[SM6.10] Complete TriangleObjectPositions implementation (#8097)

- Lowering for all TriangleObjectPositions variants with tests (renaming
HLSL intrinsics to match spec)
- BuiltInTrianglePositions struct definition and tests
- Validation logic with expected pass and fail tests

Closes #8030 #8031 #8032 #8033

Author: simoll

include/dxc/HLSL/HLOperations.h

@@ -160,6 +160,10 @@ const unsigned kOpcodeIdx = 0;
 // Used to initialize values that have no valid index in the HL overload.
 const unsigned kInvalidIdx = UINT32_MAX;
 
+// HL intrinsic operations place the SRet return pointer at operand index 1,
+// since the first argument is reserved for the HL OpCode.
+const unsigned kIOP_SRetOpIdx = 1;
+
 // Matrix store.
 const unsigned kMatStoreDstPtrOpIdx = 1;
 const unsigned kMatStoreValOpIdx = 2;

include/dxc/HlslIntrinsicOp.h

@@ -78,7 +78,7 @@ enum class IntrinsicOp {
   IOP_ReportHit = 67,
   IOP_SetMeshOutputCounts = 68,
   IOP_TraceRay = 69,
-  IOP_TriangleObjectPosition = 401,
+  IOP_TriangleObjectPositions = 401,
   IOP_WaveActiveAllEqual = 70,
   IOP_WaveActiveAllTrue = 71,
   IOP_WaveActiveAnyTrue = 72,
@@ -339,7 +339,7 @@ enum class IntrinsicOp {
   MOP_CandidateProceduralPrimitiveNonOpaque = 298,
   MOP_CandidateTriangleBarycentrics = 299,
   MOP_CandidateTriangleFrontFace = 300,
-  MOP_CandidateTriangleObjectPosition = 402,
+  MOP_CandidateTriangleObjectPositions = 402,
   MOP_CandidateTriangleRayT = 301,
   MOP_CandidateType = 302,
   MOP_CandidateWorldToObject3x4 = 303,
@@ -360,7 +360,7 @@ enum class IntrinsicOp {
   MOP_CommittedStatus = 317,
   MOP_CommittedTriangleBarycentrics = 318,
   MOP_CommittedTriangleFrontFace = 319,
-  MOP_CommittedTriangleObjectPosition = 403,
+  MOP_CommittedTriangleObjectPositions = 403,
   MOP_CommittedWorldToObject3x4 = 320,
   MOP_CommittedWorldToObject4x3 = 321,
   MOP_Proceed = 322,
@@ -398,7 +398,7 @@ enum class IntrinsicOp {
   MOP_DxHitObject_MakeNop = 358,
   MOP_DxHitObject_SetShaderTableIndex = 388,
   MOP_DxHitObject_TraceRay = 389,
-  MOP_DxHitObject_TriangleObjectPosition = 404,
+  MOP_DxHitObject_TriangleObjectPositions = 404,
   IOP_DxMaybeReorderThread = 359,
   MOP_Count = 328,
   MOP_FinishedCrossGroupSharing = 329,

include/dxc/dxcapi.internal.h

@@ -134,11 +134,13 @@ enum LEGAL_INTRINSIC_COMPTYPES {
                           // i32, i16, u32, u16,
                           // int8_4packed, uint8_4packed
 
+  LICOMPTYPE_BUILTIN_TRIANGLE_POSITIONS = 54,
+
 #ifdef ENABLE_SPIRV_CODEGEN
-  LICOMPTYPE_VK_BUFFER_POINTER = 54,
-  LICOMPTYPE_COUNT = 55
+  LICOMPTYPE_VK_BUFFER_POINTER = 55,
+  LICOMPTYPE_COUNT = 56
 #else
-  LICOMPTYPE_COUNT = 54
+  LICOMPTYPE_COUNT = 55
 #endif
 };
 

lib/DxilValidation/DxilValidation.cpp

@@ -2350,17 +2350,20 @@ static void ValidateDxilOperationCallInProfile(CallInst *CI,
     DxilInst_HitObject_TraceRay HOTraceRay(CI);
   } break;
 
