* [PATCH] Fix PR 58960
@ 2014-01-30 5:42 Andrey Belevantsev
2014-02-07 5:25 ` Andrey Belevantsev
2014-02-13 16:31 ` Vladimir Makarov
0 siblings, 2 replies; 3+ messages in thread
From: Andrey Belevantsev @ 2014-01-30 5:42 UTC (permalink / raw)
To: GCC Patches; +Cc: Vladimir N. Makarov
[-- Attachment #1: Type: text/plain, Size: 1184 bytes --]
Hello,
As detailed in http://gcc.gnu.org/bugzilla/show_bug.cgi?id=58960#c6, we
fail to use the DF liveness info in the register pressure sensitive
scheduling for the new blocks as we do not properly compute it in this
case. The patch fixes this by avoiding to use the sched-pressure for the
new regions, as currently these are only ia64 recovery blocks and supposed
to be cold. In the case we'd get other cases of the new blocks, this may
be reconsidered. The other options of computing the DF info sketched at
the above link do not seem plausible for this stage.
Bootstrapped and tested on ia64, also tested by Andreas Schwab on ia64 (see
PR log). OK for trunk?
Andrey
2013-01-30 Andrey Belevantsev <abel@ispras.ru>
PR rtl-optimization/58960
* haifa-sched.c (alloc_global_sched_pressure_data): New, factored out from ...
(sched_init) ... here.
(free_global_sched_pressure_data): New, factored out from ...
(sched_finish): ... here.
* sched-int.h (free_global_sched_pressure_data): Declare.
* sched-rgn.c (nr_regions_initial): New static global.
(haifa_find_rgns): Initialize it.
(schedule_region): Disable sched-pressure for the newly generated regions.
[-- Attachment #2: pr58960.diff --]
[-- Type: text/x-patch, Size: 5078 bytes --]
diff --git a/gcc/haifa-sched.c b/gcc/haifa-sched.c
index 4d51984..45398bf 100644
--- a/gcc/haifa-sched.c
+++ b/gcc/haifa-sched.c
@@ -6553,6 +6553,53 @@ setup_sched_dump (void)
? stderr : dump_file);
}
+/* Allocate data for register pressure sensitive scheduling. */
+static void
+alloc_global_sched_pressure_data (void)
+{
+ if (sched_pressure != SCHED_PRESSURE_NONE)
+ {
+ int i, max_regno = max_reg_num ();
+
+ if (sched_dump != NULL)
+ /* We need info about pseudos for rtl dumps about pseudo
+ classes and costs. */
+ regstat_init_n_sets_and_refs ();
+ ira_set_pseudo_classes (true, sched_verbose ? sched_dump : NULL);
+ sched_regno_pressure_class
+ = (enum reg_class *) xmalloc (max_regno * sizeof (enum reg_class));
+ for (i = 0; i < max_regno; i++)
+ sched_regno_pressure_class[i]
+ = (i < FIRST_PSEUDO_REGISTER
+ ? ira_pressure_class_translate[REGNO_REG_CLASS (i)]
+ : ira_pressure_class_translate[reg_allocno_class (i)]);
+ curr_reg_live = BITMAP_ALLOC (NULL);
+ if (sched_pressure == SCHED_PRESSURE_WEIGHTED)
+ {
+ saved_reg_live = BITMAP_ALLOC (NULL);
+ region_ref_regs = BITMAP_ALLOC (NULL);
+ }
+ }
+}
+
+/* Free data for register pressure sensitive scheduling. */
+void
+free_global_sched_pressure_data (void)
+{
+ if (sched_pressure != SCHED_PRESSURE_NONE)
+ {
+ if (regstat_n_sets_and_refs != NULL)
+ regstat_free_n_sets_and_refs ();
+ if (sched_pressure == SCHED_PRESSURE_WEIGHTED)
+ {
+ BITMAP_FREE (region_ref_regs);
+ BITMAP_FREE (saved_reg_live);
+ }
+ BITMAP_FREE (curr_reg_live);
+ free (sched_regno_pressure_class);
+ }
+}
+
/* Initialize some global state for the scheduler. This function works
with the common data shared between all the schedulers. It is called
from the scheduler specific initialization routine. */
@@ -6656,29 +6703,7 @@ sched_init (void)
if (targetm.sched.init_global)
