#define DEBUG_TYPE "gvn"
#include "llvm/Transforms/Scalar.h"
-#include "llvm/GlobalVariable.h"
-#include "llvm/IRBuilder.h"
-#include "llvm/IntrinsicInst.h"
-#include "llvm/LLVMContext.h"
-#include "llvm/Metadata.h"
#include "llvm/ADT/DenseMap.h"
#include "llvm/ADT/DepthFirstIterator.h"
#include "llvm/ADT/Hashing.h"
+#include "llvm/ADT/SetVector.h"
#include "llvm/ADT/SmallPtrSet.h"
#include "llvm/ADT/Statistic.h"
#include "llvm/Analysis/AliasAnalysis.h"
+#include "llvm/Analysis/CFG.h"
#include "llvm/Analysis/ConstantFolding.h"
-#include "llvm/Analysis/Dominators.h"
#include "llvm/Analysis/InstructionSimplify.h"
#include "llvm/Analysis/Loads.h"
#include "llvm/Analysis/MemoryBuiltins.h"
#include "llvm/Analysis/MemoryDependenceAnalysis.h"
#include "llvm/Analysis/PHITransAddr.h"
#include "llvm/Analysis/ValueTracking.h"
-#include "llvm/Assembly/Writer.h"
+#include "llvm/IR/DataLayout.h"
+#include "llvm/IR/Dominators.h"
+#include "llvm/IR/GlobalVariable.h"
+#include "llvm/IR/IRBuilder.h"
+#include "llvm/IR/IntrinsicInst.h"
+#include "llvm/IR/LLVMContext.h"
+#include "llvm/IR/Metadata.h"
#include "llvm/Support/Allocator.h"
#include "llvm/Support/CommandLine.h"
#include "llvm/Support/Debug.h"
#include "llvm/Support/PatternMatch.h"
-#include "llvm/Target/TargetData.h"
#include "llvm/Target/TargetLibraryInfo.h"
#include "llvm/Transforms/Utils/BasicBlockUtils.h"
#include "llvm/Transforms/Utils/SSAUpdater.h"
+#include <vector>
using namespace llvm;
using namespace PatternMatch;
if (e.varargs[0] > e.varargs[1])
std::swap(e.varargs[0], e.varargs[1]);
}
-
+
if (CmpInst *C = dyn_cast<CmpInst>(I)) {
// Sort the operand value numbers so x<y and y>x get the same value number.
CmpInst::Predicate Predicate = C->getPredicate();
II != IE; ++II)
e.varargs.push_back(*II);
}
-
+
return e;
}
valueNumbering.insert(std::make_pair(V, num));
}
-uint32_t ValueTable::lookup_or_add_call(CallInst* C) {
+uint32_t ValueTable::lookup_or_add_call(CallInst *C) {
if (AA->doesNotAccessMemory(C)) {
Expression exp = create_expression(C);
- uint32_t& e = expressionNumbering[exp];
+ uint32_t &e = expressionNumbering[exp];
if (!e) e = nextValueNumber++;
valueNumbering[C] = e;
return e;
} else if (AA->onlyReadsMemory(C)) {
Expression exp = create_expression(C);
- uint32_t& e = expressionNumbering[exp];
+ uint32_t &e = expressionNumbering[exp];
if (!e) {
e = nextValueNumber++;
valueNumbering[C] = e;
valueNumbering[V] = nextValueNumber;
return nextValueNumber++;
}
-
+
Instruction* I = cast<Instruction>(V);
Expression exp;
switch (I->getOpcode()) {
case Instruction::LShr:
case Instruction::AShr:
case Instruction::And:
- case Instruction::Or :
+ case Instruction::Or:
case Instruction::Xor:
case Instruction::ICmp:
case Instruction::FCmp:
//===----------------------------------------------------------------------===//
namespace {
+ class GVN;
+ struct AvailableValueInBlock {
+ /// BB - The basic block in question.
+ BasicBlock *BB;
+ enum ValType {
+ SimpleVal, // A simple offsetted value that is accessed.
+ LoadVal, // A value produced by a load.
+ MemIntrin, // A memory intrinsic which is loaded from.
+ UndefVal // A UndefValue representing a value from dead block (which
+ // is not yet physically removed from the CFG).
+ };
+
+ /// V - The value that is live out of the block.
+ PointerIntPair<Value *, 2, ValType> Val;
+
+ /// Offset - The byte offset in Val that is interesting for the load query.
+ unsigned Offset;
+
+ static AvailableValueInBlock get(BasicBlock *BB, Value *V,
+ unsigned Offset = 0) {
+ AvailableValueInBlock Res;
+ Res.BB = BB;
+ Res.Val.setPointer(V);
+ Res.Val.setInt(SimpleVal);
+ Res.Offset = Offset;
+ return Res;
+ }
+
+ static AvailableValueInBlock getMI(BasicBlock *BB, MemIntrinsic *MI,
+ unsigned Offset = 0) {
+ AvailableValueInBlock Res;
+ Res.BB = BB;
+ Res.Val.setPointer(MI);
+ Res.Val.setInt(MemIntrin);
+ Res.Offset = Offset;
+ return Res;
+ }
+
+ static AvailableValueInBlock getLoad(BasicBlock *BB, LoadInst *LI,
+ unsigned Offset = 0) {
+ AvailableValueInBlock Res;
+ Res.BB = BB;
+ Res.Val.setPointer(LI);
+ Res.Val.setInt(LoadVal);
+ Res.Offset = Offset;
+ return Res;
+ }
+
+ static AvailableValueInBlock getUndef(BasicBlock *BB) {
+ AvailableValueInBlock Res;
+ Res.BB = BB;
+ Res.Val.setPointer(0);
+ Res.Val.setInt(UndefVal);
+ Res.Offset = 0;
+ return Res;
+ }
+
+ bool isSimpleValue() const { return Val.getInt() == SimpleVal; }
+ bool isCoercedLoadValue() const { return Val.getInt() == LoadVal; }
+ bool isMemIntrinValue() const { return Val.getInt() == MemIntrin; }
+ bool isUndefValue() const { return Val.getInt() == UndefVal; }
+
+ Value *getSimpleValue() const {
+ assert(isSimpleValue() && "Wrong accessor");
+ return Val.getPointer();
+ }
+
+ LoadInst *getCoercedLoadValue() const {
+ assert(isCoercedLoadValue() && "Wrong accessor");
+ return cast<LoadInst>(Val.getPointer());
+ }
+
+ MemIntrinsic *getMemIntrinValue() const {
+ assert(isMemIntrinValue() && "Wrong accessor");
+ return cast<MemIntrinsic>(Val.getPointer());
+ }
+
+ /// MaterializeAdjustedValue - Emit code into this block to adjust the value
+ /// defined here to the specified type. This handles various coercion cases.
+ Value *MaterializeAdjustedValue(Type *LoadTy, GVN &gvn) const;
+ };
class GVN : public FunctionPass {
bool NoLoads;
MemoryDependenceAnalysis *MD;
DominatorTree *DT;
- const TargetData *TD;
+ const DataLayout *DL;
const TargetLibraryInfo *TLI;
+ SetVector<BasicBlock *> DeadBlocks;
ValueTable VN;
-
+
/// LeaderTable - A mapping from value numbers to lists of Value*'s that
/// have that value number. Use findLeader to query it.
struct LeaderTableEntry {
Value *Val;
- BasicBlock *BB;
+ const BasicBlock *BB;
LeaderTableEntry *Next;
};
DenseMap<uint32_t, LeaderTableEntry> LeaderTable;
BumpPtrAllocator TableAllocator;
-
+
SmallVector<Instruction*, 8> InstrsToErase;
+
+ typedef SmallVector<NonLocalDepResult, 64> LoadDepVect;
+ typedef SmallVector<AvailableValueInBlock, 64> AvailValInBlkVect;
+ typedef SmallVector<BasicBlock*, 64> UnavailBlkVect;
+
public:
static char ID; // Pass identification, replacement for typeid
explicit GVN(bool noloads = false)
}
bool runOnFunction(Function &F);
-
+
/// markInstructionForDeletion - This removes the specified instruction from
/// our various maps and marks it for deletion.
void markInstructionForDeletion(Instruction *I) {
VN.erase(I);
InstrsToErase.push_back(I);
}
-
- const TargetData *getTargetData() const { return TD; }
+
+ const DataLayout *getDataLayout() const { return DL; }
DominatorTree &getDominatorTree() const { return *DT; }
AliasAnalysis *getAliasAnalysis() const { return VN.getAliasAnalysis(); }
MemoryDependenceAnalysis &getMemDep() const { return *MD; }
private:
/// addToLeaderTable - Push a new Value to the LeaderTable onto the list for
/// its value number.
- void addToLeaderTable(uint32_t N, Value *V, BasicBlock *BB) {
+ void addToLeaderTable(uint32_t N, Value *V, const BasicBlock *BB) {
LeaderTableEntry &Curr = LeaderTable[N];
if (!Curr.Val) {
Curr.Val = V;
Curr.BB = BB;
return;
}
-
+
LeaderTableEntry *Node = TableAllocator.Allocate<LeaderTableEntry>();
Node->Val = V;
Node->BB = BB;
Node->Next = Curr.Next;
Curr.Next = Node;
}
-
+
/// removeFromLeaderTable - Scan the list of values corresponding to a given
/// value number, and remove the given instruction if encountered.
void removeFromLeaderTable(uint32_t N, Instruction *I, BasicBlock *BB) {
Prev = Curr;
Curr = Curr->Next;
}
-
+
if (Prev) {
Prev->Next = Curr->Next;
} else {
// This transformation requires dominator postdominator info
virtual void getAnalysisUsage(AnalysisUsage &AU) const {
- AU.addRequired<DominatorTree>();
+ AU.addRequired<DominatorTreeWrapperPass>();
AU.addRequired<TargetLibraryInfo>();
if (!NoLoads)
AU.addRequired<MemoryDependenceAnalysis>();
AU.addRequired<AliasAnalysis>();
- AU.addPreserved<DominatorTree>();
+ AU.addPreserved<DominatorTreeWrapperPass>();
AU.addPreserved<AliasAnalysis>();
}
-
- // Helper fuctions
- // FIXME: eliminate or document these better
+
+ // Helper fuctions of redundant load elimination
bool processLoad(LoadInst *L);
- bool processInstruction(Instruction *I);
bool processNonLocalLoad(LoadInst *L);
+ void AnalyzeLoadAvailability(LoadInst *LI, LoadDepVect &Deps,
+ AvailValInBlkVect &ValuesPerBlock,
+ UnavailBlkVect &UnavailableBlocks);
+ bool PerformLoadPRE(LoadInst *LI, AvailValInBlkVect &ValuesPerBlock,
+ UnavailBlkVect &UnavailableBlocks);
+
+ // Other helper routines
+ bool processInstruction(Instruction *I);
bool processBlock(BasicBlock *BB);
void dump(DenseMap<uint32_t, Value*> &d);
bool iterateOnFunction(Function &F);
bool performPRE(Function &F);
- Value *findLeader(BasicBlock *BB, uint32_t num);
+ Value *findLeader(const BasicBlock *BB, uint32_t num);
void cleanupGlobalSets();
void verifyRemoved(const Instruction *I) const;
bool splitCriticalEdges();
+ BasicBlock *splitCriticalEdges(BasicBlock *Pred, BasicBlock *Succ);
unsigned replaceAllDominatedUsesWith(Value *From, Value *To,
- BasicBlock *Root);
- bool propagateEquality(Value *LHS, Value *RHS, BasicBlock *Root);
+ const BasicBlockEdge &Root);
+ bool propagateEquality(Value *LHS, Value *RHS, const BasicBlockEdge &Root);
+ bool processFoldableCondBr(BranchInst *BI);
+ void addDeadBlock(BasicBlock *BB);
+ void assignValNumForDeadCode();
};
char GVN::ID = 0;
INITIALIZE_PASS_BEGIN(GVN, "gvn", "Global Value Numbering", false, false)
INITIALIZE_PASS_DEPENDENCY(MemoryDependenceAnalysis)
-INITIALIZE_PASS_DEPENDENCY(DominatorTree)
+INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfo)
INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
INITIALIZE_PASS_END(GVN, "gvn", "Global Value Numbering", false, false)
+#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
void GVN::dump(DenseMap<uint32_t, Value*>& d) {
errs() << "{\n";
for (DenseMap<uint32_t, Value*>::iterator I = d.begin(),
}
errs() << "}\n";
}
+#endif
/// IsValueFullyAvailableInBlock - Return true if we can prove that the value
/// we're analyzing is fully available in the specified block. As we go, keep
// Mark as unavailable.
