llvm-project/flang/docs/Parsing.md

<!--===- docs/Parsing.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
  
-->

# The F18 Parser

```eval_rst
.. contents::
   :local:
```

This program source code implements a parser for the Fortran programming
language.

The draft ISO standard for Fortran 2018 dated July 2017 was used as the
primary definition of the language.  The parser also accepts many features
from previous versions of the standard that are no longer part of the Fortran
2018 language.

It also accepts many features that have never been part of any version
of the standard Fortran language but have been supported by previous
implementations and are known or suspected to remain in use.  As a
general principle, we want to recognize and implement any such feature
so long as it does not conflict with requirements of the current standard
for Fortran.

The parser is implemented in standard ISO C++ and requires the 2017
edition of the language and library.  The parser constitutes a reentrant
library with no mutable or constructed static data.  Best modern C++
programming practices are observed to ensure that the ownership of
dynamic memory is clear, that value rather than object semantics are
defined for the data structures, that most functions are free from
invisible side effects, and that the strictest available type checking
is enforced by the C++ compiler when the Fortran parser is built.
Class inheritance is rare and dynamic polymorphism is avoided in favor
of modern discriminated unions.  To the furthest reasonable extent, the
parser has been implemented in a declarative fashion that corresponds
closely to the text of the Fortran language standard.

The several major modules of the Fortran parser are composed into a
top-level Parsing class, by means of which one may drive the parsing of a
source file and receive its parse tree and error messages.  The interfaces
of the Parsing class correspond to the two major passes of the parser,
which are described below.

## Prescanning and Preprocessing

The first pass is performed by an instance of the Prescanner class,
with help from an instance of Preprocessor.

The prescanner generates the "cooked character stream", implemented
by a CookedSource class instance, in which:
* line ends have been normalized to single ASCII LF characters (UNIX newlines)
* all `INCLUDE` files have been expanded
* all continued Fortran source lines have been unified
* all comments and insignificant spaces have been removed
* fixed form right margins have been clipped
* extra blank card columns have been inserted into character literals
  and Hollerith constants
* preprocessing directives have been implemented
* preprocessing macro invocations have been expanded
* legacy `D` lines in fixed form source have been omitted or included
* except for the payload in character literals, Hollerith constants,
  and character and Hollerith edit descriptors, all letters have been
  normalized to lower case
* all original non-ASCII characters in Hollerith constants have been
  decoded and re-encoded into UTF-8

Lines in the cooked character stream can be of arbitrary length.

The purpose of the cooked character stream is to enable the implementation
of a parser whose sole concern is the recognition of the Fortran language
from productions that closely correspond to the grammar that is presented
in the Fortran standard, without having to deal with the complexity of
all of the source-level concerns in the preceding list.

The implementation of the preprocessor interacts with the prescanner by
means of _token sequences_.  These are partitionings of input lines into
contiguous virtual blocks of characters, and are the only place in this
Fortran compiler in which we have reified a tokenization of the program
source; the parser proper does not have a tokenizer.  The prescanner
builds these token sequences out of source lines and supplies them
to the preprocessor, which interprets directives and expands macro
invocations.  The token sequences returned by the preprocessor are then
marshaled to constitute the cooked character stream that is the output of
the prescanner.

The preprocessor and prescanner can both instantiate new temporary
instances of the Prescanner class to locate, open, and process any
include files.

The tight interaction and mutual design of the prescanner and preprocessor
enable a principled implementation of preprocessing for the Fortran
language that implements a reasonable facsimile of the C language
preprocessor that is fully aware of Fortran's source forms, line
continuation mechanisms, case insensitivity, token syntax, &c.

The preprocessor always runs.  There's no good reason for it not to.

The content of the cooked character stream is available and useful
for debugging, being as it is a simple value forwarded from the first major
pass of the compiler to the second.

## Source Provenance

The prescanner constructs a chronicle of every file that is read by the
parser, viz. the original source file and all others that it directly
or indirectly includes.  One copy of the content of each of these files
is mapped or read into the address space of the parser.  Memory mapping
is used initially, but files with DOS line breaks or a missing terminal
newline are immediately normalized in a buffer when necessary.

