[OpenCL] Make global ctor init function a kernel

[Anastasia]

Summary

We need to be able to enqueue internal function that initializes global constructors on the host side. Therefore it has to be converted to a kernel.

Note, supporting destruction would need some more work. However, it seems global destruction has little meaning without any dynamic resource allocation on the device and program scope variables are destroyed by the runtime when program is released.

[Anastasia]

lib/CodeGen/CodeGenModule.cpp

1231

CFG -> CGF

[kpet]

lib/CodeGen/CGDeclCXX.cpp

583

functions
should be -> must be?
to the kernel -> to a kernel

584

in order that they may be launched by the host?

590

Agree that global destructors aren't that attractive a feature in most contexts. There are quite a few runtime issues with them too. We can think about this later.

593

Shouldn't the calling convention be SPIR_KERNEL if you want the function to be host-visible?

lib/CodeGen/CodeGenModule.h

1322

kernels generated implicitly
if a kernel | when a kernel

[Anastasia]

• Improved comments
• Switched to SPIR kernel

[kpet]

The comments could use a bit more polishing but nothing that justifies making another iteration IMHO. LGTM!

[Anastasia]

In D61488#1496154, @kpet wrote:

The comments could use a bit more polishing but nothing that justifies making another iteration IMHO. LGTM!

Sure, feel free to put up a patch if you like. :)

Thanks!

lib/CodeGen/CGDeclCXX.cpp

CodeGenModule::EmitCXXGlobalInitFunc() {
   }
 
   llvm::Function *Fn = CreateGlobalInitOrDestructFunction(
       FTy, llvm::Twine("_GLOBAL__sub_I_", FileName), FI);
 
   CodeGenFunction(*this).GenerateCXXGlobalInitFunc(Fn, CXXGlobalInits);
   AddGlobalCtor(Fn);
 
+  // In OpenCL global init functions must be converted to kernels in order to
+  // be able to launch them from the host.
+  // FIXME: Some more work might be needed to handle destructors correctly.
+  // Current initialization function makes use of function pointers callbacks.
+  // We can't support function pointers especially between host and device.
+  // However it seems global destruction has little meaning without any
+  // dynamic resource allocation on the device and program scope variables are
+  // destroyed by the runtime when program is released.
+  if (getLangOpts().OpenCL) {
+    GenOpenCLArgMetadata(Fn);
+    Fn->setCallingConv(llvm::CallingConv::SPIR_KERNEL);
+  }
+
   CXXGlobalInits.clear();
 }
 
 void CodeGenModule::EmitCXXGlobalDtorFunc() {
   if (CXXGlobalDtors.empty())
     return;
 
   llvm::FunctionType *FTy = llvm::FunctionType::get(VoidTy, false);

lib/CodeGen/CodeGenFunction.cpp

CodeGenFunction::DecodeAddrUsedInPrologue(llvm::Value *F,
   auto *GOTAsInt = Builder.CreateAdd(PCRelAsInt, FuncAsInt, "global_addr.int");
   auto *GOTAddr = Builder.CreateIntToPtr(GOTAsInt, Int8PtrPtrTy, "global_addr");
 
   // Load the original pointer through the global.
   return Builder.CreateLoad(Address(GOTAddr, getPointerAlign()),
                             "decoded_addr");
 }
 
