[libc++] Decrease instantiation cost of __constexpr_memmove

[philnik777]

Using if constexpr in __constexpr_memmove makes the instantiation three times faster for the same type, since it avoids a bunch of class instantiations and SFINAE for constexpr support that's never actually used. Given that __constexpr_memmove is used quite a bit through std::copy and is instantiated multiple times when just including <__string/char_traits.h> this can provide a nice compile time speedup for a very simple change.

[llvmbot]

@llvm/pr-subscribers-libcxx

Author: Nikolas Klauser (philnik777)

Changes

Using if constexpr in __constexpr_memmove makes the instantiation three times faster for the same type, since it avoids a bunch of class instantiations and SFINAE for constexpr support that's nenver actually used. Given that __constexpr_memmove is used quite a bit through std::copy and is instantiated multiple times when just including <__string/char_traits.h> this can provide a nice compile time speedup for a very simple change.

---

Full diff: https://github.com/llvm/llvm-project/pull/125109.diff

1 Files Affected:

• (modified) libcxx/include/__string/constexpr_c_functions.h (+12-9)

diff --git a/libcxx/include/__string/constexpr_c_functions.h b/libcxx/include/__string/constexpr_c_functions.h
index 0bc128b68b5799..fbe7e10d440ce1 100644
--- a/libcxx/include/__string/constexpr_c_functions.h
+++ b/libcxx/include/__string/constexpr_c_functions.h
@@ -204,23 +204,26 @@ _LIBCPP_HIDE_FROM_ABI _LIBCPP_CONSTEXPR_SINCE_CXX20 _Tp& __assign_trivially_copy
   return __dest;
 }
 
-template <class _Tp, class _Up, __enable_if_t<__is_always_bitcastable<_Up, _Tp>::value, int> = 0>
+template <class _Tp, class _Up>
 _LIBCPP_HIDE_FROM_ABI _LIBCPP_CONSTEXPR_SINCE_CXX14 _Tp*
 __constexpr_memmove(_Tp* __dest, _Up* __src, __element_count __n) {
+  static_assert(__is_always_bitcastable<_Up, _Tp>::value);
   size_t __count = static_cast<size_t>(__n);
   if (__libcpp_is_constant_evaluated()) {
 #ifdef _LIBCPP_COMPILER_CLANG_BASED
-    if (is_same<__remove_cv_t<_Tp>, __remove_cv_t<_Up> >::value) {
+    if _LIBCPP_CONSTEXPR (is_same<__remove_cv_t<_Tp>, __remove_cv_t<_Up> >::value) {
       ::__builtin_memmove(__dest, __src, __count * sizeof(_Tp));
       return __dest;
-    }
+    } else
 #endif
-    if (std::__is_pointer_in_range(__src, __src + __count, __dest)) {
-      for (; __count > 0; --__count)
-        std::__assign_trivially_copyable(__dest[__count - 1], __src[__count - 1]);
-    } else {
-      for (size_t __i = 0; __i != __count; ++__i)
-        std::__assign_trivially_copyable(__dest[__i], __src[__i]);
+    {
+      if (std::__is_pointer_in_range(__src, __src + __count, __dest)) {
+        for (; __count > 0; --__count)
+          std::__assign_trivially_copyable(__dest[__count - 1], __src[__count - 1]);
+      } else {
+        for (size_t __i = 0; __i != __count; ++__i)
+          std::__assign_trivially_copyable(__dest[__i], __src[__i]);
+      }
     }
   } else if (__count > 0) {
     ::__builtin_memmove(__dest, __src, (__count - 1) * sizeof(_Tp) + __datasizeof_v<_Tp>);

[mordante]

Thanks, LGTM!

Just curious, do you have any numbers on the compilation performance improvement?

FYI I fixed a minor typo in never in the commit message.

[philnik777]

Thanks, LGTM!

Just curious, do you have any numbers on the compilation performance improvement?

IIRC for every instantiation where the condition is true you save ~0.8ms. Not huge on its own, but given that it's instantiated quite a bit it's probably a nice improvement across an entire code base.

FYI I fixed a minor typo in never in the commit message.

Thanks!