Merge branch 'main' to staging-sm6.9

.github/CODEOWNERS

@@ -1 +1,2 @@
-* @microsoft/hlsl-release
+# Uncomment the next line in release branches after ask-mode begins
+# * @microsoft/hlsl-release

CMakeLists.txt

@@ -17,15 +17,6 @@ if(POLICY CMP0022)
   cmake_policy(SET CMP0022 NEW) # automatic when 2.8.12 is required
 endif()
 
-if (POLICY CMP0051)
-  # CMake 3.1 and higher include generator expressions of the form
-  # $<TARGETLIB:obj> in the SOURCES property.  These need to be
-  # stripped everywhere that access the SOURCES property, so we just
-  # defer to the OLD behavior of not including generator expressions
-  # in the output for now.
-  cmake_policy(SET CMP0051 OLD)
-endif()
-
 if(CMAKE_VERSION VERSION_LESS 3.1.20141117)
   set(cmake_3_2_USES_TERMINAL)
 else()

CONTRIBUTING.md

@@ -40,10 +40,32 @@ Before submitting a feature or substantial code contribution please discuss it w
 
 ### Coding guidelines
 
-The coding, style, and general engineering guidelines follow those described in the docs/CodingStandards.rst. For additional guidelines in code specific to HLSL, see the docs/HLSLChanges.rst file.
+The coding, style, and general engineering guidelines follow those described in the [LLVM Coding Standards](docs/CodingStandards.rst). For additional guidelines in code specific to HLSL, see the [HLSL Changes](docs/HLSLChanges.rst) docs.
 
 DXC has adopted a clang-format requirement for all incoming changes to C and C++ files. PRs to DXC should have the *changed code* clang formatted to the LLVM style, and leave the remaining portions of the file unchanged. This can be done using the `git-clang-format` tool or IDE driven workflows. A GitHub action will run on all PRs to validate that the change is properly formatted.
 
+#### Applying LLVM Standards
+
+All new code contributed to DXC should follow the LLVM coding standards.
+
+Note that the LLVM Coding Standards have a golden rule:
+
+> **If you are extending, enhancing, or bug fixing already implemented code, use the style that is already being used so that the source is uniform and easy to follow.**
+
+The golden rule should continue to be applied to places where DXC is self-consistent. A good example is DXC's common use of `PascalCase` instead of `camelCase` for APIs in some parts of the HLSL implementation. In any place where DXC is not self-consistent new code should follow the LLVM Coding Standard.
+
+A good secondary rule to follow is:
+
+> **When in doubt, follow LLVM.**
+
+Adopting LLVM's coding standards provides a consistent set of rules and guidelines to hold all contributions to. This allows patch authors to clearly understand the expectations placed on contributions, and allows reviewers to have a bar to measure contributions against. Aligning with LLVM by default ensures the path of least resistance for everyone.
+
+Since many of the LLVM Coding Standards are not enforced automatically we rely on code reviews to provide feedback and ensure contributions align with the expected coding standards. Since we rely on reviewers for enforcement and humans make mistakes, please keep in mind:
+
+> **Code review is a conversation.**
+
+It is completely reasonable for a patch author to question feedback and provide additional context about why something was done the way it was. Reviewers often see narrow slices in diffs rather than the full context of a file or part of the compiler, so they may not always provide perfect feedback. This is especially true with the application of the "golden rule" since it depends on understanding a wider context.
+
 ### Documenting Pull Requests
 
 Pull request descriptions should have the following format:

docs/SPIR-V.rst

@@ -320,6 +320,7 @@ Supported extensions
 * SPV_KHR_maximal_reconvergence
 * SPV_KHR_float_controls
 * SPV_NV_shader_subgroup_partitioned
+* SPV_KHR_quad_control
 
