On 12/06/2012 01:12 PM, Vitaly Davidovich wrote:
Understood - I'm just trying to make sure I don't have the wrong mental
model of this in my mind. Taking pathology aside (e.g. code cache is full
when time to recompile, poor tiering interaction, etc), I'd expect the fast
inlined path to be restored assuming call site type profile (of the ones we
care about) doesn't change after other subclasses are loaded. Would be
good if someone can correct that if it's inaccurate. Apologies if this is
slightly hijacking the thread, but it's Remi's fault :).
:)
Usually, it's inline just fine because the code just stores the
ThreadLocal in a static field, the VM profiles the code and you get the
fast path.
But some codes use a ThreadLocal in a non static field or use a map of
ThreadLocal so the profile becomes bloated, and there is no inlining
anymore.
If all loaded codes always use the same ThreadLocal (using
ThreadLocal.withSupplier()) then, the call to the supplier is not
inlined but you don't care because it's not part of the fast path of get.
cheers,
Rémi
Sent from my phone
On Dec 6, 2012 7:03 AM, "Doug Lea" <d...@cs.oswego.edu> wrote:
On 12/06/12 06:56, Vitaly Davidovich wrote:
Doug,
When you see the fast to slow ThreadLocal transition due to class loading
invalidating inlined get(), do you not then see it get restored back to
fast
mode since the receiver type in your call sites is still the monomorphic
ThreadLocal (and not the unrelated subclasses)? Just trying to understand
what
Rémi and you are saying.
The possible outcomes are fairly non-deterministic, depending
on hotspot's mood about recompiles, tiered-compile interactions,
method size, Amddahl's law interactions, phase of moon, etc.
(In j.u.c, we have learned that our users appreciate things
being predictably fast enough rather than being
unpredictably sometimes even faster but often slower.
So when we see such cases, as with ThreadLocal, they get added
to todo list.)
-Doug
