Request for reviews (S): 7026700: regression in 6u24-rev-b23: Crash in C2 compiler in PhaseIdealLoop::build_loop_late_post

[Tom Rodriguez]

On Apr 20, 2011, at 2:51 PM, Vladimir Kozlov wrote:

> Tom Rodriguez wrote:
>> On Apr 20, 2011, at 12:11 PM, Vladimir Kozlov wrote:
>>> This is fix for jdk7.
>>> 
>>> http://cr.openjdk.java.net/~kvn/7026700/webrev
>>> 
>>> Fixed 7026700: regression in 6u24-rev-b23: Crash in C2 compiler in PhaseIdealLoop::build_loop_late_post
>>> 
>>> With EA memory slices should be created explicitly for non-static fields regardless the value of ReduceFieldZeroing flag.
>> Do these really need to guarded?  Couldn't we just always do it?  
> 
> Yes, we can do it always. Which makes raw_mem_only parameter obsolete. Should I remove it?

Sure.  The comment for raw_mem_only is incorrect anyway.

> 
>>> LoadNode::split_through_phi() should not check profitability since we want to eliminate loads from non-escaping allocations.
>> What about this logic?
>>  // Skip the split if the region dominates some control edge of the address.
>>  if (cnt == 3 && !MemNode::all_controls_dominate(address, region))
>>    return NULL;
>> Why should we give up when cnt == 3 and not others?
> 
> Yes, it is a bug, we should check all cases. I think my intention was to restrict the call to expensive all_controls_dominate() only for loops. Fixed.
> 
>>> I also want to disable memory split verification code in EA until the fix for 6984348. It produces false negative results since the verification code does not cover all cases.
>> Could you normal // comments or use #if 0 instead.  It's should have a comment explaining why it's disabled too.
> 
> Done.
> 
> I updated webrev.

Looks good.

tom

> 
> Thanks,
> Vladimir
> 
>> tom