conferences/fosdem2024/fosdem2024_ddffirst/pospopcount.s

   1 mtspr 9, 3              # move r3 to CTR
   2 setvl 3,0,8,0,1,1       # MVL=8, VL=r3=MIN(MVL,CTR)
   3 # load VL bytes (update r4 addr) but width=8 (dw=8)
   4 addi 6, 0, 0            # set 64-bits of r6=0
   5 sv.lbzu/pi/dw=8 *6, 1(4)
   6 # gather performs transpose (gets us to positional)
   7 gbbd 8,6
   8 # now bits are turned around, popcount and sum them
   9 setvl 0,0,8,0,1,1       # set MVL=VL=8
  10 sv.popcntd/sw=8 *24,*8  # do (transposed) popcnt
  11 sv.add *16,*16,*24      # accumulate in results
  12 # branch back if CTR non-zero. works even when VL=8
  13 sv.bc/all 16, *0, -0x28 # reduce CTR by VL