https://gcc.gnu.org/bugzilla/show_bug.cgi?id=90174
--- Comment #9 from Feng Xue <fxue at os dot amperecomputing.com> --- (In reply to Feng Xue from comment #5) > > I would say that top-down algorithm behaves better than bottom-up one. I > > implemented Callahan-Koblentz (bottom-up algorithm) in GCC about 10 years > > ago and it behaved worse than the current one. I think adding an additional > Intuitively, it's better to give preference to regions with higher frequency > to let them pick up registers earlier and with more freedom, which matches > with bottom-up coloring strategy. > > > pass over regions probably would solve the problem but it will complicate > > already complicated RA (which is about 60K lines now). In any case I'll > > think about this problem. > > > > The problem you are mentioning in this code is quite insignificant (one > > memory access in the top most region). > But it will be significant inside multi-level loop. Actually, the problem is > exposed by a 8-level loop in a real Fortran application. > > > > > I also checked the attachments. What I see is GCC generates a better big > > loop body (a loop with high register pressure). GCC generated loop body has > > 50 x86-64 insns with 22 memory accesses vs 56 with 26 memory access for > > LLVM. > As far as how to spilling live range at loop boundary is concerned, gcc is > not very efficient. To make loop body not be the focus, when we remove two > live-range variables in original case, we can get same amount of memory > accesses for gcc and llvm, both of which generates 9 register spills in loop > body. I attached modified case and assembly files. For the modified case, allocation of llvm and gcc remain nearly the same for loop body, but as to 'lv0', llvm gets better spill/split result over gcc.
