manylinux2014/centos cross-compilation relying on symbols not present on musl systems

[flavorjones]

Please describe the bug

See sparklemotion/nokogiri#2142 for background and conversation about this.

When compiling C code that calls isnan or isinf, code compiled on Debian systems previously implemented this via macros (not function calls). But on the manylinux2014 system, unresolved symbols appear in the shared library for __isinf@@GLIBC_2.2.5 and __isnan@@GLIBC_2.2.5. These can be resolved on Debian linux systems, but not on musl linux systems.

Help us reproduce what you're seeing

Cross-compile nokogiri without the libxml2 patch patches/libxml2/0009-avoid-isnan-isinf.patch, and run the test suite in an alpine image. You can also provide --prevent-strip to nokogiri's extconf.rb to avoid stripping debugging symols.

Specifically, here's what I do to reproduce this:

• add --prevent-strip to the cross-config-options in nokogiri's rakelib/cross-ruby.rake
• delete patches/libxml2/0009-avoid-isnan-isinf.patch
• run rake gem:x86_64-linux to build a gem using manylinux2014 in the pkg/ subdir
• fire up an alpine image via: docker run -it --mount=type=bind,source=$(pwd),target=/nokogiri flavorjones/nokogiri-test:alpine /bin/bash
• inside that alpine container:
  • cd /nokogiri
  • gem install --local pkg/nokogiri*linux*gem
  • rake test or rake test:gdb
  • also run nm /usr/local/bundle/gems/nokogiri-1.11.0.rc4-x86_64-linux/lib/nokogiri/3.0/nokogiri.so to see the unresolved symbols

I can help with this if you need more information.

Expected behavior

I think I expect the centos system to generate the same unresolved symbols as the previous debian system did. If it's possible to set CPP flags to make this work, I'd be happy to set them -- I've tried a bunch of things like -D_ISOC99_SOURCE and -D_XOPEN_SOURCE=600 but wasn't able to change the fact that centos's glibc seems to turn it into a function call under the hood.

[flavorjones]

Updated desc with more details about how to reproduce.

[larskanis]

Thank you @flavorjones for your great analysis and the patch!

I took a look at the libc definitions and found this.

Definition in /usr/include/math.h in Centos-7 is essentially this:

#  define isnan(x) \
     (sizeof (x) == sizeof (float)                                            \
      ? __isnanf (x)                                                          \
      : sizeof (x) == sizeof (double)                                         \
      ? __isnan (x) : __isnanl (x))

And it defines the glibc library reference /usr/include/bits/mathcalls.h kind of:

int __isnan(double) __attribute__ ((__const__));

Definition in /usr/include/math.h in Ubuntu-20.04 makes use of gcc built-in function, so that it doesn't use external references:

#  define isnan(x) __builtin_isnan (x)

In musl it is defined without external function calls

#define isnan(x) ( \
	sizeof(x) == sizeof(float) ? (__FLOAT_BITS(x) & 0x7fffffff) > 0x7f800000 : \
	sizeof(x) == sizeof(double) ? (__DOUBLE_BITS(x) & -1ULL>>1) > 0x7ffULL<<52 : \
	__fpclassifyl(x) == FP_NAN)