[flang][MLIR][OpenMP] Extend delayed privatization for allocatables #84033
Conversation
llvmbot
added
flang
|
@llvm/pr-subscribers-flang-openmp @llvm/pr-subscribers-flang-fir-hlfir Author: Kareem Ergawy (ergawy) ChangesAdds delayed privatization support for allocatables. In order to explain the problem this diff is trying to solve, it would be useful to see an example of Consider the following Fortran code: subroutine delayed_privatization_allocatable
implicit none
integer, allocatable :: var1
!$omp parallel firstprivate(var1)
var1 = 10
!$omp end parallel
end subroutineYou would get something like this (again no delayed privatization yet): ...
%3:2 = hlfir.declare %0 {fortran_attrs = #fir.var_attrs<allocatable>, ...} : ...
omp.parallel {
// Allocation logic
...
%9 = fir.load %3#<!-- -->1 : !fir.ref<!fir.box<!fir.heap<i32>>>
...
fir.if %12 {
...
} else {
...
}
%13:2 = hlfir.declare %8 {fortran_attrs = #fir.var_attrs<allocatable>, ...} : ...
// Copy logic
%14 = fir.load %13#<!-- -->0 : !fir.ref<!fir.box<!fir.heap<i32>>>
...
fir.if %17 {
%24 = fir.load %3#<!-- -->0 : !fir.ref<!fir.box<!fir.heap<i32>>>
...
}
...
omp.terminator
}Note that for both the original and the private declaration, the allocation and copy logic use both elements (e.g. Some possible solutions are:
A "variable shadow" is a composite value of a special type:
We don't introduce any special opertions to deal with shadow values. Rather we treat them similar to other composites. If you want to get a certain component, use Patch is 44.24 KiB, truncated to 20.00 KiB below, full version: https://github.com/llvm/llvm-project/pull/84033.diff 19 Files Affected:
diff --git a/flang/include/flang/Lower/SymbolMap.h b/flang/include/flang/Lower/SymbolMap.h
index a55e4b133fe0a8..395c22c4b94f5a 100644
--- a/flang/include/flang/Lower/SymbolMap.h
+++ b/flang/include/flang/Lower/SymbolMap.h
@@ -15,6 +15,7 @@
#include "flang/Common/reference.h"
#include "flang/Optimizer/Builder/BoxValue.h"
+#include "flang/Optimizer/Builder/VariableShadow.h"
#include "flang/Optimizer/Dialect/FIRType.h"
#include "flang/Optimizer/Dialect/FortranVariableInterface.h"
#include "flang/Optimizer/Support/Matcher.h"
@@ -77,7 +78,8 @@ struct SymbolBox : public fir::details::matcher<SymbolBox> {
using VT =
std::variant<Intrinsic, FullDim, Char, CharFullDim, PointerOrAllocatable,
- Box, fir::FortranVariableOpInterface, None>;
+ Box, fir::FortranVariableOpInterface,
+ hlfir::FortranVariableShadow, None>;
//===--------------------------------------------------------------------===//
// Constructors
@@ -101,6 +103,7 @@ struct SymbolBox : public fir::details::matcher<SymbolBox> {
[](const fir::FortranVariableOpInterface &x) {
return fir::FortranVariableOpInterface(x).getBase();
},
+ [](const hlfir::FortranVariableShadow &x) { return x.getBase(); },
[](const auto &x) { return x.getAddr(); });
}
diff --git a/flang/include/flang/Optimizer/Builder/BoxValue.h b/flang/include/flang/Optimizer/Builder/BoxValue.h
index 2fed2d48a7a080..151a197cc29b7f 100644
--- a/flang/include/flang/Optimizer/Builder/BoxValue.h
+++ b/flang/include/flang/Optimizer/Builder/BoxValue.h
@@ -13,6 +13,8 @@
#ifndef FORTRAN_OPTIMIZER_BUILDER_BOXVALUE_H
#define FORTRAN_OPTIMIZER_BUILDER_BOXVALUE_H
+#include "flang/Optimizer/Builder/Todo.h"
+#include "flang/Optimizer/Builder/VariableShadow.h"
#include "flang/Optimizer/Dialect/FIRType.h"
#include "flang/Optimizer/Support/FatalError.h"
#include "flang/Optimizer/Support/Matcher.h"
