Repository navigation
Description
Description
With the tracing JIT, PHP 8.4, 8.5 and 8.6 compute wrong values in this function once its loop is compiled. $avail goes from 30 to
54 in the first iteration and stays 54, and $codes[0] is always 4, so every element must be (0x123456789ABCDEF >> 54) & 15 = 4.
The two inner while loops after the computation never run (their condition is false), but the bug needs them.
<?php
declare(strict_types=1);
function f(int $bits, int $avail, array $codes, int $n): array
{
$out = [];
for ($i = 0; $i < $n; ++$i) {
while ($avail <= 48) {
$avail += 8;
}
$code = $codes[0];
$val = ($bits >> $avail) & ((1 << $code) - 1);
if ($code > 16) {
while ($avail <= 48) {
}
}
if ($i < $n - 1) {
while ($avail <= 48) {
}
}
$out[] = $val;
}
return $out;
}
echo implode(',', array_unique(f(0x123456789ABCDEF, 30, [4], 300))), "\n";php -d opcache.enable_cli=1 -d opcache.jit_buffer_size=64M -d opcache.jit=tracing -d opcache.file_update_protection=0 repro.php
Resulted in this output:
4,6
But I expected this output instead:
4
It is already fixed on master (8.7.0-dev). git bisect between b30219a (2026-09-24, wrong) and master 9c86e04 (correct) names
the first fixed commit as 54ec8c4 "Update IR" (IR commit 281a66a87). The relevant part is dstogov/ir GH-302 ("x86-64 de-SSA temporary
reuse"), fixed upstream by ea66717 "Fix GH-302: DESSA temporary reuse during resolving the remaining windmill blades" and 70ae5f1 "Fix
GH-302: possible clobbering of second operand":
| build | tracing JIT |
|---|---|
| 8.4.27, 8.5.12 (releases) | wrong |
| PHP-8.4 head c0e0774 (8.4.29-dev), PHP-8.5 head 9afb52d (8.5.14-dev), PHP-8.6 head 820ec86 (8.6.1-dev) | wrong |
| master 9c86e04 (8.7.0-dev) | correct |
PHP-8.6 with master's ext/opcache/jit/ir from 54ec8c4^ |
wrong |
PHP-8.6 with master's ext/opcache/jit/ir from 54ec8c4 |
correct |
PHP-8.6 with master's ext/opcache/jit/ir from 54ec8c4^ plus only ea66717 and 70ae5f1 |
correct |
PHP-8.6 with only the ir_ra.c dead-PHI change of 54ec8c4, or only the ir_sccp.c change |
wrong |
The two GH-302 commits do not apply cleanly to the IR copies in PHP-8.4 and PHP-8.5 (older IR), so those need an adapted backport.
GH-302 itself mentions no PHP impact; it was found in rv32emu.
Other details:
- Needs register allocation and the type-inference optimization level:
opcache.jit= tracing / 1254 / 1255 / 1253 / 1154 wrong;
1054 (no register allocation), 1252, 1251, function / 1205 and off are correct. - Only one trace is compiled (the outer
forloop). The first wrong element is at iteration 66; 234 of 300 are wrong. - Each of these hides it:
$codes[0]replaced by a plainintparameter,& ((1 << $code) - 1)replaced by& 15, or either
laterwhileloop removed. - Found in a real program: the bitstream reader of a pure-PHP Zstandard decoder (refill loops
while ($bitsAvailable <= 48 && $cursor >= $start) { $container = ($container << 8) | ord($src[$cursor--]); $bitsAvailable += 8; })
decodes valid frames wrongly on 8.4 and 8.5 with the default tracing JIT. The function above was reduced from it.
PHP Version
PHP 8.4.27, 8.5.12 (cli, NTS, distribution packages); branch heads PHP-8.4, PHP-8.5, PHP-8.6 as of 2026-11-13
Operating System
Linux x86_64 (6.17); branch heads built in a debian:trixie container
Activity
Fix phpGH-25456: Tracing JIT wrong values after an inner loop