On Sun, 2012-09-02 at 12:32 +0200, Georg-Johann Lay wrote: > Richard Sandiford schrieb: > > Oleg Endo <oleg.e...@t-online.de> writes: > >> On Sat, 2012-09-01 at 10:10 +0100, Richard Sandiford wrote: > >> > >>> Thanks for doing this. We should perhaps add the address space too, > >>> but if you don't feel like redoing the whole patch, that can wait until > >>> someone wants it. > >> I just had a look at the address space thing... > >> There are already target hook overloads with address space parameters, > >> like legitimate_address_p and legitimize_address. Paolo suggested to do > >> something similar a while ago: > >> http://gcc.gnu.org/ml/gcc-patches/2009-06/msg01191.html > >> > >> .. but somehow this never made it into mainline?! > >> > >> Maybe it would be better to ... > >> a) Add an 'address_cost' overload to the existing onces. > >> (at least it would be consistent...) > >> b) Remove the overloads altogether and add the address space arg to the > >> original funcs. > >> > >> Personally, I'd probably favor b). Some targets (e.g. rl78) already > >> override the address space overloads and just ignore the address space > >> arg. Either way, maybe it's better to do it in a separate patch. This > >> looks a bit like a pandora's box :) > > > > Well, I certainly wasn't suggesting you had to do anything with > > overloaded hooks (which I agree are unfortunate). It's just that > > rtlanal.c:address_cost already gets an address space argument, so if > > we're going through all the ports adding the mode argument to the hook, > > it seemed a pity not to add the address space argument at the same time. > > Agreed. > > It's not possible to determine the costs from the address and the mode > alone, cf. HImode and the address spaces of the avr port for example. > > Costs are very different depending no the address space and strangely > the address space isnot part of the address but only of the surrounding > mem...
OKOK -- I'll do it :) (within the next couple of days) Cheers, Oleg
