]> git.ipfire.org Git - thirdparty/gcc.git/commitdiff
loop-split improvements, part 2
authorJan Hubicka <jh@suse.cz>
Fri, 28 Jul 2023 07:48:34 +0000 (09:48 +0200)
committerJan Hubicka <jh@suse.cz>
Fri, 28 Jul 2023 07:51:42 +0000 (09:51 +0200)
this patch fixes profile update in the first case of loop splitting.
The pass still gives up on very basic testcases:

__attribute__ ((noinline,noipa))
void test1 (int n)
{
  if (n <= 0 || n > 100000)
    return;
  for (int i = 0; i <= n; i++)
    {
      if (i < n)
do_something ();
      if (a[i])
do_something2();
    }
}

Here I needed to do the conditoinal that enforces sane value range of n.
The reason is that it gives up on:
      !number_of_iterations_exit (loop1, exit1, &niter, false, true)
and without the conditonal we get assumption that n>=0 and not INT_MAX.
I think from overflow we shold derive that INT_MAX test is not needed and since
the loop does nothing for n<0 it is also just an paranoia.

I am not sure how to fix this though :(.  In general the pass does not really
need to compute iteration count.  It only needs to know what direction the IVs
go so it can detect tests that fires in first part of iteration space.

Rich, any idea what the correct test should be?

In testcase:
  for (int i = 0; i < 200; i++)
    if (i < 150)
      do_something ();
    else
      do_something2 ();
the old code did wrong update of the exit condition probabilities.
We know that first loop iterates 150 times and the second loop 50 times
and we get it by simply scaling loop body by the probability of inner test.

With the patch we now get:

  <bb 2> [count: 1000]:

  <bb 3> [count: 150000]:    <- loop 1 correctly iterates 149 times
  # i_10 = PHI <i_7(8), 0(2)>
  do_something ();
  i_7 = i_10 + 1;
  if (i_7 <= 149)
    goto <bb 8>; [99.33%]
  else
    goto <bb 17>; [0.67%]

  <bb 8> [count: 149000]:
  goto <bb 3>; [100.00%]

  <bb 16> [count: 1000]:
  # i_15 = PHI <i_18(17)>

  <bb 9> [count: 49975]:    <- loop 2 should iterate 50 times but
       we are slightly wrong
  # i_3 = PHI <i_15(16), i_14(13)>
  do_something2 ();
  i_14 = i_3 + 1;
  if (i_14 != 200)
    goto <bb 13>; [98.00%]
  else
    goto <bb 7>; [2.00%]

  <bb 13> [count: 48975]:
  goto <bb 9>; [100.00%]

  <bb 17> [count: 1000]:   <- this test is always true becuase it is
      reached form bb 3
  # i_18 = PHI <i_7(3)>
  if (i_18 != 200)
    goto <bb 16>; [99.95%]
  else
    goto <bb 7>; [0.05%]

  <bb 7> [count: 1000]:
  return;

The reason why we are slightly wrong is the condtion in bb17 that
is always true but the pass does not konw it.

Rich any idea how to do that?  I think connect_loops should work out
the cas where the loop exit conditon is never satisfied at the time
the splitted condition fails for first time.

Before patch on hmmer we get a lot of mismatches:
Profile report here claims:
dump id |static mismat|dynamic mismatch                                     |
        |in count     |in count                  |time                      |
lsplit  |      5    +5|   8151850567  +8151850567531506481006       +57.9%|
ldist   |      9    +4|  15345493501  +7193642934606848841056       +14.2%|
ifcvt   |     10    +1|  15487514871   +142021370689469797790       +13.6%|
vect    |     35   +25|  17558425961  +2070911090517375405715       -25.0%|
cunroll |     42    +7|  16898736178   -659689783452445796198        -4.9%|
loopdone|     33    -9|   2678017188 -14220718990330969127663             |
tracer  |     34    +1|   2678018710        +1522| 330613415364        +0.0%|
fre     |     33    -1|   2676980249     -1038461330465677073        -0.0%|
expand  |     28    -5|   2497468467   -179511782|--------------------------|

With patch

lsplit  |      0      |            0             | 328723360744        -2.3%|
ldist   |      0      |            0             | 396193562452       +20.6%|
ifcvt   |      1    +1|     71010686    +71010686478743508522       +20.8%|
vect    |     14   +13|    697518955   +626508269299398068323       -37.5%|
cunroll |     13    -1|    489349408   -208169547257777839725       -10.5%|
loopdone|     11    -2|    402558559    -86790849201010712702             |
tracer  |     13    +2|    402977200      +418641| 200651036623        +0.0%|
fre     |     13      |    402622146      -355054| 200344398654        -0.2%|
expand  |     11    -2|    333608636    -69013510|--------------------------|

So no mismatches for lsplit and ldist and also lsplit thinks it improves speed by
2.3% rather than regressig it by 57%.

