Chromium Code Reviews| Index: docs/language/informal/assert-in-initializer-list.md |
| diff --git a/docs/language/informal/assert-in-initializer-list.md b/docs/language/informal/assert-in-initializer-list.md |
| new file mode 100644 |
| index 0000000000000000000000000000000000000000..d323be10de42a98c0ff997f9dc599cb9c384f1a0 |
| --- /dev/null |
| +++ b/docs/language/informal/assert-in-initializer-list.md |
| @@ -0,0 +1,92 @@ |
| +# Asserts in Initializer List |
| +[lrn@google.com](mailto:lrn@google.com) |
| +Version 1.1 |
| +Status: Accepted, Informally specified |
| + |
| +(See: http://dartbug.com/24841, http://dartbug.com/27141) |
| + |
| +In some cases, you want to validate your inputs before creating an instance, even in a const constructor. To allow that, we have tested the possibility of allowing assert statements in the initializer list of a generative constructor. |
| + |
| +We started by implementing the feature in the VM behind a flag, with at syntax support from the analyzer and the formatter. |
| + |
| +This was as successful experiment, and the feature is actively being used by the Flutter project, so now we promote the experimental feature to a language feature. |
| + |
| +The feature should now be implemented by all Dart tools and be formally specified in the Dart language specification. |
|
floitsch
2017/06/08 11:44:59
Remove paragraph.
|
| + |
| + |
| +## Syntax |
| + |
| +The syntax will be changed to allow an assert statement without trailing semicolon (just the `assert(condition[, message])`) to appear as an item in the initializer list. |
|
floitsch
2017/06/08 11:44:59
will be -> is
|
| +Example: |
| + |
| +```dart |
| + C(x, y) : this.foo = x, assert(x < y), this.bar = y; |
| +``` |
| + |
| +The assert can occur anywhere in the list where an initializing assignment can. |
| + |
| +That is, the grammar changes so that *superCallOrFieldIntitializer* can also produce an assert. |
| + |
| +For simplicity, we add a new production for the assert-without-the-semicolon, and reuse that in both the initializer list and the *assertStatement*. |
| + |
| +> *superCallOrFieldInitializer*: |
| +> **super** arguments |
| +> | **super** ‘.’ identifier arguments |
| +> | fieldInitializer |
| +> | assertion |
| +> ; |
| +> |
| +> assertion: **assert** ‘(' expression (‘,' expression)? ‘)' ; |
| +> |
| +> assertStatement: assertion ‘;' ; |
| + |
| +The *superCallOrFieldInitializer* production will probably change name too, perhaps to *initializerListEntry*, but that's not important for the behavior. |
| + |
| +## Semantics |
| + |
| +The initializer list assert works the same way as an assert statement in a function body (with special treatment for asserts in a const constructor's initializer list, see next section). The assert expressions are evaluated in the initializer list scope, which does not have access to `this`, exactly the same way that an assert statement would be evaluated in the same scope. The runtime behavior is effectively: |
| + |
| +1. evaluate the condition expression (in the initializer list scope) to a result, `o`. |
| +1. If `o` implements `Function`, call it with zero arguments and let `r` be the return value, |
| +1. otherwise let `r` = `o`. |
| +1. Perform boolean conversion on `r`. This throws if `r` is not an instance of `bool`. |
| +1. if `r` isn't `true`, |
| + a. if there is a message expression, evaluate that to a value `m` |
| + b. otherwise let `m` be `null` |
| + c. then throw an `AssertionError` with `m` as message. |
| + |
| +Statically, like in an assertion statement, it's a warning if the static type of the condition expression isn't assignable to either `bool` or `bool Function()`. |
| + |
| +Here step 2, 4 and 5a may throw before reaching step 5c, in which case that is the effect of the assert. |
| + |
| + |
| +The assert statement is evaluated at its position in the initializer list, relative to the left-to-right evaluation of initializer list entries. |
| + |
| +As usual, assert statements have no effect unless asserts are enabled (e.g., by running in checked mode). |
| + |
| + |
| +## Const Semantics |
| + |
| +If the constructor is a const constructor, the condition and message expressions in the assert must be potentially compile-time constant expressions. If any of them aren't, it is a compile-time error, the same way a non-potentially compile-time constant initializer expression in the initializer list is. |
| + |
| +Further, the condition expression should not evaluate to a function, since we can't call functions at compile time. We can't prevent it from evaluating to a function, but the function cannot not be called. To account for this, the behavior above is changed for const constructor initializer list asserts: |
| + |
| +*Step 2 above is dropped for an assert in a const constructor initializer list.* |
| + |
| +The change is entirely syntax driven - an assert inside a const constructor initializer list does not test whether the expression is a function, not even when the constructor is invoked using `new`. |
| +This change from the current specification is needed because asserts previously couldn't occur in a (potentially) const context[^1]. |
| + |
| +During a const constructor invocation (that is, when the const constructor is invoked using the `const` prefix), if the assert fails, either due to boolean conversion when `r` is not a boolean value or due to assertion failure when `r` is `false`, it is treated like any other compile-time throw in a compile-time constant expression, and it causes a compile-time error. |
| + |
| + |
| +## Revisions |
| + |
| +1.0 Initial specification. |
| + |
| +1.1 Handle second expression in asserts as well, add grammar rules. |
| + |
| + |
| +## Notes |
| + |
| +[^1]: |
| + If we ever add "const functions" which can be "called" in a const context, then we may allow them here, but other functions are still compile time errors. |