EntryVal = 0;
- for (succ_iterator I = succ_begin(Entry), E = succ_end(Entry); I != E; ++I)
- BBWorklist.push_back(*I);
+ BBWorklist.append(succ_begin(Entry), succ_end(Entry));
} while (!BBWorklist.empty());
return false;
/// CoerceAvailableValueToLoadType will succeed.
static bool CanCoerceMustAliasedValueToLoad(Value *StoredVal,
Type *LoadTy,
- const TargetData &TD) {
+ const DataLayout &DL) {
// If the loaded or stored value is an first class array or struct, don't try
// to transform them. We need to be able to bitcast to integer.
if (LoadTy->isStructTy() || LoadTy->isArrayTy() ||
StoredVal->getType()->isStructTy() ||
StoredVal->getType()->isArrayTy())
return false;
-
+
// The store has to be at least as big as the load.
- if (TD.getTypeSizeInBits(StoredVal->getType()) <
- TD.getTypeSizeInBits(LoadTy))
+ if (DL.getTypeSizeInBits(StoredVal->getType()) <
+ DL.getTypeSizeInBits(LoadTy))
return false;
-
+
return true;
}
-
/// CoerceAvailableValueToLoadType - If we saw a store of a value to memory, and
/// then a load from a must-aliased pointer of a different type, try to coerce
/// InsertPt is the place to insert new instructions.
///
/// If we can't do it, return null.
-static Value *CoerceAvailableValueToLoadType(Value *StoredVal,
+static Value *CoerceAvailableValueToLoadType(Value *StoredVal,
Type *LoadedTy,
Instruction *InsertPt,
- const TargetData &TD) {
- if (!CanCoerceMustAliasedValueToLoad(StoredVal, LoadedTy, TD))
+ const DataLayout &DL) {
+ if (!CanCoerceMustAliasedValueToLoad(StoredVal, LoadedTy, DL))
return 0;
-
+
// If this is already the right type, just return it.
Type *StoredValTy = StoredVal->getType();
-
- uint64_t StoreSize = TD.getTypeSizeInBits(StoredValTy);
- uint64_t LoadSize = TD.getTypeSizeInBits(LoadedTy);
-
+
+ uint64_t StoreSize = DL.getTypeSizeInBits(StoredValTy);
+ uint64_t LoadSize = DL.getTypeSizeInBits(LoadedTy);
+
// If the store and reload are the same size, we can always reuse it.
if (StoreSize == LoadSize) {
// Pointer to Pointer -> use bitcast.
- if (StoredValTy->isPointerTy() && LoadedTy->isPointerTy())
+ if (StoredValTy->getScalarType()->isPointerTy() &&
+ LoadedTy->getScalarType()->isPointerTy())
return new BitCastInst(StoredVal, LoadedTy, "", InsertPt);
-
+
// Convert source pointers to integers, which can be bitcast.
- if (StoredValTy->isPointerTy()) {
- StoredValTy = TD.getIntPtrType(StoredValTy->getContext());
+ if (StoredValTy->getScalarType()->isPointerTy()) {
+ StoredValTy = DL.getIntPtrType(StoredValTy);
StoredVal = new PtrToIntInst(StoredVal, StoredValTy, "", InsertPt);
}
-
+
Type *TypeToCastTo = LoadedTy;
- if (TypeToCastTo->isPointerTy())
- TypeToCastTo = TD.getIntPtrType(StoredValTy->getContext());
-
+ if (TypeToCastTo->getScalarType()->isPointerTy())
+ TypeToCastTo = DL.getIntPtrType(TypeToCastTo);
+
if (StoredValTy != TypeToCastTo)
StoredVal = new BitCastInst(StoredVal, TypeToCastTo, "", InsertPt);
-
+
// Cast to pointer if the load needs a pointer type.
- if (LoadedTy->isPointerTy())
+ if (LoadedTy->getScalarType()->isPointerTy())
StoredVal = new IntToPtrInst(StoredVal, LoadedTy, "", InsertPt);
-
+
return StoredVal;
}
-
+
// If the loaded value is smaller than the available value, then we can
// extract out a piece from it. If the available value is too small, then we
// can't do anything.
assert(StoreSize >= LoadSize && "CanCoerceMustAliasedValueToLoad fail");
-
+
// Convert source pointers to integers, which can be manipulated.
- if (StoredValTy->isPointerTy()) {
- StoredValTy = TD.getIntPtrType(StoredValTy->getContext());
+ if (StoredValTy->getScalarType()->isPointerTy()) {
+ StoredValTy = DL.getIntPtrType(StoredValTy);
StoredVal = new PtrToIntInst(StoredVal, StoredValTy, "", InsertPt);
}
-
+
// Convert vectors and fp to integer, which can be manipulated.
if (!StoredValTy->isIntegerTy()) {
StoredValTy = IntegerType::get(StoredValTy->getContext(), StoreSize);
StoredVal = new BitCastInst(StoredVal, StoredValTy, "", InsertPt);
}
-
+
// If this is a big-endian system, we need to shift the value down to the low
// bits so that a truncate will work.
- if (TD.isBigEndian()) {
+ if (DL.isBigEndian()) {
Constant *Val = ConstantInt::get(StoredVal->getType(), StoreSize-LoadSize);
StoredVal = BinaryOperator::CreateLShr(StoredVal, Val, "tmp", InsertPt);
}
-
+
// Truncate the integer to the right size now.
Type *NewIntTy = IntegerType::get(StoredValTy->getContext(), LoadSize);
StoredVal = new TruncInst(StoredVal, NewIntTy, "trunc", InsertPt);
-
+
if (LoadedTy == NewIntTy)
return StoredVal;
-
+
// If the result is a pointer, inttoptr.
- if (LoadedTy->isPointerTy())
+ if (LoadedTy->getScalarType()->isPointerTy())
return new IntToPtrInst(StoredVal, LoadedTy, "inttoptr", InsertPt);
-
+
// Otherwise, bitcast.
return new BitCastInst(StoredVal, LoadedTy, "bitcast", InsertPt);
}
static int AnalyzeLoadFromClobberingWrite(Type *LoadTy, Value *LoadPtr,
Value *WritePtr,
uint64_t WriteSizeInBits,
- const TargetData &TD) {
+ const DataLayout &DL) {
// If the loaded or stored value is a first class array or struct, don't try
// to transform them. We need to be able to bitcast to integer.
if (LoadTy->isStructTy() || LoadTy->isArrayTy())
return -1;
-
+
int64_t StoreOffset = 0, LoadOffset = 0;
- Value *StoreBase = GetPointerBaseWithConstantOffset(WritePtr, StoreOffset,TD);
- Value *LoadBase = GetPointerBaseWithConstantOffset(LoadPtr, LoadOffset, TD);
+ Value *StoreBase = GetPointerBaseWithConstantOffset(WritePtr,StoreOffset,&DL);
+ Value *LoadBase = GetPointerBaseWithConstantOffset(LoadPtr, LoadOffset, &DL);
if (StoreBase != LoadBase)
return -1;
-
+
// If the load and store are to the exact same address, they should have been
// a must alias. AA must have gotten confused.
// FIXME: Study to see if/when this happens. One case is forwarding a memset
abort();
}
#endif
-
+
// If the load and store don't overlap at all, the store doesn't provide
// anything to the load. In this case, they really don't alias at all, AA
// must have gotten confused.
- uint64_t LoadSize = TD.getTypeSizeInBits(LoadTy);
-
+ uint64_t LoadSize = DL.getTypeSizeInBits(LoadTy);
+
if ((WriteSizeInBits & 7) | (LoadSize & 7))
return -1;
uint64_t StoreSize = WriteSizeInBits >> 3; // Convert to bytes.
LoadSize >>= 3;
-
-
+
+
bool isAAFailure = false;
if (StoreOffset < LoadOffset)
isAAFailure = StoreOffset+int64_t(StoreSize) <= LoadOffset;
#endif
return -1;
}
-
+
// If the Load isn't completely contained within the stored bits, we don't
// have all the bits to feed it. We could do something crazy in the future
// (issue a smaller load then merge the bits in) but this seems unlikely to be
if (StoreOffset > LoadOffset ||
StoreOffset+StoreSize < LoadOffset+LoadSize)
return -1;
-
+
// Okay, we can do this transformation. Return the number of bytes into the
// store that the load is.
return LoadOffset-StoreOffset;
-}
+}
/// AnalyzeLoadFromClobberingStore - This function is called when we have a
/// memdep query of a load that ends up being a clobbering store.
static int AnalyzeLoadFromClobberingStore(Type *LoadTy, Value *LoadPtr,
StoreInst *DepSI,
- const TargetData &TD) {
+ const DataLayout &DL) {
// Cannot handle reading from store of first-class aggregate yet.
if (DepSI->getValueOperand()->getType()->isStructTy() ||
DepSI->getValueOperand()->getType()->isArrayTy())
return -1;
Value *StorePtr = DepSI->getPointerOperand();
- uint64_t StoreSize =TD.getTypeSizeInBits(DepSI->getValueOperand()->getType());
+ uint64_t StoreSize =DL.getTypeSizeInBits(DepSI->getValueOperand()->getType());
return AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr,
- StorePtr, StoreSize, TD);
+ StorePtr, StoreSize, DL);
}
/// AnalyzeLoadFromClobberingLoad - This function is called when we have a
/// memdep query of a load that ends up being clobbered by another load. See if
/// the other load can feed into the second load.
static int AnalyzeLoadFromClobberingLoad(Type *LoadTy, Value *LoadPtr,
- LoadInst *DepLI, const TargetData &TD){
+ LoadInst *DepLI, const DataLayout &DL){
// Cannot handle reading from store of first-class aggregate yet.
if (DepLI->getType()->isStructTy() || DepLI->getType()->isArrayTy())
return -1;
-
+
Value *DepPtr = DepLI->getPointerOperand();
- uint64_t DepSize = TD.getTypeSizeInBits(DepLI->getType());
- int R = AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, DepPtr, DepSize, TD);
+ uint64_t DepSize = DL.getTypeSizeInBits(DepLI->getType());
+ int R = AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, DepPtr, DepSize, DL);
if (R != -1) return R;
-
+
// If we have a load/load clobber an DepLI can be widened to cover this load,
// then we should widen it!
int64_t LoadOffs = 0;
const Value *LoadBase =
- GetPointerBaseWithConstantOffset(LoadPtr, LoadOffs, TD);
- unsigned LoadSize = TD.getTypeStoreSize(LoadTy);
-
+ GetPointerBaseWithConstantOffset(LoadPtr, LoadOffs, &DL);
+ unsigned LoadSize = DL.getTypeStoreSize(LoadTy);
+
unsigned Size = MemoryDependenceAnalysis::
- getLoadLoadClobberFullWidthSize(LoadBase, LoadOffs, LoadSize, DepLI, TD);
+ getLoadLoadClobberFullWidthSize(LoadBase, LoadOffs, LoadSize, DepLI, DL);
if (Size == 0) return -1;
-
- return AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, DepPtr, Size*8, TD);
+
+ return AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, DepPtr, Size*8, DL);
}
static int AnalyzeLoadFromClobberingMemInst(Type *LoadTy, Value *LoadPtr,
MemIntrinsic *MI,
- const TargetData &TD) {
+ const DataLayout &DL) {
// If the mem operation is a non-constant size, we can't handle it.
