Implement work_group_static / work_group_scratch_memory (#15061)

The patch partially implements `work_group_static` and update proposal.

Implemented:

- `work_group_static` to handle static allocation in kernel.
- `get_dynamic_work_group_memory` to handle runtime allocation, but only
on CUDA

`work_group_static` is implemented by exposing `SYCLScope(WorkGroup)`,
allowing the class to be decorated by the attribute and uses the same
mechanism during lowering to place the variable in local memory.

`get_dynamic_work_group_memory` uses a new builtin function,
`__sycl_dynamicLocalMemoryPlaceholder `, which is lowered into
referencing a 0 sized array GV when targeting NVPTX. The approach for
SPIR will need to differ from this lowering.

UR change oneapi-src/unified-runtime#1968,
oneapi-src/unified-runtime#2403

---------

Signed-off-by: Lukas Sommer <[email protected]>
Signed-off-by: Victor Lomuller <[email protected]>
Co-authored-by: Lukas Sommer <[email protected]>
Co-authored-by: Atharva Dubey <[email protected]>
Co-authored-by: Marcos Maronas <[email protected]>
Co-authored-by: Callum Fare <[email protected]>
Co-authored-by: Martin Morrison-Grant <[email protected]>

Author: 6 people

clang/include/clang/Basic/Attr.td

@@ -1682,12 +1682,12 @@ def SYCLIntelESimdVectorize : InheritableAttr {
 }
 
 def SYCLScope : Attr {
-  // No spelling, as this attribute can't be created in the source code.
-  let Spellings = [];
+  let Spellings = [CXX11<"__sycl_detail__", "wg_scope">];
   let Args = [EnumArgument<"level", "Level", /*is_string=*/false,
                                              ["work_group", "work_item"],
-                                             ["WorkGroup", "WorkItem"]>];
-  let Subjects = SubjectList<[Function, Var]>;
+                                             ["WorkGroup", "WorkItem"],
+                                             /*optional=*/true>];
+  let Subjects = SubjectList<[Function, Var, CXXRecord]>;
   let LangOpts = [SYCLIsDevice];
 
   let AdditionalMembers = [{
@@ -1700,7 +1700,7 @@ def SYCLScope : Attr {
     }
   }];
 
-  let Documentation = [InternalOnly];
+  let Documentation = [SYCLWGScopeDocs];
 }
 
 def SYCLDeviceIndirectlyCallable : InheritableAttr {

clang/include/clang/Basic/AttrDocs.td

@@ -4271,6 +4271,18 @@ function pointer for the specified function.
   }];
 }
 
+def SYCLWGScopeDocs : Documentation {
+  let Category = DocCatFunction;
+  let Heading = "__sycl_detail__::wg_scope";
+  let Content = [{
+This attribute can only be applied to records with a trivial default constructor and destructor.
+Types with this attribute cannot be used for non-static data members.
+It indicates that any block and namespace scope variable of a type holding this attribute
+will be allocated in local memory. For variables allocated in block scope, they behave
+as implicitly declared as static.
+  }];
+}
+
 def SYCLDeviceDocs : Documentation {
   let Category = DocCatFunction;
   let Heading = "sycl_device";

clang/include/clang/Basic/DiagnosticSemaKinds.td

@@ -12556,6 +12556,11 @@ def err_sycl_external_global : Error<
 def warn_sycl_kernel_too_big_args : Warning<
   "size of kernel arguments (%0 bytes) may exceed the supported maximum "
   "of %1 bytes on some devices">, InGroup<SyclStrict>, ShowInSystemHeader;
+def err_sycl_wg_scope : Error<
+  "SYCL work group scope only applies to class with a trivial "
+  "%select{default constructor|destructor}0">;
+def err_sycl_field_with_wg_scope : Error<
+  "non-static data member is of a type with a SYCL work group scope attribute applied to it">;
 def err_sycl_virtual_types : Error<
   "no class with a vtable can be used in a SYCL kernel or any code included in the kernel">;
 def note_sycl_recursive_function_declared_here: Note<"function implemented using recursion declared here">;

clang/include/clang/Sema/SemaSYCL.h

@@ -267,6 +267,7 @@ class SemaSYCL : public SemaBase {
 
   void CheckSYCLKernelCall(FunctionDecl *CallerFunc,
                            ArrayRef<const Expr *> Args);
+  void CheckSYCLScopeAttr(CXXRecordDecl *Decl);
 
   /// Creates a SemaDiagnosticBuilder that emits the diagnostic if the current
   /// context is "used as device code".
@@ -478,6 +479,7 @@ class SemaSYCL : public SemaBase {
                                                   const ParsedAttr &AL);
   void handleSYCLIntelMaxWorkGroupsPerMultiprocessor(Decl *D,
                                                      const ParsedAttr &AL);
+  void handleSYCLScopeAttr(Decl *D, const ParsedAttr &AL);
 
   void checkSYCLAddIRAttributesFunctionAttrConflicts(Decl *D);
 

clang/lib/CodeGen/CGDecl.cpp

@@ -217,7 +217,11 @@ void CodeGenFunction::EmitVarDecl(const VarDecl &D) {
   if (D.getType().getAddressSpace() == LangAS::opencl_local)
     return CGM.getOpenCLRuntime().EmitWorkGroupLocalVarDecl(*this, D);
 
