On Wed, 20 May 2026 14:04:12 GMT, Roland Westrelin <[email protected]> wrote:
> EA runs a first time, finds a non escaping allocation, splits unique > types, creates extra phi nodes, then `LoadFlat` nodes for that > allocation are expanded. EA runs a second time. At a `Region`, it > finds 2 `Phi` for alias 19: > > 1866 Phi === 1206 1867 1858 [[ 1948 1304 1855 1317 1250 1936 1263 ]] > #memory Memory: > @aryptr:flat:atomic:instptr:compiler/valhalla/inlinetypes/TestFlatArrayMemoryPhi$MyValue:BotPTR:exact > *,iid=bot[int:2] > (java/lang/Cloneable,java/io/Serializable):NotNull:exact:flat():atomic+any,iid=44, > name=b1, idx=19; !orig=[1208],[1938],[1553],[1239] !jvms: > TestFlatArrayMemoryPhi::test @ bci:34 (line 58) > 2073 Phi === 1206 1149 1858 [[ 1948 2061 ]] #memory Memory: > @aryptr:flat:atomic:instptr:compiler/valhalla/inlinetypes/TestFlatArrayMemoryPhi$MyValue:BotPTR:exact > *,iid=bot[int:2] > (java/lang/Cloneable,java/io/Serializable):NotNull:exact:flat():atomic+any,iid=44, > name=b1, idx=19; !orig=2060,[1208],[1938],[1553],[1239] !jvms: > TestFlatArrayMemoryPhi::test @ bci:34 (line 58) > > > The first one was created by EA on its first run. The second was > created after the first round of EA: it's the result of the > transformation of `(Phi (MergeMem ))` into `(MergeMem (Phi ..) (Phi ..)`. > > The second `Phi` references another bottom memory `Phi` (1149) and > that confuses the logic of EA. > > The reason, that second `Phi` is created after EA is that initially > the bottom `Phi` has a `LoadFlat` as input but it gets removed because > it's from the non escaping allocation and the `Phi`'s input becomes a > `MergeMem`. To fix that, I propose that the logic that finds the > memory state for a particular alias during EA skips over `LoadFloat` > nodes: either a `LoadFloat` is for the current non escaping allocation > and it's going away anyway or it's from some other allocation and it's > unrelated. > > --------- > - [x] I confirm that I make this contribution in accordance with the [OpenJDK > Interim AI Policy](https://openjdk.org/legal/ai). Testing is all clean. ------------- PR Comment: https://git.openjdk.org/valhalla/pull/2459#issuecomment-4501603788
