On Wed, Sep 14, 2016 at 12:11:50PM -0600, Jeff Law wrote:
> On 09/14/2016 07:04 AM, Segher Boessenkool wrote:
> >>I'd
> >>probably start by fixing the dataflow issues and see if that fixes the
> >>regrename thing as a side effect.
> >
> >Have you seen https://gcc.gnu.org/ml/gcc-patches/2016-07/msg00091.html ?
> I missed it.
Yeah thought so, too much stuff in flight here.
> My interpretation....
>
> The uses at each "strange" exit fixing the first issue with
> shrink-wrapping definitely sounds like we've got a dataflow problem of
> some sort.
>
> If you think about it, conceptually we want the return insn to make the
> callee saved registers "used" so that DCE, regrename and friends don't
> muck with them. The fact that we don't is as much never having to care
> all that much until recently.
(There is no return insn at those exits; these are exits *without*
successor block, not the exit block).
It is puzzling to me why adding USEs for just the registers that *are*
separately shrink-wrapped does not work; only all those that *could* be
shrink-wrapped does. Do you have any idea about that?
> I continue to wonder if we could add something similar to
> CALL_INSN_FUNCTION_USAGE where we attach uses for all the call-saved
> registers onto a return insn. We would attach them at the end of
> prologue/epilogue generation and only attach those where were live
> somewhere in the code.
Maybe adding the new insns to the {pro,epi}logue_insn_hash will help
something. Or maybe it will blow up spectacularly. Will know in a
bit.
> By deferring that step until after prologue/epilogue generation we
> shouldn't cause unnecessary register saves/restores.
Hrm. I'll see about that as well.
> I pondered just doing it for the separately wrapped components on that
> particular path, but I've yet to convince myself that's actually correct.
If that is not correct, how is the status quo correct? That is what
puzzles me above, too.
> Bernd knows the regrename code better than anyone. Is there any way the
> two of you could work together to try and track down what's going on in
> the hash_map_rand case? Even throwing in some more debug stuff might
> help narrow things down since it's renaming something to a non-volatile,
> non-separately shrink wrapped register that's causing problems.
Okay with me, I could certainly use his help. I'll try the above things
first though, so not before friday.
> Can we agree that there's a set of targets that will improve and a set
> that are harmed? And that to enable regrename by default we need to
> either better describe the pipeline characteristics we're optimizing for
> or a well defined way for targets to turn it off?
There is a well-defined way to turn it off, via common/config/*/*-common.c ,
TARGET_OPTION_OPTIMIZATION_TABLE. We disagree on whether most targets will
want it enabled, i.e. whether it should (eventually) be enabled by default.
Segher
