On Mon, Jan 05, 2015 at 10:23:57PM +0100, Jakub Jelinek wrote:
> On Mon, Jan 05, 2015 at 03:16:46PM -0500, John David Anglin wrote:
> > I think there may be one situation after reload that's not handled
> > by the above. frame_read is only used for const calls. What about
> > the situation where we have a non const sibcall and the frame pointer
> > isn't eliminated?
>
> After reload DSE is run after threading prologues/epilogues, end
> the prologue/epilogue sequences usually contain wild reads, e.g.
> (mem:BLK (scratch)) in some insn etc.
> Do you have some particular testcase in mind?
>
> That said, DSE after reload is much more limited than the DSE before reload,
> so is less important, so perhaps even
> if ((HARD_FRAME_POINTER_IS_ARG_POINTER || reload_completed)
> && SIBLING_CALL_P (insn))
> {
> add_wild_read (bb_info);
>
> return;
>
> }
> might be good enough.
Or you could e.g. do the
if (HARD_FRAME_POINTER_IS_ARG_POINTER
&& !reload_completed
&& SIBLING_CALL_P (insn))
{ add_wild_read (bb_info); return; }
case first, then compute const_call and memset_call,
if (const_call || memset_call)
{ current_big_block }
else if (reload_completed && SIBLING_CALL_P (insn))
add_wild_read (bb_info);
else
add_non_frame_wild_read (bb_info);
That way, you would not punish const/memset calls unnecessarily after reload
when they are sibling calls.
Jakub
