On Mon, 19 Jan 2026 14:51:48 GMT, Christian Hagedorn <[email protected]> wrote:
> When using compiler replay with `--enable-preview` with array store/load > profiling, we only create `ObjArrayKlass` objects instead of subclass objects > (i.e. `RefArrayKlass` or `FlatArrayKlass`). The reason is that we are > directly resolving klasses with `SystemDirectory::resolve_or_fail()`: > https://github.com/openjdk/valhalla/blob/a4fb7ebd5af316d3e99d10bffb44d5be3aab5548/src/hotspot/share/ci/ciReplay.cpp#L557 > > This method will call `InstanceKlass::array_klass()` at some point which > directly creates an `ObjArrayKlass` object. This let's the replayed > compilation fail when trying to speculate on an value class array which > expects a subclass of `ObjArrayKlass`. > > The fix I propose is to make sure that we are always creating subclass > objects by explicitly using `ArrayProperties::DEFAULT` when parsing an > `ArrayKlass` during replay compilation. > > I added a compiler replay test which triggers the same assert as shown in the > report. > > Thanks, > Christian src/hotspot/share/ci/ciReplay.cpp line 562: > 560: // Create ref or flat array klass. > 561: k = > ObjArrayKlass::cast(k)->klass_with_properties(ArrayKlass::ArrayProperties::DEFAULT, > THREAD); > 562: } This will always create the default refined klass. Shouldn't we keep track of which refined klass was actually referenced in the replay file and then create that particular one here? ------------- PR Review Comment: https://git.openjdk.org/valhalla/pull/1931#discussion_r2705210950
