On Thu, Mar 18, 2021 at 09:31:03AM -0600, Jeff Law wrote:
> > > OK. But it'd sure be nice to be able to do something like force a value
> > > of
> > > MOVE_MAX using a --param to make this kind of hack unnecessary.
> > I fear such a param would be quite dangerous, dunno what would happen if
> > somebody chose a length that can't be backed up by some integral or SIMD
> > type. Maybe for the gimple-fold.c case
> > tree type = lang_hooks.types.type_for_size (ilen * 8, 1);
> > if (type
> > && is_a <scalar_int_mode> (TYPE_MODE (type), &mode)
> > && GET_MODE_SIZE (mode) * BITS_PER_UNIT == ilen * 8
> > would fail (so we couldn't handle that way the 16 byte case anyway on all
> > targets), but there are other parts of the compiler that use MOVE_MAX.
> >
> > I think maybe better would be to instead improve the optimization so that
> > it would work even with the non-lowered memcpy calls. But that would be
> > a GCC12 thing probably.
>
> In my mind it'd only be for the testsuite and would be documented as such.
> I wouldn't want users twiddling it. We could do the same with BRANCH_COST
For BRANCH_COST we already have -mbranch-cost= option on a couple of targets
and then --param=logical-op-non-short-circuit={0,1} to override the
gimplification behavior.
Jakub
