[clang] [Clang][HIP] Skip internalization for non-LTO device links (PR #225859)

[Joseph Huber via cfe-commits]

================
@@ -53,19 +53,19 @@ void AMDGCN::Linker::constructLldCommand(Compilation &C, const JobAction &JA,
                                          const llvm::opt::ArgList &Args) const {
   // Construct lld command.
   // The output from ld.lld is an HSA code object file.
-  ArgStringList LldArgs{"-flavor",
-                        "gnu",
-                        "-m",
-                        "elf64_amdgpu",
-                        "--no-undefined",
-                        "-shared",
-                        "-plugin-opt=-amdgpu-internalize-symbols"};
+  const ToolChain &TC = getToolChain();
+  LTOKind LTOMode = TC.getLTOMode(Args, Action::OFK_HIP);
+  ArgStringList LldArgs{"-flavor",        "gnu",    "-m", "elf64_amdgpu",
+                        "--no-undefined", "-shared"};
+  // Native object references are invisible when LTO compiles bitcode libraries
+  // in a non-LTO link, so library definitions must remain external.
+  if (LTOMode != LTOK_None)
+    LldArgs.push_back("-plugin-opt=-amdgpu-internalize-symbols");
----------------
jhuber6 wrote:

This is the part that confuses me. I *really* dislike the internalization pass here, and LTO is supposed to make it unnecessary if you set up your visibility correctly. I.e. -fvisibility=hidden. I don't fully remember what this pass does, but I can imagine it breaking things I will rely on, like the RPC client symbol that is marked `[[gnu::visibility("protected})]]` being internalized and no longer readable from HSA.

https://github.com/llvm/llvm-project/pull/225859