Fold compound expressions with literal arguments into literals

[anandthakker]

When parsing an expression, if its arguments are all Literals, then evaluate it immediately and yield a Literal instead.

Yields a modest perf improvement in tile_layout_dds:

• before: 51.37629 - 52.20937 ms / iteration
• after: 47.81951 - 49.86712 ms / iteration

More importantly, this means we get validation errors at compile time rather than evaluation time for the common case of color curves defined like [curve, [linear], [zoom], 0, "red", 10, "bleu mispelled"]

benchmark	master 2200752	expressions-fold-constant 0567a8f
map-load	111 ms	119 ms
style-load	128 ms	168 ms
buffer	1,167 ms	1,120 ms
tile_layout_dds	1,329 ms	1,255 ms
fps	60 fps	60 fps
frame-duration	7.1 ms, 1% > 16ms	6.9 ms, 2% > 16ms
query-point	0.98 ms	1.24 ms
query-box	73.43 ms	75.72 ms
geojson-setdata-small	17 ms	0 ms
geojson-setdata-large	88 ms	175 ms
filter	Done	Done

[anandthakker]

Example generated code for a color curve with "red" as the stop output value (excerpt of the curve/linear-color integration test output)

before

function e4() {   return "red" }
var v5 = [e4];
function e6() {   return $this.toColor(v5) }

after

var v4 = (new $this.Color(1, 0, 0, 1));
function e5() {   return v4 }

[mourner]

Very cool!

[anandthakker]

Rebased, generalized, and updated benchmarks.

[jfirebaugh]

What's the reason for these changes?

[anandthakker]

Breaking a new dependency cycle, which was introduced by parse_expression now depending on evaluation_context. I guess it's better to break it there instead (especially since we're already doing so in that module)

[jfirebaugh]

Nit: return ctx.addVariable(value);

[jfirebaugh]

This looks similar to isFeatureConstant/isZoomConstant. Can they share more code?

[anandthakker]

👍 I almost did that, and then stopped myself because I thought I was prematurely abstracting.

[anandthakker]

Oh, actually the code itself is a bit different. isConstant only looks one level deep (because that's sufficient for the way we're using it). We could consolidate the base case code, though

[anandthakker]

@jfirebaugh updated to address your notes/questions