#include "llvm/Analysis/ValueTracking.h"
#include "llvm/Assembly/Writer.h"
#include "llvm/Target/TargetData.h"
+#include "llvm/Target/TargetLibraryInfo.h"
#include "llvm/Transforms/Utils/BasicBlockUtils.h"
#include "llvm/Transforms/Utils/SSAUpdater.h"
#include "llvm/ADT/DenseMap.h"
#include "llvm/Support/CommandLine.h"
#include "llvm/Support/Debug.h"
#include "llvm/Support/IRBuilder.h"
+#include "llvm/Support/PatternMatch.h"
using namespace llvm;
+using namespace PatternMatch;
STATISTIC(NumGVNInstr, "Number of instructions deleted");
STATISTIC(NumGVNLoad, "Number of loads deleted");
STATISTIC(NumGVNPRE, "Number of instructions PRE'd");
STATISTIC(NumGVNBlocks, "Number of blocks merged");
+STATISTIC(NumGVNSimpl, "Number of instructions simplified");
+STATISTIC(NumGVNEqProp, "Number of equalities propagated");
STATISTIC(NumPRELoad, "Number of loads PRE'd");
static cl::opt<bool> EnablePRE("enable-pre",
namespace {
struct Expression {
uint32_t opcode;
- const Type *type;
+ Type *type;
SmallVector<uint32_t, 4> varargs;
Expression(uint32_t o = ~2U) : opcode(o) { }
uint32_t nextValueNumber;
Expression create_expression(Instruction* I);
+ Expression create_extractvalue_expression(ExtractValueInst* EI);
uint32_t lookup_or_add_call(CallInst* C);
public:
ValueTable() : nextValueNumber(1) { }
// ValueTable Internal Functions
//===----------------------------------------------------------------------===//
-
Expression ValueTable::create_expression(Instruction *I) {
Expression e;
e.type = I->getType();
OI != OE; ++OI)
e.varargs.push_back(lookup_or_add(*OI));
- if (CmpInst *C = dyn_cast<CmpInst>(I))
+ if (CmpInst *C = dyn_cast<CmpInst>(I)) {
e.opcode = (C->getOpcode() << 8) | C->getPredicate();
- else if (ExtractValueInst *E = dyn_cast<ExtractValueInst>(I)) {
- for (ExtractValueInst::idx_iterator II = E->idx_begin(), IE = E->idx_end();
- II != IE; ++II)
- e.varargs.push_back(*II);
} else if (InsertValueInst *E = dyn_cast<InsertValueInst>(I)) {
for (InsertValueInst::idx_iterator II = E->idx_begin(), IE = E->idx_end();
II != IE; ++II)
return e;
}
+Expression ValueTable::create_extractvalue_expression(ExtractValueInst *EI) {
+ assert(EI != 0 && "Not an ExtractValueInst?");
+ Expression e;
+ e.type = EI->getType();
+ e.opcode = 0;
+
+ IntrinsicInst *I = dyn_cast<IntrinsicInst>(EI->getAggregateOperand());
+ if (I != 0 && EI->getNumIndices() == 1 && *EI->idx_begin() == 0 ) {
+ // EI might be an extract from one of our recognised intrinsics. If it
+ // is we'll synthesize a semantically equivalent expression instead on
+ // an extract value expression.
+ switch (I->getIntrinsicID()) {
+ case Intrinsic::sadd_with_overflow:
+ case Intrinsic::uadd_with_overflow:
+ e.opcode = Instruction::Add;
+ break;
+ case Intrinsic::ssub_with_overflow:
+ case Intrinsic::usub_with_overflow:
+ e.opcode = Instruction::Sub;
+ break;
+ case Intrinsic::smul_with_overflow:
+ case Intrinsic::umul_with_overflow:
+ e.opcode = Instruction::Mul;
+ break;
+ default:
+ break;
+ }
+
+ if (e.opcode != 0) {
+ // Intrinsic recognized. Grab its args to finish building the expression.
