Patchwork Redesign jump threading profile updates

login
register
mail settings
Submitter Teresa Johnson
Date March 26, 2014, 11:44 p.m.
Message ID <CAAe5K+XMKuBvrA2zbSdA38nq+KiezseL6z_4KayvV322VmJtZQ@mail.gmail.com>
Download mbox | patch
Permalink /patch/334134/
State New
Headers show

Comments

Teresa Johnson - March 26, 2014, 11:44 p.m.
Recently I discovered that the profile updates being performed by jump
threading were incorrect in many cases, particularly in the case where
the threading path contains a joiner. Some of the duplicated
blocks/edges were not getting any counts, leading to incorrect
function splitting and other downstream optimizations, and there were
other insanities as well. After making a few attempts to fix the
handling I ended up completely redesigning the profile update code,
removing a few places throughout the code where it was attempting to
do some updates.

The biggest complication (see the large comment and example above the
new routine compute_path_counts) is that we duplicate a conditional
jump in the joiner case, possibly multiple times for multiple jump
thread paths through that joiner, and it isn't trivial to figure out
what probability to assign each of the duplicated successor edges (and
the original after threading). Each jump thread path may need to have
a different probability of staying on path through the joiner in order
to keep the counts going out of the threading path sane.

The patch below was bootstrapped and tested on
x86_64-unknown-linux-gnu, and also tested with a profiledbootstrap. I
additionally tested with cpu2006, confirming that the amount of
resulting cycle samples in the split cold sections reduced, and
through manual inspection that many different cases were now correct.
I also measured performance with cpu2006, running each benchmark
multiple times on a Westmere and see some speedups (453.povray 1-2%,
403.gcc 1-1.5%, and noisy but positive speedups in 471.omnetpp and
483.xalancbmk).

Looks like my mailer is corrupting the spacing, which makes it harder
to look at the CFG examples in the big header comment block I added.
So I have also included the patch as an attachment.

Ok for stage 1?

Thanks,
Teresa

 2014-03-26  Teresa Johnson  <tejohnson@google.com>

        * tree-ssa-threadupdate.c (struct ssa_local_info_t): New
        duplicate_blocks bitmap.
        (remove_ctrl_stmt_and_useless_edges): Ditto.
        (create_block_for_threading): Ditto.
        (compute_path_counts): New function.
        (update_profile): Ditto.
        (deduce_freq): Ditto.
        (recompute_probabilities): Ditto.
        (update_joiner_offpath_counts): Ditto.
        (ssa_fix_duplicate_block_edges): Update profile info.
        (ssa_create_duplicates): Pass new parameter.
        (ssa_redirect_edges): Remove old profile update.
        (thread_block_1): New duplicate_blocks bitmap,
        remove old profile update.
        (thread_single_edge): Pass new parameter.

         Else we want to remove the control statement & edges, then create
@@ -515,6 +961,8 @@ ssa_fix_duplicate_block_edges (struct redirection_
          edge victim;
          edge e2;

+          gcc_assert (has_joiner);
+
          /* This updates the PHIs at the destination of the duplicate
             block.  */
          update_destination_phis (local_info->bb, rd->dup_blocks[count]);
@@ -528,14 +976,13 @@ ssa_fix_duplicate_block_edges (struct redirection_
             threading path.  */
          if (!any_remaining_duplicated_blocks (path, i))
            {
-             e2 = redirect_edge_and_branch (victim, path->last ()->e->dest);
-             e2->count = path->last ()->e->count;
+             e2 = redirect_edge_and_branch (victim, elast->dest);
              /* If we redirected the edge, then we need to copy PHI arguments
                 at the target.  If the edge already existed (e2 != victim
                 case), then the PHIs in the target already have the correct
                 arguments.  */
              if (e2 == victim)
-               copy_phi_args (e2->dest, path->last ()->e, e2);
+               copy_phi_args (e2->dest, elast, e2);
            }
          else
            {
@@ -562,7 +1009,31 @@ ssa_fix_duplicate_block_edges (struct redirection_
                    }
                }
            }
-         count++;
+
+         /* Update the counts and frequency of both the original block
+            and path edge, and the duplicates.  The path duplicate's
+            incoming count and frequency are the totals for all edges
+            incoming to this jump threading path computed earlier.
+            And we know that the duplicated path will have path_out_count
+            flowing out of it (i.e. along the duplicated path out of the
+            duplicated joiner).  */
+         update_profile (epath, e2, path_in_count, path_out_count,
+                         path_in_freq);
+
+         /* Next we need to update the counts of the original and duplicated
+            edges from the joiner that go off path.  */
+         update_joiner_offpath_counts (epath, e2->src, path_in_count,
+                                        path_out_count);
+
+         /* Finally, we need to set the probabilities on the duplicated
+            edges out of the duplicated joiner (e2->src).  The probabilities
+            along the original path will all be updated below after we finish
+            processing the whole path.  */
+         recompute_probabilities (e2->src);
+
+         /* Record the frequency flowing to the downstream duplicated
+            path blocks.  */
+         cur_path_freq = EDGE_FREQUENCY (e2);
        }
       else if ((*path)[i]->type == EDGE_COPY_SRC_BLOCK)
        {
@@ -570,9 +1041,55 @@ ssa_fix_duplicate_block_edges (struct redirection_
          create_edge_and_update_destination_phis (rd, rd->dup_blocks[count]);
          if (count == 1)
            single_succ_edge (rd->dup_blocks[1])->aux = NULL;
+
+         /* Update the counts and frequency of both the original block
+            and path edge, and the duplicates.  Since we are now after
+            any joiner that may have existed on the path, the count
+            flowing along the duplicated threaded path is path_out_count.
+            If we didn't have a joiner, then cur_path_freq was the sum
+            of the total frequencies along all incoming edges to the
+            thread path (path_in_freq).  If we had a joiner, it would have
+            been updated at the end of that handling to the edge frequency
+            along the duplicated joiner path edge.  */
+         update_profile (epath, EDGE_SUCC (rd->dup_blocks[count], 0),
+                         path_out_count, path_out_count,
+                         cur_path_freq);
+       }
+      else
+        {
+         /* No copy case.  In this case we don't have an equivalent block
+            on the duplicated thread path to update, but we do need
+            to remove the portion of the counts/freqs that were moved
+            to the duplicated path from the counts/freqs flowing through
+            this block on the original path.  Since all the no-copy edges
+            are after any joiner, the removed count is the same as
+            path_out_count.
+
+            If we didn't have a joiner, then cur_path_freq was the sum
+            of the total frequencies along all incoming edges to the
+            thread path (path_in_freq).  If we had a joiner, it would have
+            been updated at the end of that handling to the edge frequency
+            along the duplicated joiner path edge.  */
+            update_profile (epath, NULL, path_out_count, path_out_count,
+                            cur_path_freq);
+       }
+
+      /* Increment the index into the duplicated path when we processed
+         a duplicated block.  */
+      if ((*path)[i]->type == EDGE_COPY_SRC_JOINER_BLOCK
+          || (*path)[i]->type == EDGE_COPY_SRC_BLOCK)
+      {
          count++;
-       }
+      }
     }
+
+  /* Now walk orig blocks and update their probabilities, since the
+     counts and freqs should be updated properly by above loop.  */
+  for (unsigned int i = 1; i < path->length (); i++)
+    {
+      edge epath = (*path)[i]->e;
+      recompute_probabilities (epath->src);
+    }
 }

