On Thu, 17 Oct 2024 23:11:07 GMT, Joe Darcy <da...@openjdk.org> wrote:
> Port of Float16 from java.lang in the lworld+fp16 branch to > jdk.incubabor.vector. > Comparing with #21490 we can see that there are more than minimum number of > intrinsics I recommended above, but (crucially) the intrinsics are decoupled > from the box type, and refactored into `Float16Math.java`. > > https://github.com/openjdk/jdk/pull/21490/files#diff-105a2bf4929174c594b5bcc54f749e93d9fca1b5371ca301fff02badd0e8da5a > > For example, see the max intrinsics line 53. These intrinsics are better > structured for the JIT, because there is no extraneous boxing to deal with. > (Boxes involve null checks and heap allocation and other details which are > just extra overhead.) > > Apart from intrinsic classification, a simpler thing to do that may help the > box methods to fold up correctly would be a `@ForceInline` annotation. But > for the simple methods in this PR that is usually overkill. I'm agnostic toward how the intrinsification of Float16 is implemented. I've removed the commented-out intrinsics annotations in a subsequent push. As noted, for review purposes I wanted to have a least one iteration of Float16 in the incubator that was as close as possible to the version of Float16 that has been in use in the lworld+fp16 branch. ------------- PR Comment: https://git.openjdk.org/jdk/pull/21574#issuecomment-2425177653
