merge main into staging-sm6.9

CMakeLists.txt

@@ -762,9 +762,7 @@ if (LLVM_INCLUDE_DOCS)
   add_subdirectory(docs)
 endif()
 
-if (LLVM_BUILD_DOCS)
-  add_hlsl_hctgen(DxilDocs OUTPUT docs/DXIL.rst CODE_TAG) # HLSL Change
-endif()
+add_hlsl_hctgen(DxilDocs OUTPUT docs/DXIL.rst CODE_TAG) # HLSL Change
 
 add_subdirectory(cmake/modules)
 

include/dxc/HlslIntrinsicOp.h

@@ -336,7 +336,34 @@ enum class IntrinsicOp {
   MOP_TraceRayInline = 325,
   MOP_WorldRayDirection = 326,
   MOP_WorldRayOrigin = 327,
+  MOP_DxHitObject_FromRayQuery = 363,
+  MOP_DxHitObject_GetAttributes = 364,
+  MOP_DxHitObject_GetGeometryIndex = 365,
+  MOP_DxHitObject_GetHitKind = 366,
+  MOP_DxHitObject_GetInstanceID = 367,
+  MOP_DxHitObject_GetInstanceIndex = 368,
+  MOP_DxHitObject_GetObjectRayDirection = 369,
+  MOP_DxHitObject_GetObjectRayOrigin = 370,
+  MOP_DxHitObject_GetObjectToWorld3x4 = 371,
+  MOP_DxHitObject_GetObjectToWorld4x3 = 372,
+  MOP_DxHitObject_GetPrimitiveIndex = 373,
+  MOP_DxHitObject_GetRayFlags = 374,
+  MOP_DxHitObject_GetRayTCurrent = 375,
+  MOP_DxHitObject_GetRayTMin = 376,
+  MOP_DxHitObject_GetShaderTableIndex = 377,
+  MOP_DxHitObject_GetWorldRayDirection = 378,
+  MOP_DxHitObject_GetWorldRayOrigin = 379,
+  MOP_DxHitObject_GetWorldToObject3x4 = 380,
+  MOP_DxHitObject_GetWorldToObject4x3 = 381,
+  MOP_DxHitObject_Invoke = 382,
+  MOP_DxHitObject_IsHit = 383,
+  MOP_DxHitObject_IsMiss = 384,
+  MOP_DxHitObject_IsNop = 385,
+  MOP_DxHitObject_LoadLocalRootTableConstant = 386,
+  MOP_DxHitObject_MakeMiss = 387,
   MOP_DxHitObject_MakeNop = 358,
+  MOP_DxHitObject_SetShaderTableIndex = 388,
+  MOP_DxHitObject_TraceRay = 389,
   IOP_DxMaybeReorderThread = 359,
   MOP_Count = 328,
   MOP_FinishedCrossGroupSharing = 329,
@@ -369,7 +396,7 @@ enum class IntrinsicOp {
   IOP_usign = 355,
   MOP_InterlockedUMax = 356,
   MOP_InterlockedUMin = 357,
-  Num_Intrinsics = 363,
+  Num_Intrinsics = 390,
 };
 inline bool HasUnsignedIntrinsicOpcode(IntrinsicOp opcode) {
   switch (opcode) {

include/dxc/dxcapi.internal.h

@@ -131,12 +131,13 @@ enum LEGAL_INTRINSIC_COMPTYPES {
   LICOMPTYPE_THREAD_NODE_OUTPUT_RECORDS = 50,
 
   LICOMPTYPE_HIT_OBJECT = 51,
+  LICOMPTYPE_RAY_QUERY = 52,
 
 #ifdef ENABLE_SPIRV_CODEGEN
-  LICOMPTYPE_VK_BUFFER_POINTER = 52,
-  LICOMPTYPE_COUNT = 53
+  LICOMPTYPE_VK_BUFFER_POINTER = 53,
+  LICOMPTYPE_COUNT = 54
 #else
-  LICOMPTYPE_COUNT = 52
+  LICOMPTYPE_COUNT = 53
 #endif
 };
 

