[cfe-dev] different gcc/clang behavior for Internal visibility

[Andrii Nakryiko via cfe-dev]

On Mon, May 10, 2021 at 12:52 PM Fangrui Song <maskray at google.com> wrote:
>
> On 2021-05-10, Y Song via cfe-dev wrote:
> >Hi,
> >
> >The bpf linker project tries to explore to use INTERNAL visibility as in
> >   https://lore.kernel.org/bpf/20210507054119.270888-1-andrii@kernel.org/
> >
> >But we found clang actually changed user "internal" visibility to "hidden".
> >For example, I have the following example,
> >
> >$ cat t.c
> >int __attribute__((visibility("internal"))) foo() { return 0; }
> >$ clang -c t.c && llvm-readelf -s t.o | grep foo
> >     3: 0000000000000000     8 FUNC    GLOBAL HIDDEN      2 foo
> >$ gcc -c t.c && llvm-readelf -s t.o | grep foo
> >     8: 0000000000000000    11 FUNC    GLOBAL INTERNAL    1 foo
> >$
> >
> >Looks like this is caused by clang Attr.td,
> >
> >diff --git a/clang/include/clang/Basic/Attr.td
> >b/clang/include/clang/Basic/Attr.td
> >index 5e04f32187cd..4559a1bcfe42 100644
> >--- a/clang/include/clang/Basic/Attr.td
> >+++ b/clang/include/clang/Basic/Attr.td
> >@@ -2776,7 +2776,7 @@ def Visibility : InheritableAttr {
> >   let Spellings = [GCC<"visibility">];
> >   let Args = [EnumArgument<"Visibility", "VisibilityType",
> >                            ["default", "hidden", "internal", "protected"],
> >-                           ["Default", "Hidden", "Hidden", "Protected"]>];
> >+                           ["Default", "Hidden", "Internal", "Protected"]>];
> >   let MeaningfulToClassTemplateDefinition = 1;
> >   let Documentation = [Undocumented];
> > }
> >@@ -2786,7 +2786,7 @@ def TypeVisibility : InheritableAttr {
> >   let Spellings = [Clang<"type_visibility">];
> >   let Args = [EnumArgument<"Visibility", "VisibilityType",
> >                            ["default", "hidden", "internal", "protected"],
> >-                           ["Default", "Hidden", "Hidden", "Protected"]>];
> >+                           ["Default", "Hidden", "Internal", "Protected"]>];
> > //  let Subjects = [Tag, ObjCInterface, Namespace];
> >   let Documentation = [Undocumented];
> > }
> >
> >One of early commits,
> >
> >commit 570024a8d9b4a4aa4a35f077a0a65003dc7b71fe
> >Author: Eli Friedman <eli.friedman at gmail.com>
> >Date:   Thu Aug 5 06:57:20 2010 +0000
> >
> >    Implement #pragma GCC visibility.
> >
> >    llvm-svn: 110315
> >
> >I see
> >
> >+    else if (VisType->isStr("internal"))
> >+      type = VisibilityAttr::HiddenVisibility; // FIXME
> >
> >Do we have any plan to support Internal visibility?
>
> I don't think there is value supporting STV_INTERNAL.
>
> STV_INTERNAL as we see today in the ELF specification was requested by SGI.
>
> > The meaning of this visibility attribute may be defined by processor
> > supplements to further constrain hidden symbols. A processor
> > supplement's definition should be such that generic tools can safely
> > treat internal symbols as hidden.  An internal symbol contained in a
> > relocatable object must be either removed or converted to STB_LOCAL
> > binding by the link-editor when the relocatable object is included in an
> > executable file or shared object.
>
> I recall from reading somewhere that it is used by its propritery
> compiler for some LTO like optimizations. STV_INTERNAL is identical to
> STV_HIDDEN in GNU/Solaris (and likely HP-UX).
>
> For the Linux kernel, my suggestion is to just use __attribute__((visibility("hidden"))).
> Then there will be no confusion that "internal" translates to STV_HIDDEN.

We are considering using both STV_HIDDEN and STV_INTERNAL for BPF
target and we want to be able to distinguish between the two. Given
ELF defines all four (STV_DEFAULT, STV_PROTECTED, STV_HIDDEN,
STV_INTERNAL), it would be great to be able to actually express
STV_INTERNAL. Are there any specific problems with passing-through
STV_INTERNAL down to ELF symbol visibility?