Re: [PATCH] range-op-float: Fix up -ffinite-math-only range extension and don't extend into infinities [PR109008]

[Aldy Hernandez <aldyh@redhat.com>]

On 3/10/23 09:53, Richard Biener wrote:
> On Fri, 10 Mar 2023, Jakub Jelinek wrote:

Coming back to this...

>   /* Temporarily disable -ffinite-math-only, so that frange::set doesn't
>      reduce the range back to real_min_representable (type) as lower bound
>      or real_max_representable (type) as upper bound.  */
>   bool save_flag_finite_math_only = flag_finite_math_only;
>   flag_finite_math_only = false;
>   ret.set (type, lb, ub);
>   if (lhs.kind () != VR_VARYING)
>     {
>       ret.clear_nan ();
>       ret.union_ (lhs);
>     }
>   flag_finite_math_only = save_flag_finite_math_only;

It looks like what you want to do is be able to create a range with a 
known NAN state, but without the setter reducing the range to 
min/max_representable.

How about we enhance the API to provide:

1. Constructor with a known NAN state.
2. Setter with a flag to keep it from canonicalizing into 
min/max_representable.

The flag in 2 could in the future be saved in the frange object to keep 
union and friends from further canonicalization.

So the above could be written as:

	// Construct [lb, ub] with a known NAN state.
	frange tmp (lb, ub, lhs.get_nan_state ());

	// Set RET without dropping/reducing the range to MIN/MAX.
	ret.set (tmp, FRANGE_FLAG_NO_REPRESENTABLE_REDUCTION);

An alternative is to allow the setter to set everything:

	ret.set (type, lb, ub,
		lhs.get_nan_state (),
		FRANGE_FLAG_NO_REPRESENTABLE_REDUCTION);

Would this work?  I'd be happy to whip up something this week, or if 
preferred, leave it to the next release.

Aldy