Vectorize find_first_of for 4 and 8 byte elements (#4587)

Co-authored-by: Stephan T. Lavavej <stl@nuwen.net>

benchmarks/src/find_first_of.cpp

@@ -33,18 +33,14 @@ void bm(benchmark::State& state) {
     }
 }
 
-#define ARGS               \
-    Args({2, 3})           \
-        ->Args({7, 4})     \
-        ->Args({9, 3})     \
-        ->Args({22, 5})    \
-        ->Args({58, 2})    \
-        ->Args({102, 4})   \
-        ->Args({325, 1})   \
-        ->Args({1011, 11}) \
-        ->Args({3056, 7});
-
-BENCHMARK(bm<uint8_t>)->ARGS;
-BENCHMARK(bm<uint16_t>)->ARGS;
+void common_args(auto bm) {
+    bm->Args({2, 3})->Args({7, 4})->Args({9, 3})->Args({22, 5})->Args({58, 2});
+    bm->Args({102, 4})->Args({325, 1})->Args({1011, 11})->Args({3056, 7});
+}
+
+BENCHMARK(bm<uint8_t>)->Apply(common_args);
+BENCHMARK(bm<uint16_t>)->Apply(common_args);
+BENCHMARK(bm<uint32_t>)->Apply(common_args);
+BENCHMARK(bm<uint64_t>)->Apply(common_args);
 
 BENCHMARK_MAIN();

stl/inc/algorithm

@@ -62,6 +62,10 @@ const void* __stdcall __std_find_first_of_trivial_1(
     const void* _First1, const void* _Last1, const void* _First2, const void* _Last2) noexcept;
 const void* __stdcall __std_find_first_of_trivial_2(
     const void* _First1, const void* _Last1, const void* _First2, const void* _Last2) noexcept;
+const void* __stdcall __std_find_first_of_trivial_4(
+    const void* _First1, const void* _Last1, const void* _First2, const void* _Last2) noexcept;
+const void* __stdcall __std_find_first_of_trivial_8(
+    const void* _First1, const void* _Last1, const void* _First2, const void* _Last2) noexcept;
 
 __declspec(noalias) _Min_max_1i __stdcall __std_minmax_1i(const void* _First, const void* _Last) noexcept;
 __declspec(noalias) _Min_max_1u __stdcall __std_minmax_1u(const void* _First, const void* _Last) noexcept;
@@ -202,6 +206,12 @@ _Ty1* _Find_first_of_vectorized(
     } else if constexpr (sizeof(_Ty1) == 2) {
         return const_cast<_Ty1*>(
             static_cast<const _Ty1*>(::__std_find_first_of_trivial_2(_First1, _Last1, _First2, _Last2)));
+    } else if constexpr (sizeof(_Ty1) == 4) {
+        return const_cast<_Ty1*>(
+            static_cast<const _Ty1*>(::__std_find_first_of_trivial_4(_First1, _Last1, _First2, _Last2)));
+    } else if constexpr (sizeof(_Ty1) == 8) {
+        return const_cast<_Ty1*>(
+            static_cast<const _Ty1*>(::__std_find_first_of_trivial_8(_First1, _Last1, _First2, _Last2)));
     } else {
         static_assert(false, "Unexpected size");
     }
@@ -230,9 +240,7 @@ _INLINE_VAR constexpr ptrdiff_t _Threshold_find_first_of = 16;
 
 // Can we activate the vector algorithms for find_first_of?
 template <class _It1, class _It2, class _Pr>
-constexpr bool _Vector_alg_in_find_first_of_is_safe =
-    _Equal_memcmp_is_safe<_It1, _It2, _Pr> // can replace value comparison with bitwise comparison
-    && sizeof(_Iter_value_t<_It1>) <= 2; // pcmpestri compatible size
+constexpr bool _Vector_alg_in_find_first_of_is_safe = _Equal_memcmp_is_safe<_It1, _It2, _Pr>;
 
 // Can we activate the vector algorithms for replace?
 template <class _Iter, class _Ty1>