return false;
}
-/// Return true if the Value is a gc reference type which is potentially used
-/// after the instruction 'loc'. This is only used with the edge reachability
-/// liveness code. Note: It is assumed the V dominates loc.
-static bool isLiveGCReferenceAt(Value &V, Instruction *loc, DominatorTree &DT,
- LoopInfo *LI) {
- if (!isGCPointerType(V.getType()))
- return false;
-
- if (V.use_empty())
- return false;
-
- // Given assumption that V dominates loc, this may be live
- return true;
+// Return true if this type is one which a) is a gc pointer or contains a GC
+// pointer and b) is of a type this code expects to encounter as a live value.
+// (The insertion code will assert that a type which matches (a) and not (b)
+// is not encountered.)
+static bool isHandledGCPointerType(Type *T) {
+ // We fully support gc pointers
+ if (isGCPointerType(T))
+ return true;
+ // We partially support vectors of gc pointers. The code will assert if it
+ // can't handle something.
+ if (auto VT = dyn_cast<VectorType>(T))
+ if (isGCPointerType(VT->getElementType()))
+ return true;
+ return false;
}
#ifndef NDEBUG
-static bool isAggWhichContainsGCPtrType(Type *Ty) {
+/// Returns true if this type contains a gc pointer whether we know how to
+/// handle that type or not.
+static bool containsGCPtrType(Type *Ty) {
+ if(isGCPointerType(Ty))
+ return true;
if (VectorType *VT = dyn_cast<VectorType>(Ty))
return isGCPointerType(VT->getScalarType());
if (ArrayType *AT = dyn_cast<ArrayType>(Ty))
- return isGCPointerType(AT->getElementType()) ||
- isAggWhichContainsGCPtrType(AT->getElementType());
+ return containsGCPtrType(AT->getElementType());
if (StructType *ST = dyn_cast<StructType>(Ty))
return std::any_of(ST->subtypes().begin(), ST->subtypes().end(),
[](Type *SubType) {
- return isGCPointerType(SubType) ||
- isAggWhichContainsGCPtrType(SubType);
+ return containsGCPtrType(SubType);
});
return false;
}
+
+// Returns true if this is a type which a) is a gc pointer or contains a GC
+// pointer and b) is of a type which the code doesn't expect (i.e. first class
+// aggregates). Used to trip assertions.
+static bool isUnhandledGCPointerType(Type *Ty) {
+ return containsGCPtrType(Ty) && !isHandledGCPointerType(Ty);
+}
#endif
+/// Return true if the Value is a gc reference type which is potentially used
+/// after the instruction 'loc'. This is only used with the edge reachability
+/// liveness code. Note: It is assumed the V dominates loc.
+static bool isLiveGCReferenceAt(Value &V, Instruction *Loc, DominatorTree &DT,
+ LoopInfo *LI) {
+ if (!isHandledGCPointerType(V.getType()))
+ return false;
+
+ if (V.use_empty())
+ return false;
+
+ // Given assumption that V dominates loc, this may be live
+ return true;
+}
+
// Conservatively identifies any definitions which might be live at the
// given instruction. The analysis is performed immediately before the
// given instruction. Values defined by that instruction are not considered
// Are there any gc pointer arguments live over this point? This needs to be
// special cased since arguments aren't defined in basic blocks.
for (Argument &arg : F->args()) {
- assert(!isAggWhichContainsGCPtrType(arg.getType()) &&
+ assert(!isUnhandledGCPointerType(arg.getType()) &&
"support for FCA unimplemented");
if (is_live_gc_reference(arg)) {
break;
}
- assert(!isAggWhichContainsGCPtrType(inst.getType()) &&
+ assert(!isUnhandledGCPointerType(inst.getType()) &&
"support for FCA unimplemented");
if (is_live_gc_reference(inst)) {
result.liveset = liveset;
}
-/// True iff this value is the null pointer constant (of any pointer type)
-static bool LLVM_ATTRIBUTE_UNUSED isNullConstant(Value *V) {
- return isa<Constant>(V) && isa<PointerType>(V->getType()) &&
- cast<Constant>(V)->isNullValue();
+/// If we can trivially determine that this vector contains only base pointers,
+/// return the base instruction.
+static Value *findBaseOfVector(Value *I) {
+ assert(I->getType()->isVectorTy() &&
+ cast<VectorType>(I->getType())->getElementType()->isPointerTy() &&
+ "Illegal to ask for the base pointer of a non-pointer type");
+
+ // Each case parallels findBaseDefiningValue below, see that code for
+ // detailed motivation.
+
+ if (isa<Argument>(I))
+ // An incoming argument to the function is a base pointer
+ return I;
+
+ // We shouldn't see the address of a global as a vector value?
+ assert(!isa<GlobalVariable>(I) &&
+ "unexpected global variable found in base of vector");
+
+ // inlining could possibly introduce phi node that contains
+ // undef if callee has multiple returns
+ if (isa<UndefValue>(I))
+ // utterly meaningless, but useful for dealing with partially optimized
+ // code.
