token_get_all(TOKEN_PARSE): class constant named with a reserved keyword is no longer retagged to T_STRING when the same keyword appears in its initializer (regression in 8.3)

Beginne in ext/tokenizer/tokenizer.c, insbesondere bei der Verarbeitung von ON_FEEDBACK und dem Pfad extract_token_id_to_replace(), und reproduziere anschließend das Beispiel token_get_all($code, TOKEN_PARSE) aus dem Issue. Vergleiche die Tokenfolge für class Foo { public const NEW = Bar::NEW; } zwischen PHP 8.2 und PHP 8.3+. Erledigt ist die Aufgabe, wenn sowohl das Vorkommen in der Deklaration als auch das im Initializer T_STRING erzeugen, ohne die anderen Trigger-Fälle zu regressieren.

Description

Description

Since PHP 8.3, token_get_all($code, TOKEN_PARSE) fails to retag a reserved
keyword used as a class constant name to T_STRING when the same keyword
also appears as a class-constant fetch inside the initializer of that constant.

The code itself parses, lints and runs fine — only the tokenizer output is wrong.

<?php
$code = '<?php class Foo { public const NEW = Bar::NEW; }';

foreach (token_get_all($code, TOKEN_PARSE) as $t) {
    if (is_array($t) && strcasecmp($t[1], 'new') === 0) {
        echo token_name($t[0]), ' [', $t[1], ']', PHP_EOL;
    }
}

Resulted in this output (PHP 8.3, 8.4, 8.5):

T_NEW [NEW]        <-- declaration: NOT retagged (bug)
T_STRING [NEW]     <-- usage after :: : retagged correctly

But I expected this output instead (and this is what PHP <= 8.2 produces):

T_STRING [NEW]
T_STRING [NEW]

https://3v4l.org/r4Yni — shows the regression starts exactly at 8.3.0 (all 8.2.x
produce the expected output; 8.3.0 through 8.5.10 produce the wrong one).

Trigger conditions

The retagging only fails when the same reserved word occurs both as the
declared constant name and as a :: constant fetch in the initializer:

Code	Declared name token on 8.3+
class F { const NEW = 1; }	T_STRING ✅
class F { const NEW = Bar::BAZ; }	T_STRING ✅
class F { const FOO = Bar::NEW; }	(usage) T_STRING ✅
class F { const NEW = Bar::NEW; }	T_NEW ❌
class F { const NEW = [self::NEW]; }	T_NEW ❌

Reproduced with other semi-reserved keywords as well (the issue is not specific
to new).

Suspected cause

ext/tokenizer/tokenizer.c handles the parser feedback for
"keyword used as identifier" by searching the already-collected token list
backwards by token text and retagging the first match:

case ON_FEEDBACK: {
    HashTable *tokens_ht = ctx->tokens;
    zval *token_zv, *id_zv = NULL;
    ZEND_HASH_REVERSE_FOREACH_VAL(tokens_ht, token_zv) {
        id_zv = extract_token_id_to_replace(token_zv, text, length);
        if (id_zv) {
            break;
        }
    } ZEND_HASH_FOREACH_END();
    ...

Up to PHP 8.2 the class-constant name was reduced (and the feedback emitted)
before the initializer was scanned, so the reverse-by-text search found the
correct token. With typed class constants (PHP 8.3, #10444) the grammar has
to disambiguate the type from the constant name, the reduction of the name
happens later, and by the time the feedback fires the initializer tokens have
already been scanned — so the reverse search finds the initializer occurrence
of the same text instead of the declaration, leaving the declared name tagged
as a keyword.

PHP Version

• PHP 8.2.33 — not affected (expected output)
• PHP 8.3 (php:8.3-cli) — affected
• PHP 8.4 (php:8.4-cli) — affected
• PHP 8.5.10 (php:8.5-cli) — affected

Operating System

Reproduced on Debian-based official Docker images — the behaviour is platform
independent.

PHP Version

PHP 8.5.9 (cli) (built: Jul 28 2026 13:06:52) (NTS)
Copyright (c) The PHP Group
Built by Homebrew
Zend Engine v4.5.9, Copyright (c) Zend Technologies
    with Zend OPcache v8.5.9, Copyright (c), by Zend Technologies

Operating System

No response