[RISCV][FMV] Support target_clones

clang/include/clang/Basic/DiagnosticFrontendKinds.td

@@ -374,4 +374,8 @@ def warn_missing_symbol_graph_dir : Warning<
 def err_ast_action_on_llvm_ir : Error<
   "cannot apply AST actions to LLVM IR file '%0'">,
   DefaultFatal;
+
+def err_os_unsupport_riscv_target_clones : Error<
+  "target_clones is currently only supported on Linux">;
+
 }

clang/include/clang/Basic/TargetInfo.h

@@ -1481,7 +1481,8 @@ class TargetInfo : public TransferrableTargetInfo,
   /// Identify whether this target supports multiversioning of functions,
   /// which requires support for cpu_supports and cpu_is functionality.
   bool supportsMultiVersioning() const {
-    return getTriple().isX86() || getTriple().isAArch64();
+    return getTriple().isX86() || getTriple().isAArch64() ||
+           getTriple().isRISCV();
   }
 
   /// Identify whether this target supports IFuncs.

clang/lib/AST/ASTContext.cpp

@@ -13744,6 +13744,16 @@ void ASTContext::getFunctionFeatureMap(llvm::StringMap<bool> &FeatureMap,
       Features.insert(Features.begin(),
                       Target->getTargetOpts().FeaturesAsWritten.begin(),
                       Target->getTargetOpts().FeaturesAsWritten.end());
+    } else if (Target->getTriple().isRISCV()) {
+      StringRef VersionStr = TC->getFeatureStr(GD.getMultiVersionIndex());
+      if (VersionStr != "default") {
+        ParsedTargetAttr ParsedAttr = Target->parseTargetAttr(VersionStr);
+        Features.insert(Features.begin(), ParsedAttr.Features.begin(),
+                        ParsedAttr.Features.end());
+      }
+      Features.insert(Features.begin(),
+                      Target->getTargetOpts().FeaturesAsWritten.begin(),
+                      Target->getTargetOpts().FeaturesAsWritten.end());
     } else {
       StringRef VersionStr = TC->getFeatureStr(GD.getMultiVersionIndex());
       if (VersionStr.starts_with("arch="))

clang/lib/Basic/Targets/RISCV.cpp

@@ -257,7 +257,7 @@ bool RISCVTargetInfo::initFeatureMap(
 
   // If a target attribute specified a full arch string, override all the ISA
   // extension target features.
-  const auto I = llvm::find(FeaturesVec, "__RISCV_TargetAttrNeedOverride");
+  const auto I = llvm::find(FeaturesVec, "+__RISCV_TargetAttrNeedOverride");
   if (I != FeaturesVec.end()) {
     std::vector<std::string> OverrideFeatures(std::next(I), FeaturesVec.end());
 
@@ -367,6 +367,12 @@ bool RISCVTargetInfo::handleTargetFeatures(std::vector<std::string> &Features,
   return true;
 }
 
+bool RISCVTargetInfo::isValidFeatureName(StringRef Feature) const {
+  if (Feature.starts_with("__RISCV_TargetAttrNeedOverride"))
+    return true;
+  return llvm::RISCVISAInfo::isSupportedExtensionFeature(Feature);
+}
+
 bool RISCVTargetInfo::isValidCPUName(StringRef Name) const {
   bool Is64Bit = getTriple().isArch64Bit();
   return llvm::RISCV::parseCPU(Name, Is64Bit);
@@ -391,7 +397,7 @@ void RISCVTargetInfo::fillValidTuneCPUList(
 
 static void handleFullArchString(StringRef FullArchStr,
                                  std::vector<std::string> &Features) {
-  Features.push_back("__RISCV_TargetAttrNeedOverride");
+  Features.push_back("+__RISCV_TargetAttrNeedOverride");
   auto RII = llvm::RISCVISAInfo::parseArchString(
       FullArchStr, /* EnableExperimentalExtension */ true);
   if (llvm::errorToBool(RII.takeError())) {

clang/lib/Basic/Targets/RISCV.h

@@ -106,6 +106,8 @@ class RISCVTargetInfo : public TargetInfo {
   bool handleTargetFeatures(std::vector<std::string> &Features,
                             DiagnosticsEngine &Diags) override;
 