@@ -479,7 +481,8 @@ class ExtendedValue : public details::matcher<ExtendedValue> {
public:
using VT =
std::variant<UnboxedValue, CharBoxValue, ArrayBoxValue, CharArrayBoxValue,
- ProcBoxValue, BoxValue, MutableBoxValue, PolymorphicValue>;
+ ProcBoxValue, BoxValue, MutableBoxValue, PolymorphicValue,
+ hlfir::FortranVariableShadow>;
ExtendedValue() : box{UnboxedValue{}} {}
template <typename A, typename = std::enable_if_t<
@@ -516,6 +519,11 @@ class ExtendedValue : public details::matcher<ExtendedValue> {
[](const fir::CharBoxValue &box) -> unsigned { return 0; },
[](const fir::ProcBoxValue &box) -> unsigned { return 0; },
[](const fir::PolymorphicValue &box) -> unsigned { return 0; },
+ [](const hlfir::FortranVariableShadow &box) -> unsigned {
+ TODO(box.getValue().getLoc(),
+ "rank(): FortranVariableShadow");
+ return 0;
+ },
[](const auto &box) -> unsigned { return box.rank(); });
}
diff --git a/flang/include/flang/Optimizer/Builder/HLFIRTools.h b/flang/include/flang/Optimizer/Builder/HLFIRTools.h
index 170e134baef614..1de6e6b2d340da 100644
--- a/flang/include/flang/Optimizer/Builder/HLFIRTools.h
+++ b/flang/include/flang/Optimizer/Builder/HLFIRTools.h
@@ -14,6 +14,7 @@
#define FORTRAN_OPTIMIZER_BUILDER_HLFIRTOOLS_H
#include "flang/Optimizer/Builder/BoxValue.h"
+#include "flang/Optimizer/Builder/VariableShadow.h"
#include "flang/Optimizer/Dialect/FIROps.h"
#include "flang/Optimizer/Dialect/FortranVariableInterface.h"
#include "flang/Optimizer/HLFIR/HLFIRDialect.h"
@@ -55,8 +56,12 @@ class Entity : public mlir::Value {
}
Entity(fir::FortranVariableOpInterface variable)
: mlir::Value(variable.getBase()) {}
+ Entity(FortranVariableShadow variableShadow)
+ : mlir::Value(variableShadow.getValue()) {}
+
bool isValue() const { return isFortranValue(*this); }
bool isVariable() const { return !isValue(); }
+ bool isMutableShadow() const;
bool isMutableBox() const { return hlfir::isBoxAddressType(getType()); }
bool isProcedurePointer() const {
return fir::isBoxProcAddressType(getType());
@@ -74,6 +79,10 @@ class Entity : public mlir::Value {
/// Is this an assumed ranked entity?
bool isAssumedRank() const { return getRank() == -1; }
+ bool isFortranVariableShadow() const {
+ return this->getType().isa<fir::ShadowType>();
+ }
+
/// Return the rank of this entity or -1 if it is an assumed rank.
int getRank() const {
mlir::Type type = fir::unwrapPassByRefType(fir::unwrapRefType(getType()));
@@ -152,6 +161,14 @@ class Entity : public mlir::Value {
return this->getDefiningOp<fir::FortranVariableOpInterface>();
}
+ fir::ShadowType getIfVariableShadow() const {
+ fir::ExtractValueOp op = this->getDefiningOp<fir::ExtractValueOp>();
+ if (!op)
+ return fir::ShadowType();
+
+ return op.getAdt().getType().dyn_cast<fir::ShadowType>();
+ }
+
// Return a "declaration" operation for this variable if visible,
// or the "declaration" operation of the allocatable/pointer this
// variable was dereferenced from (if it is visible).
@@ -174,8 +191,13 @@ class Entity : public mlir::Value {
}
bool isAllocatable() const {
- auto varIface = getIfVariableInterface();
- return varIface ? varIface.isAllocatable() : false;
+ if (auto varIface = getIfVariableInterface())
+ return varIface.isAllocatable();
+
+ if (auto shadowType = getIfVariableShadow())
+ return shadowType.getAllocatable();
+
+ return false;
}
bool isPointer() const {
@@ -185,7 +207,7 @@ class Entity : public mlir::Value {
// Get the entity as an mlir SSA value containing all the shape, type
// parameters and dynamic shape information.
