[flang-commits] [flang] 26d3655 - [flang][docs] Add lowering design doc for polymorphic entities
Valentin Clement via flang-commits
flang-commits at lists.llvm.org
Mon Aug 15 09:13:23 PDT 2022
Author: Valentin Clement
Date: 2022-08-15T18:13:15+02:00
New Revision: 26d3655a1557566d9d5731cd2584033ee845d3f5
URL: https://github.com/llvm/llvm-project/commit/26d3655a1557566d9d5731cd2584033ee845d3f5
DIFF: https://github.com/llvm/llvm-project/commit/26d3655a1557566d9d5731cd2584033ee845d3f5.diff
LOG: [flang][docs] Add lowering design doc for polymorphic entities
This document aims to give insights at the representation of polymorphic
entities in FIR and how polymorphic related constructs and features are lowered
to FIR.
Reviewed By: klausler, razvanlupusoru
Differential Revision: https://reviews.llvm.org/D131515
Added:
flang/docs/PolymorphicEntities.md
Modified:
Removed:
################################################################################
diff --git a/flang/docs/PolymorphicEntities.md b/flang/docs/PolymorphicEntities.md
new file mode 100644
index 0000000000000..a354791729566
--- /dev/null
+++ b/flang/docs/PolymorphicEntities.md
@@ -0,0 +1,871 @@
+# Polymorphic Entities
+
+A polymorphic entity is a data entity that can be of
diff erent type during the
+execution of a program.
+
+This document aims to give insights at the representation of polymorphic
+entities in FIR and how polymorphic related constructs and features are lowered
+to FIR.
+
+## Fortran standard
+
+Here is a list of the sections and constraints of the Fortran standard involved
+for polymorphic entities.
+
+- 7.3.2.1 - 7.3.2.2: TYPE specifier (TYPE(*))
+ - C708
+ - C709
+ - C710
+ - C711
+- 7.3.2.3: CLASS specifier
+- 7.5.4.5: The passed-object dummy argument
+ - C760
+- 9.7.1: ALLOCATE statement
+ - C933
+- 9.7.2: NULLIFY statement
+ - When a NULLIFY statement is applied to a polymorphic pointer (7.3.2.3),
+ its dynamic type becomes the same as its declared type.
+- 10.2.2.3: Data pointer assignment
+- 11.1.3: ASSOCIATE construct
+- 11.1.11: SELECT TYPE construct
+ - C1157
+ - C1158
+ - C1159
+ - C1160
+ - C1161
+ - C1162
+ - C1163
+ - C1164
+ - C1165
+- 16.9.76 EXTENDS_TYPE_OF (A, MOLD)
+- 16.9.165 SAME_TYPE_AS (A, B)
+- 16.9.184 STORAGE_SIZE (A [, KIND])
+- C.10.5 Polymorphic Argument Association (15.5.2.9)
+
+---
+
+## Representation in FIR
+
+### Polymorphic entities `CLASS(type1)`
+
+A polymorphic entity is represented as a class type in FIR. In the example below
+the dummy argument `p` is passed to the subroutine `foo` as a polymorphic entity
+with the extensible type `point`. The type information captured in the class is
+the best statically available at compile time.
+`!fir.class` is a new type introduced for polymorphic entities. It's similar to
+a box type but allows the distinction between a monomorphic and a polymorphic
+descriptor.
+A specific `BoxTypeInterface` (TypeInterface) can be introduced to share the
+same API for both types where it is necessary. `!fir.class` and `!fir.box` can
+also be based on a same `BaseBoxType` similar to the `BaseMemRefType` done for
+MemRef.
+
+**Fortran**
+```fortran
+type point
+ real :: x, y
+end type point
+
+type, extends(point) :: point_3d
+ real :: z
+end type
+
+subroutine foo(p)
+ class(point) :: p
+ ! code of the subroutine
+end subroutine
+```
+
+**FIR**
+```c
+func.func @foo(%p : !fir.class<!fir.type<_QTpoint{x:f32,y:f32}>>)
+```
+
+### Unlimited polymorphic entities `CLASS(*)`
+
+The unlimited polymorphic entity is represented as a class type with `*`.
+
+**Fortran**
+```fortran
+subroutine bar(x)
+ class(*) :: x
+ ! code of the subroutine
+end subroutine
+```
+
+**FIR**
+```c
+func.func @bar(%x : !fir.class<*>)
+```
+
+### Assumed-type `TYPE(*)`
+
+Assumed type is added in Fortran 2018 and it is available only for dummy
+arguments. It's mainly used for interfaces to non-Fortran code and is similar
+to C's `void`. It's not part of polymorphic entities directly but it's not
+currently implemented in flang.
