[Flang] Add code-object-version option (#72638)

Information about code object version can be configured by the user for
AMD GPU target and it needs to be placed in LLVM IR generated by Flang.

Information about code object version in MLIR generated by the parser
can be reused by other tools. There is no need to specify extra flags if
we want to invoke MLIR tools (like fir-opt) separately.

Changes in comparison to a8ac93:
 * added information about required targets for test
   flang/test/Driver/driver-help.f90

clang/include/clang/Basic/TargetOptions.h

@@ -78,17 +78,9 @@ class TargetOptions {
   /// \brief If enabled, allow AMDGPU unsafe floating point atomics.
   bool AllowAMDGPUUnsafeFPAtomics = false;
 
-  /// \brief Enumeration value for AMDGPU code object version, which is the
-  /// code object version times 100.
-  enum CodeObjectVersionKind {
-    COV_None,
-    COV_2 = 200, // Unsupported.
-    COV_3 = 300, // Unsupported.
-    COV_4 = 400,
-    COV_5 = 500,
-  };
   /// \brief Code object version for AMDGPU.
-  CodeObjectVersionKind CodeObjectVersion = CodeObjectVersionKind::COV_None;
+  llvm::CodeObjectVersionKind CodeObjectVersion =
+      llvm::CodeObjectVersionKind::COV_None;
 
   /// \brief Enumeration values for AMDGPU printf lowering scheme
   enum class AMDGPUPrintfKind {

clang/include/clang/Driver/Options.td

@@ -4721,9 +4721,9 @@ defm amdgpu_ieee : BoolOption<"m", "amdgpu-ieee",
 
 def mcode_object_version_EQ : Joined<["-"], "mcode-object-version=">, Group<m_Group>,
   HelpText<"Specify code object ABI version. Defaults to 4. (AMDGPU only)">,
-  Visibility<[ClangOption, CC1Option]>,
+  Visibility<[ClangOption, FlangOption, CC1Option, FC1Option]>,
   Values<"none,4,5">,
-  NormalizedValuesScope<"TargetOptions">,
+  NormalizedValuesScope<"llvm::CodeObjectVersionKind">,
   NormalizedValues<["COV_None", "COV_4", "COV_5"]>,
   MarshallingInfoEnum<TargetOpts<"CodeObjectVersion">, "COV_4">;
 

clang/lib/CodeGen/CGBuiltin.cpp

@@ -17588,7 +17588,7 @@ Value *EmitAMDGPUWorkGroupSize(CodeGenFunction &CGF, unsigned Index) {
 
   auto Cov = CGF.getTarget().getTargetOpts().CodeObjectVersion;
 
-  if (Cov == clang::TargetOptions::COV_None) {
+  if (Cov == CodeObjectVersionKind::COV_None) {
     StringRef Name = "__oclc_ABI_version";
     auto *ABIVersionC = CGF.CGM.getModule().getNamedGlobal(Name);
     if (!ABIVersionC)
@@ -17606,7 +17606,7 @@ Value *EmitAMDGPUWorkGroupSize(CodeGenFunction &CGF, unsigned Index) {
 
     Value *IsCOV5 = CGF.Builder.CreateICmpSGE(
         ABIVersion,
-        llvm::ConstantInt::get(CGF.Int32Ty, clang::TargetOptions::COV_5));
+        llvm::ConstantInt::get(CGF.Int32Ty, CodeObjectVersionKind::COV_5));
 
