Re: [PATCH] Unswitching outer loops.

[Richard Biener <richard dot guenther at gmail dot com>]

On Wed, Sep 30, 2015 at 12:46 PM, Yuri Rumyantsev <ysrumyan@gmail.com> wrote:
> Hi Richard,
>
> I re-designed outer loop unswitching using basic idea of 23855 patch -
> hoist invariant guard if loop is empty without guard. Note that this
> was added to loop unswitching pass with simple modifications - using
> another loop iterator etc.
>
> Bootstrap and regression testing did not show any new failures.
> What is your opinion?

Overall it looks good.  Some comments below - a few more testcases would
be nice as well.

+  /* Loop must not be infinite.  */
+  if (!finite_loop_p (loop))
+    return false;

why's that?

+  body = get_loop_body_in_dom_order (loop);
+  for (i = 0; i < loop->num_nodes; i++)
+    {
+      if (body[i]->loop_father != loop)
+       continue;
+      if (!empty_bb_without_guard_p (loop, body[i]))

I wonder if there is a better way to iterate over the interesting
blocks and PHIs
we need to check for side-effects (and thus we maybe can avoid gathering
the loop in DOM order).

+      FOR_EACH_SSA_TREE_OPERAND (name, stmt, op_iter, SSA_OP_DEF)
+       {
+         if (may_be_used_outside

may_be_used_outside can be hoisted above the loop.  I wonder if we can take
advantage of loop-closed SSA form here (and the fact we have a single exit
from the loop).  Iterating over exit dest PHIs and determining whether the
exit edge DEF is inside the loop part it may not be should be enough.

+  gcc_assert (single_succ_p (pre_header));

that should be always true.

+  gsi_remove (&gsi, false);
+  bb = guard->dest;
+  remove_edge (guard);
+  /* Update dominance for destination of GUARD.  */
+  if (EDGE_COUNT (bb->preds) == 0)
+    {
+      basic_block s_bb;
+      gcc_assert (single_succ_p (bb));
+      s_bb = single_succ (bb);
+      delete_basic_block (bb);
+      if (single_pred_p (s_bb))
+       set_immediate_dominator (CDI_DOMINATORS, s_bb, single_pred (s_bb));

all this massaging should be simplified by leaving it to CFG cleanup by
simply adjusting the CONDs condition to always true/false.  There is
gimple_cond_make_{true,false} () for this (would be nice to have a variant
taking a bool).

+  new_edge = make_edge (pre_header, exit->dest, flags);
+  if (fix_dom_of_exit)
+    set_immediate_dominator (CDI_DOMINATORS, exit->dest, pre_header);
+  update_stmt (gsi_stmt (gsi));

the update_stmt should be not necessary, it's done by gsi_insert_after already.

+  /* Add NEW_ADGE argument for all phi in post-header block.  */
+  bb = exit->dest;
+  for (gphi_iterator gsi = gsi_start_phis (bb);
+       !gsi_end_p (gsi); gsi_next (&gsi))
+    {
+      gphi *phi = gsi.phi ();
+      /* edge_iterator ei; */
+      tree arg;
+      if (virtual_operand_p (gimple_phi_result (phi)))
+       {
+         arg = PHI_ARG_DEF_FROM_EDGE (phi, loop_preheader_edge (loop));
+         add_phi_arg (phi, arg, new_edge, UNKNOWN_LOCATION);
+       }
+      else
+       {
+         /* Use exit edge argument.  */
+         arg = PHI_ARG_DEF_FROM_EDGE (phi, exit);
+         add_phi_arg (phi, arg, new_edge, UNKNOWN_LOCATION);

Hum.  How is it ok to use the exit edge argument for the edge that skips
the loop?  Why can't you always use the pre-header edge value?
That is, if we have

 for(i=0;i<m;++i)
   {
     if (n > 0)
    {
     for (;;)
       {
       }
     }
   }
  ... = i;

then i is used after the loop and the correct value to use if
n > 0 is false is '0'.  Maybe this way we can also relax
what check_exit_phi does?  IMHO the only restriction is
if sth defined inside the loop before the header check for
the inner loop is used after the loop.

Thanks,
Richard.