-  // Clustered Geometry intrinsics
+  // Clustered Geometry & Triangle Object Positions intrinsics
   case DXIL::OpCode::RayQuery_CandidateClusterID:
-  case DXIL::OpCode::RayQuery_CommittedClusterID: {
+  case DXIL::OpCode::RayQuery_CommittedClusterID:
+  case DXIL::OpCode::RayQuery_CandidateTriangleObjectPosition:
+  case DXIL::OpCode::RayQuery_CommittedTriangleObjectPosition: {
     // Validate rayQueryHandle is not undef
     Value *RayQueryHandle = CI->getArgOperand(1);
     if (isa<UndefValue>(RayQueryHandle))
       ValCtx.EmitInstrError(CI, ValidationRule::InstrNoReadingUninitialized);
     break;
   }
 
-  case DXIL::OpCode::HitObject_ClusterID: {
+  case DXIL::OpCode::HitObject_ClusterID:
+  case DXIL::OpCode::HitObject_TriangleObjectPosition: {
     // Validate HitObject is not undef
     Value *HitObject = CI->getArgOperand(1);
     if (isa<UndefValue>(HitObject))

lib/HLSL/HLOperationLower.cpp

@@ -6016,6 +6016,108 @@ Value *TranslateNoArgVectorOperation(CallInst *CI, IntrinsicOp IOP,
   return retVal;
 }
 
