Re: [PATCH 8/9] shrink-wrap: shrink-wrapping for separate components

[Jeff Law]

On 09/14/2016 07:38 AM, Segher Boessenkool wrote:

On Mon, Sep 12, 2016 at 12:02:50PM -0600, Jeff Law wrote:

As a final optimisation, if a block needs a prologue and its immediate
dominator has the block as a post-dominator, the dominator gets the
prologue as well.

So why not just put it in the idom and not in the dominated block?


That's what it does :-)

Then I must have mis-parsed.  Thanks for clarifying.


"As a final optimisation, if a block needs a prologue and its immediate
dominator has the block as a post-dominator, ***that immediate dominator***
gets the prologue as well."

That is clearer I hope :-)

It is :-)




Hmm, then explain again why DCE is mucking up?  I don't immediately see
how EPILOGUE_BEG notes come into play with DCE.  It seems to rely on the
DF data and AFAICT DF only cares about the EPILOGUE_BEG note in
can_move_insns_across which shouldn't be used by DCE.


The register restore *is* dead code, but we need to have the same CFI
for all convergent paths.
OK.   I think I was conflating multiple issues.  So we need to keep the 
restore alive so that we have the same CFI across those paths, even 
though it appears dead on one or more paths.


I think this points us back to what you were experimenting with to 
address the regrename problems -- specifically creating "uses" at those 
key points.  That solves the DCE problem as well as one of the regrename 
problems, right?




Whether or not an edge needs a prologue or epilogue is a function not
just of the state at the head or tail of the edge, but instead is a
function of global dataflow propagation?  Thus we can't disqualify until
after we've done the dataflow propagation?  Right?


We can figure out before we decide what blocks need what components, what
edges can not get a prologue or epilogue for which components.  This
complicates the selection algorithm a whole lot, for not much gain that
I have seen so far, so I just give up in the cases that end up "bad".
OK.  I'll drop it :-)  It was more a mental exercise in understanding 
then something I think needed to change.


Jeff

Re: [PATCH 8/9] shrink-wrap: shrink-wrapping for separate components

[Segher Boessenkool]

On Mon, Sep 12, 2016 at 12:02:50PM -0600, Jeff Law wrote:
> >>>As a final optimisation, if a block needs a prologue and its immediate
> >>>dominator has the block as a post-dominator, the dominator gets the
> >>>prologue as well.
> >>So why not just put it in the idom and not in the dominated block?
> >
> >That's what it does :-)
> Then I must have mis-parsed.  Thanks for clarifying.

"As a final optimisation, if a block needs a prologue and its immediate
dominator has the block as a post-dominator, ***that immediate dominator***
gets the prologue as well."

That is clearer I hope :-)

> Hmm, then explain again why DCE is mucking up?  I don't immediately see 
> how EPILOGUE_BEG notes come into play with DCE.  It seems to rely on the 
> DF data and AFAICT DF only cares about the EPILOGUE_BEG note in 
> can_move_insns_across which shouldn't be used by DCE.

The register restore *is* dead code, but we need to have the same CFI
for all convergent paths.

> >>Consider using auto_sbitmap rather than manually managing
> >>allocation/releasing of the per-block structures.  In fact, can't all of
> >>SW become a class and we lose the explicit init/fini routines in favor
> >>of a ctor/dtor?
> >
> >Yes, you can always add indirection.  I do not think the code becomes
> >more readable that way (quite the opposite).  Explicit is *good*.
> The GCC project is moving away from this kind of explicit 
> allocation/deallocation and more towards a RAII.  Unless there is a 
> clear need for the explicit allocation/deallocation, please put this 
> stuff into a class with an appropriate ctor/dtor.
> 
> FWIW, I was a big opponent of how much stuff happens "behind your back" 
> with some languages (including C++).  But over the last few years my 
> personal stance has softened considerably after seeing how cleanly RAII 
> solves certain problems.

We then still cannot get rid of SW, which is a convenience macro to do
a nasty cast on bb->aux.  If bb->aux was some pretty class hierarchy,
easy to use and all that, I would of course agree with your suggestion.
But as it is it is just a bare pointer, so the less we hide the safer
it is.