ConstantInt *SizeCst = dyn_cast<ConstantInt>(MI->getLength());
if (SizeCst == 0) return -1;
// of the memset..
if (MI->getIntrinsicID() == Intrinsic::memset)
return AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, MI->getDest(),
- MemSizeInBits, TD);
-
+ MemSizeInBits, DL);
+
// If we have a memcpy/memmove, the only case we can handle is if this is a
// copy from constant memory. In that case, we can read directly from the
// constant memory.
MemTransferInst *MTI = cast<MemTransferInst>(MI);
-
+
Constant *Src = dyn_cast<Constant>(MTI->getSource());
if (Src == 0) return -1;
-
- GlobalVariable *GV = dyn_cast<GlobalVariable>(GetUnderlyingObject(Src, &TD));
+
+ GlobalVariable *GV = dyn_cast<GlobalVariable>(GetUnderlyingObject(Src, &DL));
if (GV == 0 || !GV->isConstant()) return -1;
-
+
// See if the access is within the bounds of the transfer.
int Offset = AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr,
- MI->getDest(), MemSizeInBits, TD);
+ MI->getDest(), MemSizeInBits, DL);
if (Offset == -1)
return Offset;
-
+
+ unsigned AS = Src->getType()->getPointerAddressSpace();
// Otherwise, see if we can constant fold a load from the constant with the
// offset applied as appropriate.
Src = ConstantExpr::getBitCast(Src,
- llvm::Type::getInt8PtrTy(Src->getContext()));
- Constant *OffsetCst =
+ Type::getInt8PtrTy(Src->getContext(), AS));
+ Constant *OffsetCst =
ConstantInt::get(Type::getInt64Ty(Src->getContext()), (unsigned)Offset);
Src = ConstantExpr::getGetElementPtr(Src, OffsetCst);
- Src = ConstantExpr::getBitCast(Src, PointerType::getUnqual(LoadTy));
- if (ConstantFoldLoadFromConstPtr(Src, &TD))
+ Src = ConstantExpr::getBitCast(Src, PointerType::get(LoadTy, AS));
+ if (ConstantFoldLoadFromConstPtr(Src, &DL))
return Offset;
return -1;
}
-
+
/// GetStoreValueForLoad - This function is called when we have a
/// memdep query of a load that ends up being a clobbering store. This means
/// before we give up.
static Value *GetStoreValueForLoad(Value *SrcVal, unsigned Offset,
Type *LoadTy,
- Instruction *InsertPt, const TargetData &TD){
+ Instruction *InsertPt, const DataLayout &DL){
LLVMContext &Ctx = SrcVal->getType()->getContext();
-
- uint64_t StoreSize = (TD.getTypeSizeInBits(SrcVal->getType()) + 7) / 8;
- uint64_t LoadSize = (TD.getTypeSizeInBits(LoadTy) + 7) / 8;
-
+
+ uint64_t StoreSize = (DL.getTypeSizeInBits(SrcVal->getType()) + 7) / 8;
+ uint64_t LoadSize = (DL.getTypeSizeInBits(LoadTy) + 7) / 8;
+
IRBuilder<> Builder(InsertPt->getParent(), InsertPt);
-
+
// Compute which bits of the stored value are being used by the load. Convert
// to an integer type to start with.
- if (SrcVal->getType()->isPointerTy())
- SrcVal = Builder.CreatePtrToInt(SrcVal, TD.getIntPtrType(Ctx));
+ if (SrcVal->getType()->getScalarType()->isPointerTy())
+ SrcVal = Builder.CreatePtrToInt(SrcVal,
+ DL.getIntPtrType(SrcVal->getType()));
if (!SrcVal->getType()->isIntegerTy())
SrcVal = Builder.CreateBitCast(SrcVal, IntegerType::get(Ctx, StoreSize*8));
-
+
// Shift the bits to the least significant depending on endianness.
unsigned ShiftAmt;
- if (TD.isLittleEndian())
+ if (DL.isLittleEndian())
ShiftAmt = Offset*8;
else
ShiftAmt = (StoreSize-LoadSize-Offset)*8;
-
+
if (ShiftAmt)
SrcVal = Builder.CreateLShr(SrcVal, ShiftAmt);
-
+
if (LoadSize != StoreSize)
SrcVal = Builder.CreateTrunc(SrcVal, IntegerType::get(Ctx, LoadSize*8));
-
- return CoerceAvailableValueToLoadType(SrcVal, LoadTy, InsertPt, TD);
+
+ return CoerceAvailableValueToLoadType(SrcVal, LoadTy, InsertPt, DL);
}
/// GetLoadValueForLoad - This function is called when we have a
static Value *GetLoadValueForLoad(LoadInst *SrcVal, unsigned Offset,
Type *LoadTy, Instruction *InsertPt,
GVN &gvn) {
- const TargetData &TD = *gvn.getTargetData();
+ const DataLayout &DL = *gvn.getDataLayout();
// If Offset+LoadTy exceeds the size of SrcVal, then we must be wanting to
// widen SrcVal out to a larger load.
- unsigned SrcValSize = TD.getTypeStoreSize(SrcVal->getType());
- unsigned LoadSize = TD.getTypeStoreSize(LoadTy);
+ unsigned SrcValSize = DL.getTypeStoreSize(SrcVal->getType());
+ unsigned LoadSize = DL.getTypeStoreSize(LoadTy);
if (Offset+LoadSize > SrcValSize) {
assert(SrcVal->isSimple() && "Cannot widen volatile/atomic load!");
assert(SrcVal->getType()->isIntegerTy() && "Can't widen non-integer load");
NewLoadSize = NextPowerOf2(NewLoadSize);
Value *PtrVal = SrcVal->getPointerOperand();
-
+
// Insert the new load after the old load. This ensures that subsequent
// memdep queries will find the new load. We can't easily remove the old
// load completely because it is already in the value numbering table.
IRBuilder<> Builder(SrcVal->getParent(), ++BasicBlock::iterator(SrcVal));
- Type *DestPTy =
+ Type *DestPTy =
IntegerType::get(LoadTy->getContext(), NewLoadSize*8);
- DestPTy = PointerType::get(DestPTy,
- cast<PointerType>(PtrVal->getType())->getAddressSpace());
+ DestPTy = PointerType::get(DestPTy,
+ PtrVal->getType()->getPointerAddressSpace());
Builder.SetCurrentDebugLocation(SrcVal->getDebugLoc());
PtrVal = Builder.CreateBitCast(PtrVal, DestPTy);
LoadInst *NewLoad = Builder.CreateLoad(PtrVal);
DEBUG(dbgs() << "GVN WIDENED LOAD: " << *SrcVal << "\n");
DEBUG(dbgs() << "TO: " << *NewLoad << "\n");
-
+
// Replace uses of the original load with the wider load. On a big endian
// system, we need to shift down to get the relevant bits.
Value *RV = NewLoad;
- if (TD.isBigEndian())
+ if (DL.isBigEndian())
RV = Builder.CreateLShr(RV,
NewLoadSize*8-SrcVal->getType()->getPrimitiveSizeInBits());
RV = Builder.CreateTrunc(RV, SrcVal->getType());
SrcVal->replaceAllUsesWith(RV);
-
+
// We would like to use gvn.markInstructionForDeletion here, but we can't
// because the load is already memoized into the leader map table that GVN
// tracks. It is potentially possible to remove the load from the table,
gvn.getMemDep().removeInstruction(SrcVal);
SrcVal = NewLoad;
}
-
- return GetStoreValueForLoad(SrcVal, Offset, LoadTy, InsertPt, TD);
+
+ return GetStoreValueForLoad(SrcVal, Offset, LoadTy, InsertPt, DL);
}
/// memdep query of a load that ends up being a clobbering mem intrinsic.
static Value *GetMemInstValueForLoad(MemIntrinsic *SrcInst, unsigned Offset,
Type *LoadTy, Instruction *InsertPt,
- const TargetData &TD){
+ const DataLayout &DL){
LLVMContext &Ctx = LoadTy->getContext();
- uint64_t LoadSize = TD.getTypeSizeInBits(LoadTy)/8;
+ uint64_t LoadSize = DL.getTypeSizeInBits(LoadTy)/8;
IRBuilder<> Builder(InsertPt->getParent(), InsertPt);
-
+
// We know that this method is only called when the mem transfer fully
// provides the bits for the load.
if (MemSetInst *MSI = dyn_cast<MemSetInst>(SrcInst)) {
Value *Val = MSI->getValue();
if (LoadSize != 1)
Val = Builder.CreateZExt(Val, IntegerType::get(Ctx, LoadSize*8));
-
+
Value *OneElt = Val;
-
+
// Splat the value out to the right number of bits.
for (unsigned NumBytesSet = 1; NumBytesSet != LoadSize; ) {
// If we can double the number of bytes set, do it.
NumBytesSet <<= 1;
continue;
}
-
+
// Otherwise insert one byte at a time.
Value *ShVal = Builder.CreateShl(Val, 1*8);
Val = Builder.CreateOr(OneElt, ShVal);
++NumBytesSet;
}
-
- return CoerceAvailableValueToLoadType(Val, LoadTy, InsertPt, TD);
+
+ return CoerceAvailableValueToLoadType(Val, LoadTy, InsertPt, DL);
}
-
+
// Otherwise, this is a memcpy/memmove from a constant global.
MemTransferInst *MTI = cast<MemTransferInst>(SrcInst);
Constant *Src = cast<Constant>(MTI->getSource());
+ unsigned AS = Src->getType()->getPointerAddressSpace();
// Otherwise, see if we can constant fold a load from the constant with the
// offset applied as appropriate.
