[lworld] Integrated: 8366973: [lworld] Deopt/recompilation cycle due to call to method with unloaded return type

[Tobias Hartmann]

On Tue, 9 Sep 2025 08:38:06 GMT, Tobias Hartmann <thartmann at openjdk.org> wrote:

> In some benchmarks we hit an endless deopt/recompilation cycle because of this bailout in C2 when we compile a method with an unloaded return type:
> https://github.com/openjdk/valhalla/blob/858be30119bd5bc37a69d16042523f53bea71a36/src/hotspot/share/ci/ciTypeFlow.cpp#L766-L772
> 
> It's a regression / limitation from [JDK-8301007](https://bugs.openjdk.org/browse/JDK-8301007) that I had planned to fix later with [JDK-8284443](https://bugs.openjdk.org/browse/JDK-8284443) because I thought it wouldn't have an impact on performance. Turns out it does, so I'll fix it with this issue similar to what we do in `PhaseMacroExpand::expand_mh_intrinsic_return`.
> 
> For testing, I added a new `PreloadClasses` debug flag that's enabled by default. Disabling it will disable preloading of all classes from the LoadableDescriptors attributes. Running testing with this caught some more issues that I fixed in this PR as well. I'll add it to our (internal) stress testing job separately.
> 
> I also fixed some inconsistency in the preload logging and adjusted the corresponding tests.
> 
> Thanks,
> Tobias

This pull request has now been integrated.

Changeset: 9014ec21
Author:    Tobias Hartmann <thartmann at openjdk.org>
URL:       https://git.openjdk.org/valhalla/commit/9014ec219423a83cf1fb63c496b86fca05c40a8f
Stats:     319 lines in 27 files changed: 207 ins; 18 del; 94 mod

8366973: [lworld] Deopt/recompilation cycle due to call to method with unloaded return type

-------------

PR: https://git.openjdk.org/valhalla/pull/1554