nextValueNumber++;
}
+ } else if (AA->onlyReadsMemory(C)) {
+ Expression e = create_expression(C);
+
+ if (expressionNumbering.find(e) == expressionNumbering.end()) {
+ expressionNumbering.insert(std::make_pair(e, nextValueNumber));
+ valueNumbering.insert(std::make_pair(V, nextValueNumber));
+ return nextValueNumber++;
+ }
+
+ Instruction* local_dep = MD->getDependency(C);
+
+ if (local_dep == MemoryDependenceAnalysis::None) {
+ valueNumbering.insert(std::make_pair(V, nextValueNumber));
+ return nextValueNumber++;
+ } else if (local_dep != MemoryDependenceAnalysis::NonLocal) {
+ if (!isa<CallInst>(local_dep)) {
+ valueNumbering.insert(std::make_pair(V, nextValueNumber));
+ return nextValueNumber++;
+ }
+
+ CallInst* local_cdep = cast<CallInst>(local_dep);
+
+ if (local_cdep->getCalledFunction() != C->getCalledFunction() ||
+ local_cdep->getNumOperands() != C->getNumOperands()) {
+ valueNumbering.insert(std::make_pair(V, nextValueNumber));
+ return nextValueNumber++;
+ } else if (!C->getCalledFunction()) {
+ valueNumbering.insert(std::make_pair(V, nextValueNumber));
+ return nextValueNumber++;
+ } else {
+ 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));
+ if (c_vn != cd_vn) {
+ valueNumbering.insert(std::make_pair(V, nextValueNumber));
+ return nextValueNumber++;
+ }
+ }
+
+ uint32_t v = lookup_or_add(local_cdep);
+ valueNumbering.insert(std::make_pair(V, v));
+ return v;
+ }
+ }
+
+
+ DenseMap<BasicBlock*, Value*> deps;
+ MD->getNonLocalDependency(C, deps);
+ CallInst* cdep = 0;
+
+ for (DenseMap<BasicBlock*, Value*>::iterator I = deps.begin(),
+ E = deps.end(); I != E; ++I) {
+ if (I->second == MemoryDependenceAnalysis::None) {
+ valueNumbering.insert(std::make_pair(V, nextValueNumber));
+
+ return nextValueNumber++;
+ } else if (I->second != MemoryDependenceAnalysis::NonLocal) {
+ if (DT->properlyDominates(I->first, C->getParent())) {
+ if (CallInst* CD = dyn_cast<CallInst>(I->second))
+ cdep = CD;
+ else {
+ valueNumbering.insert(std::make_pair(V, nextValueNumber));
+ return nextValueNumber++;
+ }
+ } else {
+ valueNumbering.insert(std::make_pair(V, nextValueNumber));
+ return nextValueNumber++;
+ }
+ }
+ }
+
+ if (!cdep) {
+ valueNumbering.insert(std::make_pair(V, nextValueNumber));
+ return nextValueNumber++;
+ }
+
+ if (cdep->getCalledFunction() != C->getCalledFunction() ||
+ cdep->getNumOperands() != C->getNumOperands()) {
+ valueNumbering.insert(std::make_pair(V, nextValueNumber));
+ return nextValueNumber++;
+ } else if (!C->getCalledFunction()) {
+ valueNumbering.insert(std::make_pair(V, nextValueNumber));
+ return nextValueNumber++;
+ } else {
+ 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(cdep->getOperand(i));
+ if (c_vn != cd_vn) {
+ valueNumbering.insert(std::make_pair(V, nextValueNumber));
+ return nextValueNumber++;
+ }
+ }
+
+ uint32_t v = lookup_or_add(cdep);
+ valueNumbering.insert(std::make_pair(V, v));
+ return v;
+ }
+
} else {
valueNumbering.insert(std::make_pair(V, nextValueNumber));
return nextValueNumber++;
}
//===----------------------------------------------------------------------===//
-// ValueNumberedSet Class
+// GVN Pass
//===----------------------------------------------------------------------===//
-namespace {
-class VISIBILITY_HIDDEN ValueNumberedSet {
- private:
- SmallPtrSet<Value*, 8> contents;
- BitVector numbers;
- public:
- ValueNumberedSet() { numbers.resize(1); }
- ValueNumberedSet(const ValueNumberedSet& other) {
- numbers = other.numbers;
- contents = other.contents;
- }
-
- typedef SmallPtrSet<Value*, 8>::iterator iterator;
