[SYCL] Warn when number of kernel args exceeds maximum available on GPU (#2361)

Emit a warning when number of resulting kernel arguments exceeds 2k -
maximum available number of kernel arguments on GPU device. Emit a
warning only in GPU AOT mode since other devices don't have such
limitation.

Author: Fznamznon

clang/include/clang/Basic/DiagnosticSemaKinds.td

@@ -10997,6 +10997,11 @@ def err_sycl_restrict : Error<
   "|use a const static or global variable that is neither zero-initialized "
   "nor constant-initialized"
   "}0">;
+def warn_sycl_kernel_too_many_args : Warning<
+  "kernel argument count (%0) exceeds supported maximum of %1 on GPU">,
+  InGroup<SyclStrict>;
+def note_sycl_kernel_args_count : Note<"array elements and fields of a "
+  "class/struct may be counted separately">;
 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/lib/Sema/SemaSYCL.cpp

@@ -56,6 +56,7 @@ enum KernelInvocationKind {
 
 const static std::string InitMethodName = "__init";
 const static std::string FinalizeMethodName = "__finalize";
+constexpr unsigned GPUMaxKernelArgsNum = 2000;
 
 namespace {
 
@@ -1657,6 +1658,83 @@ class SyclKernelDeclCreator : public SyclKernelFieldHandler {
   using SyclKernelFieldHandler::leaveStruct;
 };
 
+class SyclKernelNumArgsChecker : public SyclKernelFieldHandler {
+  SourceLocation KernelLoc;
+  unsigned NumOfParams = 0;
+
+  bool handleSpecialType(QualType FieldTy) {
+    const CXXRecordDecl *RecordDecl = FieldTy->getAsCXXRecordDecl();
+    assert(RecordDecl && "The accessor/sampler must be a RecordDecl");
+    CXXMethodDecl *InitMethod = getMethodByName(RecordDecl, InitMethodName);
+    assert(InitMethod && "The accessor/sampler must have the __init method");
+    NumOfParams += InitMethod->getNumParams();
+    return true;
+  }
+
+public:
+  SyclKernelNumArgsChecker(Sema &S, SourceLocation Loc)
+      : SyclKernelFieldHandler(S), KernelLoc(Loc) {}
+
+  ~SyclKernelNumArgsChecker() {
+    if (SemaRef.Context.getTargetInfo().getTriple().getSubArch() ==
+        llvm::Triple::SPIRSubArch_gen) {
+      if (NumOfParams > GPUMaxKernelArgsNum) {
+        SemaRef.Diag(KernelLoc, diag::warn_sycl_kernel_too_many_args)
+            << NumOfParams << GPUMaxKernelArgsNum;
+        SemaRef.Diag(KernelLoc, diag::note_sycl_kernel_args_count);
+      }
+    }
+  }
+
+  bool handleSyclAccessorType(FieldDecl *FD, QualType FieldTy) final {
+    return handleSpecialType(FieldTy);
+  }
+
+  bool handleSyclAccessorType(const CXXRecordDecl *, const CXXBaseSpecifier &,
+                              QualType FieldTy) final {
+    return handleSpecialType(FieldTy);
+  }
+
+  bool handleSyclSamplerType(FieldDecl *FD, QualType FieldTy) final {
+    return handleSpecialType(FieldTy);
+  }
+
+  bool handleSyclSamplerType(const CXXRecordDecl *, const CXXBaseSpecifier &BS,
+                             QualType FieldTy) final {
+    return handleSpecialType(FieldTy);
+  }
+
+  bool handlePointerType(FieldDecl *FD, QualType FieldTy) final {
+    NumOfParams++;
+    return true;
+  }
+
+  bool handleScalarType(FieldDecl *FD, QualType FieldTy) final {
+    NumOfParams++;
+    return true;
+  }
+
+  bool handleUnionType(FieldDecl *FD, QualType FieldTy) final {
+    return handleScalarType(FD, FieldTy);
+  }
+
+  bool handleSyclHalfType(FieldDecl *FD, QualType FieldTy) final {
+    NumOfParams++;
+    return true;
+  }
+
+  bool handleSyclStreamType(FieldDecl *FD, QualType FieldTy) final {
+    NumOfParams++;
+    return true;
+  }
+  bool handleSyclStreamType(const CXXRecordDecl *, const CXXBaseSpecifier &,
+                            QualType FieldTy) final {
+    NumOfParams++;
+    return true;
+  }
+  using SyclKernelFieldHandler::handleSyclHalfType;
+};
+
 class SyclKernelBodyCreator : public SyclKernelFieldHandler {
   SyclKernelDeclCreator &DeclCreator;
   llvm::SmallVector<Stmt *, 16> BodyStmts;
@@ -2351,6 +2429,7 @@ void Sema::CheckSYCLKernelCall(FunctionDecl *KernelFunc, SourceRange CallLoc,
 
