public inbox for gcc-patches@gcc.gnu.org
 help / color / mirror / Atom feed
From: Yuri Rumyantsev <ysrumyan@gmail.com>
To: Richard Biener <richard.guenther@gmail.com>,
	gcc-patches <gcc-patches@gcc.gnu.org>,
		Igor Zamyatin <izamyatin@gmail.com>
Subject: Re: [PATCH] Unswitching outer loops.
Date: Wed, 30 Sep 2015 11:40:00 -0000	[thread overview]
Message-ID: <CAEoMCqQzyKtyT5N_LCsO3w9W65Tz4RcWXFp52ek=Z3r06AFdJQ@mail.gmail.com> (raw)
In-Reply-To: <CAFiYyc2O9i690A0LZ0+jEOP8nkyz8Btc0YAb469aMgnRaVsmsQ@mail.gmail.com>

[-- Attachment #1: Type: text/plain, Size: 10321 bytes --]

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?

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.

[-- Attachment #2: patch.new --]
[-- Type: application/octet-stream, Size: 14901 bytes --]

diff --git a/gcc/testsuite/gcc.dg/loop-unswitch-2.c b/gcc/testsuite/gcc.dg/loop-unswitch-2.c
new file mode 100644
index 0000000..012c07b
--- /dev/null
+++ b/gcc/testsuite/gcc.dg/loop-unswitch-2.c
@@ -0,0 +1,16 @@
+/* { dg-do compile } */
+/* { dg-options "-O2 -funswitch-loops -fdump-tree-unswitch-details" } */
+
+void foo (float **a, float **b, float *c, int n, int m, int l)
+{
+  int i,j,k;
+  float s;
+  for (i=0; i<l; i++)
+    {
+      for (j=0; j<n; j++)
+	for (k=0; k<m; k++)
+	  c[i] += a[i][k] * b[k][j];
+    }
+}
+
+/* { dg-final { scan-tree-dump-times "guard hoisted" 2 "unswitch" } } */ 
diff --git a/gcc/tree-ssa-loop-unswitch.c b/gcc/tree-ssa-loop-unswitch.c
index a273638..9385503 100644
--- a/gcc/tree-ssa-loop-unswitch.c
+++ b/gcc/tree-ssa-loop-unswitch.c
@@ -39,6 +39,8 @@ along with GCC; see the file COPYING3.  If not see
 #include "params.h"
 #include "tree-pass.h"
 #include "tree-inline.h"
+#include "gimple-iterator.h"
+#include "cfghooks.h"
 
 /* This file implements the loop unswitching, i.e. transformation of loops like
 
@@ -79,6 +81,12 @@ along with GCC; see the file COPYING3.  If not see
 static struct loop *tree_unswitch_loop (struct loop *, basic_block, tree);
 static bool tree_unswitch_single_loop (struct loop *, int);
 static tree tree_may_unswitch_on (basic_block, struct loop *);
+static bool tree_unswitch_outer_loop (struct loop *);
+static edge find_loop_guard (struct loop *);
+static bool empty_bb_without_guard_p (struct loop *, basic_block);
+static bool used_outside_loop_p (struct loop *, tree);
+static void hoist_guard (struct loop *, edge);
+static bool check_exit_phi (struct loop *);
 
 /* Main entry point.  Perform loop unswitching on all suitable loops.  */
 
@@ -87,42 +95,15 @@ tree_ssa_unswitch_loops (void)
 {
   struct loop *loop;
   bool changed = false;
-  HOST_WIDE_INT iterations;
 
