The Grand Re-Evaluation

[aiwenar]

..., or a better way of handling values

Improve how arguments to various directives are evaluated.

Fixes #95, fixes Connexions/cnx-recipes#604

How it works currently

Every directive which accepts evaluated values either has its own evaluating logic, or uses Oven.eval_string_value. This poses a number of problems:

• Code evaluating values is re-implemented with very few changes in every function which accepts evaluated values. This in particular causes most CSS functions to be implemented three times.
• Functions which accept a default value evaluate said value as string, which causes problems when used in content: directive, see for example Connexions/cnx-recipes#604.

Additionally since evaluation happens immediately when processing directives target-* functions can only be implemented by pretending to be a string (oven.TargetVal).

How it works with this PR

All directives now use a single shared value evaluator: expr.evaluate (through a convenient wrapper Oven.evaluate). CSS functions are now implemented only once, and instead of returning a concrete Python value return instead a typed value (css.Value): combination of a CSS type (css.Type) and some transparent value. To extract the concrete value from a typed value callers of expr.evaluate specify what type they expect, and values are converted to that type, if possible.

This solves problems such as Connexions/cnx-recipes#604, as now attr() is free to return either a css.String, when the attribute exists, or a css.DocumentFragment returned by the default value of content().

To solve the special status of target-*() functions, conversion from typed values to concrete values has been delayed until Oven.bake. This allows target-*() functions to be just evaluations of normal functions delayed until after the entire tree has been processed, and has an (partially) unintended consequence of making target-*() functions usable in every context in which their normal counterpart is. For example, first-letter(target-string(#ref, name)) or target-string(#ref, name, target-string(#other-ref, other-name)) are now correct expressions and work just as one would expect.

Additional changes this PR makes

To simplify parsing of CSS values a new class, css.Parser, has been introduced, and all current manual parsers rewritten using it. This has substantially simplified implementations of nearly all directives (just take a look at do_string_set).

Failed tests

• inside: I don't know why this happens
• default: In CSS Values and Units Module Level 4, attr() takes a qualified name as attribute name, and that's how I implemented it now. Previously however cnx-easybake accepted anything which evaluated to a string. If this change is intentional implementation of attr should be changed, if not, the test case.
• string_set and unicode: These tests suffer from first-letter is an identity when applied to result of the string() function #95. I assume that means tests need to be corrected, but I restrained from doing so until someone can confirm this.
• uuid: UUIDs are now randomly generated rather than mocked. I have not been able to find out why.

[reedstrm]

@aiwenar sorry, didn't see this request 4 days ago. Starting a review now, but based on the description alone, this is a massive improvement! Will get it reviewed, merged and run through the battery of regression tests w/ Helene ASAP.

[brittweinstein]

@helenemccarron can we close this?

[helenemccarron]

I think @reedstrm was going to review, this would be useful to get this merged, I think but I do not know if we have the resources do so as of now. Is there a way to transfer knowledge so we can take advantage of this work>