Check --emit flag presence

[2dav]

Miri hopelessly fails with no MIR available for DefId(2:9560 ~ core[7538]::fmt::num::DEC_DIGITS_LUT) when --emit rustc flag is set.

RUSTFLAGS="--emit=llvm-ir" cargo miri run

with the help from discord community it figures out to be system global RUSTFLAGS="--emit=llvm-ir" what cause it, took two days.
Since miri doesn't play with --emit flag at all, its presence could be checked at the start to give reasonable error message in case.

[RalfJung]

Thanks for the report!

Interesting, I wonder why that flag causes these problems. We do explicitly unset RUSTFLAGS when building the sysroot and yet the error looks like the sysroot build went wrong so something is odd here.

Do regular builds still work? I kind of assumed cargo would get mad at you for messing with which artifacts rustc is producing.

[RalfJung]

Okay the error even happens when using a previously built sysroot, so it's not the sysroot that is going wrong.

When I try this is test-cargo-miri, I get

error: internal compiler error: compiler/rustc_monomorphize/src/collector.rs:971:9: no MIR available for DefId(1:10229 ~ std[8367]::panicking::HOOK_LOCK)
[...]
query stack during panic:
#0 [collect_and_partition_mono_items] collect_and_partition_mono_items
end of query stack
error: could not compile `byteorder`

@oli-obk any idea why setting --emit during the crate build means MIR from the sysroot cannot be found any more, but only when going through Miri... (which always has all MIR)? Why would it even ask for the MIR of a static?

Or maybe the error is lying and the actual problem is that this is a check-only sysroot, i.e. without LLVM bitcode -- so trying to emit specifically LLVM IR in that sysroot fails? That would make sense, but the error complains about missing MIR, not missing bitcode.

[bjorn3]

The ICE happens at https://github.com/rust-lang/rust/blob/100f12d17026fccfc5d80527b5976dd66b228b13/compiler/rustc_monomorphize/src/collector.rs#L971 I think normally is_reachable_non_generic would return true and thus should_codegen_locally would return false. Because is_reachable_non_generic doesn't return true, it checks that MIR is available. is_reachable_non_generic returns false because we have a shortcut at https://github.com/rust-lang/rust/blob/45e2c2881d11324d610815bfff097e25c412199e/compiler/rustc_codegen_ssa/src/back/symbol_export.rs#L49 where for crates that don't get codegend reachable_non_generics_provider returns an empty map. This then ends up in the crate metadata.

The fix for this would be to early error when trying to compile with sess.opts.output_types.should_codegen() returning true, while it returned false for any dependency.

[RalfJung]

The fix for this would be to early error when trying to compile with sess.opts.output_types.should_codegen() returning true, while it returned false for any dependency.

This would have to happen in rustc, not Miri, right?

[bjorn3]

Yes