+ return I;
+
+ // Due to inheritance, this must be _after_ the global variable and undef
+ // checks
+ if (Constant *Con = dyn_cast<Constant>(I)) {
+ assert(!isa<GlobalVariable>(I) && !isa<UndefValue>(I) &&
+ "order of checks wrong!");
+ assert(Con->isNullValue() && "null is the only case which makes sense");
+ return Con;
+ }
+
+ if (isa<LoadInst>(I))
+ return I;
+
+ // Note: This code is currently rather incomplete. We are essentially only
+ // handling cases where the vector element is trivially a base pointer. We
+ // need to update the entire base pointer construction algorithm to know how
+ // to track vector elements and potentially scalarize, but the case which
+ // would motivate the work hasn't shown up in real workloads yet.
+ llvm_unreachable("no base found for vector element");
}
/// Helper function for findBasePointer - Will return a value which either a)
assert(I->getType()->isPointerTy() &&
"Illegal to ask for the base pointer of a non-pointer type");
- // There are instructions which can never return gc pointer values. Sanity
- // check
- // that this is actually true.
- assert(!isa<InsertElementInst>(I) && !isa<ExtractElementInst>(I) &&
- !isa<ShuffleVectorInst>(I) && "Vector types are not gc pointers");
- assert((!isa<Instruction>(I) || isa<InvokeInst>(I) ||
- !cast<Instruction>(I)->isTerminator()) &&
- "With the exception of invoke terminators don't define values");
- assert(!isa<StoreInst>(I) && !isa<FenceInst>(I) &&
- "Can't be definitions to start with");
- assert(!isa<ICmpInst>(I) && !isa<FCmpInst>(I) &&
- "Comparisons don't give ops");
- // There's a bunch of instructions which just don't make sense to apply to
- // a pointer. The only valid reason for this would be pointer bit
- // twiddling which we're just not going to support.
- assert((!isa<Instruction>(I) || !cast<Instruction>(I)->isBinaryOp()) &&
- "Binary ops on pointer values are meaningless. Unless your "
- "bit-twiddling which we don't support");
-
- if (Argument *Arg = dyn_cast<Argument>(I)) {
+ // This case is a bit of a hack - it only handles extracts from vectors which
+ // trivially contain only base pointers. See note inside the function for
+ // how to improve this.
+ if (auto *EEI = dyn_cast<ExtractElementInst>(I)) {
+ Value *VectorOperand = EEI->getVectorOperand();
+ Value *VectorBase = findBaseOfVector(VectorOperand);
+ (void)VectorBase;
+ assert(VectorBase && "extract element not known to be a trivial base");
+ return EEI;
+ }
+
+ if (isa<Argument>(I))
// An incoming argument to the function is a base pointer
// We should have never reached here if this argument isn't an gc value
- assert(Arg->getType()->isPointerTy() &&
- "Base for pointer must be another pointer");
- return Arg;
- }
+ return I;
- if (GlobalVariable *global = dyn_cast<GlobalVariable>(I)) {
+ if (isa<GlobalVariable>(I))
// base case
- assert(global->getType()->isPointerTy() &&
- "Base for pointer must be another pointer");
- return global;
- }
+ return I;
// inlining could possibly introduce phi node that contains
// undef if callee has multiple returns
- if (UndefValue *undef = dyn_cast<UndefValue>(I)) {
- assert(undef->getType()->isPointerTy() &&
- "Base for pointer must be another pointer");
- return undef; // utterly meaningless, but useful for dealing with
- // partially optimized code.
- }
+ if (isa<UndefValue>(I))
+ // utterly meaningless, but useful for dealing with
+ // partially optimized code.
+ return I;
// Due to inheritance, this must be _after_ the global variable and undef
// checks
- if (Constant *con = dyn_cast<Constant>(I)) {
+ if (Constant *Con = dyn_cast<Constant>(I)) {
assert(!isa<GlobalVariable>(I) && !isa<UndefValue>(I) &&
"order of checks wrong!");
// Note: Finding a constant base for something marked for relocation
// off a potentially null value and have proven it null. We also use
// null pointers in dead paths of relocation phis (which we might later
// want to find a base pointer for).
- assert(con->getType()->isPointerTy() &&
- "Base for pointer must be another pointer");
- assert(con->isNullValue() && "null is the only case which makes sense");
- return con;
+ assert(isa<ConstantPointerNull>(Con) &&
+ "null is the only case which makes sense");
+ return Con;
}
if (CastInst *CI = dyn_cast<CastInst>(I)) {
- Value *def = CI->stripPointerCasts();
- assert(def->getType()->isPointerTy() &&
- "Base for pointer must be another pointer");
+ Value *Def = CI->stripPointerCasts();
// If we find a cast instruction here, it means we've found a cast which is
// not simply a pointer cast (i.e. an inttoptr). We don't know how to
// handle int->ptr conversion.
- assert(!isa<CastInst>(def) && "shouldn't find another cast here");
- return findBaseDefiningValue(def);
+ assert(!isa<CastInst>(Def) && "shouldn't find another cast here");
+ return findBaseDefiningValue(Def);
}
- if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
- if (LI->getType()->isPointerTy()) {
- Value *Op = LI->getOperand(0);
- (void)Op;
- // Has to be a pointer to an gc object, or possibly an array of such?
- assert(Op->getType()->isPointerTy());
- return LI; // The value loaded is an gc base itself
- }
- }
- if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(I)) {
- Value *Op = GEP->getOperand(0);
- if (Op->getType()->isPointerTy()) {
- return findBaseDefiningValue(Op); // The base of this GEP is the base
- }
- }
+ if (isa<LoadInst>(I))
+ return I; // The value loaded is an gc base itself
- if (AllocaInst *alloc = dyn_cast<AllocaInst>(I)) {
- // An alloca represents a conceptual stack slot. It's the slot itself
- // that the GC needs to know about, not the value in the slot.