     // Indexing the implicit kernarg segment.
     Value *ImplicitGEP = CGF.Builder.CreateConstGEP1_32(
@@ -17621,7 +17621,7 @@ Value *EmitAMDGPUWorkGroupSize(CodeGenFunction &CGF, unsigned Index) {
         Address(Result, CGF.Int16Ty, CharUnits::fromQuantity(2)));
   } else {
     Value *GEP = nullptr;
-    if (Cov == clang::TargetOptions::COV_5) {
+    if (Cov == CodeObjectVersionKind::COV_5) {
       // Indexing the implicit kernarg segment.
       GEP = CGF.Builder.CreateConstGEP1_32(
           CGF.Int8Ty, EmitAMDGPUImplicitArgPtr(CGF), 12 + Index * 2);

clang/lib/CodeGen/CodeGenModule.cpp

@@ -847,7 +847,7 @@ void CodeGenModule::Release() {
     // Emit amdgpu_code_object_version module flag, which is code object version
     // times 100.
     if (getTarget().getTargetOpts().CodeObjectVersion !=
-        TargetOptions::COV_None) {
+        llvm::CodeObjectVersionKind::COV_None) {
       getModule().addModuleFlag(llvm::Module::Error,
                                 "amdgpu_code_object_version",
                                 getTarget().getTargetOpts().CodeObjectVersion);

clang/lib/CodeGen/Targets/AMDGPU.cpp

@@ -368,7 +368,7 @@ void AMDGPUTargetCodeGenInfo::emitTargetGlobals(
     return;
 
   if (CGM.getTarget().getTargetOpts().CodeObjectVersion ==
-      clang::TargetOptions::COV_None)
+      llvm::CodeObjectVersionKind::COV_None)
     return;
 
   auto *Type = llvm::IntegerType::getIntNTy(CGM.getModule().getContext(), 32);

clang/lib/Driver/ToolChains/Flang.cpp

@@ -277,6 +277,14 @@ static void processVSRuntimeLibrary(const ToolChain &TC, const ArgList &Args,
   }
 }
 
+void Flang::AddAMDGPUTargetArgs(const ArgList &Args,
+                                ArgStringList &CmdArgs) const {
+  if (Arg *A = Args.getLastArg(options::OPT_mcode_object_version_EQ)) {
+    StringRef Val = A->getValue();
+    CmdArgs.push_back(Args.MakeArgString("-mcode-object-version=" + Val));
+  }
+}
+
 void Flang::addTargetOptions(const ArgList &Args,
                              ArgStringList &CmdArgs) const {
   const ToolChain &TC = getToolChain();
@@ -300,6 +308,9 @@ void Flang::addTargetOptions(const ArgList &Args,
 
   case llvm::Triple::r600:
   case llvm::Triple::amdgcn:
+    getTargetFeatures(D, Triple, Args, CmdArgs, /*ForAs*/ false);
+    AddAMDGPUTargetArgs(Args, CmdArgs);
+    break;
   case llvm::Triple::riscv64:
   case llvm::Triple::x86_64:
     getTargetFeatures(D, Triple, Args, CmdArgs, /*ForAs*/ false);

clang/lib/Driver/ToolChains/Flang.h

@@ -63,6 +63,13 @@ class LLVM_LIBRARY_VISIBILITY Flang : public Tool {
   void AddAArch64TargetArgs(const llvm::opt::ArgList &Args,
                             llvm::opt::ArgStringList &CmdArgs) const;
 
+  /// Add specific options for AMDGPU target.
+  ///
+  /// \param [in] Args The list of input driver arguments
+  /// \param [out] CmdArgs The list of output command arguments
+  void AddAMDGPUTargetArgs(const llvm::opt::ArgList &Args,
+                           llvm::opt::ArgStringList &CmdArgs) const;
+
   /// Extract offload options from the driver arguments and add them to
   /// the command arguments.
   /// \param [in] C The current compilation for the driver invocation

flang/include/flang/Frontend/CodeGenOptions.h

@@ -85,6 +85,10 @@ class CodeGenOptions : public CodeGenOptionsBase {
     RK_WithPattern, // Remark pattern specified via '-Rgroup=regexp'.
   };
 
