On Tue, 28 Mar 2023, Jakub Jelinek wrote:
> On Tue, Mar 28, 2023 at 08:57:12AM +0000, Richard Biener wrote:
> > Hmm, but canonicalize_math_p () should be false after vectorization?
> >
> > When we moved the pass we should have made sure to put the
> > PROP_gimple_opt_math property set to pass_expand_powcabs instead.
>
> Which pass is the one that actually canonicalizes the math such
> that we want to keep its choices for later?
> I must say I don't know the details why the sincos path has been
> even moved.
The pass was split into sincos + pass_expand_powcabs - canonicalization
happens through folding and pass_expand_powcabs expands
pow (x, 2) to x * x. Folding would make x * x to pow (x, 2)
so it's important to set the property after pass_expand_powcabs.
I think we moved sincos because vectorizing cexpi never materialized(?)
but maybe I misremember.
> > Now, the sqrt (sqrt ()) canonicalization to pow (.., 1./4) is
> > probably invalid anyway, not sure if we can add a user-written
> > vector sqrt testcase that would trigger during the canonicalization
>
> How do we write user written vector sqrt?
I'm not sure ;)
> > phase. There are other uses of build_real that have the same
> > issue - what's your conclusion this is never a problem there?
>
> Looking through build_real* calls in match.pd, others are
> either on simplifications of some expressions with REAL_CST operand
> (so can't be vector then), or using
> LOG*/EXP*/CBRT*/SIN*/ATAN*/COS*/POW*/CABS*/HYPOT*/POWI*/SIGNBIT*
> calls in the expression being simplified, or this case.
> I think no target provides vector optabs for those or they don't
> have internal fns at all.
> Maybe
> (simplify
> /* signbit(x) -> x<0 if x doesn't have signed zeros. */
> (SIGNBIT @0)
> (if (!HONOR_SIGNED_ZEROS (@0))
> (convert (lt @0 { build_real (TREE_TYPE (@0), dconst0); }))))
> ?
Maybe.
But as said, the fix is probably to move the pass property.
Richard.
