On Wed, Jan 08, 2014 at 05:54:55PM +0100, Uros Bizjak wrote:
> This caused PR59724 on alpha:
> 
> 20021116-1.c: In function ‘foo’:
> 20021116-1.c:31:1: error: NOTE_INSN_BASIC_BLOCK is missing for block 9
>  }
>  ^
> 20021116-1.c:31:1: error: insn outside basic block
> (jump_insn 94 52 93 9 (return) 20021116-1.c:31 -1
>      (nil)
>  -> return)

Ugh, indeed.  The problem is that try_head_merge_bb really wants
flow_find_head_matching_sequence to count all (non-note) insns, not
just active insns, because otherwise as in the above testcase we
can have e.g. 2 active insns followed by one non-active, all matching
(flow_find_head_matching_sequence returns 2) and on another edge
just 2 active insns and nothing else matching.  2 == 2, so the caller
thinks it doesn't matter which one is shorter, but we have the insn range
of 3 insns together.

So, this patch just reverts the try_head_merge_bb changes and makes
flow_find_head_matching_sequence behave the old way when called from
try_head_merge_bb, i.e. count all non-note insns, and only when called
from ifcvt.c count just active insns.  Plus the ifcvt.c change ensures
we don't mistakenly call it with stop_after == 0 (which wouldn't actually
stop).

Bootstrapped/regtested on x86_64-linux and i686-linux, Uros is testing it
on Alpha.  Ok for trunk?

2014-01-08  Jakub Jelinek  <ja...@redhat.com>

        PR rtl-optimization/59724
        * ifcvt.c (cond_exec_process_if_block): Don't call
        flow_find_head_matching_sequence with 0 longest_match.
        * cfgcleanup.c (flow_find_head_matching_sequence): Count even
        non-active insns if !stop_after.
        (try_head_merge_bb): Revert 2014-01-07 changes.

--- gcc/ifcvt.c.jj      2014-01-08 10:23:20.000000000 +0100
+++ gcc/ifcvt.c 2014-01-08 18:46:17.017715169 +0100
@@ -522,7 +522,10 @@ cond_exec_process_if_block (ce_if_block
          n_insns -= 2 * n_matching;
        }
 
-      if (then_start && else_start)
+      if (then_start
+         && else_start
+         && then_n_insns > n_matching
+         && else_n_insns > n_matching)
        {
          int longest_match = MIN (then_n_insns - n_matching,
                                   else_n_insns - n_matching);
--- gcc/cfgcleanup.c.jj 2014-01-07 08:54:05.772736321 +0100
+++ gcc/cfgcleanup.c    2014-01-08 18:41:14.433307914 +0100
@@ -1421,7 +1421,8 @@ flow_find_cross_jump (basic_block bb1, b
 /* Like flow_find_cross_jump, except start looking for a matching sequence from
    the head of the two blocks.  Do not include jumps at the end.
    If STOP_AFTER is nonzero, stop after finding that many matching
-   instructions.  */
+   instructions.  If STOP_AFTER is zero, count all INSN_P insns, if it is
+   non-zero, only count active insns.  */
 
 int
 flow_find_head_matching_sequence (basic_block bb1, basic_block bb2, rtx *f1,
@@ -1493,7 +1494,7 @@ flow_find_head_matching_sequence (basic_
 
          beforelast1 = last1, beforelast2 = last2;
          last1 = i1, last2 = i2;
-         if (active_insn_p (i1))
+         if (!stop_after || active_insn_p (i1))
            ninsns++;
        }
 
@@ -2408,7 +2409,9 @@ try_head_merge_bb (basic_block bb)
       max_match--;
       if (max_match == 0)
        return false;
-      e0_last_head = prev_active_insn (e0_last_head);
+      do
+       e0_last_head = prev_real_insn (e0_last_head);
+      while (DEBUG_INSN_P (e0_last_head));
     }
 
   if (max_match == 0)
@@ -2428,14 +2431,16 @@ try_head_merge_bb (basic_block bb)
       basic_block merge_bb = EDGE_SUCC (bb, ix)->dest;
       rtx head = BB_HEAD (merge_bb);
 
-      if (!active_insn_p (head))
-       head = next_active_insn (head);
+      while (!NONDEBUG_INSN_P (head))
+       head = NEXT_INSN (head);
       headptr[ix] = head;
       currptr[ix] = head;
 
       /* Compute the end point and live information  */
       for (j = 1; j < max_match; j++)
-       head = next_active_insn (head);
+       do
+         head = NEXT_INSN (head);
+       while (!NONDEBUG_INSN_P (head));
       simulate_backwards_to_point (merge_bb, live, head);
       IOR_REG_SET (live_union, live);
     }


        Jakub

Reply via email to