Compile fails with clang 16

[jmontleon]

I am seeing the following trying to build on Fedora 38 ppc64;e and s390x. I do not have this problem on x86_64 or aarch64.

I believe it is coming from the change in behavior to Wimplicit-function-declaration. https://www.redhat.com/en/blog/new-warnings-and-errors-clang-16

[ 11%] Building C object confu-deps/cpuinfo/CMakeFiles/cpuinfo.dir/src/api.c.o
cd /builddir/build/BUILD/pytorch-v2.0.1.gitf81f9093/redhat-linux-build/confu-deps/cpuinfo && /usr/bin/clang -DCPUINFO_LOG_LEVEL=2 -I/builddir/build/BUILD/pytorch-v2.0.1.gitf81f9093/third_party/cpuinfo/src -I/builddir/build/BUILD/pytorch-v2.0.1.gitf81f9093/third_party/cpuinfo/include -I/builddir/build/BUILD/pytorch-v2.0.1.gitf81f9093/third_party/cpuinfo/deps/clog/include -O2 -flto=thin -fexceptions -g -grecord-gcc-switches -pipe -Wall -Werror=format-security -Wp,-U_FORTIFY_SOURCE,-D_FORTIFY_SOURCE=3 -Wp,-D_GLIBCXX_ASSERTIONS --config /usr/lib/rpm/redhat/redhat-hardened-clang.cfg -fstack-protector-strong   -m64 -march=z13 -mtune=z14 -fasynchronous-unwind-tables -fstack-clash-protection -DNDEBUG -std=c99 -fPIC -MD -MT confu-deps/cpuinfo/CMakeFiles/cpuinfo.dir/src/api.c.o -MF CMakeFiles/cpuinfo.dir/src/api.c.o.d -o CMakeFiles/cpuinfo.dir/src/api.c.o -c /builddir/build/BUILD/pytorch-v2.0.1.gitf81f9093/third_party/cpuinfo/src/api.c
/builddir/build/BUILD/pytorch-v2.0.1.gitf81f9093/third_party/cpuinfo/src/api.c:319:23: error: call to undeclared function 'syscall'; ISO C99 and later do not support implicit function declarations [-Wimplicit-function-declaration]
                if CPUINFO_UNLIKELY(syscall(__NR_getcpu, &cpu, NULL, NULL) != 0) {
                                    ^
/builddir/build/BUILD/pytorch-v2.0.1.gitf81f9093/third_party/cpuinfo/src/api.c:338:23: error: call to undeclared function 'syscall'; ISO C99 and later do not support implicit function declarations [-Wimplicit-function-declaration]
                if CPUINFO_UNLIKELY(syscall(__NR_getcpu, &cpu, NULL, NULL) != 0) {
                                    ^
2 errors generated.

A full build log can be found here: https://download.copr.fedorainfracloud.org/results/jmontleon/pytorch/fedora-38-s390x/06239794-pytorch/builder-live.log.gz

[Maratyszcza]

Why is syscall undeclared on Red Hat?

[Maratyszcza]

Also, cpuinfo doesn't support ppc64 and s390x

[jmontleon]

Is there a way to build pytorch without it?

[Maratyszcza]

I don't know, you should ask PyTorch devs

[jmontleon]

Did I file this in the wrong place? cpuinfo is included in the third_party deps of pytorch and this repo is within the pytorch org. I had assumed you all talk occasionally?

The only option I see available in pytorch is -DUSE_SYSTEM_PYTORCH to use a version provided by the system rather than bundled in the third party repo.

And the cmake warning when building on unsupported architectures:

CMake Warning at third_party/cpuinfo/CMakeLists.txt:81 (MESSAGE):
  Target processor architecture "s390x" is not supported in cpuinfo.  cpuinfo
  will compile, but cpuinfo_initialize() will always fail.

[jmontleon]

To you question, "Why is syscall undeclared on Red Hat?", I don't think this is distribution specific. As mentioned I don't have this problem on amd64 or aarch64 using the same version of clang. Rather it seems a difference in architectures.

One of the explanations I see is: "#define _GNU_SOURCE before including unistd.h or any other header as syscall(2) is not POSIX." https://stackoverflow.com/questions/42469583/problems-calling-syscall-function-in-c

It's also present in the example within the man page https://man7.org/linux/man-pages/man2/syscall.2.html#EXAMPLES

Putting this define at the top of the file does seem to resolve the build failure. I don't have a strong enough understanding at present to say why it's necessary for some architectures and not others.

[bhbruce]

@Maratyszcza syscall is not a standard function in posix.

It is implemented in glibc. It looks like the Clang has strict constraints (compared to GCC) ISO C99 and later do not support implicit function declarations [-Wimplicit-function-declaration].

The man page https://linux.die.net/man/2/syscall also show that we should add #define _GNU_SOURCE before including headers.