- assert(alloc->getType()->isPointerTy() &&
- "Base for pointer must be another pointer");
- return alloc;
- }
+ if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(I))
+ // The base of this GEP is the base
+ return findBaseDefiningValue(GEP->getPointerOperand());
if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
switch (II->getIntrinsicID()) {
+ case Intrinsic::experimental_gc_result_ptr:
default:
// fall through to general call handling
break;
case Intrinsic::experimental_gc_result_float:
case Intrinsic::experimental_gc_result_int:
llvm_unreachable("these don't produce pointers");
- case Intrinsic::experimental_gc_result_ptr:
- // This is just a special case of the CallInst check below to handle a
- // statepoint with deopt args which hasn't been rewritten for GC yet.
- // TODO: Assert that the statepoint isn't rewritten yet.
- return II;
case Intrinsic::experimental_gc_relocate: {
// Rerunning safepoint insertion after safepoints are already
// inserted is not supported. It could probably be made to work,
// We assume that functions in the source language only return base
// pointers. This should probably be generalized via attributes to support
// both source language and internal functions.
- if (CallInst *call = dyn_cast<CallInst>(I)) {
- assert(call->getType()->isPointerTy() &&
- "Base for pointer must be another pointer");
- return call;
- }
- if (InvokeInst *invoke = dyn_cast<InvokeInst>(I)) {
- assert(invoke->getType()->isPointerTy() &&
- "Base for pointer must be another pointer");
- return invoke;
- }
+ if (isa<CallInst>(I) || isa<InvokeInst>(I))
+ return I;
// I have absolutely no idea how to implement this part yet. It's not
// neccessarily hard, I just haven't really looked at it yet.
assert(!isa<LandingPadInst>(I) && "Landing Pad is unimplemented");
- if (AtomicCmpXchgInst *cas = dyn_cast<AtomicCmpXchgInst>(I)) {
+ if (isa<AtomicCmpXchgInst>(I))
// A CAS is effectively a atomic store and load combined under a
// predicate. From the perspective of base pointers, we just treat it
- // like a load. We loaded a pointer from a address in memory, that value
- // had better be a valid base pointer.
- return cas->getPointerOperand();
- }
- if (AtomicRMWInst *atomic = dyn_cast<AtomicRMWInst>(I)) {
- assert(AtomicRMWInst::Xchg == atomic->getOperation() &&
- "All others are binary ops which don't apply to base pointers");
- // semantically, a load, store pair. Treat it the same as a standard load
- return atomic->getPointerOperand();
- }
+ // like a load.
+ return I;
+
+ assert(!isa<AtomicRMWInst>(I) && "Xchg handled above, all others are "
+ "binary ops which don't apply to pointers");
// The aggregate ops. Aggregates can either be in the heap or on the
// stack, but in either case, this is simply a field load. As a result,
// this is a defining definition of the base just like a load is.
- if (ExtractValueInst *ev = dyn_cast<ExtractValueInst>(I)) {
- return ev;
- }
+ if (isa<ExtractValueInst>(I))
+ return I;
// We should never see an insert vector since that would require we be
// tracing back a struct value not a pointer value.
// return a value which dynamically selects from amoung several base
// derived pointers (each with it's own base potentially). It's the job of
// the caller to resolve these.
- if (SelectInst *select = dyn_cast<SelectInst>(I)) {
- return select;
- }
-
- return cast<PHINode>(I);
+ assert((isa<SelectInst>(I) || isa<PHINode>(I)) &&
+ "missing instruction case in findBaseDefiningValing");
+ return I;
}
/// Returns the base defining value for this value.
-static Value *findBaseDefiningValueCached(Value *I, DefiningValueMapTy &cache) {
- Value *&Cached = cache[I];
+static Value *findBaseDefiningValueCached(Value *I, DefiningValueMapTy &Cache) {
+ Value *&Cached = Cache[I];
if (!Cached) {
Cached = findBaseDefiningValue(I);
}
- assert(cache[I] != nullptr);
+ assert(Cache[I] != nullptr);
if (TraceLSP) {
- errs() << "fBDV-cached: " << I->getName() << " -> " << Cached->getName()
+ dbgs() << "fBDV-cached: " << I->getName() << " -> " << Cached->getName()
<< "\n";
}
return Cached;
/// Return a base pointer for this value if known. Otherwise, return it's
/// base defining value.
-static Value *findBaseOrBDV(Value *I, DefiningValueMapTy &cache) {
- Value *def = findBaseDefiningValueCached(I, cache);
- auto Found = cache.find(def);
- if (Found != cache.end()) {
+static Value *findBaseOrBDV(Value *I, DefiningValueMapTy &Cache) {
+ Value *Def = findBaseDefiningValueCached(I, Cache);
+ auto Found = Cache.find(Def);
+ if (Found != Cache.end()) {
// Either a base-of relation, or a self reference. Caller must check.
return Found->second;
}
// Only a BDV available
- return def;
+ return Def;
}
/// Given the result of a call to findBaseDefiningValue, or findBaseOrBDV,
/// is it known to be a base pointer? Or do we need to continue searching.
