Make explicit -fno-semantic-interposition (in -fpic mode) infer dso_local

-fno-semantic-interposition is currently the CC1 default. (The opposite
disables some interprocedural optimizations.) However, it does not infer
dso_local: on most targets accesses to ExternalLinkage functions/variables
defined in the current module still need PLT/GOT.

This patch makes explicit -fno-semantic-interposition infer dso_local,
so that PLT/GOT can be eliminated if targets implement local aliases
for AsmPrinter::getSymbolPreferLocal (currently only x86).

Currently we check whether the module flag "SemanticInterposition" is 0.
If yes, infer dso_local. In the future, we can infer dso_local unless
"SemanticInterposition" is 1: frontends other than clang will also
benefit from the optimization if they don't bother setting the flag.
(There will be risks if they do want ELF interposition: they need to set
"SemanticInterposition" to 1.)

clang/include/clang/Basic/LangOptions.def

@@ -301,6 +301,7 @@ ENUM_LANGOPT(TypeVisibilityMode, Visibility, 3, DefaultVisibility,
 LANGOPT(SetVisibilityForExternDecls, 1, 0,
         "apply global symbol visibility to external declarations without an explicit visibility")
 BENIGN_LANGOPT(SemanticInterposition        , 1, 0, "semantic interposition")
+BENIGN_LANGOPT(ExplicitNoSemanticInterposition, 1, 0, "explicitly no semantic interposition")
 ENUM_LANGOPT(StackProtector, StackProtectorMode, 2, SSPOff,
              "stack protector mode")
 ENUM_LANGOPT(TrivialAutoVarInit, TrivialAutoVarInitKind, 2, TrivialAutoVarInitKind::Uninitialized,

clang/include/clang/Driver/Options.td

@@ -3371,7 +3371,7 @@ defm ipa_cp : BooleanFFlag<"ipa-cp">,
     Group<clang_ignored_gcc_optimization_f_Group>;
 defm ivopts : BooleanFFlag<"ivopts">, Group<clang_ignored_gcc_optimization_f_Group>;
 def fsemantic_interposition : Flag<["-"], "fsemantic-interposition">, Group<f_Group>, Flags<[CC1Option]>;
-def fno_semantic_interposition: Flag<["-"], "fno-semantic-interposition">, Group<f_Group>;
+def fno_semantic_interposition: Flag<["-"], "fno-semantic-interposition">, Group<f_Group>, Flags<[CC1Option]>;
 defm non_call_exceptions : BooleanFFlag<"non-call-exceptions">, Group<clang_ignored_f_Group>;
 defm peel_loops : BooleanFFlag<"peel-loops">, Group<clang_ignored_gcc_optimization_f_Group>;
 defm permissive : BooleanFFlag<"permissive">, Group<clang_ignored_f_Group>;

clang/lib/CodeGen/CodeGenModule.cpp

@@ -493,6 +493,9 @@ void CodeGenModule::Release() {
   if (Context.getLangOpts().SemanticInterposition)
     // Require various optimization to respect semantic interposition.
     getModule().setSemanticInterposition(1);
+  else if (Context.getLangOpts().ExplicitNoSemanticInterposition)
+    // Allow dso_local on applicable targets.
+    getModule().setSemanticInterposition(0);
 
   if (CodeGenOpts.EmitCodeView) {
     // Indicate that we want CodeView in the metadata.

clang/lib/Driver/ToolChains/Clang.cpp

@@ -4471,10 +4471,12 @@ void Clang::ConstructJob(Compilation &C, const JobAction &JA,
     CmdArgs.push_back(A->getValue());
   }
 
-  if (Args.hasFlag(options::OPT_fsemantic_interposition,
-                   options::OPT_fno_semantic_interposition, false) &&
-      RelocationModel != llvm::Reloc::Static && !IsPIE)
-    CmdArgs.push_back("-fsemantic-interposition");
+  // The default is -fno-semantic-interposition. We render it just because we
+  // require explicit -fno-semantic-interposition to infer dso_local.
+  if (Arg *A = Args.getLastArg(options::OPT_fsemantic_interposition,
+                               options::OPT_fno_semantic_interposition))
+    if (RelocationModel != llvm::Reloc::Static && !IsPIE)
+      A->render(Args, CmdArgs);
 
