I am currently looking at the performance page.

• https://github.com/simdjson/simdjson/blob/master/doc/performance.md#free-padding

which contains this function.

bool need_allocation(const char *buf, size_t len) {
  return ((reinterpret_cast<uintptr_t>(buf + len - 1) % page_size())
    + simdjson::SIMDJSON_PADDING > static_cast<uintptr_t>(page_size()));
}

I am writing some unit tests.

• system_page_size is 4096 (4k)
• buf starts at 0x0
• len = 4033
• SIMDJSON_PADDING = 64 bytes

The last address of such an array is 0 + 4033 - 1 = 4032.

Add the padding, we get 4096. 4096, if interpreted as a pointer to an address, is in the next page. So the function should return true.

Instead, it returns false.

In other words, the operator should be >= and not >.

One possible solution occurs to me. While SIMDJSON_PADDING is defined to be 64 bytes, the actual amount of padding required is only 63 bytes.

Why? Because the SIMD machine reads blocks of 64 bytes at a time, and if our array of data to be processed ends on the final byte before a new block of 64 starts, we do not need any additional padding. If the array ended on the first address of a new block of 64 bytes, then an additional 63 bytes would be required.

Can anyone confirm what is the expected behavior?