GTK bug?

Hi,

I've problems with loading GTK. Every time I tried to
load GTK, (1) I got errors (at the end mail) when compiling
some classes. However, simple GTK programs works fine
for serveral minutes - (2) after that, everything freezes
and gst starts to consume 100% CPU. Do you have same
problems or it is just problem of my machine?

Jan

----
janfrog@bruxa:...OmniBrowser$ 5 > gst
"Global garbage collection... done"
GNU Smalltalk ready

st> PackageLoader fileInPackage:'GTK'
Loading package GTK
Recompiling classes...
Recompiling class: GTK.GtkRequisition class
Recompiling selector: #sizeof
Object: GtkRequisition error: cannot redefine CStruct/CUnion
Smalltalk.Error(Smalltalk.Exception)>>signal
Smalltalk.Error(Smalltalk.Exception)>>signal:
GtkRequisition class(Smalltalk.Object)>>error:
GtkRequisition class(Smalltalk.CCompound
class)>>declaration:inject:into:
GtkRequisition class(Smalltalk.CStruct class)>>declaration:
Smalltalk.CStruct class(Smalltalk.CCompound
class)>>subclass:declaration:classVariableNames:poolDictionaries:category:
Smalltalk.UndefinedObject>>executeStatements
Recompiling classes...
Object: GSList error: cannot redefine CStruct/CUnion
Smalltalk.Error(Smalltalk.Exception)>>signal
Smalltalk.Error(Smalltalk.Exception)>>signal:
GSList class(Smalltalk.Object)>>error:
GSList class(Smalltalk.CCompound class)>>declaration:inject:into:
GSList class(Smalltalk.CStruct class)>>declaration:
Smalltalk.CStruct class(Smalltalk.CCompound
class)>>subclass:declaration:classVariableNames:poolDictionaries:category:
Smalltalk.UndefinedObject>>executeStatements
Recompiling classes...
Recompiling class: GTK.GdkEventButton class
Recompiling selector: #sizeof
Object: GdkEventButton error: cannot redefine CStruct/CUnion
Smalltalk.Error(Smalltalk.Exception)>>signal
Smalltalk.Error(Smalltalk.Exception)>>signal:
GdkEventButton class(Smalltalk.Object)>>error:
GdkEventButton class(Smalltalk.CCompound
class)>>declaration:inject:into:
GdkEventButton class(Smalltalk.CStruct class)>>declaration:
Smalltalk.CStruct class(Smalltalk.CCompound
class)>>subclass:declaration:classVariableNames:poolDictionaries:category:
Smalltalk.UndefinedObject>>executeStatements
Recompiling classes...
Recompiling class: GTK.GdkEventMotion class
Recompiling selector: #sizeof
Object: GdkEventMotion error: cannot redefine CStruct/CUnion
Smalltalk.Error(Smalltalk.Exception)>>signal
Smalltalk.Error(Smalltalk.Exception)>>signal:
GdkEventMotion class(Smalltalk.Object)>>error:
GdkEventMotion class(Smalltalk.CCompound
class)>>declaration:inject:into:
GdkEventMotion class(Smalltalk.CStruct class)>>declaration:
Smalltalk.CStruct class(Smalltalk.CCompound
class)>>subclass:declaration:classVariableNames:poolDictionaries:category:
Smalltalk.UndefinedObject>>executeStatements
Recompiling classes...
Recompiling class: GTK.GdkEventConfigure class
Recompiling selector: #sizeof
Object: GdkEventConfigure error: cannot redefine CStruct/CUnion
Smalltalk.Error(Smalltalk.Exception)>>signal
Smalltalk.Error(Smalltalk.Exception)>>signal:
GdkEventConfigure class(Smalltalk.Object)>>error:
GdkEventConfigure class(Smalltalk.CCompound
class)>>declaration:inject:into:
GdkEventConfigure class(Smalltalk.CStruct class)>>declaration:
Smalltalk.CStruct class(Smalltalk.CCompound
class)>>subclass:declaration:classVariableNames:poolDictionaries:category:
Smalltalk.UndefinedObject>>executeStatements
"Global garbage collection... done, heap grown"
PackageLoader
st>





