On 11/19/26 21:34, Peter0x44 wrote:
+{
+  const char *name = setlocale (LC_NUMERIC, NULL);
+  return name != NULL && strcmp (name, saved_name) == 0;
+}
       if (!old_locale_ctr++)
     {
-      old_locale = setlocale (LC_NUMERIC, NULL);
-      setlocale (LC_NUMERIC, "C");
+      /* The next setlocale call may free this string; keep a copy.  */
+      const char *cur_locale = setlocale (LC_NUMERIC, NULL);
+      if (cur_locale != NULL && strcmp (cur_locale, "C") != 0)
+        {
+          old_locale = strdup (cur_locale);
+          setlocale (LC_NUMERIC, "C");
+        }
     }

The strdup here probably should be checked in case it fails.

There's another bug in this code (not related to this patch)

The counter’s increments and decrements don’t match:

data_transfer_init_worker increments old_locale_ctr only for FORM_FORMATTED. finalize_transfer decrements it for every transfer, including unformatted ones and transfers that failed before the increment (for example "Cannot READ after a nonadvancing WRITE"). It also returns early for child DTIO and for the namelist/unformatted error, without decrementing.

One unformatted WRITE drops the counter to -1. After that, ! old_locale_ctr++ is never true, and formatted I/O stops switching to “C” at all. Going the other way, the early returns can leave the counter stuck above zero, so the locale is never restored.

I missed this email and have accidentally pushed the changes to master branch. I have reverted it for now.

Thanks for the review.

Reply via email to