Bug 148914 - GCAwareJITStubRoutines should be GC cells
Summary: GCAwareJITStubRoutines should be GC cells
Status: NEW
Alias: None
Product: WebKit
Classification: Unclassified
Component: JavaScriptCore (show other bugs)
Version: Other
Hardware: All All
: P2 Normal
Assignee: Nobody
URL:
Keywords:
Depends on: 148915
Blocks:
  Show dependency treegraph
 
Reported: 2015-09-06 13:01 PDT by Filip Pizlo
Modified: 2015-09-08 11:29 PDT (History)
0 users

See Also:


Attachments

Note You need to log in before you can comment on or make changes to this bug.
Description Filip Pizlo 2015-09-06 13:01:57 PDT
This would allow us to say that one stub routine requires another stub routine to be alive.
Comment 1 Filip Pizlo 2015-09-08 11:29:42 PDT
This could also be addressed by having stub routines hold RefPtr's to each other.  For example, in the case of a getter, we have the following objects:

1) PolymorphicAccess
2) AccessCase
3) CallLinkInfo
4) JITStubRoutine for PolymorphicAccess
5) JITStubRoutine for CallLinkInfo

Currently we would have the following ownership/ref rules:

(1) owns (2)
(2) owns (3)
(1) refs (4)
(3) refs (5)

Everything would be fine if we also added:

(5) refs (4)

This feels like it might be a bit ad-hoc and certainly not as flexible if these were GC cells.  But, it would work!

In particular:
- If (5) gets deleted, then (3)->(5) must have been cleared first.
- If (4) gets deleted, then (1)->(4) and (5)->(4) must have been cleared first.
- If the owning stub gets destroyed, then everyone gets deleted.

Basically, there is no case where things leak and there is no case where a dangling pointer is left behind.

This could be implemented by giving CallLinkInfo a pointer to its owning JIT stub routine, or more generally, its "JIT owner".  This would mean that we'd also have:

(3) refs (4).

Then, when the CallLinkInfo needs to spawn a stub routine, it can ensure that this new stub routine refs the stub that owned the CallLinkInfo.