+static Value *ConstructBuiltInTrianglePositionsFromFloat9(
+    Value *float9Vec, StructType *hlslStructTy, IRBuilder<> &Builder) {
+  Type *f32Ty = Type::getFloatTy(Builder.getContext());
+  Type *float3Ty = VectorType::get(f32Ty, 3);
+  Value *result = UndefValue::get(hlslStructTy);
+
+  // Build p0, p1, p2 from vector elements 0-2, 3-5, 6-8
+  for (unsigned field = 0; field < 3; field++) {
+    Value *float3 = UndefValue::get(float3Ty);
+    for (unsigned i = 0; i < 3; i++) {
+      Value *elem = Builder.CreateExtractElement(float9Vec, field * 3 + i);
+      float3 = Builder.CreateInsertElement(float3, elem, i);
+    }
+    result = Builder.CreateInsertValue(result, float3, field);
+  }
+
+  return result;
+}
+
+Value *TranslateTriangleObjectPositions(
+    CallInst *CI, IntrinsicOp IOP, OP::OpCode opcode,
+    HLOperationLowerHelper &helper, HLObjectOperationLowerHelper *pObjHelper,
+    bool &Translated) {
+  hlsl::OP *hlslOP = &helper.hlslOP;
+  IRBuilder<> Builder(CI);
+
+  Value *outputPtr = CI->getArgOperand(HLOperandIndex::kIOP_SRetOpIdx);
+  StructType *hlslStructTy =
+      cast<StructType>(outputPtr->getType()->getPointerElementType());
+
+  Type *f32Ty = Type::getFloatTy(CI->getContext());
+  Function *dxilFunc = hlslOP->GetOpFunc(opcode, f32Ty);
+  Constant *opArg = hlslOP->GetU32Const((unsigned)opcode);
+
+  Value *dxilCall = Builder.CreateCall(dxilFunc, {opArg});
+
+  Value *structValue = ConstructBuiltInTrianglePositionsFromFloat9(
+      dxilCall, hlslStructTy, Builder);
+  Builder.CreateStore(structValue, outputPtr);
+
+  return nullptr;
+}
+
+Value *TranslateRayQueryTriangleObjectPositions(
+    CallInst *CI, IntrinsicOp IOP, OP::OpCode opcode,
+    HLOperationLowerHelper &helper, HLObjectOperationLowerHelper *pObjHelper,
+    bool &Translated) {
+  hlsl::OP *hlslOP = &helper.hlslOP;
+
+  Value *handle = CI->getArgOperand(HLOperandIndex::kHandleOpIdx);
+  StructType *hlslStructTy =
+      cast<StructType>(CI->getType()->getPointerElementType());
+
+  Function *F = CI->getParent()->getParent();
+  IRBuilder<> AllocaBuilder(&F->getEntryBlock(), F->getEntryBlock().begin());
+  AllocaInst *resultAlloca = AllocaBuilder.CreateAlloca(hlslStructTy);
+
+  IRBuilder<> Builder(CI);
+
+  Type *f32Ty = Type::getFloatTy(CI->getContext());
+  Function *dxilFunc = hlslOP->GetOpFunc(opcode, f32Ty);
+  Constant *opArg = hlslOP->GetU32Const((unsigned)opcode);
+
+  Value *dxilCall = Builder.CreateCall(dxilFunc, {opArg, handle});
+
+  Value *structValue = ConstructBuiltInTrianglePositionsFromFloat9(
+      dxilCall, hlslStructTy, Builder);
+  Builder.CreateStore(structValue, resultAlloca);
+
+  return resultAlloca;
+}
+
+Value *TranslateHitObjectTriangleObjectPositions(
+    CallInst *CI, IntrinsicOp IOP, OP::OpCode opcode,
+    HLOperationLowerHelper &helper, HLObjectOperationLowerHelper *pObjHelper,
+    bool &Translated) {
+  hlsl::OP *hlslOP = &helper.hlslOP;
+
+  Value *hitObjectPtr = CI->getArgOperand(HLOperandIndex::kHandleOpIdx);
+  StructType *hlslStructTy =
+      cast<StructType>(CI->getType()->getPointerElementType());
+
+  Function *F = CI->getParent()->getParent();
+  IRBuilder<> AllocaBuilder(&F->getEntryBlock(), F->getEntryBlock().begin());
+  AllocaInst *resultAlloca = AllocaBuilder.CreateAlloca(hlslStructTy);
+
+  IRBuilder<> Builder(CI);
+  Value *hitObject = Builder.CreateLoad(hitObjectPtr);
+
+  Type *f32Ty = Type::getFloatTy(CI->getContext());
+  Function *dxilFunc = hlslOP->GetOpFunc(opcode, f32Ty);
+  Constant *opArg = hlslOP->GetU32Const((unsigned)opcode);
+
+  Value *dxilCall = Builder.CreateCall(dxilFunc, {opArg, hitObject});
+
+  Value *structValue = ConstructBuiltInTrianglePositionsFromFloat9(
+      dxilCall, hlslStructTy, Builder);
+  Builder.CreateStore(structValue, resultAlloca);
+
+  return resultAlloca;
+}
+
 template <typename ColElemTy>
 static void GetMatrixIndices(Constant *&Rows, Constant *&Cols, bool Is3x4,
                              LLVMContext &Ctx) {
@@ -7543,13 +7645,16 @@ constexpr IntrinsicLower gLowerTable[] = {
     {IntrinsicOp::MOP_DxHitObject_GetClusterID, TranslateHitObjectScalarGetter,
      DXIL::OpCode::HitObject_ClusterID},
 
-    {IntrinsicOp::IOP_TriangleObjectPosition, EmptyLower,
+    {IntrinsicOp::IOP_TriangleObjectPositions, TranslateTriangleObjectPositions,
      DXIL::OpCode::TriangleObjectPosition},
-    {IntrinsicOp::MOP_CandidateTriangleObjectPosition, EmptyLower,
+    {IntrinsicOp::MOP_CandidateTriangleObjectPositions,
+     TranslateRayQueryTriangleObjectPositions,
      DXIL::OpCode::RayQuery_CandidateTriangleObjectPosition},
-    {IntrinsicOp::MOP_CommittedTriangleObjectPosition, EmptyLower,
+    {IntrinsicOp::MOP_CommittedTriangleObjectPositions,
+     TranslateRayQueryTriangleObjectPositions,
      DXIL::OpCode::RayQuery_CommittedTriangleObjectPosition},
-    {IntrinsicOp::MOP_DxHitObject_TriangleObjectPosition, EmptyLower,
+    {IntrinsicOp::MOP_DxHitObject_TriangleObjectPositions,
+     TranslateHitObjectTriangleObjectPositions,
      DXIL::OpCode::HitObject_TriangleObjectPosition},
 
