parser: Remove empty multiline string parts earlier

[infinisil]

Motivation

Consider these two expressions:

# a.nix
[
  ''
    ${"string"}''
]

# b.nix
[
  ''${"string"}''
]

They parse differently:

$ nix-instantiate --parse a.nix
[ (("" + "string")) ]
$ nix-instantiate --parse b.nix
[ ("string") ]

The first one also ends up allocating an extra thunk to be evaluated, but they evaluate to the same:

$ NIX_SHOW_STATS=1 NIX_SHOW_STATS_PATH=stats.json nix-instantiate --eval --strict a.nix
[ "string" ]
$ jq .nrThunks stats.json
1

$ NIX_SHOW_STATS=1 NIX_SHOW_STATS_PATH=stats.json nix-instantiate --eval --strict b.nix
[ "string" ]
$ jq .nrThunks stats.json
0

There is no need for this inconsistency, so this PR changes it to treat both the same (a.nix now acts the same as b.nix). This is done by preventing the propagation of empty strings to be concatenated in the parser already.

Context

I've found this when I tried to verify that nix-instantiate --parse doesn't change upon applying a formatter on Nixpkgs. Turns out it does change at least because of this.

Priorities and Process

This work is sponsored by Antithesis ✨

Add 👍 to pull requests you find important.

The Nix maintainer team uses a GitHub project board to schedule and track reviews.

[roberth]

LGTM

[infinisil]

Hold on, I think I found a minor problem with this, will follow-up

[infinisil]

Alright should be good now.

[infinisil]

The problem I discovered is fixed by this. This is necessary because newly es2 can be entirely empty, in which case we should not create an ExprConcatStrings, but just return an empty string directly. Not doing so would give inconsistent results from

''''

(which would newly create an ExprConcatStrings, allocating a thunk)

and

''
''

(which would be turned into "")

This is now covered in the test.

[infinisil]

Applied and rebased with your suggestion :)