On Fri, 20 Sep 2024 15:26:36 GMT, Roman Kennke <rken...@openjdk.org> wrote:
>> I tried to reproduce for a few hours now using a custom testcase, with no >> success. >> I am pretty sure that this can happen, that is why I added this code. >> Originally I had an assert there asserting that index is not used. I do >> remember that this happens very rarely, and I don't remember the exact >> condition. Looking at the possible operands in opclass memory, I think this >> can only happen when we load an nKlass from an address of the form [rX, rY], >> i.e. the address in rX indexed by rY. This is an odd thing to happen for >> loadNKlass, I think, because rY should always be klass_offset_in_bytes. >> Maybe this is possible when we get odd address merges where we get a PhiNode >> as the offset/index? I don't know. >> I agree, this *might* lead to surprising problems with implicit >> null-checking, if it is expected that the first instruction in loadNKlass >> provokes the SIGSEGV. A way around this would be to declare an opclass that >> is a subset of 'memory' that excludes all operands with index, and match on >> that. I think this would force the lea as a separate instruction and ensure >> that we never see such a thing in loadNKlass. However, I would not feel very >> confident to do that without a reproducer. Let me dig a little further. >> >> For reference, here is my unsuccessful reproducer: >> https://gist.github.com/rkennke/8a57610d74fcde07a9390f268ec35738 > > Something like this is what I have in mind. It seems to pass tier1 tests. I > still haven't managed to reproduce the path that requires an index register, > though. > https://github.com/rkennke/jdk/commit/2c4a7877e4ef94017c8155578d8cfc9342441656 Thanks for the update! If there is a path requiring an index register, I would agree on limiting the memory opclass to exclude indices as you suggest. ------------- PR Review Comment: https://git.openjdk.org/jdk/pull/20677#discussion_r1772945253
