Le 05/03/2016 09:14, clipka a écrit :
> Am 04.03.2016 um 19:51 schrieb Le_Forgeron:
>> Le 04/03/2016 19:41, William F Pokorny a écrit :
> 
>> Nope, even with a produced "right" picture, valgrind reports a definitive problem
of uninitialised data.
> 
> If valgrind was right, that would certainly explain the problems.
> Unfortunately it's mistaken.
> 
> Valgrind's report indicates that Bezier_Node_Struct::Radius_Squared is
> not (or not always) initialized properly when
> BicubicPatch::bezier_tree_walker() is called.
> 
> But if that is the case, then adding
> 
>     Node->Radius_Squared = -1;
> 
> to BicubicPatch::create_new_bezier_node() and adding
> 
>     POV_ASSERT(Node->Radius_Squared >= 0);
> 
> to the beginning of BicubicPatch::bezier_tree_walker() should force the
> code to bomb.
> 
> It doesn't. Valgrind must be seeing ghosts.
> 

I do not say that the Radius_Squared is not initialized...

but adding

memset(Node, 0, sizeof(BEZIER_NODE));

in create_new_bezier_node does remove the issue with valgrind.

Now, remember that we are using -O3 as default level of optimisation,
which means the execution can be reported on line X while the compiler
is actually performing some operations of line X-1 or X+1.

And we are dealing with an hybride code: the declaration of Bezier_Node_Struct is a
100% C++ thing,
with an implicit constructor that would call Véctor3d constructor for Center,
yet create_new_bezier_node is a C thing, calling a memory allocator and casting the
type.

The other members of Bezier_Node_Struct are still C thing, so they are not
initialised.

Nevertheless, for Center, the code that computes it has been "optimised".
Instead of using a temporary variable, initialised to 0 (xc, yc, zc), to add the
various contributions
and then copying the result in Center at the end, the function now add directly the
various contributions
to Center. But what if... you start with *garbage* in Center ? Garbage in, Garbage
out.
(it's in bezier_bounding_sphere() )

And to finish it (give me a hammer, that's the last nail on the coffin):
 If I replace the memset over the whole structure with just (in addition to Data_Ptr =
NULL):

 Node->Center = Vector3d();

valgrind become as happy as with the full initialisation.


Even a bigger hammer: if I use memset with a value of 0xFF (instead of 0), I get a
100% background image.
Reliably.

Post a reply to this message