Reapply 'Implement target_clones multiversioning'

See discussion in D51650, this change was a little aggressive in an
error while doing a 'while we were here', so this removes that error
condition, as it is apparently useful.

This reverts commit bb49346.

clang/include/clang/AST/Decl.h

@@ -1840,7 +1840,8 @@ enum class MultiVersionKind {
   None,
   Target,
   CPUSpecific,
-  CPUDispatch
+  CPUDispatch,
+  TargetClones
 };
 
 /// Represents a function declaration or definition.
@@ -2459,6 +2460,10 @@ class FunctionDecl : public DeclaratorDecl,
   /// the target functionality.
   bool isTargetMultiVersion() const;
 
+  /// True if this function is a multiversioned dispatch function as a part of
+  /// the target-clones functionality.
+  bool isTargetClonesMultiVersion() const;
+
   /// \brief Get the associated-constraints of this function declaration.
   /// Currently, this will either be a vector of size 1 containing the
   /// trailing-requires-clause or an empty vector.

clang/include/clang/Basic/Attr.td

@@ -2677,6 +2677,40 @@ def Target : InheritableAttr {
   }];
 }
 
+def TargetClones : InheritableAttr {
+  let Spellings = [GCC<"target_clones">];
+  let Args = [VariadicStringArgument<"featuresStrs">];
+  let Documentation = [TargetClonesDocs];
+  let Subjects = SubjectList<[Function], ErrorDiag>;
+  let AdditionalMembers = [{
+    StringRef getFeatureStr(unsigned Index) const {
+      return *(featuresStrs_begin() + Index);
+    }
+    // 'default' is always moved to the end, so it isn't considered
+    // when mangling the index.
+    unsigned getMangledIndex(unsigned Index) const {
+      if (getFeatureStr(Index) == "default")
+        return std::count_if(featuresStrs_begin(), featuresStrs_end(),
+                              [](StringRef S) { return S != "default"; });
+
+      return std::count_if(featuresStrs_begin(), featuresStrs_begin() + Index,
+                           [](StringRef S) { return S != "default"; });
+    }
+
+    // True if this is the first of this version to appear in the config string.
+    // This is used to make sure we don't try to emit this function multiple
+    // times.
+    bool isFirstOfVersion(unsigned Index) const {
+      StringRef FeatureStr(getFeatureStr(Index));
+      return 0 == std::count_if(
+                      featuresStrs_begin(), featuresStrs_begin() + Index,
+                      [FeatureStr](StringRef S) { return S == FeatureStr; });
+    }
+  }];
+}
+
+def : MutualExclusions<[TargetClones, Target, CPUDispatch, CPUSpecific]>;
+
 def MinVectorWidth : InheritableAttr {
   let Spellings = [Clang<"min_vector_width">];
   let Args = [UnsignedArgument<"VectorWidth">];

clang/include/clang/Basic/AttrDocs.td

@@ -2233,6 +2233,40 @@ Additionally, a function may not become multiversioned after its first use.
 }];
 }
 
+def TargetClonesDocs : Documentation {
+  let Category = DocCatFunction;
+  let Content = [{
+Clang supports the ``target_clones("OPTIONS")`` attribute. This attribute may be
+attached to a function declaration and causes function multiversioning, where
+multiple versions of the function will be emitted with different code
+generation options.  Additionally, these versions will be resolved at runtime
+based on the priority of their attribute options. All ``target_clone`` functions
+are considered multiversioned functions.
+
+All multiversioned functions must contain a ``default`` (fallback)
+implementation, otherwise usages of the function are considered invalid.
+Additionally, a function may not become multiversioned after its first use.
+
+The options to ``target_clones`` can either be a target-specific architecture
+(specified as ``arch=CPU``), or one of a list of subtarget features.
+
+Example "subtarget features" from the x86 backend include: "mmx", "sse", "sse4.2",
+"avx", "xop" and largely correspond to the machine specific options handled by
+the front end.
+
+The versions can either be listed as a comma-separated sequence of string
+literals or as a single string literal containing a comma-separated list of
+versions.  For compatibility with GCC, the two formats can be mixed.  For
+example, the following will emit 4 versions of the function:
+
+  .. code-block:: c++
+
+    __attribute__((target_clones("arch=atom,avx2","arch=ivybridge","default")))
+    void foo() {}
+
+}];
+}
+
 def MinVectorWidthDocs : Documentation {
   let Category = DocCatFunction;
   let Content = [{

clang/include/clang/Basic/DiagnosticGroups.td

@@ -1275,9 +1275,14 @@ def : DiagGroup<"spirv-compat", [SpirCompat]>; // Alias.
 // Warning for the GlobalISel options.
 def GlobalISel : DiagGroup<"global-isel">;
 
