jz 0.5.0 → 0.6.0

package/src/ir.js

@@ -21,19 +21,14 @@
  */
 
 import { ctx, err, inc, PTR, LAYOUT } from './ctx.js'
-import { T, VAL, valTypeOf, lookupValType, repOf, repOfGlobal, objLiteralSchemaId } from './analyze.js'
+import { ptrBoxPrefixBigInt, atomNanHex, nanPrefixHex } from '../layout.js'
+import { I32_MIN, I32_MAX, isI32, isLiteralStr, isFuncRef } from './ast.js'
+import { VAL, lookupValType, repOf, repOfGlobal } from './reps.js'
+import { valTypeOf } from './kind.js'
+import { T } from './ast.js'
+import { objLiteralSchemaId } from './static.js'
 
-// === Numeric range ===
-
-/** Signed-32-bit range. Used everywhere a number value must round-trip through
- *  wasm `i32` (literal constants, default-arg folding, exprType inference). */
-export const I32_MIN = -2147483648
-export const I32_MAX = 2147483647
-
-/** True when `v` is a finite integer that fits in i32 *and* isn't -0 (which i32
- *  cannot represent). Callers that don't care about -0 can compare against
- *  I32_MIN/I32_MAX directly. */
-export const isI32 = (v) => Number.isInteger(v) && v >= I32_MIN && v <= I32_MAX && !Object.is(v, -0)
+export { I32_MIN, I32_MAX, isI32, isLiteralStr, isFuncRef }
 
 // === Type helpers ===
 
@@ -42,9 +37,7 @@ export const typed = (node, type) => (node.type = type, node)
 
 /** NaN-box prefix for a pointer of VAL kind K with aux bits: `0x7FF8 | type<<47 | aux<<32`. */
 function ptrBoxPrefix(ptrType, aux = 0) {
-  return (0x7FF8n << 48n)
-    | ((BigInt(ptrType) & 0xFn) << 47n)
-    | ((BigInt(aux) & 0x7FFFn) << 32n)
+  return ptrBoxPrefixBigInt(ptrType, aux)
 }
 
 /** Build f64 NaN-boxed pointer IR from an i32 offset node of known kind.
@@ -102,7 +95,67 @@ export const asPtrOffset = (n, ptrKind) => {
 }
 
 /** Coerce emitted IR to a target WASM param type ('i32' | 'i64' | 'f64'). */
-export const asParamType = (n, t) => t === 'i32' ? asI32(n) : t === 'i64' ? asI64(n) : asF64(n)
+export const asParamType = (n, t) => t === 'i32' ? asI32(n) : t === 'i64' ? asI64(n) : t === 'v128' ? n : asF64(n)
+
+/**
+ * Narrow an f64 arithmetic tree under ToInt32 — the general int-accumulator path.
+ *
+ * ToInt32 is reduction mod 2^32, and {+, −, ×} form a RING under that modulus:
+ * operands may wrap to i32 eagerly and the final result still equals ToInt32 of
+ * the JS value — PROVIDED the original f64 computation was exact (no rounding).
+ * Exactness is tracked structurally: every interior node's worst-case magnitude
+ * (`maxAbs`, real un-wrapped value) must stay below 2^53. Leaves are peeled
+ * `f64.convert_i32_*` wrappers (≤2^31/2^32) and integer constants.
+ *
+ * `/` is NOT a ring op (fractions): it narrows only at the ToInt32 ROOT, with a
+ * FAITHFUL numerator (i32 value == JS value — wrapped sums excluded) and a
+ * constant integer divisor. i32.div_s truncates toward zero exactly like
+ * ToInt32 of the f64 quotient (error < ulp/2 < distance-to-integer for any i32
+ * numerator); c ∈ {0,−1,1} are diverted (trap / INT_MIN trap / identity).
+ *
+ * Returns {node (i32-typed), maxAbs, faithful} or null — callers use `.node`.
+ */
+const narrowI32 = (x, isRoot) => {
+  if (!Array.isArray(x)) return null
+  if (x.type === 'i32') return { node: x, maxAbs: 2 ** 31, faithful: true }
+  const op = x[0]
+  if (op === 'f64.convert_i32_s' || op === 'f64.convert_i32_u')
+    // Peel — same as toI32's peephole. _u values ∈ [0, 2^32): the re-tag IS the
+    // wrap (ring-compatible), but the i32 view differs from the JS value above
+    // 2^31, so _u is not faithful.
+    return {
+      node: Array.isArray(x[1]) ? typed(x[1], 'i32') : x[1],
+      maxAbs: op === 'f64.convert_i32_s' ? 2 ** 31 : 2 ** 32,
+      faithful: op === 'f64.convert_i32_s',
+    }
+  if (op === 'f64.const' && typeof x[1] === 'number' && Number.isInteger(x[1]) && Math.abs(x[1]) < 2 ** 52)
+    return { node: typed(['i32.const', x[1] | 0], 'i32'), maxAbs: Math.abs(x[1]), faithful: Math.abs(x[1]) < 2 ** 31 }
+  if (op === 'f64.add' || op === 'f64.sub' || op === 'f64.mul') {
+    const a = narrowI32(x[1]), b = narrowI32(x[2])
+    if (!a || !b) return null
+    const maxAbs = op === 'f64.mul' ? a.maxAbs * b.maxAbs : a.maxAbs + b.maxAbs
+    if (maxAbs >= 2 ** 53) return null
+    const iop = op === 'f64.add' ? 'i32.add' : op === 'f64.sub' ? 'i32.sub' : 'i32.mul'
+    return { node: typed([iop, a.node, b.node], 'i32'), maxAbs, faithful: false }
+  }
+  if (op === 'f64.neg') {
+    const a = narrowI32(x[1])
+    if (!a) return null
+    return { node: typed(['i32.sub', ['i32.const', 0], a.node], 'i32'), maxAbs: a.maxAbs, faithful: false }
+  }
+  if (op === 'f64.div' && isRoot) {
+    const a = narrowI32(x[1])
+    if (!a || !a.faithful) return null
+    const c = Array.isArray(x[2]) && x[2][0] === 'f64.const' && typeof x[2][1] === 'number' ? x[2][1] : null
+    if (c == null || !Number.isInteger(c) || c === 0 || c === 1 || Math.abs(c) >= 2 ** 31) return null
+    // c = −1 would trap on INT_MIN; 0 − x wraps INT_MIN → INT_MIN, matching ToInt32(2^31).
+    const node = c === -1
+      ? typed(['i32.sub', ['i32.const', 0], a.node], 'i32')
+      : typed(['i32.div_s', a.node, ['i32.const', c]], 'i32')
+    return { node, maxAbs: 2 ** 31, faithful: c !== -1 }
+  }
+  return null
+}
 
