Tweak enum serialization to generate better LLVM IR and more compact code

[alessandrod]

Hello!

I'm looking at reducing the size of some programs that use borsh, and I've noticed that deriving BorshSerialize on enums can sometimes trigger some pretty bad inlining from rustc/LLVM.

This PR changes enum serialization from:

  match self {
    Enum::Variant1(field1, field2, ...) => {
        // serialize variant idx
        let variant_idx: u8 = 0;
        writer.write_all(&variant_idx.to_le_bytes())?;

        // serialize variant fields
        BorshSerialize::serialize(field1, writer)?;
        BorshSerialize::serialize(field2, writer)?;
        ...
    }
    Enum::Variant2(field1, field2, ...) => {
        // serialize variant idx
        let variant_idx: u8 = 1;
        writer.write_all(&variant_idx.to_le_bytes())?;

        // serialize variant fields
        BorshSerialize::serialize(field1, writer)?;
        BorshSerialize::serialize(field2, writer)?;
        ...

    }
    ...
  }

To:

  let variant_idx: u8 = match self {
    Enum::Variant1(..) => 0,
    Enum::Variant2(..) => 1,
    ...
  };
  // serialize variant_idx
  writer.write_all(&variant_idx.to_le_bytes())?;

  match self {
    Enum::Variant1(field1, field2, ...) => {
        // serialize variant fields
        BorshSerialize::serialize(field1, writer)?;
        BorshSerialize::serialize(field2, writer)?;
        ...
    }
    Enum::Variant2(field1, field2, ...) => {
        // serialize variant fields
        BorshSerialize::serialize(field1, writer)?;
        BorshSerialize::serialize(field2, writer)?;
        ...

    }
    ...
  }

The latter generates better LLVM IR, and avoids writer.write_all(&variant_idx.to_le_bytes())? from being inlined into each match branch.

Here's a test case that shows the issue: https://gist.github.com/alessandrod/f192c59f6b75159733d2cb6c60a78f1b

Cargo bloat before:

 File  .text    Size Crate Name
13.8%  36.5%  6.1KiB bopt? <bopt::LargeEnum as borsh::ser::BorshSerialize>::serialize
 4.6%  12.0%  2.0KiB   std core::fmt::Formatter::pad_integral
 3.2%   8.5%  1.4KiB   std <&T as core::fmt::Display>::fmt
 3.1%   8.2%  1.4KiB   std <core::fmt::Arguments as core::fmt::Display>::fmt
 2.7%   7.0%  1.2KiB   std core::fmt::num::imp::<impl core::fmt::Display for u64>::fmt
10.5%  27.8%  4.7KiB       And 33 smaller methods. Use -n N to show more.
37.9% 100.0% 16.9KiB       .text section size, the file size is 44.4KiB

Cargo bloat after:

 File  .text    Size Crate Name
 6.8%  19.0%  2.0KiB   std core::fmt::Formatter::pad_integral
 4.8%  13.3%  1.4KiB   std <&T as core::fmt::Display>::fmt
 4.7%  12.9%  1.4KiB   std <core::fmt::Arguments as core::fmt::Display>::fmt
 4.0%  11.0%  1.2KiB   std core::fmt::num::imp::<impl core::fmt::Display for u64>::fmt
 2.9%   8.1%    888B   std <core::str::iter::CharIndices as core::iter::traits::iterator::Iterator>::next
12.9%  35.7%  3.8KiB       And 31 smaller methods. Use -n N to show more.
36.1% 100.0% 10.7KiB       .text section size, the file size is 29.7KiB

[matklad]

excellent find, thanks!