RFR (S) 8073479: Replace obj.getClass hacks with Objects.requireNonNull

[Vitaly Davidovich]

So I'm in agreement with you as a whole.  The profile pollution issue is a
bigger concern than just requireNonNull; it means the generally useful (and
seemingly benign) practice of refactoring common code into a single place
can lead to performance regressions.  I do hope that's rectified somehow.

sent from my phone
On Feb 20, 2015 7:27 AM, "Aleksey Shipilev" <aleksey.shipilev at oracle.com>
wrote:

> On 02/19/2015 09:36 PM, Vitaly Davidovich wrote:
> > The problem is microbenchmarking this may not tell the true story (like
> > all microbenchmarks).  In this case the explicit test and branch is
> > probably perfectly predicted because there's not much code.  In a bigger
> > application, where the history fell out of the buffer, it'll depend on
> > what static prediction is used.  Or, this branch will evict another
> > useful one from the history buffer.  Not saying you're wrong in your
> > suggestion to go ahead with this change, but something to keep in mind.
>
> Noted, however the argument can be made that:
>  a) that the same thing applies to getClass()-style checks for public
> APIs. Once anybody calls the API method with nulls, the getClass() check
> would inflate into explicit branch, and then you face the same branch
> predictor behavior;
>  b) the branch misprediction for otherwise stable branch is by
> construction transient, and CPUs should re-learn quickly if we keep on
> calling the method. If we don't, then the cost of branch mispredict is
> not important to begin with. This should be taken with a grain of salt,
> because only a selected few (definitely not me, quite probably no one
> except HW engineers) know how exactly branch predictors work.
>
> That said, the reasonable way you can address these (micro)things in
> compiler is to intrinsify Object.requireNonNull to resist the profile
> pollution and emit the implicit null check unconditionally. But I submit
> that class library should not care about low-probability mispredict, and
> do the thing that clearly declares intent. The performance cost of
> profile pollution, is negligible for our cases.
>
> -Aleksey.
>
>
>