clojure.walk and 11 analyzer namespaces are compiled without intrinsics

[skydread1]

Problem

The intrinsic registry is a plain def, so compiling its own namespace empties it:

;; magic-compiler/src/magic/analyzer/intrinsics.clj:11
(def intrinsic-forms (atom {}))

build.clj compiles that namespace at position 25 and magic.intrinsics, which fills it, at 38. The 12 entries in between are emitted with an empty registry, clojure.walk among them.

walk.clj:45 is (instance? clojure.lang.IMapEntry form). The committed DLL:

ldsfld    Var ...clojure_core$instance_QMARK_
call      Var::getRawRoot()
castclass IFn
ldtoken   clojure.lang.IMapEntry
callvirt  IFn::invoke(object, object)
unbox.any Boolean

With the registry intact it is one isinst clojure.lang.IMapEntry. walk__0 is 745 bytes instead of 607.

Valid IL either way, so nothing fails and no drift check fires. The cost is a Var deref and a boxed Type per call, in walk / prewalk / postwalk and in 11 analyzer passes.

Not new: the 2022-03-26 build of clojure.walk.clj.dll has the intrinsic, the 2022-11-19 build does not.

Suggestion

- (def intrinsic-forms (atom {}))
+ (defonce intrinsic-forms (atom {}))

magic.util already does this for the gensym map (util.clj:10). Reordering build.clj will not work: magic.analyzer requires clojure.walk and magic.intrinsics requires magic.analyzer, so walk has to come first.

Takes effect on bootstrap pass 2, and should regenerate all 12 DLLs.

[skydread1]

Compile time is unchanged, contrary to Problem: 20s/21s before, 22s/21s after. The win is run time: postwalk 83ms to 68ms, keywordize-keys 153ms to 112ms.

Registering a var also disables its inline (analyzer.clj:373), and most intrinsics decline on untyped args, so those sites lose both: 65 inlined RT::nth calls became boxed Var dispatches, total IL +1.9%. Pre-existing, and these 12 DLLs were the only ones exempt, since everything else in the tree has always compiled with a full registry. Ramsey marked it in 2017 (TODO reapply inlining ... at intrinsics.clj:31); closing it would improve every DLL, so it gets its own issue.

Side effect: refresh.clj's exclusion list drops from 21 entries to 9. Only 3 of the 12 removed have a source under src/stdlib (string, set, walk); the other 9 are clojure.tools.analyzer.*, whose sources sit in deps/maven/ and were skipped regardless. So check-drift byte-verifies 28 stdlib DLLs instead of 25.