Reason for the subpar performance of the lookup validation algorithm in Java [discussion] #10
Description
This Java code can be 'fast' but not nearly at the speed of a C implementation.
I believe that the reason is this code section:
ByteVector byte1High = prev1.lanewise(LSHR, 4).selectFrom(lut.byte1HighLookup());
ByteVector byte1Low = prev1.and((byte) 0x0f).selectFrom(lut.byte1LowLookup());
ByteVector byte2High = input.lanewise(LSHR, 4).selectFrom(lut.byte2HighLookup());
At a glance, selectFrom looks like a standard vectorized table lookup like vpshufb or vtbl. But I think it is not, unfortunately. It appears to generate a long flow of instructions.
So the people working on the Java Vector API appear to have assumed that instructions like vpshufb or vtbl were of secondary importance and did not need to be exposed to the programmer. I believe that it is a mistake.
The C# approach is quite different. In .NET 8, you have Ssse3.Shuffle or AdvSimd.Arm64.VectorTableLookup for example.
See https://mail.openjdk.org/pipermail/panama-dev/2024-June/020476.html