Supposedly, the new boot loader code sets up the CPU to expect different memory sizes. Although the data word width is the same, the bank sizes are different.

The data abort message could be caused by the kernel when it attempts to access a memory structure that does not match the physical memory. Hence, there's a hardware vector call that sends the program counter to the handler for that CPU error condition. If the kernel has no handler installed to handle the condition, the CPU stalls.

Sounds to me like your physical memory map seems to have changed sometime. What happens if you try to run the memory tests at boot time? Repeatedly?