 /* Hash table traversal callback routine to create duplicate blocks.  */
@@ -598,7 +1115,8 @@ ssa_create_duplicates (struct redirection_data **s
       if ((*path)[i]->type == EDGE_COPY_SRC_BLOCK
          || (*path)[i]->type == EDGE_COPY_SRC_JOINER_BLOCK)
        {
-         create_block_for_threading ((*path)[i]->e->src, rd, 1);
+         create_block_for_threading ((*path)[i]->e->src, rd, 1,
+                                      &local_info->duplicate_blocks);
          break;
        }
     }
@@ -607,7 +1125,8 @@ ssa_create_duplicates (struct redirection_data **s
      use the template to create a new block.  */
   if (local_info->template_block == NULL)
     {
-      create_block_for_threading ((*path)[1]->e->src, rd, 0);
+      create_block_for_threading ((*path)[1]->e->src, rd, 0,
+                                  &local_info->duplicate_blocks);
       local_info->template_block = rd->dup_blocks[0];

       /* We do not create any outgoing edges for the template.  We will
@@ -616,7 +1135,8 @@ ssa_create_duplicates (struct redirection_data **s
     }
   else
     {
-      create_block_for_threading (local_info->template_block, rd, 0);
+      create_block_for_threading (local_info->template_block, rd, 0,
+                                  &local_info->duplicate_blocks);

       /* Go ahead and wire up outgoing edges and update PHIs for the duplicate
         block.   */
@@ -686,19 +1206,6 @@ ssa_redirect_edges (struct redirection_data **slot
            fprintf (dump_file, "  Threaded jump %d --> %d to %d\n",
                     e->src->index, e->dest->index, rd->dup_blocks[0]->index);

-         rd->dup_blocks[0]->count += e->count;
-
-         /* Excessive jump threading may make frequencies large enough so
-            the computation overflows.  */
-         if (rd->dup_blocks[0]->frequency < BB_FREQ_MAX * 2)
-           rd->dup_blocks[0]->frequency += EDGE_FREQUENCY (e);
-
-         /* In the case of threading through a joiner block, the outgoing
-            edges from the duplicate block were updated when they were
-            redirected during ssa_fix_duplicate_block_edges.  */
-         if ((*path)[1]->type != EDGE_COPY_SRC_JOINER_BLOCK)
-           EDGE_SUCC (rd->dup_blocks[0], 0)->count += e->count;
-
          /* Redirect the incoming edge (possibly to the joiner block) to the
             appropriate duplicate block.  */
          e2 = redirect_edge_and_branch (e, rd->dup_blocks[0]);
@@ -781,6 +1288,8 @@ thread_block_1 (basic_block bb, bool noloop_only,
   ssa_local_info_t local_info;
   struct loop *loop = bb->loop_father;

+  local_info.duplicate_blocks = BITMAP_ALLOC (NULL);
+
   /* To avoid scanning a linear array for the element we need we instead
      use a hash table.  For normal code there should be no noticeable
      difference.  However, if we have a block with a large number of
@@ -865,10 +1374,6 @@ thread_block_1 (basic_block bb, bool noloop_only,
            continue;
        }

-      if (e->dest == e2->src)
-       update_bb_profile_for_threading (e->dest, EDGE_FREQUENCY (e),
-                                        e->count, (*THREAD_PATH (e))[1]->e);
-
       /* Insert the outgoing edge into the hash table if it is not
         already in the hash table.  */
       lookup_redirection_data (e, INSERT);
@@ -921,6 +1426,9 @@ thread_block_1 (basic_block bb, bool noloop_only,
       && bb == bb->loop_father->header)
     set_loop_copy (bb->loop_father, NULL);

+  BITMAP_FREE (local_info.duplicate_blocks);
+  local_info.duplicate_blocks = NULL;
+
   /* Indicate to our caller whether or not any jumps were threaded.  */
   return local_info.jumps_threaded;
 }
@@ -987,7 +1495,7 @@ thread_single_edge (edge e)
   npath->safe_push (x);
   rd.path = npath;

-  create_block_for_threading (bb, &rd, 0);
+  create_block_for_threading (bb, &rd, 0, NULL);
   remove_ctrl_stmt_and_useless_edges (rd.dup_blocks[0], NULL);
   create_edge_and_update_destination_phis (&rd, rd.dup_blocks[0]);
Jeff Law - April 17, 2014, 5:39 a.m.
On 03/26/14 17:44, Teresa Johnson wrote:
> Recently I discovered that the profile updates being performed by jump
> threading were incorrect in many cases, particularly in the case where
> the threading path contains a joiner. Some of the duplicated
> blocks/edges were not getting any counts, leading to incorrect
> function splitting and other downstream optimizations, and there were
> other insanities as well. After making a few attempts to fix the
> handling I ended up completely redesigning the profile update code,
> removing a few places throughout the code where it was attempting to
> do some updates.
The profile updates in that code is a mess.  It's never really been 
looked at in any systematic way, what's there is ad-hoc and usually in 
response to someone mentioning the profile data was incorrectly updated. 
   As we rely more and more on that data the ad-hoc updating is going to 
cause us more and more pain.

So any work in this space is going to be greatly appreciated.

I'll have to look at this in some detail.  But I wanted you to know I 
was aware of the work and it's in my queue.

Thanks!

jeff
Teresa Johnson - April 17, 2014, 1:23 p.m.
On Wed, Apr 16, 2014 at 10:39 PM, Jeff Law <law@redhat.com> wrote:
> On 03/26/14 17:44, Teresa Johnson wrote:
>>
>> Recently I discovered that the profile updates being performed by jump
>> threading were incorrect in many cases, particularly in the case where
>> the threading path contains a joiner. Some of the duplicated
>> blocks/edges were not getting any counts, leading to incorrect
>> function splitting and other downstream optimizations, and there were
>> other insanities as well. After making a few attempts to fix the
>> handling I ended up completely redesigning the profile update code,
>> removing a few places throughout the code where it was attempting to
>> do some updates.
>
> The profile updates in that code is a mess.  It's never really been looked
> at in any systematic way, what's there is ad-hoc and usually in response to
> someone mentioning the profile data was incorrectly updated.   As we rely
> more and more on that data the ad-hoc updating is going to cause us more and
> more pain.
>
> So any work in this space is going to be greatly appreciated.
>
> I'll have to look at this in some detail.  But I wanted you to know I was
> aware of the work and it's in my queue.

