Allow promoted readonly properties to be reassigned once in constructor

[nicolas-grekas]

I find it quite annoying that readonly properties play badly with CPP (constructor property promotion).

Doing simple processing of any argument before assigning it to a readonly property forces opting out of CPP.

This PR allows setting once a readonly property in the body of a constructor after the property was previously set using CPP.

Before this PR (CPP not possible):

class Point {
    public readonly float $x;
    public readonly float $y;

    public function __construct(
        float $x = 0.0,
        float $y = 0.0,
    ) {
        $this->x = abs($x);
        $this->y = abs($y);
    }
}

After this PR:

class Point {
    public function __construct(
        public readonly float $x = 0.0,
        public readonly float $y = 0.0,
    ) {
        $this->x = abs($x);
        $this->y = abs($y);
    }
}

This allows keeping properties declaration in its compact form.

I'll submit an RFC of course but would also welcome early feedback here before.

[nicolas-grekas]

Implementation is now ready.
Team work with Claude Code opus 4.5 💪
The CI is green (the failures are about CI tools that couldn't be installed, unrelated to the patch)

[nicolas-grekas]

RFC published at https://wiki.php.net/rfc/promoted_readonly_constructor_reassign

[TimWolla]

A number of comments for the tests. Didn't yet look at all of them.

[iluuu1994]

@nicolas-grekas It looks like the function JIT is failing (this JIT was previously not tested in PRs). I started a run of the real-time benchmark. I've tried this already last week but it failed for some reason, not sure if the failure was related to the changes. The valgrind benchmark shows a regression of 0.3% for Symfony/Wordpress. Not huge but not insignificant (nowadays we fight a lot for these small percentages).

[github-actions]

AWS x86_64 (c6id.metal)

Attribute	Value
Environment	aws
Instance type	c6id.metal
Architecture	x86_64
CPU	Intel(R) Xeon(R) Platinum 8375C CPU @ 2.90GHz, 64 cores @ 2900 MHz
CPU settings	disabled deeper C-states, disabled turbo boost, disabled hyper-threading
RAM	251 GB
Kernel	6.12.66-88.122.amzn2023.x86_64
OS	Amazon Linux 2023.10.20260202
GCC	14.2.1
Time	2026-03-11 11:50:28 UTC
Job details	https://github.com/php/php-src/actions/runs/22951117105 (Artifacts)
Changeset	https://github.com/php/php-src/compare/78c64aadad..ecec79b58a

Laravel 12.11.0 demo app - 50 iterations, 50 warmups, 100 requests (sec)

PHP	Min	Max	Std dev	Rel std dev %	Mean	Mean diff %	Median	Median diff %	Skewness	Z-stat	P-value	Memory
PHP - baseline@78c64aa	0.39440	0.39718	0.00049	0.12%	0.39483	0.00%	0.39472	0.00%	4.010	0.000	1.000	25.26 MB
PHP - reinit-readonly-constructor	0.39263	0.39601	0.00081	0.21%	0.39368	-0.29%	0.39354	-0.30%	0.807	6.408	0.000	25.27 MB

Symfony 2.8.0 demo app - 50 iterations, 50 warmups, 100 requests (sec)

PHP	Min	Max	Std dev	Rel std dev %	Mean	Mean diff %	Median	Median diff %	Skewness	Z-stat	P-value	Memory
PHP - baseline@78c64aa	0.66638	0.66771	0.00029	0.04%	0.66690	0.00%	0.66685	0.00%	0.619	0.000	1.000	25.31 MB
PHP - reinit-readonly-constructor	0.67190	0.68407	0.00200	0.30%	0.67318	0.94%	0.67249	0.85%	3.869	-8.614	0.000	25.32 MB

Wordpress 6.9 main page - 50 iterations, 20 warmups, 20 requests (sec)

PHP	Min	Max	Std dev	Rel std dev %	Mean	Mean diff %	Median	Median diff %	Skewness	Z-stat	P-value	Memory
PHP - baseline@78c64aa	0.59052	0.59439	0.00053	0.09%	0.59151	0.00%	0.59144	0.00%	3.408	0.000	1.000	25.40 MB
PHP - reinit-readonly-constructor	0.59402	0.59683	0.00050	0.08%	0.59488	0.57%	0.59475	0.56%	2.081	-8.593	0.000	25.29 MB

bench.php - 50 iterations, 20 warmups, 2 requests (sec)

