On Thu, Feb 14, 2019 at 05:30:52PM +0100, Eric Botcazou wrote: > > Yeah, something like that. It will need some serious testing, to make > > sure we don't regress (including not regressing what that patch that took > > them away was meant to do). I can arrange some testing, will you do the > > patch though? > > I can do the patch and also (correctness) testing for 32-bit Linux.
Performance testing is important here, of course. > Another issue is the extent of the patch: practically speaking, putting back > the '*' modifier before all the 'd' constraints would be sufficient, but the > current setting is a bit inconsistent|*] so this could also be adjusted. > > [*] For example, in the *movdi_internal32 pattern, 2 'wi' constraints have it > but not the other 2. Likewise for "wv'. I think we should change as little as possible for 7/8/9 here, because this is pretty fragile :-( But for 10, yes, let's get more sanity. (We'll have the "enabled" attribute for GCC 10, btw). Segher