To make the process as smooth as possible, this document outlines all the things you need to do to upgrade to 0.19.
- Command Line
elm.json
- Changes
--optimize
- Compiler Performance
- Parse Errors
- Stricter Record Update Syntax
- Removed User-Defined Operators
Note: You can try out
elm-upgrade
which automates some of the 0.18 to 0.19 changes. It is also in an alpha stage, and Aaron has said it makes sense to talk things through here.
There is now just one elm
binary at the command line. The terminal commands are now:
# 0.19 # 0.18
elm make # elm-make
elm repl # elm-repl
elm reactor # elm-reactor
elm install # elm-package install
elm publish # elm-package publish
elm bump # elm-package bump
elm diff # elm-package diff
elm-package.json
becomes elm.json
which is specialized for applications and packages. For example, it helps you lock your dependencies in applications and get broad dependency ranges in packages.
See the full outlines here:
elm.json
for applicationselm.json
for packages
Both are quite similar to the elm-package.json
format, and elm-upgrade
can help you with this.
String.toInt : String -> Maybe Int
(notResult
anymore)String.toFloat : String -> Maybe Float
(notResult
anymore)Basics.toString
becomesDebug.toString
,String.fromInt
, andString.fromFloat
.Basics.rem 451 10
becomesremainderBy 10 451
451 % 10
becomesmodBy 10 451
(,)
becomesTuple.pair
style : List (String, String) -> Attribute msg
becomesString -> String -> Attribute msg
Html.beginnerProgram
becomesBrowser.sandbox
.Html.program
becomesBrowser.element
andBrowser.document
.
Json.Encode
andJson.Decode
moved toelm/json
Time
andDate
moved toelm/time
with a significantly improved APIRandom
moved toelm/random
with a better implementation and a few new functionsRegex
moved toelm/regex
with a much clearer README
elm-lang/*
moved toelm/*
evancz/url-parser
moved toelm/url
with a simpler and more flexible APIelm-tools/elm-parser
moved toelm/parser
with speed boost when compiling with the--optimize
flagelm/browser
combines and simplifies the following 0.18 packages:elm-lang/navigation
with smoother APIselm-lang/dom
with ability to get node positions and dimensions.elm-lang/mouse
with decoderselm-lang/window
elm-lang/keyboard
uses decoders like thiselm-lang/page-visibility
elm-lang/animation-frame
uncurry
curry
flip
(!)
Prefer named helper functions in these cases.
You can now compile with elm make --optimize
which enables things like:
- Reliable field name shortening in compiled assets
- Unbox things like
type Height = Height Float
to just be a float at runtime - Unbox
Char
values - Use more compact names for
type
constructors in compiled assets.
Some of these optimizations require "forgetting information" that is useful while debugging, so the Debug
module becomes unavailable when you add the --optimize
flag. The idea being that you want to be shipping code with this flag (like -O2
in C) but not compiling with it all day in development.
I did a bunch of performance optimizations for the compiler itself. For example:
- I rewrote the parser to be very significantly faster (partly by allocating very little!)
- I revamped how type inference looks up the type of foreign variables to be
O(1)
rather thanO(log(n))
- I redid how code is generated to allow DCE with declarations as the level of granuality
- Packages are downloaded once per user and saved in
~/.elm/
- Packages are built once for any given set of dependencies, so they do not contribute to build times of fresh projects.
Point is, the compiler is very significantly faster!
Part of rewriting the parser was making nicer parse errors. Many people only really see them when getting started, and rather than saying "man, these are terrible" they think "man, programming is hard" leading to a big underreporting of quality issues here. Anyway, please explore that a bit and see if you run into anything odd!
It used to be possible for { r | x = v }
to change the type of field x
. This is no longer possible. This greatly improves the quality of error messages in many cases.
You can still change the type of a field, but you must reconstruct the record with the record literal syntax, or with a record constructor.
The idea is that 99.9% of uses get a much better experience with type errors, whereas 0.1% of uses become somewhat more verbose. As someone who had a bit of code that changed record types, I have found this to be a really excellent trade.
It is no longer possible to define custom operators. For example, someone defined:
(|-~->) : (a -> a1_1 -> a3) -> (a2 -> a1_1) -> a -> a2 -> a3
They are still able to define that function, but it will need a human readable name that explains what it is meant to do. The reasoning behind this decision is outlined in detail in this document.
toString
— A relatively common bug was to show anInt
in the UI, and then later that value changes to something else.toString
would just show wrong information until someone noticed. The newString.fromInt
andString.fromFloat
ensure that cannot happen. Furthermore, more elaborate types almost certainly need localization or internationalization, which should be handled differently anyway.