fmt of non-decimal radix untangled

[pascaldekloe]

Have the implementation match its decimal counterpart.

• Digit table instead of conversion functions
• Correct buffer size per radix
• Elimination of dead code for negative
• No trait abstraction for integers

Original Performance

    fmt::write_10ints_bin                                                393.03ns/iter      +/- 1.41
    fmt::write_10ints_hex                                                316.84ns/iter      +/- 1.49
    fmt::write_10ints_oct                                                327.16ns/iter      +/- 0.46

Patched Performance

    fmt::write_10ints_bin                                                392.31ns/iter      +/- 3.05
    fmt::write_10ints_hex                                                302.41ns/iter      +/- 5.48
    fmt::write_10ints_oct                                                322.01ns/iter      +/- 3.82

r? tgross35

[pascaldekloe]

https://rust.godbolt.org/z/K1hWxa7dP

[bors]

☔ The latest upstream changes (presumably #144044) made this pull request unmergeable. Please resolve the merge conflicts.

[pascaldekloe]

Can you either have a look or reassign @tgross35? Got a followup pending on this one...

[tgross35]

I'll reassign for now, but if nobody beats me to it I'll take a look on Monday

r? libs

(leaving myself assigned so it stays on my list)

[tgross35]

By the way; if you are interested, we could use some int (and float) formatting and parsing benchmarks at https://github.com/rust-lang/rustc-perf/blob/2120e3b7b8996e96858b88edefea371679a3d415/collector/runtime-benchmarks/fmt/src/main.rs (I just learned that is possible), which would mean they get run as part of our pretty extensive perf infra rather than you needing to post the results of local runs.

[pascaldekloe]

Interesting indeed @tgross35. The specialized benchmarks we have at the moment (such as write_u8_min and write_u64_max) won't scale. Think about how many benchmarks we need for fmt::LowerExp with edge cases on alignment and precision. I am curious to hear what you think of the _10ints_ concept in this patch. The idea is to have a balanced mix of common cases.