On Thu, 29 May 2025 13:08:25 GMT, Kim Barrett <kbarr...@openjdk.org> wrote:
>> 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? Yes, I find it much cleaner when intrinsic is non-virtual. > Or maybe that's better done as a later cleanup? Up to you. I'm fine with the current PR if you remove the comment referring to C2. ------------- PR Review Comment: https://git.openjdk.org/jdk/pull/24315#discussion_r2114543162