lib/HLSL/HLOperationLower.cpp

@@ -6197,6 +6197,77 @@ Value *TranslateMaybeReorderThread(CallInst *CI, IntrinsicOp IOP,
                                    bool &Translated) {
   return nullptr; // TODO: Merge SER DXIL patches
 }
+
+Value *TranslateHitObjectFromRayQuery(CallInst *CI, IntrinsicOp IOP,
+                                      OP::OpCode OpCode,
+                                      HLOperationLowerHelper &Helper,
+                                      HLObjectOperationLowerHelper *pObjHelper,
+                                      bool &Translated) {
+  return UndefValue::get(CI->getType()); // TODO: Merge SER DXIL patches
+}
+
+Value *TranslateHitObjectTraceRay(CallInst *CI, IntrinsicOp IOP,
+                                  OP::OpCode OpCode,
+                                  HLOperationLowerHelper &Helper,
+                                  HLObjectOperationLowerHelper *pObjHelper,
+                                  bool &Translated) {
+  return UndefValue::get(CI->getType()); // TODO: Merge SER DXIL patches
+}
+
+Value *TranslateHitObjectInvoke(CallInst *CI, IntrinsicOp IOP,
+                                OP::OpCode OpCode,
+                                HLOperationLowerHelper &Helper,
+                                HLObjectOperationLowerHelper *pObjHelper,
+                                bool &Translated) {
+  return nullptr; // TODO: Merge SER DXIL patches
+}
+
+Value *TranslateHitObjectGetAttributes(CallInst *CI, IntrinsicOp IOP,
+                                       OP::OpCode OpCode,
+                                       HLOperationLowerHelper &Helper,
+                                       HLObjectOperationLowerHelper *pObjHelper,
+                                       bool &Translated) {
+  return UndefValue::get(CI->getType()); // TODO: Merge SER DXIL patches
+}
+
+Value *TranslateHitObjectScalarGetter(CallInst *CI, IntrinsicOp IOP,
+                                      OP::OpCode OpCode,
+                                      HLOperationLowerHelper &Helper,
+                                      HLObjectOperationLowerHelper *pObjHelper,
+                                      bool &Translated) {
+  return UndefValue::get(CI->getType()); // TODO: Merge SER DXIL patches
+}
+
+Value *TranslateHitObjectVectorGetter(CallInst *CI, IntrinsicOp IOP,
+                                      OP::OpCode OpCode,
+                                      HLOperationLowerHelper &Helper,
+                                      HLObjectOperationLowerHelper *pObjHelper,
+                                      bool &Translated) {
+  return UndefValue::get(CI->getType()); // TODO: Merge SER DXIL patches
+}
+
+Value *TranslateHitObjectMatrixGetter(CallInst *CI, IntrinsicOp IOP,
+                                      OP::OpCode OpCode,
+                                      HLOperationLowerHelper &Helper,
+                                      HLObjectOperationLowerHelper *pObjHelper,
+                                      bool &Translated) {
+  return UndefValue::get(CI->getType()); // TODO: Merge SER DXIL patches
+}
+
+Value *TranslateHitObjectLoadLocalRootTableConstant(
+    CallInst *CI, IntrinsicOp IOP, OP::OpCode OpCode,
+    HLOperationLowerHelper &Helper, HLObjectOperationLowerHelper *pObjHelper,
+    bool &Translated) {
+  return UndefValue::get(CI->getType()); // TODO: Merge SER DXIL patches
+}
+
+Value *TranslateHitObjectSetShaderTableIndex(
+    CallInst *CI, IntrinsicOp IOP, OP::OpCode OpCode,
+    HLOperationLowerHelper &Helper, HLObjectOperationLowerHelper *pObjHelper,
+    bool &Translated) {
+  return UndefValue::get(CI->getType()); // TODO: Merge SER DXIL patches
+}
+
 } // namespace
 