-  /* Go through inner loops (only original ones).  */
-  FOR_EACH_LOOP (loop, LI_ONLY_INNERMOST)
+  /* Go through all loops starting from innermost.  */
+  FOR_EACH_LOOP (loop, LI_FROM_INNERMOST)
     {
-      if (dump_file && (dump_flags & TDF_DETAILS))
-        fprintf (dump_file, ";; Considering loop %d\n", loop->num);
-
-      /* Do not unswitch in cold regions. */
-      if (optimize_loop_for_size_p (loop))
-        {
-          if (dump_file && (dump_flags & TDF_DETAILS))
-            fprintf (dump_file, ";; Not unswitching cold loops\n");
-          continue;
-        }
-
-      /* The loop should not be too large, to limit code growth. */
-      if (tree_num_loop_insns (loop, &eni_size_weights)
-          > (unsigned) PARAM_VALUE (PARAM_MAX_UNSWITCH_INSNS))
-        {
-          if (dump_file && (dump_flags & TDF_DETAILS))
-            fprintf (dump_file, ";; Not unswitching, loop too big\n");
-          continue;
-        }
-
-      /* If the loop is not expected to iterate, there is no need
-	 for unswitching.  */
-      iterations = estimated_loop_iterations_int (loop);
-      if (iterations >= 0 && iterations <= 1)
-	{
-          if (dump_file && (dump_flags & TDF_DETAILS))
-            fprintf (dump_file, ";; Not unswitching, loop is not expected to iterate\n");
-          continue;
-	}
-
-      changed |= tree_unswitch_single_loop (loop, 0);
+      if (!loop->inner)
+	/* Unswitch innermost loop.  */
+	changed |= tree_unswitch_single_loop (loop, 0);
+      else
+	changed |= tree_unswitch_outer_loop (loop);
     }
 
   if (changed)
@@ -216,6 +197,39 @@ tree_unswitch_single_loop (struct loop *loop, int num)
   tree cond = NULL_TREE;
   gimple stmt;
   bool changed = false;
+  HOST_WIDE_INT iterations;
+
+  /* Perform initial tests if unswitch is eligible.  */
+  if (num == 0)
+    {
+      /* Do not unswitch in cold regions. */
+      if (optimize_loop_for_size_p (loop))
+	{
+	  if (dump_file && (dump_flags & TDF_DETAILS))
+	    fprintf (dump_file, ";; Not unswitching cold loops\n");
+	  return false;
+	}
+
+      /* The loop should not be too large, to limit code growth. */
+      if (tree_num_loop_insns (loop, &eni_size_weights)
+	  > (unsigned) PARAM_VALUE (PARAM_MAX_UNSWITCH_INSNS))
+	{
+	  if (dump_file && (dump_flags & TDF_DETAILS))
+	    fprintf (dump_file, ";; Not unswitching, loop too big\n");
+	  return false;
+	}
+
+      /* If the loop is not expected to iterate, there is no need
+	 for unswitching.  */
+      iterations = estimated_loop_iterations_int (loop);
+      if (iterations >= 0 && iterations <= 1)
+	{
+	  if (dump_file && (dump_flags & TDF_DETAILS))
+	    fprintf (dump_file, ";; Not unswitching, loop is not expected"
+		     " to iterate\n");
+	  return false;
+	}
+    }
 
   i = 0;
   bbs = get_loop_body (loop);
@@ -403,6 +417,359 @@ tree_unswitch_loop (struct loop *loop,
 		       REG_BR_PROB_BASE - prob_true, false);
 }
 
