Bug 5091921: Sign flip issues in loop optimizer
tom.rodriguez at oracle.com
Mon Mar 21 14:06:10 PDT 2011
Yes, there's no indication that 7024475 is related to the sign flip issue. I've been working on other issues for the last week but I'm back looking at this. I can see that the outer loop bound is getting getting reset to its initial value when we start the OSR code. I'm still tracking down why it occurs.
On Mar 21, 2011, at 1:34 PM, Dawid Weiss wrote:
> Ah, clear now. Thanks Vladimir.
> On Mon, Mar 21, 2011 at 6:03 PM, Vladimir Kozlov
> <vladimir.kozlov at oracle.com> wrote:
>> your problem (7024475), it seems, is not related to the sign flip problem.
>> It is a regression after 6384206 changes in 6u14, as Tom pointed.
>> The sign flip problem exists from day one of Hotspot VM.
>> Dawid Weiss wrote:
>>>> since, as easy to reproduce it, it is also easy to avoid it by modifying
>>>> java program.
>>> Can you expand on this, please? While the issue in Lucene could be
>>> worked around by manually unrolling the loop, I don't see a trivial
>>> fix for the case I was hitting, where was a nesting of two loops with
>>> correlated counters.
More information about the hotspot-compiler-dev