> Thanks.
>
> ChangeLog:
> 2015-09-30  Yuri Rumyantsev  <ysrumyan@gmail.com>
>
> * tree-ssa-loop-unswitch.c: Include "gimple-iterator.h" and
> "cfghooks.h", add prototypes for introduced new functions.
> (tree_ssa_unswitch_loops): Use from innermost loop iterator, move all
> checks on ability of loop unswitching to tree_unswitch_single_loop;
> invoke tree_unswitch_single_loop or tree_unswitch_outer_loop depending
> on innermost loop check.
> (tree_unswitch_single_loop): Add all required checks on ability of
> loop unswitching under zero recursive level guard.
> (tree_unswitch_outer_loop): New function.
> (find_loop_guard): Likewise.
> (empty_bb_without_guard_p): Likewise.
> (used_outside_loop_p): Likewise.
> (hoist_guard): Likewise.
> (check_exit_phi): Likewise.
>
>    gcc/testsuite/ChangeLog:
> * gcc.dg/loop-unswitch-2.c: New test.
>
> 2015-09-16 11:26 GMT+03:00 Richard Biener <richard.guenther@gmail.com>:
>> Yeah, as said, the patch wasn't fully ready and it also felt odd to do
>> this hoisting in loop header copying.  Integrating it
>> with LIM would be a better fit eventually.
>>
>> Note that we did agree to go forward with your original patch just
>> making it more "generically" perform outer loop
>> unswitching.  Did you explore that idea further?
>>
>>
>>
>> On Tue, Sep 15, 2015 at 6:00 PM, Yuri Rumyantsev <ysrumyan@gmail.com> wrote:
>>> Thanks Richard.
>>>
>>> I found one more issue that could not be fixed simply. In 23855 you
>>> consider the following test-case:
>>> void foo(int *ie, int *je, double *x)
>>> {
>>>   int i, j;
>>>   for (j=0; j<*je; ++j)
>>>     for (i=0; i<*ie; ++i)
>>>       x[i+j] = 0.0;
>>> }
>>> and proposed to hoist up a check on *ie out of loop. It requires
>>> memref alias analysis since in general x and ie can alias (if their
>>> types are compatible - int *ie & int * x). Such analysis is performed
>>> by pre or lim passes. Without such analysis we can not hoist a test on
>>> non-zero for *ie out of loop using 238565 patch.
>>>  The second concern is that proposed copy header algorithm changes
>>> loop structure significantly and it is not accepted by vectorizer
>>> since latch is not empty (such transformation assumes loop peeling for
>>> one iteration. So I can propose to implement simple guard hoisting
>>> without copying header and tail blocks (if it is possible).
>>>
>>> I will appreciate you for any advice or help since without such
>>> hoisting we are not able to perform outer loop vectorization for
>>> important benchmark.
>>> and
>>>
>>> 2015-09-15 14:22 GMT+03:00 Richard Biener <richard.guenther@gmail.com>:
>>>> On Thu, Sep 3, 2015 at 6:32 PM, Yuri Rumyantsev <ysrumyan@gmail.com> wrote:
>>>>> Hi Richard,
>>>>>
>>>>> I started learning, tuning and debugging patch proposed in 23855 and
>>>>> discovered thta it does not work properly.
>>>>> So I wonder is it tested patch and it should work?
>>>>
>>>> I don't remember, but as it wasn't committed it certainly wasn't ready.
>>>>
>>>>> Should it accept for hoisting the following loop nest
>>>>>   for (i=0; i<n; i++) {
>>>>>     s = 0;
>>>>>     for (j=0; j<m; j++)
>>>>>       s += a[i] * b[j];
>>>>>     c[i] = s;
>>>>>   }
>>>>> Note that i-loop will nit be empty if m is equal to 0.
>>>>
>>>> if m is equal to 0 then we still have the c[i] = s store, no?  Of course
>>>> we could unswitch the outer loop on m == 0 but simple hoisting wouldn't work.
>>>>
>>>> Richard.
>>>>
