[jorgsowa]

For a leading-numeric string offset such as "1x", the VM warns Illegal string offset "1x" and uses the leading number, also under ??:

function f(string $s) { return $s["1x"] ?? "default"; }
var_dump(f("abc"));
// VM:  Warning: Illegal string offset "1x" + string(1) "b"
// JIT: string(7) "default", no warning

This is the behaviour specified by the Saner Numeric Strings RFC and pinned in Zend/tests/offsets/string_container_offset_behaviour.phpt. The RFC change (b224878) updated the JIT's R helper but not zend_jit_fetch_dim_str_is_helper, which kept a strict parse and silently returned null. The IS helper now parses the offset like the VM, warns on trailing data, and holds a reference to the string in case the error handler releases it.

zend_may_throw_ex() did not count this warning for FETCH_DIM_IS either. With ?? this has no effect, because COALESCE always consumes the result. The inner fetch of a nested isset($s["1x"][0]) / empty() is different:

• DCE removed it when the result was unused, losing the warning even without JIT.
• Function JIT skipped saving the call chain around it, leaking already-sent arguments when the error handler throws.

Code	VM (no opcache)	opcache, JIT off	opcache, function JIT
$s["1x"] ?? "d"	warning, "b"	warning, "b"	before: "d", no warning after: warning, "b"
$x = isset($s["1x"][0]); (result used)	warning, true	warning, true	before: false, no warning after: warning, true
isset($s["1x"][0]); (result unused)	warning	before: no warning after: warning	before: no warning after: warning
consume($t, isset($s["1x"][0])) with a throwing error handler	exception	exception	before: exception plus a leaked $t after: exception

[iliaal]

The undefined-variable warning here can free str the same way: with global $s; return $s[$undef] ?? 1; and an error handler doing $GLOBALS['s'] = null, function JIT then reads the freed string (valgrind; PHP-8.4 has it without this PR too). Could the new guard cover this case as well, like the VM's IS_UNDEF branch does? The VM's own trailing-data warning in zend_fetch_dimension_address_read() has no guard either and shows the same invalid read without opcache.

[jorgsowa]

I created separate PR for this bug fix: #25329

It's valid bug, but I would like to keep all fixes separately. It's better for clearer git history.

[iliaal]

The new refcount guard has no test: with a handler doing $GLOBALS['s'] = null, $s["1x"] ?? "default" under function JIT reads the freed string without it.