+ assert(I->getNumArgOperands() == 2 &&
+ "Expect two args for recognised intrinsics.");
+ e.varargs.push_back(lookup_or_add(I->getArgOperand(0)));
+ e.varargs.push_back(lookup_or_add(I->getArgOperand(1)));
+ return e;
+ }
+ }
+
+ // Not a recognised intrinsic. Fall back to producing an extract value
+ // expression.
+ e.opcode = EI->getOpcode();
+ for (Instruction::op_iterator OI = EI->op_begin(), OE = EI->op_end();
+ OI != OE; ++OI)
+ e.varargs.push_back(lookup_or_add(*OI));
+
+ for (ExtractValueInst::idx_iterator II = EI->idx_begin(), IE = EI->idx_end();
+ II != IE; ++II)
+ e.varargs.push_back(*II);
+
+ return e;
+}
+
//===----------------------------------------------------------------------===//
// ValueTable External Functions
//===----------------------------------------------------------------------===//
// Non-local case.
const MemoryDependenceAnalysis::NonLocalDepInfo &deps =
MD->getNonLocalCallDependency(CallSite(C));
- // FIXME: call/call dependencies for readonly calls should return def, not
- // clobber! Move the checking logic to MemDep!
+ // FIXME: Move the checking logic to MemDep!
CallInst* cdep = 0;
// Check to see if we have a single dominating call instruction that is
// identical to C.
for (unsigned i = 0, e = deps.size(); i != e; ++i) {
const NonLocalDepEntry *I = &deps[i];
- // Ignore non-local dependencies.
if (I->getResult().isNonLocal())
continue;
- // We don't handle non-depedencies. If we already have a call, reject
+ // We don't handle non-definitions. If we already have a call, reject
// instruction dependencies.
- if (I->getResult().isClobber() || cdep != 0) {
+ if (!I->getResult().isDef() || cdep != 0) {
cdep = 0;
break;
}
case Instruction::ExtractElement:
case Instruction::InsertElement:
case Instruction::ShuffleVector:
- case Instruction::ExtractValue:
case Instruction::InsertValue:
case Instruction::GetElementPtr:
exp = create_expression(I);
break;
+ case Instruction::ExtractValue:
+ exp = create_extractvalue_expression(cast<ExtractValueInst>(I));
+ break;
default:
valueNumbering[V] = nextValueNumber;
return nextValueNumber++;
MemoryDependenceAnalysis *MD;
DominatorTree *DT;
const TargetData *TD;
-
+ const TargetLibraryInfo *TLI;
+
ValueTable VN;
/// LeaderTable - A mapping from value numbers to lists of Value*'s that
// This transformation requires dominator postdominator info
virtual void getAnalysisUsage(AnalysisUsage &AU) const {
AU.addRequired<DominatorTree>();
+ AU.addRequired<TargetLibraryInfo>();
if (!NoLoads)
AU.addRequired<MemoryDependenceAnalysis>();
AU.addRequired<AliasAnalysis>();
void cleanupGlobalSets();
void verifyRemoved(const Instruction *I) const;
bool splitCriticalEdges();
+ unsigned replaceAllDominatedUsesWith(Value *From, Value *To,
+ BasicBlock *Root);
+ bool propagateEquality(Value *LHS, Value *RHS, BasicBlock *Root);
};
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(TargetLibraryInfo)
INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
INITIALIZE_PASS_END(GVN, "gvn", "Global Value Numbering", false, false)
/// CanCoerceMustAliasedValueToLoad - Return true if
/// CoerceAvailableValueToLoadType will succeed.
static bool CanCoerceMustAliasedValueToLoad(Value *StoredVal,
- const Type *LoadTy,
+ Type *LoadTy,
const TargetData &TD) {
// 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 we can't do it, return null.
static Value *CoerceAvailableValueToLoadType(Value *StoredVal,
- const Type *LoadedTy,
+ Type *LoadedTy,
Instruction *InsertPt,
const TargetData &TD) {
if (!CanCoerceMustAliasedValueToLoad(StoredVal, LoadedTy, TD))
return 0;
// If this is already the right type, just return it.
