Re: [PING][PATCH] Fix PR 28684

[Richard Guenther <rguenther at suse dot de>]

On Mon, 6 Nov 2006, Revital1 Eres wrote:

> 
> > In some of the cases you do
> >
> > @@ -1107,7 +1107,7 @@
> >
> >      case MINUS_EXPR:
> >        /* - (A - B) -> B - A  */
> > -      if ((! FLOAT_TYPE_P (type) || flag_unsafe_math_optimizations)
> > +      if ((! FLOAT_TYPE_P (type) || flag_associative_math)
> >       && reorder_operands_p (TREE_OPERAND (t, 0), TREE_OPERAND (t,
> > 1)))
> >     return fold_build2 (MINUS_EXPR, type,
> >               TREE_OPERAND (t, 1), TREE_OPERAND (t, 0));
> >
> > or
> >
> > @@ -971,7 +971,7 @@
> >
> >      case MINUS_EXPR:
> >        /* We can't turn -(A-B) into B-A when we honor signed zeros.  */
> > -      return (! FLOAT_TYPE_P (type) || flag_unsafe_math_optimizations)
> > +      return (! FLOAT_TYPE_P (type) || flag_associative_math)
> >          && reorder_operands_p (TREE_OPERAND (t, 0),
> >                  TREE_OPERAND (t, 1));
> >
> > where the comment already suggests this should read ||
> !HONOR_SIGNED_ZEROS
> > (TYPE_MODE (type)).  There are machines without signed zeros out there
> > which will benefit from this even without flag_associative_math.
> >
> > Richard.
> >
> 
> But the transformation is re-association; Can we apply it without the flag
> (just by checking
> that we are not honoring signed zeros)?

It depends on how we violate IEEE math by doing the re-association.  If
we re-associate a + (b + c) to (a + b) + c we may change rounding 
behavior, if we re-order a + b to b + a we don't change rounding and
behavior with signed zeros doesn't change.  If we re-order (not 
reassociate!) - (a - b) to b - a, we change rounding behavior (at least
with rounding towards infinity for example) _and_ change behavior with
signed zeros.

So you are right, we need flag_associative_math here due to the rounding
issues (which is basically all flag_associative_math is about, in addition
to the signed zero stuff).

It's just all nicely twisted, so we need some very good documentation on
the flags.

flag_reciprocal_math is all about transforming a / b to a * (1/b), so
if we here do something like a * (b / c) to (a * b) / c, that is 
re-association, not reciprocal stuff.

Richard.

--
Richard Guenther <rguenther@suse.de>
Novell / SUSE Labs