LLVM ctpop.v16i8 intrinsic breaks on s390x

[TheIronBorn]

This works as a minimal LLVM example: https://godbolt.org/z/_nnrkF

Unfortunately godbolt's nightly Rust is broken so I can't reproduce it in Rust: https://github.com/mattgodbolt/compiler-explorer/issues/1165

u8x16::count_ones/zeros will likely need a workaround until fixed.

Reported to LLVM: https://bugs.llvm.org/show_bug.cgi?id=39730

(ctlz.v16i8 & cttz.v16i8 break as well)

[nikic]

This is the IR generated by rustc:

define void @_ZN7example5ctpop17h99089f7f1b127411E(<16 x i8>* noalias nocapture sret dereferenceable(16), <16 x i8>* noalias nocapture dereferenceable(16) %x) unnamed_addr #0 {
  %arg = alloca <16 x i8>, align 16
  %1 = load <16 x i8>, <16 x i8>* %x, align 16
  store <16 x i8> %1, <16 x i8>* %arg, align 16
  call void @llvm.ctpop.v16i8(<16 x i8>* noalias nocapture sret dereferenceable(16) %0, <16 x i8>* noalias nocapture dereferenceable(16) %arg)
  br label %bb1

bb1: ; preds = %start
  ret void
}

Which is clearly wrong. Using the unadjusted ABI generates the correct code.

I'm not sure where exactly that places the bug -- possibly rustc should always assume unadjusted when linking LLVM intrinsics, as nothing else really makes sense in that context, and it leaves less room for error in target ABI handling?

[gnzlbg]

possibly rustc should always assume unadjusted when linking LLVM intrinsics, as nothing else really makes sense in that context

I think it makes sense to fix this upstream - cc @alexcrichton @rkruppe

[alexcrichton]

Sounds like a good idea to me!