New release portability issue

[Dennis Clarke]

--------------------- apology ---------------
Sorry for the resend, myself and Thunderbird are having problems also.
---------------------------------------------
|
| Initially sent 19 Feb 2025 :
|
|     On an old Fujitsu SPARC64 server running Solaris 10 there is
| no builtin compiler function for __builtin_popcount. This seems to
| be something new in libcrux_mlkem768_sha3.h header. There was no
| issue with openssh-9.8p1.
|
|     The LLVM folks have neat little bit of code that can do the bit
| shifting and adding :
|
|
| 
https://github.com/llvm-mirror/compiler-rt/blob/master/lib/builtins/popcountdi2.c
|
|    It would be better to use inline asm where the SPARC64 hardware
| does indeed have the popc instruction.
---------------------------------------------
|
| There was a reply from Damien Miller 20 Feb 2025
|
|
| This is weird, because this code has not changed significantly
| since 9.8p1. It's possible that configure might be behaving
| differently, could you try comparing the output of ./configure
| run on each version?
|
================= once more from the top ================

Thank you for the reply and that is a good idea. I shall do that this
morning once the coffee machine begins to work.

In the meanwhile there is the problem with libcrux_mlkem768_sha3.h
which is new. It did not exist in 9.8p1 at all. Neither did the need
for that GCC builtin function.

I have the two versions code extracted here :

sparc64$ ls -lad openssh-9.8p1 openssh-9.9p2
drwxr-xr-x   7 1000     1000         331 Jul  1  2024 openssh-9.8p1
drwxr-xr-x   7 1000     1000         334 Feb 18 08:15 openssh-9.9p2

That new header does not exist in 9.9p2 :

sparc64$ grep -n '__builtin_popcount' openssh-9.8p1/libcrux_mlkem768_sha3.h
grep: openssh-9.8p1/libcrux_mlkem768_sha3.h: No such file or directory
sparc64$

The need for that GCC builting does not exist in 9.8p1 :

sparc64$ find ./openssh-9.8p1 -type f | xargs grep '__builtin_popcount'
sparc64$

It certainly does exist in 9.9p2 :

sparc64$ find ./openssh-9.9p2 -type f | xargs grep '__builtin_popcount'
./openssh-9.9p2/libcrux_mlkem768_sha3.h:  return __builtin_popcount(x0);
sparc64$

That can be a bit of a show stopper for a platform without a modern GCC
and in fact, to be even more of a problem, there is no reasonable GCC
version released in half a decade or more. I may be able to port version
gcc-9.5.0 onto the old Fujitsu SPARC64 server. Maybe. However the vendor
Oracle Studio 12.6 developer tools certainly does not recognize that GCC
specific builtin function.

I can easily write a bit-shift and count routine but it would be orders
of magnitude slower than the popc opcode asm inline. Which is likely
what is needed on SPARC64. For example we have this in the GMP math lib
code :

https://gmplib.org/repo/gmp/file/tip/mpn/sparc64/ultrasparct3/popcount.asm

000001  dnl  SPARC v9 mpn_popcount for T3/T4.
000002
000003  dnl  Contributed to the GNU project by David Miller.
000004
000005  dnl  Copyright 2013 Free Software Foundation, Inc.
000006
000007  dnl  This file is part of the GNU MP Library.
000008  dnl
000009  dnl  The GNU MP Library is free software; you can redistribute 
it and/or modify
000010  dnl  it under the terms of either:
000011  dnl
000012  dnl    * the GNU Lesser General Public License as published by 
the Free
000013  dnl      Software Foundation; either version 3 of the License, 
or (at your
000014  dnl      option) any later version.
000015  dnl
000016  dnl  or
000017  dnl
000018  dnl    * the GNU General Public License as published by the Free 
Software
000019  dnl      Foundation; either version 2 of the License, or (at 
your option) any
000020  dnl      later version.
000021  dnl
000022  dnl  or both in parallel, as here.
000023  dnl
000024  dnl  The GNU MP Library is distributed in the hope that it will 
be useful, but
000025  dnl  WITHOUT ANY WARRANTY; without even the implied warranty of 
MERCHANTABILITY
000026  dnl  or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General 
Public License
000027  dnl  for more details.
000028  dnl
000029  dnl  You should have received copies of the GNU General Public 
License and the
000030  dnl  GNU Lesser General Public License along with the GNU MP 
Library.  If not,
000031  dnl  see https://www.gnu.org/licenses/.
000032
000033  include(`../config.m4')
000034
000035  C                  cycles/limb
000036  C UltraSPARC T3:        15
000037  C UltraSPARC T4:         2.5
000038
000039  C INPUT PARAMETERS
000040  define(`up',   `%o0')
000041  define(`n',    `%o1')
000042  define(`pcnt', `%o5')
000043
000044  ASM_START()
000045          REGISTER(%g2,#scratch)
000046          REGISTER(%g3,#scratch)
000047  PROLOGUE(mpn_popcount)
000048          subcc   n, 1, n
000049          be      L(final_one)
000050           clr    pcnt
000051  L(top):
000052          ldx     [up + 0], %g1
000053          sub     n, 2, n
000054          ldx     [up + 8], %o4
000055          add     up, 16, up
000056          popc    %g1, %g2
000057          popc    %o4, %g3
000058          add     pcnt, %g2, pcnt
000059          brgz    n, L(top)
000060           add    pcnt, %g3, pcnt
000061          brlz,pt n, L(done)
000062           nop
000063  L(final_one):
000064          ldx     [up + 0], %g1
000065          popc    %g1, %g2
000066          add     pcnt, %g2, pcnt
000067  L(done):
000068          retl
000069           mov    pcnt, %o0
000070  EPILOGUE()

However a bit shift and count process in old C90 would get the job done
also. Very very slowly.


-- 
--
Dennis Clarke
RISC-V/SPARC/PPC/ARM/CISC
UNIX and Linux spoken