61         v128_t va0 = wasm_v128_load(a0);  in xnn_f32_igemm_minmax_ukernel_1x8s4__wasmsimd_x86()  local
68         vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0, vb0123c0));  in xnn_f32_igemm_minmax_ukernel_1x8s4__wasmsimd_x86()
69         vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0, vb4567c0));  in xnn_f32_igemm_minmax_ukernel_1x8s4__wasmsimd_x86()
71         va0 = wasm_v32x4_shuffle(va0, va0, 1, 2, 3, 0);  in xnn_f32_igemm_minmax_ukernel_1x8s4__wasmsimd_x86()
76         vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0, vb0123c1));  in xnn_f32_igemm_minmax_ukernel_1x8s4__wasmsimd_x86()
77         vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0, vb4567c1));  in xnn_f32_igemm_minmax_ukernel_1x8s4__wasmsimd_x86()
79         va0 = wasm_v32x4_shuffle(va0, va0, 1, 2, 3, 0);  in xnn_f32_igemm_minmax_ukernel_1x8s4__wasmsimd_x86()
84         vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0, vb0123c2));  in xnn_f32_igemm_minmax_ukernel_1x8s4__wasmsimd_x86()
85         vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0, vb4567c2));  in xnn_f32_igemm_minmax_ukernel_1x8s4__wasmsimd_x86()
87         va0 = wasm_v32x4_shuffle(va0, va0, 1, 2, 3, 0);  in xnn_f32_igemm_minmax_ukernel_1x8s4__wasmsimd_x86()
92         vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0, vb0123c3));  in xnn_f32_igemm_minmax_ukernel_1x8s4__wasmsimd_x86()
93         vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0, vb4567c3));  in xnn_f32_igemm_minmax_ukernel_1x8s4__wasmsimd_x86()
105           const v128_t va0 = wasm_v32x4_load_splat(a0);  in xnn_f32_igemm_minmax_ukernel_1x8s4__wasmsimd_x86()  local
108           vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0, vb0123));  in xnn_f32_igemm_minmax_ukernel_1x8s4__wasmsimd_x86()
109           vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0, vb4567));  in xnn_f32_igemm_minmax_ukernel_1x8s4__wasmsimd_x86()