[PATCH] Fix PR78413

[PATCH] Fix PR78413

[Bill Schmidt]

Hi,

The if-conversion patch for PR77848 missed a case where an outer loop
should not be versioned for vectorization; this was caught by an assert
in tests recorded in https://gcc.gnu.org/bugzilla/show_bug.cgi?id=78413.
This patch fixes the problem by ensuring that both inner and outer loop
latches have a single predecessor before versioning an outer loop.

Bootstrapped and tested on powerpc64le-unknown-linux-gnu with no
regressions.  Is this ok for trunk?

Thanks,
Bill


[gcc]

2016-11-18  Bill Schmidt  <wschmidt@linux.vnet.ibm.com>

	PR tree-optimization/78413
	* tree-if-conv.c (versionable_outer_loop_p): Require that both
	inner and outer loop latches have single predecessors.

[gcc/testsuite]

2016-11-18  Bill Schmidt  <wschmidt@linux.vnet.ibm.com>

	PR tree-optimization/78413
	* gcc.dg/tree-ssa/pr78413.c: New test.


Index: gcc/testsuite/gcc.dg/tree-ssa/pr78413.c
===================================================================
--- gcc/testsuite/gcc.dg/tree-ssa/pr78413.c	(revision 0)
+++ gcc/testsuite/gcc.dg/tree-ssa/pr78413.c	(working copy)
@@ -0,0 +1,35 @@
+/* PR78413.  These previously failed in tree if-conversion due to a loop
+   latch with multiple predecessors that the code did not anticipate.  */
+/* { dg-do compile } */
+/* { dg-options "-O3 -ffast-math" } */
+
+extern long long int llrint(double x);
+int a;
+double b;
+__attribute__((cold)) void decode_init() {
+  int c, d = 0;
+  for (; d < 12; d++) {
+    if (d)
+      b = 0;
+    c = 0;
+    for (; c < 6; c++)
+      a = b ? llrint(b) : 0;
+  }
+}
+
+struct S {
+  _Bool bo;
+};
+int a, bb, c, d;
+void fn1() {
+  do
+    do
+      do {
+	struct S *e = (struct S *)1;
+	do
+	  bb = a / (e->bo ? 2 : 1);
+	while (bb);
+      } while (0);
+    while (d);
+  while (c);
+}
Index: gcc/tree-if-conv.c
===================================================================
--- gcc/tree-if-conv.c	(revision 242551)
+++ gcc/tree-if-conv.c	(working copy)
@@ -2575,6 +2575,8 @@ version_loop_for_if_conversion (struct loop *loop)
     - The loop has a single exit.
     - The loop header has a single successor, which is the inner
       loop header.
+    - Each of the inner and outer loop latches have a single
+      predecessor.
     - The loop exit block has a single predecessor, which is the
       inner loop's exit block.  */
 
@@ -2586,7 +2588,9 @@ versionable_outer_loop_p (struct loop *loop)
       || loop->inner->next
       || !single_exit (loop)
       || !single_succ_p (loop->header)
-      || single_succ (loop->header) != loop->inner->header)
+      || single_succ (loop->header) != loop->inner->header
+      || !single_pred_p (loop->latch)
+      || !single_pred_p (loop->inner->latch))
     return false;
   
   basic_block outer_exit = single_pred (loop->latch);

Re: [PATCH] Fix PR78413

[Markus Trippelsdorf]

On 2016.11.18 at 10:27 -0600, Bill Schmidt wrote:
> ===================================================================
> --- gcc/testsuite/gcc.dg/tree-ssa/pr78413.c	(revision 0)
> +++ gcc/testsuite/gcc.dg/tree-ssa/pr78413.c	(working copy)
> @@ -0,0 +1,35 @@
> +/* PR78413.  These previously failed in tree if-conversion due to a loop
> +   latch with multiple predecessors that the code did not anticipate.  */
> +/* { dg-do compile } */
> +/* { dg-options "-O3 -ffast-math" } */

Please add -fno-strict-aliasing, otherwise fn1() wouldn't ICE.

> +extern long long int llrint(double x);
> +int a;
> +double b;
> +__attribute__((cold)) void decode_init() {
> +  int c, d = 0;
> +  for (; d < 12; d++) {
> +    if (d)
> +      b = 0;
> +    c = 0;
> +    for (; c < 6; c++)
> +      a = b ? llrint(b) : 0;
> +  }
> +}
> +
> +struct S {
> +  _Bool bo;
> +};
> +int a, bb, c, d;
> +void fn1() {
> +  do
> +    do
> +      do {
> +	struct S *e = (struct S *)1;
> +	do
> +	  bb = a / (e->bo ? 2 : 1);
> +	while (bb);
> +      } while (0);
> +    while (d);
> +  while (c);
> +}

-- 
Markus

Re: [PATCH] Fix PR78413

[Bill Schmidt]

