Vector bit reversal

[ben-marshall]

In the context of #17 (Vector extension: Carry-less Multiply) - How best to handle needing to reverse bits-in-bytes to get the input in the correct form?

• In Markku's analysis for the scalar implementation of GCM, he used the Bitmanip "grev" instruction to reverse the bits in every byte of the input. I can't see an existing instruction in the V-spec which does this efficiently. Adding an instruction like

  vbrev.vv vrt    // Bit reverse, destructive.
  

  Which reverses the bits in every scalar element based on SEW, and requiring support for SEW=8 for the crypto extension seems reasonable?

• There are other (non-obvious) ways to do bit reversal quickly which look amenable to vectorising, which we can fall back on.

[grnewell]

The best alternative method I have seen so far, adapted for RVV from one of the methods in the link in the main post (this one) , uses masking and shifting and takes 15 base vector extension instructions to reverse the order of the 8 bits in all the bytes of all the elements in a vector (not counting loading and other set-up overhead)

[ben-marshall]

Andy's proposal for vector grev can be found here. These will be transcribed into the draft spec.