VersionConflictResolutionIntegrationTest.groovy

Introduce check on replaced modules before short-circuit selection

When we perform short-circuit selection for a module (where its selected version is `null`), we need to check that

this module actually doesn't participate in module replacements. If it does, we need to fall back on classic conflict

resolution. Otherwise, the following situation might occur:

- `a` is replaced by `b`

- we add `a:1` to the graph. No version is selected for `a`, so we select `a:1`

- we add `b:1` to the graph. No version is selected for `b`, so we select `b:1`

- we end up with both `a` and `b` in the result

Instead, we now follow this process:

- `a` is replaced by `b`

- we add `a:1` to the graph. No version is selected for `a`, but `a` participates in replacements. We don't short-circuit,

but in the end we select `a:1`

- we add `b:1` to the graph. No version is selected for `b`, but `b` participates in replacements. We fall back to conflict

resolution, which sees that `a` needs to be replaced by `b`.

- only `b:1` ends up in the result

Improve version range selection

This commit fixes a couple of bugs in range selection. It refactors the code, making it clearer what the different states

of a module can be. In particular, the previous version included both `Conflict` and `Evicted`, but in practice there was

no difference in using one or the other!

Now the selection uses four different states (`Selected`, `Selectable`, `Evicted` and `Orphan`) which allows us to

discriminate between cases a node was evicted in conflict resolution, or when it was simplify preferred over another

possible choice.

Fixup tests

Add test case showing that conflicts on different modules are handled separately

Introduce a way to restart selection of a module

This commit allows restarting selection of a module from a conflict resolution rule. The idea is to be able to restart

selection when "new information" is available. One use case is the range selection of a module. If, during graph

visit, a different range is seen, then we know we can potentially have disjoint ranges, in which case classic

resolution occurs, or intersecting ranges, in which case we need to select the highest version in range. This commit

prepares a different solution to the problem by allowing us to restart selection once we determine we have two

intersecting ranges.

It's worth noting that this commit requires temporary disabling of the tests, because the rule doesn't enforce this

behavior: it will now create an infinite loop, without ever selecting a version. The reason is precisely that we

don't make this "new information" available to selection. This will be implemented in a subsequent commit.

Add more integration tests to highlight the behavior with `failOnVersionConflict`

Add test coverage for version range conflict resolution

Simplify `DynamicVersionResolver`

This commit merges `RepositoryResolveState` with `ComponentSelectionContext`. This makes the code easier to read, by avoiding

incomprehensible back references. Unit tests are also clearer since they now clearly show which versions were tested, rejected

and not found.

Changed the way that the target configuration for a project dependency is calculated, to always select only from the configurations of the target project that have attributes attached, regardless of whether the consumer has defined any attributes or not. Fall back to `default` only when the target project has no configurations with attributes defined.

This introduces a breaking change when consuming the output of a project with the Java plugin applied from a project that does not have the Java plugin applied. Previously, the `default` configuration would be selected, now the `runtimeElements` configuration is selected. However, this is consistent with the case where the consuming project is also using the Java plugin or is using the Android plugin. It also means that custom configurations in these projects will select the same thing as the runtime classpath configuration.

Changed traversal of dependency graph results so that the `ArtifactSet` instances included in the result are visited in consumer first order, rather than the nodes, so that the artifact sets can be collected directly into a list in the correct order rather than collecting and retaining an additional map from node -> artifact set to apply the ordering.

Reworked dependency result artifact/file ordering to address some issues and to make 'consumer first' the default ordering.

Specifically, nodes in the graph are now sorted by component, rather than sorting by node. When sorting by node for a component with multiple nodes in the graph (eg a direct dependency on a maven module from a Gradle project plus a transitive dependency via another maven module, but also possible in other ways) it was possible for a particular component to appear _before_ all of its consumers.

This change affects the arbitrary order for files/artifacts that form a cycle.

Enable parallel resolution of the dependency graph

This commit introduces parallel resolution of a dependency graph. When the graph is resolved:

- pick a node from the queue (starting from the root node)

- for its outgoing edges, resolve them serially, or concurrently if:

- all of the edges reference nodes which metadata hasn't been resolved yet

- or there are more than 8 edges

- then add discovered edges _in order_ add the end of the queue

This means that not _all_ graph resolution is done concurrently: there's a heuristic to tell if it's worth doing it,

and similarly, there are operations which still need to be done serially in order to have guarantees on the shape

of the graph (typically, to make sure that files are visible in a certain order).

Use doLast instead of left shift operator

Made this change in preparation for deprecating the left shift operator.

+review REVIEW-6236

Use component id rather than module version id in the 'required by' paths in resolve exceptions. This provides a better description for local components, and also allows for components that don't have a module version id.

Moved `ResolveTestFixture` into a `org.gradle.integtests.fixtures` subpackage

This has the net effect of making this class visible to the Tooling API integration

tests, since the `org.gradle.integtests.fixtures` package is explicitly allowed

in the filtering classloader used by the TAPI integration tests.

Fixed a bunch of tests to work when the machine's line separator is not \n.

Renamed subprojects/core-impl to subprojects/dependency-management.