Re: Interfacing with external code using libgccjit

[David Malcolm <dmalcolm@redhat.com>]

On Thu, 2023-06-15 at 21:59 +0200, Anselm Schüler wrote:
> Hi, I’m very new to libgccjit and was hoping to use it to implement a
> DSL that abstracts over some high-precision mathematics.

Hi!

> However, to do that, I need to interface with a high-precision
> library 
> such as GMP or MPFR.
> I’m wondering if that’s possible. 

It ought to be possible, but you need some way to express the
declarations from the library's headers via the ligccjit API; we don't
yet have a convenient way to take a header and express it in libgccjit
form.  As you note, doing it "by hand" is laborious; sorry that we 
don't yet have an automated way of doing this.  Possible approaches
would be to use a library that can parse C headers, such as:
  https://github.com/vnmakarov/mir
or perhaps using libabigail to work with the ABI representation:
  https://sourceware.org/libabigail/
(which would be parsing the debug data)


> Can I use 
> gcc_jit_context_new_function with GCC_JIT_FUNCTION_IMPORTED to define
> such functions from e.g. MPFR? 

Yes.


> If so, is the function retrieved from 
> the binary itself or does it need to be installed on the target 
> machine? 

It's equivalent to parsing a declaration from a header file; it doesn't
bring in the definition of the function.

> Do I need to pass -lmpfr to libgccjit or does it suffice to 
> use it during the compilation of the host program?

I believe you'd need to pass "-lmpfr" to
gcc_jit_context_add_driver_option.


> Additionally, it seems to be necessary to redefine all the types 
> defined by the library in order to be able to pass and receive them.
> Is 
> there some facility for automating this? It seems rather laborious.
> 

Yes, as noted above, this is currently a major drawback, and an
automated way of doing this would be very helpful.

Hope this is helpful
Dave