Compilation failure on LLVM 19 with `*-gnullvm`

[kleisauke]

The logic at:

cc-rs/src/lib.rs

Lines 1930 to 1935 in dcd8ed3

	if cmd.is_like_clang() && target.contains("windows") {
	// Disambiguate mingw and msvc on Windows. Problem is that
	// depending on the origin clang can default to a mismatchig
	// run-time.
	cmd.push_cc_arg(format!("--target={}", target).into());
	}

(introduced in PR #825)

Breaks with LLVM 19 (rust-lang/rust#127513) when targeting *-pc-windows-gnullvm due to PR llvm/llvm-project#78655.

Details

The following warnings were emitted during compilation:

warning: compiler_builtins@0.1.109: clang: error: version 'llvm' in target triple 'x86_64-pc-windows-gnullvm' is invalid

error: failed to run custom build command for `compiler_builtins v0.1.109`

Caused by:
[...]

  --- stderr


  error occurred: Command "x86_64-w64-mingw32-clang" "-Oz" "--target=x86_64-pc-windows-gnullvm" "-ffunction-sections" "-fdata-sections" "-m64" "--target=x86_64-pc-windows-gnullvm" "-Os" "-g" "-gcodeview" "-fdata-sections" "-ffunction-sections" "-fno-builtin" "-fvisibility=hidden" "-ffreestanding" "-DVISIBILITY_HIDDEN" "-o" "/var/tmp/tmp-librsvg-x86_64-w64-mingw32/librsvg-2.58.92.build_/target/x86_64-pc-windows-gnullvm/release/build/compiler_builtins-f90e4ff268ddda14/out/978315636d0ac771-absvdi2.o" "-c" "/data/mxe/usr/x86_64-pc-linux-gnu/lib/rustlib/src/rust/src/llvm-project/compiler-rt/lib/builtins/absvdi2.c" with args x86_64-w64-mingw32-clang did not execute successfully (status code exit status: 1).

/cc @mati865 FYI.

[NobodyXu]

cc @ChrisDenton Do we want to remove the llvm suffix here?

[kleisauke]

Hmm, it looks like --target=x86_64-pc-windows-gnullvm was added twice to the clang invocation. This patch seems to fix it for me:

Details

--- a/src/lib.rs
+++ b/src/lib.rs
@@ -1939,7 +1939,10 @@ impl Build {
                     cmd.push_opt_unless_duplicate(format!("-O{}", opt_level).into());
                 }
 
-                if cmd.is_like_clang() && target.contains("windows") {
+                if cmd.is_like_clang()
+                    && target.contains("windows")
+                    && !target.ends_with("-gnullvm")
+                {
                     // Disambiguate mingw and msvc on Windows. Problem is that
                     // depending on the origin clang can default to a mismatchig
                     // run-time.
@@ -2190,6 +2193,10 @@ impl Build {
                         }
 
                         cmd.push_cc_arg(format!("--target={}", target).into());
+                    } else if target.ends_with("-gnullvm") {
+                        cmd.push_cc_arg(
+                            format!("--target={}", target.strip_suffix("llvm").unwrap()).into(),
+                        );
                     } else {
                         cmd.push_cc_arg(format!("--target={}", target).into());
                     }
@@ -3467,6 +3474,7 @@ impl Build {
                     "hexagon-unknown-linux-musl" => Some("hexagon-linux-musl"),
                     "i586-unknown-linux-musl" => Some("musl"),
                     "i686-pc-windows-gnu" => Some("i686-w64-mingw32"),
+                    "i686-pc-windows-gnullvm" => Some("i686-w64-mingw32"),
                     "i686-uwp-windows-gnu" => Some("i686-w64-mingw32"),
                     "i686-unknown-linux-gnu" => self.find_working_gnu_prefix(&[
                         "i686-linux-gnu",

(the i686-pc-windows-gnullvm change should probably be split into a separate commit)

[NobodyXu]

Thanks, that looks good, can you open a PR?

I would review it and also ask @ChrisDenton to review it, since they are more familiar with windows than me.

