RFR: 8367530: The exhaustiveness errors could be improved [v8]

[Maurizio Cimadamore]

On Thu, 13 Nov 2025 18:56:40 GMT, Jan Lahoda <jlahoda at openjdk.org> wrote:

>> Consider code like:
>> 
>> package test;
>> public class Test {
>>     private int test(Root r) {
>>         return switch (r) {
>>             case Root(R2(R1 _), R2(R1 _)) -> 0;
>>             case Root(R2(R1 _), R2(R2 _)) -> 0;
>>             case Root(R2(R2 _), R2(R1 _)) -> 0;
>>         };
>>     }
>>     sealed interface Base {}
>>     record R1() implements Base {}
>>     record R2(Base b1) implements Base {}
>>     record Root(R2 b2, R2 b3) {}
>> }
>> ``` 
>> 
>> This is missing a case for `Root(R2(R2 _), R2(R2 _))`. javac will produce an error correctly, but the error is not very helpful:
>> 
>> $ javac test/Test.java
>> .../test/Test.java:4: error: the switch expression does not cover all possible input values
>>         return switch (r) {
>>                ^
>> 1 error
>> 
>> 
>> The goal of this PR is to improve the error, at least in some cases to something along these lines:
>> 
>> $ javac test/Test.java 
>> .../test/Test.java:4: error: the switch expression does not cover all possible input values
>>         return switch (r) {
>>                ^
>>   missing patterns: 
>>     test.Test.Root(test.Test.R2(test.Test.R2 _), test.Test.R2(test.Test.R2 _))
>> 1 error
>> 
>> 
>> The (very simplified) way it works in a recursive (or induction) way:
>> - start with defining the missing pattern as the binding pattern for the selector type. This would certainly exhaust the switch.
>> - for a current missing pattern, try to enhance it:
>>     - if the current type is a sealed type, try to expand to its (direct) permitted subtypes. Remove those that are not needed.
>>     - if the current (binding pattern) type is a record type, expand it to a record type, generate all possible combinations of its component types based on sealed hierarchies. Remove those that are not needed.
>> 
>> This approach relies heavily on our ability to compute exhaustiveness, which is evaluated repeatedly in the process.
>> 
>> There are some cases where the algorithm does not produce ideal results (see the tests), but overall seems much better than what we have now.
>> 
>> Another significant limitation is the speed of the process. Evaluating exhaustiveness is not a fast process, and this algorithm evaluates exhaustiveness repeatedly, potentially for many combinations of patterns (esp. for record patterns). So part of the proposal here is to have a time deadline for the computation. The default is 5s, and can be changed by `-XDexhaustivityTimeout=<timeout-in-ms>`.
>> 
>> There's also an open possibility for select tools to...
>
> Jan Lahoda has updated the pull request incrementally with one additional commit since the last revision:
> 
>   Fixing trailing whitespaces.

I think this is impressive work. In "normal" situations (e.g. switches not too big, or too nested) I can easily imagine the new diagnostics to be a life saver.

There's of course a lot of tinkering and followup work that might be possible, to improve the performance of the analysis, or to fine tune the expansion more to the shape of the code.

For now, I think I see two more general issues that stick out:
* the early flattening to string, which bypasses the diagnostic formatter -- but that should be easy to fix
* the timeout-based strategy. We don't have anything like that anywhere else in the compiler. I think it would be preferrable to have a "variable rate" of accuracy, and maybe limit how the analysis is ran, unless the user really wants to discover every detail. But I'm not sure it's always possible. Cutting on recursion might be a good way to put a ceiling on complexity. Another avenue might be to refuse to expand sealed types that have more than N ermitted subclasses.

-------------

PR Review: https://git.openjdk.org/jdk/pull/27256#pullrequestreview-3465174562