- mlir::Value getBase() const { return *this; }
+ mlir::Value getBase() const;
// Get the entity as a FIR base. This may not carry the shape and type
// parameters information, and even when it is a box with shape information.
@@ -231,6 +253,9 @@ translateToExtendedValue(mlir::Location loc, fir::FirOpBuilder &builder,
fir::ExtendedValue
translateToExtendedValue(mlir::Location loc, fir::FirOpBuilder &builder,
fir::FortranVariableOpInterface fortranVariable);
+fir::ExtendedValue
+translateToExtendedValue(mlir::Location loc, fir::FirOpBuilder &builder,
+ hlfir::FortranVariableShadow fortranVariableShadow);
/// Generate declaration for a fir::ExtendedValue in memory.
fir::FortranVariableOpInterface
diff --git a/flang/include/flang/Optimizer/Builder/VariableShadow.h b/flang/include/flang/Optimizer/Builder/VariableShadow.h
new file mode 100644
index 00000000000000..af83c4e34abb60
--- /dev/null
+++ b/flang/include/flang/Optimizer/Builder/VariableShadow.h
@@ -0,0 +1,47 @@
+//===-- VariableShadow.h ----------------------------------------*- C++ -*-===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+//
+// Coding style: https://mlir.llvm.org/getting_started/DeveloperGuide/
+//
+//===----------------------------------------------------------------------===//
+
+#include "flang/Optimizer/Builder/FIRBuilder.h"
+#include "flang/Optimizer/Dialect/FortranVariableInterface.h"
+#include "mlir/IR/Value.h"
+
+#ifndef FORTRAN_OPTIMIZER_BUILDER_VARIABLESHADOW_H
+#define FORTRAN_OPTIMIZER_BUILDER_VARIABLESHADOW_H
+
+namespace hlfir {
+
+class FortranVariableShadow {
+public:
+ FortranVariableShadow(fir::FirOpBuilder &builder,
+ mlir::BlockArgument shadowingVal,
+ fir::FortranVariableOpInterface shadowedVariable);
+
+ mlir::BlockArgument getValue() const { return shadowingVal; }
+
+ mlir::Value getBase() const { return base; }
+
+ mlir::Value getFirBase() const { return firBase; }
+
+ fir::FortranVariableOpInterface getShadowedVariable() const {
+ return shadowedVariable;
+ }
+
+private:
+ mlir::BlockArgument shadowingVal;
+ mlir::Value base;
+ mlir::Value firBase;
+ fir::FortranVariableOpInterface shadowedVariable;
+};
+
+} // namespace hlfir
+
+#endif // FORTRAN_OPTIMIZER_BUILDER_VARIABLESHADOW_H
diff --git a/flang/include/flang/Optimizer/Dialect/FIRTypes.td b/flang/include/flang/Optimizer/Dialect/FIRTypes.td
index 2a2f50720859e7..d83b216bce7b30 100644
--- a/flang/include/flang/Optimizer/Dialect/FIRTypes.td
+++ b/flang/include/flang/Optimizer/Dialect/FIRTypes.td
@@ -576,6 +576,65 @@ def fir_VoidType : FIR_Type<"Void", "void"> {
let genStorageClass = 0;
}
+def fir_VariableShadowType : FIR_Type<"Shadow", "shadow"> {
+ let summary = "Type for Fortran-variables shadow values";
+
+ let description = [{
+ A type to encapulate required data for a Fortran-variable shadow value. A
+ shadow type encapsulates the base type as well as the original base type. It
+ also encapsulates information other information about the shadowed value,
+ such as whether it is allocatable or not.