 // Resource Handle.
@@ -6908,6 +6979,63 @@ IntrinsicLower gLowerTable[] = {
      DXIL::OpCode::NumOpCodes},
     {IntrinsicOp::MOP_GetBufferContents, UnsupportedVulkanIntrinsic,
      DXIL::OpCode::NumOpCodes},
+    {IntrinsicOp::MOP_DxHitObject_FromRayQuery, TranslateHitObjectFromRayQuery,
+     DXIL::OpCode::HitObject_FromRayQuery},
+    {IntrinsicOp::MOP_DxHitObject_GetAttributes,
+     TranslateHitObjectGetAttributes, DXIL::OpCode::HitObject_Attributes},
+    {IntrinsicOp::MOP_DxHitObject_GetGeometryIndex,
+     TranslateHitObjectScalarGetter, DXIL::OpCode::HitObject_GeometryIndex},
+    {IntrinsicOp::MOP_DxHitObject_GetHitKind, TranslateHitObjectScalarGetter,
+     DXIL::OpCode::HitObject_HitKind},
+    {IntrinsicOp::MOP_DxHitObject_GetInstanceID, TranslateHitObjectScalarGetter,
+     DXIL::OpCode::HitObject_InstanceID},
+    {IntrinsicOp::MOP_DxHitObject_GetInstanceIndex,
+     TranslateHitObjectScalarGetter, DXIL::OpCode::HitObject_InstanceIndex},
+    {IntrinsicOp::MOP_DxHitObject_GetObjectRayDirection,
+     TranslateHitObjectVectorGetter,
+     DXIL::OpCode::HitObject_ObjectRayDirection},
+    {IntrinsicOp::MOP_DxHitObject_GetObjectRayOrigin,
+     TranslateHitObjectVectorGetter, DXIL::OpCode::HitObject_ObjectRayOrigin},
+    {IntrinsicOp::MOP_DxHitObject_GetObjectToWorld3x4,
+     TranslateHitObjectMatrixGetter, DXIL::OpCode::HitObject_ObjectToWorld3x4},
+    {IntrinsicOp::MOP_DxHitObject_GetObjectToWorld4x3,
+     TranslateHitObjectMatrixGetter, DXIL::OpCode::HitObject_ObjectToWorld3x4},
+    {IntrinsicOp::MOP_DxHitObject_GetPrimitiveIndex,
+     TranslateHitObjectScalarGetter, DXIL::OpCode::HitObject_PrimitiveIndex},
+    {IntrinsicOp::MOP_DxHitObject_GetRayFlags, TranslateHitObjectScalarGetter,
+     DXIL::OpCode::HitObject_RayFlags},
+    {IntrinsicOp::MOP_DxHitObject_GetRayTCurrent,
+     TranslateHitObjectScalarGetter, DXIL::OpCode::HitObject_RayTCurrent},
+    {IntrinsicOp::MOP_DxHitObject_GetRayTMin, TranslateHitObjectScalarGetter,
+     DXIL::OpCode::HitObject_RayTMin},
+    {IntrinsicOp::MOP_DxHitObject_GetShaderTableIndex,
+     TranslateHitObjectScalarGetter, DXIL::OpCode::HitObject_ShaderTableIndex},
+    {IntrinsicOp::MOP_DxHitObject_GetWorldRayDirection,
+     TranslateHitObjectVectorGetter, DXIL::OpCode::HitObject_WorldRayDirection},
+    {IntrinsicOp::MOP_DxHitObject_GetWorldRayOrigin,
+     TranslateHitObjectVectorGetter, DXIL::OpCode::HitObject_WorldRayOrigin},
+    {IntrinsicOp::MOP_DxHitObject_GetWorldToObject3x4,
+     TranslateHitObjectMatrixGetter, DXIL::OpCode::HitObject_WorldToObject3x4},
+    {IntrinsicOp::MOP_DxHitObject_GetWorldToObject4x3,
+     TranslateHitObjectMatrixGetter, DXIL::OpCode::HitObject_WorldToObject3x4},
+    {IntrinsicOp::MOP_DxHitObject_Invoke, TranslateHitObjectInvoke,
+     DXIL::OpCode::HitObject_Invoke},
+    {IntrinsicOp::MOP_DxHitObject_IsHit, TranslateHitObjectScalarGetter,
+     DXIL::OpCode::HitObject_IsHit},
+    {IntrinsicOp::MOP_DxHitObject_IsMiss, TranslateHitObjectScalarGetter,
+     DXIL::OpCode::HitObject_IsMiss},
+    {IntrinsicOp::MOP_DxHitObject_IsNop, TranslateHitObjectScalarGetter,
+     DXIL::OpCode::HitObject_IsNop},
+    {IntrinsicOp::MOP_DxHitObject_LoadLocalRootTableConstant,
+     TranslateHitObjectLoadLocalRootTableConstant,
+     DXIL::OpCode::HitObject_LoadLocalRootTableConstant},
+    {IntrinsicOp::MOP_DxHitObject_MakeMiss, TranslateHitObjectMake,
+     DXIL::OpCode::HitObject_MakeMiss},
+    {IntrinsicOp::MOP_DxHitObject_SetShaderTableIndex,
+     TranslateHitObjectSetShaderTableIndex,
+     DXIL::OpCode::HitObject_SetShaderTableIndex},
+    {IntrinsicOp::MOP_DxHitObject_TraceRay, TranslateHitObjectTraceRay,
+     DXIL::OpCode::HitObject_TraceRay},
 };
 } // namespace
 static_assert(

