collections: dramatically speed up Vec::reserve with magic

[erickt]

(I don't understand why this works, and so I don't quite trust this yet. I'm pushing it up to see if anyone else can replicate this performance increase)

Somehow llvm is able to optimize this version of Vec::reserve into dramatically faster than the old version. In micro-benchmarks this was 2-10 times faster. It also reduce my Rust compile time from 41 minutes to 27 minutes.

Closes #19281.

[pcwalton]

cc @thestinger or @eddyb if you want to investigate this

[aturon]

It also reduce my Rust compile time from 41 minutes to 27 minutes.

!!!

[nikomatsakis]

I say we land it. I suspect the gains will be markedly lower than advertised, but the code reads nicer anyway. :)

[bstrie]

Let's land it, but let's file a bug to investigate the reasons why LLVM is so finicky here. This is too spooky to just sweep under the rug.

[japaric]

I didn't see any noticeable difference in build time between this and master:

# master
make -j8  2074.12s user 23.50s system 131% cpu 26:34.38 total
# This PR on top of master
make -j8  2240.88s user 25.49s system 131% cpu 28:39.00 total

(some per crate-ish times)

My build times have always been around or below 30 minutes (without using -Ccodegen_units), so this looks normal.

We could still merge this and see if there's any change on the bots. (or we could check it on the try branch first?)

[erickt]

Yeah, on a second build I wasn't able to get the benefit either :(. I think my without-patch build was just abnormally slow rather than this patch being abnormally fast.