What I meant is that making the cache fully-associative would result in faster implementation than N-way associative with non-power-of-2 cache line size or bucket count because of the additional circuitry involved in selecting right bucket (integer modulus and so on).
Interesting trick is using different values for cache index and tag (logical address for index and physical address as tag is common, but that's mostly about doing cache lookup concurrently with TLB lookup). In theory one could use result of some hash function as index, which would certainly change the aliasing behavior (although it is not exactly certain that such change would be for the better) and again, there is the additional cost of calculating said hash function (in combinatorial logic and with reasonably small delay).
Interesting trick is using different values for cache index and tag (logical address for index and physical address as tag is common, but that's mostly about doing cache lookup concurrently with TLB lookup). In theory one could use result of some hash function as index, which would certainly change the aliasing behavior (although it is not exactly certain that such change would be for the better) and again, there is the additional cost of calculating said hash function (in combinatorial logic and with reasonably small delay).