[mati865]

I think using Rust triple is unfortunate. Many other triples won't match LLVM, like Win7 triple if we stick to Windows.

That said I don't know if it's possible to obtain LLVM triple from rustc so adding workarounds for all the triples might be the only way.

[ChrisDenton]

On nightly you can use rustc --print target-spec-json but that's not stable yet. A stable way I can think of is to perhaps infer it from the cfg without looking at the Rust target string at all. E.g. the LLVM target is x86_64-pc-windows-msvc if target_arch="x86_64" and target_os="windows" and target_env="msvc".

I guess a test could compare the cc-rs inferred LLVM targets against the nightly target-spec-json to make sure they match. But that's slightly tricker as targets can appear and disappear so there might not be an exact match for any given target.

[NobodyXu]

cc @kleisauke @mati865 @ChrisDenton if it is available in nightly, then we can pre-generate it like we did for riscv:

cc-rs/dev-tools/gen-target-info/src/main.rs

Line 36 in c2fd8ec

generate_riscv_arch_mapping(&mut f, &target_specs);

gen-target-info is run on CI every week and create a PR if something changes, and it also can be run locally.

The generated info would then be committed and included in crate publish.

[mati865]

Yeah, it's been available in night for over a year. I don't know how gen-target-info works though.

[NobodyXu]

Yeah, it's been available in night for over a year. I don't know how gen-target-info works though.

Basically, it creates a rust source file in src, which is checked into git and included in cargo-publish.

It's usually run in CI per week, or on the developer machine.

If the generated source file differs, then you need to commit and open a PR.

[mati865]

Sounds good but I won't be able to look into it probably until the weekend.

[mati865]

@kleisauke how can I test this issue easily?
I think #1176 should address this issue, but I'd like to test it first.
Cross-compiling librsvg from Linux fails on missing cairo pkgconf files. Is it possible for you to try it?

[kleisauke]

@mati865 Thanks for opening PR #1176! AFAIK, you can reproduce this error when building a crate with -Zbuild-std, as that tries to build compiler-builtins.

Testing this with librsvg is a bit difficult because you have to cross-compile its dependencies as well. Also, you need to disable any raw-dylib usage internally in Rust via these (now removed) patches:
compiler/rustc_codegen_cranelift/patches/0030-stdlib-Revert-use-raw-dylib-for-Windows-futex-APIs.patch
compiler/rustc_codegen_cranelift/patches/0029-stdlib-rawdylib-processprng.patch

Otherwise, you'll encounter these errors:
https://github.com/libvips/build-win64-mxe/blob/896ccdbf57ff3fcbfa4f535e7cdc81d72d30c836/build/plugins/llvm-mingw/rust.mk#L35-L41
(I think MSVC also has a similar error, see e.g. issue mesonbuild/meson#13236)

Note that in the starting post --target=x86_64-pc-windows-gnullvm was added twice to the clang invocation, due to:

cc-rs/src/lib.rs

Lines 1941 to 1946 in 5863b31

	if cmd.is_like_clang() && target.contains("windows") {
	// Disambiguate mingw and msvc on Windows. Problem is that
	// depending on the origin clang can default to a mismatchig
	// run-time.
	cmd.push_cc_arg(format!("--target={}", target).into());
	}

(perhaps this whole block can be removed?)

[mati865]

Oh, this complicates things as I won't be able to easily test the fix with -Zbuild-std.

Import errors were discussed a bit in msys2/MINGW-packages#21017, it affects all windows targets other than windows-gnu.

Note that in the starting post --target=x86_64-pc-windows-gnullvm was added twice to the clang invocation

I doubt adding the argument twice causes any problems; the second occurrence should override the first one. It can be probably removed, but I'm not planning to spend time on testing it.

[kleisauke]

I can confirm PR #1176 fixes this, tested with commit libvips/build-win64-mxe@5b050d1.

Import errors were discussed a bit in msys2/MINGW-packages#21017, it affects all windows targets other than windows-gnu.

Interesting, thanks for linking and investigating that issue.

[mati865]

Great, thanks!