Generics

From a high-level perspective generics are supported in LFortran by three main elements:

1. Requirements declaring deferred types and their abstract functions

2. Templates implementing generic symbols (functions, subroutines, derived types) using deferred types and abstract functions described by requirements

3. Instantiations by passing concrete types and functions into templates

Requirements

Requirements declare deferred types (generic types) and its associated functions, similar to typeclasses in Haskell and traits in Rust. For example, the signature for a generic monoid of any type can be represented by the following requirement:

requirement monoid(T, op)
  ! declaring a deferred type (generic type)
  type, deferred :: T
  ! declaring a function associated with the deferred type
  function op(x, y) result(z)
    type(T), intent(in) :: x, y
    type(T) :: z
  end function
  function empty() result(z)
    type(T) :: z
  end function
end requirement

The monoid requirement declares a deferred type T and two functions,op that takes arguments of type T and return a value with the same type T and empty that returns a value of type T. The deferred type T is internally interpreted as a variable T typed as TypeParameter T (for brevity we will write it just as T). Functions declared in requirements are abstract (without function body).

On ASR level, the requirement is represented by the symbol Requirement. ASR for requirements are built during symbol table visit in the function visit_Requirement. monoid’s symbol table would contain a variable T with the type T, a function op with the type T * T -> T, and a function empty with the type () -> T.

A requirement is checked to see if all the parameters have a corresponding symbol declared inside of it. A warning is generated if a parameter is declared but no symbol is found with the same name.

On its own, a requirement does not do any computation. It is also not compiled into the target language when the ASR is compiled.

Templates

Templates take the role of a scope for generic functions. A template uses generic types and its associated type signatures obtained from requirements to check generic operations.

As a running example, we will consider a generic function for n-times multiplication for argument of any type. This function can be represented as:

module generics_example
  ! same requirement as before
  requirement monoid(T, op, empty)
    ...
  end requirement

  ! the template starts from here
  template array_t(S, op_temp, empty_temp)
    require :: monoid(S, op_temp, empty_temp)
  contains
    ! below is the generic function
    function array_sum(arr) result(r)
      type(S), intent(in) :: arr(:)
      type(S) :: r
      integer :: n, i
      n = size(arr)
      r = empty_temp(0)
      if (n > 0) then
        r = arr(1)
        do i = 2, n
          res = op_temp(r, arr(i))
        end do
      end if
    end function
  end template
end module

The template array_t contains the generic function array_sum that takes an array arr of type S. The template later on will be added to the parent symbol table as a Template symbol with the name array_t.

Typing context has to be obtained to type the parameter arr and the addition operation op_temp(r, arr(i)). Such typing context can be made available within the scope of the template by using a requirement. Here, the Require statement require :: monoid(S, op_temp, empty_temp) builds the types of the template’s parameters S, op_temp, and empty_temp based on the types of the symbols in monoid.

This corresponds to the visit_UnitRequire that can be found in visit_Template during symbol table visit. Calling a require statement copies the symbols from the requiremement and replaces their names with the argument names given by the require statement. In this case, the symbol T, op, and empty from the requirement monoid are replaced by S, op_temp, and empty_temp that are passed as arguments. This replacement is done by the function rename_symbol.

Eventually the require statement adds into array_t’s symbol table two symbols, S as a variable with type S, op_temp and empty_temp as functions with types S * S -> S and () -> S based on monoid’s symbol table.

Symbol table visit checks the variable declarations in array_sum. Since S is available in the symbol table now, both variables arr and r can be typed. Later during body visit, with type S, op_temp, and empty_temp in the symbol table, the function call op_temp(r, arr(i)) in array_sum would be checked in the same way as non-generic functions.

ASR representation of templates are also not compiled into the target language.

Instantiations

Generic functions need to be instantiated with concrete types and functions to be used in run-time. The process of instantiation replaces the generic types in the function definition with concrete types (such as integer, real) and replace the abstract functions with implemented functions.

