RFR: 8299748: java/util/zip/Deinflate.java failing on s390x

Amit Kumar duke at openjdk.org
Thu Feb 2 08:28:31 UTC 2023


On Wed, 1 Feb 2023 08:46:12 GMT, Alan Bateman <alanb at openjdk.org> wrote:

>> Hi Alan, here is output: 
>> 
>> level:0, strategy: 2, dowrap: true                                                                                                
>> is finished: true                                                
>> is finished: true
>> is finished: true
>> is finished: true                                                                                                                 is finished: true
>> level:1, strategy: 0, dowrap: false               
>> is finished: false     
>> m=525312, n=497972, len=524288, eq=false                                                                                          STDERR:    
>> java.lang.RuntimeException: De/inflater failed:java.util.zip.Deflater at 23a0858b                                               
>> 
>> 
>> 
>> But I guess this behaviour could be explained (by zEDC). On s390x, there is additional instruction present and it could be enabled by setting `DFLTCC=1`. And for accelerator the size of compressed data could go`2 times` the size of actual data. Again this is not deterministic as well, i.e. for same data there could be different valid deflate stream.
>
>> level:1, strategy: 0, dowrap: false               
>> is finished: false     
> 
> Thanks for checking that. So "is finished: false" is telling us that not all of the input was compressed. So I think the right thing is to do the deflate in a loop until there is more input to compress, the inflate probably needs the same. Your original proposal was to make the output buffer larger and I suspect that is just working around a threshold or block size used by the accelerator.

Hi @AlanBateman ,
with latest changes (doing inflate/deinflate) test passes over s390 and x86 as well. Please take a look now.

-------------

PR: https://git.openjdk.org/jdk/pull/12283


More information about the core-libs-dev mailing list