> >>For the PPC R0 vs LR is the only thing that causes disqualification
> >>right?
> >
> >Currently, yes.
> >
> >>Can't that be handled when we build the set of components we
> >>want to insert for each edge/block?  Is there some advantage to handling
> >>disqualifications after all the potential insertion points have been
> >>handled?
> >
> >We do not know if an edge needs a prologue, epilogue, or neither, until
> >we have decided whether *both* ends of that edge want the component active
> >or not.
> Right.  Hmm, maybe I'm not asking the question clearly.
> 
> Whether or not an edge needs a prologue or epilogue is a function not 
> just of the state at the head or tail of the edge, but instead is a 
> function of global dataflow propagation?  Thus we can't disqualify until 
> after we've done the dataflow propagation?  Right?

We can figure out before we decide what blocks need what components, what
edges can not get a prologue or epilogue for which components.  This
complicates the selection algorithm a whole lot, for not much gain that
I have seen so far, so I just give up in the cases that end up "bad".

It is not easy at all to see what edges will need to get a *logue,
because not always both blocks that edge connects are in the same
dominator subtree (or tree even, for an epilogue-aware placement
algorithm, but this patch doesn't do that yet; it's a more minor
optimisation, only reduces code size a little).


Segher

Re: [PATCH 8/9] shrink-wrap: shrink-wrapping for separate components

[Jeff Law]

On 09/09/2016 03:57 PM, Segher Boessenkool wrote:

On Thu, Sep 08, 2016 at 12:34:07PM -0600, Jeff Law wrote:

On 07/31/2016 07:42 PM, Segher Boessenkool wrote:


Deciding what blocks should run with a certain component active so that
the total cost of executing the prologues (and epilogues) is optimal, is
not a computationally feasible problem.

Really?  It's just a dataflow problem is it not and one that ought to
converge very quickly I'd think.  Or is it more a function of having to
run it on so many independent components?  I'm still pondering the value
of having every GPR be an independent component :-)


The cost function (as a function of which BBs runs with a certain component
enabled) is not monotonic: the cost for having a component active for a
subset of BBs can be higher, the same, or equal to the cost for all such
nodes (where "cost" means "how often do we execute this prologue").

Understood.   You covered this reasonably well in another reply.  Thanks.


Cross jumping is rather simplistic, so I'm not surprised that it doesn't
catch all this stuff.


I hoped it would, so I could have so much simpler code.  Sniff.
In general, I think our ability to identify and de-duplicate code is 
poor at best, especially at the RTL level.  It's never been a major area 
of focus.   So, yea.  Sniff.






As a final optimisation, if a block needs a prologue and its immediate
dominator has the block as a post-dominator, the dominator gets the
prologue as well.

So why not just put it in the idom and not in the dominated block?


That's what it does :-)

Then I must have mis-parsed.  Thanks for clarifying.





void
thread_prologue_and_epilogue_insns (void)
{
+  if (optimize > 1)
+{
+  df_live_add_problem ();
+  df_live_set_all_dirty ();
+}

Perhaps conditional on separate shrink wrapping?


Actually, I think we need to do this once more, one of the times always
(also when only using "regular" shrink-wrapping), because otherwise the
info in the dump file is out of whack.  Everything is okay once the next
pass starts, of course.  I'll have a look.
OK.  It struck me as a bit odd, so just a verify on your side that it 
was intended is fine with me.






@@ -5932,6 +5936,13 @@ thread_prologue_and_epilogue_insns (void)

  try_shrink_wrapping (_edge, prologue_seq);

+  df_analyze ();
+  try_shrink_wrapping_separate (entry_edge->dest);
+  if (crtl->shrink_wrapped_separate)
+prologue_seq = make_prologue_seq ();

Perhaps push the df_analyze call into try_shrink_wrapping_separate?


Yeah possibly.

Your call.





ANd if it was successful, do you need to update the DF information?  Is
that perhaps the cause of some of the issues we're seeing with DCE,
regrename, the scheduler?


