Hi,
Yosi pointed out how to do it faster.
Usually we work on messages that are not too small - at least, say, 64 bytes.
So if we break our original 64-byte packet into 2 parts - say, the first 32 bytes and the last 32 bytes - can work on 2 32-byte messages independently, recombining the results once in the end.
Initial conditions for the packets will be r0 = 0 (CRC32 of 0 message) and some precalculated const CRC32_32.
LOOP CRC32 LC0 = p4; // p4 = (32 - 1) in this case
p3 = p0;
r0 = 0;
r1 = CRC32_32; // CRC32 initial value corresponging to x^(8*32) message polynomial
p3 += 32; //
r5 = r0 >> 24 || r2 = b[p0++]; // r1 - input byte, 8 MSB of CRC interacting with input
r2 = r2 ^ r5;
r3 = (r7+r2) << 2 ; // r7 = CRC table base
i1 = r3; // ptr into CRC update table, size 256*4 bytes
r5 = r1 >> 24 || r2 = b[p3++];
r2 = r2 ^ r5;
r3 = (r7+r2) << 2 ; // r7 = CRC table base
i2 = r3;
r5 = r0 << 8; // 24 LSB of old CRC interacting with CRC update
r4 = r1 << 8 || r3 = [i1]; // r6 is 32-bit CRC corresponding to input byte^MSbyte of CRC
LOOP_BEGIN CRC32;
r0 = r5 ^ r3; // update r0
nop; // still 1 stall here...
r5 = r0 >> 24 || r2 = b[p0++] || r6 = [i2]; // r1 - input byte, 8 MSB of CRC interacting with input
r2 = r2 ^ r5;
r2 = (r7+r2) << 2 ; // r7 = CRC table base
i1 = r2; // ptr into CRC update table, size 256*4 bytes
r1 = r4 ^ r6; // update r1
r5 = r1 >> 24 || r2 = b[p3++];
r2 = r2 ^ r5;
r2 = (r7+r2) << 2 ; // r7 = CRC table base
i2 = r2;
r5 = r0 << 8; // 24 LSB of old CRC interacting with CRC update
r4 = r1 << 8 || r3 = [i1]; // r6 is 32-bit CRC corresponding to input byte^MSbyte of CRC
LOOP_END CRC32;
r0 = r0 ^ r1; //final r0 update
// back at ~ 13 cycles/2 bytes = 6.5 c/byte = 0.8 c/bit
