void BlockState::processWrite(RLEContext &Ctx, SILInstruction *I, SILValue Mem,
SILValue Val, RLEKind Kind) {
// Initialize the LSLocation.
LSLocation L(Mem);
// If we cant figure out the Base or Projection Path for the write,
// process it as an unknown memory instruction.
if (!L.isValid()) {
// we can ignore unknown store instructions if we are computing the
// AvailSetMax.
if (!isComputeAvailSetMax(Kind)) {
processUnknownWriteInst(Ctx, I, Kind);
}
return;
}
// Expand the given location and val into individual fields and process
// them as separate writes.
LSLocationList Locs;
LSLocation::expand(L, &I->getModule(), Locs, Ctx.getTE());
if (isComputeAvailSetMax(Kind)) {
for (unsigned i = 0; i < Locs.size(); ++i) {
updateMaxAvailForwardSetForWrite(Ctx, Ctx.getLocationBit(Locs[i]));
}
return;
}
// Are we computing the genset and killset ?
if (isComputeAvailGenKillSet(Kind)) {
for (unsigned i = 0; i < Locs.size(); ++i) {
updateGenKillSetForWrite(Ctx, Ctx.getLocationBit(Locs[i]));
}
return;
}
// Are we computing available set ?
if (isComputeAvailSet(Kind)) {
for (unsigned i = 0; i < Locs.size(); ++i) {
updateForwardSetForWrite(Ctx, Ctx.getLocationBit(Locs[i]));
}
return;
}
// Are we computing available value or performing RLE?
if (isComputeAvailValue(Kind) || isPerformingRLE(Kind)) {
LSValueList Vals;
LSValue::expand(Val, &I->getModule(), Vals, Ctx.getTE());
for (unsigned i = 0; i < Locs.size(); ++i) {
updateForwardSetAndValForWrite(Ctx, Ctx.getLocationBit(Locs[i]),
Ctx.getValueBit(Vals[i]));
}
return;
}
llvm_unreachable("Unknown RLE compute kind");
}
BlockState::ValueState BlockState::getValueStateAtEndOfBlock(RLEContext &Ctx,
LSLocation &L) {
LSLocationList Locs;
LSLocation::expand(L, &BB->getModule(), Locs, Ctx.getTE());
// Find number of covering value and concrete values for the locations
// expanded from the given location.
unsigned CSCount = 0, CTCount = 0;
ValueTableMap &OTM = getForwardValOut();
for (auto &X : Locs) {
LSValue &V = Ctx.getLSValue(OTM[Ctx.getLSLocationBit(X)]);
if (V.isCoveringValue()) {
++CSCount;
continue;
}
++CTCount;
}
if (CSCount == Locs.size())
return ValueState::CoverValues;
if (CTCount == Locs.size())
return ValueState::ConcreteValues;
return ValueState::CoverAndConcreteValues;
}
void BlockState::processRead(RLEContext &Ctx, SILInstruction *I, SILValue Mem,
SILValue Val, RLEKind Kind) {
// Initialize the LSLocation.
LSLocation L(Mem);
// If we cant figure out the Base or Projection Path for the read, simply
// ignore it for now.
if (!L.isValid())
return;
// Expand the given LSLocation and Val into individual fields and process
// them as separate reads.
LSLocationList Locs;
LSLocation::expand(L, &I->getModule(), Locs, Ctx.getTE());
// Are we computing available set ?.
if (isComputeAvailSet(Kind)) {
for (auto &X : Locs) {
if (isTrackingLSLocation(Ctx.getLSLocationBit(X)))
continue;
updateForwardSetForRead(Ctx, Ctx.getLSLocationBit(X));
}
return;
}
// Are we computing available values ?.
bool CanForward = true;
if (isComputeAvailValue(Kind) || isPerformRLE(Kind)) {
LSValueList Vals;
LSValue::expand(Val, &I->getModule(), Vals, Ctx.getTE());
for (unsigned i = 0; i < Locs.size(); ++i) {
if (isTrackingLSLocation(Ctx.getLSLocationBit(Locs[i])))
continue;
updateForwardValForRead(Ctx, Ctx.getLSLocationBit(Locs[i]),
Ctx.getLSValueBit(Vals[i]));
CanForward = false;
}
}
// Simply return if we are not performing RLE or we do not have all the
// values available to perform RLE.
if (!isPerformRLE(Kind) || !CanForward)
return;
// Lastly, forward value to the load.
