Add CSS value selectors to TagHelper attributes

[NTaylorMullen]

Add the ability to do:

HtmlTargetElement(Attributes = "href^=~/") // Value starts with "~/"
HtmlTargetElement(Attributes = "href$=;") // Value ends with ";"
HtmlTargetElement(Attributes = "href=something") // Value equals "something"

This would enable our UrlResolutionTagHelper in MVC to only apply when it sees a ~/.

[NTaylorMullen]

/cc @Eilon @rynowak

Talked with @DamianEdwards and we need to do this in the name of perf for UrlResolutionTagHelper for RC2.

[rynowak]

See: aspnet/Mvc#3516 (comment)

[Eilon]

Is there something not super complicated that we can do for v1? I'd rather not invent/implement a language for this...

[DamianEdwards]

Spoke with @davidfowl and for now perhaps we just support a callback model instead. e.g.

[HtmlTargetElement("a", Attributes = "href", PredicateMethodName = "OnEmitAsTagHelper")]
public class MyTagHelper : TagHelper
{
    public static bool OnEmitAsTagHelper(string elementName, TagHelperAttributeList attributes)
    {
        return attributes["href"].StartsWith("~);
    }
}

[DamianEdwards]

Or maybe we create a new context type for this stage, e.g. TagHelperBuilderContext, which we can grow over time as we add more support for Tag Helpers to participate in view compilation?

[dougbu]

There are at least three perf issues related to the UrlResolutionTagHelper:

1. It targets a number of elements that would not be targeted otherwise, leading to "extra" TagHelperExecutionContext, TagHelperOutput, ... allocations. Pretty much any solution that doesn't require an instance of the tag helper should address this part of the problem.
2. The tag helper system allocates large-sized TagHelperAttributes and related arrays / lists and so on. Recent changes such as Lazily initialize TagHelperAttributeLists. #683 (still in PR) should help here. But any callback approach would mean these TagHelperAttributes are still allocated.
   • Hmm, might want to remeasure after Lazily initialize TagHelperAttributeLists. #683 goes in.
3. The support for attribute values containing C# expressions that start with "~/" requires early evaluation of IHtmlContent values. We could reduce this overhead significantly if we removed support for entirely dynamic attributes and IHtmlContent included a string FirstSegment { get; } property for UrlResolutionTagHelper to inspect. The solutions mentioned above share removal of entirely dynamic attributes with this suggestion.

Do we have current numbers showing whether 1, 2, or 3 is the primary issue for significantly reduced performance when UrlResolutionTagHelper is enabled? (2 is the one that I see near the top for many scenarios. I've heard others point toward 1. But I'm not sure of the post-#683 situation.)

[DamianEdwards]

The attributes in this case would be allocated during view compilation, so that's not an issue. The example was just pseudo code, so it's likely a different type to pass the attributes to the predicate is better.

As for attributes that contain expressions, we need to do the simplest thing for now that preserves the pre-Tag Helpers behavior.

[NTaylorMullen]

The work you're suggesting is near equivalent workload to implementing a CSS selector on the Attributes property. Seeing that we have a whole host of codegen related extension points slated for post v1 I think the OnEmitAsTagHelper type method would make more sense there and a CSS selector here.

[DamianEdwards]

I think we should do some rapid experiments here to validate some of the ideas.

[NTaylorMullen]

Which ideas are you referring to?

[NTaylorMullen]

@DamianEdwards and I talked offline and we're gonna go with the CSS selector on Attributes =. I assume this feature is approved for RC2?

[DamianEdwards]

Hack something up so we can assess whether it actually helps the perf issue.