tools/clang/lib/AST/ASTContextHLSL.cpp

@@ -1390,19 +1390,27 @@ CXXRecordDecl *hlsl::DeclareVkBufferPointerType(ASTContext &context,
       DeclarationName(&context.Idents.get("Get")), true);
   CanQualType canQualType =
       recordDecl->getTypeForDecl()->getCanonicalTypeUnqualified();
-  CreateConstructorDeclarationWithParams(
+  auto *copyConstructorDecl = CreateConstructorDeclarationWithParams(
       context, recordDecl, context.VoidTy,
       {context.getRValueReferenceType(canQualType)}, {"bufferPointer"},
-      context.DeclarationNames.getCXXConstructorName(canQualType), false);
-  CreateConstructorDeclarationWithParams(
+      context.DeclarationNames.getCXXConstructorName(canQualType), false, true);
+  auto *addressConstructorDecl = CreateConstructorDeclarationWithParams(
       context, recordDecl, context.VoidTy, {context.UnsignedIntTy}, {"address"},
-      context.DeclarationNames.getCXXConstructorName(canQualType), false);
+      context.DeclarationNames.getCXXConstructorName(canQualType), false, true);
+  hlsl::CreateFunctionTemplateDecl(
+      context, recordDecl, copyConstructorDecl,
+      Builder.getTemplateDecl()->getTemplateParameters()->begin(), 2);
+  hlsl::CreateFunctionTemplateDecl(
+      context, recordDecl, addressConstructorDecl,
+      Builder.getTemplateDecl()->getTemplateParameters()->begin(), 2);
 
   StringRef OpcodeGroup = GetHLOpcodeGroupName(HLOpcodeGroup::HLIntrinsic);
   unsigned Opcode = static_cast<unsigned>(IntrinsicOp::MOP_GetBufferContents);
   methodDecl->addAttr(
       HLSLIntrinsicAttr::CreateImplicit(context, OpcodeGroup, "", Opcode));
   methodDecl->addAttr(HLSLCXXOverloadAttr::CreateImplicit(context));
+  copyConstructorDecl->addAttr(HLSLCXXOverloadAttr::CreateImplicit(context));
+  addressConstructorDecl->addAttr(HLSLCXXOverloadAttr::CreateImplicit(context));
 
