On Thu, Jun 27, 2019 at 1:31 PM Jan Beulich <jbeul...@suse.com> wrote:
>
> >>> On 27.06.19 at 13:02, <ubiz...@gmail.com> wrote:
> > On Thu, Jun 27, 2019 at 12:47 PM Jan Beulich <jbeul...@suse.com> wrote:
> >>
> >> >>> On 27.06.19 at 12:22, <ubiz...@gmail.com> wrote:
> >> > On Thu, Jun 27, 2019 at 11:10 AM Jan Beulich <jbeul...@suse.com> wrote:
> >> >>
> >> >> >>> On 27.06.19 at 11:03, wrote:
> >> >> > With just an "m" constraint misaligned memory operands won't be forced
> >> >> > into a register, and hence cause #GP. So far this was guaranteed only
> >> >> > in the case that CVT{,T}PD2DQ were chosen (which looks to be the case
> >> >> > on
> >> >> > x86-64 only).
> >> >> >
> >> >> > Instead of switching the second alternative to Bm, use just m on the
> >> >> > first and replace nonimmediate_operand by vector_operand.
> >> >>
> >> >> While doing this and the others where I'm also replacing Bm by uses of
> >> >> vector_operand, I've started wondering whether Bm couldn't (and then
> >> >> shouldn't) be dropped altogether, replacing it everywhere by "m"
> >> >> combined with vector_operand (or vector_memory_operand when
> >> >> register operands aren't allowed anyway).
> >> >
> >> > No. Register allocator will propagate unaligned memory in non-AVX
> >> > case, which is not allowed with vector_operand.
> >>
> >> I'm afraid I don't understand: Unaligned SIMD memory accesses will
> >> generally fault in non-AVX mode, so such propagation would seem
> >> wrong to me and hence would seem to be correctly not allowed.
> >> Furthermore both vector_operand and Bm resolve to the same
> >> vector_memory_operand. The TARGET_AVX check actually is inside
> >> vector_memory_operand, i.e. affects both the same way.
> >
> > "Bm" *prevents* propagation of unaligned access for non-AVX targets.
> > As said, register allocator does not care for operand predicates (it
> > only looks at operand constraints), so it will propagate unaligned
> > access with "m" operand. To avoid propagation, "Bm" should and does
> > use vector_memory_operand constraint internally.
>
> Okay, I think I got it now (also because of your reply on the other
> thread). It means in the patch here I need to retain Bm rather than
> dropping it, too, and additionally use it on the other alternative.
The correct solution is a bit more complicated. I don't know if these
instructions tolerate unaligned operand in non-AVX case. If they
don't, then vector_operand should be used and the first alternative
should be split to avx and non-avx part, where non-avx part uses Bm
constraint.
Uros.
