High Level Overview¶
LFortran is structured around two independent modules, AST and ASR, both of which are standalone (completely independent of the rest of LFortran) and users are encouraged to use them independently for other applications and build tools on top:
Abstract Syntax Tree (AST), module
lfortran.ast: Represents any Fortran source code, strictly based on syntax, no semantic is included. The AST module can convert itself to Fortran source code.
Abstract Semantic Representation (ASR), module
lfortran.asr: Represents a valid Fortran source code, all semantic is included. Invalid Fortran code is not allowed (an error will be given). The ASR module can convert itself to an AST.
The LFortran compiler is then composed of the following independent stages:
Parsing: converts Fortran source code to an AST
Semantic: converts an AST to an ASR
High level optimizations: optimize ASR to a possibly faster/simpler ASR (things like inlining functions, eliminating redundant expressions or statements, etc.)
LLVM IR code generation and lower level optimizations: converts an ASR to an LLVM IR. This stage also does all other optimizations that do not produce an ASR, but still make sense to do before passing to LLVM IR.
Machine code generation: LLVM then does all its optimizations and generates machine code (such as a binary executable, a library, an object file, or it is loaded and executed using JIT as part of the interactive LFortran session or in a Jupyter kernel).
LFortran is structured as a library, and so one can for example use the parser to obtain an AST and do something with it, or one can then use the semantic analyzer to obtain ASR and do something with it. One can generate the ASR directly (e.g., from SymPy) and then either convert to AST and to a Fortran source code, or use LFortran to compile it to machine code directly. In other words, one can use LFortran to easily convert between the three equivalent representations:
Fortran source code
Abstract Syntax Tree (AST)
Abstract Semantic Representation (ASR)
They are all equivalent in the following sense:
Any ASR can always be converted to an equivalent AST
Any AST can always be converted to an equivalent Fortran source code
Any Fortran source code can always be either converted to an equivalent AST or one gets a syntax error
Any AST can always be either converted to an equivalent ASR or one gets a semantic error
So when a conversion can be done, they are equivalent, and the conversion can always be done unless the code is invalid.
ASR Design Details¶
The ASR is designed to have the following features:
ASR is still semantically equivalent to the original Fortran code (it did not lose any semantic information). ASR can be converted to AST, and AST to Fortran source code which is functionally equivalent to the original.
ASR is as simple as possible: it does not contain any information that could not be inferred from ASR.
The ASR C++ classes (down the road) are designed similarly to SymEngine: they are constructed once and after that they are immutable. The constructor checks in Debug more that all the requirements are met (e.g., that all Variables in a Function have a dummy argument set, that explicit-shape arrays are not allocatable and all other Fortran requirements to make it a valid code), but in Release mode it quickly constructs the class without checks. Then there are builder classes that construct the ASR C++ classes to meet requirements (checked in Debug mode) and the builder gives an error message if a code is not a valid Fortran code, and if it doesn't give an error message, then the ASR C++ classes are constructed correctly. Thus by construction, the ASR classes always contain valid Fortran code and the rest of LFortran can depend on it.
Information that is lost when parsing source to AST: whitespace, multiline/single line if statement distinction, case sensitivity of keywords.
Information that is lost when going from AST to ASR:
detailed syntax how variables were defined and the order of type attributes (whether array dimension is using the
dimension attribute, or parentheses at the variable; or how many variables there are per declaration line or their order), as ASR only represents the aggregated type information in the symbol table.
ASR is the simplest way to generate Fortran code, as one does not have to worry about the detailed syntax (as in AST) about how and where things are declared. One specifies the symbol table for a module, then for each symbol (functions, global variables, types, ...) one specifies the local variables and if this is an interface then one needs to specify where one can find an implementation, otherwise a body is supplied with statements, those nodes are almost the same as in AST, except that each variable is just a reference to a symbol in the symbol table (so by construction one cannot have undefined variables). The symbol table for each node such as Function or Module also references its parent (for example a function references a module, a module references the global scope).
The ASR can be directly converted to an AST without gathering any other information. And the AST directly to Fortran source code.
The ASR is always representing a semantically valid Fortran code. This is enforced by checks in the ASR C++ constructors (in Debug build). When an ASR is used, one can assume it is valid.
Fortran 2008 standard chapter 2 "Fortran concepts" specifies that Fortran code is a collection of program units (either all in one file, or in separate files), where each program unit is one of:
module or submodule
function or subroutine
Note: It can also be a block data program unit, that is used to provide initial values for data objects in named common blocks, but we do not recommend the use of common blocks (use modules instead).
We extend the Fortran language by introducing a global scope, which is not only the list of program units (as in F2008) but can also include statements, declarations, use statements and expressions. We define global scope as a collection of the following items:
module or submodule
function or subroutine
In addition, if a variable is not defined in an assignment statement (such as
x = 5+3) then the type of the variable is inferred from the right hand side
x = 5+3 would be of type
y = 5._dp
would be of type
real(dp)). This rule only applies at the top level of
global scope. Types must be fully specified inside main programs, modules,
functions and subroutines, just like in F2008.
The global scope has its own symbol table. The main program and module/submodule do not see any symbols from this symbol table. But functions, subroutines, statements and expressions at the top level of global scope use and operate on this symbol table.
The global scope has the following symbols predefined in the symbol table:
the usual standard set of Fortran functions (such as
dpdouble precision symbol, so that one can use
5._dpfor double precision.
Each item in the global scope is interpreted as follows: main program is compiled into an executable with the same name and executed; modules, functions and subroutines are compiled and loaded; use statement and declaration adds those symbols with the proper type into the global scope symbol table, but do not generate any code; statement is wrapped into an anonymous subroutine with no arguments, compiled, loaded and executed; expression is wrapped into an anonymous function with no arguments returning the expression, compiled, loaded, executed and the return value is returned to the user.
The global scope is always interpreted, item by item, per the previous
paragraph. It is meant to allow interactive usage, experimentations and
writing simple scripts. Code in global scope must be interpreted using
lfortran. For more complex (production) code it is recommended to turn it
into modules and programs (by wrapping loose statements into subroutines or
functions and by adding type declarations) and compile it with
any other Fortran compiler.
Here are some examples of valid code in global scope:
a = 5 print *, a
a = 5 subroutine p() print *, a end subroutine call p()
module a implicit none integer :: i end module use a, only: i i = 5
x = [1, 2, 3] y = [1, 2, 1] call plot(x, y, "o-")
The LFortran extension of Fortran was chosen in a way so as to minimize the number of changes. In particular, only the top level of the global scope has relaxed some of the Fortran rules (such as making specifying types optional) so as to allow simple and quick interactive usage, but inside functions, subroutines, modules or programs this relaxation does not apply.
The number of changes were kept to minimum in order to make it straightforward to turn code at global scope into standard compliant Fortran code using programs and modules, so that it can be compiled by any Fortran compiler.