Case Study: Main function should return 'Unit'

Introduction

The MainFunctionReturnUnitInspection from the Kotlin IntelliJ IDEA plugin checks whether the main()-like function has the Unit return type. If the return type is implicit, it's easy to return something different by mistake. The inspection helps to avoid the issue.

In the following function, the author most likely intended to create a main() function. But in fact, it returns a PrintStream.

fun main(args: Array<String>) = System.err.apply {
    args.forEach(::println)
}

Both the K1 and K2 inspection implementations use a named function visitor. The difference lies in the processFunction() implementation.

internal class MainFunctionReturnUnitInspection : LocalInspectionTool() {
    override fun buildVisitor(holder: ProblemsHolder, isOnTheFly: Boolean): PsiElementVisitor {
        return namedFunctionVisitor { processFunction(it, holder) }
    }
}

K1 Implementation

Here is the old, K1-based implementation of processFunction().

if (function.name != "main") return
val descriptor = function.descriptor as? FunctionDescriptor ?: return
val mainFunctionDetector = MainFunctionDetector(function.languageVersionSettings) { it.resolveToDescriptorIfAny() }
if (!mainFunctionDetector.isMain(descriptor, checkReturnType = false)) return
if (descriptor.returnType?.let { KotlinBuiltIns.isUnit(it) } == true) return
holder.registerProblem(...)

Let's look at the implementation step by step.

if (function.name != "main") return

Here we check whether the function is named main to avoid running further checking logic that depends on semantic code analysis. It is crucial, as PSI-based checks are much faster than semantic ones.

val descriptor = function.descriptor as? FunctionDescriptor ?: return

Then, by using function.descriptor we get a DeclarationDescriptor, a container for semantic declaration information. In the case of a KtNamedFunction, it should be a FunctionDescriptor, holding the function's return type.

The function.descriptor helper calls resolveToDescriptorIfAny(), which itself delegates to analyze() to get the BindingContext. The obtained BindingContext provides bindings between elements in the code and their semantic representation (such as, KtDeclaration to DeclarationDescriptor).

val mainFunctionDetector = MainFunctionDetector(function.languageVersionSettings) { it.resolveToDescriptorIfAny() }
if (!mainFunctionDetector.isMain(descriptor, checkReturnType = false)) return

Additionally, we use the MainFunctionDetector compiler utility to check whether a descriptor indeed looks like a main() function. Specifically for the IDE use-cases, MainFunctionDetector provides an overload that allows us to skip the return type check. The MainFunctionDetector expects a descriptor, so we pass the one we've obtained from the KtNamedFunction.

if (descriptor.returnType?.let { KotlinBuiltIns.isUnit(it) } == true) return

Finally, we check whether the function's return type is Unit. We use the KotlinBuiltIns class from the compiler which has a number of static methods which check whether a type is a specific Kotlin built-in type.

If the return type is not Unit, the function is not a proper main() one, so we register a problem.

K2 Implementation

Then, let's compare the old implementation with the K2-based one.

val detectorConfiguration = KotlinMainFunctionDetector.Configuration(checkResultType = false)

if (!PsiOnlyKotlinMainFunctionDetector.isMain(function, detectorConfiguration)) return
if (!function.hasDeclaredReturnType() && function.hasBlockBody()) return
if (!KotlinMainFunctionDetector.getInstance().isMain(function, detectorConfiguration)) return

analyze(function) {
    if (!function.symbol.returnType.isUnitType) {
        holder.registerProblem(...)
    }
}

Starting from the beginning, one can notice that the K2 implementation is more efficient.

if (!PsiOnlyKotlinMainFunctionDetector.isMain(function, detectorConfiguration)) return

Here we not only check the function name, but we use the PsiOnlyKotlinMainFunctionDetector to check whether the function looks like a main() one using only PSI-based checks. In particular, the detector checks also the presence of type/value parameters and the function location in the file.

if (!function.hasDeclaredReturnType() && function.hasBlockBody()) return

Next, we check whether the function has a declared return type. If it doesn't, and if the function has a block body, we assume that the return type is Unit, so we skip further checks.

if (!KotlinMainFunctionDetector.getInstance().isMain(function, detectorConfiguration)) return

Then, we use another implementation of the KotlinMainFunctionDetector – this one performs semantic checks. For the K2 mode, it will be SymbolBasedKotlinMainFunctionDetector.

We use the same detectorConfiguration as before to skip the return type check, as we want to catch main()-like functions with a wrong return type.

analyze(function) {
    if (!function.symbol.returnType.isUnitType) {
        holder.registerProblem(...)
    }
}

Finally, we perform the return type check by ourselves using the Analysis API.

The Analysis API requires all work with resolved declarations to be performed inside the analyze {} block. Inside the lambda, we get access to various helpers providing semantic information about declarations visible from the use-site module. In our case, it would be the module containing the function (as we passed it to analyze()).

We need to check whether the function has a Unit return type. One can get it from a KaFunctionSymbol, a semantic representation of a function, similar to FunctionDescriptor in the old compiler.

The symbol property available inside the analysis block maps a KtDeclaration to a KaSymbol. Unlike K1, the Analysis API guarantees there will be a symbol for every declaration, so we do not need an extra ?: return. Also, symbol is overloaded for different kinds of declarations, so we can avoid explicit casting to a KaFunctionSymbol.

Just like in the old implementation, we check whether the return type is Unit, and if it's not, we register a problem.

Last modified: 11 September 2024