-static bool isKnownBaseResult(Value *v) {
- if (!isa<PHINode>(v) && !isa<SelectInst>(v)) {
+static bool isKnownBaseResult(Value *V) {
+ if (!isa<PHINode>(V) && !isa<SelectInst>(V)) {
// no recursion possible
return true;
}
- if (cast<Instruction>(v)->getMetadata("is_base_value")) {
+ if (isa<Instruction>(V) &&
+ cast<Instruction>(V)->getMetadata("is_base_value")) {
// This is a previously inserted base phi or select. We know
// that this is a base value.
return true;
}
PhiState(Value *b) : status(Base), base(b) {}
PhiState() : status(Unknown), base(nullptr) {}
- PhiState(const PhiState &other) : status(other.status), base(other.base) {
- assert(status != Base || base);
- }
Status getStatus() const { return status; }
Value *getBase() const { return base; }
// have reached conflict state. The current version seems too conservative.
bool progress = true;
- size_t oldSize = 0;
while (progress) {
- oldSize = states.size();
+#ifndef NDEBUG
+ size_t oldSize = states.size();
+#endif
progress = false;
// We're only changing keys in this loop, thus safe to keep iterators
for (auto Pair : states) {
}
// Insert Phis for all conflicts
- // Only changing keys in 'states', thus safe to keep iterators
+ // We want to keep naming deterministic in the loop that follows, so
+ // sort the keys before iteration. This is useful in allowing us to
+ // write stable tests. Note that there is no invalidation issue here.
+ SmallVector<Value*, 16> Keys;
+ Keys.reserve(states.size());
for (auto Pair : states) {
- Instruction *v = cast<Instruction>(Pair.first);
- PhiState state = Pair.second;
+ Value *V = Pair.first;
+ Keys.push_back(V);
+ }
+ std::sort(Keys.begin(), Keys.end(), order_by_name);
+ // TODO: adjust naming patterns to avoid this order of iteration dependency
+ for (Value *V : Keys) {
+ Instruction *v = cast<Instruction>(V);
+ PhiState state = states[V];
assert(!isKnownBaseResult(v) && "why did it get added?");
assert(!state.isUnknown() && "Optimistic algorithm didn't complete!");
- if (state.isConflict()) {
- if (isa<PHINode>(v)) {
- int num_preds =
- std::distance(pred_begin(v->getParent()), pred_end(v->getParent()));
- assert(num_preds > 0 && "how did we reach here");
- PHINode *phi = PHINode::Create(v->getType(), num_preds, "base_phi", v);
- NewInsertedDefs.insert(phi);
- // Add metadata marking this as a base value
- auto *const_1 = ConstantInt::get(
- Type::getInt32Ty(
- v->getParent()->getParent()->getParent()->getContext()),
- 1);
- auto MDConst = ConstantAsMetadata::get(const_1);
- MDNode *md = MDNode::get(
- v->getParent()->getParent()->getParent()->getContext(), MDConst);
- phi->setMetadata("is_base_value", md);
- states[v] = PhiState(PhiState::Conflict, phi);
- } else if (SelectInst *sel = dyn_cast<SelectInst>(v)) {
- // The undef will be replaced later
- UndefValue *undef = UndefValue::get(sel->getType());
- SelectInst *basesel = SelectInst::Create(sel->getCondition(), undef,
- undef, "base_select", sel);
- NewInsertedDefs.insert(basesel);
- // Add metadata marking this as a base value
- auto *const_1 = ConstantInt::get(
- Type::getInt32Ty(
- v->getParent()->getParent()->getParent()->getContext()),
- 1);
- auto MDConst = ConstantAsMetadata::get(const_1);
- MDNode *md = MDNode::get(
- v->getParent()->getParent()->getParent()->getContext(), MDConst);
- basesel->setMetadata("is_base_value", md);
- states[v] = PhiState(PhiState::Conflict, basesel);
- } else
- llvm_unreachable("unknown conflict type");
+ if (!state.isConflict())
+ continue;
+
+ if (isa<PHINode>(v)) {
+ int num_preds =
+ std::distance(pred_begin(v->getParent()), pred_end(v->getParent()));
+ assert(num_preds > 0 && "how did we reach here");
+ PHINode *phi = PHINode::Create(v->getType(), num_preds, "base_phi", v);
+ NewInsertedDefs.insert(phi);
+ // Add metadata marking this as a base value
+ auto *const_1 = ConstantInt::get(
+ Type::getInt32Ty(
+ v->getParent()->getParent()->getParent()->getContext()),
+ 1);
+ auto MDConst = ConstantAsMetadata::get(const_1);
+ MDNode *md = MDNode::get(
+ v->getParent()->getParent()->getParent()->getContext(), MDConst);
+ phi->setMetadata("is_base_value", md);
+ states[v] = PhiState(PhiState::Conflict, phi);
+ } else {
+ SelectInst *sel = cast<SelectInst>(v);
+ // The undef will be replaced later
+ UndefValue *undef = UndefValue::get(sel->getType());
+ SelectInst *basesel = SelectInst::Create(sel->getCondition(), undef,
+ undef, "base_select", sel);
+ NewInsertedDefs.insert(basesel);
+ // Add metadata marking this as a base value
+ auto *const_1 = ConstantInt::get(
+ Type::getInt32Ty(
+ v->getParent()->getParent()->getParent()->getContext()),
+ 1);
+ auto MDConst = ConstantAsMetadata::get(const_1);
+ MDNode *md = MDNode::get(
+ v->getParent()->getParent()->getParent()->getContext(), MDConst);
+ basesel->setMetadata("is_base_value", md);
+ states[v] = PhiState(PhiState::Conflict, basesel);
}
}
assert(!isKnownBaseResult(v) && "why did it get added?");
assert(!state.isUnknown() && "Optimistic algorithm didn't complete!");
- if (state.isConflict()) {
- if (PHINode *basephi = dyn_cast<PHINode>(state.getBase())) {
- PHINode *phi = cast<PHINode>(v);
- unsigned NumPHIValues = phi->getNumIncomingValues();
- for (unsigned i = 0; i < NumPHIValues; i++) {
- Value *InVal = phi->getIncomingValue(i);
- BasicBlock *InBB = phi->getIncomingBlock(i);
-
- // If we've already seen InBB, add the same incoming value
- // we added for it earlier. The IR verifier requires phi
- // nodes with multiple entries from the same basic block
- // to have the same incoming value for each of those
- // entries. If we don't do this check here and basephi
- // has a different type than base, we'll end up adding two
- // bitcasts (and hence two distinct values) as incoming
- // values for the same basic block.
