When I was debugging the case in JDK-8340646, I found the offset layout of class `TestConvertImplicitNullCheck ` on macOS_aarch64 is different from the layout on linux.

With the debug vm flag `-XX:+PrintFieldLayout`, on macOS_aarch64, I watched:

Instance fields:
 @0 12/- RESERVED
 @12 "afff" C 2/2 REGULAR
 @14 "a001" C 2/2 REGULAR
 @16 "a002" C 2/2 REGULAR
 @18 "a003" C 2/2 REGULAR
 @20 "a004" C 2/2 REGULAR
 @22 "a005" C 2/2 REGULAR
 ...
 @8196 "affc" C 2/2 REGULAR
 @8198 "affd" C 2/2 REGULAR
 @8200 "affe" C 2/2 REGULAR
 @8202 "a000" C 2/2 REGULAR

But on Linux, I watched:

Instance fields:
 @0 12/- RESERVED
 @12 "a000" C 2/2 REGULAR
 @14 "a001" C 2/2 REGULAR
 @16 "a002" C 2/2 REGULAR
 @18 "a003" C 2/2 REGULAR
 @20 "a004" C 2/2 REGULAR
...
 @8196 "affc" C 2/2 REGULAR
 @8198 "affd" C 2/2 REGULAR
 @8200 "affe" C 2/2 REGULAR
 @8202 "afff" C 2/2 REGULAR

`afff` and `a000` have different orders here.

The fields do get sorted according to size, see `FieldGroup::sort_by_size()` in `share/classfile/fieldLayoutBuilder.cpp`. Perhaps the sorting algorithm is not stable on macOS when reordering items with the same size.