RFR(XS):8216580:X86: Fix generation of VNNI vector code by allowing adjacent LoadS nodes to be isomorphic

Tue Mar 5 00:34:36 UTC 2019

Hi Vladimir

I have tested the patch with compiler tests on VNNI h/w and it passed.
While doing tests in jdk, I noticed that the checks should be guarded against NULL.
So I have added those checks:
if(s1_ctrl != NULL && s2_ctrl != NULL) { ...  
The webrev is here:
http://cr.openjdk.java.net/~vdeshpande/8216580/webrev.03/
I have also rebased the patch on jdk/jdk.

Regards,
Vivek

-----Original Message-----
From: Vladimir Kozlov [mailto:vladimir.kozlov at oracle.com] 
Sent: Friday, March 1, 2019 1:47 PM
To: Deshpande, Vivek R <vivek.r.deshpande at intel.com>; 'Tobias Hartmann' <tobias.hartmann at oracle.com>; 'hotspot-compiler-dev at openjdk.java.net compiler' <hotspot-compiler-dev at openjdk.java.net>
Cc: Viswanathan, Sandhya <sandhya.viswanathan at intel.com>; Raj, Guru <guru.raj at intel.com>
Subject: Re: RFR(XS):8216580:X86: Fix generation of VNNI vector code by allowing adjacent LoadS nodes to be isomorphic

My testing passed. I think you can push after you finish testing.
Please, re-base you changes to jdk/jdk repository before push. I see that webrev.02 was prepared vs jdk/jdk12 which is wrong.

