[Vm-dev] newDelta >0 failed in sqUnixMemory

[David T. Lewis]

On Sat, Jan 31, 2009 at 11:15:02AM +0100, Damien Cassou wrote:
>  
> On Fri, Jan 30, 2009 at 6:29 PM, Andreas Raab <andreas.raab at gmx.de> wrote:
> >>> fprintf(stderr, "oldLimit=%p\ndelta=%d\nnewSize=%d\nnewDelta=%d\n");
> >>>
> >>> oldLimit=0x80562b8
> >>> delta=-1077506096
> >>> newSize=0
> >>> newDelta=423027548
> >>> squeak: /tmp/squeak-svn/platforms/unix/vm/sqUnixMemory.c:175:
> >>> uxGrowMemoryBy: Assertion `newDelta >= 0' failed.
> >>>
> >>> Does it help?
> >>
> >> Yes. I don't think newSize is supposed to be 0 here.
> >>
> >>      int newSize=  min(valign(oldLimit - heap + delta), heapLimit);
> >>      int newDelta= newSize - heapSize;
> >>
> >> I'm not sure what is wrong, but apparently it's in the expression
> >> "min(valign(oldLimit - heap + delta), heapLimit)".
> >
> > Could be a CSE bug. Considering that min(a,b) is usually defined as "a < b ?
> > a : b" or so, a compiler might do something with the valign(...) expression.
> > I would try to move the expression out of there and see if that helps.
> 
> Nobody noticed, but I did a very stupid mistake. Look at the call to
> fprintf that I copy/pasted in my previous mail. I forgot all the
> parameters :-). I don't how it is possible that gcc still compiled the
> code...
> 
> I moved valign out:
> 
>       int align = valign(oldLimit - heap + delta);
>       int newSize= min(align, heapLimit);
> 
> And printed the values:
> 
>  if(newDelta < 0)
> 	fprintf(stderr,
> "oldLimit=%p\ndelta=%d\nnewSize=%d\nnewDelta=%d\nalign=%d\n",oldLimit,delta,newSize,newDelta,align);
> 
> And got:
> 
> oldLimit=0x791c8fe8
> delta=0
> newSize=23293952
> newDelta=-4096
> align=23293952
> squeak: /tmp/squeak-svn/platforms/unix/vm/sqUnixMemory.c:175:
> uxGrowMemoryBy: Assertion `newDelta >= 0' failed.
> Aborted
> 
> Does that make sense to anyone?

Something is going wrong in the call to valign. I can't quite
spot it, but note that valign(x) is {(x) & pageMask}, where pageMask
is 0xfffff000. Your newDelta result is -4096, which also happens
to be 0xfffff000.

Maybe as a next step you can break out the calculation further:
       int argToAlign = oldLimit - heap + delta;
       int align = valign(argToAlign);
       int newSize= min(align, heapLimit);

Also printing the values of heap and heapLimit might give a clue.

Just a hunch, but I would guess that if you changed valign() from a
macro to a function that it would somehow decide to start working.

Dave