memcpy optimisations for openrisc (linux)

memcpy_testing.txt

# The test is running memcpy to copy an `int src[4096]` to another `int dest[4096]`
# compile test
or1k-elf-gcc -O2 -g -o memcpy_test.elf -mboard=de0_nano memcpy_test.c

# Run sime with trace the grep for loads and stores and sort by counts of each
sim  -t -f arch/openrisc/or1ksim.cfg memcpy_test.elf | grep -e '\(l.l\|l.s\)' | cut -d' ' -f 4 | sort | uniq -c | sort -n | tail

# Hello world is a baseline for how many operations happen without any memcpy test
Hello World (baseline)
    105 l.slli
    169 l.sfltu
    180 l.sfne
    484 l.lwz
   1106 l.sw
   7358 l.lbz
   7439 l.sfeqi

# After unrolling the load stores are same but, less branches
byte unroll
     98 l.slli
    164 l.sfne    (16K less)
    169 l.sfltu
    461 l.lwz
   1101 l.sw
   2114 l.sfnei     
   2542 l.sfeqi
  16402 l.sb
  18838 l.lbz

# Doing word copies there are less load stores (4096 compared to 16000)
word
   69 l.sfnei
     97 l.slli
    164 l.sfne
    169 l.sfltu
   2454 l.lbz
   2539 l.sfeqi
   4096 l.sfgtuir6,0x3
   4554 l.lwz
   5194 l.sw

# Doing word copy plus loop unrolls there are less branches
word unroll
     30 l.sub
     58 l.srli
     67 l.sfnei
     99 l.slli
    169 l.sfltu
    676 l.sfne
   2454 l.lbz
   2542 l.sfeqi     4000 less branches
   4557 l.lwz
   5197 l.sw