On Nov 4, 2016, at 2:35 AM, Georg-Johann Lay <a...@gjlay.de> wrote: > >> .word .L3-(.L2) >> >> which seems wrong, given all the other code-gen. I think this has to be >> gs(.L3)-(gs(.L2)), no? I tried to get binutils to do that for me directly >> with a .word and it seemed resistant; which is unfortunate.
> Using gs, the expression would be gs(.L3)-(gs(.L2)) as gs implements word > addresses, not byte addresses. As I said, I tried that and could not get it to work. >> The compile only test cases work fine now, so nothing needs to be done for >> them. > > ICE is not a nice approach I never saw an ICE with current release binutils and top of tree gcc. If you want me to see an ICE and comment on it, I would need the command line apparently. > , IMO. I'd prefer some more graceful error using code locations of the > labels (which are not available). My preference is code-gen. :-) Anyway, the change I posted creates nice graceful errors, if one wants.
