[HIP][NFC] Refactor managed var codegen

[yxsamliu]

Refactor managed variable handling in codegen so that the transformation is done separately from registration.

This will allow the new driver to register the managed var in the linker wrapper.

[llvmbot]

@llvm/pr-subscribers-clang-codegen

@llvm/pr-subscribers-clang

Author: Yaxun (Sam) Liu (yxsamliu)

Changes

Refactor managed variable handling in codegen so that the transformation is done separately from registration.

This will allow the new driver to register the managed var in the linker wrapper.

---

Full diff: https://github.com/llvm/llvm-project/pull/85976.diff

1 Files Affected:

• (modified) clang/lib/CodeGen/CGCUDANV.cpp (+9-18)

diff --git a/clang/lib/CodeGen/CGCUDANV.cpp b/clang/lib/CodeGen/CGCUDANV.cpp
index d3f2573fd5e38a..b756318c46a900 100644
--- a/clang/lib/CodeGen/CGCUDANV.cpp
+++ b/clang/lib/CodeGen/CGCUDANV.cpp
@@ -605,20 +605,10 @@ llvm::Function *CGNVCUDARuntime::makeRegisterGlobalsFn() {
       uint64_t VarSize =
           CGM.getDataLayout().getTypeAllocSize(Var->getValueType());
       if (Info.Flags.isManaged()) {
-        auto *ManagedVar = new llvm::GlobalVariable(
-            CGM.getModule(), Var->getType(),
-            /*isConstant=*/false, Var->getLinkage(),
-            /*Init=*/Var->isDeclaration()
-                ? nullptr
-                : llvm::ConstantPointerNull::get(Var->getType()),
-            /*Name=*/"", /*InsertBefore=*/nullptr,
-            llvm::GlobalVariable::NotThreadLocal);
-        ManagedVar->setDSOLocal(Var->isDSOLocal());
-        ManagedVar->setVisibility(Var->getVisibility());
-        ManagedVar->setExternallyInitialized(true);
-        ManagedVar->takeName(Var);
-        Var->setName(Twine(ManagedVar->getName() + ".managed"));
-        replaceManagedVar(Var, ManagedVar);
+        assert(Var->getName().ends_with(".managed") &&
+               "HIP managed variables not transformed");
+        auto *ManagedVar = CGM.getModule().getNamedGlobal(
+            Var->getName().drop_back(StringRef(".managed").size()));
         llvm::Value *Args[] = {
             &GpuBinaryHandlePtr,
             ManagedVar,
@@ -1093,7 +1083,9 @@ void CGNVCUDARuntime::transformManagedVars() {
               : llvm::ConstantPointerNull::get(Var->getType()),
           /*Name=*/"", /*InsertBefore=*/nullptr,
           llvm::GlobalVariable::NotThreadLocal,
-          CGM.getContext().getTargetAddressSpace(LangAS::cuda_device));
+          CGM.getContext().getTargetAddressSpace(CGM.getLangOpts().CUDAIsDevice
+                                                     ? LangAS::cuda_device
+                                                     : LangAS::Default));
       ManagedVar->setDSOLocal(Var->isDSOLocal());
       ManagedVar->setVisibility(Var->getVisibility());
       ManagedVar->setExternallyInitialized(true);
@@ -1102,7 +1094,7 @@ void CGNVCUDARuntime::transformManagedVars() {
       Var->setName(Twine(ManagedVar->getName()) + ".managed");
       // Keep managed variables even if they are not used in device code since
       // they need to be allocated by the runtime.
-      if (!Var->isDeclaration()) {
+      if (CGM.getLangOpts().CUDAIsDevice && !Var->isDeclaration()) {
         assert(!ManagedVar->isDeclaration());
         CGM.addCompilerUsedGlobal(Var);
         CGM.addCompilerUsedGlobal(ManagedVar);
@@ -1160,9 +1152,8 @@ void CGNVCUDARuntime::createOffloadingEntries() {
 
 // Returns module constructor to be added.
 llvm::Function *CGNVCUDARuntime::finalizeModule() {
+  transformManagedVars();
   if (CGM.getLangOpts().CUDAIsDevice) {
-    transformManagedVars();
-
     // Mark ODR-used device variables as compiler used to prevent it from being
     // eliminated by optimization. This is necessary for device variables
     // ODR-used by host functions. Sema correctly marks them as ODR-used no

[Artem-B]

LGTM, sans the "NFC" part in the description.

[Artem-B]

This does not look like "NFC" as we now perform the transform for the host compilation, too.

I assume we do have existing tests covering generation of the variables.

[jhuber6]

I'm guessing we also don't have a test for __managed__ in the external test suite?

[yxsamliu]

we did the equivalent transformation previously during registration of managed var by calling replaceManagedVar directly, so it is actually NFC. We have lit test https://github.com/llvm/llvm-project/blob/main/clang/test/CodeGenCUDA/managed-var.cu that covers transformation in both device and host code.

We also have execution tests in hip-tests https://github.com/ROCm/hip-tests/blob/d0088ee34f1bd507e793c6fc5a148488d577c11a/catch/TypeQualifiers/hipManagedKeyword.cc#L28

[yxsamliu]

we do not have test for managed var in llvm-test-suite