While looking at bug 1844868, I noticed string deduplication during nursery GC affects some of the Speedometer 3 sub tests, especially Charts-*.

A perf comparison for a patch to disable deduplication shows some 3-6% improvements. Also some regressions but it's probably worth investigating what's going on.

There is also bug 1837937

I've always expected string deduplication to slow things down, but back when I was looking for it, I was surprised to find that none of the benchmarks seemed to be affected. This is a useful discovery.

I guess for right now, I'm hoping that I can make up the time by making better use of that deduplication time by storing the hash generated during it. But that's total speculation. I should have numbers for it pretty soon, at least.

Otherwise, I'll need to look closer and see if either pretenuring or just skipping deduplication for some subset of strings would fix this.