Rustc strips all symbols on MacOS when strip = "debuginfo" is specified, but not when strip = "symbols" is specified

[orlp]

Consider this basic example:

fn main() {
    std::thread::sleep(std::time::Duration::from_secs_f64(1.0));
}

When compiled with cargo build and the following Cargo.toml:

[profile.dev]
debug = false
strip = "symbols"

we see the following profile with samply record:

That is, all symbols are still there. However, when we instead build with

[profile.dev]
debug = false
strip = "debuginfo"

we see the following profile:

I believe this is a bug, as the documentation for strip specifies that symbols is supposed to be a more aggressive stripping than debuginfo, which is supposed to leave backtrace information mostly intact. We see the opposite behavior.

---

The following lines of code are suspect:

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

Lines 1103 to 1118 in bf6f8a4

	if sess.target.is_like_osx {
	let stripcmd = "rust-objcopy";
	match (strip, crate_type) {
	(Strip::Debuginfo, _) => {
	strip_symbols_with_external_utility(sess, stripcmd, out_filename, &["-S"])
	}
	// Per the manpage, `-x` is the maximum safe strip level for dynamic libraries. (#93988)
	(Strip::Symbols, CrateType::Dylib | CrateType::Cdylib | CrateType::ProcMacro) => {
	strip_symbols_with_external_utility(sess, stripcmd, out_filename, &["-x"])
	}
	(Strip::Symbols, _) => {
	strip_symbols_with_external_utility(sess, stripcmd, out_filename, &[])
	}
	(Strip::None, _) => {}
	}
	}

-S on rust-objcopy is documented as such:

  -S                      Alias for --strip-all

It does not seem appropriate to be used in the Debuginfo branch, but omitted in the Symbols branch.

[orlp]

What made it extra annoying to debug this issue is that the cargo docs lie about the default for strip. It says strip = "none" is the default, but this is not true. The default depends on whether or not debug = false is specified.

[orlp]

This is a regression from stable to nightly, tracked down to #131405.

[apiraino]

@rustbot label +T-bootstrap

cc @davidtwco @jieyouxu since #131405

[jieyouxu]

The regression is actually a T-compiler concern, because we used the wrong flags for rust-objcopy in cg_ssa.