   return Builder.completeDefinition();
 }

tools/clang/lib/Sema/SemaExprCXX.cpp

@@ -1057,26 +1057,51 @@ Sema::BuildCXXTypeConstructExpr(TypeSourceInfo *TInfo,
     Expr *Arg = Exprs[0];
 #ifdef ENABLE_SPIRV_CODEGEN
     if (hlsl::IsVKBufferPointerType(Ty) && Arg->getType()->isIntegerType()) {
-      for (auto *ctor : Ty->getAsCXXRecordDecl()->ctors()) {
-        if (auto *functionType = ctor->getType()->getAs<FunctionProtoType>()) {
-          if (functionType->getNumParams() != 1 ||
-              !functionType->getParamType(0)->isIntegerType())
-            continue;
-
-          CanQualType argType = Arg->getType()->getCanonicalTypeUnqualified();
-          if (!Arg->isRValue()) {
-            Arg = ImpCastExprToType(Arg, argType, CK_LValueToRValue).get();
-          }
-          if (argType != Context.UnsignedLongLongTy) {
-            Arg = ImpCastExprToType(Arg, Context.UnsignedLongLongTy,
-                                    CK_IntegralCast)
-                      .get();
-          }
-          return CXXConstructExpr::Create(
-              Context, Ty, TyBeginLoc, ctor, false, {Arg}, false, false, false,
-              false, CXXConstructExpr::ConstructionKind::CK_Complete,
-              SourceRange(LParenLoc, RParenLoc));
+      typedef DeclContext::specific_decl_iterator<FunctionTemplateDecl> ft_iter;
+      auto *recordDecl = Ty->getAsCXXRecordDecl();
+      auto *specDecl = cast<ClassTemplateSpecializationDecl>(recordDecl);
+      auto *templatedDecl =
+          specDecl->getSpecializedTemplate()->getTemplatedDecl();
+      auto functionTemplateDecls =
+          llvm::iterator_range<ft_iter>(ft_iter(templatedDecl->decls_begin()),
+                                        ft_iter(templatedDecl->decls_end()));
+      for (auto *ftd : functionTemplateDecls) {
+        auto *fd = ftd->getTemplatedDecl();
+        if (fd->getNumParams() != 1 ||
+            !fd->getParamDecl(0)->getType()->isIntegerType())
+          continue;
+
+        void *insertPos;
+        auto templateArgs = ftd->getInjectedTemplateArgs();
+        auto *functionDecl = ftd->findSpecialization(templateArgs, insertPos);
+        if (!functionDecl) {
+          DeclarationNameInfo DInfo(ftd->getDeclName(),
+                                    recordDecl->getLocation());
+          auto *templateArgList = TemplateArgumentList::CreateCopy(
+              Context, templateArgs.data(), templateArgs.size());
+          functionDecl = CXXConstructorDecl::Create(
+              Context, recordDecl, Arg->getLocStart(), DInfo, Ty, TInfo, false,
+              false, false, false);
+          functionDecl->setFunctionTemplateSpecialization(ftd, templateArgList,
+                                                          insertPos);
+        } else if (functionDecl->getDeclKind() != Decl::Kind::CXXConstructor) {
+          continue;
+        }
+
+        CanQualType argType = Arg->getType()->getCanonicalTypeUnqualified();
+        if (!Arg->isRValue()) {
+          Arg = ImpCastExprToType(Arg, argType, CK_LValueToRValue).get();
+        }
+        if (argType != Context.UnsignedLongLongTy) {
+          Arg = ImpCastExprToType(Arg, Context.UnsignedLongLongTy,
+                                  CK_IntegralCast)
+                    .get();
         }
+        return CXXConstructExpr::Create(
+            Context, Ty, TyBeginLoc, cast<CXXConstructorDecl>(functionDecl),
+            false, {Arg}, false, false, false, false,
+            CXXConstructExpr::ConstructionKind::CK_Complete,
+            SourceRange(LParenLoc, RParenLoc));
       }
     }
 #endif

tools/clang/lib/Sema/SemaHLSL.cpp

@@ -580,9 +580,9 @@ const UINT g_uBasicKindProps[] = {
     0, // AR_OBJECT_PROCEDURAL_PRIMITIVE_HIT_GROUP,
     0, // AR_OBJECT_RAYTRACING_PIPELINE_CONFIG1,
 
-    BPROP_OBJECT, // AR_OBJECT_RAY_QUERY,
-    BPROP_OBJECT, // AR_OBJECT_HEAP_RESOURCE,
-    BPROP_OBJECT, // AR_OBJECT_HEAP_SAMPLER,
+    LICOMPTYPE_RAY_QUERY, // AR_OBJECT_RAY_QUERY,
+    BPROP_OBJECT,         // AR_OBJECT_HEAP_RESOURCE,
+    BPROP_OBJECT,         // AR_OBJECT_HEAP_SAMPLER,
 
     BPROP_OBJECT | BPROP_RWBUFFER | BPROP_TEXTURE, // AR_OBJECT_RWTEXTURE2DMS
     BPROP_OBJECT | BPROP_RWBUFFER |
@@ -1135,6 +1135,9 @@ static const ArBasicKind g_ResourceCT[] = {AR_OBJECT_HEAP_RESOURCE,
 
 static const ArBasicKind g_RayDescCT[] = {AR_OBJECT_RAY_DESC, AR_BASIC_UNKNOWN};
 