+
+ For example, if we want to shadow a value produced a `hlfir.declare`
+ instruction similar to the following:
+ ```
+ %1:2 = hlfir.declare %0
+ {fortran_attrs = #fir.var_attrs<allocatable>}
+ : (!fir.ref<!fir.box<!fir.heap<i32>>>)
+ -> (!fir.ref<!fir.box<!fir.heap<i32>>>, !fir.ref<!fir.box<!fir.heap<i32>>>)
+ ```
+
+ We would need to somehow create a value of the type:
+ ```
+ !fir.shadow<!fir.ref<!fir.box<!fir.heap<i32>>>,
+ !fir.ref<!fir.box<!fir.heap<i32>>>,
+ allocatable : true>
+ ```
+
+ The mechanism to create this value can be, for example, through a
+ combincation `fir.undefined`, `fir.insert_value` instructions similar to what
+ would be done for other composite types like `tuple`s:
+ ```
+ %2 = fir.undefined !fir.shadow<!fir.ref<!fir.box<!fir.heap<i32>>>,
+ !fir.ref<!fir.box<!fir.heap<i32>>>,
+ allocatable : true>
+ %3 = fir.insert_value %2, %1#0, [0 : index] : <... type info omitted ...>
+ %4 = fir.insert_value %3, %3#1, [1 : index] : <... type info omitted ...>
+ ```
+
+ From that point on, you can use `%4` as the shadow of the originally declared
+ variable being shadowed `%1`.
+
+ For more info about the concept of variable shadows, check `VariableShadow.h`.
+ }];
+
+ let parameters = (ins "mlir::Type":$baseTy,
+ "mlir::Type":$firBaseTy,
+ "bool":$allocatable);
+
+ let assemblyFormat = [{
+ `<` $baseTy `,` $firBaseTy `,` `allocatable` `:` $allocatable `>`
+ }];
+
+ let extraClassDeclaration = [{
+ mlir::Type getType(unsigned idx) const {
+ if (idx == 0) return getBaseTy();
+ if (idx == 1) return getFirBaseTy();
+ return mlir::Type();
+ }
+ }];
+}
+
// Whether a type is a BaseBoxType
def IsBaseBoxTypePred
: CPred<"$_self.isa<::fir::BaseBoxType>()">;
@@ -590,13 +649,17 @@ def AnyRealLike : TypeConstraint<Or<[FloatLike.predicate,
fir_RealType.predicate]>, "any real">;
def AnyIntegerType : Type<AnyIntegerLike.predicate, "any integer">;
+def IsShadowTypePred
+ : CPred<"$_self.isa<::fir::ShadowType>()">;
+
// Composable types
// Note that we include both fir_ComplexType and AnyComplex, so we can use both
// the FIR ComplexType and the MLIR ComplexType (the former is used to represent
// Fortran complex and the latter for C++ std::complex).
def AnyCompositeLike : TypeConstraint<Or<[fir_RecordType.predicate,
fir_SequenceType.predicate, fir_ComplexType.predicate, AnyComplex.predicate,
- fir_VectorType.predicate, IsTupleTypePred, fir_CharacterType.predicate]>,
+ fir_VectorType.predicate, IsTupleTypePred, fir_CharacterType.predicate,
+ IsShadowTypePred]>,
"any composite">;
// Reference types
diff --git a/flang/lib/Lower/Bridge.cpp b/flang/lib/Lower/Bridge.cpp
index 153ce0623ab3ae..3d2ff056aed650 100644
--- a/flang/lib/Lower/Bridge.cpp
+++ b/flang/lib/Lower/Bridge.cpp
@@ -468,6 +468,10 @@ class FirConverter : public Fortran::lower::AbstractConverter {
return hlfir::translateToExtendedValue(getCurrentLocation(),
getFirOpBuilder(), x);
},
+ [&](const hlfir::FortranVariableShadow &x) -> fir::ExtendedValue {
+ return hlfir::translateToExtendedValue(getCurrentLocation(),
+ getFirOpBuilder(), x);
+ },
[](const auto &box) -> fir::ExtendedValue { return box; });
}
@@ -1009,6 +1013,16 @@ class FirConverter : public Fortran::lower::AbstractConverter {
return v.match(
[&](const Fortran::lower::SymbolBox::Intrinsic &)
-> Fortran::lower::SymbolBox { return v; },
+ [&](const hlfir::FortranVariableShadow &box)
+ -> Fortran::lower::SymbolBox {
+ auto exv =
+ hlfir::translateToExtendedValue(toLocation(), *builder, box);
+ return exv.match(
+ [](mlir::Value x) -> Fortran::lower::SymbolBox {
+ return Fortran::lower::SymbolBox::Intrinsic{x};
+ },
+ [](auto x) -> Fortran::lower::SymbolBox { return x; });
+ },
[](const auto &) -> Fortran::lower::SymbolBox { return {}; });
return {};
@@ -1066,6 +1080,7 @@ class FirConverter : public Fortran::lower::AbstractConverter {
assert(lowerToHighLevelFIR());
hlfir::Entity lhs{dst};
hlfir::Entity rhs{src};
+
// Temporary_lhs is set to true in hlfir.assign below to avoid user
// assignment to be used and finalization to be called on the LHS.