+  /// \brief Code object version for AMDGPU.
+  llvm::CodeObjectVersionKind CodeObjectVersion =
+      llvm::CodeObjectVersionKind::COV_4;
+
   /// Optimization remark with an optional regular expression pattern.
   struct OptRemark {
     RemarkKind Kind = RemarkKind::RK_Missing;

flang/lib/Frontend/CompilerInvocation.cpp

@@ -268,6 +268,17 @@ static void parseCodeGenArgs(Fortran::frontend::CodeGenOptions &opts,
       opts.PrepareForThinLTO = true;
   }
 
+  if (const llvm::opt::Arg *a = args.getLastArg(
+          clang::driver::options::OPT_mcode_object_version_EQ)) {
+    llvm::StringRef s = a->getValue();
+    if (s == "5")
+      opts.CodeObjectVersion = llvm::CodeObjectVersionKind::COV_5;
+    if (s == "4")
+      opts.CodeObjectVersion = llvm::CodeObjectVersionKind::COV_4;
+    if (s == "none")
+      opts.CodeObjectVersion = llvm::CodeObjectVersionKind::COV_None;
+  }
+
   // -f[no-]save-optimization-record[=<format>]
   if (const llvm::opt::Arg *a =
           args.getLastArg(clang::driver::options::OPT_opt_record_file))

flang/lib/Frontend/FrontendActions.cpp

@@ -244,8 +244,7 @@ static void setMLIRDataLayout(mlir::ModuleOp &mlirModule,
   mlirModule->setAttr(mlir::DLTIDialect::kDataLayoutAttrName, dlSpec);
 }
 
-static void addDepdendentLibs(mlir::ModuleOp &mlirModule,
-                              CompilerInstance &ci) {
+static void addDependentLibs(mlir::ModuleOp &mlirModule, CompilerInstance &ci) {
   const std::vector<std::string> &libs =
       ci.getInvocation().getCodeGenOpts().DependentLibs;
   if (libs.empty()) {
@@ -264,6 +263,68 @@ static void addDepdendentLibs(mlir::ModuleOp &mlirModule,
   }
 }
 
+// Add to MLIR code target specific items which are dependent on target
+// configuration specified by the user.
+// Clang equivalent function: AMDGPUTargetCodeGenInfo::emitTargetGlobals
+static void addAMDGPUSpecificMLIRItems(mlir::ModuleOp &mlirModule,
+                                       CompilerInstance &ci) {
+  const TargetOptions &targetOpts = ci.getInvocation().getTargetOpts();
+  const llvm::Triple triple(targetOpts.triple);
+  const llvm::StringRef codeObjectVersionGlobalOpName = "__oclc_ABI_version";
+
+  // TODO: Share address spaces enumeration between Clang and Flang.
+  // Currently this enumeration is defined in Clang specific class
+  // defined in file: clang/lib/Basic/Targets/AMDGPU.h .
+  // and we need to move it to LLVM directory.
+  const int constantAddressSpace = 4;
+
+  if (!triple.isAMDGPU()) {
+    return;
+  }
+  const CodeGenOptions &codeGenOpts = ci.getInvocation().getCodeGenOpts();
+  if (codeGenOpts.CodeObjectVersion == llvm::CodeObjectVersionKind::COV_None) {
+    return;
+  }
+
+  mlir::ConversionPatternRewriter builder(mlirModule.getContext());
+  unsigned oclcABIVERsion = codeGenOpts.CodeObjectVersion;
+  auto int32Type = builder.getI32Type();
+
+  std::optional<mlir::LLVM::GlobalOp> originalGV;
+
+  mlirModule.walk([&originalGV, codeObjectVersionGlobalOpName](
+                      mlir::LLVM::GlobalOp globalOp) {
+    if (globalOp.getName() == codeObjectVersionGlobalOpName)
+      originalGV = globalOp;
+  });
+  if (originalGV.has_value()) {
+    mlir::LLVM::GlobalOp originalGVOp = originalGV.value();
+    if (originalGVOp.getLinkage() != mlir::LLVM::Linkage::External) {
+      return;
+    }
+    // Update the variable if it is already present in MLIR but it was marked
+    // as external linkage variable
+    originalGVOp.setLinkage(mlir::LLVM::Linkage::WeakODR);
+    originalGVOp.setValueAttr(
+        builder.getIntegerAttr(int32Type, oclcABIVERsion));
+    originalGVOp.setUnnamedAddr(mlir::LLVM::UnnamedAddr::Local);
+    originalGVOp.setAddrSpace(constantAddressSpace);
+    originalGVOp.setVisibility_(mlir::LLVM::Visibility::Hidden);
+    return;
+  }
+
+  mlir::LLVM::GlobalOp covInfo = builder.create<mlir::LLVM::GlobalOp>(
+      /* Location */ mlirModule.getLoc(), /* Type */ int32Type,
+      /* IsConstant */ true, /* Linkage */ mlir::LLVM::Linkage::WeakODR,
+      /* Name */ codeObjectVersionGlobalOpName,
+      /* Value */ builder.getIntegerAttr(int32Type, oclcABIVERsion));
+  covInfo.setUnnamedAddr(mlir::LLVM::UnnamedAddr::Local);
+  covInfo.setAddrSpace(constantAddressSpace);
+  covInfo.setVisibility_(mlir::LLVM::Visibility::Hidden);
+  builder.setInsertionPointToStart(mlirModule.getBody());
+  builder.insert(covInfo);
+}
+
 bool CodeGenAction::beginSourceFileAction() {
   llvmCtx = std::make_unique<llvm::LLVMContext>();
   CompilerInstance &ci = this->getInstance();
@@ -365,8 +426,10 @@ bool CodeGenAction::beginSourceFileAction() {
   Fortran::parser::Program &parseTree{*ci.getParsing().parseTree()};
   lb.lower(parseTree, ci.getInvocation().getSemanticsContext());
 
