> On 2015.01.20 at 21:04 +0100, Jan Hubicka wrote: > > this patch fixes ICE in ipa_merge_profiles on speculative edges. > > > > Bootstrapped/regtested x86_64-linux, comitted. Also tested by Markus on > > Firefox build. > > This needs one additional fix. See below. > > Otherwise it will crash: > > lto1: internal compiler error: Segmentation fault > 0xa12f6f crash_signal > ../../gcc/gcc/toplev.c:381 > 0x88b190 ipa_merge_profiles(cgraph_node*, cgraph_node*) > ../../gcc/gcc/ipa-utils.c:637 > 0x603722 lto_cgraph_replace_node > ../../gcc/gcc/lto/lto-symtab.c:124 > 0x604cf3 lto_symtab_merge_symbols_1 > ../../gcc/gcc/lto/lto-symtab.c:619 > 0x604cf3 lto_symtab_merge_symbols() > ../../gcc/gcc/lto/lto-symtab.c:647 > 0x5fa52e read_cgraph_and_symbols > ../../gcc/gcc/lto/lto.c:3109 > 0x5fa52e lto_main() > ../../gcc/gcc/lto/lto.c:3436 > Please submit a full bug report, > with preprocessed source if appropriate.
I see, thanks. It means that we do have comdats that diverge in their call statements. I guess we ought to just cancel merging in that case, otherwise the profile can be completely off :( I will need to write some call target compare then. So far the code assumes that if #of BBs match then the bodies match and proceeds with merging. Martin, perhaps we can re-use some of ipa-icf infrastructure here to quickly check that at least CFG shapes and call targets match? I will commit this hack for now; hopefully it is only infrequent side case. Honza
