error: failed to parse bitcode for LTO module: Invalid cast

[connerebbinghaus]

Compiling the following in release on the recent nightlies causes a crash. Note that compiling in debug succeeds.

Code

#![feature(portable_simd)]
use std::simd::*;
use std::arch::x86_64::*;

#[target_feature(enable = "sse4.1")]
pub unsafe fn fast_round_sse(i: f32x8) -> f32x8 {
    let a = i.to_array();
    let [low, high]: [[f32; 4]; 2] =
        unsafe { std::mem::transmute::<[f32; 8], [[f32; 4]; 2]>(a) };
        
    let low = f32x4::from(_mm_round_ps::<{_MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC}>(f32x4::from_array(low).into()));
    let high = f32x4::from(_mm_round_ps::<{_MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC}>(f32x4::from_array(high).into()));

    let a: [f32; 8] =
        unsafe { std::mem::transmute::<[[f32; 4]; 2], [f32; 8]>([low.to_array(), high.to_array()]) };
    f32x8::from_array(a)
}

//#[target_feature(enable = "sse4.1")]
//pub unsafe fn round_many(b: &mut [f32x8]) {
//    for a in b {
//        *a = fast_round_sse(*a);
//    }
//}

Meta

rustc --version --verbose:

rustc 1.71.0-nightly (fec9adcdb 2023-04-21)
binary: rustc
commit-hash: fec9adcdbc21469ef105162cc8cabf81c72d06be
commit-date: 2023-04-21
host: x86_64-unknown-linux-gnu
release: 1.71.0-nightly
LLVM version: 16.0.2

Error output

error: failed to parse bitcode for LTO module: Invalid cast (Producer: 'LLVM16.0.2-rust-1.71.0-nightly' Reader: 'LLVM 16.0.2-rust-1.71.0-nightly')

error: could not compile `ice_repro` (lib) due to previous error

If the round_many function is uncommented, rustc instead crashes with an illegal instruction

error: could not compile `ice_repro` (lib)

Caused by:
  process didn't exit successfully: `rustc --crate-name ice_repro --edition=2021 src/lib.rs --error-format=json --json=diagnostic-rendered-ansi,artifacts,future-incompat --diagnostic-width=236 --crate-type lib --emit=dep-info,metadata,link -C opt-level=3 -C embed-bitcode=no -C metadata=720812225537f937 -C extra-filename=-720812225537f937 --out-dir /home/conner/ice_repro/target/release/deps -L dependency=/home/conner/ice_repro/target/release/deps` (signal: 4, SIGILL: illegal instruction)

Bisect

searched nightlies: from nightly-2023-04-17 to nightly-2023-04-23
regressed nightly: nightly-2023-04-18
searched commit range: d0f204e...7908a1d
regressed commit: 5546cb6

bisected with cargo-bisect-rustc v0.6.6

Host triple: x86_64-unknown-linux-gnu
Reproduce with:

cargo bisect-rustc -- build --release 

[saethlin]

This bisects to a MIR inlining PR, but the actual bug here is related to lowering of repr(simd). It's possible to write code that gets lowered to invalid LLVM IR without MIR inlining, but the code required is rather contrived.

This has been extensively discussed on the Zulip: https://rust-lang.zulipchat.com/#narrow/stream/257879-project-portable-simd/topic/splat.20no.20longer.20compiles.20for.20release.20builds I'm surprised nobody else opened an issue for this. Thanks!