-
- iterator begin() { return contents.begin(); }
- iterator end() { return contents.end(); }
-
- bool insert(Value* v) { return contents.insert(v); }
- void insert(iterator I, iterator E) { contents.insert(I, E); }
- void erase(Value* v) { contents.erase(v); }
- unsigned count(Value* v) { return contents.count(v); }
- size_t size() { return contents.size(); }
-
- void set(unsigned i) {
- if (i >= numbers.size())
- numbers.resize(i+1);
-
- numbers.set(i);
- }
-
- void operator=(const ValueNumberedSet& other) {
- contents = other.contents;
- numbers = other.numbers;
- }
-
- void reset(unsigned i) {
- if (i < numbers.size())
- numbers.reset(i);
- }
-
- bool test(unsigned i) {
- if (i >= numbers.size())
- return false;
-
- return numbers.test(i);
- }
-
- void clear() {
- contents.clear();
- numbers.clear();
+
+namespace llvm {
+ template<> struct DenseMapInfo<uint32_t> {
+ static inline uint32_t getEmptyKey() { return ~0; }
+ static inline uint32_t getTombstoneKey() { return ~0 - 1; }
+ static unsigned getHashValue(const uint32_t& Val) { return Val * 37; }
+ static bool isPod() { return true; }
+ static bool isEqual(const uint32_t& LHS, const uint32_t& RHS) {
+ return LHS == RHS;
}
-};
+ };
}
-//===----------------------------------------------------------------------===//
-// GVN Pass
-//===----------------------------------------------------------------------===//
+namespace {
+ struct VISIBILITY_HIDDEN ValueNumberScope {
+ ValueNumberScope* parent;
+ DenseMap<uint32_t, Value*> table;
+
+ ValueNumberScope(ValueNumberScope* p) : parent(p) { }
+ };
+}
namespace {
private:
ValueTable VN;
-
- DenseMap<BasicBlock*, ValueNumberedSet> availableOut;
+ DenseMap<BasicBlock*, ValueNumberScope*> localAvail;
typedef DenseMap<Value*, SmallPtrSet<Instruction*, 4> > PhiMapType;
PhiMapType phiMap;
// This transformation requires dominator postdominator info
virtual void getAnalysisUsage(AnalysisUsage &AU) const {
- AU.setPreservesCFG();
AU.addRequired<DominatorTree>();
AU.addRequired<MemoryDependenceAnalysis>();
AU.addRequired<AliasAnalysis>();
- AU.addRequired<TargetData>();
+
+ AU.addPreserved<DominatorTree>();
AU.addPreserved<AliasAnalysis>();
AU.addPreserved<MemoryDependenceAnalysis>();
- AU.addPreserved<TargetData>();
}
// Helper fuctions
// FIXME: eliminate or document these better
- Value* find_leader(ValueNumberedSet& vals, uint32_t v) ;
- void val_insert(ValueNumberedSet& s, Value* v);
bool processLoad(LoadInst* L,
- DenseMap<Value*, LoadInst*>& lastLoad,
- SmallVector<Instruction*, 4>& toErase);
+ DenseMap<Value*, LoadInst*> &lastLoad,
+ SmallVectorImpl<Instruction*> &toErase);
bool processInstruction(Instruction* I,
- ValueNumberedSet& currAvail,
DenseMap<Value*, LoadInst*>& lastSeenLoad,
- SmallVector<Instruction*, 4>& toErase);
+ SmallVectorImpl<Instruction*> &toErase);
bool processNonLocalLoad(LoadInst* L,
- SmallVector<Instruction*, 4>& toErase);
- bool processMemCpy(MemCpyInst* M, MemCpyInst* MDep,
- SmallVector<Instruction*, 4>& toErase);
- bool performCallSlotOptzn(MemCpyInst* cpy, CallInst* C,
- SmallVector<Instruction*, 4>& toErase);
+ SmallVectorImpl<Instruction*> &toErase);
+ bool processBlock(DomTreeNode* DTN);
Value *GetValueForBlock(BasicBlock *BB, LoadInst* orig,
DenseMap<BasicBlock*, Value*> &Phis,
bool top_level = false);
- void dump(DenseMap<BasicBlock*, Value*>& d);
+ void dump(DenseMap<uint32_t, Value*>& d);
bool iterateOnFunction(Function &F);
Value* CollapsePhi(PHINode* p);
bool isSafeReplacement(PHINode* p, Instruction* inst);
+ bool performPRE(Function& F);