-  // Add dependent libraries
-  addDepdendentLibs(*mlirModule, ci);
+  // Add target specific items like dependent libraries, target specific
+  // constants etc.
+  addDependentLibs(*mlirModule, ci);
+  addAMDGPUSpecificMLIRItems(*mlirModule, ci);
 
   // run the default passes.
   mlir::PassManager pm((*mlirModule)->getName(),

flang/test/Driver/code-object-version.f90

@@ -0,0 +1,9 @@
+! REQUIRES: x86-registered-target, amdgpu-registered-target
+! RUN: not %flang  -target amdgcn-amd-amdhsa -target-cpu gfx908 -mcode-object-version=3  -S  %s -o \
+! RUN:   /dev/null 2>&1 | FileCheck  --check-prefix=INVALID_VERSION %s
+
+! RUN: %flang  -target  x86_64-unknown-linux-gnu -mcode-object-version=3  -S  %s -o \
+! RUN:   /dev/null 2>&1 | FileCheck  --check-prefix=UNUSED_PARAM %s
+
+! INVALID_VERSION: error: invalid integral value '3' in '-mcode-object-version=3'
+! UNUSED_PARAM: warning: argument unused during compilation: '-mcode-object-version=3' [-Wunused-command-line-argument]

flang/test/Driver/driver-help-hidden.f90

@@ -114,6 +114,8 @@
 ! CHECK-NEXT: -I <dir>                Add directory to the end of the list of include search paths
 ! CHECK-NEXT: -L <dir>                Add directory to library search path
 ! CHECK-NEXT: -march=<value>          For a list of available architectures for the target use '-mcpu=help'
+! CHECK-NEXT: -mcode-object-version=<value>
+! CHECK-NEXT:                         Specify code object ABI version. Defaults to 4. (AMDGPU only)
 ! CHECK-NEXT: -mcpu=<value>           For a list of available CPUs for the target use '-mcpu=help'
 ! CHECK-NEXT: -mllvm=<arg>            Alias for -mllvm
 ! CHECK-NEXT: -mllvm <value>          Additional arguments to forward to LLVM's option processing

flang/test/Driver/driver-help.f90

@@ -100,6 +100,8 @@
 ! HELP-NEXT: -I <dir>                Add directory to the end of the list of include search paths
 ! HELP-NEXT: -L <dir>                Add directory to library search path
 ! HELP-NEXT: -march=<value>          For a list of available architectures for the target use '-mcpu=help'
+! HELP-NEXT: -mcode-object-version=<value>
+! HELP-NEXT:                         Specify code object ABI version. Defaults to 4. (AMDGPU only)
 ! HELP-NEXT: -mcpu=<value>           For a list of available CPUs for the target use '-mcpu=help'
 ! HELP-NEXT: -mllvm=<arg>            Alias for -mllvm
 ! HELP-NEXT: -mllvm <value>          Additional arguments to forward to LLVM's option processing
@@ -232,6 +234,8 @@
 ! HELP-FC1-NEXT: -init-only              Only execute frontend initialization
 ! HELP-FC1-NEXT: -I <dir>                Add directory to the end of the list of include search paths
 ! HELP-FC1-NEXT: -load <dsopath>         Load the named plugin (dynamic shared object)
+! HELP-FC1-NEXT: -mcode-object-version=<value>
+! HELP-FC1-NEXT:                         Specify code object ABI version. Defaults to 4. (AMDGPU only)
 ! HELP-FC1-NEXT: -menable-no-infs        Allow optimization to assume there are no infinities.
 ! HELP-FC1-NEXT: -menable-no-nans        Allow optimization to assume there are no NaNs.
 ! HELP-FC1-NEXT: -mllvm <value>          Additional arguments to forward to LLVM's option processing

flang/test/Lower/AMD/code-object-version.f90

@@ -0,0 +1,13 @@
+!REQUIRES: amdgpu-registered-target
+!RUN: %flang_fc1 -emit-hlfir -triple amdgcn-amd-amdhsa -target-cpu gfx908 %s -o - | FileCheck  --check-prefix=COV_DEFAULT %s
+!RUN: %flang_fc1 -emit-hlfir -triple amdgcn-amd-amdhsa -target-cpu gfx908 -mcode-object-version=none %s -o - | FileCheck  --check-prefix=COV_NONE %s
+!RUN: %flang_fc1 -emit-hlfir -triple amdgcn-amd-amdhsa -target-cpu gfx908 -mcode-object-version=4 %s -o - | FileCheck  --check-prefix=COV_4 %s
+!RUN: %flang_fc1 -emit-hlfir -triple amdgcn-amd-amdhsa -target-cpu gfx908 -mcode-object-version=5 %s -o - | FileCheck  --check-prefix=COV_5 %s
+
+!COV_DEFAULT: llvm.mlir.global weak_odr hidden local_unnamed_addr constant @__oclc_ABI_version(400 : i32) {addr_space = 4 : i32} : i32
+!COV_NONE-NOT: llvm.mlir.global weak_odr hidden local_unnamed_addr constant @__oclc_ABI_version(400 : i32) {addr_space = 4 : i32} : i32
+!COV_4: llvm.mlir.global weak_odr hidden local_unnamed_addr constant @__oclc_ABI_version(400 : i32) {addr_space = 4 : i32} : i32
+!COV_5: llvm.mlir.global weak_odr hidden local_unnamed_addr constant @__oclc_ABI_version(500 : i32) {addr_space = 4 : i32} : i32
+subroutine target_simple
+end subroutine target_simple
+

llvm/include/llvm/Target/TargetOptions.h

@@ -121,6 +121,16 @@ namespace llvm {
     Never,
   };
 
+  /// \brief Enumeration value for AMDGPU code object version, which is the
+  /// code object version times 100.
+  enum CodeObjectVersionKind {
+    COV_None,
+    COV_2 = 200, // Unsupported.
+    COV_3 = 300, // Unsupported.
+    COV_4 = 400,
+    COV_5 = 500,
+  };
+
   class TargetOptions {
   public:
     TargetOptions()