+/* Unswitch outer loops by hoisting invariant guard on
+   inner loop without code duplication.  */
+static bool
+tree_unswitch_outer_loop (struct loop *loop)
+{
+  edge exit, guard;
+
+  gcc_assert (loop->inner);
+  if (loop->inner->next)
+    return false;
+  /* Accept loops with single exit only.  */
+  exit = single_exit (loop);
+  if (!exit)
+    return false;
+  /* Check that phi argument of exit edge is not defined inside loop.  */
+  if (!check_exit_phi (loop))
+    return false;
+  /* Loop must not be infinite.  */
+  if (!finite_loop_p (loop))
+    return false;
+  guard = find_loop_guard (loop);
+  if (guard)
+    {
+      hoist_guard (loop, guard);
+      update_ssa (TODO_update_ssa);
+      return true;
+    }
+  return false;
+}
+
+/* Checks if the body of the LOOP is within an invariant guard.  If this
+   is the case, returns the edge that jumps over the real body of the loop,
+   otherwise returns NULL.  */
+
+static edge
+find_loop_guard (struct loop *loop)
+{
+  basic_block header = loop->header;
+  edge guard_edge, te, fe;
+  /* bitmap processed, known_invariants;*/
+  basic_block *body = NULL;
+  unsigned i;
+  tree use;
+  ssa_op_iter iter;
+
+  /* We check for the following situation:
+
+     while (1)
+       {
+	 [header]]
+         loop_phi_nodes;
+	 something1;
+	 if (cond1)
+	   body;
+	 nvar = phi(orig, bvar) ... for all variables changed in body;
+	 [guard_end]
+	 something2;
+	 if (cond2)
+	   break;
+	 something3;
+       }
+
+     where:
+
+     1) cond1 is loop invariant
+     2) If cond1 is false, then the loop is essentially empty; i.e.,
+	a) nothing in something1, something2 and something3 has side
+	   effects
+	b) anything defined in something1, something2 and something3
+	   is not used outside of the loop.  */
+
+  while (single_succ_p (header))
+    header = single_succ (header);
+  if (!last_stmt (header)
+      || gimple_code (last_stmt (header)) != GIMPLE_COND)
+    return NULL;
+
+  extract_true_false_edges_from_block (header, &te, &fe);
+  if (!flow_bb_inside_loop_p (loop, te->dest)
+      || !flow_bb_inside_loop_p (loop, fe->dest))
+    return NULL;
+
+  if (just_once_each_iteration_p (loop, te->dest)
+      || (single_succ_p (te->dest)
+	  && just_once_each_iteration_p (loop, single_succ (te->dest))))
+    {
+      if (just_once_each_iteration_p (loop, fe->dest))
+	return NULL;
+      guard_edge = te;
+    }
+  else if (just_once_each_iteration_p (loop, fe->dest)
+	   || (single_succ_p (fe->dest)
+	       && just_once_each_iteration_p (loop, single_succ (fe->dest))))
+    guard_edge = fe;
+  else
+    return NULL;
+
+  if (dump_file && (dump_flags & TDF_DETAILS))
+    fprintf (dump_file,
+	     "Considering guard %d -> %d in loop %d\n",
+	     guard_edge->src->index, guard_edge->dest->index, loop->num);
+  /* Check if condition operands do not have definitions inside loop since
+     any bb copying is not performed.  */
+  FOR_EACH_SSA_TREE_OPERAND (use, last_stmt (header), iter, SSA_OP_USE)
+    {
+      gimple def = SSA_NAME_DEF_STMT (use);
+      basic_block def_bb = gimple_bb (def);
+      if (def_bb
+          && flow_bb_inside_loop_p (loop, def_bb))
+	{
+	  if (dump_file && (dump_flags & TDF_DETAILS))
+	    fprintf (dump_file, "  guard operands have definitions"
+				" inside loop\n");
+	  return NULL;
+	}
+    }
+
+  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]))
+	{
+	  if (dump_file && (dump_flags & TDF_DETAILS))
+	    fprintf (dump_file, "  block %d has side effects\n", body[i]->index);
+	  guard_edge = NULL;
+	  goto end;
+	}
+    }
+
+  if (dump_file && (dump_flags & TDF_DETAILS))
+    fprintf (dump_file, "  suitable to hoist\n");
+end:
+  if (body)
+    free (body);
+  return guard_edge;
+}
+
+/* Returns true if
+   1) no statement in BB has side effects
+   2) assuming that edge GUARD is always taken, all definitions in BB
+      are noy used outside of the loop.
+   KNOWN_INVARIANTS is a set of ssa names we know to be invariant, and
+   PROCESSED is a set of ssa names for that we already tested whether they
+   are invariant or not.  */
+
+static bool
+empty_bb_without_guard_p (struct loop *loop, basic_block bb)
+{
+  basic_block exit_bb = single_exit (loop)->src;
+  bool may_be_used_outside = (bb == exit_bb
+			      || !dominated_by_p (CDI_DOMINATORS, bb, exit_bb));
+  tree name;
+  ssa_op_iter op_iter;
+
+  /* Phi nodes do not have side effects, but their results might be used
+     outside of the loop.  */