+ Value* lookupNumber(BasicBlock* BB, uint32_t num);
};
char GVN::ID = 0;
static RegisterPass<GVN> X("gvn",
"Global Value Numbering");
-STATISTIC(NumGVNInstr, "Number of instructions deleted");
-STATISTIC(NumGVNLoad, "Number of loads deleted");
-
-/// find_leader - Given a set and a value number, return the first
-/// element of the set with that value number, or 0 if no such element
-/// is present
-Value* GVN::find_leader(ValueNumberedSet& vals, uint32_t v) {
- if (!vals.test(v))
- return 0;
-
- for (ValueNumberedSet::iterator I = vals.begin(), E = vals.end();
- I != E; ++I)
- if (v == VN.lookup(*I))
- return *I;
-
- assert(0 && "No leader found, but present bit is set?");
- return 0;
-}
-
-/// val_insert - Insert a value into a set only if there is not a value
-/// with the same value number already in the set
-void GVN::val_insert(ValueNumberedSet& s, Value* v) {
- uint32_t num = VN.lookup(v);
- if (!s.test(num))
- s.insert(v);
-}
-
-void GVN::dump(DenseMap<BasicBlock*, Value*>& d) {
+void GVN::dump(DenseMap<uint32_t, Value*>& d) {
printf("{\n");
- for (DenseMap<BasicBlock*, Value*>::iterator I = d.begin(),
+ for (DenseMap<uint32_t, Value*>::iterator I = d.begin(),
E = d.end(); I != E; ++I) {
- if (I->second == MemoryDependenceAnalysis::None)
- printf("None\n");
- else
+ printf("%d\n", I->first);
I->second->dump();
}
printf("}\n");
DenseMap<BasicBlock*, Value*>::iterator V = Phis.find(BB);
if (V != Phis.end() && !top_level) return V->second;
+ // If the block is unreachable, just return undef, since this path
+ // can't actually occur at runtime.
+ if (!getAnalysis<DominatorTree>().isReachableFromEntry(BB))
+ return Phis[BB] = UndefValue::get(orig->getType());
+
BasicBlock* singlePred = BB->getSinglePredecessor();
if (singlePred) {
Value *ret = GetValueForBlock(singlePred, orig, Phis);
// 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 = new PHINode(orig->getType(), orig->getName()+".rle",
- BB->begin());
+ PHINode *PN = PHINode::Create(orig->getType(), orig->getName()+".rle",
+ BB->begin());
PN->reserveOperandSpace(std::distance(pred_begin(BB), pred_end(BB)));
if (Phis.count(BB) == 0)
/// processNonLocalLoad - Attempt to eliminate a load whose dependencies are
/// non-local by performing PHI construction.
bool GVN::processNonLocalLoad(LoadInst* L,
- SmallVector<Instruction*, 4>& toErase) {
+ SmallVectorImpl<Instruction*> &toErase) {
MemoryDependenceAnalysis& MD = getAnalysis<MemoryDependenceAnalysis>();
// Find the non-local dependencies of the load
/// 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,
- SmallVector<Instruction*, 4>& toErase) {
+bool GVN::processLoad(LoadInst *L, DenseMap<Value*, LoadInst*> &lastLoad,
+ SmallVectorImpl<Instruction*> &toErase) {
if (L->isVolatile()) {
lastLoad[L->getPointerOperand()] = L;
return false;
return deletedLoad;
}
-/// performCallSlotOptzn - takes a memcpy and a call that it depends on,
-/// and checks for the possibility of a call slot optimization by having
-/// the call write its result directly into the destination of the memcpy.
-bool GVN::performCallSlotOptzn(MemCpyInst* cpy, CallInst* C,
- SmallVector<Instruction*, 4>& toErase) {
- // The general transformation to keep in mind is
- //
- // call @func(..., src, ...)
- // memcpy(dest, src, ...)
- //
- // ->
- //
- // memcpy(dest, src, ...)
- // call @func(..., dest, ...)
- //
- // Since moving the memcpy is technically awkward, we additionally check that
- // src only holds uninitialized values at the moment of the call, meaning that
- // the memcpy can be discarded rather than moved.