// This may or may not be correct but mimics the current behaviour
diff --git a/flang/lib/Lower/ConvertExpr.cpp b/flang/lib/Lower/ConvertExpr.cpp
index d157db2cde4961..9b69bb10967349 100644
--- a/flang/lib/Lower/ConvertExpr.cpp
+++ b/flang/lib/Lower/ConvertExpr.cpp
@@ -981,6 +981,9 @@ class ScalarExprLowering {
},
[&](const fir::ProcBoxValue &toBox) {
TODO(loc, "procedure pointer component in derived type assignment");
+ },
+ [&](const hlfir::FortranVariableShadow &toBox) {
+ TODO(loc, "FortranVariableShadow in derived type assignment");
});
}
return res;
diff --git a/flang/lib/Lower/OpenMP/DataSharingProcessor.cpp b/flang/lib/Lower/OpenMP/DataSharingProcessor.cpp
index 717b8cc0276a30..be88d30e12d306 100644
--- a/flang/lib/Lower/OpenMP/DataSharingProcessor.cpp
+++ b/flang/lib/Lower/OpenMP/DataSharingProcessor.cpp
@@ -15,7 +15,9 @@
#include "Utils.h"
#include "flang/Lower/PFTBuilder.h"
#include "flang/Lower/SymbolMap.h"
+#include "flang/Optimizer/Builder/HLFIRTools.h"
#include "flang/Optimizer/Builder/Todo.h"
+#include "flang/Optimizer/HLFIR/HLFIROps.h"
#include "flang/Semantics/tools.h"
#include "mlir/Dialect/OpenMP/OpenMPDialect.h"
@@ -357,6 +359,28 @@ void DataSharingProcessor::doPrivatize(const Fortran::semantics::Symbol *sym) {
bool isFirstPrivate =
sym->test(Fortran::semantics::Symbol::Flag::OmpFirstPrivate);
+ mlir::Type privatizerArgType = hsb.match(
+ [&](const fir::FortranVariableOpInterface &box) -> mlir::Type {
+ hlfir::Entity variable(box);
+ return fir::ShadowType::get(
+ &converter.getMLIRContext(), variable.getBase().getType(),
+ variable.getFirBase().getType(), variable.isAllocatable());
+ },
+ [&](const auto &box) -> mlir::Type { return symType; });
+
+ auto addSymbol = [&](mlir::Region ®ion, unsigned argIdx, bool force) {
+ hsb.match(
+ [&](const fir::FortranVariableOpInterface &box) {
+ hlfir::Entity variable(box);
+ hlfir::FortranVariableShadow variableShadow(
+ firOpBuilder, region.getArgument(argIdx), box);
+ symTable->addSymbol(*sym, variableShadow, force);
+ },
+ [&](const auto &box) {
+ symTable->addSymbol(*sym, region.getArgument(argIdx), force);
+ });
+ };
+
mlir::omp::PrivateClauseOp privatizerOp = [&]() {
auto moduleOp = firOpBuilder.getModule();
auto uniquePrivatizerName = fir::getTypeAsString(
@@ -373,7 +397,7 @@ void DataSharingProcessor::doPrivatize(const Fortran::semantics::Symbol *sym) {
firOpBuilder.setInsertionPoint(&moduleOp.getBodyRegion().front(),
moduleOp.getBodyRegion().front().begin());
auto result = firOpBuilder.create<mlir::omp::PrivateClauseOp>(
- symLoc, uniquePrivatizerName, symType,
+ symLoc, uniquePrivatizerName, privatizerArgType,
isFirstPrivate ? mlir::omp::DataSharingClauseType::FirstPrivate
: mlir::omp::DataSharingClauseType::Private);
@@ -383,15 +407,44 @@ void DataSharingProcessor::doPrivatize(const Fortran::semantics::Symbol *sym) {
{
mlir::Region &allocRegion = result.getAllocRegion();
mlir::Block *allocEntryBlock = firOpBuilder.createBlock(
- &allocRegion, /*insertPt=*/{}, symType, symLoc);
