Re: [Qemu-devel] [PATCH 0/3] target-arm: support setend instruction

[Paolo Bonzini]

Il 21/04/2014 18:08, Peter Maydell ha scritto:
> Thinking more about this, we'd be wrong already for bigendian
> if these were wrong. The bit about the linux-user/main.c strex/ldrex
> still stands, though, because those use get_user*/put_user* which
> use the TARGET_BIGENDIAN specified endianness.

Ok.

> The commpage
> code may also need updating to correctly handle the effects of being
> called with the wrong endianness (though we can probably get away
> with declaring that to be lunacy and not well defined...)

Yeah, I think this is lunacy. :)

> If we take an exception while the E bit is set we'll now incorrectly
> start executing the exception handler in big-endian mode.
> Minimally, we need to force "always little endian". Ideally, we
> would implement SCTLR.EE correctly.

I'll look into SCTLR.

> Also worth checking: in linux-user mode, if the guest executes
> SETEND and then takes a signal, what does the Linux ABI say
> should be the value of CPSR.E on entry to the signal handler,
> and do we get it right?

No, my patch does not even try.  Thanks for pointing out.  The kernel 
sets CPSR.E for armeb and clears it otherwise.

Paolo

> [Tangentially: we really need a vdso for aarch64, at which point we
> might as well do the aarch32 commpage with one too. IIRC RTH had
> some patches for x86-64 vdso which I should fish out of the archives...]
>
> thanks
> -- PMM