Great, thanks for the update! I realize that it is not a trivial
change so it would take some time to get through. Hopefully it should
address the ongoing profile fixup issues.
Teresa

>
> Thanks!
>
> jeff
Teresa Johnson - May 27, 2014, 2:10 p.m.
On Thu, Apr 17, 2014 at 6:23 AM, Teresa Johnson <tejohnson@google.com> wrote:
> On Wed, Apr 16, 2014 at 10:39 PM, Jeff Law <law@redhat.com> wrote:
>> On 03/26/14 17:44, Teresa Johnson wrote:
>>>
>>> Recently I discovered that the profile updates being performed by jump
>>> threading were incorrect in many cases, particularly in the case where
>>> the threading path contains a joiner. Some of the duplicated
>>> blocks/edges were not getting any counts, leading to incorrect
>>> function splitting and other downstream optimizations, and there were
>>> other insanities as well. After making a few attempts to fix the
>>> handling I ended up completely redesigning the profile update code,
>>> removing a few places throughout the code where it was attempting to
>>> do some updates.
>>
>> The profile updates in that code is a mess.  It's never really been looked
>> at in any systematic way, what's there is ad-hoc and usually in response to
>> someone mentioning the profile data was incorrectly updated.   As we rely
>> more and more on that data the ad-hoc updating is going to cause us more and
>> more pain.
>>
>> So any work in this space is going to be greatly appreciated.
>>
>> I'll have to look at this in some detail.  But I wanted you to know I was
>> aware of the work and it's in my queue.
>
> Great, thanks for the update! I realize that it is not a trivial
> change so it would take some time to get through. Hopefully it should
> address the ongoing profile fixup issues.
> Teresa

Ping.
Thanks,
Teresa

>
>>
>> Thanks!
>>
>> jeff
>
>
>
> --
> Teresa Johnson | Software Engineer | tejohnson@google.com | 408-460-2413
Teresa Johnson - July 7, 2014, 9:22 p.m.
Ping. Jeff, any update on when you can look at this?
Thanks,
Teresa

On Tue, May 27, 2014 at 7:10 AM, Teresa Johnson <tejohnson@google.com> wrote:
> On Thu, Apr 17, 2014 at 6:23 AM, Teresa Johnson <tejohnson@google.com> wrote:
>> On Wed, Apr 16, 2014 at 10:39 PM, Jeff Law <law@redhat.com> wrote:
>>> On 03/26/14 17:44, Teresa Johnson wrote:
>>>>
>>>> Recently I discovered that the profile updates being performed by jump
>>>> threading were incorrect in many cases, particularly in the case where
>>>> the threading path contains a joiner. Some of the duplicated
>>>> blocks/edges were not getting any counts, leading to incorrect
>>>> function splitting and other downstream optimizations, and there were
>>>> other insanities as well. After making a few attempts to fix the
>>>> handling I ended up completely redesigning the profile update code,
>>>> removing a few places throughout the code where it was attempting to
>>>> do some updates.
>>>
>>> The profile updates in that code is a mess.  It's never really been looked
>>> at in any systematic way, what's there is ad-hoc and usually in response to
>>> someone mentioning the profile data was incorrectly updated.   As we rely
>>> more and more on that data the ad-hoc updating is going to cause us more and
>>> more pain.
>>>
>>> So any work in this space is going to be greatly appreciated.
>>>
>>> I'll have to look at this in some detail.  But I wanted you to know I was
>>> aware of the work and it's in my queue.
>>
>> Great, thanks for the update! I realize that it is not a trivial
>> change so it would take some time to get through. Hopefully it should
>> address the ongoing profile fixup issues.
>> Teresa
>
> Ping.
> Thanks,
> Teresa
>
>>
>>>
>>> Thanks!
>>>
>>> jeff
>>
>>
>>
>> --
>> Teresa Johnson | Software Engineer | tejohnson@google.com | 408-460-2413
>
>
>
> --
> Teresa Johnson | Software Engineer | tejohnson@google.com | 408-460-2413
Jeff Law - July 7, 2014, 9:24 p.m.
On 07/07/14 15:22, Teresa Johnson wrote:
> Ping. Jeff, any update on when you can look at this?
Hoping to do so on that long plane ride to the UK for Cauldron :-)

SLC->LAX->Heathrow  Lots of uninterrupted time to look at those changes.


Jeff
Jeff Law - July 23, 2014, 2:29 a.m.
On 03/26/14 17:44, Teresa Johnson wrote:
> Recently I discovered that the profile updates being performed by jump
> threading were incorrect in many cases, particularly in the case where
> the threading path contains a joiner. Some of the duplicated
> blocks/edges were not getting any counts, leading to incorrect
> function splitting and other downstream optimizations, and there were
> other insanities as well. After making a few attempts to fix the
> handling I ended up completely redesigning the profile update code,
> removing a few places throughout the code where it was attempting to
> do some updates.
>
> The biggest complication (see the large comment and example above the
> new routine compute_path_counts) is that we duplicate a conditional
> jump in the joiner case, possibly multiple times for multiple jump
> thread paths through that joiner, and it isn't trivial to figure out
> what probability to assign each of the duplicated successor edges (and
> the original after threading). Each jump thread path may need to have
> a different probability of staying on path through the joiner in order
> to keep the counts going out of the threading path sane.
>
> The patch below was bootstrapped and tested on
> x86_64-unknown-linux-gnu, and also tested with a profiledbootstrap. I
> additionally tested with cpu2006, confirming that the amount of
> resulting cycle samples in the split cold sections reduced, and
> through manual inspection that many different cases were now correct.
> I also measured performance with cpu2006, running each benchmark
> multiple times on a Westmere and see some speedups (453.povray 1-2%,
> 403.gcc 1-1.5%, and noisy but positive speedups in 471.omnetpp and
> 483.xalancbmk).
>
> Looks like my mailer is corrupting the spacing, which makes it harder
> to look at the CFG examples in the big header comment block I added.
> So I have also included the patch as an attachment.
>
> Ok for stage 1?
>
> Thanks,
> Teresa
>
>   2014-03-26  Teresa Johnson  <tejohnson@google.com>
>
>          * tree-ssa-threadupdate.c (struct ssa_local_info_t): New
>          duplicate_blocks bitmap.
>          (remove_ctrl_stmt_and_useless_edges): Ditto.
>          (create_block_for_threading): Ditto.
>          (compute_path_counts): New function.
>          (update_profile): Ditto.
>          (deduce_freq): Ditto.
>          (recompute_probabilities): Ditto.
>          (update_joiner_offpath_counts): Ditto.
>          (ssa_fix_duplicate_block_edges): Update profile info.
>          (ssa_create_duplicates): Pass new parameter.
>          (ssa_redirect_edges): Remove old profile update.
>          (thread_block_1): New duplicate_blocks bitmap,
>          remove old profile update.
>          (thread_single_edge): Pass new parameter.
First off, sorry this took so long to get reviewed.

Most of what's going on in here is similar to something I sketched out, 
but never coded up a while back -- with the significant difference that 
you're handling joiner blocks as well.