+// A warning group for the GNU extension to allow mixed specifier types for
+// target-clones multiversioning.
+def TargetClonesMixedSpecifiers : DiagGroup<"target-clones-mixed-specifiers">;
+
 // A warning group specifically for warnings related to function
 // multiversioning.
-def FunctionMultiVersioning : DiagGroup<"function-multiversion">;
+def FunctionMultiVersioning
+    : DiagGroup<"function-multiversion", [TargetClonesMixedSpecifiers]>;
 
 def NoDeref : DiagGroup<"noderef">;
 

clang/include/clang/Basic/DiagnosticSemaKinds.td

@@ -2981,7 +2981,8 @@ def err_invalid_branch_protection_spec : Error<
   "invalid or misplaced branch protection specification '%0'">;
 def warn_unsupported_target_attribute
     : Warning<"%select{unsupported|duplicate|unknown}0%select{| architecture|"
-              " tune CPU}1 '%2' in the 'target' attribute string; 'target' "
+              " tune CPU}1 '%2' in the '%select{target|target_clones}3' "
+              "attribute string; '%select{target|target_clones}3' "
               "attribute ignored">,
       InGroup<IgnoredAttributes>;
 def err_attribute_unsupported
@@ -9864,6 +9865,8 @@ def warn_duplicate_attribute_exact : Warning<
 def warn_duplicate_attribute : Warning<
   "attribute %0 is already applied with different arguments">,
   InGroup<IgnoredAttributes>;
+def err_disallowed_duplicate_attribute : Error<
+  "attribute %0 cannot appear more than once on a declaration">;
 
 def warn_sync_fetch_and_nand_semantics_change : Warning<
   "the semantics of this intrinsic changed with GCC "
@@ -11254,21 +11257,26 @@ def err_multiversion_duplicate : Error<
   "multiversioned function redeclarations require identical target attributes">;
 def err_multiversion_noproto : Error<
   "multiversioned function must have a prototype">;
-def err_multiversion_disallowed_other_attr : Error<
-  "attribute '%select{target|cpu_specific|cpu_dispatch}0' multiversioning cannot be combined"
-  " with attribute %1">;
+def err_multiversion_disallowed_other_attr
+    : Error<"attribute "
+            "'%select{|target|cpu_specific|cpu_dispatch|target_clones}0' "
+            "multiversioning cannot be combined"
+            " with attribute %1">;
 def err_multiversion_mismatched_attrs
     : Error<"attributes on multiversioned functions must all match, attribute "
             "%0 %select{is missing|has different arguments}1">;
 def err_multiversion_diff : Error<
   "multiversioned function declaration has a different %select{calling convention"
   "|return type|constexpr specification|inline specification|linkage|"
   "language linkage}0">;
-def err_multiversion_doesnt_support : Error<
-  "attribute '%select{target|cpu_specific|cpu_dispatch}0' multiversioned functions do not "
-  "yet support %select{function templates|virtual functions|"
-  "deduced return types|constructors|destructors|deleted functions|"
-  "defaulted functions|constexpr functions|consteval function}1">;
+def err_multiversion_doesnt_support
+    : Error<"attribute "
+            "'%select{|target|cpu_specific|cpu_dispatch|target_clones}0' "
+            "multiversioned functions do not "
+            "yet support %select{function templates|virtual functions|"
+            "deduced return types|constructors|destructors|deleted functions|"
+            "defaulted functions|constexpr functions|consteval "
+            "function|lambdas}1">;
 def err_multiversion_not_allowed_on_main : Error<
   "'main' cannot be a multiversioned function">;
 def err_multiversion_not_supported : Error<
@@ -11285,6 +11293,19 @@ def warn_multiversion_duplicate_entries : Warning<
 def warn_dispatch_body_ignored : Warning<
   "body of cpu_dispatch function will be ignored">,
   InGroup<FunctionMultiVersioning>;
+def err_target_clone_must_have_default
+    : Error<"'target_clones' multiversioning requires a default target">;
+def err_target_clone_doesnt_match
+    : Error<"'target_clones' attribute does not match previous declaration">;
+def warn_target_clone_mixed_values
+    : ExtWarn<
+          "mixing 'target_clones' specifier mechanisms is permitted for GCC "
+          "compatibility; use a comma separated sequence of string literals, "
+          "or a string literal containing a comma-separated list of versions">,
+      InGroup<TargetClonesMixedSpecifiers>;
+def warn_target_clone_duplicate_options
+    : Warning<"version list contains duplicate entries">,
+      InGroup<FunctionMultiVersioning>;
 