+  bool isValidFeatureName(StringRef Feature) const override;
+
   bool hasBitIntType() const override { return true; }
 
   bool hasBFloat16Type() const override { return true; }

clang/lib/CodeGen/CGBuiltin.cpp

@@ -62,6 +62,7 @@
 #include "llvm/Support/MathExtras.h"
 #include "llvm/Support/ScopedPrinter.h"
 #include "llvm/TargetParser/AArch64TargetParser.h"
+#include "llvm/TargetParser/RISCVTargetParser.h"
 #include "llvm/TargetParser/X86TargetParser.h"
 #include <optional>
 #include <sstream>
@@ -14174,6 +14175,16 @@ Value *CodeGenFunction::EmitAArch64CpuInit() {
   return Builder.CreateCall(Func);
 }
 
+Value *CodeGenFunction::EmitRISCVCpuInit() {
+  llvm::FunctionType *FTy = llvm::FunctionType::get(VoidTy, false);
+  llvm::FunctionCallee Func =
+      CGM.CreateRuntimeFunction(FTy, "__init_riscv_features_bit");
+  cast<llvm::GlobalValue>(Func.getCallee())->setDSOLocal(true);
+  cast<llvm::GlobalValue>(Func.getCallee())
+      ->setDLLStorageClass(llvm::GlobalValue::DefaultStorageClass);
+  return Builder.CreateCall(Func);
+}
+
 Value *CodeGenFunction::EmitX86CpuInit() {
   llvm::FunctionType *FTy = llvm::FunctionType::get(VoidTy,
                                                     /*Variadic*/ false);
@@ -14226,6 +14237,73 @@ CodeGenFunction::EmitAArch64CpuSupports(ArrayRef<StringRef> FeaturesStrs) {
   return Result;
 }
 
+Value *CodeGenFunction::EmitRISCVCpuSupports(ArrayRef<StringRef> FeaturesStrs,
+                                             unsigned &MaxGroupIDUsed) {
+
+  const unsigned FeatureBitSize = 2;
+  llvm::ArrayType *ArrayOfInt64Ty =
+      llvm::ArrayType::get(Int64Ty, FeatureBitSize);
+  llvm::Type *StructTy = llvm::StructType::get(Int32Ty, ArrayOfInt64Ty);
+  llvm::Constant *RISCVFeaturesBits =
+      CGM.CreateRuntimeVariable(StructTy, "__riscv_feature_bits");
+  cast<llvm::GlobalValue>(RISCVFeaturesBits)->setDSOLocal(true);
+
+  auto LoadFeatureBit = [&](unsigned Index) {
+    // Create GEP then load.
+    Value *IndexVal = llvm::ConstantInt::get(Int32Ty, Index);
+    std::vector<llvm::Value *> GEPIndices = {llvm::ConstantInt::get(Int32Ty, 0),
+                                             llvm::ConstantInt::get(Int32Ty, 1),
+                                             IndexVal};
+    Value *Ptr =
+        Builder.CreateInBoundsGEP(StructTy, RISCVFeaturesBits, GEPIndices);
+    Value *FeaturesBit =
+        Builder.CreateAlignedLoad(Int64Ty, Ptr, CharUnits::fromQuantity(8));
+    return FeaturesBit;
+  };
+
+  SmallVector<unsigned long long> RequireFeatureBits =
+      llvm::RISCV::getRequireFeatureBitMask(FeaturesStrs);
+  Value *Result = Builder.getTrue();
+  for (unsigned i = 0; i < RequireFeatureBits.size(); i++) {
+    if (!RequireFeatureBits[i])
+      continue;
+    MaxGroupIDUsed = i;
+    Value *Mask = Builder.getInt64(RequireFeatureBits[i]);
+    Value *Bitset = Builder.CreateAnd(LoadFeatureBit(i), Mask);
+    Value *Cmp = Builder.CreateICmpEQ(Bitset, Mask);
+    Result = Builder.CreateAnd(Result, Cmp);
+  }
+
+  return Result;
+}
+
+Value *CodeGenFunction::EmitRISCVFeatureBitsLength(unsigned MaxGroupIDUsed) {
+
+  const unsigned FeatureBitSize = 2;
+  llvm::ArrayType *ArrayOfInt64Ty =
+      llvm::ArrayType::get(Int64Ty, FeatureBitSize);
+  llvm::Type *StructTy = llvm::StructType::get(Int32Ty, ArrayOfInt64Ty);
+  llvm::Constant *RISCVFeaturesBits =
+      CGM.CreateRuntimeVariable(StructTy, "__riscv_feature_bits");
+  cast<llvm::GlobalValue>(RISCVFeaturesBits)->setDSOLocal(true);
+
+  auto LoadMaxGroupID = [&]() {
+    std::vector<llvm::Value *> GEPIndices = {
+        llvm::ConstantInt::get(Int32Ty, 0), llvm::ConstantInt::get(Int32Ty, 0)};
+    llvm::Value *Ptr =
+        Builder.CreateInBoundsGEP(StructTy, RISCVFeaturesBits, GEPIndices);
+    Value *Length =
+        Builder.CreateAlignedLoad(Int64Ty, Ptr, CharUnits::fromQuantity(8));
+    return Length;
+  };
+
+  Value *UsedMaxGroupID = Builder.getInt64(MaxGroupIDUsed);
+  Value *GroupIDResult =
+      Builder.CreateICmpULT(UsedMaxGroupID, LoadMaxGroupID());
+
+  return GroupIDResult;
+}
+
 Value *CodeGenFunction::EmitX86BuiltinExpr(unsigned BuiltinID,
                                            const CallExpr *E) {
   if (BuiltinID == Builtin::BI__builtin_cpu_is)