-
- // Deliberately get the source and destination with bitcasts stripped away,
- // because we'll need to do type comparisons based on the underlying type.
- Value* cpyDest = cpy->getDest();
- Value* cpySrc = cpy->getSource();
- CallSite CS = CallSite::get(C);
-
- // We need to be able to reason about the size of the memcpy, so we require
- // that it be a constant.
- ConstantInt* cpyLength = dyn_cast<ConstantInt>(cpy->getLength());
- if (!cpyLength)
- return false;
-
- // Require that src be an alloca. This simplifies the reasoning considerably.
- AllocaInst* srcAlloca = dyn_cast<AllocaInst>(cpySrc);
- if (!srcAlloca)
- return false;
-
- // Check that all of src is copied to dest.
- TargetData& TD = getAnalysis<TargetData>();
-
- ConstantInt* srcArraySize = dyn_cast<ConstantInt>(srcAlloca->getArraySize());
- if (!srcArraySize)
- return false;
-
- uint64_t srcSize = TD.getABITypeSize(srcAlloca->getAllocatedType()) *
- srcArraySize->getZExtValue();
-
- if (cpyLength->getZExtValue() < srcSize)
- return false;
-
- // Check that accessing the first srcSize bytes of dest will not cause a
- // trap. Otherwise the transform is invalid since it might cause a trap
- // to occur earlier than it otherwise would.
- if (AllocaInst* A = dyn_cast<AllocaInst>(cpyDest)) {
- // The destination is an alloca. Check it is larger than srcSize.
- ConstantInt* destArraySize = dyn_cast<ConstantInt>(A->getArraySize());
- if (!destArraySize)
- return false;
-
- uint64_t destSize = TD.getABITypeSize(A->getAllocatedType()) *
- destArraySize->getZExtValue();
-
- if (destSize < srcSize)
- return false;
- } else if (Argument* A = dyn_cast<Argument>(cpyDest)) {
- // If the destination is an sret parameter then only accesses that are
- // outside of the returned struct type can trap.
- if (!A->hasStructRetAttr())
- return false;
-
- const Type* StructTy = cast<PointerType>(A->getType())->getElementType();
- uint64_t destSize = TD.getABITypeSize(StructTy);
-
- if (destSize < srcSize)
- return false;
- } else {
- return false;
- }
-
- // Check that src is not accessed except via the call and the memcpy. This
- // guarantees that it holds only undefined values when passed in (so the final
- // memcpy can be dropped), that it is not read or written between the call and
- // the memcpy, and that writing beyond the end of it is undefined.
-
- SmallVector<User*, 8> srcUseList(srcAlloca->use_begin(),
- srcAlloca->use_end());
- while (!srcUseList.empty()) {
- User* UI = srcUseList.back();
- srcUseList.pop_back();
-
- if (isa<GetElementPtrInst>(UI) || isa<BitCastInst>(UI)) {
- for (User::use_iterator I = UI->use_begin(), E = UI->use_end();
- I != E; ++I)
- srcUseList.push_back(*I);
- } else if (UI != C && UI != cpy) {
- return false;
- }
- }
-
- // Since we're changing the parameter to the callsite, we need to make sure
- // that what would be the new parameter dominates the callsite.
- DominatorTree& DT = getAnalysis<DominatorTree>();
- if (Instruction* cpyDestInst = dyn_cast<Instruction>(cpyDest))
- if (!DT.dominates(cpyDestInst, C))
- return false;
-
- // In addition to knowing that the call does not access src in some
- // unexpected manner, for example via a global, which we deduce from
- // the use analysis, we also need to know that it does not sneakily
- // access dest. We rely on AA to figure this out for us.
- AliasAnalysis& AA = getAnalysis<AliasAnalysis>();
- if (AA.getModRefInfo(C, cpy->getRawDest(), srcSize) !=
- AliasAnalysis::NoModRef)
- return false;
-
- // All the checks have passed, so do the transformation.
- for (unsigned i = 0; i < CS.arg_size(); ++i)
- if (CS.getArgument(i) == cpySrc) {
- if (cpySrc->getType() != cpyDest->getType())
- cpyDest = CastInst::createPointerCast(cpyDest, cpySrc->getType(),
- cpyDest->getName(), C);
- CS.setArgument(i, cpyDest);
- }
-
- // Drop any cached information about the call, because we may have changed
- // its dependence information by changing its parameter.
