LLVM error on High Sierra

[fmckeogh]

Building cortex-m-rt on nightly, beta and stable in Mac OS High Sierra (tested on two machines) results in the following error:

LLVM ERROR: Global variable '__RESET_VECTOR' has an invalid section specifier '.vector_table.reset_vector': mach-o section specifier requires a segment whose length is between 1 and 16 characters.

It build correctly on my Ubuntu 18.04 machine.

[japaric]

mach-o section specifier requires a segment whose length is between 1 and 16 characters

Are you compiling the cortex-m-rt crate for the host (for mac)?

If that's the case we can drop the link_section when compiling for mac; trying to link a cortex-m program compiled for the host (mac / Linux) is going to fail anyway.

[fmckeogh]

I'm so sorry for wasting your time, I'm the biggest idiot. I forgot to set the default target in .cargo/config.

[jonas-schievink]

The cortex-m crate goes out of its way to build at least on Linux, would be nice if all the core crates consistently built on Linux, macOS, and Windows

[japaric]

we ran into a similar issue in the defmt project.
we found out that on macOS the longest linker section name can be .section,0123456789ABCDEF (note: comma after .section not period and a max of 16 characters after the comma)
this issue can probably be fixed by adding a cfg_attr(is_macos, link_section = "..") to the relevant static variables

[therealprof]

Indeed, I've encountered it frequently.

[berkus]

What does this fix do? It replaces section name with exactly the same but taking half a kilobyte more space in source code; am I missing something there?

[therealprof]

@berkus Not sure what you mean. What are you seeing and on which system? This is changing the section names for macOS only because it has a weird length restriction. It should not affect anything except to prevent linker errors on macOS.

[berkus]

@therealprof ok, lets go with the way of example, shall we.

#[cfg_attr(target_os = "macos", link_section = ".vector_table,interrupts")]
#[cfg_attr(not(target_os = "macos"), link_section = ".vector_table.interrupts")]

On macOS the section name would be .vector_table.interrupts, however on any other system it would be .vector_table.interrupts which is EXACTLY THE SAME.

So now you've reduced readability, understandability and maintainability of the code to provide exactly the same section names for every system out there.

What am I missing? I could have missed something, to be precise am looking at this commit.

Here it is, condensed:

#[cfg_attr(target_os = "macos", link_section = ".HardFault,user")]
#[cfg_attr(not(target_os = "macos"), link_section = ".HardFault.user")] // NB: same name

#[cfg_attr(target_os = "macos", link_section = ".vector_table,reset_vector")]
#[cfg_attr(not(target_os = "macos"), link_section = ".vector_table.reset_vector")] // NB: same name

#[cfg_attr(target_os = "macos", link_section = ".HardFault,default")]
#[cfg_attr(not(target_os = "macos"), link_section = ".HardFault.default")] // NB: same name

#[cfg_attr(target_os = "macos", link_section = ".vector_table,exceptions")]
#[cfg_attr(not(target_os = "macos"), link_section = ".vector_table.exceptions")] // NB: same name

#[cfg_attr(target_os = "macos", link_section = ".vector_table,interrupts")]
#[cfg_attr(not(target_os = "macos"), link_section = ".vector_table.interrupts")] // NB: same name

[berkus]

Ahha, now I see the comma vs period, my bad!

[berkus]

@therealprof btw, could it be using the comma on all systems, not only macos?

[therealprof]

Not sure sorry, I'm actually curious why this is supposed to work on macOS (NB: I haven't actually tested it myself). Maybe some internal linker magic to substitute the "." back?

[berkus]

@therealprof on mac this is two names, ".vector_table" for the segment and "reset_vector" for the section name.

I will test on linux if a section name with comma works, the ELF file format spec says why not so the only limiting factor would be linker's parser.

[adamgreig]

We could probably just remove the link_section attribute entirely on MacOS, right? It's not going to actually do anything as I understand it, since we can't build final libraries or executables targeting MacOS with cortex-m-rt and it doesn't use our linker script (where the names matter) on MacOS either; this change just lets people on MacOS run cargo check/test without a target specified. I wouldn't want to swap the ARM builds or linker script to using commas for section names just so that Macs can check the code when compiling natively, though.