Re: [PING][PATCH][PR65443] Add transform_to_exit_first_loop_alt

[Tom de Vries <Tom_deVries@mentor.com>]

[-- Attachment #1: Type: text/plain, Size: 6357 bytes --]

On 20-04-15 14:25, Richard Biener wrote:
> On Wed, 15 Apr 2015, Tom de Vries wrote:
>
>> On 03-04-15 14:39, Tom de Vries wrote:
>>> On 27-03-15 15:10, Tom de Vries wrote:
>>>> Hi,
>>>>
>>>> this patch fixes PR65443, a todo in the parloops pass for function
>>>> transform_to_exit_first_loop:
>>>> ...
>>>>      TODO: the common case is that latch of the loop is empty and
>>>> immediately
>>>>      follows the loop exit.  In this case, it would be better not to copy
>>>> the
>>>>      body of the loop, but only move the entry of the loop directly before
>>>> the
>>>>      exit check and increase the number of iterations of the loop by one.
>>>>      This may need some additional preconditioning in case NIT = ~0.
>>>> ...
>>>>
>>>> The current implementation of transform_to_exit_first_loop transforms the
>>>> loop
>>>> by moving and duplicating the loop body. This patch transforms the loop
>>>> into the
>>>> same normal form, but without the duplication, if that's possible
>>>> (determined by
>>>> try_transform_to_exit_first_loop_alt).
>>>>
>>>> The actual transformation, done by transform_to_exit_first_loop_alt
>>>> transforms
>>>> this loop:
>>>> ...
>>>>        bb preheader:
>>>>        ...
>>>>        goto <bb header>
>>>>
>>>>        <bb header>:
>>>>        ...
>>>>        if (ivtmp < n)
>>>>          goto <bb latch>;
>>>>        else
>>>>          goto <bb exit>;
>>>>
>>>>        <bb latch>:
>>>>        ivtmp = ivtmp + inc;
>>>>        goto <bb header>
>>>> ...
>>>>
>>>> into this one:
>>>> ...
>>>>        bb preheader:
>>>>        ...
>>>>        goto <bb newheader>
>>>>
>>>>        <bb header>:
>>>>        ...
>>>>        goto <bb latch>;
>>>>
>>>>        <bb newheader>:
>>>>        if (ivtmp < n + 1)
>>>>          goto <bb header>;
>>>>        else
>>>>          goto <bb exit>;
>>>>
>>>>        <bb latch>:
>>>>        ivtmp = ivtmp + inc;
>>>>        goto <bb newheader>
>>>> ...
>>>>
>>>
>>> Updated patch, now using redirect_edge_var_map and flush_pending_stmts.
>>>
>>> Bootstrapped and reg-tested on x86_64.
>>>
>>> OK for stage1 trunk?
>>>
>>
>> Ping.
>

Hi Richard,

thanks for the review.

> +static void
> +replace_imm_uses (tree val, imm_use_iterator *imm_iter)
> +{
> +  use_operand_p use_p;
> +
> +  FOR_EACH_IMM_USE_ON_STMT (use_p, *imm_iter)
> +    SET_USE (use_p, val);
>
> Use propagate_value.  Why this odd interface passing a imm_iter?!
>

The replace_imm_uses function is common code factored out of 
replace_uses_in_bb_by and replace_uses_in_bbs_by. I'm not sure what is odd about 
the interface of the replace_imm_uses function, but I've removed the function.

I tried using propagate_value, but that one doesn't like virtuals.

> In fact most of the "repair SSA" in transform_to_exit_first_loop_alt
> looks too ad-hoc to me ... it also looks somewhat familiar to other
> code ...
>
> Unfortunately the comment before the function isn't in SSA form
> so it's hard to follow the transform.
>

I've added the ssa bits in the transformation comment before the function, and 
updated variable names and comments in the function.

> I consider the parloops code bitrotten, no repair possible, so
> I might be convinced to not care about new spaghetti in there...
>
> +  /* Fix up loop structure.  TODO: Check whether this is sufficient.  */
> +  loop->header = new_header;
> +
>
> no, surely not.  Number of iterations (and estimates) are off
> after the transform

The loop is cancelled at the end of gen_parallel_loop, and the estimations are 
removed.  We only seem to be using a limited set of the loop data until the loop 
is cancelled. Updated comment accordingly.