-static void removeImageAccessQualifier(std::string& TyName) {
-  std::string ReadOnlyQual("__read_only");
-  std::string::size_type ReadOnlyPos = TyName.find(ReadOnlyQual);
-  if (ReadOnlyPos != std::string::npos)
-    // "+ 1" for the space after access qualifier.
-    TyName.erase(ReadOnlyPos, ReadOnlyQual.size() + 1);
-  else {
-    std::string WriteOnlyQual("__write_only");
-    std::string::size_type WriteOnlyPos = TyName.find(WriteOnlyQual);
-    if (WriteOnlyPos != std::string::npos)
-      TyName.erase(WriteOnlyPos, WriteOnlyQual.size() + 1);
-    else {
-      std::string ReadWriteQual("__read_write");
-      std::string::size_type ReadWritePos = TyName.find(ReadWriteQual);
-      if (ReadWritePos != std::string::npos)
-        TyName.erase(ReadWritePos, ReadWriteQual.size() + 1);
-    }
-  }
-}
-
-// Returns the address space id that should be produced to the
-// kernel_arg_addr_space metadata. This is always fixed to the ids
-// as specified in the SPIR 2.0 specification in order to differentiate
-// for example in clGetKernelArgInfo() implementation between the address
-// spaces with targets without unique mapping to the OpenCL address spaces
-// (basically all single AS CPUs).
-static unsigned ArgInfoAddressSpace(LangAS AS) {
-  switch (AS) {
-  case LangAS::opencl_global:   return 1;
-  case LangAS::opencl_constant: return 2;
-  case LangAS::opencl_local:    return 3;
-  case LangAS::opencl_generic:  return 4; // Not in SPIR 2.0 specs.
-  default:
-    return 0; // Assume private.
-  }
-}
-
-// OpenCL v1.2 s5.6.4.6 allows the compiler to store kernel argument
-// information in the program executable. The argument information stored
-// includes the argument name, its type, the address and access qualifiers used.
-static void GenOpenCLArgMetadata(const FunctionDecl *FD, llvm::Function *Fn,
-                                 CodeGenModule &CGM, llvm::LLVMContext &Context,
-                                 CGBuilderTy &Builder, ASTContext &ASTCtx) {
-  // Create MDNodes that represent the kernel arg metadata.
-  // Each MDNode is a list in the form of "key", N number of values which is
-  // the same number of values as their are kernel arguments.
-
-  const PrintingPolicy &Policy = ASTCtx.getPrintingPolicy();
-
-  // MDNode for the kernel argument address space qualifiers.
-  SmallVector<llvm::Metadata *, 8> addressQuals;
-
-  // MDNode for the kernel argument access qualifiers (images only).
-  SmallVector<llvm::Metadata *, 8> accessQuals;
-
-  // MDNode for the kernel argument type names.
-  SmallVector<llvm::Metadata *, 8> argTypeNames;
-
-  // MDNode for the kernel argument base type names.
-  SmallVector<llvm::Metadata *, 8> argBaseTypeNames;
-
-  // MDNode for the kernel argument type qualifiers.
-  SmallVector<llvm::Metadata *, 8> argTypeQuals;
-
-  // MDNode for the kernel argument names.
-  SmallVector<llvm::Metadata *, 8> argNames;
-
-  for (unsigned i = 0, e = FD->getNumParams(); i != e; ++i) {
-    const ParmVarDecl *parm = FD->getParamDecl(i);
-    QualType ty = parm->getType();
-    std::string typeQuals;
-
-    if (ty->isPointerType()) {
-      QualType pointeeTy = ty->getPointeeType();
-
-      // Get address qualifier.
-      addressQuals.push_back(llvm::ConstantAsMetadata::get(Builder.getInt32(
-        ArgInfoAddressSpace(pointeeTy.getAddressSpace()))));
-
-      // Get argument type name.
-      std::string typeName =
-          pointeeTy.getUnqualifiedType().getAsString(Policy) + "*";
-
-      // Turn "unsigned type" to "utype"
-      std::string::size_type pos = typeName.find("unsigned");
-      if (pointeeTy.isCanonical() && pos != std::string::npos)
-        typeName.erase(pos+1, 8);
-
-      argTypeNames.push_back(llvm::MDString::get(Context, typeName));
-
-      std::string baseTypeName =
-          pointeeTy.getUnqualifiedType().getCanonicalType().getAsString(
-              Policy) +
-          "*";
-
-      // Turn "unsigned type" to "utype"
-      pos = baseTypeName.find("unsigned");
-      if (pos != std::string::npos)
-        baseTypeName.erase(pos+1, 8);
-
