Re: libgomp: Make GCC 5 OpenACC offloading executables work

[Thomas Schwinge <thomas@codesourcery.com>]

Hi!

On Wed, 11 May 2016 11:38:39 -0400, Nathan Sidwell <nathan@acm.org> wrote:
> On 05/11/16 10:22, Bernd Schmidt wrote:
> > On 05/11/2016 03:46 PM, Thomas Schwinge wrote:
> >>> What we now got, doesn't work, for several reasons.  GCC 5 OpenACC
> >>> offloading executables will just run into SIGSEGV.
> >
> > I'm tempted to say, let's just wait until someone actually reports that in
> > bugzilla. Offloading in gcc-5 was broken enough that I expect no one was
> > actually using it. There's really very little point in carrying compatibility
> > crud around.
> 
> I agree.  This would simply be enabling a poorly performing binary, rather than 
> encouraging a shiny new one.

I conceptually agree to that.  (If we're serious about that, then we can
remove more code, such as the legacy libgomp entry point itself -- a
"missing symbol: [...]" is still vaguely better than a SIGSEGV.)  Yet,
what I fixed here, is just what Jakub and Nathan agreed upon in
<http://news.gmane.org/find-root.php?message_id=%3C20150924084034.GC1847%40tucnak.redhat.com%3E>:
"GCC 5 compiled offloaded OpenACC/PTX code will always do host fallback".
Currently such code will always result in a SIGSEGV, which the patch
fixes.  (And, given that we now have this patch, it seems "unfair" to
"wait until someone actually reports that in bugzilla".)  Another option,
instead of having such legacy entry pointw do host fallback, is to
instead call gomp_fatal with a message like "re-compile your code with a
newer GCC version".


Grüße
 Thomas