diff --git a/lib/DXIL/DxilOperations.cpp b/lib/DXIL/DxilOperations.cpp
index b786fee9fc..d0d6152da0 100644
--- a/lib/DXIL/DxilOperations.cpp
+++ b/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
diff --git a/lib/HLSL/HLOperations.cpp b/lib/HLSL/HLOperations.cpp
index 2cb3c489e8..9ee0be6eca 100644
--- a/lib/HLSL/HLOperations.cpp
+++ b/lib/HLSL/HLOperations.cpp
@@ -531,6 +531,7 @@ static AttributeSet GetHLFunctionAttributes(LLVMContext &C,
     case IntrinsicOp::IOP_GroupMemoryBarrier:
     case IntrinsicOp::IOP_AllMemoryBarrierWithGroupSync:
     case IntrinsicOp::IOP_AllMemoryBarrier:
+    case IntrinsicOp::IOP_SetMeshOutputCounts:
       addAttr(Attribute::NoDuplicate);
       break;
     }
diff --git a/tools/clang/test/CodeGenDXIL/hlsl/intrinsics/SetMeshOutputCounts-noduplicate.hlsl b/tools/clang/test/CodeGenDXIL/hlsl/intrinsics/SetMeshOutputCounts-noduplicate.hlsl
new file mode 100644
index 0000000000..cbe8a1757e
--- /dev/null
+++ b/tools/clang/test/CodeGenDXIL/hlsl/intrinsics/SetMeshOutputCounts-noduplicate.hlsl
@@ -0,0 +1,18 @@
+// RUN: %dxc /T ms_6_5 -E main -fcgl %s | FileCheck %s
+
+// Verify that SetMeshOutputCounts is emitted with the noduplicate function
+// attribute during AST->IR generation. This prevents passes such as
+// JumpThreading from cloning the call when its arguments are computed in a
+// scalar branch (see GitHub issue #8104).
+
+// CHECK: call void @"dx.hl.op.nd.void (i32, i32, i32)"(i32 68,
+// CHECK: declare void @"dx.hl.op.nd.void (i32, i32, i32)"(i32, i32, i32) [[ATTR:#[0-9]+]]
+// CHECK: attributes [[ATTR]] = { noduplicate nounwind }
+
+struct Payload { uint nv; uint np; };
+
+[outputtopology("triangle")]
+[numthreads(1, 1, 1)]
+void main(in payload Payload pl) {
+  SetMeshOutputCounts(pl.nv, pl.np);
+}
diff --git a/tools/clang/test/CodeGenDXIL/passes/SetMeshOutputCounts-noduplicate-dxilgen.ll b/tools/clang/test/CodeGenDXIL/passes/SetMeshOutputCounts-noduplicate-dxilgen.ll
new file mode 100644
index 0000000000..a179f7079b
--- /dev/null
+++ b/tools/clang/test/CodeGenDXIL/passes/SetMeshOutputCounts-noduplicate-dxilgen.ll
@@ -0,0 +1,57 @@
+; RUN: %dxopt %s -hlsl-passes-resume -dxilgen -S | FileCheck %s
+
+; Verify that DXIL Op Lowering preserves the noduplicate attribute on the
+; SetMeshOutputCounts intrinsic when generating the dx.op.setMeshOutputCounts
+; DXIL operation. See GitHub issue #8104.
