This allows us to accurately track the source spans for parenthesized
expressions, which in turn allows us to print accurate error
indications.
Adding a new class for this more accurately represents the structure
of the expression, but it also involves an extra allocation during
parsing and an extra level of nesting during evaluation which could
have a small but real performance impact.
We could alternatively add a package-internal setter for
Expression.span, and update the source span for parenthesized
expressions after they're initially parsed. However, this has its own
downsides: it adds complexity and mutability to the object model; and
many expression classes currently use lazily-generated spans, so
making them settable would require adding extra slots on those
classes.
I decided to go with the extra class because it only adds overhead
when parentheses are actually used in practice, as opposed to adding
overhead to every list/color/etc. The runtime overhead is also likely
to be mitigated if at any point we add a constant-folding step.
When a stylesheet is imported, the parsed stylesheet object is cached
based on its canonical URL. However, the stylesheet.span.sourceUrl was
based on the text of the import that was used to load that stylesheet.
The idea was to make the source URL in stack traces look nicer, but it
meant that relative URLs could be resolved based on the old importer's
URL before being sent to the new importer, which caused bugs.
Now stylesheet.span.sourceUrl is always the canonical URL of the
stylesheet, and thus safe to cache. We then use the import cache to
convert the canonical URL to a human-friendly URL at the point at
which we generate stack traces.
This also deprecates support for relative canonical URLs. The
semantics of these URLs were always unclear, and with the new change
in import internals the old behavior doesn't make much sense. It's
preserved for backwards-compatibility, but deprecated.
We were previously trimming at bath the compound selector level and
the selector list level. This now only trims at the selector list
level.
The _trim() function also took a list of selector lists, on the
idea (from Ruby Sass) that it could avoid trimming selectors generated
from the same extension that were very unlikely to be redundant. In
practice, though, the fact that we weaved together selector lists at
the compound level meant that we didn't have meaningful
known-non-redundant lists like Ruby Sass did, so this wasn't useful.
This has no behavioral effect, although it does seem to improve
performance slightly for extend-heavy stylesheets.
