Merge from 'sycl' to 'sycl-web' (#10)

Author: iclsrc

clang/include/clang/Basic/DiagnosticSemaKinds.td

@@ -9758,6 +9758,8 @@ def warn_opencl_generic_address_space_arg : Warning<
   "passing non-generic address space pointer to %0"
   " may cause dynamic conversion affecting performance">,
   InGroup<Conversion>, DefaultIgnore;
+def err_bad_union_kernel_param_members : Error<
+  "%0 cannot be used inside a union kernel parameter">;
 
 // OpenCL v2.0 s6.13.6 -- Builtin Pipe Functions
 def err_opencl_builtin_pipe_first_arg : Error<

clang/lib/Driver/ToolChains/Clang.cpp

@@ -4113,8 +4113,10 @@ void Clang::ConstructJob(Compilation &C, const JobAction &JA,
                      false))
       CmdArgs.push_back("-fsycl-explicit-simd");
 
+    // Default value for FPGA is false, for all other targets is true.
     if (!Args.hasFlag(options::OPT_fsycl_early_optimizations,
-                      options::OPT_fno_sycl_early_optimizations, true))
+                      options::OPT_fno_sycl_early_optimizations,
+                      Triple.getSubArch() != llvm::Triple::SPIRSubArch_fpga))
       CmdArgs.push_back("-fno-sycl-early-optimizations");
     else if (RawTriple.isSPIR()) {
       // Set `sycl-opt` option to configure LLVM passes for SPIR target

clang/lib/Sema/SemaSYCL.cpp

@@ -831,6 +831,9 @@ class KernelObjVisitor {
     else if (ElementTy->isStructureOrClassType())
       VisitRecord(Owner, ArrayField, ElementTy->getAsCXXRecordDecl(),
                   handlers...);
+    else if (ElementTy->isUnionType())
+      VisitUnion(Owner, ArrayField, ElementTy->getAsCXXRecordDecl(),
+                 handlers...);
     else if (ElementTy->isArrayType())
       VisitArrayElements(ArrayField, ElementTy, handlers...);
     else if (ElementTy->isScalarType())
@@ -858,6 +861,65 @@ class KernelObjVisitor {
   void VisitRecord(const CXXRecordDecl *Owner, ParentTy &Parent,
                    const CXXRecordDecl *Wrapper, Handlers &... handlers);
 