> and loop->latch might also need updating.
>

That one actually stays the same. Updated comment accordingly.

> "Safest" is to simply schedule a fixup (but you'll lose any
> loop annotation in that process).
>

Since the loop is cancelled, AFAIU we don't need that, but... You're referring 
to the annotations mentioned in replace_loop_annotate_in_block? Losing the 
annotations seems to be a generic problem of scheduling such a fixup, not 
particular to this patch. Do you have a concern related to the patch and these 
annotations?

> +  /* Figure out whether nit + 1 overflows.  */
> +  if (TREE_CODE (nit) == INTEGER_CST)
> +    {
> +      if (!tree_int_cst_equal (nit, TYPE_MAXVAL (nit_type)))
>
> in case nit_type is a pointer type TYPE_MAXVAL will be NULL I think.
>

We've done ivcanon just before this point, so AFAIU we can assume we're dealing 
with an unsigned integer.

> Is the whole exercise for performance?

I'm trying to fix the todo in the code for parloops, which is about performance, 
though I don't expect a large gain.

OTOH, my main focus is a related oacc kernels issue. For an oacc kernels region, 
the entire loop is enclosed in a kernels region. Peeling of the last iteration 
means I have to guard that iteration to run on only one thread, which currently 
isn't done, so in that sense it's a correctness issue as well.

> In that case using an
> entirely new, unsigned IV, that runs from 0 to niter should be easiest and

Introducting such an IV would mean that we don't have to update the IV, but 
still we have to connect the new IV to the uses of the old IV.

The current special handling of the IV variable is actually not that elaborate:
...
+      /* Replace ivtmp_a with ivtmp_c in condition 'if (ivtmp_a < n)'.  */
+      replace_uses_in_bb_by (res_a, res_c, new_header);
...
So I'm not sure it's easier.

just run-time guard that niter == +INF case?

That doesn't work out nicely for the oacc kernels case. I need to know 
compile-time wheter I can parallelize or not.

> For the graphite case, can't you make graphite generate the
> loops exit-first in the first place?
>

Perhaps, but that doesn't make a difference for the oacc kernels case.

> The testcases are just correctness ones?  That is, there isn't
> any testcase that checks whether the new code is exercised?
> (no extra debugging dumping?)
>

There are 3 new test patterns, each with a libgomp/testsuite/libgomp.c and a 
gcc/testsuite/gcc.dg variant, so 6 new test-cases in total. The 
gcc/testsuite/gcc.dg variant checks that the new code is exercised. The 
libgomp/testsuite/libgomp.c variant checks that the new code generates correct code.

Reposting updated patch (for brevity, without testcases) below.

Thanks,
- Tom


[-- Attachment #2: 0002-Add-transform_to_exit_first_loop_alt.patch --]
[-- Type: text/x-patch, Size: 14010 bytes --]

Add transform_to_exit_first_loop_alt

2015-04-14  Tom de Vries  <tom@codesourcery.com>

	PR tree-optimization/65443
	* tree-parloops.c (replace_imm_uses, replace_uses_in_bb_by)
	(replace_uses_in_bbs_by, transform_to_exit_first_loop_alt)
	(try_transform_to_exit_first_loop_alt): New function.
	(transform_to_exit_first_loop): Use
	try_transform_to_exit_first_loop_alt.
---
 gcc/tree-parloops.c | 405 ++++++++++++++++++++++++++++++++++++++++++++++++++--
 1 file changed, 393 insertions(+), 12 deletions(-)

diff --git a/gcc/tree-parloops.c b/gcc/tree-parloops.c
index 080d35e..8168632 100644
--- a/gcc/tree-parloops.c
+++ b/gcc/tree-parloops.c
@@ -78,6 +78,7 @@ along with GCC; see the file COPYING3.  If not see
 #include "plugin-api.h"
 #include "ipa-ref.h"
 #include "cgraph.h"
+#include "tree-ssa.h"
 
