File pr34641.patch of Package gcc41

Subject: GCC 4.1 Fix für GCC Bugzilla 34641

Hallo,

ich musste gerade feststellen, dass wir mit 4.1 beim Kernelbuild 
neuerdings in das gleiche Problem laufen.  Ich hatte dazu ulkigerweise 
schon einen 4.1 Patch auf der Platte rumliegen, also vermute ich, dass 
ich euch das schoneinmal angetragen hatte.  Ich konnte allerdings keinen 
passenden Bugzilla für Sles 10 finden.  Könntest du kurz checken, ob der 
für Sles 10 SP 3 eingeplant ist?

Dank

-Andreas-


Index: gcc/reload.c
===================================================================
--- gcc/reload.c.orig	2006-06-22 18:11:03.000000000 +0200
+++ gcc/reload.c	2009-11-20 13:41:41.000000000 +0100
@@ -943,29 +943,33 @@ push_reload (rtx in, rtx out, rtx *inloc
   if (outmode == VOIDmode && out != 0)
     outmode = GET_MODE (out);
 
-  /* If IN is a pseudo register everywhere-equivalent to a constant, and
-     it is not in a hard register, reload straight from the constant,
-     since we want to get rid of such pseudo registers.
-     Often this is done earlier, but not always in find_reloads_address.  */
+  /* If find_reloads and friends until now missed to replace a pseudo
+     with a constant of reg_equiv_constant something went wrong
+     beforehand.
+     Note that it can't simply be done here if we missed it earlier
+     since the constant might need to be pushed into the literal pool
+     and the resulting memref would probably need further
+     reloading.  */
   if (in != 0 && REG_P (in))
     {
       int regno = REGNO (in);
 
-      if (regno >= FIRST_PSEUDO_REGISTER && reg_renumber[regno] < 0
-	  && reg_equiv_constant[regno] != 0)
-	in = reg_equiv_constant[regno];
+      gcc_assert (regno < FIRST_PSEUDO_REGISTER
+		  || reg_renumber[regno] >= 0
+		  || reg_equiv_constant[regno] == NULL_RTX);
     }
 
-  /* Likewise for OUT.  Of course, OUT will never be equivalent to
-     an actual constant, but it might be equivalent to a memory location
-     (in the case of a parameter).  */
+  /* reg_equiv_constant only contains constants which are obviously
+     not appropriate as destination.  So if we would need to replace
+     the destination pseudo with a constant we are in real
+     trouble.  */
   if (out != 0 && REG_P (out))
     {
       int regno = REGNO (out);
 
-      if (regno >= FIRST_PSEUDO_REGISTER && reg_renumber[regno] < 0
-	  && reg_equiv_constant[regno] != 0)
-	out = reg_equiv_constant[regno];
+      gcc_assert (regno < FIRST_PSEUDO_REGISTER
+		  || reg_renumber[regno] >= 0
+		  || reg_equiv_constant[regno] == NULL_RTX);
     }
 
   /* If we have a read-write operand with an address side-effect,
@@ -4746,15 +4750,12 @@ find_reloads_address (enum machine_mode
     {
       regno = REGNO (ad);
 
-      /* If the register is equivalent to an invariant expression, substitute
-	 the invariant, and eliminate any eliminable register references.  */
-      tem = reg_equiv_constant[regno];
-      if (tem != 0
-	  && (tem = eliminate_regs (tem, mode, insn))
-	  && strict_memory_address_p (mode, tem))
+      if (reg_equiv_constant[regno] != 0)
 	{
-	  *loc = ad = tem;
-	  return 0;
+	  find_reloads_address_part (reg_equiv_constant[regno], loc,
+				     BASE_REG_CLASS,
+				     GET_MODE (ad), opnum, type, ind_levels);
+	  return 1;
 	}
 
       tem = reg_equiv_memory_loc[regno];