The virtual input stream, which marshals every appearance of every file
and every expansion of every macro invocation, is not materialized as
an actual stream of bytes.  There is, however, a mapping from each byte
position in this virtual input stream back to whence it came (maintained
by an instance of the AllSources class).  Offsets into this virtual input
stream constitute values of the Provenance class.  Provenance values,
and contiguous ranges thereof, are used to describe and delimit source
positions for messaging.

Further, every byte in the cooked character stream supplied by the
prescanner to the parser can be inexpensively mapped to its provenance.
Simple `const char *` pointers to characters in the cooked character
stream, or to contiguous ranges thereof, are used as source position
indicators within the parser and in the parse tree.

## Messages

Message texts, and snprintf-like formatting strings for constructing
messages, are instantiated in the various components of the parser with
C++ user defined character literals tagged with `_err_en_US`, `_warn_en_US`,
`port_en_US`, and `_en_US` to signify severity and language; the default
language is the dialect of English used in the United States.

All "fatal" errors that do not immediately abort compilation but do
prevent the generation of binary and module files are `_err_en_US`.
Warnings about detected flaws in the program that probably indicate
problems worth attention are `_warn_en_US`.
Non-conforming extensions, legacy features, and obsolescent or deleted
features will raise `_port_en_US` messages when those are enabled.
Other messages have a simple `_en_US` suffix, including all messages
that are explanatory attachments.

As described above, messages are associated with
source code positions by means of provenance values.

## The Parse Tree

Each of the ca. 450 numbered requirement productions in the standard
Fortran language grammar, as well as the productions implied by legacy
extensions and preserved obsolescent features, maps to a distinct class
in the parse tree so as to maximize the efficacy of static type checking
by the C++ compiler.

A transcription of the Fortran grammar appears with production requirement
numbers in the commentary before these class definitions, so that one
may easily refer to the standard (or to the parse tree definitions while
reading that document).

Three paradigms collectively implement most of the parse tree classes:
* *wrappers*, in which a single data member `v` has been encapsulated
  in a new type
* *tuples* (or product types), in which several values of arbitrary
  types have been encapsulated in a single data member `t` whose type
  is an instance of `std::tuple<>`
* *discriminated unions* (or sum types), in which one value whose type is
  a dynamic selection from a set of distinct types is saved in a data
  member `u` whose type is an instance of `std::variant<>`

The use of these patterns is a design convenience, and exceptions to them
are not uncommon wherever it made better sense to write custom definitions.

Parse tree entities should be viewed as values, not objects; their
addresses should not be abused for purposes of identification.  They are
assembled with C++ move semantics during parse tree construction.
Their default and copy constructors are deliberately deleted in their
declarations. 

The std::list<> data type is used in the parse tree to reliably store pointers
to other relevant entries in the tree. Since the tree lists are moved and
spliced at certain points std::list<> provides the necessary guarantee of the
stability of pointers into these lists.

There is a general purpose library by means of which parse trees may
be traversed.

## Parsing

This compiler attempts to recognize the entire cooked character stream
(see above) as a Fortran program.  It records the reductions made during
a successful recognition as a parse tree value.  The recognized grammar
is that of a whole source file, not just of its possible statements,
so the parser has no global state that tracks the subprogram hierarchy
or the structure of their nested block constructs.  The parser performs
no semantic analysis along the way, deferring all of that work to the
next pass of the compiler.

The resulting parse tree therefore necessarily contains ambiguous parses
that cannot be resolved without recourse to a symbol table.  Most notably,
leading assignments to array elements can be misrecognized as statement
function definitions, and array element references can be misrecognized
as function calls.  The semantic analysis phase of the compiler performs
local rewrites of the parse tree once it can be disambiguated by symbols
and types.

Formally speaking, this parser is based on recursive descent with
localized backtracking (specifically, it will not backtrack into a
successful reduction to try its other alternatives).  It is not generated
as a table or code from a specification of the Fortran grammar; rather, it
_is_ the grammar, as declaratively respecified in C++ constant expressions
using a small collection of basic token recognition objects and a library
of "parser combinator" template functions that compose them to form more
complicated recognizers and their correspondences to the construction
of parse tree values.

## Unparsing

Parse trees can be converted back into free form Fortran source code.
This formatter is not really a classical "pretty printer", but is
more of a data structure dump whose output is suitable for compilation
by another compiler.  It is also used for testing the parser, since a
reparse of an unparsed parse tree should be an identity function apart from
source provenance.