PHP	Min	Max	Std dev	Rel std dev %	Mean	Mean diff %	Median	Median diff %	Skewness	Z-stat	P-value	Memory
PHP - baseline@78c64aa	0.44512	0.44688	0.00045	0.10%	0.44592	0.00%	0.44589	0.00%	0.216	0.000	1.000	25.40 MB
PHP - reinit-readonly-constructor	0.44578	0.45047	0.00085	0.19%	0.44811	0.49%	0.44805	0.49%	-0.182	-8.297	0.000	25.29 MB

[iluuu1994]

@nicolas-grekas Sorry for being late to the conversation. Have you considered implementing this solely in the slow readonly error path? I.e. just before emitting the "must not not modify readonly property" error, you could check whether the property is being modified within the constructor. This would allow multiple assignments within the constructor, which I find acceptable. But maybe you can even disallow multiple assignments within the constructor through another property flag.

[nicolas-grekas]

A "slow readonly error path only" implementation wouldn't preserve the desired semantics, because it cannot distinguish these two cases without extra transient state:

• First constructor run after newInstanceWithoutConstructor, where one reassignment should be allowed.
• A later direct __construct() call on an already-constructed object, which must still fail.

Patch updated with perf changes, hopefully removing that previous 0.3%. Let me know.

[iluuu1994]

I think there might be ways to achieve this behavior with less complexity. For example, the emitted assignments for readonly in the constructor could be preceded by a check that asserts properties are uninitialized when the constructor is called. This prevents multiple constructor calls. Then, the slow path could add a property flag to indicate the property has been assigned a second time, to prevent further reassignments within the constructor.

[nicolas-grekas]

LLM-assisted response - might help reason about your proposal @iluuu1994 , and doesn't mean you're wrong:

An uninitialized check at constructor entry would help with repeated direct __construct() calls, but the feature still needs runtime state in the property write paths to allow exactly one reassignment from the active CPP-owning constructor chain, including methods/closures called from it, while rejecting writes on other objects. In practice this cannot live only in the slow path, because typed-property writes also go through the VM fast path and JIT helpers. We also need clone semantics to remain separate. So this approach still ends up needing context-sensitive state in the write paths, which is essentially what the current design is doing.

[github-actions]

AWS x86_64 (c6id.metal)

Attribute	Value
Environment	aws
Instance type	c6id.metal
Architecture	x86_64
CPU	Intel(R) Xeon(R) Platinum 8375C CPU @ 2.90GHz, 64 cores @ 2900 MHz
CPU settings	disabled deeper C-states, disabled turbo boost, disabled hyper-threading
RAM	251 GB
Kernel	6.12.66-88.122.amzn2023.x86_64
OS	Amazon Linux 2023.10.20260202
GCC	14.2.1
Time	2026-03-11 18:09:15 UTC
Job details	https://github.com/php/php-src/actions/runs/22967396824 (Artifacts)
Changeset	https://github.com/php/php-src/compare/13b83a46cf..e201e69e00

Laravel 12.11.0 demo app - 50 iterations, 50 warmups, 100 requests (sec)

PHP	Min	Max	Std dev	Rel std dev %	Mean	Mean diff %	Median	Median diff %	Skewness	Z-stat	P-value	Memory
PHP - baseline@13b83a4	0.38979	0.39327	0.00056	0.14%	0.39054	0.00%	0.39042	0.00%	2.866	0.000	1.000	25.17 MB
PHP - reinit-readonly-constructor	0.38989	0.39166	0.00030	0.08%	0.39031	-0.06%	0.39024	-0.05%	2.097	2.899	0.004	25.30 MB

Symfony 2.8.0 demo app - 50 iterations, 50 warmups, 100 requests (sec)

PHP	Min	Max	Std dev	Rel std dev %	Mean	Mean diff %	Median	Median diff %	Skewness	Z-stat	P-value	Memory
PHP - baseline@13b83a4	0.66816	0.67189	0.00069	0.10%	0.66882	0.00%	0.66862	0.00%	2.590	0.000	1.000	25.21 MB
PHP - reinit-readonly-constructor	0.66689	0.67495	0.00141	0.21%	0.66945	0.09%	0.66939	0.12%	1.108	-3.395	0.001	25.21 MB

Wordpress 6.9 main page - 50 iterations, 20 warmups, 20 requests (sec)

