Hi!

Many thanks, Alexandra and Tobias for working this out together!


On 2020-06-23T06:50:26-0300, Alexandre Oliva <oliva@adacore.com> wrote:
> On Jun  9, 2020, Thomas Schwinge <thomas@codesourcery.com> wrote:
>
>> Previously, for '-foffload=nvptx-none -foffload=-fdump-rtl-mach
>> -save-temps -o ./nvptx-merged-loop.exe', GCC produced the expected
>> 'nvptx-merged-loop.o.307r.mach'.
>
> I believe the patch I've just installed fixes the UNRESOLVED results
> caused by not finding dump files.

Yes, confirmed, thanks!

A few small issues remain, for those I'll respond elsewhere in this
thread.


>> Consider 'libgomp.oacc-c-c++-common/pr85381-2.c':
>
>>     /* { dg-additional-options "-save-temps" } */
>
>>     /* { dg-final { scan-assembler-times "bar.sync" 2 } } */
>
>> This expects to scan the PTX offloading compilation assembler code (not
>> host code!), expecting that nvptx offloading code assembly is produced
>> after the host code, and thus overwrites the latter file.  (Yes, that's
>> certainly ugly/fragile...)
>
> I'm afraid this will need further adjusting in the testsuite, as we'll
> store the nvptx asm saved aux output in a separate file.
> scan-assembler-times will no longer work for this purpose, we'll need
> something that knows how to find the offloaded asm.

So, that's (now?) easy enough to repair.  I've pushed "[testsuite]
Replace fragile 'scan-assembler' with 'scan-offload-rtl' in
'libgomp.oacc-c-c++-common/pr85381*.c'", see attached.


Grüße
 Thomas


-----------------
Mentor Graphics (Deutschland) GmbH, Arnulfstraße 201, 80634 München / Germany
Registergericht München HRB 106955, Geschäftsführer: Thomas Heurung, Alexander Walter