[libcxx-commits] [libcxx] [libc++] Optimize the std::mismatch tail (PR #83440)

[Louis Dionne via libcxx-commits]

================
@@ -84,8 +87,36 @@ __mismatch(_Tp* __first1, _Tp* __last1, _Tp* __first2, _Pred& __pred, _Proj1& __
       __first1 += __unroll_count * __vec_size;
       __first2 += __unroll_count * __vec_size;
     }
+
+    // check the remaining 0-3 vectors
+    while (static_cast<size_t>(__last1 - __first1) >= __vec_size) {
+      if (auto __cmp_res = std::__load_vector<__vec>(__first1) == std::__load_vector<__vec>(__first2);
+          !std::__all_of(__cmp_res)) {
+        auto __offset = std::__find_first_not_set(__cmp_res);
+        return {__first1 + __offset, __first2 + __offset};
+      }
+      __first1 += __vec_size;
+      __first2 += __vec_size;
+    }
+
+    if (__last1 - __first1 == 0)
+      return {__first1, __first2};
+
+    // Check if we can load elements in fron of the current pointer. If that's the case load a vector at
----------------
ldionne wrote:

```suggestion
    // Check if we can load elements in front of the current pointer. If that's the case load a vector at
```

https://github.com/llvm/llvm-project/pull/83440