 Vulkan specific attributes
 --------------------------
@@ -4008,6 +4009,8 @@ Quad          ``QuadReadAcrossX()``        ``OpGroupNonUniformQuadSwap``
 Quad          ``QuadReadAcrossY()``        ``OpGroupNonUniformQuadSwap``
 Quad          ``QuadReadAcrossDiagonal()`` ``OpGroupNonUniformQuadSwap``
 Quad          ``QuadReadLaneAt()``         ``OpGroupNonUniformQuadBroadcast``
+Quad          ``QuadAny()``                ``OpGroupNonUniformQuadAnyKHR``
+Quad          ``QuadAll()``                ``OpGroupNonUniformQuadAllKHR``
 N/A           ``WaveMatch()``              ``OpGroupNonUniformPartitionNV``
 Multiprefix   ``WaveMultiPrefixSum()``     ``OpGroupNonUniform*Add``           ``PartitionedExclusiveScanNV``
 Multiprefix   ``WaveMultiPrefixProduct()`` ``OpGroupNonUniform*Mul``           ``PartitionedExclusiveScanNV``
@@ -4016,6 +4019,11 @@ Multiprefix   ``WaveMultiPrefixBitOr()``   ``OpGroupNonUniformLogicalOr``      `
 Multiprefix   ``WaveMultiPrefixBitXor()``  ``OpGroupNonUniformLogicalXor``     ``PartitionedExclusiveScanNV``
 ============= ============================ =================================== ==============================
 
+``QuadAny`` and ``QuadAll`` will use the ``OpGroupNonUniformQuadAnyKHR`` and
+``OpGroupNonUniformQuadAllKHR`` instructions if the ``SPV_KHR_quad_control``
+extension is enabled. If it is not, they will fall back to constructing the
+value using multiple calls to ``OpGroupNonUniformQuadBroadcast``.
+
 The Implicit ``vk`` Namespace
 =============================
 

include/dxc/DXIL/DxilConstants.h

@@ -487,9 +487,6 @@ inline bool IsFeedbackTexture(DXIL::ResourceKind ResourceKind) {
 // Enumeration for operations specified by DXIL
 enum class OpCode : unsigned {
   //
-  RawBufferVectorLoad = 303, // reads from a raw buffer and structured buffer
-  RawBufferVectorStore =
-      304,           // writes to a RWByteAddressBuffer or RWStructuredBuffer
   Reserved0 = 226,   // Reserved
   Reserved1 = 227,   // Reserved
   Reserved10 = 236,  // Reserved
@@ -901,8 +898,11 @@ enum class OpCode : unsigned {
   GetDimensions = 72,   // gets texture size information
   RawBufferLoad = 139,  // reads from a raw buffer and structured buffer
   RawBufferStore = 140, // writes to a RWByteAddressBuffer or RWStructuredBuffer
-  TextureLoad = 66,     // reads texel data without any filtering or sampling
-  TextureStore = 67,    // reads texel data without any filtering or sampling
+  RawBufferVectorLoad = 303, // reads from a raw buffer and structured buffer
+  RawBufferVectorStore =
+      304,           // writes to a RWByteAddressBuffer or RWStructuredBuffer
+  TextureLoad = 66,  // reads texel data without any filtering or sampling
+  TextureStore = 67, // reads texel data without any filtering or sampling
   TextureStoreSample = 225, // stores texel data at specified sample index
 
   // Sampler Feedback
@@ -1046,7 +1046,6 @@ enum class OpCode : unsigned {
   NumOpCodes_Dxil_1_6 = 222,
   NumOpCodes_Dxil_1_7 = 226,
   NumOpCodes_Dxil_1_8 = 258,
-  NumOpCodes_Dxil_1_9 = 305,
 
   NumOpCodes = 305 // exclusive last value of enumeration
 };
@@ -1060,8 +1059,6 @@ enum class OpCode : unsigned {
 // Groups for DXIL operations with equivalent function templates
 enum class OpCodeClass : unsigned {
   //
-  RawBufferVectorLoad,
-  RawBufferVectorStore,
   Reserved,
 
   // Amplification shader instructions
@@ -1284,6 +1281,8 @@ enum class OpCodeClass : unsigned {
   GetDimensions,
   RawBufferLoad,
   RawBufferStore,
+  RawBufferVectorLoad,
+  RawBufferVectorStore,
   TextureLoad,
   TextureStore,
   TextureStoreSample,
@@ -1361,7 +1360,6 @@ enum class OpCodeClass : unsigned {
   NumOpClasses_Dxil_1_6 = 149,
   NumOpClasses_Dxil_1_7 = 153,
   NumOpClasses_Dxil_1_8 = 174,
-  NumOpClasses_Dxil_1_9 = 179,
 
   NumOpClasses = 179 // exclusive last value of enumeration
 };
@@ -1422,6 +1420,12 @@ const unsigned kRawBufferLoadElementOffsetOpIdx = 3;
 const unsigned kRawBufferLoadMaskOpIdx = 4;
 const unsigned kRawBufferLoadAlignmentOpIdx = 5;
 