 // three-way comparison operator diagnostics
 def err_implied_comparison_category_type_not_found : Error<

clang/include/clang/Sema/Sema.h

@@ -4351,6 +4351,10 @@ class Sema final {
   llvm::Error isValidSectionSpecifier(StringRef Str);
   bool checkSectionName(SourceLocation LiteralLoc, StringRef Str);
   bool checkTargetAttr(SourceLocation LiteralLoc, StringRef Str);
+  bool checkTargetClonesAttrString(SourceLocation LiteralLoc, StringRef Str,
+                                   const StringLiteral *Literal,
+                                   bool &HasDefault, bool &HasCommas,
+                                   SmallVectorImpl<StringRef> &Strings);
   bool checkMSInheritanceAttrOnDefinition(
       CXXRecordDecl *RD, SourceRange Range, bool BestCase,
       MSInheritanceModel SemanticSpelling);

clang/lib/AST/ASTContext.cpp

@@ -11800,6 +11800,15 @@ void ASTContext::getFunctionFeatureMap(llvm::StringMap<bool> &FeatureMap,
                     Target->getTargetOpts().FeaturesAsWritten.begin(),
                     Target->getTargetOpts().FeaturesAsWritten.end());
     Target->initFeatureMap(FeatureMap, getDiagnostics(), TargetCPU, Features);
+  } else if (const auto *TC = FD->getAttr<TargetClonesAttr>()) {
+    std::vector<std::string> Features;
+    StringRef VersionStr = TC->getFeatureStr(GD.getMultiVersionIndex());
+    if (VersionStr.startswith("arch="))
+      TargetCPU = VersionStr.drop_front(sizeof("arch=") - 1);
+    else if (VersionStr != "default")
+      Features.push_back((StringRef{"+"} + VersionStr).str());
+
+    Target->initFeatureMap(FeatureMap, getDiagnostics(), TargetCPU, Features);
   } else {
     FeatureMap = Target->getTargetOpts().FeatureMap;
   }

clang/lib/AST/Decl.cpp

@@ -3271,6 +3271,8 @@ MultiVersionKind FunctionDecl::getMultiVersionKind() const {
     return MultiVersionKind::CPUDispatch;
   if (hasAttr<CPUSpecificAttr>())
     return MultiVersionKind::CPUSpecific;
+  if (hasAttr<TargetClonesAttr>())
+    return MultiVersionKind::TargetClones;
   return MultiVersionKind::None;
 }
 
@@ -3286,6 +3288,10 @@ bool FunctionDecl::isTargetMultiVersion() const {
   return isMultiVersion() && hasAttr<TargetAttr>();
 }
 
+bool FunctionDecl::isTargetClonesMultiVersion() const {
+  return isMultiVersion() && hasAttr<TargetClonesAttr>();
+}
+
 void
 FunctionDecl::setPreviousDeclaration(FunctionDecl *PrevDecl) {
   redeclarable_base::setPreviousDecl(PrevDecl);

clang/lib/CodeGen/CodeGenModule.cpp

@@ -1266,6 +1266,20 @@ static bool isUniqueInternalLinkageDecl(GlobalDecl GD,
          (CGM.getFunctionLinkage(GD) == llvm::GlobalValue::InternalLinkage);
 }
 
