undefined | Better HN

0 pointsJoker_vD2y ago0 comments

If by unified memory model you mean "flat address space" then no, it's not. The moment you need two (or more) dynamically-sized arrays, you need to implement realloc with memmov in the unfortunate case. In a world where each array could have it's own segment, this problem doesn't arise because they simply cannot intersect and realloc boils down into increasing/decreasing a segment's extent.

0 comments

3 comments · 2 top-level

grotorea2y ago· 1 in thread

Wouldn't that just move the "magic" down to the kernel or the MMU which has to do the realloc and memmov hidden from the programmer instead of the programmer being able to chose when and how to do it?

Joker_vDOP2y ago

Yes, but here's the thing: the virtual memory system already does this kind of "magic", maintaining the mapping of virtual addresses onto the physical ones, and it doesn't need to actually move data between physical pages at all since the mapping is discontinuous.

gpderetta2y ago

On the good[1] old times of x86 segments, you still couldn't assign a segment to each array as segments were a limited resource. Something still has to do the mapping of the segments to physical linearly addressable memory.

[1] spoiler alert: they were not good.

j / k navigate · click thread line to collapse