-      argBaseTypeNames.push_back(llvm::MDString::get(Context, baseTypeName));
-
-      // Get argument type qualifiers:
-      if (ty.isRestrictQualified())
-        typeQuals = "restrict";
-      if (pointeeTy.isConstQualified() ||
-          (pointeeTy.getAddressSpace() == LangAS::opencl_constant))
-        typeQuals += typeQuals.empty() ? "const" : " const";
-      if (pointeeTy.isVolatileQualified())
-        typeQuals += typeQuals.empty() ? "volatile" : " volatile";
-    } else {
-      uint32_t AddrSpc = 0;
-      bool isPipe = ty->isPipeType();
-      if (ty->isImageType() || isPipe)
-        AddrSpc = ArgInfoAddressSpace(LangAS::opencl_global);
-
-      addressQuals.push_back(
-          llvm::ConstantAsMetadata::get(Builder.getInt32(AddrSpc)));
-
-      // Get argument type name.
-      std::string typeName;
-      if (isPipe)
-        typeName = ty.getCanonicalType()->getAs<PipeType>()->getElementType()
-                     .getAsString(Policy);
-      else
-        typeName = ty.getUnqualifiedType().getAsString(Policy);
-
-      // Turn "unsigned type" to "utype"
-      std::string::size_type pos = typeName.find("unsigned");
-      if (ty.isCanonical() && pos != std::string::npos)
-        typeName.erase(pos+1, 8);
-
-      std::string baseTypeName;
-      if (isPipe)
-        baseTypeName = ty.getCanonicalType()->getAs<PipeType>()
-                          ->getElementType().getCanonicalType()
-                          .getAsString(Policy);
-      else
-        baseTypeName =
-          ty.getUnqualifiedType().getCanonicalType().getAsString(Policy);
-
-      // Remove access qualifiers on images
-      // (as they are inseparable from type in clang implementation,
-      // but OpenCL spec provides a special query to get access qualifier
-      // via clGetKernelArgInfo with CL_KERNEL_ARG_ACCESS_QUALIFIER):
-      if (ty->isImageType()) {
-        removeImageAccessQualifier(typeName);
-        removeImageAccessQualifier(baseTypeName);
-      }
-
-      argTypeNames.push_back(llvm::MDString::get(Context, typeName));
-
-      // Turn "unsigned type" to "utype"
-      pos = baseTypeName.find("unsigned");
-      if (pos != std::string::npos)
-        baseTypeName.erase(pos+1, 8);
-
-      argBaseTypeNames.push_back(llvm::MDString::get(Context, baseTypeName));
-
-      if (isPipe)
-        typeQuals = "pipe";
-    }
-
-    argTypeQuals.push_back(llvm::MDString::get(Context, typeQuals));
-
-    // Get image and pipe access qualifier:
-    if (ty->isImageType()|| ty->isPipeType()) {
-      const Decl *PDecl = parm;
-      if (auto *TD = dyn_cast<TypedefType>(ty))
-        PDecl = TD->getDecl();
-      const OpenCLAccessAttr *A = PDecl->getAttr<OpenCLAccessAttr>();
-      if (A && A->isWriteOnly())
-        accessQuals.push_back(llvm::MDString::get(Context, "write_only"));
-      else if (A && A->isReadWrite())
-        accessQuals.push_back(llvm::MDString::get(Context, "read_write"));
-      else
-        accessQuals.push_back(llvm::MDString::get(Context, "read_only"));
-    } else
-      accessQuals.push_back(llvm::MDString::get(Context, "none"));
-
-    // Get argument name.
-    argNames.push_back(llvm::MDString::get(Context, parm->getName()));
-  }
-
-  Fn->setMetadata("kernel_arg_addr_space",
-                  llvm::MDNode::get(Context, addressQuals));
-  Fn->setMetadata("kernel_arg_access_qual",
-                  llvm::MDNode::get(Context, accessQuals));
-  Fn->setMetadata("kernel_arg_type",
-                  llvm::MDNode::get(Context, argTypeNames));
-  Fn->setMetadata("kernel_arg_base_type",
-                  llvm::MDNode::get(Context, argBaseTypeNames));
-  Fn->setMetadata("kernel_arg_type_qual",
-                  llvm::MDNode::get(Context, argTypeQuals));
-  if (CGM.getCodeGenOpts().EmitOpenCLArgMetadata)
-    Fn->setMetadata("kernel_arg_name",
-                    llvm::MDNode::get(Context, argNames));
-}
-
 void CodeGenFunction::EmitOpenCLKernelMetadata(const FunctionDecl *FD,
                                                llvm::Function *Fn)
 {
   if (!FD->hasAttr<OpenCLKernelAttr>())
     return;
 
