Richard Henderson <[email protected]> writes:
> On 10/26/2016 08:47 PM, David Gibson wrote:
>>> > +void helper_vprtybq(ppc_avr_t *r, ppc_avr_t *b)
>>> > +{
>>> > + int i;
>>> > + uint8_t s = 0;
>>> > + for (i = 0; i < 16; i++) {
>>> > + s ^= (b->u8[i] & 1);
>>> > + }
>>> > + r->u64[LO_IDX] = (!s) ? 0 : 1;
>>> > + r->u64[HI_IDX] = 0;
>>> > +}
>>> > +
>> I think you can implement these better. First mask with 0x01010101
>> (of the appropriate length) to extract the LSB bits of each byte.
>> Then XOR the two halves together, then quarters and so forth,
>> ln2(size) times to arrive at the parity. This is similar to the usual
>> Hamming weight implementation.
>>
>
> You don't even have to mask with 0x01010101 to start. Just fold halves til
> you
> get to the byte level and then mask with 1.
Right, it does reduce number of operations:
+#define SIZE_MASK(x) ((1ULL << (x)) - 1)
+static uint64_t vparity(uint64_t f1, uint64_t f2, int size)
+{
+ uint64_t res = f1 ^ f2;
+ if (size == 8) return res;
+ return vparity(res & SIZE_MASK(size/2), res >> (size/2), size/2);
+}
+
Regards
Nikunj