 /** Coerce node to i32 with wrapping (JS `|0` semantics: values > 2^31 wrap to negative).
  *  Per ECMAScript ToInt32, NaN and ±∞ map to 0. `i64.trunc_sat_f64_s` handles NaN
@@ -120,18 +173,22 @@ export const toI32 = n => {
   }
   if (Array.isArray(n) && n[0] === 'f64.const' && typeof n[1] === 'number') {
     const v = n[1]
-    return typed(['i32.const', Number.isFinite(v) ? v | 0 : 0], 'i32')
+    return typed(['i32.const', Number.isFinite(v) ? v | 0 : 0], 'i32')   // JS `|0` is ToInt32
   }
+  // General int-arithmetic narrowing: an exact-int f64 tree of {+,−,×,neg,/C}
+  // computes in i32 (mod-2^32 ring) — no trunc/guard at all.
+  const nw = narrowI32(n, true)
+  if (nw) return nw.node
   // Leaf nodes are cheap to duplicate; for everything else, evaluate once via local.tee.
   const isLeaf = Array.isArray(n) && n.length <= 2 &&
     (n[0] === 'f64.const' || n[0] === 'local.get' || n[0] === 'global.get')
+  // `i32.wrap_i64(i64.trunc_sat_f64_s x)` is exact ToInt32 for |x| < 2^63 (the
+  // overwhelming common range), maps NaN/−∞→0, and +∞ is guarded to 0 by the
+  // select. For |x| ≥ 2^63 it saturates rather than wrapping mod 2^32 — a
+  // deliberately-allowed asm.js-style boundary (no per-`|0` helper/guard cost).
   const wrap = x => typed(['i32.wrap_i64', ['i64.trunc_sat_f64_s', x]], 'i32')
   if (isLeaf) {
-    return typed(['select',
-      wrap(n),
-      ['i32.const', 0],
-      ['f64.ne', n, ['f64.const', Infinity]]
-    ], 'i32')
+    return typed(['select', wrap(n), ['i32.const', 0], ['f64.ne', n, ['f64.const', Infinity]]], 'i32')
   }
   const t = temp('inf')
   return typed(['select',
@@ -166,16 +223,16 @@ export const fromI64 = n => {
  *  See module/symbol.js for the broader reserved-atom-id scheme.
  *  Distinct from 0, NaN, and all pointers. Triggers default params.
  *  At the JS boundary, null and undefined preserve their identity for interop. */
-export const NULL_NAN = '0x' + (LAYOUT.NAN_PREFIX_BITS | (1n << BigInt(LAYOUT.AUX_SHIFT))).toString(16).toUpperCase().padStart(16, '0')
-export const UNDEF_NAN = '0x' + (LAYOUT.NAN_PREFIX_BITS | (2n << BigInt(LAYOUT.AUX_SHIFT))).toString(16).toUpperCase().padStart(16, '0')
+export const NULL_NAN = atomNanHex(1)
+export const UNDEF_NAN = atomNanHex(2)
 /** Boxed-boolean carrier. `false`/`true` are reserved atoms — materialized only
  *  where boolean identity is observed (typeof/String/JSON/host boundary); in
  *  branch/arithmetic position booleans stay raw i32/f64 0/1. The atomId encodes
  *  the truth value in its low bit (4=false, 5=true), so `aux & 1` recovers 0/1
  *  and `4 | bit` boxes it — see boolBoxIR / unboxBoolIR. */
 export const BOOL_ATOM_BASE = 4
-export const FALSE_NAN = '0x' + (LAYOUT.NAN_PREFIX_BITS | (4n << BigInt(LAYOUT.AUX_SHIFT))).toString(16).toUpperCase().padStart(16, '0')
-export const TRUE_NAN = '0x' + (LAYOUT.NAN_PREFIX_BITS | (5n << BigInt(LAYOUT.AUX_SHIFT))).toString(16).toUpperCase().padStart(16, '0')
+export const FALSE_NAN = atomNanHex(4)
+export const TRUE_NAN = atomNanHex(5)
 /** WAT-template-ready sentinel expressions for use in stdlib template strings.
  *  `f64.const nan:0xHEX` is 3 bytes shorter than `f64.reinterpret_i64 (i64.const ...)`. */
 export const NULL_WAT = `(f64.const nan:${NULL_NAN})`
@@ -211,13 +268,20 @@ export function unboxBoolIR(f64expr) {
 
 /** Max arity of inline closure slots. Closures are compiled with signature
  *  (env f64, argc i32, a0..a{MAX-1} f64) → f64 — no per-call heap alloc.
- *  Calls with more args than MAX error; rest-param closures receive at most
- *  MAX rest args when invoked via spread with >MAX dynamic elements. */
+ *  Direct (non-spread) calls with more args than MAX error. Spread calls are
+ *  unbounded: the spread site publishes the full args-array offset in
+ *  $__closure_spill, and a rest-param callee reads args[MAX..argc-1] from it
+ *  (see module/function.js spread path + compile/index.js rest collection). */
 export const MAX_CLOSURE_ARITY = 8
 
 /** Matches WASM instructions that require a memory section. */
-// FIXME: can be simpler regex, just test second part - load vs store vs grow vs size vs memory
-export const MEM_OPS = /\b(i32\.load|i32\.store|f64\.load|f64\.store|f32\.load|f32\.store|i64\.load|i64\.store|memory\.size|memory\.grow|i32\.load8|i32\.load16|i32\.store8|i32\.store16)\b/
+// Any instruction that touches linear memory ⇒ the module must declare memory.
+// Matches every `memory.*` op (size/grow/copy/fill/init) and every typed load/store
+// incl. width suffixes (load8_u, store16, i64.load32_s, v128.load, …). The old
+// hand-enumerated list silently missed memory.copy/fill, v128.load/store and
+// i64.store8/16/32 (all used in stdlib) — a body using only those would wrongly
+// report no-memory. Broad-but-precise: only `memory.` and `<type>.load|store` match.
+export const MEM_OPS = /\b(memory\.\w+|(?:i32|i64|f32|f64|v128)\.(?:load|store)\w*)\b/
 
 export const WASM_OPS = new Set(['block','loop','if','then','else','br','br_if','call','call_indirect','return','return_call','throw','try_table','catch','nop','drop','unreachable','select','result','mut','param','func','module','memory','table','elem','data','type','import','export','local','global','ref'])
 export const SPREAD_MUTATORS = new Set(['push', 'add', 'set', 'unshift'])
@@ -250,15 +314,22 @@ export function mkPtrIR(type, aux, offset) {
   return typed(['call', '$__mkptr', tIR, aIR, oIR], 'f64')
 }
 
-/** Offset extraction for a NaN-boxed pointer, specialized on known value type.
- *  For non-ARRAY pointer kinds we skip `__ptr_offset` entirely — arrays are the
- *  only type that can forward on reallocation, everyone else returns the raw low 32 bits.
+/** Offset extraction for a NaN-boxed pointer.
+ *  Goes through `__ptr_offset`, which chases the relocation-forwarding chain
+ *  (cap == -1 sentinel at off-4 → relocated offset at off-8). The chase is a
+ *  single load+compare for any live (non-forwarded) header, so it is a no-op for
+ *  fixed-shape receivers (OBJECT/TYPED/…) whose cap word is never -1.
+ *
+ *  We do NOT skip it for "non-ARRAY" static types: that shortcut was unsound on
+ *  two counts. (1) ARRAY is not the only growable container — HASH/SET/MAP relocate
+ *  too. (2) jz value types are not always precise: a binding inferred OBJECT (a
+ *  polymorphic parameter, a widened union) can hold a relocated ARRAY at runtime.
+ *  Writing through its stale pre-relocation base then clobbers whatever now occupies
+ *  that freed region — a memory-safety hazard that must not depend on inference
+ *  precision. Memory safety is unconditional; the forwarding follow stays.
  *  If the node is already an unboxed pointer (ptrKind), return it directly. */
 export function ptrOffsetIR(valIR, valType) {
   if (valIR.ptrKind != null && valIR.ptrKind !== VAL.ARRAY) return valIR
-  if (valType != null && valType !== VAL.ARRAY) {
-    return ['i32.wrap_i64', ['i64.reinterpret_f64', valIR]]
-  }
   inc('__ptr_offset')
   return ['call', '$__ptr_offset', ['i64.reinterpret_f64', valIR]]
 }
@@ -282,48 +353,73 @@ export function ptrTypeIR(valIR, valType) {
     ['i64.const', 0xF]]]
 }
 
+// SELF-HOST CONTRACT: f64 slot BITS travel as canonical '0x'+16-hex STRINGS.
+// A BigInt crossing a function return / array element / object slot is
+// kind-erased in the kernel (raw i64 bits are untagged) and every subsequent
+// op on it misdispatches; BigInt64Array/BigUint64Array views and
+// DataView.{get,set}BigUint64 are a legacy f64-value shim there. Strings are
+// tagged and survive every boundary; BigInt math happens only inside single
+// expressions. (Same contract as wat/optimize.js's i64 VALUE CONTRACT.)
 const _F64_BITS_BUF = new ArrayBuffer(8)
 const _F64_BITS_F = new Float64Array(_F64_BITS_BUF)
-const _F64_BITS_U = new BigUint64Array(_F64_BITS_BUF)
+const _F64_BITS_U32 = new Uint32Array(_F64_BITS_BUF)  // LE halves: [0]=lo, [1]=hi
+const _hx8 = (u) => (u >>> 0).toString(16).padStart(8, '0')
 
