#include "llvm/Value.h"
#include "llvm/ADT/DenseMap.h"
#include "llvm/ADT/DepthFirstIterator.h"
-#include "llvm/ADT/ScopedHashTable.h"
#include "llvm/ADT/SmallPtrSet.h"
#include "llvm/ADT/SmallVector.h"
-#include "llvm/ADT/SparseBitVector.h"
#include "llvm/ADT/Statistic.h"
#include "llvm/Analysis/Dominators.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/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
void setAliasAnalysis(AliasAnalysis* A) { AA = A; }
void setMemDep(MemoryDependenceAnalysis* M) { MD = M; }
void setDomTree(DominatorTree* D) { DT = D; }
+ uint32_t getNextUnusedValueNumber() { return nextValueNumber; }
};
}
// 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) {
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; }
+ 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;
};
}
-typedef ScopedHashTable<uint32_t, Value*> ValueNumberMap;
-typedef ScopedHashTableScope<uint32_t, Value*> ValueNumberScope;
+namespace {
+ struct VISIBILITY_HIDDEN ValueNumberScope {
+ ValueNumberScope* parent;
+ DenseMap<uint32_t, Value*> table;
+
+ ValueNumberScope(ValueNumberScope* p) : parent(p) { }
+ };
+}
namespace {
private:
ValueTable VN;
-
- DenseMap<BasicBlock*, ValueNumberScope> availableOut;
- ValueNumberMap BaseMap;
+ 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.addPreserved<DominatorTree>();
AU.addPreserved<AliasAnalysis>();
AU.addPreserved<MemoryDependenceAnalysis>();
}
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");
-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);
return deletedLoad;
}
+Value* GVN::lookupNumber(BasicBlock* BB, uint32_t num) {
+ DenseMap<BasicBlock*, ValueNumberScope*>::iterator I = localAvail.find(BB);
+ if (I == localAvail.end())
+ return 0;
+
+ ValueNumberScope* locals = I->second;
+
+ while (locals) {
+ DenseMap<uint32_t, Value*>::iterator I = locals->table.find(num);
+ if (I != locals->table.end())
+ return I->second;
+ else
+ locals = locals->parent;
+ }
+
+ return 0;
+}
+
/// 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))
- return processLoad(L, lastSeenLoad, toErase);
+ if (LoadInst* L = dyn_cast<LoadInst>(I)) {
+ 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))
+ if (isa<AllocationInst>(I) || isa<TerminatorInst>(I)) {
+ localAvail[I->getParent()]->table.insert(std::make_pair(num, I));
return false;
-
- unsigned num = VN.lookup_or_add(I);
+ }
// Collapse PHI nodes
if (PHINode* p = dyn_cast<PHINode>(I)) {
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 (BaseMap.begin(num) != BaseMap.end()) {
- Value* repl = *BaseMap.begin(num);
+
+ // 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()) {
- BaseMap.insert(num, I);
+ } else {
+ localAvail[I->getParent()]->table.insert(std::make_pair(num, I));
}
return false;
bool GVN::processBlock(DomTreeNode* DTN) {
BasicBlock* BB = DTN->getBlock();
- ValueNumberScope NewScope(BaseMap);
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);
toErase.clear();
}
- for (DomTreeNode::iterator I = DTN->begin(), E = DTN->end(); I != E; ++I)
- changed_function |= processBlock(*I);
-
return changed_function;
}
+/// 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;
+ }
+
+ uint32_t valno = VN.lookup(BI);
+
+ // 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 {
+ Value* V = lookupNumber(PREPred, VN.lookup(op));
+ if (!V) {
+ success = false;
+ break;
+ } else
+ PREInstr->setOperand(i, V);
+ }
+ }
+
+ // 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;
+ }
+ }
+
+ 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();
- availableOut.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
- return processBlock(DT.getRootNode());
+ 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;
}