diff mbox series

middle-end: Fix dominator information with loop duplication PR114197

Message ID 20240301221437.1981406-1-ewlu@rivosinc.com
State New
Headers show
Series middle-end: Fix dominator information with loop duplication PR114197 | expand

Commit Message

Edwin Lu March 1, 2024, 10:14 p.m. UTC
When adding the new_preheader to the cfg, only the new_preheader's dominator
information is updated. If one of the new basic block's children was part
of the original cfg and adding new_preheader to the cfg introduces another path
to that child, the child's dominator information will not be updated. This may
cause verify_dominator's assertion to fail.

Force recalculating dominators for all duplicated basic blocks and their
successors when updating new_preheader's dominator information.

	PR 114197

gcc/ChangeLog:

	* tree-vect-loop-manip.cc (slpeel_tree_duplicate_loop_to_edge_cfg):
	Recalculate dominator info when adding new_preheader to cfg

gcc/testsuite/ChangeLog:

	* gcc.dg/vect/pr114197.c: New test.

Signed-off-by: Edwin Lu <ewlu@rivosinc.com>
---
 gcc/testsuite/gcc.dg/vect/pr114197.c | 18 ++++++++++++++++++
 gcc/tree-vect-loop-manip.cc          | 17 ++++++++++++++++-
 2 files changed, 34 insertions(+), 1 deletion(-)
 create mode 100644 gcc/testsuite/gcc.dg/vect/pr114197.c

Comments

Richard Biener March 4, 2024, 10:42 a.m. UTC | #1
On Fri, Mar 1, 2024 at 11:16 PM Edwin Lu <ewlu@rivosinc.com> wrote:
>
> When adding the new_preheader to the cfg, only the new_preheader's dominator
> information is updated. If one of the new basic block's children was part
> of the original cfg and adding new_preheader to the cfg introduces another path
> to that child, the child's dominator information will not be updated. This may
> cause verify_dominator's assertion to fail.
>
> Force recalculating dominators for all duplicated basic blocks and their
> successors when updating new_preheader's dominator information.

We're already doing this (which IMO is bad), via the
iterate_fix_dominators call.
You're adding another bunch of similar things and I think the use of
recompute_dominator isn't safe as it assumes all predecessors have valid
dominator info.

I'll have a look.

Richard.

>         PR 114197
>
> gcc/ChangeLog:
>
>         * tree-vect-loop-manip.cc (slpeel_tree_duplicate_loop_to_edge_cfg):
>         Recalculate dominator info when adding new_preheader to cfg
>
> gcc/testsuite/ChangeLog:
>
>         * gcc.dg/vect/pr114197.c: New test.
>
> Signed-off-by: Edwin Lu <ewlu@rivosinc.com>
> ---
>  gcc/testsuite/gcc.dg/vect/pr114197.c | 18 ++++++++++++++++++
>  gcc/tree-vect-loop-manip.cc          | 17 ++++++++++++++++-
>  2 files changed, 34 insertions(+), 1 deletion(-)
>  create mode 100644 gcc/testsuite/gcc.dg/vect/pr114197.c
>
> diff --git a/gcc/testsuite/gcc.dg/vect/pr114197.c b/gcc/testsuite/gcc.dg/vect/pr114197.c
> new file mode 100644
> index 00000000000..b1fb807729c
> --- /dev/null
> +++ b/gcc/testsuite/gcc.dg/vect/pr114197.c
> @@ -0,0 +1,18 @@
> +/* { dg-do compile } */
> +/* { dg-additional-options "-O3" } */
> +
> +
> +#pragma pack(push)
> +struct a {
> +  volatile signed b : 8;
> +};
> +#pragma pack(pop)
> +int c;
> +static struct a d = {5};
> +void e() {
> +f:
> +  for (c = 8; c < 55; ++c)
> +    if (!d.b)
> +      goto f;
> +}
> +
> diff --git a/gcc/tree-vect-loop-manip.cc b/gcc/tree-vect-loop-manip.cc
> index f72da915103..0f3a489e78c 100644
> --- a/gcc/tree-vect-loop-manip.cc
> +++ b/gcc/tree-vect-loop-manip.cc
> @@ -1840,7 +1840,22 @@ slpeel_tree_duplicate_loop_to_edge_cfg (class loop *loop, edge loop_exit,
>         }
>
>        if (was_imm_dom || duplicate_outer_loop)
> -       set_immediate_dominator (CDI_DOMINATORS, exit_dest, new_exit->src);
> +       {
> +         set_immediate_dominator (CDI_DOMINATORS, exit_dest, new_exit->src);
> +
> +         /* Update the dominator info for children of duplicated bbs.  */
> +         for (unsigned i = 0; i < scalar_loop->num_nodes; i++)
> +           {
> +             basic_block dom_bb = NULL;
> +             edge e;
> +             edge_iterator ei;
> +             FOR_EACH_EDGE (e, ei, new_bbs[i]->succs)
> +               {
> +                 dom_bb = recompute_dominator (CDI_DOMINATORS, e->dest);
> +                 set_immediate_dominator (CDI_DOMINATORS, e->dest, dom_bb);
> +               }
> +           }
> +       }
>
>        /* And remove the non-necessary forwarder again.  Keep the other
>           one so we have a proper pre-header for the loop at the exit edge.  */
> --
> 2.34.1
>
diff mbox series

Patch

diff --git a/gcc/testsuite/gcc.dg/vect/pr114197.c b/gcc/testsuite/gcc.dg/vect/pr114197.c
new file mode 100644
index 00000000000..b1fb807729c
--- /dev/null
+++ b/gcc/testsuite/gcc.dg/vect/pr114197.c
@@ -0,0 +1,18 @@ 
+/* { dg-do compile } */
+/* { dg-additional-options "-O3" } */
+
+
+#pragma pack(push)
+struct a {
+  volatile signed b : 8;
+};
+#pragma pack(pop)
+int c;
+static struct a d = {5};
+void e() {
+f:
+  for (c = 8; c < 55; ++c)
+    if (!d.b)
+      goto f;
+}
+
diff --git a/gcc/tree-vect-loop-manip.cc b/gcc/tree-vect-loop-manip.cc
index f72da915103..0f3a489e78c 100644
--- a/gcc/tree-vect-loop-manip.cc
+++ b/gcc/tree-vect-loop-manip.cc
@@ -1840,7 +1840,22 @@  slpeel_tree_duplicate_loop_to_edge_cfg (class loop *loop, edge loop_exit,
 	}
 
       if (was_imm_dom || duplicate_outer_loop)
-	set_immediate_dominator (CDI_DOMINATORS, exit_dest, new_exit->src);
+	{
+	  set_immediate_dominator (CDI_DOMINATORS, exit_dest, new_exit->src);
+
+	  /* Update the dominator info for children of duplicated bbs.  */
+	  for (unsigned i = 0; i < scalar_loop->num_nodes; i++)
+	    {
+	      basic_block dom_bb = NULL;
+	      edge e;
+	      edge_iterator ei;
+	      FOR_EACH_EDGE (e, ei, new_bbs[i]->succs)
+		{
+		  dom_bb = recompute_dominator (CDI_DOMINATORS, e->dest);
+		  set_immediate_dominator (CDI_DOMINATORS, e->dest, dom_bb);
+		}
+	    }
+	}
 
       /* And remove the non-necessary forwarder again.  Keep the other
          one so we have a proper pre-header for the loop at the exit edge.  */