wip: generalize shape mechanism off the hardcoded vec3 shape (jolt-t34)

Removes the {:r :g :b} hardcoding. ANY constant key set is now a shape:
- inference: a struct type from a map LITERAL carries :shape (its canonical
  str-sorted key vector — completeness); joins/access-inferred structs lack
  it, so they never get a bare index. The literal node and lookup subjects
  carry the shape; the back end derives the index from it.
- backend: emit-map turns any shape-tagged const-key map into a shape tuple;
  emit-kw-lookup reads the field by bare index when the complete shape is
  proven, else by the value's own descriptor (so a shape-rec whose :shape was
  dropped by a join still reads correctly).
- runtime: core-get and core-assoc handle shape-recs.

Status: CORRECT for direct field access, container round-trips, and assoc
(minimal repros pass). NOT yet complete — the full ray tracer still hits an
uncovered path (a shape-rec reaching a map op without coverage: keys/vals/
count/seq/equality/print/jolt-call/dissoc/contains?/the interpreter's
coll-lookup all still need shape-rec branches). And the perf win needs
COMPLETENESS PRESERVATION through joins/containers (merge-fields/cap drop
:shape today, so nested vec3 access falls to the descriptor path, slower than
a struct get) — without it the general version is slower than the vec3
prototype.

All behind JOLT_SHAPE (off by default). Gate green with the flag off, suite
4718. This preserves the general design; the transparency layer + completeness
preservation are the remaining multi-session work.
This commit is contained in:
Yogthos 2026-06-13 17:01:35 -04:00
parent d33fb85041
commit e377c223b6
3 changed files with 93 additions and 75 deletions

View file

@ -806,21 +806,26 @@
;; k, if known, else :any.
(defn- struct-safe? [t] (struct-type? t))
(defn- field-type [t k] (if (struct-type? t) (get (sfields t) k :any) :any))
;; vec3 shape detection (jolt-t34, prototype): a struct type whose key set is
;; exactly {:r :g :b}. The back end represents such maps as shape tuples, so a
;; lookup on a value PROVEN to be this shape can read by bare index with no
;; runtime check. Scoped to the one shape for the prototype.
(defn- vec3-shape? [t]
(and (struct-type? t)
(let [fs (sfields t)]
(and (get fs :r) (get fs :g) (get fs :b) (= 3 (count (keys fs)))))))
;; the structural hint for a subject: :shape when provably the vec3 shape (bare
;; indexed read), else :struct when raw-get-safe.
(defn- struct-hint [t] (if (vec3-shape? t) :shape :struct))
;; Shape (hidden class, jolt-t34). A struct type built from a map LITERAL carries
;; its complete layout — :shape, the canonical (str-sorted) key vector. The back
;; end represents such a map as a shape tuple and reads a field by bare index.
;; A struct type from a JOIN or from field-access inference has no :shape
;; (incomplete: the full key set isn't proven), so it keeps the dynamic path —
;; never a bare index. No shape is hardcoded; any constant key set is one.
(defn- shape-order