+
+Assumed-type is represented as `!fir.type<*>`.
+
+### SELECT TYPE construct
+
+The `SELECT TYPE` construct select for execution at most one of its constituent
+block. The selection is based on the dynamic type of the selector.
+
+**Fortran**
+```fortran
+type point
+ real :: x, y
+end type point
+type, extends(point) :: point_3d
+ real :: z
+end type point_3d
+type, extends(point) :: color_point
+ integer :: color
+end type color_point
+
+type(point), target :: p
+type(point_3d), target :: p3
+type(color_point), target :: c
+class(point), pointer :: p_or_c
+p_or_c => c
+select type ( a => p_or_c )
+class is (point)
+ print*, a%x, a%y
+type is (point_3d)
+ print*, a%x, a%y, a%z
+class default
+ print*,
+end select
+```
+
+From the Fortran standard:
+> A `TYPE IS` type guard statement matches the selector if the dynamic type
+and kind type parameter values of the selector are the same as those specified
+by the statement. A `CLASS IS` type guard statement matches the selector if the
+dynamic type of the selector is an extension of the type specified by the
+statement and the kind type parameter values specified by the statement are the
+same as the corresponding type parameter values of the dynamic type of the
+selector.
+
+In the example above the `CLASS IS` type guard is matched.
+
+The construct is lowered to a specific FIR operation `fir.select_type`. It is
+similar to other FIR "select" operations such as `fir.select` and
+`fir.select_rank`. The dynamic type of the selector value is matched against a
+list of type descriptor. The `TYPE IS` type guard statement is represented by a
+`#fir.type_is` attribute and the `CLASS IS` type guard statement is represented
+by a `#fir.class_is` attribute.
+The `CLASS DEFAULT` type guard statement is represented by a `unit` attribute.
+
+**FIR**
+```
+fir.select_type %p : !fir.class<!fir.type<_QTpoint{x:f32,y:f32}>> [
+ #fir.class_is<!fir.type<_QTpoint{x:f32,y:f32}>>, ^bb1,
+ #fir.type_is<!fir.type<_QTpoint_3d{x:f32,y:f32,z:f32}>>, ^bb2,
+ unit, ^bb3]
+```
+
+Lowering of the `fir.select_type` operation will produce a if-then-else ladder.
+The testing of the dynamic type of the selector is done by calling runtime
+functions.
+
+The runtime has two functions to compare dynamic types . Note that this two
+functions _ignore_ the values of `KIND` type parameters. A version of these
+functions that does not _ignore_ the value of the `KIND` type parameters will
+be implemented for the `SELECT TYPE` type guards testing.
+
+Currently available functions for the `EXTENDS_TYPE_OF` and `SAME_TYPE_AS`
+intrinsics (`flang/include/flang/Evaluate/type.h`).
+```cpp
+std::optional<bool> ExtendsTypeOf(const DynamicType &) const;
+std::optional<bool> SameTypeAs(const DynamicType &) const;
+```
+
+**FIR** (lower level FIR/MLIR after conversion to an if-then-else ladder)
+```
+module {
+ func @f(%arg0: !fir.class<*>) -> i32 {
+ %c4_i32 = arith.constant 4 : i32
+ %c8_i32 = arith.constant 8 : i32
+ %c16_i32 = arith.constant 16 : i32
+ %0 = fir.gentypedesc !fir.tdesc<!fir.type<!fir.type<_QTpoint{x:f32,y:f32}>>>
+ %1 = fir.convert %arg0 : (!fir.class<!fir.type<_QTpoint{x:f32,y:f32}>>) -> !fir.box<none>
+ %2 = fir.convert %0 : (!fir.tdesc<!fir.type<!fir.type<_QTpoint{x:f32,y:f32}>>>) -> !fir.ref<none>
+ %3 = fir.call @ExtendsTypeOfWithKind(%1, %2) : (!fir.box<none>, !fir.ref<none>) -> i1
+ cond_br %3, ^bb2(%c4_i32 : i32), ^bb1
+ ^bb1: // pred: ^bb0
+ %4 = fir.gentypedesc !fir.type<_QTpoint_3d{x:f32,y:f32,z:f32}>
+ %5 = fir.convert %arg0 : (!fir.class<!fir.type<_QTpoint{x:f32,y:f32}>>) -> !fir.box<none>
+ %6 = fir.convert %4 : (!fir.tdesc<!fir.type<_QTpoint_3d{x:f32,y:f32,z:f32}>>) -> !fir.ref<none>
+ %7 = fir.call @SameTypeAsWithKind(%5, %6) : (!fir.box<none>, !fir.ref<none>) -> i1
+ cond_br %7, ^bb4(%c16_i32 : i32), ^bb3
+ ^bb2(%8: i32): // pred: ^bb0
+ return %8 : i32
+ ^bb3: // pred: ^bb1
+ br ^bb5(%c8_i32 : i32)
+ ^bb4(%9: i32): // pred: ^bb1
+ %10 = arith.addi %9, %9 : i32
+ return %10 : i32
+ ^bb5(%11: i32): // pred: ^bb3
+ %12 = arith.muli %11, %11 : i32
+ return %12 : i32
+ }
+ func private @ExactSameTypeAsWithKind(!fir.box<none>, !fir.ref<none>) -> i1
+ func private @SameTypeAsWithKind(!fir.box<none>, !fir.ref<none>) -> i1
+}
+```
+
+Note: some dynamic type checks can be inlined for performance. Type check with
+intrinsic types when dealing with unlimited polymorphic entities is an ideal
+candidate for inlined checks.
