Hi Martin, On 2018-09-03 06:01 AM, Martin Jambor wrote: > Hi, > > On Fri, Aug 31 2018, Michael Ploujnikov wrote: >> I've done some more digging into the current uses of >> numbered_clone_function_name and checked if any tests fail if I change >> it to suffixed_function_name: >> >> - gcc/cgraphclones.c: DECL_NAME (new_decl) = numbered_clone_function_name >> (thunk->decl, "artificial_thunk"); >> - no new tests fail, inconclusive >> - and despite the comment on redirect_callee_duplicating_thunks >> about "one or more" duplicates it doesn't seem like >> duplicate_thunk_for_node would be called more than once for each >> node, assuming each node is named uniquely, but I'm far from an >> expert in this area > > The comment means that if there is a chain of thunks, the method clones > all of them. Nevertheless, you need name numbering here for the same > reason why you need them for constprop. > > >> - gcc/omp-expand.c: DECL_NAME (kern_fndecl) = >> numbered_clone_function_name (kern_fndecl, "kernel"); >> - no new tests fail, inconclusive >> - I didn't see (and couldn't figure out a way to get) any of the >> existing omp/acc tests actually exercise this codeptah > > I guess this one should not need it. Build with > --enable-offload-targets=hsa and run gomp.exp to try yourself. I can > run run-time HSA tests for you if you want. > > Martin >
I've tried building with numbered_clone_function_name replaced by suffixed_function_name and with --enable-offload-targets=hsa and didn't see any errors in gomp.exp. I don't have a readily available HSA setup so if you could do a quick test, I would really appreciate it! - Michael
signature.asc
Description: OpenPGP digital signature
