[gregharmon]

Summary

Allow opcache.file_cache to persist scripts while JIT is enabled. The file contains bytecode and stable VM handlers; it does not contain generated machine code or process-local JIT metadata.

When a file-cached script is loaded into shared memory, rebuild its JIT state. Scripts loaded into process memory continue to run without JIT, including opcache.file_cache_only mode.

Changes

• Restore original VM handlers from JIT metadata while serializing a private copy. For script-load JIT, rebuild only handlers that point into generated JIT code.
• Clear the function-info slot in the serialized copy and remove the unconditional JIT write guard.
• Initialize the JIT translation table and rebuild JIT after loading bytecode from disk into shared memory.
• Add a PHPT regression test for script-load, first-exec, hot-counter, and tracing JIT. It checks the file write, loads the cached bytecode in a separate process after changing the source, and checks that a shared-memory load generates JIT code again.

Validation

• Built PHP 8.7.0-dev on macOS arm64 with OPcache JIT enabled.
• Passed gh17770_file_cache.phpt, gh16979_check_file_cache_function.phpt, and gh18136.phpt.
• file_cache_frameless_handler.phpt skipped locally because the minimal build did not include dl_test.
• Cross-platform CI is running on this PR.

[gregharmon]

Addressed the Windows CI failure in the new GH-17770 regression test.

Update in ext/opcache/tests/gh17770_jit_file_cache_write.phpt:

• Switched opcache.file_cache to a quoted dedicated path: {PWD}/gh17770_cache
• Pre-creates that directory in --SKIPIF--
• Keeps the test hermetic by cleaning the dedicated cache directory in --CLEAN--
• Uses opcache_is_script_cached_in_file_cache() for direct script-level assertion

This keeps the regression intent unchanged while avoiding startup/path issues on Windows workers.

[dstogov]

This preserves the intended model: file cache stores OPcache bytecode/metadata only; JIT machine code is not persisted and is rebuilt at runtime as usual.

JIT compilation takes time. And it's not going to be profitable in case we re-do it on each request.
I suppose, file_cache without JIT should be faster when with JIT.

Probably, this change may be helpful for some cases - e.g rare opcache resets, but it's probably not a big problem to recompile php from sources. In case of often resets, enabling JIT will probably increase CPU load and for file_cache_only increase it significantly.

[iluuu1994]

@dstogov How would you like to proceed here? This seems like a reasonable change to me. File cache can be used to reduce cache stampede, this applies even when JIT is enabled. But if you're against this change, the PR should be closed.

[dstogov]

I'm not completely against this PR.
I think JIT should be disabled when file_cache_only is set.

[dstogov]

I'm not completely against this PR. I think JIT should be disabled when file_cache_only is set.

Better, JIT should be disabled for scripts loaded from file-cache, but not loaded into shared memory.

[splitice]

Is there some reason this can't be merged. It seems like it does the correct thing.

A more comprehensive solution might be to seperate JIT metadata from the opcache opcode data entirely, this could potentially even allow on disk storage of JIT (useful for same machine, multiple invokation e.g CLI).

[splitice]

@gregharmon I tested this against a PHP 5.6 base (unfortunately jit is pretty broken in 5.5 under our test matrix) against a variety of different tools (phan, phpstan, phpunit, etc) and internal projects.

Ended up re-working this concept slightly as I was recording crashes.

AI refused to review my work due to lack of Daybreak approval (working on it) so potentially it contains a defect I havent been able to spot.

This patch is still under review on our end.

0001-jit-with-file-cache-v2.patch

In PR form I would strip the comments but I retained for @gregharmon's understanding.

[splitice]

Also is there some project of PHP JIT regression tests? It feels like this kind of issue should be automatedly tested.

[gregharmon]

I have Daybreak, working on an update.

[splitice]

😅 My PHP opcode knowledge is a bit rusty clearly.

I'll give this a run through the known failures from last time.

[splitice]

That passes against my matrix.