+
+---
+
+## Dynamic dispatch
+
+Dynamic dispatch is the process of selecting which implementation of a
+polymorphic procedure to call at runtime. The runtime already has information
+to be used in this process (more information can be found here:
+[RuntimeTypeInfo.md](RuntimeTypeInfo.md)).
+
+The declaration of the data structures are present in
+`flang/runtime/type-info.h`.
+
+In the example below, there is a basic type `shape` with two type extensions
+`triangle` and `rectangle`.
+The two type extensions override the `get_area` type-bound procedure.
+
+**UML**
+```
+
+ |---------------------|
+ | Shape |
+ |---------------------|
+ | + color:integer |
+ | + isFilled:logical |
+ |---------------------|
+ | + init() |
+ | + get_area():real |
+ |---------------------|
+ /\
+ /__\
+ |
+ |---------------------------------------------------|
+ | |
+ | |
+|---------------------| |---------------------|
+| triangle | | rectangle |
+|---------------------| |---------------------|
+| + base:real | | + length:real |
+| + height:real | | + width:real |
+|---------------------| |---------------------|
+| + get_area():real | | + get_area():real |
+|---------------------| |---------------------|
+
+```
+
+**Fortran**
+```fortran
+module geometry
+type :: shape
+ integer :: color
+ logical :: isFilled
+contains
+ procedure :: get_area => get_area_shape
+ procedure :: init => init_shape
+end type shape
+
+type, extends(shape) :: triangle
+ real :: base
+ real :: height
+contains
+ procedure :: get_area => get_area_triangle
+end type triangle
+
+type, extends(shape) :: rectangle
+ real :: length
+ real :: width
+contains
+ procedure :: get_area => get_area_rectangle
+end type rectangle
+
+type shape_array
+ class(shape), allocatable :: item
+end type
+
+contains
+
+function get_area_shape(this)
+ real :: get_area_shape
+ class(shape) :: this
+ get_area_shape = 0.0
+end function
+
+subroutine init_shape(this, color)
+ class(shape) :: this
+ integer :: color
+ this%color = color
+ this%isFilled = .false.
+end subroutine
+
+function get_area_triangle(this)
+ real :: get_area_triangle
+ class(triangle) :: this
+ get_area_triangle = (this%base * this%height) / 2
+end function
+
+function get_area_rectangle(this)
+ real :: get_area_rectangle
+ class(rectangle) :: this
+ get_area_rectangle = this%length * this%width
+end function
+
+function get_all_area(shapes)
+ real :: get_all_area
+ type(shape_array) :: shapes(:)
+ real :: sum
+ integer :: i
+
+ get_all_area = 0.0
+
+ do i = 1, size(shapes)
+ get_all_area = get_all_area + shapes(i)%item%get_area()
+ end do
+end function
+
+subroutine set_base_values(sh, v1, v2)
+ class(shape) :: sh
+ real, intent(in) :: v1, v2
+
+ select type (sh)
+ type is (triangle)
+ sh%base = v1
+ sh%height = v2
+ type is (rectangle)
+ sh%length = v1
+ sh%width = v2
+ class default
+ print*,'Cannot set values'
+ end select
+end subroutine
+
+end module
+
+program foo
+ use geometry
+
+ real :: area
+
+ type(shape_array), dimension(2) :: shapes
+
+ allocate (triangle::shapes(1)%item)
+ allocate (rectangle::shapes(2)%item)
+
+ do i = 1, size(shapes)
+ call shapes(i)%item%init(i)
+ end do
+
+ call set_base_values(shapes(1)%item, 2.0, 1.5)
+ call set_base_values(shapes(2)%item, 5.0, 4.5)
+
+ area = get_all_area(shapes)
+
+ print*, area
+
+ deallocate(shapes(1)%item)
+ deallocate(shapes(2)%item)
+end program
+```
+
+The `fir.dispatch` operation is used to perform a dynamic dispatch. This
+operation is comparable to the `fir.call` operation but for polymorphic
+entities.