     {IntrinsicOp::IOP___builtin_LinAlg_CopyConvertMatrix, EmptyLower,

lib/Transforms/Scalar/ScalarReplAggregatesHLSL.cpp

@@ -1500,6 +1500,11 @@ static bool isUDTIntrinsicArg(CallInst *CI, unsigned OpIdx) {
     if (OpIdx == HLOperandIndex::kCallShaderPayloadOpIdx)
       return true;
     break;
+  case IntrinsicOp::IOP_TriangleObjectPositions:
+    // Not UDT exactly, but sret parameter that needs the same treatment.
+    if (OpIdx == HLOperandIndex::kIOP_SRetOpIdx)
+      return true;
+    break;
   case IntrinsicOp::MOP_DxHitObject_FromRayQuery:
     if (NumOps == HLOperandIndex::kHitObjectFromRayQuery_WithAttrs_NumOp &&
         OpIdx ==

tools/clang/lib/AST/HlslTypes.cpp

@@ -136,7 +136,8 @@ bool IsHLSLBuiltinRayAttributeStruct(clang::QualType QT) {
   if (const RecordType *RT = dyn_cast<RecordType>(Ty)) {
     const RecordDecl *RD = RT->getDecl();
     if (RD->getName() == "BuiltInTriangleIntersectionAttributes" ||
-        RD->getName() == "RayDesc")
+        RD->getName() == "RayDesc" ||
+        RD->getName() == "BuiltInTrianglePositions")
       return true;
   }
   return false;

tools/clang/lib/Sema/SemaHLSL.cpp

@@ -211,6 +211,7 @@ enum ArBasicKind {
   AR_OBJECT_ACCELERATION_STRUCT,
   AR_OBJECT_USER_DEFINED_TYPE,
   AR_OBJECT_TRIANGLE_INTERSECTION_ATTRIBUTES,
+  AR_OBJECT_BUILTIN_TRIANGLE_POSITIONS,
 
   // subobjects
   AR_OBJECT_STATE_OBJECT_CONFIG,
@@ -567,6 +568,7 @@ const UINT g_uBasicKindProps[] = {
     LICOMPTYPE_ACCELERATION_STRUCT, // AR_OBJECT_ACCELERATION_STRUCT
     LICOMPTYPE_USER_DEFINED_TYPE,   // AR_OBJECT_USER_DEFINED_TYPE
     0, // AR_OBJECT_TRIANGLE_INTERSECTION_ATTRIBUTES
+    LICOMPTYPE_BUILTIN_TRIANGLE_POSITIONS, // AR_OBJECT_BUILTIN_TRIANGLE_POSITIONS
 
     // subobjects
     0, // AR_OBJECT_STATE_OBJECT_CONFIG,
@@ -1134,6 +1136,9 @@ 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_BuiltInTrianglePositionsCT[] = {
+    AR_OBJECT_BUILTIN_TRIANGLE_POSITIONS, AR_BASIC_UNKNOWN};
+
 static const ArBasicKind g_RayQueryCT[] = {AR_OBJECT_RAY_QUERY,
                                            AR_BASIC_UNKNOWN};
 
@@ -1300,15 +1305,16 @@ const ArBasicKind *g_LegalIntrinsicCompTypes[] = {
     g_SInt16Or32OnlyCT,     // LICOMPTYPE_SINT16_OR_32_ONLY
     g_AnySamplerCT,         // LICOMPTYPE_ANY_SAMPLER
 
