http://gcc.gnu.org/bugzilla/show_bug.cgi?id=54713
--- Comment #5 from Richard Guenther <rguenth at gcc dot gnu.org> 2012-09-28
08:36:28 UTC ---
(In reply to comment #4)
> Created attachment 28293 [details]
> gcc48-pr54713.patch
>
> Updated patch. I had to make the CONSTRUCTOR checking less strict, in
> particular there are CONSTRUCTORS with (consecutive) indexes as well as NULL
> indexes, there are CONSTRUCTORS with less than nunits elements even when the
> elements are scalar.
Ugh, less than nunits is probably ok for GENERIC but in GIMPLE we really
want trailing zeros explicit ... :/ Thus, if we need to "gimplify"
vector constructors we can as well clear their indexes?
> But even with all this I'm still getting gfortran.dg/loc_2.f90 -O3 ICEs,
> apparently SLP creates vector CONSTRUCTOR with type integer(kind=4) vector,
> but
> elements logical(kind=4) (i.e. same TYPE_MODE, but INTEGER_TYPE vs.
> BOOLEAN_TYPE. Shall I make the verification even less strict (check TYPE_MODE
> instead of useless_type_conversion?), or shall we fix the vectorizer?
The vectorizer needs to be fixed. See the comment in
get_vectype_for_scalar_type_and_size:
/* For vector types of elements whose mode precision doesn't
match their types precision we use a element type of mode
precision. The vectorization routines will have to make sure
they support the proper result truncation/extension.
We also make sure to build vector types with INTEGER_TYPE
component type only. */
> Shouldn't I split the patch into the bugfixes + testcases and submit it before
> the verification, so that the latter can be done more slowly?
Yes, I think separate patches would be prefered.