- MemoryDependenceAnalysis& MD = getAnalysis<MemoryDependenceAnalysis>();
- MD.dropInstruction(C);
-
- // Remove the memcpy
- MD.removeInstruction(cpy);
- toErase.push_back(cpy);
-
- return true;
-}
-
-/// processMemCpy - perform simplication of memcpy's. If we have memcpy A which
-/// copies X to Y, and memcpy B which copies Y to Z, then we can rewrite B to be
-/// a memcpy from X to Z (or potentially a memmove, depending on circumstances).
-/// This allows later passes to remove the first memcpy altogether.
-bool GVN::processMemCpy(MemCpyInst* M, MemCpyInst* MDep,
- SmallVector<Instruction*, 4>& toErase) {
- // We can only transforms memcpy's where the dest of one is the source of the
- // other
- if (M->getSource() != MDep->getDest())
- return false;
-
- // Second, the length of the memcpy's must be the same, or the preceeding one
- // must be larger than the following one.
- ConstantInt* C1 = dyn_cast<ConstantInt>(MDep->getLength());
- ConstantInt* C2 = dyn_cast<ConstantInt>(M->getLength());
- if (!C1 || !C2)
- return false;
-
- uint64_t DepSize = C1->getValue().getZExtValue();
- uint64_t CpySize = C2->getValue().getZExtValue();
-
- if (DepSize < CpySize)
- return false;
-
- // Finally, we have to make sure that the dest of the second does not
- // alias the source of the first
- AliasAnalysis& AA = getAnalysis<AliasAnalysis>();
- if (AA.alias(M->getRawDest(), CpySize, MDep->getRawSource(), DepSize) !=
- AliasAnalysis::NoAlias)
- return false;
- else if (AA.alias(M->getRawDest(), CpySize, M->getRawSource(), CpySize) !=
- AliasAnalysis::NoAlias)
- return false;
- else if (AA.alias(MDep->getRawDest(), DepSize, MDep->getRawSource(), DepSize)
- != AliasAnalysis::NoAlias)
- return false;
-
- // If all checks passed, then we can transform these memcpy's
- Function* MemCpyFun = Intrinsic::getDeclaration(
- M->getParent()->getParent()->getParent(),
- M->getIntrinsicID());
-
- std::vector<Value*> args;
- args.push_back(M->getRawDest());
- args.push_back(MDep->getRawSource());
- args.push_back(M->getLength());
- args.push_back(M->getAlignment());
+Value* GVN::lookupNumber(BasicBlock* BB, uint32_t num) {
+ DenseMap<BasicBlock*, ValueNumberScope*>::iterator I = localAvail.find(BB);
+ if (I == localAvail.end())
+ return 0;
- CallInst* C = new CallInst(MemCpyFun, args.begin(), args.end(), "", M);
+ ValueNumberScope* locals = I->second;
- MemoryDependenceAnalysis& MD = getAnalysis<MemoryDependenceAnalysis>();
- if (MD.getDependency(C) == MDep) {
- MD.dropInstruction(M);
- toErase.push_back(M);
- return true;
+ while (locals) {
+ DenseMap<uint32_t, Value*>::iterator I = locals->table.find(num);
+ if (I != locals->table.end())
+ return I->second;
+ else
+ locals = locals->parent;
}
- MD.removeInstruction(C);
- toErase.push_back(C);
- return false;
+ return 0;
}
/// processInstruction - When calculating availability, handle an instruction
/// by inserting it into the appropriate sets
-bool GVN::processInstruction(Instruction* I,
- ValueNumberedSet& currAvail,
- DenseMap<Value*, LoadInst*>& lastSeenLoad,
- SmallVector<Instruction*, 4>& toErase) {
+bool GVN::processInstruction(Instruction *I,
+ DenseMap<Value*, LoadInst*> &lastSeenLoad,
+ SmallVectorImpl<Instruction*> &toErase) {
if (LoadInst* L = dyn_cast<LoadInst>(I)) {
- return processLoad(L, lastSeenLoad, toErase);
- } else if (MemCpyInst* M = dyn_cast<MemCpyInst>(I)) {
- MemoryDependenceAnalysis& MD = getAnalysis<MemoryDependenceAnalysis>();
-
- // The are two possible optimizations we can do for memcpy:
- // a) memcpy-memcpy xform which exposes redundance for DSE
- // b) call-memcpy xform for return slot optimization
- Instruction* dep = MD.getDependency(M);
- if (dep == MemoryDependenceAnalysis::None ||
- dep == MemoryDependenceAnalysis::NonLocal)
- return false;
- if (MemCpyInst *MemCpy = dyn_cast<MemCpyInst>(dep))
- return processMemCpy(M, MemCpy, toErase);
- if (CallInst* C = dyn_cast<CallInst>(dep))
- return performCallSlotOptzn(M, C, toErase);
- return false;
+ bool changed = processLoad(L, lastSeenLoad, toErase);
+
+ if (!changed) {
+ unsigned num = VN.lookup_or_add(L);
+ localAvail[I->getParent()]->table.insert(std::make_pair(num, L));
+ }
+
+ return changed;
}
+ uint32_t nextNum = VN.getNextUnusedValueNumber();
unsigned num = VN.lookup_or_add(I);
+ // Allocations are always uniquely numbered, so we can save time and memory
+ // by fast failing them.
