Reproducible .wasm file builds

[kennu]

While I was building .wasm files with various languages (AssemblyScript, C, C++, C#, Go, Grain, Haskell, Motoko, Rust, Zig), I noticed that Swift is the only one that generates a different .wasm file after every build.

In practice, if you store .wasm files in Git or otherwise rely on change detection, you get a new version after every build, even if the source code didn't change.

This is not a big problem for me, but I thought I'd mention it, since it appears to be unique to Swift. Please feel free to close the issue if it's out of scope of the project.

[kennu]

With a hex dump, the difference is 16 bytes somewhere deep in the middle of the .wasm files:

diff 1.txt 2.txt       
295713c295713
< 22154160 20 fc 5f a5 fa 0f ef 97 26 dd cf b7 b2 d7 e9 60
---
> 22154160 e5 3b 7a d5 5b e6 4a 9e 7d 2f f0 d9 03 35 15 c5

[kateinoigakukun]

Thank you for reporting :) The differential section is .swift_modhash, which is a metadata section for incremental compilation, but I'm not sure why it changes for every build.
As far as I checked, Swift toolchain for Linux also produce different hash for each build, so I guess it's not wasm specific problem.
But we should keep this track on apple/swift repository 👍

[ephemer]

I'm curious whether this also occurs during release builds?

[gregcotten]

@ephemer

Answer: yes

[kateinoigakukun]

--gc-sections support resolved this issue #5128 because it enabled GCing .swift_modhash section which is always unused at runtime. I've confirmed swift build emits exactly same binaries for the same inputs with the latest nightly toolchain https://github.com/swiftwasm/swift/releases/tag/swift-wasm-DEVELOPMENT-SNAPSHOT-2024-03-01-a