On Friday, 16 July 2021 11:31:29 CEST Richard Biener wrote:
> On Fri, Jul 16, 2021 at 10:57 AM Matthias Kretz <[email protected]> wrote:
> > On Wednesday, 14 July 2021 10:14:55 CEST Richard Biener wrote:
> > > I think implementing it similar to how we do __builtin_shufflevector
> > > would
> > > be easily possible. PAREN_EXPR is a tree code.
> >
> > Like this? If you like it, I'll write the missing documentation and do
> > real
> > regression testing.
>
> Yes, like this. Now, __builtin_noassoc (a + b + c) might suggest that
> it prevents a + b + c from being re-associated - but it does not.
> PAREN_EXPR is a barrier for association, so for 'a + b + c + PAREN_EXPR <d
> + e + f>' the a+b+c and d+e+f chains will not mix but they individually can
> be re-associated. That said __builtin_noassoc might be a bad name,
> maybe __builtin_assoc_barrier is better?
Yes, I agree with renaming it. And assoc_barrier sounds intuitive to me.
> To fully prevent association of a a + b + d + e chain you need at least
> two PAREN_EXPRs, for example (a+b) + (d+e) would do.
>
> One could of course provide __builtin_noassoc (a+b+c+d) with the
> implied semantics and insert PAREN_EXPRs around all operands
> when lowering it.
I wouldn't want to go there. __builtin_noassoc(f(x, y, z))? We probably both
agree that it would be a no-op, but it reads like f should be evaluated with -
fno-associative-math.
> Not sure what's more useful in practice - directly exposing the middle-end
> PAREN_EXPR or providing a way to mark a whole expression as to be
> not re-associated? Maybe both?
I think this is a tool for specialists. Give them the low-level tool and
they'll build whatever higher level abstractions they need on top of it. Like
float sum_noassoc(RangeOfFloats auto x) {
float sum = 0;
for (float v : x)
sum = __builtin_assoc_barrier(v + x);
return sum;
}
--
──────────────────────────────────────────────────────────────────────────
Dr. Matthias Kretz https://mattkretz.github.io
GSI Helmholtz Centre for Heavy Ion Research https://gsi.de
std::experimental::simd https://github.com/VcDevel/std-simd
──────────────────────────────────────────────────────────────────────────
