On Fri, Aug 16, 2024 at 01:35:01PM -0500, Nathan Bossart wrote:
> On Fri, Aug 16, 2024 at 09:00:00PM +0300, Alexander Lakhin wrote:
>> #6 0x00005576cf627c68 in bms_singleton_member (a=0x5576d09f7fb0) at bitmapset.c:691
>> 691 if (result >= 0 || HAS_MULTIPLE_ONES(w))
>
> At a glance, this appears to be caused by the RIGHTMOST_ONE macro:
>
> #define RIGHTMOST_ONE(x) ((signedbitmapword) (x) & -((signedbitmapword) (x)))
I believe hand-rolling the two's complement calculation should be
sufficient to avoid depending on -fwrapv here. godbolt.org indicates that
it produces roughly the same code, too.
diff --git a/src/backend/nodes/bitmapset.c b/src/backend/nodes/bitmapset.c
index cd05c642b0..d37a997c0e 100644
--- a/src/backend/nodes/bitmapset.c
+++ b/src/backend/nodes/bitmapset.c
@@ -67,7 +67,7 @@
* we get zero.
*----------
*/
-#define RIGHTMOST_ONE(x) ((signedbitmapword) (x) & -((signedbitmapword) (x)))
+#define RIGHTMOST_ONE(x) ((bitmapword) (x) & (~((bitmapword) (x)) + 1))
#define HAS_MULTIPLE_ONES(x) ((bitmapword) RIGHTMOST_ONE(x) != (x))
--
nathan
