[flang][hlfir] do not propagate polymorphic temporary as allocatables #142609
+516
−680
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Polymorphic temporary are currently propagated as fir.ref<fir.class<fir.heap<>>> because their allocation may be delayed to the hlfir.assign copy (using realloc). This patch moves away from this and directly allocate the temp and propagate it as a fir.class. The polymorphic temporaries creating is also simplified by avoiding the need to call the runtime to setup the descriptor altogether (the runtime is still call for the allocation currently because alloca/allocmem do not support polymorphism). Rational: - this lower complexity and reduce the need to add codes to special case polymorphic handling - I want to move towards a point where where we can use fir.alloca/fir.allocmem to allocate polymorphic temporary so that LLVM can better understand the shape and size of the data after inlining. - I want to strengthen hlfir.declare box value result to not have the allocatable attribute to have the declare symbol attribute and the runtime attribute inside the descriptor in sync and avoid situation where the runtime thinks it can reallocate things that are not allocatable at that point of the program. This will fix a bug with SELECT TYPE lowering.
llvmbot
added
flang
|
@llvm/pr-subscribers-flang-codegen @llvm/pr-subscribers-flang-fir-hlfir Author: None (jeanPerier) ChangesPolymorphic temporary are currently propagated as fir.ref<fir.class<fir.heap<>>> because their allocation may be delayed to the hlfir.assign copy (using realloc). This patch moves away from this and directly allocate the temp and propagate it as a fir.class. The polymorphic temporaries creating is also simplified by avoiding the need to call the runtime to setup the descriptor altogether (the runtime is still call for the allocation currently because alloca/allocmem do not support polymorphism). Rational:
Patch is 171.65 KiB, truncated to 20.00 KiB below, full version: https://github.com/llvm/llvm-project/pull/142609.diff 32 Files Affected:
diff --git a/flang/include/flang/Optimizer/Builder/FIRBuilder.h b/flang/include/flang/Optimizer/Builder/FIRBuilder.h
index 9382d77a8d67b..6e617d50f0be5 100644
--- a/flang/include/flang/Optimizer/Builder/FIRBuilder.h
+++ b/flang/include/flang/Optimizer/Builder/FIRBuilder.h
@@ -271,14 +271,14 @@ class FirOpBuilder : public mlir::OpBuilder, public mlir::OpBuilder::Listener {
/// Sample genDeclare callback for createArrayTemp() below.
/// It creates fir.declare operation using the given operands.
/// \p memref is the base of the allocated temporary,
- /// which may be !fir.ref<!fir.array<>> or !fir.ref<!fir.box/class<>>.
+ /// which may be !fir.ref<!fir.array<>> or !fir.box/class<>.
static mlir::Value genTempDeclareOp(fir::FirOpBuilder &builder,
mlir::Location loc, mlir::Value memref,
llvm::StringRef name, mlir::Value shape,
llvm::ArrayRef<mlir::Value> typeParams,
fir::FortranVariableFlagsAttr attrs);
- /// Create a temporary array with the given \p arrayType,
+ /// Create a temporary with the given \p baseType,
/// \p shape, \p extents and \p typeParams. An optional
/// \p polymorphicMold specifies the entity which dynamic type
/// has to be used for the allocation.
@@ -291,16 +291,26 @@ class FirOpBuilder : public mlir::OpBuilder, public mlir::OpBuilder::Listener {
/// If \p useStack is true, the function will try to do the allocation
/// in stack memory (which is not always possible currently).
/// The first return value is the base of the temporary object,
- /// which may be !fir.ref<!fir.array<>> or !fir.ref<!fir.box/class<>>.
+ /// which may be !fir.ref<!fir.array<>> or !fir.box/class<>.
/// The second return value is true, if the actual allocation
/// was done in heap memory.
+ std::pair<mlir::Value, bool> createAndDeclareTemp(
+ mlir::Location loc, mlir::Type baseType, mlir::Value shape,
+ llvm::ArrayRef<mlir::Value> extents,
+ llvm::ArrayRef<mlir::Value> typeParams,
+ const std::function<decltype(genTempDeclareOp)> &genDeclare,
+ mlir::Value polymorphicMold, bool useStack, llvm::StringRef tmpName);
+ /// Create and declare an array temporary.
std::pair<mlir::Value, bool>
createArrayTemp(mlir::Location loc, fir::SequenceType arrayType,
mlir::Value shape, llvm::ArrayRef<mlir::Value> extents,
llvm::ArrayRef<mlir::Value> typeParams,
const std::function<decltype(genTempDeclareOp)> &genDeclare,
mlir::Value polymorphicMold, bool useStack = false,
- llvm::StringRef tmpName = ".tmp.array");
+ llvm::StringRef tmpName = ".tmp.array") {
+ return createAndDeclareTemp(loc, arrayType, shape, extents, typeParams,
+ genDeclare, polymorphicMold, useStack, tmpName);
+ }
/// Create an LLVM stack save intrinsic op. Returns the saved stack pointer.
/// The stack address space is fetched from the data layout of the current
diff --git a/flang/include/flang/Optimizer/Builder/MutableBox.h b/flang/include/flang/Optimizer/Builder/MutableBox.h
index 39657ddaf6e03..e1cb0f9852ba7 100644
--- a/flang/include/flang/Optimizer/Builder/MutableBox.h
+++ b/flang/include/flang/Optimizer/Builder/MutableBox.h
@@ -181,6 +181,18 @@ mlir::Value genIsNotAllocatedOrAssociatedTest(fir::FirOpBuilder &builder,
mlir::Value genNullBoxStorage(fir::FirOpBuilder &builder, mlir::Location loc,
mlir::Type boxTy);
+/// Generate an unallocated box of the given \p boxTy with the
+/// bounds, type parameters, and dynamic type set according to the
+/// parameters.
+/// \p shape may be null for scalars, and \p polymorphicMold may be null for
+/// statically typed entities. This box can then be directly passed to the
+/// runtime for allocation.
