On 16/11/26 17:53, Peter Maydell wrote:
On Mon, 16 Nov 2026 at 16:31, Pierrick Bouvier
<[email protected]> wrote:

On 11/16/2026 2:00 AM, Peter Maydell wrote:
On Mon, 16 Nov 2026 at 08:53, Philippe Mathieu-Daudé
<[email protected]> wrote:

On 13/11/26 19:55, Peter Maydell wrote:
As a more general question, is this something we would want
for architectures where both endiannesses are relevant (e.g.
mips, ppc64)? At the moment we consistently for all targets
handle big vs little endian as "it's a separate target with
its own executable". That's convenient because you can at
compile time hardwire in things like variations in struct
layout between the two. We could switch to doing that kind
of thing at runtime, but what is the benefit we're aiming
for by doing that?

I did the other conversions (Microblaze, Xtensa, PPC, MIPS, SH4)
and wanted to test the water first where this would get more attention,
with the series involving less changes.

My motivation is MIPS where we currently have 6 binaries and I'd like
to add binaries to cover more ABIs without having a combinatory
explosion and have 18 linux-user MIPS binaries. With this approach I
reduced the current MIPS set to 3 binaries and add 2 more, ending with
5 covering more ABIs instead of currently 6.

In order to get there I need the few core changes introduced in this
series, which affect all user binaries. I have the idea ARM users are
the first ones interested by no performance change thus is where
I started the benchmark to prove this approach have no impact in emulation.

The TargetInfo structure change (one initializer LTO-protected and
all accessors via read-only exposed structure) is also what Richard
was very keen to have (page size support planned) instead of the
current approach.

I'm thus returning the question as: why would we not want to do this?

Mostly because what we have today works, and the idea of "the
target ABI is known at compile-time" is quite deeply embedded in the
way the linux-user code is written (both in terms of ifdefs and
also C types being defined appropriately for one specific ABI).

While we could go one step further and manage distincts ABIs at
runtime, I don't see the point so far, and that might have some
performance impact (although I'm not that sure). While endianness
is part of the kernel ABI, the changes are less invasive because
we already did them on system path.

The last part is what we had for system mode, and precisely what we have
been able to eliminate while working on the single-binary.

Yes, but it's taken a lot of work to get to that point. For
system mode we had some clear reasons why that work was
necessary and worthwhile. For user-mode it's less clear to me.

On that point, I already did the work -- and test / benchmark --
so it isn't an issue ;) The expected review part isn't huge neither.

Keeping the zoo of MIPS ABIs under control seems the most
promising reason.

For aarch64 in particular I would prefer to deprecate-and-drop
big-endian support (probably starting with the deprecation for
12.0, rather than pushing it into 11.2), since the kernel just
did that in 7.4.
Tangent to this, it makes sense to deprecate quickly enough on
QEMU side when a host-related project we use deprecates / removes
it; when this is guest related I don't see the urge, in particular
when the maintenance is almost zero cost. For the particular
aarch64_be I agree what we lacks is coverage testing; but for
other features I'd prefer we keep them as long as there is interest
within QEMU users, independent of external projects change like
the Linux kernel.

Reply via email to