WebAssembly targets: change in default target-features

⚓ Rust    📅 2024-09-24    👤 freedit    👁️ 22      

freedit

The Rust compiler has recently upgraded to using LLVM 19 and this change accompanies some updates to the default set of target features enabled for WebAssembly targets of the Rust compiler. Beta Rust today, which will become Rust 1.82 on 2024-10-17, reflects all of these changes and can be used for testing.

WebAssembly is an evolving standard where extensions are being added over time through a proposals process. WebAssembly proposals reach maturity, get merged into the specification itself, get implemented in engines, and remain this way for quite some time before producer toolchains (e.g. LLVM) update to enable these sufficiently-mature proposals by default. In LLVM 19 this has happened with the multi-value and reference-types proposals for the LLVM/Rust target features multivalue and reference-types. These are now enabled by default in LLVM and transitively means that it's enabled by default for Rust as well.

WebAssembly targets for Rust now have improved documentation about WebAssembly proposals and their corresponding target features. This post is going to review these changes and go into depth about what's changing in LLVM.

WebAssembly Proposals and Compiler Target Features

WebAssembly proposals are the formal means by which the WebAssembly standard itself is evolved over time. Most proposals need toolchain integration in one form or another, for example new flags in LLVM or the Rust compiler. The -Ctarget-feature=... mechanism is used to implement this today. This is a signal to LLVM and the Rust compiler which WebAssembly proposals are enabled or disabled.

There is a loose coupling between the name of a proposal (often the name of the github repository of the proposal) and the feature name LLVM/Rust use. For example there is the multi-value proposal but a multivalue feature.

The lifecycle of the implementation of a feature in Rust/LLVM typically looks like:

  1. A new WebAssembly proposal is created in a new repository, for example WebAssembly/foo.
  2. Eventually Rust/LLVM implement the proposal under -Ctarget-feature=+foo
  3. Eventually the upstream proposal is merged into the specification, and WebAssembly/foo becomes an archived repository
  4. Rust/LLVM enable the -Ctarget-feature=+foo feature by default but typically retain the ability to disable it as well.

The reference-types and multivalue target features in Rust are at step (4) here now and this post is explaining the consequences of doing so.

Enabling Reference Types by Default

The reference-types proposal to WebAssembly introduced a few new concepts to WebAssembly, notably the externref type which is a host-defined GC resource that WebAssembly cannot access but can pass around. Rust does not have support for the WebAssembly externref type and LLVM 19 does not change that. WebAssembly modules produced from Rust will continue to not use the externref type nor have a means of being able to do so. This may be enabled in the future (e.g. a hypothetical core::arch::wasm32::Externref type or similar), but it will mostly likely only be done on an opt-in basis and will not affect preexisting code by default.

Also included in the reference-types proposal, however, was the ability to have multiple WebAssembly tables in a single module. In the original version of the WebAssembly specification only a single table was allowed and this restriction was relaxed with the reference-types proposal. WebAssembly tables are used by LLVM and Rust to implement indirect function calls. For example function pointers in WebAssembly are actually table indices and indirect function calls are a WebAssembly call_indirect instruction with this table index.

With the reference-types proposal the binary encoding of call_indirect instructions was updated. Prior to the reference-types proposal call_indirect was encoded with a fixed zero byte in its instruction (required to be exactly 0x00). This fixed zero byte was relaxed to a 32-bit LEB to indicate which table the call_indirect instruction was using. For those unfamiliar LEB is a way of encoding multi-byte integers in a smaller number of bytes for smaller integers. For example the 32-bit integer 0 can be encoded as 0x00 with a LEB. LEBs are flexible to additionally allow "overlong" encodings so the integer 0 can additionally be encoded as 0x80 0x00.

LLVM's support of separate compilation of source code to a WebAssembly binary means that when an object file is emitted it does not know the final index of the table that is going to be used in the final binary. Before reference-types there was only one option, table 0, so 0x00 was always used when encoding call_indirect instructions. After reference-types, however, LLVM will emit an over-long LEB of the form 0x80 0x80 0x80 0x80 0x00 which is the maximal length of a 32-bit LEB. This LEB is then filled in by the linker with a relocation to the actual table index that is used by the final module.

When putting all of this together, it means that with LLVM 19, which has the reference-types feature enabled by default, any WebAssembly module with an indirect function call (which is almost always the case for Rust code) will produce a WebAssembly binary that cannot be decoded by engines and tooling that do not support the reference-types proposal. It is expected that this change will have a low impact due to the age of the reference-types proposal and breadth of implementation in engines. Given the multitude of WebAssembly engines, however, it's recommended that any WebAssembly users test out Rust 1.82 beta and see if the produced module still runs on their engine of choice.

LLVM, Rust, and Multiple Tables

One interesting point worth mentioning is that despite the reference-types proposal enabling multiple tables in WebAssembly modules this is not actually taken advantage of at this time by either LLVM or Rust. WebAssembly modules emitted will still have at most one table of functions. This means that the over-long 5-byte encoding of index 0 as 0x80 0x80 0x80 0x80 0x00 is not actually necessary at this time. LLD, LLVM's linker for WebAssembly, wants to process all LEB relocations in a similar manner which currently forces this 5-byte encoding of zero. For example when a function calls another function the call instruction encodes the target function index as a 5-byte LEB which is filled in by the linker. There is quite often more than one function so the 5-byte encoding enables all possible function indices to be encoded.

In the future LLVM might start using multiple tables as well. For example LLVM may have a mode in the future where there's a table-per-function type instead of a single heterogenous table. This can enable engines to implement call_indirect more efficiently. This is not implemented at this time, however.