clang/lib/CodeGen/CodeGenFunction.cpp

@@ -2850,10 +2850,113 @@ void CodeGenFunction::EmitMultiVersionResolver(
   case llvm::Triple::aarch64:
     EmitAArch64MultiVersionResolver(Resolver, Options);
     return;
+  case llvm::Triple::riscv32:
+  case llvm::Triple::riscv64:
+    EmitRISCVMultiVersionResolver(Resolver, Options);
+    return;
 
   default:
-    assert(false && "Only implemented for x86 and AArch64 targets");
+    assert(false && "Only implemented for x86, AArch64 and RISC-V targets");
+  }
+}
+
+void CodeGenFunction::EmitRISCVMultiVersionResolver(
+    llvm::Function *Resolver, ArrayRef<MultiVersionResolverOption> Options) {
+
+  if (getContext().getTargetInfo().getTriple().getOS() !=
+      llvm::Triple::OSType::Linux) {
+    CGM.getDiags().Report(diag::err_os_unsupport_riscv_target_clones);
+    return;
+  }
+
+  llvm::BasicBlock *CurBlock = createBasicBlock("resolver_entry", Resolver);
+  Builder.SetInsertPoint(CurBlock);
+  EmitRISCVCpuInit();
+
+  bool SupportsIFunc = getContext().getTargetInfo().supportsIFunc();
+  bool HasDefault = false;
+  int DefaultIndex = 0;
+  // Check the each candidate function.
+  for (unsigned Index = 0; Index < Options.size(); Index++) {
+
+    if (Options[Index].Conditions.Features[0].starts_with("default")) {
+      HasDefault = true;
+      DefaultIndex = Index;
+      continue;
+    }
+
+    Builder.SetInsertPoint(CurBlock);
+
+    std::vector<std::string> TargetAttrFeats =
+        getContext()
+            .getTargetInfo()
+            .parseTargetAttr(Options[Index].Conditions.Features[0])
+            .Features;
+
+    if (!TargetAttrFeats.empty()) {
+      // If this function doens't need override, then merge with module level
+      // target features. Otherwise, remain the current target features.
+      auto I = llvm::find(TargetAttrFeats, "+__RISCV_TargetAttrNeedOverride");
+      if (I == TargetAttrFeats.end())
+        TargetAttrFeats.insert(TargetAttrFeats.begin(),
+                               Target.getTargetOpts().FeaturesAsWritten.begin(),
+                               Target.getTargetOpts().FeaturesAsWritten.end());
+      else
+        TargetAttrFeats.erase(I);
+
+      // Only consider +<extension-feature>.
+      llvm::SmallVector<StringRef, 8> PlusTargetAttrFeats;
+      for (StringRef Feat : TargetAttrFeats) {
+        if (!getContext().getTargetInfo().isValidFeatureName(
+                Feat.substr(1).str()))
+          continue;
+        if (Feat.starts_with("+"))
+          PlusTargetAttrFeats.push_back(Feat.substr(1));
+      }
+
+      llvm::BasicBlock *SecondCond =