 /** Return i64 bit pattern (BigInt) of a pure-literal IR node, or null if non-literal. */
 export function extractF64Bits(node) {
   if (!Array.isArray(node)) return null
   if (node[0] === 'f64.const') {
-    if (typeof node[1] === 'number') { _F64_BITS_F[0] = node[1]; return _F64_BITS_U[0] }
+    if (typeof node[1] === 'number') { _F64_BITS_F[0] = node[1]; return '0x' + _hx8(_F64_BITS_U32[1]) + _hx8(_F64_BITS_U32[0]) }
     if (typeof node[1] === 'string' && node[1].startsWith('nan:')) {
-      try { return BigInt(node[1].slice(4)) | 0x7FF0000000000000n } catch { return null }
+      try {
+        const v = BigInt(node[1].slice(4)) | 0x7ff0000000000000n
+        return '0x' + v.toString(16).padStart(16, '0')
+      } catch { return null }
     }
     return null
   }
   if (node[0] === 'f64.reinterpret_i64' && Array.isArray(node[1]) && node[1][0] === 'i64.const' && typeof node[1][1] === 'string') {
     const s = node[1][1]
     if (s.startsWith('-')) {
-      const abs = s.slice(1)
-      try { return ((1n << 64n) - BigInt(abs)) & 0xFFFFFFFFFFFFFFFFn } catch { return null }
+      // Two's complement WITHOUT a 2^64 term: (-1 − |v|) + 1 ≡ 2^64 − |v| both
+      // natively and on the kernel's mod-2^64 carrier (1n<<64n is unrepresentable
+      // there and would silently corrupt).
+      try {
+        const v = (0xffffffffffffffffn - BigInt(s.slice(1)) + 1n) & 0xffffffffffffffffn
+        return '0x' + v.toString(16).padStart(16, '0')
+      } catch { return null }
     }
-    try { return BigInt(s) } catch { return null }
+    try {
+      const v = BigInt(s)
+      return '0x' + v.toString(16).padStart(16, '0')
+    } catch { return null }
   }
   return null
 }
 
-/** Append `slots` (BigInt i64 each) to ctx.runtime.data 8-byte aligned, return raw byte offset of first slot.
- *  Slots that look like NaN-boxed pointers are recorded in `ctx.runtime.staticPtrSlots` so the
- *  prefix-strip pass can patch their embedded offsets. */
+/** Append `slots` ('0x'+16-hex bit strings, see contract above) to
+ *  ctx.runtime.data 8-byte aligned, return raw byte offset of first slot.
+ *  Slots that look like NaN-boxed pointers are recorded in
+ *  `ctx.runtime.staticPtrSlots` so the prefix-strip pass can patch their
+ *  embedded offsets. Writes go through u32 halves — DataView's BigInt
+ *  accessors are unfaithful in the self-host kernel. */
 export function appendStaticSlots(slots, headerBytes = 0) {
   if (!ctx.runtime.data) ctx.runtime.data = ''
   while (ctx.runtime.data.length % 8 !== 0) ctx.runtime.data += '\0'
   const off = ctx.runtime.data.length
   const u8 = new Uint8Array(headerBytes + slots.length * 8)
   const dv = new DataView(u8.buffer)
-  for (let i = 0; i < slots.length; i++) dv.setBigUint64(headerBytes + i * 8, slots[i], true)
+  for (let i = 0; i < slots.length; i++) {
+    const h = slots[i]
+    dv.setUint32(headerBytes + i * 8, parseInt(h.slice(10), 16) >>> 0, true)
+    dv.setUint32(headerBytes + i * 8 + 4, parseInt(h.slice(2, 10), 16) >>> 0, true)
+  }
   let chunk = ''
   for (let i = 0; i < u8.length; i++) chunk += String.fromCharCode(u8[i])
   ctx.runtime.data += chunk
   if (!ctx.runtime.staticPtrSlots) ctx.runtime.staticPtrSlots = []
   for (let i = 0; i < slots.length; i++) {
-    const bits = slots[i]
-    if (((bits >> 48n) & 0xFFF8n) === BigInt(LAYOUT.NAN_PREFIX)) {
+    if ((parseInt(slots[i].slice(2, 6), 16) & 0xFFF8) === LAYOUT.NAN_PREFIX) {
       ctx.runtime.staticPtrSlots.push(off + i * 8)
     }
   }
@@ -335,8 +431,8 @@ export function appendStaticSlots(slots, headerBytes = 0) {
 /** Check if emitted node is a compile-time constant. */
 export const isLit = n => (n[0] === 'i32.const' || n[0] === 'f64.const') && typeof n[1] === 'number'
 export const litVal = n => n[1]
-const isNullLit = n => Array.isArray(n) && n.length === 2 && n[0] == null && n[1] == null
-const isUndefLit = n => Array.isArray(n) && n.length === 0
+export const isNullLit = n => Array.isArray(n) && n.length === 2 && n[0] == null && n[1] == null
+export const isUndefLit = n => Array.isArray(n) && n.length === 0
 export const isNullishLit = n => isNullLit(n) || isUndefLit(n)
 
 /** Side-effect-free (safe for WASM select). */
@@ -348,46 +444,173 @@ const PURE_OPS = new Set(['i32.const', 'f64.const', 'local.get', 'global.get',
   'i32.eq', 'i32.ne', 'i32.lt_s', 'i32.gt_s', 'i32.le_s', 'i32.ge_s', 'i32.eqz'])
 export const isPureIR = n => Array.isArray(n) && PURE_OPS.has(n[0]) && n.slice(1).every(c => !Array.isArray(c) || isPureIR(c))
 
-/** Check if AST node is a literal string: ['str', 'value']. */
-export const isLiteralStr = idx => Array.isArray(idx) && idx[0] === 'str' && typeof idx[1] === 'string'
+/** Ops whose f64 result is always a plain number (never a NaN-boxed pointer).
+ *  Used by toNumF64 to skip the __to_num wrapper when the value is provably numeric.
+ *  NOTE: f64.const is NOT included — it may encode a NaN-boxed pointer. */
+const PURE_F64_OPS = new Set([
+  'f64.add', 'f64.sub', 'f64.mul', 'f64.div', 'f64.neg', 'f64.abs', 'f64.sqrt',
+  'f64.min', 'f64.max', 'f64.ceil', 'f64.floor', 'f64.trunc', 'f64.nearest', 'f64.copysign',
+  'f64.convert_i32_s', 'f64.convert_i32_u', 'f64.promote_f32',
+])
+
+/** True iff `r` provably yields a plain f64 NUMBER (never a NaN-boxed pointer or
+ *  nullish sentinel). A `block`/`if` is numeric only when its value-producing tail
+ *  is — so `o.a?.b` (a block whose result is a property value or undef sentinel)
+ *  is correctly NOT numeric, while `cond ? n*2 : n*3` is. Conservative: any shape
+ *  not provably numeric (property gets, user calls, local.get, f64.const nan:…)
+ *  returns false, so the caller keeps the __to_num coercion. */
+const isNumericIR = (r) => {
+  if (!Array.isArray(r)) return false
+  const op = r[0]
+  if (PURE_F64_OPS.has(op)) return true
+  if (op === 'call' && typeof r[1] === 'string' && (r[1].startsWith('$math.') || r[1] === '$__time_ms')) return true
+  if (op === 'f64.const') return typeof r[1] === 'number'   // 'nan:…' carrier ⇒ pointer/sentinel
+  if (op === 'block') return isNumericIR(r[r.length - 1])   // block value = its tail expr
+  if (op === 'if') {                                        // both arms must be numeric
+    const thenArm = r.find(x => Array.isArray(x) && x[0] === 'then')
+    const elseArm = r.find(x => Array.isArray(x) && x[0] === 'else')
+    return !!thenArm && !!elseArm &&
+      isNumericIR(thenArm[thenArm.length - 1]) && isNumericIR(elseArm[elseArm.length - 1])
+  }
+  return false
+}
 
-/** Resolve compile-time value type from AST node (literal → type, name → lookup). */
+/** Resolve compile-time value type from AST node (literal → name → lookup). */
 export const resolveValType = (node, valTypeOf, lookupValType) =>
   valTypeOf(node) ?? (typeof node === 'string' ? lookupValType(node) : null)
 
-/** Check if AST node is a string reference to a known function name. */
-export const isFuncRef = (node, funcNames) => typeof node === 'string' && funcNames.has(node)
-
 /** Check if (a, op, b) is a postfix pattern: [op, name] and [, 1] literal. */
 export const isPostfix = (a, op, b) => Array.isArray(a) && a[0] === op && Array.isArray(b) && b[0] == null && b[1] === 1
 
 /** Emit a numeric constant with correct i32/f64 typing.
  *  `-0` is f64-only (i32 has no signed zero) — preserve the sign by emitting f64. */
 export const emitNum = v => isI32(v)
-  ? typed(['i32.const', v], 'i32') : typed(['f64.const', v], 'f64')
+  ? typed(['i32.const', v], 'i32')
+  // Emit NaN via the `nan` token, not the raw JS number: a numeric NaN literal in
+  // the IR loses its quiet-mantissa bit (0x7FF8→0x7FF0, i.e. becomes Infinity) when
+  // the self-host kernel marshals the IR back across the wasm→host boundary. The
+  // `nan` token assembles to the canonical 0x7FF8 number-NaN unambiguously.
+  : typed(['f64.const', v !== v ? 'nan' : v], 'f64')
 
