Remove incorrect forward-declaration of lgamma_r header

[georgthegreat]

libc++ declares lgamma_r as:

extern "C" double lgamma_r(double, int *);

while modern glibc uses the following:

__MATHCALL (lgamma,_r, (_Mdouble_, int *__signgamp));

This causes (I did not digged to the very bottom of it) the following compilation error when compiling with nvcc:

libcxx/include/__random/binomial_distribution.h(117): error: linkage specification is incompatible with previous "lgamma_r"
/usr/include/x86_64-linux-gnu/bits/mathcalls.h(271): here

Using original declaration solves the issue.

[llvmbot]

@llvm/pr-subscribers-libcxx

Author: Yuriy Chernyshov (georgthegreat)

Changes

libc++ declares lgamma_r as:

extern "C" double lgamma_r(double, int *);

while modern glibc uses the following:

__MATHCALL (lgamma,_r, (_Mdouble_, int *__signgamp));

This causes (I did not digged to the very bottom of it) the following compilation error when compiling with nvcc:

libcxx/include/__random/binomial_distribution.h(117): error: linkage specification is incompatible with previous "lgamma_r"
/usr/include/x86_64-linux-gnu/bits/mathcalls.h(271): here

---

Full diff: https://github.com/llvm/llvm-project/pull/80979.diff

1 Files Affected:

• (modified) libcxx/include/__random/binomial_distribution.h (-4)

diff --git a/libcxx/include/__random/binomial_distribution.h b/libcxx/include/__random/binomial_distribution.h
index e8774bb8d67ee..54f321af63d50 100644
--- a/libcxx/include/__random/binomial_distribution.h
+++ b/libcxx/include/__random/binomial_distribution.h
@@ -97,10 +97,6 @@ class _LIBCPP_TEMPLATE_VIS binomial_distribution {
   }
 };
 
-#ifndef _LIBCPP_MSVCRT_LIKE
-extern "C" double lgamma_r(double, int*);
-#endif
-
 inline _LIBCPP_HIDE_FROM_ABI double __libcpp_lgamma(double __d) {
 #if defined(_LIBCPP_MSVCRT_LIKE)
   return lgamma(__d);

[georgthegreat]

This error is triggered upon #including <random> after <cmath>.

#include <iostream>
#include <cmath>
#include <random>

int main() {
    int state = 0;

    long double x;
    std::cin >> x;
    auto val = lgamma_r(x, &state);

    std::cout << "Gamma(" << x << ") = " << val << std::endl;

    return 0;
}

[github-actions]

✅ With the latest revision this PR passed the C/C++ code formatter.

[ldionne]

This fails on the Apple CI with:

/Users/libcxx-buildkite-agent/libcxx.buildkite-agent/builds/f1-1-macminivault-com/llvm-project/libcxx-ci/build/apple-system/arm64/include/c++/v1/__random/binomial_distribution.h:105:10: error: use of undeclared identifier 'lgamma_r'
  return lgamma_r(__d, &__sign);
         ^

[georgthegreat]

We have this patch ported locally and it works just fine (though our CI tests against macos-version-min=11.0).
Maybe this specific CI turns _REENTRANT off somehow.

There is no cuda compiler under MacOS, so we may simply forward-declare it under __APPLE__

[ldionne]

This seems to break when we use modules (see the CI job).

[georgthegreat]

Unfortunately, I have no experience with modules and can not quite understand the reason for hte failure.
Could you, please, help me to debug this?

[ldionne]

Could you rebase this on top of latest main so we can have a look at the modules issues if they still exist?

[georgthegreat]

@ldionne, it turns out that this was fixed in 84ce230

I am closing this PR as there is no need in it anymore.