Running a program is a closed world — every namespace is required, then it runs
to completion — so make it direct-link by default (inlining, record shapes, the
inference's specialization), and for a -m/-M entry auto-enable the whole-program
cross-namespace inference pass. A decomposed multi-namespace program was ~3.7x
slower than the same code in one namespace purely because per-namespace
inference can't see a caller in a not-yet-loaded namespace; this closes that for
the common case with no flags and no hints.
Interactive modes (repl, -e, nrepl-server) stay indirect/open — they have to let
you redefine vars, which direct-linking seals against. Opt-outs:
JOLT_NO_DIRECT_LINK forces the open path even for a program run (hot-reload,
runtime redefinition); JOLT_NO_WHOLE_PROGRAM keeps direct-linking but per-ns;
JOLT_DIRECT_LINK / JOLT_WHOLE_PROGRAM still force-on. Namespaces required inside
-main (after the batch pass) fall back to per-ns inference.
The success checker (RFC 0006) rides on the inference for free, but a casual
program run shouldn't spam type warnings just because it now direct-links, so its
default-on is suppressed when direct-linking was auto-enabled (:direct-link-auto?);
an explicit JOLT_DIRECT_LINK or JOLT_TYPE_CHECK still turns it on. whole-program-
test and devirt-test opt their per-ns baseline out of the new auto-default.
Docs: RFC 0005 gains 'Compilation modes and defaults' + 'Cross-namespace
inference'; RFC 0004 documents cross-ns/param hints; self-hosting-compiler and
--help updated. Full gate green.
direct-var? now treats a cfunction root the same as a function root, so a
call/ref to a native fn (clojure.math/sqrt et al.) embeds the value instead of
a per-call cell deref. This was the hot indirection in the ray tracer — sqrt
runs every bounce — and it applies in every direct-link build, not just
whole-program.
const-link? is new and whole-program-only: in a closed world every non-dynamic
var has a stable root, so embed it as a constant (quoted unless it's already
callable) rather than reading the cell each reference. Covers what direct-var?
can't — ^:redef vars (reloading is off under the flag), data defs, and record
type/ctor roots. Dynamic vars stay indirect; a nil (not-yet-defined) root stays
indirect and the whole-program re-emit picks it up once the root is in place.
Measured on the records ray tracer: hot-path indirect refs (sqrt + data vars)
gone; the only indirect refs left are cold defrecord self-references. whole-
program-test now also checks a ^:redef fn and a data def so the per-ns vs
whole-program comparison guards const-link soundness.
JOLT_WHOLE_PROGRAM (requires direct-linking) defers the per-namespace inference
and runs ONE fixpoint over every user unit at once, so param types propagate
across namespace boundaries — a non-inlined fn's record params get proven from
its callers in another unit, which the per-ns pass can't see. Sound only under
the closed-world assumption (no later eval/redefinition) the flag asserts; slow,
memory-heavy builds are the documented trade-off (the reason it's opt-in).
infer-unit! now takes one ns-name OR a list; infer-program! gathers all recorded
user namespaces and runs the existing fixpoint over the union (re-emit was already
ns-agnostic — keyed by var-key, callee-first). The evaluator defers + records each
unit under the flag; run-main triggers infer-program! after all requires, before
-main. Off by default — per-ns behaviour unchanged.
Measured: a recursive (non-inlined) cross-ns record reader runs 1.66x faster
(8.9s -> 5.3s) — params proven -> bare-index reads. NOTE: small accessor fns are
INLINED cross-ns and records carry GLOBAL declared shapes, so most record reads
are already proven without this pass; the win is for non-inlined hot fns, and it's
the foundation for future whole-program work (devirtualization, unboxing).
Adds whole-program-test (subprocess soundness: per-ns and whole-program produce
identical results on a cross-ns record program).
