48       v128_t va0 = wasm_v128_load(a0);  in xnn_f32_gemm_minmax_ukernel_1x8s4__wasmsimd_x86()  local
55       vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0, vb0123c0));  in xnn_f32_gemm_minmax_ukernel_1x8s4__wasmsimd_x86()
56       vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0, vb4567c0));  in xnn_f32_gemm_minmax_ukernel_1x8s4__wasmsimd_x86()
58       va0 = wasm_v32x4_shuffle(va0, va0, 1, 2, 3, 0);  in xnn_f32_gemm_minmax_ukernel_1x8s4__wasmsimd_x86()
63       vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0, vb0123c1));  in xnn_f32_gemm_minmax_ukernel_1x8s4__wasmsimd_x86()
64       vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0, vb4567c1));  in xnn_f32_gemm_minmax_ukernel_1x8s4__wasmsimd_x86()
66       va0 = wasm_v32x4_shuffle(va0, va0, 1, 2, 3, 0);  in xnn_f32_gemm_minmax_ukernel_1x8s4__wasmsimd_x86()
71       vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0, vb0123c2));  in xnn_f32_gemm_minmax_ukernel_1x8s4__wasmsimd_x86()
72       vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0, vb4567c2));  in xnn_f32_gemm_minmax_ukernel_1x8s4__wasmsimd_x86()
74       va0 = wasm_v32x4_shuffle(va0, va0, 1, 2, 3, 0);  in xnn_f32_gemm_minmax_ukernel_1x8s4__wasmsimd_x86()
79       vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0, vb0123c3));  in xnn_f32_gemm_minmax_ukernel_1x8s4__wasmsimd_x86()
80       vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0, vb4567c3));  in xnn_f32_gemm_minmax_ukernel_1x8s4__wasmsimd_x86()
88         const v128_t va0 = wasm_v32x4_load_splat(a0);  in xnn_f32_gemm_minmax_ukernel_1x8s4__wasmsimd_x86()  local
95         vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0, vb0123));  in xnn_f32_gemm_minmax_ukernel_1x8s4__wasmsimd_x86()
96         vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0, vb4567));  in xnn_f32_gemm_minmax_ukernel_1x8s4__wasmsimd_x86()