Re: [PATCH v4 17/21] LoongArch Linux User Emulation

[Richard Henderson]

On 9/8/21 11:50 AM, Song Gao wrote:
> Hi Richard,
>
> On 09/05/2021 06:04 PM, Richard Henderson wrote:
> > > +struct sigframe {
> > > +    uint32_t sf_ass[4];             /* argument save space for o32 */
> >
> > Surely there is no "o32" for loongarch?
>
> Yes, qemu only support 64bit. but the kernel has 'o32'.  Should we have to be 
> consistent with the kernel?

Yes, you need to be consistent with the kernel.  However... the kernel is not yet 
upstream, and therefore the ABI is (officially) still malleable.

Anyway, this padding was copied from mips o32, and should not have been.  Looking at the 
loongarch gcc sources, REG_PARM_STACK_SPACE is always 0, and thus the 4 words reserved 
here are never used.

I see that gcc/libgcc/config/loongarch/linux-unwind.h builds in knowledge of these unused 
words.  I also see that linux/arch/loongarch/vdso/sigreturn.S does not provide correct 
unwind info.  Certainly the kernel vdso should be fixed, so that code within gcc should 
not be needed.  At which point the ABI for the signal frame is entirely private to the kernel.


r~