+mlir::Value getAndEstablishBoxStorage(fir::FirOpBuilder &builder,
+ mlir::Location loc,
+ fir::BaseBoxType boxTy, mlir::Value shape,
+ llvm::ArrayRef<mlir::Value> typeParams,
+ mlir::Value polymorphicMold);
+
} // namespace fir::factory
#endif // FORTRAN_OPTIMIZER_BUILDER_MUTABLEBOX_H
diff --git a/flang/include/flang/Optimizer/Dialect/FIRType.h b/flang/include/flang/Optimizer/Dialect/FIRType.h
index 52b14f15f89bd..ab6254da19e06 100644
--- a/flang/include/flang/Optimizer/Dialect/FIRType.h
+++ b/flang/include/flang/Optimizer/Dialect/FIRType.h
@@ -47,6 +47,9 @@ class BaseBoxType : public mlir::Type {
/// Returns the element type of this box type.
mlir::Type getEleTy() const;
+ /// Get the raw address type of the memory described by the box.
+ mlir::Type getBaseAddressType() const;
+
/// Unwrap element type from fir.heap, fir.ptr and fir.array.
mlir::Type unwrapInnerType() const;
@@ -56,6 +59,9 @@ class BaseBoxType : public mlir::Type {
/// Is this a box for a pointer?
bool isPointer() const;
+ /// Is this a box describing volatile memory?
+ bool isVolatile() const;
+
/// Return the same type, except for the shape, that is taken the shape
/// of shapeMold.
BaseBoxType getBoxTypeWithNewShape(mlir::Type shapeMold) const;
diff --git a/flang/lib/Optimizer/Builder/FIRBuilder.cpp b/flang/lib/Optimizer/Builder/FIRBuilder.cpp
index 68a1cc7a3aee6..584f3c8ee310e 100644
--- a/flang/lib/Optimizer/Builder/FIRBuilder.cpp
+++ b/flang/lib/Optimizer/Builder/FIRBuilder.cpp
@@ -366,55 +366,29 @@ mlir::Value fir::FirOpBuilder::createHeapTemporary(
name, dynamicLength, dynamicShape, attrs);
}
-std::pair<mlir::Value, bool> fir::FirOpBuilder::createArrayTemp(
- mlir::Location loc, fir::SequenceType arrayType, mlir::Value shape,
+std::pair<mlir::Value, bool> fir::FirOpBuilder::createAndDeclareTemp(
+ mlir::Location loc, mlir::Type baseType, mlir::Value shape,
llvm::ArrayRef<mlir::Value> extents, llvm::ArrayRef<mlir::Value> typeParams,
const std::function<decltype(FirOpBuilder::genTempDeclareOp)> &genDeclare,
mlir::Value polymorphicMold, bool useStack, llvm::StringRef tmpName) {
if (polymorphicMold) {
// Create *allocated* polymorphic temporary using the dynamic type
- // of the mold and the provided shape/extents. The created temporary
- // array will be written element per element, that is why it has to be
- // allocated.
- mlir::Type boxHeapType = fir::HeapType::get(arrayType);
- mlir::Value alloc = fir::factory::genNullBoxStorage(
- *this, loc, fir::ClassType::get(boxHeapType));
- fir::FortranVariableFlagsAttr declAttrs =
- fir::FortranVariableFlagsAttr::get(
- getContext(), fir::FortranVariableFlagsEnum::allocatable);
-
- mlir::Value base = genDeclare(*this, loc, alloc, tmpName,
- /*shape=*/nullptr, typeParams, declAttrs);
-
- int rank = extents.size();
- fir::runtime::genAllocatableApplyMold(*this, loc, alloc, polymorphicMold,
- rank);
- if (!extents.empty()) {
- mlir::Type idxTy = getIndexType();
- mlir::Value one = createIntegerConstant(loc, idxTy, 1);
- unsigned dim = 0;
- for (mlir::Value extent : extents) {
- mlir::Value dimIndex = createIntegerConstant(loc, idxTy, dim++);
- fir::runtime::genAllocatableSetBounds(*this, loc, alloc, dimIndex, one,
- extent);
- }
- }
- if (!typeParams.empty()) {
- // We should call AllocatableSetDerivedLength() here.
- // TODO: does the mold provide the length parameters or
- // the operation itself or should they be in sync?
- TODO(loc, "polymorphic type with length parameters");
- }
- fir::runtime::genAllocatableAllocate(*this, loc, alloc);
-
+ // of the mold and the provided shape/extents.
+ auto boxType = fir::ClassType::get(fir::HeapType::get(baseType));
+ mlir::Value boxAddress = fir::factory::getAndEstablishBoxStorage(
+ *this, loc, boxType, shape, typeParams, polymorphicMold);
+ fir::runtime::genAllocatableAllocate(*this, loc, boxAddress);
+ mlir::Value box = create<fir::LoadOp>(loc, boxAddress);
+ mlir::Value base =
+ genDeclare(*this, loc, box, tmpName, /*shape=*/mlir::Value{},
+ typeParams, fir::FortranVariableFlagsAttr{});
return {base, /*isHeapAllocation=*/true};
}
mlir::Value allocmem;
if (useStack)
- allocmem = createTemporary(loc, arrayType, tmpName, extents, typeParams);
+ allocmem = createTemporary(loc, baseType, tmpName, extents, typeParams);
else
- allocmem =
- createHeapTemporary(loc, arrayType, tmpName, extents, typeParams);
+ allocmem = createHeapTemporary(loc, baseType, tmpName, extents, typeParams);
mlir::Value base = genDeclare(*this, loc, allocmem, tmpName, shape,
typeParams, fir::FortranVariableFlagsAttr{});
return {base, !useStack};
diff --git a/flang/lib/Optimizer/Builder/HLFIRTools.cpp b/flang/lib/Optimizer/Builder/HLFIRTools.cpp
index f24dc2caeedfc..ce82cad153893 100644
--- a/flang/lib/Optimizer/Builder/HLFIRTools.cpp
+++ b/flang/lib/Optimizer/Builder/HLFIRTools.cpp
@@ -1331,63 +1331,36 @@ bool hlfir::elementalOpMustProduceTemp(hlfir::ElementalOp elemental) {
std::pair<hlfir::Entity, mlir::Value>
hlfir::createTempFromMold(mlir::Location loc, fir::FirOpBuilder &builder,
hlfir::Entity mold) {
+ assert(!mold.isAssumedRank() &&
+ "cannot create temporary from assumed-rank mold");
llvm::SmallVector<mlir::Value> lenParams;
hlfir::genLengthParameters(loc, builder, mold, lenParams);
llvm::StringRef tmpName{".tmp"};
- mlir::Value alloc;
- mlir::Value isHeapAlloc;
- mlir::Value shape{};
- fir::FortranVariableFlagsAttr declAttrs;
- if (mold.isPolymorphic()) {
- // Create unallocated polymorphic temporary using the dynamic type
- // of the mold. The static type of the temporary matches
- // the static type of the mold, but then the dynamic type
- // of the mold is applied to the temporary's descriptor.
