Undefined weak symbols

[ilg-ul]

On Linux, if a weak symbol is not defined but used in an if statement, it returns zero, and so the program knows at runtime that the symbol is not defined.

On macOS, according to my tests, the Apple linker does not accept undefined weak symbols.

Can you confirm that this is the normal behaviour of the current Apple linker?

Is there any linker option that could change this behaviour? (I could not find anything related)

Thank you!

[iains]

Default ELF linker semantics are different from default Mach-O linker semantics. ELF allows any undefined symbols, Mach-O does not (note that there are ELF linker options to make it reject missing symbols and, as I am going to write below, options to allow missing symbols on Mach-O).

In both cases, weak referenced symbols do operate at runtime in the same way (if the symbol is not present in the dynamically-linked exe, it will be 0).

For Mach-O, using current linkers, you have two ways to deal with this;

1. you tell the linker that it's OK for a set of symbols to be undefined at static-link time -U,_symbol .. or (if there are many) by providing the list in a file [ man ld ].
2. you tell the linker that it's OK for any symbol to be undefined at static link time -undefined dynamic_lookup

the second is a large sledgehammer and will not only allow the symbols you want to be undefined, but any others that get accidentally misspelled etc.

There are times when the second is useful (mostly in implementing plugins or when too lazy to specify the correct set).

additional: dynamic lookup is incompatible with some relocation mechanisms used on Arm64 (according to linker warnings) .. so perhaps it is wise to consider mechanisms to avoid that sledgehammer...

[iains]

In the embedded world where I come from, weak is used in some frameworks to define default functions which can be reimplemented by the application. This creates a lot of issues since such code cannot be modularised easily, and the whole build must be monolithic.

Design of frameworks and/or applications is outside the scope of the "implementation" if the toolchain &c. The toolchain provides facilities - it's up to the systems' engineers to decide how to use those.

It is part of a suite of tests, here the interesting part is the missing func, which on Linux returns 0, while on macOS fails to link.

Which is the question that is answered above; this is only a consequence of differing defaults for the static linker between (most) ELF and Mach-O based systems, it makes no difference to the actual runtime behaviour.

[iains]

In the embedded world where I come from, weak is used in some frameworks to define default functions which can be reimplemented by the application. This creates a lot of issues since such code cannot be modularised easily, and the whole build must be monolithic.

Although - as you describe it here - this is the "replaceable function" case (3) not the "maybe provided, maybe not" (1) - although the same functionality could probably be designed to use the "maybe provided" mechanism - in which case the build would not need to be monolithic.

... but now we're digressing :)
... the answer remains "macOS can do the same as Linux, you just need to provide the relevant information to the static linker". There are tests in the GCC testsuite that do exactly this.

[ilg-ul]

The point here is neither the design of embedded frameworks (which are a mess) nor the need for monolithic builds (which is the usual approach for embedded).

The question is related to this specific test, does it make any sense to enable it for GCC on macOS? Is it of any use? (for me it looks like a nuisance, since I had to explicitly mark it as expected to fail).

[iains]

The question is related to this specific test, does it make any sense to enable it for GCC on macOS? Is it of any use? (for me it looks like a nuisance, since I had to explicitly mark it as expected to fail).

I'm not sure why it is not appropriate to fix the test to work as intended? - all you need to do is to add -Wl,-U,_symbolname to the link flags - for a single missing symbol.

[ilg-ul]

all you need to do is to add -Wl,-U,_symbolname

That was useful, I added -Wl,-U,_func and the test passed.

Thank you, I think the question was answered.