[PATCH] D51664: [IR] Lazily number instructions for local dominance queries

[Reid Kleckner via Phabricator via llvm-commits]

rnk added a comment.

In https://reviews.llvm.org/D51664#1243954, @sanjoy wrote:

> > If I get no results, at least 1% RSS is an upper bound on increased LTO memory usage. I'm happy to trade that for 40% shorter compile time of the slowest TUs in clang.
>
> That I agree with.  :)


I'm only patient enough to get two runs, before and after, and max RSS before the patch was 4816464 kb, and after, 4867836 kb. That's an increase of 1.06%. In absolute terms, ~50MB of wasted memory for Instruction positions feels high.

> Though, as I said on the llvm-dev thread, we may be able to get the best of both worlds by using something like the waymarking algorithm.

That's probably possible, but I'm a little afraid to try to steal bits from ilist next/prev pointers. The complexity cost of stealing those bits from ilist is likely to be more than is really worth it, and it's not clear if we want to use linked lists over the long term. I would rather get right complexity first, and then make a follow-up change to try to reduce the memory usage. I don't want to let the perfect be the enemy of the good.


https://reviews.llvm.org/D51664