[vector] RFR: Fix FMA intrinsic

[Vladimir Ivanov]

Haven't pushed that one yet. Any objections to the proposed fix?

Best regards,
Vladimir Ivanov

On 18/05/2018 01:43, Vladimir Ivanov wrote:
> Yeah, spotted that inconsistency when running the tests :-)
> 
> Here's what I ended up with:
>    http://cr.openjdk.java.net/~vlivanov/panama/vector/fma/webrev.01/
> 
> It passes scalar ops in all cases except VectorCast nodes.
> 
> Best regards,
> Vladimir Ivanov
> 
> On 5/17/18 11:45, Lupusoru, Razvan A wrote:
>> The fix for fma looks fine. That said, it seems that you removed all 
>> vector opcodes from VectorNode::opcode. This is problematic because 
>> none of the other intrinsic implementations that call that method are 
>> updated to pass in opc instead of the "vopc".
>>
>> - Create a second version of VectorNode::make that takes in vopc 
>> instead of opc. Update the version that takes in opc to simply call 
>> VectorNode::opcode and then your new version of VectorNode::make
>> - Update all intrinsic implementations to ensure that vopc is passed 
>> in to the second variant.
>> OR
>> - Update all calls to VectorNode::make to pass in opc instead of 
>> sopc/vopc.
>>
>> I prefer the first variant. Thanks!
>>
>> --Razvan
>>
>> -----Original Message-----
>> From: panama-dev [mailto:panama-dev-bounces at openjdk.java.net] On 
>> Behalf Of Vladimir Ivanov
>> Sent: Thursday, May 17, 2018 1:04 AM
>> To: panama-dev at openjdk.java.net
>> Subject: [vector] RFR: Fix FMA intrinsic
>>
>> http://cr.openjdk.java.net/~vlivanov/panama/vector/fma/webrev.00/
>>
>> There are 2 bug being fixed:
>>     (1) 3rd argument of a ternary operation is always null in 
>> intrinsified version;
>>
>>     (2) VectorNode::make() accepts vector ops, but misses 
>> Op_FmaVF/Op_FmaVD in ternary case. Instead of adding them into the 
>> list, I propose to always pass scalar ops into the methods and get rid 
>> of the special cases for vector ops.
>>
>> Thanks!
>>
>> Best regards,
>> Vladimir Ivanov
>>