See above.  The DF info is correct when the next pass starts (or ends,
I do not remember; the dump file does show correct information).
Hmm, then explain again why DCE is mucking up?  I don't immediately see 
how EPILOGUE_BEG notes come into play with DCE.  It seems to rely on the 
DF data and AFAICT DF only cares about the EPILOGUE_BEG note in 
can_move_insns_across which shouldn't be used by DCE.








Related,
consider using an enum rather than magic constants in the target bits (I
noted seeing component #0 as a magic constant in the ppc implementation).


But 0 is the hard register used there :-)

Oh, missed that.  Nevermind :-)


So it seems like there's a toplevel list of components that's owned by
the target and state at each block components that are owned by the
generic code.  That's fine.  Just make sure we doc that the toplevel
list of components is allocated by the backend (and where does it get
freed?)


Every sbitmap is owned by the separate shrink-wrapping pass, not by the
target.  As the documentation says:

I'm specifically referring to:

+@deftypefn {Target Hook} sbitmap 
TARGET_SHRINK_WRAP_GET_SEPARATE_COMPONENTS (void)


+@deftypefn {Target Hook} sbitmap TARGET_SHRINK_WRAP_COMPONENTS_FOR_BB 
(basic_block)


Which in the rs6000 implementation allocate and return an sbitmap.  I 
don't see anywhere we could reasonably free them in the target with the 
existing hooks, so it seems like they have to be free'd by the generic 
code. So ownership of those sbitmaps isn't particularly clear.






Consider using auto_sbitmap rather than manually managing
allocation/releasing of the per-block structures.  In fact, can't all of
SW become a class and we lose the explicit init/fini routines in favor
of a ctor/dtor?


Yes, you can always add indirection.  I do not think the code becomes
more readable that way (quite the opposite).  Explicit is *good*.
The GCC project is moving away from this kind of explicit 
allocation/deallocation and more towards a RAII.  Unless there is a 
clear need for the explicit allocation/deallocation, please put this 
stuff into a class with an appropriate ctor/dtor.


FWIW, I was a big opponent of how much stuff happens "behind your back" 
with some languages (including C++).  But over the 

Re: [PATCH 8/9] shrink-wrap: shrink-wrapping for separate components

[Segher Boessenkool]

On Thu, Sep 08, 2016 at 12:34:07PM -0600, Jeff Law wrote:
> On 07/31/2016 07:42 PM, Segher Boessenkool wrote:
> >
> >Deciding what blocks should run with a certain component active so that
> >the total cost of executing the prologues (and epilogues) is optimal, is
> >not a computationally feasible problem.
> Really?  It's just a dataflow problem is it not and one that ought to 
> converge very quickly I'd think.  Or is it more a function of having to 
> run it on so many independent components?  I'm still pondering the value 
> of having every GPR be an independent component :-)

The cost function (as a function of which BBs runs with a certain component
enabled) is not monotonic: the cost for having a component active for a
subset of BBs can be higher, the same, or equal to the cost for all such
nodes (where "cost" means "how often do we execute this prologue").

> >Now all that is left is inserting prologues and epilogues on all edges
> >that jump into resp. out of the "active" set of blocks.  Often we need
> >to insert some components' prologues (or epilogues) on all edges into
> >(or out of) a block.  In theory cross-jumping can unify all such, but
> >in practice that often fails; besides, that is a lot of work.  So in
> >this case we insert the prologue and epilogue components at the "head"
> >or "tail" of a block, instead.
> Cross jumping is rather simplistic, so I'm not surprised that it doesn't 
> catch all this stuff.

I hoped it would, so I could have so much simpler code.  Sniff.

> >As a final optimisation, if a block needs a prologue and its immediate
> >dominator has the block as a post-dominator, the dominator gets the
> >prologue as well.
> So why not just put it in the idom and not in the dominated block? 

That's what it does :-)


> > void
> > thread_prologue_and_epilogue_insns (void)
> > {
> >+  if (optimize > 1)
> >+{
> >+  df_live_add_problem ();
> >+  df_live_set_all_dirty ();
> >+}
> Perhaps conditional on separate shrink wrapping?

