add support for mod_init_funcs (was compat: rust-ctor fails to trigger with Cranelift on MacOS, but works with LLVM)

[stewartadam]

I noticed when switching to the cranelift backend that my typetag types stopped working:
dtolnay/typetag#35

I'm not sure what exactly is to blame, but figured it warranted reporting as a compat issue. Through some subsequent debugging, I've identified noticed an issue specific to Cranelift on MacOS where the rust-ctor crate to fails to trigger #[ctor]-annotated code on module initialization: mmastrac/rust-ctor#339

Other thoughts:

• There was some discussion in Linking to a static library does not always include exported symbols rust#99721 that it could be due to issues with rustc and static linking, but given this is cranelift-specific (switching to LLVM causes typetag and rust-ctor to work as expected) I'm skeptical.
• The inventory crate uses a technique similar to ctor so I bet Inventory crate does not work with cranelift #1577 is also bit by the same root cause affecting this issue.

[bjorn3]

Looks like on Mach-O constructor sections should get the S_MOD_INIT_FUNC_POINTERS flag. It seems like LLVM sets this flag for __DATA,__mod_init_func even if you miss the mod_init_funcs flag specified in the #[link_section]. I suspect this is a side-effect of https://github.com/llvm/llvm-project/blob/2e8e254d18f51b6ca898bf0b1e4d12109b5b16c7/llvm/lib/CodeGen/TargetLoweringObjectFileImpl.cpp#L1292-L1304 combined with merging the flags of identically named sections. The rust-ctor crate should probably use __DATA,__mod_init_func,mod_init_funcs as #[link_section]. That alone won't fix this issue though as handling for mod_init_funcs hasn't been added at

rustc_codegen_cranelift/src/constant.rs

Lines 396 to 433 in 87e71fe

	let (segment_name, section_name) = if tcx.sess.target.is_like_darwin {
	// See https://github.com/llvm/llvm-project/blob/main/llvm/lib/MC/MCSectionMachO.cpp
	let mut parts = section_name.as_str().split(',');
	let Some(segment_name) = parts.next() else {
	tcx.dcx().fatal(format!(
	"#[link_section = \"{}\"] is not valid for macos target: must be segment and section separated by comma",
	section_name
	));
	};
	let Some(section_name) = parts.next() else {
	tcx.dcx().fatal(format!(
	"#[link_section = \"{}\"] is not valid for macos target: must be segment and section separated by comma",
	section_name
	));
	};
	if section_name.len() > 16 {
	tcx.dcx().fatal(format!(
	"#[link_section = \"{}\"] is not valid for macos target: section name bigger than 16 bytes",
	section_name
	));
	}
	let section_type = parts.next().unwrap_or("regular");
	if section_type != "regular" && section_type != "cstring_literals" {
	tcx.dcx().fatal(format!(
	"#[link_section = \"{}\"] is not supported: unsupported section type {}",
	section_name, section_type,
	));
	}
	let _attrs = parts.next();
	if parts.next().is_some() {
	tcx.dcx().fatal(format!(
	"#[link_section = \"{}\"] is not valid for macos target: too many components",
	section_name
	));
	}
	// FIXME(bytecodealliance/wasmtime#8901) set S_CSTRING_LITERALS section type when
	// cstring_literals is specified
	(segment_name, section_name)

yet. This also needs a Cranelift change to be able to set this flag in the object file in the first place.

[mmastrac]

ctor now emits the full flag (fixed in mmastrac/rust-ctor#344).