The generated compute shader source is much larger than original one.
This was found when running dEQP-GLES31.functional.ssbo.layout.random.nested_structs_arrays_instance_arrays.22 against virgl. On some version of mesa, it need 20+ seconds to finish. While on host, it only require 4 seconds to finish. For example: if host is on this version:
https://cgit.freedesktop.org/mesa/mesa/commit/?id=a920979d4f30a48a23f8ff375ce05fa8a947dd96
While this can be improved with a newer version of host mesa, but I still want to ask: while the original glsl only have 139 lines, the generated glsl send to host is 7k+ lines. Is this something expected? Can we improve this? origin.csvirgl.cs