Thanks to visit codestin.com
Credit goes to github.com

Skip to content

YJIT: call alloc func directly in new#13219

Open
jhawthorn wants to merge 2 commits intoruby:masterfrom
jhawthorn:yjit_alloc_func
Open

YJIT: call alloc func directly in new#13219
jhawthorn wants to merge 2 commits intoruby:masterfrom
jhawthorn:yjit_alloc_func

Conversation

@jhawthorn
Copy link
Member

@jhawthorn jhawthorn commented Apr 30, 2025 •
edited
Loading

This removes an indirect branch by calling the allocator function from the class directly.

Benchmark 1: ./miniruby_master --yjit -e '20_000_000.times { Object.new }'
  Time (mean ± σ):      1.203 s ±  0.020 s    [User: 1.187 s, System: 0.011 s]
  Range (min … max):    1.168 s …  1.230 s    10 runs

Benchmark 2: ./miniruby_allocator --yjit -e '20_000_000.times { Object.new }'
  Time (mean ± σ):     967.5 ms ±  22.4 ms    [User: 951.6 ms, System: 11.5 ms]
  Range (min … max):   949.9 ms … 1026.0 ms    10 runs

Summary
  ./miniruby_allocator --yjit -e '20_000_000.times { Object.new }' ran
    1.24 ± 0.04 times faster than ./miniruby_master --yjit -e '20_000_000.times { Object.new }'

cc @tenderlove

@jhawthorn
YJIT: call alloc func directly in new
3d7525a 
This removes an indirect branch by calling the allocator function from
the class directly.
@matzbot matzbot requested a review from a team April 30, 2025 23:21
k0kubun
k0kubun reviewed Apr 30, 2025
View reviewed changes
yjit/src/codegen.rs

let allocator = unsafe { rb_get_alloc_func(comptime_recv) };
let allocator = allocator.unwrap_or(rb_obj_alloc);
let obj = asm.ccall(allocator as _, vec![comptime_recv.into()]);
Copy link
Member

@k0kubun k0kubun Apr 30, 2025 •
edited
Loading

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

When an allocator function is found, it seems to skip the if (rb_obj_class(obj) != rb_class_real(klass)) { rb_raise(rb_eTypeError, "wrong instance allocation"); } check in class_call_alloc_func. Is the error unreachable?

Copy link
Member

@tenderlove tenderlove Apr 30, 2025 •
edited
Loading

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Wouldn't this only be a problem for T_DATA objects? I don't think they use rb_class_allocate_instance as their allocator.

Copy link
Member Author

@jhawthorn jhawthorn Apr 30, 2025 •
edited
Loading

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I think it's only reachable from a custom allocator which is implemented incorrectly (there's no reason to ever return an object with the wrong class), so I'm not sure if it's necessary to keep.

If it is I suppose we could allowlist just the most common builtin allocators (or even just the T_OBJECT allocator).

Copy link
Member

@k0kubun k0kubun Apr 30, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I honestly didn't know when it happens (it'd be nice to have a comment there) and thus asked a question. If the check is useless (on opt_new's rb_obj_alloc call) on the interpreter too, it's fine.

@launchable-app

This comment has been minimized.

Sign in to view
@tekknolagi
Copy link
Contributor

tekknolagi commented Oct 22, 2025 •
edited
Loading

I think we can revive this if you are still interested. We do something very similar in ZJIT after @byroot's shapes changes

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

@tenderlove tenderlove tenderlove left review comments

@k0kubun k0kubun k0kubun left review comments

Assignees

No one assigned

Labels

None yet

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

4 participants

@jhawthorn @tekknolagi @tenderlove @k0kubun