Allow null values for jsonencode

[pselle]

This addresses hashicorp/terraform#23062, where jsonencode does not allow null values, but it is documented in Terraform that it does: https://www.terraform.io/docs/configuration/functions/jsonencode.html.

Terraform currently uses this zclconf edition of the go-cty library, but if this change is unwarranted here, perhaps it would be a better fit to migrate Terraform to https://github.com/hashicorp/go-cty and make changes there.

[codecov]

Codecov Report

Merging #54 into master will increase coverage by 0.00%.
The diff coverage is 100.00%.

@@           Coverage Diff           @@
##           master      #54   +/-   ##
=======================================
  Coverage   70.53%   70.54%           
=======================================
  Files          79       79           
  Lines        7151     7153    +2     
=======================================
+ Hits         5044     5046    +2     
  Misses       1664     1664           
  Partials      443      443           

Impacted Files	Coverage Δ
cty/function/stdlib/json.go	91.66% <100.00%> (+0.75%)	⬆️

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 6e20bc2...0c4042c. Read the comment docs.

[apparentlymart]

Thanks for looking at this!

My intuition here was that jsonencode(null) would return the string "null" rather than actually returning a null value, because then the result can be used anywhere that JSON syntax is expected. I think if this returned a "real" null then it would end up just pushing the error further downstream, by assigning a null to an argument that isn't expecting it. 🤔 What do you think?

[pselle]

Null is an allowed value according to the JSON spec, and it's what the Terraform docs say is expected behavior (encoding null results in null).

To be honest, reading the issue I wondered if this was a case where "this is fixed by updating the docs" but it appears null is a valid case. If we look at other languages, json.dumps in Python translates Python's None (since there's no null in Python) to null in JSON.

So, I think this is the right move, and returning the string null would be less correct that returning null.

[alisdair]

It doesn't seem right to me that we're returning a cty.NullVal here. My understanding is that any non-error return type of jsonencode should always be a cty.String, which is a JSON representation of the passed value.

In this case, I think that means cty.StringVal("null"), i.e. a JSON document which consists only of the JavaScript null value.

For example, for JavaScript:

> JSON.stringify(null)
'null'

Python:

>>> json.dumps(None)
'null'

Ruby:

> JSON.generate(nil)
=> "null"

All of these are strings.

[pselle]

@alisdair Thanks! Your comment helped me understand what I imagine was Martin's suggestion. I've committed an update.

[alisdair]

This looks right to me!

[apparentlymart]

Ahh yes, sorry... that was indeed what I was meaning to say. Sorry for being unclear! 😖

This looks great, and I'm going to merge it now.