+Call to `NON_OVERRIDABLE` type-bound procedure are resolved at compile time and
+a `fir.call` operation is emitted instead of a `fir.dispatch`.
+When the type of a polymorphic entity can be fully determined at compile
+time, a `fir.dispatch` op can even be converted to a `fir.call` op. This will
+be discussed in more detailed later in the document in the devirtualization
+section.
+
+**FIR**
+Here is simple example of the `fir.dispatch` operation. The operation specify
+the binding name of the type-bound procedure to be called and pass the
+descriptor as argument. If the `NOPASS` attribute is set then the descriptor is
+not passed as argument when lowered. If `PASS(arg-name)` is specified, the
+`fir.pass` attribute is added to point to the PASS argument in the
+`fir.dispatch` operation. `fir.nopass` attribute is added for the `NOPASS`. The
+descriptor still need to be present in the `fir.dispatch` operation for the
+dynamic dispatch. The CodeGen will then omit the descriptor in the argument
+of the generated call.
+
+The dispatch explanation focus only on the call to `get_area()` as seen in the
+example.
+
+**Fortran**
+```fortran
+get_all_area = get_all_area + shapes(i)%item%get_area()
+```
+
+**FIR**
+```c
+%1 = fir.convert %0 : (!fir.ref<!fir.class<!fir.type<_QMgeometryTtriangle{color:i32,isFilled:!fir.logical<4>,base:f32,height:f32>>>) -> !fir.ref<!fir.box<none>>
+%2 = fir.dispatch "get_area"(%1) : (!fir.ref<!fir.box<none>>) -> f32
+```
+
+The type information is stored in the `f18Addendum` of the descriptor. The
+format is defined in `flang/runtime/type-info.h` and part of its representation
+in LLVM IR is shown below. The binding is comparable to a vtable. Each derived
+type has a complete type-bound procedure table in which all of the bindings of
+its ancestor types appear first.
+
+**LLVMIR**
+
+Representation of the derived type information with the bindings.
+```c
+%_QM__fortran_type_infoTderivedtype = type { { ptr, i64, i32, i8, i8, i8, i8, [1 x [3 x i64]], ptr, [1 x i64] }, { ptr, i64, i32, i8, i8, i8, i8 }, i64, { ptr, i64, i32, i8, i8, i8, i8, ptr, [1 x i64] }, { ptr, i64, i32, i8, i8, i8, i8, [1 x [3 x i64]] }, { ptr, i64, i32, i8, i8, i8, i8, [1 x [3 x i64]] }, { ptr, i64, i32, i8, i8, i8, i8, [1 x [3 x i64]], ptr, [1 x i64] }, { ptr, i64, i32, i8, i8, i8, i8, [1 x [3 x i64]], ptr, [1 x i64] }, { ptr, i64, i32, i8, i8, i8, i8, [1 x [3 x i64]], ptr, [1 x i64] }, i32, i8, i8, i8, i8, [4 x i8] }
+%_QM__fortran_type_infoTbinding = type { %_QM__fortran_builtinsT__builtin_c_funptr, { ptr, i64, i32, i8, i8, i8, i8 } }
+%_QM__fortran_builtinsT__builtin_c_funptr = type { i64 }
+```
+
+The `fir.dispatch` is then lowered to use the runtime information to extract the
+correct function from the vtable and to perform the actual call. Here is
+what it can look like in pseudo LLVM IR code.
+
+**LLVMIR**
+```c
+// Retrieve the bindings (vtable) from the type information from the descriptor
+%1 = call %_QM__fortran_type_infoTbinding* @_FortranAGetBindings(%desc)
+// Retrieve the position of the specific bindings in the table
+%2 = call i32 @_FortranAGetBindingOffset(%1, "get_area")
+// Get the binding from the table
+%3 = getelementptr %_QM__fortran_type_infoTbinding, %_QM__fortran_type_infoTbinding* %1, i32 0, i32 %2
+// Get the function pointer from the binding
+%4 = getelementptr %_QM__fortran_builtinsT__builtin_c_funptr, %_QM__fortran_type_infoTbinding %3, i32 0, i32 0
+// Cast func pointer
+%5 = inttoptr i64 %4 to <procedure pointer>
+// Load the function
+%6 = load f32(%_QMgeometryTshape*)*, %5
+// Perform the actual function call
+%7 = call f32 %6(%_QMgeometryTshape* %shape)
+```
+
+_Note:_ functions `@_FortranAGetBindings` and `@_FortranAGetBindingOffset` are
+not available in the runtime and will need to be implemented.