Update is still not perfect since we do not work out that the second loop
never iterates.

Ifcft wrecks profile by desing since it insert conditonals with both arms 100%
that will be eliminated later after vect.  It is not clear to me what happens
in vect though.

Bootstrapped/regtested x86_64-linux, comitted.

gcc/ChangeLog:

PR middle-end/106923
* tree-ssa-loop-split.cc (connect_loops): Change probability
of the test preconditioning second loop to very_likely.
(fix_loop_bb_probability): Handle correctly case where
on of the arms of the conditional is empty.
(split_loop): Fold the test guarding first condition to
see if it is constant true; Set correct entry block
probabilities of the split loops; determine correct loop
eixt probabilities.

gcc/testsuite/ChangeLog:

PR middle-end/106293
* gcc.dg/tree-prof/loop-split-1.c: New test.
* gcc.dg/tree-prof/loop-split-2.c: New test.
* gcc.dg/tree-prof/loop-split-3.c: New test.

gcc/testsuite/gcc.dg/tree-prof/loop-split-1.c [new file with mode: 0644]
gcc/testsuite/gcc.dg/tree-prof/loop-split-2.c [new file with mode: 0644]
gcc/testsuite/gcc.dg/tree-prof/loop-split-3.c [new file with mode: 0644]
gcc/tree-ssa-loop-split.cc

diff --git a/gcc/testsuite/gcc.dg/tree-prof/loop-split-1.c b/gcc/testsuite/gcc.dg/tree-prof/loop-split-1.c
new file mode 100644 (file)
index 0000000..46f7ae6
--- /dev/null
@@ -0,0 +1,33 @@
+/* { dg-options "-O3 -fdump-tree-lsplit-details-blocks -fdump-tree-optimized-details-blocks" } */
+
+
+void
+__attribute__ ((noinline,noipa))
+do_something()
+{
+}
+void
+__attribute__ ((noinline,noipa))
+do_something2()
+{
+}
+
+__attribute__ ((noinline,noipa))
+void test1 (int n, int ga)
+{
+  for (int i = 0; i < 200; i++)
+         if (i < 150)
+                 do_something ();
+         else
+                 do_something2 ();
+}
+int
+main(int, char **)
+{
+       for (int i = 0 ; i < 1000; i++)
+         test1(10, 10);
+       return 0;
+}
+/* { dg-final-use-not-autofdo { scan-tree-dump-times "Loop split" 1 "lsplit" } } */
+/* { dg-final-use-not-autofdo { scan-tree-dump-times "Invalid sum" 0 "lsplit" } } */
+/* { dg-final-use-not-autofdo { scan-tree-dump-times "Invalid sum" 0 "optimized" } } */
diff --git a/gcc/testsuite/gcc.dg/tree-prof/loop-split-2.c b/gcc/testsuite/gcc.dg/tree-prof/loop-split-2.c
new file mode 100644 (file)
index 0000000..de2c9f5
--- /dev/null
@@ -0,0 +1,36 @@
+/* { dg-options "-O3 -fdump-tree-lsplit-details-blocks -fdump-tree-optimized-details-blocks" } */
+
+int M = 100;
+
+void
+__attribute__ ((noinline,noipa))
+do_something()
+{
+}
+void
+__attribute__ ((noinline,noipa))
+do_something2()
+{
+}
+
+__attribute__ ((noinline,noipa))
+void test1 (int n)
+{
+  if (n <= 0 || n > 100000)
+        return; 
+  for (int i = 0; i <= n; i++)
+          if (i < n)
+                  do_something ();
+          else
+                  do_something2 ();
+}
+int
+main(int, char **)
+{
+        for (int i = 0 ; i < 1000; i++)
+          test1(M);
+        return 0;
+}
+/* { dg-final-use-not-autofdo { scan-tree-dump-times "Loop split" 1 "lsplit" } } */
+/* { dg-final-use-not-autofdo { scan-tree-dump-times "Invalid sum" 0 "lsplit" } } */
+/* { dg-final-use-not-autofdo { scan-tree-dump-times "Invalid sum" 0 "optimized" } } */
diff --git a/gcc/testsuite/gcc.dg/tree-prof/loop-split-3.c b/gcc/testsuite/gcc.dg/tree-prof/loop-split-3.c
new file mode 100644 (file)
index 0000000..a88bc1f
--- /dev/null
@@ -0,0 +1,41 @@
+/* { dg-options "-O3 -fdump-tree-lsplit-details-blocks -fdump-tree-optimized-details-blocks" } */
+
+int M = 100;
+int a[1000];
+
+void
+__attribute__ ((noinline,noipa))
+do_something()
+{
+}
+void
+__attribute__ ((noinline,noipa))
+do_something2()
+{
+}
+
+__attribute__ ((noinline,noipa))
+void test1 (int n)
+{
+  if (n <= 0 || n > 100000)
+        return; 
+  for (int i = 0; i <= n; i++)
+    {
+      if (i < n)
+       do_something ();
+      if (a[i])
+       do_something2();
+    }
+}
+int
+main(int, char **)
+{
+  for (int i = 0 ; i < 1000; i+=3)
+    a[i]=1;
+  for (int i = 0 ; i < 1000; i++)
+    test1(M);
+  return 0;
+}
+/* { dg-final-use-not-autofdo { scan-tree-dump-times "Loop split" 1 "lsplit" } } */
+/* { dg-final-use-not-autofdo { scan-tree-dump-times "Invalid sum" 0 "lsplit" } } */
+/* { dg-final-use-not-autofdo { scan-tree-dump-times "Invalid sum" 0 "optimized" } } */
index f441f3fe1b519ee210568db3da9e3e4fcae0557d..70cd0aaefa737a276a64664eb4eefeb20684a354 100644 (file)
@@ -41,6 +41,7 @@ along with GCC; see the file COPYING3.  If not see
 #include "cfghooks.h"
 #include "gimple-fold.h"
 #include "gimplify-me.h"
+#include "print-tree.h"
 