-    g_ByteAddressBufferCT,       // LICOMPTYPE_BYTEADDRESSBUFFER
-    g_RWByteAddressBufferCT,     // LICOMPTYPE_RWBYTEADDRESSBUFFER
-    g_NodeRecordOrUAVCT,         // LICOMPTYPE_NODE_RECORD_OR_UAV
-    g_AnyOutputRecordCT,         // LICOMPTYPE_ANY_NODE_OUTPUT_RECORD
-    g_GroupNodeOutputRecordsCT,  // LICOMPTYPE_GROUP_NODE_OUTPUT_RECORDS
-    g_ThreadNodeOutputRecordsCT, // LICOMPTYPE_THREAD_NODE_OUTPUT_RECORDS
-    g_DxHitObjectCT,             // LICOMPTYPE_HIT_OBJECT
-    g_RayQueryCT,                // LICOMPTYPE_RAY_QUERY
-    g_LinAlgCT,                  // LICOMPTYPE_LINALG
+    g_ByteAddressBufferCT,        // LICOMPTYPE_BYTEADDRESSBUFFER
+    g_RWByteAddressBufferCT,      // LICOMPTYPE_RWBYTEADDRESSBUFFER
+    g_NodeRecordOrUAVCT,          // LICOMPTYPE_NODE_RECORD_OR_UAV
+    g_AnyOutputRecordCT,          // LICOMPTYPE_ANY_NODE_OUTPUT_RECORD
+    g_GroupNodeOutputRecordsCT,   // LICOMPTYPE_GROUP_NODE_OUTPUT_RECORDS
+    g_ThreadNodeOutputRecordsCT,  // LICOMPTYPE_THREAD_NODE_OUTPUT_RECORDS
+    g_DxHitObjectCT,              // LICOMPTYPE_HIT_OBJECT
+    g_RayQueryCT,                 // LICOMPTYPE_RAY_QUERY
+    g_LinAlgCT,                   // LICOMPTYPE_LINALG
+    g_BuiltInTrianglePositionsCT, // LICOMPTYPE_BUILTIN_TRIANGLE_POSITIONS
 #ifdef ENABLE_SPIRV_CODEGEN
     g_VKBufferPointerCT, // LICOMPTYPE_VK_BUFFER_POINTER
 #endif
@@ -1379,6 +1385,7 @@ static const ArBasicKind g_ArBasicKindsAsTypes[] = {
 
     AR_OBJECT_WAVE, AR_OBJECT_RAY_DESC, AR_OBJECT_ACCELERATION_STRUCT,
     AR_OBJECT_TRIANGLE_INTERSECTION_ATTRIBUTES,
+    AR_OBJECT_BUILTIN_TRIANGLE_POSITIONS,
 
     // subobjects
     AR_OBJECT_STATE_OBJECT_CONFIG, AR_OBJECT_GLOBAL_ROOT_SIGNATURE,
@@ -1487,6 +1494,7 @@ static const uint8_t g_ArBasicKindsTemplateCount[] = {
     0, // AR_OBJECT_RAY_DESC
     0, // AR_OBJECT_ACCELERATION_STRUCT
     0, // AR_OBJECT_TRIANGLE_INTERSECTION_ATTRIBUTES
+    0, // AR_OBJECT_BUILTIN_TRIANGLE_POSITIONS
 
     0, // AR_OBJECT_STATE_OBJECT_CONFIG,
     0, // AR_OBJECT_GLOBAL_ROOT_SIGNATURE,
@@ -1636,6 +1644,7 @@ static const SubscriptOperatorRecord g_ArBasicKindsSubscripts[] = {
     {0, MipsFalse, SampleFalse}, // AR_OBJECT_RAY_DESC
     {0, MipsFalse, SampleFalse}, // AR_OBJECT_ACCELERATION_STRUCT
     {0, MipsFalse, SampleFalse}, // AR_OBJECT_TRIANGLE_INTERSECTION_ATTRIBUTES
+    {0, MipsFalse, SampleFalse}, // AR_OBJECT_BUILTIN_TRIANGLE_POSITIONS
 
