Deferred errors
Deferring more kinds of error, and the questions deferral leaves open
What deferred errors do, which errors are deferred, and debug and release builds are described in
Deferred errors in the tour, and how the compiler defers them in
Deferred errors in the Deep dives section. This page keeps the
wider deferral and questions still open. It is part of the compilation
workstream, and its issues carry the compilation label with type checker or wasm backend as they apply.
Plan
Wider deferral, if it is wanted:
- Type definitions. An error in one, such as a field of an unknown type, is not deferred. A later version could make the type the error type.
- Imports. An import that is not found, an import cycle or an error inside an imported module is not deferred. A later version could defer the names the import brings in.
- Syntax errors, with parser recovery (compilation): a top-level definition that does not parse. The principles say syntax errors stop a build.
Tests
When there is a test runner, it should run every test in a debug build: a test that reaches a deferred error fails
with that error’s message, and the others pass or fail on their own merits. This is what makes the feature useful
for red-green iteration. The proposed runner, nanyx test, does this; see
Testing.
Open questions
- Blame. With constraints solved together, the error is reported at whichever constraint fails, which is not always the expression the programmer got wrong. The trap is at the same place as the diagnostic, so it is consistent, but it may trap later (or earlier) than the real mistake. Is that acceptable, or does deferral want blame that follows annotations, as some bidirectional checkers do?
- Generic definitions. A generic definition with an error in its body is instantiated once per use by the
backend. Its traps are the same in each instance; confirm that monomorphisation does not need the definition’s
type to be free of the error type (
Unifier.errorType) anywhere else.