[LA64_DYNAREC] Speed up crc validation

[ksco]

No description provided.

[ksco]

This is only slightly slower than the exact match method, which is good enough.

[ptitSeb]

Well, the exact match will always be faster in case of mismatch...
But how this new one is faster? is it just loop unrolling?

[ksco]

No, it's more than that. Every iteration of the original 1-way CRC loop has a data dependency on the previous iteration, and CRC is quite slow. A multi-way CRC loop will cancel most of those dependencies.

[ptitSeb]

Oh, I see. The point is to avoid starving the pipeline with dependant load+crc pairs

[ksco]

For 12 elements, it's like

1 -> 2 -> 3 -> 4 -> 5 -> 6 -> 7 -> 8 -> 9 -> 10 -> 11 -> 12

v.s.

1 -> 7
  2 -> 8
    3 -> 9
      4 -> 10
        5 -> 11
          6 -> 12

[ksco]

I tested 4-way, 6-way, 8-way, 6 way seems to be the best, but I guess that depends on the hardware.