https://gcc.gnu.org/bugzilla/show_bug.cgi?id=80536
--- Comment #6 from Marek Polacek <mpolacek at gcc dot gnu.org> ---
Yeah, it helps with this particular testcase (and I agree we want to get rid of
that fold() call in save_expr -- I'll take care of it), but I wonder if this
issue is something separate: starting with r230506 we can generate expressions
such as
i * ((unsigned long) -0 + 1) * 2
and given how fold() works we don't get to simplify the sub-expression
"(unsigned long) -0 + 1)" so this expression isn't folded to "i * 2". So I
wonder if we want something like
--- a/gcc/convert.c
+++ b/gcc/convert.c
@@ -524,7 +524,13 @@ convert_to_integer_1 (tree type, tree expr, bool dofold)
return expr;
return build2_loc (EXPR_LOCATION (expr), COMPOUND_EXPR, TREE_TYPE (t),
TREE_OPERAND (expr, 0), t);
- }
+ }
+
+ /* -0 is 0, so get rid of the NEGATE_EXPR. */
+ if (0 && ex_form == NEGATE_EXPR
+ && TREE_CODE (TREE_OPERAND (expr, 0)) == INTEGER_CST
+ && integer_zerop (TREE_OPERAND (expr, 0)))
+ return convert_to_integer_maybe_fold (type, TREE_OPERAND (expr, 0),
dofold);
/* Convert e.g. (long)round(d) -> lround(d). */
/* If we're converting to char, we may encounter differing behavior