Actually, I think we need to do this once more, one of the times always
(also when only using "regular" shrink-wrapping), because otherwise the
info in the dump file is out of whack.  Everything is okay once the next
pass starts, of course.  I'll have a look.

> >@@ -5932,6 +5936,13 @@ thread_prologue_and_epilogue_insns (void)
> >
> >   try_shrink_wrapping (_edge, prologue_seq);
> >
> >+  df_analyze ();
> >+  try_shrink_wrapping_separate (entry_edge->dest);
> >+  if (crtl->shrink_wrapped_separate)
> >+prologue_seq = make_prologue_seq ();
> Perhaps push the df_analyze call into try_shrink_wrapping_separate?

Yeah possibly.

> ANd if it was successful, do you need to update the DF information?  Is 
> that perhaps the cause of some of the issues we're seeing with DCE, 
> regrename, the scheduler?

See above.  The DF info is correct when the next pass starts (or ends,
I do not remember; the dump file does show correct information).

> Consider allowing the target to provide a mapping from the component to 
> a symbolic name of some kind and using that in the dumps.

Yeah that will help, once we do more than just GPRs anyway -- not everyone
remembers the GCC register #s for all registers ;-)

> Related, 
> consider using an enum rather than magic constants in the target bits (I 
> noted seeing component #0 as a magic constant in the ppc implementation).

But 0 is the hard register used there :-)

And since we now are C++, I cannot use enums as integers.  Sigh.

The meaning of "0" should of course be documented.

> So it seems like there's a toplevel list of components that's owned by 
> the target and state at each block components that are owned by the 
> generic code.  That's fine.  Just make sure we doc that the toplevel 
> list of components is allocated by the backend (and where does it get 
> freed?)

Every sbitmap is owned by the separate shrink-wrapping pass, not by the
target.  As the documentation says:

+@deftypefn {Target Hook} void TARGET_SHRINK_WRAP_SET_HANDLED_COMPONENTS (sbitma
+Mark the components in the parameter as handled, so that the
+@code{prologue} and @code{epilogue} named patterns know to ignore those
+components.  The target code should not hang on to the @code{sbitmap}, it
+will be deleted after this call.
+@end deftypefn

> Consider using auto_sbitmap rather than manually managing 
> allocation/releasing of the per-block structures.  In fact, can't all of 
> SW become a class and we lose the explicit init/fini routines in favor 
> of a ctor/dtor?

Yes, you can always add indirection.  I do not think the code becomes
more readable that way (quite the opposite).  Explicit is *good*.

The init/fini code is small, and that is not an accident.

> >+  /* Find which prologue resp. epilogue components are needed for all
> >+ predecessor edges to this block.  */
> Nit.  Avoid ".resp".  I actually had to look that abbreviation up :-) 
> Being a bit more verbose in the comments would be preferred over ".resp".

Re: [PATCH 8/9] shrink-wrap: shrink-wrapping for separate components

[Jeff Law]

On 07/31/2016 07:42 PM, Segher Boessenkool wrote:


Deciding what blocks should run with a certain component active so that
the total cost of executing the prologues (and epilogues) is optimal, is
not a computationally feasible problem.
Really?  It's just a dataflow problem is it not and one that ought to 
converge very quickly I'd think.  Or is it more a function of having to 
run it on so many independent components?  I'm still pondering the value 
of having every GPR be an independent component :-)


ISTM this adds a fair amount of complexity to the implementation in that 
prologues and epilogues for a particular component can run more than 
once.  Can you give a concrete example where this happens so that we can 
all understand it better?


If we keep this aspect of the implementation it seems like a note in the 
developer section would be in order.  It's certainly non-intuitive.


I only glanced over the code that seems related to this aspect of the 
implementation.  If we decide to go forward, I'd like to look at it 
again more closely.