+static void AppendTargetClonesMangling(const CodeGenModule &CGM,
+                                       const TargetClonesAttr *Attr,
+                                       unsigned VersionIndex,
+                                       raw_ostream &Out) {
+  Out << '.';
+  StringRef FeatureStr = Attr->getFeatureStr(VersionIndex);
+  if (FeatureStr.startswith("arch="))
+    Out << "arch_" << FeatureStr.substr(sizeof("arch=") - 1);
+  else
+    Out << FeatureStr;
+
+  Out << '.' << Attr->getMangledIndex(VersionIndex);
+}
+
 static std::string getMangledNameImpl(CodeGenModule &CGM, GlobalDecl GD,
                                       const NamedDecl *ND,
                                       bool OmitMultiVersionMangling = false) {
@@ -1319,6 +1333,10 @@ static std::string getMangledNameImpl(CodeGenModule &CGM, GlobalDecl GD,
       case MultiVersionKind::Target:
         AppendTargetMangling(CGM, FD->getAttr<TargetAttr>(), Out);
         break;
+      case MultiVersionKind::TargetClones:
+        AppendTargetClonesMangling(CGM, FD->getAttr<TargetClonesAttr>(),
+                                   GD.getMultiVersionIndex(), Out);
+        break;
       case MultiVersionKind::None:
         llvm_unreachable("None multiversion type isn't valid here");
       }
@@ -1983,8 +2001,9 @@ bool CodeGenModule::GetCPUAndFeaturesAttributes(GlobalDecl GD,
   FD = FD ? FD->getMostRecentDecl() : FD;
   const auto *TD = FD ? FD->getAttr<TargetAttr>() : nullptr;
   const auto *SD = FD ? FD->getAttr<CPUSpecificAttr>() : nullptr;
+  const auto *TC = FD ? FD->getAttr<TargetClonesAttr>() : nullptr;
   bool AddedAttr = false;
-  if (TD || SD) {
+  if (TD || SD || TC) {
     llvm::StringMap<bool> FeatureMap;
     getContext().getFunctionFeatureMap(FeatureMap, GD);
 
@@ -3226,6 +3245,12 @@ void CodeGenModule::EmitMultiVersionFunctionDefinition(GlobalDecl GD,
     for (unsigned I = 0; I < Spec->cpus_size(); ++I)
       EmitGlobalFunctionDefinition(GD.getWithMultiVersionIndex(I), nullptr);
     // Requires multiple emits.
+  } else if (FD->isTargetClonesMultiVersion()) {
+    auto *Clone = FD->getAttr<TargetClonesAttr>();
+    for (unsigned I = 0; I < Clone->featuresStrs_size(); ++I)
+      if (Clone->isFirstOfVersion(I))
+        EmitGlobalFunctionDefinition(GD.getWithMultiVersionIndex(I), nullptr);
+    EmitTargetClonesResolver(GD);
   } else
     EmitGlobalFunctionDefinition(GD, GV);
 }
@@ -3307,6 +3332,63 @@ llvm::GlobalValue::LinkageTypes getMultiversionLinkage(CodeGenModule &CGM,
   return llvm::GlobalValue::WeakODRLinkage;
 }
 