 /* This file implements two kinds of loop splitting.
 
@@ -355,7 +356,7 @@ connect_loops (class loop *loop1, class loop *loop2)
       new_e->flags |= EDGE_TRUE_VALUE;
     }
 
-  new_e->probability = profile_probability::likely ();
+  new_e->probability = profile_probability::very_likely ();
   skip_e->probability = new_e->probability.invert ();
 
   return new_e;
@@ -496,6 +497,8 @@ fix_loop_bb_probability (class loop *loop1, class loop *loop2, edge true_edge,
   unsigned j;
   for (j = 0; j < loop1->num_nodes; j++)
     if (bbs1[j] == loop1->latch
+       /* Watch for case where the true conditional is empty.  */
+       || !single_pred_p (true_edge->dest)
        || !dominated_by_p (CDI_DOMINATORS, bbs1[j], true_edge->dest))
       bbs1[j]->count
        = bbs1[j]->count.apply_probability (true_edge->probability);
@@ -507,6 +510,8 @@ fix_loop_bb_probability (class loop *loop1, class loop *loop2, edge true_edge,
   bbs2 = get_loop_body (loop2);
   for (j = 0; j < loop2->num_nodes; j++)
     if (bbs2[j] == loop2->latch
+       /* Watch for case where the flase conditional is empty.  */
+       || !single_pred_p (bbi_copy)
        || !dominated_by_p (CDI_DOMINATORS, bbs2[j], bbi_copy))
       bbs2[j]->count
        = bbs2[j]->count.apply_probability (true_edge->probability.invert ());
@@ -564,6 +569,7 @@ split_loop (class loop *loop1)
   for (i = 0; i < loop1->num_nodes; i++)
     if ((guard_iv = split_at_bb_p (loop1, bbs[i], &border, &iv)))
       {
+       profile_count entry_count = loop_preheader_edge (loop1)->count ();
        /* Handling opposite steps is not implemented yet.  Neither
           is handling different step sizes.  */
        if ((tree_int_cst_sign_bit (iv.step)
@@ -610,7 +616,8 @@ split_loop (class loop *loop1)
        if (stmts2)
          gsi_insert_seq_on_edge_immediate (loop_preheader_edge (loop1),
                                            stmts2);
-       tree cond = build2 (guard_code, boolean_type_node, guard_init, border);
+       tree cond = fold_build2 (guard_code, boolean_type_node,
+                                guard_init, border);
        if (!initial_true)
          cond = fold_build1 (TRUTH_NOT_EXPR, boolean_type_node, cond);
 