>>>>> 2015-08-03 10:27 GMT+03:00 Richard Biener <richard.guenther@gmail.com>:
>>>>>> On Fri, Jul 31, 2015 at 1:17 PM, Yuri Rumyantsev <ysrumyan@gmail.com> wrote:
>>>>>>> Hi Richard,
>>>>>>>
>>>>>>> I learned your updated patch for 23825 and it is more general in
>>>>>>> comparison with my.
>>>>>>> I'd like to propose you a compromise - let's consider my patch only
>>>>>>> for force-vectorize outer loop only to allow outer-loop
>>>>>>> vecctorization.
>>>>>>
>>>>>> I don't see why we should special-case that if the approach in 23825
>>>>>> is sensible.
>>>>>>
>>>>>>> Note that your approach will not hoist invariant
>>>>>>> guards if loops contains something else except for inner-loop, i.e. it
>>>>>>> won't be empty for taken branch.
>>>>>>
>>>>>> Yes, it does not perform unswitching but guard hoisting.  Note that this
>>>>>> is originally Zdenek Dvoraks patch.
>>>>>>
>>>>>>> I also would like to answer on your last question - CFG cleanup is
>>>>>>> invoked to perform deletion of single-argument phi nodes from tail
>>>>>>> block through substitution - such phi's prevent outer-loop
>>>>>>> vectorization. But it is clear that such transformation can be done
>>>>>>> other pass.
>>>>>>
>>>>>> Hmm, I wonder why the copy_prop pass after unswitching does not
>>>>>> get rid of them?
>>>>>>
>>>>>>> What is your opinion?
>>>>>>
>>>>>> My opinion is that if we want to enhance unswitching to catch this
>>>>>> (or similar) cases then we should make it a lot more general than
>>>>>> your pattern-matching approach.  I see nothing that should prevent
>>>>>> us from considering unswitching non-innermost loops in general.
>>>>>> It should be only a cost consideration to not do non-innermost loop
>>>>>> unswitching (in addition to maybe a --param specifying the maximum
>>>>>> depth of a loop nest to unswitch).
>>>>>>
>>>>>> So my first thought when seeing your patch still holds - the patch
>>>>>> looks very much too specific.
>>>>>>
>>>>>> Richard.
>>>>>>
>>>>>>> Yuri.
>>>>>>>
>>>>>>> 2015-07-28 13:50 GMT+03:00 Richard Biener <richard.guenther@gmail.com>:
>>>>>>>> On Thu, Jul 23, 2015 at 4:45 PM, Yuri Rumyantsev <ysrumyan@gmail.com> wrote:
>>>>>>>>> Hi Richard,
>>>>>>>>>
>>>>>>>>> I checked that both test-cases from 23855 are sucessfully unswitched
>>>>>>>>> by proposed patch. I understand that it does not catch deeper loop
>>>>>>>>> nest as
>>>>>>>>>    for (i=0; i<10; i++)
>>>>>>>>>      for (j=0;j<n;j++)
>>>>>>>>>         for (k=0;k<20;k++)
>>>>>>>>>   ...
>>>>>>>>> but duplication of middle-loop does not look reasonable.
>>>>>>>>>
>>>>>>>>> Here is dump for your second test-case:
>>>>>>>>>
>>>>>>>>> void foo(int *ie, int *je, double *x)
>>>>>>>>> {
>>>>>>>>>   int i, j;
>>>>>>>>>   for (j=0; j<*je; ++j)
>>>>>>>>>     for (i=0; i<*ie; ++i)
>>>>>>>>>       x[i+j] = 0.0;
>>>>>>>>> }
>>>>>>>>> grep -i unswitch t6.c.119t.unswitch
>>>>>>>>> ;; Unswitching outer loop
>>>>>>>>
>>>>>>>> I was saying that why go with a limited approach when a patch (in
>>>>>>>> unknown state...)
>>>>>>>> is available that does it more generally?  Also unswitching is quite
>>>>>>>> expensive compared
>>>>>>>> to "moving" the invariant condition.
>>>>>>>>
>>>>>>>> In your patch:
>>>>>>>>
>>>>>>>> +  if (!nloop->force_vectorize)
>>>>>>>> +    nloop->force_vectorize = true;
>>>>>>>> +  if (loop->safelen != 0)
>>>>>>>> +    nloop->safelen = loop->safelen;
>>>>>>>>