-  if (D.getAttr<SYCLScopeAttr>() && D.getAttr<SYCLScopeAttr>()->isWorkGroup())
+  SYCLScopeAttr *ScopeAttr = D.getAttr<SYCLScopeAttr>();
+  if (!ScopeAttr)
+    if (auto *RD = D.getType()->getAsCXXRecordDecl())
+      ScopeAttr = RD->getAttr<SYCLScopeAttr>();
+  if (ScopeAttr && ScopeAttr->isWorkGroup())
     return CGM.getSYCLRuntime().emitWorkGroupLocalVarDecl(*this, D);
 
   assert(D.hasLocalStorage());

clang/lib/CodeGen/CGSYCLRuntime.cpp

@@ -96,7 +96,10 @@ void CGSYCLRuntime::emitWorkGroupLocalVarDecl(CodeGenFunction &CGF,
                                               const VarDecl &D) {
 #ifndef NDEBUG
   SYCLScopeAttr *Scope = D.getAttr<SYCLScopeAttr>();
-  assert(Scope && Scope->isWorkGroup() && "work group scope expected");
+  if (!Scope)
+    if (auto *RD = D.getType()->getAsCXXRecordDecl())
+      Scope = RD->getAttr<SYCLScopeAttr>();
+  assert((Scope && Scope->isWorkGroup()) && "work group scope expected");
 #endif // NDEBUG
   // generate global variable in the address space selected by the clang CodeGen
   // (should be local)

clang/lib/CodeGen/CodeGenModule.cpp

@@ -5789,6 +5789,9 @@ LangAS CodeGenModule::GetGlobalVarAddressSpace(const VarDecl *D) {
 
   if (LangOpts.SYCLIsDevice && D) {
     auto *Scope = D->getAttr<SYCLScopeAttr>();
+    if (!Scope)
+      if (auto *RD = D->getType()->getAsCXXRecordDecl())
+        Scope = RD->getAttr<SYCLScopeAttr>();
     if (Scope && Scope->isWorkGroup())
       return LangAS::sycl_local;
   }

clang/lib/Sema/Sema.cpp

@@ -1849,7 +1849,9 @@ class DeferredDiagnosticsEmitter
         if (!S.SYCL().checkAllowedSYCLInitializer(VD) &&
             !S.SYCL()
                  .isTypeDecoratedWithDeclAttribute<
-                     SYCLGlobalVariableAllowedAttr>(VD->getType())) {
+                     SYCLGlobalVariableAllowedAttr>(VD->getType()) &&
+            !S.SYCL().isTypeDecoratedWithDeclAttribute<SYCLScopeAttr>(
+                VD->getType())) {
           S.Diag(Loc, diag::err_sycl_restrict)
               << SemaSYCL::KernelConstStaticVariable;
           return;

clang/lib/Sema/SemaDecl.cpp

@@ -7876,6 +7876,8 @@ NamedDecl *Sema::ActOnVariableDeclarator(
     // attribute.
     if (SCSpec == DeclSpec::SCS_static && !R.isConstant(Context) &&
         !SYCL().isTypeDecoratedWithDeclAttribute<SYCLGlobalVariableAllowedAttr>(
+            NewVD->getType()) &&
+        !SYCL().isTypeDecoratedWithDeclAttribute<SYCLScopeAttr>(
             NewVD->getType()))
       SYCL().DiagIfDeviceCode(D.getIdentifierLoc(), diag::err_sycl_restrict)
           << SemaSYCL::KernelNonConstStaticDataVariable;
@@ -18578,6 +18580,14 @@ FieldDecl *Sema::CheckFieldDecl(DeclarationName Name, QualType T,
     InvalidDecl = true;
   }
 
+  if (LangOpts.SYCLIsDevice) {
+    const CXXRecordDecl *RD = T->getAsCXXRecordDecl();
+    if (RD && RD->hasAttr<SYCLScopeAttr>()) {
+      Diag(Loc, diag::err_sycl_field_with_wg_scope);
+      InvalidDecl = true;
+    }
+  }
+
   if (LangOpts.OpenCL) {
     // OpenCL v1.2 s6.9b,r & OpenCL v2.0 s6.12.5 - The following types cannot be
     // used as structure or union field: image, sampler, event or block types.

clang/lib/Sema/SemaDeclAttr.cpp

@@ -6717,6 +6717,9 @@ ProcessDeclAttribute(Sema &S, Scope *scope, Decl *D, const ParsedAttr &AL,
   case ParsedAttr::AT_SYCLDevice:
     S.SYCL().handleSYCLDeviceAttr(D, AL);
     break;
+  case ParsedAttr::AT_SYCLScope:
+    S.SYCL().handleSYCLScopeAttr(D, AL);
+    break;
   case ParsedAttr::AT_SYCLDeviceIndirectlyCallable:
     S.SYCL().handleSYCLDeviceIndirectlyCallableAttr(D, AL);
     break;