On 10/1/26 06:43, Uros Bizjak wrote:
On Thu, Oct 1, 2026 at 12:46 AM Oleg Tolmatcev <[email protected]> wrote:
For an MS ABI va_list, ix86_gimplify_va_arg calls
std_gimplify_va_arg_expr, which asks the pass_by_reference hook whether
the argument was passed by reference. Without cumulative args, the hook
assumes the default ABI, SysV on Linux, so va_arg reads an argument
passed by reference as if it were passed by value.
Split the ABI-specific decision out of ix86_pass_by_reference. For an
indirect MS ABI argument, follow ia64_gimplify_va_arg: use the standard
va_arg helper to read its pointer and then dereference it. Use the
explicit SysV ABI for a SysV va_list so the opposite target-default
mismatch is handled as well.
gcc/ChangeLog:
* config/i386/i386.cc (ix86_pass_by_reference_abi): New function,
split out of ...
(ix86_pass_by_reference): ... this.
(ix86_gimplify_va_arg): Derive argument indirection from the ABI of
the va_list type. Read indirect MS ABI arguments as pointers.
gcc/testsuite/ChangeLog:
* gcc.target/i386/va-arg-abi-attribute.c: New test.
OK.
For future contributions, please state how you tested the patch, see
[1]. Also, if you don't have write access, please point out lack of
write access in your initial submission, as also advised in [1].
[1] https://gcc.gnu.org/contribute.html#patches
Thanks,
Uros.
Can these be pushed or are there any further review? I also don't see
the v2 second patch anywhere.
