PR 8146

[Janis Johnson]

On Fri, Nov 01, 2002 at 07:03:43PM -0500, Paul Jarc wrote:
> Joe Buck <jbuck@synopsys.com> wrote:
> > I suppose the next question is to find out what's different.  The easiest
> > way would be to capture the assembly for the three .o files that come out
> > different and compare that.
> 
> Ok, so I should build 2.95.3 using 3.2, but just one build instead of
> a full bootstrap.  Then I'll use that 2.95.3 and a normal 2.95.3 to
> compile each file to assembly, and then I'll post that assembly.
> Right?  So how do I do a build without a bootstrap?
> 
> Or do you want the assembly produced by 2.95.3 and by 3.2?  I've got
> that for 2.95.3's cccp.c: <URL:http://multivac.cwru.edu/prj/8146/>
> 
> The other two differing files, f/data.o and f/target.o, are never
> compiled by the host compiler during a bootstrap.  So 3.2's output for
> 2.95.3's sources for those isn't relevant, right?

What's needed is to understand the difference in the code generated by
the stage1 compiler built by GCC 3.2-branch and the stage2 compiler
built by the stage1 compiler; both were built from the same source and
so should generate exactly the same code.

I get the comparison failures for the two Fortran files even when both
of those compilers are built with -g, which makes me think there's an
uninitialized variable somewhere in the 2.95.3 sources.  The first dump
of RTL differs.  I don't have time to look into this, but that might 
get someone else going.

Janis