+  // Base case, only calls these when filtered.
+  template <typename... FilteredHandlers, typename ParentTy>
+  std::enable_if_t<(sizeof...(FilteredHandlers) > 0)>
+  VisitUnion(FilteredHandlers &... handlers, const CXXRecordDecl *Owner,
+             ParentTy &Parent, const CXXRecordDecl *Wrapper) {
+    (void)std::initializer_list<int>{
+        (handlers.enterUnion(Owner, Parent), 0)...};
+    VisitRecordHelper(Wrapper, Wrapper->fields(), handlers...);
+    (void)std::initializer_list<int>{
+        (handlers.leaveUnion(Owner, Parent), 0)...};
+  }
+
+  // Handle empty base case.
+  template <typename ParentTy>
+  void VisitUnion(const CXXRecordDecl *Owner, ParentTy &Parent,
+                  const CXXRecordDecl *Wrapper) {}
+
+  template <typename... FilteredHandlers, typename ParentTy,
+            typename CurHandler, typename... Handlers>
+  std::enable_if_t<!CurHandler::VisitUnionBody &&
+                   (sizeof...(FilteredHandlers) > 0)>
+  VisitUnion(FilteredHandlers &... filtered_handlers,
+             const CXXRecordDecl *Owner, ParentTy &Parent,
+             const CXXRecordDecl *Wrapper, CurHandler &cur_handler,
+             Handlers &... handlers) {
+    VisitUnion<FilteredHandlers...>(filtered_handlers..., Owner, Parent,
+                                    Wrapper, handlers...);
+  }
+
+  template <typename... FilteredHandlers, typename ParentTy,
+            typename CurHandler, typename... Handlers>
+  std::enable_if_t<CurHandler::VisitUnionBody &&
+                   (sizeof...(FilteredHandlers) > 0)>
+  VisitUnion(FilteredHandlers &... filtered_handlers,
+             const CXXRecordDecl *Owner, ParentTy &Parent,
+             const CXXRecordDecl *Wrapper, CurHandler &cur_handler,
+             Handlers &... handlers) {
+    VisitUnion<FilteredHandlers..., CurHandler>(
+        filtered_handlers..., cur_handler, Owner, Parent, Wrapper, handlers...);
+  }
+
+  // Add overloads without having filtered-handlers
+  // to handle leading-empty argument packs.
+  template <typename ParentTy, typename CurHandler, typename... Handlers>
+  std::enable_if_t<!CurHandler::VisitUnionBody>
+  VisitUnion(const CXXRecordDecl *Owner, ParentTy &Parent,
+             const CXXRecordDecl *Wrapper, CurHandler &cur_handler,
+             Handlers &... handlers) {
+    VisitUnion(Owner, Parent, Wrapper, handlers...);
+  }
+
+  template <typename ParentTy, typename CurHandler, typename... Handlers>
+  std::enable_if_t<CurHandler::VisitUnionBody>
+  VisitUnion(const CXXRecordDecl *Owner, ParentTy &Parent,
+             const CXXRecordDecl *Wrapper, CurHandler &cur_handler,
+             Handlers &... handlers) {
+    VisitUnion<CurHandler>(cur_handler, Owner, Parent, Wrapper, handlers...);
+  }
+
   template <typename... Handlers>
   void VisitRecordHelper(const CXXRecordDecl *Owner,
                          clang::CXXRecordDecl::base_class_const_range Range,
@@ -943,6 +1005,11 @@ class KernelObjVisitor {
           CXXRecordDecl *RD = FieldTy->getAsCXXRecordDecl();
           VisitRecord(Owner, Field, RD, handlers...);
         }
+      } else if (FieldTy->isUnionType()) {
+        if (KF_FOR_EACH(handleUnionType, Field, FieldTy)) {
+          CXXRecordDecl *RD = FieldTy->getAsCXXRecordDecl();
+          VisitUnion(Owner, Field, RD, handlers...);
+        }
       } else if (FieldTy->isReferenceType())
         KF_FOR_EACH(handleReferenceType, Field, FieldTy);
       else if (FieldTy->isPointerType())
@@ -982,6 +1049,7 @@ class SyclKernelFieldHandler {
   SyclKernelFieldHandler(Sema &S) : SemaRef(S) {}
 
 public:
+  static constexpr const bool VisitUnionBody = false;
   // Mark these virtual so that we can use override in the implementer classes,
   // despite virtual dispatch never being used.
 
@@ -1006,6 +1074,7 @@ class SyclKernelFieldHandler {
   }
   virtual bool handleSyclHalfType(FieldDecl *, QualType) { return true; }
   virtual bool handleStructType(FieldDecl *, QualType) { return true; }
+  virtual bool handleUnionType(FieldDecl *, QualType) { return true; }
   virtual bool handleReferenceType(FieldDecl *, QualType) { return true; }
   virtual bool handlePointerType(FieldDecl *, QualType) { return true; }
   virtual bool handleArrayType(FieldDecl *, QualType) { return true; }
@@ -1025,6 +1094,8 @@ class SyclKernelFieldHandler {
   virtual bool leaveStruct(const CXXRecordDecl *, const CXXBaseSpecifier &) {
     return true;
   }
+  virtual bool enterUnion(const CXXRecordDecl *, FieldDecl *) { return true; }
+  virtual bool leaveUnion(const CXXRecordDecl *, FieldDecl *) { return true; }
 
   // The following are used for stepping through array elements.
 