   llvm::LLVMContext &Context = getLLVMContext();
 
-  GenOpenCLArgMetadata(FD, Fn, CGM, Context, Builder, getContext());
+  CGM.GenOpenCLArgMetadata(Fn, FD, this);
 
   if (const VecTypeHintAttr *A = FD->getAttr<VecTypeHintAttr>()) {
     QualType HintQTy = A->getTypeHint();
     const ExtVectorType *HintEltQTy = HintQTy->getAs<ExtVectorType>();
     bool IsSignedInteger =
         HintQTy->isSignedIntegerType() ||
         (HintEltQTy && HintEltQTy->getElementType()->isSignedIntegerType());
     llvm::Metadata *AttrMDArgs[] = {

lib/CodeGen/CodeGenModule.h

public:
   /// Get the declaration of std::terminate for the platform.
   llvm::FunctionCallee getTerminateFn();
 
   llvm::SanitizerStatReport &getSanStats();
 
   llvm::Value *
   createOpenCLIntToSamplerConversion(const Expr *E, CodeGenFunction &CGF);
 
+  /// OpenCL v1.2 s5.6.4.6 allows the compiler to store kernel argument
+  /// information in the program executable. The argument information stored
+  /// includes the argument name, its type, the address and access qualifiers
+  /// used. This helper can be used to generate metadata for source code kernel
+  /// function as well as generated implicitly kernels. If a kernel is generated
+  /// implicitly null value has to be passed to the last two parameters,
+  /// otherwise all parameters must have valid non-null values.
+  /// \param FN is a pointer to IR function being generated.
+  /// \param FD is a pointer to function declaration if any.
+  /// \param CGF is a pointer to CodeGenFunction that generates this function.
+  void GenOpenCLArgMetadata(llvm::Function *Fn, const FunctionDecl *FD=nullptr,  
+                                 CodeGenFunction *CGF=nullptr);
+
   /// Get target specific null pointer.
   /// \param T is the LLVM type of the null pointer.
   /// \param QT is the clang QualType of the null pointer.
   llvm::Constant *getNullPointer(llvm::PointerType *T, QualType QT);
 
 private:
   llvm::Constant *GetOrCreateLLVMFunction(
       StringRef MangledName, llvm::Type *Ty, GlobalDecl D, bool ForVTable,

lib/CodeGen/CodeGenModule.cpp

void CodeGenModule::SetLLVMFunctionAttributes(GlobalDecl GD,
                                               llvm::Function *F) {
   unsigned CallingConv;
   llvm::AttributeList PAL;
   ConstructAttributeList(F->getName(), Info, GD, PAL, CallingConv, false);
   F->setAttributes(PAL);
   F->setCallingConv(static_cast<llvm::CallingConv::ID>(CallingConv));
 }
 
