Re: [Gm2] Re: Over-aggressive Optimizations with gcc 4.1.1

[Gaius Mulley]

John O Goyo <address@hidden> writes:

> Greetings:
> 
> Gaius Mulley wrote:
> > John O Goyo <address@hidden> writes:
> >
> >> Greetings:
> >>
> >> I am trying gm2 with gcc 4.1.1 under Solaris 10 on Sparc (ilp32).  As
> >> Gaius said, gm2 builds, most of check-gm2 passes, gm2.paranoid fails,
> >> and it installs (except the man page).  However, optimization does
> >> some bizarre things.
> >>
> >> Witness.
> >>
> [...]
> > Also thanks for reporting that 4.1.1 and gm2 builds on Solaris 10.
> > This is the first time I've heard this combination - and it gets as
> > far as my LP64 Athlon GNU/Linux (different endianness iirc).
> 
> I see the same results using 4.1.1 on Solaris 10 with i32lp64 with the
> usual extra test failures on adding addresses.  (As a side note, I
> find that trying to build gm2 with i32lp64 with my resident gcc-3.x
> fails with the gm2 libraries compiled into ELF32.   However, if I
> build gcc-4.1.1 first and then build gm2 with gcc-4.1.1, the gm2
> libraries are correctly compiled into ELF64.  I do not understand what
> happens here. The only difference in the configuration options is
> adding gm2 to the languages.)

Hi John,

I saw something similar to this - and certainly if the default path
points to a different gcc version (when building gm2) then the linking
can become confused.  Hence I've altered gm2 and subcomponents so that
it explicitly calls the version of gcc it needs.  It appears to have
cleared up a number of link errors (incorrect spec .... messages)
when performing "make check-gm2".

It might also have solved the problem you are seeing above..?

Also worth mentioning is that I've added more permutations of options
to the regression tests - with the result that "make check-gm2" takes
5 times as long, sadly.  However it does include your test in
regression format (which fails currently) when linking to a
addition.mod when compiled with -O2..

Best wishes,
Gaius