 // === Temp locals ===
 
-/** Allocate a temp local, returns name without $. Optional tag aids WAT readability.
- *  Skips names already registered (by analyzeBody().locals from prepare-generated
- *  names) to avoid collisions that would silently override the pre-analyzed type. */
-export function temp(tag = '') {
+/** Allocate a fresh local name with the given tag, registered as `type`. The
+ *  selfhost compiler doesn't yet handle exported-const arrow factories returning
+ *  closures, so the three temp() helpers stay as `function` declarations and
+ *  delegate to this shared core. */
+function freshLocal(type, tag) {
   let name
   do { name = `${T}${tag}${ctx.func.uniq++}` } while (ctx.func.locals.has(name))
-  ctx.func.locals.set(name, 'f64')
+  ctx.func.locals.set(name, type)
   return name
 }
-export function tempI32(tag = '') {
-  let name
-  do { name = `${T}${tag}${ctx.func.uniq++}` } while (ctx.func.locals.has(name))
-  ctx.func.locals.set(name, 'i32')
-  return name
+export function temp    (tag = '') { return freshLocal('f64', tag) }
+export function tempI32 (tag = '') { return freshLocal('i32', tag) }
+export function tempI64 (tag = '') { return freshLocal('i64', tag) }
+
+// === IR scaffolds ===
+
+/** Wrap a sequence of statements as a typed `(block (result <type>) …)`.
+ *  Default result is `f64` (the value-type for most jz emissions).
+ *  Shorthand for the `typed(['block', ['result', T], …stmts], T)` pattern that
+ *  appears in nearly every emitter — keeps call sites focused on the body. */
+export const block64 = (...stmts) => typed(['block', ['result', 'f64'], ...stmts], 'f64')
+export const blockTyped = (type, ...stmts) => typed(['block', ['result', type], ...stmts], type)
+
+/** Allocate an f64 temp, set it to `val`, run `body(name)` and yield its result.
+ *  `body` may return either a single IR node (used as the block result) or an
+ *  array of nodes whose last expression becomes the result. Eliminates the
+ *  repetitive `const t = temp(); …['local.set', $t, val]; …['local.get', $t]`
+ *  scaffold around tee-and-use patterns. */
+export function withTemp(val, body, tag = '') {
+  const t = temp(tag)
+  const out = body(t)
+  const tail = Array.isArray(out) && out.every(n => Array.isArray(n)) ? out : [out]
+  return block64(['local.set', `$${t}`, val], ...tail)
 }
-export function tempI64(tag = '') {
-  let name
-  do { name = `${T}${tag}${ctx.func.uniq++}` } while (ctx.func.locals.has(name))
-  ctx.func.locals.set(name, 'i64')
-  return name
+
+/** Whole-fn structural refcount: walks `fn`, counting how many times each
+ *  array node is referenced. Used by optimizer passes to skip shared subtrees
+ *  (watr CSE may leave them) — mutating a node with refcount > 1 would also
+ *  affect references outside the current region. Single-pass O(N). */
+export function buildRefcount(fn) {
+  const refcount = new Map()
+  const walk = (node) => {
+    if (!Array.isArray(node)) return
+    const n = (refcount.get(node) || 0) + 1
+    refcount.set(node, n)
+    if (n > 1) return  // already counted children below
+    for (let i = 0; i < node.length; i++) walk(node[i])
+  }
+  walk(fn)
+  return refcount
+}
+
+/** Pick the next free `$__<prefix><id>` local-name id by collecting all
+ *  existing ids in a single walk. Replaces the per-pass
+ *  `while (fn.some(... === $__prefixK)) k++` (O(K·N)) with one O(N) scan. */
+export function nextLocalId(fn, prefix) {
+  const seen = new Set()
+  const needle = `$__${prefix}`
+  const walk = (n) => {
+    if (!Array.isArray(n)) return
+    if (n[0] === 'local' && typeof n[1] === 'string' && n[1].startsWith(needle)) {
+      const tail = n[1].slice(needle.length)
+      if (/^\d+$/.test(tail)) seen.add(+tail)
+    }
+    for (let i = 0; i < n.length; i++) walk(n[i])
+  }
+  walk(fn)
+  let id = 0
+  while (seen.has(id)) id++
+  return id
+}
+
+/** Single-kind ptr-tag predicate: `__ptr_type(bits) == ptr`. Takes the f64
+ *  carrier expression and the PTR constant. Use this when guarding one branch;
+ *  use `dispatchByPtrType` for multi-case forks. Stamps `inc('__ptr_type')`. */
+export function ptrTypeEq(f64Expr, ptr) {
+  inc('__ptr_type')
+  return typed(['i32.eq', ['call', '$__ptr_type', ['i64.reinterpret_f64', f64Expr]], ['i32.const', ptr]], 'i32')
+}
+
+/** ToPrimitive sidecar probe (ES2024 7.1.1): an own `valueOf`/`toString` data
+ *  property shadows the builtin. Reads the dynamic-prop sidecar slot keyed by
+ *  `nameIR` (an emitted i64 string key) off receiver `objIR`; if it holds a
+ *  closure, yields `onOverride($p)`, else `onFallback($o)` (both f64). Shared by
+ *  the member-READ path (module/core.js — onOverride returns the closure value,
+ *  onFallback calls the arity-≤1 builtin) and the method-CALL path (emit.js —
+ *  onOverride invokes the closure, onFallback calls the builtin method). */
+export function sidecarOverride(objIR, nameIR, onOverride, onFallback) {
+  const o = temp('vo'), p = temp('vp')
+  inc('__dyn_get_expr', '__ptr_type')
+  return block64(
+    ['local.set', `$${o}`, asF64(objIR)],
+    ['local.set', `$${p}`, ['f64.reinterpret_i64',
+      ['call', '$__dyn_get_expr', ['i64.reinterpret_f64', ['local.get', `$${o}`]], nameIR]]],
+    ['if', ['result', 'f64'],
+      ptrTypeEq(['local.get', `$${p}`], PTR.CLOSURE),
+      ['then', onOverride(p, o)],
+      ['else', onFallback(o)]])
+}
+
+/** Dispatch on `__ptr_type(bits)` — emits a right-leaning if/else chain over
+ *  PTR constants. `cases` is `[[PTR.X, ir], …]`; `fallback` is the else IR.
+ *  `resultType` defaults to `'f64'`; pass `null` for a void dispatch (e.g.
+ *  pure memory-writing branches). Centralizes the
+ *  `i32.eq (call $__ptr_type bits) (i32.const PTR.X)` pattern so emitters
+ *  dispatching by pointer kind stay declarative. */
+export function dispatchByPtrType(typeLocal, cases, fallback, resultType = 'f64') {
+  let out = fallback
+  const head = resultType ? ['if', ['result', resultType]] : ['if']
+  for (let i = cases.length - 1; i >= 0; i--) {
+    const [ptr, ir] = cases[i]
+    out = [...head,
+      ['i32.eq', ['local.get', `$${typeLocal}`], ['i32.const', ptr]],
+      ['then', ir],
+      ['else', out]]
+  }
+  return out
 }
 
 // === Numeric helpers ===
@@ -395,23 +618,17 @@ export function tempI64(tag = '') {
 /** WASM has no f64.rem — implement as a - trunc(a/b) * b.
  *  Both `a` and `b` appear twice in the expansion; cache non-pure operands
  *  in locals so side effects (e.g. assignments) only execute once. */
-export const f64rem = (a, b) => {
-  const pa = isPureIR(a), pb = isPureIR(b)
-  if (pa && pb) return typed(['f64.sub', a, ['f64.mul', ['f64.trunc', ['f64.div', a, b]], b]], 'f64')
-  const ta = pa ? null : temp(), tb = pb ? null : temp()
-  const ga = pa ? a : ['local.get', `$${ta}`], gb = pb ? b : ['local.get', `$${tb}`]
-  const pre = []
-  if (!pa) pre.push(['local.set', `$${ta}`, a])
-  if (!pb) pre.push(['local.set', `$${tb}`, b])
-  return typed(['block', ['result', 'f64'], ...pre,
-    ['f64.sub', ga, ['f64.mul', ['f64.trunc', ['f64.div', ga, gb]], gb]]], 'f64')
-}
+// JS `%` on the f64 path. Delegates to the exact `__rem` (binary fmod) stdlib —
+// the textbook `a - b*trunc(a/b)` is inexact for large a/b and wrong on the
+// ±Inf / 0 / NaN edges. The i32.rem_s fast path in emit.js handles the common
+// integer-with-nonzero-literal-divisor case; everything else lands here.
+export const f64rem = (a, b) => (inc('__rem'), typed(['call', '$__rem', a, b], 'f64'))
 
