Re: [patch V4 26/36] rseq: Optimize event setting

[Thomas Gleixner]

On Thu, Sep 11 2025 at 13:15, Mathieu Desnoyers wrote:
> On 2025-09-11 12:06, Thomas Gleixner wrote:
>> On Thu, Sep 11 2025 at 10:03, Mathieu Desnoyers wrote:
>>> On 2025-09-08 17:32, Thomas Gleixner wrote:
>>>>    static inline void rseq_handle_notify_resume(struct pt_regs *regs)
>>>>    {
>>>> -	if (current->rseq.event.has_rseq)
>>>> +	/* '&' is intentional to spare one conditional branch */
>>>> +	if (current->rseq.event.sched_switch & current->rseq.event.has_rseq)
>>>
>>> I wonder.. except for the corner case of rseq unregistration,
>>> when can we have sched_switch set but not has_rseq ?
>>>
>>> We could remove a load from the fast path and the AND if we
>>> clear the sched_switch flag on rseq unregistration.
>> 
>> We probably could. Though I doubt it matters much and I opted for
>> correctness instead of premature optimization.
> Note that this concerns code that documents an explicit " & " as
> intentional to spare a branch, which led me to assume that optimizing
> it was important.

Correct. Optimization and correctness are not mutually exclusive and it
was an obvious optimization to spare the second branch which the
compiler is forced to emit with &&.

A load/AND from them same cache line is definitely cheaper than an extra
conditional branch.

I'm happy to look into that once the dust has settled...

Thanks,

        tglx