> > No, that is a different issue from toplevel asm, that is global > > register variables which this patch set as far as I know does not fix. > > Global reg vars should be handled similar though, by constraining > partitioning. I think it is easier to just put info about global vars into struct function so we can compile unit with mixed globals. I am not sure we want to do cross-module inlining between modules with different global registers though.
Honza.