In #173438 I added a FIR specific loop invariant code motion pass. During the review, Tom pointed out certain limitations about OpenMP dialect operations that should be taken into consideration during transformations such as LICM: https://github.com/llvm/llvm-project/pull/173438#discussion_r2657612148 I also found issues with hoisting operations out of `acc.loop` operations in certain conditions (see the added test in `licm.fir`). I am proposing a new operation interface that will allow to control movement of operations during MLIR transformations. In particular, I propose two methods (there might be more): * op.canMoveOutOf(cand) - returns true, if it is allowed to move 'cand' operation out of 'op'. * op.canMoveFromDescendant(descendant, cand) - return true, if it is allowed to move 'cand' out of 'descendant' and into 'op'. I used the new interface to get rid of explicit OpenMP interfaces checks in Flang's LICM, and I also used it for `acc.loop` operation (though, I provided conservative initial implementation). The new interface is part of FIR dialect, but I think it would better fit into the core MLIR set of interfaces so that the checks that I make in Flang's LICM are actually done in `mlir::moveLoopInvariantCode`. Moreover, other code movement transformations that may appear in MLIR may also need to use such an interface. I would like to get some feedback on whether it is reasonable to move the interface to core MLIR.
103 lines
4.3 KiB
C++
103 lines
4.3 KiB
C++
//===-- FIROpenMPOpsInterfaces.cpp ----------------------------------------===//
|
|
//
|
|
// 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
|
|
//
|
|
//===----------------------------------------------------------------------===//
|
|
/// \file
|
|
/// This file implements FIR operation interfaces, which may be attached
|
|
/// to OpenMP dialect operations.
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
#include "flang/Optimizer/Dialect/FIROperationMoveOpInterface.h"
|
|
#include "flang/Optimizer/OpenMP/Support/RegisterOpenMPExtensions.h"
|
|
#include "mlir/Dialect/OpenMP/OpenMPDialect.h"
|
|
|
|
namespace {
|
|
/// Helper template that must be specialized for each operation.
|
|
/// The methods are declared just for documentation.
|
|
template <typename OP, typename Enable = void>
|
|
struct OperationMoveModel {
|
|
// Returns true if it is allowed to move the given 'candidate'
|
|
// operation from the 'descendant' operation into operation 'op'.
|
|
// If 'candidate' is nullptr, then the caller is querying whether
|
|
// any operation from any descendant can be moved into 'op' operation.
|
|
bool canMoveFromDescendant(mlir::Operation *op, mlir::Operation *descendant,
|
|
mlir::Operation *candidate) const;
|
|
|
|
// Returns true if it is allowed to move the given 'candidate'
|
|
// operation out of operation 'op'. If 'candidate' is nullptr,
|
|
// then the caller is querying whether any operation can be moved
|
|
// out of 'op' operation.
|
|
bool canMoveOutOf(mlir::Operation *op, mlir::Operation *candidate) const;
|
|
};
|
|
|
|
// Helpers to check if T is one of Ts.
|
|
template <typename T, typename... Ts>
|
|
struct is_any_type : std::disjunction<std::is_same<T, Ts>...> {};
|
|
|
|
template <typename T, typename... Ts>
|
|
struct is_any_omp_op
|
|
: std::integral_constant<
|
|
bool, is_any_type<typename std::remove_cv<T>::type, Ts...>::value> {};
|
|
|
|
template <typename T, typename... Ts>
|
|
constexpr bool is_any_omp_op_v = is_any_omp_op<T, Ts...>::value;
|
|
|
|
/// OperationMoveModel specialization for OMP_LOOP_WRAPPER_OPS.
|
|
template <typename OP>
|
|
struct OperationMoveModel<
|
|
OP,
|
|
typename std::enable_if<is_any_omp_op_v<OP, OMP_LOOP_WRAPPER_OPS>>::type>
|
|
: public fir::OperationMoveOpInterface::ExternalModel<
|
|
OperationMoveModel<OP>, OP> {
|
|
bool canMoveFromDescendant(mlir::Operation *op, mlir::Operation *descendant,
|
|
mlir::Operation *candidate) const {
|
|
// Operations cannot be moved from descendants of LoopWrapperInterface
|
|
// operation into the LoopWrapperInterface operation.
|
|
return false;
|
|
}
|
|
bool canMoveOutOf(mlir::Operation *op, mlir::Operation *candidate) const {
|
|
// The LoopWrapperInterface operations are only supposed to contain
|
|
// a loop operation, and it is probably okay to move operations
|
|
// from the descendant loop operation out of the LoopWrapperInterface
|
|
// operation. For now, return false to be conservative.
|
|
return false;
|
|
}
|
|
};
|
|
|
|
/// OperationMoveModel specialization for OMP_OUTLINEABLE_OPS.
|
|
template <typename OP>
|
|
struct OperationMoveModel<
|
|
OP, typename std::enable_if<is_any_omp_op_v<OP, OMP_OUTLINEABLE_OPS>>::type>
|
|
: public fir::OperationMoveOpInterface::ExternalModel<
|
|
OperationMoveModel<OP>, OP> {
|
|
bool canMoveFromDescendant(mlir::Operation *op, mlir::Operation *descendant,
|
|
mlir::Operation *candidate) const {
|
|
// Operations can be moved from descendants of OutlineableOpenMPOpInterface
|
|
// operation into the OutlineableOpenMPOpInterface operation.
|
|
return true;
|
|
}
|
|
bool canMoveOutOf(mlir::Operation *op, mlir::Operation *candidate) const {
|
|
// Operations cannot be moved out of OutlineableOpenMPOpInterface operation.
|
|
return false;
|
|
}
|
|
};
|
|
|
|
// Helper to call attachInterface<OperationMoveModel> for all Ts
|
|
// (types of operations).
|
|
template <typename... Ts>
|
|
void attachInterfaces(mlir::MLIRContext *ctx) {
|
|
(Ts::template attachInterface<OperationMoveModel<Ts>>(*ctx), ...);
|
|
}
|
|
} // anonymous namespace
|
|
|
|
void fir::omp::registerOpInterfacesExtensions(mlir::DialectRegistry ®istry) {
|
|
registry.addExtension(
|
|
+[](mlir::MLIRContext *ctx, mlir::omp::OpenMPDialect *dialect) {
|
|
attachInterfaces<OMP_LOOP_WRAPPER_OPS>(ctx);
|
|
attachInterfaces<OMP_OUTLINEABLE_OPS>(ctx);
|
|
});
|
|
}
|