 /* This pass tries to distribute iterations of loops into several threads.
    The implementation is straightforward -- for each loop we test whether its
@@ -1487,17 +1488,386 @@ create_loop_fn (location_t loc)
   return decl;
 }
 
-/* Moves the exit condition of LOOP to the beginning of its header, and
-   duplicates the part of the last iteration that gets disabled to the
-   exit of the loop.  NIT is the number of iterations of the loop
-   (used to initialize the variables in the duplicated part).
+/* Replace uses of NAME by VAL in block BB.  */
 
-   TODO: the common case is that latch of the loop is empty and immediately
-   follows the loop exit.  In this case, it would be better not to copy the
-   body of the loop, but only move the entry of the loop directly before the
-   exit check and increase the number of iterations of the loop by one.
-   This may need some additional preconditioning in case NIT = ~0.
-   REDUCTION_LIST describes the reductions in LOOP.  */
+static void
+replace_uses_in_bb_by (tree name, tree val, basic_block bb)
+{
+  gimple use_stmt;
+  imm_use_iterator imm_iter;
+
+  FOR_EACH_IMM_USE_STMT (use_stmt, imm_iter, name)
+    {
+      if (gimple_bb (use_stmt) != bb)
+	continue;
+
+      use_operand_p use_p;
+      FOR_EACH_IMM_USE_ON_STMT (use_p, imm_iter)
+	SET_USE (use_p, val);
+    }
+}
+
+/* Replace uses of NAME by VAL in blocks BBS.  */
+
+static void
+replace_uses_in_bbs_by (tree name, tree val, bitmap bbs)
+{
+  gimple use_stmt;
+  imm_use_iterator imm_iter;
+
+  FOR_EACH_IMM_USE_STMT (use_stmt, imm_iter, name)
+    {
+      if (!bitmap_bit_p (bbs, gimple_bb (use_stmt)->index))
+	continue;
+
+      use_operand_p use_p;
+      FOR_EACH_IMM_USE_ON_STMT (use_p, imm_iter)
+	SET_USE (use_p, val);
+    }
+}
+
+/* Do transformation from:
+
+     <bb preheader>:
+     ...
+     goto <bb header>
+
+     <bb header>:
+     ivtmp_a = PHI <ivtmp_init (preheader), ivtmp_b (latch)>
+     sum_a = PHI <sum_init (preheader), sum_b (latch)>
+     ...
+     use (ivtmp_a)
+     ...
+     sum_b = sum_a + sum_update
+     ...
+     if (ivtmp_a < n)
+       goto <bb latch>;
+     else
+       goto <bb exit>;
+
+     <bb latch>:
+     ivtmp_b = ivtmp_a + 1;
+     goto <bb header>
+
+     <bb exit>:
+     sum_z = PHI <sum_b (cond[1])>
+
+     [1] Where <bb cond> is single_pred (bb latch); In the simplest case,
+	 that's <bb header>.
+
+   to:
+
+     <bb preheader>:
+     ...
+     goto <bb newheader>
+
+     <bb header>:
+     ivtmp_a = PHI <ivtmp_c (latch)>
+     sum_a = PHI <sum_c (latch)>
+     ...
+     use (ivtmp_a)
+     ...
+     sum_b = sum_a + sum_update
+     ...
+     goto <bb latch>;
+
+     <bb newheader>:
+     ivtmp_c = PHI <ivtmp_init (preheader), ivtmp_b (latch)>
+     sum_c = PHI <sum_init (preheader), sum_b (latch)>
+     if (ivtmp_c < n + 1)
+       goto <bb header>;
+     else
+       goto <bb exit>;
+
+     <bb latch>:
+     ivtmp_b = ivtmp_a + 1;
+     goto <bb newheader>
+
+     <bb exit>:
+     sum_z = PHI <sum_c (newheader)>
+
+
+   In unified diff format:
+
+      <bb preheader>:
+      ...
+-     goto <bb header>
++     goto <bb newheader>
+
+      <bb header>:
+-     ivtmp_a = PHI <ivtmp_init (preheader), ivtmp_b (latch)>
+-     sum_a = PHI <sum_init (preheader), sum_b (latch)>
++     ivtmp_a = PHI <ivtmp_c (latch)>