Src = ConstantExpr::getBitCast(Src,
- llvm::Type::getInt8PtrTy(Src->getContext()));
- Constant *OffsetCst =
- ConstantInt::get(Type::getInt64Ty(Src->getContext()), (unsigned)Offset);
+ Type::getInt8PtrTy(Src->getContext(), AS));
+ Constant *OffsetCst =
+ ConstantInt::get(Type::getInt64Ty(Src->getContext()), (unsigned)Offset);
Src = ConstantExpr::getGetElementPtr(Src, OffsetCst);
- Src = ConstantExpr::getBitCast(Src, PointerType::getUnqual(LoadTy));
- return ConstantFoldLoadFromConstPtr(Src, &TD);
+ Src = ConstantExpr::getBitCast(Src, PointerType::get(LoadTy, AS));
+ return ConstantFoldLoadFromConstPtr(Src, &DL);
}
-namespace {
-
-struct AvailableValueInBlock {
- /// BB - The basic block in question.
- BasicBlock *BB;
- enum ValType {
- SimpleVal, // A simple offsetted value that is accessed.
- LoadVal, // A value produced by a load.
- MemIntrin // A memory intrinsic which is loaded from.
- };
-
- /// V - The value that is live out of the block.
- PointerIntPair<Value *, 2, ValType> Val;
-
- /// Offset - The byte offset in Val that is interesting for the load query.
- unsigned Offset;
-
- static AvailableValueInBlock get(BasicBlock *BB, Value *V,
- unsigned Offset = 0) {
- AvailableValueInBlock Res;
- Res.BB = BB;
- Res.Val.setPointer(V);
- Res.Val.setInt(SimpleVal);
- Res.Offset = Offset;
- return Res;
- }
-
- static AvailableValueInBlock getMI(BasicBlock *BB, MemIntrinsic *MI,
- unsigned Offset = 0) {
- AvailableValueInBlock Res;
- Res.BB = BB;
- Res.Val.setPointer(MI);
- Res.Val.setInt(MemIntrin);
- Res.Offset = Offset;
- return Res;
- }
-
- static AvailableValueInBlock getLoad(BasicBlock *BB, LoadInst *LI,
- unsigned Offset = 0) {
- AvailableValueInBlock Res;
- Res.BB = BB;
- Res.Val.setPointer(LI);
- Res.Val.setInt(LoadVal);
- Res.Offset = Offset;
- return Res;
- }
-
- bool isSimpleValue() const { return Val.getInt() == SimpleVal; }
- bool isCoercedLoadValue() const { return Val.getInt() == LoadVal; }
- bool isMemIntrinValue() const { return Val.getInt() == MemIntrin; }
-
- Value *getSimpleValue() const {
- assert(isSimpleValue() && "Wrong accessor");
- return Val.getPointer();
- }
-
- LoadInst *getCoercedLoadValue() const {
- assert(isCoercedLoadValue() && "Wrong accessor");
- return cast<LoadInst>(Val.getPointer());
- }
-
- MemIntrinsic *getMemIntrinValue() const {
- assert(isMemIntrinValue() && "Wrong accessor");
- return cast<MemIntrinsic>(Val.getPointer());
- }
-
- /// MaterializeAdjustedValue - Emit code into this block to adjust the value
- /// defined here to the specified type. This handles various coercion cases.
- Value *MaterializeAdjustedValue(Type *LoadTy, GVN &gvn) const {
- Value *Res;
- if (isSimpleValue()) {
- Res = getSimpleValue();
- if (Res->getType() != LoadTy) {
- const TargetData *TD = gvn.getTargetData();
- assert(TD && "Need target data to handle type mismatch case");
- Res = GetStoreValueForLoad(Res, Offset, LoadTy, BB->getTerminator(),
- *TD);
-
- DEBUG(dbgs() << "GVN COERCED NONLOCAL VAL:\nOffset: " << Offset << " "
- << *getSimpleValue() << '\n'
- << *Res << '\n' << "\n\n\n");
- }
- } else if (isCoercedLoadValue()) {
- LoadInst *Load = getCoercedLoadValue();
- if (Load->getType() == LoadTy && Offset == 0) {
- Res = Load;
- } else {
- Res = GetLoadValueForLoad(Load, Offset, LoadTy, BB->getTerminator(),
- gvn);
-
- DEBUG(dbgs() << "GVN COERCED NONLOCAL LOAD:\nOffset: " << Offset << " "
- << *getCoercedLoadValue() << '\n'
- << *Res << '\n' << "\n\n\n");
- }
- } else {
- const TargetData *TD = gvn.getTargetData();
- assert(TD && "Need target data to handle type mismatch case");
- Res = GetMemInstValueForLoad(getMemIntrinValue(), Offset,
- LoadTy, BB->getTerminator(), *TD);
- DEBUG(dbgs() << "GVN COERCED NONLOCAL MEM INTRIN:\nOffset: " << Offset
- << " " << *getMemIntrinValue() << '\n'
- << *Res << '\n' << "\n\n\n");
- }
- return Res;
- }
-};
-
-} // end anonymous namespace
/// ConstructSSAForLoadSet - Given a set of loads specified by ValuesPerBlock,
/// construct SSA form, allowing us to eliminate LI. This returns the value
/// that should be used at LI's definition site.
-static Value *ConstructSSAForLoadSet(LoadInst *LI,
+static Value *ConstructSSAForLoadSet(LoadInst *LI,
SmallVectorImpl<AvailableValueInBlock> &ValuesPerBlock,
GVN &gvn) {
// Check for the fully redundant, dominating load case. In this case, we can
// just use the dominating value directly.
- if (ValuesPerBlock.size() == 1 &&
+ if (ValuesPerBlock.size() == 1 &&
gvn.getDominatorTree().properlyDominates(ValuesPerBlock[0].BB,
- LI->getParent()))
+ LI->getParent())) {
+ assert(!ValuesPerBlock[0].isUndefValue() && "Dead BB dominate this block");
return ValuesPerBlock[0].MaterializeAdjustedValue(LI->getType(), gvn);
+ }
// Otherwise, we have to construct SSA form.
SmallVector<PHINode*, 8> NewPHIs;
SSAUpdater SSAUpdate(&NewPHIs);
SSAUpdate.Initialize(LI->getType(), LI->getName());
-
+
Type *LoadTy = LI->getType();
-
+
for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i) {
const AvailableValueInBlock &AV = ValuesPerBlock[i];
BasicBlock *BB = AV.BB;
-
+
if (SSAUpdate.HasValueForBlock(BB))
continue;
SSAUpdate.AddAvailableValue(BB, AV.MaterializeAdjustedValue(LoadTy, gvn));
}
-
+
// Perform PHI construction.
Value *V = SSAUpdate.GetValueInMiddleOfBlock(LI->getParent());
-
+
// If new PHI nodes were created, notify alias analysis.
- if (V->getType()->isPointerTy()) {
+ if (V->getType()->getScalarType()->isPointerTy()) {
AliasAnalysis *AA = gvn.getAliasAnalysis();
-
+
for (unsigned i = 0, e = NewPHIs.size(); i != e; ++i)
AA->copyValue(LI, NewPHIs[i]);
-
+
// Now that we've copied information to the new PHIs, scan through
// them again and inform alias analysis that we've added potentially
// escaping uses to any values that are operands to these PHIs.
return V;
}
+Value *AvailableValueInBlock::MaterializeAdjustedValue(Type *LoadTy, GVN &gvn) const {
+ Value *Res;
+ if (isSimpleValue()) {
+ Res = getSimpleValue();
+ if (Res->getType() != LoadTy) {
+ const DataLayout *DL = gvn.getDataLayout();
+ assert(DL && "Need target data to handle type mismatch case");
+ Res = GetStoreValueForLoad(Res, Offset, LoadTy, BB->getTerminator(),
+ *DL);
+
+ DEBUG(dbgs() << "GVN COERCED NONLOCAL VAL:\nOffset: " << Offset << " "
+ << *getSimpleValue() << '\n'
+ << *Res << '\n' << "\n\n\n");
+ }
+ } else if (isCoercedLoadValue()) {
+ LoadInst *Load = getCoercedLoadValue();
+ if (Load->getType() == LoadTy && Offset == 0) {
+ Res = Load;
+ } else {
+ Res = GetLoadValueForLoad(Load, Offset, LoadTy, BB->getTerminator(),
+ gvn);
+
+ DEBUG(dbgs() << "GVN COERCED NONLOCAL LOAD:\nOffset: " << Offset << " "
+ << *getCoercedLoadValue() << '\n'
+ << *Res << '\n' << "\n\n\n");
+ }
+ } else if (isMemIntrinValue()) {
+ const DataLayout *DL = gvn.getDataLayout();
+ assert(DL && "Need target data to handle type mismatch case");
+ Res = GetMemInstValueForLoad(getMemIntrinValue(), Offset,
+ LoadTy, BB->getTerminator(), *DL);
+ DEBUG(dbgs() << "GVN COERCED NONLOCAL MEM INTRIN:\nOffset: " << Offset
+ << " " << *getMemIntrinValue() << '\n'
+ << *Res << '\n' << "\n\n\n");
+ } else {
+ assert(isUndefValue() && "Should be UndefVal");
+ DEBUG(dbgs() << "GVN COERCED NONLOCAL Undef:\n";);
+ return UndefValue::get(LoadTy);
+ }
+ return Res;
+}
+
static bool isLifetimeStart(const Instruction *Inst) {
if (const IntrinsicInst* II = dyn_cast<IntrinsicInst>(Inst))
return II->getIntrinsicID() == Intrinsic::lifetime_start;
return false;
}
-/// processNonLocalLoad - Attempt to eliminate a load whose dependencies are
-/// non-local by performing PHI construction.
-bool GVN::processNonLocalLoad(LoadInst *LI) {
- // Find the non-local dependencies of the load.
- SmallVector<NonLocalDepResult, 64> Deps;
- AliasAnalysis::Location Loc = VN.getAliasAnalysis()->getLocation(LI);
- MD->getNonLocalPointerDependency(Loc, true, LI->getParent(), Deps);
- //DEBUG(dbgs() << "INVESTIGATING NONLOCAL LOAD: "
- // << Deps.size() << *LI << '\n');
-
- // 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.
- unsigned NumDeps = Deps.size();
- if (NumDeps > 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 (NumDeps == 1 &&
- !Deps[0].getResult().isDef() && !Deps[0].getResult().isClobber()) {
- DEBUG(
- dbgs() << "GVN: non-local load ";
- WriteAsOperand(dbgs(), LI);
- dbgs() << " has unknown dependencies\n";
- );
- return false;
- }
+void GVN::AnalyzeLoadAvailability(LoadInst *LI, LoadDepVect &Deps,
+ AvailValInBlkVect &ValuesPerBlock,
+ UnavailBlkVect &UnavailableBlocks) {
// 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
// dependencies that produce an unknown value for the load (such as a call
// that could potentially clobber the load).
- SmallVector<AvailableValueInBlock, 64> ValuesPerBlock;
- SmallVector<BasicBlock*, 64> UnavailableBlocks;
-
+ unsigned NumDeps = Deps.size();
for (unsigned i = 0, e = NumDeps; i != e; ++i) {
BasicBlock *DepBB = Deps[i].getBB();
MemDepResult DepInfo = Deps[i].getResult();
+ if (DeadBlocks.count(DepBB)) {
+ // Dead dependent mem-op disguise as a load evaluating the same value
+ // as the load in question.
