> So I don't think I have a good feel for the advantages and disadvantages > of doing this. Robin's analysis of the aarch64 changes was nice and > detailed though. I think the one that worries me most is the addressing > mode one. fwprop is probably the first chance we get to propagate adds > into addresses, and virtual register elimination means that some of > those opportunities won't show up in gimple. > > There again, virtual register elimination wouldn't be the reason for > the ld4_s8.c failure. Perhaps there's something missing in expand. > > Other than that, I think my main question is: why just unary operations? > Is the underlying assumption that we only want to propagate a maximum of > one register? If so, then I think we should check for that directly, by > iterating over subrtxes.
The main reason for stopping at unary operations was to limit the scope and change as little as possible (not restricting the change to one register). I'm currently testing a v2 that iterates over subrtxs. > Perhaps we should allow the optimisation without register-pressure > information if (a) the source register and destination register are > in the same pressure class and (b) all uses of the destination are > being replaced. (FWIW, rtl-ssa should make it easier to try to > replace all definitions at once, with an all-or-nothing choice, > if we ever wanted to do that.) I presume you're referring to replacing one register (dest) in all using insns? Source and destination are somewhat overloaded in fwprop context because I'm thinking of the "to be replaced" register as dest when it's actually the replacement register. AFAICT fwprop currently iterates over insns, going through all their uses and trying if an individual use can be substituted. Do you suggest to change this general iteration order to iterate over the defs of an insn and then try to replace all the uses at once (e.g. using ssa->change_insns)? When keeping the current order, wouldn't we need to store all potential changes instead of committing them and later apply them in bulk, e.g. grouped by use? This order would also help to pick the propagation with the most number of uses (i.e. propagation potential) but maybe I'm misunderstanding? Regards Robin
