Implement SVM J2I thunk validation and guarantee resolved virtual dispatch under SVM on x86 #14542
Conversation
|
It's likely that corresponding changes should be just as straightforward in other code generators, but those changes are not included in this PR |
dsouzai
added
the
depends:omr
|
Oh also, you need to update the minor version for JITServer, as a new relocation record has been added (see https://github.com/eclipse-openj9/openj9/blob/master/doc/compiler/misc/BestPractices.md#jitserver) |
jdmpapin
force-pushed
the
svm-j2i-thunk
branch
from
fcd9757
to
76bb70e
Compare
|
Removed the |
This makes the J2I thunk loading and registration logic available to be called for the J2I thunk from method SVM validation that will be defined in a subsequent commit. The calls to pthread_jit_write_protect_np() are removed because although this method is writing to the code cache, the write protection is already handled in applyRelocationAtAllOffsets().
This validation ensures that the J2I thunk appropriate to the method specified by the given method ID exists and is registered with the VM, and uses the address of the J2I thunk to define the given thunk ID.
This will allow the IL generator to generate a resolved virtual call node for invokeinterface when calling a non-final method of Object, e.g. toString().
jdmpapin
force-pushed
the
svm-j2i-thunk
branch
from
76bb70e
to
c3f145a
Compare
|
Rebased to fix the |
|
jenkins test sanity.functional+aot xlinux jdk17 |
|
jenkins test sanity all jdk17 |
1 similar comment
|
jenkins test sanity all jdk17 |
|
s390x failure because of #13128
|
|
jenkins test sanity.functional+aot xlinux jdk17 |
|
The AOT build was aborted because I guess it took too long to get scheduled/run -_- |
|
jenkins test sanity.functional+aot xlinux jdk17 |
|
@IBMJimmyk, since we discussed doing the same change on Power, see the final (x86) commit in this PR. If Power code gen already only generates a plain call through the vtable for resolved virtual calls, it should be similar @joransiu, I haven't discussed this with you, but a similar statement applies to z in case you want to improve code generation for this case in SVM-based AOT |
|
I opened #14638 to fix build failures on AArch64 macOS. |
This commit adds calls to pthread_jit_write_protect_np() for AArch64 macOS, which were removed by eclipse-openj9#14542. Signed-off-by: KONNO Kazuhiro <[email protected]>
The J2I thunk from method validation ensures that the J2I thunk appropriate to the method specified by the given method ID exists and is registered with the VM, and uses the address of the J2I thunk to define the given thunk ID.
Then with this validation, it's very straightforward to guarantee resolved virtual dispatch under SVM on x86. This will allow the IL generator to generate a resolved virtual call node for
invokeinterfacewhen calling a non-final method ofObject, e.g.toString().This PR requires eclipse/omr#6358 to be merged first