[Bug target/96061] New: Please support 'no_caller_saved_registers' attribute on aarch64 (or maybe on all targets)

[Bug target/96061] New: Please support 'no_caller_saved_registers' attribute on aarch64 (or maybe on all targets)

[alex.popov at linux dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=96061

            Bug ID: 96061
           Summary: Please support 'no_caller_saved_registers' attribute
                    on aarch64 (or maybe on all targets)
           Product: gcc
           Version: unknown
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: target
          Assignee: unassigned at gcc dot gnu.org
          Reporter: alex.popov at linux dot com
  Target Milestone: ---

This is an enhancement request.

GCC currently supports 'no_caller_saved_registers' attribute only for x86.
Clang supports it for all target architectures.

It would be very useful for Linux kernel, if GCC supported
'no_caller_saved_registers' attribute for at least aarch64. In particular, that
would allow to avoid useless work with caller-saved registers caused by
STACKLEAK GCC plugin instrumentation for the Linux kernel on aarch64.

This kernel patch uses 'no_caller_saved_registers' to fix the described issue
on x86:
https://www.openwall.com/lists/kernel-hardening/2020/06/24/13
(the patch is accepted for the mainline kernel)

That GCC enhancement can be useful for _mcount (gprof) as well.

Thanks!

[Bug target/96061] Please support 'no_caller_saved_registers' attribute on aarch64 (or maybe on all targets)

[hjl.tools at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=96061

H.J. Lu <hjl.tools at gmail dot com> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
   Last reconfirmed|                            |2020-07-04
             Status|UNCONFIRMED                 |NEW
     Ever confirmed|0                           |1

--- Comment #1 from H.J. Lu <hjl.tools at gmail dot com> ---
(In reply to Alexander Popov from comment #0)
> This is an enhancement request.
> 
> GCC currently supports 'no_caller_saved_registers' attribute only for x86.
> Clang supports it for all target architectures.

We are working on -fzero-call-used-regs:

https://gitlab.com/x86-gcc/gcc/-/tree/users/qinzhao/caller-saved/master


They are related.

[Bug target/96061] Please support 'no_caller_saved_registers' attribute on aarch64 (or maybe on all targets)

[alex.popov at linux dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=96061

--- Comment #2 from Alexander Popov <alex.popov at linux dot com> ---
> We are working on -fzero-call-used-regs:

H.J. Lu, thanks for the information!

However, this flag can't be used per-function, like the attribute.
So supporting 'no_caller_saved_registers' attribute for aarch64 (and maybe
other architectures) is desired anyway.

Thanks!

[Bug target/96061] Please support 'no_caller_saved_registers' attribute on aarch64 (or maybe on all targets)

[hjl.tools at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=96061

--- Comment #3 from H.J. Lu <hjl.tools at gmail dot com> ---
(In reply to Alexander Popov from comment #2)
> > We are working on -fzero-call-used-regs:
> 
> H.J. Lu, thanks for the information!
> 
> However, this flag can't be used per-function, like the attribute.

Incorrect.  There is also a zero_call_used_regs attribute.

[Bug target/96061] Please support 'no_caller_saved_registers' attribute on aarch64 (or maybe on all targets)

[alex.popov at linux dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=96061

--- Comment #4 from Alexander Popov <alex.popov at linux dot com> ---
> Incorrect.  There is also a zero_call_used_regs attribute.

Thanks, now I've found it in your tree:

```
You can control this behavior for a specific function by using the function
attribute @code{zero_call_used_regs}.
```

Could you explain the connection between your attribute for zeroing registers
and 'no_caller_saved_registers' attribute?

[Bug target/96061] Please support 'no_caller_saved_registers' attribute on aarch64 (or maybe on all targets)

[hjl.tools at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=96061

--- Comment #5 from H.J. Lu <hjl.tools at gmail dot com> ---
(In reply to Alexander Popov from comment #4)
> > Incorrect.  There is also a zero_call_used_regs attribute.
> 
> Thanks, now I've found it in your tree:
> 
> ```
> You can control this behavior for a specific function by using the function
> attribute @code{zero_call_used_regs}.
> ```
> 
> Could you explain the connection between your attribute for zeroing
> registers and 'no_caller_saved_registers' attribute?

We are porting

https://gitlab.com/x86-gcc/gcc/-/tree/users/hjl/caller-saved/old

from x86 back-end to middle-end.  no_caller_saved_registers changes what
registers zero_call_used_regs can do.  Move no_caller_saved_registers to
middle-end has a direct impact on zero_call_used_regs implementation in
middle-end.

[Bug target/96061] Please support 'no_caller_saved_registers' attribute on aarch64 (or maybe on all targets)

[alex.popov at linux dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=96061

--- Comment #6 from Alexander Popov <alex.popov at linux dot com> ---
>> Could you explain the connection between your attribute for zeroing
>> registers and 'no_caller_saved_registers' attribute?
> 
> We are porting
> 
> https://gitlab.com/x86-gcc/gcc/-/tree/users/hjl/caller-saved/old
> 
> from x86 back-end to middle-end.  no_caller_saved_registers changes what
> registers zero_call_used_regs can do.  Move no_caller_saved_registers to
> middle-end has a direct impact on zero_call_used_regs implementation in
> middle-end.

Ok. So, as I can understand, your work will enable 'no_caller_saved_registers'
attribute for all architectures. Am I right?

If so, good luck to you! Looking forward to your results.

[Bug target/96061] Please support 'no_caller_saved_registers' attribute on aarch64 (or maybe on all targets)

[hjl.tools at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=96061

--- Comment #7 from H.J. Lu <hjl.tools at gmail dot com> ---
(In reply to Alexander Popov from comment #6)
> >> Could you explain the connection between your attribute for zeroing
> >> registers and 'no_caller_saved_registers' attribute?
> > 
> > We are porting
> > 
> > https://gitlab.com/x86-gcc/gcc/-/tree/users/hjl/caller-saved/old
> > 
> > from x86 back-end to middle-end.  no_caller_saved_registers changes what
> > registers zero_call_used_regs can do.  Move no_caller_saved_registers to
> > middle-end has a direct impact on zero_call_used_regs implementation in
> > middle-end.
> 
> Ok. So, as I can understand, your work will enable
> 'no_caller_saved_registers' attribute for all architectures. Am I right?

No.  We are implementing zero_call_used_regs in middle-end.  Someone
can piggyback no_caller_saved_registers on our work.

[Bug target/96061] Please support 'no_caller_saved_registers' attribute on aarch64 (or maybe on all targets)

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=96061

Andrew Pinski <pinskia at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Severity|normal                      |enhancement