Replacing 32-bit loop counter with 64-bit causes strange performance degradation with _mm_popcnt_u64 on Intel CPU

Hello everyone! I'm writing a C++ program to count set bits in a large data array. I'm using the _mm_popcnt_u64 intrinsic from the SSE4.2 instruction set. The code is simple: I iterate through a uint64_t array and accumulate the result. Everything works fine, but I noticed a strange thing: if I change the loop counter type from 32-bit (unsigned int) to 64-bit (size_t), performance drops by 2-3 times! The data is the same, the array size is the same, the compiler is the same (MSVC 2019, flags /O2 /arch:AVX2). Here's the simplified code: #include #include #include #include int main() { const size_t N = 1'000'000; std::vector data(N, 0xFFFFFFFFFFFFFFFF); // all bits = 1 // Version 1: 32-bit counter unsigned int sum32 = 0; auto t1 =