numerical differences in x87 fpu code generated by clang-14 and clang-15

[anbe42]

As I've reported in Debian: https://bugs.debian.org/1031500

I've extracted a function (_Z8_cl_fmodff) from pocl generated bytecode (pocl_cl_fmod.ll, part of the pocl OpenCL implementation of float fmod(float,float)) that produces different numeric results when compiled with lvm-14 and llvm-15. I've added a small C main to demonstrate that (main.c, Makefile).
These numerical differences from pocl-3.1/llvm-15 cause test failures in pyopencl, libgpuarray.

This happens when generating 32-bit x87 fpu code with -O2 --target=i386-unknown-linux-gnu -march=i686.

$ ./test_pocl_cl_fmod_llvm14
a=0 b=0.1 c=0 d=0 abserr=0 relerr=0
a=0.1 b=0.1221239 c=0.1 d=0.1 abserr=0 relerr=0
a=0.2 b=0.14424779 c=0.055752218 d=0.055752218 abserr=0 relerr=0
a=0.3 b=0.16637169 c=0.13362832 d=0.13362832 abserr=0 relerr=0
a=0.4 b=0.18849558 c=0.023008853 d=0.023008853 abserr=0 relerr=0
a=0.5 b=0.21061946 c=0.078761071 d=0.078761071 abserr=0 relerr=0
a=0.6 b=0.23274337 c=0.13451329 d=0.13451329 abserr=0 relerr=0
a=0.7 b=0.25486726 c=0.19026548 d=0.19026548 abserr=0 relerr=0
a=0.8 b=0.27699116 c=0.24601769 d=0.24601769 abserr=0 relerr=0
a=0.9 b=0.29911503 c=0.0026548803 d=0.0026548803 abserr=0 relerr=0

$ ./test_pocl_cl_fmod_llvm15
a=0 b=0.1 c=0 d=0 abserr=0 relerr=0
a=0.1 b=0.1221239 c=0.1 d=0.1 abserr=0 relerr=0
a=0.2 b=0.14424779 c=0.055752218 d=0.055752218 abserr=0 relerr=0
a=0.3 b=0.16637169 c=0.13362832 d=0.13362832 abserr=0 relerr=0
a=0.4 b=0.18849558 c=0.023008853 d=0.023008853 abserr=0 relerr=0
a=0.5 b=0.21061946 c=0.078761071 d=0.078761071 abserr=0 relerr=0
a=0.6 b=0.23274337 c=0.13451327 d=0.13451329 abserr=1.49012e-08 relerr=6.4024e-08
a=0.7 b=0.25486726 c=0.19026548 d=0.19026548 abserr=0 relerr=0
a=0.8 b=0.27699116 c=0.24601772 d=0.24601769 abserr=-2.98023e-08 relerr=-1.07593e-07 FAIL
a=0.9 b=0.29911503 c=0.0026548505 d=0.0026548803 abserr=2.98023e-08 relerr=9.9635e-08

c = _Z8_cl_fmodff(a, b)  # from pocl
d = fmodf(a, b)          # from libc, as reference

Please get the files from https://bugs.debian.org/1031500, GitHub refused any attachment type I tried.

[llvmbot]

@llvm/issue-subscribers-backend-x86

[DimitryAndric]

This regressed (or was maybe exposed) in 0af3e6a ("[InstCombine] Sink instructions with multiple users in a successor block") by @wweiandrew (cc @nikic @preames):

$ ./pr60837-1-llvmorg-15-init-04973-gc236b41e451
a=0 b=0.1 c=0 d=0 abserr=0 relerr=0
a=0.1 b=0.1221239 c=0.1 d=0.1 abserr=0 relerr=0
a=0.2 b=0.14424779 c=0.055752218 d=0.055752218 abserr=0 relerr=0
a=0.3 b=0.16637169 c=0.13362832 d=0.13362832 abserr=0 relerr=0
a=0.4 b=0.18849558 c=0.023008853 d=0.023008853 abserr=0 relerr=0
a=0.5 b=0.21061946 c=0.078761071 d=0.078761071 abserr=0 relerr=0
a=0.6 b=0.23274337 c=0.13451329 d=0.13451329 abserr=0 relerr=0
a=0.7 b=0.25486726 c=0.19026548 d=0.19026548 abserr=0 relerr=0
a=0.8 b=0.27699116 c=0.24601769 d=0.24601769 abserr=0 relerr=0
a=0.9 b=0.29911503 c=0.0026548803 d=0.0026548803 abserr=0 relerr=0

$ ./pr60837-1-llvmorg-15-init-04974-g0af3e6a22da
a=0 b=0.1 c=0 d=0 abserr=0 relerr=0
a=0.1 b=0.1221239 c=0.1 d=0.1 abserr=0 relerr=0
a=0.2 b=0.14424779 c=0.055752218 d=0.055752218 abserr=0 relerr=0
a=0.3 b=0.16637169 c=0.13362832 d=0.13362832 abserr=0 relerr=0
a=0.4 b=0.18849558 c=0.023008853 d=0.023008853 abserr=0 relerr=0
a=0.5 b=0.21061946 c=0.078761071 d=0.078761071 abserr=0 relerr=0
a=0.6 b=0.23274337 c=0.13451327 d=0.13451329 abserr=1.49012e-08 relerr=6.4024e-08
a=0.7 b=0.25486726 c=0.19026548 d=0.19026548 abserr=0 relerr=0
a=0.8 b=0.27699116 c=0.24601772 d=0.24601769 abserr=-2.98023e-08 relerr=-1.07593e-07 FAIL

Resulting assembly for fmod.ll, compiled with -m32 -O2, using llvmorg-15-init-04973-gc236b41e451 and llvmorg-15-init-04974-g0af3e6a22da:

fmod-llvmorg-15-init-04973-gc236b41e451.s.txt
fmod-llvmorg-15-init-04974-g0af3e6a22da.s.txt

[efriedma-quic]

This looks like the general issue of "we don't set the precision correctly for x87 math"? As previously reported in #44218 #42470 #40569 #13862