+// RawBufferVectorLoad.
+const unsigned kRawBufferVectorLoadHandleOpIdx = 1;
+const unsigned kRawBufferVectorLoadIndexOpIdx = 2;
+const unsigned kRawBufferVectorLoadElementOffsetOpIdx = 3;
+const unsigned kRawBufferVectorLoadAlignmentOpIdx = 4;
+
 // RawBufferStore
 const unsigned kRawBufferStoreHandleOpIdx = 1;
 const unsigned kRawBufferStoreIndexOpIdx = 2;
@@ -1433,6 +1437,13 @@ const unsigned kRawBufferStoreVal3OpIdx = 7;
 const unsigned kRawBufferStoreMaskOpIdx = 8;
 const unsigned kRawBufferStoreAlignmentOpIdx = 9;
 
+// RawBufferVectorStore
+const unsigned kRawBufferVectorStoreHandleOpIdx = 1;
+const unsigned kRawBufferVectorStoreIndexOpIdx = 2;
+const unsigned kRawBufferVectorStoreElementOffsetOpIdx = 3;
+const unsigned kRawBufferVectorStoreValOpIdx = 4;
+const unsigned kRawBufferVectorStoreAlignmentOpIdx = 5;
+
 // TextureStore.
 const unsigned kTextureStoreHandleOpIdx = 1;
 const unsigned kTextureStoreCoord0OpIdx = 2;

include/dxc/DXIL/DxilOperations.h

@@ -162,9 +162,9 @@ class OP {
 
   static bool IsDxilOpExtendedOverload(OpCode C);
 
-  // Return true if the overload name for this operation may be constructed
-  // based on a type name that may not represent the same type in different
-  // modules.
+  // Return true if the overload name suffix for this operation may be
+  // constructed based on a user-defined or user-influenced type name
+  // that may not represent the same type in different linked modules.
   static bool MayHaveNonCanonicalOverload(OpCode OC);
 
 private:

include/dxc/DxilPIXPasses/DxilPIXPasses.h

@@ -27,6 +27,7 @@ ModulePass *createDxilDebugInstrumentationPass();
 ModulePass *createDxilShaderAccessTrackingPass();
 ModulePass *createDxilPIXAddTidToAmplificationShaderPayloadPass();
 ModulePass *createDxilPIXDXRInvocationsLogPass();
+ModulePass *createDxilNonUniformResourceIndexInstrumentationPass();
 
 void initializeDxilAddPixelHitInstrumentationPass(llvm::PassRegistry &);
 void initializeDxilDbgValueToDbgDeclarePass(llvm::PassRegistry &);
@@ -41,5 +42,7 @@ void initializeDxilShaderAccessTrackingPass(llvm::PassRegistry &);
 void initializeDxilPIXAddTidToAmplificationShaderPayloadPass(
     llvm::PassRegistry &);
 void initializeDxilPIXDXRInvocationsLogPass(llvm::PassRegistry &);
+void initializeDxilNonUniformResourceIndexInstrumentationPass(
+    llvm::PassRegistry &);
 
 } // namespace llvm

include/dxc/HlslIntrinsicOp.h

@@ -231,6 +231,9 @@ enum class IntrinsicOp {
   IOP_VkReadClock = 223,
   IOP_Vkext_execution_mode = 224,
   IOP_Vkext_execution_mode_id = 225,
+  IOP_Vkreinterpret_pointer_cast = 360,
+  IOP_Vkstatic_pointer_cast = 361,
+  MOP_GetBufferContents = 362,
   MOP_Append = 226,
   MOP_RestartStrip = 227,
   MOP_CalculateLevelOfDetail = 228,
@@ -366,7 +369,7 @@ enum class IntrinsicOp {
   IOP_usign = 355,
   MOP_InterlockedUMax = 356,
   MOP_InterlockedUMin = 357,
-  Num_Intrinsics = 360,
+  Num_Intrinsics = 363,
 };
 inline bool HasUnsignedIntrinsicOpcode(IntrinsicOp opcode) {
   switch (opcode) {

include/dxc/WinAdapter.h

@@ -51,7 +51,8 @@
 #define _countof(a) (sizeof(a) / sizeof(*(a)))
 
