Shrink DecodeError from 48 to 32 bytes on 64-bit arch

[maciejhirsz]

This is done by two tricks:

1. Packing the entire description of an enum with it's type name and allowed variants behind a &'static ref.
2. Reducing DecodeError::CStringNulError to just the position of the null byte.

Aside from the extra struct being added, this is a breaking change due to type names for many manual impls being reduced to just the identifier name (derived impls were already just the identifier here). This is unfortunately necessary as std::any::type_name is not const, and even if it were using it with an external generic type would make it unusable in const environment, meaning it's impossible to construct a static AllowedEnum struct with it. CStringNulError I think is exotic enough and rare enough that dropping information here makes sense, if not we could also Box the internal error (std already allocates there anyway).

If these changes are unacceptable, it's still possible to reduce the size of DecodeError to 40 bytes by just putting the AllowedEnumVariants behind a &'static ref, as that enum alone is 3 words wide.

[codecov]

Codecov Report

Merging #553 (fd9c82e) into trunk (18d5249) will increase coverage by 0.05%.
The diff coverage is 42.85%.

@@            Coverage Diff             @@
##            trunk     #553      +/-   ##
==========================================
+ Coverage   54.44%   54.49%   +0.05%     
==========================================
  Files          49       50       +1     
  Lines        4423     4426       +3     
==========================================
+ Hits         2408     2412       +4     
+ Misses       2015     2014       -1     

Impacted Files	Coverage Δ
derive/src/derive_enum.rs	0.00% <ø> (ø)
src/de/impls.rs	58.51% <0.00%> (ø)
src/de/mod.rs	26.37% <0.00%> (ø)
src/error.rs	0.00% <ø> (ø)
tests/derive.rs	97.76% <ø> (-0.02%)	⬇️
tests/issues/issue_498.rs	80.00% <ø> (ø)
src/features/impl_std.rs	66.34% <40.00%> (+0.48%)	⬆️
tests/error_size.rs	100.00% <100.00%> (ø)

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 18d5249...fd9c82e. Read the comment docs.

[maciejhirsz]

Actually I think I miscalculated, and we can get 32 bytes wide DecodeError by just having the AllowedEnumVariants behind a ref: type_name is 2 words, allowed (AllowedEnumVariants) behind a ref is 1 word, found is half a word and can fit alongside the tag -> 4 words total.

Gonna redo that.

[VictorKoenders]

Can we add a test to validate the size of all Error types in bincode? That way we know if they change and what the actual value is

[maciejhirsz]

FYI: this is the variant that makes EncodeError 40 bytes wide with std on:

    /// The encoder tried to encode a `SystemTime`, but it was before `SystemTime::UNIX_EPOCH`
    #[cfg(feature = "std")]
    InvalidSystemTime {
        /// The error that was thrown by the SystemTime
        inner: std::time::SystemTimeError,
        /// The SystemTime that caused the error
        time: std::time::SystemTime,
    },

Could remove time, or box one or the other.

[VictorKoenders]

Interesting, on x86_64-pc-windows-gnu + std, the EncodeError is 32 bytes

But why

[maciejhirsz]

Interesting, on x86_64-pc-windows-gnu + std, the EncodeError is 32 bytes

But why

Kind of makes sense, SystemTime is platform specific and it's more precise on Linux than Windows. What is interesting is that SystemTimeError is a newtype for Duration, which has the same layout all platforms.

I've boxed time in that variant, that should keep the whole thing behave well on all platforms with 32-byte EncodeError.

[VictorKoenders]

Oh I thought we were compiling against -msvc as well, and I was very confused that -gnu was different.

Boxing the SystemTime seems like a good solution to me

[VictorKoenders]

Thanks!