[libcxx-commits] [libcxx] [libc++] Vectorize mismatch (PR #73255)

[Denis Yaroshevskiy via libcxx-commits]

================
@@ -11,23 +11,89 @@
 #define _LIBCPP___ALGORITHM_MISMATCH_H
 
 #include <__algorithm/comp.h>
+#include <__algorithm/simd_utils.h>
+#include <__algorithm/unwrap_iter.h>
 #include <__config>
-#include <__iterator/iterator_traits.h>
+#include <__functional/identity.h>
+#include <__type_traits/invoke.h>
+#include <__type_traits/is_constant_evaluated.h>
+#include <__type_traits/is_equality_comparable.h>
+#include <__type_traits/operation_traits.h>
+#include <__utility/move.h>
 #include <__utility/pair.h>
 
 #if !defined(_LIBCPP_HAS_NO_PRAGMA_SYSTEM_HEADER)
 #  pragma GCC system_header
 #endif
 
+_LIBCPP_PUSH_MACROS
+#include <__undef_macros>
+
 _LIBCPP_BEGIN_NAMESPACE_STD
 
+template <class _Iter1, class _Sent1, class _Iter2, class _Pred, class _Proj1, class _Proj2>
+_LIBCPP_NODISCARD _LIBCPP_HIDE_FROM_ABI _LIBCPP_CONSTEXPR_SINCE_CXX20 pair<_Iter1, _Iter2>
+__mismatch_loop(_Iter1 __first1, _Sent1 __last1, _Iter2 __first2, _Pred& __pred, _Proj1& __proj1, _Proj2& __proj2) {
+  while (__first1 != __last1) {
+    if (!std::__invoke(__pred, std::__invoke(__proj1, *__first1), std::__invoke(__proj2, *__first2)))
+      break;
+    ++__first1;
+    ++__first2;
+  }
+  return std::make_pair(std::move(__first1), std::move(__first2));
+}
+
+template <class _Iter1, class _Sent1, class _Iter2, class _Pred, class _Proj1, class _Proj2>
+_LIBCPP_NODISCARD _LIBCPP_HIDE_FROM_ABI _LIBCPP_CONSTEXPR_SINCE_CXX20 pair<_Iter1, _Iter2>
+__mismatch(_Iter1 __first1, _Sent1 __last1, _Iter2 __first2, _Pred& __pred, _Proj1& __proj1, _Proj2& __proj2) {
+  return std::__mismatch_loop(__first1, __last1, __first2, __pred, __proj1, __proj2);
+}
+
+#if _LIBCPP_VECTORIZE_ALGORIHTMS
+
+template <class _Tp,
+          class _Pred,
+          class _Proj1,
+          class _Proj2,
+          __enable_if_t<is_integral<_Tp>::value && __desugars_to<__equal_tag, _Pred, _Tp, _Tp>::value &&
+                            __is_identity<_Proj1>::value && __is_identity<_Proj2>::value,
+                        int> = 0>
+_LIBCPP_NODISCARD _LIBCPP_HIDE_FROM_ABI _LIBCPP_CONSTEXPR_SINCE_CXX20 pair<_Tp*, _Tp*>
+__mismatch(_Tp* __first1, _Tp* __last1, _Tp* __first2, _Pred& __pred, _Proj1& __proj1, _Proj2& __proj2) {
+  constexpr size_t __unroll_count = 4;
+  constexpr size_t __vec_size     = __native_vector_size<_Tp>;
+  using __vec                     = __simd_vector<_Tp, __vec_size>;
+  while (!__libcpp_is_constant_evaluated() && static_cast<size_t>(__last1 - __first1) >= __unroll_count * __vec_size) {
+    __vec __lhs[__unroll_count];
+    __vec __rhs[__unroll_count];
+
+    for (size_t __i = 0; __i != __unroll_count; ++__i) {
+      __lhs[__i] = std::__load_vector<__vec>(__first1 + __i * __vec_size);
+      __rhs[__i] = std::__load_vector<__vec>(__first2 + __i * __vec_size);
+    }
+
+    for (size_t __i = 0; __i != __unroll_count; ++__i) {
+      if (auto __cmp_res = __lhs[__i] == __rhs[__i]; !std::__all_of(__cmp_res)) {
----------------
DenisYaroshevskiy wrote:

I really don't like the dependency here.

1) have a look if maybe comparing all into `bool  __res[__unroll_count];` would be helpful.

2) or all of them togehter and then do one `__all_of`

Try which is best.

Look at std::memcmp:
https://codebrowser.dev/glibc/glibc/sysdeps/x86_64/multiarch/memcmpeq-avx2.S.html#220

https://github.com/llvm/llvm-project/pull/73255