On Mon, Nov 19, 2012 at 10:50 PM, Eric Botcazou wrote: >> The root cause is the bad REG_EQUAL note. I think the most robust >> solution is to make the webizer re-compute notes before renaming. >> Patch for that is attached. > > Thanks for the analysis. However... > >> Ciao! >> Steven >> >> >> PR rtl-optimization/55006 >> * web.c (web_main): Add the DF_NOTE problem, and explain whatfor. >> >> Index: web.c >> =================================================================== >> --- web.c (revision 193454) >> +++ web.c (working copy) >> @@ -335,9 +335,16 @@ web_main (void) >> unsigned int uses_num = 0; >> rtx insn; >> >> + /* Add the flags and problems we need. The DF_EQ_NOTES flag is set so >> + that uses of registers in REG_EQUAL and REG_EQUIV notes are included >> + in the web that their DEF belongs to, so that these uses are also >> + properly renamed. The DF_NOTE problem is added to make sure that >> + all notes are up-to-date and valid: Re-computing the notes problem >> + also cleans up all dead REG_EQUAL notes. */ >> df_set_flags (DF_NO_HARD_REGS + DF_EQ_NOTES); >> df_set_flags (DF_RD_PRUNE_DEAD_DEFS); >> df_chain_add_problem (DF_UD_CHAIN); >> + df_note_add_problem (); >> df_analyze (); >> df_set_flags (DF_DEFER_INSN_RESCAN); > > ... that's not very satisfactory, as web doesn't use the DF_NOTE problem, so > adding it just to clean things up in the other kind of notes is weird.
True. > Can't we arrange to clean up the REG_EQUAL/REG_EQUIV notes when we use them, > i.e. when DF_EQ_NOTES is set? That could be done, yes. Cleaning up the REG_EQ* notes requires liveness at the insn level, so it'd require a bigger re-organization of the code. Perhaps adding a new pass (conditional on DF_EQ_NOTES) over all insn in df_lr_finalize, tracking liveness and calling df_remove_dead_eq_notes on each insn. But most passes that use the REG_EQ* notes don't set the DF_EQ_NOTES flag. Perhaps df_note_add_problem should imply setting DF_EQ_NOTES? Ciao! Steven
