Support ignoring the client's local SCC when compiling a JITServer AOT cache store #18937
Conversation
|
Attn @mpirvu. Maybe @dsouzai would like to look as well. This shouldn't change any behaviour, except that in This is the bulk of the work needed in #18301 to support AOT cache store compilations when the client doesn't have a local SCC. Once everything is ready, the client will be able to set |
cjjdespres
force-pushed
the
offset-override
branch
from
b5a3201 to
0abad4e
Compare
|
Force-pushed to bring in the ROM class packing changes that were just merged. |
The _useServerOffsets field will be set when a client wants its local SCC (whether or not it exists) to be ignored during JITServer AOT cache compilations. Currently the client always sets this field to false. Signed-off-by: Christian Despres <[email protected]>
Signed-off-by: Christian Despres <[email protected]>
cjjdespres
force-pushed
the
offset-override
branch
from
0abad4e to
754230f
Compare
|
Force-pushed just to rebase onto master. |
The methods of TR_J9JITServerSharedCache and the handling of the well-known classes object in the SVM and AOT relocation header initialization have been modified so that it we are compiling an AOT cache store and are ignoring the client's local SCC, we now consult the JITServer AOT cache and use the idAndType of its records in place of local SCC offsets. Various client-server messages have been modified to accommodate the changes in server compilation. The client will either test for a local SCC before accessing it, or the server will send an explicit boolean parameter indicating that a local SCC offset is required. Signed-off-by: Christian Despres <[email protected]>
Signed-off-by: Christian Despres <[email protected]>
Signed-off-by: Christian Despres <[email protected]>
cjjdespres
force-pushed
the
offset-override
branch
from
754230f to
2275743
Compare
|
jenkins test sanity plinuxjit,xlinuxjit,zlinuxjit,alinux64jit jdk17 |
|
jenkins compile aix,win jdk21 |
|
The failure on aarch64 is the well known CRIU restore one, hence merging. |
| // This is an out-of-process compilation; check the cache in the client session first | ||
| // This is an out-of-process compilation. | ||
|
|
||
| // If we're ignoring the client's SCC, we can skip client consultation here |
There was a problem hiding this comment.
It's still beneficial (at least by avoiding the global lock) to cache the WKC record in the client session and check the cached version before calling JITServerAOTCache::getWellKnownClassesRecord().
A new
_useServerOffsetsfield inVMInfohas been added. It will be set when a client wants its local SCC (whether or not it exists) to be ignored during JITServer AOT cache compilations. Currently the client always sets this field to false.The methods of
TR_J9JITServerSharedCacheand the handling of the well-known classes object in the SVM and AOT relocation header initialization have been modified so that it we are compiling an AOT cache store and are ignoring the client's local SCC, we now consult the JITServer AOT cache and use theidAndTypeof its records in place of local SCC offsets.Various client-server messages have been modified to accommodate the changes in server compilation. The client will either test for a local SCC before accessing it, or the server will send an explicit boolean parameter indicating that a local SCC offset is required.