make: fix linking

[kaspar030]

Ever wondered why exactly files needed to be put into "UNDEF" in order to link correctly?

When considering archives, ld only considers objects in archives if there is at least one symbol in them needed elsewhere. Also, it doesn't seem to consider a symbol if it already found a weak symbol before.

In the current make system, this leads to:

• weak symbols not overridden by strong symbols unless the object is passed to the linker outside of an archive (in "UNDEF")
• weak symbols don't get overridden unless at least one symbol of that object is used somewhere (this is why overridden ISRs are only actually used if some other function of that file is used, e.g., isr_timer0 will only be used instead of the weak-default dummy handler if e.g., timer_init() is called)

Both feel hacky to me.

This PR encloses processing of module-archives in "--whole-archive", causing all objects to be always considered for linking. That way, the UNDEF-objects are not needed anymore.

Unfortunately this always overrides all defined isr defined in periph_*, as they are currently all always built. So to not get somewhat bloated binaries, this depends on #5065 #7241.

Also, the config module breaks, so #5756 is also required. (#5756 merged)

[Kijewski]

Ever wondered why exactly files needed to be put into "UNDEF" in order to link correctly?

Dunno if that's still true, but once upon a time (probably in 2014) GCC considered all the interrupt handlers as unused, and removed them. Turns out boards are rather boring if you don't have any external interrupts. Since then there was a lot of work to make -fLTO and --gc-sections work, so probably this problem was fixed. too.

[jnohlgard]

@Kijewski The issue with missing ISRs or interrupt vectors is solved by marking the functions with __attribute__((used)). Like you wrote, it should be fixed with the various LTO improvements lately.

[haukepetersen]

would you mind rebasing, as #5756 is merged by now? Thanx

I really like this PR! If the only thing is now to wait for #5065, I again call for help fixing the last Makefile.buildtest issue in that PR so we can merge this one also.

[kaspar030]

• rebased

[miri64]

@kaspar030 what makes this still WIP?

[kaspar030]

@miri64 I need to re-check that this is still working as expected once #5065 is in.

[miri64]

Feature freeze -> postponed

[jnohlgard]

@cladmi I created a proof of concept for switching to using ld -r instead of ar for the modules. Will post a PR later this week.

[cladmi]

@gebart Cool :) Using ld -r for modules could also help if trying to dynamically load elf objects on a node and I could be interested in that.

The work done by @kaspar030 will help for the modules that would break if not handled as libraries. Do you remember if other modules required to be moved to EXTRA_LIBS ?

[jnohlgard]

@cladmi see #8711 for a proof of concept

[kaspar030]

Closing in favour of #8711.