On Fri, Oct 8, 2021 at 12:16 AM Richard Biener <rguent...@suse.de> wrote:
>
> On Tue, 5 Oct 2021, H.J. Lu wrote:
>
> > On Tue, Oct 5, 2021 at 3:07 AM Richard Biener <rguent...@suse.de> wrote:
> > >
> > > On Mon, 4 Oct 2021, H.J. Lu wrote:
> > >
> > > > commit adedd5c173388ae505470df152b9cb3947339566
> > > > Author: Jakub Jelinek <ja...@redhat.com>
> > > > Date: Tue May 3 13:37:25 2016 +0200
> > > >
> > > > re PR target/49244 (__sync or __atomic builtins will not emit 'lock
> > > > bts/btr/btc')
> > > >
> > > > optimized bit test on atomic builtin return with lock bts/btr/btc. But
> > > > it works only for unsigned integers since atomic builtins operate on the
> > > > 'uintptr_t' type. It fails on bool:
> > > >
> > > > _1 = atomic builtin;
> > > > _4 = (_Bool) _1;
> > > >
> > > > and signed integers:
> > > >
> > > > _1 = atomic builtin;
> > > > _2 = (int) _1;
> > > > _5 = _2 & (1 << N);
> > > >
> > > > Improve bit test on atomic builtin return by converting:
> > > >
> > > > _1 = atomic builtin;
> > > > _4 = (_Bool) _1;
> > > >
> > > > to
> > > >
> > > > _1 = atomic builtin;
> > > > _5 = _1 & (1 << 0);
> > > > _4 = (_Bool) _5;
> > > >
> > > > and converting:
> > > >
> > > > _1 = atomic builtin;
> > > > _2 = (int) _1;
> > > > _5 = _2 & (1 << N);
> > > >
> > > > to
> > > > _1 = atomic builtin;
> > > > _6 = _1 & (1 << N);
> > > > _5 = (int) _6;
> > >
> > > Why not do this last bit with match.pd patterns (and independent on
> > > whether _1 is defined by an atomic builtin)? For the first suggested
> >
> > The full picture is
> >
> > _1 = _atomic_fetch_or_* (ptr_6, mask, _3);
> > _2 = (int) _1;
> > _5 = _2 & mask;
> >
> > to
> >
> > _1 = _atomic_fetch_or_* (ptr_6, mask, _3);
> > _6 = _1 & mask;
> > _5 = (int) _6;
> >
> > It is useful only if 2 masks are the same.
> >
> > > transform that's likely going to be undone by folding, no?
> > >
> >
> > The bool case is
> >
> > _1 = __atomic_fetch_or_* (ptr_6, 1, _3);
> > _4 = (_Bool) _1;
> >
> > to
> >
> > _1 = __atomic_fetch_or_* (ptr_6, 1, _3);
> > _5 = _1 & 1;
> > _4 = (_Bool) _5;
> >
> > Without __atomic_fetch_or_*, the conversion isn't needed.
> > After the conversion, optimize_atomic_bit_test_and will
> > immediately optimize the code sequence to
> >
> > _6 = .ATOMIC_BIT_TEST_AND_SET (&v, 0, 0, 0);
> > _4 = (_Bool) _6;
> >
> > and there is nothing to fold after it.
>
> Hmm, I see - so how about instead teaching the code that
> produces the .ATOMIC_BIT_TEST_AND_SET the alternate forms instead
> of doing the intermediate step separately?
>
The old algorithm is
1. Check gimple forms. Return if the form isn't supported.
2. Do transformation.
My current approach treats the gimple forms accepted by the
old algorithm as canonical forms and changes the algorithm
to
1. If gimple forms aren't canonical, then
a. If gimple forms can't be transformed to canonical forms,
return;
b. Transform to canonical form.
endif
2. Check gimple forms. Return if the form isn't supported.
3. Do transformation.
The #2 check is redundant when gimple forms have been
transformed to canonical forms.
I can change my patch to
1. If gimple forms aren't canonical, then
a. If gimple forms can't be transformed to canonical forms,
return;
b. Transform to canonical form.
else
Check gimple forms. Return if the form isn't supported.
endif
2. Do transformation.
The advantage of canonical forms is that we don't have to
transform all different forms.
Does it sound OK?
Thanks.
--
H.J.
