Inlining + target_feature broken in powerpc64

[gnzlbg]

godbolt

#![feature(repr_simd, powerpc_target_feature)]
#![allow(non_camel_case_types)]

#[repr(simd)] pub struct u32x4(u32, u32, u32, u32);

impl u32x4 {
    #[inline]
    // #[inline(always)]
    fn splat(x: u32) -> Self {
        u32x4(x, x, x, x)
    }
}

#[target_feature(enable = "altivec")]
pub unsafe fn splat_u32x4(x: u32) -> u32x4 {
    u32x4::splat(x)
}

with #[inline] that code produces a function call within splat_u32x4 (b example::u32x4::splat) to u32x4::splat, which is not eliminated, even though this method is module private. With #[inline(always)], u32x4::splat is inlined into splat_u32x4, and no code for u32x4::splat is generated.

#[inline] should not be needed here, much less #[inline(always)], yet without #[inline(always)] this produces bad codegen.

Removing the #[target_feature] attribute from splat_u32x4 fixes the issue, no #[inline] necessary: godbolt. So there must be some interaction between inlining and target features going on here.

[sayantn]

This is something on the LLVM side. See this comment. This happens in all microarchitectures except for ARM and x86 I think.

[sayantn]

This was resolved in llvm/llvm-project#117493 and llvm/llvm-project#126562
godbolt

[sayantn]

There are still some optimization issues in darwin and ios
https://godbolt.org/z/dePb98rvT