ogxd / gxhash

In get_partial_safe, it's possible to use aligned loads by declaring the buffer as MaybeUninit<State> and then casting the pointer for std::ptr::copy and zeroing the rest of the buffer, instead of declaring the buffer as an u8 array.

gxhash/src/gxhash/platform/x86_128.rs

Lines 35 to 39 in b2b9d24

    
           let mut buffer = [0i8; VECTOR_SIZE]; 
        
           // Copy data into the buffer 
        
           std::ptr::copy(data as *const i8, buffer.as_mut_ptr(), len); 
        
           // Load the buffer into a __m256i vector 
        
           let partial_vector = _mm_loadu_epi8(buffer.as_ptr());

From what I can see there is a compiler optimization that stack allocates [0i8; VECTOR_SIZE] instead of heap allocating (probably because VECTOR_SIZE is a constant), so MaybeUninit<State> may not be faster.

About to close this one unless someone has some snippet to propose?

I tried using a struct which contains only the byte array and is marked as #[repr(align(16))]. I have not tested the performance yet, but this should still allocate on the stack and force 16-byte alignment.

Closing this as proposed solution does not provides significant performance gains nor simplifies the code. Feel free to open another issue if you have something to suggest.

	let mut buffer = [0i8; VECTOR_SIZE];
	// Copy data into the buffer
	std::ptr::copy(data as *const i8, buffer.as_mut_ptr(), len);
	// Load the buffer into a __m256i vector
	let partial_vector = _mm_loadu_epi8(buffer.as_ptr());

Use aligned loads in get_partial_safe