-
- if (mold.isArray())
- hlfir::genShape(loc, builder, mold);
-
- // Create polymorphic allocatable box on the stack.
- mlir::Type boxHeapType = fir::HeapType::get(fir::unwrapRefType(
- mlir::cast<fir::BaseBoxType>(mold.getType()).getEleTy()));
- // The box must be initialized, because AllocatableApplyMold
- // may read its contents (e.g. for checking whether it is allocated).
- alloc = fir::factory::genNullBoxStorage(builder, loc,
- fir::ClassType::get(boxHeapType));
- // The temporary is unallocated even after AllocatableApplyMold below.
- // If the temporary is used as assignment LHS it will be automatically
- // allocated on the heap, as long as we use Assign family
- // runtime functions. So set MustFree to true.
- isHeapAlloc = builder.createBool(loc, true);
- declAttrs = fir::FortranVariableFlagsAttr::get(
- builder.getContext(), fir::FortranVariableFlagsEnum::allocatable);
- } else if (mold.isArray()) {
- mlir::Type sequenceType =
- hlfir::getFortranElementOrSequenceType(mold.getType());
+ mlir::Value shape{};
+ llvm::SmallVector<mlir::Value> extents;
+ if (mold.isArray()) {
shape = hlfir::genShape(loc, builder, mold);
- auto extents = hlfir::getIndexExtents(loc, builder, shape);
- alloc = builder.createHeapTemporary(loc, sequenceType, tmpName, extents,
- lenParams);
- isHeapAlloc = builder.createBool(loc, true);
- } else {
- alloc = builder.createTemporary(loc, mold.getFortranElementType(), tmpName,
- /*shape=*/std::nullopt, lenParams);
- isHeapAlloc = builder.createBool(loc, false);
- }
- auto declareOp =
- builder.create<hlfir::DeclareOp>(loc, alloc, tmpName, shape, lenParams,
- /*dummy_scope=*/nullptr, declAttrs);
- if (mold.isPolymorphic()) {
- int rank = mold.getRank();
- // TODO: should probably read rank from the mold.
- if (rank < 0)
- TODO(loc, "create temporary for assumed rank polymorphic");
- fir::runtime::genAllocatableApplyMold(builder, loc, alloc,
- mold.getFirBase(), rank);
+ extents = hlfir::getExplicitExtentsFromShape(shape, builder);
}
- return {hlfir::Entity{declareOp.getBase()}, isHeapAlloc};
+ bool useStack = !mold.isArray() && !mold.isPolymorphic();
+ auto genTempDeclareOp =
+ [](fir::FirOpBuilder &builder, mlir::Location loc, mlir::Value memref,
+ llvm::StringRef name, mlir::Value shape,
+ llvm::ArrayRef<mlir::Value> typeParams,
+ fir::FortranVariableFlagsAttr attrs) -> mlir::Value {
+ auto declareOp =
+ builder.create<hlfir::DeclareOp>(loc, memref, name, shape, typeParams,
+ /*dummy_scope=*/nullptr, attrs);
+ return declareOp.getBase();
+ };
+
+ auto [base, isHeapAlloc] = builder.createAndDeclareTemp(
+ loc, mold.getElementOrSequenceType(), shape, extents, lenParams,
+ genTempDeclareOp, mold.isPolymorphic() ? mold.getBase() : nullptr,
+ useStack, tmpName);
+ return {hlfir::Entity{base}, builder.createBool(loc, isHeapAlloc)};
}
hlfir::Entity hlfir::createStackTempFromMold(mlir::Location loc,
diff --git a/flang/lib/Optimizer/Builder/MutableBox.cpp b/flang/lib/Optimizer/Builder/MutableBox.cpp
index e6d630412ec34..f20b5e37dffd8 100644
--- a/flang/lib/Optimizer/Builder/MutableBox.cpp
+++ b/flang/lib/Optimizer/Builder/MutableBox.cpp
@@ -349,10 +349,7 @@ mlir::Value fir::factory::createUnallocatedBox(
const bool isAssumedRank = baseBoxType.isAssumedRank();
if (isAssumedRank)
baseBoxType = baseBoxType.getBoxTypeWithNewShape(/*rank=*/0);
- auto baseAddrType = baseBoxType.getEleTy();
- if (!fir::isa_ref_type(baseAddrType))
- baseAddrType =
- builder.getRefType(baseAddrType, fir::isa_volatile_type(baseBoxType));
+ auto baseAddrType = baseBoxType.getBaseAddressType();
auto type = fir::unwrapRefType(baseAddrType);
auto eleTy = fir::unwrapSequenceType(type);
if (auto recTy = mlir::dyn_cast<fir::RecordType>(eleTy))
@@ -982,3 +979,20 @@ mlir::Value fir::factory::genNullBoxStorage(fir::FirOpBuilder &builder,
builder.create<fir::StoreOp>(loc, nullBox, boxStorage);
return boxStorage;
}
+
+mlir::Value fir::factory::getAndEstablishBoxStorage(
+ fir::FirOpBuilder &builder, mlir::Location loc, fir::BaseBoxType boxTy,
+ mlir::Value shape, llvm::ArrayRef<mlir::Value> typeParams,
+ mlir::Value polymorphicMold) {
+ mlir::Value boxStorage = builder.createTemporary(loc, boxTy);
+ mlir::Value nullAddr =
+ builder.createNullConstant(loc, boxTy.getBaseAddressType());
+ mlir::Value box =
+ builder.create<fir::EmboxOp>(loc, boxTy, nullAddr, shape,
+ /*emptySlice=*/mlir::Value{},
+ fir::factory::elideLengthsAlreadyInType(
+ boxTy.unwrapInnerType(), typeParams),
+ polymorphicMold);
+ builder.create<fir::StoreOp>(loc, box, boxStorage);
+ return boxStorage;
+}