+
+- `@_FortranAGetBindings` retrieves the bindings from the descriptor. The
+ descriptor holds the type information that holds the bindings.
+- `@_FortranAGetBindingOffset` retrieves the procedure offset in the bindings
+ based on the binding name provided.
+
+Retrieving the binding table and the offset are done separately so multiple
+dynamic dispatch on the same polymorphic entities can be optimized (the binding
+table is retrieved only once for multiple call).
+
+### Passing polymorphic entities as argument
+
+**Fortran**
+```fortran
+TYPE t1
+END TYPE
+TYPE, EXTENDS(t1) :: t2
+END TYPE
+```
+
+1) Dummy argument is fixed type and actual argument is fixed type.
+ - `TYPE(t1)` to `TYPE(t1)`: Nothing special to take into consideration.
+2) Dummy argument is polymorphic and actual argument is fixed type. In these
+ cases, the actual argument need to be boxed to be passed to the
+ subroutine/function since those are expecting a descriptor.
+ ```c
+ func.func @_QMmod1Ps(%arg0: !fir.class<!fir.type<_QMmod1Tshape{x:i32,y:i32}>>)
+ func.func @_QQmain() {
+ %0 = fir.alloca !fir.type<_QMmod1Tshape{x:i32,y:i32}> {uniq_name = "_QFEsh"}
+ %1 = fir.embox %0 : (!fir.ref<!fir.type<_QMmod1Tshape{x:i32,y:i32}>>) -> !fir.class<!fir.type<_QMmod1Tshape{x:i32,y:i32}>>
+ fir.call @_QMmod1Ps(%1) : (!fir.class<!fir.type<_QMmod1Tshape{x:i32,y:i32}>>) -> ()
+ return
+ }
+ ```
+ - `TYPE(t1)` to `CLASS(t1)`
+ - `TYPE(t2)` to `CLASS(t1)`
+ - `TYPE(t1)` to `CLASS(t2)` - Invalid
+ - `TYPE(t2)` to `CLASS(t2)`
+3) Actual argument is polymorphic and dummy argument is fixed type. These case
+ are restricted to the declared type of the polymorphic entities.
+ - The simple case is when the actual argument is a scalar
+ polymorphic entity passed to a non-PDT. The caller just extract the
+ base address from the descriptor and pass it to the function.
+ - In other cases, the caller needs to perform a copyin/copyout since it
+ cannot just extract the base address of the `CLASS(T)` because it is
+ likely not contiguous.
+ - `CLASS(t1)` to `TYPE(t1)`
+ - `CLASS(t2)` to `TYPE(t1)` - Invalid
+ - `CLASS(t1)` to `TYPE(t2)` - Invalid
+ - `CLASS(t2)` to `TYPE(t2)`
+4) Both actual and dummy arguments are polymorphic. These particular cases are
+ straight forward. The function expect polymorphic entities already.
+ The boxed type is passed without change.
+ - `CLASS(t1)` to `CLASS(t1)`
+ - `CLASS(t2)` to `CLASS(t1)`
+ - `CLASS(t1)` to `CLASS(t2)` - Invalid
+ - `CLASS(t2)` to `CLASS(t2)`
+
+### User-Defined Derived Type Input/Output
+
+User-Defined Derived Type Input/Output allows to define how a derived-type
+is read or written from/to a file.
+
+There are 4 basic subroutines that can be defined:
+- Formatted READ
+- Formatted WRITE
+- Unformatted READ
+- Unformatted WRITE
+
+Here are their respective interfaces:
+
+**Fortran**
+```fortran
+subroutine read_formatted(dtv, unit, iotype, v_list, iostat, iomsg)
+subroutine write_formatted(dtv, unit, iotype, v_list, iostat, iomsg)
+subroutine read_unformatted(dtv, unit, iotype, v_list, iostat, iomsg)
+subroutine write_unformatted(dtv, unit, iotype, v_list, iostat, iomsg)
+```
+
+When defined on a derived-type, these specific type-bound procedures are stored
+as special bindings in the type descriptor (see `SpecialBinding` in
+`flang/runtime/type-info.h`).