Everything looks to be well thought through and documented in the code 
at a level I wish existed throughout GCC.

The only thing I see missing is regression tests.  I don't think you 
need to do anything huge here, but it ought to be possible to set up 
relatively simple cases which show the probabilities/counts being 
updated properly.

Otherwise it looks excellent.  It's pre-approved once you've added some 
kind of testing and fixed the nits noted below.


> +   In the aboe example, after all jump threading is complete, we will
s/aboe/above/


> +  struct el *next, *el;
> +  bitmap in_edge_srcs = BITMAP_ALLOC (NULL);
> +  for (el = rd->incoming_edges; el; el = next)
> +    {
> +      next = el->next;
> +      bitmap_set_bit (in_edge_srcs, el->e->src->index);
> +    }
Please add vertical whitespace after this loop, but before declaring 
variables for the next loop.

Jeff
Teresa Johnson - July 23, 2014, 9:08 p.m.
On Tue, Jul 22, 2014 at 7:29 PM, Jeff Law <law@redhat.com> wrote:
> On 03/26/14 17:44, Teresa Johnson wrote:
>>
>> Recently I discovered that the profile updates being performed by jump
>> threading were incorrect in many cases, particularly in the case where
>> the threading path contains a joiner. Some of the duplicated
>> blocks/edges were not getting any counts, leading to incorrect
>> function splitting and other downstream optimizations, and there were
>> other insanities as well. After making a few attempts to fix the
>> handling I ended up completely redesigning the profile update code,
>> removing a few places throughout the code where it was attempting to
>> do some updates.
>>
>> The biggest complication (see the large comment and example above the
>> new routine compute_path_counts) is that we duplicate a conditional
>> jump in the joiner case, possibly multiple times for multiple jump
>> thread paths through that joiner, and it isn't trivial to figure out
>> what probability to assign each of the duplicated successor edges (and
>> the original after threading). Each jump thread path may need to have
>> a different probability of staying on path through the joiner in order
>> to keep the counts going out of the threading path sane.
>>
>> The patch below was bootstrapped and tested on
>> x86_64-unknown-linux-gnu, and also tested with a profiledbootstrap. I
>> additionally tested with cpu2006, confirming that the amount of
>> resulting cycle samples in the split cold sections reduced, and
>> through manual inspection that many different cases were now correct.
>> I also measured performance with cpu2006, running each benchmark
>> multiple times on a Westmere and see some speedups (453.povray 1-2%,
>> 403.gcc 1-1.5%, and noisy but positive speedups in 471.omnetpp and
>> 483.xalancbmk).
>>
>> Looks like my mailer is corrupting the spacing, which makes it harder
>> to look at the CFG examples in the big header comment block I added.
>> So I have also included the patch as an attachment.
>>
>> Ok for stage 1?
>>
>> Thanks,
>> Teresa
>>
>>   2014-03-26  Teresa Johnson  <tejohnson@google.com>
>>
>>          * tree-ssa-threadupdate.c (struct ssa_local_info_t): New
>>          duplicate_blocks bitmap.
>>          (remove_ctrl_stmt_and_useless_edges): Ditto.
>>          (create_block_for_threading): Ditto.
>>          (compute_path_counts): New function.
>>          (update_profile): Ditto.
>>          (deduce_freq): Ditto.
>>          (recompute_probabilities): Ditto.
>>          (update_joiner_offpath_counts): Ditto.
>>          (ssa_fix_duplicate_block_edges): Update profile info.
>>          (ssa_create_duplicates): Pass new parameter.
>>          (ssa_redirect_edges): Remove old profile update.
>>          (thread_block_1): New duplicate_blocks bitmap,
>>          remove old profile update.
>>          (thread_single_edge): Pass new parameter.
>
> First off, sorry this took so long to get reviewed.
>
> Most of what's going on in here is similar to something I sketched out, but
> never coded up a while back -- with the significant difference that you're
> handling joiner blocks as well.
>
> Everything looks to be well thought through and documented in the code at a
> level I wish existed throughout GCC.
>
> The only thing I see missing is regression tests.  I don't think you need to
> do anything huge here, but it ought to be possible to set up relatively
> simple cases which show the probabilities/counts being updated properly.
>
> Otherwise it looks excellent.  It's pre-approved once you've added some kind
> of testing and fixed the nits noted below.

Thanks! I will fix the issues you note below and create some test
cases before I commit.
Teresa

>
>
>
>> +   In the aboe example, after all jump threading is complete, we will
>
> s/aboe/above/
>
>
>
>> +  struct el *next, *el;
>> +  bitmap in_edge_srcs = BITMAP_ALLOC (NULL);
>> +  for (el = rd->incoming_edges; el; el = next)
>> +    {
>> +      next = el->next;
>> +      bitmap_set_bit (in_edge_srcs, el->e->src->index);
>> +    }
>
> Please add vertical whitespace after this loop, but before declaring
> variables for the next loop.
>
> Jeff
>
Teresa Johnson - Aug. 2, 2014, 5:10 a.m.
On Wed, Jul 23, 2014 at 2:08 PM, Teresa Johnson <tejohnson@google.com> wrote:
> On Tue, Jul 22, 2014 at 7:29 PM, Jeff Law <law@redhat.com> wrote:
>> On 03/26/14 17:44, Teresa Johnson wrote:
>>>
>>> Recently I discovered that the profile updates being performed by jump
>>> threading were incorrect in many cases, particularly in the case where
>>> the threading path contains a joiner. Some of the duplicated
>>> blocks/edges were not getting any counts, leading to incorrect
>>> function splitting and other downstream optimizations, and there were
>>> other insanities as well. After making a few attempts to fix the
>>> handling I ended up completely redesigning the profile update code,
>>> removing a few places throughout the code where it was attempting to
>>> do some updates.
>>>
>>> The biggest complication (see the large comment and example above the
>>> new routine compute_path_counts) is that we duplicate a conditional
>>> jump in the joiner case, possibly multiple times for multiple jump
>>> thread paths through that joiner, and it isn't trivial to figure out
>>> what probability to assign each of the duplicated successor edges (and
>>> the original after threading). Each jump thread path may need to have
>>> a different probability of staying on path through the joiner in order
>>> to keep the counts going out of the threading path sane.
>>>
>>> The patch below was bootstrapped and tested on
>>> x86_64-unknown-linux-gnu, and also tested with a profiledbootstrap. I
>>> additionally tested with cpu2006, confirming that the amount of
>>> resulting cycle samples in the split cold sections reduced, and
>>> through manual inspection that many different cases were now correct.
>>> I also measured performance with cpu2006, running each benchmark
>>> multiple times on a Westmere and see some speedups (453.povray 1-2%,
>>> 403.gcc 1-1.5%, and noisy but positive speedups in 471.omnetpp and
>>> 483.xalancbmk).
>>>
>>> Looks like my mailer is corrupting the spacing, which makes it harder
>>> to look at the CFG examples in the big header comment block I added.
>>> So I have also included the patch as an attachment.
>>>
>>> Ok for stage 1?
>>>
>>> Thanks,
>>> Teresa
>>>
>>>   2014-03-26  Teresa Johnson  <tejohnson@google.com>
>>>
>>>          * tree-ssa-threadupdate.c (struct ssa_local_info_t): New
>>>          duplicate_blocks bitmap.
>>>          (remove_ctrl_stmt_and_useless_edges): Ditto.
>>>          (create_block_for_threading): Ditto.
>>>          (compute_path_counts): New function.
>>>          (update_profile): Ditto.
>>>          (deduce_freq): Ditto.
>>>          (recompute_probabilities): Ditto.
>>>          (update_joiner_offpath_counts): Ditto.
>>>          (ssa_fix_duplicate_block_edges): Update profile info.
>>>          (ssa_create_duplicates): Pass new parameter.
>>>          (ssa_redirect_edges): Remove old profile update.
>>>          (thread_block_1): New duplicate_blocks bitmap,
>>>          remove old profile update.
>>>          (thread_single_edge): Pass new parameter.
>>
>> First off, sorry this took so long to get reviewed.
>>
>> Most of what's going on in here is similar to something I sketched out, but
>> never coded up a while back -- with the significant difference that you're
>> handling joiner blocks as well.
>>
>> Everything looks to be well thought through and documented in the code at a
>> level I wish existed throughout GCC.
>>
>> The only thing I see missing is regression tests.  I don't think you need to
>> do anything huge here, but it ought to be possible to set up relatively
>> simple cases which show the probabilities/counts being updated properly.
>>
>> Otherwise it looks excellent.  It's pre-approved once you've added some kind
>> of testing and fixed the nits noted below.
>
> Thanks! I will fix the issues you note below and create some test
> cases before I commit.

