[PATCH 0/4] target/arm vector improvements

[Richard Henderson]

The first patch has been seen before.

  https://patchwork.ozlabs.org/patch/1115039/

It had a bug and I didn't fix it right away and then forgot.
Fixed now; I had mixed up the operand ordering for aarch32.

The next 3 are something that I noticed while doing other stuff.

In particular, pmull is used heavily during https transfers.
While cloning a repository, the old code peaks at 27% of the
total runtime, as measured by perf top.  The new code does
not quite reach 3% repeating the same clone.

In addition, the new helper functions are in the form that
will be required for the implementation of SVE2.

The comment in patch 2 about ARMv8.4-DIT is perhaps a stretch,
but re-reading the pmull instruction description in the current
ARM ARM brought it to mind.

Since TCG is officially not in the security domain, it's
probably not a bug to just claim to support DIT without
actually doing anything to ensure the algorithms used are in
fact timing independent of the data.

On the other hand, I expect the bit distribution of stuff
going through these sort of hashing algorithms to approach
50% 1's and 0's, so I also don't think we gain anything on
average to terminate the loop early.

Thoughts on DIT specifically?


r~


Richard Henderson (4):
  target/arm: Vectorize USHL and SSHL
  target/arm: Convert PMUL.8 to gvec
  target/arm: Convert PMULL.64 to gvec
  target/arm: Convert PMULL.8 to gvec

 target/arm/helper-sve.h    |   2 +
 target/arm/helper.h        |  21 ++-
 target/arm/translate.h     |   6 +
 target/arm/neon_helper.c   | 117 -------------
 target/arm/translate-a64.c |  83 ++++-----
 target/arm/translate.c     | 350 ++++++++++++++++++++++++++++++++-----
 target/arm/vec_helper.c    | 211 ++++++++++++++++++++++
 7 files changed, 562 insertions(+), 228 deletions(-)

-- 
2.17.1