[vwnc] VW crash on large method invocation

Hi Eliot! Thanks for the detailed explanation. I know this is a (very) rare case, however since it is potentially dangerous and a trivial solution doesn't seem to be very complex (just throw an error if, after compiling, the size of the compiled code is greater than X) I can't help wondering why it isn't fixed. At least It would have saved me some debugging time :)

Cheers!
Andrés

Eliot Miranda escribió:

On Thu, Aug 21, 2008 at 7:13 AM, Andres Fortier <
[hidden email]> wrote:

Hi Reinout, thanks for the tip and explanation. However I still find
it quite weird that I get no warnings or errors when compiling...

If you run in a console window (assuming you're on WIndows) the VM should
print some form of translation failure on stdout.

Steve and I have discussed the following solution often, but never got round
to implementing, because this is a rare problem.  The VM should maintain a
limit on the size of native method it will generate and provide some
send-back error path, much like doesNotUnderstand: (but more similar in
intent to illegalBytecode), so that when one tries to run a method that
would generate too much code the VM sends, e.g. methodTooBig: aMessage, to
the receiver, and methodTooBig: interprets the method using the simulation
machinery.  Overall this is probably better then translating because it
allows the VM to maintain a better working set of translated methods that
the interpreted method will call, and the VM doesn't waste time translating
huge methods.  The simulation machinery can probably be sped-up considerably
using perform: and an array of selectors for the bytecode set.


Thanks again!

Andrés

Reinout Heeck escribió:

On Windows your code crashes the VM too (translation error).

Looking at your method I see that it contains about 60000 bytes of
bytecode. Assuming the JIT creates machine code that is a factor 4 (just
a figure) larger it seems that you might be blowing the JITs capability
to use short branches or indicate compiled code length in two bytes or
something else like that.

The ratios in the VM are about 10 to 1 for CISC code (x86) through 20 to 1
for RISC code such as SPARC.   The VM's branch code *should* cope.  It does
an optimistic pass using short branches and if this fails sets a "use long
branches" flag, and redoes the compilation with long branches (but there
could easily be a bug here).

The pattern to use with VW is to create large literal arrays and use
small code to turn that into your desired object. This is how windowSpec
methods work for instance.


So if you write your code like this

data
  ^#(
     #(-7.19265 -10.0)
     #(-4.39122 -10.0)
     #(-1.59041 -10.0)
     ...
     #(-4.2767 6.19155)
     #(-7.35022 -6.39121)
  )
     collect: [:pair | pair first @ pair last]


you end up with a method that contains only 5 byte codes and 1 literal
-- nowhere near any VM limits :-)


HTH,

Reinout
-------




Andres Fortier wrote:

Hi list,
         today I bumped into a weird problem (VW 7.4.1 running on
linux). I was preparing some datasets to perform tests and profiles.
These datasets where gathered from files and automatically converted to
methods, generating something like:

data

| col |
col:=OrderedCollection new.
col add: e1.
col add: e2.
...
col add: eN.
^col.

However, when evaluating the method to get the collection, VW just juts
down; no warnings or errors, just quit. By doing some trial-and-error
I've found out that the limit for the collection seems to be 3952
elements (in particular Point elements); if it is larger than that it
breaks the image.
I'm sending a file-out attached. Evaluating "Explode data" should shut
down the image

Any hints about this problem are appreciated.

Cheers,
           andrés


------------------------------------------------------------------------

_______________________________________________
vwnc mailing list
[hidden email]
http://lists.cs.uiuc.edu/mailman/listinfo/vwnc

_______________________________________________
vwnc mailing list
[hidden email]
http://lists.cs.uiuc.edu/mailman/listinfo/vwnc

_______________________________________________
vwnc mailing list
[hidden email]
http://lists.cs.uiuc.edu/mailman/listinfo/vwnc