85         v128_t va0 = wasm_v128_load(a0);  in xnn_f32_igemm_minmax_ukernel_3x8s4__wasmsimd_arm()  local
96         vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0, vb0123c0));  in xnn_f32_igemm_minmax_ukernel_3x8s4__wasmsimd_arm()
99         vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0, vb4567c0));  in xnn_f32_igemm_minmax_ukernel_3x8s4__wasmsimd_arm()
103         va0 = wasm_v32x4_shuffle(va0, va0, 1, 2, 3, 0);  in xnn_f32_igemm_minmax_ukernel_3x8s4__wasmsimd_arm()
110         vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0, vb0123c1));  in xnn_f32_igemm_minmax_ukernel_3x8s4__wasmsimd_arm()
113         vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0, vb4567c1));  in xnn_f32_igemm_minmax_ukernel_3x8s4__wasmsimd_arm()
117         va0 = wasm_v32x4_shuffle(va0, va0, 1, 2, 3, 0);  in xnn_f32_igemm_minmax_ukernel_3x8s4__wasmsimd_arm()
124         vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0, vb0123c2));  in xnn_f32_igemm_minmax_ukernel_3x8s4__wasmsimd_arm()
127         vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0, vb4567c2));  in xnn_f32_igemm_minmax_ukernel_3x8s4__wasmsimd_arm()
131         va0 = wasm_v32x4_shuffle(va0, va0, 1, 2, 3, 0);  in xnn_f32_igemm_minmax_ukernel_3x8s4__wasmsimd_arm()
138         vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0, vb0123c3));  in xnn_f32_igemm_minmax_ukernel_3x8s4__wasmsimd_arm()
141         vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0, vb4567c3));  in xnn_f32_igemm_minmax_ukernel_3x8s4__wasmsimd_arm()
155           const v128_t va0 = wasm_v32x4_load_splat(a0);  in xnn_f32_igemm_minmax_ukernel_3x8s4__wasmsimd_arm()  local
162           vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0, vb0123));  in xnn_f32_igemm_minmax_ukernel_3x8s4__wasmsimd_arm()
163           vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0, vb4567));  in xnn_f32_igemm_minmax_ukernel_3x8s4__wasmsimd_arm()