Repository navigation
Update JNI Target relocation record to have offset to the reloLocation. - #6326
Conversation
mstoodle
left a comment
There was a problem hiding this comment.
the changes look ok to me, given what's in the OpenJ9 pull request.
| { | ||
| uint8_t *startOfInstruction = self()->getBinaryEncoding(); | ||
| uint8_t *startOfImmediate = cursor; | ||
| intptr_t diff = reinterpret_cast<intptr_t>(startOfImmediate) - reinterpret_cast<intptr_t>(startOfInstruction); |
There was a problem hiding this comment.
did you want to add the < 256 assertion here as well as in OpenJ9 when the actual relocation record is created? I could understand if not (you could view as a requirement imposed by a downstream component) but given it's JNI and likely not to be reused by anything else, maybe catching a problem here is desirable? I'll leave it as optional, because I can see it both ways.
There was a problem hiding this comment.
I was thinking about it, but I think I mainly didn't add it here it so that there if, for whatever reason (perhaps on a different platform) the offset had to be increased to fit in a bigger type, then there would have to be an OMR PR just to update the assert. As such I figured it's probably best to just have the assert close to where the data in the relo header is actually stored.
|
jenkins build all |
|
hm, guess I could have only tested x86 and z...oh well :( |
|
Windows build error is because of a warning that's treated as an error, namely: This is because of the line MSVC does not like the fact that |
On x86, the JNI Target relocation puts the incorrect address when generating a Runtime Assumption. However, the address that the relocation infra itself needs to patch is correct. Thus, the relo record has to hold not just the location to patch but also the location to register the assumption against. This is the same on most platforms, but different on x86. Signed-off-by: Irwin D'Souza <dsouzai.gh@gmail.com>
|
I think you could also just |
|
jenkins build all |
On x86, the JNI Target relocation puts the incorrect address when
generating a Runtime Assumption. However, the address that the
relocation infra itself needs to patch is correct. Thus, the relo record
has to hold not just the location to patch but also the location to
register the assumption against. This is the same on most platforms, but
different on x86.
eclipse-openj9/openj9#14421 depends on this change; OpenJ9 will be broken without a coordinated merge.