Currently, custom derives inside the main juniper crate are supported by
an ugly hack using the __juniper_use_everything macro.
This commit adds new custom derive variants that are for
main juniper crate internal use only (GraphQL{Enum,InputObject}Internal.
All custom derives inside the juniper crate are refactored to use the
new '*Internal' derives.
This allows us to
* remove the use_everything macro,
* simplify the generated code for custom derives
* support the Rust 2018 edition
Introduce an abstraction for scalar values
Before this change, possible scalar values were hard coded to be representable
by one of the following types: `i32`, `f64`, `String` or `bool`. This
restricts the types of custom scalar values that can be defined. For
example, it was not possible to define a scalar value that represents an
`i64` without mapping it to a string (which would be inefficient).
One solution to fix the example above would simply be to change the
internal representation to allow it to represent an `i64`, but this would
only fix the problem for one type (until someone wants to support
`i128` for example). Also this would make juniper not follow the
GraphQL standard closely.
This commit takes another approach, by making the exact "internal"
representation of scalar values swappable (in such a way that a downstream crate could provide its own representation tailored to their needs). This allows juniper to provide a default type that only
contains the types described in the standard whereas other crates could define custom scalars for their needs.
To accomplish this we need to change several things in the current implementation:
* Add some traits that abstract the behavior of such a scalar value representation
* Change `Value` and `InputValue` to have a scalar variant (with a
generic type) instead of hard coded variants for the standard
types. This implies adding a generic parameter to both enums that
needs to be added in the whole crate.
* Change the parser to allow deciding between different types of
scalar values. The problem is basically that the original parser
implementation had no way to know whether a parsed integer number is
a `i32` or a `i64` (for example). To fix this we added some knowledge
of the existing schema to the parser.
* Fix some macros and derives to follow the new behavior.
This commit also contains an unrelated change about the way `juniper_codegen`
resolves items from `juniper`. The `_internal` flag is removed and
the resolution is replaced by a macro.
The scalar parsing strategy is as follows:
* Pass optional type information all the way down in the parser. If a
field/type/… does note exist, just do not pass down the type
information.
* The lexer now distinguishes between several fundamental scalar types (`String`, `Float`, `Int`). It does not try to actually parse those values, instead it just annotates them that this is a floating point number, an integer number, or a string value, etc.
* If type information exists while parsing a scalar value, try the following:
1. Try parsing the value using that type information.
2. If that fails try parsing the value using the inferred type information from the lexer.
* If no type information exists, try parsing the scalar value using the inferred type from the lexer,
All macros support the introduced scalar value abstraction. It is now possible to specify if a certain implementation should be based on a specific scalar value representation or be generic about the exact representation. All macros now default to the `DefaultScalarValue` type provided by
`juniper` if no scalar value representation is specified. This is done with usability and backwards compatibility in mind.
Finally, we allow specifying the scalar value representations via an attribute
(`#[graphql(scalar = "Type")]`). A default generic implementation
is provided.
Fields can now be skipped with the `#[graphql(skip)]` annotation. Note this
doesn't really make sense for GraphQLInputObjects so this isn't supported there.
Fixes https://github.com/graphql-rust/juniper/issues/220.
If a field has a description, we properly generate a meta field definition like:
```rust
{
let field = registry.field::<String>("id", &());
let field = field.description("Whatever");
field
}
```
If a field does not have a description, we were generating something like:
```rust
{
let field = registry.field::<String>("id", &());
let field = field; // <--- Note not needed
field
}
```
This of course works (and was likely optimized out by the compiler), but bloats
generated code for no benefit.
This change merely makes the second case generate:
```rust
{
let field = registry.field::<String>("id", &());
field
}
```
Fixes https://github.com/graphql-rust/juniper/issues/185.
Description attribute was available, but not actually used for derive
Object, InputObject and enum and the fields on input object.
Description is now properly set in each of those cases.
Also, tests for all three derives now properly verify the meta
information (name and description) for the derived GraphQLType.