+          createBasicBlock("resovler_cond", Resolver);
+
+      Builder.SetInsertPoint(SecondCond);
+      unsigned MaxGroupIDUsed = 0;
+      llvm::Value *SecondCondition =
+          EmitRISCVCpuSupports(PlusTargetAttrFeats, MaxGroupIDUsed);
+
+      Builder.SetInsertPoint(CurBlock);
+      llvm::Value *FirstCondition = EmitRISCVFeatureBitsLength(MaxGroupIDUsed);
+
+      llvm::BasicBlock *RetBlock =
+          createBasicBlock("resolver_return", Resolver);
+      CGBuilderTy RetBuilder(*this, RetBlock);
+      CreateMultiVersionResolverReturn(CGM, Resolver, RetBuilder,
+                                       Options[Index].Function, SupportsIFunc);
+      llvm::BasicBlock *ElseBlock = createBasicBlock("resolver_else", Resolver);
+
+      Builder.SetInsertPoint(CurBlock);
+      Builder.CreateCondBr(FirstCondition, SecondCond, ElseBlock);
+
+      Builder.SetInsertPoint(SecondCond);
+      Builder.CreateCondBr(SecondCondition, RetBlock, ElseBlock);
+
+      CurBlock = ElseBlock;
+    }
   }
+
+  // Finally, emit the default one.
+  if (HasDefault) {
+    Builder.SetInsertPoint(CurBlock);
+    CreateMultiVersionResolverReturn(
+        CGM, Resolver, Builder, Options[DefaultIndex].Function, SupportsIFunc);
+    return;
+  }
+
+  // If no generic/default, emit an unreachable.
+  Builder.SetInsertPoint(CurBlock);
+  llvm::CallInst *TrapCall = EmitTrapCall(llvm::Intrinsic::trap);
+  TrapCall->setDoesNotReturn();
+  TrapCall->setDoesNotThrow();
+  Builder.CreateUnreachable();
+  Builder.ClearInsertionPoint();
 }
 
 void CodeGenFunction::EmitAArch64MultiVersionResolver(

clang/lib/CodeGen/CodeGenFunction.h

@@ -5254,6 +5254,9 @@ class CodeGenFunction : public CodeGenTypeCache {
   void
   EmitAArch64MultiVersionResolver(llvm::Function *Resolver,
                                   ArrayRef<MultiVersionResolverOption> Options);
+  void
+  EmitRISCVMultiVersionResolver(llvm::Function *Resolver,
+                                ArrayRef<MultiVersionResolverOption> Options);
 
 private:
   QualType getVarArgType(const Expr *Arg);
@@ -5278,6 +5281,10 @@ class CodeGenFunction : public CodeGenTypeCache {
   FormAArch64ResolverCondition(const MultiVersionResolverOption &RO);
   llvm::Value *EmitAArch64CpuSupports(const CallExpr *E);
   llvm::Value *EmitAArch64CpuSupports(ArrayRef<StringRef> FeatureStrs);
+  llvm::Value *EmitRISCVCpuInit();
+  llvm::Value *EmitRISCVCpuSupports(ArrayRef<StringRef> FeatureStrs,
+                                    unsigned &MaxGroupIDUsed);
+  llvm::Value *EmitRISCVFeatureBitsLength(unsigned MaxGroupIDUsed);
 };
 