-
- int blockIndex = basephi->getBasicBlockIndex(InBB);
- if (blockIndex != -1) {
- Value *oldBase = basephi->getIncomingValue(blockIndex);
- basephi->addIncoming(oldBase, InBB);
+ if (!state.isConflict())
+ continue;
+
+ if (PHINode *basephi = dyn_cast<PHINode>(state.getBase())) {
+ PHINode *phi = cast<PHINode>(v);
+ unsigned NumPHIValues = phi->getNumIncomingValues();
+ for (unsigned i = 0; i < NumPHIValues; i++) {
+ Value *InVal = phi->getIncomingValue(i);
+ BasicBlock *InBB = phi->getIncomingBlock(i);
+
+ // If we've already seen InBB, add the same incoming value
+ // we added for it earlier. The IR verifier requires phi
+ // nodes with multiple entries from the same basic block
+ // to have the same incoming value for each of those
+ // entries. If we don't do this check here and basephi
+ // has a different type than base, we'll end up adding two
+ // bitcasts (and hence two distinct values) as incoming
+ // values for the same basic block.
+
+ int blockIndex = basephi->getBasicBlockIndex(InBB);
+ if (blockIndex != -1) {
+ Value *oldBase = basephi->getIncomingValue(blockIndex);
+ basephi->addIncoming(oldBase, InBB);
#ifndef NDEBUG
- Value *base = findBaseOrBDV(InVal, cache);
- if (!isKnownBaseResult(base)) {
- // Either conflict or base.
- assert(states.count(base));
- base = states[base].getBase();
- assert(base != nullptr && "unknown PhiState!");
- assert(NewInsertedDefs.count(base) &&
- "should have already added this in a prev. iteration!");
- }
-
- // In essense this assert states: the only way two
- // values incoming from the same basic block may be
- // different is by being different bitcasts of the same
- // value. A cleanup that remains TODO is changing
- // findBaseOrBDV to return an llvm::Value of the correct
- // type (and still remain pure). This will remove the
- // need to add bitcasts.
- assert(base->stripPointerCasts() == oldBase->stripPointerCasts() &&
- "sanity -- findBaseOrBDV should be pure!");
-#endif
- continue;
- }
-
- // Find either the defining value for the PHI or the normal base for
- // a non-phi node
Value *base = findBaseOrBDV(InVal, cache);
if (!isKnownBaseResult(base)) {
// Either conflict or base.
assert(states.count(base));
base = states[base].getBase();
assert(base != nullptr && "unknown PhiState!");
+ assert(NewInsertedDefs.count(base) &&
+ "should have already added this in a prev. iteration!");
}
- assert(base && "can't be null");
- // Must use original input BB since base may not be Instruction
- // The cast is needed since base traversal may strip away bitcasts
- if (base->getType() != basephi->getType()) {
- base = new BitCastInst(base, basephi->getType(), "cast",
- InBB->getTerminator());
- NewInsertedDefs.insert(base);
- }
- basephi->addIncoming(base, InBB);
+
+ // In essense this assert states: the only way two
+ // values incoming from the same basic block may be
+ // different is by being different bitcasts of the same
+ // value. A cleanup that remains TODO is changing
+ // findBaseOrBDV to return an llvm::Value of the correct
+ // type (and still remain pure). This will remove the
+ // need to add bitcasts.
+ assert(base->stripPointerCasts() == oldBase->stripPointerCasts() &&
+ "sanity -- findBaseOrBDV should be pure!");
+#endif
+ continue;
}
- assert(basephi->getNumIncomingValues() == NumPHIValues);
- } else if (SelectInst *basesel = dyn_cast<SelectInst>(state.getBase())) {
- SelectInst *sel = cast<SelectInst>(v);
- // Operand 1 & 2 are true, false path respectively. TODO: refactor to
- // something more safe and less hacky.
- for (int i = 1; i <= 2; i++) {
- Value *InVal = sel->getOperand(i);
- // Find either the defining value for the PHI or the normal base for
- // a non-phi node
- Value *base = findBaseOrBDV(InVal, cache);
- if (!isKnownBaseResult(base)) {
- // Either conflict or base.
