> > > * f95-lang.c (gfc_init_builtin_functions): Build __builtin_unreachable. > > > > I wonder if other languages need similar adjustment? > > I also wondered ;) Only Fortran triggered, I will take a look. > > > > + /* Now destroy the loop. First try to do so by cancelling the > > + patch from exit conditional if we identified good candidate. > > + > > > > you mean 'path' here, right? Similar in other places. > > Yeah. > > > > + /* Unloop destroys the latch edge. */ > > + unloop (loop, &irred_invalidated); > > + if (irred_invalidated) > > + mark_irreducible_loops (); > > > > this makes the whole thing quadratic in the number of loops. > > Please pass down a flag and handle this in the place we > > update SSA form (thus, once per function / unroll iteration). > > This was in my first version of the patch. Then I noticed that remove_path > also relies on irreducible loops to be computed and does the exactly same > logic as I do. > So we would need to extend the &irred_invalidated to remove_path as well.
Actually thinking about it, I perhaps can just split the loopback edge and output there the unreachable call letting cfgcleanup to do all the hard work. In a way I like it better when the complette unroller cancels the loops and has chance to cascade, but the cascading needs SSA update and cleanup of unneeded PHI nodes anyway. Honza
