fix(compiler/rustc_target): set correct linker flags for wasm32v1-none

[StackOverflowExcept1on]

Resolves #145491

r? @alexcrichton

https://github.com/llvm/llvm-project/blob/llvmorg-20.1.8/lld/wasm/Writer.cpp#L573-L691

https://releases.llvm.org/20.1.0/docs/ReleaseNotes.html#changes-to-the-webassembly-backend

[rustbot]

These commits modify compiler targets.
(See the Target Tier Policy.)

[bjorn3]

Doing it at the place I indicated (the WasmLd implementation) has the advantage that it works for all wasm targets and also when the user passes -Ctarget-cpu to override the target default.

[StackOverflowExcept1on]

Yes, that's probably good idea, but we need WasmTM->CPU to be set from the command line.

[bjorn3]

wasm-ld can't notice any difference between setting this cli arg in the target spec and setting it in the WasmLd implementation I linked to.

[StackOverflowExcept1on]

Yes, you're right. I checked in the debugger and it's also worth adding --mllvm=-mattr=+mutable-globals to change WasmTM->TargetFS (features set).

[bjorn3]

This still doesn't put the --mllvm=-mcpu=... in compiler/rustc_codegen_ssa/src/back/linker.rs.

[StackOverflowExcept1on]

Can you show demo code with RUSTC_LOG=debug where -C target-feature won't work? It seemed ok to me. The only problem is that with an empty file in the target_features section the features enabled by the user won't be displayed (since .bc files don't save features).

[StackOverflowExcept1on]

Perhaps you mean something like -C target-features=-mutable-globals and -C target-cpu=generic, but I don't understand why user would want to do that.

[bjorn3]

For wasm32v1-none it is maybe not all that relevant given that you probably don't want to enable any extra features, but for wasm32-unknown-unknown or wasm32-wasip1/2 if I use for example -Clinker-plugin-lto -Ctarget-features=+simd128 I would expect all functions to be compiled with simd128 enabled (including standard library functions) even if the leaf crate is empty (and thus nothing otherwise forces simd128 to be enabled). Supporting this requires passing --mllvm=-mcpu= and --mllvm=-mattr= in

rust/compiler/rustc_codegen_ssa/src/back/linker.rs

Line 1480 in b779120

fn push_linker_plugin_lto_args(&mut self) {

, at which point wasm32v1-none should behave correctly too without having to change the target spec afaik.

[StackOverflowExcept1on]

So, you are suggesting to explicitly pass cpu and features to wasm-ld in push_linker_plugin_lto_args depending on what the user is using?

[bjorn3]

Yes, with a fallback to whatever target cpu and features is in the target spec. This is also what we do for gcc for example:

rust/compiler/rustc_codegen_ssa/src/back/linker.rs

Line 426 in b779120

&format!("-plugin-opt=mcpu={}", self.target_cpu),

Actually maybe getting the target features right would be non-trivial as that isn't plumbed through yet and not used for gcc either. So maybe only set the target cpu in linker.rs and just add a FIXME for the target features (and if you do need to set the target features just do it in the target spec)

[alexcrichton]

r? @bjorn3

I think you may understand the intricacies here better than I