- const Type *StoredValTy = StoredVal->getType();
+ Type *StoredValTy = StoredVal->getType();
- uint64_t StoreSize = TD.getTypeStoreSizeInBits(StoredValTy);
- uint64_t LoadSize = TD.getTypeStoreSizeInBits(LoadedTy);
+ uint64_t StoreSize = TD.getTypeSizeInBits(StoredValTy);
+ uint64_t LoadSize = TD.getTypeSizeInBits(LoadedTy);
// If the store and reload are the same size, we can always reuse it.
if (StoreSize == LoadSize) {
StoredVal = new PtrToIntInst(StoredVal, StoredValTy, "", InsertPt);
}
- const Type *TypeToCastTo = LoadedTy;
+ Type *TypeToCastTo = LoadedTy;
if (TypeToCastTo->isPointerTy())
TypeToCastTo = TD.getIntPtrType(StoredValTy->getContext());
}
// Truncate the integer to the right size now.
- const Type *NewIntTy = IntegerType::get(StoredValTy->getContext(), LoadSize);
+ Type *NewIntTy = IntegerType::get(StoredValTy->getContext(), LoadSize);
StoredVal = new TruncInst(StoredVal, NewIntTy, "trunc", InsertPt);
if (LoadedTy == NewIntTy)
/// Check this case to see if there is anything more we can do before we give
/// up. This returns -1 if we have to give up, or a byte number in the stored
/// value of the piece that feeds the load.
-static int AnalyzeLoadFromClobberingWrite(const Type *LoadTy, Value *LoadPtr,
+static int AnalyzeLoadFromClobberingWrite(Type *LoadTy, Value *LoadPtr,
Value *WritePtr,
uint64_t WriteSizeInBits,
const TargetData &TD) {
/// AnalyzeLoadFromClobberingStore - This function is called when we have a
/// memdep query of a load that ends up being a clobbering store.
-static int AnalyzeLoadFromClobberingStore(const Type *LoadTy, Value *LoadPtr,
+static int AnalyzeLoadFromClobberingStore(Type *LoadTy, Value *LoadPtr,
StoreInst *DepSI,
const TargetData &TD) {
// Cannot handle reading from store of first-class aggregate yet.
/// 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(const Type *LoadTy, Value *LoadPtr,
+static int AnalyzeLoadFromClobberingLoad(Type *LoadTy, Value *LoadPtr,
LoadInst *DepLI, const TargetData &TD){
// Cannot handle reading from store of first-class aggregate yet.
if (DepLI->getType()->isStructTy() || DepLI->getType()->isArrayTy())
-static int AnalyzeLoadFromClobberingMemInst(const Type *LoadTy, Value *LoadPtr,
+static int AnalyzeLoadFromClobberingMemInst(Type *LoadTy, Value *LoadPtr,
MemIntrinsic *MI,
const TargetData &TD) {
// If the mem operation is a non-constant size, we can't handle it.
llvm::Type::getInt8PtrTy(Src->getContext()));
Constant *OffsetCst =
ConstantInt::get(Type::getInt64Ty(Src->getContext()), (unsigned)Offset);
- Src = ConstantExpr::getGetElementPtr(Src, &OffsetCst, 1);
+ Src = ConstantExpr::getGetElementPtr(Src, OffsetCst);
Src = ConstantExpr::getBitCast(Src, PointerType::getUnqual(LoadTy));
if (ConstantFoldLoadFromConstPtr(Src, &TD))
return Offset;
/// mustalias. Check this case to see if there is anything more we can do
/// before we give up.