Thanks,
Vladimir

On 3/1/19 11:53 AM, Deshpande, Vivek R wrote:
> Hi Vladimir
> 
> Thanks for the review. I am also working on testing it on the VNNI enabled h/w.
> 
> Regards,
> Vivek
> 
> -----Original Message-----
> From: Vladimir Kozlov [mailto:vladimir.kozlov at oracle.com]
> Sent: Friday, March 1, 2019 10:02 AM
> To: Deshpande, Vivek R <vivek.r.deshpande at intel.com>; 'Tobias 
> Hartmann' <tobias.hartmann at oracle.com>; 
> 'hotspot-compiler-dev at openjdk.java.net compiler' 
> <hotspot-compiler-dev at openjdk.java.net>
> Cc: Viswanathan, Sandhya <sandhya.viswanathan at intel.com>; Raj, Guru 
> <guru.raj at intel.com>
> Subject: Re: RFR(XS):8216580:X86: Fix generation of VNNI vector code 
> by allowing adjacent LoadS nodes to be isomorphic
> 
> This looks good. I assume you did full testing of these new changes on VNNI machine. I will submit testing on what we have.
> 
> Thanks,
> Vladimir
> 
> On 2/28/19 5:23 PM, Deshpande, Vivek R wrote:
>> Hi Vladimir
>>
>> Thanks for your inputs. I have made the changes according to your suggestion.
>> The webrev is here:
>> http://cr.openjdk.java.net/~vdeshpande/8216580/webrev.02/
>> This addresses the questions you had raised.
>> With this patch the checks are applied to all the nodes but returns true only in case of muladds2i.
>>
>> Regards,
>> Vivek
>>
>> -----Original Message-----
>> From: Vladimir Kozlov [mailto:vladimir.kozlov at oracle.com]
>> Sent: Wednesday, February 13, 2019 12:29 PM
>> To: Deshpande, Vivek R <vivek.r.deshpande at intel.com>; 'Tobias 
>> Hartmann' <tobias.hartmann at oracle.com>; 
>> 'hotspot-compiler-dev at openjdk.java.net compiler'
>> <hotspot-compiler-dev at openjdk.java.net>
>> Cc: Viswanathan, Sandhya <sandhya.viswanathan at intel.com>; Raj, Guru 
>> <guru.raj at intel.com>
>> Subject: Re: RFR(XS):8216580:X86: Fix generation of VNNI vector code 
>> by allowing adjacent LoadS nodes to be isomorphic
>>
>> Hi Vivek,
>>
>> Most of new checks are loop invariant: !s1_ctrl_inv and
>> !s1_ctrl->is_RangeCheck()
>>
>> I think you don't need to search for is_muladds2i() if those checks return false.
>>
>> Most general question is: why it should apply only to muladds2i nodes only? Can we do the same for others?
>>
>> Thanks,
>> Vladimir
>>
>> On 2/8/19 2:17 PM, Deshpande, Vivek R wrote:
>>> Hi Vladimir
>>>
>>> Would you please take a look at this patch.
>>>
>>> The Adjacent LoadS have different control RangeCheck node for accesses of type a[2i] and a[2i+1].
>>> This patch allows those nodes to be isomorphic as they belong same counted loop and MulAddS2I nodes.
>>>     
>>> Webrev:
>>> http://cr.openjdk.java.net/~vdeshpande/8216580/webrev.01/
>>> Bug ID:
>>> https://bugs.openjdk.java.net/browse/JDK-8216580
>>>
>>> Regards,
>>> Vivek
>>>
>>> -----Original Message-----
>>> From: Deshpande, Vivek R
>>> Sent: Monday, January 28, 2019 9:45 AM
>>> To: Tobias Hartmann <tobias.hartmann at oracle.com>; 
>>> hotspot-compiler-dev at openjdk.java.net compiler 
>>> <hotspot-compiler-dev at openjdk.java.net>; Vladimir Kozlov 
>>> <vladimir.kozlov at oracle.com>
>>> Cc: Viswanathan, Sandhya <sandhya.viswanathan at intel.com>; Raj, Guru 
>>> <guru.raj at intel.com>
>>> Subject: RE: RFR(XS):8216580:X86: Fix generation of VNNI vector code 
>>> by allowing adjacent LoadS nodes to be isomorphic
>>>
>>> Hi Vladimir
>>>
>>> Would you please take a look at the patch.
>>> The Adjacent LoadS have different control RangeCheck node for accesses of type a[2i] and a[2i+1].
>>> This patch allows those nodes to be isomorphic as they belong same counted loop and MulAddS2I nodes.
>>>     
>>> Webrev:
>>> http://cr.openjdk.java.net/~vdeshpande/8216580/webrev.01/
>>>
>>> Regards,
>>> Vivek
>>>
>>> -----Original Message-----
>>> From: Tobias Hartmann [mailto:tobias.hartmann at oracle.com]
>>> Sent: Tuesday, January 15, 2019 2:57 AM
>>> To: Deshpande, Vivek R <vivek.r.deshpande at intel.com>; 
