JIT issues and todolist

Hi all,

I will not have time for another couple of weeks for the JIT but just
in case someone is interested to continue or has some insight here is
a list of known issues:


1.) The Delay test is failing as the spOffset/sp of a context is -1.
This means that in some jitted code there is an unbalanced push/pop
in the stack.

2.) The maybe releasing of native code is failing with snapshot/resume
of the image. Passing an '-i' will 'fix' the segfault. The interesting
bit is that this happens _after_ the fixup must have set context
native_ip to 0. So I wonder if this could actually be the fix:

diff --git a/libgst/interp-jit.inl b/libgst/interp-jit.inl
index 0e8fa27..18dc40e 100644
--- a/libgst/interp-jit.inl
+++ b/libgst/interp-jit.inl
@@ -353,6 +353,7 @@ refresh_native_ips (OOP contextOOP)
          virtualIP = TO_INT (context->ipOffset);
          native_ip =
            _gst_map_virtual_ip (context->method, receiverClass, virtualIP);
+          context->method->flags |= F_XLAT_REACHABLE;
 
 #ifndef OPTIMIZE
          if (!native_ip)


3.) The debug semaphore does not work, I think it should either set
the exception flag or the emit_interrupt_check should make a NULL test
for it. I am not sure how this is supposed to work yet.


4.) Porting to GNU lighting 1.9/2.0 would be nice. This way I could
have a JIT on ARM as well. I have had this wish/hope for quite some time
already but thanks to another Paulo GNU lightning can support that now.


cheers

        holger


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

On Sun, Jun 09, 2013 at 09:18:00PM +0200, Holger Hans Peter Freyther wrote:
The issue is not the spOffset -1 for the context part but actually the
inlined primitive for Object>>#basicAt:. I have #ifdef 0/#endif out the
code to force to go through the normal primitive code and the test is
passing now.

I fail to create such a BlockContext (I assume it is going through
the Process>>#queueInterrupt or signalInterrupt:). Paolo do you think
you could help me to build a shorter re-producer?

Would you mind if we disable the inlining with #ifdef 0 right now and
re-enable with some more time?



Even on the interpreter:
An instance of BlockContext
  parent: BlockClosure>>ensure: (BlkClosure.st:271)
  nativeIP: 0
  ip: 8
  sp: -1
  receiver: UndefinedObject
  method: [] in UndefinedObject>>executeStatements
  outerContext: nil
[] in BlockClosure>>asContext: (BlkClosure.st:181)
BlockContext class>>fromClosure:parent: (BlkContext.st:68)
optimized [] in UndefinedObject>>executeStatements (tests/delays.st:40)
BlockClosure>>ensure: (BlkClosure.st:271)
[] in Delay>>value:onTimeoutDo: (Delay.st:315)
BlockClosure>>on:do: (BlkClosure.st:196)
Delay>>value:onTimeoutDo: (Delay.st:316)
UndefinedObject>>executeStatements (tests/delays.st:40)


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

On Sun, Jun 16, 2013 at 08:27:25AM +0200, Holger Hans Peter Freyther wrote:
> On Sun, Jun 09, 2013 at 09:18:00PM +0200, Holger Hans Peter Freyther wrote:

> Even on the interpreter:
> An instance of BlockContext
>   parent: BlockClosure>>ensure: (BlkClosure.st:271)
>   nativeIP: 0
>   ip: 8
>   sp: -1
>   receiver: UndefinedObject
>   method: [] in UndefinedObject>>executeStatements
>   outerContext: nil

It is really a "wonder" that this is working at all...

VMpr_Object_basicAt
   int arg2 = -1;
   index_oop(oop1, arg2);


index_oop(oop, 4294967295)
   object = OOP_TO_OBJ(oop);
   intptr_t instanceSpec = GET_IN..
   return index_oop_spec(oop, object, index, instanceSpec);

index_oop_spec(oop, object, 4294967295, 57407)

   if (4294967295 < 1)
     return NULL

   index = 4294967295 - 1;
   ...
   case GST_ISP_POINTER:
     maxIndex = 15;
     index += instanceSpec >> ISP_NUMFIXEDFIELDS (5)
     return object->data[5]

...

Is this signed => unsigned conversion wanted or is it just some
luck that this is working out okay?

holger

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