[LLD] LTO internalizing global marked with 'STV_PROTECTED'

[jhuber6]

I'm trying to compile the following files targeting the AMDGPU architecture.

target triple = "amdgcn-amd-amdhsa"

@__llvm_libc_heap_ptr = external hidden addrspace(1) global ptr, align 8

; Function Attrs: convergent mustprogress noinline nounwind optnone
define protected amdgpu_kernel void @kernel(ptr noundef %out) #0 {
entry:
  %out.addr = alloca ptr, align 8, addrspace(5)
  %out.addr.ascast = addrspacecast ptr addrspace(5) %out.addr to ptr
  store ptr %out, ptr %out.addr.ascast, align 8
  %0 = load ptr, ptr addrspacecast (ptr addrspace(1) @__llvm_libc_heap_ptr to ptr), align 8
  %1 = load ptr, ptr %out.addr.ascast, align 8
  store ptr %0, ptr %1, align 8
  ret void
}

attributes #0 = { convergent mustprogress noinline nounwind optnone "frame-pointer"="all" "no-builtin-printf" "no-trapping-math"="true" "stack-protector-buffer-size"="8" }

!llvm.module.flags = !{!0, !1, !2, !3}
!llvm.ident = !{!4}

!0 = !{i32 1, !"amdhsa_code_object_version", i32 500}
!1 = !{i32 1, !"wchar_size", i32 4}
!2 = !{i32 8, !"PIC Level", i32 2}
!3 = !{i32 7, !"frame-pointer", i32 2}
!4 = !{!"clang version 21.0.0git"}

target triple = "amdgcn-amd-amdhsa"

@__llvm_libc_heap_ptr = protected local_unnamed_addr addrspace(1) global ptr null, align 8
@__llvm_libc_heap_size = protected local_unnamed_addr addrspace(1) global i64 0, align 8

!llvm.module.flags = !{!0, !1, !2, !3, !4}
!llvm.ident = !{!5}

!0 = !{i32 1, !"wchar_size", i32 4}
!1 = !{i32 8, !"PIC Level", i32 2}
!2 = !{i32 7, !"frame-pointer", i32 2}
!3 = !{i32 1, !"ThinLTO", i32 0}
!4 = !{i32 1, !"EnableSplitLTOUnit", i32 1}
!5 = !{!"clang version 21.0.0git"}

If I take these two files and run them through ld.lld I notice that the reference to the protected variable gets internalized. In the resolutions.txt file, it claims that it is not visible to the regular object because it is not exported. But, this is a -shared build and a protected symbol should not be internalized like this.

ld.lld foo.o bar.o -shared -save-temps

Reading the temporary outputs shows the internalization happening.

$ opt -S a.out.0.0.preopt.bc
@__llvm_libc_heap_size = protected local_unnamed_addr addrspace(1) global i64 0, align 8
@__llvm_libc_heap_ptr = protected local_unnamed_addr addrspace(1) global ptr null, align 8
...
$ opt -S a.out.0.2.internalize.bc
@__llvm_libc_heap_size = protected addrspace(1) global i64 0, align 8
@__llvm_libc_heap_ptr = internal addrspace(1) global ptr null, align 8
...

[jhuber6]

Apologies, I just realized after making this issue that the issue is that my external declaration does not have protected visibility on it. If you make that protected it retains as expected.

It's extremely unintuitive, but I guess that's just how linkers decide to do things.

[jhuber6]

Honestly, I'm just going to reopen this because it has very unfortunate interactions with the -fvisibility= option in clang. That doesn't apply to extern definitions, which means that this will happen if you use -fvisibility=, see https://godbolt.org/z/d3zoWzq4h. It seems unintuitive that external references need to be explicitly marked or else they will silently be internalized.