   CmdArgs.push_back("-mthread-model");
   if (Arg *A = Args.getLastArg(options::OPT_mthread_model)) {

clang/lib/Frontend/CompilerInvocation.cpp

@@ -3049,6 +3049,9 @@ static void ParseLangArgs(LangOptions &Opts, ArgList &Args, InputKind IK,
   }
 
   Opts.SemanticInterposition = Args.hasArg(OPT_fsemantic_interposition);
+  // An explicit -fno-semantic-interposition infers dso_local.
+  Opts.ExplicitNoSemanticInterposition =
+      Args.hasArg(OPT_fno_semantic_interposition);
 
   // -mrtd option
   if (Arg *A = Args.getLastArg(OPT_mrtd)) {

clang/test/CodeGen/semantic-interposition.c

@@ -1,5 +1,9 @@
 // RUN: %clang_cc1 -emit-llvm -fsemantic-interposition %s -o - | FileCheck --check-prefix=INTERPOSITION %s
 // RUN: %clang_cc1 -emit-llvm %s -o - | FileCheck --check-prefix=NO %s
+/// With explicit -fno-semantic-interposition, add a module flag to inform the
+/// backend that dso_local can be inferred.
+// RUN: %clang_cc1 -emit-llvm -fno-semantic-interposition %s -o - | FileCheck --check-prefix=EXPLICIT_NO %s
 
 // INTERPOSITION: !{{[0-9]+}} = !{i32 1, !"SemanticInterposition", i32 1}
 // NO-NOT: "SemanticInterposition"
+// EXPLICIT_NO: !{{[0-9]+}} = !{i32 1, !"SemanticInterposition", i32 0}

clang/test/Driver/fsemantic-interposition.c

@@ -2,8 +2,12 @@
 // RUN: %clang -target x86_64 %s -Werror -fPIC -fsemantic-interposition -c -### 2>&1 | FileCheck %s
 // CHECK: "-fsemantic-interposition"
 
-// RUN: %clang -target x86_64 %s -Werror -fPIC -fsemantic-interposition -fno-semantic-interposition -c -### 2>&1 | FileCheck --check-prefix=NO %s
+/// Require explicit -fno-semantic-interposition to infer dso_local.
+// RUN: %clang -target x86_64 %s -Werror -fPIC -fsemantic-interposition -fno-semantic-interposition -c -### 2>&1 | FileCheck --check-prefix=EXPLICIT_NO %s
+// EXPLICIT_NO: "-fno-semantic-interposition"
+
 // RUN: %clang -target x86_64 %s -Werror -fsemantic-interposition -c -### 2>&1 | FileCheck --check-prefix=NO %s
 // RUN: %clang -target x86_64 %s -Werror -fPIC -c -### 2>&1 | FileCheck --check-prefix=NO %s
 // RUN: %clang -target x86_64 %s -Werror -fPIE -fsemantic-interposition -c -### 2>&1 | FileCheck --check-prefix=NO %s
 // NO-NOT: "-fsemantic-interposition"
+// NO-NOT: "-fno-semantic-interposition"

llvm/include/llvm/IR/GlobalValue.h

@@ -427,6 +427,7 @@ public:
   /// inlining across interposable call edges, since the callee can be
   /// replaced with something arbitrary.
   bool isInterposable() const;
+  bool canBenefitFromLocalAlias() const;
 
   bool hasExternalLinkage() const { return isExternalLinkage(getLinkage()); }
   bool hasAvailableExternallyLinkage() const {

llvm/include/llvm/IR/Module.h

@@ -857,6 +857,7 @@ public:
 
   /// Returns whether semantic interposition is to be respected.
   bool getSemanticInterposition() const;
+  bool noSemanticInterposition() const;
 
   /// Set whether semantic interposition is to be respected.
   void setSemanticInterposition(bool);

llvm/lib/CodeGen/AsmPrinter/AsmPrinter.cpp

@@ -462,10 +462,10 @@ MCSymbol *AsmPrinter::getSymbolPreferLocal(const GlobalValue &GV) const {
   // assembler would otherwise be conservative and assume a global default
   // visibility symbol can be interposable, even if the code generator already
   // assumed it.
-  if (TM.getTargetTriple().isOSBinFormatELF() &&
-      GlobalObject::isExternalLinkage(GV.getLinkage()) && GV.isDSOLocal() &&
-      !GV.isDeclaration() && !isa<GlobalIFunc>(GV) && !GV.hasComdat())
-    return getSymbolWithGlobalValueBase(&GV, "$local");
+  if (TM.getTargetTriple().isOSBinFormatELF() && GV.canBenefitFromLocalAlias())
+    if (GV.isDSOLocal() || (TM.getTargetTriple().isX86() &&
+                            GV.getParent()->noSemanticInterposition()))
+      return getSymbolWithGlobalValueBase(&GV, "$local");
   return TM.getSymbol(&GV);
 }
 

llvm/lib/IR/Globals.cpp

@@ -101,6 +101,12 @@ bool GlobalValue::isInterposable() const {
          !isDSOLocal();
 }
 