@@ -622,13 +629,71 @@ split_loop (class loop *loop1)
        initialize_original_copy_tables ();
        basic_block cond_bb;
 
+       profile_probability loop1_prob
+         = integer_onep (cond) ? profile_probability::always ()
+                               : true_edge->probability;
+       /* TODO: It is commonly a case that we know that both loops will be
+          entered.  very_likely below is the probability that second loop will
+          be entered given by connect_loops.  We should work out the common
+          case it is always true.  */
        class loop *loop2 = loop_version (loop1, cond, &cond_bb,
-                                         true_edge->probability,
-                                         true_edge->probability.invert (),
+                                         loop1_prob,
+                                         /* Pass always as we will later
+                                            redirect first loop to second
+                                            loop.  */
                                          profile_probability::always (),
                                          profile_probability::always (),
+                                         profile_probability::very_likely (),
                                          true);
        gcc_assert (loop2);
+       /* Correct probability of edge  cond_bb->preheader_of_loop2.  */
+       single_pred_edge
+               (loop_preheader_edge (loop2)->src)->probability
+                       = loop1_prob.invert ();
+
+       fix_loop_bb_probability (loop1, loop2, true_edge, false_edge);
+
+       /* Fix loop's exit probability after scaling.  */
+
+       if (entry_count.initialized_p () && entry_count.nonzero_p ())
+         {
+           edge exit_to_latch1;
+           if (EDGE_SUCC (exit1->src, 0) == exit1)
+             exit_to_latch1 = EDGE_SUCC (exit1->src, 1);
+           else
+             exit_to_latch1 = EDGE_SUCC (exit1->src, 0);
+           if (exit1->src->count.nonzero_p ())
+             {
+               /* First loop is entered loop1_prob * entry_count times
+                  and it needs to exit the same number of times.  */
+               exit1->probability
+                       = entry_count.apply_probability
+                               (loop1_prob).probability_in (exit1->src->count);
+               exit_to_latch1->probability = exit1->probability.invert ();
+               scale_dominated_blocks_in_loop (loop1, exit1->src,
+                                               exit_to_latch1->count (),
+                                               exit_to_latch1->dest->count);
+             }
+
+           edge exit_to_latch2, exit2 = single_exit (loop2);
+           if (EDGE_SUCC (exit2->src, 0) == exit2)
+             exit_to_latch2 = EDGE_SUCC (exit2->src, 1);
+           else
+             exit_to_latch2 = EDGE_SUCC (exit2->src, 0);
+           if (exit2->src->count.nonzero_p ())
+             {
+               /* Second loop is entered very_likely * entry_count times
+                  and it needs to exit the same number of times.  */
+               exit2->probability
+                       = entry_count.apply_probability
+                               (profile_probability::very_likely ())
+                               .probability_in (exit2->src->count);
+               exit_to_latch2->probability = exit2->probability.invert ();
+               scale_dominated_blocks_in_loop (loop2, exit2->src,
+                                               exit_to_latch2->count (),
+                                               exit_to_latch2->dest->count);
+             }
+         }
 
        edge new_e = connect_loops (loop1, loop2);
        connect_loop_phis (loop1, loop2, new_e);
@@ -647,16 +712,7 @@ split_loop (class loop *loop1)
        tree guard_next = PHI_ARG_DEF_FROM_EDGE (phi, loop_latch_edge (loop1));
        patch_loop_exit (loop1, guard_stmt, guard_next, newend, initial_true);
 
-       fix_loop_bb_probability (loop1, loop2, true_edge, false_edge);
-
-       /* Fix first loop's exit probability after scaling.  */
-       edge exit_to_latch1;
-       if (EDGE_SUCC (exit1->src, 0) == exit1)
-         exit_to_latch1 = EDGE_SUCC (exit1->src, 1);
-       else
-         exit_to_latch1 = EDGE_SUCC (exit1->src, 0);
-       exit_to_latch1->probability *= true_edge->probability;
-       exit1->probability = exit_to_latch1->probability.invert ();
+       /* TODO: Update any_esitmate and upper bounds.  */
 
        /* Finally patch out the two copies of the condition to be always
           true/false (or opposite).  */