> On Nov 18, 2016, at 10:33 AM, Markus Trippelsdorf <markus@trippelsdorf.de> wrote:
> 
> On 2016.11.18 at 10:27 -0600, Bill Schmidt wrote:
>> ===================================================================
>> --- gcc/testsuite/gcc.dg/tree-ssa/pr78413.c	(revision 0)
>> +++ gcc/testsuite/gcc.dg/tree-ssa/pr78413.c	(working copy)
>> @@ -0,0 +1,35 @@
>> +/* PR78413.  These previously failed in tree if-conversion due to a loop
>> +   latch with multiple predecessors that the code did not anticipate.  */
>> +/* { dg-do compile } */
>> +/* { dg-options "-O3 -ffast-math" } */
> 
> Please add -fno-strict-aliasing, otherwise fn1() wouldn't ICE.

Whoops.  Yes indeed, sorry I missed the flag difference for the second failure.

Bill

Re: [PATCH] Fix PR78413

[Richard Biener]

On Fri, Nov 18, 2016 at 5:27 PM, Bill Schmidt
<wschmidt@linux.vnet.ibm.com> wrote:
> Hi,
>
> The if-conversion patch for PR77848 missed a case where an outer loop
> should not be versioned for vectorization; this was caught by an assert
> in tests recorded in https://gcc.gnu.org/bugzilla/show_bug.cgi?id=78413.
> This patch fixes the problem by ensuring that both inner and outer loop
> latches have a single predecessor before versioning an outer loop.
>
> Bootstrapped and tested on powerpc64le-unknown-linux-gnu with no
> regressions.  Is this ok for trunk?

Ok (with the suggested testcase adjustment).

Maybe as a followup we can factor out a common predicate usable by the
vectorizer and if-conversion.  I'll see if I can do that.

Thanks,
Richard.

> Thanks,
> Bill
>
>
> [gcc]
>
> 2016-11-18  Bill Schmidt  <wschmidt@linux.vnet.ibm.com>
>
>         PR tree-optimization/78413
>         * tree-if-conv.c (versionable_outer_loop_p): Require that both
>         inner and outer loop latches have single predecessors.
>
> [gcc/testsuite]
>
> 2016-11-18  Bill Schmidt  <wschmidt@linux.vnet.ibm.com>
>
>         PR tree-optimization/78413
>         * gcc.dg/tree-ssa/pr78413.c: New test.
>
>
> Index: gcc/testsuite/gcc.dg/tree-ssa/pr78413.c
> ===================================================================
> --- gcc/testsuite/gcc.dg/tree-ssa/pr78413.c     (revision 0)
> +++ gcc/testsuite/gcc.dg/tree-ssa/pr78413.c     (working copy)
> @@ -0,0 +1,35 @@
> +/* PR78413.  These previously failed in tree if-conversion due to a loop
> +   latch with multiple predecessors that the code did not anticipate.  */
> +/* { dg-do compile } */
> +/* { dg-options "-O3 -ffast-math" } */
> +
> +extern long long int llrint(double x);
> +int a;
> +double b;
> +__attribute__((cold)) void decode_init() {
> +  int c, d = 0;
> +  for (; d < 12; d++) {
> +    if (d)
> +      b = 0;
> +    c = 0;
> +    for (; c < 6; c++)
> +      a = b ? llrint(b) : 0;
> +  }
> +}
> +
> +struct S {
> +  _Bool bo;
> +};
> +int a, bb, c, d;
> +void fn1() {
> +  do
> +    do
> +      do {
> +       struct S *e = (struct S *)1;
> +       do
> +         bb = a / (e->bo ? 2 : 1);
> +       while (bb);
> +      } while (0);
> +    while (d);
> +  while (c);
> +}
> Index: gcc/tree-if-conv.c
> ===================================================================
> --- gcc/tree-if-conv.c  (revision 242551)
> +++ gcc/tree-if-conv.c  (working copy)
> @@ -2575,6 +2575,8 @@ version_loop_for_if_conversion (struct loop *loop)
>      - The loop has a single exit.
>      - The loop header has a single successor, which is the inner
>        loop header.
> +    - Each of the inner and outer loop latches have a single
> +      predecessor.
>      - The loop exit block has a single predecessor, which is the
>        inner loop's exit block.  */
>
> @@ -2586,7 +2588,9 @@ versionable_outer_loop_p (struct loop *loop)
>        || loop->inner->next
>        || !single_exit (loop)
>        || !single_succ_p (loop->header)
> -      || single_succ (loop->header) != loop->inner->header)
> +      || single_succ (loop->header) != loop->inner->header
> +      || !single_pred_p (loop->latch)
> +      || !single_pred_p (loop->inner->latch))
>      return false;
>
>    basic_block outer_exit = single_pred (loop->latch);
>