LLVM assertion failure on Wasm exceptions

[purplesyringa]

Code

fn main() {}

:/

Meta

I'm building rustc on the current master (bcd0683) with LLVM assertions enabled. I'm compiling the program above like this:

RUSTFLAGS="-C panic=unwind" cargo +stage1 build --target wasm32-wasip1 -Z build-std

I'm using the 23.0 WASI sysroot, downloaded from their GitHub releases, if that matters.

Error output

rustc: /checkout/src/llvm-project/llvm/lib/CodeGen/SelectionDAG/SelectionDAGISel.cpp:1378: void mapWasmLandingPadIndex(MachineBasicBlock *, const CatchPadInst *): Assertion `IntrFound && "wasm.landingpad.index intrinsic not found!"' failed.

like, everywhere, on all sorts of standard crates like libc and unwind.

Disabling assertions hides the bug, and user code works fine. This also does not reproduce on rustup's nightly, probably for the same reason. I have no idea, really. Please tell me I'm doing something wrong.

For context, this seems more important because Emscripten is supposed to switch to Wasm exceptions soon-ish, and I imagine the same problem occurs there (but I haven't checked). Unlike wasm32-wasip1, unwinding is enabled there by default.

@rustbot label +A-LLVM

[workingjubilee]

cc @juntyr @hoodmane

[purplesyringa]

I also get ud2 inside LLVM even with assertions disabled on some complex code that directly uses core::intrinsics::catch_unwind. The code worked perfectly fine on nightly, it works just fine on master on x86-64, and it was mostly copied from std anyway, so I doubt (but will check twice) this isn't my bug. It seems to me like something has gone very wrong with Wasm unwinding in rustc at some point. Any pointers are appreciated.

[purplesyringa]

Downgrading my local repo to 419b3e2, which is what my nightly is based on, does not remove the assertion failures. So perhaps the only reason nightly is working is that assertions are disabled by default.

I have not yet confirmed that disabling the assertions makes the complex code (to anyone who wants to play around with this, run LITHIUM_BACKEND=itanium ci/cargo-wasi test --release --target wasm32-wasip1 in https://github.com/iex-rs/lithium/tree/dev; add +stage1 to the cargo-wasi script beforehand) compile and work on this older commit. This is a sanity check to make sure it's not just my machine that's affected. I expect that it will compile and work, indicating that the cause of the ud2 is just #135484. Perhaps changes in LLVM caused the broken IR to lead to even more brokenness?

[workingjubilee]

I could be mistaken but I don't believe we necessarily support enabling unwinding on targets that are abort-by-default?

[workingjubilee]

Do these problems actually surface for the emscripten target on nightly if you use...

hm, we didn't document it...? we should document it here, probably... https://doc.rust-lang.org/nightly/rustc/platform-support/wasm32-unknown-emscripten.html

@purplesyringa on nightly, with

--target wasm32-unknown-emscripten \
-Zbuild-std \
-Zemscripten-wasm-eh

[purplesyringa]

I have not yet confirmed that disabling the assertions makes the complex code compile and work on this older commit.

So: rustc does compile it, but it miscompiles the code. Which is a known issue (#132416) I'm trying to fix, and applying the fix leads to the ud2 all over again. :/ No idea what's going on yet.