1483871 - Consider adding an optimized JIT entry for Function.prototype.call/apply

Status: RESOLVED WONTFIX

(Core :: JavaScript Engine, enhancement, P3)

Description

[Jan de Mooij [:jandem]]

A bit like bug 1483869: I think we could add an "optimized JIT entry" (a feature added for Wasm natives) to the fun_call/fun_apply JSNatives.

This JIT entry point would just be a trampoline where we call the target function if it's scripted, or call into C++ if it's not. Once we have that, we can remove the Baseline ICs for call/apply and the similar Ion call optimizations.

Comment 1

[Ted Campbell [:tcampbell]]

If we go this route can we please wrap the pointers in a new data structure and make this a little more explicit? There is already enough hidden complexity with these "optimized JIT entries" having surprising coupling to the GC relocation mechanism.

Comment 2

[Matthew Gaudet (he/him) [:mgaudet]]

Iain... I feel like we have something like this these days via CacheIR. Close as WORKSFORME?

Comment 3

[Iain Ireland [:iain]]

I think this bug was opened before we had CacheIR-based call ICs in baseline.

We have a variety of different kinds of support in CacheIR now. Notably, using specialized CacheIR, we can inline through fun_call and some versions of fun_apply. The trampoline wouldn't help with that. We also have some support for fun_call in the Ion CallGeneric trampoline.

In general, I don't think generating an IC that calls a trampoline that calls the target function is an improvement over generating a fun_call/fun_apply-specific IC that calls the target function directly.

It seems unlikely that changes in this direction would make things faster. They might simplify things slightly, but our existing complexity seems reasonably self-contained. I think we can close this.