[libc++] Optimize std::has_single_bit

[Alcaro]

Clang translates most implementations of has_single_bit to (v ^ (v-1)) > v-1 - except the one definition libc++ actually uses.

Proof of correctness: https://godbolt.org/z/d61bxW4r1

(Could also be fixed by teaching Clang to optimize better, but making source match output feels clearer to me. And it improves unoptimized performance.)

[llvmbot]

@llvm/pr-subscribers-libcxx

Author: None (Alcaro)

Changes

Clang translates most implementations of has_single_bit to (v ^ (v-1)) > v-1 - except the one definition libc++ actually uses.

Proof of correctness: https://godbolt.org/z/d61bxW4r1

(Could also be fixed by teaching Clang to optimize better, but making source match output feels clearer to me. And it improves unoptimized performance.)

---

Full diff: https://github.com/llvm/llvm-project/pull/133063.diff

1 Files Affected:

• (modified) libcxx/include/__bit/has_single_bit.h (+1-1)

diff --git a/libcxx/include/__bit/has_single_bit.h b/libcxx/include/__bit/has_single_bit.h
index 52f5853a1bc8a..ac434092a148f 100644
--- a/libcxx/include/__bit/has_single_bit.h
+++ b/libcxx/include/__bit/has_single_bit.h
@@ -25,7 +25,7 @@ _LIBCPP_BEGIN_NAMESPACE_STD
 
 template <__libcpp_unsigned_integer _Tp>
 [[nodiscard]] _LIBCPP_HIDE_FROM_ABI constexpr bool has_single_bit(_Tp __t) noexcept {
-  return __t != 0 && (((__t & (__t - 1)) == 0));
+  return (__t ^ (__t - 1)) > __t - 1;
 }
 
 _LIBCPP_END_NAMESPACE_STD

[philnik777]

It looks like you need to rebase.