+static const ArBasicKind g_RayQueryCT[] = {AR_OBJECT_RAY_QUERY,
+                                           AR_BASIC_UNKNOWN};
+
 static const ArBasicKind g_AccelerationStructCT[] = {
     AR_OBJECT_ACCELERATION_STRUCT, AR_BASIC_UNKNOWN};
 
@@ -1297,6 +1300,7 @@ const ArBasicKind *g_LegalIntrinsicCompTypes[] = {
     g_GroupNodeOutputRecordsCT,  // LICOMPTYPE_GROUP_NODE_OUTPUT_RECORDS
     g_ThreadNodeOutputRecordsCT, // LICOMPTYPE_THREAD_NODE_OUTPUT_RECORDS
     g_DxHitObjectCT,             // LICOMPTYPE_HIT_OBJECT
+    g_RayQueryCT,                // LICOMPTYPE_RAY_QUERY
 #ifdef ENABLE_SPIRV_CODEGEN
     g_VKBufferPointerCT, // LICOMPTYPE_VK_BUFFER_POINTER
 #endif

tools/clang/test/CodeGenSPIRV/vk.buffer-pointer.from-uint.hlsl

@@ -0,0 +1,46 @@
+// RUN: %dxc -spirv -Od -T cs_6_7 %s | FileCheck %s
+// RUN: %dxc -spirv -Od -T cs_6_7 -DALIGN_16 %s | FileCheck %s
+// RUN: %dxc -spirv -Od -T cs_6_7 -DNO_PC %s | FileCheck %s
+
+// Was getting bogus type errors with the defined changes
+
+#ifdef ALIGN_16
+typedef vk::BufferPointer<uint, 16> BufferType;
+#else
+typedef vk::BufferPointer<uint, 32> BufferType;
+#endif
+#ifndef NO_PC
+struct PushConstantStruct {
+  BufferType push_buffer;
+};
+[[vk::push_constant]] PushConstantStruct push_constant;
+#endif
+
+RWStructuredBuffer<uint> output;
+
+// CHECK: [[INT:%[_0-9A-Za-z]*]] = OpTypeInt 32 1
+// CHECK: [[I0:%[_0-9A-Za-z]*]] = OpConstant [[INT]] 0
+// CHECK: [[UINT:%[_0-9A-Za-z]*]] = OpTypeInt 32 0
+// CHECK: [[U0:%[_0-9A-Za-z]*]] = OpConstant [[UINT]] 0
+// CHECK: [[PPUINT:%[_0-9A-Za-z]*]] = OpTypePointer PhysicalStorageBuffer [[UINT]]
+// CHECK: [[PFPPUINT:%[_0-9A-Za-z]*]] = OpTypePointer Function [[PPUINT]]
+// CHECK: [[PUUINT:%[_0-9A-Za-z]*]] = OpTypePointer Uniform [[UINT]]
+// CHECK: [[OUTPUT:%[_0-9A-Za-z]*]] = OpVariable %{{[_0-9A-Za-z]*}} Uniform
+
+[numthreads(1, 1, 1)]
+void main() {
+  uint64_t addr = 123;
+  vk::BufferPointer<uint, 32> test = vk::BufferPointer<uint, 32>(addr);
+  output[0] = test.Get();
+}
+
+// CHECK: [[TEST:%[_0-9A-Za-z]*]] = OpVariable [[PFPPUINT]] Function
+// CHECK: [[X1:%[_0-9A-Za-z]*]] = OpConvertUToPtr [[PPUINT]]
+// CHECK: OpStore [[TEST]] [[X1]]
+// CHECK: [[X2:%[_0-9A-Za-z]*]] = OpLoad [[PPUINT]] [[TEST]] Aligned 32
+// CHECK: [[X3:%[_0-9A-Za-z]*]] = OpLoad [[UINT]] [[X2]] Aligned 4
+// CHECK: [[X4:%[_0-9A-Za-z]*]] = OpAccessChain [[PUUINT]] [[OUTPUT]] [[I0]] [[U0]]
+// CHECK: OpStore [[X4]] [[X3]]
+// CHECK: OpReturn
+// CHECK: OpFunctionEnd
+