fix: SplitDouble producing transposed output in DXIL

tools/clang/test/HLSLFileCheck/hlsl/intrinsics/cast/asuint_matrix.hlsl

@@ -0,0 +1,49 @@
+// RUN: %dxc -E main -T cs_6_0 %s | FileCheck %s
+
+// Regression test for github.com/microsoft/DirectXShaderCompiler/issues/8477:
+// asuint(double_matrix, out uint_mat lo, out uint_mat hi) was producing
+// transposed output in DXIL because TranslateDoubleAsUint wrote elements in
+// row-major order while column-major output allocas expect column-major layout.
+//
+// For a 2x2 double matrix with distinct elements d00, d01, d10, d11 (in
+// logical row-major order), the lo output must satisfy:
+//   lo[0][0] = lo_bits(d00)
+//   lo[0][1] = lo_bits(d01)
+//   lo[1][0] = lo_bits(d10)
+//   lo[1][1] = lo_bits(d11)
+//
+// Before the fix, lo[0][0] and lo[0][1] would be lo_bits(d00) and lo_bits(d10)
+// (first column) instead of the first row.  We verify that four SplitDouble
+// calls are emitted (one per matrix element) and that the compiler produces
+// valid DXIL without assertion failures.
+
+// CHECK: SplitDouble
+// CHECK: SplitDouble
+// CHECK: SplitDouble
+// CHECK: SplitDouble
+
+RWStructuredBuffer<uint> Lows : register(u0);
+RWStructuredBuffer<uint> Highs : register(u1);
+
+[numthreads(1,1,1)]
+void main()
+{
+    double2x2 m;
+    m[0][0] = asdouble(10u, 0u);
+    m[0][1] = asdouble(20u, 0u);
+    m[1][0] = asdouble(30u, 0u);
+    m[1][1] = asdouble(40u, 0u);
+
+    uint2x2 lo, hi;
+    asuint(m, lo, hi);
+
+    Lows[0] = lo[0][0];
+    Lows[1] = lo[0][1];
+    Lows[2] = lo[1][0];
+    Lows[3] = lo[1][1];
+
+    Highs[0] = hi[0][0];
+    Highs[1] = hi[0][1];
+    Highs[2] = hi[1][0];
+    Highs[3] = hi[1][1];
+}