diff --git a/flang/lib/Optimizer/CodeGen/LowerRepackArrays.cpp b/flang/lib/Optimizer/CodeGen/LowerRepackArrays.cpp
index 7fb713ff1a6c7..de97a0bbc184a 100644
--- a/flang/lib/Optimizer/CodeGen/LowerRepackArrays.cpp
+++ b/flang/lib/Optimizer/CodeGen/LowerRepackArrays.cpp
@@ -190,12 +190,10 @@ mlir::Value PackArrayConversion::allocateTempBuffer(
if (useStack && canAllocateTempOnStack(origBox))
assert(!isHeapAllocation && "temp must have been allocated on the stack");
- if (isHeapAllocation)
- if (auto baseType = mlir::dyn_cast<fir::ReferenceType>(base.getType()))
- if (mlir::isa<fir::BaseBoxType>(baseType.getEleTy()))
- return builder.create<fir::LoadOp>(loc, base);
-
mlir::Type ptrType = base.getType();
+ if (llvm::isa<fir::BaseBoxType>(ptrType))
+ return base;
+
mlir::Type tempBoxType = fir::BoxType::get(mlir::isa<fir::HeapType>(ptrType)
? ptrType
: fir::unwrapRefType(ptrType));
diff --git a/flang/lib/Optimizer/Dialect/FIRType.cpp b/flang/lib/Optimizer/Dialect/FIRType.cpp
index 1e6e95393c2f7..38c5175e27bee 100644
--- a/flang/lib/Optimizer/Dialect/FIRType.cpp
+++ b/flang/lib/Optimizer/Dialect/FIRType.cpp
@@ -1407,6 +1407,13 @@ mlir::Type BaseBoxType::getEleTy() const {
[](auto type) { return type.getEleTy(); });
}
+mlir::Type BaseBoxType::getBaseAddressType() const {
+ mlir::Type eleTy = getEleTy();
+ if (fir::isa_ref_type(eleTy))
+ return eleTy;
+ return fir::ReferenceType::get(eleTy, isVolatile());
+}
+
mlir::Type BaseBoxType::unwrapInnerType() const {
return fir::unwrapInnerType(getEleTy());
}
@@ -1492,6 +1499,12 @@ bool fir::BaseBoxType::isPointer() const {
return llvm::isa<fir::PointerType>(getEleTy());
}
+bool BaseBoxType::isVolatile() const {
+ return llvm::TypeSwitch<fir::BaseBoxType, bool>(*this)
+ .Case<fir::BoxType, fir::ClassType>(
+ [](auto type) { return type.isVolatile(); });
+}
+
//===----------------------------------------------------------------------===//
// FIROpsDialect
//===----------------------------------------------------------------------===//
diff --git a/flang/lib/Optimizer/HLFIR/IR/HLFIROps.cpp b/flang/lib/Optimizer/HLFIR/IR/HLFIROps.cpp
index 8cfca59ecdada..baa2e2180eaf6 100644
--- a/flang/lib/Optimizer/HLFIR/IR/HLFIROps.cpp
+++ b/flang/lib/Optimizer/HLFIR/IR/HLFIROps.cpp
@@ -1586,7 +1586,7 @@ void hlfir::AssociateOp::build(mlir::OpBuilder &builder,
mlir::Type firVarType;
auto sourceExprType = mlir::dyn_cast<hlfir::ExprType>(source.getType());
if (sourceExprType && sourceExprType.isPolymorphic())
- firVarType = fir::ClassType::get(fir::HeapType::get(dataType));
+ firVarType = fir::ClassType::get(dataType);
else
firVarType = fir::ReferenceType::get(dataType);
diff --git a/flang/lib/Optimizer/HLFIR/Transforms/BufferizeHLFIR.cpp b/flang/lib/Optimizer/HLFIR/Transforms/BufferizeHLFIR.cpp
index 8a36214def167..58f2b57712974 100644
--- a/flang/lib/Optimizer/HLFIR/Transforms/BufferizeHLFIR.cpp
+++ b/flang/lib/Optimizer/HLFIR/Transforms/BufferizeHLFIR.cpp
@@ -125,7 +125,9 @@ createArrayTemp(mlir::Location loc, fir::FirOpBuilder &builder,
auto [base, isHeapAlloc] = builder.createArrayTemp(
loc, sequenceType, shape, extents, lenParams, genTempDeclareOp,
polymorphicMold ? polymorphicMold->getFirBase() : nullptr);
- return {hlfir::Entity{base}, builder.createBool(loc, isHeapAlloc)};
+ hlfir::Entity temp = hlfir::Entity{base};
+ assert(!temp.isAllocatable() && "temp must have been allocated");
+ return {temp, builder.createBool(loc, isHeapAlloc)};
}
/// Copy \p source into a new temporary and package the temporary into a
@@ -134,13 +136,10 @@ static mlir::Value copyInTempAndPackage(mlir::Location loc,
fir::FirOpBuilder &builder,
hlfir::Entity source) {
auto [temp, cleanup] = hlfir::createTempFromMold(loc, builder, source);
- builder.create<hlfir::AssignOp>(loc, source, temp, temp.isAllocatable(),
+ assert(!temp.isAllocatable() && "expect temp to already be allocated");
+ builder.create<hlfir::AssignOp>(loc, source, temp, /*realloc=*/false,
/*keep_lhs_length_if_realloc=*/false,
/*temporary_lhs=*/true);
- // Dereference allocatable temporary directly to simplify processing
- // of its uses.
- if (temp.isAllocatable())
- temp = hlfir::derefPointersAndAllocatables(loc, builder, temp);
return packageBufferizedExpr(loc, builder, temp, cleanup);
}
@@ -442,18 +441,10 @@ struct AssociateOpConversion
// !fir.box<!fir.heap<!fir.type<_T{y:i32}>>> value must be
// propagated as the box address !fir.ref<!fir.type<_T{y:i32}>>.
auto adjustVar = [&](mlir::Value sourceVar, mlir::Type assocType) {
- if (mlir::isa<fir::ReferenceType>(sourceVar.getType()) &&
- mlir::isa<fir::ClassT...
