review (S) for 7032162: assert(flat != TypePtr::BOTTOM) failed: cannot alias-analyze an untyped ptr

[Tom Rodriguez]

On Apr 6, 2011, at 1:49 PM, Vladimir Kozlov wrote:

> You replaced 'require_constant' with 'true' to the call make_from_constant() so I was wondering if it may change behavior. Specifically it may produce NULL now
> if oop_constant->can_be_constant() is 'false':
> 
>      if (require_constant) {
>        if (!o->can_be_constant())  return NULL;

In the !JavaObjectInPerm world Strings should always return true for can_be_constant.  Actually I think it would be better to fix it by changing ciObject::should_be_constant instead of just patching that use. I also included Classes in the test for consistency with the old behaviour.  So I've removed the changes in parse3.cpp and done this instead.

diff -r 8e77e1f26188 src/share/vm/ci/ciObject.cpp                                                                                                     
--- a/src/share/vm/ci/ciObject.cpp                                                                                                                   
+++ b/src/share/vm/ci/ciObject.cpp                                                                                                                   
@@ -194,6 +194,16 @@ bool ciObject::can_be_constant() {
 // ciObject::should_be_constant()                                                                                                                   
 bool ciObject::should_be_constant() {                                                                                                               
   if (ScavengeRootsInCode >= 2)  return true;  // force everybody to be a constant                                                                   
+  if (!JavaObjectsInPerm) {                                                                                                                         
+    // We want Strings and Classes to be embeddable by default since                                                                                 
+    // they used to be in the perm world.  Not all Strings used to be                                                                               
+    // embeddable but there's no easy way to distinguish the interned                                                                               
+    // from the regulars ones so just treat them all that way.                                                                                       
+    ciEnv* env = CURRENT_ENV;                                                                                                                       
+    if (klass() == env->String_klass() || klass() == env->Class_klass()) {                                                                           
+      return true;                                                                                                                                   
+    }                                                                                                                                               
+  }                                                                                                                                                 
   return handle() == NULL || !is_scavengable();                                                                                                     
 }

tom

> 
> Vladimir
> 
> Tom Rodriguez wrote:
>> http://cr.openjdk.java.net/~never/7032162
>> 7032162: assert(flat != TypePtr::BOTTOM) failed: cannot alias-analyze an untyped ptr
>> Reviewed-by:
>> This appears to be an existing problem with OptimizeStringConcat where
>> missing control can allow some loads to be split up to where their
>> base is NULL which causes addresses that can't be analyzed.  The fix
>> is to use the appropriate control.  If it's truly unneeded it will be
>> optimized away.  The bug was exposed by the String not in perm changes
>> because we were no longer treating some constant strings as embeddable
>> constants.  This is fixed by handling them specially in push_constant.
>> Tested with failing test from report and full CTW with string opts.