Re: [PATCH] [PR68097] frange::set_nonnegative should not contain -NAN.

[Aldy Hernandez <aldyh@redhat.com>]

[-- Attachment #1: Type: text/plain, Size: 2030 bytes --]

On Mon, Sep 19, 2022 at 10:14 AM Richard Biener
<richard.guenther@gmail.com> wrote:
>
> On Mon, Sep 19, 2022 at 9:59 AM Aldy Hernandez <aldyh@redhat.com> wrote:
> >
> > ISTM that a specifically nonnegative range should not contain -NAN,
> > otherwise signbit_p() would return false, because we'd be unsure of the
> > sign.
> >
> > Do y'all agree?
>
> what tree_expr_nonnegative_p actually means isn't 100% clear.  For REAL_CST
> it actually looks at the sign-bit but we have
>
> (simplify
>  /* copysign(x,y) -> fabs(x) if y is nonnegative.  */
>  (COPYSIGN_ALL @0 tree_expr_nonnegative_p@1)
>  (abs @0))
>
> is abs (@0) OK for sNaNs and -NaN/+NaN?

At least for real_value's, ABS_EXPR works on NAN's.  There's no
special code dealing with them.  We just clear the sign bit:

real_arithmetic:
    case ABS_EXPR:
      *r = *op0;
      r->sign = 0;
      break;

>
> And we have
>
> /* Convert abs[u] (X)  where X is nonnegative -> (X).  */
> (simplify
>  (abs tree_expr_nonnegative_p@0)
>  @0)
>
> where at least sNaN -> qNaN would be dropped?
>
> And of course
>
> (simplify
>  /* signbit(x) -> 0 if x is nonnegative.  */
>  (SIGNBIT tree_expr_nonnegative_p@0)
>  { integer_zero_node; })
>
> that is, is tree_expr_nonnegative_p actually tree_expr_sign or
> does tree_expr_nonnegative (x) mean x >= (typeof(X)) 0
> or !(x < (typeof(X))0)?

I have no idea, but I'm happy to have frange::set_nonnegative() do
whatever you agree on.

Actually TBH, ranger only uses set_nonnegative for call's, not much else:

  if (range_of_builtin_call (r, call, src))
    ;
  else if (gimple_stmt_nonnegative_warnv_p (call, &strict_overflow_p))
    r.set_nonnegative (type);
  else if (gimple_call_nonnull_result_p (call)
       || gimple_call_nonnull_arg (call))
    r.set_nonzero (type);
  else
    r.set_varying (type);

but I guess it's good we do the right thing for correctness sake, and
if it ever gets used by someone else.

>
> That said, 'set_nonnegative' could be interpreted to be without
> NaNs?

Sounds good to me.  How's this?

Aldy

[-- Attachment #2: 0001-PR68097-Non-negative-frange-s-should-not-contain-NAN.patch --]
[-- Type: text/x-patch, Size: 1058 bytes --]

From 7eeb81a9516663d589e692b278907bb5f3d60ea0 Mon Sep 17 00:00:00 2001
From: Aldy Hernandez <aldyh@redhat.com>
Date: Sun, 18 Sep 2022 16:24:36 +0200
Subject: [PATCH] [PR68097] Non-negative frange's should not contain NANs.

	PR 68097/tree-optimization

gcc/ChangeLog:

	* value-range.cc (frange::set_nonnegative): Clear NAN.
	(range_tests_signed_zeros): New test.
---
 gcc/value-range.cc | 5 +++++
 1 file changed, 5 insertions(+)

diff --git a/gcc/value-range.cc b/gcc/value-range.cc
index 67d5d7fa90f..6e703c7eb43 100644
--- a/gcc/value-range.cc
+++ b/gcc/value-range.cc
@@ -752,6 +752,7 @@ void
 frange::set_nonnegative (tree type)
 {
   set (type, dconst0, dconstinf);
+  clear_nan ();
 }
 
 // Here we copy between any two irange's.  The ranges can be legacy or
@@ -3800,6 +3801,10 @@ range_tests_signed_zeros ()
   r1.update_nan ();
   r0.intersect (r1);
   ASSERT_TRUE (r0.known_isnan ());
+
+  // Non-negative ranges should not contain NANs.
+  r0.set_nonnegative (float_type_node);
+  ASSERT_TRUE (!r0.maybe_isnan ());
 }
 
 static void
-- 
2.37.1