setupRLE(Ctx, I, Mem);
}
bool RLEContext::gatherLocationValues(SILBasicBlock *BB, LSLocation &L,
LSLocationValueMap &Values,
ValueTableMap &VM) {
LSLocationSet CSLocs;
LSLocationList Locs;
LSLocation::expand(L, &BB->getModule(), Locs, TE);
auto *Mod = &BB->getModule();
// Find the locations that this basic block defines and the locations which
// we do not have a concrete value in the current basic block.
for (auto &X : Locs) {
Values[X] = getLSValue(VM[getLSLocationBit(X)]);
if (!Values[X].isCoveringValue())
continue;
CSLocs.insert(X);
}
// For locations which we do not have concrete values for in this basic
// block, try to reduce it to the minimum # of locations possible, this
// will help us to generate as few SILArguments as possible.
LSLocation::reduce(L, Mod, CSLocs, TE);
// To handle covering value, we need to go to the predecessors and
// materialize them there.
for (auto &X : CSLocs) {
SILValue V = computePredecessorLocationValue(BB, X);
if (!V)
return false;
// We've constructed a concrete value for the covering value. Expand and
// collect the newly created forwardable values.
LSLocationList Locs;
LSValueList Vals;
LSLocation::expand(X, Mod, Locs, TE);
LSValue::expand(V, Mod, Vals, TE);
for (unsigned i = 0; i < Locs.size(); ++i) {
Values[Locs[i]] = Vals[i];
assert(Values[Locs[i]].isValid() && "Invalid load store value");
}
}
return true;
}
SILValue BlockState::reduceValuesAtEndOfBlock(RLEContext &Ctx, LSLocation &L) {
// First, collect current available locations and their corresponding values
// into a map.
LSLocationValueMap Values;
LSLocationList Locs;
LSLocation::expand(L, &BB->getModule(), Locs, Ctx.getTE());
// Find the values that this basic block defines and the locations which
// we do not have a concrete value in the current basic block.
ValueTableMap &OTM = getForwardValOut();
for (unsigned i = 0; i < Locs.size(); ++i) {
Values[Locs[i]] = Ctx.getLSValue(OTM[Ctx.getLSLocationBit(Locs[i])]);
}
// Second, reduce the available values into a single SILValue we can use to
// forward.
SILValue TheForwardingValue;
TheForwardingValue = LSValue::reduce(L, &BB->getModule(), Values,
BB->getTerminator(), Ctx.getTE());
/// Return the forwarding value.
return TheForwardingValue;
}
SILValue LSValue::reduce(LSLocation &Base, SILModule *M,
LSLocationValueMap &Values,
SILInstruction *InsertPt,
TypeExpansionAnalysis *TE) {
// Walk bottom up the projection tree, try to reason about how to construct
// a single SILValue out of all the available values for all the memory
// locations.
//
// First, get a list of all the leaf nodes and intermediate nodes for the
// Base memory location.
LSLocationList ALocs;
ProjectionPath &BasePath = Base.getPath().getValue();
for (const auto &P :
TE->getTypeExpansionProjectionPaths(Base.getType(), M, TEKind::TENode)) {
ALocs.push_back(LSLocation(Base.getBase(), P.getValue(), BasePath));
}
// Second, go from leaf nodes to their parents. This guarantees that at the
// point the parent is processed, its children have been processed already.
for (auto I = ALocs.rbegin(), E = ALocs.rend(); I != E; ++I) {
// This is a leaf node, we have a value for it.
//
// Reached the end of the projection tree, this is a leaf node.
LSLocationList FirstLevel;
I->getFirstLevelLSLocations(FirstLevel, M);
if (FirstLevel.empty())
continue;
// If this is a class reference type, we have reached end of the type tree.
if (I->getType().getClassOrBoundGenericClass())
continue;
// This is NOT a leaf node, we need to construct a value for it.
// There is only 1 children node and its value's projection path is not
// empty, keep stripping it.
auto Iter = FirstLevel.begin();
LSValue &FirstVal = Values[*Iter];
if (FirstLevel.size() == 1 && !FirstVal.hasEmptyProjectionPath()) {
Values[*I] = FirstVal.stripLastLevelProjection();
// We have a value for the parent, remove all the values for children.
removeLSLocations(Values, FirstLevel);
continue;
}
// If there are more than 1 children and all the children nodes have
// LSValues with the same base and non-empty projection path. we can get
// away by not extracting value for every single field.