"Canonical key order for a shape: keys sorted by their string form, so two
literals with the same keys in any order intern to the same shape."
[ks] (vec (sort (fn [a b] (compare (str a) (str b))) ks)))
(defn- type-shape [t] (get t :shape))
;; tag a node (any expression, not just a :local) so the back end can specialize
;; a lookup whose SUBJECT is that node — this is what makes nested access work:
;; (:direction ray) is tagged struct, so (:r (:direction ray)) drops its guard.
(defn- mark-hint [node h] (assoc node :hint h))
;; tag a lookup subject as a struct, carrying the complete shape when known
;; (so the back end bare-indexes) — jolt-t34
(defn- mark-struct [node t]
(let [n (assoc node :hint :struct)]
(if (get t :shape) (assoc n :shape (get t :shape)) n)))
;; a value provably neither nil nor false — the back end only builds a struct
;; (vs a phm) when every value is non-nil/non-false, so a map literal is a struct
;; only when all its values have such a type. Collections are non-nil.
@ -940,7 +945,8 @@
(let [t (get tenv (get node :name))]
[(if t t :any)
(cond
(struct-safe? t) (assoc node :hint (struct-hint t))
(struct-safe? t) (let [n (assoc node :hint :struct)]
(if (type-shape t) (assoc n :shape (type-shape t)) n))
(vec-type? t) (assoc node :hint :vector)
:else node)])
(= op :map)
@ -953,10 +959,14 @@
struct? (and (> (count res) 0)
(every? (fn [pr] (scalar-const? (nth pr 0))) pairs)
(every? (fn [r] (truthy-type? (nth r 2))) res))
t (if struct?
(cap (mk-struct (reduce (fn [m r] (assoc m (nth r 3) (nth r 2))) {} res)) type-depth)
:any)]
[t (assoc node :pairs (mapv (fn [r] [(nth r 0) (nth r 1)]) res))])
base (when struct?
(cap (mk-struct (reduce (fn [m r] (assoc m (nth r 3) (nth r 2))) {} res)) type-depth))
;; a literal is a COMPLETE shape: carry its sorted key vector so the
;; back end can lay it out and bare-index lookups (jolt-t34)
shp (when (and base (struct-type? base)) (shape-order (keys (sfields base))))
t (if base (if shp (assoc base :shape shp) base) :any)
node' (assoc node :pairs (mapv (fn [r] [(nth r 0) (nth r 1)]) res))]
[t (if shp (assoc node' :shape shp) node')])
(= op :vector)
(let [irs (mapv (fn [x] (infer x tenv)) (get node :items))
ets (mapv (fn [r] (nth r 0)) irs)
@ -998,7 +1008,7 @@
(and (= :const (get fnode :op)) (keyword? (get fnode :val)) (>= n 1) (<= n 2))
(let [mr (infer (nth args 0) tenv)
mt (nth mr 0)
msub (if (struct-safe? mt) (mark-hint (nth mr 1) (struct-hint mt)) (nth mr 1))
msub (if (struct-safe? mt) (mark-struct (nth mr 1) mt) (nth mr 1))
ft (field-type mt (get fnode :val))
dr (when (= n 2) (infer (nth args 1) tenv))]
[(if dr (join ft (nth dr 0)) ft)
@ -1009,7 +1019,7 @@
(>= n 2) (= :const (get (nth args 1) :op)) (keyword? (get (nth args 1) :val)))
(let [mr (infer (nth args 0) tenv)
mt (nth mr 0)
msub (if (struct-safe? mt) (mark-hint (nth mr 1) (struct-hint mt)) (nth mr 1))
msub (if (struct-safe? mt) (mark-struct (nth mr 1) mt) (nth mr 1))
kr (infer (nth args 1) tenv)
ft (field-type mt (get (nth args 1) :val))
dr (when (= n 3) (infer (nth args 2) tenv))]

