On Thu, 29 May 2025 01:14:30 GMT, Vladimir Ivanov <vliva...@openjdk.org> wrote:
>> The comment is about why we have native `get0` at all, rather than just >> making >> `get` also be native. >> >>> [...] JDK-8271862 was about migrating to non-virtual intrinsic method. >> >> That's not a correct summary of JDK-8271862. That change was about migrating >> away from having a method that was all 3 of (1) intrinsic (2) native, and (3) >> virtual. It split that method into two parts, and allocated the properties to >> them in such a way that neither had all 3. There are several possible >> allocations of the properties to the methods that would have worked. In this >> case we already had all the machinary in place for `refersTo0` to be >> intrinsic >> and native, so the allocation that didn't require a bunch of renaming for >> consistency was to introduce `refersToImpl` to provide the virtual part. >> >> For this PR we have a method that is already intrinsic and virtual. Of >> course, >> we don't want to make it native too. Since we already have all the machinary >> in place for `get` to be intrinsic and virtual, the allocation that didn't >> require a bunch of renaming for consistency was to introduce native `get0` >> for >> use as the non-intrinsic implementation of `get`. See also here: >> https://github.com/openjdk/jdk/pull/24315#discussion_r2039137923 > > The current limitation of intrinsics support in C1/C2 is that intrinsics are > always applied in the context of some method (as part of inlining). If a > method is at the root of the compilation, it is never intrinsified. > > The problem with virtual intrinsic methods is that when devirtualization > fails, the call site ends up permanently calling into non-intrinsified > version. That's what `refersTo0` initially suffered (fixed by JDK-8271862). > `Reference::refersTo0` was virtual intrinsic method (no intrinsic when no > inlining, hence call into native) and once it was untangled (virtual > `refersToImpl` calls non-virtual intrinsic method `refersTo0`) > intrinsification happens reliably in the context of `refersToImpl`. We already have this to address that issue for the specific case of Reference.get: https://github.com/openjdk/jdk/blob/4cf729cfac57c9aec692a52c1f3f95f2403e7958/src/hotspot/share/opto/compile.cpp#L786-L792 I think if we made the non-virtual native method be the intrinsic then we could remove that special handling? Maybe that's a good enough reason to cover the cost of renaming churn. Or maybe that's better done as a later cleanup? ------------- PR Review Comment: https://git.openjdk.org/jdk/pull/24315#discussion_r2113909448
