I've been trying to speed-up a 'diff -r dir1 dir2'
by disabling the "no-throw" stuff for MSVC/clang-cl:
MSVC_INVALID_PARAMETER_HANDLING == SANE_LIBRARY_HANDLING
(suspecting it imposes some overhead. Maybe MSVC's
__try/__catch has lower overhead?)
But building with 'SANE_LIBRARY_HANDLING' crashes in
several places presumably because (in fclose.c):
#if WINDOWS_SOCKETS
/* Call the overridden close(), then the original fclose().
Note about multithread-safety: There is a race condition where some
other thread could open fd between our close and fclose. */
if (close (fd) < 0 && saved_errno == 0)
saved_errno = errno;
fclose_nothrow (fp); /* will fail with errno = EBADF,
if we did not lose a race */
---------
the 'fp' associated with the 'fd' got closed.
I did this to fix it:
--- a/fclose.c 2021-06-10 12:59:56
+++ b/fclose.c 2021-10-12 13:43:18
@@ -79,9 +79,11 @@
/* Call the overridden close(), then the original fclose().
Note about multithread-safety: There is a race condition where some
other thread could open fd between our close and fclose. */
- if (close (fd) < 0 && saved_errno == 0)
+ result = close (fd);
+ if (result < 0 && saved_errno == 0)
saved_errno = errno;
+ /* If the above 'close()' succeeded, don't close the associated 'fp' again.
*/
+ if (result)
fclose_nothrow (fp); /* will fail with errno = EBADF,
if we did not lose a race */
----------------
Does this make sense? Or are we discouraged to
disable this TRY/CATCH stuff?
--
--gv
