RFR: 8300543 Compiler Implementation for Pattern Matching for switch

[Jan Lahoda]

On Thu, 13 Apr 2023 10:59:46 GMT, ExE Boss <duke at openjdk.org> wrote:

>> Hi Christian,
>> 
>> Thanks for the comment.
>> 
>> First, the "Dumbest possible strategy" comment should be interpreted as "we would like to do something better", not that using something slow is OK permanently. There's an attempt to do performance improvements here:
>> https://github.com/openjdk/jdk/pull/9779
>> 
>> Talking of `enumSwitch` for now (some more on `typeSwitch` later), one thing what I would like allow for the long(er) term are fast(er) implementations. If we just used `String` comparison, it would be extremely difficult to get a really good performance, even for cases where e.g. all the input values are enum constants, where the method can be reduced to a map lookup (as 9779 does).
>> 
>> I think we could probably still avoid the "eager" class initialization, at the cost of using the `MutableCallSite` - first, we would produce a temporary MethodHandle, and on the first non-`null` call (when the class apparently must be initialized), we would produce the real (possibly optimized) MethodHandle.
>> 
>> I suspect (although I may be wrong) that the eager systems may not be able to work with such a call site, but those should be able to replace the bootstrap with a custom static version without changing the semantics in any way.
>> 
>> A slightly bigger problem is `typeSwitch`: a new feature is that enum constants can be part of any pattern matching switch. Like:
>> 
>> 
>> enum E {A}
>> ...
>> Object o = ...;
>> switch (o) {
>>     case E.A -> ...
>>     ...
>> }
>> 
>> 
>> For this, even `typeSwitch` permits enum constants now, and `java.lang.invoke.ConstantBootstraps.enumConstant` is used for that currently. We would probably need a wrapper to describe the enum constant that could be used by `typeSwitch` to avoid initialization (+the `MutableCallSite`, of course).
>
> Note that currently, a `switch` over an enum will already trigger initialisation of the enum class because of the need to compute the relevant mapping array by the synthetic helper class, see [JDK‑7176515] for details.
> - https://github.com/openjdk/jdk/pull/10797
> 
> [JDK‑7176515]: https://bugs.openjdk.org/browse/JDK-7176515 "[JDK‑7176515] ExceptionInInitializerError for an enum with multiple switch statements"

FWIW, a sketch on how avoiding the enum initialization might look like is here:
https://github.com/lahodaj/jdk/compare/JDK-8300543...lahodaj:jdk:JDK-8300543-lazy-enum?expand=1

more work needed to make that work for `typeSwitch`, and to combine that with #9779.

-------------

PR Review Comment: https://git.openjdk.org/jdk/pull/13074#discussion_r1165427916