RFC: Test optimizations for multiple architectures

[H.J. Lu]

On Thu, Nov 11, 2010 at 2:42 PM, Roland McGrath <roland@redhat.com> wrote:
>> --enable-test-multi-arch is intended for development build, not for production
>> build. My glibc development build environment is totally separate from the
>> production build environment. That is true for my gcc, binutils and gdb builds.
>
> People who have to maintain production builds need to run full test suites
> on the actual production code.  Otherwise it's just not testing you can
> rely on.  Doing more testing on your far-from-production development builds
> is of course better than not doing that testing.  But it's not a substitute
> for more complete testing on the real builds.
>
>> You need to be very careful about circular dependency. The current runtime
>> selection doesn't require relocations. Anything more complex may require
>> relocations at run-time.
>
> I think you may have misunderstood what I proposed.  The way ifunc works
> makes it essentially impossible to use anything in the selectors that
> requires other relocation, so we would never change it to be more complex.
> What I suggested involved no changes whatsoever in how the current ifunc
> selector code works.  The new call would just be a way to preinitialize
> those bitmaps so the selector code never runs cpuid et al.
>

What you suggested is certainly different from mine.
I want a way to test all ifunc functions in one run. You
want to give user control which implementation the
ifunc selector should return. I think they are independent
of each other.


-- 
H.J.