+
+With a derived-type the function call to `@_FortranAioOutputDescriptor` from IO
+runtime will be emitted in lowering.
+
+**Fortran**
+```fortran
+type(t) :: x
+write(10), x
+```
+
+**FIR**
+```c
+%5 = fir.call @_FortranAioBeginUnformattedOutput(%c10_i32, %4, %c56_i32) : (i32, !fir.ref<i8>, i32) -> !fir.ref<i8>
+%6 = fir.embox %2 : (!fir.ref<!fir.type<_QTt>>) -> !fir.class<!fir.type<_QTt>>
+%7 = fir.convert %6 : (!fir.class<!fir.type<_QTt>>) -> !fir.box<none>
+%8 = fir.call @_FortranAioOutputDescriptor(%5, %7) : (!fir.ref<i8>, !fir.box<none>) -> i1
+%9 = fir.call @_FortranAioEndIoStatement(%5) : (!fir.ref<i8>) -> i32
+```
+
+When dealing with polymorphic entities the call to IO runtime can stay
+unchanged. The runtime function `OutputDescriptor` can make the dynamic dispatch
+to the correct binding stored in the descriptor.
+
+### Finalization
+
+The `FINAL` specifies a final subroutine that might be executed when a data
+entity of that type is finalized. Section 7.5.6.3 defines when finalization
+occurs.
+
+Final subroutines like User-Defined Derived Type Input/Output are stored as
+special bindings in the type descriptor. The runtime is able to handle the
+finalization with a call the the `@_FortranADestroy` function
+(`flang/include/flang/Runtime/derived-api.h`).
+
+**FIR**
+```c
+%5 = fir.call @_FortranADestroy(%desc) : (!fir.box<none>) -> none
+```
+
+The `@_FortranADestroy` function will take care to call the final subroutines
+and the ones from the parent type.
+
+Appropriate call to finalization have to be lowered at the right places (7.5.6.3
+When finalization occurs).
+
+### Devirtualization
+
+Sometimes there is enough information at compile-time to avoid going through
+a dynamic dispatch for a type-bound procedure call on a polymorphic entity. To
+be able to perform this optimization directly in FIR the dispatch table is also
+present statically with the `fir.dispatch_table` and `fir.dt_entry` operations.
+
+Here is an example of these operations representing the dispatch tables for the
+same example than for the dynamic dispatch.
+
+**FIR**
+```
+fir.dispatch_table @_QMgeometryE.dt.shape {
+ fir.dt_entry init, @_QMgeometryPinit_shape
+ fir.dt_entry get_area, @_QMgeometryPget_area_shape
+}
+
+fir.dispatch_table @_QMgeometryE.dt.rectangle {
+ fir.dt_entry init, @_QMgeometryPinit_shape
+ fir.dt_entry get_area, @_QMgeometryPget_area_rectangle
+}
+
+fir.dispatch_table @_QMgeometryE.dt.triangle {
+ fir.dt_entry init, @_QMgeometryPinit_shape
+ fir.dt_entry get_area, @_QMgeometryPget_area_triangle
+}
+```
+
+With this information, an optimization pass can replace `fir.dispatch`
+operations with `fir.call` operations to the correct functions when the type is
+know at compile time.
+
+This is the case in a `type is` type-guard block as illustrated below.
+
+**Fortran**
+```fortran
+subroutine get_only_triangle_area(sh)
+ class(shape) :: sh
+ real :: area
+
+ select type (sh)
+ type is (triangle)
+ area = sh%get_area()
+ class default
+ area = 0.0
+ end select
+
+end subroutine
+```
+
+**FIR**
+
+The call to `get_area` in the `type is (triangle)` guard can be replaced.
+```c
+%3 = fir.dispatch "get_area"(%desc)
+// Replaced by
+%3 = fir.call @get_area_triangle(%desc)
+```
+
+Another example would be the one below. In this case as well, a dynamic dispatch
+is not necessary and a `fir.call` can be emitted instead.
+
+**Fortran**
+```fortran
+real :: area
+class(shape), pointer :: sh
+type(triangle), target :: tr
+
+sh => tr
+
+area = sh%get_area()
+```
+
+Note that the frontend is already replacing some of the dynamic dispatch calls
+with the correct static ones. The optimization pass is useful for cases not
+handled by the frontend and especially cases showing up after some other
+optimizations are applied.
+
+### `ALLOCATE`/`DEALLOCATE` statements
+
+The allocation and deallocation of polymorphic entities are delegated to the
+runtime.
