Allow passing Mach-O flags to set_segment_section

[madsmtm]

This enables rustc_codegen_cranelift to control these, see rust-lang/rustc_codegen_cranelift#1648.

[madsmtm]

Another design for this could be to take:

segment: String,
section: String,
macho_section_type: String,
macho_section_attributes: Vec<String>,

And then have the "pure_instructions" -> S_ATTR_PURE_INSTRUCTIONS mappings in cranelift-object? What would you prefer?

[bjorn3]

Maybe directly use object::write::SectionFlags and then or those flags with the ones cranelift-object would set by default?

[madsmtm]

But that would introduce a dependency on object to cranelift-module?

[bjorn3]

Ah, right this is cranelift-module, not cranelift-object.

[madsmtm]

A third alternative would be:

custom_segment_section: Option<String>,

And do the entire parsing inside cranelift-object.

[cfallin]

Looks reasonable to me -- thanks!

[bjorn3]

Actually shouldn't this be |=? As is this would overwrite the default flags even when no flags are passed.

[madsmtm]

I considered that, but there are no default flags for (SectionKind::Data, SectionKind::ReadOnlyData and SectionKind::ReadOnlyDataWithRel), and if there were, it would break if we tried to |= the section type.

[bjorn3]

Would skipping the flag set if macho_flags is 0 makes sense?

[madsmtm]

Maybe we could do:

let section_type = *macho_flags & SECTION_TYPE;
let section_attributes = *macho_flags & SECTION_ATTRIBUTES;
let current_attributes = (*flags & SECTION_ATTRIBUTES);
*flags = section_type | current_attributes | section_attributes;

But I feared that might be confusing too, if I request only no_dead_strip, it might be weird if I got other attributes too.

Though we should probably have an assert that *flags is 0.

[madsmtm]

I opened #13153 for adding a debug assertion instead