Add i686-pc-windows-gnullvm target

[jeremyd2019]

Suggestion

This was not added in #1883 due to oversight according to Mateusz Mikuła (@mati865):

Bummer, I totally forgot about it so it's not added to windows-rs at all.

Originally posted by Mateusz Mikuła (@mati865) in msys2/MINGW-packages#20397 (comment)

i686 requires doing the same things as in #1883 and then waiting for the new versions of windows-rs to propagate...

Originally posted by Mateusz Mikuła (@mati865) in msys2/MINGW-packages#20397 (comment)

Haven't tested but using RUSTFLAGS="--cfg windows_raw_dylib" could work without copying the libs.

Originally posted by Mateusz Mikuła (@mati865) in msys2/MINGW-packages#20397 (comment)

I feel bad about the proliferation of very similar targets though ☹️ Maybe at some point windows_raw_dylib will be on by default, and we can just say i686-pc-windows-gnullvm is only usable on rust versions that support that?

[jeremyd2019]

Specifically, I used the libraries from i686-pc-windows-gnu to unblock my build, but apparently one could also use the msvc libraries instead (I know that's what we did for aarch64 before support was added here, as there was and still is no aarch64-pc-windows-gnu). llvm is compatible with either format.

[riverar]

Last I looked into this, i686 required mingw packages that are configured/built to use DWARF-2 exception handling (instead of SJLJ). Not all hosts (e.g. macOS) at the time had packages that supported this and working with the fragmented ecosystem required a few hoop jumps. I haven't checked recently though.

Regarding libs: I'm not a fan of the idea of feeding non-llvm-generated libs to llvm tooling, feels like a ticking time bomb. It's especially unattractive when we could just use the llvm tooling as intended. (#1961 (comment), #1846 (review)). Happy to reconsider / hear everyone else's thoughts though.

[jeremyd2019]

Last I looked into this, i686 required mingw packages that are configured/built to use DWARF-2 exception handling (instead of SJLJ).

Hmm, I wonder why that would be. This crate just ships import libraries, right? Those shouldn't have any stack unwinding, or really any code to speak of, with the possible exception of the import thunks, which are a single instruction, of the form:

foo:
    jmp DWORD PTR [__imp_foo]

Even these wouldn't be present in the variant of import libraries used by LLVM or MSVC (so-called "short" import libaries), IIRC.

[riverar]

Right, the mingw issue is unrelated to libs. But related to the addition of the i686 target.

[jeremyd2019]

OK. I'm not aware of an llvm-mingw variant that uses any different exception handling mechanism, but perhaps Martin Storsjö (@mstorsjo) would know if this is something to be concerned about for -gnullvm as it apparently was for -gnu (his llvm-mingw and the CLANG32 MSYSTEM of msys2 are the toolchains I am familiar with and neither use a different exception handling configuration)

[riverar]

Oh hm, re-reviewing our cross compile tests (https://github.com/microsoft/windows-rs/blob/master/.github/workflows/cross.yml#L15-L34), I think I'm conflating gnu and gnullvm. Let's try it and see what blows up!

[jeremyd2019]

Let's try it and see what blows up!

Is that an offer or an invitation for me to make a PR? Cause I could probably figure it out just by grepping for other gnullvm references, but I don't really know what I'm doing 😁

[jeremyd2019]

I added i686-pc-windows-gnullvm to cross.yml in my fork, and it oddly didn't fail. I expected to get the same error I see trying to build rustc itself for that target, that -lwindows.0.x.y is not found.

Oh, crates/libs/targets/Cargo.toml does not have a not(target_abi = "llvm") yet, and nightly supports those target_abi conditions, so it went ahead and worked there. I guess those conditions still don't work in stable.

[mstorsjo]

Last I looked into this, i686 required mingw packages that are configured/built to use DWARF-2 exception handling (instead of SJLJ). Not all hosts (e.g. macOS) at the time had packages that supported this and working with the fragmented ecosystem required a few hoop jumps. I haven't checked recently though.

For GCC based mingw toolchains, many of them have used SJLJ for i686 for quite some time. IIRC the debian provided ones were using SJLJ until "quite recently": https://salsa.debian.org/mingw-w64-team/gcc-mingw-w64/-/commit/8da432fefb792fbe2f9fec2f15278d02e86e6281

Note how one of the reasons seems to be "to allow the Rust toolchain to provide 32-bit cross-compilers".

So for longer term distributions, I think e.g. Ubuntu LTS switched from SJLJ to Dwarf in this configuration since 22.04.

OK. I'm not aware of an llvm-mingw variant that uses any different exception handling mechanism, but perhaps Martin Storsjö (@mstorsjo) would know if this is something to be concerned about for -gnullvm as it apparently was for -gnu (his llvm-mingw and the CLANG32 MSYSTEM of msys2 are the toolchains I am familiar with and neither use a different exception handling configuration)

FWIW, we did use SJLJ for i686 for llvm-mingw earlier as well. Not for matching any other specific toolchain (we're diverging from many existing ones by using UCRT anyway), but because the Dwarf generation for Windows had a couple of bugs earlier.

First I switched from SJLJ to dwarf in 2018 after getting one bug fixed: mstorsjo/llvm-mingw@ca43ab5
Then I had to revert back to SJLJ a couple months later, after a report with another issue with dwarf: mstorsjo/llvm-mingw@f0aa7c4
Then in 2019 I was able to switch back to dwarf after getting that resolved: mstorsjo/llvm-mingw@f9ce177

I guess these qualify as ancient history at this point though :-)

[mati865]

Oh, crates/libs/targets/Cargo.toml does not have a not(target_abi = "llvm") yet, and nightly supports those target_abi conditions, so it went ahead and worked there. I guess those conditions still don't work in stable.

Yeah, target_abi will become stable starting with 1.78 but additional changes might be required for Cargo.
Some context is available here: #2515