   SyclKernelFieldChecker FieldChecker(*this);
   SyclKernelUnionChecker UnionChecker(*this);
+  SyclKernelNumArgsChecker NumArgsChecker(*this, Args[0]->getExprLoc());
   // check that calling kernel conforms to spec
   QualType KernelParamTy = KernelFunc->getParamDecl(0)->getType();
   if (KernelParamTy->isReferenceType()) {
@@ -2365,8 +2444,10 @@ void Sema::CheckSYCLKernelCall(FunctionDecl *KernelFunc, SourceRange CallLoc,
 
   KernelObjVisitor Visitor{*this};
   DiagnosingSYCLKernel = true;
-  Visitor.VisitRecordBases(KernelObj, FieldChecker, UnionChecker);
-  Visitor.VisitRecordFields(KernelObj, FieldChecker, UnionChecker);
+  Visitor.VisitRecordBases(KernelObj, FieldChecker, UnionChecker,
+                           NumArgsChecker);
+  Visitor.VisitRecordFields(KernelObj, FieldChecker, UnionChecker,
+                            NumArgsChecker);
   DiagnosingSYCLKernel = false;
   if (!FieldChecker.isValid() || !UnionChecker.isValid())
     KernelFunc->setInvalidDecl();

clang/test/SemaSYCL/num-args-overflow.cpp

@@ -0,0 +1,43 @@
+// RUN: %clang_cc1 -I %S/Inputs -fsycl -triple spir64_gen -DGPU -fsycl-is-device -fsyntax-only -verify %s
+// RUN: %clang_cc1 -I %S/Inputs -fsycl -triple spir64 -fsycl-is-device -fsyntax-only -verify %s
+// RUN: %clang_cc1 -I %S/Inputs -fsycl -triple spir64_gen -Wno-sycl-strict -fsycl-is-device -fsyntax-only -verify %s
+// RUN: %clang_cc1 -I %S/Inputs -fsycl -triple spir64_gen -Werror=sycl-strict -DERROR -fsycl-is-device -fsyntax-only -verify %s
+
+#include <sycl.hpp>
+
+template <typename Name, typename F>
+__attribute__((sycl_kernel)) void kernel(F KernelFunc) {
+  KernelFunc();
+}
+
+template <typename Name, typename F>
+void parallel_for(F KernelFunc) {
+#ifdef GPU
+  // expected-warning@+8 {{kernel argument count (2001) exceeds supported maximum of 2000 on GPU}}
+  // expected-note@+7 {{array elements and fields of a class/struct may be counted separately}}
+#elif ERROR
+  // expected-error@+5 {{kernel argument count (2001) exceeds supported maximum of 2000 on GPU}}
+  // expected-note@+4 {{array elements and fields of a class/struct may be counted separately}}
+#else
+  // expected-no-diagnostics
+#endif
+  kernel<Name>(KernelFunc);
+}
+
+using Accessor =
+    cl::sycl::accessor<int, 1, cl::sycl::access::mode::read_write, cl::sycl::access::target::global_buffer>;
+
+void use() {
+  struct S {
+    int A;
+    int B;
+    Accessor AAcc;
+    Accessor BAcc;
+    int Array[1991];
+  } Args;
+  auto L = [=]() { (void)Args; };
+#if defined(GPU) || defined(ERROR)
+  // expected-note@+2 {{in instantiation of function template specialization 'parallel_for<Foo}}
+#endif
+  parallel_for<class Foo>(L);
+}