Just an update - I found some good test cases by compiling the
c-torture tests with profile feedback with and without my patch. But
in the cases I pulled out I saw that there were still a couple profile
or probability insanities introduced by jump threading (albeit far
less than before), so I wanted to investigate before I commit. I ran
out of time this week and will not get to this until I get back from
vacation the week after next.

Teresa

> Teresa
>
>>
>>
>>
>>> +   In the aboe example, after all jump threading is complete, we will
>>
>> s/aboe/above/
>>
>>
>>
>>> +  struct el *next, *el;
>>> +  bitmap in_edge_srcs = BITMAP_ALLOC (NULL);
>>> +  for (el = rd->incoming_edges; el; el = next)
>>> +    {
>>> +      next = el->next;
>>> +      bitmap_set_bit (in_edge_srcs, el->e->src->index);
>>> +    }
>>
>> Please add vertical whitespace after this loop, but before declaring
>> variables for the next loop.
>>
>> Jeff
>>
>
>
>
> --
> Teresa Johnson | Software Engineer | tejohnson@google.com | 408-460-2413
Andrew Pinski - Aug. 2, 2014, 5:16 a.m.
On Fri, Aug 1, 2014 at 10:10 PM, Teresa Johnson <tejohnson@google.com> wrote:
> On Wed, Jul 23, 2014 at 2:08 PM, Teresa Johnson <tejohnson@google.com> wrote:
>> On Tue, Jul 22, 2014 at 7:29 PM, Jeff Law <law@redhat.com> wrote:
>>> On 03/26/14 17:44, Teresa Johnson wrote:
>>>>
>>>> Recently I discovered that the profile updates being performed by jump
>>>> threading were incorrect in many cases, particularly in the case where
>>>> the threading path contains a joiner. Some of the duplicated
>>>> blocks/edges were not getting any counts, leading to incorrect
>>>> function splitting and other downstream optimizations, and there were
>>>> other insanities as well. After making a few attempts to fix the
>>>> handling I ended up completely redesigning the profile update code,
>>>> removing a few places throughout the code where it was attempting to
>>>> do some updates.
>>>>
>>>> The biggest complication (see the large comment and example above the
>>>> new routine compute_path_counts) is that we duplicate a conditional
>>>> jump in the joiner case, possibly multiple times for multiple jump
>>>> thread paths through that joiner, and it isn't trivial to figure out
>>>> what probability to assign each of the duplicated successor edges (and
>>>> the original after threading). Each jump thread path may need to have
>>>> a different probability of staying on path through the joiner in order
>>>> to keep the counts going out of the threading path sane.
>>>>
>>>> The patch below was bootstrapped and tested on
>>>> x86_64-unknown-linux-gnu, and also tested with a profiledbootstrap. I
>>>> additionally tested with cpu2006, confirming that the amount of
>>>> resulting cycle samples in the split cold sections reduced, and
>>>> through manual inspection that many different cases were now correct.
>>>> I also measured performance with cpu2006, running each benchmark
>>>> multiple times on a Westmere and see some speedups (453.povray 1-2%,
>>>> 403.gcc 1-1.5%, and noisy but positive speedups in 471.omnetpp and
>>>> 483.xalancbmk).
>>>>
>>>> Looks like my mailer is corrupting the spacing, which makes it harder
>>>> to look at the CFG examples in the big header comment block I added.
>>>> So I have also included the patch as an attachment.
>>>>
>>>> Ok for stage 1?
>>>>
>>>> Thanks,
>>>> Teresa
>>>>
>>>>   2014-03-26  Teresa Johnson  <tejohnson@google.com>
>>>>
>>>>          * tree-ssa-threadupdate.c (struct ssa_local_info_t): New
>>>>          duplicate_blocks bitmap.
>>>>          (remove_ctrl_stmt_and_useless_edges): Ditto.
>>>>          (create_block_for_threading): Ditto.
>>>>          (compute_path_counts): New function.
>>>>          (update_profile): Ditto.
>>>>          (deduce_freq): Ditto.
>>>>          (recompute_probabilities): Ditto.
>>>>          (update_joiner_offpath_counts): Ditto.
>>>>          (ssa_fix_duplicate_block_edges): Update profile info.
>>>>          (ssa_create_duplicates): Pass new parameter.
>>>>          (ssa_redirect_edges): Remove old profile update.
>>>>          (thread_block_1): New duplicate_blocks bitmap,
>>>>          remove old profile update.
>>>>          (thread_single_edge): Pass new parameter.
>>>
>>> First off, sorry this took so long to get reviewed.
>>>
>>> Most of what's going on in here is similar to something I sketched out, but
>>> never coded up a while back -- with the significant difference that you're
>>> handling joiner blocks as well.
>>>
>>> Everything looks to be well thought through and documented in the code at a
>>> level I wish existed throughout GCC.
>>>
>>> The only thing I see missing is regression tests.  I don't think you need to
>>> do anything huge here, but it ought to be possible to set up relatively
>>> simple cases which show the probabilities/counts being updated properly.
>>>
>>> Otherwise it looks excellent.  It's pre-approved once you've added some kind
>>> of testing and fixed the nits noted below.
>>
>> Thanks! I will fix the issues you note below and create some test
>> cases before I commit.
>
> Just an update - I found some good test cases by compiling the
> c-torture tests with profile feedback with and without my patch. But
> in the cases I pulled out I saw that there were still a couple profile
> or probability insanities introduced by jump threading (albeit far
> less than before), so I wanted to investigate before I commit. I ran
> out of time this week and will not get to this until I get back from
> vacation the week after next.

Another one to try is
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=22401 .  Jeff and I are
hoping your changes fix this one too.

Thanks,
Andrew Pinski

>
> Teresa
>
>> Teresa
>>
>>>
>>>
>>>
>>>> +   In the aboe example, after all jump threading is complete, we will
>>>
>>> s/aboe/above/
>>>
>>>
>>>
>>>> +  struct el *next, *el;
>>>> +  bitmap in_edge_srcs = BITMAP_ALLOC (NULL);
>>>> +  for (el = rd->incoming_edges; el; el = next)
>>>> +    {
>>>> +      next = el->next;
>>>> +      bitmap_set_bit (in_edge_srcs, el->e->src->index);
>>>> +    }
>>>
>>> Please add vertical whitespace after this loop, but before declaring
>>> variables for the next loop.
>>>
>>> Jeff
>>>
>>
>>
>>
>> --
>> Teresa Johnson | Software Engineer | tejohnson@google.com | 408-460-2413
>
>
>
> --
> Teresa Johnson | Software Engineer | tejohnson@google.com | 408-460-2413

Patch

Index: tree-ssa-threadupdate.c
===================================================================
--- tree-ssa-threadupdate.c     (revision 208491)
+++ tree-ssa-threadupdate.c     (working copy)
@@ -229,6 +229,9 @@  struct ssa_local_info_t

   /* TRUE if we thread one or more jumps, FALSE otherwise.  */
   bool jumps_threaded;
+
+  /* Blocks duplicated for the thread.  */
+  bitmap duplicate_blocks;
 };

 /* Passes which use the jump threading code register jump threading
@@ -292,7 +295,8 @@  remove_ctrl_stmt_and_useless_edges (basic_block bb
 static void
 create_block_for_threading (basic_block bb,
                            struct redirection_data *rd,
-                           unsigned int count)
+                           unsigned int count,
+                           bitmap *duplicate_blocks)
 {
   edge_iterator ei;
   edge e;
@@ -307,6 +311,8 @@  create_block_for_threading (basic_block bb,
   /* Zero out the profile, since the block is unreachable for now.  */
   rd->dup_blocks[count]->frequency = 0;
   rd->dup_blocks[count]->count = 0;
+  if (duplicate_blocks)
+    bitmap_set_bit (*duplicate_blocks, rd->dup_blocks[count]->index);
 }

 /* Main data structure to hold information for duplicates of BB.  */
@@ -495,17 +501,457 @@  any_remaining_duplicated_blocks (vec<jump_thread_e
   return false;
 }

+
+/* Compute the amount of profile count/frequency coming into the jump threading
+   path stored in RD that we are duplicating, returned in PATH_IN_COUNT_PTR and
+   PATH_IN_FREQ_PTR, as well as the amount of counts flowing out of the
+   duplicated path, returned in PATH_OUT_COUNT_PTR.  LOCAL_INFO is used to
+   identify blocks duplicated for jump threading, which have duplicated
+   edges that need to be ignored in the analysis.  Return true if path contains
+   a joiner, false otherwise.
+
+   In the non-joiner case, this is straightforward - all the counts/frequency
+   flowing into the jump threading path should flow through the duplicated
+   block and out of the duplicated path.
+
+   In the joiner case, it is very tricky.  Some of the counts flowing into
+   the original path go offpath at the joiner.  The problem is that while
+   we know how much total count goes off-path in the original control flow,
+   we don't know how many of the counts corresponding to just the jump
+   threading path go offpath at the joiner.
+
+   For example, assume we have the following control flow and identified
+   jump threading paths:
+
+                A     B     C
+                 \    |    /
+               Ea \   |Eb / Ec
+                   \  |  /
+                    v v v
+                      J       <-- Joiner
+                     / \
+                Eoff/   \Eon
+                   /     \
+                  v       v
+                Soff     Son  <--- Normal
+                         /\
+                      Ed/  \ Ee
+                       /    \
+                      v     v
+                      D      E
+
+            Jump threading paths: A -> J -> Son -> D (path 1)
+                                  C -> J -> Son -> E (path 2)
+
+   Note that the control flow could be more complicated:
+   - Each jump threading path may have more than one incoming edge.  I.e. A and
+   Ea could represent multiple incoming blocks/edges that are included in
+   path 1.
+   - There could be EDGE_NO_COPY_SRC_BLOCK edges after the joiner (either
+   before or after the "normal" copy block).  These are not duplicated onto
+   the jump threading path, as they are single-successor.
+   - Any of the blocks along the path may have other incoming edges that
+   are not part of any jump threading path, but add profile counts along
+   the path.
+
+   In the aboe example, after all jump threading is complete, we will
+   end up with the following control flow:
+
+                A          B            C
+                |          |            |
+              Ea|          |Eb          |Ec
+                |          |            |
+                v          v            v
+               Ja          J           Jc
+               / \        / \Eon'     / \
+          Eona/   \   ---/---\--------   \Eonc
+             /     \ /  /     \           \
+            v       v  v       v          v
+           Sona     Soff      Son        Sonc
+             \                 /\         /
+              \___________    /  \  _____/
+                          \  /    \/
+                           vv      v
+                            D      E
+
+   The main issue to notice here is that when we are processing path 1
+   (A->J->Son->D) we need to figure out the outgoing edge weights to
+   the duplicated edges Ja->Sona and Ja->Soff, while ensuring that the
+   sum of the incoming weights to D remain Ed.  The problem with simply
+   assuming that Ja (and Jc when processing path 2) has the same outgoing
+   probabilities to its successors as the original block J, is that after
+   all paths are processed and other edges/counts removed (e.g. none
+   of Ec will reach D after processing path 2), we may end up with not
+   enough count flowing along duplicated edge Sona->D.
+
+   Therefore, in the case of a joiner, we keep track of all counts
+   coming in along the current path, as well as from predecessors not
+   on any jump threading path (Eb in the above example).  While we
+   first assume that the duplicated Eona for Ja->Sona has the same
+   probability as the original, we later compensate for other jump
+   threading paths that may eliminate edges.  We do that by keep track
+   of all counts coming into the original path that are not in a jump
+   thread (Eb in the above example, but as noted earlier, there could
+   be other predecessors incoming to the path at various points, such
+   as at Son).  Call this cumulative non-path count coming into the path
+   before D as Enonpath.  We then ensure that the count from Sona->D is as at
+   least as big as (Ed - Enonpath), but no bigger than the minimum
+   weight along the jump threading path.  The probabilities of both the
+   original and duplicated joiner block J and Ja will be adjusted
+   accordingly after the updates.  */
+
+static bool
+compute_path_counts (struct redirection_data *rd,
+                     ssa_local_info_t *local_info,
+                     gcov_type *path_in_count_ptr,
+                     gcov_type *path_out_count_ptr,
+                     int *path_in_freq_ptr)
+{
+  edge e = rd->incoming_edges->e;
+  vec<jump_thread_edge *> *path = THREAD_PATH (e);
+  edge elast = path->last ()->e;
+  gcov_type nonpath_count = 0;
+  bool has_joiner = false;
+  gcov_type path_in_count = 0;
+  int path_in_freq = 0;
+
+  /* Start by accumulating incoming edge counts to the path's first bb
+     into a couple buckets:
+        path_in_count: total count of incoming edges that flow into the
+                  current path.
+        nonpath_count: total count of incoming edges that are not
+                  flowing along *any* path.  These are the counts
+                  that will still flow along the original path after
+                  all path duplication is done by potentially multiple
+                  calls to this routine.
+     (any other incoming edge counts are for a different jump threading
+     path that will be handled by a later call to this routine.)
+     To make this easier, start by recording all incoming edges that flow into
+     the current path in a bitmap.  We could add up the path's incoming edge
+     counts here, but we still need to walk all the first bb's incoming edges
+     below to add up the counts of the other edges not included in this jump
+     threading path.  */
+  struct el *next, *el;
+  bitmap in_edge_srcs = BITMAP_ALLOC (NULL);
+  for (el = rd->incoming_edges; el; el = next)
+    {
+      next = el->next;
+      bitmap_set_bit (in_edge_srcs, el->e->src->index);
+    }
+  edge ein;
+  edge_iterator ei;
+  FOR_EACH_EDGE (ein, ei, e->dest->preds)
+    {
+      vec<jump_thread_edge *> *ein_path = THREAD_PATH (ein);
+      /* Simply check the incoming edge src against the set captured above.  */
+      if (ein_path
+          && bitmap_bit_p (in_edge_srcs, (*ein_path)[0]->e->src->index))
+        {
+          /* It is necessary but not sufficient that the last path edges
+             are identical.  There may be different paths that share the
+             same last path edge in the case where the last edge has a nocopy
+             source block.  */
+          gcc_assert (ein_path->last ()->e == elast);
+          path_in_count += ein->count;
+          path_in_freq += EDGE_FREQUENCY (ein);
+        }
+      else if (!ein_path)
+        {
+          /* Keep track of the incoming edges that are not on any
jump-threading
+             path.  These counts will still flow out of original path after all
+             jump threading is complete.  */
+            nonpath_count += ein->count;
+        }
+    }
+  BITMAP_FREE (in_edge_srcs);
+
+  /* Now compute the fraction of the total count coming into the first
+     path bb that is from the current threading path.  */
+  gcov_type total_count = e->dest->count;
+  /* Handle incoming profile insanities.  */
+  if (total_count < path_in_count)
+    path_in_count = total_count;
+  int onpath_scale = GCOV_COMPUTE_SCALE (path_in_count, total_count);
+
+  /* Walk the entire path to do some more computation in order to estimate
+     how much of the path_in_count will flow out of the duplicated threading
+     path.  In the non-joiner case this is straightforward (it should be
+     the same as path_in_count, although we will handle incoming profile
+     insanities by setting it equal to the minimum count along the path).
+
+     In the joiner case, we need to estimate how much of the path_in_count
+     will stay on the threading path after the joiner's conditional branch.
+     We don't really know for sure how much of the counts
+     associated with this path go to each successor of the joiner, but we'll
+     estimate based on the fraction of the total count coming into the path
+     bb was from the threading paths (computed above in onpath_scale).
+     Afterwards, we will need to do some fixup to account for other threading
+     paths and possible profile insanities.
+
+     In order to estimate the joiner case's counts we also need to update
+     nonpath_count with any additional counts coming into the path.  Other
+     blocks along the path may have additional predecessors from outside
+     the path.  */
+  gcov_type path_out_count = path_in_count;
+  gcov_type min_path_count = path_in_count;
+  for (unsigned int i = 1; i < path->length (); i++)
+    {
+      edge epath = (*path)[i]->e;
+      gcov_type cur_count = epath->count;
+      if ((*path)[i]->type == EDGE_COPY_SRC_JOINER_BLOCK)
+        {
+          has_joiner = true;
+          cur_count = apply_probability (cur_count, onpath_scale);
+        }
+      /* In the joiner case we need to update nonpath_count for any edges
+         coming into the path that will contribute to the count flowing
+         into the path successor.  */
+      if (has_joiner && epath != elast)
+      {
+        /* Look for other incoming edges after joiner.  */
+        FOR_EACH_EDGE (ein, ei, epath->dest->preds)
+          {
+            if (ein != epath
+                /* Ignore in edges from blocks we have duplicated for a
+                   threading path, which have duplicated edge counts until
+                   they are redirected by an invocation of this routine.  */
+                && !bitmap_bit_p (local_info->duplicate_blocks,
+                                  ein->src->index))
+              nonpath_count += ein->count;
+          }
+      }
+      if (cur_count < path_out_count)
+        path_out_count = cur_count;
+      if (epath->count < min_path_count)
+        min_path_count = epath->count;
+    }
+
+  /* We computed path_out_count above assuming that this path targeted
+     the joiner's on-path successor with the same likelihood as it
+     reached the joiner.  However, other thread paths through the joiner
+     may take a different path through the normal copy source block
+     (i.e. they have a different elast), meaning that they do not
+     contribute any counts to this path's elast.  As a result, it may
+     turn out that this path must have more count flowing to the on-path
+     successor of the joiner.  Essentially, all of this path's elast
+     count must be contributed by this path and any nonpath counts
+     (since any path through the joiner with a different elast will not
+     include a copy of this elast in its duplicated path).
+     So ensure that this path's path_out_count is at least the
+     difference between elast->count and nonpath_count.  Otherwise the edge
+     counts after threading will not be sane.  */
+  if (has_joiner && path_out_count < elast->count - nonpath_count)
+  {
+    path_out_count = elast->count - nonpath_count;
+    /* But neither can we go above the minimum count along the path
+       we are duplicating.  This can be an issue due to profile
+       insanities coming in to this pass.  */
+    if (path_out_count > min_path_count)
+      path_out_count = min_path_count;
+  }
+
+  *path_in_count_ptr = path_in_count;
+  *path_out_count_ptr = path_out_count;
+  *path_in_freq_ptr = path_in_freq;
+  return has_joiner;
+}
+
+
+/* Update the counts and frequencies for both an original path
+   edge EPATH and its duplicate EDUP.  The duplicate source block
+   will get a count/frequency of PATH_IN_COUNT and PATH_IN_FREQ,
+   and the duplicate edge EDUP will have a count of PATH_OUT_COUNT.  */
+static void
+update_profile (edge epath, edge edup, gcov_type path_in_count,
+                gcov_type path_out_count, int path_in_freq)
+{
+
+  /* First update the duplicated block's count / frequency.  */
+  if (edup)
+    {
+      basic_block dup_block = edup->src;
+      gcc_assert (dup_block->count == 0);
+      gcc_assert (dup_block->frequency == 0);
+      dup_block->count = path_in_count;
+      dup_block->frequency = path_in_freq;
+    }
+
+  /* Now update the original block's count and frequency in the
+     opposite manner - remove the counts/freq that will flow
+     into the duplicated block.  Handle underflow due to precision/
+     rounding issues.  */
+  epath->src->count -= path_in_count;
+  if (epath->src->count < 0)
+    epath->src->count = 0;
+  epath->src->frequency -= path_in_freq;
+  if (epath->src->frequency < 0)
+    epath->src->frequency = 0;
+
+  /* Next update this path edge's original and duplicated counts.  We know
+     that the duplicated path will have path_out_count flowing
+     out of it (in the joiner case this is the count along the duplicated path
+     out of the duplicated joiner).  This count can then be removed from the
+     original path edge.  */
+  if (edup)
+    edup->count = path_out_count;
+  epath->count -= path_out_count;
+  gcc_assert (epath->count >= 0);
+}
+
+
+/* This routine will determine what frequency is needed by an edge E
+   to ensure that its destination block meets its current block frequency.
+   This is used when we don't know E's probability.  */
+
+static int
+deduce_freq (edge e)
+{
+  edge_iterator ei;
+  edge epred;
+  int freq = e->dest->frequency;
+  FOR_EACH_EDGE (epred, ei, e->dest->preds)
+    {
+      if (epred == e)
+        continue;
+      freq -= EDGE_FREQUENCY (epred);
+    }
+  if (freq < 0)
+    freq = 0;
+  return freq;
+}
+
+
+/* The duplicate and original joiner blocks may end up with different
+   probabilities (different from both the original and from each other).
+   Recompute the probabilities here once we have updated the edge
+   counts and frequencies.  */
+
+static void
+recompute_probabilities (basic_block bb)
+{
+  edge esucc;
+  edge_iterator ei;
+  FOR_EACH_EDGE (esucc, ei, bb->succs)
+    {
+      if (bb->count)
+        esucc->probability = GCOV_COMPUTE_SCALE (esucc->count,
+                                                 bb->count);
+      else if (bb->frequency)
+        esucc->probability = GCOV_COMPUTE_SCALE (deduce_freq (esucc),
+                                                 bb->frequency);
+      if (esucc->probability > REG_BR_PROB_BASE)
+        {
+         /* Can happen with missing/guessed probabilities, since we
+            may determine that more is flowing along duplicated
+            path than joiner succ probabilities allowed.
+            Counts and freqs will be insane after jump threading,
+            at least make sure probability is sane or we will
+            get a flow verification error.
+            Not much we can do to make counts/freqs sane without
+            redoing the profile estimation.  */
+         esucc->probability = REG_BR_PROB_BASE;
+       }
+    }
+}
+
+
+/* Update the counts of the original and duplicated edges from a joiner
+   that go off path, given that we have already determined that the
+   duplicate joiner DUP_BB has incoming count PATH_IN_COUNT and
+   outgoing count along the path PATH_OUT_COUNT.  The original (on-)path
+   edge from joiner is EPATH.  */
+
+static void
+update_joiner_offpath_counts (edge epath, basic_block dup_bb,
+                              gcov_type path_in_count,
+                              gcov_type path_out_count)
+{
+  /* Compute the count that currently flows off path from the joiner.
+     In other words, the total count of joiner's out edges other than
+     epath.  Compute this by walking the successors instead of
+     subtracting epath's count from the joiner bb count, since there
+     are sometimes slight insanities where the total out edge count is
+     larger than the bb count (possibly due to rounding/truncation
+     errors).  */
+  gcov_type total_orig_off_path_count = 0;
+  edge enonpath;
+  edge_iterator ei;
+  FOR_EACH_EDGE (enonpath, ei, epath->src->succs)
+    {
+      if (enonpath == epath)
+        continue;
+      total_orig_off_path_count += enonpath->count;
+    }
+
+  /* For the path that we are duplicating, the amount that will flow
+     off path from the duplicated joiner is the delta between the
+     path's cumulative in count and the portion of that count we
+     estimated above as flowing from the joiner along the duplicated
+     path.  */
+  gcov_type total_dup_off_path_count = path_in_count - path_out_count;
+
+  /* Now do the actual updates of the off-path edges.  */
+  FOR_EACH_EDGE (enonpath, ei, epath->src->succs)
+    {
+      /* Look for edges going off of the threading path.  */
+      if (enonpath == epath)
+        continue;
+
+      /* Find the corresponding edge out of the duplicated joiner.  */
+      edge enonpathdup = find_edge (dup_bb, enonpath->dest);
+      gcc_assert (enonpathdup);
+
+      /* We can't use the original probability of the joiner's out
+         edges, since the probabilities of the original branch
+         and the duplicated branches may vary after all threading is
+         complete.  But apportion the duplicated joiner's off-path
+         total edge count computed earlier (total_dup_off_path_count)
+         among the duplicated off-path edges based on their original
+         ratio to the full off-path count (total_orig_off_path_count).
+         */
+      int scale = GCOV_COMPUTE_SCALE (enonpath->count,
+                                      total_orig_off_path_count);
+      /* Give the duplicated offpath edge a portion of the duplicated
+         total.  */
+      enonpathdup->count = apply_scale (scale,
+                                        total_dup_off_path_count);
+      /* Now update the original offpath edge count, handling underflow
+         due to rounding errors.  */
+      enonpath->count -= enonpathdup->count;
+      if (enonpath->count < 0)
+        enonpath->count = 0;
+    }
+}
+
 /* Wire up the outgoing edges from the duplicate blocks and
-   update any PHIs as needed.  */
+   update any PHIs as needed.  Also update the profile counts
+   on the original and duplicate blocks and edges.  */
 void
 ssa_fix_duplicate_block_edges (struct redirection_data *rd,
                               ssa_local_info_t *local_info)
 {
   edge e = rd->incoming_edges->e;
   vec<jump_thread_edge *> *path = THREAD_PATH (e);
+  edge elast = path->last ()->e;
+  gcov_type path_in_count = 0;
+  gcov_type path_out_count = 0;
+  int path_in_freq = 0;

+  /* First determine how much profile count to move from original
+     path to the duplicate path.  This is tricky in the presence of
+     a joiner (see comments for compute_path_counts), where some portion
+     of the path's counts will flow off-path from the joiner.  In the
+     non-joiner case the path_in_count and path_out_count should be the
+     same.  */
+  bool has_joiner = compute_path_counts (rd, local_info,
+                                         &path_in_count, &path_out_count,
+                                         &path_in_freq);
+
+  int cur_path_freq = path_in_freq;
   for (unsigned int count = 0, i = 1; i < path->length (); i++)
     {
+      edge epath = (*path)[i]->e;
+
       /* If we were threading through an joiner block, then we want
         to keep its control statement and redirect an outgoing edge.