On 09/01/2014 02:42 PM, David Laight wrote:
>> Yes unlikely() should cover the whole if statement...
>
> Actually it probably shouldn't.
> You need to look at the generated code with each different set of 'unlikely()'
> to see how gcc processes them.
> In this case, if 'rebooting' is false you want to 'fall through' on a
> statically
> predicted 'not taken' branch. You don't ever care about the second clause.
> With an 'unlikely' covering the entire statement gcc could easily add a
> forwards conditional branch (that will be mis-predicted) for the 'rebooting'
> test.
>
> (Yes, I spent a lot of time getting gcc to generate branches that were
> correctly statically predicted for some code where every cycle mattered.)
>
> David
>
Hi David,
The objdup with an 'unlikely()' covering the entire if statement is as follows:
if (unlikely(rebooting && new_index != get_nominal_index()))
return -EBUSY;
1ac: 2f 89 00 00 cmpwi cr7,r9,0 /* compare rebooting,0 */
1b0: 40 de 00 4c bne- cr7,1fc <.powernv_cpufreq_target_index+0x7c>
The '-' in the instruction bne- specifies an unlikely branch. So gcc has
processed the first clause to be identified as an unlikely branch i.e,
branch to <1fc> (to test the second clause) is unlikely on 'rebooting' not
equal to 0.
1b4: 1f ff 00 0c mulli r31,r31,12
.
. <--- Set the frequency and return --->
.
.
1fc: 3d 22 00 00 addis r9,r2,0 /* test the second clause */
200: 3d 02 00 00 addis r8,r2,0
204: 81 49 00 00 lwz r10,0(r9)
208: 81 28 00 00 lwz r9,0(r8)
20c: 7d 29 50 50 subf r9,r9,r10
210: 7f 89 f8 00 cmpw cr7,r9,r31 /* compare new_index,nominal_index */
214: 41 9e ff a0 beq+ cr7,1b4 <.powernv_cpufreq_target_index+0x34>
The '+' in the instruction beq+ specifies a likely branch. The second clause
unlikely(new_index != get_nominal_index()) is processed to
likely(new_index == get_nominal_index()).
218: 38 60 ff f0 li r3,-16 /* return -EBUSY */
21c: 4b ff ff cc b 1e8 <.powernv_cpufreq_target_index+0x68>
So unlikely() covering the entire statement will not lead to a branch
mis-prediction
for the 'rebooting' test. Having unlikely to cover both 'rebooting' and the
second
clause we can avoid the branch miss prediction for the second clause. This is
advantageous for the code path
powernv_cpufreq_target_index(policy,nominal_index)
which will be invoked by the reboot_notifier.
Thanks and Regards,
Shilpa
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majord...@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/