static Value *GetStoreValueForLoad(Value *SrcVal, unsigned Offset,
- const Type *LoadTy,
+ Type *LoadTy,
Instruction *InsertPt, const TargetData &TD){
LLVMContext &Ctx = SrcVal->getType()->getContext();
// 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), "tmp");
+ SrcVal = Builder.CreatePtrToInt(SrcVal, TD.getIntPtrType(Ctx));
if (!SrcVal->getType()->isIntegerTy())
- SrcVal = Builder.CreateBitCast(SrcVal, IntegerType::get(Ctx, StoreSize*8),
- "tmp");
+ SrcVal = Builder.CreateBitCast(SrcVal, IntegerType::get(Ctx, StoreSize*8));
// Shift the bits to the least significant depending on endianness.
unsigned ShiftAmt;
ShiftAmt = (StoreSize-LoadSize-Offset)*8;
if (ShiftAmt)
- SrcVal = Builder.CreateLShr(SrcVal, ShiftAmt, "tmp");
+ SrcVal = Builder.CreateLShr(SrcVal, ShiftAmt);
if (LoadSize != StoreSize)
- SrcVal = Builder.CreateTrunc(SrcVal, IntegerType::get(Ctx, LoadSize*8),
- "tmp");
+ SrcVal = Builder.CreateTrunc(SrcVal, IntegerType::get(Ctx, LoadSize*8));
return CoerceAvailableValueToLoadType(SrcVal, LoadTy, InsertPt, TD);
}
/// because the pointers don't mustalias. Check this case to see if there is
/// anything more we can do before we give up.
static Value *GetLoadValueForLoad(LoadInst *SrcVal, unsigned Offset,
- const Type *LoadTy, Instruction *InsertPt,
+ Type *LoadTy, Instruction *InsertPt,
GVN &gvn) {
const TargetData &TD = *gvn.getTargetData();
// If Offset+LoadTy exceeds the size of SrcVal, then we must be wanting to
unsigned SrcValSize = TD.getTypeStoreSize(SrcVal->getType());
unsigned LoadSize = TD.getTypeStoreSize(LoadTy);
if (Offset+LoadSize > SrcValSize) {
- assert(!SrcVal->isVolatile() && "Cannot widen volatile load!");
- assert(isa<IntegerType>(SrcVal->getType())&&"Can't widen non-integer load");
+ assert(SrcVal->isSimple() && "Cannot widen volatile/atomic load!");
+ assert(SrcVal->getType()->isIntegerTy() && "Can't widen non-integer load");
// If we have a load/load clobber an DepLI can be widened to cover this
// load, then we should widen it to the next power of 2 size big enough!
unsigned NewLoadSize = Offset+LoadSize;
// 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));
- const Type *DestPTy =
+ Type *DestPTy =
IntegerType::get(LoadTy->getContext(), NewLoadSize*8);
DestPTy = PointerType::get(DestPTy,
cast<PointerType>(PtrVal->getType())->getAddressSpace());
-
+ Builder.SetCurrentDebugLocation(SrcVal->getDebugLoc());
PtrVal = Builder.CreateBitCast(PtrVal, DestPTy);
LoadInst *NewLoad = Builder.CreateLoad(PtrVal);
NewLoad->takeName(SrcVal);
NewLoad->setAlignment(SrcVal->getAlignment());
-
+
DEBUG(dbgs() << "GVN WIDENED LOAD: " << *SrcVal << "\n");
DEBUG(dbgs() << "TO: " << *NewLoad << "\n");
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,
+ // but then there all of the operations based on it would need to be
+ // rehashed. Just leave the dead load around.
gvn.getMemDep().removeInstruction(SrcVal);
- //gvn.markInstructionForDeletion(SrcVal);
SrcVal = NewLoad;
}
/// GetMemInstValueForLoad - This function is called when we have a
/// memdep query of a load that ends up being a clobbering mem intrinsic.