+ ValuesPerBlock.push_back(AvailableValueInBlock::getUndef(DepBB));
+ continue;
+ }
+
if (!DepInfo.isDef() && !DepInfo.isClobber()) {
UnavailableBlocks.push_back(DepBB);
continue;
// the pointer operand of the load if PHI translation occurs. Make sure
// to consider the right address.
Value *Address = Deps[i].getAddress();
-
+
// If the dependence is to a store that writes to a superset of the bits
// read by the load, we can extract the bits we need for the load from the
// stored value.
if (StoreInst *DepSI = dyn_cast<StoreInst>(DepInfo.getInst())) {
- if (TD && Address) {
+ if (DL && Address) {
int Offset = AnalyzeLoadFromClobberingStore(LI->getType(), Address,
- DepSI, *TD);
+ DepSI, *DL);
if (Offset != -1) {
ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB,
DepSI->getValueOperand(),
}
}
}
-
+
// Check to see if we have something like this:
// load i32* P
// load i8* (P+1)
if (LoadInst *DepLI = dyn_cast<LoadInst>(DepInfo.getInst())) {
// If this is a clobber and L is the first instruction in its block, then
// we have the first instruction in the entry block.
- if (DepLI != LI && Address && TD) {
+ if (DepLI != LI && Address && DL) {
int Offset = AnalyzeLoadFromClobberingLoad(LI->getType(),
LI->getPointerOperand(),
- DepLI, *TD);
-
+ DepLI, *DL);
+
if (Offset != -1) {
ValuesPerBlock.push_back(AvailableValueInBlock::getLoad(DepBB,DepLI,
Offset));
// If the clobbering value is a memset/memcpy/memmove, see if we can
// forward a value on from it.
if (MemIntrinsic *DepMI = dyn_cast<MemIntrinsic>(DepInfo.getInst())) {
- if (TD && Address) {
+ if (DL && Address) {
int Offset = AnalyzeLoadFromClobberingMemInst(LI->getType(), Address,
- DepMI, *TD);
+ DepMI, *DL);
if (Offset != -1) {
ValuesPerBlock.push_back(AvailableValueInBlock::getMI(DepBB, DepMI,
Offset));
continue;
- }
+ }
}
}
-
+
UnavailableBlocks.push_back(DepBB);
continue;
}
Instruction *DepInst = DepInfo.getInst();
// Loading the allocation -> undef.
- if (isa<AllocaInst>(DepInst) || isMallocLikeFn(DepInst) ||
+ if (isa<AllocaInst>(DepInst) || isMallocLikeFn(DepInst, TLI) ||
// Loading immediately after lifetime begin -> undef.
isLifetimeStart(DepInst)) {
ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB,
UndefValue::get(LI->getType())));
continue;
}
-
+
if (StoreInst *S = dyn_cast<StoreInst>(DepInst)) {
// Reject loads and stores that are to the same address but are of
// different types if we have to.
if (S->getValueOperand()->getType() != LI->getType()) {
// If the stored value is larger or equal to the loaded value, we can
// reuse it.
- if (TD == 0 || !CanCoerceMustAliasedValueToLoad(S->getValueOperand(),
- LI->getType(), *TD)) {
+ if (DL == 0 || !CanCoerceMustAliasedValueToLoad(S->getValueOperand(),
+ LI->getType(), *DL)) {
UnavailableBlocks.push_back(DepBB);
continue;
}
S->getValueOperand()));
continue;
}
-
+
if (LoadInst *LD = dyn_cast<LoadInst>(DepInst)) {
// If the types mismatch and we can't handle it, reject reuse of the load.
if (LD->getType() != LI->getType()) {
// If the stored value is larger or equal to the loaded value, we can
// reuse it.
- if (TD == 0 || !CanCoerceMustAliasedValueToLoad(LD, LI->getType(),*TD)){
+ if (DL == 0 || !CanCoerceMustAliasedValueToLoad(LD, LI->getType(),*DL)){
UnavailableBlocks.push_back(DepBB);
continue;
- }
+ }
}
ValuesPerBlock.push_back(AvailableValueInBlock::getLoad(DepBB, LD));
continue;
}
-
- UnavailableBlocks.push_back(DepBB);
- continue;
- }
-
- // If we have no predecessors that produce a known value for this load, exit
- // early.
- if (ValuesPerBlock.empty()) return false;
- // If all of the instructions we depend on produce a known value for this
- // load, then it is fully redundant and we can use PHI insertion to compute
- // its value. Insert PHIs and remove the fully redundant value now.
- if (UnavailableBlocks.empty()) {
- DEBUG(dbgs() << "GVN REMOVING NONLOCAL LOAD: " << *LI << '\n');
-
- // Perform PHI construction.
- Value *V = ConstructSSAForLoadSet(LI, ValuesPerBlock, *this);
- LI->replaceAllUsesWith(V);
-
- if (isa<PHINode>(V))
- V->takeName(LI);
- if (V->getType()->isPointerTy())
- MD->invalidateCachedPointerInfo(V);
- markInstructionForDeletion(LI);
- ++NumGVNLoad;
- return true;
+ UnavailableBlocks.push_back(DepBB);
}
+}
- if (!EnablePRE || !EnableLoadPRE)
- return false;
-
+bool GVN::PerformLoadPRE(LoadInst *LI, AvailValInBlkVect &ValuesPerBlock,
+ UnavailBlkVect &UnavailableBlocks) {
// Okay, we have *some* definitions of the value. This means that the value
// is available in some of our (transitive) predecessors. Lets think about
// doing PRE of this load. This will involve inserting a new load into the
BasicBlock *LoadBB = LI->getParent();
BasicBlock *TmpBB = LoadBB;
- bool isSinglePred = false;
- bool allSingleSucc = true;
while (TmpBB->getSinglePredecessor()) {
- isSinglePred = true;
TmpBB = TmpBB->getSinglePredecessor();
if (TmpBB == LoadBB) // Infinite (unreachable) loop.
return false;
if (Blockers.count(TmpBB))
return false;
-
+
// If any of these blocks has more than one successor (i.e. if the edge we
- // just traversed was critical), then there are other paths through this
- // block along which the load may not be anticipated. Hoisting the load
+ // just traversed was critical), then there are other paths through this
+ // block along which the load may not be anticipated. Hoisting the load
// above this block would be adding the load to execution paths along
// which it was not previously executed.
if (TmpBB->getTerminator()->getNumSuccessors() != 1)
assert(TmpBB);
LoadBB = TmpBB;
- // FIXME: It is extremely unclear what this loop is doing, other than
- // artificially restricting loadpre.
- if (isSinglePred) {
- bool isHot = false;
- for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i) {
- const AvailableValueInBlock &AV = ValuesPerBlock[i];
- if (AV.isSimpleValue())
- // "Hot" Instruction is in some loop (because it dominates its dep.
- // instruction).
- if (Instruction *I = dyn_cast<Instruction>(AV.getSimpleValue()))
- 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;
- }
-
// Check to see how many predecessors have the loaded value fully
// available.
DenseMap<BasicBlock*, Value*> PredLoads;
for (unsigned i = 0, e = UnavailableBlocks.size(); i != e; ++i)
FullyAvailableBlocks[UnavailableBlocks[i]] = false;
- SmallVector<std::pair<TerminatorInst*, unsigned>, 4> NeedToSplit;
+ SmallVector<BasicBlock *, 4> CriticalEdgePred;
for (pred_iterator PI = pred_begin(LoadBB), E = pred_end(LoadBB);
PI != E; ++PI) {
BasicBlock *Pred = *PI;
return false;
}
- unsigned SuccNum = GetSuccessorNumber(Pred, LoadBB);
- NeedToSplit.push_back(std::make_pair(Pred->getTerminator(), SuccNum));
+ CriticalEdgePred.push_back(Pred);
}
}
- if (!NeedToSplit.empty()) {
- toSplit.append(NeedToSplit.begin(), NeedToSplit.end());
- return false;
- }
-
// Decide whether PRE is profitable for this load.
unsigned NumUnavailablePreds = PredLoads.size();
assert(NumUnavailablePreds != 0 &&
- "Fully available value should be eliminated above!");
-
+ "Fully available value should already be eliminated!");
+
// If this load is unavailable in multiple predecessors, reject it.
// FIXME: If we could restructure the CFG, we could make a common pred with
// all the preds that don't have an available LI and insert a new load into
if (NumUnavailablePreds != 1)
return false;
+ // Split critical edges, and update the unavailable predecessors accordingly.
+ for (SmallVectorImpl<BasicBlock *>::iterator I = CriticalEdgePred.begin(),
+ E = CriticalEdgePred.end(); I != E; I++) {
+ BasicBlock *OrigPred = *I;
+ BasicBlock *NewPred = splitCriticalEdges(OrigPred, LoadBB);
+ PredLoads.erase(OrigPred);
+ PredLoads[NewPred] = 0;
+ DEBUG(dbgs() << "Split critical edge " << OrigPred->getName() << "->"
+ << LoadBB->getName() << '\n');
+ }
+
// Check if the load can safely be moved to all the unavailable predecessors.
bool CanDoPRE = true;
SmallVector<Instruction*, 8> NewInsts;
// If all preds have a single successor, then we know it is safe to insert
// the load on the pred (?!?), so we can insert code to materialize the
// pointer if it is not available.
- PHITransAddr Address(LI->getPointerOperand(), TD);
+ PHITransAddr Address(LI->getPointerOperand(), DL);
Value *LoadPtr = 0;
- if (allSingleSucc) {
- LoadPtr = Address.PHITranslateWithInsertion(LoadBB, UnavailablePred,
- *DT, NewInsts);
- } else {
- Address.PHITranslateValue(LoadBB, UnavailablePred, DT);
- LoadPtr = Address.getAddr();
- }
+ LoadPtr = Address.PHITranslateWithInsertion(LoadBB, UnavailablePred,
+ *DT, NewInsts);
// If we couldn't find or insert a computation of this phi translated value,
// we fail PRE.
break;
}
- // 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 &&
- // FIXME: REEVALUTE THIS.
- !isSafeToLoadUnconditionally(LoadPtr,
- UnavailablePred->getTerminator(),
- LI->getAlignment(), TD)) {
- CanDoPRE = false;
- break;
- }
-
I->second = LoadPtr;
}
if (MD) MD->removeInstruction(I);
I->eraseFromParent();
}
- return false;
+ // HINT:Don't revert the edge-splitting as following transformation may
+ // also need to split these critial edges.
+ return !CriticalEdgePred.empty();
}
// Okay, we can eliminate this load by inserting a reload in the predecessor
DEBUG(if (!NewInsts.empty())
dbgs() << "INSERTED " << NewInsts.size() << " INSTS: "
<< *NewInsts.back() << '\n');
-
+
// Assign value numbers to the new instructions.
for (unsigned i = 0, e = NewInsts.size(); i != e; ++i) {
- // FIXME: We really _ought_ to insert these value numbers into their
+ // FIXME: We really _ought_ to insert these value numbers into their
// parent's availability map. However, in doing so, we risk getting into
// ordering issues. If a block hasn't been processed yet, we would be
// marking a value as AVAIL-IN, which isn't what we intend.
LI->replaceAllUsesWith(V);
if (isa<PHINode>(V))
V->takeName(LI);
- if (V->getType()->isPointerTy())
+ if (V->getType()->getScalarType()->isPointerTy())
MD->invalidateCachedPointerInfo(V);
markInstructionForDeletion(LI);
++NumPRELoad;
return true;
}
-static void patchReplacementInstruction(Value *Repl, Instruction *I) {
+/// processNonLocalLoad - Attempt to eliminate a load whose dependencies are
+/// non-local by performing PHI construction.