++     sum_a = PHI <sum_c (latch)>
+      ...
+      use (ivtmp_a)
+      ...
+      sum_b = sum_a + sum_update
+      ...
+-     if (ivtmp_a < n)
+-       goto <bb latch>;
++     goto <bb latch>;
++
++     <bb newheader>:
++     ivtmp_c = PHI <ivtmp_init (preheader), ivtmp_b (latch)>
++     sum_c = PHI <sum_init (preheader), sum_b (latch)>
++     if (ivtmp_c < n + 1)
++       goto <bb header>;
+      else
+	goto <bb exit>;
+
+      <bb latch>:
+      ivtmp_b = ivtmp_a + 1;
+-     goto <bb header>
++     goto <bb newheader>
+
+      <bb exit>:
+-     sum_z = PHI <sum_b (cond[1])>
++     sum_z = PHI <sum_c (newheader)>
+
+   Note: the example does not show any virtual phis, but these are handled more
+   or less as reductions.
+
+
+   Moves the exit condition of LOOP to the beginning of its header.
+   REDUCTION_LIST describes the reductions in LOOP.  BOUND is the new loop
+   bound.  */
+
+static void
+transform_to_exit_first_loop_alt (struct loop *loop,
+				  reduction_info_table_type *reduction_list,
+				  tree bound)
+{
+  basic_block header = loop->header;
+  basic_block latch = loop->latch;
+  edge exit = single_dom_exit (loop);
+  basic_block exit_block = exit->dest;
+  gcond *cond_stmt = as_a <gcond *> (last_stmt (exit->src));
+  tree control = gimple_cond_lhs (cond_stmt);
+  edge e;
+
+  /* Gather the bbs dominated by the exit block.  */
+  bitmap exit_dominated = BITMAP_ALLOC (NULL);
+  bitmap_set_bit (exit_dominated, exit_block->index);
+  vec<basic_block> exit_dominated_vec
+    = get_dominated_by (CDI_DOMINATORS, exit_block);
+
+  int i;
+  basic_block dom_bb;
+  FOR_EACH_VEC_ELT (exit_dominated_vec, i, dom_bb)
+    bitmap_set_bit (exit_dominated, dom_bb->index);
+
+  exit_dominated_vec.release ();
+
+  /* Create the new_header block.  */
+  basic_block new_header = split_block_before_cond_jump (exit->src);
+  edge split_edge = single_pred_edge (new_header);
+
+  /* Redirect entry edge to new_header.  */
+  edge entry = loop_preheader_edge (loop);
+  e = redirect_edge_and_branch (entry, new_header);
+  gcc_assert (e == entry);
+
+  /* Redirect post_inc_edge to new_header.  */
+  edge post_inc_edge = single_succ_edge (latch);
+  e = redirect_edge_and_branch (post_inc_edge, new_header);
+  gcc_assert (e == post_inc_edge);
+
+  /* Redirect post_cond_edge to header.  */
+  edge post_cond_edge = single_pred_edge (latch);
+  e = redirect_edge_and_branch (post_cond_edge, header);
+  gcc_assert (e == post_cond_edge);
+
+  /* Redirect split_edge to latch.  */
+  e = redirect_edge_and_branch (split_edge, latch);
+  gcc_assert (e == split_edge);
+
+  /* Set the new loop bound.  */
+  gimple_cond_set_rhs (cond_stmt, bound);
+
+  /* Repair the ssa.  */
+  vec<edge_var_map> *v = redirect_edge_var_map_vector (post_inc_edge);
+  edge_var_map *vm;
+  gphi_iterator gsi;
+  for (gsi = gsi_start_phis (header), i = 0;
+       !gsi_end_p (gsi) && v->iterate (i, &vm);
+       gsi_next (&gsi), i++)
+    {
+      gphi *phi = gsi.phi ();
+      tree res_a = PHI_RESULT (phi);
+
+      /* Create new phi.  */
+      tree res_c = copy_ssa_name (res_a, phi);
+      gphi *nphi = create_phi_node (res_c, new_header);
+
+      /* Replace ivtmp_a with ivtmp_c in condition 'if (ivtmp_a < n)'.  */
+      replace_uses_in_bb_by (res_a, res_c, new_header);
