YJIT: Introduce no_gc attribute

[k0kubun]

• Introduce no_gc attribute to skip that jit_prepare_routine_call when possible
  • If a method doesn't allocate, we should be able to skip jit_prepare_routine_call.

Kernel#class is one of such methods that are called often on railsbench ruby/ruby-bench#168.

[XrXr]

If leaf, it doesn't raise either. PC should be irrelevant.

It's relevant for allocation tracing, and we've seen that not updating PC can cause SEGV with stackprof. Also, since GC contends for the VM lock, another thread might have triggered global invalidation while it's waiting. In which case we need the patch point recording in jit_prepare_routine_call to exit as soon as possible.

By the way, allocation implies raising because of NoMemoryError

[k0kubun]

It's relevant for allocation tracing, and we've seen that not updating PC can cause SEGV with stackprof. Also, since GC contents for the VM lock, when it gets it another thread might have triggered global invalidation. In which case we need the patch point recording in jit_prepare_routine_call to exit as soon as possible.

Ah, I see. I'll keep the original jit_prepare_routine_call while still eliminating this for no_gc.

By the way, this means that we should fix #7466 too. We paired on this during your vacation, I wrote jit_prepare_routine_call first, but ended up using gen_save_sp because I couldn't explain why we'd need to update PC.

By the way, allocation implies raising because of NoMemoryError

I discussed this on ruby-lang Slack before. Koichi's take was that NoMemoryError virtually doesn't exist on Linux because it gets killed by OOM. Anyway, as long as we'll call jit_prepare_routine_call, this is no longer relevant.

[k0kubun]

Alas, CI notices String#getbyte wasn't actually leaf sometimes. I think we need something like leaf_if_arg_is: Fixnum. Anyway, for this PR, I'll try to find another example.

[k0kubun]

I added this to Kernel#class instead. It's as important as String#getbyte on railsbench.