It’s too early for this project to go after any sort of compile-time performance improvements. We are not making any decisions that would be degrading performance significantly. As for the rest, it will heavily depend on what specifically you want the compiler to do. At the moment, basic compilation performance is comparable to that of clang because the flow still spends most of the time in either clang frontend or llvm codegen. When we start doing more in MLIR, I expect this to remain more or less the case: potential overhead from doing more IR translations may be compensated by MLIR’s parallelism.
Polyhedral optimization is a whole different story. It should remain opt-in and allow for some trade-offs between the aggressiveness of the optimization and the compile time required.