Request for review (S): 7110152: assert(size_in_words <= (julong)max_jint) failed: no overflow

Bengt Rutisson bengt.rutisson at
Sun Nov 13 06:14:37 PST 2011

Hi all,

I already sent this out for review on the GC alias. It started out as a 
pure GC change, but after some updates it has become a change in 
src/share/vm/oops/arrayOop.hpp, which deserves a broader alias (thanks 
David Holmes for pointing this out). So, I am resending this review 
request to the hotspot-dev alias.

Vladimir and David Holmes have already looked at it. But more comments 
are of course welcome.

The background is that when I pushed this:
7102044: G1: VM crashes with assert(old_end != new_end) failed: don't 
call this otherwise

I changed arrayOopDesc::max_array_length() to always return max_jint in 
64 bit VMs. This was intentional and is correct as far as I can tell. 
But it turns out that a large part of the GC code passes object sizes 
around as int and not size_t. Thus, we might overflow the int when we 
add the object header.

My proposed change now reverts back to reducing the maximum array length 
to be slightly less than max_jint.


7110152 assert(size_in_words <= (julong)max_jint) failed: no overflow

Here is a reproducer that triggers the assert before my change but 
passes after:

public class MaxJIntArray {
     public static void main(String[] args) {
         final int MAX_JINT = 2147483647;
         double[] a = new double[MAX_JINT];
         System.out.println("Allocated a[" + a.length + "]");


More information about the hotspot-dev mailing list