Re: [RFC] Static property asymmetric visibility

From: Date: Tue, 11 Feb 2025 03:58:41 +0000
Subject: Re: [RFC] Static property asymmetric visibility
References: 1 2 3 4 5  Groups: php.internals 
Request: Send a blank email to [email protected] to get a copy of this message
On Tue, Jan 7, 2025, at 1:21 PM, Niels Dossche wrote:
> On 07/01/2025 19:49, Ilija Tovilo wrote:
>> I wouldn't necessarily call this a workaround, but more of a missed
>> optimization (one branch for each static property write that is
>> unlikely to mispredict). If you wish, I can have a look at separating
>> cache slots. This may lead to a slowdown due to cache priming (as only
>> R/IS and RW/W/UNSET will be shared). To be honest, I doubt either of
>> those will lead to a measurable difference in real code. I can confirm
>> this by running some benchmarks. The benefit of separate cache slots
>> is that it can be entirely handled in
>> zend_fetch_static_property_address(), which would reduce VM/ future
>> JIT changes, although the separation of the cache slots themselves
>> might be more complex (given how it's currently implemented).
>> 
>
> I think splitting cache slots will indeed be worse than what it is now.
> However, that argument in itself just doesn't really convince me 
> honestly.
> You are likely right about the branch prediction, but the CPU still has 
> to do some extra work; and it's also just more things to think about as 
> a maintainer.
> I'll think about it more, maybe I'll just abstain from voting.
>
> Kind regards
> Niels

Late follow-up here.  Ilija finally managed to squeeze in time to benchmark this.  Synthetic
benchmarks were inconclusive.  A practical benchmark using Symfony Demo showed the code from this
patch as 0.03% faster than HEAD, which is within the margin of error of the benchmark, so it's
basically a wash/no-impact.

With that settled, I will be opening the vote on this RFC tomorrow sometime, baring any last second
questions or complications.

--Larry Garfield


Thread (11 messages)

« previous php.internals (#126370) next »