bit4

Fix 32-bit address load in 64-bit AVX2 code.

3

The new code works correctly on 64-bit systems even if the address of blake2b_sigma doesn't fit into 32 bits, and it's PIC safe.

There seems to be a memory leak in stringbuf.c:sbuf_split_at(). The function returns without freeing "res" if "pos" is negative.