@@ -1047,7 +1118,6 @@ class SyclKernelFieldHandler {
 class SyclKernelFieldChecker : public SyclKernelFieldHandler {
   bool IsInvalid = false;
   DiagnosticsEngine &Diag;
-
   // Check whether the object should be disallowed from being copied to kernel.
   // Return true if not copyable, false if copyable.
   bool checkNotCopyableToKernel(const FieldDecl *FD, const QualType &FieldTy) {
@@ -1202,6 +1272,65 @@ class SyclKernelFieldChecker : public SyclKernelFieldHandler {
   }
 };
 
+// A type to check the validity of accessing accessor/sampler/stream
+// types as kernel parameters inside union.
+class SyclKernelUnionChecker : public SyclKernelFieldHandler {
+  int UnionCount = 0;
+  bool IsInvalid = false;
+  DiagnosticsEngine &Diag;
+
+public:
+  SyclKernelUnionChecker(Sema &S)
+      : SyclKernelFieldHandler(S), Diag(S.getASTContext().getDiagnostics()) {}
+  bool isValid() { return !IsInvalid; }
+  static constexpr const bool VisitUnionBody = true;
+
+  bool checkType(SourceLocation Loc, QualType Ty) {
+    if (UnionCount) {
+      IsInvalid = true;
+      Diag.Report(Loc, diag::err_bad_union_kernel_param_members) << Ty;
+    }
+    return isValid();
+  }
+
+  bool enterUnion(const CXXRecordDecl *RD, FieldDecl *FD) {
+    ++UnionCount;
+    return true;
+  }
+
+  bool leaveUnion(const CXXRecordDecl *RD, FieldDecl *FD) {
+    --UnionCount;
+    return true;
+  }
+
+  bool handleSyclAccessorType(FieldDecl *FD, QualType FieldTy) final {
+    return checkType(FD->getLocation(), FieldTy);
+  }
+
+  bool handleSyclAccessorType(const CXXBaseSpecifier &BS,
+                              QualType FieldTy) final {
+    return checkType(BS.getBeginLoc(), FieldTy);
+  }
+
+  bool handleSyclSamplerType(FieldDecl *FD, QualType FieldTy) final {
+    return checkType(FD->getLocation(), FieldTy);
+  }
+
+  bool handleSyclSamplerType(const CXXBaseSpecifier &BS,
+                             QualType FieldTy) final {
+    return checkType(BS.getBeginLoc(), FieldTy);
+  }
+
+  bool handleSyclStreamType(FieldDecl *FD, QualType FieldTy) final {
+    return checkType(FD->getLocation(), FieldTy);
+  }
+
+  bool handleSyclStreamType(const CXXBaseSpecifier &BS,
+                            QualType FieldTy) final {
+    return checkType(BS.getBeginLoc(), FieldTy);
+  }
+};
+
 // A type to Create and own the FunctionDecl for the kernel.
 class SyclKernelDeclCreator : public SyclKernelFieldHandler {
   FunctionDecl *KernelDecl;
@@ -1453,6 +1582,10 @@ class SyclKernelDeclCreator : public SyclKernelFieldHandler {
     return true;
   }
 
+  bool handleUnionType(FieldDecl *FD, QualType FieldTy) final {
+    return handleScalarType(FD, FieldTy);
+  }
+
   bool handleSyclHalfType(FieldDecl *FD, QualType FieldTy) final {
     addParam(FD, FieldTy);
     return true;
@@ -1805,6 +1938,10 @@ class SyclKernelBodyCreator : public SyclKernelFieldHandler {
     return true;
   }
 
+  bool handleUnionType(FieldDecl *FD, QualType FieldTy) final {
+    return handleScalarType(FD, FieldTy);
+  }
+
   bool enterStruct(const CXXRecordDecl *RD, const CXXBaseSpecifier &BS) final {
     CXXCastPath BasePath;
     QualType DerivedTy(RD->getTypeForDecl(), 0);
@@ -2012,6 +2149,10 @@ class SyclKernelIntHeaderCreator : public SyclKernelFieldHandler {
     return true;
   }
 
