[patch] PR 48837

[patch] PR 48837

[Zdenek Dvorak]

Hi,

when accumulator transformation is performed on a function like

foo(a)
{
  if (a > 0)
    return 1 + foo (a - 1)

  return bla();
}

this becomes

foo(a)
{
  int tmp = 0;

  while (a > 0)
    tm = 1 + tmp;

  return tmp + bla();
}

Before, bla was a tail-call, but after the optimization, it is not (since an addition
has to be performed after the result of bla is known).  However, we used to mark bla
as tail-call, leading to a misscompilation later.  Fixed by not marking tail-calls
when the transformation is performed.  Bootstrapped and regtested on i686.

Zdenek

	PR tree-optimization/48837
	* tree-tailcall.c (tree_optimize_tail_calls_1): Do not mark tailcalls
	when accumulator transformation is performed.

	* gcc.dg/pr48837.c: New testcase.

Index: tree-tailcall.c
===================================================================
--- tree-tailcall.c	(revision 173354)
+++ tree-tailcall.c	(working copy)
@@ -1021,6 +1021,14 @@ tree_optimize_tail_calls_1 (bool opt_tailcalls)
 					     integer_one_node);
     }
 
+  if (a_acc || m_acc)
+    {
+      /* When the tail call elimination using accumulators is performed,
+	 statements adding the accumulated value are inserted at all exits.
+	 This turns all other tail calls to non-tail ones.  */
+      opt_tailcalls = false;
+    }
+
   for (; tailcalls; tailcalls = next)
     {
       next = tailcalls->next;
Index: testsuite/gcc.dg/pr48837.c
===================================================================
--- testsuite/gcc.dg/pr48837.c	(revision 0)
+++ testsuite/gcc.dg/pr48837.c	(revision 0)
@@ -0,0 +1,30 @@
+/* PR tree-optimization/48837 */
+/* { dg-do run } */
+/* { dg-options "-O2" } */
+
+void abort (void);
+
+__attribute__((noinline))
+int baz(void)
+{
+  return 1;
+}
+
+inline const int *bar(const int *a, const int *b)
+{
+ return *a ? a : b;
+}
+
+int foo(int a, int b)
+{
+   return a || b ? baz() : foo(*bar(&a, &b), 1) + foo(1, 0);
+}
+
+int main(void)
+{
+ if (foo(0, 0) != 2)
+   abort();
+
+ return 0;
+}
+

Re: [patch] PR 48837

[Richard Guenther]

On Fri, May 6, 2011 at 12:44 PM, Zdenek Dvorak <rakdver@kam.mff.cuni.cz> wrote:
> Hi,
>
> when accumulator transformation is performed on a function like
>
> foo(a)
> {
>  if (a > 0)
>    return 1 + foo (a - 1)
>
>  return bla();
> }
>
> this becomes
>
> foo(a)
> {
>  int tmp = 0;
>
>  while (a > 0)
>    tm = 1 + tmp;
>
>  return tmp + bla();
> }
>
> Before, bla was a tail-call, but after the optimization, it is not (since an addition
> has to be performed after the result of bla is known).  However, we used to mark bla
> as tail-call, leading to a misscompilation later.  Fixed by not marking tail-calls
> when the transformation is performed.  Bootstrapped and regtested on i686.

Ok.

Thanks,
Richard.

> Zdenek
>
>        PR tree-optimization/48837
>        * tree-tailcall.c (tree_optimize_tail_calls_1): Do not mark tailcalls
>        when accumulator transformation is performed.
>
>        * gcc.dg/pr48837.c: New testcase.
>
> Index: tree-tailcall.c
> ===================================================================
> --- tree-tailcall.c     (revision 173354)
> +++ tree-tailcall.c     (working copy)
> @@ -1021,6 +1021,14 @@ tree_optimize_tail_calls_1 (bool opt_tailcalls)
>                                             integer_one_node);
>     }
>
> +  if (a_acc || m_acc)
> +    {
> +      /* When the tail call elimination using accumulators is performed,
> +        statements adding the accumulated value are inserted at all exits.
> +        This turns all other tail calls to non-tail ones.  */
> +      opt_tailcalls = false;
> +    }
> +
>   for (; tailcalls; tailcalls = next)
>     {
>       next = tailcalls->next;
> Index: testsuite/gcc.dg/pr48837.c
> ===================================================================
> --- testsuite/gcc.dg/pr48837.c  (revision 0)
> +++ testsuite/gcc.dg/pr48837.c  (revision 0)
> @@ -0,0 +1,30 @@
> +/* PR tree-optimization/48837 */
> +/* { dg-do run } */
> +/* { dg-options "-O2" } */
> +
> +void abort (void);
> +
> +__attribute__((noinline))
> +int baz(void)
> +{
> +  return 1;
> +}
> +
> +inline const int *bar(const int *a, const int *b)
> +{
> + return *a ? a : b;
> +}
> +
> +int foo(int a, int b)
> +{
> +   return a || b ? baz() : foo(*bar(&a, &b), 1) + foo(1, 0);
> +}
> +
> +int main(void)
> +{
> + if (foo(0, 0) != 2)
> +   abort();
> +
> + return 0;
> +}
> +
>

Re: [patch] PR 48837

[Jeff Law]

-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA1

On 05/06/11 04:44, Zdenek Dvorak wrote:
> Hi,
> 
> when accumulator transformation is performed on a function like
> 
> foo(a)
> {
>   if (a > 0)
>     return 1 + foo (a - 1)
> 
>   return bla();
> }
> 
> this becomes
> 
> foo(a)
> {
>   int tmp = 0;
> 
>   while (a > 0)
>     tm = 1 + tmp;
> 
>   return tmp + bla();
> }
> 
> Before, bla was a tail-call, but after the optimization, it is not (since an addition
> has to be performed after the result of bla is known).  However, we used to mark bla
> as tail-call, leading to a misscompilation later.  Fixed by not marking tail-calls
> when the transformation is performed.  Bootstrapped and regtested on i686.
> 
> Zdenek
> 
> 	PR tree-optimization/48837
> 	* tree-tailcall.c (tree_optimize_tail_calls_1): Do not mark tailcalls
> 	when accumulator transformation is performed.
> 
> 	* gcc.dg/pr48837.c: New testcase.
OK.

Thanks,
jeff
-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1.4.11 (GNU/Linux)
Comment: Using GnuPG with Fedora - http://enigmail.mozdev.org/

iQEcBAEBAgAGBQJNxA66AAoJEBRtltQi2kC7yysIAKZYUpU9JlyH2XwvvVslq8C1
CJ7E/akRDsBoYBS+syNsLMwkbGG0WoaFJzOd7vUmIknAHEusF6OasczsN8PD9aEB
i8xJNTZm2yxhrVZh8m/KBX96r80RwzpAhr9L1WAspiS/xpw12lRoJoh3XeKXYXWw
Z0aBL4ljCgLj6GKEyy7FbGHx0gEqaa1x7EDM1kJGCgZPAFJalJPozBiiriYL9/Th
gHqLXZ0HUhXNGql5M2S+lfZG2d30Rj1KBXDrU8EOXedHRjxb+U9+WLGsUHZtkcTI
j3//n6bjTr/YmyTe43voG3Rn6z6k0g2Eb8gF8UMvDbaSJlH9+xb6SuWLS8+mEhY=
=luE9
-----END PGP SIGNATURE-----