static Value *GetMemInstValueForLoad(MemIntrinsic *SrcInst, unsigned Offset,
- const Type *LoadTy, Instruction *InsertPt,
+ Type *LoadTy, Instruction *InsertPt,
const TargetData &TD){
LLVMContext &Ctx = LoadTy->getContext();
uint64_t LoadSize = TD.getTypeSizeInBits(LoadTy)/8;
llvm::Type::getInt8PtrTy(Src->getContext()));
Constant *OffsetCst =
ConstantInt::get(Type::getInt64Ty(Src->getContext()), (unsigned)Offset);
- Src = ConstantExpr::getGetElementPtr(Src, &OffsetCst, 1);
+ Src = ConstantExpr::getGetElementPtr(Src, OffsetCst);
Src = ConstantExpr::getBitCast(Src, PointerType::getUnqual(LoadTy));
return ConstantFoldLoadFromConstPtr(Src, &TD);
}
/// MaterializeAdjustedValue - Emit code into this block to adjust the value
/// defined here to the specified type. This handles various coercion cases.
- Value *MaterializeAdjustedValue(const Type *LoadTy, GVN &gvn) const {
+ Value *MaterializeAdjustedValue(Type *LoadTy, GVN &gvn) const {
Value *Res;
if (isSimpleValue()) {
Res = getSimpleValue();
SSAUpdater SSAUpdate(&NewPHIs);
SSAUpdate.Initialize(LI->getType(), LI->getName());
- const Type *LoadTy = LI->getType();
+ Type *LoadTy = LI->getType();
for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i) {
const AvailableValueInBlock &AV = ValuesPerBlock[i];
// escaping uses to any values that are operands to these PHIs.
for (unsigned i = 0, e = NewPHIs.size(); i != e; ++i) {
PHINode *P = NewPHIs[i];
- for (unsigned ii = 0, ee = P->getNumIncomingValues(); ii != ee; ++ii)
- AA->addEscapingUse(P->getOperandUse(2*ii));
+ for (unsigned ii = 0, ee = P->getNumIncomingValues(); ii != ee; ++ii) {
+ unsigned jj = PHINode::getOperandNumForIncomingValue(ii);
+ AA->addEscapingUse(P->getOperandUse(jj));
+ }
}
}
// If we had a phi translation failure, we'll have a single entry which is a
// clobber in the current block. Reject this early.
- if (Deps.size() == 1 && Deps[0].getResult().isClobber() &&
- Deps[0].getResult().getInst()->getParent() == LI->getParent()) {
+ if (Deps.size() == 1
+ && !Deps[0].getResult().isDef() && !Deps[0].getResult().isClobber())
+ {
DEBUG(
dbgs() << "GVN: non-local load ";
WriteAsOperand(dbgs(), LI);
- dbgs() << " is clobbered by " << *Deps[0].getResult().getInst() << '\n';
+ dbgs() << " has unknown dependencies\n";
);
return false;
}
BasicBlock *DepBB = Deps[i].getBB();
MemDepResult DepInfo = Deps[i].getResult();
+ if (!DepInfo.isDef() && !DepInfo.isClobber()) {
+ UnavailableBlocks.push_back(DepBB);
+ continue;
+ }
+
if (DepInfo.isClobber()) {
// The address being loaded in this non-local block may not be the same as
// the pointer operand of the load if PHI translation occurs. Make sure
continue;
}
+ // DepInfo.isDef() here
+
Instruction *DepInst = DepInfo.getInst();
// Loading the allocation -> undef.
for (unsigned i = 0, e = UnavailableBlocks.size(); i != e; ++i)
Blockers.insert(UnavailableBlocks[i]);
- // Lets find first basic block with more than one predecessor. Walk backwards
- // through predecessors if needed.
+ // Let's find the first basic block with more than one predecessor. Walk
+ // backwards through predecessors if needed.