+  bool handleUnionType(FieldDecl *FD, QualType FieldTy) final {
+    return handleScalarType(FD, FieldTy);
+  }
+
   bool handleSyclStreamType(FieldDecl *FD, QualType FieldTy) final {
     addParam(FD, FieldTy, SYCLIntegrationHeader::kind_std_layout);
     return true;
@@ -2105,14 +2246,14 @@ void Sema::CheckSYCLKernelCall(FunctionDecl *KernelFunc, SourceRange CallLoc,
       }
   }
 
-  SyclKernelFieldChecker Checker(*this);
-
+  SyclKernelFieldChecker FieldChecker(*this);
+  SyclKernelUnionChecker UnionChecker(*this);
   KernelObjVisitor Visitor{*this};
   DiagnosingSYCLKernel = true;
-  Visitor.VisitRecordBases(KernelObj, Checker);
-  Visitor.VisitRecordFields(KernelObj, Checker);
+  Visitor.VisitRecordBases(KernelObj, FieldChecker, UnionChecker);
+  Visitor.VisitRecordFields(KernelObj, FieldChecker, UnionChecker);
   DiagnosingSYCLKernel = false;
-  if (!Checker.isValid())
+  if (!FieldChecker.isValid() || !UnionChecker.isValid())
     KernelFunc->setInvalidDecl();
 }
 

clang/test/CodeGenSYCL/union-kernel-param-ih.cpp

@@ -0,0 +1,53 @@
+// RUN: %clang_cc1 -I %S/Inputs -fsycl -fsycl-is-device -triple spir64-unknown-unknown-sycldevice -fsycl-int-header=%t.h %s -o %t.out
+// RUN: FileCheck -input-file=%t.h %s
+
+// This test checks the integration header generated when
+// the kernel argument is union.
+
+// CHECK: #include <CL/sycl/detail/kernel_desc.hpp>
+
+// CHECK: class kernel_A;
+
+// CHECK: __SYCL_INLINE_NAMESPACE(cl) {
+// CHECK-NEXT: namespace sycl {
+// CHECK-NEXT: namespace detail {
+
+// CHECK: static constexpr
+// CHECK-NEXT: const char* const kernel_names[] = {
+// CHECK-NEXT:   "_ZTSZ4mainE8kernel_A"
+// CHECK-NEXT: };
+
+// CHECK: static constexpr
+// CHECK-NEXT: const kernel_param_desc_t kernel_signatures[] = {
+// CHECK-NEXT:   //--- _ZTSZ4mainE8kernel_A
+// CHECK-NEXT:  { kernel_param_kind_t::kind_std_layout, 12, 0 },
+// CHECK-EMPTY:
+// CHECK-NEXT:};
+
+// CHECK: static constexpr
+// CHECK-NEXT: const unsigned kernel_signature_start[] = {
+// CHECK-NEXT:  0 // _ZTSZ4mainE8kernel_A
+// CHECK-NEXT: };
+
+// CHECK: template <> struct KernelInfo<class kernel_A> {
+
+union MyUnion {
+  int FldInt;
+  char FldChar;
+  float FldArr[3];
+};
+
+template <typename name, typename Func>
+__attribute__((sycl_kernel)) void a_kernel(Func kernelFunc) {
+  kernelFunc();
+}
+
+int main() {
+
+  MyUnion obj;
+
+  a_kernel<class kernel_A>(
+      [=]() {
+        float local = obj.FldArr[2];
+      });
+}

