Use -windows-gnu for all UEFI targets

[nicholasbishop]

This restores the behavior prior to #1252, when the UEFI targets were hardcoded in src/lib.rs.

[NobodyXu]

cc @madsmtm can you take a look at this please?

[madsmtm]

Huh, I seemed to remember explicitly verifying that the code did the same thing, but I must have missed this, sorry about that :/

I'm not against doing this fix in cc if that's where it's needed, and the code looks fine. That said, I would really like some documentation on it!

I found this code comment in the rustc sources talking a bit about why -gnu is necessary, but I would expect a comment in similar regard in cc, covering at least:

• Why cc differs from rustc here (more linking failures?).
• Reasoning why it is sound for cc to use a different target than rustc (are there calling convention concerns?).
• A link to an LLVM issue or something about this, where we could work on getting proper support for Windows UEFI targets.

[nicholasbishop]

That makes sense. Unfortunately I don't have answers to most of those questions yet. The specific motivation for this PR was trying to build https://github.com/Amanieu/minicov for x86_64-unknown-uefi; that used to work, but stopped working with cc-v1.1.32. When the LLVM target is x86_64-unknown-windows, I get an error from a missing malloc.h.

Possibly the Rust UEFI aarch64/x86_64 targets should be switched to -windows-gnu, as i686 already is.

[madsmtm]

Fair enough, then perhaps a comment that we don't know why this hack is needed, and link to this PR ;)

Alternatively, if you're feeling venturous, I found https://krinkinmu.github.io/2020/10/11/efi-getting-started.html#bulding, which seems to suggest that the actual fix is to pass the -ffreestanding flag?

[madsmtm]

Actually, seems like LLVM should have support for UEFI nowadays: https://discourse.llvm.org/t/rfc-uefi-driver-support-uefi-target/73261.

So maybe pass those targets (*-unknown-uefi) to Clang instead? (Though I wouldn't block a quickfix on that).

[madsmtm]

Opened rust-lang/rust#132570

[nicholasbishop]

Thanks for opening that PR! I wasn't aware clang had added these targets.

I've updated this PR to add a comment describing the history, the current state, and a TODO for the future.

Also tweaked the way the override is set, so that it won't immediately change to follow rustc if the rustc targets are changed, since I think we would want to be intentional about changing cc-rs to match.

[madsmtm]

Looks good to me now, thanks for writing things out!