bdiv vs redc

[Torbjorn Granlund]

nisse at lysator.liu.se (Niels Möller) writes:

  > Could you perhaps write a new proposed sbpi1_bdiv_qr with the result
  > normalisation you suggest, using my proposed sbpi1_bdiv_r style?  That
  > would allow for a single outer loop, unlike the current sbpi1_bdiv_qr.
  
  Below. I'm also attaching a larger patch which also updates
  sbpi1_bdiv_q, the dc code, the t-bdiv testcode, and some of the bdiv
  callers (divexact and remove). I haven't touched the mu_bdiv code.
  
I looked at the patch briefly, and it looks good to me.

  I get lots of failures, I guess some but not all are caused by mu_bdiv
  following the old rather than the new convention (I tried setting the
  MU_BDIV thresholds to infinity, and I still have lots of failures.).

Have you checked that e.g. sbpi1_bdiv_qr gets N = QD+R?

-- 
Torbjörn