Jim White wrote:
I have a loop that spends 99+% of its time executing these two statements:
mpz_addmul_ui(BB, B, a); mpz_swap (B, BB);To my surprise, it runs 25% faster if I replace it with:
if (a==1) mpz_add (BB, B, BB); else mpz_addmul_ui(BB, B, a); mpz_swap(B, BB);It's an RCF (regular continued-fraction) algorithm, so a=1 about 41.5% of the time.
No big deal, but perhaps in future revs mpz_addmul_ui and its sister function might auto-detect this case and reroute the call?
In the Parma Polyhedra Library, we found out that, for mpz_mul(), checking the source arguments against 0 gives rise to significant speedups in some algorithms.
We thought (but we are not sure) that this is indeed a feature, and not a limitation: since only the client code can know about the frequence of special cases, let the client code take the decision. The alternative would be that GMP does a lot of checks that are possibly useless, perhaps duplicating work that has already been done by the client code.
I am glad you raised the issue here and I hope Torbjorn will correct me if I misinterpreted the GMP's policies in this respect. All the best,
Roberto