On January 17, 2015 1:37:12 AM CET, Alan Modra <amo...@gmail.com> wrote: >On Fri, Jan 16, 2015 at 11:03:24AM -0600, Segher Boessenkool wrote: >> On Fri, Jan 16, 2015 at 08:12:27PM +1030, Alan Modra wrote: >> > OK, so we need to fix this in the rs6000 backend, but it occurs to >me >> > that cprop also has a bug here. It shouldn't be touching fixed >hard >> > registers. >> >> Why not? It cannot allocate a fixed reg to a pseudo, but other than >> that there is nothing special about fixed regs; the transform is >> perfectly valid as far as I see. > >I didn't say that copying to a pseudo and using that was invalid.. >The bug I see is a mis-optimisation. Also, the asm operands case that >do_local_cprop already rules out changing is very similar to fixed >regs. Would you argue that changing asm operands is also valid? :) > >> It isn't a desirable transform in this case, but that is not true for >> fixed regs in general (just because the stack pointer is live >everywhere). > >What's the point in extending the lifetime of some pseudo when you >know the original fixed register is available everywhere? Do you have >some concrete example in mind where this "optimisation" is beneficial? > >Some ports even include pc in fixed_regs. So there are obvious >examples where regs in fixed_regs change behind the compiler's back. >Naive users might even expect to see the "current" value of those >regs. (Again, I'm not saying that it is invalid if gcc substituted an >older value.)
Just to add, we avoid doing this on the GIMPLE level as well. Richard.