 inline DominatingLLVMValue::saved_type

clang/lib/CodeGen/CodeGenModule.cpp

@@ -4233,7 +4233,10 @@ void CodeGenModule::emitMultiVersionFunctions() {
               Feats.clear();
               if (getTarget().getTriple().isAArch64())
                 TC->getFeatures(Feats, I);
-              else {
+              else if (getTarget().getTriple().isRISCV()) {
+                StringRef Version = TC->getFeatureStr(I);
+                Feats.push_back(Version);
+              } else {
                 StringRef Version = TC->getFeatureStr(I);
                 if (Version.starts_with("arch="))
                   Architecture = Version.drop_front(sizeof("arch=") - 1);

clang/lib/CodeGen/Targets/RISCV.cpp

@@ -63,9 +63,32 @@ class RISCVABIInfo : public DefaultABIInfo {
                                                CharUnits Field2Off) const;
 
   ABIArgInfo coerceVLSVector(QualType Ty) const;
+
+  using ABIInfo::appendAttributeMangling;
+  void appendAttributeMangling(TargetClonesAttr *Attr, unsigned Index,
+                               raw_ostream &Out) const override;
+  void appendAttributeMangling(StringRef AttrStr,
+                               raw_ostream &Out) const override;
 };
 } // end anonymous namespace
 
+void RISCVABIInfo::appendAttributeMangling(TargetClonesAttr *Attr,
+                                           unsigned Index,
+                                           raw_ostream &Out) const {
+  appendAttributeMangling(Attr->getFeatureStr(Index), Out);
+}
+
+void RISCVABIInfo::appendAttributeMangling(StringRef AttrStr,
+                                           raw_ostream &Out) const {
+  if (AttrStr == "default") {
+    Out << ".default";
+    return;
+  }
+
+  Out << '.';
+  Out << AttrStr;
+}
+
 void RISCVABIInfo::computeInfo(CGFunctionInfo &FI) const {
   QualType RetTy = FI.getReturnType();
   if (!getCXXABI().classifyReturnType(FI))

clang/lib/Sema/SemaDeclAttr.cpp

@@ -3152,6 +3152,31 @@ bool Sema::checkTargetClonesAttrString(
                                  /*IncludeLocallyStreaming=*/false))
         return Diag(LiteralLoc,
                     diag::err_sme_streaming_cannot_be_multiversioned);
+    } else if (TInfo.getTriple().isRISCV()) {
+      // Suppress warn_target_clone_mixed_values
+      HasCommas = false;
+
+      if (Str.starts_with("arch=")) {
+        // parseTargetAttr will parse full version string,
+        // the following split Cur string is no longer interesting.
+        if ((!Cur.starts_with("arch=")))
+          continue;
+
+        ParsedTargetAttr TargetAttr =
+            Context.getTargetInfo().parseTargetAttr(Str);
+        if (TargetAttr.Features.empty())
+          return Diag(CurLoc, diag::warn_unsupported_target_attribute)
+                 << Unsupported << None << Str << TargetClones;
+      } else if (Str == "default") {
+        DefaultIsDupe = HasDefault;
+        HasDefault = true;
+      } else {
+        return Diag(CurLoc, diag::warn_unsupported_target_attribute)
+               << Unsupported << None << Str << TargetClones;
+      }
+      if (llvm::is_contained(StringsBuffer, Str) || DefaultIsDupe)
+        Diag(CurLoc, diag::warn_target_clone_duplicate_options);
+      StringsBuffer.push_back(Str);
     } else {
       // Other targets ( currently X86 )
       if (Cur.starts_with("arch=")) {

clang/test/CodeGen/attr-target-clones-riscv-invaild.c

@@ -0,0 +1,8 @@
+// RUN: not %clang_cc1 -triple riscv64 -target-feature +i -emit-llvm -o - %s 2>&1 | FileCheck %s --check-prefix=CHECK-UNSUPPORT-OS
+
+// CHECK-UNSUPPORT-OS: error: target_clones is currently only supported on Linux
+__attribute__((target_clones("default", "arch=+c"))) int foo2(void) {
+  return 2;
+}
+
+int bar() { return foo1()+foo2(); }