https://gcc.gnu.org/bugzilla/show_bug.cgi?id=91384
--- Comment #9 from Richard Biener <rguenth at gcc dot gnu.org> ---
The local gimple transform is not right or wrong, it depends on the
surroundings
and unless we have a way to assess those in a good way doing it in one or the
other way is going to hurt either case.
The usual "fix" is to stick single-use restrictions onto the patterns involving
uses in compares, we're not very consistent with that.
Like the following
diff --git a/gcc/match.pd b/gcc/match.pd
index 43bacb4f68e..c5582021b0e 100644
--- a/gcc/match.pd
+++ b/gcc/match.pd
@@ -4503,10 +4503,11 @@ DEFINE_INT_AND_FLOAT_ROUND_FN (RINT)
&& TYPE_OVERFLOW_UNDEFINED (TREE_TYPE (@0))))
(scmp @0 @1)))
(simplify
- (cmp (negate @0) CONSTANT_CLASS_P@1)
+ (cmp (negate@2 @0) CONSTANT_CLASS_P@1)
(if (FLOAT_TYPE_P (TREE_TYPE (@0))
|| (ANY_INTEGRAL_TYPE_P (TREE_TYPE (@0))
- && TYPE_OVERFLOW_UNDEFINED (TREE_TYPE (@0))))
+ && TYPE_OVERFLOW_UNDEFINED (TREE_TYPE (@0))
+ && single_use (@2)))
(with { tree tem = const_unop (NEGATE_EXPR, TREE_TYPE (@0), @1); }
(if (tem && !TREE_OVERFLOW (tem))
(scmp @0 { tem; }))))))
