[PATCH] D18702: [WIP/LTO] Improve internalize decisions

[Rui Ueyama via llvm-commits]

ruiu added a subscriber: ruiu.

================
Comment at: ELF/LTO.cpp:78
@@ +77,3 @@
+  if (Config->Shared || Config->ExportDynamic) {
+    auto Visibility = Body->getVisibility();
+    if ((Visibility == llvm::ELF::STV_DEFAULT ||
----------------
We don't usually use `auto` if the type is obvious in a very narrow (e.g. RHS) context. We also use short names for variables with narrow scope, so I'd write

  uint8_t V = Body->getVisibility();
  return V == STV_DEFAULT || V == STV_PROTECTED;

================
Comment at: ELF/LTO.cpp:129-138
@@ -116,5 +128,12 @@
     // the IRMover will just rename the symbol.
-    // Shared libraries need to be handled slightly differently.
-    // For now, let's be conservative and just never internalize
-    // symbols when creating a shared library.
-    if (!Config->Shared && !Config->ExportDynamic && !B->isUsedInRegularObj())
+   if (B->isUsedInRegularObj()) {
+      // Case 1: The symbol is referenced from the outside. We can't
+      // internalize this symbol.
+      Keep.push_back(GV);
+    } else if (isVisible(B)) {
+      // Case 2: The symbol is exported to the symbol table. Maybe we
+      // can internalize the symbol. So we put in a different set
+      // and we defer the decision after the mover did its job.
+      Keep.push_back(GV);
+    }
+    else {
----------------
In both cases it does the same thing. Is this intended?


Repository:
  rL LLVM

http://reviews.llvm.org/D18702