On Thu, 24 Oct 2024 13:41:13 GMT, Emanuel Peter <epe...@openjdk.org> wrote:
>> To exercise non memory operand pattern we need a vector operation padding >> layer after load vector, this will always ensure that selector pick all >> register operands flavor of instruction. Since its a generic limitation, do >> you think we should float it as a separate PR? >> >> I have create an RFE https://bugs.openjdk.org/browse/JDK-8342959 for >> reference. Given that we have moved IR tests out this PR on the grounds of >> review complexity, lets not add more code here. > > Ok, we can file a separate RFE. Though I really have voiced 2 concerns: > - Making sure we always test `_mem` and `_reg` variants in the backend. See > your https://bugs.openjdk.org/browse/JDK-8342959 > - Making sure we have tests that would detect vectors that are too long. This > would require some padding between the vectors, so that we have some > untouched space - and if it does get touched we know that a vector was too > long. Does that make sense? This is I guess also a general concern - and > would have to be applied to all vector instructions. Good point on vector operations overrunning bounds. I worry about the computational increase of doing this generally for all operations (explicit or for auto vectorization i suppose). Perhaps we can focus on areas where we know this may be problematic? ------------- PR Review Comment: https://git.openjdk.org/jdk/pull/20507#discussion_r1815211324
