C interop: `f80`, `f128` and `c_longdouble`

[nxsaken]

Metadata
Point of contact	@folkertdev
Team(s)	compiler, lang, libs-api
Goal document	2026/interop-f80-f128

Summary

Rust should be able to define an ABI-compatible counterpart to every signature that C can define. This is important for C interop and translation of code from C to rust, using manual FFI, and tooling like c-bindgen and c2rust. This goal closes a long-standing gap in rust's ability to match C.

Tasks and status

---

Note: we have updated the body to match the 2026 goal. Your original text is preserved below.

Details

Metadata
Point of contact	@folkertdev
Team(s)	compiler, lang, libs-api
Goal document	2026/interop-f80-f128

Summary

Rust should be able to define an ABI-compatible counterpart to every signature that C can define. This is important for C interop and translation of code from C to rust, using manual FFI, and tooling like c-bindgen and c2rust. This goal closes a long-standing gap in rust's ability to match C.

Tasks and status

[nxsaken]

This issue is intended for status updates only.

For general questions or comments, please contact the owner(s) directly.

[folkertdev]

Key developments

We got a good amount of work done in LLVM which will make it into LLVM 23, the main items are support for window msvc and arm64ec:

• [X86][Windows] Return fp128 on the stack llvm/llvm-project#204887
• [Arm64EC] Add support for fp128 llvm/llvm-project#206980

Along with a number of smaller fixes

• [WebAssembly] support f{min, max}imum.f128 and f{min, max}imumnum.f128 llvm/llvm-project#207160
• [X86][Windows] use LowerWin64_INT128_TO_FP for u128 to f128 conversion llvm/llvm-project#207285
• [X86][AArch64][SPIRV] add more tests for ldexp/frexp llvm/llvm-project#208462

So with LLVM 23 f128 should work on tier 1 targets at least.

Work in progress

Libcalls for f128 assume that f128 = long double, which is false on the majority of targets.

• [llvm] Change fp128 lowering to use f128 functions by default llvm/llvm-project#76558
• add frexp/ldexp expansion in ExpandIrInsts llvm/llvm-project#208556

It turns out the ppc_f128 type is full of bugs

• [PowerPC][SDAG] fix bitcast on ppc_f128 swapping the two halves llvm/llvm-project#208969
• [PowerPC] fix ppc_fp128 FABS miscompile llvm/llvm-project#209286

Blockers

We've failed to get a response on the main blocker for this feature: libcalls assuming target features that are not part of the baseline. For f128 the main offender is powerpc, whose libcalls use vector registers (from the vsx target feature) for libcalls, but the baseline target does not have those registers. Hence libcalls miscompile.

We posted about it here.

llvm/llvm-project#112885 (comment)

Help wanted

For the moment we're mostly dealing with LLVM, and we'd be much helped by any additional LLVM expertise.

[folkertdev]

Key developments

I wrote an RFC for c_longdouble, for which a lang experiment was championed:

• Add core::ffi::c_longdouble rfcs#4003
• Tracking Issue for c_longdouble rust#161787
• Design meeting: Add core::ffi::c_longdouble lang-team#389

We've enabled some more f128 tests and made some overall f128 fixes:

• x86_64-win: Enable f128 on LLVM 23+ rust#158918
• run classify-runtime-const test for f128 rust#161591
• arm64ec: f128 is supported since LLVM 23 rust#160779
• rustc_target: callconv: mips64: Return structs with single f128 in FPRs rust#159967
• enable f128 tests on windows compiler-builtins#1280

And some LLVM work

• [X86] fix va_arg alignment of f128 llvm/llvm-project#218017
• [PowerPC][Clang] fix IEEE f128 complex div/mul use IBM f128 libcalls on powerpc llvm/llvm-project#218151

Work in progress

• implement VaArgSafe for f128 rust#161424
• I'm working on the X87 f80 and PPC f128 implementations, though we'll probably have to rename everything later.
• Invalid lowering of llvm.*.f128 intrinsics llvm/llvm-project#44744 is being worked on by LLVM developers, we're trying to make sure that'll work for us

Blockers

• LLVM 21 support will be dropped somewhere in October

Help wanted

• zmij, a library for floating-point-to-string conversion, got support for IEEE f128 and long double. The rust port does not support these yet. The zmij algorithm is a very promising candidate for the Display implementation of f128 and (the types aliased by) c_longdouble, providing good code size and performance.