+ &allocRegion, /*insertPt=*/{}, privatizerArgType, symLoc);
firOpBuilder.setInsertionPointToEnd(allocEntryBlock);
- symTable->addSymbol(*sym, allocRegion.getArgument(0));
+ addSymbol(allocRegion, 0, false);
symTable->pushScope();
cloneSymbol(sym);
- firOpBuilder.create<mlir::omp::YieldOp>(
- hsb.getAddr().getLoc(),
- symTable->shallowLookupSymbol(*sym).getAddr());
+
+ auto genYieldedVal = [&]() -> mlir::Value {
+ auto symBox = symTable->shallowLookupSymbol(*sym);
+ mlir::Value addr = symBox.getAddr();
+ mlir::Operation *definingOp = addr.getDefiningOp();
+ hlfir::DeclareOp declareOp =
+ mlir::dyn_cast<hlfir::DeclareOp>(definingOp);
+
+ if (declareOp == nullptr) {
+ return addr;
+ }
+ declareOp.getBase();
+ declareOp.getOriginalBase();
+ mlir::Location loc = declareOp.getLoc();
+
+ mlir::Value yieldTuple =
+ firOpBuilder.create<fir::UndefOp>(loc, privatizerArgType);
+
+ mlir::Type idxTy = firOpBuilder.getIndexType();
+
+ yieldTuple = firOpBuilder.create<fir::InsertValueOp>(
+ loc, privatizerArgType, yieldTuple, declareOp.getBase(),
+ firOpBuilder.getArrayAttr({firOpBuilder.getIntegerAttr(idxTy, 0)}));
+
+ yieldTuple = firOpBuilder.create<fir::InsertValueOp>(
+ loc, privatizerArgType, yieldTuple, declareOp.getOriginalBase(),
+ firOpBuilder.getArrayAttr({firOpBuilder.getIntegerAttr(idxTy, 1)}));
+ return yieldTuple;
+ };
+
+ firOpBuilder.create<mlir::omp::YieldOp>(hsb.getAddr().getLoc(),
+ genYieldedVal());
symTable->popScope();
}
@@ -401,18 +454,19 @@ void DataSharingProcessor::doPrivatize(const Fortran::semantics::Symbol *sym) {
// First block argument corresponding to the original/host value while
// second block argument corresponding to the privatized value.
mlir::Block *copyEntryBlock = firOpBuilder.createBlock(
- ©Region, /*insertPt=*/{}, {symType, symType}, {symLoc, symLoc});
+ ©Region, /*insertPt=*/{}, {privatizerArgType, privatizerArgType},
+ {symLoc, symLoc});
firOpBuilder.setInsertionPointToEnd(copyEntryBlock);
- symTable->addSymbol(*sym, copyRegion.getArgument(0),
- /*force=*/true);
+ addSymbol(copyRegion, 0, true);
+
symTable->pushScope();
- symTable->addSymbol(*sym, copyRegion.getArgument(1));
+ addSymbol(copyRegion, 1, false);
+
auto ip = firOpBuilder.saveInsertionPoint();
copyFirstPrivateSymbol(sym, &ip);
- firOpBuilder.create<mlir::omp::YieldOp>(
- hsb.getAddr().getLoc(),
- symTable->shallowLookupSymbol(*sym).getAddr());
+ firOpBuilder.create<mlir::omp::YieldOp>(hsb.getAddr().getLoc(),
+ copyRegion.getArgument(1));
symTable->popScope();
}
@@ -423,7 +477...
[truncated]
|
ergawy
requested review from
kiranchandramohan,
luporl,
jeanPerier,
jsjodin,
raghavendhra,
agozillon,
DominikAdamski,
kparzysz,
bhandarkar-pranav and
skatrak
|
✅ With the latest revision this PR passed the C/C++ code formatter. |
Adds delayed privatization support for allocatables. In order to explain
the problem this diff is trying to solve, it would be useful to see an
example of `firstprivate` for an alloctable and how it is currently
emitted by flang **without** delayed privatization.