clang/test/CodeGenSYCL/union-kernel-param.cpp

@@ -0,0 +1,41 @@
+// RUN: %clang_cc1 -fsycl -fsycl-is-device -I %S/Inputs -triple spir64-unknown-unknown-sycldevice -disable-llvm-passes -emit-llvm %s -o - | FileCheck %s
+
+// This test checks a kernel argument that is union with both array and non-array fields.
+
+#include "sycl.hpp"
+
+using namespace cl::sycl;
+
+union MyUnion {
+  int FldInt;
+  char FldChar;
+  float FldArr[3];
+};
+
+template <typename name, typename Func>
+__attribute__((sycl_kernel)) void a_kernel(Func kernelFunc) {
+  kernelFunc();
+}
+
+int main() {
+
+  MyUnion obj;
+
+  a_kernel<class kernel_A>(
+      [=]() {
+        float local = obj.FldArr[2];
+      });
+}
+
+// CHECK kernel_A parameters
+// CHECK: define spir_kernel void @{{.*}}kernel_A(%union.{{.*}}.MyUnion* byval(%union.{{.*}}.MyUnion) align 4 [[MEM_ARG:%[a-zA-Z0-9_]+]])
+
+// Check lambda object alloca
+// CHECK: [[LOCAL_OBJECT:%0]] = alloca %"class.{{.*}}.anon", align 4
+
+// CHECK: [[L_STRUCT_ADDR:%[a-zA-Z0-9_]+]] = getelementptr inbounds %"class.{{.*}}.anon", %"class.{{.*}}.anon"* [[LOCAL_OBJECT]], i32 0, i32 0
+// CHECK: [[MEMCPY_DST:%[0-9a-zA-Z_]+]] = bitcast %union.{{.*}}MyUnion* [[L_STRUCT_ADDR]] to i8*
+// CHECK: [[MEMCPY_SRC:%[0-9a-zA-Z_]+]] = bitcast %union.{{.*}}MyUnion* [[MEM_ARG]] to i8*
+// CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 4 [[MEMCPY_DST]], i8* align 4 [[MEMCPY_SRC]], i64 12, i1 false)
+// CHECK: [[ACC_CAST1:%[0-9]+]] = addrspacecast %"class.{{.*}}.anon"* [[LOCAL_OBJECT]] to %"class.{{.*}}.anon" addrspace(4)*
+// CHECK: call spir_func void @{{.*}}(%"class.{{.*}}.anon" addrspace(4)* [[ACC_CAST1]])

clang/test/Driver/sycl-device-optimizations.cpp

@@ -3,12 +3,16 @@
 // RUN:   | FileCheck -check-prefix=CHECK-DEFAULT %s
 // RUN:   %clang -### -fsycl -fsycl-device-only %s 2>&1 \
 // RUN:   | FileCheck -check-prefix=CHECK-DEFAULT %s
+// RUN:   %clang -### -fsycl -fintelfpga -fsycl-early-optimizations %s 2>&1 \
+// RUN:   | FileCheck -check-prefix=CHECK-DEFAULT %s
 // CHECK-DEFAULT-NOT: "-fno-sycl-early-optimizations"
 // CHECK-DEFAULT-NOT: "-disable-llvm-passes"
 
 /// Check "-fno-sycl-early-optimizations" is passed to the front-end:
 // RUN:   %clang -### -fsycl -fno-sycl-early-optimizations %s 2>&1 \
-// RUN:   | FileCheck -check-prefix=CHECK-NO-SYCL-STD-OPTS %s
+// RUN:   | FileCheck -check-prefix=CHECK-NO-SYCL-EARLY-OPTS %s
 // RUN:   %clang -### -fsycl -fsycl-device-only -fno-sycl-early-optimizations %s 2>&1 \
-// RUN:   | FileCheck -check-prefix=CHECK-NO-SYCL-STD-OPTS %s
-// CHECK-NO-SYCL-STD-OPTS: "-fno-sycl-early-optimizations"
+// RUN:   | FileCheck -check-prefix=CHECK-NO-SYCL-EARLY-OPTS %s
+// RUN:   %clang -### -fsycl -fintelfpga %s 2>&1 \
+// RUN:   | FileCheck -check-prefix=CHECK-NO-SYCL-EARLY-OPTS %s
+// CHECK-NO-SYCL-EARLY-OPTS: "-fno-sycl-early-optimizations"