[Bug other/56314] New: Please allow per-function specification of register conventions

[Bug other/56314] New: Please allow per-function specification of register conventions

[hpa at zytor dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=56314

             Bug #: 56314
           Summary: Please allow per-function specification of register
                    conventions
    Classification: Unclassified
           Product: gcc
           Version: unknown
            Status: UNCONFIRMED
          Severity: enhancement
          Priority: P3
         Component: other
        AssignedTo: unassigned@gcc.gnu.org
        ReportedBy: hpa@zytor.com


gcc allows register specifications (saved, clobbered, reserved) to be changed
on a per-file basis.  However, for optimization uses it would be much more
useful if they could be defined on a per-function basis using attributes, just
as calling conventions can be specified with attributes.

In the Linux kernel world we currently have several uses of functions with ad
hoc calling conventions.  We have to wrap them in assembly wrappers, and either
write them in assembly or put them in separate files with the calling
convention specified with compiler options.

[Bug other/56314] Please allow per-function specification of register conventions

[pinskia at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=56314

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
            Version|unknown                     |4.8.0

--- Comment #1 from Andrew Pinski <pinskia at gcc dot gnu.org> 2013-02-14 02:03:44 UTC ---
Can you show an example of what you mean by this?
Also it might be better to come up with a full semantics of such a change as
right now we don't know the full semantics of these changes.  Also is this only
for x86 (and x86_64) or for other targets too?  There are ways currently
already to spell out the regparm for x86 which might be enough for your usage
but without knowledge of the linux kernel, it is hard to know what you need.

[Bug other/56314] Please allow per-function specification of register conventions

[hpa at zytor dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=56314

--- Comment #2 from H. Peter Anvin <hpa at zytor dot com> 2013-02-14 02:23:53 UTC ---
Consider the following syntax:

__attribute__((saved(...),used(...),fixed(...)))

... as the equivalent to -fcall-saved-*, -fcall-used-*, and -ffixed-*
respectively.

In Linux/x86-64 we could then decorate the function __sw_hweight64() as:

unsigned long __attribute__((saved(rdi,rsi,rdx,rcx,r8,r9,r10,r11)))
__sw_hweight64(__u64 v);

One major use for this for us is to be able to reduce the call site impact of
certain small functions by reducing the used set of those functions.

Please let me know if you need more examples.

[Bug middle-end/56314] Please allow per-function specification of register conventions

[rguenth at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=56314

Richard Biener <rguenth at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Keywords|                            |missed-optimization, ra
             Status|UNCONFIRMED                 |NEW
   Last reconfirmed|                            |2013-02-14
                 CC|                            |vmakarov at gcc dot gnu.org
          Component|other                       |middle-end
     Ever Confirmed|0                           |1

--- Comment #3 from Richard Biener <rguenth at gcc dot gnu.org> 2013-02-14 10:29:27 UTC ---
Confirmed.

[Bug middle-end/56314] Please allow per-function specification of register conventions

[benh at kernel dot crashing.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=56314

benh at kernel dot crashing.org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |benh at kernel dot
                   |                            |crashing.org

--- Comment #4 from benh at kernel dot crashing.org 2013-02-15 06:52:31 UTC ---
I would definitely like to see that on powerpc as well.

A typical usage case for Linux is inline functions intended as a fast path
(such as spin lock) for which we want to be able to branch to an unlikely slow
path (contention case, timeout case, error case...) while limiting the impact
on the caller.

[Bug middle-end/56314] Please allow per-function specification of register conventions

[pinskia at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=56314

--- Comment #5 from Andrew Pinski <pinskia at gcc dot gnu.org> 2013-02-15 06:56:23 UTC ---
I think I would rather see IPA-RA and LTO working with the kernel before adding
these attributes as they are very fragile to get working correctly.

[Bug middle-end/56314] Please allow per-function specification of register conventions

[hpa at zytor dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=56314

--- Comment #6 from H. Peter Anvin <hpa at zytor dot com> ---
Unfortunately that's not really possible given the way the way the level does
runtime patching (which isn't going to change, sorry.) At the very least we
would need a *lot* more compiler support to give LTO all the information that
it needs; say a *very* minimum LTO would need to support ORC metadata.