_______________________________________________
help-smalltalk mailing list
[hidden email]
http://lists.gnu.org/mailman/listinfo/help-smalltalk

> I've problems with loading GTK. Every time I tried to
> load GTK, (1) I got errors (at the end mail) when compiling
> some classes. However, simple GTK programs works fine
> for serveral minutes - (2) after that, everything freezes
> and gst starts to consume 100% CPU. Do you have same
> problems or it is just problem of my machine?

(1) should be fixed by this patch.  I'll try to reproduce (2) and come back.

You can also try "kill -USR1 <pid>" on the gst process to see *where* it
freezes.

I'm interested in hearing from people who have a x86-64 linux machine.
Does it run "make check" (possibly with a failure in
DhbNumericalMethodws) or does it halt before?  My machine halts before,
but the failure is so weird (a call to ftruncate mysteriously zeros some
memory, or something like that) that I wonder if it is a kernel bug of
some sort.

Paolo



_______________________________________________
help-smalltalk mailing list
[hidden email]
http://lists.gnu.org/mailman/listinfo/help-smalltalk

Paolo Bonzini wrote:
>
>> I've problems with loading GTK. Every time I tried to
>> load GTK, (1) I got errors (at the end mail) when compiling
>> some classes. However, simple GTK programs works fine
>> for serveral minutes - (2) after that, everything freezes
>> and gst starts to consume 100% CPU. Do you have same
>> problems or it is just problem of my machine?
>
> (1) should be fixed by this patch.

Which patch?  This one.

There are multiple problems:

1) after the recent round of CObject/CType changes, CObject has a
class-instance variable and so does CStruct (which has a total of two).
  When loading GTK, some classes are initially born as subclasses of
CObject and then changed to subclasses of CStruct.  The class object was
not mutated, so that accessing the class-instance variable added by
CStruct went out of bounds; actually it accessed the size field of the
immediately neighboring object, so it was not nil and gave an error.

Since I was at it, I modified CStruct to use lazy initialization, so
that creating a CStruct with an empty field list and then adding some
fields is not a problem.

2) there was a minor GC problem in the compiler which, luckily, caused
damage immediately and not some time later.  This means it was fast to
debug.  One argument in the C function declaration was sometimes stored
in the wrong place, and `nil' would remain in the "right place".  The
primitive would fail when asked for the address of function `nil'.

Paolo

2007-09-17  Paolo Bonzini  <[hidden email]>

        * kernel/CStruct.st: Use lazy initialization for declaration.  Allow
        replacing an empty declaration.
        * kernel/Metaclass.st: Mutate the class object if the list of class-
        instance variables changes.

        * libgst/comp.c: Fix GC problem with attributes.

M  kernel/CStruct.st
M  kernel/Metaclass.st
M  libgst/comp.c

* modified files

--- orig/kernel/CStruct.st
+++ mod/kernel/CStruct.st
@@ -122,6 +122,7 @@ subclass: structName declaration: array
 
 declaration
     "Return the description of the fields in the receiver class."
+    declaration isNil ifTrue: [ declaration := #() ].
     ^declaration
 !
 
@@ -135,7 +136,7 @@ declaration: array inject: startOffset i
      old offset plus the new field's size is passed to aBlock,
      together with the new field's alignment requirements."
     | offset maxAlignment inspStr |
-    (declaration notNil and: [ declaration ~= array ])
+    (self declaration notEmpty and: [ self declaration ~= array ])
  ifTrue: [ self error: 'cannot redefine CStruct/CUnion' ].
 
     declaration := array.


--- orig/kernel/Metaclass.st
+++ mod/kernel/Metaclass.st
@@ -304,24 +304,32 @@ name: className
     shape: realShape
     ].
 
