Can GMP be replaced by trees?

[Tobias Schlüter]

With Toon's recent remark about how incorrect rounding due to excess 
precision in our frontend may corrupt results, and Steve's work which 
revealed several minor mistakes in the GMP code in arith.c, I started 
wondering if our dependency on GMP can be removed.

The reason we have the dependency on GMP in our frontend is two-fold:
1. storing constants with machine-independent precision
2. folding constant expressions with machine-independent precision and 
without loss of precision
both are of course related to the problem of cross-compilation and 
reliable results across architectures.

Now AFAIK gcc already has these mechanisms implemented: trees for 
storing values, fold_const for simplifying expressions. If these didn't 
work well, gcc wouldn't be usable as a cross-compiler. I don't know if 
they work well enough for us, but I'm inclined to believe they do.

How difficult would it be to change gfortran to use gcc's mechanisms 
instead? I don't know. I suppose the easiest way is to create a 
pseudo-GMP interface which resembles GMP as much as possible and does 
everything on trees, and then incrementally replace existing code which 
uses these interfaces to either a simplified interface to these 
operations, or to operations directly manipulating trees.

Does anybody else -- especially people who know gcc's trees not only 
superficially like I do -- think this is a worthy project? Is there 
anything GMP gives us that trees wouldn't?

I understand that our priorities should be different currently, but if 
we decide to do this, I think we should plan to do this before a 
release, so that we never introduce a dependency on GMP in the real world.

Thanks,
- Tobi