Overall operations chain of Tapenade

Overall operations chain of Tapenade

The structure of the source reflects our view of Tapenade as a chain of successive steps, that go from a set of source files to a set of differentiated files. Each step operates on one representation of the code and either enriches the given representation, or transforms it into another representation. Three sorts of representation of a program are used:

  • Files, i.e. textual files in some programming language, exist obviously at both ends of the chain. In addition to files as “text”, there are options to produce HTML files able to display output files in a GUI.

  • Abstract Syntax Trees (ASTs) exist mostly as the output of initial parsing and as the input of final pretty-printing. ASTs are also used in the central Internal Representation to hold individual code statements. Tapenade ASTs are (almost) independent from the files language (Fortran,C…). They are expressed in a special language that we call “IL” for “Imperative Language”, a sort of union of the constructs that one may find in actual languages. Code for ASTs is (mostly) in packages utils README.md and representation README.md.

  • Internal Representation (IR) is a graph-based representation of the complete given set of files. It is a call graph of flow graphs, together with a tree of Symbol Tables. Code for the IR is in package representation.

The successive steps are:

  • Parsing, that translates the set of input files into their ASTs. Code for this step is in packages front, frontend, frontf, frontc, frontcpp etc.

  • IR construction, that translates the incoming parsed ASTs into the Internal Representation. Code for this step is in package representation.

  • Analysis, that runs useful data-flow analysis on the complete IR, thus enriching the IR. Code for this step is in package analysis, see README.md.

  • Differentiation, that builds a new, differentiated IR, from the given IR. Code for this step is in package differentiation.

  • Tree regeneration, that translates the differentiated IR into new, differentiated ASTs. Code for this step is in package ir2tree.

  • Pretty-Print, that translates the differentiated ASTs into output files. Code for this step is in package prettyprint.

Two other packages exist on the side:

  • Package utils holds very basic objects used everywhere, such as trees, lists, bitsets, etc. See README.md.

  • Package gui defines a simple gui for Tapenade, and a few other graphical utilities (for debug…)

Refer to each package for more details on its implementation.

Note that the Parsing packages are not located here in fr/inria/tapenade, but in a different place. This is in part because parsers are complex, and may be defined with different tools and languages. In fact parsers are considered external to Tapenade itself, and the “official” input of Tapenade is just ASTs using formalism IL. Of course Tapenade distribution does provide parsers for Fortran, C, for convenience, but we like to think that one may define their own parser for another imperative language, make it spit out a set of IL ASTs, and connect it to the input of Tapenade. We provide the definition of the IL formalism for that purpose.