[10] RFR(S): 8181644: C1 crashes with -XX:+PrintCFGToFile

Vladimir Kozlov vladimir.kozlov at oracle.com
Wed Jun 7 15:44:15 UTC 2017

On 6/7/17 3:21 AM, Tobias Hartmann wrote:
> Hi Vladimir,
> On 06.06.2017 16:55, Vladimir Kozlov wrote:
>> May be use guarantee instead of assert in cfg_printer_output() to avoid parfait complains? Or add NULL check in all callers.
> I think parfait only cares about code available in product builds but a guarantee doesn't hurt here.

guarantee() check is generated in product build.

> I'll push this version:
> http://cr.openjdk.java.net/~thartmann/8181644/webrev.01



> Thanks,
> Tobias
>> On 6/6/17 5:47 AM, Tobias Hartmann wrote:
>>> Hi,
>>> please review the following patch:
>>> https://bugs.openjdk.java.net/browse/JDK-8181644
>>> http://cr.openjdk.java.net/~thartmann/8181644/webrev.00/
>>> C1 crashes in CFGPrinterOutput::print_LIR() when PrintCFGToFile is enabled. The problem is that a single CFGPrinterOutput object is shared between multiple C1 compiler threads leading to inconsistent values of CFGPrinterOutput::_do_print_LIR. One thread sets _do_print_LIR to 'true' while another thread expects it to be 'false' because the LIR is not available in the current phase. We crash while trying to dereference BlockBegin::_lir which is NULL.
>>> I changed the implementation to create a CFGPrinterOutput object per C1 compilation (I had to move the declaration of CFGPrinterOutput from the cpp to the hpp file). Instead of a single "output.cfg" file, now one file per compiler thread is generated ("output_tid[TID]_pid[PID].cfg").
>>> Tested manually with -Xcomp and -XX:+PrintCFGToFile.
>>> Thanks,
>>> Tobias

More information about the hotspot-compiler-dev mailing list