[FMV][AArch64] Do not emit ifunc resolver on use. (#97761)

It was raised in https://github.com/llvm/llvm-project/issues/81494 that
we are not generating correct code when there is no TU-local caller.

The suggestion was to emit a resolver:
* Whenever there is a use in the TU.
* When the TU has a definition of the default version.

See the comment for more details:

https://github.com/llvm/llvm-project/issues/81494#issuecomment-1985963497

This got addressed with https://github.com/llvm/llvm-project/pull/84405.

Generating a resolver on use means that we may end up with multiple
resolvers across different translation units. Those resolvers may not be
the same because each translation unit may contain different version
declarations (user's fault). Therefore the order of linking the final
image determines which of these weak symbols gets selected, resulting in
non consisted behavior. I am proposing to stop emitting a resolver on
use and only do so in the translation unit which contains the default
definition. This way we guarantee the existence of a single resolver.
Now, when a versioned function is used we want to emit a declaration of
the function symbol omitting the multiversion mangling.

I have added a requirement to ACLE mandating that all the function
versions are declared in the translation unit which contains the default
definition: https://github.com/ARM-software/acle/pull/328

1 files changed, 50 insertions, 47 deletions

diff --git a/clang/lib/CodeGen/CodeGenModule.cpp b/clang/lib/CodeGen/CodeGenModule.cpp
index 0c002b5..71192cb 100644
--- a/clang/lib/CodeGen/CodeGenModule.cpp
+++ b/clang/lib/CodeGen/CodeGenModule.cpp
@@ -3796,8 +3796,7 @@ void CodeGenModule::EmitGlobal(GlobalDecl GD) {
     // Forward declarations are emitted lazily on first use.
     if (!FD->doesThisDeclarationHaveABody()) {
       if (!FD->doesDeclarationForceExternallyVisibleDefinition() &&
-          (!FD->isMultiVersion() ||
-           !FD->getASTContext().getTargetInfo().getTriple().isAArch64()))
+          (!FD->isMultiVersion() || !getTarget().getTriple().isAArch64()))
         return;
 
       StringRef MangledName = getMangledName(GD);
@@ -4191,23 +4190,6 @@ llvm::GlobalValue::LinkageTypes getMultiversionLinkage(CodeGenModule &CGM,
   return llvm::GlobalValue::WeakODRLinkage;
 }
 
-static FunctionDecl *createDefaultTargetVersionFrom(const FunctionDecl *FD) {
-  auto *DeclCtx = const_cast<DeclContext *>(FD->getDeclContext());
-  TypeSourceInfo *TInfo = FD->getTypeSourceInfo();
-  StorageClass SC = FD->getStorageClass();
-  DeclarationName Name = FD->getNameInfo().getName();
-
-  FunctionDecl *NewDecl =
-      FunctionDecl::Create(FD->getASTContext(), DeclCtx, FD->getBeginLoc(),
-                           FD->getEndLoc(), Name, TInfo->getType(), TInfo, SC);
-
-  NewDecl->setIsMultiVersion();
-  NewDecl->addAttr(TargetVersionAttr::CreateImplicit(
-      NewDecl->getASTContext(), "default", NewDecl->getSourceRange()));
-
-  return NewDecl;
-}
-
 void CodeGenModule::emitMultiVersionFunctions() {
   std::vector<GlobalDecl> MVFuncsToEmit;
   MultiVersionFuncs.swap(MVFuncsToEmit);
@@ -4234,29 +4216,30 @@ void CodeGenModule::emitMultiVersionFunctions() {
       return cast<llvm::Function>(Func);
     };
 
