With a typical workflow of “just compile this source code”, this is an impossible task, because performance of algorithm implementations depends on the actual data they process (think how SQL plans queries based on table stats). Profile-guided optimization already improves performance significantly despite merely informing some heuristics. A truly optimal compiler would need to design algorithms for specific data sets (basically what people do when they profile the code).

However, there are some takes beyond the “Algol68 on PDP11” design of most languages. There’s Intel’s SPMD C compiler that extends the language to make auto-vectorization a first-class feature.

There is Halide lang specifically for image processing, which separates the goal from the algorithms used to archive it.