The generic function described in the previous section can be instantiated as a function that computes array sum by the following instantiation:

module functions
  public :: add_integer, empty_integer
contains
  function add_integer(x, y)
    integer, intent(in) :: x, y
    integer :: add_integer
    add_integer = x + y
  end function
  function empty_integer()
    integer :: empty_integer
    empty_integer = 0
  end function
end module

program instantiate_template
use functions, only: add_integer, empty_integer

instantiate array_t(integer, add_integer, empty_integer), &
  only: array_sum_integer => array_sum

! function argument for basic arithmetic operations can be replaced with an operator
instantiate array_t(integer, operator(+), empty_integer), &
  only: array_sum_integer => array_sum

! eliding function renaming would instantiating all generic functions with the same name
instantiate array_t(integer, operator(+), empty_integer)
end program

This instantiation statement

instantiate array_t(integer, add_integer, empty_integer), only: array_sum_integer => array_sum

passes the type integer, an integer addition function add_integer, and a function describing empty integer value empty_integer as arguments to template array_t, then instantiates the function array_sum as a new function named array_sum_integer. This instantiation wants to replace S in array_t with the type integer, op_temp function calls with add_integer function calls, and empty_temp function calls with empty_integer function calls.

Type Checking

Before a function is generated on ASR level by an instantiation, the compiler checks the consistency of its type substitution based on the given symbol arguments. Currently there is no notion of subtyping in LFortran, so checking is limited to exact type checks. This is done by tracking the type substitutions made by the symbol arguments and rejecting any contradicting type subsitutition. Checking is done during symbol table visit in visit_Instantiate.

We will explain this in detail through the instantiation example above. The first argument integer substitutes the type parameter S in template array_t (maps S to integer). The second argument add_integer substitutes the type S * S -> S of op_temp with integer * integer -> integer (checks S substitution). This is allowed because the function argument gives a consistent substitution with the previous argument. If it passes the type checks, the instantiated function would be generated as a function on ASR level. Successfully checked symbol arguments are put inside the variable type_subs for type substitution (S to integer) and symbol_subs for function substitution (op_temp to add_integer).

Now, suppose that we have the following instantiation where instead of integer, we pass real as an argument:

instantiate array_t(real, add_integer, empty_integer), only: array_sum_integer => array_sum

Here, the first type argument maps S to real but the second function argument maps S to integer, resulting in a contradiction. Hence ending in a type error.

This check is tracked by the two variables type_subs and symbol_subs. Function arguments are checked by the function check_restriction during symbol table visit.

Handling Operator Arguments

Instantiations can be made simpler for basic binary operations (addition, multiplication, etc.) by passing operator(<sign>) instead of pre-defined functions.

If the type substitution tracked by type_subs is consistent for a binary arithmetic operation, then a function is implicitly generated as a function argument. From the example before, we will have a function that is equivalent to the following LFortran function generated implicitly:

function ~add_intrinsic(arg0, arg1) result(ret)
  integer, intent(in) :: arg0, arg1
  integer :: ret
  ret = arg0 + arg1
end function

ASR Generation

Instantiating the function as a function on the ASR level is simply a symbol replacement process.

The generation process is split into two, generating the function signature during symbol table visit and generating the function body during body visit. The whole generation is not done during body visit because derived types generated from a template may be used to type variable declarations.

During symbol table visit, after checking arguments and placing the symbol substitution into type_subs and symbol_subs, the instantiated function is added into the symbol table and its signature is built by the function instantiate_symbol where the deferred types are replaced with concrete types. For the example above, a function equivalent to the following is generated:

elemental function array_sum_integer(arr) result(res)
  integer, intent(in) :: arr(:)   ! type replaced
  integer :: r                    ! type replaced
  integer :: n, i
end function

The substitution in type_subs and symbol_subs are preserved and then passed to the body visitor. During body visit, the body of the instantiated function is built in visit_Instantiate by the function instantiate_body, replacing any abstract functions with concrete functions:

function array_sum_integer(n, a) result(res)
  integer, intent(in) :: arr(:)
  integer :: r
  integer :: n, i
  n = size(arr)
  r = empty_integer(0)              ! function call replaced
  if (n > 0) then
    r = arr(1)
    do i = 2, n
      res = add_integer(r, arr(i))  ! function call replaced
    end do
  end if
  end do
end function

也可以看看

• Programming With Generics, for simpler explaining about using generics in LFortran