Re: CVS commit: src/sys/lib/libunwind

[Rin Okuyama <rokuyama.rk%gmail.com@localhost>]

Correction:

On 2021/06/01 7:10, Rin Okuyama wrote:
> On 2021/05/31 23:35, Joerg Sonnenberger wrote:
> > On Mon, May 31, 2021 at 12:12:24PM +0000, Rin Okuyama wrote:
> > > Bump LAST_REGISTER and LAST_RESTORE_REG to REGNO_ARM32_S31 for arm.
> >
> > This is not desirable as it significantly increases the memory use.
> > The goal here is to actually alias the single and double register in the
> > space. That's what the existing was doing.
>
> Existing code did not.
>
> With LAST_REGISTER = REGNO_ARM32_D31, libunwind gave up unwinding frames
> when it encounters s0-s31; see parseInstructions():
>
> https://nxr.netbsd.org/xref/src/sys/lib/libunwind/DwarfParser.hpp#parseInstructions

This paragraph:

> And with LAST_RESTORE_REG = REGNO_ARM32_D31, it did not copy-back contents
> of s0-s31 at all; see stepWithDwarf():
>
> https://nxr.netbsd.org/xref/src/sys/lib/libunwind/DwarfInstructions.hpp#136

is only applied to the case of LAST_REGISTER = REGNO_ARM32_S31, and
not to existing code.

However, LAST_RESTORE_REG itself has nothing to do with memory usage.
And moreover, the conclusion below is not changed at all.

Thanks,
rin

> IMO, in order to support VFP-register aliasing, we need to heavily modify
> MI codes. This adds complexities to the code, and only ARM would receive
> benefits from them.
>
> Existing code was not acceptable for me; for example, GDB aborted *every
> time* exceptions were raised. I won't revert this change at the moment.
> Please provide working alternative if you don't like mine.
>
> Thanks,
> rin
>
> > Joerg