I think a production grade compiler not only can, but must, leave performance on the table when the cost is correctness (unless the performance gain is incredibly high and the correctness loss is minimal). Correctness is not all important, but it is the most important thing. Unfortunately, compiler writers do not agree and they do silly things like "let's assume UB cannot ever happen and optimize based on that".

I do not agree in the general case. There are very useful DSL compilers which do not consider performance at all, but just compile to a target which does the optimization for them (JVM, LLVM IR or even just C)