BasicBlock *LoadBB = LI->getParent();
BasicBlock *TmpBB = LoadBB;
<< Pred->getName() << "': " << *LI << '\n');
return false;
}
+
+ if (LoadBB->isLandingPad()) {
+ DEBUG(dbgs()
+ << "COULD NOT PRE LOAD BECAUSE OF LANDING PAD CRITICAL EDGE '"
+ << Pred->getName() << "': " << *LI << '\n');
+ return false;
+ }
+
unsigned SuccNum = GetSuccessorNumber(Pred, LoadBB);
NeedToSplit.push_back(std::make_pair(Pred->getTerminator(), SuccNum));
}
}
+
if (!NeedToSplit.empty()) {
toSplit.append(NeedToSplit.begin(), NeedToSplit.end());
return false;
if (MDNode *Tag = LI->getMetadata(LLVMContext::MD_tbaa))
NewLoad->setMetadata(LLVMContext::MD_tbaa, Tag);
+ // Transfer DebugLoc.
+ NewLoad->setDebugLoc(LI->getDebugLoc());
+
// Add the newly created load.
ValuesPerBlock.push_back(AvailableValueInBlock::get(UnavailablePred,
NewLoad));
if (!MD)
return false;
- if (L->isVolatile())
+ if (!L->isSimple())
return false;
+ if (L->use_empty()) {
+ markInstructionForDeletion(L);
+ return true;
+ }
+
// ... to a pointer that has been loaded from before...
MemDepResult Dep = MD->getDependency(L);
if (Dep.isNonLocal())
return processNonLocalLoad(L);
+ if (!Dep.isDef()) {
+ DEBUG(
+ // fast print dep, using operator<< on instruction is too slow.
+ dbgs() << "GVN: load ";
+ WriteAsOperand(dbgs(), L);
+ dbgs() << " has unknown dependence\n";
+ );
+ return false;
+ }
+
Instruction *DepInst = Dep.getInst();
if (StoreInst *DepSI = dyn_cast<StoreInst>(DepInst)) {
Value *StoredVal = DepSI->getValueOperand();
return Val;
}
+/// replaceAllDominatedUsesWith - Replace all uses of 'From' with 'To' if the
+/// 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) {
+ unsigned Count = 0;
+ for (Value::use_iterator UI = From->use_begin(), UE = From->use_end();
+ UI != UE; ) {
+ Instruction *User = cast<Instruction>(*UI);
+ unsigned OpNum = UI.getOperandNo();
+ ++UI;
+
+ if (DT->dominates(Root, User->getParent())) {
+ User->setOperand(OpNum, To);
+ ++Count;
+ }
+ }
+ return Count;
+}
+
+/// 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) {
+ if (LHS == RHS) return false;
+ assert(LHS->getType() == RHS->getType() && "Equal but types differ!");
+
+ // Don't try to propagate equalities between constants.
+ if (isa<Constant>(LHS) && isa<Constant>(RHS))
+ return false;
+
+ // Make sure that any constants are on the right-hand side. In general the
+ // best results are obtained by placing the longest lived value on the RHS.
+ if (isa<Constant>(LHS))
+ std::swap(LHS, RHS);
+
+ // If neither term is constant then bail out. This is not for correctness,
+ // it's just that the non-constant case is much less useful: it occurs just
+ // as often as the constant case but handling it hardly ever results in an
+ // improvement.
+ if (!isa<Constant>(RHS))
+ return false;
+
+ // If value numbering later deduces that an instruction in the scope is equal
+ // to 'LHS' then ensure it will be turned into 'RHS'.
+ addToLeaderTable(VN.lookup_or_add(LHS), RHS, Root);
+
+ // 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
+ // never do anything if LHS has only one use.
+ bool Changed = false;
+ if (!LHS->hasOneUse()) {
+ unsigned NumReplacements = replaceAllDominatedUsesWith(LHS, RHS, Root);
+ Changed |= NumReplacements > 0;
+ NumGVNEqProp += NumReplacements;
+ }
+
+ // Now try to deduce additional equalities from this one. For example, if the
+ // known equality was "(A != B)" == "false" then it follows that A and B are
+ // equal in the scope. Only boolean equalities with an explicit true or false
+ // RHS are currently supported.
