compiler: remove unused struct fields

[andrewrk]

Unfortunately we cannot have a compile error for unused struct fields (#12854) since whether a field is used depends on semantic analysis (#3028). However, for structs that have automatic layout, it is legal for the compiler to remove fields that are not used. The compiler should do that automatically.

[silversquirl]

Wouldn't this mean no function that uses a struct with possibly-unused fields can be sent to code generation until all functions have been analyzed? Since a function analyzed later could use the field and thus affect the codegen of previous functions. Seems like the compile time performance implications would be pretty significant

[andrewrk]

That's a great point, and since it might have compiler design implications, it means it needs to be tackled earlier.

[mlugg]

Yeah, this isn't feasible as-is for precisely the reason silversquirl gave above. Moreover, any semantic analysis which requires resolving the type layout -- for instance, @sizeOf -- would require first resolving all possible usages of the type. So, we have an absurd situation where @sizeOf logically depends on the entire program -- including the code following itself -- being analyzed.

[mlugg]

If we go ahead with SpexGuy's idea to make @field and friends take an @Type(.EnumLiteral) rather than a string, then we could define a sufficient condition for omitting a field: if that identifier never occurs in any other type field, enum literal, or field access, in the entire ZCU. However, this condition is so narrow as to be entirely useless.

[andrewrk]

It would be possible to combine these two things:

• compile error for unused field (based on semantic analysis results)
• ability to declare any field unused (including in transitive dependencies)

in order to trade programmer effort for optimality. It would probably be a huge pain in the ass though.

[mlugg]

Hmm -- so for every field, you can mark it (fictional syntax) unused(is_field_unused) (default is unused(false)), and you get a compile error if it was used and marked unused or vice versa? Yeah, that could work, but I anticipate it being a huge PITA. And it makes the benefits of this proposal dubious -- at that point, just make the field type conditional, marking it void if it'll be unused.

[andrewrk]

My point is that the unused declaration would exist potentially even outside the module that declares the struct. The struct would be declared like normal (consider the motivating use case of #20977).