http://gcc.gnu.org/bugzilla/show_bug.cgi?id=54294
--- Comment #5 from Jakub Jelinek <jakub at gcc dot gnu.org> 2012-08-17 10:29:20 UTC --- Created attachment 28041 --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=28041 gcc48-pr54294.patch Seems my other PR42728 patch fixes the -fcompare-debug failure and is IMHO desirable in any case. What is going on here is that we have: (insn 14254 14253 512 423 (set (reg:DI 1196 [ ei ]) (sign_extend:DI (subreg:SI (reg:DI 1196 [ ei ]) 0))) 2 {*extendsidi2_1} (nil)) insn where (subreg:SI (reg:DI 1196 [ ei ]) 0) is replaced by (reg:SI 2915). But fwprop is df_set_flags (DF_DEFER_INSN_RESCAN);. When replacing the next insn, which without -g turns to be another set, but with -g there are several debug insns in between and next insn is the first of them, we take a shortcut. As rescan is deferred, we see both 1196 and 2915 regs as "uses" of def_insn. The shortcut performs a rtx_equal_p check on DF_REF_REG (use) which fails though, as DF_REF_REG (use) is the subreg, thus we replace the subreg in the immediately following insn. But when trying to replace it in following debug insns and with -g in the first following non-debug insn, it is not NEXT_INSN, shortcut is not taken and we do both the rtx_equal_p check (fine, DF_REF_REG is again a subreg), but also do use_killed_between including def_insn, and that calls local_ref_killed_between_p, which compares DF_REF_REGNO instead of using rtx_equal_p and therefore says that the (already non-existing) use in def_insn is killed by the set in the def_insn. So, IMHO we should possibly replace the rtx_equal_p checks with simple DF_REF_REGNO comparisons instead, which will be more conservatively correct, but we risk code quality regressions. So, perhaps another thing to do could be that when walking uses of def_insn in all_uses_available_at, we'd double check if the reg in question is really still used in the insn pattern (perhaps conditionally on whether the insn is still marked for rescanning or not).