+ if (!RHS->getType()->isIntegerTy(1))
+ // Not a boolean equality - bail out.
+ return Changed;
+ ConstantInt *CI = dyn_cast<ConstantInt>(RHS);
+ if (!CI)
+ // RHS neither 'true' nor 'false' - bail out.
+ return Changed;
+ // Whether RHS equals 'true'. Otherwise it equals 'false'.
+ bool isKnownTrue = CI->isAllOnesValue();
+ bool isKnownFalse = !isKnownTrue;
+
+ // If "A && B" is known true then both A and B are known true. If "A || B"
+ // is known false then both A and B are known false.
+ Value *A, *B;
+ if ((isKnownTrue && match(LHS, m_And(m_Value(A), m_Value(B)))) ||
+ (isKnownFalse && match(LHS, m_Or(m_Value(A), m_Value(B))))) {
+ Changed |= propagateEquality(A, RHS, Root);
+ Changed |= propagateEquality(B, RHS, Root);
+ return Changed;
+ }
+
+ // If we are propagating an equality like "(A == B)" == "true" then also
+ // propagate the equality A == B.
+ if (ICmpInst *Cmp = dyn_cast<ICmpInst>(LHS)) {
+ // Only equality comparisons are supported.
+ if ((isKnownTrue && Cmp->getPredicate() == CmpInst::ICMP_EQ) ||
+ (isKnownFalse && Cmp->getPredicate() == CmpInst::ICMP_NE)) {
+ Value *Op0 = Cmp->getOperand(0), *Op1 = Cmp->getOperand(1);
+ Changed |= propagateEquality(Op0, Op1, Root);
+ }
+ return Changed;
+ }
+
+ 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) {
+ // First off, there must not be more than one edge from Src to Dst, there
+ // should be exactly one. So keep track of the number of times Src occurs
+ // as a predecessor of Dst and fail if it's more than once. Secondly, any
+ // other predecessors of Dst should be dominated by Dst (see logic below).
+ bool SawEdgeFromSrc = false;
+ for (pred_iterator PI = pred_begin(Dst), PE = pred_end(Dst); PI != PE; ++PI) {
+ BasicBlock *Pred = *PI;
+ if (Pred == Src) {
+ // An edge from Src to Dst.
+ if (SawEdgeFromSrc)
+ // There are multiple edges from Src to Dst - fail.
+ return false;
+ SawEdgeFromSrc = true;
+ continue;
+ }
+ // If the predecessor is not dominated by Dst, then it must be possible to
+ // reach it either without passing through Src (and thus not via the edge)
+ // or by passing through Src but taking a different edge out of Src. Either
+ // way it is possible to reach Dst without passing via the edge, so fail.
+ if (!DT->dominates(Dst, *PI))
+ return false;
+ }
+ assert(SawEdgeFromSrc && "No edge between these basic blocks!");
+
+ // Every path from the entry block to Dst must at some point pass to Dst from
+ // a predecessor that is not dominated by Dst. This predecessor can only be
+ // Src, since all others are dominated by Dst. As there is only one edge from
+ // Src to Dst, the path passes by this edge.