+ if (isa<AllocationInst>(I) || isa<TerminatorInst>(I)) {
+ localAvail[I->getParent()]->table.insert(std::make_pair(num, I));
+ return false;
+ }
+
// Collapse PHI nodes
if (PHINode* p = dyn_cast<PHINode>(I)) {
Value* constVal = CollapsePhi(p);
p->replaceAllUsesWith(constVal);
toErase.push_back(p);
+ } else {
+ localAvail[I->getParent()]->table.insert(std::make_pair(num, I));
}
- // Perform value-number based elimination
- } else if (currAvail.test(num)) {
- Value* repl = find_leader(currAvail, num);
-
- if (CallInst* CI = dyn_cast<CallInst>(I)) {
- AliasAnalysis& AA = getAnalysis<AliasAnalysis>();
- if (!AA.doesNotAccessMemory(CI)) {
- MemoryDependenceAnalysis& MD = getAnalysis<MemoryDependenceAnalysis>();
- if (cast<Instruction>(repl)->getParent() != CI->getParent() ||
- MD.getDependency(CI) != MD.getDependency(cast<CallInst>(repl))) {
- // There must be an intervening may-alias store, so nothing from
- // this point on will be able to be replaced with the preceding call
- currAvail.erase(repl);
- currAvail.insert(I);
-
- return false;
- }
- }
- }
+
+ // If the number we were assigned was a brand new VN, then we don't
+ // need to do a lookup to see if the number already exists
+ // somewhere in the domtree: it can't!
+ } else if (num == nextNum) {
+ localAvail[I->getParent()]->table.insert(std::make_pair(num, I));
+ // Perform value-number based elimination
+ } else if (Value* repl = lookupNumber(I->getParent(), num)) {
// Remove it!
MemoryDependenceAnalysis& MD = getAnalysis<MemoryDependenceAnalysis>();
MD.removeInstruction(I);
I->replaceAllUsesWith(repl);
toErase.push_back(I);
return true;
- } else if (!I->isTerminator()) {
- currAvail.set(num);
- currAvail.insert(I);
+ } else {
+ localAvail[I->getParent()]->table.insert(std::make_pair(num, I));
}
return false;
//
bool GVN::runOnFunction(Function& F) {
VN.setAliasAnalysis(&getAnalysis<AliasAnalysis>());
+ VN.setMemDep(&getAnalysis<MemoryDependenceAnalysis>());
+ VN.setDomTree(&getAnalysis<DominatorTree>());
bool changed = false;
bool shouldContinue = true;
}
-// GVN::iterateOnFunction - Executes one iteration of GVN
-bool GVN::iterateOnFunction(Function &F) {
- // Clean out global sets from any previous functions
- VN.clear();
- availableOut.clear();
- phiMap.clear();
-
+bool GVN::processBlock(DomTreeNode* DTN) {
+ BasicBlock* BB = DTN->getBlock();
+
+ SmallVector<Instruction*, 8> toErase;
+ DenseMap<Value*, LoadInst*> lastSeenLoad;
bool changed_function = false;
- DominatorTree &DT = getAnalysis<DominatorTree>();
+ if (DTN->getIDom())
+ localAvail[BB] =
+ new ValueNumberScope(localAvail[DTN->getIDom()->getBlock()]);
+ else
+ localAvail[BB] = new ValueNumberScope(0);
- SmallVector<Instruction*, 4> toErase;
- DenseMap<Value*, LoadInst*> lastSeenLoad;
-
- // Top-down walk of the dominator tree
- for (df_iterator<DomTreeNode*> DI = df_begin(DT.getRootNode()),
- E = df_end(DT.getRootNode()); DI != E; ++DI) {
+ for (BasicBlock::iterator BI = BB->begin(), BE = BB->end();
+ BI != BE;) {
+ changed_function |= processInstruction(BI, lastSeenLoad, toErase);
+ if (toErase.empty()) {
+ ++BI;
+ continue;
+ }
- // Get the set to update for this block
- ValueNumberedSet& currAvail = availableOut[DI->getBlock()];
- lastSeenLoad.clear();
+ // If we need some instructions deleted, do it now.
