clang miscompiles OpenLibm on i686-*-freebsd

[Steve Kargl]

TL;DR summary: clang is broken for numerical on i686 FreeBSD.


% uname -a
FreeBSD mobile 13.0-CURRENT FreeBSD 13.0-CURRENT #0 r361834M:
Fri Jun  5 08:49:26 PDT 2020 obj/usr/src/i386.i386/sys/MOBILE  i386

% which clang
/usr/bin/clang
% clang --version
FreeBSD clang version 10.0.1 (git at github.com:llvm/llvm-project.git llvmorg-10.0.1-rc1-0-gf79cd71e145)
Target: i386-unknown-freebsd13.0
Thread model: posix
InstalledDir: /usr/bin

My testing for exp2f(x) shows

Interval tested for exp2f: [1,8]
       ulp <= 0.5:  0.936%    235635 |   0.936%    235635
0.5 <  ulp <  0.6:  0.016%      4098 |   0.953%    239733
0.6 <  ulp <  0.7:  0.024%      6147 |   0.977%    245880
0.7 <  ulp <  0.8:  0.008%      2049 |   0.985%    247929
0.8 <  ulp <  0.9:  0.016%      4098 |   1.001%    252027
0.9 <  ulp <  1.0:  0.008%      2049 |   1.010%    254076
1.0 <  ulp <  1.5:  0.334%     84009 |   1.343%    338085
1.5 <  ulp <  2.0:  0.187%     47127 |   1.531%    385212
2.0 <  ulp <  3.0:  0.839%    211047 |   2.369%    596259
3.0 <  ulp <  0.0: 97.631%  24569565 | 100.000%  25165824
Max ulp: 359512.000000 at 1.96875286e+00

which is bad (ulp means errors in the Unit of Last Place).

By default OpenLibm uses "-march=i686 -O3 -m32" on an i686 target.
The commandline used in compiling s_exp2f.c is 

   clang -fno-builtin -fno-strict-aliasing -O3 -fPIC -march=i686 -m32
   -std=c99 -Wall -I/usr/home/kargl/tmp/olibm
   -I/usr/home/kargl/tmp/olibm/include -I/usr/home/kargl/tmp/olibm/i387
   -I/usr/home/kargl/tmp/olibm/src -DASSEMBLER -D__BSD_VISIBLE
   -Wno-implicit-function-declaration -I/usr/home/kargl/tmp/olibm/ld80
   -c src/s_exp2f.c -o src/s_exp2f.c.o

If I use -O1 or -O2 I get the same above results.  If I 
change to -O0, I get

   clang -fno-builtin -fno-strict-aliasing -O0 -fPIC -march=i686 -m32
   -std=c99 -Wall -I/usr/home/kargl/tmp/olibm
   -I/usr/home/kargl/tmp/olibm/include -I/usr/home/kargl/tmp/olibm/i387
   -I/usr/home/kargl/tmp/olibm/src -DASSEMBLER -D__BSD_VISIBLE
   -Wno-implicit-function-declaration -I/usr/home/kargl/tmp/olibm/ld80
   -c src/s_exp2f.c -o src/s_exp2f.c.o

Interval tested for exp2f: [1,8]
       ulp <= 0.5:  0.056%     14072 |   0.056%     14072
0.5 <  ulp <  0.6:  0.000%         8 |   0.056%     14080
3.0 <  ulp <  0.0: 99.944%  25151744 | 100.000%  25165824
Max ulp: 22729.386719 at 1.00195301e+00

This is better, but still bad.  If I remove -m32, I get

   clang -fno-builtin -fno-strict-aliasing -O0 -fPIC -march=i686
   -std=c99 -Wall -I/usr/home/kargl/tmp/olibm
   -I/usr/home/kargl/tmp/olibm/include -I/usr/home/kargl/tmp/olibm/i387
   -I/usr/home/kargl/tmp/olibm/src -DASSEMBLER -D__BSD_VISIBLE
   -Wno-implicit-function-declaration -I/usr/home/kargl/tmp/olibm/ld80
   -c src/s_exp2f.c -o src/s_exp2f.c.o

Interval tested for exp2f: [1,8]
       ulp <= 0.5: 99.959%  25155610 |  99.959%  25155610
0.5 <  ulp <  0.6:  0.041%     10214 | 100.000%  25165824
Max ulp: 0.500980 at 1.97115958e+00

This is good, but unoptimized.  If I use -O1, -O2, or -O3 without
-m32, I get the initial bad results above.  If I change -march=i686
to -march=pentiumpro, I get the same results.

Theoretically, OpenLibm would test for SSE capabilities and choose
a better -march, but it doesn't.  This, however, is unimportant.
Any port that uses clang (or cc) that does numerical computation
and uses -march=i686 (or pentiumpro) is likely broken.

-- 
Steve