[clang] 90abdf8 - [CUDA][HIP][NFC] add CodeGenModule::shouldEmitCUDAGlobalVar (#98543)

[via cfe-commits]

Author: Yaxun (Sam) Liu
Date: 2024-07-11T21:52:04-04:00
New Revision: 90abdf83e273586a43e1270e5f0a11de5cc35383

URL: https://github.com/llvm/llvm-project/commit/90abdf83e273586a43e1270e5f0a11de5cc35383
DIFF: https://github.com/llvm/llvm-project/commit/90abdf83e273586a43e1270e5f0a11de5cc35383.diff

LOG: [CUDA][HIP][NFC] add CodeGenModule::shouldEmitCUDAGlobalVar (#98543)

Extract the logic whether to emit a global var based on CUDA/HIP
host/device related attributes to CodeGenModule::shouldEmitCUDAGlobalVar
to be used by other places.

Added: 
    

Modified: 
    clang/lib/CodeGen/CodeGenModule.cpp
    clang/lib/CodeGen/CodeGenModule.h

Removed: 
    


################################################################################
diff  --git a/clang/lib/CodeGen/CodeGenModule.cpp b/clang/lib/CodeGen/CodeGenModule.cpp
index 5c810cd332185..6c10b4a2edef8 100644
--- a/clang/lib/CodeGen/CodeGenModule.cpp
+++ b/clang/lib/CodeGen/CodeGenModule.cpp
@@ -3702,6 +3702,19 @@ template <typename AttrT> static bool hasImplicitAttr(const ValueDecl *D) {
   return D->isImplicit();
 }
 
+bool CodeGenModule::shouldEmitCUDAGlobalVar(const VarDecl *Global) const {
+  assert(LangOpts.CUDA && "Should not be called by non-CUDA languages");
+  // We need to emit host-side 'shadows' for all global
+  // device-side variables because the CUDA runtime needs their
+  // size and host-side address in order to provide access to
+  // their device-side incarnations.
+  return !LangOpts.CUDAIsDevice || Global->hasAttr<CUDADeviceAttr>() ||
+         Global->hasAttr<CUDAConstantAttr>() ||
+         Global->hasAttr<CUDASharedAttr>() ||
+         Global->getType()->isCUDADeviceBuiltinSurfaceType() ||
+         Global->getType()->isCUDADeviceBuiltinTextureType();
+}
+
 void CodeGenModule::EmitGlobal(GlobalDecl GD) {
   const auto *Global = cast<ValueDecl>(GD.getDecl());
 
@@ -3726,36 +3739,27 @@ void CodeGenModule::EmitGlobal(GlobalDecl GD) {
   // Non-constexpr non-lambda implicit host device functions are not emitted
   // unless they are used on device side.
   if (LangOpts.CUDA) {
-    if (LangOpts.CUDAIsDevice) {
+    assert((isa<FunctionDecl>(Global) || isa<VarDecl>(Global)) &&
+           "Expected Variable or Function");
+    if (const auto *VD = dyn_cast<VarDecl>(Global)) {
+      if (!shouldEmitCUDAGlobalVar(VD))
+        return;
+    } else if (LangOpts.CUDAIsDevice) {
       const auto *FD = dyn_cast<FunctionDecl>(Global);
       if ((!Global->hasAttr<CUDADeviceAttr>() ||
-           (LangOpts.OffloadImplicitHostDeviceTemplates && FD &&
+           (LangOpts.OffloadImplicitHostDeviceTemplates &&
             hasImplicitAttr<CUDAHostAttr>(FD) &&
             hasImplicitAttr<CUDADeviceAttr>(FD) && !FD->isConstexpr() &&
             !isLambdaCallOperator(FD) &&
             !getContext().CUDAImplicitHostDeviceFunUsedByDevice.count(FD))) &&
           !Global->hasAttr<CUDAGlobalAttr>() &&
-          !Global->hasAttr<CUDAConstantAttr>() &&
-          !Global->hasAttr<CUDASharedAttr>() &&
-          !Global->getType()->isCUDADeviceBuiltinSurfaceType() &&
-          !Global->getType()->isCUDADeviceBuiltinTextureType() &&
           !(LangOpts.HIPStdPar && isa<FunctionDecl>(Global) &&
             !Global->hasAttr<CUDAHostAttr>()))
         return;
-    } else {
-      // We need to emit host-side 'shadows' for all global
-      // device-side variables because the CUDA runtime needs their
-      // size and host-side address in order to provide access to
-      // their device-side incarnations.
-
-      // So device-only functions are the only things we skip.
-      if (isa<FunctionDecl>(Global) && !Global->hasAttr<CUDAHostAttr>() &&
-          Global->hasAttr<CUDADeviceAttr>())
-        return;
-
-      assert((isa<FunctionDecl>(Global) || isa<VarDecl>(Global)) &&
-             "Expected Variable or Function");
-    }
+      // Device-only functions are the only things we skip.
+    } else if (!Global->hasAttr<CUDAHostAttr>() &&
+               Global->hasAttr<CUDADeviceAttr>())
+      return;
   }
 
   if (LangOpts.OpenMP) {

diff  --git a/clang/lib/CodeGen/CodeGenModule.h b/clang/lib/CodeGen/CodeGenModule.h
index 585c4ea697fea..caa3786c033b5 100644
--- a/clang/lib/CodeGen/CodeGenModule.h
+++ b/clang/lib/CodeGen/CodeGenModule.h
@@ -563,6 +563,9 @@ class CodeGenModule : public CodeGenTypeCache {
 
   bool isTriviallyRecursive(const FunctionDecl *F);
   bool shouldEmitFunction(GlobalDecl GD);
+  // Whether a global variable should be emitted by CUDA/HIP host/device
+  // related attributes.
+  bool shouldEmitCUDAGlobalVar(const VarDecl *VD) const;
   bool shouldOpportunisticallyEmitVTables();
   /// Map used to be sure we don't emit the same CompoundLiteral twice.
   llvm::DenseMap<const CompoundLiteralExpr *, llvm::GlobalVariable *>