[libcxx-commits] [libcxx] [libc++] Decreate instantiation cost of __constexpr_memmove (PR #125109)

[Nikolas Klauser via libcxx-commits]

https://github.com/philnik777 created https://github.com/llvm/llvm-project/pull/125109

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 speedup for a very simple change.


>From 4eccfe9227bac9f982a51adf9c7bdfcb6f90a44b Mon Sep 17 00:00:00 2001
From: Nikolas Klauser <nikolasklauser at berlin.de>
Date: Thu, 30 Jan 2025 20:46:40 +0100
Subject: [PATCH] [libc++] Decreate instantiation cost of __constexpr_memmove

---
 .../include/__string/constexpr_c_functions.h  | 21 +++++++++++--------
 1 file changed, 12 insertions(+), 9 deletions(-)

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>);