This was sent to the gcc-patches list, but not to gfortran - so I'll
forward it ...
Kind regards,
Toon.
-------- Forwarded Message --------
Subject: Re: [PATCH v1] libgfortran: Copy saved LC_NUMERIC name before
switching to "C" [PR128183]
Date: Thu, 19 Nov 2026 14:47:28 +0000
From: Jonathan Yong <[email protected]>
To: [email protected]
On 10/26/26 17:08, jmarrec via Sourceware Forge wrote:
From: Julien Marrec <[email protected]>
On targets without POSIX 2008 locales (e.g. MinGW-w64), formatted I/O saves the pointer returned by setlocale
(LC_NUMERIC, NULL) and restores through it after setlocale (LC_NUMERIC, "C"). C11 allows that second call to
overwrite the string, and both Windows C runtimes invalidate it. msvcrt overwrites its static buffer with
"C", so a non-"C" LC_NUMERIC set by the program is reset to "C" by the first formatted
I/O statement. UCRT frees the buffer, so the restore reads freed memory and leaves the process in an arbitrary locale
or fails, and formatted I/O becomes several times slower.
Save a copy of the name instead, and skip the switch entirely when LC_NUMERIC is already
"C", which avoids two setlocale calls per data transfer statement in the common
case.
PR libfortran/128183
libgfortran/ChangeLog:
* io/transfer.c (data_transfer_init_worker): Save a copy of the
LC_NUMERIC name, and only switch to "C" if it is not already "C".
(finalize_transfer): Only restore if a name was saved, and free it.
* io/unit.c (old_locale): Update comment.
gcc/testsuite/ChangeLog:
* gfortran.dg/pr128183.f90: New test.
* gfortran.dg/pr128183_aux.c: New file.
Signed-off-by: Julien Marrec <[email protected]>
---
Testing:
- x86_64-pc-linux-gnu, non-bootstrap C/Fortran build, with libgfortran
configured with `ac_cv_func_uselocale=no` so that it uses the setlocale
fallback. `make check-fortran` with and without the fix: no new failures.
- The new test fails without the fix on that build (glibc frees the old name
too; valgrind reports an invalid read in finalize_transfer) and passes with it.
- x86_64-w64-mingw32: transfer.c and unit.c rebuilt with the fix and swapped
into the libgfortran.a of a GCC 16.2.0 UCRT toolchain. The reproducer passes
(it fails with the stock library), and an internal WRITE goes from 6.8 us to
1.8 us. I have not run the testsuite on MinGW-w64.
Bug and reproducers: https://gcc.gnu.org/PR128183,
https://github.com/jmarrec/libgfortran-ucrt-setlocale
I don't have write access, so please push it for me if approved. Ok for trunk?
I don't mind helping to push if at least one Fortran maintainer okays it.