elide pass-by-ref in root functions

[strake]

No description provided.

[cuviper]

Presumably this is for performance -- do you have a benchmark comparison? Can you get the same effect by just binding let a = *self; at the start, rather than nesting another function?

[strake]

Presumably this is for performance -- do you have a benchmark comparison?

here — it reliably cuts about 1/5 off the time for {i128, u128}::sqrt_rand in particular, and mostly speeds up the others a little also (on my machine).

Can you get the same effect by just binding let a = *self; at the start, rather than nesting another function?

No, the compiler already lifts the load to the start of the function — this PR elides the load altogether. The problem is the trait methods take self by reference.

[cuviper]

Ah, I did some profiling, and now I see it's a difference of inlining. The current code doesn't inline at all, but with your change it's inlining the outer &self method and calling sqrt::go by value. That's what lets it avoid the reference and load altogether.

Thanks!

bors r+

[bors]

Build succeeded

• continuous-integration/travis-ci/push