What's a kernel-style transformation? That term is unfamiliar to me? Are you referring to e.g.: optimizing stencil kernels?
You're correct that classical semantics-preserving compiler optimizations are very restricted, when not user guided as in the OP's tool. They do not have any form of specification weaker than the input code. This is why most synthesis tools do take something higher level than code.
Then again, some really do have the workflow "figure out existing code's algorithm, and pick a better one." The first thing in this space that comes to mind is verified lifting, being pursued by my collaborators at UW.
You're correct that classical semantics-preserving compiler optimizations are very restricted, when not user guided as in the OP's tool. They do not have any form of specification weaker than the input code. This is why most synthesis tools do take something higher level than code.
Then again, some really do have the workflow "figure out existing code's algorithm, and pick a better one." The first thing in this space that comes to mind is verified lifting, being pursued by my collaborators at UW.