[aarch64-port-dev ] RFR(S): 8214922: Add vectorization support for fmin/fmax

[Vladimir Ivanov]

Also, it was reported that the assumptions which motivated JDK-8078563 
don't always hold.

There's an RFE to investigate backing it out:
   "C2: Consider enabling auto-vectorization for simple reductions"
   https://bugs.openjdk.java.net/browse/JDK-8188313

Best regards,
Vladimir Ivanov

On 21/12/2018 10:29, Vladimir Ivanov wrote:
> Andrew,
> 
> Haven't checked AArch64, bu there's some logic to enable reduction 
> vectorization only for more complex code shapes (see [1] for some details).
> 
> I was able to trigger auto-vectorization on x86 for the following code 
> shape:
>      int r = 0;
>      for (int i = 0; i < MAX; i++) {
>          r += A[i]*B[i];
>      }
> 
> Best regards,
> Vladimir Ivanov
> 
> [1] https://bugs.openjdk.java.net/browse/JDK-8078563
> 
> On 21/12/2018 09:02, Andrew Haley wrote:
>> Hi Roland,
>>
>> I'm looking at the test code and I can't figure out why trivial 
>> reduction doesn't
>> seem to be working. In fact, I can't even get this to work:
>>
>>      @Benchmark
>>      public float testPlusReduceFloat(BenchmarkState state) {
>>          float[] fa = state.fa;
>>          float sum = 0;
>>          for (int i = 0; i < fa.length; i++) {
>>              sum += fa[i];
>>          }
>>          return sum;
>>      }
>>
>> Should this work? And if not, do you have any idea of what code I 
>> could use to
>> test the SIMD reduction?
>>