 /** Resolve the slot index of a ToPrimitive method (`valueOf`/`toString`) on an
  *  OBJECT operand — from a schema-bound variable or an inline object literal.
  *  Returns -1 when the method is absent. */
 function primMethodIdx(node, name) {
-  if (typeof node === 'string') return ctx.schema.find(node, name)
+  if (typeof node === 'string') return ctx.schema.slotOf(node, name)
   const sid = objLiteralSchemaId(node)
   const props = sid != null ? ctx.schema.list[sid] : null
   return props ? props.indexOf(name) : -1
@@ -448,18 +665,40 @@ function toPrimitiveChain(node, v, order) {
   return typed(['block', blk, ...body], 'i64')
 }
 
+const cloneIR = (n) => Array.isArray(n) ? n.map(cloneIR) : n
+
+/** ToNumber for a runtime value that may carry a nullish sentinel: null→+0, undefined→NaN,
+ *  anything else → itself. `valIR` must be side-effect-free (a local read) — it is duplicated,
+ *  so each occurrence gets a fresh clone. Used for bindings flagged in ctx.func.maybeNullish;
+ *  a real number isn't either sentinel, so it falls through the `else` unchanged. */
+const coerceNullishToNum = (valIR) => typed(
+  ['if', ['result', 'f64'],
+    ['i64.eq', ['i64.reinterpret_f64', cloneIR(valIR)], ['i64.const', NULL_NAN]],
+    ['then', ['f64.const', 0]],
+    ['else', ['if', ['result', 'f64'],
+      ['i64.eq', ['i64.reinterpret_f64', cloneIR(valIR)], ['i64.const', UNDEF_NAN]],
+      ['then', ['f64.const', 'nan']],
+      ['else', cloneIR(valIR)]]]],
+  'f64')
+
 /** Coerce an emitted IR value to a plain f64 Number per JS `ToNumber`.
  *  Skips coercion when static type proves the value is already numeric
  *  (i32 node, compile-time literal, known VAL.NUMBER/VAL.BIGINT). When the full
  *  string-parsing `__to_num` isn't loaded (no string module → no strings can
  *  exist) nullish *literals* still fold statically (null→+0, undefined→NaN);
- *  non-literal values pass through uncoerced. */
+ *  non-literal values pass through uncoerced — except bindings flagged
+ *  maybeNullish, which get a runtime nullish coerce (null-flow correctness). */
 export function toNumF64(node, v) {
   // An i32 node carrying `.ptrKind` is an *unboxed pointer* (object/array local),
   // not a number — skipping coercion would reinterpret pointer bits as an f64.
   // Only a plain i32 (loop counter, `x|0`) is genuinely already-numeric.
   if ((v.type === 'i32' && v.ptrKind == null) || isLit(v)) return asF64(v)
-  const vt = keyValType(node)
+  // A binding assigned a nullish literal may hold null/undefined here — coerce per ToNumber
+  // (null→+0, undefined→NaN); a real number falls through unchanged. Only flagged bindings pay
+  // this, so the numeric kernels jz optimizes for (which never assign null) stay untouched.
+  if (typeof node === 'string' && ctx.func.maybeNullish?.has(node)) return coerceNullishToNum(asF64(v))
+  const vt = valTypeOf(node)
+  if (vt === VAL.BOOL) return typed(['f64.convert_i32_s', truthyIR(v)], 'f64')
   if (vt === VAL.NUMBER || vt === VAL.BIGINT) return asF64(v)
   if (vt === VAL.DATE) {
     const ptr = v.ptrKind === VAL.DATE
@@ -473,7 +712,7 @@ export function toNumF64(node, v) {
   // yield non-primitives a TypeError is thrown. The chosen primitive still
   // flows through `__to_num` so a string return ("−7") is parsed. An abrupt
   // completion (throwing method) propagates through the closure call.
-  if (vt === VAL.OBJECT && ctx.closure.call && ctx.schema.find) {
+  if (vt === VAL.OBJECT && ctx.closure.call && ctx.schema.slotOf) {
     const prim = toPrimitiveChain(node, v, ['valueOf', 'toString'])
     if (prim) {
       // No `__to_num` helper → the program provably has no strings, so the
@@ -497,6 +736,19 @@ export function toNumF64(node, v) {
     if (v[0] === 'f64.convert_i32_s' || v[0] === 'f64.convert_i32_u') return v
     if (v[0] === 'call' && v[1] === '$__time_ms') return v
   }
+  // f64 arithmetic ops and math intrinsics never produce NaN-boxed pointers — the
+  // result is always a plain f64 number. Skip __to_num for these, eliminating the
+  // call overhead that dominates tight numeric kernels (floatbeats, matrix loops).
+  // A `block`/`if` qualifies only when its value-producing tail is provably numeric
+  // (`isNumericIR`): `cond ? n*2 : n*3` skips, but `o.a?.b` (block yielding a
+  // property value / undef sentinel) does NOT — else `o.a?.b > 6` would compare the
+  // boxed string's NaN bits (NaN > 6 → false). User function calls are excluded too
+  // (may return dynamic-property strings); only $math.* is provably numeric.
+  if (v.type === 'f64' && Array.isArray(v) && (
+    PURE_F64_OPS.has(v[0]) ||
+    (v[0] === 'call' && typeof v[1] === 'string' && v[1].startsWith('$math.')) ||
+    ((v[0] === 'block' || v[0] === 'if') && isNumericIR(v))
+  )) return v
   if (!ctx.core.stdlib['__to_num']) {
     // No full ToNumber helper loaded — the program provably has no strings.
     // A nullish *literal* still coerces (null→+0, undefined→NaN) — fold it
@@ -525,8 +777,8 @@ export function toNumF64(node, v) {
  *  `__to_str` so a numeric return is rendered. A throwing method propagates as
  *  an abrupt completion through the closure call. */
 export function toStrI64(node, v) {
-  const vt = keyValType(node)
-  if (vt === VAL.OBJECT && ctx.closure.call && ctx.schema.find) {
+  const vt = valTypeOf(node)
+  if (vt === VAL.OBJECT && ctx.closure.call && ctx.schema.slotOf) {
     const prim = toPrimitiveChain(node, v, ['toString', 'valueOf'])
     if (prim) {
       inc('__to_str')
@@ -540,8 +792,37 @@ export function toStrI64(node, v) {
 /** Convert already-emitted WASM node to i32 boolean. NaN is falsy (like JS).
  *  Peepholes: i32 → as-is; `f64.convert_i32_*(x)` → x (i32 conversion never NaN);
  *  nested `__is_truthy(x)` → x (already 0/1); literal f64 const folds to 0/1. */
+// f64 ops whose result is always a plain NUMBER (never a NaN-boxed carrier) and can
+// be NaN — their truthiness must test NaN by value, not by bit pattern (see truthyIR).
+const NUM_F64_TRUTHY_OPS = new Set([
+  'f64.add', 'f64.sub', 'f64.mul', 'f64.div', 'f64.neg', 'f64.abs', 'f64.sqrt',
+  'f64.min', 'f64.max', 'f64.ceil', 'f64.floor', 'f64.trunc', 'f64.nearest', 'f64.copysign',
+])
+
+const numericTruthy = e => {
+  const t = temp('tb')
+  const g = () => typed(['local.get', `$${t}`], 'f64')
+  return typed(['block', ['result', 'i32'],
+    ['local.set', `$${t}`, e],
+    ['i32.and', ['f64.ne', g(), ['f64.const', 0]], ['f64.eq', g(), g()]]], 'i32')
+}
+
+// i32 ops whose result is already a 0/1 boolean (comparisons + eqz) — safe to use
+// directly as a truthiness without a redundant `!= 0`.
+const I32_BOOL_OPS = new Set(['i32.eq', 'i32.ne', 'i32.lt_s', 'i32.lt_u', 'i32.gt_s', 'i32.gt_u',
+  'i32.le_s', 'i32.le_u', 'i32.ge_s', 'i32.ge_u', 'i32.eqz'])
+
 export function truthyIR(e) {
-  if (e.type === 'i32') return e
+  // An i32 *constant* is a concrete number, not a known 0/1 boolean — fold it to its
+  // truthiness (nonzero → 1).
+  if (Array.isArray(e) && e[0] === 'i32.const') return typed(['i32.const', e[1] ? 1 : 0], 'i32')
+  if (e.type === 'i32') {
+    // A comparison/eqz result is already 0/1 → use directly. Any *other* i32 may be a
+    // concrete narrowed integer (e.g. `Boolean(n)` where n is an i32 number), which is
+    // NOT a 0/1 boolean — normalize via `!= 0` so its truthiness is correct.
+    if (Array.isArray(e) && I32_BOOL_OPS.has(e[0])) return e
+    return typed(['i32.ne', e, ['i32.const', 0]], 'i32')
+  }
   // Unboxed pointer offsets: truthy iff non-zero offset.
   if (e.ptrKind != null) return typed(['i32.ne', e, ['i32.const', 0]], 'i32')
   if (Array.isArray(e)) {
@@ -563,7 +844,7 @@ export function truthyIR(e) {
     // all other NaN-boxed pointers (SSO strings, heap ptrs, etc.) are truthy.
     if (e[0] === 'f64.reinterpret_i64' && Array.isArray(e[1]) && e[1][0] === 'i64.const') {
       const bits = String(e[1][1])
-      const FALSY = new Set([UNDEF_NAN, NULL_NAN, FALSE_NAN, '0x7FF8000000000000', '0x7FFA400000000000'])
+      const FALSY = new Set([UNDEF_NAN, NULL_NAN, FALSE_NAN, nanPrefixHex(), '0x7FFA400000000000'])
       return typed(['i32.const', FALSY.has(bits) ? 0 : 1], 'i32')
     }
     // Fresh pointer constructors never produce nullish. Treat as always truthy.
@@ -581,8 +862,25 @@ export function truthyIR(e) {
           vt === VAL.CLOSURE || vt === VAL.TYPED || vt === VAL.BUFFER || vt === VAL.REGEX || vt === VAL.DATE) {
         return typed(['i32.eqz', isNullish(e)], 'i32')
       }
+      // A plain NUMBER is truthy iff non-zero AND not NaN. `f64.eq x x` tests NaN by
+      // VALUE (false for ANY NaN bits), so this is correct on every platform — unlike
+      // __is_truthy, which bit-compares the canonical number-NaN and so mis-reads
+      // x86's sign-set 0xFFF8.. NaN (from f64.div(0,0) / %) as a truthy box. (local.get
+      // is pure → duplicated, not teed.) Bigint carriers are reinterpret/i64 shapes
+      // and never reach here as VAL.NUMBER.
+      if (vt === VAL.NUMBER) {
+        const g = () => typed(['local.get', e[1]], 'f64')
+        return typed(['i32.and', ['f64.ne', g(), ['f64.const', 0]], ['f64.eq', g(), g()]], 'i32')
+      }
     }
+    // Direct number-producing f64 expression (arithmetic, or the `%` / __rem helper):
+    // same NaN-safe test, single-evaluated through a temp (the value may be a call).
+    if (NUM_F64_TRUTHY_OPS.has(e[0]) || (e[0] === 'call' && e[1] === '$__rem')) return numericTruthy(e)
   }
+  // Composite IR tagged by emit as a definite NUMBER. Use value-based NaN
+  // truthiness; opaque f64 carriers (strings/objects/bigints/nullish/booleans)
+  // remain on __is_truthy so NaN-boxed payloads stay truthy/falsy by tag.
+  if (e.valKind === VAL.NUMBER) return numericTruthy(e)
   inc('__is_truthy')
   return typed(['call', '$__is_truthy', asI64(e)], 'i32')
 }
@@ -590,10 +888,6 @@ export const toBoolFromEmitted = truthyIR
 
