Add noduplicate attribute to SetMeshOutputCounts and DispatchMesh intrinsics

Agent-Logs-Url: https://github.com/microsoft/DirectXShaderCompiler/sessions/0efbd2d6-46d4-4a20-89eb-cb685105780b

Co-authored-by: chrisbieneman <[email protected]>

Author: Copilot

lib/DXIL/DxilOperations.cpp

@@ -1510,7 +1510,7 @@ static const OP::OpCodeProperty CoreOps_OpCodeProps[] = {
      "SetMeshOutputCounts",
      OCC::SetMeshOutputCounts,
      "setMeshOutputCounts",
-     Attribute::None,
+     Attribute::NoDuplicate,
      0,
      {},
      {}}, // Overloads: v
@@ -1552,7 +1552,7 @@ static const OP::OpCodeProperty CoreOps_OpCodeProps[] = {
      "DispatchMesh",
      OCC::DispatchMesh,
      "dispatchMesh",
-     Attribute::None,
+     Attribute::NoDuplicate,
      1,
      {{0x100}},
      {{0x0}}}, // Overloads: u

lib/HLSL/HLOperations.cpp

@@ -531,6 +531,8 @@ static AttributeSet GetHLFunctionAttributes(LLVMContext &C,
     case IntrinsicOp::IOP_GroupMemoryBarrier:
     case IntrinsicOp::IOP_AllMemoryBarrierWithGroupSync:
     case IntrinsicOp::IOP_AllMemoryBarrier:
+    case IntrinsicOp::IOP_SetMeshOutputCounts:
+    case IntrinsicOp::IOP_DispatchMesh:
       addAttr(Attribute::NoDuplicate);
       break;
     }

tools/clang/test/CodeGenHLSL/mesh-val/sinkSetMeshOutputCounts.hlsl

@@ -0,0 +1,50 @@
+// RUN: %dxc -E main -T ms_6_5 %s | FileCheck %s
+
+// Test that SetMeshOutputCounts has noduplicate attribute, preventing the
+// optimizer from sinking/duplicating the call into branches that compute
+// the count values.
+
+// CHECK: dx.op.setMeshOutputCounts
+// CHECK: noduplicate
+
+#define MAX_VERT 32
+#define MAX_PRIM 16
+#define NUM_THREADS 32
+
+struct MeshPerVertex {
+    float4 position : SV_Position;
+};
+
+struct MeshPayload {
+    float normal;
+};
+
+[numthreads(NUM_THREADS, 1, 1)]
+[outputtopology("triangle")]
+void main(
+            out indices uint3 primIndices[MAX_PRIM],
+            out vertices MeshPerVertex verts[MAX_VERT],
+            in payload MeshPayload mpl,
+            in uint tig : SV_GroupIndex,
+            in uint vid : SV_ViewID
+         )
+{
+    // Compute counts in a branch - optimizer used to sink SetMeshOutputCounts
+    // into each branch, producing two copies and failing validation.
+    uint nverts, nprims;
+    if (vid % 2) {
+        nverts = MAX_VERT;
+        nprims = MAX_PRIM;
+    } else {
+        nverts = MAX_VERT / 2;
+        nprims = MAX_PRIM / 2;
+    }
+    SetMeshOutputCounts(nverts, nprims);
+
+    if (tig < nverts) {
+        verts[tig].position = float4(0, 0, 0, 1);
+    }
+    if (tig < nprims) {
+        primIndices[tig] = uint3(tig * 3, tig * 3 + 1, tig * 3 + 2);
+    }
+}

utils/hct/hctdb.py

@@ -4156,7 +4156,7 @@ def UFI(name, **mappings):
             "SetMeshOutputCounts",
             "Mesh shader intrinsic SetMeshOutputCounts",
             "v",
-            "",
+            "nd",
             [
                 retvoid_param,
                 db_dxil_param(2, "i32", "numVertices", "number of output vertices"),
@@ -4234,7 +4234,7 @@ def UFI(name, **mappings):
             "DispatchMesh",
             "Amplification shader intrinsic DispatchMesh",
             "u",
-            "",
+            "nd",
             [
                 retvoid_param,
                 db_dxil_param(2, "i32", "threadGroupCountX", "thread group count x"),