[RFC PATCH 0/8] fpu: experimental conversion of float128_addsub

[Alex Bennée]

Hi Richard,

This is the current state of my experiment to convert a 128 bit float
function (in this case addsub). The actual conversion was fairly
simple (basically a copy & paste with some tweaks for using the
unint128 inline functions). However I ran into a number of stumbles
with the int128.h support including casting of values like ~(Uint128)0
and messing around to handle things like missing __builtin support for
clz. I suspect having some of the #defines expand into uint128_*
functions plays some part in the 4x growth in code compared to the old
version. However the drop in performance is a lot less than that.

In terms of total code churn we replace each deleted line in
softfloat.c with 2 lines of new code although I suspect if we pressed
on we could reduce the diffstat deficit. Debugging the actual failures
was relatively painless with rr and the new code - perhaps because I
just find it easier to follow.

I've included your early patches as that happened to be the state of
my tree when I branched off. If we want to go forward with a more
complete conversion I guess we would need:

  - a more complete int128.h conversion (including fallback for non 
CONFIG_INT128)
  - seeing if some of the resulting code bloat can be reduced
  - seeing what scope there is for commonality of special case handling

I'm not a fan of having so much duplication but at least I personally
find the code is more readable.

Alex Bennée (3):
  int128.h: add bunch of uint128 utility functions (INCOMPLETE)
  tests/fp: add quad support to the benchmark utility
  softfloat: implement addsub_floats128 using Uint128 and new style code

Richard Henderson (5):
  softfloat: Use mulu64 for mul64To128
  softfloat: Use int128.h for some operations
  softfloat: Tidy a * b + inf return
  softfloat: Add float_cmask and constants
  softfloat: Inline pick_nan_muladd into its caller

 include/fpu/softfloat-macros.h |  80 ++--
 include/qemu/int128.h          | 122 ++++++
 fpu/softfloat.c                | 697 ++++++++++++++++++++-------------
 tests/fp/fp-bench.c            |  88 ++++-
 fpu/softfloat-specialize.c.inc |  39 ++
 5 files changed, 711 insertions(+), 315 deletions(-)

-- 
2.20.1