ARM VFP calling convention detection uses LLVM triple suffix instead of llvm_floatabi, misses Apple ARM

[SebTardif]

Bug

In compiler/rustc_target/src/callconv/arm.rs line 96, VFP (hardware floating-point) calling convention detection uses cx.target_spec().llvm_target.ends_with("hf") as a proxy for hard-float ABI. Apple ARM targets (armv7s-apple-ios, armv7k-apple-watchos) set llvm_floatabi: Some(FloatAbi::Hard) via the base spec (line 130-131 of spec/base/apple/mod.rs), but their LLVM triples end with "-ios" or "-watchos", not "hf".

This causes HFA (Homogeneous Floating-point Aggregate) arguments to be passed on the stack instead of in VFP registers on these targets, breaking ABI compatibility with C.

Impact

FFI functions on Apple 32-bit ARM targets may use wrong calling convention for structs like struct { f32, f32, f32, f32 }. These are Tier 3 targets.

Fix

PR #474

[SebTardif]

Upstream context and cross-reference

The VFP detection via llvm_target.ends_with("hf") was introduced by @cuviper (Josh Stone) in commit efc6764 ("Support homogeneous aggregates for hard-float ARM", 2017) and refined by @AminArria (Irina Popa) in commit fb15d44 (2018). The ends_with("hf") check was correct at the time because all hard-float ARM targets used the "eabihf" triple suffix.

Apple ARM targets were later added with llvm_floatabi: Some(FloatAbi::Hard) in compiler/rustc_target/src/spec/base/apple/mod.rs (line 130-131), but their LLVM triples follow the Apple convention (armv7s-apple-ios, armv7k-apple-watchos) which does not end with "hf". The VFP detection was never updated to use the structured llvm_floatabi field.

@workingjubilee (Jubilee Young) recently touched this code in commit 72ecde2 (2025-05-21), updating fn_abi.conv to use the new CanonAbi::Arm(ArmCall::Aapcs) type. The ends_with("hf") check survived unchanged.

These are Tier 3 targets, so impact is limited, but HFA structs like struct { f32, f32, f32, f32 } would be passed on the stack instead of in VFP registers, differing from Clang's behavior.

Fix: PR #474