[truncated]
|
|
@llvm/pr-subscribers-flang-openmp Author: None (jeanPerier) ChangesPolymorphic temporary are currently propagated as fir.ref<fir.class<fir.heap<>>> because their allocation may be delayed to the hlfir.assign copy (using realloc). This patch moves away from this and directly allocate the temp and propagate it as a fir.class. The polymorphic temporaries creating is also simplified by avoiding the need to call the runtime to setup the descriptor altogether (the runtime is still call for the allocation currently because alloca/allocmem do not support polymorphism). Rational:
Patch is 171.65 KiB, truncated to 20.00 KiB below, full version: https://github.com/llvm/llvm-project/pull/142609.diff 32 Files Affected:
diff --git a/flang/include/flang/Optimizer/Builder/FIRBuilder.h b/flang/include/flang/Optimizer/Builder/FIRBuilder.h
index 9382d77a8d67b..6e617d50f0be5 100644
--- a/flang/include/flang/Optimizer/Builder/FIRBuilder.h
+++ b/flang/include/flang/Optimizer/Builder/FIRBuilder.h
@@ -271,14 +271,14 @@ class FirOpBuilder : public mlir::OpBuilder, public mlir::OpBuilder::Listener {
/// Sample genDeclare callback for createArrayTemp() below.
/// It creates fir.declare operation using the given operands.
/// \p memref is the base of the allocated temporary,
- /// which may be !fir.ref<!fir.array<>> or !fir.ref<!fir.box/class<>>.
+ /// which may be !fir.ref<!fir.array<>> or !fir.box/class<>.
static mlir::Value genTempDeclareOp(fir::FirOpBuilder &builder,
mlir::Location loc, mlir::Value memref,
llvm::StringRef name, mlir::Value shape,
llvm::ArrayRef<mlir::Value> typeParams,
fir::FortranVariableFlagsAttr attrs);
- /// Create a temporary array with the given \p arrayType,
+ /// Create a temporary with the given \p baseType,
/// \p shape, \p extents and \p typeParams. An optional
/// \p polymorphicMold specifies the entity which dynamic type
/// has to be used for the allocation.
@@ -291,16 +291,26 @@ class FirOpBuilder : public mlir::OpBuilder, public mlir::OpBuilder::Listener {
/// If \p useStack is true, the function will try to do the allocation
/// in stack memory (which is not always possible currently).
/// The first return value is the base of the temporary object,
- /// which may be !fir.ref<!fir.array<>> or !fir.ref<!fir.box/class<>>.
+ /// which may be !fir.ref<!fir.array<>> or !fir.box/class<>.
/// The second return value is true, if the actual allocation
/// was done in heap memory.
+ std::pair<mlir::Value, bool> createAndDeclareTemp(
+ mlir::Location loc, mlir::Type baseType, mlir::Value shape,
+ llvm::ArrayRef<mlir::Value> extents,
+ llvm::ArrayRef<mlir::Value> typeParams,
+ const std::function<decltype(genTempDeclareOp)> &genDeclare,
+ mlir::Value polymorphicMold, bool useStack, llvm::StringRef tmpName);
+ /// Create and declare an array temporary.
std::pair<mlir::Value, bool>
createArrayTemp(mlir::Location loc, fir::SequenceType arrayType,
mlir::Value shape, llvm::ArrayRef<mlir::Value> extents,
llvm::ArrayRef<mlir::Value> typeParams,
const std::function<decltype(genTempDeclareOp)> &genDeclare,
mlir::Value polymorphicMold, bool useStack = false,
- llvm::StringRef tmpName = ".tmp.array");
+ llvm::StringRef tmpName = ".tmp.array") {
+ return createAndDeclareTemp(loc, arrayType, shape, extents, typeParams,
+ genDeclare, polymorphicMold, useStack, tmpName);
+ }
/// Create an LLVM stack save intrinsic op. Returns the saved stack pointer.
/// The stack address space is fetched from the data layout of the current
diff --git a/flang/include/flang/Optimizer/Builder/MutableBox.h b/flang/include/flang/Optimizer/Builder/MutableBox.h
index 39657ddaf6e03..e1cb0f9852ba7 100644
--- a/flang/include/flang/Optimizer/Builder/MutableBox.h
+++ b/flang/include/flang/Optimizer/Builder/MutableBox.h
@@ -181,6 +181,18 @@ mlir::Value genIsNotAllocatedOrAssociatedTest(fir::FirOpBuilder &builder,
mlir::Value genNullBoxStorage(fir::FirOpBuilder &builder, mlir::Location loc,
mlir::Type boxTy);
+/// Generate an unallocated box of the given \p boxTy with the
+/// bounds, type parameters, and dynamic type set according to the
+/// parameters.
+/// \p shape may be null for scalars, and \p polymorphicMold may be null for
+/// statically typed entities. This box can then be directly passed to the
+/// runtime for allocation.
+mlir::Value getAndEstablishBoxStorage(fir::FirOpBuilder &builder,
+ mlir::Location loc,
+ fir::BaseBoxType boxTy, mlir::Value shape,
+ llvm::ArrayRef<mlir::Value> typeParams,
+ mlir::Value polymorphicMold);
+
} // namespace fir::factory
#endif // FORTRAN_OPTIMIZER_BUILDER_MUTABLEBOX_H
diff --git a/flang/include/flang/Optimizer/Dialect/FIRType.h b/flang/include/flang/Optimizer/Dialect/FIRType.h
index 52b14f15f89bd..ab6254da19e06 100644
--- a/flang/include/flang/Optimizer/Dialect/FIRType.h
+++ b/flang/include/flang/Optimizer/Dialect/FIRType.h
@@ -47,6 +47,9 @@ class BaseBoxType : public mlir::Type {
/// Returns the element type of this box type.
mlir::Type getEleTy() const;
+ /// Get the raw address type of the memory described by the box.
+ mlir::Type getBaseAddressType() const;
+
/// Unwrap element type from fir.heap, fir.ptr and fir.array.
mlir::Type unwrapInnerType() const;
@@ -56,6 +59,9 @@ class BaseBoxType : public mlir::Type {
/// Is this a box for a pointer?
bool isPointer() const;
+ /// Is this a box describing volatile memory?
+ bool isVolatile() const;
+
/// Return the same type, except for the shape, that is taken the shape
/// of shapeMold.