+static void removeImageAccessQualifier(std::string& TyName) {
+  std::string ReadOnlyQual("__read_only");
+  std::string::size_type ReadOnlyPos = TyName.find(ReadOnlyQual);
+  if (ReadOnlyPos != std::string::npos)
+    // "+ 1" for the space after access qualifier.
+    TyName.erase(ReadOnlyPos, ReadOnlyQual.size() + 1);
+  else {
+    std::string WriteOnlyQual("__write_only");
+    std::string::size_type WriteOnlyPos = TyName.find(WriteOnlyQual);
+    if (WriteOnlyPos != std::string::npos)
+      TyName.erase(WriteOnlyPos, WriteOnlyQual.size() + 1);
+    else {
+      std::string ReadWriteQual("__read_write");
+      std::string::size_type ReadWritePos = TyName.find(ReadWriteQual);
+      if (ReadWritePos != std::string::npos)
+        TyName.erase(ReadWritePos, ReadWriteQual.size() + 1);
+    }
+  }
+}
+
+// Returns the address space id that should be produced to the
+// kernel_arg_addr_space metadata. This is always fixed to the ids
+// as specified in the SPIR 2.0 specification in order to differentiate
+// for example in clGetKernelArgInfo() implementation between the address
+// spaces with targets without unique mapping to the OpenCL address spaces
+// (basically all single AS CPUs).
+static unsigned ArgInfoAddressSpace(LangAS AS) {
+  switch (AS) {
+  case LangAS::opencl_global:   return 1;
+  case LangAS::opencl_constant: return 2;
+  case LangAS::opencl_local:    return 3;
+  case LangAS::opencl_generic:  return 4; // Not in SPIR 2.0 specs.
+  default:
+    return 0; // Assume private.
+  }
+}
+
+void CodeGenModule::GenOpenCLArgMetadata(llvm::Function *Fn,
+                                         const FunctionDecl *FD,
+                                         CodeGenFunction *CGF) {
+  assert(((FD && CGF) || (!FD && !CGF)) &&
+         "Incorrect use - FD and CGF should either be both null or not!");
+  // Create MDNodes that represent the kernel arg metadata.
+  // Each MDNode is a list in the form of "key", N number of values which is
+  // the same number of values as their are kernel arguments.
+
+  const PrintingPolicy &Policy = Context.getPrintingPolicy();
+
+  // MDNode for the kernel argument address space qualifiers.
+  SmallVector<llvm::Metadata *, 8> addressQuals;
+
+  // MDNode for the kernel argument access qualifiers (images only).
+  SmallVector<llvm::Metadata *, 8> accessQuals;
+
+  // MDNode for the kernel argument type names.
+  SmallVector<llvm::Metadata *, 8> argTypeNames;
+
+  // MDNode for the kernel argument base type names.
+  SmallVector<llvm::Metadata *, 8> argBaseTypeNames;
+
+  // MDNode for the kernel argument type qualifiers.
+  SmallVector<llvm::Metadata *, 8> argTypeQuals;
+
+  // MDNode for the kernel argument names.
+  SmallVector<llvm::Metadata *, 8> argNames;
+
+  if (FD && CGF)
+    for (unsigned i = 0, e = FD->getNumParams(); i != e; ++i) {
+      const ParmVarDecl *parm = FD->getParamDecl(i);
+      QualType ty = parm->getType();
+      std::string typeQuals;
+
+      if (ty->isPointerType()) {
+        QualType pointeeTy = ty->getPointeeType();
+
+        // Get address qualifier.
+        addressQuals.push_back(
+            llvm::ConstantAsMetadata::get(CGF->Builder.getInt32(
+                ArgInfoAddressSpace(pointeeTy.getAddressSpace()))));
+
+        // Get argument type name.
+        std::string typeName =
+            pointeeTy.getUnqualifiedType().getAsString(Policy) + "*";
+
+        // Turn "unsigned type" to "utype"
+        std::string::size_type pos = typeName.find("unsigned");
+        if (pointeeTy.isCanonical() && pos != std::string::npos)
+          typeName.erase(pos + 1, 8);
+
+        argTypeNames.push_back(llvm::MDString::get(VMContext, typeName));
+
+        std::string baseTypeName =
+            pointeeTy.getUnqualifiedType().getCanonicalType().getAsString(
+                Policy) +
+            "*";
+
+        // Turn "unsigned type" to "utype"
+        pos = baseTypeName.find("unsigned");
+        if (pos != std::string::npos)
+          baseTypeName.erase(pos + 1, 8);
+
+        argBaseTypeNames.push_back(
+            llvm::MDString::get(VMContext, baseTypeName));
+
+        // Get argument type qualifiers:
+        if (ty.isRestrictQualified())
+          typeQuals = "restrict";
+        if (pointeeTy.isConstQualified() ||
+            (pointeeTy.getAddressSpace() == LangAS::opencl_constant))
+          typeQuals += typeQuals.empty() ? "const" : " const";
+        if (pointeeTy.isVolatileQualified())
+          typeQuals += typeQuals.empty() ? "volatile" : " volatile";
+      } else {
+        uint32_t AddrSpc = 0;