- assert(states.count(base));
- base = states[base].getBase();
- assert(base != nullptr && "unknown PhiState!");
- }
- assert(base && "can't be null");
- // Must use original input BB since base may not be Instruction
- // The cast is needed since base traversal may strip away bitcasts
- if (base->getType() != basesel->getType()) {
- base = new BitCastInst(base, basesel->getType(), "cast", basesel);
- NewInsertedDefs.insert(base);
- }
- basesel->setOperand(i, base);
+
+ // Find either the defining value for the PHI or the normal base for
+ // a non-phi node
+ Value *base = findBaseOrBDV(InVal, cache);
+ if (!isKnownBaseResult(base)) {
+ // Either conflict or base.
+ assert(states.count(base));
+ base = states[base].getBase();
+ assert(base != nullptr && "unknown PhiState!");
}
- } else
- llvm_unreachable("unexpected conflict type");
+ assert(base && "can't be null");
+ // Must use original input BB since base may not be Instruction
+ // The cast is needed since base traversal may strip away bitcasts
+ if (base->getType() != basephi->getType()) {
+ base = new BitCastInst(base, basephi->getType(), "cast",
+ InBB->getTerminator());
+ NewInsertedDefs.insert(base);
+ }
+ basephi->addIncoming(base, InBB);
+ }
+ assert(basephi->getNumIncomingValues() == NumPHIValues);
+ } else {
+ SelectInst *basesel = cast<SelectInst>(state.getBase());
+ SelectInst *sel = cast<SelectInst>(v);
+ // Operand 1 & 2 are true, false path respectively. TODO: refactor to
+ // something more safe and less hacky.
+ for (int i = 1; i <= 2; i++) {
+ Value *InVal = sel->getOperand(i);
+ // Find either the defining value for the PHI or the normal base for
+ // a non-phi node
+ Value *base = findBaseOrBDV(InVal, cache);
+ if (!isKnownBaseResult(base)) {
+ // Either conflict or base.
+ assert(states.count(base));
+ base = states[base].getBase();
+ assert(base != nullptr && "unknown PhiState!");
+ }
+ assert(base && "can't be null");
+ // Must use original input BB since base may not be Instruction
+ // The cast is needed since base traversal may strip away bitcasts
+ if (base->getType() != basesel->getType()) {
+ base = new BitCastInst(base, basesel->getType(), "cast", basesel);
+ NewInsertedDefs.insert(base);
+ }
+ basesel->setOperand(i, base);
+ }
}
}
DenseMap<llvm::Value *, llvm::Value *> &PointerToBase,
DominatorTree *DT, DefiningValueMapTy &DVCache,
DenseSet<llvm::Value *> &NewInsertedDefs) {
- for (Value *ptr : live) {
+ // For the naming of values inserted to be deterministic - which makes for
+ // much cleaner and more stable tests - we need to assign an order to the
+ // live values. DenseSets do not provide a deterministic order across runs.
+ SmallVector<Value*, 64> Temp;
+ Temp.insert(Temp.end(), live.begin(), live.end());
+ std::sort(Temp.begin(), Temp.end(), order_by_name);
+ for (Value *ptr : Temp) {
Value *base = findBasePointer(ptr, DVCache, NewInsertedDefs);
assert(base && "failed to find base pointer");
PointerToBase[ptr] = base;
// If you see this trip and like to live really dangerously, the code should
// be correct, just with idioms the verifier can't handle. You can try
// disabling the verifier at your own substaintial risk.
- assert(!isNullConstant(base) && "the relocation code needs adjustment to "
- "handle the relocation of a null pointer "
- "constant without causing false positives "
- "in the safepoint ir verifier.");
+ assert(!isa<ConstantPointerNull>(base) &&
+ "the relocation code needs adjustment to handle the relocation of "
+ "a null pointer constant without causing false positives in the "
+ "safepoint ir verifier.");
}
}
/// statepointToken - statepoint instruction to which relocates should be
/// bound.
/// Builder - Llvm IR builder to be used to construct new calls.
-void CreateGCRelocates(ArrayRef<llvm::Value *> liveVariables,
- const int liveStart,
- ArrayRef<llvm::Value *> basePtrs,
- Instruction *statepointToken, IRBuilder<> Builder) {
-
+static void CreateGCRelocates(ArrayRef<llvm::Value *> liveVariables,
+ const int liveStart,
+ ArrayRef<llvm::Value *> basePtrs,
+ Instruction *statepointToken,
+ IRBuilder<> Builder) {
SmallVector<Instruction *, 64> NewDefs;
NewDefs.reserve(liveVariables.size());
Function &F, DominatorTree &DT, ArrayRef<Value *> live,
ArrayRef<struct PartiallyConstructedSafepointRecord> records) {
#ifndef NDEBUG
- int initialAllocaNum = 0;
-
- // record initial number of allocas
- for (inst_iterator itr = inst_begin(F), end = inst_end(F); itr != end;
- itr++) {
- if (isa<AllocaInst>(*itr))
- initialAllocaNum++;
- }
+ // record initial number of (static) allocas; we'll check we have the same
+ // number when we get done.