BaseBoxType getBoxTypeWithNewShape(mlir::Type shapeMold) const;
diff --git a/flang/lib/Optimizer/Builder/FIRBuilder.cpp b/flang/lib/Optimizer/Builder/FIRBuilder.cpp
index 68a1cc7a3aee6..584f3c8ee310e 100644
--- a/flang/lib/Optimizer/Builder/FIRBuilder.cpp
+++ b/flang/lib/Optimizer/Builder/FIRBuilder.cpp
@@ -366,55 +366,29 @@ mlir::Value fir::FirOpBuilder::createHeapTemporary(
name, dynamicLength, dynamicShape, attrs);
}
-std::pair<mlir::Value, bool> fir::FirOpBuilder::createArrayTemp(
- mlir::Location loc, fir::SequenceType arrayType, mlir::Value shape,
+std::pair<mlir::Value, bool> fir::FirOpBuilder::createAndDeclareTemp(
+ mlir::Location loc, mlir::Type baseType, mlir::Value shape,
llvm::ArrayRef<mlir::Value> extents, llvm::ArrayRef<mlir::Value> typeParams,
const std::function<decltype(FirOpBuilder::genTempDeclareOp)> &genDeclare,
mlir::Value polymorphicMold, bool useStack, llvm::StringRef tmpName) {
if (polymorphicMold) {
// Create *allocated* polymorphic temporary using the dynamic type
- // of the mold and the provided shape/extents. The created temporary
- // array will be written element per element, that is why it has to be
- // allocated.
- mlir::Type boxHeapType = fir::HeapType::get(arrayType);
- mlir::Value alloc = fir::factory::genNullBoxStorage(
- *this, loc, fir::ClassType::get(boxHeapType));
- fir::FortranVariableFlagsAttr declAttrs =
- fir::FortranVariableFlagsAttr::get(
- getContext(), fir::FortranVariableFlagsEnum::allocatable);
-
- mlir::Value base = genDeclare(*this, loc, alloc, tmpName,
- /*shape=*/nullptr, typeParams, declAttrs);
-
- int rank = extents.size();
- fir::runtime::genAllocatableApplyMold(*this, loc, alloc, polymorphicMold,
- rank);
- if (!extents.empty()) {
- mlir::Type idxTy = getIndexType();
- mlir::Value one = createIntegerConstant(loc, idxTy, 1);
- unsigned dim = 0;
- for (mlir::Value extent : extents) {
- mlir::Value dimIndex = createIntegerConstant(loc, idxTy, dim++);
- fir::runtime::genAllocatableSetBounds(*this, loc, alloc, dimIndex, one,
- extent);
- }
- }
- if (!typeParams.empty()) {
- // We should call AllocatableSetDerivedLength() here.
- // TODO: does the mold provide the length parameters or
- // the operation itself or should they be in sync?
- TODO(loc, "polymorphic type with length parameters");
- }
- fir::runtime::genAllocatableAllocate(*this, loc, alloc);
-
+ // of the mold and the provided shape/extents.
+ auto boxType = fir::ClassType::get(fir::HeapType::get(baseType));
+ mlir::Value boxAddress = fir::factory::getAndEstablishBoxStorage(
+ *this, loc, boxType, shape, typeParams, polymorphicMold);
+ fir::runtime::genAllocatableAllocate(*this, loc, boxAddress);
+ mlir::Value box = create<fir::LoadOp>(loc, boxAddress);
+ mlir::Value base =
+ genDeclare(*this, loc, box, tmpName, /*shape=*/mlir::Value{},
+ typeParams, fir::FortranVariableFlagsAttr{});
return {base, /*isHeapAllocation=*/true};
}
mlir::Value allocmem;
if (useStack)
- allocmem = createTemporary(loc, arrayType, tmpName, extents, typeParams);
+ allocmem = createTemporary(loc, baseType, tmpName, extents, typeParams);
else
- allocmem =
- createHeapTemporary(loc, arrayType, tmpName, extents, typeParams);
+ allocmem = createHeapTemporary(loc, baseType, tmpName, extents, typeParams);
mlir::Value base = genDeclare(*this, loc, allocmem, tmpName, shape,
typeParams, fir::FortranVariableFlagsAttr{});
return {base, !useStack};
diff --git a/flang/lib/Optimizer/Builder/HLFIRTools.cpp b/flang/lib/Optimizer/Builder/HLFIRTools.cpp
index f24dc2caeedfc..ce82cad153893 100644
--- a/flang/lib/Optimizer/Builder/HLFIRTools.cpp
+++ b/flang/lib/Optimizer/Builder/HLFIRTools.cpp
@@ -1331,63 +1331,36 @@ bool hlfir::elementalOpMustProduceTemp(hlfir::ElementalOp elemental) {
std::pair<hlfir::Entity, mlir::Value>
hlfir::createTempFromMold(mlir::Location loc, fir::FirOpBuilder &builder,
hlfir::Entity mold) {
+ assert(!mold.isAssumedRank() &&
+ "cannot create temporary from assumed-rank mold");
llvm::SmallVector<mlir::Value> lenParams;
hlfir::genLengthParameters(loc, builder, mold, lenParams);
llvm::StringRef tmpName{".tmp"};
- mlir::Value alloc;
- mlir::Value isHeapAlloc;
- mlir::Value shape{};
- fir::FortranVariableFlagsAttr declAttrs;
- if (mold.isPolymorphic()) {
- // Create unallocated polymorphic temporary using the dynamic type
- // of the mold. The static type of the temporary matches
- // the static type of the mold, but then the dynamic type
- // of the mold is applied to the temporary's descriptor.
-
- if (mold.isArray())
- hlfir::genShape(loc, builder, mold);
-
- // Create polymorphic allocatable box on the stack.
- mlir::Type boxHeapType = fir::HeapType::get(fir::unwrapRefType(
- mlir::cast<fir::BaseBoxType>(mold.getType()).getEleTy()));
- // The box must be initialized, because AllocatableApplyMold
- // may read its contents (e.g. for checking whether it is allocated).
- alloc = fir::factory::genNullBoxStorage(builder, loc,
- fir::ClassType::get(boxHeapType));
- // The temporary is unallocated even after AllocatableApplyMold below.
- // If the temporary is used as assignment LHS it will be automatically
- // allocated on the heap, as long as we use Assign family
- // runtime functions. So set MustFree to true.
