[PATCH PR67909 PR67947]

[PATCH PR67909 PR67947]

[Yuri Rumyantsev]

[-- Attachment #1: Type: text/plain, Size: 497 bytes --]

Hi All,

Here is a simple patch for unswitching outer loop through guard-edge
hoisting. The check that guard-edge is around the inner loop was
missed.

Bootstrapping and regression testing did not show new failures.

Is it OK for trunk?

ChangeLog:
2014-10-13  Yuri Rumyantsev  <ysrumyan@gmail.com>

PR tree-optimization/67909, 67947
* tree-ssa-loop-unswitch.c (find_loop_guard): Add check that GUARD_EDGE
really skip the inner loop.

gcc/testsuite/ChangeLog
* gcc.dg/torture/pr67947.c: New test.

[-- Attachment #2: patch --]
[-- Type: application/octet-stream, Size: 1604 bytes --]

diff --git a/gcc/testsuite/gcc.dg/torture/pr67947.c b/gcc/testsuite/gcc.dg/torture/pr67947.c
new file mode 100644
index 0000000..5664c48
--- /dev/null
+++ b/gcc/testsuite/gcc.dg/torture/pr67947.c
@@ -0,0 +1,30 @@
+/* { dg-additional-options "-O3" } */
+
+#include <stdlib.h>
+
+int a;
+int c;
+__attribute__((noinline, noclone)) void foo (int x)
+{
+  if (x == 0)
+    c++;
+}
+
+int
+main (int argc, char* argv[])
+{
+  int j, k, b = 0;
+  if (argc == 0)
+    b = 1;
+  for (j = 0; j < 3; j++)
+    for (k = 0; k < 1; k++)
+      {
+	foo (0);
+	if (b)
+	  for (k = -1; a;)
+	    ;
+      }
+  if (c != 3)
+    abort ();
+  return 0;
+}
diff --git a/gcc/tree-ssa-loop-unswitch.c b/gcc/tree-ssa-loop-unswitch.c
index 4328d6a..ba8a29e 100644
--- a/gcc/tree-ssa-loop-unswitch.c
+++ b/gcc/tree-ssa-loop-unswitch.c
@@ -466,7 +466,6 @@ 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;
@@ -524,6 +523,15 @@ find_loop_guard (struct loop *loop)
   else
     return NULL;
 
+  /* Guard edge must skip inner loop.  */
+  if (!dominated_by_p (CDI_DOMINATORS, loop->inner->header,
+      guard_edge == fe ? te->dest : fe->dest))
+    {
+      if (dump_file && (dump_flags & TDF_DETAILS))
+	fprintf (dump_file, "Guard edge %d --> %d is not around the loop!\n",guard_edge->src->index,guard_edge->dest->index);
+      return NULL;
+    }
+
   if (dump_file && (dump_flags & TDF_DETAILS))
     fprintf (dump_file,
 	     "Considering guard %d -> %d in loop %d\n",

Re: [PATCH PR67909 PR67947]

[H.J. Lu]

On Tue, Oct 13, 2015 at 4:57 AM, Yuri Rumyantsev <ysrumyan@gmail.com> wrote:
> Hi All,
>
> Here is a simple patch for unswitching outer loop through guard-edge
> hoisting. The check that guard-edge is around the inner loop was
> missed.
>
> Bootstrapping and regression testing did not show new failures.
>
> Is it OK for trunk?
>
> ChangeLog:
> 2014-10-13  Yuri Rumyantsev  <ysrumyan@gmail.com>
>
> PR tree-optimization/67909, 67947
> * tree-ssa-loop-unswitch.c (find_loop_guard): Add check that GUARD_EDGE
> really skip the inner loop.
>
> gcc/testsuite/ChangeLog
> * gcc.dg/torture/pr67947.c: New test.

+  /* Guard edge must skip inner loop.  */
+  if (!dominated_by_p (CDI_DOMINATORS, loop->inner->header,
+      guard_edge == fe ? te->dest : fe->dest))
        ^^^^  It should line up with "CDI_DOMINATORS".

+ fprintf (dump_file, "Guard edge %d --> %d is not around the
loop!\n",guard_edge->src->index,guard_edge->dest->index);

Please break lone line.

-- 
H.J.

Re: [PATCH PR67909 PR67947]

[Yuri Rumyantsev]

[-- Attachment #1: Type: text/plain, Size: 1149 bytes --]

Here is updated patch with splitting long line.
The patch is attached.

Yuri.