+  if (may_be_used_outside)
+    {
+      for (gphi_iterator gsi = gsi_start_phis (bb);
+	   !gsi_end_p (gsi); gsi_next (&gsi))
+	{
+	  gphi *phi = gsi.phi ();
+	  name = PHI_RESULT (phi);
+	  if (virtual_operand_p (name))
+	    continue;
+
+	  if (used_outside_loop_p (loop, name))
+	    return false;
+	}
+    }
+
+  for (gimple_stmt_iterator gsi = gsi_start_bb (bb);
+       !gsi_end_p (gsi); gsi_next (&gsi))
+    {
+      gimple stmt = gsi_stmt (gsi);
+      if (gimple_has_side_effects (stmt))
+	return false;
+
+      if (gimple_vdef(stmt))
+	return false;
+
+      FOR_EACH_SSA_TREE_OPERAND (name, stmt, op_iter, SSA_OP_DEF)
+	{
+	  if (may_be_used_outside
+	      && used_outside_loop_p (loop, name))
+	    return false;
+	}
+    }
+  return true;
+}
+
+/* Return true if NAME is used outside of LOOP.  */
+
+static bool
+used_outside_loop_p (struct loop *loop, tree name)
+{
+  imm_use_iterator it;
+  use_operand_p use;
+
+  FOR_EACH_IMM_USE_FAST (use, it, name)
+    {
+      gimple stmt = USE_STMT (use);
+      if (!flow_bb_inside_loop_p (loop, gimple_bb (stmt)))
+	return true;
+    }
+
+  return false;
+}
+
+/* Moves the check of GUARD outside of LOOP.  */
+
+static void
+hoist_guard (struct loop *loop, edge guard)
+{
+  edge exit = single_exit (loop);
+  edge preh = loop_preheader_edge (loop);
+  basic_block pre_header = preh->src;
+  basic_block bb;
+  edge te, fe, e, tmpe, new_edge;
+  gimple stmt;
+  basic_block guard_bb = guard->src;
+  gimple_stmt_iterator gsi;
+  int flags = 0;
+  bool fix_dom_of_exit;
+
+  gcc_assert (single_succ_p (pre_header));
+  bb = get_immediate_dominator (CDI_DOMINATORS, exit->dest);
+  fix_dom_of_exit = flow_bb_inside_loop_p (loop, bb);
+  /* Hoist GUARD out of loop.  */
+  gsi = gsi_last_bb (guard_bb);
+  stmt = gsi_stmt (gsi);
+  gcc_assert (gimple_code (stmt) == GIMPLE_COND);
+  extract_true_false_edges_from_block (guard_bb, &te, &fe);
+  tmpe = (guard == te) ? fe : te;
+  tmpe->flags &=~ (EDGE_TRUE_VALUE | EDGE_FALSE_VALUE);
+  flags = tmpe->flags;
+  tmpe->flags |= EDGE_FALLTHRU;
+  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));
+      else
+	{
+	  basic_block dom_bb = recompute_dominator (CDI_DOMINATORS, s_bb);
+	  gcc_assert (dom_bb);
+	  set_immediate_dominator (CDI_DOMINATORS, s_bb, dom_bb);
+	}
+    }
+  else if (single_pred_p (bb))
+    set_immediate_dominator (CDI_DOMINATORS, bb, single_pred (bb));
+  else
+    {
+      basic_block dom_bb = recompute_dominator (CDI_DOMINATORS, bb);;
+      gcc_assert (dom_bb);
+      set_immediate_dominator (CDI_DOMINATORS, bb, dom_bb);
+    }
+  /* Insert guard to PRE_HEADER.  */
+  if (!empty_block_p (pre_header))
+    gsi = gsi_last_bb (pre_header);
+  else
+    gsi = gsi_start_bb (pre_header);
+  gsi_insert_after (&gsi, stmt, GSI_NEW_STMT);
+  /* Create new loop pre-header.  */
+  e = split_block (pre_header, last_stmt (pre_header));
+  gcc_assert (loop_preheader_edge (loop)->src == e->dest);
+  if (guard == fe)
+    {
+      e->flags = EDGE_TRUE_VALUE;
+      flags |= EDGE_FALSE_VALUE;
+    }
+  else
+    {
+      e->flags = EDGE_FALSE_VALUE;
+      flags |= EDGE_TRUE_VALUE;
+    }
+  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));
+  /* 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);
+	}
+    }
+
+  mark_virtual_operands_for_renaming (cfun);
+  update_ssa (TODO_update_ssa);
+  if (dump_file && (dump_flags & TDF_DETAILS))
+    fprintf (dump_file, "  guard hoisted.\n");
+}
+
+/* Return true if phi argument for exit edge can be used
+   for edge around loop.  */
+
+static bool
+check_exit_phi (struct loop *loop)
+{
+  edge exit = single_exit (loop);
+  basic_block pre_header = loop_preheader_edge (loop)->src;
+
+  for (gphi_iterator gsi = gsi_start_phis (exit->dest);
+       !gsi_end_p (gsi); gsi_next (&gsi))
+    {
+      gphi *phi = gsi.phi ();
+      tree arg;
+      gimple def;
+      basic_block def_bb;
+      if (virtual_operand_p (gimple_phi_result (phi)))
+	continue;
+      arg = PHI_ARG_DEF_FROM_EDGE (phi, exit);
+      if (TREE_CODE (arg) != SSA_NAME)
+	continue;
+      def = SSA_NAME_DEF_STMT (arg);
+      if (!def)
+	continue;
+      def_bb = gimple_bb (def);
+      if (!def_bb)
+	continue;
+      if (!dominated_by_p (CDI_DOMINATORS, pre_header, def_bb))
+	/* Definition inside loop!  */
+	return false;
+      /* Check loop closed phi invariant.  */
+      if (!flow_bb_inside_loop_p (def_bb->loop_father, pre_header))
+	return false;
+    }
+  return true;
+}
+
 /* Loop unswitching pass.  */
 
