ZJIT: Clear jit entry from iseqs after TracePoint activation #14407
Conversation
st0012
force-pushed
the
zjit-disable-when-tracepoint-is-enabled
branch
from
4a292de to
6a4b985
Compare
zjit/src/codegen.rs
Outdated
| // Do not test the JIT code in HIR tests | ||
| if cfg!(test) { | ||
| return std::ptr::null(); | ||
| } | ||
|
|
||
| // Check if c_call or c_return tracing is enabled - if so, don't compile | ||
| if unsafe { rb_c_method_tracing_currently_enabled(ec) } { |
There was a problem hiding this comment.
Why check that here? If you're trying to mimic what YJIT does, it should be a side exit on HIR instructions for C method calls (or dynamic dispatch that can be C method calls). We generally don't want to fail the compilation (return null()).
It only impacts JIT entries after TracePoint is enabled, which would miss returning from a non-leaf call to JIT code since we have no such patch points yet, but I guess it's fine to start from supporting just that. |
st0012
force-pushed
the
zjit-disable-when-tracepoint-is-enabled
branch
3 times, most recently
from
82465f8 to
328ef59
Compare
st0012
changed the title
st0012
force-pushed
the
zjit-disable-when-tracepoint-is-enabled
branch
from
328ef59 to
07eb3b6
Compare
Thanks for the pointer. I tried to add a new patch point for that in #14420 but it's currently not working 😞 I'll log more info about it there |
zjit/src/gc.rs
Outdated
| @@ -222,3 +222,14 @@ pub fn remove_gc_offsets(payload_ptr: *mut IseqPayload, removed_range: &Range<Co | |||
| fn ranges_overlap<T>(left: &Range<T>, right: &Range<T>) -> bool where T: PartialOrd { | |||
| left.start < right.end && right.start < left.end | |||
| } | |||
|
|
|||
| /// Iterate over all existing ISEQs | |||
| pub fn for_each_iseq<F: FnMut(IseqPtr)>(mut callback: F) { | |||
There was a problem hiding this comment.
How about putting this in cruby.rs? Unless gc.rs uses it by itself, it has nothing to do with GC.
st0012
force-pushed
the
zjit-disable-when-tracepoint-is-enabled
branch
from
07eb3b6 to
2a5a9bf
Compare
This will allow TracePoint events to be triggered when calling previously JITed methods from the interpreter.
I also opened #14420 to find ways to add a new patch point for TracePoint but it's not working atm.