-    bool HasDefaultDecl = !FD->isTargetVersionMultiVersion();
-    bool ShouldEmitResolver =
-        !getContext().getTargetInfo().getTriple().isAArch64();
+    // For AArch64, a resolver is only emitted if a function marked with
+    // target_version("default")) or target_clones() is present and defined
+    // in this TU. For other architectures it is always emitted.
+    bool ShouldEmitResolver = !getTarget().getTriple().isAArch64();
     SmallVector<CodeGenFunction::MultiVersionResolverOption, 10> Options;
 
     getContext().forEachMultiversionedFunctionVersion(
         FD, [&](const FunctionDecl *CurFD) {
           llvm::SmallVector<StringRef, 8> Feats;
+          bool IsDefined = CurFD->doesThisDeclarationHaveABody();
 
           if (const auto *TA = CurFD->getAttr<TargetAttr>()) {
             TA->getAddedFeatures(Feats);
             llvm::Function *Func = createFunction(CurFD);
             Options.emplace_back(Func, TA->getArchitecture(), Feats);
           } else if (const auto *TVA = CurFD->getAttr<TargetVersionAttr>()) {
-            bool HasDefaultDef = TVA->isDefaultVersion() &&
-                                 CurFD->doesThisDeclarationHaveABody();
-            HasDefaultDecl |= TVA->isDefaultVersion();
-            ShouldEmitResolver |= (CurFD->isUsed() || HasDefaultDef);
+            if (TVA->isDefaultVersion() && IsDefined)
+              ShouldEmitResolver = true;
             TVA->getFeatures(Feats);
             llvm::Function *Func = createFunction(CurFD);
             Options.emplace_back(Func, /*Architecture*/ "", Feats);
           } else if (const auto *TC = CurFD->getAttr<TargetClonesAttr>()) {
-            ShouldEmitResolver |= CurFD->doesThisDeclarationHaveABody();
+            if (IsDefined)
+              ShouldEmitResolver = true;
             for (unsigned I = 0; I < TC->featuresStrs_size(); ++I) {
               if (!TC->isFirstOfVersion(I))
                 continue;
@@ -4282,13 +4265,6 @@ void CodeGenModule::emitMultiVersionFunctions() {
     if (!ShouldEmitResolver)
       continue;
 
-    if (!HasDefaultDecl) {
-      FunctionDecl *NewFD = createDefaultTargetVersionFrom(FD);
-      llvm::Function *Func = createFunction(NewFD);
-      llvm::SmallVector<StringRef, 1> Feats;
-      Options.emplace_back(Func, /*Architecture*/ "", Feats);
-    }
-
     llvm::Constant *ResolverConstant = GetOrCreateMultiVersionResolver(GD);
     if (auto *IFunc = dyn_cast<llvm::GlobalIFunc>(ResolverConstant)) {
       ResolverConstant = IFunc->getResolver();
@@ -4339,6 +4315,14 @@ void CodeGenModule::emitMultiVersionFunctions() {
     emitMultiVersionFunctions();
 }
 
+static void replaceDeclarationWith(llvm::GlobalValue *Old,
+                                   llvm::Constant *New) {
+  assert(cast<llvm::Function>(Old)->isDeclaration() && "Not a declaration");
+  New->takeName(Old);
+  Old->replaceAllUsesWith(New);
+  Old->eraseFromParent();
+}
+
 void CodeGenModule::emitCPUDispatchDefinition(GlobalDecl GD) {
   const auto *FD = cast<FunctionDecl>(GD.getDecl());
   assert(FD && "Not a FunctionDecl?");
@@ -4443,12 +4427,9 @@ void CodeGenModule::emitCPUDispatchDefinition(GlobalDecl GD) {
     // Fix up function declarations that were created for cpu_specific before
     // cpu_dispatch was known
     if (!isa<llvm::GlobalIFunc>(IFunc)) {
-      assert(cast<llvm::Function>(IFunc)->isDeclaration());
       auto *GI = llvm::GlobalIFunc::create(DeclTy, 0, Linkage, "", ResolverFunc,
                                            &getModule());
-      GI->takeName(IFunc);
-      IFunc->replaceAllUsesWith(GI);
-      IFunc->eraseFromParent();
+      replaceDeclarationWith(IFunc, GI);
       IFunc = GI;
     }
 
@@ -4478,7 +4459,8 @@ void CodeGenModule::AddDeferredMultiVersionResolverToEmit(GlobalDecl GD) {
 }
 