Consider the following Fortran code:
```fortran
subroutine delayed_privatization_allocatable
implicit none
integer, allocatable :: var1
!$omp parallel firstprivate(var1)
var1 = 10
!$omp end parallel
end subroutine
```
You would get something like this (again no delayed privatization yet):
```mlir
...
%3:2 = hlfir.declare %0 {fortran_attrs = #fir.var_attrs<allocatable>, ...} : ...
omp.parallel {
// Allocation logic
...
%9 = fir.load %3#1 : !fir.ref<!fir.box<!fir.heap<i32>>>
...
fir.if %12 {
...
} else {
...
}
%13:2 = hlfir.declare %8 {fortran_attrs = #fir.var_attrs<allocatable>, ...} : ...
// Copy logic
%14 = fir.load %13#0 : !fir.ref<!fir.box<!fir.heap<i32>>>
...
fir.if %17 {
%24 = fir.load %3#0 : !fir.ref<!fir.box<!fir.heap<i32>>>
...
}
...
omp.terminator
}
```
Note that for both the original and the private declaration, the
allocation and copy logic use both elements (e.g. `%3#0` and `%3#1`).
This poses the following problem for delayed privatization: how can
capture both values in order to pass them to the outlined privatizer?
The main issue is that `hlfir.declare` returns 2 SSA values that not
encapsulated together under a single composite value.
Some possible solutions are:
1. Change the return type of `hlfir.declare`. However, this would be
very invasive and therefore does not sound like a good idea.
2. Use the built-in `tuple` type to "pack" both values together. This is
reasonable but not very flexible since we won't be able to express
other information about the encapsulated values such as whether it is
alloctable.
3. Introduce a new concept/type that allows us to do the encapsulation
in a more flexible way. This is the "variable shadow" concept
introduced by this PR.
A "variable shadow" is a composite value of a special type:
`!fir.shadow` that has some special properties:
* It can be constructed only from `BlockArgument`s.
* It always has the type `!fir.shadow`.
* It packs the required information needed by the delayed privatizer
that correspond to the result of `hlfir.declare` operation.
We don't introduce any special opertions to deal with shadow values.
Rather we treat them similar to other composites. If you want to get a
certain component, use `fir.extract_value`. If you want to populate a
certain component, use `fir.insert_value`.
ergawy
force-pushed
the
delayed_privatization_11
branch
from
4e0b26e to
7d98b66
Compare
We have not made specific changes for HLFIR lowering in the privatisation code and could have inadvertently used both |
I am not sure to be honest. Since the 2 values can be of different types in some cases, I assumed both need to be maintained. I also guess it goes deeper than This is just from what I learned in the past 2 weeks staring at this part of the codebase, so I might be over complicating things for sure. |
|
I think using the following pattern will get you the first value and it should mostly be sufficient. |
|
The point I was trying to make was that the appearance of the Were the failures all in OpenMP tests? It will be good to get an opinion from @jeanPerier on whether creating clones |
An experimental diff to discuss using only the first component of `hlfir.declare`'s result for allocatables. This is just to have a more informed discussion on llvm#84033. This is not meant to be merged, at least for now.
