http://gcc.gnu.org/bugzilla/show_bug.cgi?id=56309
--- Comment #5 from Uros Bizjak <ubizjak at gmail dot com> 2013-02-14 08:22:33 UTC --- For reference, some quotes from Honza: -- PR54073-- The decision on whether to use cmov or jmp was always tricky on x86 architectures. Cmov increase dependency chains, register pressure (both values needs to be loaded in) and has long opcode. So jump sequence, if well predicted, flows better through the out-of-order core. If badly predicted it is, of course, a disaster. I think more modern CPUs solved the problems with long latency of cmov, but the dependency chains are still there. [...] We should do something about rnflow. I will look into that. The usual wisdom is that lacking profile feedback one should handle non-loop branhces as inpredctable and loop branches as predictable, so all handled by ifconvert fals to the first category. With profile feedback one can see branch probability and if it is close to 0 or REG_BR_PROB_BASE tread the branch as predictable. We handle this with predictable_edge_p parameter passed to BRANCH_COST (that by itself is a gross, but for years we was not able to come with something saner) -- /PR54073 -- -- PR53046 -- Well, as I wrote to the other PR, the main problem of cmov is extension of dependency chain. For well predicted sequence with conditional jump there is no update of rbs so the loop executes faster, because the loads/stores/comparisons executes "in parallel". The load in the next iteration can then happen speculatively before the condition from previous iteration is resolved. With cmov in it, there is dependence on rbx for all the other computations in the loop. I guess there is no localy available information suggesting suggesting that the particular branch is well predictable, at least without profile feedback (where we won't disable the conversion anyway). [...] -- /PR53046 --