+bool GVN::processNonLocalLoad(LoadInst *LI) {
+ // Step 1: Find the non-local dependencies of the load.
+ LoadDepVect Deps;
+ AliasAnalysis::Location Loc = VN.getAliasAnalysis()->getLocation(LI);
+ MD->getNonLocalPointerDependency(Loc, true, LI->getParent(), Deps);
+
+ // 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.
+ unsigned NumDeps = Deps.size();
+ if (NumDeps > 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 (NumDeps == 1 &&
+ !Deps[0].getResult().isDef() && !Deps[0].getResult().isClobber()) {
+ DEBUG(
+ dbgs() << "GVN: non-local load ";
+ LI->printAsOperand(dbgs());
+ dbgs() << " has unknown dependencies\n";
+ );
+ return false;
+ }
+
+ // Step 2: Analyze the availability of the load
+ AvailValInBlkVect ValuesPerBlock;
+ UnavailBlkVect UnavailableBlocks;
+ AnalyzeLoadAvailability(LI, Deps, ValuesPerBlock, UnavailableBlocks);
+
+ // If we have no predecessors that produce a known value for this load, exit
+ // early.
+ if (ValuesPerBlock.empty())
+ return false;
+
+ // Step 3: Eliminate fully redundancy.
+ //
+ // If all of the instructions we depend on produce a known value for this
+ // load, then it is fully redundant and we can use PHI insertion to compute
+ // its value. Insert PHIs and remove the fully redundant value now.
+ if (UnavailableBlocks.empty()) {
+ DEBUG(dbgs() << "GVN REMOVING NONLOCAL LOAD: " << *LI << '\n');
+
+ // Perform PHI construction.
+ Value *V = ConstructSSAForLoadSet(LI, ValuesPerBlock, *this);
+ LI->replaceAllUsesWith(V);
+
+ if (isa<PHINode>(V))
+ V->takeName(LI);
+ if (V->getType()->getScalarType()->isPointerTy())
+ MD->invalidateCachedPointerInfo(V);
+ markInstructionForDeletion(LI);
+ ++NumGVNLoad;
+ return true;
+ }
+
+ // Step 4: Eliminate partial redundancy.
+ if (!EnablePRE || !EnableLoadPRE)
+ return false;
+
+ return PerformLoadPRE(LI, ValuesPerBlock, UnavailableBlocks);
+}
+
+
+static void patchReplacementInstruction(Instruction *I, Value *Repl) {
// Patch the replacement so that it is not more restrictive than the value
// being replaced.
BinaryOperator *Op = dyn_cast<BinaryOperator>(I);
ReplInst->setMetadata(Kind, MDNode::getMostGenericRange(IMD, ReplMD));
break;
case LLVMContext::MD_prof:
- llvm_unreachable("MD_prof in a non terminator instruction");
+ llvm_unreachable("MD_prof in a non-terminator instruction");
break;
case LLVMContext::MD_fpmath:
ReplInst->setMetadata(Kind, MDNode::getMostGenericFPMath(IMD, ReplMD));
}
}
-static void patchAndReplaceAllUsesWith(Value *Repl, Instruction *I) {
- patchReplacementInstruction(Repl, I);
+static void patchAndReplaceAllUsesWith(Instruction *I, Value *Repl) {
+ patchReplacementInstruction(I, Repl);
I->replaceAllUsesWith(Repl);
}
markInstructionForDeletion(L);
return true;
}
-
+
// ... to a pointer that has been loaded from before...
MemDepResult Dep = MD->getDependency(L);
// If we have a clobber and target data is around, see if this is a clobber
// that we can fix up through code synthesis.
- if (Dep.isClobber() && TD) {
+ if (Dep.isClobber() && DL) {
// Check to see if we have something like this:
// store i32 123, i32* %P
// %A = bitcast i32* %P to i8*
if (StoreInst *DepSI = dyn_cast<StoreInst>(Dep.getInst())) {
int Offset = AnalyzeLoadFromClobberingStore(L->getType(),
L->getPointerOperand(),
- DepSI, *TD);
+ DepSI, *DL);
if (Offset != -1)
AvailVal = GetStoreValueForLoad(DepSI->getValueOperand(), Offset,
- L->getType(), L, *TD);
+ L->getType(), L, *DL);
}
-
+
// Check to see if we have something like this:
// load i32* P
// load i8* (P+1)
// we have the first instruction in the entry block.
if (DepLI == L)
return false;
-
+
int Offset = AnalyzeLoadFromClobberingLoad(L->getType(),
L->getPointerOperand(),
- DepLI, *TD);
+ DepLI, *DL);
if (Offset != -1)
AvailVal = GetLoadValueForLoad(DepLI, Offset, L->getType(), L, *this);
}
-
+
// If the clobbering value is a memset/memcpy/memmove, see if we can forward
// a value on from it.
if (MemIntrinsic *DepMI = dyn_cast<MemIntrinsic>(Dep.getInst())) {
int Offset = AnalyzeLoadFromClobberingMemInst(L->getType(),
L->getPointerOperand(),
- DepMI, *TD);
+ DepMI, *DL);
if (Offset != -1)
- AvailVal = GetMemInstValueForLoad(DepMI, Offset, L->getType(), L, *TD);
+ AvailVal = GetMemInstValueForLoad(DepMI, Offset, L->getType(), L, *DL);
}
-
+
if (AvailVal) {
DEBUG(dbgs() << "GVN COERCED INST:\n" << *Dep.getInst() << '\n'
<< *AvailVal << '\n' << *L << "\n\n\n");
-
+
// Replace the load!
L->replaceAllUsesWith(AvailVal);
- if (AvailVal->getType()->isPointerTy())
+ if (AvailVal->getType()->getScalarType()->isPointerTy())
MD->invalidateCachedPointerInfo(AvailVal);
markInstructionForDeletion(L);
++NumGVNLoad;
return true;
}
}
-
+
// If the value isn't available, don't do anything!
if (Dep.isClobber()) {
DEBUG(
// fast print dep, using operator<< on instruction is too slow.
dbgs() << "GVN: load ";
- WriteAsOperand(dbgs(), L);
+ L->printAsOperand(dbgs());
Instruction *I = Dep.getInst();
dbgs() << " is clobbered by " << *I << '\n';
);
DEBUG(
// fast print dep, using operator<< on instruction is too slow.
dbgs() << "GVN: load ";
- WriteAsOperand(dbgs(), L);
+ L->printAsOperand(dbgs());
dbgs() << " has unknown dependence\n";
);
return false;
Instruction *DepInst = Dep.getInst();
if (StoreInst *DepSI = dyn_cast<StoreInst>(DepInst)) {
Value *StoredVal = DepSI->getValueOperand();
-
+
// The store and load are to a must-aliased pointer, but they may not
// actually have the same type. See if we know how to reuse the stored
// value (depending on its type).
if (StoredVal->getType() != L->getType()) {
- if (TD) {
+ if (DL) {
StoredVal = CoerceAvailableValueToLoadType(StoredVal, L->getType(),
- L, *TD);
+ L, *DL);
if (StoredVal == 0)
return false;
-
+
DEBUG(dbgs() << "GVN COERCED STORE:\n" << *DepSI << '\n' << *StoredVal
<< '\n' << *L << "\n\n\n");
}
- else
+ else
return false;
}
// Remove it!
L->replaceAllUsesWith(StoredVal);
- if (StoredVal->getType()->isPointerTy())
+ if (StoredVal->getType()->getScalarType()->isPointerTy())
MD->invalidateCachedPointerInfo(StoredVal);
markInstructionForDeletion(L);
++NumGVNLoad;
if (LoadInst *DepLI = dyn_cast<LoadInst>(DepInst)) {
Value *AvailableVal = DepLI;
-
+
// The loads are of a must-aliased pointer, but they may not actually have
// the same type. See if we know how to reuse the previously loaded value
// (depending on its type).
if (DepLI->getType() != L->getType()) {
- if (TD) {
+ if (DL) {
AvailableVal = CoerceAvailableValueToLoadType(DepLI, L->getType(),
- L, *TD);
+ L, *DL);
if (AvailableVal == 0)
return false;
-
+
DEBUG(dbgs() << "GVN COERCED LOAD:\n" << *DepLI << "\n" << *AvailableVal
<< "\n" << *L << "\n\n\n");
}
- else
+ else
return false;
}
-
+
// Remove it!
- patchAndReplaceAllUsesWith(AvailableVal, L);
- if (DepLI->getType()->isPointerTy())
+ patchAndReplaceAllUsesWith(L, AvailableVal);
+ if (DepLI->getType()->getScalarType()->isPointerTy())
MD->invalidateCachedPointerInfo(DepLI);
markInstructionForDeletion(L);
++NumGVNLoad;
// 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 (isa<AllocaInst>(DepInst) || isMallocLikeFn(DepInst)) {
+ if (isa<AllocaInst>(DepInst) || isMallocLikeFn(DepInst, TLI)) {
L->replaceAllUsesWith(UndefValue::get(L->getType()));
markInstructionForDeletion(L);
++NumGVNLoad;
return true;
}
-
+
// If this load occurs either right after a lifetime begin,
// then the loaded value is undefined.
if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(DepInst)) {
return false;
}
-// findLeader - In order to find a leader for a given value number at a
+// findLeader - In order to find a leader for a given value number at a
// specific basic block, we first obtain the list of all Values for that number,
-// and then scan the list to find one whose block dominates the block in
+// and then scan the list to find one whose block dominates the block in
// question. This is fast because dominator tree queries consist of only
// a few comparisons of DFS numbers.
-Value *GVN::findLeader(BasicBlock *BB, uint32_t num) {
+Value *GVN::findLeader(const BasicBlock *BB, uint32_t num) {
LeaderTableEntry Vals = LeaderTable[num];
if (!Vals.Val) return 0;
-
+
Value *Val = 0;
if (DT->dominates(Vals.BB, BB)) {
Val = Vals.Val;
if (isa<Constant>(Val)) return Val;
}
-
+
LeaderTableEntry* Next = Vals.Next;
while (Next) {
if (DT->dominates(Next->BB, BB)) {
if (isa<Constant>(Next->Val)) return Next->Val;
if (!Val) Val = Next->Val;
}
-
+
Next = Next->Next;
}
/// use is dominated by the given basic block. Returns the number of uses that
/// were replaced.
unsigned GVN::replaceAllDominatedUsesWith(Value *From, Value *To,
- BasicBlock *Root) {
+ const BasicBlockEdge &Root) {
unsigned Count = 0;
for (Value::use_iterator UI = From->use_begin(), UE = From->use_end();
UI != UE; ) {
Use &U = (UI++).getUse();
- // If From occurs as a phi node operand then the use implicitly lives in the
- // corresponding incoming block. Otherwise it is the block containing the
- // user that must be dominated by Root.