 /// If a dispatcher for the specified mangled name is not in the module, create
-/// and return an llvm Function with the specified type.
+/// and return it. The dispatcher is either an llvm Function with the specified
+/// type, or a global ifunc.
 llvm::Constant *CodeGenModule::GetOrCreateMultiVersionResolver(GlobalDecl GD) {
   const auto *FD = cast<FunctionDecl>(GD.getDecl());
   assert(FD && "Not a FunctionDecl?");
@@ -4506,8 +4488,15 @@ llvm::Constant *CodeGenModule::GetOrCreateMultiVersionResolver(GlobalDecl GD) {
     ResolverName += ".resolver";
   }
 
-  // If the resolver has already been created, just return it.
-  if (llvm::GlobalValue *ResolverGV = GetGlobalValue(ResolverName))
+  // If the resolver has already been created, just return it. This lookup may
+  // yield a function declaration instead of a resolver on AArch64. That is
+  // because we didn't know whether a resolver will be generated when we first
+  // encountered a use of the symbol named after this resolver. Therefore,
+  // targets which support ifuncs should not return here unless we actually
+  // found an ifunc.
+  llvm::GlobalValue *ResolverGV = GetGlobalValue(ResolverName);
+  if (ResolverGV &&
+      (isa<llvm::GlobalIFunc>(ResolverGV) || !getTarget().supportsIFunc()))
     return ResolverGV;
 
   const CGFunctionInfo &FI = getTypes().arrangeGlobalDeclaration(GD);
@@ -4533,7 +4522,8 @@ llvm::Constant *CodeGenModule::GetOrCreateMultiVersionResolver(GlobalDecl GD) {
                                   "", Resolver, &getModule());
     GIF->setName(ResolverName);
     SetCommonAttributes(FD, GIF);
-
+    if (ResolverGV)
+      replaceDeclarationWith(ResolverGV, GIF);
     return GIF;
   }
 
@@ -4542,6 +4532,8 @@ llvm::Constant *CodeGenModule::GetOrCreateMultiVersionResolver(GlobalDecl GD) {
   assert(isa<llvm::GlobalValue>(Resolver) &&
          "Resolver should be created for the first time");
   SetCommonAttributes(FD, cast<llvm::GlobalValue>(Resolver));
+  if (ResolverGV)
+    replaceDeclarationWith(ResolverGV, Resolver);
   return Resolver;
 }
 
@@ -4571,6 +4563,7 @@ llvm::Constant *CodeGenModule::GetOrCreateLLVMFunction(
     ForDefinition_t IsForDefinition) {
   const Decl *D = GD.getDecl();
 
+  std::string NameWithoutMultiVersionMangling;
   // Any attempts to use a MultiVersion function should result in retrieving
   // the iFunc instead. Name Mangling will handle the rest of the changes.
   if (const FunctionDecl *FD = cast_or_null<FunctionDecl>(D)) {
@@ -4592,14 +4585,24 @@ llvm::Constant *CodeGenModule::GetOrCreateLLVMFunction(
 
     if (FD->isMultiVersion()) {
       UpdateMultiVersionNames(GD, FD, MangledName);
-      if (FD->getASTContext().getTargetInfo().getTriple().isAArch64() &&
-          !FD->isUsed())
-        AddDeferredMultiVersionResolverToEmit(GD);
-      else if (!IsForDefinition)
-        return GetOrCreateMultiVersionResolver(GD);
+      if (!IsForDefinition) {
+        // On AArch64 we do not immediatelly emit an ifunc resolver when a
+        // function is used. Instead we defer the emission until we see a
+        // default definition. In the meantime we just reference the symbol
+        // without FMV mangling (it may or may not be replaced later).
+        if (getTarget().getTriple().isAArch64()) {
+          AddDeferredMultiVersionResolverToEmit(GD);
+          NameWithoutMultiVersionMangling = getMangledNameImpl(
+              *this, GD, FD, /*OmitMultiVersionMangling=*/true);
+        } else
+          return GetOrCreateMultiVersionResolver(GD);
+      }
     }
   }
 
+  if (!NameWithoutMultiVersionMangling.empty())
+    MangledName = NameWithoutMultiVersionMangling;
+
   // Lookup the entry, lazily creating it if necessary.
   llvm::GlobalValue *Entry = GetGlobalValue(MangledName);
   if (Entry) {