8.4 tracing JIT phpseclib failures

[danog]

Description

The following phpseclib testsuites fail on x86-64-v4 with tracing JIT only:

• tests/Unit/Crypt/RC2Test.php

Reproducer: https://github.com/danog/jit_bugs, bugs/phpseclib.sh tracing /app/tests/Unit/Crypt/RC2Test.php

ping @dstogov

PHP Version

nightly

Operating System

No response

[nielsdos]

I debugged this for a while and it seems like values from different CVs are ending up into different CVs. I think it's related to spill conflicts of CVs with ZEND_ASSIGN, but it's complicated to understand. Always returning 1 for IS_CV in zend_jit_spilling_may_cause_conflict does prevent the issue, so that hints that it's maybe indeed spill related.

[dstogov]

@nielsdos Thanks for the initial analyses. This is definitely related to spill conflicts (but related to VAR and POST_INC).

These kind of conflicts are caused by binding of spill slots to PHP VM stack locations. (See slide 53 at https://www.researchgate.net/publication/374470404_IR_JIT_Framework_a_base_for_the_next_generation_JIT_for_PHP). It's possible (especially after IR optimizations) that other variables are loades/stored from/to the same memory locations.

I think, I found a fix for this case, but it doesn't solve the problem in general.