diff --git a/src/gallium/auxiliary/gallivm/lp_bld_gather.c b/src/gallium/auxiliary/gallivm/lp_bld_gather.c
index ccd0376..c2e777c 100644
--- a/src/gallium/auxiliary/gallivm/lp_bld_gather.c
+++ b/src/gallium/auxiliary/gallivm/lp_bld_gather.c
@@ -419,40 +419,58 @@ lp_build_gather(struct gallivm_state *gallivm,
     * cast to 2x32f type, so the fetch is always int and on top of that
     * we avoid the vec pad and use scalar zext due the above mentioned
     * issue).
     * Note this is optimized for x86 sse2 and up backend. Could be tweaked
     * for other archs if necessary...
     */
    if (((src_width % 32) == 0) && ((src_width % dst_type.width) == 0) &&
        (dst_type.length > 1)) {
       /* use vector fetch (if dst_type is vector) */
       vec_fetch = TRUE;
       if (dst_type.floating) {
          fetch_type = lp_type_float_vec(dst_type.width, src_width);
       } else {
          fetch_type = lp_type_int_vec(dst_type.width, src_width);
       }
       /* intentionally not using lp_build_vec_type here */
       src_type = LLVMVectorType(lp_build_elem_type(gallivm, fetch_type),
                                 fetch_type.length);
       fetch_dst_type = fetch_type;
       fetch_dst_type.length = dst_type.length;
+    } else if (((src_width % 16) == 0) && (dst_type.length > 1)) {
+      vec_fetch = TRUE;
+
+      if (dst_type.floating) {
+         fetch_type = lp_type_float_vec(16, src_width);
+      } else {
+         fetch_type = lp_type_int_vec(16, src_width);
+      }
+
+      src_type = LLVMVectorType(lp_build_elem_type(gallivm, fetch_type),
+                                fetch_type.length);
+      if (dst_type.floating) {
+         fetch_dst_type = lp_type_float_vec(16, dst_type.width * dst_type.length);
+      } else {
+         fetch_dst_type = lp_type_int_vec(16, dst_type.width * dst_type.length);
+      }
+      fetch_dst_type.length = dst_type.length * 2;
+
     } else {
       /* use scalar fetch */
       vec_fetch = FALSE;
       if (dst_type.floating && ((src_width == 32) || (src_width == 64))) {
          fetch_type = lp_type_float(src_width);
       } else {
          fetch_type = lp_type_int(src_width);
       }
       src_type = lp_build_vec_type(gallivm, fetch_type);
       fetch_dst_type = fetch_type;
       fetch_dst_type.width = dst_type.width * dst_type.length;
    }
 
    if (length == 1) {
       /* Scalar */
       res = lp_build_gather_elem_vec(gallivm, length,
                                      src_width, src_type, fetch_dst_type,
                                      aligned, base_ptr, offsets, 0,
                                      vector_justify);
       return LLVMBuildBitCast(gallivm->builder, res,