Generating OpVectorShuffle instead of OpCompositeConstruct

Question

According to the SPIR-V specification the opcode OpVectorShuffle seems to be versatile enough to take two input vectors (and a list of constants).

When trying to combine two vectors / extract unaligned subsequence from two vectors i.e.

#version 450
    
layout (local_size_x = 2, local_size_y = 8, local_size_z = 1) in;
layout(set = 0, binding = 0) buffer InputBuffer { vec4 input_buffer[3]; };
layout(set = 0, binding = 1) buffer OutputBuffer { vec4 result_buffer[1]; };
    
void main()
{
    vec4 a = input_buffer[0];
    vec4 b = input_buffer[1];
    vec4 w = input_buffer[2];
    result_buffer[0] = vec4(a.w, b.x, b.y, b.z) * w.xxxx;
}

I'm getting

      %4 = OpFunction %void None %3
      %5 = OpLabel
     %19 = OpAccessChain %_ptr_StorageBuffer_v4float %15 %int_0 %int_0
     %20 = OpLoad %v4float %19
     %23 = OpAccessChain %_ptr_StorageBuffer_v4float %15 %int_0 %int_1
     %24 = OpLoad %v4float %23
     %27 = OpAccessChain %_ptr_StorageBuffer_v4float %15 %int_0 %int_2
     %28 = OpLoad %v4float %27
     %36 = OpCompositeExtract %float %20 3
     %39 = OpCompositeExtract %float %24 0
     %41 = OpCompositeExtract %float %24 1
     %44 = OpCompositeExtract %float %24 2
     %45 = OpCompositeConstruct %v4float %36 %39 %41 %44
     %47 = OpVectorShuffle %v4float %28 %28 0 0 0 0
     %48 = OpFMul %v4float %45 %47
     %49 = OpAccessChain %_ptr_StorageBuffer_v4float %33 %int_0 %int_0
           OpStore %49 %48
           OpReturn
           OpFunctionEnd

IMO this seems the perfect candidate to use the

    %45 = OpVectorShuffle %v4float %20 %24 3 4 5 6

The command to compile is

>/usr/local/bin/glslc -S -O -fshader-stage=compute ./conv.glsl --target-env=vulkan1.1 -o asm.txt

Would there be a way to convince the compiler (spirv-tools v2021.4, glslang 11.1.0-316-g600c5037) to further optimize or to express the idea better?