 // === Value-type classification ===
 
-export function keyValType(node) {
-  return typeof node === 'string' ? lookupValType(node) : valTypeOf(node)
-}
-
 export function usesDynProps(vt) {
   return vt === VAL.ARRAY || vt === VAL.STRING || vt === VAL.CLOSURE
     || vt === VAL.TYPED || vt === VAL.SET || vt === VAL.MAP || vt === VAL.REGEX
@@ -607,15 +901,27 @@ export function needsDynShadow(target) {
   // access (fn.parse, i32.parse aliases) sees the same value as schema slots.
   const vt = typeof target === 'string' ? (ctx.func.localReps?.get(target)?.val || ctx.scope.globalValTypes?.get(target)) : null
   if (vt === 'closure' || usesDynProps(vt)) return true
+  // A module-wide dynamic-key access (`obj[expr]`) means ANY object may later be
+  // read through the dyn-props hash (__dyn_get_any), so every object literal is
+  // built with a shadow. Mutation sites (Object.assign, `o.k = v`) must mirror
+  // into that same shadow or a subsequent hash read returns a stale slot value.
+  // Honor anyDynKey for NAMED targets too — not just anonymous (target == null)
+  // literals — so construct-time shadowing and mutate-time mirroring agree. They
+  // desynced before: a named literal shadowed via anyDynKey, but its assign saw
+  // only dynKeyVars (which holds the *dynamically-keyed* vars, not this binding).
+  if (ctx.types?.anyDynKey) return true
   const dyn = ctx.types?.dynKeyVars
-  if (target == null) return ctx.types?.anyDynKey ?? false
-  return dyn ? dyn.has(target) : false
+  return target != null && dyn ? dyn.has(target) : false
 }
 
 // === Variable storage abstraction ===
 // Centralizes the boxed/global/local 3-way dispatch (used by =, ++/--, +=, etc.)
 
 /** Check if name is a module-scope global (not shadowed by local/param). */
+/** Bound in the current function frame — a declared local or a parameter. */
+export const isBoundName = name =>
+  ctx.func.locals?.has(name) || ctx.func.current?.params?.some(p => p.name === name)
+
 export function isGlobal(name) {
   return ctx.scope.globals.has(name) && !ctx.func.locals?.has(name) && !ctx.func.current?.params?.some(p => p.name === name)
 }
@@ -630,16 +936,43 @@ export function boxedAddr(name) {
   return ['local.get', `$${ctx.func.boxed.get(name)}`]
 }
 
