correctness and comprehensiveness of compilers [was: Re: C compiletime]

[tm_gccmail@mail.kloo.net]

On Wed, 18 Jun 2003, Vladimir Makarov wrote:

> Karel Gardas wrote:
> > 
> > On Wed, 18 Jun 2003, Scott Robert Ladd wrote:
> > 
> > Just one example: all last three releases of Intel C++ compiler for Linux
> > (I mean 6.0, 7.0 and 7.1) miscompile MICO (http://www.mico.org) such that
> > its IDL compiler is either crashing (6.0 + 7.0) or loops infinitely (7.1).
> > 
> 
>   I don't know about MICO test.  But it is amazing for me how such
> respectable company as Intel could make a release (7.1) which can not
> compile correctly some SPECINT2000 tests (vpr and gcc) with -O2 and -O3
> for P4 (their mainstream processor).  I found that a week ago using the
> demo CD which we got on the gcc summit.
> 
> Vlad

I've noticed hardware vendor's compilers usually are pretty buggy. They're
basically vehicles for getting good numbers on certain benchmarks, and
often contain optimizations useless in the real world and/or fail to
contain basic optimizations necessary for good performance on real-world
programs.

(Evidently they don't use icc for SPEC...)

For example, the Renesas H8 compiler generates the worst code ever for
switch() that I've ever seen - it generates a completely linear sequence
of compares and branches - no binary searching at all. Guess what happens
when you compile a bytecode interpreter which decodes bytecodes using a
huge switch() statement with this compiler...

Toshi