ben-manes / caffeine / #5497

100%

master: 100%

push

github

Refactor the node and cache code generators

Replace the untyped Set<List<Object>> combinations and the parallel
@Nullable Feature[] featureByIndex array with a cartesian product of
Optional<Feature> dimensions. Eviction is modeled as one mutually-exclusive
dimension (none, size, or weight), which removes the
MAXIMUM_WEIGHT-supersedes-MAXIMUM_SIZE fixup.

Share the driving machinery by composition: a FactoryGenerator is configured
with the matrix dimensions, the class-name encoding, and a SpecFactory that
builds a single specialization, so the node and local-cache generators no
longer duplicate generate()/writeJavaFile()/reformat() and the parent/child
derivation. Each generator is now just its configuration -- a dimension list, a
rule list, an encode alphabet, and a makeSpec method reference.

Make the class-name encoding independent of set iteration order. The name is a
prefix-coded concatenation of feature codes in a canonical order; that order
was previously carried implicitly by ImmutableSet insertion order (guarded by
two apologetic @SuppressWarnings, since an enum-set collector would reorder the
features and silently corrupt the local-cache names whose matrix order differs
from the enum order). The canonical order is now derived from the single
dimension list and applied by sorting at the naming boundary, so no Set or
collector choice can change the output. Since RuleContext already re-wraps the
feature sets into enum order, the rule layer was never order-sensitive; the
parent class name is computed centrally and handed to makeSpec, so both
suppressions are removed. A checkState asserts the prefix-coding invariant at
generation time.

The generated sources are byte-identical.

4034 of 4042 branches covered (99.8%)

8218 of 8218 relevant lines covered (100.0%)

1.0 hits per line