[RFA/RFC, IRA] Fix PR bootstrap/37859

[Adam Nemet <anemet at caviumnetworks dot com>]

This is a bootstrap failure on mips64octeon-linux-gnu.  try_forward_edges gets
miscompiled by IRA which causes an ICE in stage3.

This is the simplified CFG (I also have the dot file for the complete CFG with
the relevant blocks annotated if someone is interested):

        +-----> 110
       	|      	 |
 	|  a61   .
        |  r696  .
        |  h-1   .
	|        |
	|       15 <----------+
	|        |            |
	|        .  a256      |
	|        .  r719   r719 = xmalloc (..)
        |        .  h30       |
	|        |            |
	|       82 -----------+
        |        |   +--------+
	|        |   | a231   |
	|        | <-+ r711   |
	|       87 --+ h30    |
	|        |   |        |
	|      	 |   +--------+
	+------ 88

Arrows point downward unless marked otherwise.  For each loop I included an
allocno and the corresponding pseudo register and hard register.  These
allocnos all correspond to the same original pseudo (r284).

The bug occurs because we omit the move from r711 to r696 on edge 87->88 and
hence forget that we ever allocated the threaded_edges array.

The reason the move is omitted becase not_modified_p in ira-emit.c returns
true for a231 and a61.

  /* Return TRUE if move of SRC_ALLOCNO to DEST_ALLOCNO does not change
     value of the destination.  One possible reason for this is the
     situation when SRC_ALLOCNO is not modified in the corresponding
     loop.  */
  static bool
  not_modified_p (ira_allocno_t src_allocno, ira_allocno_t dest_allocno)
  {
    int regno, orig_regno;
    ira_allocno_t a;
    ira_loop_tree_node_t node;
  
    ira_assert (ALLOCNO_CAP_MEMBER (src_allocno) == NULL
  	      && ALLOCNO_CAP_MEMBER (dest_allocno) == NULL);
    orig_regno = ALLOCNO_REGNO (src_allocno);
    regno = REGNO (ALLOCNO_REG (dest_allocno));
    for (node = ALLOCNO_LOOP_TREE_NODE (src_allocno);
         node != NULL;
         node = node->parent)
      if ((a = node->regno_allocno_map[orig_regno]) == NULL)
        break;
      else if (REGNO (ALLOCNO_REG (a)) == (unsigned) regno)
        return true;
      else if (bitmap_bit_p (node->modified_regnos, orig_regno))
        return false;
    return node != NULL;
  }

In the first iteration since r711 is unchanged in the loop starting at bb87 we
go on.  In the second iteration we're looking at the same allocnos so we
return true.  This is wrong however because the value of the allocno could
have changed in this loop.

My solution is to check whether the value is unmodified before checking
equivalence of the allocnos.

Testing is still ongoing but boostrapping has already finished on
x86_64-linux-gnu and has passed stage2-3 comparison on mips64octeon-linux-gnu.

I will report back with the result of the regtesting but in the meantime I'd
like to get some feedback if my thinking is correct.

Adam


	PR bootstrap/37859
	* ira-emit.c (not_modified_p): Only return true if register is not
	modified in the loop.

Index: ira-emit.c
===================================================================
--- ira-emit.c	(revision 141180)
+++ ira-emit.c	(working copy)
@@ -280,10 +280,10 @@ not_modified_p (ira_allocno_t src_allocn
        node = node->parent)
     if ((a = node->regno_allocno_map[orig_regno]) == NULL)
       break;
-    else if (REGNO (ALLOCNO_REG (a)) == (unsigned) regno)
-      return true;
     else if (bitmap_bit_p (node->modified_regnos, orig_regno))
       return false;
+    else if (REGNO (ALLOCNO_REG (a)) == (unsigned) regno)
+      return true;
   return node != NULL;
 }