Over the years, SwiftLintFramework had become a fairly massive monolith,
containing over 400 source files with both core infrastructure and
rules.
Architecturally, the rules should rely on the core infrastructure but
not the other way around. There are two exceptions to this:
`custom_rules` and `superfluous_disable_command` which need special
integration with the linter infrastructure.
Now the time has come to formalize this architecture and one way to do
that is to move the core SwiftLint functionality out of
SwiftLintFramework and into a new SwiftLintCore module that the rules
can depend on.
Beyond enforcing architectural patterns, this also has the advantage of
speeding up incremental compilation by skipping rebuilding the core
functionality when iterating on rules.
Because the core functionality is always useful when building rules, I'm
opting to import SwiftLintCore in SwiftLintFramework as `@_exported` so
that it's implicitly available to all files in SwiftLintFramework
without needing to import it directly.
In a follow-up I'll also split the built-in rules and the extra rules
into their own modules. More modularization is possible from there, but
not planned.
The bulk of this PR just moves files from `Source/SwiftLintFramework/*`
to `Source/SwiftLintCore/*`. There are some other changes that can't be
split up into their own PRs:
* Change jazzy to document the SwiftLintCore module instead of
SwiftLintFramework.
* Change imports in unit tests to reflect where code was moved to.
* Update `sourcery` make rule to reflect where code was moved to.
* Create a new `coreRules` array and register those rules with the
registry. This allows the `custom_rules` and
`superfluous_disable_command` rule implementations to remain internal
to the SwiftLintCore module, preventing more implementation details
from leaking across architectural layers.
* Move `RuleRegistry.registerAllRulesOnce()` out of the type declaration
and up one level so it can access rules defined downstream from
SwiftLintCore.
This will allow for registering rules that aren't compiled as part of
SwiftLintFramework.
Specifically this will allow us to split the built-in and extra rules
into separate modules, leading to faster incremental compilation when
working on rules since the rest of the framework won't need to be
rebuilt on every compilation.
* Revert "Skip tests requiring runfiles support when testing with rules_xcodeproj (#4694)"
This reverts commit 040096a641.
* Enable all tests when using rules_xcodeproj
Turns out we don't need runfiles support
(https://github.com/buildbuddy-io/rules_xcodeproj/issues/828) since the
tests run outside the sandbox in Xcode, so we can get the path to these
files from the `BUILD_WORKSPACE_DIRECTORY` environment variable that
`bazel test` sets.
A long-standing limitation with SourceKit's "editor open" request is
that we weren't able to get certain tokens, such as braces, brackets and
parentheses.
This meant that this code block would be counted as two lines:
```swift
print(
"hi"
)
```
because the trailing `)` would be treated as a whitespace line.
This meant that our "body length" family of rules that measure the
effective line count of declarations like functions, types or closures
would often significantly under-count the number of content lines in a
body.
Now with SwiftSyntax, we can get all tokens, including the ones
SourceKit was previously ignoring, so we can get much more accurate line
counts when ignoring whitespace and comments.
In addition, we weren't very thorough in how we measured body length.
As an exercise, how many lines long would you say the body of this
function is?
```swift
func hello() {
print("hello")
}
```
Does the body span one line or three lines?
I propose that we consistently ignore the left and right brace lines
when calculating the body line count of these scopes so that we measure
body line counts like this:
```swift
// 1 line
{ print("foo") }
// 1 line
{
}
// 1 line
{
print("foo")
}
// 2 lines
{
let sum = 1 + 2
print(sum)
}
```
Now with those changes in place, in order to keep the default
configuration thresholds to similar levels as before, we need to adjust
them slightly. Here's what I'm suggesting:
|Rule|Before|After|
|-|-|-|
|closure_body_length|20/100|30/100|
|function_body_length|40/100|50/100|
|type_body_length|200/350|250/350|
This is a pretty significant breaking change and I suspect we'll hear
from users who are surprised that some of their declarations now exceed
the rule limits, but I believe this new approach to calculating body
lines is more correct and intuitive compared to what we've had until
now.
OSSCheck is also going to report a bazillion changes with this, which is
expected given the scope of this change.
JP Simard