|
Thanks a lot for the quick response and guidance @kiranchandramohan. Also, typical newbie misdirection on my side 😄. Just for the sake of discussion, I implemented a (very) dirty hack to use only the first component for privatization logic (allocation, copying, and deallocation) without delayed privatization in the picture to simplify the discussion. This is implemented in #84123. If using the first component is always safe and does not break any semantics I am not aware of, then I can productize: #84123 (i.e. find a less hacky solution) and then move to supporting delayed privatization for allocatables in a later PR. |
|
I am not very familiar with the OpenMP outlining. Here are my quick thoughts:
The story of In the context of OpenMP outlining, if you need to capture a descriptor, it should be I think I may be able to modify converter.getSymbolAddress to use |
|
Thanks @jeanPerier for your thoughts. From what you are describing:
I tested a small sample: subroutine foo()
interface
subroutine bar(x)
character(*) :: x(3:)
end subroutine
end interface
character(100) :: x(2:42)
call bar(x)
end subroutineAnd I thought that the func.func @_QPfoo() {
%c100 = arith.constant 100 : index
%c2 = arith.constant 2 : index
%c41 = arith.constant 41 : index
%0 = fir.alloca !fir.array<41x!fir.char<1,100>> {bindc_name = "x", uniq_name = "_QFfooEx"}
%1 = fir.shape_shift %c2, %c41 : (index, index) -> !fir.shapeshift<1>
%2:2 = hlfir.declare %0(%1) typeparams %c100 {uniq_name = "_QFfooEx"} : (!fir.ref<!fir.array<41x!fir.char<1,100>>>, !fir.shapeshift<1>, index) -> (!fir.box<!fir.array<41x!fir.char<1,100>>>, !fir.ref<!fir.array<41x!fir.char<1,100>>>)
%3 = fir.convert %2#0 : (!fir.box<!fir.array<41x!fir.char<1,100>>>) -> !fir.box<!fir.array<?x!fir.char<1,?>>>
fir.call @_QPbar(%3) fastmath<contract> : (!fir.box<!fir.array<?x!fir.char<1,?>>>) -> ()
return
}So, |
|
For the purpose of OpenMP private cloning and copying, it always safe to use #0 even for assume-shape. Is that correct? Yes, that is correct. Any usage of "#1" should be seen as optimizations to improve codegeneration at the FIR level. It is always correct to start with "#0".
The original value is not a fir.box in that case, so what I wrote cannot apply: one cannot save the creation of a descriptor and use Anyway, what I wrote above turns out to be incorrect since de2be3e that I missed and renders the usages of |
|
Awesome. The picture is a lot clearer now! Then, I will abandon this diff and try to find a way to cleanly use |
A follow-up question, based on you analysis of de2be3e, if it turns out the commit is actually not problematic, does it mean we can change |
Modifies the privatization logic so that the emitted code only used the HLFIR base (i.e. SSA value `#0` returned from `hlfir.declare`). Before that, that emitted privatization logic was a mix of using `#0` and `llvm#1` which leads to some difficulties trying to move to delayed privatization (see the discussion on llvm#84033).
Modifies the privatization logic so that the emitted code only used the HLFIR base (i.e. SSA value `#0` returned from `hlfir.declare`). Before that, that emitted privatization logic was a mix of using `#0` and `#1` which leads to some difficulties trying to move to delayed privatization (see the discussion on #84033).
|
Abandoning in favor of #84740. |
Adds delayed privatization support for allocatables. In order to explain the problem this diff is trying to solve, it would be useful to see an example of
firstprivatefor an alloctable and how it is currently emitted by flang without delayed privatization.Consider the following Fortran code:
You would get something like this (again no delayed privatization yet):
... %3:2 = hlfir.declare %0 {fortran_attrs = #fir.var_attrs<allocatable>, ...} : ... omp.parallel { // Allocation logic ... %9 = fir.load %3#1 : !fir.ref<!fir.box<!fir.heap<i32>>> ... fir.if %12 { ... } else { ... } %13:2 = hlfir.declare %8 {fortran_attrs = #fir.var_attrs<allocatable>, ...} : ... // Copy logic %14 = fir.load %13#0 : !fir.ref<!fir.box<!fir.heap<i32>>> ... fir.if %17 { %24 = fir.load %3#0 : !fir.ref<!fir.box<!fir.heap<i32>>> ... } ... omp.terminator }Note that for both the original and the private declaration, the allocation and copy logic use both elements (e.g.
%3#0and%3#1). This poses the following problem for delayed privatization: how can capture both values in order to pass them to the outlined privatizer? The main issue is thathlfir.declarereturns 2 SSA values that not encapsulated together under a single composite value.Some possible solutions are:
hlfir.declare. However, this would be very invasive and therefore does not sound like a good idea.tupletype to "pack" both values together. This is reasonable but not very flexible since we won't be able to express other information about the encapsulated values such as whether it is alloctable.A "variable shadow" is a composite value of a special type:
!fir.shadowthat has some special properties:BlockArguments.!fir.shadow.hlfir.declareoperation.We don't introduce any special opertions to deal with shadow values. Rather we treat them similar to other composites. If you want to get a certain component, use
fir.extract_value. If you want to populate a certain component, usefir.insert_value.