[libcxx] Use lgamma rather than lgamma_r with LLVM libc

[petrhosek]

lgamma_r is currently only available on GPU targets.

[llvmbot]

@llvm/pr-subscribers-libcxx

@llvm/pr-subscribers-libc

Author: Petr Hosek (petrhosek)

Changes

lgamma_r is currently only available on GPU targets.

---

Full diff: https://github.com/llvm/llvm-project/pull/109556.diff

1 Files Affected:

• (modified) libcxx/include/__random/binomial_distribution.h (+1-1)

diff --git a/libcxx/include/__random/binomial_distribution.h b/libcxx/include/__random/binomial_distribution.h
index 3f19746bae238c..9538c15e2dc97b 100644
--- a/libcxx/include/__random/binomial_distribution.h
+++ b/libcxx/include/__random/binomial_distribution.h
@@ -103,7 +103,7 @@ extern "C" double lgamma_r(double, int*);
 #endif
 
 inline _LIBCPP_HIDE_FROM_ABI double __libcpp_lgamma(double __d) {
-#if defined(_LIBCPP_MSVCRT_LIKE)
+#if defined(_LIBCPP_MSVCRT_LIKE) || defined(__LLVM_LIBC__)
   return lgamma(__d);
 #else
   int __sign;

[jhuber6]

But we don't have lgamma either for non-GPU targets, so how does this work?

[petrhosek]

But we don't have lgamma either for non-GPU targets, so how does this work?

libc++ provides it's own implementation which uses __builtin_lgamma, see:

llvm-project/libcxx/include/__math/gamma.h

Lines 26 to 38 in eaedbbc

	inline _LIBCPP_HIDE_FROM_ABI float lgamma(float __x) _NOEXCEPT { return __builtin_lgammaf(__x); }
	template <class = int>
	_LIBCPP_HIDE_FROM_ABI double lgamma(double __x) _NOEXCEPT {
	return __builtin_lgamma(__x);
	}
	inline _LIBCPP_HIDE_FROM_ABI long double lgamma(long double __x) _NOEXCEPT { return __builtin_lgammal(__x); }
	template <class _A1, __enable_if_t<is_integral<_A1>::value, int> = 0>
	inline _LIBCPP_HIDE_FROM_ABI double lgamma(_A1 __x) _NOEXCEPT {
	return __builtin_lgamma((double)__x);
	}

[jhuber6]

But we don't have lgamma either for non-GPU targets, so how does this work?

libc++ provides it's own implementation which uses __builtin_lgamma, see:

llvm-project/libcxx/include/__math/gamma.h

Lines 26 to 38 in eaedbbc

	inline _LIBCPP_HIDE_FROM_ABI float lgamma(float __x) _NOEXCEPT { return __builtin_lgammaf(__x); }
	template <class = int>
	_LIBCPP_HIDE_FROM_ABI double lgamma(double __x) _NOEXCEPT {
	return __builtin_lgamma(__x);
	}
	inline _LIBCPP_HIDE_FROM_ABI long double lgamma(long double __x) _NOEXCEPT { return __builtin_lgammal(__x); }
	template <class _A1, __enable_if_t<is_integral<_A1>::value, int> = 0>
	inline _LIBCPP_HIDE_FROM_ABI double lgamma(_A1 __x) _NOEXCEPT {
	return __builtin_lgamma((double)__x);
	}

__builtin_lgamma isn't handled anywhere in the backend AFAIK, it will almost certainly be reduced to an unresolved libcall.

[ldionne]

Please rebase onto main to re-trigger CI. The CI instability should be resolved now.

[ldionne]

LGTM w/ green CI (please rebase).

[ldionne]

I rebased. This can be landed once CI is green.

[jhuber6]

I'm fine with this in general, since I think once the LLVM libc implements lgamma it is probably fine to not provide the signgam external global. This isn't mentioned in the standard, so it's probably an extension GNU / POSIX made than then had to make lgamma_r to escape from once people started realizing global state was bad. I'm just saying that neither lgamma nor lgamma_r are provided right now, the builtin will just make the symbol undefined (but I guess you can compile it?) It's the same as if you just put double lgamma(double) in the file without ever definint it.