//
// Simply create a new node with all the aggregated base value, i.e.
// stripping off the last level projection.
bool HasIdenticalValueBase = true;
SILValue FirstBase = FirstVal.getBase();
Iter = std::next(Iter);
for (auto EndIter = FirstLevel.end(); Iter != EndIter; ++Iter) {
LSValue &V = Values[*Iter];
HasIdenticalValueBase &= (FirstBase == V.getBase());
}
if (FirstLevel.size() > 1 && HasIdenticalValueBase &&
!FirstVal.hasEmptyProjectionPath()) {
Values[*I] = FirstVal.stripLastLevelProjection();
// We have a value for the parent, remove all the values for children.
removeLSLocations(Values, FirstLevel);
continue;
}
// In 3 cases do we need aggregation.
//
// 1. If there is only 1 child and we cannot strip off any projections,
// that means we need to create an aggregation.
//
// 2. There are multiple children and they have the same base, but empty
// projection paths.
//
// 3. Children have values from different bases, We need to create
// extractions and aggregation in this case.
//
llvm::SmallVector<SILValue, 8> Vals;
for (auto &X : FirstLevel) {
Vals.push_back(Values[X].materialize(InsertPt));
}
SILBuilder Builder(InsertPt);
// We use an auto-generated SILLocation for now.
// TODO: make the sil location more precise.
NullablePtr<swift::SILInstruction> AI =
Projection::createAggFromFirstLevelProjections(
Builder, RegularLocation::getAutoGeneratedLocation(), I->getType(),
Vals);
// This is the Value for the current node.
ProjectionPath P;
Values[*I] = LSValue(SILValue(AI.get()), P);
removeLSLocations(Values, FirstLevel);
// Keep iterating until we have reach the top-most level of the projection
// tree.
// i.e. the memory location represented by the Base.
}
assert(Values.size() == 1 && "Should have a single location this point");
// Finally materialize and return the forwarding SILValue.
return Values.begin()->second.materialize(InsertPt);
}
void
LSValue::reduceInner(LSLocation &Base, SILModule *M, LSLocationValueMap &Values,
SILInstruction *InsertPt) {
// If this is a class reference type, we have reached end of the type tree.
if (Base.getType(M).getClassOrBoundGenericClass())
return;
// This is a leaf node, we must have a value for it.
LSLocationList NextLevel;
Base.getNextLevelLSLocations(NextLevel, M);
if (NextLevel.empty())
return;
// This is not a leaf node, reduce the next level node one by one.
for (auto &X : NextLevel) {
LSValue::reduceInner(X, M, Values, InsertPt);
}
// This is NOT a leaf node, we need to construct a value for it.
auto Iter = NextLevel.begin();
LSValue &FirstVal = Values[*Iter];
// There is only 1 children node and its value's projection path is not
// empty, keep stripping it.
if (NextLevel.size() == 1 && !FirstVal.hasEmptyProjectionPath()) {
Values[Base] = FirstVal.stripLastLevelProjection();
// We have a value for the parent, remove all the values for children.
removeLSLocations(Values, NextLevel);
return;
}
bool HasIdenticalBase = true;
SILValue FirstBase = FirstVal.getBase();
for (auto &X : NextLevel) {
HasIdenticalBase &= (FirstBase == Values[X].getBase());
}
// This is NOT a leaf node and it has multiple children, but they have the
// same value base.
if (NextLevel.size() > 1 && HasIdenticalBase) {
if (!FirstVal.hasEmptyProjectionPath()) {
Values[Base] = FirstVal.stripLastLevelProjection();
// We have a value for the parent, remove all the values for children.
removeLSLocations(Values, NextLevel);
return;
}
}
// In 3 cases do we need aggregation.
//
// 1. If there is only 1 child and we cannot strip off any projections,
// that means we need to create an aggregation.
//
// 2. There are multiple children and they have the same base, but empty
// projection paths.
//
// 3. Children have values from different bases, We need to create
// extractions and aggregation in this case.
//
llvm::SmallVector<SILValue, 8> Vals;
for (auto &X : NextLevel) {
Vals.push_back(Values[X].materialize(InsertPt));
}
SILBuilder Builder(InsertPt);
Builder.setCurrentDebugScope(InsertPt->getFunction()->getDebugScope());
// We use an auto-generated SILLocation for now.
