Re: [PATCH] gcc: Keep TImode libcalls on the direct-return convention [PR78799]

ср, 12 авг. 2026 г. в 09:44, LIU Hao <[email protected]>:
>
> 在 2026-8-9 22:43, Oleg Tolmatcev 写道:
> > Libgcc helper libcalls such as __udivti3 still use the historical
> > TImode direct-return convention. Preserve that convention when
> > TARGET_RETURN_IN_MEMORY is queried for a libcall so the caller's
> > expectation continues to match libgcc on x86_64-w64-mingw32.
> >
> > gcc/ChangeLog:
> >
> >       PR target/78799
> >       * config/i386/i386.cc (ix86_return_in_memory): Keep TImode
> >       libcalls on the direct-return convention.
> >       * testsuite/gcc.target/i386/pr78799-2.c: New test.
> >
> > Signed-off-by: Oleg Tolmatcev <[email protected]>
> > ---
> >   gcc/config/i386/i386.cc                   |  8 ++++++-
> >   gcc/testsuite/gcc.target/i386/pr78799-2.c | 28 +++++++++++++++++++++++
> >   2 files changed, 35 insertions(+), 1 deletion(-)
> >   create mode 100644 gcc/testsuite/gcc.target/i386/pr78799-2.c
>
> I have two questions about this change:
>
> First is that `__udivti3` seems defined in libgcc2.c and returns `__int128` in memory, so this change
> would result in a mismatch of how the value is returned, wouldn't it?
>
> Second is that the commit message says 'the caller's expectation continues to match libgcc on
> x86_64-w64-mingw32', but for GCC 16- (before e63a873f1df7493b8b657acafb451bf40166044e) the value is
> returned in XMM0. Does this change expect the result in XMM0 or RDX:RAX? If the result should go into
> RDX:RAX then it's still an ABI break with GCC 16 and therefore makes less sense.

Thanks for the review. Good catch. It is indeed unnecessary and I am
withdrawing it.My tests were crashing
because of a mismatch between GCC 17 and libgcc from GCC 16 and my
analysis was wrong.

Thanks again for the review.

Oleg