gcc 9.3.0 -fno-plt -fuse-ld=gold results into bad instructions on ppc64le

[nehaljwani]

Steps to reproduce (on a powerpc box):

# conda create -yp ./dev gxx_linux-ppc64le -c conda-forge
# conda activate ./dev
# cat <<EOF > wth.cpp
#include <iostream>
int main() {
   return 0;
}
EOF
# $CXX wth.cpp -g -fno-plt -fuse-ld=gold
./a.out
Illegal instruction

Another pruned example (from mongodb build system):

#include <cstring>
#include <cstdint>
#include <cstdlib>

int main() {
    using Native = __vector signed char;
    const size_t size = sizeof(Native);
    const Native bits = { 120, 112, 104, 96, 88, 80, 72, 64, 56, 48, 40, 32, 24, 16, 8, 0 };

    uint8_t inputBuf[size];
    std::memset(inputBuf, 0xFF, sizeof(inputBuf));

    return 0;
}

This is also reproducible using the cross compilers + qemu-static + binfmt_misc + QEMU_LD_PREFIX. The error is simply:

qemu: uncaught target signal 4 (Illegal instruction) - core dumped
[1]    641165 illegal hardware instruction (core dumped)  ./a.out

Not sure if this has been reported upstream, as I am able to reproduce this only with this combo: -fno-plt + -fuse-ld=gold. If I remove either, the problem goes away.

I saw this problem happen with mongodb and mysql. @hmaarrfk found that downgrading the compiler version to 8 resolves the issue, so I ended up doing that on both these feedstocks.

cc @isuruf @beckermr

[nehaljwani]

I'm planning to report it upstream, waiting on BZ account creation at the moment.

@mingwandroid have you ever come across this problem before?

[mingwandroid]

I haven't come across it, no. I am working on compilers at the moment though. I am looking at 10.2 as the base-line for GCC.

[xhochy]

@nehaljwani did you ever get to report this upstream?

[nehaljwani]

At the time when I hit the problem, I couldn't reproduce it with a compiler built from source (not using the recipe), so couldn't report it upstream.

[hmaarrfk]

I hit this conda-forge/fish-feedstock#29 here too.