+ return true;
+}
/// processInstruction - When calculating availability, handle an instruction
/// by inserting it into the appropriate sets
// 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, DT)) {
+ if (Value *V = SimplifyInstruction(I, TD, TLI, DT)) {
I->replaceAllUsesWith(V);
if (MD && V->getType()->isPointerTy())
MD->invalidateCachedPointerInfo(V);
markInstructionForDeletion(I);
+ ++NumGVNSimpl;
return true;
}
return false;
}
- // For conditions branches, we can perform simple conditional propagation on
+ // 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()))
return false;
-
+
Value *BranchCond = BI->getCondition();
- uint32_t CondVN = VN.lookup_or_add(BranchCond);
-
+
BasicBlock *TrueSucc = BI->getSuccessor(0);
BasicBlock *FalseSucc = BI->getSuccessor(1);
-
- if (TrueSucc->getSinglePredecessor())
- addToLeaderTable(CondVN,
- ConstantInt::getTrue(TrueSucc->getContext()),
- TrueSucc);
- if (FalseSucc->getSinglePredecessor())
- addToLeaderTable(CondVN,
- ConstantInt::getFalse(TrueSucc->getContext()),
- FalseSucc);
-
- return false;
+ BasicBlock *Parent = BI->getParent();
+ bool Changed = false;
+
+ if (isOnlyReachableViaThisEdge(Parent, TrueSucc, DT))
+ Changed |= propagateEquality(BranchCond,
+ ConstantInt::getTrue(TrueSucc->getContext()),
+ TrueSucc);
+
+ if (isOnlyReachableViaThisEdge(Parent, FalseSucc, DT))
+ Changed |= propagateEquality(BranchCond,
+ ConstantInt::getFalse(FalseSucc->getContext()),
+ FalseSucc);
+
+ return Changed;
}
-
+
+ // For switches, propagate the case values into the case destinations.
+ if (SwitchInst *SI = dyn_cast<SwitchInst>(I)) {
+ Value *SwitchCond = SI->getCondition();
+ BasicBlock *Parent = SI->getParent();
+ bool Changed = false;
+ for (unsigned i = 1, e = SI->getNumCases(); i != e; ++i) {
+ BasicBlock *Dst = SI->getSuccessor(i);
+ if (isOnlyReachableViaThisEdge(Parent, Dst, DT))
+ Changed |= propagateEquality(SwitchCond, SI->getCaseValue(i), Dst);
+ }
+ return Changed;
+ }
+
// Instructions with void type don't return a value, so there's
// no point in trying to find redudancies in them.
if (I->getType()->isVoidTy()) return false;
MD = &getAnalysis<MemoryDependenceAnalysis>();
DT = &getAnalysis<DominatorTree>();
TD = getAnalysisIfAvailable<TargetData>();
+ TLI = &getAnalysis<TargetLibraryInfo>();
VN.setAliasAnalysis(&getAnalysis<AliasAnalysis>());
VN.setMemDep(MD);
VN.setDomTree(DT);
// Nothing to PRE in the entry block.
if (CurrentBlock == &F.getEntryBlock()) continue;
+ // Don't perform PRE on a landing pad.
+ if (CurrentBlock->isLandingPad()) continue;
+
for (BasicBlock::iterator BI = CurrentBlock->begin(),
BE = CurrentBlock->end(); BI != BE; ) {
Instruction *CurInst = BI++;
PREInstr->insertBefore(PREPred->getTerminator());
PREInstr->setName(CurInst->getName() + ".pre");
+ PREInstr->setDebugLoc(CurInst->getDebugLoc());
predMap[PREPred] = PREInstr;
VN.add(PREInstr, ValNo);
++NumGVNPRE;
VN.add(Phi, ValNo);
addToLeaderTable(ValNo, Phi, CurrentBlock);
-
+ Phi->setDebugLoc(CurInst->getDebugLoc());
CurInst->replaceAllUsesWith(Phi);
if (Phi->getType()->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.
- for (unsigned ii = 0, ee = Phi->getNumIncomingValues(); ii != ee; ++ii)
- VN.getAliasAnalysis()->addEscapingUse(Phi->getOperandUse(2*ii));
+ for (unsigned ii = 0, ee = Phi->getNumIncomingValues(); ii != ee;
+ ++ii) {
+ unsigned jj = PHINode::getOperandNumForIncomingValue(ii);
+ VN.getAliasAnalysis()->addEscapingUse(Phi->getOperandUse(jj));
+ }
if (MD)
MD->invalidateCachedPointerInfo(Phi);