 namespace {

  parent reply	other threads:[~2015-09-30 10:46 UTC|newest]

Thread overview: 17+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2015-07-10 10:03 Yuri Rumyantsev
2015-07-14 11:07 ` Richard Biener
2015-07-23 15:21   ` Yuri Rumyantsev
2015-07-28 11:00     ` Richard Biener
2015-07-31 12:07       ` Yuri Rumyantsev
2015-07-31 15:54         ` Jeff Law
2015-08-03  7:27         ` Richard Biener
     [not found]           ` <CAEoMCqSorkh1WmFtVB_huC2hbcVr8uc1EYaRaNVe1g+5hVuzPw@mail.gmail.com>
     [not found]             ` <CAFiYyc1nCCyF-4BH2hPWkKpmXnaQFQ34RMM5TTuHjZxZ25crrA@mail.gmail.com>
     [not found]               ` <CAEoMCqSRsER9ZGgnX9eJgZJyN4EwkpxzWWk1FHRxWNiEW0HVCg@mail.gmail.com>
     [not found]                 ` <CAFiYyc2O9i690A0LZ0+jEOP8nkyz8Btc0YAb469aMgnRaVsmsQ@mail.gmail.com>
2015-09-30 11:40                   ` Yuri Rumyantsev [this message]
2015-10-05 10:57                     ` Richard Biener
2015-10-05 13:13                       ` Yuri Rumyantsev
2015-10-06  7:59                         ` Richard Biener
2015-10-06 11:41                           ` Yuri Rumyantsev
2015-10-06 12:21                             ` Richard Biener
2015-10-07  9:53                               ` Yuri Rumyantsev
2015-10-07 15:26                                 ` Yuri Rumyantsev
2015-10-08 12:31                                   ` Richard Biener
2015-10-09 19:05                                 ` H.J. Lu

Reply instructions:

You may reply publicly to this message via plain-text email
using any one of the following methods:

* Save the following mbox file, import it into your mail client,
  and reply-to-all from there: mbox

  Avoid top-posting and favor interleaved quoting:
  https://en.wikipedia.org/wiki/Posting_style#Interleaved_style

* Reply using the --to, --cc, and --in-reply-to
  switches of git-send-email(1):

  git send-email \
    --in-reply-to='CAEoMCqQzyKtyT5N_LCsO3w9W65Tz4RcWXFp52ek=Z3r06AFdJQ@mail.gmail.com' \
    --to=ysrumyan@gmail.com \
    --cc=gcc-patches@gcc.gnu.org \
    --cc=izamyatin@gmail.com \
    --cc=richard.guenther@gmail.com \
    /path/to/YOUR_REPLY

  https://kernel.org/pub/software/scm/git/docs/git-send-email.html

* If your mail client supports setting the In-Reply-To header
  via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for read-only IMAP folder(s) and NNTP newsgroup(s).