PHP	Min	Max	Std dev	Rel std dev %	Mean	Mean diff %	Median	Median diff %	Skewness	Z-stat	P-value	Memory
PHP - baseline@13b83a4	0.58611	0.58938	0.00067	0.11%	0.58684	0.00%	0.58664	0.00%	2.273	0.000	1.000	25.23 MB
PHP - reinit-readonly-constructor	0.58966	0.59257	0.00059	0.10%	0.59030	0.59%	0.59013	0.60%	2.491	-8.614	0.000	25.23 MB

bench.php - 50 iterations, 20 warmups, 2 requests (sec)

PHP	Min	Max	Std dev	Rel std dev %	Mean	Mean diff %	Median	Median diff %	Skewness	Z-stat	P-value	Memory
PHP - baseline@13b83a4	0.44419	0.44721	0.00057	0.13%	0.44561	0.00%	0.44559	0.00%	0.029	0.000	1.000	25.23 MB
PHP - reinit-readonly-constructor	0.44464	0.44692	0.00055	0.12%	0.44570	0.02%	0.44562	0.01%	0.171	-0.672	0.501	25.23 MB

[github-actions]

AWS x86_64 (c6id.metal)

Attribute	Value
Environment	aws
Instance type	c6id.metal
Architecture	x86_64
CPU	Intel(R) Xeon(R) Platinum 8375C CPU @ 2.90GHz, 64 cores @ 2900 MHz
CPU settings	disabled deeper C-states, disabled turbo boost, disabled hyper-threading
RAM	251 GB
Kernel	6.12.66-88.122.amzn2023.x86_64
OS	Amazon Linux 2023.10.20260202
GCC	14.2.1
Time	2026-03-12 16:23:41 UTC
Job details	https://github.com/php/php-src/actions/runs/23012277036 (Artifacts)
Changeset	https://github.com/php/php-src/compare/13b83a46cf..182e462af1

Laravel 12.11.0 demo app - 50 iterations, 50 warmups, 100 requests (sec)

PHP	Min	Max	Std dev	Rel std dev %	Mean	Mean diff %	Median	Median diff %	Skewness	Z-stat	P-value	Memory
PHP - baseline@13b83a4	0.38807	0.39007	0.00044	0.11%	0.38862	0.00%	0.38849	0.00%	1.503	0.000	1.000	25.17 MB
PHP - reinit-readonly-constructor	0.38589	0.38828	0.00051	0.13%	0.38643	-0.56%	0.38628	-0.57%	2.067	8.531	0.000	25.18 MB

Symfony 2.8.0 demo app - 50 iterations, 50 warmups, 100 requests (sec)

PHP	Min	Max	Std dev	Rel std dev %	Mean	Mean diff %	Median	Median diff %	Skewness	Z-stat	P-value	Memory
PHP - baseline@13b83a4	0.66498	0.66707	0.00043	0.06%	0.66550	0.00%	0.66542	0.00%	1.436	0.000	1.000	25.21 MB
PHP - reinit-readonly-constructor	0.66116	0.66500	0.00070	0.11%	0.66197	-0.53%	0.66170	-0.56%	2.144	8.600	0.000	25.28 MB

Wordpress 6.9 main page - 50 iterations, 20 warmups, 20 requests (sec)

PHP	Min	Max	Std dev	Rel std dev %	Mean	Mean diff %	Median	Median diff %	Skewness	Z-stat	P-value	Memory
PHP - baseline@13b83a4	0.58473	0.58783	0.00050	0.08%	0.58538	0.00%	0.58527	0.00%	2.589	0.000	1.000	25.27 MB
PHP - reinit-readonly-constructor	0.58502	0.58911	0.00095	0.16%	0.58571	0.06%	0.58538	0.02%	2.533	-2.065	0.039	25.28 MB

bench.php - 50 iterations, 20 warmups, 2 requests (sec)

PHP	Min	Max	Std dev	Rel std dev %	Mean	Mean diff %	Median	Median diff %	Skewness	Z-stat	P-value	Memory
PHP - baseline@13b83a4	0.44436	0.44817	0.00062	0.14%	0.44551	0.00%	0.44545	0.00%	1.416	0.000	1.000	25.27 MB
PHP - reinit-readonly-constructor	0.44430	0.44686	0.00062	0.14%	0.44553	0.00%	0.44551	0.01%	0.088	-0.327	0.743	25.23 MB