flang/docs/C++17.md

<!--===- docs/C++17.md

   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

-->

## C++14/17 features used in f18

The C++ dialect used in this project constitutes a subset of the
standard C++ programming language and library features.
We want our dialect to be compatible with the LLVM C++ language
subset that will be in use at the time that we integrate with that
project.
We also want to maximize portability, future-proofing,
compile-time error checking, and use of best practices.

To that end, we have a C++ style guide (q.v.) that lays
out the details of how our C++ code should look and gives
guidance about feature usage.

We have chosen to use some features of the recent C++17
language standard in f18.
The most important of these are:
* sum types (discriminated unions) in the form of `std::variant`
* `using` template parameter packs
* generic lambdas with `auto` argument types
* product types in the form of `std::tuple`
* `std::optional`

(`std::tuple` is actually a C++11 feature, but I include it
in this list because it's not particularly well known.)

### Sum types

First, some background information to explain the need for sum types
in f18.

Fortran is notoriously problematic to lex and parse, as tokenization
depends on the state of the partial parse;
the language has no reserved words in the sense that C++ does.
Fortran parsers implemented with distinct lexing and parsing phases
(generated by hand or with tools) need to implement them as
coroutines with complicated state, and experience has shown that
it's hard to get them right and harder to extend them as the language
evolves.

Alternatively, with the use of backtracking, one can parse Fortran with
a unified lexer/parser.
We have chosen to do so because it is simpler and should reduce
both initial bugs and long-term maintenance.

Specifically, f18's parser uses the technique of recursive descent with
backtracking.
It is constructed as the incremental composition of pure parsing functions
that each, when given a context (location in the input stream plus some state),
either _succeeds_ or _fails_ to recognize some piece of Fortran.
On success, they return a new state and some semantic value, and this is
usually an instance of a C++ `struct` type that encodes the semantic
content of a production in the Fortran grammar.

This technique allows us to specify both the Fortran grammar and the
representation of successfully parsed programs with C++ code
whose functions and data structures correspond closely to the productions
of Fortran.

The specification of Fortran uses a form of BNF with alternatives,
optional elements, sequences, and lists.  Each of these constructs
in the Fortran grammar maps directly in the f18 parser to both
the means of combining other parsers as alternatives, &c., and to
the declarations of the parse tree data structures that represent
the results of successful parses.
Move semantics are used in the parsing functions to acquire and
combine the results of sub-parses into the result of a larger
parse.

To represent nodes in the Fortran parse tree, we need a means of
handling sum types for productions that have multiple alternatives.
The bounded polymorphism supplied by the C++17 `std::variant` fits
those needs exactly.
For example, production R502 in Fortran defines the top-level
program unit of Fortran as being a function, subroutine, module, &c.
The `struct ProgramUnit` in the f18 parse tree header file
represents each program unit with a member that is a `std::variant`
over the six possibilities.
Similarly, the parser for that type in the f18 grammar has six alternatives,
each of which constructs an instance of `ProgramUnit` upon the result of
parsing a `Module`, `FunctionSubprogram`, and so on.

Code that performs semantic analysis on the result of a successful
parse is typically implemented with overloaded functions.
A function instantiated on `ProgramUnit` will use `std::visit` to
identify the right alternative and perform the right actions.
The call to `std::visit` must pass a visitor that can handle all
of the possibilities, and f18 will fail to build if one is missing.

Were we unable to use `std::variant` directly, we would likely
have chosen to implement a local `SumType` replacement; in the
absence of C++17's abilities of `using` a template parameter pack
and allowing `auto` arguments in anonymous lambda functions,
it would be less convenient to use.

The other options for polymorphism in C++ at the level of C++11
would be to:
* loosen up compile-time type safety and use a unified parse tree node
  representation with an enumeration type for an operator and generic
  subtree pointers, or
* define the sum types for the parse tree as abstract base classes from
  which each particular alternative would derive, and then use virtual
  functions (or the forbidden `dynamic_cast`) to identify alternatives
  during analysis

### Product types

Many productions in the Fortran grammar describe a sequence of various
sub-parses.
For example, R504 defines the things that may appear in the "specification
part" of a subprogram in the order in which they are allowed: `USE`
statements, then `IMPORT` statements, and so on.

The parse tree node that represents such a thing needs to incorporate
the representations of those parses, of course.
It turns out to be convenient to allow these data members to be anonymous
components of a `std::tuple` product type.
This type facilitates the automation of code that walks over all of the
members in a type-safe fashion and avoids the need to invent and remember
needless member names -- the components of a `std::tuple` instance can
be identified and accessed in terms of their types, and those tend to be
distinct.

So we use `std::tuple` for such things.
It has also been handy for template metaprogramming that needs to work
with lists of types.

### `std::optional`

This simple little type is used wherever a value might or might not be
present.
It is especially useful for function results and
rvalue reference arguments.
It corresponds directly to the optional elements in the productions
of the Fortran grammar.
It is also used as a wrapper around a parse tree node type to define the
results of the various parsing functions, where presence of a value
signifies a successful recognition and absence denotes a failed parse.
It is used in data structures in place of nullable pointers to
avoid indirection as well as the possible confusion over whether a pointer
is allowed to be null.