Do not require AVX when building with SAGE_FAT_BINARY

[vbraun]

Setting CFLAGS to -mno-avx -mno-avx2 -mno-bmi2 should help the produced binaries work on older processors.

Move from sagemath/binary-pkg#31

CC: @culler @dimpase @embray @kiwifb @kliem @orlitzky @mkoeppe @slel @kwankyu

Component: build

Author: Samuel Lelièvre

Branch/Commit: u/soehms/adapt_fat_binary_32434 @ 3c61bf2

Reviewer: Sebastian Oehms, ...

Issue created by migration from https://trac.sagemath.org/ticket/32434

[slel]

comment:1

Thanks for opening this ticket. What files need changing?

Ran git grep SAGE_FAT_BINARY
and git grep fat-binary; still clueless.

Probably a subset of these files:

$ git grep SAGE_FAT_BINARY | cut -f 1 -d : | sort | uniq
build/bin/sage-build-env-config.in
build/pkgs/argon2_cffi/spkg-install.in
build/pkgs/atlas/configuration.py
build/pkgs/atlas/patches/atlas-config
build/pkgs/curl/spkg-install.in
build/pkgs/ecm/SPKG.rst
build/pkgs/ecm/spkg-install.in
build/pkgs/fflas_ffpack/spkg-install.in
build/pkgs/gcc/spkg-configure.m4
build/pkgs/gf2x/spkg-install.in
build/pkgs/givaro/spkg-install.in
build/pkgs/gmp/spkg-install.in
build/pkgs/libffi/spkg-install.in
build/pkgs/linbox/spkg-install.in
build/pkgs/m4ri/spkg-install.in
build/pkgs/mpir/spkg-install.in
build/pkgs/ntl/spkg-install.in
build/pkgs/numpy/spkg-install.in
build/pkgs/openblas/spkg-install.in
build/pkgs/primecount/spkg-install.in
build/pkgs/r/spkg-install.in
configure.ac
docker/Dockerfile
src/doc/en/installation/source.rst
src/sage_setup/command/sage_build_ext.py

[slel]

Description changed:

--- 
+++ 
@@ -1,5 +1,4 @@
 Setting `CFLAGS to -mno-avx -mno-avx2 -mno-bmi2` should help the produced binaries work on older processors.
-
 
 Move from https://github.com/sagemath/binary-pkg/issues/31
 

[orlitzky]

comment:2

Warning: we shouldn't add anything to CFLAGS without first checking that the user's compiler supports it.

IMO a better long-term solution is to ensure that all of our SPKGs respect the user's CFLAGS, CXXFLAGS, et cetera, and don't add unnecessary flags to them on-the-fly. All Gentoo packages do that, so you can likely steal patches/workarounds from our package repository or from the sage-on-gentoo overlay.

[slel]

comment:3

Would this change in src/sage_setup/command/sage_build_ext.py
help in any way?

 def check_flags(self):
     """
     Sanity check the compiler flags used to build the extensions
     """
     forbidden = None
     if os.environ.get("SAGE_FAT_BINARY") == "yes":
         # When building with SAGE_FAT_BINARY=yes, we should not
         # enable CPU features which do not exist on every CPU.
         # Such flags usually come from other libraries adding the
         # flags to the pkgconfig configuration.  So if you hit these
         # errors, the problem is most likely with some external
         # library and not with Sage.
         import re
-        forbidden = re.compile(r"-march=|-mpcu=|-msse3|-msse4|-mpopcnt|-mavx")
+        forbidden = re.compile(r"-march=|-mpcu=|-msse3|-msse4|-mpopcnt"
+                               r"|-mavx|-mavx2|-mbmi2")

     if forbidden is not None:
         errors = 0
         for ext in self.extensions:
             flags = ext.extra_compile_args
             for flag in flags:
                 if forbidden.match(flag):
                     log.error("%s uses forbidden flag '%s'", ext.name, flag)
                     errors += 1
         if errors:
             raise RuntimeError("forbidden flags used")

[orlitzky]

comment:4

Replying to @slel:

Would this change in src/sage_setup/command/sage_build_ext.py
help in any way?

I think it strips those flags when building the sage library (which is the right thing to do in any case), but probably doesn't solve the issue in the original bug report because it's most likely coming from some SPKG and not the sage library itself.

OpenBLAS is a prime suspect, since 0.3.13 will use certain machine-specific instructions when no TARGET is specified: OpenMathLib/OpenBLAS#3056

On sage-support, Nathan said that the mac app has to set TARGET=CORE2 to support macs from 2013. I think to fix the same problem, we need to decide how old of a computer should be able to run the sage binaries, and set a TARGET appropriately when SAGE_FAT_BINARY is used. (Of course, this is all based on the guess that OpenBLAS is to blame. Can anyone reproduce the problem?) If no one else has an opinion, TARGET=CORE2 sounds fine to me. Building from source is always an option.

[fchapoton]

comment:5

bump to 9.6

[williamstein]

comment:7

I would just like this ticket to get prioritized more highly. It does impact users and is probably easy to fix. That's all.

[mkoeppe]

comment:8

Won't happen without people who need SAGE_FAT_BINARY doing the work and setting up tests.

[williamstein]

comment:9

It occurs to me that suggestions like use "CFLAGS to -mno-avx -mno-avx2 -mno-bmi2" aren't really in the spirit of "SAGE_FAT_BINARY". They are in the spirit of "make a binary that is very slow but works on a maximum amount of hardware". Maybe we can't have it both ways, but simply disabling AVX, etc. is probably going to result in a much slower Sage binary.