- BasicBlock *UsingBlock;
- if (PHINode *PN = dyn_cast<PHINode>(U.getUser()))
- UsingBlock = PN->getIncomingBlock(U);
- else
- UsingBlock = cast<Instruction>(U.getUser())->getParent();
-
- if (DT->dominates(Root, UsingBlock)) {
+ if (DT->dominates(Root, U)) {
U.set(To);
++Count;
}
return Count;
}
+/// isOnlyReachableViaThisEdge - There is an edge from 'Src' to 'Dst'. Return
+/// true if every path from the entry block to 'Dst' passes via this edge. In
+/// particular 'Dst' must not be reachable via another edge from 'Src'.
+static bool isOnlyReachableViaThisEdge(const BasicBlockEdge &E,
+ DominatorTree *DT) {
+ // While in theory it is interesting to consider the case in which Dst has
+ // more than one predecessor, because Dst might be part of a loop which is
+ // only reachable from Src, in practice it is pointless since at the time
+ // GVN runs all such loops have preheaders, which means that Dst will have
+ // been changed to have only one predecessor, namely Src.
+ const BasicBlock *Pred = E.getEnd()->getSinglePredecessor();
+ const BasicBlock *Src = E.getStart();
+ assert((!Pred || Pred == Src) && "No edge between these basic blocks!");
+ (void)Src;
+ return Pred != 0;
+}
+
/// propagateEquality - The given values are known to be equal in every block
/// dominated by 'Root'. Exploit this, for example by replacing 'LHS' with
/// 'RHS' everywhere in the scope. Returns whether a change was made.
-bool GVN::propagateEquality(Value *LHS, Value *RHS, BasicBlock *Root) {
+bool GVN::propagateEquality(Value *LHS, Value *RHS,
+ const BasicBlockEdge &Root) {
SmallVector<std::pair<Value*, Value*>, 4> Worklist;
Worklist.push_back(std::make_pair(LHS, RHS));
bool Changed = false;
+ // For speed, compute a conservative fast approximation to
+ // DT->dominates(Root, Root.getEnd());
+ bool RootDominatesEnd = isOnlyReachableViaThisEdge(Root, DT);
while (!Worklist.empty()) {
std::pair<Value*, Value*> Item = Worklist.pop_back_val();
LVN = RVN;
}
}
- assert((!isa<Instruction>(RHS) ||
- DT->properlyDominates(cast<Instruction>(RHS)->getParent(), Root)) &&
- "Instruction doesn't dominate scope!");
// If value numbering later sees that an instruction in the scope is equal
// to 'LHS' then ensure it will be turned into 'RHS'. In order to preserve
// if RHS is an instruction (if an instruction in the scope is morphed into
// LHS then it will be turned into RHS by the next GVN iteration anyway, so
// using the leader table is about compiling faster, not optimizing better).
- if (!isa<Instruction>(RHS))
- addToLeaderTable(LVN, RHS, Root);
+ // The leader table only tracks basic blocks, not edges. Only add to if we
+ // have the simple case where the edge dominates the end.
+ if (RootDominatesEnd && !isa<Instruction>(RHS))
+ addToLeaderTable(LVN, RHS, Root.getEnd());
// Replace all occurrences of 'LHS' with 'RHS' everywhere in the scope. As
// LHS always has at least one use that is not dominated by Root, this will
// If the number we were assigned was brand new then there is no point in
// looking for an instruction realizing it: there cannot be one!
if (Num < NextNum) {
- Value *NotCmp = findLeader(Root, Num);
+ Value *NotCmp = findLeader(Root.getEnd(), Num);
if (NotCmp && isa<Instruction>(NotCmp)) {
unsigned NumReplacements =
replaceAllDominatedUsesWith(NotCmp, NotVal, Root);
}
// Ensure that any instruction in scope that gets the "A < B" value number
// is replaced with false.
- addToLeaderTable(Num, NotVal, Root);
+ // The leader table only tracks basic blocks, not edges. Only add to if we
+ // have the simple case where the edge dominates the end.
+ if (RootDominatesEnd)
+ addToLeaderTable(Num, NotVal, Root.getEnd());
continue;
}
return Changed;
}
-/// isOnlyReachableViaThisEdge - There is an edge from 'Src' to 'Dst'. Return
-/// true if every path from the entry block to 'Dst' passes via this edge. In
-/// particular 'Dst' must not be reachable via another edge from 'Src'.
-static bool isOnlyReachableViaThisEdge(BasicBlock *Src, BasicBlock *Dst,
- DominatorTree *DT) {
- // While in theory it is interesting to consider the case in which Dst has
- // more than one predecessor, because Dst might be part of a loop which is
- // only reachable from Src, in practice it is pointless since at the time
- // GVN runs all such loops have preheaders, which means that Dst will have
- // been changed to have only one predecessor, namely Src.
- BasicBlock *Pred = Dst->getSinglePredecessor();
- assert((!Pred || Pred == Src) && "No edge between these basic blocks!");
- (void)Src;
- return Pred != 0;
-}
-
/// processInstruction - When calculating availability, handle an instruction
/// by inserting it into the appropriate sets
bool GVN::processInstruction(Instruction *I) {
// to value numbering it. Value numbering often exposes redundancies, for
// example if it determines that %y is equal to %x then the instruction
// "%z = and i32 %x, %y" becomes "%z = and i32 %x, %x" which we now simplify.
- if (Value *V = SimplifyInstruction(I, TD, TLI, DT)) {
+ if (Value *V = SimplifyInstruction(I, DL, TLI, DT)) {
I->replaceAllUsesWith(V);
- if (MD && V->getType()->isPointerTy())
+ if (MD && V->getType()->getScalarType()->isPointerTy())
MD->invalidateCachedPointerInfo(V);
markInstructionForDeletion(I);
++NumGVNSimpl;
// For conditional branches, we can perform simple conditional propagation on
// the condition value itself.
if (BranchInst *BI = dyn_cast<BranchInst>(I)) {
- if (!BI->isConditional() || isa<Constant>(BI->getCondition()))
+ if (!BI->isConditional())
return false;
- Value *BranchCond = BI->getCondition();
+ if (isa<Constant>(BI->getCondition()))
+ return processFoldableCondBr(BI);
+ Value *BranchCond = BI->getCondition();
BasicBlock *TrueSucc = BI->getSuccessor(0);
BasicBlock *FalseSucc = BI->getSuccessor(1);
+ // Avoid multiple edges early.
+ if (TrueSucc == FalseSucc)
+ return false;
+
BasicBlock *Parent = BI->getParent();
bool Changed = false;
- if (isOnlyReachableViaThisEdge(Parent, TrueSucc, DT))
- Changed |= propagateEquality(BranchCond,
- ConstantInt::getTrue(TrueSucc->getContext()),
- TrueSucc);
+ Value *TrueVal = ConstantInt::getTrue(TrueSucc->getContext());
+ BasicBlockEdge TrueE(Parent, TrueSucc);
+ Changed |= propagateEquality(BranchCond, TrueVal, TrueE);
- if (isOnlyReachableViaThisEdge(Parent, FalseSucc, DT))
- Changed |= propagateEquality(BranchCond,
- ConstantInt::getFalse(FalseSucc->getContext()),
- FalseSucc);
+ Value *FalseVal = ConstantInt::getFalse(FalseSucc->getContext());
+ BasicBlockEdge FalseE(Parent, FalseSucc);
+ Changed |= propagateEquality(BranchCond, FalseVal, FalseE);
return Changed;
}
Value *SwitchCond = SI->getCondition();
BasicBlock *Parent = SI->getParent();
bool Changed = false;
+
+ // Remember how many outgoing edges there are to every successor.
+ SmallDenseMap<BasicBlock *, unsigned, 16> SwitchEdges;
+ for (unsigned i = 0, n = SI->getNumSuccessors(); i != n; ++i)
+ ++SwitchEdges[SI->getSuccessor(i)];
+
for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end();
i != e; ++i) {
BasicBlock *Dst = i.getCaseSuccessor();
- if (isOnlyReachableViaThisEdge(Parent, Dst, DT))
- Changed |= propagateEquality(SwitchCond, i.getCaseValue(), Dst);
+ // If there is only a single edge, propagate the case value into it.
+ if (SwitchEdges.lookup(Dst) == 1) {
+ BasicBlockEdge E(Parent, Dst);
+ Changed |= propagateEquality(SwitchCond, i.getCaseValue(), E);
+ }
}
return Changed;
}
// Instructions with void type don't return a value, so there's
// no point in trying to find redundancies in them.
if (I->getType()->isVoidTy()) return false;
-
+
uint32_t NextNum = VN.getNextUnusedValueNumber();
unsigned Num = VN.lookup_or_add(I);
addToLeaderTable(Num, I, I->getParent());
return false;
}
-
+
// Perform fast-path value-number based elimination of values inherited from
// dominators.
Value *repl = findLeader(I->getParent(), Num);
addToLeaderTable(Num, I, I->getParent());
return false;
}
-
+
// Remove it!
- patchAndReplaceAllUsesWith(repl, I);
- if (MD && repl->getType()->isPointerTy())
+ patchAndReplaceAllUsesWith(I, repl);
+ if (MD && repl->getType()->getScalarType()->isPointerTy())
MD->invalidateCachedPointerInfo(repl);
markInstructionForDeletion(I);
return true;
/// runOnFunction - This is the main transformation entry point for a function.
bool GVN::runOnFunction(Function& F) {
+ if (skipOptnoneFunction(F))
+ return false;
+
if (!NoLoads)
MD = &getAnalysis<MemoryDependenceAnalysis>();
- DT = &getAnalysis<DominatorTree>();
- TD = getAnalysisIfAvailable<TargetData>();
+ DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
+ DataLayoutPass *DLP = getAnalysisIfAvailable<DataLayoutPass>();
+ DL = DLP ? &DLP->getDataLayout() : 0;
TLI = &getAnalysis<TargetLibraryInfo>();
VN.setAliasAnalysis(&getAnalysis<AliasAnalysis>());
VN.setMemDep(MD);
// optimization opportunities.
for (Function::iterator FI = F.begin(), FE = F.end(); FI != FE; ) {
BasicBlock *BB = FI++;
-
+
bool removedBlock = MergeBlockIntoPredecessor(BB, this);
if (removedBlock) ++NumGVNBlocks;
while (ShouldContinue) {
DEBUG(dbgs() << "GVN iteration: " << Iteration << "\n");
ShouldContinue = iterateOnFunction(F);
- if (splitCriticalEdges())
- ShouldContinue = true;
Changed |= ShouldContinue;
++Iteration;
}
if (EnablePRE) {
+ // Fabricate val-num for dead-code in order to suppress assertion in
+ // performPRE().
+ assignValNumForDeadCode();
bool PREChanged = true;
while (PREChanged) {
PREChanged = performPRE(F);
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();
+ // Do not cleanup DeadBlocks in cleanupGlobalSets() as it's called for each
+ // iteration.