+// '$'-prefixed name memo. readVar/writeVar run per IR node; rebuilding the
+// `$name` string each time costs an alloc+copy in the self-host kernel AND
+// produces a fresh instance per use — making watr's name-keyed lookups
+// content-compare. The memo returns ONE canonical instance per name, so
+// construction is a map hit and every downstream comparison is bit-eq.
+// Module-level: in-kernel it lives per instance (arena strings are immortal),
+// natively it is a plain cross-compile cache; the name vocabulary is bounded.
+const DOLLAR = new Map()
+export const dollar = (name) => {
+  let v = DOLLAR.get(name)
+  if (v === undefined) { v = '$' + name; DOLLAR.set(name, v) }
+  return v
+}
+
 /** Read variable value: boxed → f64.load, global → global.get, local → local.get.
  *  Unboxed pointer locals (repOf(name).ptrKind) tag the returned node with `.ptrKind`
  *  so downstream coercions know it's an i32 offset, not a numeric. */
 export function readVar(name) {
-  if (ctx.func.boxed?.has(name))
+  if (ctx.func.boxed?.has(name)) {
+    // i32-narrowed cell (closure-capture narrowing — see analyzeFuncForEmit's
+    // cellTypes): the cell stores a raw i32, load it directly.
+    if (ctx.func.cellTypes?.has(name)) return typed(['i32.load', boxedAddr(name)], 'i32')
     return typed(['f64.load', boxedAddr(name)], 'f64')
+  }
   if (isGlobal(name)) {
-    const node = typed(['global.get', `$${name}`], ctx.scope.globalTypes.get(name) || 'f64')
+    const gt = ctx.scope.globalTypes.get(name) || 'f64'
+    const node = typed(['global.get', dollar(name)], gt)
     const grep = repOfGlobal(name)
-    if (grep?.ptrKind != null) {
+    if (gt === 'f64' && (lookupValType(name) === VAL.NUMBER || grep?.val === VAL.NUMBER)) node.valKind = VAL.NUMBER
+    // ptrKind tags a raw i32 pointer offset — meaningful only for an i32-STORED
+    // global (a typed-array/buffer carrier unboxed by unboxConstTypedGlobals). An
+    // f64 global holds a NaN-boxed value: object/array reads unbox at the access
+    // site via the schema/reinterpret path, never an i32 reinterpret of the storage.
+    // Attaching ptrKind to an f64 global makes `asF64` box the f64 *as if it were an
+    // i32* (i64.extend_i32_u on a global.get of type f64 → invalid wasm). Gate on the
+    // storage type so the tag follows the declared ABI.
+    if (gt === 'i32' && grep?.ptrKind != null) {
       node.ptrKind = grep.ptrKind
       if (grep.ptrAux != null) node.ptrAux = grep.ptrAux
     }
@@ -657,7 +990,8 @@ export function readVar(name) {
     return t === 'i32' ? typed(['i32.const', rep.intConst], 'i32')
                        : typed(['f64.const', rep.intConst], 'f64')
   }
-  const node = typed(['local.get', `$${name}`], t)
+  const node = typed(['local.get', dollar(name)], t)
+  if (t === 'f64' && (lookupValType(name) === VAL.NUMBER || rep?.val === VAL.NUMBER)) node.valKind = VAL.NUMBER
   // Proven uint32 accumulator local (narrowUint32): a later asF64 must widen with
   // convert_i32_u (the i32 bit pattern is an unsigned value), not _s. `.wrapSafe`
   // marks it as the always-ToUint32-sunk kind so the arithmetic widening guards
@@ -676,22 +1010,29 @@ export function readVar(name) {
 export function writeVar(name, valIR, void_) {
   if (ctx.func.boxed?.has(name)) {
     const addr = boxedAddr(name)
-    const v = asF64(valIR)
-    if (void_) return typed(['block', ['f64.store', addr, v]], 'void')
-    const t = temp()
-    return typed(['block', ['result', 'f64'],
+    // i32-narrowed cell: store the raw i32 (mirrors the integer-global write
+    // gate below — the storage type decides the coercion).
+    const i32Cell = ctx.func.cellTypes?.has(name)
+    const st = i32Cell ? 'i32.store' : 'f64.store'
+    const v = i32Cell ? asI32(valIR) : asF64(valIR)
+    if (void_) return typed(['block', [st, addr, v]], 'void')
+    const t = i32Cell ? tempI32() : temp()
+    return typed(['block', ['result', i32Cell ? 'i32' : 'f64'],
       ['local.set', `$${t}`, v],
-      ['f64.store', addr, ['local.get', `$${t}`]],
-      ['local.get', `$${t}`]], 'f64')
+      [st, addr, ['local.get', `$${t}`]],
+      ['local.get', `$${t}`]], i32Cell ? 'i32' : 'f64')
   }
   if (isGlobal(name)) {
-    const v = asF64(valIR)
-    if (void_) return typed(['block', ['global.set', `$${name}`, v]], 'void')
-    const t = temp()
-    return typed(['block', ['result', 'f64'],
+    // Scalar globals are f64 by default, but integer-global inference (plan.js)
+    // narrows purpose-focused counters/sizes to i32 — coerce the write to match.
+    const gt = ctx.scope.globalTypes.get(name) || 'f64'
+    const v = gt === 'i32' ? asI32(valIR) : asF64(valIR)
+    if (void_) return typed(['block', ['global.set', dollar(name), v]], 'void')
+    const t = gt === 'i32' ? tempI32() : temp()
+    return typed(['block', ['result', gt],
       ['local.set', `$${t}`, v],
-      ['global.set', `$${name}`, ['local.get', `$${t}`]],
-      ['local.get', `$${t}`]], 'f64')
+      ['global.set', dollar(name), ['local.get', `$${t}`]],
+      ['local.get', `$${t}`]], gt)
   }
   const t = ctx.func.locals.get(name) || 'f64'
   const ptrKind = repOf(name)?.ptrKind
@@ -703,10 +1044,10 @@ export function writeVar(name, valIR, void_) {
       ? valIR
       : typed(['i32.wrap_i64', ['i64.reinterpret_f64', asF64(valIR)]], 'i32')
   } else {
-    coerced = t === 'f64' ? asF64(valIR) : asI32(valIR)
+    coerced = t === 'v128' ? valIR : t === 'f64' ? asF64(valIR) : asI32(valIR)
   }
-  if (void_) return typed(['local.set', `$${name}`, coerced], 'void')
-  const teeNode = typed(['local.tee', `$${name}`, coerced], t)
+  if (void_) return typed(['local.set', dollar(name), coerced], 'void')
+  const teeNode = typed(['local.tee', dollar(name), coerced], t)
   if (ptrKind != null) teeNode.ptrKind = ptrKind
   return teeNode
 }
@@ -716,57 +1057,52 @@ export function writeVar(name, valIR, void_) {
  *  unboxed pointer locals are proven non-null by unboxablePtrs.
  *  Inlines directly: (i32.or (i64.eq bits NULL_NAN) (i64.eq bits UNDEF_NAN))
  *  rather than calling $__is_nullish — saves WASM call dispatch in V8 JIT. */
-export const isNullish = (f64expr) => {
-  if (f64expr.ptrKind != null) return typed(['i32.const', 0], 'i32')
+// Shared peephole for the NaN-box sentinel checks. When the operand's bits are
+// statically known — an unboxed pointer (never an atom → 0), a numeric `f64.const`
+// (never an atom → 0), or a boxed `(f64.const nan:…)` / `(f64.reinterpret_i64
+// (i64.const …))` literal — resolve `onBits(bitsHex)` / 0 at compile time; else
+// hand the expr to `fallback` for the runtime test. One place owns the literal set.
+const constI32 = (b) => typed(['i32.const', b ? 1 : 0], 'i32')
+const matchF64Bits = (f64expr, onBits, fallback) => {
+  if (f64expr.ptrKind != null) return constI32(0)
   if (Array.isArray(f64expr)) {
     if (f64expr[0] === 'f64.const') {
-      // Check for NaN-boxed sentinel: (f64.const nan:0x...) — matches NULL_IR/UNDEF_IR form.
       const lit = String(f64expr[1])
-      if (lit.startsWith('nan:')) {
-        const bits = lit.slice(4)
-        return typed(['i32.const', (bits === NULL_NAN || bits === UNDEF_NAN) ? 1 : 0], 'i32')
-      }
-      return typed(['i32.const', 0], 'i32')  // numeric literal — never nullish
+      return lit.startsWith('nan:') ? onBits(lit.slice(4)) : constI32(0)
     }
-    if (f64expr[0] === 'f64.reinterpret_i64' && Array.isArray(f64expr[1]) && f64expr[1][0] === 'i64.const') {
-      const bits = String(f64expr[1][1])
-      return typed(['i32.const', (bits === NULL_NAN || bits === UNDEF_NAN) ? 1 : 0], 'i32')
-    }
-  }
-  // Inline the 3-op nullish test. Tee into a temp i64 so the value is computed once.
-  // For simple (local.get $x) we can just reinterpret twice — V8 CSEs it.
-  if (Array.isArray(f64expr) && f64expr[0] === 'local.get') {
-    const bits = ['i64.reinterpret_f64', f64expr]
-    return typed(['i32.or',
-      ['i64.eq', bits, ['i64.const', NULL_NAN]],
-      ['i64.eq', ['i64.reinterpret_f64', f64expr], ['i64.const', UNDEF_NAN]]], 'i32')
+    if (f64expr[0] === 'f64.reinterpret_i64' && Array.isArray(f64expr[1]) && f64expr[1][0] === 'i64.const')
+      return onBits(String(f64expr[1][1]))
   }
-  // Non-trivial expr: fall back to the helper — keeps binary size stable & preserves eval once.
-  inc('__is_nullish')
-  return typed(['call', '$__is_nullish', ['i64.reinterpret_f64', f64expr]], 'i32')
+  return fallback(f64expr)
 }
 