+ NumGVNInstr += toErase.size();
+
+ // Avoid iterator invalidation.
+ bool AtStart = BI == BB->begin();
+ if (!AtStart)
+ --BI;
+
+ for (SmallVector<Instruction*, 4>::iterator I = toErase.begin(),
+ E = toErase.end(); I != E; ++I)
+ (*I)->eraseFromParent();
- BasicBlock* BB = DI->getBlock();
+ if (AtStart)
+ BI = BB->begin();
+ else
+ ++BI;
+
+ toErase.clear();
+ }
- // A block inherits AVAIL_OUT from its dominator
- if (DI->getIDom() != 0)
- currAvail = availableOut[DI->getIDom()->getBlock()];
+ return changed_function;
+}
- for (BasicBlock::iterator BI = BB->begin(), BE = BB->end();
- BI != BE; ) {
- changed_function |= processInstruction(BI, currAvail,
- lastSeenLoad, toErase);
+/// performPRE - Perform a purely local form of PRE that looks for diamond
+/// control flow patterns and attempts to perform simple PRE at the join point.
+bool GVN::performPRE(Function& F) {
+ bool changed = false;
+ SmallVector<std::pair<TerminatorInst*, unsigned>, 4> toSplit;
+ for (df_iterator<BasicBlock*> DI = df_begin(&F.getEntryBlock()),
+ DE = df_end(&F.getEntryBlock()); DI != DE; ++DI) {
+ BasicBlock* CurrentBlock = *DI;
+
+ // Nothing to PRE in the entry block.
+ if (CurrentBlock == &F.getEntryBlock()) continue;
+
+ for (BasicBlock::iterator BI = CurrentBlock->begin(),
+ BE = CurrentBlock->end(); BI != BE; ) {
+ if (isa<AllocationInst>(BI) || isa<TerminatorInst>(BI) ||
+ isa<PHINode>(BI) || BI->mayReadFromMemory() ||
+ BI->mayWriteToMemory()) {
+ BI++;
+ continue;
+ }
- NumGVNInstr += toErase.size();
+ uint32_t valno = VN.lookup(BI);
- // Avoid iterator invalidation
- ++BI;
-
- for (SmallVector<Instruction*, 4>::iterator I = toErase.begin(),
- E = toErase.end(); I != E; ++I)
- (*I)->eraseFromParent();
-
- toErase.clear();
+ // Look for the predecessors for PRE opportunities. We're
+ // only trying to solve the basic diamond case, where
+ // a value is computed in the successor and one predecessor,
+ // but not the other. We also explicitly disallow cases
+ // where the successor is its own predecessor, because they're
+ // more complicated to get right.
+ unsigned numWith = 0;
+ unsigned numWithout = 0;
+ BasicBlock* PREPred = 0;
+ DenseMap<BasicBlock*, Value*> predMap;
+ for (pred_iterator PI = pred_begin(CurrentBlock),
+ PE = pred_end(CurrentBlock); PI != PE; ++PI) {
+ // We're not interested in PRE where the block is its
+ // own predecessor, on in blocks with predecessors
+ // that are not reachable.
+ if (*PI == CurrentBlock) {
+ numWithout = 2;
+ break;
+ } else if (!localAvail.count(*PI)) {
+ numWithout = 2;
+ break;
+ }
+
+ DenseMap<uint32_t, Value*>::iterator predV =
+ localAvail[*PI]->table.find(valno);
+ if (predV == localAvail[*PI]->table.end()) {
+ PREPred = *PI;
+ numWithout++;
+ } else if (predV->second == BI) {
+ numWithout = 2;
+ } else {
+ predMap[*PI] = predV->second;
+ numWith++;
+ }
+ }
+
+ // Don't do PRE when it might increase code size, i.e. when
+ // we would need to insert instructions in more than one pred.
