RKSimon added a comment.
Please can you create a llvm side parallel patch that updates the relevant
fast-isel tests
================
Comment at: clang/lib/Headers/avx512vlbwintrin.h:1501
+ (__v8qi){0, 0, 0, 0, 0, 0, 0, 0}, 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11,
+ 12, 13, 14, 15);
}
----------------
mike.dvoretsky wrote:
> RKSimon wrote:
> > Are we happy with using illegal types like this? What about flipping the
> > shuffle and convert?
> >
> > ```
> > return (__m128i)__builtin_convertvector(
> > __builtin_shufflevector((__v8hi)__A,
> > (__v8hi){0, 0, 0, 0, 0, 0, 0,
> > 0},
> > 0, 1, 2, 3, 4, 5, 6, 7, 8, 9,
> > 10, 11, 12, 13, 14, 15), __v16qi);
> > ```
> This would bring its own issues, since in the cvtepi64_epi8 cases the inner
> shuffle would produce vectors of 16 64-bit values. There would be no extra
> typedef, but in the back-end these would be split in type legalization,
> making it harder to fold them into VPMOV instructions.
Yeah, neither solution is particularly clean. Please keep it as is.
https://reviews.llvm.org/D48712
_______________________________________________
cfe-commits mailing list
cfe-commits@lists.llvm.org
http://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
