WIP: Initial changes to support macOS build

gushromp commented 2 years ago

What works:

Building most things on x86_64
Building itself (make compare)
Using default GC and making debug builds (AFAIK, arc/orc don't work on Linux yet either?)
Some release builds (release build of nlvm self-build works fine)

What doesn't:

M1/arm64 yet, though I hope this is just because of ABI incompatibilities in the posix module which I'll try to iron out. In the meantime you can run in x86_64 via Rosetta just fine.
Release builds crash with strange segfault on some things I've tested (i.e. unittest when it tries to initialize itself):
```
proc ensureInitialized() =
if formatters.len == 0: # <- segfault happens here
  formatters = @[OutputFormatter(defaultConsoleFormatter())]
```
I'm still trying to investigate and hopefully fix this issue.

Notes:

I've temporarily changed the Nim submodule URL to my own fork, to make this branch work. If/when we merge this PR, we should revert it :)
If you run into a segfault or weird behavior/crashes, it's probably because I was too lazy to go through the entire posix module and change every imported struct definition by hand. It should probably be possible to repurpose the detect tool to do this boring work automatically for us.
For some reason, the Mac variant of POSIX constants isn't using the autogenerated values and thus I opted to temporarily define most of those in the ll bitcode. But it would probably be more elegant/performant to remove it from there and just import macosx_consts.nim in the posix module instead.
I hate POSIX.

gushromp commented 2 years ago

get the non-mac changes merged in a separate PR - ie there's a number of nice fixes and improvements here already, like the handling of the C++ flags, versions etc that prepare us for multi-platform support

I'll be glad to do so, but please help me understand the scope of the changes you'd like to see in a separate PR - since most of those are when defined(linux)/when defined(macosx) etc. Does that mean you'd like me to leave the conditional compilation/shell script stuff for linux only in that other PR?

add mac to CI in a matrix build - to begin with, mac should not result in a hard fail (it should continue to show a green checkmark even if mac fails, but it's good to run it at least, to get an idea)

Will do!

negotiate a fix for the constant issue - the important part is that it's autogenerated - then it can live either in upstream or in the .ll file, either is fine

The posix constants fall back to posix_other_consts.nim on Mac by default whereas for Linux it uses the output of detect. I'm not sure why this is the case, maybe @Araq can shed some light on this? Either way I'm more than happy to use detect for both Arm64/X86_64 on Mac and use its output for these constants so we can remove them from the bitcode :)

The other thing to look out for is ABI sizes - this is a general concern: if the size of an object is unknown or wrong, it'll result in hard-to-track issues - maybe that's what your crash is about: some stack allocation that is to small because a type is poorly described.

They're hard to track down in general - compiling with -d:nimAbiCheck (or something similar) can sometimes help.

Yeah this has been the most painful part of getting this to work :) I didn't know about the checkAbi flag, thanks for letting me know, I'll try it out.