+The corresponding function signatures can be found in
+`flang/include/flang/Runtime/allocatable.h` and in
+`flang/include/flang/Runtime/pointer.h` for pointer allocation.
+
+`ALLOCATE`
+
+The `ALLOCATE` statement is lowered to runtime calls as shown in the example
+below.
+
+**Fortran**
+```fortran
+allocate(triangle::shapes(1)%item)
+allocate(rectangle::shapes(2)%item)
+```
+
+**FIR**
+```c
+%0 = fir.alloca !fir.class<!fir.type<_QMgeometryTtriangle{color:i32,isFilled:!fir.logical<4>,base:f32,height:f32>>
+%1 = fir.alloca !fir.class<!fir.type<_QMgeometryTtriangle{color:i32,isFilled:!fir.logical<4>,base:f32,height:f32}>>
+%3 = fir.convert %0 : (!fir.ref<!fir.class<!fir.type<_QMgeometryTtriangle{color:i32,isFilled:!fir.logical<4>,base:f32,height:f32>>>) -> !fir.ref<!fir.box<none>>
+%4 = fir.gentypedesc !fir.type<_QMgeometryTtriangle{color:i32,isFilled:!fir.logical<4>,base:f32,height:f32}>>
+%5 = fir.call @_FortranAAllocatableInitDerived(%3, %4)
+
+%6 = fir.convert %1 : (!fir.ref<!fir.class<_QMgeometryTtriangle{color:i32,isFilled:!fir.logical<4>,base:f32,height:f32}>>>) -> !fir.ref<!fir.box<none>>
+%7 = fir.gentypedesc !fir.type<_QMgeometryTtriangle{color:i32,isFilled:!fir.logical<4>,base:f32,height:f32}>> %8 = fir.call @_FortranAAllocatableInitDerived(%6, %7)
+```
+
+For pointer allocation, the `PointerAllocate` function is used.
+
+`DEALLOCATE`
+
+The `DEALLOCATE` statement is lowered to a runtime call to
+`AllocatableDeallocate` and `PointerDeallocate` for pointers.
+
+**Fortran**
+```fortran
+deallocate(shapes(1)%item)
+deallocate(shapes(2)%item)
+```
+
+**FIR**
+```c
+%8 = fir.call @_FortranAAllocatableDeallocate(%desc1)
+%9 = fir.call @_FortranAAllocatableDeallocate(%desc2)
+```
+
+### `EXTENDS_TYPE_OF`/`SAME_TYPE_AS` intrinsics
+
+`EXTENDS_TYPE_OF` and `SAME_TYPE_AS` intrinsics have implementation in the
+runtime. Respectively `SameTypeAs` and `ExtendsTypeOf` in
+`flang/include/flang/Evaluate/type.h`.
+
+Both intrinsic functions are lowered to their respective runtime calls.
+
+### Assignment / Pointer assignment
+
+Intrinsic assignment of an object to another is already implemented in the
+runtime. The function `@_FortranAAsssign` performs the correct operations.
+
+Available in `flang/include/flang/Runtime/assign.h`.
+
+### User defined assignment and operator
+
+**Fortran**
+```fortran
+module mod1
+type t1
+contains
+ procedure :: assign_t1
+ generic :: assignment(=) => assign_t1
+end type t1
+
+type, extends(t1) :: t2
+end type
+
+contains
+
+subroutine assign_t1(to, from)
+ class(t1), intent(inout) :: to
+ class(t1), intent(in) :: from
+ ! Custom code for the assignment
+end subroutine
+
+subroutine assign_t2(to, from)
+ class(t2), intent(inout) :: to
+ class(t2), intent(in) :: from
+ ! Custom code for the assignment
+end subroutine
+
+end module
+
+program main
+use mod
+
+class(t1), allocatable :: v1
+class(t1), allocatable :: v2
+
+allocate(t2::v1)
+allocate(t2::v2)
+
+v2 = v1
+
+end program
+```
+
+In the example above the assignment `v2 = v1` is done by a call to `assign_t1`.
+This is resolved at compile time since `t2` could not have a generic type-bound
+procedure for assignment with an interface that is not distinguishable. This
+is the same for user defined operators.
+
+### `NULLIFY`
+
+When a `NULLIFY` statement is applied to a polymorphic pointer (7.3.2.3), its
+dynamic type becomes the same as its declared type.
+
+The `NULLIFY` statement is lowered to a call to the corresponding runtime
+function `PointerNullifyDerived` in `flang/include/flang/Runtime/pointer.h`.
+
+### Impact on existing FIR operations dealing with descriptors
+
+Currently, FIR has a couple of operations taking descriptors as inputs or
+producing descriptors as outputs. These operations might need to deal with the
+dynamic type of polymorphic entities.
