Describe the bug
When running the "Richards" test, a high proportion of cache misses are coming from the inline cache.
This is because the IC goes through a GcRefCell borrow to get entries:
|
pub(crate) entries: GcRefCell<ArrayVec<CacheEntry, PIC_CAPACITY>>, |
When these are borrowed here, its nearly always a cache miss.
We have 2 options.
- As the hot path is through the some branch the mutable borrow isn't even needed, as thats only used when a weakShape has been collected, so most of these calls don't even need a mutable borrow, a shared one should be fine and offer less over head.
A quick fix could be something like this:
pub(crate) fn get(&self, shape: &Shape) -> Option<(Shape, Slot)> {
if self.megamorphic.get() { return None; }
let entries = self.entries.borrow(); // shared borrow — no write to borrow flag
let shape_addr = shape.to_addr_usize();
for (i, entry) in entries.iter().enumerate() {
if let Some(upgraded) = entry.shape.upgrade() {
if upgraded.to_addr_usize() == shape_addr {
return Some((upgraded, entry.slot));
}
} else {
// stale entry found — drop shared borrow, take exclusive, clean up
drop(entries);
let mut entries = self.entries.borrow_mut();
entries.swap_remove(i);
return None; // shape didn't match anyway (it was stale)
}
}
None
}2: Do we need to wrap in GcRefCell at all? CacheEntry only holds a weakShape, so probably doesn't require the wrapping in GcRefCell as by its very nature a weakShape wouldn't keep the value alive anyway. Could we just swap this to a Cell and have this still work?
Describe the bug
When running the "Richards" test, a high proportion of cache misses are coming from the inline cache.
This is because the IC goes through a GcRefCell borrow to get entries:
boa/core/engine/src/vm/inline_cache/mod.rs
Line 34 in 1af45b2
When these are borrowed here, its nearly always a cache miss.
We have 2 options.
A quick fix could be something like this:
2: Do we need to wrap in GcRefCell at all? CacheEntry only holds a weakShape, so probably doesn't require the wrapping in GcRefCell as by its very nature a weakShape wouldn't keep the value alive anyway. Could we just swap this to a
Celland have this still work?