+ int InitialAllocaNum = 0;
+ for (auto I = F.getEntryBlock().begin(), E = F.getEntryBlock().end();
+ I != E; I++)
+ if (isa<AllocaInst>(*I))
+ InitialAllocaNum++;
#endif
// TODO-PERF: change data structures, reserve
// store must be inserted after load, otherwise store will be in alloca's
// use list and an extra load will be inserted before it
StoreInst *store = new StoreInst(def, alloca);
- if (isa<Instruction>(def)) {
- store->insertAfter(cast<Instruction>(def));
+ if (Instruction *inst = dyn_cast<Instruction>(def)) {
+ if (InvokeInst *invoke = dyn_cast<InvokeInst>(inst)) {
+ // InvokeInst is a TerminatorInst so the store need to be inserted
+ // into its normal destination block.
+ BasicBlock *normalDest = invoke->getNormalDest();
+ store->insertBefore(normalDest->getFirstNonPHI());
+ } else {
+ assert(!inst->isTerminator() &&
+ "The only TerminatorInst that can produce a value is "
+ "InvokeInst which is handled above.");
+ store->insertAfter(inst);
+ }
} else {
assert((isa<Argument>(def) || isa<GlobalVariable>(def) ||
- (isa<Constant>(def) && cast<Constant>(def)->isNullValue())) &&
+ isa<ConstantPointerNull>(def)) &&
"Must be argument or global");
store->insertAfter(cast<Instruction>(alloca));
}
}
#ifndef NDEBUG
- for (inst_iterator itr = inst_begin(F), end = inst_end(F); itr != end;
- itr++) {
- if (isa<AllocaInst>(*itr))
- initialAllocaNum--;
- }
- assert(initialAllocaNum == 0 && "We must not introduce any extra allocas");
+ for (auto I = F.getEntryBlock().begin(), E = F.getEntryBlock().end();
+ I != E; I++)
+ if (isa<AllocaInst>(*I))
+ InitialAllocaNum--;
+ assert(InitialAllocaNum == 0 && "We must not introduce any extra allocas");
#endif
}
assert(liveset.size() == PointerToBase.size());
}
+/// Remove any vector of pointers from the liveset by scalarizing them over the
+/// statepoint instruction. Adds the scalarized pieces to the liveset. It
+/// would be preferrable to include the vector in the statepoint itself, but
+/// the lowering code currently does not handle that. Extending it would be
+/// slightly non-trivial since it requires a format change. Given how rare
+/// such cases are (for the moment?) scalarizing is an acceptable comprimise.
+static void splitVectorValues(Instruction *StatepointInst,
+ StatepointLiveSetTy& LiveSet, DominatorTree &DT) {
+ SmallVector<Value *, 16> ToSplit;
+ for (Value *V : LiveSet)
+ if (isa<VectorType>(V->getType()))
+ ToSplit.push_back(V);
+
+ if (ToSplit.empty())
+ return;
+
+ Function &F = *(StatepointInst->getParent()->getParent());
+
+ DenseMap<Value*, AllocaInst*> AllocaMap;
+ // First is normal return, second is exceptional return (invoke only)
+ DenseMap<Value*, std::pair<Value*,Value*>> Replacements;
+ for (Value *V : ToSplit) {
+ LiveSet.erase(V);
+
+ AllocaInst *Alloca = new AllocaInst(V->getType(), "",
+ F.getEntryBlock().getFirstNonPHI());
+ AllocaMap[V] = Alloca;
+
+ VectorType *VT = cast<VectorType>(V->getType());
+ IRBuilder<> Builder(StatepointInst);
+ SmallVector<Value*, 16> Elements;
+ for (unsigned i = 0; i < VT->getNumElements(); i++)
+ Elements.push_back(Builder.CreateExtractElement(V, Builder.getInt32(i)));
+ LiveSet.insert(Elements.begin(), Elements.end());
+
+ auto InsertVectorReform = [&](Instruction *IP) {
+ Builder.SetInsertPoint(IP);
+ Builder.SetCurrentDebugLocation(IP->getDebugLoc());
+ Value *ResultVec = UndefValue::get(VT);
+ for (unsigned i = 0; i < VT->getNumElements(); i++)
+ ResultVec = Builder.CreateInsertElement(ResultVec, Elements[i],
+ Builder.getInt32(i));
+ return ResultVec;
+ };
+
+ if (isa<CallInst>(StatepointInst)) {
+ BasicBlock::iterator Next(StatepointInst);
+ Next++;
+ Instruction *IP = &*(Next);
+ Replacements[V].first = InsertVectorReform(IP);
+ Replacements[V].second = nullptr;
+ } else {
+ InvokeInst *Invoke = cast<InvokeInst>(StatepointInst);
+ // We've already normalized - check that we don't have shared destination
+ // blocks
+ BasicBlock *NormalDest = Invoke->getNormalDest();
+ assert(!isa<PHINode>(NormalDest->begin()));
+ BasicBlock *UnwindDest = Invoke->getUnwindDest();
+ assert(!isa<PHINode>(UnwindDest->begin()));
+ // Insert insert element sequences in both successors
+ Instruction *IP = &*(NormalDest->getFirstInsertionPt());
+ Replacements[V].first = InsertVectorReform(IP);
+ IP = &*(UnwindDest->getFirstInsertionPt());
+ Replacements[V].second = InsertVectorReform(IP);
+ }
+ }
+ for (Value *V : ToSplit) {
+ AllocaInst *Alloca = AllocaMap[V];
+
+ // Capture all users before we start mutating use lists