     {0, MipsFalse, SampleFalse}, // AR_OBJECT_STATE_OBJECT_CONFIG,
     {0, MipsFalse, SampleFalse}, // AR_OBJECT_GLOBAL_ROOT_SIGNATURE,
@@ -1803,6 +1812,7 @@ static const char *g_ArBasicTypeNames[] = {
     "RaytracingAccelerationStructure",
     "user defined type",
     "BuiltInTriangleIntersectionAttributes",
+    "BuiltInTrianglePositions",
 
     // subobjects
     "StateObjectConfig",
@@ -2813,6 +2823,31 @@ AddBuiltInTriangleIntersectionAttributes(ASTContext &context,
   return attributesDecl;
 }
 
+// struct BuiltInTrianglePositions
+// {
+//   float3 p0, p1, p2;
+// };
+static CXXRecordDecl *CreateBuiltInTrianglePositions(ASTContext &context,
+                                                     QualType float3Ty) {
+  DeclContext *curDC = context.getTranslationUnitDecl();
+  IdentifierInfo &positionsId = context.Idents.get(
+      StringRef("BuiltInTrianglePositions"), tok::TokenKind::identifier);
+  CXXRecordDecl *positionsDecl = CXXRecordDecl::Create(
+      context, TagTypeKind::TTK_Struct, curDC, NoLoc, NoLoc, &positionsId,
+      nullptr, DelayTypeCreationTrue);
+  positionsDecl->addAttr(
+      FinalAttr::CreateImplicit(context, FinalAttr::Keyword_final));
+  positionsDecl->startDefinition();
+  // float3 p0, p1, p2;
+  CreateSimpleField(context, positionsDecl, "p0", float3Ty);
+  CreateSimpleField(context, positionsDecl, "p1", float3Ty);
+  CreateSimpleField(context, positionsDecl, "p2", float3Ty);
+  positionsDecl->completeDefinition();
+  positionsDecl->setImplicit(true);
+  curDC->addDecl(positionsDecl);
+  return positionsDecl;
+}
+
 //
 // Subobjects
 
@@ -3605,6 +3640,10 @@ class HLSLExternalSource : public ExternalSemaSource {
       case LICOMPTYPE_HIT_OBJECT:
         paramTypes.push_back(GetBasicKindType(AR_OBJECT_HIT_OBJECT));
         break;
+      case LICOMPTYPE_BUILTIN_TRIANGLE_POSITIONS:
+        paramTypes.push_back(
+            GetBasicKindType(AR_OBJECT_BUILTIN_TRIANGLE_POSITIONS));
+        break;
 #ifdef ENABLE_SPIRV_CODEGEN
       case LICOMPTYPE_VK_BUFFER_POINTER: {
         const ArBasicKind *match =
@@ -3824,6 +3863,10 @@ class HLSLExternalSource : public ExternalSemaSource {
             LookupVectorType(HLSLScalarType::HLSLScalarType_float, 2);
         recordDecl =
             AddBuiltInTriangleIntersectionAttributes(*m_context, float2Type);
+      } else if (kind == AR_OBJECT_BUILTIN_TRIANGLE_POSITIONS) {
+        QualType float3Ty =
+            LookupVectorType(HLSLScalarType::HLSLScalarType_float, 3);
+        recordDecl = CreateBuiltInTrianglePositions(*m_context, float3Ty);
       } else if (IsSubobjectBasicKind(kind)) {
         switch (kind) {
         case AR_OBJECT_STATE_OBJECT_CONFIG:
@@ -4861,6 +4904,7 @@ class HLSLExternalSource : public ExternalSemaSource {
     case AR_OBJECT_RAY_DESC:
     case AR_OBJECT_HIT_OBJECT:
     case AR_OBJECT_TRIANGLE_INTERSECTION_ATTRIBUTES:
+    case AR_OBJECT_BUILTIN_TRIANGLE_POSITIONS:
     case AR_OBJECT_RWTEXTURE2DMS:
     case AR_OBJECT_RWTEXTURE2DMS_ARRAY: