* [PATCH] Speedup DF dataflow solver
@ 2023-02-14 14:21 Richard Biener
2023-02-15 10:33 ` Jakub Jelinek
0 siblings, 1 reply; 2+ messages in thread
From: Richard Biener @ 2023-02-14 14:21 UTC (permalink / raw)
To: gcc-patches; +Cc: Jakub Jelinek
The following makes sure to process blocks that follow the current
block in the iteration order in the same iteration and only postpone
blocks that would be visited earlier to the next iteration.
For the all.i testcase in PR26854 at -O2 this shaves off 50% of
the time to solve the DF RD problem, other problems also improve
but not as drastically.
Bootstrapped on x86_64-unknown-linux-gnu, testing in progress.
OK for trunk?
Thanks,
Richard.
PR middle-end/26854
* df-core.cc (df_worklist_propagate_forward): Put later
blocks on worklist and only earlier blocks on pending.
(df_worklist_propagate_backward): Likewise.
(df_worklist_dataflow_doublequeue): Change the iteration
to process new blocks in the same iteration if that
maintains the iteration order.
---
gcc/df-core.cc | 54 ++++++++++++++++++++++++++++++++------------------
1 file changed, 35 insertions(+), 19 deletions(-)
diff --git a/gcc/df-core.cc b/gcc/df-core.cc
index e5ae9ab9348..38f69ac5743 100644
--- a/gcc/df-core.cc
+++ b/gcc/df-core.cc
@@ -874,7 +874,8 @@ make_pass_df_finish (gcc::context *ctxt)
/* Helper function for df_worklist_dataflow.
Propagate the dataflow forward.
Given a BB_INDEX, do the dataflow propagation
- and set bits on for successors in PENDING
+ and set bits on for successors in PENDING for earlier
+ and WORKLIST for later in bbindex_to_postorder
if the out set of the dataflow has changed.
AGE specify time when BB was visited last time.
@@ -890,10 +891,11 @@ make_pass_df_finish (gcc::context *ctxt)
static bool
df_worklist_propagate_forward (struct dataflow *dataflow,
- unsigned bb_index,
- unsigned *bbindex_to_postorder,
- bitmap pending,
- sbitmap considered,
+ unsigned bb_index,
+ unsigned *bbindex_to_postorder,
+ bitmap worklist,
+ bitmap pending,
+ sbitmap considered,
vec<int> &last_change_age,
int age)
{
@@ -924,7 +926,13 @@ df_worklist_propagate_forward (struct dataflow *dataflow,
unsigned ob_index = e->dest->index;
if (bitmap_bit_p (considered, ob_index))
- bitmap_set_bit (pending, bbindex_to_postorder[ob_index]);
+ {
+ if (bbindex_to_postorder[bb_index]
+ < bbindex_to_postorder[ob_index])
+ bitmap_set_bit (worklist, bbindex_to_postorder[ob_index]);
+ else
+ bitmap_set_bit (pending, bbindex_to_postorder[ob_index]);
+ }
}
return true;
}
@@ -937,10 +945,11 @@ df_worklist_propagate_forward (struct dataflow *dataflow,
static bool
df_worklist_propagate_backward (struct dataflow *dataflow,
- unsigned bb_index,
- unsigned *bbindex_to_postorder,
- bitmap pending,
- sbitmap considered,
+ unsigned bb_index,
+ unsigned *bbindex_to_postorder,
+ bitmap worklist,
+ bitmap pending,
+ sbitmap considered,
vec<int> &last_change_age,
int age)
{
@@ -971,7 +980,13 @@ df_worklist_propagate_backward (struct dataflow *dataflow,
unsigned ob_index = e->src->index;
if (bitmap_bit_p (considered, ob_index))
- bitmap_set_bit (pending, bbindex_to_postorder[ob_index]);
+ {
+ if (bbindex_to_postorder[bb_index]
+ < bbindex_to_postorder[ob_index])
+ bitmap_set_bit (worklist, bbindex_to_postorder[ob_index]);
+ else
+ bitmap_set_bit (pending, bbindex_to_postorder[ob_index]);
+ }
}
return true;
}
@@ -1021,36 +1036,37 @@ df_worklist_dataflow_doublequeue (struct dataflow *dataflow,
and pending is for the next. */
while (!bitmap_empty_p (pending))
{
- bitmap_iterator bi;
- unsigned int index;
-
std::swap (pending, worklist);
- EXECUTE_IF_SET_IN_BITMAP (worklist, 0, index, bi)
+ do
{
+ unsigned index = bitmap_first_set_bit (worklist);
+ bitmap_clear_bit (worklist, index);
+
unsigned bb_index;
dcount++;
- bitmap_clear_bit (pending, index);
bb_index = blocks_in_postorder[index];
prev_age = last_visit_age[index];
if (dir == DF_FORWARD)
changed = df_worklist_propagate_forward (dataflow, bb_index,
bbindex_to_postorder,
- pending, considered,
+ worklist, pending,
+ considered,
last_change_age,
prev_age);
else
changed = df_worklist_propagate_backward (dataflow, bb_index,
bbindex_to_postorder,
- pending, considered,
+ worklist, pending,
+ considered,
last_change_age,
prev_age);
last_visit_age[index] = ++age;
if (changed)
last_change_age[index] = age;
}
- bitmap_clear (worklist);
+ while (!bitmap_empty_p (worklist));
}
BITMAP_FREE (worklist);
--
2.35.3
^ permalink raw reply [flat|nested] 2+ messages in thread
* Re: [PATCH] Speedup DF dataflow solver
2023-02-14 14:21 [PATCH] Speedup DF dataflow solver Richard Biener
@ 2023-02-15 10:33 ` Jakub Jelinek
0 siblings, 0 replies; 2+ messages in thread
From: Jakub Jelinek @ 2023-02-15 10:33 UTC (permalink / raw)
To: Richard Biener; +Cc: gcc-patches
On Tue, Feb 14, 2023 at 03:21:53PM +0100, Richard Biener wrote:
> The following makes sure to process blocks that follow the current
> block in the iteration order in the same iteration and only postpone
> blocks that would be visited earlier to the next iteration.
>
> For the all.i testcase in PR26854 at -O2 this shaves off 50% of
> the time to solve the DF RD problem, other problems also improve
> but not as drastically.
>
> Bootstrapped on x86_64-unknown-linux-gnu, testing in progress.
>
> OK for trunk?
>
> Thanks,
> Richard.
>
> PR middle-end/26854
> * df-core.cc (df_worklist_propagate_forward): Put later
> blocks on worklist and only earlier blocks on pending.
> (df_worklist_propagate_backward): Likewise.
> (df_worklist_dataflow_doublequeue): Change the iteration
> to process new blocks in the same iteration if that
> maintains the iteration order.
LGTM.
Jakub
^ permalink raw reply [flat|nested] 2+ messages in thread
end of thread, other threads:[~2023-02-15 10:33 UTC | newest]
Thread overview: 2+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2023-02-14 14:21 [PATCH] Speedup DF dataflow solver Richard Biener
2023-02-15 10:33 ` Jakub Jelinek
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).