Re: [tip:x86/asm] x86: Fix and improve percpu_cmpxchg{8,16}b_double()

["Jan Beulich" <JBeulich@suse.com>]

>>> On 15.12.11 at 23:52, Christoph Lameter <cl@linux.com> wrote:
> On Thu, 15 Dec 2011, tip-bot for Jan Beulich wrote:
>> The __dummy variable was pointless (and needlessly initialized
>> in the 2x32-bit case), given that local copies of the inputs
>> already exist.
> 
> Hmm... I had some failures if I did not specify that dummy in the
> inline asm. Does this work for all gcc versions?

You need to have the output go somewhere, but using the already
existing __o2 for this purpose is possible and sufficient.

>> The 2x64-bit variant forced the address of the first object into
>> %rsi, even though this is needed only for the call to the
>> emulation function. The real cmpxchg16b can operate on an
>> memory.
> 
> Yup. Good idea to code the load into the alternative code path to avoid
> the cmpxchg of the primary code path to be restricted to %si register.
> 
> You dropped the padding with NOPs. Are the instructions on both paths
> always the same length?

leaq and cmpxchg16b use the same operand, so their modrm
encoding (including eventual SIB and immediate) are the same.

leaq is REX+opcode+modrm...
cmpxchg16b (SEG+)REX+0x0f+opcode+modrm

so the latter is two bytes longer than the former (one byte in UP).

call being five bytes vs setz %al being 3 bytes makes it that
lea+call are one byte longer than cmpxchg16b+setz in the UP case,
but I think this is tolerable. If not, the operand of lea could be
made %rip-relative for the case where the operand is a direct
access (i.e. become %1 instead of %P1).

Jan