For users who want a minimally-sized WebAssembly module (e.g. if you're in a web context and sending bytes over the wire) it's recommended to use an optimization tool such as wasm-opt to shrink the size of the output of LLVM. Even before this change with reference-types it's recommended to do this as wasm-opt can typically optimize LLVM's default output even further. When optimizing a module through wasm-opt these 5-byte encodings of index 0 are all shrunk to a single byte.

Enabling Multi-Value by Default

The second feature enabled by default in LLVM 19 is multivalue. The multi-value proposal to WebAssembly enables functions to have more than one return value for example. WebAssembly instructions are additionally allowed to have more than one return value as well. This proposal is one of the first to get merged into the WebAssembly specification after the original MVP and has been implemented in many engines for quite some time.

The consequences of enabling this feature by default in LLVM are more minor for Rust, however, than enabling the reference-types feature by default. LLVM's default C ABI for WebAssembly code is not changing even when multivalue is enabled. Additionally Rust's extern "C" ABI for WebAssembly is not changing either and continues to match LLVM's (or strives to, differences to LLVM are considered bugs to fix). Despite this though the change has the possibility of still affecting Rust users.

Rust for some time has supported an extern "wasm" ABI on Nightly which was an experimental means of exposing the ability of defining a function in Rust which returned multiple values (e.g. used the multi-value proposal). Due to infrastructural changes and refactorings in LLVM itself this feature of Rust has been removed and is no longer supported on Nightly at all. As a result there is no longer any possible method of writing a function in Rust that returns multiple values at the WebAssembly function type level.

In summary this change is expected to not affect any Rust code in the wild unless you were using the Nightly feature of extern "wasm" in which case you'll be forced to drop support for that and use extern "C" instead. Supporting WebAssembly multi-return functions in Rust is a broader topic than this post can cover, but at this time it's an area that's ripe for contribution from suitably motivated contributors.

Aside: ABI Stability and WebAssembly

While on the topic of ABIs and the multivalue feature it's perhaps worth also going over a bit what ABIs mean for WebAssembly. The current definition of the extern "C" ABI for WebAssembly is documented in the tool-conventions repository and this is what Clang implements for C code as well. LLVM implements enough support for lowering to WebAssembly as well to support all of this. The extern "Rust ABI is not stable on WebAssembly, as is the case for all Rust targets, and is subject to change over time. There is no reference documentation at this time for what extern "Rust" is on WebAssembly.

The extern "C" ABI, what C code uses by default as well, is difficult to change because stability is often required across different compiler versions. For example WebAssembly code compiled with LLVM 18 might be expected to work with code compiled by LLVM 20. This means that changing the ABI is a daunting task that requires version fields, explicit markers, etc, to help prevent mismatches.

The extern "Rust" ABI, however, is subject to change over time. A great example of this could be that when the multivalue feature is enabled the extern "Rust" ABI could be redefined to use the multiple-return-values that WebAssembly would then support. This would enable much more efficient returns of values larger than 64-bits. Implementing this would require support in LLVM though which is not currently present.

This all means that actually using multiple-returns in functions, or the WebAssembly feature that the multivalue enables, is still out on the horizon and not implemented. First LLVM will need to implement complete lowering support to generate WebAssembly functions with multiple returns, and then extern "Rust" can be change to use this when fully supported. In the yet-further-still future C code might be able to change, but that will take quite some time due to its cross-version-compatibility story.

Enabling Future Proposals to WebAssembly

This is not the first time that a WebAssembly proposal has gone from off-by-default to on-by-default in LLVM, nor will it be the last. For example LLVM already enables the sign-extension proposal by default which MVP WebAssembly did not have. It's expected that in the not-too-distant future the nontrapping-fp-to-int proposal will likely be enabled by default. These changes are currently not made with strict criteria in mind (e.g. N engines must have this implemented for M years), and there may be breakage that happens.

If you're using a WebAssembly engine that does not support the modules emitted by Rust 1.82 beta and LLVM 19 then your options are:

  • Try seeing if the engine you're using has any updates available to it. You might be using an older version which didn't support a feature but a newer version supports the feature.
  • Open an issue to raise awareness that a change is causing breakage. This could either be done on your engine's repository, the Rust repository, or the WebAssembly tool-conventions repository. It's recommended to first search to confirm there isn't already an open issue though.
  • Recompile your code with features disabled, more on this in the next section.

The general assumption behind enabling new features by default is that it's a relatively hassle-free operation for end users while bringing performance benefits for everyone (e.g. nontrapping-fp-to-int will make float-to-int conversions more optimal). If updates end up causing hassle it's best to flag that early on so rollout plans can be adjusted if needed.

Disabling on-by-default WebAssembly proposals

For a variety of reasons you might be motivated to disable on-by-default WebAssembly features: for example maybe your engine is difficult to update or doesn't support a new feature. Disabling on-by-default features is unfortunately not the easiest task. It is notably not sufficient to use -Ctarget-features=-sign-ext to disable a feature for just your own project's compilation because the Rust standard library, shipped in precompiled form, is still compiled with the feature enabled.

To disable on-by-default WebAssembly proposal it's required that you use Cargo's -Zbuild-std feature. For example:

$ export RUSTFLAGS=-Ctarget-cpu=mvp
$ cargo +nightly build -Zbuild-std=panic_abort,std --target wasm32-unknown-unknown

This will recompiled the Rust standard library in addition to your own code with the "MVP CPU" which is LLVM's placeholder for all WebAssembly proposals disabled. This will disable sign-ext, reference-types, multi-value, etc.

🏷️ Rust_feed