request(S): 7066339: Tiered: policy should make consistent decisions about osr levels (original) (raw)

Igor Veresov igor.veresov at oracle.com
Sun Jul 17 14:32:03 PDT 2011

I modified the fix a little bit. With -Xcomp we had inconsistent behavior with tiered and TieredStopAtLevel. Everything was compiled at level 3 regardless of TieredStopAtLevel flag value. I've added CompilationPolicy::initial_compile_level() to make it policy-dependent.

Webrev: http://cr.openjdk.java.net/~iveresov/7066339/webrev.00/

Sorry for the noise.

igor

On 7/14/11 1:04 PM, Igor Veresov wrote:

Thanks Vladimir and Tom!

igor

On 7/14/11 12:42 PM, Tom Rodriguez wrote: Looks ok.

tom On Jul 13, 2011, at 6:49 PM, Igor Veresov wrote:

After my fix for 7058689 it is possible that the AdvancedThresholdPolicy will make a different decision as to what compilation level for osr is appropriate in loopevent() and callevent() from what happens in methodbackbranchevent(). In methodbackbranchevent() in case the OSR compilation level is CompLevellimitedprofile we change it to CompLevelfullprofile but we don't do the same in callevent() and loopevent() to determine whether the existing OSR methods are current. The solution would be to move this level fixup in form of disabling the tier 3 delay feedback in common() for OSRs.

There was also bad interaction between the in-queue level change optimization with TieredStopAtLevel option set to 3, which caused compilation loops between level 2 and 3. Now it will be disabled if TieredStopAtLevel< 4. Also moved handling of TieredStopAtLevel to common(), because when it was in compile() it could lead to unnecessary OSR compiles with TieredStopAtLevel< 4. Webrev: http://cr.openjdk.java.net/~iveresov/7066339/webrev.00/ Thanks, igor

More information about the hotspot-compiler-dev mailing list