[9] RFR(S): 8178033: C1 crashes with -XX:UseAVX = 3: "not a mov [reg+offs], reg instruction"
Vladimir Kozlov
vladimir.kozlov at oracle.com
Wed Apr 5 16:07:21 UTC 2017
Good.
Thanks,
Vladimir
On 4/4/17 10:25 PM, Tobias Hartmann wrote:
> Thanks Vladimir and Michael!
>
> Here's the new webrev:
> http://cr.openjdk.java.net/~thartmann/8178033/webrev.01/
>
> Best regards,
> Tobias
>
> On 04.04.2017 18:31, Berg, Michael C wrote:
>> I agree to both, for the use of supports_evex() under 0x62 in locate_operand() and in instruction_start() with the new code or at least alternatively UseAVX > 2.
>>
>> Regards,
>> Michael
>>
>> -----Original Message-----
>> From: Vladimir Kozlov [mailto:vladimir.kozlov at oracle.com]
>> Sent: Tuesday, April 04, 2017 9:19 AM
>> To: hotspot-compiler-dev at openjdk.java.net
>> Cc: Berg, Michael C <michael.c.berg at intel.com>
>> Subject: Re: [9] RFR(S): 8178033: C1 crashes with -XX:UseAVX = 3: "not a mov [reg+offs], reg instruction"
>>
>> I think it is better to use VM_Version::supports_evex() in assert. May be change similar assert in Assembler::locate_operand() too.
>> Note, assembler instructions VM_Version::supports_evex() check too.
>>
>> CCing to Michael to review this change.
>>
>> Thanks,
>> Vladimir
>>
>> On 4/4/17 3:38 AM, Tobias Hartmann wrote:
>>> Hi,
>>>
>>> please review the following patch:
>>> https://bugs.openjdk.java.net/browse/JDK-8178033
>>> http://cr.openjdk.java.net/~thartmann/8178033/webrev.00/
>>>
>>> JDK-8076276 [1] added support for AVX512 but NativeMovRegMem::instruction_start() was not updated to recognize/skip the 4-bytes prefix for EVEX instructions. As a result, C1 crashes in NativeMovRegMem::verify() because the instruction_address() points to the EVEX prefix (0x62) which is not a valid MovRegMem instruction.
>>>
>>> Similar to the VEX prefixes, we should skip the corresponding number of bytes such that the instruction address points to the prefixed opcode (see also 'case 0x62' in Assembler::locate_operand()).
>>>
>>> Tested with replay compilation and RBT (running).
>>>
>>> Thanks,
>>> Tobias
>>>
>>> [1] https://bugs.openjdk.java.net/browse/JDK-8076276
>>>
More information about the hotspot-compiler-dev
mailing list