targetm.sched.init_global (sched_dump, sched_verbose, get_max_uid () + 1);
- if (sched_pressure != SCHED_PRESSURE_NONE)
- {
- int i, max_regno = max_reg_num ();
-
- if (sched_dump != NULL)
- /* We need info about pseudos for rtl dumps about pseudo
- classes and costs. */
- regstat_init_n_sets_and_refs ();
- ira_set_pseudo_classes (true, sched_verbose ? sched_dump : NULL);
- sched_regno_pressure_class
- = (enum reg_class *) xmalloc (max_regno * sizeof (enum reg_class));
- for (i = 0; i < max_regno; i++)
- sched_regno_pressure_class[i]
- = (i < FIRST_PSEUDO_REGISTER
- ? ira_pressure_class_translate[REGNO_REG_CLASS (i)]
- : ira_pressure_class_translate[reg_allocno_class (i)]);
- curr_reg_live = BITMAP_ALLOC (NULL);
- if (sched_pressure == SCHED_PRESSURE_WEIGHTED)
- {
- saved_reg_live = BITMAP_ALLOC (NULL);
- region_ref_regs = BITMAP_ALLOC (NULL);
- }
- }
+ alloc_global_sched_pressure_data ();
curr_state = xmalloc (dfa_state_size);
}
@@ -6777,18 +6802,7 @@ void
sched_finish (void)
{
haifa_finish_h_i_d ();
- if (sched_pressure != SCHED_PRESSURE_NONE)
- {
- if (regstat_n_sets_and_refs != NULL)
- regstat_free_n_sets_and_refs ();
- if (sched_pressure == SCHED_PRESSURE_WEIGHTED)
- {
- BITMAP_FREE (region_ref_regs);
- BITMAP_FREE (saved_reg_live);
- }
- BITMAP_FREE (curr_reg_live);
- free (sched_regno_pressure_class);
- }
+ free_global_sched_pressure_data ();
free (curr_state);
if (targetm.sched.finish_global)
diff --git a/gcc/sched-int.h b/gcc/sched-int.h
index 3b1106f..b5481b9 100644
--- a/gcc/sched-int.h
+++ b/gcc/sched-int.h
@@ -1337,6 +1337,7 @@ extern void debug_ds (ds_t);
extern void initialize_live_range_shrinkage (void);
extern void finish_live_range_shrinkage (void);
extern void sched_init_region_reg_pressure_info (void);
+extern void free_global_sched_pressure_data (void);
extern int haifa_classify_insn (const_rtx);
extern void get_ebb_head_tail (basic_block, basic_block, rtx *, rtx *);
extern int no_real_insns_p (const_rtx, const_rtx);
diff --git a/gcc/sched-rgn.c b/gcc/sched-rgn.c
index 406dc1f..0573b6a 100644
--- a/gcc/sched-rgn.c
+++ b/gcc/sched-rgn.c
@@ -79,6 +79,9 @@ static int is_cfg_nonregular (void);
/* Number of regions in the procedure. */
int nr_regions = 0;
+/* Same as above before adding any new regions. */
+static int nr_regions_initial = 0;
+
/* Table of region descriptions. */
region *rgn_table = NULL;
@@ -1064,6 +1067,7 @@ haifa_find_rgns (void)
BLOCK_TO_BB (bb->index) = 0;
}
+ nr_regions_initial = nr_regions;
free (max_hdr);
free (degree);
free (stack);
@@ -2991,6 +2995,15 @@ schedule_region (int rgn)
rgn_n_insns = 0;
+ /* Do not support register pressure sensitive scheduling for the new regions
+ as we don't update the liveness info for them. */
+ if (rgn >= nr_regions_initial)
+ {
+ if (sched_pressure != SCHED_PRESSURE_NONE)
+ free_global_sched_pressure_data ();
+ sched_pressure = SCHED_PRESSURE_NONE;
+ }
+
rgn_setup_region (rgn);
/* Don't schedule region that is marked by
^ permalink raw reply [flat|nested] 3+ messages in thread
* Re: [PATCH] Fix PR 58960
2014-01-30 5:42 [PATCH] Fix PR 58960 Andrey Belevantsev
@ 2014-02-07 5:25 ` Andrey Belevantsev
2014-02-13 16:31 ` Vladimir Makarov
1 sibling, 0 replies; 3+ messages in thread
From: Andrey Belevantsev @ 2014-02-07 5:25 UTC (permalink / raw)
To: GCC Patches; +Cc: Vladimir N. Makarov
On 30.01.2014 9:42, Andrey Belevantsev wrote:
> Hello,
Ping.