View file

@ -355,15 +355,13 @@
# - ^:struct / ^Record hinted subject: skip the guard, bare get (~20 vs ~36ns).
# - hinted + JOLT_CHECK_HINTS: keep the guard but THROW on the tagged arm, so a
# lying hint surfaces a clear error (dev aid; off by default, no perf cost).
# vec3 shape layout: descriptor at 0, then sorted keys [:b :g :r] at 1,2,3 (jolt-t34)
(def- vec3-shape-idx {:b 1 :g 2 :r 3})
(defn- emit-kw-lookup [subj-node m-expr k d-expr]
# the subject is a struct (raw-get-safe) when hinted so — by an explicit
# ^:struct/^Record hint on a local, OR by inference tagging ANY subject
# expression it proved to be a struct (jolt-d6u/RFC 0005), which is what lets
# nested access like (:r (:direction ray)) drop its guard.
(def hinted (and subj-node (or (= :struct (subj-node :hint)) (= :shape (subj-node :hint)))))
(def checked (and (= :struct (subj-node :hint)) (os/getenv "JOLT_CHECK_HINTS")))
(def hinted (and subj-node (= :struct (subj-node :hint))))
(def checked (and hinted (os/getenv "JOLT_CHECK_HINTS")))
(def m (if (symbol? m-expr) m-expr (jsym)))
(def wrap (fn [body] (if (symbol? m-expr) body ['let [m m-expr] body])))
(def err (when checked
@ -371,14 +369,22 @@
k " " (subj-node :name) ") — value carries :jolt/type "
"(a phm/sorted/transient/lazy-seq), not the plain "
"struct/record the ^:struct/^Record hint asserts")]))
(def sidx (and (os/getenv "JOLT_SHAPE") (get vec3-shape-idx k)))
# subject PROVEN to be the vec3 shape (jolt-t34): read by bare index, no check.
(def shaped (and sidx (= :shape (subj-node :hint))))
# otherwise, when JOLT_SHAPE is on and the key is a vec3 field but the shape
# isn't proven, runtime-check (tuple? + struct? of elem 0, inline opcodes).
# Subject carries a complete :shape (jolt-t34) => it is provably a shape-rec;
# the field reads by bare index. The shape is the inference's canonical key
# vector, so the index is the key's position in it, +1 for the descriptor.
(def sidx
(when (and (os/getenv "JOLT_SHAPE") subj-node (subj-node :shape))
(def sk (let [s (subj-node :shape)] (if (pv/pvec? s) (pv/pv->array s) s)))
(var pos nil) (var i 0)
(each kk sk (when (= kk k) (set pos i)) (++ i))
(when pos (+ pos 1))))
# sidx: complete shape proven -> bare index (fastest). Otherwise, with shapes
# on, a raw-get-safe value may still be a shape-rec (its :shape dropped by a
# join), so read the index from the value's own descriptor when it is a tuple;
# a real struct takes the bare get. (jolt-t34)
(defn get-or-shape [getexpr]
(cond shaped ['in m sidx]
sidx ['if ['and ['tuple? m] ['struct? ['in m 0]]] ['in m sidx] getexpr]
(cond sidx ['in m sidx]
(os/getenv "JOLT_SHAPE") ['if ['tuple? m] ['in m ['+ 1 [[['in m 0] :idx] k]]] getexpr]
getexpr))
(if (nil? d-expr)
(let [fast (get-or-shape ['get m k])]
@ -481,17 +487,15 @@
(unless (and (= :const (k :op))
(or (keyword? kv) (string? kv) (number? kv) (boolean? kv)))
(set fast false)))
# Shape-record fast path (jolt-t34, JOLT_SHAPE): a vec3-shaped {:r :g :b} map
# literal becomes a shape tuple (≈2x cheaper than a struct). Scoped to this
# one shape for the prototype so other maps (hit-info/ray/material) stay
# structs and assoc on them is unaffected. Values are emitted in the shape's
# canonical (sorted-key) order; element 0 is the compile-time shape descriptor.
# Shape-record fast path (jolt-t34, JOLT_SHAPE): a const-key map literal the
# inference tagged with a complete :shape becomes a shape tuple (≈2x cheaper
# than a struct). The shape is the inference's canonical (str-sorted) key
# vector — ANY constant key set, no hardcoding. Values are emitted in that
# order; element 0 is the compile-time shape descriptor.
(def shape-keys
(when (and fast (os/getenv "JOLT_SHAPE"))
(def ks @[])
(each pair pairs (array/push ks ((norm-node (in (vview pair) 0)) :val)))
(def srt (sorted ks))
(when (deep= srt @[:b :g :r]) srt)))
(when (and fast (os/getenv "JOLT_SHAPE") (node :shape))
(def sk (node :shape))
(if (pv/pvec? sk) (pv/pv->array sk) (if (or (tuple? sk) (array? sk)) sk nil))))
(if shape-keys
(do
(def desc (shape-for shape-keys))

