[PATCH] D109295: [X86] Handle inverted inputs when matching VPTERNLOG from 2 binary ops.

[Pengfei Wang via Phabricator via llvm-commits]

pengfei added a comment.

In D109295#2985347 <https://reviews.llvm.org/D109295#2985347>, @LuoYuanke wrote:

> In D109295#2985246 <https://reviews.llvm.org/D109295#2985246>, @pengfei wrote:
>
>> I found another example:
>>
>>   define dso_local <4 x i64> @foo2(<4 x i64> %0, <4 x i64> %1, <4 x i64> %2) {
>>     %4 = xor <4 x i64> %2, <i64 -1, i64 -1, i64 -1, i64 -1>
>>     %5 = or <4 x i64> %4, %1
>>     %6 = or <4 x i64> %0, %1
>>     %7 = and <4 x i64> %5, %6
>>     ret <4 x i64> %7
>>   }
>>
>> Can we simply it to below in the approach?
>>
>>   vpor    %ymm1, %ymm0, %ymm0
>>   vpternlogq      $208, %ymm2, %ymm1, %ymm0
>>   retq
>
> Seem no with current vpternlog framework. We currently only support A op1 (B op2 C). Not figured out how to extend the framework to accept more operators as long as there is 3 source bit.

I meant simplified from current generation:

  vpcmpeqd        %ymm3, %ymm3, %ymm3
  vpternlogq      $222, %ymm2, %ymm1, %ymm3
  vpternlogq      $200, %ymm1, %ymm3, %ymm0
  retq

We can save one vpternlogq.


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D109295/new/

https://reviews.llvm.org/D109295