[Bug middle-end/43631] var-tracking inserts notes with non-NULL BLOCK_FOR_INSN in between basic blocks

[steven at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=43631



--- Comment #14 from Steven Bosscher <steven at gcc dot gnu.org> 2012-12-07 
11:59:15 UTC ---

(In reply to comment #13)

> Please hold on with that, seems BB_END is BARRIER, which is wrong.

> Starting with distilling a testcase...



Perhaps Alex' is right that the real problem is in how BLOCK_FOR_INSN

is handled in add_insn_after and add_insn_before.  Jakub, does the

following make sense to you?



Index: emit-rtl.c

===================================================================

--- emit-rtl.c  (revision 194229)

+++ emit-rtl.c  (working copy)

@@ -3845,19 +3845,25 @@ add_insn_after (rtx insn, rtx after, bas

       gcc_assert (stack);

     }



-  if (!BARRIER_P (after)

-      && !BARRIER_P (insn)

-      && (bb = BLOCK_FOR_INSN (after)))

+  /* If the new insn is a barrier, bb should be NULL because a barrier

+     is never inside a basic block.  Also avoid emitting double barriers.  */

+  if (BARRIER_P (insn))

+    gcc_checking_assert (bb == NULL && !BARRIER_P (after));

+  /* Otherwise, try to deduce a basic block from AFTER.  */

+  else if (!bb

+          && !BARRIER_P (after)

+          && BLOCK_FOR_INSN (after) != NULL

+          && after != BB_END (BLOCK_FOR_INSN (after)))

+    bb = BLOCK_FOR_INSN (after);

+

+  if (bb)

     {

       set_block_for_insn (insn, bb);

       if (INSN_P (insn))

        df_insn_rescan (insn);

-      /* Should not happen as first in the BB is always

-        either NOTE or LABEL.  */

-      if (BB_END (bb) == after

-         /* Avoid clobbering of structure when creating new BB.  */

-         && !BARRIER_P (insn)

-         && !NOTE_INSN_BASIC_BLOCK_P (insn))

+      /* Move BB_END from AFTER to INSN unless INSN is a note (a note cannot

+        be the end a basic block).  */

+      if (BB_END (bb) == after && !NOTE_P (insn))

        BB_END (bb) = insn;

     }