Post-monomorphization MIR optimizations

[Aaron1011]

See #44285

Currently, all MIR optimizations are run before we perform monomorphization. This causes us to miss a number of optimization opportunities:

• Learning the layout of generic arguments (e.g. if the substituted type is a ZST)
• Learning about extra trait impls (Copy, !Drop) that affect code generation (see Suboptimal invoke-related codegen for default trait impls #43150 (comment))
• As a result of the above issue, emitting unnecessary landing pads.

[hanna-kruppe]

As far as I know, all MIR optimizations today and those alluded to here can be performed just as well or better by LLVM and the main reason MIR optimizations are worthwhile nonetheless is because they can be quicker due to not redoing work for every monomorphization.

Doing changes to monomorphized MIR would not just give up that competitive advantage, it also comes with other costs. Currently there's no such thing as monomorphized MIR, codegen just walks the generic MIR and applies substitutions only as needed to generate monomorphic LLVM IR. If a specialized copy of the MIR is created for each monomorphization (which seems necessary to run passes that modify each monomorphization differently), that's extra allocations and copies that may not pay off.

[bluss]

Would it be beneficial to use monomorphized constants to eliminate dead branches and dead code? For example, in the following, skip emitting everything that's used by other_method::<T>.

fn foo<T>() {
    if some_const_boolean_property::<T>() {
        /* do nothing */
    } else {
        other_method::<T>();
    }
}

[bjorn3]

@rustbot modify labels: +A-mir-opt

[tmiasko]

This is already happening when using MIR inliner (mir-opt-level >= 2).