+
+- `fir.load`/`fir.store`
+ - Currently a `fir.load` of a `fir.box` is a special case. In the code
+ generation no copy is made. This could be problematic with polymorphic
+ entities. When a `fir.load` is performed on a `fir.class` type, the dynamic
+ can be copied.
+
+ **Fortran**
+ ```fortran
+ module mod1
+ class(shape), pointer :: a
+ contains
+ subroutine sub1(a, b)
+ class(shape) :: b
+ associate (b => a)
+ ! Some more code
+ end associate
+ end subroutine
+ end module
+ ```
+
+ In the example above, the dynamic type of `a` and `b` might be
diff erent. The
+ dynamic type of `a` must be copied when it is associated on `b`.
+
+ **FIR**
+ ```c
+ // fir.load must copy the dynamic type from the pointer `a`
+ %0 = fir.address_of(@_QMmod1Ea) : !fir.ref<!fir.class<!fir.ptr<!fir.type<_QMmod1Tshape{x:i32,y:i32}>>>>
+ %1 = fir.load %0 : !fir.ref<!fir.class<!fir.ptr<!fir.type<_QMmod1Tshape{x:i32,y:i32}>>>>
+ ```
+
+- `fir.embox`
+ - The embox operation is used to create a descriptor from a reference. With
+ polymorphic entities, it is used to create a polymorphic descriptor from
+ a derived type. The declared type of the descriptor and the derived type
+ are identical. The dynamic type of the descriptor must be set when it is
+ created. This is already handled by lowering.
+
+- `fir.rebox`
+ - The rebox operation is used to create a new descriptor from a another
+ descriptor with new optional dimension. If the original descriptor is a
+ polymorphic entities its dynamic type must be propagated to the new
+ descriptor.
+ ```
+ %0 = fir.slice %c10, %c33, %c2 : (index, index, index) -> !fir.slice<1>
+ %1 = fir.shift %c0 : (index) -> !fir.shift<1>
+ %2 = fir.rebox %x(%1)[%0] : (!fir.class<!fir.array<?x!fir.type<>>>, !fir.shift<1>, !fir.slice<1>) -> !fir.class<!fir.array<?x!fir.type<>>>
+ ```
+---
+
+# Testing
+
+- Lowering part is tested with LIT tests in tree
+- Polymorphic entities involved a lot of runtime information so executable
+ tests will be useful for full testing.
+
+---
+
+# Current TODOs
+Current list of TODOs in lowering:
+- `flang/lib/Lower/Allocatable.cpp:465` not yet implemented: SOURCE allocation
+- `flang/lib/Lower/Allocatable.cpp:468` not yet implemented: MOLD allocation
+- `flang/lib/Lower/Allocatable.cpp:471` not yet implemented: polymorphic entity allocation
+- `flang/lib/Lower/Bridge.cpp:448` not yet implemented: create polymorphic host associated copy
+- `flang/lib/Lower/Bridge.cpp:2185` not yet implemented: assignment to polymorphic allocatable
+- `flang/lib/Lower/Bridge.cpp:2288` not yet implemented: pointer assignment involving polymorphic entity
+- `flang/lib/Lower/Bridge.cpp:2316` not yet implemented: pointer assignment involving polymorphic entity
+- `flang/lib/Lower/CallInterface.cpp:795` not yet implemented: support for polymorphic types
+- `flang/lib/Lower/ConvertType.cpp:237` not yet implemented: support for polymorphic types
+
+Current list of TODOs in code generation:
+
+- `flang/lib/Optimizer/CodeGen/CodeGen.cpp:897` not yet implemented: fir.dispatch codegen
+- `flang/lib/Optimizer/CodeGen/CodeGen.cpp:911` not yet implemented: fir.dispatch_table codegen
+- `flang/lib/Optimizer/CodeGen/CodeGen.cpp:924` not yet implemented: fir.dt_entry codegen
+- `flang/lib/Optimizer/CodeGen/CodeGen.cpp:2651` not yet implemented: fir.gentypedesc codegen
+
+---
+
+Resources:
+- [1] https://www.pgroup.com/blogs/posts/f03-oop-part1.htm
+- [2] https://www.pgroup.com/blogs/posts/f03-oop-part2.htm
+- [3] https://www.pgroup.com/blogs/posts/f03-oop-part3.htm
+- [4] https://www.pgroup.com/blogs/posts/f03-oop-part4.htm
+- [5] Modern Fortran explained
More information about the flang-commits
mailing list