+void CodeGenModule::EmitTargetClonesResolver(GlobalDecl GD) {
+  const auto *FD = cast<FunctionDecl>(GD.getDecl());
+  assert(FD && "Not a FunctionDecl?");
+  const auto *TC = FD->getAttr<TargetClonesAttr>();
+  assert(TC && "Not a target_clones Function?");
+
+  QualType CanonTy = Context.getCanonicalType(FD->getType());
+  llvm::Type *DeclTy = getTypes().ConvertType(CanonTy);
+
+  if (const auto *CXXFD = dyn_cast<CXXMethodDecl>(FD)) {
+    const CGFunctionInfo &FInfo = getTypes().arrangeCXXMethodDeclaration(CXXFD);
+    DeclTy = getTypes().GetFunctionType(FInfo);
+  }
+
+  llvm::Function *ResolverFunc;
+  if (getTarget().supportsIFunc()) {
+    auto *IFunc = cast<llvm::GlobalIFunc>(
+        GetOrCreateMultiVersionResolver(GD, DeclTy, FD));
+    ResolverFunc = cast<llvm::Function>(IFunc->getResolver());
+  } else
+    ResolverFunc =
+        cast<llvm::Function>(GetOrCreateMultiVersionResolver(GD, DeclTy, FD));
+
+  SmallVector<CodeGenFunction::MultiVersionResolverOption, 10> Options;
+  for (unsigned VersionIndex = 0; VersionIndex < TC->featuresStrs_size();
+       ++VersionIndex) {
+    if (!TC->isFirstOfVersion(VersionIndex))
+      continue;
+    StringRef Version = TC->getFeatureStr(VersionIndex);
+    StringRef MangledName =
+        getMangledName(GD.getWithMultiVersionIndex(VersionIndex));
+    llvm::Constant *Func = GetGlobalValue(MangledName);
+    assert(Func &&
+           "Should have already been created before calling resolver emit");
+
+    StringRef Architecture;
+    llvm::SmallVector<StringRef, 1> Feature;
+
+    if (Version.startswith("arch="))
+      Architecture = Version.drop_front(sizeof("arch=") - 1);
+    else if (Version != "default")
+      Feature.push_back(Version);
+
+    Options.emplace_back(cast<llvm::Function>(Func), Architecture, Feature);
+  }
+
+  const TargetInfo &TI = getTarget();
+  std::stable_sort(
+      Options.begin(), Options.end(),
+      [&TI](const CodeGenFunction::MultiVersionResolverOption &LHS,
+            const CodeGenFunction::MultiVersionResolverOption &RHS) {
+        return TargetMVPriority(TI, LHS) > TargetMVPriority(TI, RHS);
+      });
+  CodeGenFunction CGF(*this);
+  CGF.EmitMultiVersionResolver(ResolverFunc, Options);
+}
+
 void CodeGenModule::emitMultiVersionFunctions() {
   std::vector<GlobalDecl> MVFuncsToEmit;
   MultiVersionFuncs.swap(MVFuncsToEmit);
@@ -3511,8 +3593,25 @@ llvm::Constant *CodeGenModule::GetOrCreateMultiVersionResolver(
   // Since this is the first time we've created this IFunc, make sure
   // that we put this multiversioned function into the list to be
   // replaced later if necessary (target multiversioning only).
-  if (!FD->isCPUDispatchMultiVersion() && !FD->isCPUSpecificMultiVersion())
+  if (FD->isTargetMultiVersion())
     MultiVersionFuncs.push_back(GD);
+  else if (FD->isTargetClonesMultiVersion()) {
+    // In target_clones multiversioning, make sure we emit this if used.
+    auto DDI =
+        DeferredDecls.find(getMangledName(GD.getWithMultiVersionIndex(0)));
+    if (DDI != DeferredDecls.end()) {
+      addDeferredDeclToEmit(GD);
+      DeferredDecls.erase(DDI);
+    } else {
+      // Emit the symbol of the 1st variant, so that the deferred decls know we
+      // need it, otherwise the only global value will be the resolver/ifunc,
+      // which end up getting broken if we search for them with GetGlobalValue'.
+      GetOrCreateLLVMFunction(
+          getMangledName(GD.getWithMultiVersionIndex(0)), DeclTy, FD,
+          /*ForVTable=*/false, /*DontDefer=*/true,
+          /*IsThunk=*/false, llvm::AttributeList(), ForDefinition);
+    }
+  }
 
   if (getTarget().supportsIFunc()) {
     llvm::Type *ResolverType = llvm::FunctionType::get(

clang/lib/CodeGen/CodeGenModule.h

@@ -1500,6 +1500,7 @@ class CodeGenModule : public CodeGenTypeCache {
   void EmitAliasDefinition(GlobalDecl GD);
   void emitIFuncDefinition(GlobalDecl GD);
   void emitCPUDispatchDefinition(GlobalDecl GD);
+  void EmitTargetClonesResolver(GlobalDecl GD);
   void EmitObjCPropertyImplementations(const ObjCImplementationDecl *D);
   void EmitObjCIvarInitializations(ObjCImplementationDecl *D);