mark f128 as reliable on powerpc64 with +vsx

[folkertdev]

This feature is only enabled by default for powerpc64le, and not enabled on the big-endian targets. Even with vsx, the default long double is ppcf128 on the big-endian targets, so LLVM calls the incorrect libcall. Also some libcalls just don't exist.

I tested this with and without rustflags = ["-C", "target-feature=+vsx"] in the bootstrap.toml:

./x.py test library/core --target powerpc64-unknown-linux-gnu
./x.py test library/core --target powerpc64le-unknown-linux-gnu

These all hit one failure, in the docs test for library/core/src/keyword_docs.rs where become is used: powerpc does not support guaranteed tail calls. Anyway, that is not relevant for f128 support.

Sadly powerpc-unknown-linux-gnu with +vsx hits the issue fixed in llvm/llvm-project#216613, I've requested a backport for the fix.

r? beetrees or @tgross35
cc @Gelbpunkt

[beetrees]

This will result in target_has_reliable_f128 being different e.g. if linking core/std compiled without vsx to a test compiled with -Ctarget-feature=+vsx, but that's not how compiler tests are ran in practice and this is an internal cfg so I think that's fine.

@bors r+

[rust-bors]

📌 Commit b540652 has been approved by beetrees

It is now in the queue for this repository.

[beetrees]

@bors r-

Even with vsx, the default long double is ppcf128 on the big-endian targets, so LLVM calls the incorrect libcall.

Seems to work on compiler explorer?

[rust-bors]

This pull request was unapproved.

View changes since this unapproval

[folkertdev]

I think LLVM just emits that symbol, but it does not actually exist

#![feature(f128)]

#[unsafe(no_mangle)]
pub fn square(num: f128) -> f128 {
    num.sin()
}

fn main() {
    dbg!(square(3.14));
    dbg!(3.14f32.sin());
}

with cargo +stage1 run --target powerpc64-unknown-linux-gnu gives

error: linking with `powerpc64-linux-gnu-gcc` failed: exit status: 1
  |
  = note:  "powerpc64-linux-gnu-gcc" "-m64" "/home/folkertdev/rust/playground/target/powerpc64-unknown-linux-gnu/debug/build/playground/fc020cdbc196014d/out/rustc1Mxegi/symbols.o" "<15 object files omitted>" "-Wl,--as-needed" "-Wl,-Bstatic" "<sysroot>/lib/rustlib/powerpc64-unknown-linux-gnu/lib/{libstd-*,libpanic_unwind-*,libobject-*,libmemchr-*,libaddr2line-*,libgimli-*,libcfg_if-*,librustc_demangle-*,libstd_detect-*,libhashbrown-*,librustc_std_workspace_alloc-*,libminiz_oxide-*,libadler2-*,libunwind-*,liblibc-*,librustc_std_workspace_core-*,liballoc-*,libcore-*,libcompiler_builtins-*}.rlib" "-Wl,-Bdynamic" "-lgcc_s" "-lutil" "-lrt" "-lpthread" "-lm" "-ldl" "-lc" "-L" "/home/folkertdev/rust/playground/target/powerpc64-unknown-linux-gnu/debug/build/playground/fc020cdbc196014d/out/rustc1Mxegi/raw-dylibs" "-Wl,--eh-frame-hdr" "-Wl,-z,noexecstack" "-L" "<sysroot>/lib/rustlib/powerpc64-unknown-linux-gnu/lib" "-o" "/home/folkertdev/rust/playground/target/powerpc64-unknown-linux-gnu/debug/build/playground/fc020cdbc196014d/out/playground" "-Wl,--gc-sections" "-pie" "-Wl,-z,relro,-z,now" "-nodefaultlibs"
  = note: some arguments are omitted. use `--verbose` to show all linker arguments
  = note: /usr/lib/gcc-cross/powerpc64-linux-gnu/13/../../../../powerpc64-linux-gnu/bin/ld: /home/folkertdev/rust/playground/target/powerpc64-unknown-linux-gnu/debug/build/playground/fc020cdbc196014d/out/playground.86r5npe4uuwy526gz589nbf7b.02b4q5j.rcgu.o: in function `<f128>::sin':
          /home/folkertdev/rust/rust/library/std/src/num/f128.rs:381:(.text._RNvMNtCs2SUpeEcTk1l_3std4f128C4f1283sinCs33bJLyDsXdB_10playground+0x14): undefined reference to `sinf128'
          collect2: error: ld returned 1 exit status

That is different from the x86 issue though (which means I must be misunderstanding the x86 problem in LLVM, or it just does not actually apply here and really is x86-specific). I'll adjust the comment.

[folkertdev]

@rustbot ready

[beetrees]

A quick check on compiler explorer suggests LLVM has emitted the correct libcalls for x86_64 since LLVM 19. I think the main blocker now on most major targets is implementing the symbols in compiler builtins. That does lead me to wonder whether the blanket (_, Os::Linux) if target_pointer_width == 64 => true is accurate, if they're not available on big-endian 64-bit PowerPC. Regardless, this change is an improvement over the status quo, so @bors r+

[rust-bors]

📌 Commit a3ad87e has been approved by beetrees

It is now in the queue for this repository.

[folkertdev]

A quick check on compiler explorer suggests LLVM has emitted the correct libcalls for x86_64 since LLVM 19.

Right, but only on gnu targets. I ran into this before

llvm/llvm-project#44744 (comment)

For powerpc even with the musl targets the names are correct. Just, musl does not provide any f128 implementations as far as I'm aware.

[tgross35]

I guess AIX has VSX but is still broken llvm/llvm-project#101545, seems like this may have broken builds on the target https://github.com/rust-lang/libc/actions/runs/35495229282/job/106036887284?pr=5411