This repo exercises store forwarding. Modern processors can forward a memory store to a subsequent load from the same address. This is called “store-to-load forwarding”, and it improves performance because the load does not have to wait for the data to be written to the cache and then read back again. For reference, see Agner Fog's guide.
The example of a forward store is:
movaps xmmword ptr [rsp], xmm0
mov eax, dword ptr [rsp + 2]It stores 16 bytes and reads 4 bytes with the offset 2.
Benchmarks are represented in a store_X_load_Y_offset_Z where we store X bytes, load Y bytes with the offset Z (obviously, Y + Z
In the end it means you should avoid in libraries loading last bytes or loading bytes across the boundary of 8 bytes -- it has bad performance on Intel and Arm. AMD has been quite good overall. Examples for that:
TBD: more processors, 32 byte loads, etc.