In the following code in MutableNUMASpace::ensure_parsability:

            size_t touched_words = ... pointer_delta(s->end(), s->top()) ...; // where end() and top() are HeapWords.
[...]
 0 HeapWord *crossing_start = align_up((HeapWord*)cur_top, os::vm_page_size());
 1 HeapWord *crossing_end = align_down((HeapWord*)(cur_top + touched_words), os::vm_page_size());
 2 if (crossing_start != crossing_end) {
 3 // If object header crossed a small page boundary we mark the area
 4 // as invalid rounding it to a page_size().
 5 HeapWord *start = MAX2(align_down((HeapWord*)cur_top, page_size()), s->bottom());
 6 HeapWord *end = MIN2(align_up((HeapWord*)(cur_top + touched_words), page_size()), s->end());
 7 invalid = MemRegion(start, end);

cur_top is an intptr_t that gets added something in word_granularity in lines 1 and 6. That looks wrong, particularly because in the other uses of align_up and align_down cur_top is cast to HeapWord* first.

I.e. cur_top + touched_words doesn't use pointer arithmetic, so the the resulting values are too small.

This was found while working on JDK-8178489.