style: Switch to FxHash for the style system.

[emilio]

This improves stylist rebuild times by a pretty neat amount, see bug 1386443.

---

This change is 

[highfive]

Heads up! This PR modifies the following files:

• @bholley: components/style/Cargo.toml, components/style/gecko/data.rs, components/style/selector_map.rs, components/style/lib.rs, components/style/invalidation/media_queries.rs and 7 more
• @canaltinova: components/style/Cargo.toml, components/style/gecko/data.rs, components/style/selector_map.rs, components/style/lib.rs, components/style/invalidation/media_queries.rs and 7 more

[highfive]

Warning

• These commits modify style and layout code, but no tests are modified. Please consider adding a test!

[emilio]

r? @bholley

[emilio]

@bors-servo try

[bors-servo]

⌛ Trying commit fbbf5c0 with merge 8774284...

[emilio]

(also, @bzbarsky may want to review)

[bors-servo]

💔 Test failed - linux-dev

[mbrubeck]

It looks like the only use of fnv in Stylo is now in the selectors crate. Should we convert that too, and reduce the number of crates in the Stylo build?

[bzbarsky]

The patch seems pretty straightforward. My question is where the free lunch comes from.

That is, given the assumption that FnvHash is competently written, if it's this much slower that must be because it's able to do something that FxHash doesn't do. Presumably we don't care about that something, but do we know what it is and why we don't care about it?

[bholley]

Is it just that the hashing algorithm is faster? I thought we mostly didn't use the hashing algorithm, since we have precomputed hashes on nsIAtom. Unless we have significant traffic on non-atom-keyed hashmaps? If so, what are they?

As for selectors, seems like we should either remove or compile out the dependency on FnvHash, since we don't actually use the codepath in stylo (we use precomputed hash). That could be a followup E-Easy bug though.

[emilio]

Is it just that the hashing algorithm is faster? I thought we mostly didn't use the hashing algorithm, since we have precomputed hashes on nsIAtom. Unless we have significant traffic on non-atom-keyed hashmaps? If so, what are they?

We're still hashing that hash, afaict.

And re @bzbarsky's questions, I think I'd rather point to rust-lang/rust#37229, which has already pretty much all the discussion I can imagine :)

[bors-servo]

☔ The latest upstream changes (presumably #17988) made this pull request unmergeable. Please resolve the merge conflicts.

[emilio]

I'm working on something better than this :)