Release builds accidentally strip "unused" namespace requires that register specs + multimethods (`:target :esm-files`).

[commiterate]

Release builds accidentally strip "unused" namespace requires that register specs when using :target :esm-files.

Reproduction: https://github.com/commiterate/test-clojurescript-esm (the Notes section in the readme has more details)

For example, suppose we have:

ns_x.cljc

(ns ns-x
  (:require
    #?(:cljs
       [cljs.spec.alpha :as spec :include-macros true]
       :default
       [clojure.spec.alpha :as spec])))

(spec/def ::x number?)

ns_y.cljc

(ns ns-y
  (:require
    #?(:cljs
       [cljs.spec.alpha :as spec :include-macros true]
       :default
       [clojure.spec.alpha :as spec])
    [ns-x :as ns-x]))

(spec/fdef y
  :args (spec/cat :keys (spec/keys* :req-un [::ns-x/x]))
  :ret number?)

(defn y
  "y = 2x"
  [& {:keys [x]}]
  (* 2 x))

The ns-x namespace import in ns-y will be stripped from the compiled JS code and result in a "spec not found" error when importing the ns-y module.

[commiterate]

It looks like the workaround is to assign specs to variables. For example:

ns_x.cljc

(ns ns-x
  (:require
    #?(:cljs
       [cljs.spec.alpha :as spec :include-macros true]
       :default
       [clojure.spec.alpha :as spec])))

(def x-spec
  (spec/def ::x number?))

ns_y.cljc

(ns ns-y
  (:require
    #?(:cljs
       [cljs.spec.alpha :as spec :include-macros true]
       :default
       [clojure.spec.alpha :as spec])
    [ns-x :as ns-x]))

(def y-fspec
  (spec/fdef y
    :args (spec/cat :keys (spec/keys* :req-un [ns-x/x-spec]))
    :ret number?))

(defn y
  "y = 2x"
  [& {:keys [x]}]
  (* 2 x))

spec's def takes a keyword + spec form, associates the 2 in the global registry, then returns the keyword.

That means ns-x/x-spec is :ns-x/x. This can be used in spec forms like keys* which requires keywords for its parameters (e.g. :req, :opt).

spec's fdef does the same and returns a qualified function symbol but I don't think many people use this symbol in other places. It might be a good idea to still assign function specs to variables so they can be pruned by dead code elimination.

[commiterate]

Related: #1102

[commiterate]

Hmm the workaround won't work for purely side-effecting namespaces that register multimethods like Emmy's emmy.numbers namespace.

https://github.com/mentat-collective/emmy/blob/v0.32.0/src/emmy/numbers.cljc

The multimethods are defined in emmy.generic with the implementations in various namespaces like emmy.numbers. That means we need to do:

(ns my-ns
  (:require
    [emmy.generic :as g]
    ; Only contains multimethod implementations.
    ; Require will be dropped by advanced optimizations.
    [emmy.numbers]))

(def const (g/+ 1 1))

It doesn't seem like :force-imports will work since the :modules map will be replaced.

shadow-cljs/src/main/shadow/build/targets/esm_files.clj

Lines 54 to 58 in 4973aea

	;; will be replaced after compilation
	(build-api/configure-modules
	{:main
	{:entries []
	:depends-on #{}}})

[thheller]

I'm sorry I missed this.

Unfortunately the reproduction includes way too much stuff that should be entirely irrelevant to reproduce this. I can't even run it since I do not have bun and will not install it. If you could reduce your reproduction to use no external dependencies that would greatly speed up me looking into this.

[commiterate]

No worries!

Minimized the toolchain in the reproduction repository.

The main branch doesn't contain the spec workaround (to show the original spec issue) while the spec_workaround branch does (to show the multimethod issue).

The Notes section of their respective readmes show sample executions + errors.