- isHeapAlloc = builder.createBool(loc, true);
- declAttrs = fir::FortranVariableFlagsAttr::get(
- builder.getContext(), fir::FortranVariableFlagsEnum::allocatable);
- } else if (mold.isArray()) {
- mlir::Type sequenceType =
- hlfir::getFortranElementOrSequenceType(mold.getType());
+ mlir::Value shape{};
+ llvm::SmallVector<mlir::Value> extents;
+ if (mold.isArray()) {
shape = hlfir::genShape(loc, builder, mold);
- auto extents = hlfir::getIndexExtents(loc, builder, shape);
- alloc = builder.createHeapTemporary(loc, sequenceType, tmpName, extents,
- lenParams);
- isHeapAlloc = builder.createBool(loc, true);
- } else {
- alloc = builder.createTemporary(loc, mold.getFortranElementType(), tmpName,
- /*shape=*/std::nullopt, lenParams);
- isHeapAlloc = builder.createBool(loc, false);
- }
- auto declareOp =
- builder.create<hlfir::DeclareOp>(loc, alloc, tmpName, shape, lenParams,
- /*dummy_scope=*/nullptr, declAttrs);
- if (mold.isPolymorphic()) {
- int rank = mold.getRank();
- // TODO: should probably read rank from the mold.
- if (rank < 0)
- TODO(loc, "create temporary for assumed rank polymorphic");
- fir::runtime::genAllocatableApplyMold(builder, loc, alloc,
- mold.getFirBase(), rank);
+ extents = hlfir::getExplicitExtentsFromShape(shape, builder);
}
- return {hlfir::Entity{declareOp.getBase()}, isHeapAlloc};
+ bool useStack = !mold.isArray() && !mold.isPolymorphic();
+ auto genTempDeclareOp =
+ [](fir::FirOpBuilder &builder, mlir::Location loc, mlir::Value memref,
+ llvm::StringRef name, mlir::Value shape,
+ llvm::ArrayRef<mlir::Value> typeParams,
+ fir::FortranVariableFlagsAttr attrs) -> mlir::Value {
+ auto declareOp =
+ builder.create<hlfir::DeclareOp>(loc, memref, name, shape, typeParams,
+ /*dummy_scope=*/nullptr, attrs);
+ return declareOp.getBase();
+ };
+
+ auto [base, isHeapAlloc] = builder.createAndDeclareTemp(
+ loc, mold.getElementOrSequenceType(), shape, extents, lenParams,
+ genTempDeclareOp, mold.isPolymorphic() ? mold.getBase() : nullptr,
+ useStack, tmpName);
+ return {hlfir::Entity{base}, builder.createBool(loc, isHeapAlloc)};
}
hlfir::Entity hlfir::createStackTempFromMold(mlir::Location loc,
diff --git a/flang/lib/Optimizer/Builder/MutableBox.cpp b/flang/lib/Optimizer/Builder/MutableBox.cpp
index e6d630412ec34..f20b5e37dffd8 100644
--- a/flang/lib/Optimizer/Builder/MutableBox.cpp
+++ b/flang/lib/Optimizer/Builder/MutableBox.cpp
@@ -349,10 +349,7 @@ mlir::Value fir::factory::createUnallocatedBox(
const bool isAssumedRank = baseBoxType.isAssumedRank();
if (isAssumedRank)
baseBoxType = baseBoxType.getBoxTypeWithNewShape(/*rank=*/0);
- auto baseAddrType = baseBoxType.getEleTy();
- if (!fir::isa_ref_type(baseAddrType))
- baseAddrType =
- builder.getRefType(baseAddrType, fir::isa_volatile_type(baseBoxType));
+ auto baseAddrType = baseBoxType.getBaseAddressType();
auto type = fir::unwrapRefType(baseAddrType);
auto eleTy = fir::unwrapSequenceType(type);
if (auto recTy = mlir::dyn_cast<fir::RecordType>(eleTy))
@@ -982,3 +979,20 @@ mlir::Value fir::factory::genNullBoxStorage(fir::FirOpBuilder &builder,
builder.create<fir::StoreOp>(loc, nullBox, boxStorage);
return boxStorage;
}
+
+mlir::Value fir::factory::getAndEstablishBoxStorage(
+ fir::FirOpBuilder &builder, mlir::Location loc, fir::BaseBoxType boxTy,
+ mlir::Value shape, llvm::ArrayRef<mlir::Value> typeParams,
+ mlir::Value polymorphicMold) {
+ mlir::Value boxStorage = builder.createTemporary(loc, boxTy);
+ mlir::Value nullAddr =
+ builder.createNullConstant(loc, boxTy.getBaseAddressType());
+ mlir::Value box =
+ builder.create<fir::EmboxOp>(loc, boxTy, nullAddr, shape,
+ /*emptySlice=*/mlir::Value{},
+ fir::factory::elideLengthsAlreadyInType(
+ boxTy.unwrapInnerType(), typeParams),
+ polymorphicMold);
+ builder.create<fir::StoreOp>(loc, box, boxStorage);
+ return boxStorage;
+}
diff --git a/flang/lib/Optimizer/CodeGen/LowerRepackArrays.cpp b/flang/lib/Optimizer/CodeGen/LowerRepackArrays.cpp
index 7fb713ff1a6c7..de97a0bbc184a 100644
--- a/flang/lib/Optimizer/CodeGen/LowerRepackArrays.cpp
+++ b/flang/lib/Optimizer/CodeGen/LowerRepackArrays.cpp
@@ -190,12 +190,10 @@ mlir::Value PackArrayConversion::allocateTempBuffer(
if (useStack && canAllocateTempOnStack(origBox))
assert(!isHeapAllocation && "temp must have been allocated on the stack");
- if (isHeapAllocation)
- if (auto baseType = mlir::dyn_cast<fir::ReferenceType>(base.getType()))
- if (mlir::isa<fir::BaseBoxType>(baseType.getEleTy()))
- return builder.create<fir::LoadOp>(loc, base);
-
mlir::Type ptrType = base.getType();
+ if (llvm::isa<fir::BaseBoxType>(ptrType))
+ return base;
+
mlir::Type tempBoxType = fir::BoxType::get(mlir::isa<fir::HeapType>(ptrType)
? ptrType
: fir::unwrapRefType(ptrType));
diff --git a/flang/lib/Optimizer/Dialect/FIRType.cpp b/flang/lib/Optimizer/Dialect/FIRType.cpp
index 1e6e95393c2f7..38c5175e27bee 100644
--- a/flang/lib/Optimizer/Dialect/FIRType.cpp
+++ b/flang/lib/Optimizer/Dialect/FIRType.cpp