>>>>>>>> I see no guard on force_vectorize so = true looks bogus here.  Please just use
>>>>>>>> copy_loop_info.
>>>>>>>>
>>>>>>>> +  if (integer_nonzerop (cond_new))
>>>>>>>> +    gimple_cond_set_condition_from_tree (cond_stmt, boolean_true_node);
>>>>>>>> +  else if (integer_zerop (cond_new))
>>>>>>>> +    gimple_cond_set_condition_from_tree (cond_stmt, boolean_false_node);
>>>>>>>>
>>>>>>>> gimple_cond_make_true/false (cond_stmt);
>>>>>>>>
>>>>>>>> btw, seems odd that we have to recompute which loop is the true / false variant
>>>>>>>> when we just fed a guard condition to loop_version.  Can't we statically
>>>>>>>> determine whether loop or nloop has the in-loop condition true or false?
>>>>>>>>
>>>>>>>> +  /* Clean-up cfg to remove useless one-argument phi in exit block of
>>>>>>>> +     outer-loop.  */
>>>>>>>> +  cleanup_tree_cfg ();
>>>>>>>>
>>>>>>>> I know unswitching is already O(number-of-unswitched-loops * size-of-function)
>>>>>>>> because it updates SSA form after each individual unswitching (and it does that
>>>>>>>> because it invokes itself recursively on unswitched loops).  But do you really
>>>>>>>> need to invoke CFG cleanup here?
>>>>>>>>
>>>>>>>> Richard.
>>>>>>>>
>>>>>>>>> Yuri.
>>>>>>>>>
>>>>>>>>> 2015-07-14 14:06 GMT+03:00 Richard Biener <richard.guenther@gmail.com>:
>>>>>>>>>> On Fri, Jul 10, 2015 at 12:02 PM, Yuri Rumyantsev <ysrumyan@gmail.com> wrote:
>>>>>>>>>>> Hi All,
>>>>>>>>>>>
>>>>>>>>>>> Here is presented simple transformation which tries to hoist out of
>>>>>>>>>>> outer-loop a check on zero trip count for inner-loop. This is very
>>>>>>>>>>> restricted transformation since it accepts outer-loops with very
>>>>>>>>>>> simple cfg, as for example:
>>>>>>>>>>>     acc = 0;
>>>>>>>>>>>    for (i = 1; i <= m; i++) {
>>>>>>>>>>>       for (j = 0; j < n; j++)
>>>>>>>>>>>          if (l[j] == i) { v[j] = acc; acc++; };
>>>>>>>>>>>       acc <<= 1;
>>>>>>>>>>>    }
>>>>>>>>>>>
>>>>>>>>>>> Note that degenerative outer loop (without inner loop) will be
>>>>>>>>>>> completely deleted as dead code.
>>>>>>>>>>> The main goal of this transformation was to convert outer-loop to form
>>>>>>>>>>> accepted by outer-loop vectorization (such test-case is also included
>>>>>>>>>>> to patch).
>>>>>>>>>>>
>>>>>>>>>>> Bootstrap and regression testing did not show any new failures.
>>>>>>>>>>>
>>>>>>>>>>> Is it OK for trunk?
>>>>>>>>>>
>>>>>>>>>> I think this is
>>>>>>>>>>
>>>>>>>>>> https://gcc.gnu.org/bugzilla/show_bug.cgi?id=23855
>>>>>>>>>>
>>>>>>>>>> as well.  It has a patch adding a invariant loop guard hoisting
>>>>>>>>>> phase to loop-header copying.  Yeah, it needs updating to
>>>>>>>>>> trunk again I suppose.  It's always non-stage1 when I come
>>>>>>>>>> back to that patch.
>>>>>>>>>>
>>>>>>>>>> Your patch seems to be very specific and only handles outer
>>>>>>>>>> loops of innermost loops.
>>>>>>>>>>
>>>>>>>>>> Richard.
>>>>>>>>>>
>>>>>>>>>>> ChangeLog:
>>>>>>>>>>> 2015-07-10  Yuri Rumyantsev  <ysrumyan@gmail.com>
>>>>>>>>>>>
>>>>>>>>>>> * tree-ssa-loop-unswitch.c: Include "tree-cfgcleanup.h" and
>>>>>>>>>>> "gimple-iterator.h", add prototype for tree_unswitch_outer_loop.
>>>>>>>>>>> (tree_ssa_unswitch_loops): Add invoke of tree_unswitch_outer_loop.
>>>>>>>>>>> (tree_unswitch_outer_loop): New function.
>>>>>>>>>>>
>>>>>>>>>>> gcc/testsuite/ChangeLog:
>>>>>>>>>>> * gcc.dg/tree-ssa/unswitch-outer-loop-1.c: New test.
>>>>>>>>>>> * gcc.dg/vect/vect-outer-simd-3.c: New test.