+
+; CHECK: call void @dx.op.setMeshOutputCounts(i32 168,
+; CHECK: declare void @dx.op.setMeshOutputCounts(i32, i32, i32) [[ATTR:#[0-9]+]]
+; CHECK: attributes [[ATTR]] = { noduplicate nounwind }
+
+target datalayout = "e-m:e-p:32:32-i1:32-i8:32-i16:32-i32:32-i64:64-f16:32-f32:32-f64:64-n8:16:32:64"
+target triple = "dxil-ms-dx"
+
+%struct.Payload = type { i32, i32 }
+
+; Function Attrs: nounwind
+define void @main(%struct.Payload* %pl) #0 {
+  %1 = getelementptr inbounds %struct.Payload, %struct.Payload* %pl, i32 0, i32 1
+  %2 = load i32, i32* %1, align 4
+  %3 = getelementptr inbounds %struct.Payload, %struct.Payload* %pl, i32 0, i32 0
+  %4 = load i32, i32* %3, align 4
+  call void @"dx.hl.op.nd.void (i32, i32, i32)"(i32 68, i32 %4, i32 %2)
+  ret void
+}
+
+; Function Attrs: noduplicate nounwind
+declare void @"dx.hl.op.nd.void (i32, i32, i32)"(i32, i32, i32) #1
+
+attributes #0 = { nounwind }
+attributes #1 = { noduplicate nounwind }
+
+!pauseresume = !{!0}
+!dx.version = !{!1}
+!dx.valver = !{!2}
+!dx.shaderModel = !{!3}
+!dx.typeAnnotations = !{!4, !8}
+!dx.entryPoints = !{!13}
+!dx.fnprops = !{!14}
+!dx.options = !{!15, !16}
+
+!0 = !{!"hlsl-hlemit", !"hlsl-hlensure"}
+!1 = !{i32 1, i32 5}
+!2 = !{i32 1, i32 10}
+!3 = !{!"ms", i32 6, i32 5}
+!4 = !{i32 0, %struct.Payload undef, !5}
+!5 = !{i32 8, !6, !7}
+!6 = !{i32 6, !"nv", i32 3, i32 0, i32 7, i32 5}
+!7 = !{i32 6, !"np", i32 3, i32 4, i32 7, i32 5}
+!8 = !{i32 1, void (%struct.Payload*)* @main, !9}
+!9 = !{!10, !12}
+!10 = !{i32 1, !11, !11}
+!11 = !{}
+!12 = !{i32 13, !11, !11}
+!13 = !{void (%struct.Payload*)* @main, !"main", null, null, null}
+!14 = !{void (%struct.Payload*)* @main, i32 13, i32 1, i32 1, i32 1, i32 0, i32 0, i8 2, i32 8}
+!15 = !{i32 64}
+!16 = !{i32 -1}
diff --git a/tools/clang/test/CodeGenDXIL/passes/SetMeshOutputCounts-noduplicate-jumpthreading.ll b/tools/clang/test/CodeGenDXIL/passes/SetMeshOutputCounts-noduplicate-jumpthreading.ll
new file mode 100644
index 0000000000..e5defa90b2
--- /dev/null
+++ b/tools/clang/test/CodeGenDXIL/passes/SetMeshOutputCounts-noduplicate-jumpthreading.ll
@@ -0,0 +1,44 @@
+; RUN: %dxopt %s -jump-threading -S | FileCheck %s
+
+; Verify that the JumpThreading pass respects the noduplicate attribute on
+; the dx.op.setMeshOutputCounts DXIL operation. Prior to GitHub issue #8104,
+; SetMeshOutputCounts did not have noduplicate, and JumpThreading would
+; clone the call into multiple predecessor blocks when its arguments were
+; PHIs whose incoming values came from a scalar branch, causing
+; "SetMeshOutputCounts cannot be called multiple times" validation
+; failures.
+
+; The noduplicate attribute on the call site (and the declaration) must
+; prevent JumpThreading from duplicating the call.
+
+; CHECK-LABEL: define void @test
+; CHECK: call void @dx.op.setMeshOutputCounts(i32 168,
+; CHECK-NOT: call void @dx.op.setMeshOutputCounts
+
+target datalayout = "e-m:e-p:32:32-i1:32-i8:32-i16:32-i32:32-i64:64-f16:32-f32:32-f64:64-n8:16:32:64"
+target triple = "dxil-ms-dx"
+
+define void @test(i1 %c, i32 %a, i32 %b) {
+entry:
+  br i1 %c, label %then, label %merge
+
+then:
+  br label %merge
+
+merge:
+  %nv = phi i32 [ %a, %then ], [ 0, %entry ]
+  %np = phi i32 [ %b, %then ], [ 0, %entry ]
+  call void @dx.op.setMeshOutputCounts(i32 168, i32 %nv, i32 %np)
+  br i1 %c, label %t2, label %t3
+
+t2:
+  ret void
+
+t3:
+  ret void
+}
+
+; Function Attrs: noduplicate nounwind
+declare void @dx.op.setMeshOutputCounts(i32, i32, i32) #0
+
+attributes #0 = { noduplicate nounwind }
diff --git a/utils/hct/hctdb.py b/utils/hct/hctdb.py
index 634b62cc6f..1403d78719 100644
--- a/utils/hct/hctdb.py
+++ b/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"),