@@ -1407,6 +1407,13 @@ mlir::Type BaseBoxType::getEleTy() const {
[](auto type) { return type.getEleTy(); });
}
+mlir::Type BaseBoxType::getBaseAddressType() const {
+ mlir::Type eleTy = getEleTy();
+ if (fir::isa_ref_type(eleTy))
+ return eleTy;
+ return fir::ReferenceType::get(eleTy, isVolatile());
+}
+
mlir::Type BaseBoxType::unwrapInnerType() const {
return fir::unwrapInnerType(getEleTy());
}
@@ -1492,6 +1499,12 @@ bool fir::BaseBoxType::isPointer() const {
return llvm::isa<fir::PointerType>(getEleTy());
}
+bool BaseBoxType::isVolatile() const {
+ return llvm::TypeSwitch<fir::BaseBoxType, bool>(*this)
+ .Case<fir::BoxType, fir::ClassType>(
+ [](auto type) { return type.isVolatile(); });
+}
+
//===----------------------------------------------------------------------===//
// FIROpsDialect
//===----------------------------------------------------------------------===//
diff --git a/flang/lib/Optimizer/HLFIR/IR/HLFIROps.cpp b/flang/lib/Optimizer/HLFIR/IR/HLFIROps.cpp
index 8cfca59ecdada..baa2e2180eaf6 100644
--- a/flang/lib/Optimizer/HLFIR/IR/HLFIROps.cpp
+++ b/flang/lib/Optimizer/HLFIR/IR/HLFIROps.cpp
@@ -1586,7 +1586,7 @@ void hlfir::AssociateOp::build(mlir::OpBuilder &builder,
mlir::Type firVarType;
auto sourceExprType = mlir::dyn_cast<hlfir::ExprType>(source.getType());
if (sourceExprType && sourceExprType.isPolymorphic())
- firVarType = fir::ClassType::get(fir::HeapType::get(dataType));
+ firVarType = fir::ClassType::get(dataType);
else
firVarType = fir::ReferenceType::get(dataType);
diff --git a/flang/lib/Optimizer/HLFIR/Transforms/BufferizeHLFIR.cpp b/flang/lib/Optimizer/HLFIR/Transforms/BufferizeHLFIR.cpp
index 8a36214def167..58f2b57712974 100644
--- a/flang/lib/Optimizer/HLFIR/Transforms/BufferizeHLFIR.cpp
+++ b/flang/lib/Optimizer/HLFIR/Transforms/BufferizeHLFIR.cpp
@@ -125,7 +125,9 @@ createArrayTemp(mlir::Location loc, fir::FirOpBuilder &builder,
auto [base, isHeapAlloc] = builder.createArrayTemp(
loc, sequenceType, shape, extents, lenParams, genTempDeclareOp,
polymorphicMold ? polymorphicMold->getFirBase() : nullptr);
- return {hlfir::Entity{base}, builder.createBool(loc, isHeapAlloc)};
+ hlfir::Entity temp = hlfir::Entity{base};
+ assert(!temp.isAllocatable() && "temp must have been allocated");
+ return {temp, builder.createBool(loc, isHeapAlloc)};
}
/// Copy \p source into a new temporary and package the temporary into a
@@ -134,13 +136,10 @@ static mlir::Value copyInTempAndPackage(mlir::Location loc,
fir::FirOpBuilder &builder,
hlfir::Entity source) {
auto [temp, cleanup] = hlfir::createTempFromMold(loc, builder, source);
- builder.create<hlfir::AssignOp>(loc, source, temp, temp.isAllocatable(),
+ assert(!temp.isAllocatable() && "expect temp to already be allocated");
+ builder.create<hlfir::AssignOp>(loc, source, temp, /*realloc=*/false,
/*keep_lhs_length_if_realloc=*/false,
/*temporary_lhs=*/true);
- // Dereference allocatable temporary directly to simplify processing
- // of its uses.
- if (temp.isAllocatable())
- temp = hlfir::derefPointersAndAllocatables(loc, builder, temp);
return packageBufferizedExpr(loc, builder, temp, cleanup);
}
@@ -442,18 +441,10 @@ struct AssociateOpConversion
// !fir.box<!fir.heap<!fir.type<_T{y:i32}>>> value must be
// propagated as the box address !fir.ref<!fir.type<_T{y:i32}>>.
auto adjustVar = [&](mlir::Value sourceVar, mlir::Type assocType) {
- if (mlir::isa<fir::ReferenceType>(sourceVar.getType()) &&
- mlir::isa<fir::ClassT...
[truncated]
|
Co-authored-by: Asher Mancinelli <ashermancinelli@gmail.com>
|
✅ With the latest revision this PR passed the C/C++ code formatter. |
jeanPerier
commented
Jun 4, 2025
ashermancinelli
approved these changes
Jun 5, 2025
vzakhari
approved these changes
Jun 5, 2025
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Polymorphic temporary are currently propagated as fir.ref<fir.class<fir.heap<>>> because their allocation may be delayed to the hlfir.assign copy (using realloc).
This patch moves away from this and directly allocate the temp and propagate it as a fir.class.
The polymorphic temporaries creating is also simplified by avoiding the need to call the runtime to setup the descriptor altogether (the runtime is still call for the allocation currently because alloca/allocmem do not support polymorphism).
Rational:
this lowers complexity and reduce the need to add code to special case polymorphic handling.
I want to move towards a point where where we can use fir.alloca/fir.allocmem to allocate polymorphic temporary so that LLVM can better understand the shape and size of the data after inlining.
I want to hlfir.declare box value result to not have the allocatable attribute. The hlfir.declare fortran attributes and the runtime attribute inside the descriptor should be in sync I think to avoid situation where the runtime thinks it can reallocate things that are not allocatable at that point of the program. This will fix a bug with SELECT TYPE lowering.