review request (M): 6912064: type profiles need to be more thorough for dynamic language support
Vladimir Kozlov
vladimir.kozlov at oracle.com
Fri Jul 16 09:58:05 PDT 2010
Actually it was not last mail, I replied:
Vladimir Kozlov wrote:
> I see, I missed that for instanceof case you just compare object's klass
> with profiled klass and don't check that it is subtype of require_klass
> (which is NULL for this case). Fine.
About last changes. Why you not do 'stopped()' check after that call?:
2668 cast_obj = maybe_cast_profiled_receiver(not_null_obj, data, tk->klass());
Vladimir
John Rose wrote:
> I got a pre-review for 6912064 in January. These changes still work well; I just used them successfully with a performance tuning experiment.
>
> There was a corner case bug which required an additional 'stopped()' check in graphKit. Here's the update:
> http://cr.openjdk.java.net/~jrose/6912064/webrev.02/
>
> The four lines in this file starting "// Profile disagrees with this path." are new:
> http://cr.openjdk.java.net/~jrose/6912064/webrev.02/src/share/vm/opto/graphKit.cpp.udiff.html
>
> That is the only difference from January.
>
> Any further comments on this change set?
>
> -- John
>
> FTR, here is the previous communication. (Yes, it's been a while.)
>
> From: John Rose <John.Rose at Sun.COM>
> Date: January 26, 2010 7:14:13 PM PST
> To: Vladimir Kozlov <Vladimir.Kozlov at Sun.COM>
> Cc: Igor Veresov <Igor.Veresov at Sun.COM>, hotspot-compiler-dev compiler <hotspot-compiler-dev at openjdk.java.net>
> Subject: Re: for pre-review (M): 6912064: type profiles need to be more thorough for dynamic language support
>
> On Jan 26, 2010, at 6:59 PM, Vladimir Kozlov wrote:
>
>> Then why you generate the klass check with uncommon trap for instanceof
>> for cases when profiling shows that it will fail?
>
> For code like this:
>
> if (x instanceof String)
> doString((String) x);
> else if (x instanceof Number)
> doNumber((Number) x);
>
> If the 'x' is always an Integer, I want to fold all the tests up from the first instanceof, not the second.
>
> The profile will show when 'x' is of a monomorphic type. That is more interesting than whether the first check happens to go true or false.
>
> -- John
>
>
More information about the hotspot-compiler-dev
mailing list