On Fri, Feb 26, 2016 at 01:10:17PM -0800, Richard Henderson wrote: > On 02/26/2016 01:01 PM, Segher Boessenkool wrote: > >> How do you imagine the rs6000 change will regress codegen? > > > > Combine of sequences with double-length adds. > > What sort of test case are you imagining here? The trivial tests I've looked > at have all been optimal (before and after).
There are many different combinations, for example (32-bit code) long long add_s42(long long a, int b) { return a + ((long long) b << 32) + 42; } and we generate optimal code on BE for all of those (not on LE yet, and there is the issue with open-coded carry chains, and and and, and things shift around. Also all the scc things. But I really should submit some testcases). Segher