+ if (numWithout != 1 || numWith == 0) {
+ BI++;
+ continue;
+ }
+
+ // We can't do PRE safely on a critical edge, so instead we schedule
+ // the edge to be split and perform the PRE the next time we iterate
+ // on the function.
+ unsigned succNum = 0;
+ for (unsigned i = 0, e = PREPred->getTerminator()->getNumSuccessors();
+ i != e; ++i)
+ if (PREPred->getTerminator()->getSuccessor(i) == PREPred) {
+ succNum = i;
+ break;
+ }
+
+ if (isCriticalEdge(PREPred->getTerminator(), succNum)) {
+ toSplit.push_back(std::make_pair(PREPred->getTerminator(), succNum));
+ changed = true;
+ BI++;
+ continue;
+ }
+
+ // Instantiate the expression the in predecessor that lacked it.
+ // Because we are going top-down through the block, all value numbers
+ // 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 = BI->clone();
+ bool success = true;
+ for (unsigned i = 0; i < BI->getNumOperands(); ++i) {
+ Value* op = BI->getOperand(i);
+ if (isa<Argument>(op) || isa<Constant>(op) || isa<GlobalValue>(op))
+ PREInstr->setOperand(i, op);
+ else if (!lookupNumber(PREPred, VN.lookup(op))) {
+ success = false;
+ break;
+ } else
+ PREInstr->setOperand(i, lookupNumber(PREPred, VN.lookup(op)));
+ }
+
+ // Fail out if we encounter an operand that is not available in
+ // the PRE predecessor. This is typically because of loads which
+ // are not value numbered precisely.
+ if (!success) {
+ delete PREInstr;
+ BI++;
+ continue;
+ }
+
+ PREInstr->insertBefore(PREPred->getTerminator());
+ PREInstr->setName(BI->getName() + ".pre");
+ predMap[PREPred] = PREInstr;
+ VN.add(PREInstr, valno);
+ NumGVNPRE++;
+
+ // Update the availability map to include the new instruction.
+ localAvail[PREPred]->table.insert(std::make_pair(valno, PREInstr));
+
+ // Create a PHI to make the value available in this block.
+ PHINode* Phi = PHINode::Create(BI->getType(),
+ BI->getName() + ".pre-phi",
+ CurrentBlock->begin());
+ for (pred_iterator PI = pred_begin(CurrentBlock),
+ PE = pred_end(CurrentBlock); PI != PE; ++PI)
+ Phi->addIncoming(predMap[*PI], *PI);
+
+ VN.add(Phi, valno);
+ localAvail[CurrentBlock]->table[valno] = Phi;
+
+ BI->replaceAllUsesWith(Phi);
+ VN.erase(BI);
+
+ Instruction* erase = BI;
+ BI++;
+ erase->eraseFromParent();
+
+ changed = true;
}
}
- return changed_function;
+ for (SmallVector<std::pair<TerminatorInst*, unsigned>, 4>::iterator
+ I = toSplit.begin(), E = toSplit.end(); I != E; ++I)
+ SplitCriticalEdge(I->first, I->second, this);
+
+ return changed;
+}
+
+// GVN::iterateOnFunction - Executes one iteration of GVN
+bool GVN::iterateOnFunction(Function &F) {
+ // Clean out global sets from any previous functions
+ VN.clear();
+ phiMap.clear();
+
+ for (DenseMap<BasicBlock*, ValueNumberScope*>::iterator
+ I = localAvail.begin(), E = localAvail.end(); I != E; ++I)
+ delete I->second;
+ localAvail.clear();
+
+ DominatorTree &DT = getAnalysis<DominatorTree>();
+
+ // Top-down walk of the dominator tree
+ bool changed = false;
+ for (df_iterator<DomTreeNode*> DI = df_begin(DT.getRootNode()),
+ DE = df_end(DT.getRootNode()); DI != DE; ++DI)
+ changed |= processBlock(*DI);
+
+ if (EnablePRE)
+ changed |= performPRE(F);
+
+ return changed;
}
projects / oota-llvm.git / blobdiff