[master] RFR: Relax array elements alignment

[David Holmes]

On Wed, 23 Feb 2022 13:03:03 GMT, Roman Kennke <rkennke at openjdk.org> wrote:

> PR #40 eliminated the Klass* word. A whole class of objects would not benefit from it, though: array elements are aligned at 8-byte-boundaries, even for element types where this is not necessary (byte, bool, char, short, int, float, compressed-objs). This means that array elements would still start at byte#16, after 8 bytes header and 4 bytes length and another 4 bytes unused gap.
> 
> This change improves the array element alignment and relaxes it such that elements can align at 4-byte boundaries.
> 
> Note about C2 changes: the code that extracts the state field (either byte, int or long array) assumed that all arrays would be equal, and asks for header size of T_INT. This would give wrong offset for T_LONG arrays.
> 
> I'm adding a test that checks correct array elements base offsets.
> 
> GHA seems to see 3 new test failures (timeouts), but only on Windows(x64). I'm not quite sure how this is related, and probably needs some testing, once I get hold of a windows machine:
> 
> compiler/c2/irTests/TestAutoVectorization2DArray.java 
> compiler/c2/irTests/TestIRAbs.java 
> compiler/c2/irTests/TestStripMiningDropsSafepoint.java 
> 
> Testing:
>  - [x] tier1 (x86_64)
>  - [x] tier1 (x86_32)
>  - [x] tier1 (aarch64)
>  - [x] tier2 (x86_64)
>  - [x] tier2 (x86_32)
>  - [x] tier2 (aarch64)
>  - [x] tier3 (x86_64)
>  - [x] tier3 (x86_32)
>  - [x] tier3 (aarch64)

I'm not at all familiar with the code being changed here but I just want to flag something that is probably obvious, but array accesses are required to be atomic (except for double and long) so I hope these alignment changes don't break that.

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

PR: https://git.openjdk.java.net/lilliput/pull/41