On Thu, May 16, 2019 at 12:03:14PM +0100, Iain Sandoe wrote:
> I did a quick check...
>
> dfp.exp most (all?) fail despite
>
> /* { dg-require-effective-target powerpc_p9vector_ok } */
>
> with errors like this…
>
> error: decimal floating point not supported for this target
Okay, so the test should probably work if the test would actually try to
use DFP, not just do a machine insn in asm (maybe just using a DFP constant
or variable as input to the asm is enough).
> a number (large enough) of the bfp.exp tests fail despite
>
> /* { dg-require-effective-target powerpc_p9vector_ok } */
>
> with things like ...
> .../gcc.target/powerpc/bfp/scalar-extract-exp-5.c:13:15: error: unknown type
> name '__ieee128'; did you mean '__int128’?
>
> It could be that these are missing an require-effective-target-float128.
Yeah, something like that. It looks like we have no selector for exactly
__ieee128 yet.
> Other differences are rather spread around the testsuite, so I’ve not
> re-checked.
>
> note that circa 1000 tests are attempted with the new assembler that were
> unsupported with cctools.
> at least half of those fail.
Oh wow.
> > (I am motivated to have
> > parity between the cctools and newer assemblers in coverage on
> > Darwin for now, and then to try expanding the horizons when the basics
> > are working well).
>
>
> It’s helpful to me right now that tests that are UNSUPPORTED with the cctools
> assembler are not attempted with the newer one.
> That makes a<->b comparisons easier, and helps highlight the cases where
> tests fail because the new assembler has better error checking rather than
> spurious attempts to do things that Darwin can't.
If you can make the selector fail on Darwin (rather than specific testcases),
this will all be fine, not too much churn.
> In the longer term, when the testsuite noise is manageable - we can try
> backing these things out one at a time and see what new fails we get - and
> either fix the cases individually (or put the blanket provision back, of
> course).
Sounds perfect. Thanks!
Segher
