Re: [RFC PATCH, i386]: Use %r15 for REAL_PIC_OFFSET_TABLE_REGNUM on x86_64

["Leif Ekblad" <leif at rdos dot net>]

In the case of cpuid, the code is hardly performance sensitive, and probably 
runs only at startup. An alternative solution for the broken code here is to 
move the result from rbx to another register, and to save/restore rbx. 
Currently, this is the only place in libgcc and newlib affected by this 
problem.

Leif Ekblad

----- Original Message ----- 
From: "Andi Kleen" <andi@firstfloor.org>
To: "Uros Bizjak" <ubizjak@gmail.com>
Cc: <gcc-patches@gcc.gnu.org>; "Richard Henderson" <rth@redhat.com>; "Jakub 
Jelinek" <jakub@redhat.com>; <jh@suse.cz>; "H.J. Lu" <hjl.tools@gmail.com>
Sent: Monday, December 24, 2012 10:32 PM
Subject: Re: [RFC PATCH, i386]: Use %r15 for REAL_PIC_OFFSET_TABLE_REGNUM on 
x86_64


> Uros Bizjak <ubizjak@gmail.com> writes:
>
> > Hello!
> >
> > Currently, we use %rbx as REAL_PIC_OFFSET_TABLE_REGNUM on x86_64.
> > Since this register gets marked as fixed reg in
> > ix86_conditional_register_usage, we get into troubles with insns that
> > use %rbx (cmpxchg, cpuid). According to x86_64 psABI, we are free to
> > use any register, so attached patch changes %rbx with %r15 (also
> > following the example in the psABI). This patch has no implications on
> > small code model (that doesn't use REAL_PIC_OFFSET_TABLE_REGNUM
> > anyway), but on medium and large code model fixes usage of cpuid.h
> > (please see PR 55712 [1]) and avoids a pair of xchgs around cmpxchg or
> > cpuid instructions.
>
> So everyone who worked around this and use r15 now broke.
>
> It would be probably better to just teach the register allocator
> to spill those registers as needed, to handle some asm() statement.
> Not sure how hard that would be.
>
> -Andi
>
> --
> ak@linux.intel.com -- Speaking for myself only