1.88.0 build fails on powerpc64le-linux-gnu with IEEE long double

[doko42]

The 1.88.0 build fails on powerpc64le-linux-gnu with IEEE long double enabled. It still succeeds with with IBM long double. The 1.83.0 build succeeds with both long double configurations.

a failing build log can be found at https://launchpad.net/ubuntu/+source/boost1.88/1.88.0-1ubuntu2/+build/30693987/+files/buildlog_ubuntu-questing-ppc64el.boost1.88_1.88.0-1ubuntu2_BUILDING.txt.gz

however it's non-obvious what exactly is failing ... any hint?

[avithemad]

There seems to be two issues in this bug

1. The build silently fails when charconv library build fails (the logs say that the targets have been skipped, but the exit status from the build process is 1 meaning not successful, and therefore the over build fails). If we run the build with b2 again, we can see the error particularly in the charconv library
2. Redefinition of functions in charconv library. When ieeelongdouble is enabled, the types for long double and __float128 is essentially the same during compilation and we see the following error

/home/avinashj/src/boost_1.88_src/libs/charconv/build/../src/to_chars_float_impl.hpp:849:17: error: redefinition of ‘boost::charconv::to_chars_result boost::charconv::detail::to_chars_float_impl(char*, char*, Real, boost::charconv::chars_format, int) [with Real = long double]’
  849 | to_chars_result to_chars_float_impl(char* first, char* last, __float128 value, chars_format fmt, int precision) noexcept
      |                 ^~~~~~~~~~~~~~~~~~~
/home/avinashj/src/boost_1.88_src/libs/charconv/build/../src/to_chars_float_impl.hpp:776:17: note: ‘boost::charconv::to_chars_result boost::charconv::detail::to_chars_float_impl(char*, char*, Real, boost::charconv::chars_format, int) [with Real = long double]’ previously declared here
  776 | to_chars_result to_chars_float_impl(char* first, char* last, long double value, chars_format fmt, int precision) noexcept
      |                 ^~~~~~~~~~~~~~~~~~~

specifically the macros BOOST_CHARCONV_HAS_QUADMATH and BOOST_CHARCONV_UNSUPPORTED_LONG_DOUBLE are defined and undefined respectively, and therefore both definitions move on for compilation after pre-processing which causes this specific error. Small note: In some architectures like IA-64 __float128 is an alias for long double (more details here: https://gcc.gnu.org/onlinedocs/gcc/Floating-Types.html.), so this error would happen there as well. For example, in x86 platform if -mlong-double-128 compiler option is used (however here error happens at assembly stage rather than at compile stage).

[avithemad]

@doko42 , please let me know in case we need to investigate this further for why both macros are turned on during compilation.

[doko42]

that would be nice for a first successful build

[sdarwin]

The errors mention "charconv", @mborland might know.

[mborland]

@doko42 Can you please give https://github.com/boostorg/charconv/pull/278 a try if you have a chance? I added an extra configuration macro that checks for the number of mantissa bits in a long double on PPC64 which should tell us if you're using IEEE754 ldbl or IBM128 ldbl.

[doko42]

please see https://launchpad.net/~doko/+archive/ubuntu/toolchain/+sourcepub/17487487/+listing-archive-extra it looks that the build is progressing further, but it still fails

[mborland]

Sorry for the noise; the linked PR should be good now. I have local dockers running PPC64LE with IBM128 and IEEE128 and both build and run fine with the updated PR.

[doko42]

it's running longer, now having a 5.7MB compressed build log, but unfortunately again, without any visible error message. please see https://launchpad.net/~doko/+archive/ubuntu/toolchain/+build/31057161

[mborland]

Any chance you could try building boost from source on that platform using either B2 or Cmake? It should give us the error messages.

[doko42]

requested in https://bugs.debian.org/1110432, I'm not the original maintainer

[mborland]

Thanks. In the meantime I pushed another attempt at a fix to the linked PR. Once again it runs fine locally on QEMU ppc64le with Docker, but I have not been able to replicate the original bug.

[avithemad]

@mborland Perhaps you can try adding "-mabi=ieeelongdouble" to the CFLAGS and CXXFLAGS, and try to compile to see this error. I think by default it is just using the ibmlongdouble and so you are not able to reproduce it.

[doko42]

https://launchpad.net/ubuntu/+source/boost1.88/1.88.0-1ubuntu4 builds fine

the applied patch is http://launchpadlibrarian.net/809518464/boost1.88_1.88.0-1ubuntu3_1.88.0-1ubuntu4.diff.gz

[mborland]

I have updated the linked PR to match the applied patch.

[mborland]

Merged for 1.89 release