On Fri, 2014-01-10 at 13:38 -0500, Adam Jensen wrote:

> Is the mcode variant preferred over the gcc variant? What's the most
> likely cause of ghdl-gcc failing tests while ghdl-mcode passes the
> tests?

Generally no, neither is preferred; they serve different purposes and
each has its strengths.

One recent user found the gcc version choked with huge auto-generated
VHDL source files (along the lines of gate-level netlists), consuming
gigabytes of swap space during the (gcc backend) optimisation passes
then falling over with allocation errors, while the mcode version sailed
through the same test with 10% or less of the memory foootprint.

Another user routinely links to C code (and I have linked to Ada) using
the VHPI interface; as the mcode compiler supports neither of these
languages, the gcc version is essential to him.

> How significant is the run-time performance difference between the two?
> (Personally, I am interested in the run-time performance of [large]
> models with complex data structures - not just RTL and I especially
> don't care about net-list simulation performance). Has anyone performed
> any side-by-side characterizations of the two? 

I have performed precisely ONE such comparison and to my surprise, the
penalty using the mcode compiler was only 20% (vs gcc at -O or -O1; at
the time, -O2 and -O3 failed to compile; since fixed).

More such comparisons could be useful.

- Brian


_______________________________________________
Ghdl-discuss mailing list
[email protected]
https://mail.gna.org/listinfo/ghdl-discuss

Reply via email to