+ DeadBlocks.clear();
return Changed;
}
// (and incrementing BI before processing an instruction).
assert(InstrsToErase.empty() &&
"We expect InstrsToErase to be empty across iterations");
+ if (DeadBlocks.count(BB))
+ return false;
+
bool ChangedFunction = false;
for (BasicBlock::iterator BI = BB->begin(), BE = BB->end();
if (!AtStart)
--BI;
- for (SmallVector<Instruction*, 4>::iterator I = InstrsToErase.begin(),
+ for (SmallVectorImpl<Instruction *>::iterator I = InstrsToErase.begin(),
E = InstrsToErase.end(); I != E; ++I) {
DEBUG(dbgs() << "GVN removed: " << **I << '\n');
if (MD) MD->removeInstruction(*I);
- (*I)->eraseFromParent();
DEBUG(verifyRemoved(*I));
+ (*I)->eraseFromParent();
}
InstrsToErase.clear();
/// control flow patterns and attempts to perform simple PRE at the join point.
bool GVN::performPRE(Function &F) {
bool Changed = false;
- DenseMap<BasicBlock*, Value*> predMap;
+ SmallVector<std::pair<Value*, BasicBlock*>, 8> predMap;
for (df_iterator<BasicBlock*> DI = df_begin(&F.getEntryBlock()),
DE = df_end(&F.getEntryBlock()); DI != DE; ++DI) {
BasicBlock *CurrentBlock = *DI;
if (P == CurrentBlock) {
NumWithout = 2;
break;
- } else if (!DT->dominates(&F.getEntryBlock(), P)) {
+ } else if (!DT->isReachableFromEntry(P)) {
NumWithout = 2;
break;
}
Value* predV = findLeader(P, ValNo);
if (predV == 0) {
+ predMap.push_back(std::make_pair(static_cast<Value *>(0), P));
PREPred = P;
++NumWithout;
} else if (predV == CurInst) {
+ /* CurInst dominates this predecessor. */
NumWithout = 2;
+ break;
} else {
- predMap[P] = predV;
+ predMap.push_back(std::make_pair(predV, P));
++NumWith;
}
}
// we would need to insert instructions in more than one pred.
if (NumWithout != 1 || NumWith == 0)
continue;
-
+
// Don't do PRE across indirect branch.
if (isa<IndirectBrInst>(PREPred->getTerminator()))
continue;
// the PRE predecessor. This is typically because of loads which
// are not value numbered precisely.
if (!success) {
- delete PREInstr;
DEBUG(verifyRemoved(PREInstr));
+ delete PREInstr;
continue;
}
PREInstr->insertBefore(PREPred->getTerminator());
PREInstr->setName(CurInst->getName() + ".pre");
PREInstr->setDebugLoc(CurInst->getDebugLoc());
- predMap[PREPred] = PREInstr;
VN.add(PREInstr, ValNo);
++NumGVNPRE;
addToLeaderTable(ValNo, PREInstr, PREPred);
// Create a PHI to make the value available in this block.
- pred_iterator PB = pred_begin(CurrentBlock), PE = pred_end(CurrentBlock);
- PHINode* Phi = PHINode::Create(CurInst->getType(), std::distance(PB, PE),
+ PHINode* Phi = PHINode::Create(CurInst->getType(), predMap.size(),
CurInst->getName() + ".pre-phi",
CurrentBlock->begin());
- for (pred_iterator PI = PB; PI != PE; ++PI) {
- BasicBlock *P = *PI;
- Phi->addIncoming(predMap[P], P);
+ for (unsigned i = 0, e = predMap.size(); i != e; ++i) {
+ if (Value *V = predMap[i].first)
+ Phi->addIncoming(V, predMap[i].second);
+ else
+ Phi->addIncoming(PREInstr, PREPred);
}
VN.add(Phi, ValNo);
addToLeaderTable(ValNo, Phi, CurrentBlock);
Phi->setDebugLoc(CurInst->getDebugLoc());
CurInst->replaceAllUsesWith(Phi);
- if (Phi->getType()->isPointerTy()) {
+ if (Phi->getType()->getScalarType()->isPointerTy()) {
// Because we have added a PHI-use of the pointer value, it has now
// "escaped" from alias analysis' perspective. We need to inform
// AA of this.
unsigned jj = PHINode::getOperandNumForIncomingValue(ii);
VN.getAliasAnalysis()->addEscapingUse(Phi->getOperandUse(jj));
}
-
+
if (MD)
MD->invalidateCachedPointerInfo(Phi);
}
DEBUG(dbgs() << "GVN PRE removed: " << *CurInst << '\n');
if (MD) MD->removeInstruction(CurInst);
- CurInst->eraseFromParent();
DEBUG(verifyRemoved(CurInst));
+ CurInst->eraseFromParent();
Changed = true;
}
}
return Changed;
}
+/// Split the critical edge connecting the given two blocks, and return
+/// the block inserted to the critical edge.
+BasicBlock *GVN::splitCriticalEdges(BasicBlock *Pred, BasicBlock *Succ) {
+ BasicBlock *BB = SplitCriticalEdge(Pred, Succ, this);
+ if (MD)
+ MD->invalidateCachedPredecessors();
+ return BB;
+}
+
/// splitCriticalEdges - Split critical edges found during the previous
/// iteration that may enable further optimization.
bool GVN::splitCriticalEdges() {
/// iterateOnFunction - Executes one iteration of GVN
bool GVN::iterateOnFunction(Function &F) {
cleanupGlobalSets();
-
+
// Top-down walk of the dominator tree
bool Changed = false;
#if 0
RE = RPOT.end(); RI != RE; ++RI)
Changed |= processBlock(*RI);
#else
+ // Save the blocks this function have before transformation begins. GVN may
+ // split critical edge, and hence may invalidate the RPO/DT iterator.
+ //
+ std::vector<BasicBlock *> BBVect;
+ BBVect.reserve(256);
for (df_iterator<DomTreeNode*> DI = df_begin(DT->getRootNode()),
DE = df_end(DT->getRootNode()); DI != DE; ++DI)
- Changed |= processBlock(DI->getBlock());
+ BBVect.push_back(DI->getBlock());
+
+ for (std::vector<BasicBlock *>::iterator I = BBVect.begin(), E = BBVect.end();
+ I != E; I++)
+ Changed |= processBlock(*I);
#endif
return Changed;
I = LeaderTable.begin(), E = LeaderTable.end(); I != E; ++I) {
const LeaderTableEntry *Node = &I->second;
assert(Node->Val != Inst && "Inst still in value numbering scope!");
-
+
while (Node->Next) {
Node = Node->Next;
assert(Node->Val != Inst && "Inst still in value numbering scope!");
}
}
}
+
+// BB is declared dead, which implied other blocks become dead as well. This
+// function is to add all these blocks to "DeadBlocks". For the dead blocks'
+// live successors, update their phi nodes by replacing the operands
+// corresponding to dead blocks with UndefVal.
+//
+void GVN::addDeadBlock(BasicBlock *BB) {
+ SmallVector<BasicBlock *, 4> NewDead;
+ SmallSetVector<BasicBlock *, 4> DF;
+
+ NewDead.push_back(BB);
+ while (!NewDead.empty()) {
+ BasicBlock *D = NewDead.pop_back_val();
+ if (DeadBlocks.count(D))
+ continue;
+
+ // All blocks dominated by D are dead.
+ SmallVector<BasicBlock *, 8> Dom;
+ DT->getDescendants(D, Dom);
+ DeadBlocks.insert(Dom.begin(), Dom.end());
+
+ // Figure out the dominance-frontier(D).
+ for (SmallVectorImpl<BasicBlock *>::iterator I = Dom.begin(),
+ E = Dom.end(); I != E; I++) {
+ BasicBlock *B = *I;
+ for (succ_iterator SI = succ_begin(B), SE = succ_end(B); SI != SE; SI++) {
+ BasicBlock *S = *SI;
+ if (DeadBlocks.count(S))
+ continue;
+
+ bool AllPredDead = true;
+ for (pred_iterator PI = pred_begin(S), PE = pred_end(S); PI != PE; PI++)
+ if (!DeadBlocks.count(*PI)) {
+ AllPredDead = false;
+ break;
+ }
+
+ if (!AllPredDead) {
+ // S could be proved dead later on. That is why we don't update phi
+ // operands at this moment.
+ DF.insert(S);
+ } else {
+ // While S is not dominated by D, it is dead by now. This could take
+ // place if S already have a dead predecessor before D is declared
+ // dead.
+ NewDead.push_back(S);
+ }
+ }
+ }
+ }
+
+ // For the dead blocks' live successors, update their phi nodes by replacing
+ // the operands corresponding to dead blocks with UndefVal.
+ for(SmallSetVector<BasicBlock *, 4>::iterator I = DF.begin(), E = DF.end();
+ I != E; I++) {
+ BasicBlock *B = *I;
+ if (DeadBlocks.count(B))
+ continue;
+
+ SmallVector<BasicBlock *, 4> Preds(pred_begin(B), pred_end(B));
+ for (SmallVectorImpl<BasicBlock *>::iterator PI = Preds.begin(),
+ PE = Preds.end(); PI != PE; PI++) {
+ BasicBlock *P = *PI;
+
+ if (!DeadBlocks.count(P))
+ continue;
+
+ if (isCriticalEdge(P->getTerminator(), GetSuccessorNumber(P, B))) {
+ if (BasicBlock *S = splitCriticalEdges(P, B))
+ DeadBlocks.insert(P = S);
+ }
+
+ for (BasicBlock::iterator II = B->begin(); isa<PHINode>(II); ++II) {
+ PHINode &Phi = cast<PHINode>(*II);
+ Phi.setIncomingValue(Phi.getBasicBlockIndex(P),
+ UndefValue::get(Phi.getType()));
+ }
+ }
+ }
+}
+
+// If the given branch is recognized as a foldable branch (i.e. conditional
+// branch with constant condition), it will perform following analyses and
+// transformation.
+// 1) If the dead out-coming edge is a critical-edge, split it. Let
+// R be the target of the dead out-coming edge.
+// 1) Identify the set of dead blocks implied by the branch's dead outcoming
+// edge. The result of this step will be {X| X is dominated by R}
+// 2) Identify those blocks which haves at least one dead prodecessor. The
+// result of this step will be dominance-frontier(R).
+// 3) Update the PHIs in DF(R) by replacing the operands corresponding to
+// dead blocks with "UndefVal" in an hope these PHIs will optimized away.
+//
+// Return true iff *NEW* dead code are found.
+bool GVN::processFoldableCondBr(BranchInst *BI) {
+ if (!BI || BI->isUnconditional())
+ return false;
+
+ ConstantInt *Cond = dyn_cast<ConstantInt>(BI->getCondition());
+ if (!Cond)
+ return false;
+
+ BasicBlock *DeadRoot = Cond->getZExtValue() ?
+ BI->getSuccessor(1) : BI->getSuccessor(0);
+ if (DeadBlocks.count(DeadRoot))
+ return false;
+
+ if (!DeadRoot->getSinglePredecessor())
+ DeadRoot = splitCriticalEdges(BI->getParent(), DeadRoot);
+
+ addDeadBlock(DeadRoot);
+ return true;
+}
+
+// performPRE() will trigger assert if it come across an instruciton without
+// associated val-num. As it normally has far more live instructions than dead
+// instructions, it makes more sense just to "fabricate" a val-number for the
+// dead code than checking if instruction involved is dead or not.
+void GVN::assignValNumForDeadCode() {
+ for (SetVector<BasicBlock *>::iterator I = DeadBlocks.begin(),
+ E = DeadBlocks.end(); I != E; I++) {
+ BasicBlock *BB = *I;
+ for (BasicBlock::iterator II = BB->begin(), EE = BB->end();
+ II != EE; II++) {
+ Instruction *Inst = &*II;
+ unsigned ValNum = VN.lookup_or_add(Inst);
+ addToLeaderTable(ValNum, Inst, BB);
+ }
+ }
+}