+
+      /* Replace ivtmp/sum_b with ivtmp/sum_c in header phi.  */
+      add_phi_arg (phi, res_c, post_cond_edge, UNKNOWN_LOCATION);
+
+      /* Replace sum_b with sum_c in exit phi.  Loop-closed ssa does not hold
+	 for virtuals, so we cannot get away with exit_block only.  */
+      tree res_b = redirect_edge_var_map_def (vm);
+      replace_uses_in_bbs_by (res_b, res_c, exit_dominated);
+
+      struct reduction_info *red = reduction_phi (reduction_list, phi);
+      gcc_assert (virtual_operand_p (res_a)
+		  || res_a == control
+		  || red != NULL);
+
+      if (red)
+	{
+	  /* Register the new reduction phi.  */
+	  red->reduc_phi = nphi;
+	  gimple_set_uid (red->reduc_phi, red->reduc_version);
+	}
+    }
+  gcc_assert (gsi_end_p (gsi) && !v->iterate (i, &vm));
+  BITMAP_FREE (exit_dominated);
+
+  /* Set the preheader argument of the new phis to ivtmp/sum_init.  */
+  flush_pending_stmts (entry);
+
+  /* Set the latch arguments of the new phis to ivtmp/sum_b.  */
+  flush_pending_stmts (post_inc_edge);
+
+  /* Register the reduction exit phis.  */
+  for (gphi_iterator gsi = gsi_start_phis (exit_block);
+       !gsi_end_p (gsi);
+       gsi_next (&gsi))
+    {
+      gphi *phi = gsi.phi ();
+      tree res_z = PHI_RESULT (phi);
+      if (virtual_operand_p (res_z))
+	continue;
+
+      tree res_c = PHI_ARG_DEF_FROM_EDGE (phi, exit);
+      gimple reduc_phi = SSA_NAME_DEF_STMT (res_c);
+      struct reduction_info *red = reduction_phi (reduction_list, reduc_phi);
+      if (red != NULL)
+	red->keep_res = phi;
+    }
+
+  /* We're going to cancel the loop at the end of gen_parallel_loop, but until
+     then we're still using some fields, so only bother about fields that are
+     still used: header and latch.
+     The loop has a new header bb, so we update it.  The latch bb stays the
+     same.  */
+  loop->header = new_header;
+
+  /* Recalculate dominance info.  */
+  free_dominance_info (CDI_DOMINATORS);
+  calculate_dominance_info (CDI_DOMINATORS);
+}
+
+/* Tries to moves the exit condition of LOOP to the beginning of its header
+   without duplication of the loop body.  NIT is the number of iterations of the
+   loop.  REDUCTION_LIST describes the reductions in LOOP.  Return true if
+   transformation is successful.  */
+
+static bool
+try_transform_to_exit_first_loop_alt (struct loop *loop,
+				      reduction_info_table_type *reduction_list,
+				      tree nit)
+{
+  /* Check whether the latch contains a single statement.  */
+  if (!gimple_seq_singleton_p (bb_seq (loop->latch)))
+    return true;
+
+  /* Check whether the latch contains the loop iv increment.  */
+  edge back = single_succ_edge (loop->latch);
+  edge exit = single_dom_exit (loop);
+  gcond *cond_stmt = as_a <gcond *> (last_stmt (exit->src));
+  tree control = gimple_cond_lhs (cond_stmt);
+  gphi *phi = as_a <gphi *> (SSA_NAME_DEF_STMT (control));
+  tree inc_res = gimple_phi_arg_def (phi, back->dest_idx);
+  if (gimple_bb (SSA_NAME_DEF_STMT (inc_res)) != loop->latch)
+    return false;
+
+  /* Check whether there's no code between the loop condition and the latch.  */
+  if (!single_pred_p (loop->latch)
+      || single_pred (loop->latch) != exit->src)
+    return false;
+
+  tree alt_bound = NULL_TREE;
+  tree nit_type = TREE_TYPE (nit);
+
