Arginfo: add and use known strings for attributes

[DanielEScherzer]

No description provided.

[DanielEScherzer]

This is the first part of my follow-up to #18780 dealing with reducing the work of registering attributes on constants (and other things) - here, if a string is known the known string is used for attribute name and values, and 8.0, 8.1, 8.2, 8.3, 8.4, and 8.5 are all added as known strings.

---

the same strings are allocated twice

It should probably be safe to just intern them all. They are always allocated on start-up, so the memory will be used either way. Interning should just reduce it, no?

Yes I agree.
If you pinky promise that you're gonna look at it before 8.5's release I'll allow it.

Originally posted by @nielsdos in #18780 (comment)

[kamil-tekiela]

Why no match?

[DanielEScherzer]

Because gen_stub supports running on PHP 7.4, I have a todo to update that eventually

[ndossche]

Looks like this makes a degradation on Symfony's Valgrind instruction count. It'd be interesting to know why.

[DanielEScherzer]

Looks like this makes a degradation on Symfony's Valgrind instruction count. It'd be interesting to know why.

Hmm, curious

For 64e2832 which is the last commit before this PR, https://github.com/php/php-src/actions/runs/16128700585/job/45511592322 reports

Details

{
    "branch": "master",
    "Zend\/bench.php": {
        "instructions": "2232991601"
    },
    "Zend\/bench.php JIT": {
        "instructions": "600974941"
    },
    "Symfony Demo 2.2.3": {
        "instructions": "39045323"
    },
    "Symfony Demo 2.2.3 JIT": {
        "instructions": "33525633"
    },
    "Wordpress 6.2": {
        "instructions": "122389752"
    },
    "Wordpress 6.2 JIT": {
        "instructions": "93235272"
    }
}

But for this PR at 99ec395, https://github.com/php/php-src/actions/runs/16155780300/job/45597800753 reports

Details

{
    "branch": "19075\/merge",
    "Zend\/bench.php": {
        "instructions": "2232998193"
    },
    "Zend\/bench.php JIT": {
        "instructions": "600982807"
    },
    "Symfony Demo 2.2.3": {
        "instructions": "39068192"
    },
    "Symfony Demo 2.2.3 JIT": {
        "instructions": "33683655"
    },
    "Wordpress 6.2": {
        "instructions": "122444747"
    },
    "Wordpress 6.2 JIT": {
        "instructions": "93289268"
    }
}

So basically (differences in bold)

Benchmark	Old (master)	New (this PR)
Zend/bench.php	2232991601	2232998193
Zend/bench.php JIT	600974941	600982807
Symfony Demo 2.2.3	39045323	39068192
Symfony Demo 2.2.3 JIT	33525633	33683655
Wordpress 6.2	122389752	122444747
Wordpress 6.2 JIT	93235272	93289268

So this isn't just for symfony, this patch appears to consistently increase the instruction count?

I suspect the reason is that the addition of the new known permanent strings means that any searches for a permanent string now have a couple extra strings to check against - do we care? The memory usage should be lower even if the instruction count is higher

[ndossche]

This could just be startup overhead that is being influenced.
Note btw that the "Symfony Demo 2.2.3 JIT" often fluctuates heavily so that one in particular can likely be ignored.

[DanielEScherzer]

Okay - do you consider the instruction count change a blocker?

[ndossche]

Not really, but it would still be good to confirm with php-cgi's -T argument to see if this is actually startup overhead.

[DanielEScherzer]

Not really, but it would still be good to confirm with php-cgi's -T argument to see if this is actually startup overhead.

I couldn't figure out how to do this, would you be willing to check?

[ndossche]

I'll check tomorrow

[ndossche]

Conceptually this is okay.
I measured locally on Symfony and obtained (for -T10,50):
Before: 29,953,445,510
After: 29,956,118,247

This difference is 1.00008923x, or 100.008922996%, which is lower than CI reports and I think that this is just noise.
We can check the realtime benchmark after this is merged to see the impact on memory usage and on real time performance