View file

@ -495,6 +495,42 @@
(error "Vector arg to map conj must be a pair")))
result)))))))))))))
# --- shape records (hidden classes, jolt-t34) -------------------------------
# A "shape record" is a cheap fixed-layout representation for a map literal
# whose keys are a known compile-time set (e.g. a vec3 {:r :g :b}). It is a
# Janet tuple [SHAPE v0 v1 ...] where SHAPE is an interned descriptor struct
# {:jolt/shape KEYS :idx {k->pos}}. Construction is a tuple (≈2x cheaper than a
# struct), const-keyword lookup compiles to an index, and general map ops below
# recognize it via shape-rec? and treat it as a map — so it is transparent
# wherever it flows. Created only by the backend when JOLT_SHAPE is on and the
# inference proves the shape; the runtime support here is always present so a
# shape value is handled correctly regardless.
(def- shape-cache @{}) # sorted-keys-tuple -> interned shape descriptor
(defn shape-for
"Interned shape descriptor for an ordered key vector (keys in layout order)."
[keyv]
(def kk (tuple ;keyv))
(or (get shape-cache kk)
(let [idx @{}]
(var i 0) (each k keyv (put idx k i) (++ i))
(def desc (struct :jolt/shape kk :idx (table/to-struct idx)))
(put shape-cache kk desc)
desc)))
(defn shape-rec? [x]
(and (tuple? x) (> (length x) 0)
(struct? (in x 0)) (not (nil? (in (in x 0) :jolt/shape)))))
(defn shape-keys [rec] ((in rec 0) :jolt/shape))
(defn shape-get [rec k default]
(def pos (get ((in rec 0) :idx) k))
(if (nil? pos) default (in rec (+ pos 1))))
(defn shape-assoc [rec k v]
# assoc on a known key keeps the layout; a new key falls back to a struct
(def desc (in rec 0))
(def pos (get (desc :idx) k))
(if (nil? pos)
(let [t @{}] (each kk (desc :jolt/shape) (put t kk (shape-get rec kk nil))) (put t k v) (table/to-struct t))
(let [a (array ;rec)] (put a (+ pos 1) v) (tuple ;a))))
(defn core-assoc [m & kvs]
(when (odd? (length kvs))
(error "assoc expects an even number of key/value arguments"))
@ -504,6 +540,10 @@
(and (struct? m) (get m :jolt/type)))
(error (string "assoc requires a map or vector, got " (type m))))
(cond
(shape-rec? m)
(do (var result m) (var i 0)
(while (< i (length kvs)) (set result (shape-assoc result (kvs i) (kvs (+ i 1)))) (+= i 2))
result)
(core-sorted-map? m) ((sorted-op m :assoc) m kvs)
(phm? m)
(do (var result m) (var i 0) (while (< i (length kvs)) (set result (phm-assoc result (kvs i) (kvs (+ i 1)))) (+= i 2)) result)
@ -560,42 +600,6 @@
(do (var result @{}) (each k (keys m) (var in-ks false) (each k2 ks (if (deep= k k2) (do (set in-ks true) (break)))) (if (not in-ks) (put result k (m k))))
(if (struct? m) (table/to-struct result) result))))
# --- shape records (hidden classes, jolt-t34) -------------------------------
# A "shape record" is a cheap fixed-layout representation for a map literal
# whose keys are a known compile-time set (e.g. a vec3 {:r :g :b}). It is a
# Janet tuple [SHAPE v0 v1 ...] where SHAPE is an interned descriptor struct
# {:jolt/shape KEYS :idx {k->pos}}. Construction is a tuple (≈2x cheaper than a
# struct), const-keyword lookup compiles to an index, and general map ops below
# recognize it via shape-rec? and treat it as a map — so it is transparent
# wherever it flows. Created only by the backend when JOLT_SHAPE is on and the
# inference proves the shape; the runtime support here is always present so a
# shape value is handled correctly regardless.
(def- shape-cache @{}) # sorted-keys-tuple -> interned shape descriptor
(defn shape-for
"Interned shape descriptor for an ordered key vector (keys in layout order)."
[keyv]
(def kk (tuple ;keyv))
(or (get shape-cache kk)
(let [idx @{}]
(var i 0) (each k keyv (put idx k i) (++ i))
(def desc (struct :jolt/shape kk :idx (table/to-struct idx)))
(put shape-cache kk desc)
desc)))
(defn shape-rec? [x]
(and (tuple? x) (> (length x) 0)
(struct? (in x 0)) (not (nil? (in (in x 0) :jolt/shape)))))
(defn shape-keys [rec] ((in rec 0) :jolt/shape))
(defn shape-get [rec k default]
(def pos (get ((in rec 0) :idx) k))
(if (nil? pos) default (in rec (+ pos 1))))
(defn shape-assoc [rec k v]
# assoc on a known key keeps the layout; a new key falls back to a struct
(def desc (in rec 0))
(def pos (get (desc :idx) k))
(if (nil? pos)
(let [t @{}] (each kk (desc :jolt/shape) (put t kk (shape-get rec kk nil))) (put t k v) (table/to-struct t))
(let [a (array ;rec)] (put a (+ pos 1) v) (tuple ;a))))
(defn core-get [m k &opt default]
(default default nil)
(if (nil? m) default