+bool GlobalValue::canBenefitFromLocalAlias() const {
+  // See AsmPrinter::getSymbolPreferLocal().
+  return GlobalObject::isExternalLinkage(getLinkage()) && !isDeclaration() &&
+         !isa<GlobalIFunc>(this) && !hasComdat();
+}
+
 unsigned GlobalValue::getAlignment() const {
   if (auto *GA = dyn_cast<GlobalAlias>(this)) {
     // In general we cannot compute this at the IR level, but we try.

llvm/lib/IR/Module.cpp

@@ -600,6 +600,13 @@ void Module::setSemanticInterposition(bool SI) {
   addModuleFlag(ModFlagBehavior::Error, "SemanticInterposition", SI);
 }
 
+bool Module::noSemanticInterposition() const {
+  // Conservatively require an explicit zero value for now.
+  Metadata *MF = getModuleFlag("SemanticInterposition");
+  auto *Val = cast_or_null<ConstantAsMetadata>(MF);
+  return Val && cast<ConstantInt>(Val->getValue())->getZExtValue() == 0;
+}
+
 void Module::setOwnedMemoryBuffer(std::unique_ptr<MemoryBuffer> MB) {
   OwnedMemoryBuffer = std::move(MB);
 }

llvm/lib/Target/TargetMachine.cpp

@@ -193,6 +193,14 @@ bool TargetMachine::shouldAssumeDSOLocal(const Module &M,
     // Check if we can use copy relocations.
     if (!(GV && GV->isThreadLocal()) && RM == Reloc::Static)
       return true;
+  } else if (TT.isOSBinFormatELF()) {
+    // If dso_local allows AsmPrinter::getSymbolPreferLocal to use a local
+    // alias, set the flag. We cannot set dso_local for other global values,
+    // because otherwise direct accesses to a probably interposable symbol (even
+    // if the codegen assumes not) will be rejected by the linker.
+    if (!GV || !GV->canBenefitFromLocalAlias())
+      return false;
+    return TT.isX86() && M.noSemanticInterposition();
   }
 
   // ELF & wasm support preemption of other symbols.

llvm/test/CodeGen/X86/semantic-interposition-infer-dsolocal.ll

@@ -0,0 +1,46 @@
+; RUN: llc -mtriple=x86_64 -relocation-model=pic < %s | FileCheck %s
+
+;; With a module flag SemanticInterposition=0, infer dso_local flags even if PIC.
+;; Local aliases will be generated for applicable variables and functions.
+
+@var = global i32 0, align 4
+
+@ifunc = ifunc i32 (), bitcast (i32 ()* ()* @ifunc_resolver to i32 ()*)
+
+define i32 @ifunc_impl() {
+entry:
+  ret i32 0
+}
+
+define i32 ()* @ifunc_resolver() {
+entry:
+  ret i32 ()* @ifunc_impl
+}
+
+declare i32 @external()
+
+define i32 @func() {
+  ret i32 0
+}
+
+;; Don't set dso_local on declarations or ifuncs.
+define i32 @foo() {
+; CHECK: movl .Lvar$local(%rip), %ebp
+; CHECK: callq external@PLT
+; CHECK: callq ifunc@PLT
+; CHECK: callq .Lfunc$local{{$}}
+entry:
+  %0 = load i32, i32* @var, align 4
+  %call = tail call i32 @external()
+  %add = add nsw i32 %call, %0
+  %call1 = tail call i32 @ifunc()
+  %add2 = add nsw i32 %add, %call1
+  %call2 = tail call i32 @func()
+  %add3 = add nsw i32 %add, %call2
+  ret i32 %add3
+}
+
+!llvm.module.flags = !{!0, !1}
+
+!0 = !{i32 1, !"SemanticInterposition", i32 0}
+!1 = !{i32 7, !"PIC Level", i32 2}