+  /* Figure out whether nit + 1 overflows.  */
+  if (TREE_CODE (nit) == INTEGER_CST)
+    {
+      if (!tree_int_cst_equal (nit, TYPE_MAXVAL (nit_type)))
+	{
+	  alt_bound = fold_build2_loc (UNKNOWN_LOCATION, PLUS_EXPR, nit_type,
+				       nit, build_one_cst (nit_type));
+
+	  gcc_assert (TREE_CODE (alt_bound) == INTEGER_CST);
+	}
+      else
+	{
+	  /* Todo: Figure out if we can trigger this, if it's worth to handle
+	     optimally, and if we can handle it optimally.  */
+	}
+    }
+  else
+    {
+      gcc_assert (TREE_CODE (nit) == SSA_NAME);
+
+      gimple def = SSA_NAME_DEF_STMT (nit);
+
+      if (def
+	  && is_gimple_assign (def)
+	  && gimple_assign_rhs_code (def) == PLUS_EXPR)
+	{
+	  tree op1 = gimple_assign_rhs1 (def);
+	  tree op2 = gimple_assign_rhs2 (def);
+	  if (integer_minus_onep (op1))
+	    alt_bound = op2;
+	  else if (integer_minus_onep (op2))
+	    alt_bound = op1;
+	}
+
+      /* There is a number of test-cases for which we don't get an alt_bound
+	 here: they're listed here, with the lhs of the last stmt as the nit:
+
+	 libgomp.graphite/force-parallel-1.c:
+	 _21 = (signed long) N_6(D);
+	 _19 = _21 + -1;
+	 _7 = (unsigned long) _19;
+
+	 libgomp.graphite/force-parallel-2.c:
+	 _33 = (signed long) N_9(D);
+	 _16 = _33 + -1;
+	 _37 = (unsigned long) _16;
+
+	 libgomp.graphite/force-parallel-5.c:
+	 <bb 6>:
+	 # graphite_IV.5_46 = PHI <0(5), graphite_IV.5_47(11)>
+	 <bb 7>:
+	 _33 = (unsigned long) graphite_IV.5_46;
+
+	 g++.dg/tree-ssa/pr34355.C:
+	 _2 = (unsigned int) i_9;
+	 _3 = 4 - _2;
+
+	 gcc.dg/pr53849.c:
+	 _5 = d.0_11 + -2;
+	 _18 = (unsigned int) _5;
+
+	 We will be able to handle some of these cases, if we can determine when
+	 it's safe to look past casts.  */
+    }
+
+  if (alt_bound == NULL_TREE)
+    return false;
+
+  transform_to_exit_first_loop_alt (loop, reduction_list, alt_bound);
+  return true;
+}
+
+/* Moves the exit condition of LOOP to the beginning of its header.  NIT is the
+   number of iterations of the loop.  REDUCTION_LIST describes the reductions in
+   LOOP.  */
 
 static void
 transform_to_exit_first_loop (struct loop *loop,
@@ -1882,8 +2252,19 @@ gen_parallel_loop (struct loop *loop,
   /* Base all the induction variables in LOOP on a single control one.  */
   canonicalize_loop_ivs (loop, &nit, true);
 
-  /* Ensure that the exit condition is the first statement in the loop.  */
-  transform_to_exit_first_loop (loop, reduction_list, nit);
+  /* Ensure that the exit condition is the first statement in the loop.
+     The common case is that latch of the loop is empty (apart from the
+     increment) and immediately follows the loop exit test.  Attempt to move the
+     entry of the loop directly before the exit check and increase the number of
+     iterations of the loop by one.  */
+  if (!try_transform_to_exit_first_loop_alt (loop, reduction_list, nit))
+    {
+      /* Fall back on the method that handles more cases, but duplicates the
+	 loop body: move the exit condition of LOOP to the beginning of its
+	 header, and duplicate the part of the last iteration that gets disabled
+	 to the exit of the loop.  */
+      transform_to_exit_first_loop (loop, reduction_list, nit);
+    }
 
   /* Generate initializations for reductions.  */
   if (reduction_list->elements () > 0)
-- 
1.9.1