Request for reviews (S): 6973308: Missing zero length check before repne scas in check_klass_subtype_slow_path()

[Tom Rodriguez]

Looks good.

tom

On Jul 29, 2010, at 4:39 PM, Vladimir Kozlov wrote:

> I updated webrev with additional changes in assembler_x86.cpp
> to verify r12_heapbase even in zero based case.
> These changes exposed this bug.
> 
> Thanks,
> Vladimir
> 
> Vladimir Kozlov wrote:
>> http://cr.openjdk.java.net/~kvn/6973308/webrev
>> Fixed 6973308: Missing zero length check before repne scas in check_klass_subtype_slow_path()
>> Z flag value will not be set by repne_scan(); if RCX == 0 since
>> 'repne' instruction does not change flags (only scas instruction
>> which is repeated sets flags).
>> Setting Z = 1 before repne_scan() in check_klass_subtype_slow_path()
>> will cause exception java.lang.IncompatibleClassChangeError:
>>     // pointer were to sneak in here.  Note that we have already loaded the
>>     // Klass::super_check_offset from the super_klass in the fast path,
>>     // so if there is a null in that register, we are already in the afterlife.
>> +    push(rax);
>> +    xorq(rax,rax); // Set Z = 1
>> +    pop(rax);
>>     repne_scanl();
>> % java -Xmx128M -jar GCBasher.jar -time:300
>> Exception in thread "main" java.lang.IncompatibleClassChangeError
>>        at java.lang.StringCoding$StringDecoder.decode(StringCoding.java:153)
>> Solution:
>> Instead of checking RCX for zero set Z = 0 (not equal) before 'repne'
>> to indicate that class was not found when RCX == 0.
>> Also fixed printed free stack size in hs_err files.
>> Added ability to abort VM only on the exception with specific message.
>> Tested with failed case.