 // If it is GCC, there is no UUID support and we must emulate it.
-#ifndef __clang__
+// Clang support depends on the -fms-extensions compiler flag.
+#if !defined(__clang__) || !defined(_MSC_EXTENSIONS)
 #define __EMULATE_UUID 1
 #endif // __clang__
 

include/dxc/dxcapi.internal.h

@@ -7,6 +7,9 @@
 //                                                                           //
 // Provides non-public declarations for the DirectX Compiler component.      //
 //                                                                           //
+// Modifications Copyright(C) 2025 Advanced Micro Devices, Inc.              //
+// All rights reserved.                                                      //
+//                                                                           //
 ///////////////////////////////////////////////////////////////////////////////
 
 #ifndef __DXC_API_INTERNAL__
@@ -35,6 +38,7 @@ typedef struct ID3D10Blob ID3D10Blob;
 static const BYTE INTRIN_TEMPLATE_FROM_TYPE = 0xff;
 static const BYTE INTRIN_TEMPLATE_VARARGS = 0xfe;
 static const BYTE INTRIN_TEMPLATE_FROM_FUNCTION = 0xfd;
+static const BYTE INTRIN_TEMPLATE_FROM_FUNCTION_2 = 0xfc;
 
 // Use this enumeration to describe allowed templates (layouts) in intrinsics.
 enum LEGAL_INTRINSIC_TEMPLATES {
@@ -128,7 +132,12 @@ enum LEGAL_INTRINSIC_COMPTYPES {
 
   LICOMPTYPE_HIT_OBJECT = 51,
 
+#ifdef ENABLE_SPIRV_CODEGEN
+  LICOMPTYPE_VK_BUFFER_POINTER = 52,
+  LICOMPTYPE_COUNT = 53
+#else
   LICOMPTYPE_COUNT = 52
+#endif
 };
 
 static const BYTE IA_SPECIAL_BASE = 0xf0;

lib/DXIL/DxilOperations.cpp

@@ -2633,6 +2633,8 @@ const OP::OpCodeProperty OP::m_OpCodeProps[(unsigned)OP::OpCode::NumOpCodes] = {
      0,
      {},
      {}}, // Overloads: v
+
+    // Resources
     {OC::RawBufferVectorLoad,
      "RawBufferVectorLoad",
      OCC::RawBufferVectorLoad,
@@ -2816,8 +2818,7 @@ bool OP::IsOverloadLegal(OpCode opCode, Type *pType) {
   if (OpProps.NumOverloadDims > 1) {
     StructType *ST = dyn_cast<StructType>(pType);
     // Make sure multi-overload is well-formed.
-    if (!ST || ST->hasName() ||
-        ST->getNumContainedTypes() != OpProps.NumOverloadDims)
+    if (!ST || ST->hasName() || ST->getNumElements() != OpProps.NumOverloadDims)
       return false;
     for (unsigned I = 0; I < ST->getNumElements(); ++I)
       Types[I] = ST->getElementType(I);
@@ -5757,6 +5758,8 @@ Function *OP::GetOpFunc(OpCode opCode, Type *pOverloadType) {
     A(pV);
     A(pI32);
     break;
+
+    // Resources
   case OpCode::RawBufferVectorLoad:
     RRT(pETy);
     A(pI32);
@@ -6224,11 +6227,12 @@ bool OP::IsResRetType(llvm::Type *Ty) {
     if (Ty == ResTy)
       return true;
   }
+  // Check for vector overload which isn't cached in m_pResRetType.
   StructType *ST = cast<StructType>(Ty);
-  if (!ST->hasName() || ST->getNumContainedTypes() < 2)
+  if (!ST->hasName() || ST->getNumElements() < 2 ||
+      !ST->getElementType(0)->isVectorTy())
     return false;
-  return Ty == GetResRetType(ST->getContainedType(0));
-  return false;
+  return Ty == GetResRetType(ST->getElementType(0));
 }
 
 Type *OP::GetResRetType(Type *pOverloadType) {

lib/DxilPIXPasses/CMakeLists.txt

@@ -20,6 +20,7 @@ add_llvm_library(LLVMDxilPIXPasses
   PixPassHelpers.cpp
   DxilPIXAddTidToAmplificationShaderPayload.cpp
   DxilPIXDXRInvocationsLog.cpp
+  DxilNonUniformResourceIndexInstrumentation.cpp
 
   ADDITIONAL_HEADER_DIRS
   ${LLVM_MAIN_INCLUDE_DIR}/llvm/IR