+        bool isPipe = ty->isPipeType();
+        if (ty->isImageType() || isPipe)
+          AddrSpc = ArgInfoAddressSpace(LangAS::opencl_global);
+
+        addressQuals.push_back(
+            llvm::ConstantAsMetadata::get(CGF->Builder.getInt32(AddrSpc)));
+
+        // Get argument type name.
+        std::string typeName;
+        if (isPipe)
+          typeName = ty.getCanonicalType()
+                         ->getAs<PipeType>()
+                         ->getElementType()
+                         .getAsString(Policy);
+        else
+          typeName = ty.getUnqualifiedType().getAsString(Policy);
+
+        // Turn "unsigned type" to "utype"
+        std::string::size_type pos = typeName.find("unsigned");
+        if (ty.isCanonical() && pos != std::string::npos)
+          typeName.erase(pos + 1, 8);
+
+        std::string baseTypeName;
+        if (isPipe)
+          baseTypeName = ty.getCanonicalType()
+                             ->getAs<PipeType>()
+                             ->getElementType()
+                             .getCanonicalType()
+                             .getAsString(Policy);
+        else
+          baseTypeName =
+              ty.getUnqualifiedType().getCanonicalType().getAsString(Policy);
+
+        // Remove access qualifiers on images
+        // (as they are inseparable from type in clang implementation,
+        // but OpenCL spec provides a special query to get access qualifier
+        // via clGetKernelArgInfo with CL_KERNEL_ARG_ACCESS_QUALIFIER):
+        if (ty->isImageType()) {
+          removeImageAccessQualifier(typeName);
+          removeImageAccessQualifier(baseTypeName);
+        }
+
+        argTypeNames.push_back(llvm::MDString::get(VMContext, typeName));
+
+        // Turn "unsigned type" to "utype"
+        pos = baseTypeName.find("unsigned");
+        if (pos != std::string::npos)
+          baseTypeName.erase(pos + 1, 8);
+
+        argBaseTypeNames.push_back(
+            llvm::MDString::get(VMContext, baseTypeName));
+
+        if (isPipe)
+          typeQuals = "pipe";
+      }
+
+      argTypeQuals.push_back(llvm::MDString::get(VMContext, typeQuals));
+
+      // Get image and pipe access qualifier:
+      if (ty->isImageType() || ty->isPipeType()) {
+        const Decl *PDecl = parm;
+        if (auto *TD = dyn_cast<TypedefType>(ty))
+          PDecl = TD->getDecl();
+        const OpenCLAccessAttr *A = PDecl->getAttr<OpenCLAccessAttr>();
+        if (A && A->isWriteOnly())
+          accessQuals.push_back(llvm::MDString::get(VMContext, "write_only"));
+        else if (A && A->isReadWrite())
+          accessQuals.push_back(llvm::MDString::get(VMContext, "read_write"));
+        else
+          accessQuals.push_back(llvm::MDString::get(VMContext, "read_only"));
+      } else
+        accessQuals.push_back(llvm::MDString::get(VMContext, "none"));
+
+      // Get argument name.
+      argNames.push_back(llvm::MDString::get(VMContext, parm->getName()));
+    }
+
+  Fn->setMetadata("kernel_arg_addr_space",
+                  llvm::MDNode::get(VMContext, addressQuals));
+  Fn->setMetadata("kernel_arg_access_qual",
+                  llvm::MDNode::get(VMContext, accessQuals));
+  Fn->setMetadata("kernel_arg_type",
+                  llvm::MDNode::get(VMContext, argTypeNames));
+  Fn->setMetadata("kernel_arg_base_type",
+                  llvm::MDNode::get(VMContext, argBaseTypeNames));
+  Fn->setMetadata("kernel_arg_type_qual",
+                  llvm::MDNode::get(VMContext, argTypeQuals));
+  if (getCodeGenOpts().EmitOpenCLArgMetadata)
+    Fn->setMetadata("kernel_arg_name",
+                    llvm::MDNode::get(VMContext, argNames));
+}
+
 /// Determines whether the language options require us to model
 /// unwind exceptions.  We treat -fexceptions as mandating this
 /// except under the fragile ObjC ABI with only ObjC exceptions
 /// enabled.  This means, for example, that C with -fexceptions
 /// enables this.
 static bool hasUnwindExceptions(const LangOptions &LangOpts) {
   // If exceptions are completely disabled, obviously this is false.
   if (!LangOpts.Exceptions) return false;

test/CodeGenOpenCLCXX/global_init.cl

+// RUN: %clang_cc1 %s -triple spir -cl-std=c++ -emit-llvm -O0 -o - | FileCheck %s
+
+struct S {
+  S() {}
+};
+
+S s;
+
+//CHECK: define internal spir_kernel void @_GLOBAL__sub_I_{{.*}}!kernel_arg_addr_space [[ARGMD:![0-9]+]] !kernel_arg_access_qual [[ARGMD]] !kernel_arg_type [[ARGMD]] !kernel_arg_base_type [[ARGMD]] !kernel_arg_type_qual [[ARGMD]]
+// Check that parameters are empty.
+//CHECK: [[ARGMD]] = !{}