Hello LLVM developer community,
I’m a graduate student of the Technical University of Eindhoven with a potential
project for GSOC in mind that I would like to discuss.
The project is related to integrating support for the tool SWEET  into LLVM.
Please allow me to explain briefly what this SWEET tool is about and why I
think that integrating support for the tool is an interesting addition to LLVM.
SWEET is a tool developed by a research team at the Mälardalen University in
Västerås, Sweden. The tool allows one to compute the execution time of the
longest path of execution in a piece of code, this is called the Worst Case
Execution Time. SWEET achieves this by using static analysis, meaning, that the
tool does not execute the code but looks at the code’s structure and variables.
Computing the longest path of your code is really important in certain areas
of embedded systems, where the application code running on an embedded processor
has to meet a certain deadline. An example would be in aerospace, where the
longest path in cycles is typically computed by hand (!). Miss computing the
WCET could have disastrous results (e.g. an exploding rocket, car crash)
Now is were it gets interesting. SWEET has a novel approach to supporting new
targets via its interface language called ALF . Typically, authors of
WCET static analysis tools have to implement support for specific target and
do not allow a third party to develop support for a new target. They also tend
to ask a large amount of money for supporting a new target. Did I mention
already that SWEET is open-source?
With this ALF interface language, all that is required to add a new target is
decompiling a given binary to ALF. Or: compiling it straight to ALF in the first
place. The GSOC project would be about making it easier to add for a certain
target a conversion to ALF. I suppose this would involve adding some sort of
abstraction in the code emission stage or in the MC layer. My knowledge of
the LLVM code base does not allow me to answer this question in full detail,
this is also one of the reasons why I am asking you for an opinion.
Note that going from IR to ALF is too high level and this leaves out important
details of the target. This can provide other advantages and has been attempted
before for the ALFBackend project . The ALFBackend project already contains
some classes that models ALF and provides features for generating ALF code. So
there is a potential re-use.
I think I should also mention that I am in in contact with the research team
responsible for developing SWEET and with an aerospace software company that
is to write software for an embedded processor  that is to be used in the
Ariane-6 for the trust vector control. This could be a first use case, and a target
for proof of concept.
I would love to hear what you think.