[MOD-16730] Fix SIGSEGV in FP32/L2 SSE kernel: unaligned load in residual-3 path

src/VecSim/spaces/L2/L2_SSE_FP32.h

@@ -31,23 +31,29 @@ float FP32_L2SqrSIMD16_SSE(const void *pVect1v, const void *pVect2v, size_t dime
     if constexpr (residual % 4) {
         __m128 v1, v2, diff;
         if constexpr (residual % 4 == 3) {
-            // Load 3 floats and set the last one to 0
-            v1 = _mm_loadr_ps(pVect1); // load 4 floats
-            v2 = _mm_loadr_ps(pVect2);
-            // sets the last float of v1 to the last of v2, so the diff is 0.
-            v1 = _mm_move_ss(v1, v2);
+            // Load 4 floats (unaligned). Vectors are not guaranteed to be 16-byte aligned here
+            // (block stride is dim * 4 and no alignment hint is set when dim % 4 != 0), so an
+            // aligned load (_mm_loadr_ps / movaps) would fault.
+            v1 = _mm_loadu_ps(pVect1);
+            v2 = _mm_loadu_ps(pVect2);
+            diff = _mm_sub_ps(v1, v2);
+            // Rotate the out-of-residual 4th element into the low lane and zero it - that
+            // element is processed by the next step.
+            diff = _mm_shuffle_ps(diff, diff, _MM_SHUFFLE(2, 1, 0, 3));
+            diff = _mm_move_ss(diff, _mm_setzero_ps());
         } else if constexpr (residual % 4 == 2) {
             // Load 2 floats and set the last two to 0
             v1 = _mm_loadh_pi(_mm_setzero_ps(), (__m64 *)pVect1);
             v2 = _mm_loadh_pi(_mm_setzero_ps(), (__m64 *)pVect2);
+            diff = _mm_sub_ps(v1, v2);
         } else if constexpr (residual % 4 == 1) {
             // Load 1 float and set the last three to 0
             v1 = _mm_load_ss(pVect1);
             v2 = _mm_load_ss(pVect2);
+            diff = _mm_sub_ps(v1, v2);
         }
         pVect1 += residual % 4;
         pVect2 += residual % 4;
-        diff = _mm_sub_ps(v1, v2);
         sum = _mm_mul_ps(diff, diff);
     }
 

tests/unit/test_spaces.cpp

@@ -581,6 +581,28 @@ TEST_F(SpacesTest, GetDistFuncSQ8FP16Asymmetric) {
 }
 
 #ifdef CPU_FEATURES_ARCH_X86_64
+#ifdef OPT_SSE
+// Regression test for MOD-16730: the FP32 L2 SSE kernel's residual % 4 == 3 path used
+// _mm_loadr_ps (movaps), which faults on non-16-byte-aligned addresses. Vectors are not
+// guaranteed such alignment: VecSimAllocator::allocate() returns malloc + 8 (allocation
+// header), and the dispatcher sets no alignment hint when dim % 4 != 0. This test feeds the
+// kernel buffers at that exact placement (16-aligned base + 8).
+TEST_F(SpacesTest, FP32_L2Sqr_SSE_misaligned_residual3) {
+    constexpr size_t dim = 19; // dim % 16 == 3 -> residual 3 path
+    alignas(16) static char raw1[16 + dim * sizeof(float)];
+    alignas(16) static char raw2[16 + dim * sizeof(float)];
+    float *v1 = reinterpret_cast<float *>(raw1 + 8); // address == 8 (mod 16)
+    float *v2 = reinterpret_cast<float *>(raw2 + 8);
+    for (size_t i = 0; i < dim; i++) {
+        v1[i] = float(i);
+        v2[i] = float(i) + 1.5f;
+    }
+    float baseline = FP32_L2Sqr(v1, v2, dim);
+    dist_func_t<float> arch_opt_func = spaces::Choose_FP32_L2_implementation_SSE(dim);
+    ASSERT_EQ(baseline, arch_opt_func(v1, v2, dim));
+}
+#endif // OPT_SSE
+
 TEST_F(SpacesTest, smallDimChooser) {
     // Verify that small dimensions gets the no optimization function.
     for (size_t dim = 1; dim < 8; dim++) {