Use macro for implicit summoning in typeclass companion objects for Cats/Algebra/Spire

[ctongfei]

I wrote a macro in this gist. This removes the boilerplate every companion object of typeclasses has (the def apply[A](implicit A: Ev[A]) = A method), and is faster: the method invocation is replaced by the actual typeclass instance filled by the macro.

[ctongfei]

And results in smaller jar size: the apply method no longer needs to be specialized.

[edmundnoble]

Have you seen https://github.com/non/imp? It seems to do most of what you propose, except for the ImplicitSummoner traits. Paging @non, because the reason why we don't use imp in cats is escaping me.

[johnynek]

As @non notes in the issues, the JVM can already optimize this and there is little evidence this helps: https://github.com/DarkDimius/imp-bench

…

On Sat, May 13, 2017 at 20:24 Edmund Noble ***@***.***> wrote: Have you seen https://github.com/non/imp? It seems to do most of what you propose, except for the ImplicitSummoner traits. Paging @non <https://github.com/non>, because the reason why we don't use imp in cats is escaping me. — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#1667 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AAEJdv5_ppAikPJRMgBhK7_ksZb-yUV9ks5r5p4MgaJpZM4NaLX_> .

[edmundnoble]

@johnynek I just, just ran benchmarks to test if the JIT is doing this. If it is, we have to consider that not every piece of code is JITted and it never is on its first runs. JMH is giving us potentially unrealistic results because of this; my benchmarks have CompilerControl.Mode.Exclude which will disallow the JIT. I posted these results at DarkDimius/imp-bench#4:

[info] Benchmark                                 Mode  Cnt   Score   Error  Units
[info] addable.ImplVsImplicitlyAddable.baseline  avgt   30  82.945 ± 2.765  ns/op
[info] addable.ImplVsImplicitlyAddable.explicit  avgt   30  73.834 ± 1.009  ns/op
[info] addable.ImplVsImplicitlyAddable.imply     avgt   30  74.103 ± 0.975  ns/op

[ctongfei]

@edmundnoble I was not aware of the imp project before. But still we could use my ImplicitSummoner traits to reduce some boilerplates.

[diesalbla]

Considering that the use of macros to generate boilerplate, such as done in #2925, it seems unlikely that this will be taken up.

[rossabaker]

I agree, and the complexity only gets worse during the Scala 3 transition.