4 files changed, 19 insertions, 1 deletions
diff --git a/clang/lib/CodeGen/CGDeclCXX.cpp b/clang/lib/CodeGen/CGDeclCXX.cpp
index bf16b7bec4b..5b172a3480b 100644
--- a/clang/lib/CodeGen/CGDeclCXX.cpp
+++ b/clang/lib/CodeGen/CGDeclCXX.cpp
@@ -437,7 +437,7 @@ CodeGenModule::EmitCXXGlobalVarDeclInitFunc(const VarDecl *D,
   // that are of class type, cannot have a non-empty constructor. All
   // the checks have been done in Sema by now. Whatever initializers
   // are allowed are empty and we just need to ignore them here.
-  if (getLangOpts().CUDA && getLangOpts().CUDAIsDevice &&
+  if (getLangOpts().CUDAIsDevice && !getLangOpts().GPUAllowDeviceInit &&
       (D->hasAttr<CUDADeviceAttr>() || D->hasAttr<CUDAConstantAttr>() ||
        D->hasAttr<CUDASharedAttr>()))
     return;
@@ -608,6 +608,11 @@ CodeGenModule::EmitCXXGlobalInitFunc() {
     Fn->setCallingConv(llvm::CallingConv::SPIR_KERNEL);
   }
 
+  if (getLangOpts().HIP) {
+    Fn->setCallingConv(llvm::CallingConv::AMDGPU_KERNEL);
+    Fn->addFnAttr("device-init");
+  }
+
   CXXGlobalInits.clear();
 }
 
diff --git a/clang/lib/Driver/ToolChains/HIP.cpp b/clang/lib/Driver/ToolChains/HIP.cpp
index ad9384df6a2..d84a454359a 100644
--- a/clang/lib/Driver/ToolChains/HIP.cpp
+++ b/clang/lib/Driver/ToolChains/HIP.cpp
@@ -292,6 +292,10 @@ void HIPToolChain::addClangTargetOptions(
                          false))
     CC1Args.push_back("-fgpu-rdc");
 
+  if (DriverArgs.hasFlag(options::OPT_fgpu_allow_device_init,
+                         options::OPT_fno_gpu_allow_device_init, false))
+    CC1Args.push_back("-fgpu-allow-device-init");
+
   // Default to "hidden" visibility, as object level linking will not be
   // supported for the foreseeable future.
   if (!DriverArgs.hasArg(options::OPT_fvisibility_EQ,
diff --git a/clang/lib/Frontend/CompilerInvocation.cpp b/clang/lib/Frontend/CompilerInvocation.cpp
index 665695ec3b1..767a0718b24 100644
--- a/clang/lib/Frontend/CompilerInvocation.cpp
+++ b/clang/lib/Frontend/CompilerInvocation.cpp
@@ -2528,6 +2528,13 @@ static void ParseLangArgs(LangOptions &Opts, ArgList &Args, InputKind IK,
     Opts.CUDADeviceApproxTranscendentals = 1;
 
   Opts.GPURelocatableDeviceCode = Args.hasArg(OPT_fgpu_rdc);
+  if (Args.hasArg(OPT_fgpu_allow_device_init)) {
+    if (Opts.HIP)
+      Opts.GPUAllowDeviceInit = 1;
+    else
+      Diags.Report(diag::warn_ignore_hip_only_option)
+          << Args.getLastArg(OPT_fgpu_allow_device_init)->getAsString(Args);
+  }
   Opts.HIPUseNewLaunchAPI = Args.hasArg(OPT_fhip_new_launch_api);
 
   if (Opts.ObjC) {
diff --git a/clang/lib/Sema/SemaCUDA.cpp b/clang/lib/Sema/SemaCUDA.cpp
index d0ddfd040c9..0c61057e107 100644
--- a/clang/lib/Sema/SemaCUDA.cpp
+++ b/clang/lib/Sema/SemaCUDA.cpp
@@ -492,6 +492,8 @@ void Sema::checkAllowedCUDAInitializer(VarDecl *VD) {
   const Expr *Init = VD->getInit();
   if (VD->hasAttr<CUDADeviceAttr>() || VD->hasAttr<CUDAConstantAttr>() ||
       VD->hasAttr<CUDASharedAttr>()) {
+    if (LangOpts.GPUAllowDeviceInit)
+      return;
     assert(!VD->isStaticLocal() || VD->hasAttr<CUDASharedAttr>());
     bool AllowedInit = false;
     if (const CXXConstructExpr *CE = dyn_cast<CXXConstructExpr>(Init))