Re: bitop multiplication overflow semantics

[Mike Pall <mikelu-1301@...>]

Cosmin Apreutesei wrote:
> Thanks Mike, in the meantime I've done some quick tests and the
> results are surprising:
> 
> 1) Lua implementation[1] -- 230 MB/s

The speed doubles with this (avoids conversions to/from doubles):

  local tobit = bit.tobit
  ...
  for i=0,nblocks-1 do
    h1 = tobit(mmul(...) + 0xe6546b64)
  end

Some inefficiencies remain wrt. conversions and a missing strength
reduction of the multiply. I've fixed that in git HEAD and I'm now
getting the promised 2 GB/s.

--Mike