This is the mail archive of the
libc-ports@sources.redhat.com
mailing list for the libc-ports project.
Re: [PATCH] arm: Save and restore VFP arguments registers in PLT trampolines
On Wed, Apr 11, 2012 at 10:21 PM, Meador Inge <meadori@codesourcery.com> wrote:
>>> OK?
>>>
>>> Signed-off-by: Meador Inge <meadori@codesourcery.com>
>>>
>>> 2012-04-11 ?Meador Inge ?<meadori@codesourcery.com>
>>>
>>> ? ? ? ?* sysdeps/arm/dl-trampoline.S (_dl_runtime_resolve): Save and
>>> ? ? ? ?restore VFP argument registers.
>>> ? ? ? ?(_dl_runtime_profile): Likewise.
>>
>> Why do they need to be saved/restored?
>>
>> Are VFP registers caller saves?
>
> Per the AAPCS d0-d7 are used for passing floating-point parameters.
> The same way we save and restore r0-r3 we must save and restore d0-d7
> when building for __ARM_NEON__.
Good.
>> Is it because we are *about* to call the real function, but instead
>> control is transferred to the lazy resolution code which might itself
>> use VFP registers and clobber the existing values in the VFP registers
>> that were intended for the eventually resolved function?
>
> Exactly.
Good.
>> If this is the case... then how did this ever work in the past? Were we lucky?
>
> I guess so. ?It is tricky to reproduce. ?Joseph created a fairly complicated
> test case that I used to analyze the behavior.
The ld-audit interface might have exposed this easily.
The patch looks good to me, but you'll need Joseph as machine
maintainer for ARM to sign off on it.
I assume, because you didn't say, that the testsuite ran without
regression for a hardfp build?
Cheers,
Carlos.