Tests with `harness = false` aren't run

[alecmocatta]

This in Cargo.toml results in foo.rs's fn main not being invoked by cargo miri test.

[[test]]
name = "foo"
harness = false

[RalfJung]

What does harness = false do? I never saw that.^^

[alecmocatta]

It runs your fn main instead of the libtest harness that would invoke #[test]s.

[RalfJung]

Okay so likely for those we should invoke plain rustc and not add the --test argument?

[alecmocatta]

I don't quite follow how miri invokes rustc, but that sounds right, with --cfg test instead of --test. Though I'd have thought the cargo check invocation should invoke rustc correctly given harness/no harness. I'd need to grok what's going on better...

[RalfJung]

When cargo-miri finds a test target, it builds it with --test:

miri/cargo-miri/bin.rs

Lines 419 to 420 in ca9e988

	(MiriCommand::Test, "test") => {
	cmd.arg("--test").arg(target.name);

[alecmocatta]

Ah, that's a different --test to what I thought you meant. That shouldn't need to be changed. Scanning the code, tests with harness = false should just work! Not sure why they aren't. I'll try and take a further look later today.

[RalfJung]

AFAIK rustc --test means that the test harness is added to this crate.

Another possible cause of this problem could be how these tests are represented in cargo metadata.

[alecmocatta]

As I understand it, that line is passing --test test_name to cargo, which will then invoke the wrapped rustc with --test or --cfg test depending on harness/no harness, which will then be forwarded to rustc. So that's being dealt with correctly as far as I can tell. A couple of things missing (that aren't the cause of this issue) are passing TESTNAME (as in cargo test TESTNAME) and --bench (if invoked via cargo bench) to the interpreted binary.