+    "Now add/remove pool dictionaries.  FIXME: They may affect name binding,
+     so we should probably recompile everything if they change."
     aClass sharedPoolDictionaries isNil
- ifTrue: [ aClass setSharedPools: sharedPoolNames ]
+ ifTrue: [
+    aClass setSharedPools: sharedPoolNames ]
  ifFalse: [
     sharedPoolNames do: [ :dict |
- (aClass sharedPoolDictionaries includes: dict)
-    ifFalse: [ aClass addSharedPool: dict ]
-    ].
+ (aClass sharedPoolDictionaries includes: dict) ifFalse: [
+    aClass addSharedPool: dict ] ].
 
     aClass sharedPoolDictionaries copy do: [ :dict |
- (sharedPoolNames includes: dict)
-    ifFalse: [
+ (sharedPoolNames includes: dict) ifFalse: [
  aClass removeSharedPool: dict.
- needToRecompileMetaclasses := true
-    ]
-    ]
+ needToRecompileMetaclasses := true ] ]
  ].
 
     (aClass superclass ~~ superclass) ifTrue: [
+ "Mutate the class if the set of class-instance variables changes."
+ (self superclass allInstVarNames ~= superclass class allInstVarNames)
+    ifTrue: [
+ aClass class
+    updateInstanceVars:
+ superclass class allInstVarNames,
+ aClass class instVarNames
+    shape: aClass class shape ].
+
  "Fix references between classes..."
  aClass superclass removeSubclass: aClass.
  superclass addSubclass: aClass.
@@ -333,6 +341,7 @@ name: className
  superclass class addSubclass: self.
  self superclass: superclass class.
  needToRecompileMetaclasses := true.
+
     ].
 
     aClass category: categoryName.
@@ -477,7 +486,6 @@ growClassInstance
     | newClass numInstVars |
     newClass := self new.
     numInstVars := self instSize.
-    numInstVars printNl.
     1 to: numInstVars - 1 do:
  [ :i | newClass instVarAt: i put:
    (instanceClass instVarAt: i) ].


--- orig/libgst/comp.c
+++ mod/libgst/comp.c
@@ -2372,20 +2372,21 @@ _gst_make_attribute (tree_node attribute
        i++, keyword = keyword->v_list.next)
     {
       tree_node value = keyword->v_list.value;
+      OOP result;
       if (value->nodeType != TREE_CONST_EXPR)
  {
-          OOP result = _gst_execute_statements (NULL, value, UNDECLARED_NONE,
-        true);
-          value = _gst_make_oop_constant (&value->location, result);
+          result = _gst_execute_statements (NULL, value, UNDECLARED_NONE, true);
           if (!result)
     {
       _gst_had_error = true;
       return _gst_nil_oop;
     }
  }
+      else
+ result = _gst_make_constant_oop (value);
 
       argsArray = OOP_TO_OBJ (argsArrayOOP);
-      argsArray->data[i] = _gst_make_constant_oop (value);
+      argsArray->data[i] = result;
     }
 
   messageOOP = _gst_message_new_args (selectorOOP, argsArrayOOP);




_______________________________________________
help-smalltalk mailing list
[hidden email]
http://lists.gnu.org/mailman/listinfo/help-smalltalk

> I'm interested in hearing from people who have a x86-64 linux machine.
> Does it run "make check" (possibly with a failure in
> DhbNumericalMethodws) or does it halt before?  My machine halts before,
> but the failure is so weird (a call to ftruncate mysteriously zeros some
> memory, or something like that) that I wonder if it is a kernel bug of
> some sort.

For the record, it must be.  It does not reproduce on Stefan Schmiedl's
machine (thanks!).

Paolo


_______________________________________________
help-smalltalk mailing list
[hidden email]
http://lists.gnu.org/mailman/listinfo/help-smalltalk