NullablePtr<swift::SILInstruction> AI =
Projection::createAggFromFirstLevelProjections(
Builder, RegularLocation::getAutoGeneratedLocation(),
Base.getType(M).getObjectType(),
Vals);
// This is the Value for the current base.
ProjectionPath P(Base.getType(M));
Values[Base] = LSValue(SILValue(AI.get()), P);
removeLSLocations(Values, NextLevel);
}
void DSEContext::processWrite(SILInstruction *I, SILValue Val, SILValue Mem,
DSEKind Kind) {
auto *S = getBlockState(I);
// Construct a LSLocation to represent the memory read by this instruction.
// NOTE: The base will point to the actual object this inst is accessing,
// not this particular field.
//
// e.g. %1 = alloc_stack $S
// %2 = struct_element_addr %1, #a
// store %3 to %2 : $*Int
//
// Base will point to %1, but not %2. Projection path will indicate which
// field is accessed.
//
// This will make comparison between locations easier. This eases the
// implementation of intersection operator in the data flow.
LSLocation L;
if (BaseToLocIndex.find(Mem) != BaseToLocIndex.end()) {
L = BaseToLocIndex[Mem];
} else {
SILValue UO = getUnderlyingObject(Mem);
L = LSLocation(UO, ProjectionPath::getProjectionPath(UO, Mem));
}
// If we can't figure out the Base or Projection Path for the store
// instruction, simply ignore it.
if (!L.isValid())
return;
// Expand the given Mem into individual fields and process them as separate
// writes.
bool Dead = true;
LSLocationList Locs;
LSLocation::expand(L, Mod, Locs, TE);
llvm::SmallBitVector V(Locs.size());
// Are we computing max store set.
if (isComputeMaxStoreSet(Kind)) {
for (auto &E : Locs) {
// Update the max store set for the basic block.
processWriteForMaxStoreSet(S, getLocationBit(E));
}
return;
}
// Are we computing genset and killset.
if (isBuildingGenKillSet(Kind)) {
for (auto &E : Locs) {
// Only building the gen and kill sets here.
processWriteForGenKillSet(S, getLocationBit(E));
}
// Data flow has not stabilized, do not perform the DSE just yet.
return;
}
// We are doing the actual DSE.
assert(isPerformingDSE(Kind) && "Invalid computation kind");
unsigned idx = 0;
for (auto &E : Locs) {
// This is the last iteration, compute BBWriteSetOut and perform the dead
// store elimination.
if (processWriteForDSE(S, getLocationBit(E)))
V.set(idx);
Dead &= V.test(idx);
++idx;
}
// Fully dead store - stores to all the components are dead, therefore this
// instruction is dead.
if (Dead) {
DEBUG(llvm::dbgs() << "Instruction Dead: " << *I << "\n");
S->DeadStores.insert(I);
++NumDeadStores;
return;
}
// Partial dead store - stores to some locations are dead, but not all. This
// is a partially dead store. Also at this point we know what locations are
// dead.
llvm::DenseSet<LSLocation> Alives;
if (V.any()) {
// Take out locations that are dead.
for (unsigned i = 0; i < V.size(); ++i) {
if (V.test(i))
continue;
// This location is alive.
Alives.insert(Locs[i]);
}
// Try to create as few aggregated stores as possible out of the locations.
LSLocation::reduce(L, Mod, Alives);
// Oops, we have too many smaller stores generated, bail out.
if (Alives.size() > MaxPartialStoreCount)
return;
// At this point, we are performing a partial dead store elimination.
//
// Locations here have a projection path from their Base, but this
// particular instruction may not be accessing the base, so we need to
// *rebase* the locations w.r.t. to the current instruction.
SILValue B = Locs[0].getBase();
Optional<ProjectionPath> BP = ProjectionPath::getProjectionPath(B, Mem);
// Strip off the projection path from base to the accessed field.
for (auto &X : Alives) {
X.removePathPrefix(BP);
}
// We merely setup the remaining live stores, but do not materialize in IR
// yet, These stores will be materialized before the algorithm exits.
for (auto &X : Alives) {
SILValue Value = X.getPath()->createExtract(Val, I, true);
SILValue Addr = X.getPath()->createExtract(Mem, I, false);
S->LiveAddr.insert(Addr);
S->LiveStores[Addr] = Value;
}
// Lastly, mark the old store as dead.
DEBUG(llvm::dbgs() << "Instruction Partially Dead: " << *I << "\n");
S->DeadStores.insert(I);
++NumPartialDeadStores;
}
}