>
> As detailed in http://gcc.gnu.org/bugzilla/show_bug.cgi?id=58960#c6, we
> fail to use the DF liveness info in the register pressure sensitive
> scheduling for the new blocks as we do not properly compute it in this
> case. The patch fixes this by avoiding to use the sched-pressure for the
> new regions, as currently these are only ia64 recovery blocks and supposed
> to be cold. In the case we'd get other cases of the new blocks, this may
> be reconsidered. The other options of computing the DF info sketched at
> the above link do not seem plausible for this stage.
>
> Bootstrapped and tested on ia64, also tested by Andreas Schwab on ia64 (see
> PR log). OK for trunk?
>
> Andrey
>
> 2013-01-30 Andrey Belevantsev <abel@ispras.ru>
>
> PR rtl-optimization/58960
>
> * haifa-sched.c (alloc_global_sched_pressure_data): New, factored out
> from ...
> (sched_init) ... here.
> (free_global_sched_pressure_data): New, factored out from ...
> (sched_finish): ... here.
> * sched-int.h (free_global_sched_pressure_data): Declare.
> * sched-rgn.c (nr_regions_initial): New static global.
> (haifa_find_rgns): Initialize it.
> (schedule_region): Disable sched-pressure for the newly generated
> regions.
^ permalink raw reply [flat|nested] 3+ messages in thread
* Re: [PATCH] Fix PR 58960
2014-01-30 5:42 [PATCH] Fix PR 58960 Andrey Belevantsev
2014-02-07 5:25 ` Andrey Belevantsev
@ 2014-02-13 16:31 ` Vladimir Makarov
1 sibling, 0 replies; 3+ messages in thread
From: Vladimir Makarov @ 2014-02-13 16:31 UTC (permalink / raw)
To: Andrey Belevantsev, GCC Patches
On 01/30/2014 12:42 AM, Andrey Belevantsev wrote:
> Hello,
>
> As detailed in http://gcc.gnu.org/bugzilla/show_bug.cgi?id=58960#c6,
> we fail to use the DF liveness info in the register pressure sensitive
> scheduling for the new blocks as we do not properly compute it in this
> case. The patch fixes this by avoiding to use the sched-pressure for
> the new regions, as currently these are only ia64 recovery blocks and
> supposed to be cold. In the case we'd get other cases of the new
> blocks, this may be reconsidered. The other options of computing the
> DF info sketched at the above link do not seem plausible for this stage.
>
> Bootstrapped and tested on ia64, also tested by Andreas Schwab on ia64
> (see PR log). OK for trunk?
>
>
The patch is ok. Andrey, thanks for working on the PR and sorry for the
delay with the approval.
>
> 2013-01-30 Andrey Belevantsev <abel@ispras.ru>
>
> PR rtl-optimization/58960
>
> * haifa-sched.c (alloc_global_sched_pressure_data): New, factored
> out from ...
> (sched_init) ... here.
> (free_global_sched_pressure_data): New, factored out from ...
> (sched_finish): ... here.
> * sched-int.h (free_global_sched_pressure_data): Declare.
> * sched-rgn.c (nr_regions_initial): New static global.
> (haifa_find_rgns): Initialize it.
> (schedule_region): Disable sched-pressure for the newly generated
> regions.
^ permalink raw reply [flat|nested] 3+ messages in thread
end of thread, other threads:[~2014-02-13 16:31 UTC | newest]
Thread overview: 3+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2014-01-30 5:42 [PATCH] Fix PR 58960 Andrey Belevantsev
2014-02-07 5:25 ` Andrey Belevantsev
2014-02-13 16:31 ` Vladimir Makarov
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).