18     uint8_t* output)  in xnn_x8_zip_xm_ukernel__neon()  argument
24   uint8_t* last_output = (uint8_t*) ((uintptr_t) output + (m - 4));  in xnn_x8_zip_xm_ukernel__neon()
47 …      vst1_lane_u32(__builtin_assume_aligned(output, 1), vreinterpret_u32_u16(vxyzw_lo.val[0]), 0);  in xnn_x8_zip_xm_ukernel__neon()
48         output = (uint8_t*) ((uintptr_t) output + m);  in xnn_x8_zip_xm_ukernel__neon()
50 …      vst1_lane_u32(__builtin_assume_aligned(output, 1), vreinterpret_u32_u16(vxyzw_lo.val[0]), 1);  in xnn_x8_zip_xm_ukernel__neon()
51         output = (uint8_t*) ((uintptr_t) output + m);  in xnn_x8_zip_xm_ukernel__neon()
53 …      vst1_lane_u32(__builtin_assume_aligned(output, 1), vreinterpret_u32_u16(vxyzw_lo.val[1]), 0);  in xnn_x8_zip_xm_ukernel__neon()
54         output = (uint8_t*) ((uintptr_t) output + m);  in xnn_x8_zip_xm_ukernel__neon()
56 …      vst1_lane_u32(__builtin_assume_aligned(output, 1), vreinterpret_u32_u16(vxyzw_lo.val[1]), 1);  in xnn_x8_zip_xm_ukernel__neon()
57         output = (uint8_t*) ((uintptr_t) output + m);  in xnn_x8_zip_xm_ukernel__neon()
59 …      vst1_lane_u32(__builtin_assume_aligned(output, 1), vreinterpret_u32_u16(vxyzw_hi.val[0]), 0);  in xnn_x8_zip_xm_ukernel__neon()
60         output = (uint8_t*) ((uintptr_t) output + m);  in xnn_x8_zip_xm_ukernel__neon()
62 …      vst1_lane_u32(__builtin_assume_aligned(output, 1), vreinterpret_u32_u16(vxyzw_hi.val[0]), 1);  in xnn_x8_zip_xm_ukernel__neon()
63         output = (uint8_t*) ((uintptr_t) output + m);  in xnn_x8_zip_xm_ukernel__neon()
65 …      vst1_lane_u32(__builtin_assume_aligned(output, 1), vreinterpret_u32_u16(vxyzw_hi.val[1]), 0);  in xnn_x8_zip_xm_ukernel__neon()
66         output = (uint8_t*) ((uintptr_t) output + m);  in xnn_x8_zip_xm_ukernel__neon()
68 …      vst1_lane_u32(__builtin_assume_aligned(output, 1), vreinterpret_u32_u16(vxyzw_hi.val[1]), 1);  in xnn_x8_zip_xm_ukernel__neon()
69         output = (uint8_t*) ((uintptr_t) output + m);  in xnn_x8_zip_xm_ukernel__neon()
96           vst1_lane_u32(__builtin_assume_aligned(output, 1), vxyzw0, 0);  in xnn_x8_zip_xm_ukernel__neon()
97           output = (uint8_t*) ((uintptr_t) output + m);  in xnn_x8_zip_xm_ukernel__neon()
99           vst1_lane_u32(__builtin_assume_aligned(output, 1), vxyzw0, 1);  in xnn_x8_zip_xm_ukernel__neon()
100           output = (uint8_t*) ((uintptr_t) output + m);  in xnn_x8_zip_xm_ukernel__neon()
102           vst1_lane_u32(__builtin_assume_aligned(output, 1), vxyzw1, 0);  in xnn_x8_zip_xm_ukernel__neon()
103           output = (uint8_t*) ((uintptr_t) output + m);  in xnn_x8_zip_xm_ukernel__neon()
105           vst1_lane_u32(__builtin_assume_aligned(output, 1), vxyzw1, 1);  in xnn_x8_zip_xm_ukernel__neon()
106           output = (uint8_t*) ((uintptr_t) output + m);  in xnn_x8_zip_xm_ukernel__neon()
113           vst1_lane_u32(__builtin_assume_aligned(output, 1), vxyzw0, 0);  in xnn_x8_zip_xm_ukernel__neon()
114           output = (uint8_t*) ((uintptr_t) output + m);  in xnn_x8_zip_xm_ukernel__neon()
116           vst1_lane_u32(__builtin_assume_aligned(output, 1), vxyzw0, 1);  in xnn_x8_zip_xm_ukernel__neon()
117           output = (uint8_t*) ((uintptr_t) output + m);  in xnn_x8_zip_xm_ukernel__neon()
122           vst1_lane_u32(__builtin_assume_aligned(output, 1), vxyzw0, 0);  in xnn_x8_zip_xm_ukernel__neon()
123           output = (uint8_t*) ((uintptr_t) output + m);  in xnn_x8_zip_xm_ukernel__neon()
126       output = (uint8_t*) ((uintptr_t) output + output_increment);  in xnn_x8_zip_xm_ukernel__neon()
127       if (output > last_output) {  in xnn_x8_zip_xm_ukernel__neon()
128         output = last_output;  in xnn_x8_zip_xm_ukernel__neon()
133     uint8_t* o = output;  in xnn_x8_zip_xm_ukernel__neon()