Comment 5 for bug 2020119

2GB might be too little, 105MB with unicode strings will take 420MB, multiple by two, 840MB, then during the GC might have to use twice as much space, so 1680MB, and that just the strings.

And it can be more than 2x the size of the working set, as the gc might be triggered in the middle of the third iteration, so you'll have more than two files in memory.

Probably just set A to NIL before reading a new file, use base-strings, etc.