Reply To: switches

There’s a –Os setting to specialize for space. At present it shuts off cloning speculable operations up into inbound control flow arcs, except in the special case where the EBB has only one op (commonly a retn), so there is no size bloat because the cloned op replaces the otherwise necessary branch.

We don’t unroll loops at present, and actually have had quite a bit of argument with the LLVM code that really really wants to unroll. There is one case where we know we want to unroll, in loop pipelines in which there is enough hardware functionality to do more than one iteration concurrently and there’s no loop-carried dependencies. However, the specializer isn’t smart enough yet to do that.