+ SmallVector<Instruction*, 16> Users;
+ for (User *U : V->users())
+ Users.push_back(cast<Instruction>(U));
+
+ for (Instruction *I : Users) {
+ if (auto Phi = dyn_cast<PHINode>(I)) {
+ for (unsigned i = 0; i < Phi->getNumIncomingValues(); i++)
+ if (V == Phi->getIncomingValue(i)) {
+ LoadInst *Load = new LoadInst(Alloca, "",
+ Phi->getIncomingBlock(i)->getTerminator());
+ Phi->setIncomingValue(i, Load);
+ }
+ } else {
+ LoadInst *Load = new LoadInst(Alloca, "", I);
+ I->replaceUsesOfWith(V, Load);
+ }
+ }
+
+ // Store the original value and the replacement value into the alloca
+ StoreInst *Store = new StoreInst(V, Alloca);
+ if (auto I = dyn_cast<Instruction>(V))
+ Store->insertAfter(I);
+ else
+ Store->insertAfter(Alloca);
+
+ // Normal return for invoke, or call return
+ Instruction *Replacement = cast<Instruction>(Replacements[V].first);
+ (new StoreInst(Replacement, Alloca))->insertAfter(Replacement);
+ // Unwind return for invoke only
+ Replacement = cast_or_null<Instruction>(Replacements[V].second);
+ if (Replacement)
+ (new StoreInst(Replacement, Alloca))->insertAfter(Replacement);
+ }
+
+ // apply mem2reg to promote alloca to SSA
+ SmallVector<AllocaInst*, 16> Allocas;
+ for (Value *V : ToSplit)
+ Allocas.push_back(AllocaMap[V]);
+ PromoteMemToReg(Allocas, DT);
+}
+
static bool insertParsePoints(Function &F, DominatorTree &DT, Pass *P,
SmallVectorImpl<CallSite> &toUpdate) {
#ifndef NDEBUG
SmallVector<Value *, 64> DeoptValues;
for (Use &U : StatepointCS.vm_state_args()) {
Value *Arg = cast<Value>(&U);
- if (isGCPointerType(Arg->getType()))
+ assert(!isUnhandledGCPointerType(Arg->getType()) &&
+ "support for FCA unimplemented");
+ if (isHandledGCPointerType(Arg->getType()))
DeoptValues.push_back(Arg);
}
insertUseHolderAfter(CS, DeoptValues, holders);
// site.
findLiveReferences(F, DT, P, toUpdate, records);
+ // Do a limited scalarization of any live at safepoint vector values which
+ // contain pointers. This enables this pass to run after vectorization at
+ // the cost of some possible performance loss. TODO: it would be nice to
+ // natively support vectors all the way through the backend so we don't need
+ // to scalarize here.
+ for (size_t i = 0; i < records.size(); i++) {
+ struct PartiallyConstructedSafepointRecord &info = records[i];
+ Instruction *statepoint = toUpdate[i].getInstruction();
+ splitVectorValues(cast<Instruction>(statepoint), info.liveset, DT);
+ }
+
// B) Find the base pointers for each live pointer
/* scope for caching */ {
// Cache the 'defining value' relation used in the computation and
if (!shouldRewriteStatepointsIn(F))
return false;
- // Gather all the statepoints which need rewritten.
+ DominatorTree &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree();
+
+ // Gather all the statepoints which need rewritten. Be careful to only
+ // consider those in reachable code since we need to ask dominance queries
+ // when rewriting. We'll delete the unreachable ones in a moment.
SmallVector<CallSite, 64> ParsePointNeeded;
+ bool HasUnreachableStatepoint = false;
for (Instruction &I : inst_range(F)) {
// TODO: only the ones with the flag set!
- if (isStatepoint(I))
- ParsePointNeeded.push_back(CallSite(&I));
+ if (isStatepoint(I)) {
+ if (DT.isReachableFromEntry(I.getParent()))
+ ParsePointNeeded.push_back(CallSite(&I));
+ else
+ HasUnreachableStatepoint = true;
+ }
}
+ bool MadeChange = false;
+
+ // Delete any unreachable statepoints so that we don't have unrewritten
+ // statepoints surviving this pass. This makes testing easier and the
+ // resulting IR less confusing to human readers. Rather than be fancy, we
+ // just reuse a utility function which removes the unreachable blocks.
+ if (HasUnreachableStatepoint)
+ MadeChange |= removeUnreachableBlocks(F);
+
// Return early if no work to do.
if (ParsePointNeeded.empty())
- return false;
+ return MadeChange;
+
+ // As a prepass, go ahead and aggressively destroy single entry phi nodes.
+ // These are created by LCSSA. They have the effect of increasing the size
+ // of liveness sets for no good reason. It may be harder to do this post
+ // insertion since relocations and base phis can confuse things.
+ for (BasicBlock &BB : F)
+ if (BB.getUniquePredecessor()) {
+ MadeChange = true;
+ FoldSingleEntryPHINodes(&BB);
+ }
- DominatorTree &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree();
- return insertParsePoints(F, DT, this, ParsePointNeeded);
+ MadeChange |= insertParsePoints(F, DT, this, ParsePointNeeded);
+ return MadeChange;
}