>>> hotspot-compiler-dev at openjdk.java.net compiler 
>>> <hotspot-compiler-dev at openjdk.java.net>
>>> Cc: Vladimir Kozlov <vladimir.kozlov at oracle.com>; Viswanathan, 
>>> Sandhya <sandhya.viswanathan at intel.com>; Raj, Guru 
>>> <guru.raj at intel.com>
>>> Subject: Re: RFR(XS):8216580:X86: Fix generation of VNNI vector code 
>>> by allowing adjacent LoadS nodes to be isomorphic
>>>
>>> Hi Vivek,
>>>
>>> please add parentheses around the == comparison in lines 1225,1226.
>>>
>>> Otherwise this looks reasonable to me but I'm not too familiar with that code.
>>>
>>> Best regards,
>>> Tobias
>>>
>>> On 12.01.19 01:03, Deshpande, Vivek R wrote:
>>>> Hi Tobias
>>>>
>>>> The webrev for the bug JDK-821650 is here:
>>>> http://cr.openjdk.java.net/~vdeshpande/8216580/webrev.00/
>>>> This fixes generation of vector code by allowing adjacent LoadS nodes to be isomorphic when they have different control RangeCheck nodes for a[i] and a[i+1] accesses in same MulAddS2I node.
>>>> Could you please review it.
>>>>
>>>> Regards,
>>>> Vivek
>>>>
>>>> -----Original Message-----
>>>> From: Deshpande, Vivek R
>>>> Sent: Friday, January 11, 2019 11:38 AM
>>>> To: 'Tobias Hartmann' <tobias.hartmann at oracle.com>; 
>>>> hotspot-compiler-dev at openjdk.java.net compiler 
>>>> <hotspot-compiler-dev at openjdk.java.net>
>>>> Cc: Vladimir Kozlov <vladimir.kozlov at oracle.com>; Viswanathan, 
>>>> Sandhya <sandhya.viswanathan at intel.com>; Raj, Guru 
>>>> <guru.raj at intel.com>
>>>> Subject: RE: RFR(S):8216050:X86: Fix for Superword optimization 
>>>> fails with assert(0 <= i && i < _len) failed: illegal index
>>>>
>>>> Hi Tobias
>>>>
>>>> Thanks for reviewing the patch.
>>>> I have made the changes according to your suggestion.
>>>> In this webrev:
>>>> http://cr.openjdk.java.net/~vdeshpande/8216050/webrev.01/
>>>> I have fix for the crash reported in the 8216050.
>>>>
>>>> The lower cost is needed for generation of vpdpwssd instruction, by combining AddVI and MulAddVS2VI.
>>>> For other instructions pmaddwd and vpmaddwd, they get generated on platforms upto skylake with default cost.
>>>>
>>>> I have updated the bug also with the link to webrev.
>>>>
>>>> I have created a different bug JDK-8216580 for
>>>>     3) Fix generation of vector code by allowing adjacent LoadS nodes to be isomorphic when they have different control RangeCheck nodes
>>>>         for a[i] and a[i+1] accesses in same MulAddS2I node
>>>>
>>>> Thank you.
>>>> Regards,
>>>> Vivek
>>>>
>>>> -----Original Message-----
>>>> From: Tobias Hartmann [mailto:tobias.hartmann at oracle.com]
>>>> Sent: Friday, January 11, 2019 4:49 AM
>>>> To: Deshpande, Vivek R <vivek.r.deshpande at intel.com>; 
>>>> hotspot-compiler-dev at openjdk.java.net compiler 
>>>> <hotspot-compiler-dev at openjdk.java.net>
>>>> Cc: Vladimir Kozlov <vladimir.kozlov at oracle.com>; Viswanathan, 
>>>> Sandhya <sandhya.viswanathan at intel.com>; Raj, Guru 
>>>> <guru.raj at intel.com>
>>>> Subject: Re: RFR(S):8216050:X86: Fix for Superword optimization 
>>>> fails with assert(0 <= i && i < _len) failed: illegal index
>>>>
>>>> Hi Vivek,
>>>>
>>>> On 11.01.19 07:58, Deshpande, Vivek R wrote:
>>>>> 1) Fix for the crash by matching the operand by swapping to right positions.
>>>>
>>>> Looks good but the change to loopopts.cpp:530 screwed up the indentation around the ifs, please fix.
>>>>
>>>>> 2) Cost based generation of vpdpwssd instruction.
>>>>
>>>> Other instructions added by JDK-8214751 still miss a cost definition, for example:
>>>> http://hg.openjdk.java.net/jdk/jdk/rev/4bb6e0871bf7#l5.20
>>>>
>>>>> 3) Fix generation of vector code by allowing adjacent LoadS nodes 
>>>>> to be isomorphic when they have different control RangeCheck nodes
>>>>>        for a[i] and a[i+1] accesses in same MulAddS2I node
>>>>
>>>> This is unrelated to the original bug, right? If so, this should be integrated with a separate RFE.
>>>>
>>>> Thanks,
>>>> Tobias
>>>>