Demangle using fixed-length disambiguators

[teor2345]

Most UI tests expect a 16-character disambiguator, but rustc-demangle omits leading zeroes.
This can cause test failures on ~1/16 of the changes to demangling tests.

For example: rust-lang/rust#160050 (comment)

This PR adds a Disambiguator wrapper type, so it is impossible to format it incorrectly.

I used https://www.dcode.fr/base62-encoding to generate the encoded examples, you need to select the second alphabet (lowercase first).

[bjorn3]

write!(out, "{:016x}", self.0) would work too, right?

[teor2345]

Unfortunately this is core-only code:

rustc-demangle/src/v0.rs

Lines 5 to 12 in f36e298

	macro_rules! write {
	($($ignored:tt)*) => {
	compile_error!(
	"use `self.print(value)` or `fmt::Trait::fmt(&value, self.out)`, \
	instead of `write!(self.out, \"{...}\", value)`"
	)
	};
	}

[bjorn3]

write!() is in core. I don't know if the reason given for avoiding write!() in #50 is still fully valid given that we had a change to the fmt implementation since:

The second commit is mostly stylistic, trying to enforce a pattern of "not using write! in demangling logic", and only using fmt::Trait::fmt(&value, self.out) - which I'm guessing we do want (for performance reasons?).

In any case I guess the current code is fine.

[teor2345]

I have a slight preference for the standard fixed-width code, it's more maintainable and less likely to have bugs.

But my custom code benchmarks at:

test v0::tests::bench_demangle_crate_with_zero_disambiguator ... bench:         173.72 ns/iter (+/- 8.86)
test v0::tests::bench_short_crate_disambiguator_bug          ... bench:       3,842.77 ns/iter (+/- 69.09)

And format_args! (the expansion of write!) is slightly slower:

test v0::tests::bench_demangle_crate_with_zero_disambiguator ... bench:         175.11 ns/iter (+/- 6.15)
test v0::tests::bench_short_crate_disambiguator_bug          ... bench:       4,005.31 ns/iter (+/- 240.34)

Specifically, it expands to out.write_fmt(core::format_args!("{:016x}", self.0)).

I'm using a MacOS M1 laptop, so I don't know if these results carry across. But it seems like custom code may still be a winner over write!. I'll add a comment to explain this, and push an update.

[teor2345]

My benchmarking code is here:
https://github.com/teor2345/rustc-demangle/tree/short-crate-hash-benchmark

No other tests format the crate hash, which explains why we never picked up this bug.