Now all that is left is inserting prologues and epilogues on all edges
that jump into resp. out of the "active" set of blocks.  Often we need
to insert some components' prologues (or epilogues) on all edges into
(or out of) a block.  In theory cross-jumping can unify all such, but
in practice that often fails; besides, that is a lot of work.  So in
this case we insert the prologue and epilogue components at the "head"
or "tail" of a block, instead.
Cross jumping is rather simplistic, so I'm not surprised that it doesn't 
catch all this stuff.




As a final optimisation, if a block needs a prologue and its immediate
dominator has the block as a post-dominator, the dominator gets the
prologue as well.
So why not just put it in the idom and not in the dominated block? 
Doesn't this just end up running that component's prologue twice?




2016-06-07  Segher Boessenkool  

* function.c (thread_prologue_and_epilogue_insns): Recompute the
live info.  Call try_shrink_wrapping_separate.  Compute the
prologue_seq afterwards, if it has possibly changed.  Compute the
split_prologue_seq and epilogue_seq later, too.
* shrink-wrap.c: #include cfgbuild.h.
(dump_components): New function.
(struct sw): New struct.
(SW): New function.
(init_separate_shrink_wrap): New function.
(fini_separate_shrink_wrap): New function.
(place_prologue_for_one_component): New function.
(spread_components): New function.
(disqualify_problematic_components): New function.
(emit_common_heads_for_components): New function.
(emit_common_tails_for_components): New function.
(insert_prologue_epilogue_for_components): New function.
(try_shrink_wrapping_separate): New function.
* shrink-wrap.h: Declare try_shrink_wrapping_separate.
---
 gcc/function.c|  15 +-
 gcc/shrink-wrap.c | 715 ++
 gcc/shrink-wrap.h |   1 +
 3 files changed, 729 insertions(+), 2 deletions(-)

diff --git a/gcc/function.c b/gcc/function.c
index bba0705..390e9a6 100644
--- a/gcc/function.c
+++ b/gcc/function.c
@@ -5912,6 +5912,12 @@ make_epilogue_seq (void)
 void
 thread_prologue_and_epilogue_insns (void)
 {
+  if (optimize > 1)
+{
+  df_live_add_problem ();
+  df_live_set_all_dirty ();
+}

Perhaps conditional on separate shrink wrapping?


@@ -5922,9 +5928,7 @@ thread_prologue_and_epilogue_insns (void)
   edge entry_edge = single_succ_edge (ENTRY_BLOCK_PTR_FOR_FN (cfun));
   edge orig_entry_edge = entry_edge;

-  rtx_insn *split_prologue_seq = make_split_prologue_seq ();
   rtx_insn *prologue_seq = make_prologue_seq ();
-  rtx_insn *epilogue_seq = make_epilogue_seq ();

   /* Try to perform a kind of shrink-wrapping, making sure the
  prologue/epilogue is emitted only around those parts of the
@@ -5932,6 +5936,13 @@ thread_prologue_and_epilogue_insns (void)

   try_shrink_wrapping (_edge, prologue_seq);

+  df_analyze ();
+  try_shrink_wrapping_separate (entry_edge->dest);
+  if (crtl->shrink_wrapped_separate)
+prologue_seq = make_prologue_seq ();

Perhaps push the df_analyze call into try_shrink_wrapping_separate?

ANd if it was successful, do you need to update the DF information?  Is 
that perhaps the cause of some of the issues we're seeing with DCE, 
regrename, the scheduler?





diff --git a/gcc/shrink-wrap.c b/gcc/shrink-wrap.c
index b85b1c3..643e375 100644
--- a/gcc/shrink-wrap.c
+++ b/gcc/shrink-wrap.c
@@ -34,6 +34,7 @@ along with GCC; see the file COPYING3.  If not see
 #include "output.h"
 #include "tree-pass.h"
 #include "cfgrtl.h"
+#include "cfgbuild.h"
 #include "params.h"
 #include "bb-reorder.h"
 #include "shrink-wrap.h"
@@ -1006,3 +1007,717 @@ try_shrink_wrapping (edge *entry_edge, rtx_insn 
*prologue_seq)
   BITMAP_FREE (bb_with);
   free_dominance_info