RFR: 8349457: [leyden] Workaround Leyden problems with -XX:-TieredCompilation

[Aleksey Shipilev]

On Wed, 5 Feb 2025 13:11:07 GMT, Andrew Dinn <adinn at openjdk.org> wrote:

>> Currently, supplying `-XX:-TieredCompilation` would break with Leyden workflows, because the archive-backed code cache heap would only be created if `SegmentedCodeCache` is enabled. With default compilation policy, this happens normally when tiered compilation is enabled, but not without it.
>> 
>> There are other minor peculiarities that we need to handle, which are exposed with `runtime/cds` + `-XX:-TieredCompilation` runs.
>> 
>> This prevents us from testing C2-only Leyden configs and studying compilation dynamics without tiered policy getting in the way.
>> 
>> Additional testing:
>>  - [x] Ad-hoc Leyden experiments now work with `-XX:-TieredCompilation`: the C2-only code is being stored and loaded successfully from the CDS archive
>>  - [x] `runtime/cds` now passes with `-XX:-TieredCompilation`
>>  - [x] `runtime/cds` still passes by default (i.e. with tiered compilation)
>
> src/hotspot/share/compiler/precompiler.cpp line 243:
> 
>> 241:     }
>> 242: 
>> 243:     if (TieredCompilation) {
> 
> How does this work as far as consistency between training/assembly run and production run is concerned? Do we need to enforce the same setting across these runs or can we e.g. train and then (assuming we fork) assemble with -TieredCompilation but run in production with +TieredCompilation -- or vice versa?

I tested different combinations of `{+|-}TieredCompilation` with training and production runs, and it works well. I believe the normal mechanisms make it right. No C1 code? Nothing loaded. C1 code in archive, but C1 is disabled? We are not loading it then. 

In fact, there seem to be a slight performance edge in training with C2 only, and then running with C1+C2: no penalty on loading lots of C1 compiled code, and quicker transition from intepreter to "real" C1 in only a few places where it is needed! Compiler dynamics is weird.

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

PR Review Comment: https://git.openjdk.org/leyden/pull/32#discussion_r1942995741