On Sat, May 16, 2015 at 09:55:08PM -0400, Hans-Peter Nilsson wrote: > > With a plus or minus combine would always write it as a mult. > > I don't think any other pass would create this combination. I > > haven't tested it though. > > Ports probably also generate that internally at various RTL > passes, something that takes a bit more than an at-a-glance code > inspection.
Yeah, true. Hopefully port maintainers will know about what the port does, or at least notice the ICEs. > > > I've been trying to come up with valid reasons there'd be ever > > > be canonicalization by multiplication, but failed so I guess > > > I'll rip it out. > > > > The "unreachable" thing should quickly tell you if that guess is wrong. > > Not something you want to leave in a production compiler, of course. > > I think you misunderstood; I mean that I pondered > "philosophical" reasons to change the canonical representation; > if there was some reasoning, current or looking forward, where > we'd "add back" mult as non-address-canonical RTL. Ah. Yes, same here -- I don't see any situation where having mult instead of shift outside of a mem would be more convenient. > Actually, there are two things you could help with: > > - (pointer-to) step-by-step instructions to recreate the Linux > (kernel :) build, as those you did before for a multiple of > targets. I'd like to know I fixed your observations. I used <http://git.infradead.org/users/segher/buildall.git>; it has a README. I see that doc is a little out of date, I'll update. > - add a preferred canonicalization function to do conversion > to/from memory-address-canonical RTL. Like > fwprop.c:canonicalize_address (just not static :) and maybe also > a canonicalize_nonaddress. At the moment, ports call > replace_equiv_address (family of functions) when changing > address RTL, but that code-path (at a glance) doesn't > canonicalize whatever non-address-canonical RTL you throw at it. > Maybe it should? Maybe validize_mem is nicer? It depends much what you really want to change. Does simplify_rtx not do what you want for "nonaddress"? Segher