2015-10-13 15:38 GMT+03:00 H.J. Lu <hjl.tools@gmail.com>:
> On Tue, Oct 13, 2015 at 4:57 AM, Yuri Rumyantsev <ysrumyan@gmail.com> wrote:
>> Hi All,
>>
>> Here is a simple patch for unswitching outer loop through guard-edge
>> hoisting. The check that guard-edge is around the inner loop was
>> missed.
>>
>> Bootstrapping and regression testing did not show new failures.
>>
>> Is it OK for trunk?
>>
>> ChangeLog:
>> 2014-10-13  Yuri Rumyantsev  <ysrumyan@gmail.com>
>>
>> PR tree-optimization/67909, 67947
>> * tree-ssa-loop-unswitch.c (find_loop_guard): Add check that GUARD_EDGE
>> really skip the inner loop.
>>
>> gcc/testsuite/ChangeLog
>> * gcc.dg/torture/pr67947.c: New test.
>
> +  /* Guard edge must skip inner loop.  */
> +  if (!dominated_by_p (CDI_DOMINATORS, loop->inner->header,
> +      guard_edge == fe ? te->dest : fe->dest))
>         ^^^^  It should line up with "CDI_DOMINATORS".
>
> + fprintf (dump_file, "Guard edge %d --> %d is not around the
> loop!\n",guard_edge->src->index,guard_edge->dest->index);
>
> Please break lone line.
>
> --
> H.J.

[-- Attachment #2: patch1 --]
[-- Type: application/octet-stream, Size: 1610 bytes --]

diff --git a/gcc/testsuite/gcc.dg/torture/pr67947.c b/gcc/testsuite/gcc.dg/torture/pr67947.c
new file mode 100644
index 0000000..5664c48
--- /dev/null
+++ b/gcc/testsuite/gcc.dg/torture/pr67947.c
@@ -0,0 +1,30 @@
+/* { dg-additional-options "-O3" } */
+
+#include <stdlib.h>
+
+int a;
+int c;
+__attribute__((noinline, noclone)) void foo (int x)
+{
+  if (x == 0)
+    c++;
+}
+
+int
+main (int argc, char* argv[])
+{
+  int j, k, b = 0;
+  if (argc == 0)
+    b = 1;
+  for (j = 0; j < 3; j++)
+    for (k = 0; k < 1; k++)
+      {
+	foo (0);
+	if (b)
+	  for (k = -1; a;)
+	    ;
+      }
+  if (c != 3)
+    abort ();
+  return 0;
+}
diff --git a/gcc/tree-ssa-loop-unswitch.c b/gcc/tree-ssa-loop-unswitch.c
index 4328d6a..b69b995 100644
--- a/gcc/tree-ssa-loop-unswitch.c
+++ b/gcc/tree-ssa-loop-unswitch.c
@@ -466,7 +466,6 @@ 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;
@@ -524,6 +523,16 @@ find_loop_guard (struct loop *loop)
   else
     return NULL;
 
+  /* Guard edge must skip inner loop.  */
+  if (!dominated_by_p (CDI_DOMINATORS, loop->inner->header,
+      guard_edge == fe ? te->dest : fe->dest))
+    {
+      if (dump_file && (dump_flags & TDF_DETAILS))
+	fprintf (dump_file, "Guard edge %d --> %d is not around the loop!\n",
+		 guard_edge->src->index, guard_edge->dest->index);
+      return NULL;
+    }
+
   if (dump_file && (dump_flags & TDF_DETAILS))
     fprintf (dump_file,
 	     "Considering guard %d -> %d in loop %d\n",

Re: [PATCH PR67909 PR67947]

[Richard Biener]

On Tue, Oct 13, 2015 at 2:49 PM, Yuri Rumyantsev <ysrumyan@gmail.com> wrote:
> Here is updated patch with splitting long line.
> The patch is attached.

Ok with aligning the guard_edge == ... line properly

Thanks,
Richard.

> Yuri.
>
> 2015-10-13 15:38 GMT+03:00 H.J. Lu <hjl.tools@gmail.com>:
>> On Tue, Oct 13, 2015 at 4:57 AM, Yuri Rumyantsev <ysrumyan@gmail.com> wrote:
>>> Hi All,
>>>
>>> Here is a simple patch for unswitching outer loop through guard-edge
>>> hoisting. The check that guard-edge is around the inner loop was
>>> missed.
>>>
>>> Bootstrapping and regression testing did not show new failures.
>>>
>>> Is it OK for trunk?
>>>
>>> ChangeLog:
>>> 2014-10-13  Yuri Rumyantsev  <ysrumyan@gmail.com>
>>>
>>> PR tree-optimization/67909, 67947
>>> * tree-ssa-loop-unswitch.c (find_loop_guard): Add check that GUARD_EDGE
>>> really skip the inner loop.
>>>
>>> gcc/testsuite/ChangeLog
>>> * gcc.dg/torture/pr67947.c: New test.
>>
>> +  /* Guard edge must skip inner loop.  */
>> +  if (!dominated_by_p (CDI_DOMINATORS, loop->inner->header,
>> +      guard_edge == fe ? te->dest : fe->dest))
>>         ^^^^  It should line up with "CDI_DOMINATORS".
>>
>> + fprintf (dump_file, "Guard edge %d --> %d is not around the
>> loop!\n",guard_edge->src->index,guard_edge->dest->index);
>>
>> Please break lone line.
>>
>> --
>> H.J.