+export const isNullish = (f64expr) => matchF64Bits(f64expr,
+  bits => constI32(bits === NULL_NAN || bits === UNDEF_NAN),
+  (e) => {
+    // (local.get $x): inline the test, reinterpreting twice (V8 CSEs it). Other
+    // exprs call $__is_nullish — keeps binary size stable and evaluates once.
+    if (Array.isArray(e) && e[0] === 'local.get') {
+      const bits = ['i64.reinterpret_f64', e]
+      return typed(['i32.or',
+        ['i64.eq', bits, ['i64.const', NULL_NAN]],
+        ['i64.eq', ['i64.reinterpret_f64', e], ['i64.const', UNDEF_NAN]]], 'i32')
+    }
+    inc('__is_nullish')
+    return typed(['call', '$__is_nullish', ['i64.reinterpret_f64', e]], 'i32')
+  })
+
 /** Check if f64 expr is exactly `undefined` (UNDEF_NAN). Returns i32.
  *  Used by default-param semantics — only `undefined` (or missing arg) triggers
  *  the default; `null` should pass through. */
-export const isUndef = (f64expr) => {
-  if (f64expr.ptrKind != null) return typed(['i32.const', 0], 'i32')
-  if (Array.isArray(f64expr)) {
-    if (f64expr[0] === 'f64.const') {
-      const lit = String(f64expr[1])
-      if (lit.startsWith('nan:')) {
-        const bits = lit.slice(4)
-        return typed(['i32.const', bits === UNDEF_NAN ? 1 : 0], 'i32')
-      }
-      return typed(['i32.const', 0], 'i32')
-    }
-    if (f64expr[0] === 'f64.reinterpret_i64' && Array.isArray(f64expr[1]) && f64expr[1][0] === 'i64.const') {
-      const bits = String(f64expr[1][1])
-      return typed(['i32.const', bits === UNDEF_NAN ? 1 : 0], 'i32')
-    }
-  }
-  return typed(['i64.eq', ['i64.reinterpret_f64', f64expr], ['i64.const', UNDEF_NAN]], 'i32')
-}
+export const isUndef = (f64expr) => matchF64Bits(f64expr,
+  bits => constI32(bits === UNDEF_NAN),
+  (e) => typed(['i64.eq', ['i64.reinterpret_f64', e], ['i64.const', UNDEF_NAN]], 'i32'))
+
+/** Check if f64 expr is exactly `null` (NULL_NAN). Returns i32.
+ *  Strict `=== null` must match only null — not undefined (use isUndef for that). */
+export const isNull = (f64expr) => matchF64Bits(f64expr,
+  bits => constI32(bits === NULL_NAN),
+  (e) => typed(['i64.eq', ['i64.reinterpret_f64', e], ['i64.const', NULL_NAN]], 'i32'))
 
 /** Mask that clears the boolean atom's truth bit, mapping TRUE_NAN→FALSE_NAN.
  *  `(bits & BOOL_ATOM_MASK) === FALSE_NAN` recognizes both in one i64.and+i64.eq. */
@@ -774,16 +1110,11 @@ const BOOL_ATOM_MASK = '0x' + BigInt.asUintN(64, ~(1n << BigInt(LAYOUT.AUX_SHIFT
 
 /** Check if f64 expr is a boxed-boolean atom (TRUE_NAN or FALSE_NAN). Returns i32.
  *  Single-eval: masks the truth bit and compares to FALSE_NAN once. */
-export const isBoolAtom = (f64expr) => {
-  if (f64expr.ptrKind != null) return typed(['i32.const', 0], 'i32')
-  if (Array.isArray(f64expr) && f64expr[0] === 'f64.const' && String(f64expr[1]).startsWith('nan:')) {
-    const b = String(f64expr[1]).slice(4)
-    return typed(['i32.const', (b === TRUE_NAN || b === FALSE_NAN) ? 1 : 0], 'i32')
-  }
-  return typed(['i64.eq',
-    ['i64.and', ['i64.reinterpret_f64', f64expr], ['i64.const', BOOL_ATOM_MASK]],
-    ['i64.const', FALSE_NAN]], 'i32')
-}
+export const isBoolAtom = (f64expr) => matchF64Bits(f64expr,
+  bits => constI32(bits === TRUE_NAN || bits === FALSE_NAN),
+  (e) => typed(['i64.eq',
+    ['i64.and', ['i64.reinterpret_f64', e], ['i64.const', BOOL_ATOM_MASK]],
+    ['i64.const', FALSE_NAN]], 'i32'))
 
 // === Array layout helpers — routed through the array carrier (abi/array.js) ===
 
@@ -813,7 +1144,7 @@ export function elemStore(ptr, i, val) {
  *  re-loading from ptr-8.
  *  Optional `ptrLocal`: caller already has the resolved ARRAY data pointer in
  *  an i32 local. Reuses it instead of calling __ptr_offset again. */
-export function arrayLoop(arrExpr, bodyFn, lenLocal, ptrLocal) {
+export function arrayLoop(arrExpr, bodyFn, lenLocal, ptrLocal, reverse) {
   const arr = ptrLocal ? null : temp('aa'), ptr = ptrLocal ?? tempI32('ap'), i = tempI32('ai'), item = temp('av')
   const len = lenLocal ?? tempI32('al')
   const id = ctx.func.uniq++
@@ -827,13 +1158,18 @@ export function arrayLoop(arrExpr, bodyFn, lenLocal, ptrLocal) {
   }
   if (!lenLocal) setup.push(
     ['local.set', `$${len}`, ['i32.load', ['i32.sub', ['local.get', `$${ptr}`], ['i32.const', 8]]]])
+  // Forward: i 0→len-1. Reverse (findLast*): i len-1→0, same elem indices.
+  const start = reverse ? ['i32.sub', ['local.get', `$${len}`], ['i32.const', 1]] : ['i32.const', 0]
+  const done = reverse ? ['i32.lt_s', ['local.get', `$${i}`], ['i32.const', 0]]
+                       : ['i32.ge_s', ['local.get', `$${i}`], ['local.get', `$${len}`]]
+  const step = ['i32.const', reverse ? -1 : 1]
   setup.push(
-    ['local.set', `$${i}`, ['i32.const', 0]],
+    ['local.set', `$${i}`, start],
     ['block', `$brk${id}`, ['loop', `$loop${id}`,
-      ['br_if', `$brk${id}`, ['i32.ge_s', ['local.get', `$${i}`], ['local.get', `$${len}`]]],
+      ['br_if', `$brk${id}`, done],
       ['local.set', `$${item}`, elemLoad(ptr, i)],
       ...bodyFn(ptr, len, i, typed(['local.get', `$${item}`], 'f64')),
-      ['local.set', `$${i}`, ['i32.add', ['local.get', `$${i}`], ['i32.const', 1]]],
+      ['local.set', `$${i}`, ['i32.add', ['local.get', `$${i}`], step]],
       ['br', `$loop${id}`]]])
   return setup
 }
@@ -885,6 +1221,7 @@ export function loopTop() {
 export const flat = ir => {
   if (ir == null) return []
   if (!Array.isArray(ir)) return [ir]  // bare 'drop', 'nop', etc.
+  if (ir.length === 0) return []
   if (typeof ir[0] === 'string' || ir[0] == null) return [ir]  // single instruction: ['op', ...args] or [null, val]
   return ir  // multi-instruction: [instr1, instr2, ...]
 }
@@ -906,6 +1243,38 @@ export function findBodyStart(fn) {
   return fn.length
 }
 
+/** Debug-mode structural check of a `(func …)` IR node. Catches the bug classes
+ *  that otherwise surface as OPAQUE watr errors several phases later — `Duplicate
+ *  local $x`, `Unknown local $x` — but here pinned to the exact name (and, via the
+ *  caller, the phase + function) that produced them, so a codegen/optimizer bug is
+ *  localized at its source instead of at watr. Self-contained: validates every
+ *  `local.{get,set,tee}` against the function header's param/local declarations,
+ *  and rejects a duplicate declaration. Returns an error string, or null if clean.
+ *  (Call-target and type-tag validation need the module symbol table + a type pass;
+ *  deferred — locals are the common codegen-bug class and need nothing external.) */
+export function verifyFn(fn) {
+  if (!Array.isArray(fn) || fn[0] !== 'func') return null
+  const bodyStart = findBodyStart(fn)
+  const declared = new Set()
+  for (let i = 2; i < bodyStart; i++) {
+    const c = fn[i]
+    if (!Array.isArray(c) || (c[0] !== 'param' && c[0] !== 'local') || typeof c[1] !== 'string') continue
+    if (declared.has(c[1])) return `duplicate local/param ${c[1]}`
+    declared.add(c[1])
+  }
+  let bad = null
+  const walk = (n) => {
+    if (bad || !Array.isArray(n)) return
+    const op = n[0]
+    if ((op === 'local.get' || op === 'local.set' || op === 'local.tee') && typeof n[1] === 'string' && !declared.has(n[1])) {
+      bad = `${op} of undeclared local ${n[1]}`; return
+    }
+    for (let i = 1; i < n.length; i++) walk(n[i])
+  }
+  for (let i = bodyStart; i < fn.length; i++) walk(fn[i])
+  return bad
+}
+
 /**
  * Tail-call rewrite: walks tail positions of an emitted IR tree and replaces
  * direct `(call $name args...)` ops with `(return_call $name args...)`.