@asmlift/core 0.5.0 → 0.6.0

package/src/structure/analysis.ts

@@ -1,12 +1,29 @@
 // asmlift structurer — the ANALYSIS phase. Pure derivation over the lifted fn — nothing here
 // mutates the IR or depends on naming/emission state:
 //   • use-site registry — every use of a value, POSITIONED (op + block + index);
-//   • per-block SSA value liveness (backward dataflow) — consumed by the coalescing
-//     interference check in structure.ts;
-//   • the effect-ordering model — which call/load defs must MATERIALIZE as named temps at
-//     their own program position instead of inlining at their use.
-import { globalCellOf, mayWriteGlobal } from '../ir/alias';
-import { Block, Fn, Op, Value, successorsOf } from '../ir/core';
+//   • per-block SSA value liveness (backward dataflow) — consumed by structure.ts and
+//     structure/namecoalesce.ts;
+//   • the effect-ordering model — which defs must MATERIALIZE as named temps at their own
+//     program position instead of inlining at their use (calls/loads for effect order, plus
+//     the pure defs the homing rules claim);
+//   • the HOMING-AXIS ENUMERATION GATES — one export per candidate axis, each answering "does
+//     this function hold a value the axis would home at all" so rank.ts can skip an axis whose
+//     candidate would only duplicate the default. Each mirrors its axis's scope inside `analyze`
+//     and states where it DIVERGES from it, in which direction, and what that costs.
+import { disjointConstSlots, globalCellOf, mayWriteGlobal } from '../ir/alias';
+import {
+  Block,
+  Fn,
+  Op,
+  Successor,
+  Value,
+  defOpMap,
+  dominators,
+  mergeClasses,
+  predecessors,
+  successorsOf,
+} from '../ir/core';
+import { EFFECTFUL_OPS, ORDER_SENSITIVE_OPS, REEVAL_UNSAFE_OPS } from '../ir/opcodes';
 
 export interface UseSite {
   blk: Block;
@@ -14,6 +31,635 @@ export interface UseSite {
   op: Op;
 }
 
+/** the ops whose operands[0] is a memory-access BASE — the address-home axis's slot model */
+const MEM_BASE_OPS = new Set(['load', 'store', 'aload', 'astore']);
+
+/** Is the op's own value an ADDRESS — a pointer or array whose standalone rendering must carry a
+ *  cast? Homed, `add(p, 8)` renders `(u16 *)(gPtr + 8)`: the cast lands outside the sum, so a byte
+ *  offset becomes element arithmetic and the address moves. The VALUE-side counterpart of
+ *  `coneHoldsAddr`, which is a different question rather than a weaker one — that walk crosses
+ *  reads, so a rule whose clientele IS values over a load cannot ask it, and a pointer loaded from
+ *  memory has no gaddr of its own for it to find. Each caller's refusal says which half it buys. */
+function rendersAsAddress(op: Op): boolean {
+  const t = op.results[0]?.type;
+  return t?.kind === 'ptr' || t?.kind === 'array';
+}
+
+/** Values C evaluates only under a `&&`/`||`: the SECOND operand of every `logic_and`/`logic_or`
+ *  and its transitive operand cone. raise/shortcircuit.ts lifts that cone into the block ABOVE the
+ *  branch on the contract that the structurer inlines it back under C's own short circuit, so the
+ *  def block of anything in it is a FOLD ARTIFACT — a rule that names one of these values there
+ *  emits it above the guard the source wrote (`p != 0 && *p != 0` becoming `v0 = *p;` above its own
+ *  null check). Asked by the def-block placement rule and by the merge-feed-home scope. */
+function shortCircuitGuardedValues(fn: Fn, defOf: Map<Value, Op>): Set<Value> {
+  const guarded = new Set<Value>();
+  const work: Value[] = [];
+  for (const b of fn.blocks) {
+    for (const op of b.ops) {
+      const rhs = op.opcode === 'logic_and' || op.opcode === 'logic_or' ? op.operands[1] : undefined;
+      if (rhs !== undefined) {
+        work.push(rhs);
+      }
+    }
+  }
+  while (work.length) {
+    const v = work.pop()!;
+    if (!guarded.has(v)) {
+      guarded.add(v);
+      work.push(...(defOf.get(v)?.operands ?? []));
+    }
+  }
+  return guarded;
+}
+
+/** Any gaddr/laddr in the op's operand cone (the op included). Rendered standalone, an address
+ *  computation over one loses the memAccess's inline byte-stride cast — the value changes, so a
+ *  homing rule asking this refuses the cone (the cast-aware machinery in l3/basecse.ts,
+ *  scopebase.ts and nearbase.ts serves those bases instead). The walk deliberately crosses loads —
+ *  a gaddr reachable only through a load's address keeps its cast at that load's own deref, so
+ *  over-refusal there costs a candidate, never soundness. That over-refusal is why the
+ *  derived-read-home axis asks its own pair instead (an address in the cone OUTSIDE a read, plus
+ *  `rendersAsAddress` on the value): reads over named globals are its whole clientele. */
+function coneHoldsAddr(op0: Op, defOf: Map<Value, Op>): boolean {
+  const seen = new Set<Value>();
+  const cone = [op0];
+  while (cone.length) {
+    const d = cone.pop()!;
+    if (d.opcode === 'gaddr' || d.opcode === 'laddr') {
+      return true;
+    }
+    for (const x of d.operands) {
+      if (!seen.has(x)) {
+        seen.add(x);
+        const dd = defOf.get(x);
+        if (dd) {
+          cone.push(dd);
+        }
+      }
+    }
+  }
+  return false;
+}
+
+/** THE ADDRESS-HOME AXIS'S SCOPE: every value whose MERGE CLASS (ir/core.ts `mergeClasses`) is
+ *  used only as the base of memory accesses, and at 2+ of them.
+ *
+ *  WHY THE CLASS AND NOT THE VALUE. The question the axis asks is about a REGISTER — did the
+ *  machine derive one address and dereference it at several sites — and a register that survives a
+ *  branch merge is spelled in functional-form SSA as an edge argument plus a block parameter. So a
+ *  base each arm derives and the join then reads is N+1 SSA values with one base use apiece, none
+ *  of which reaches the 2-access threshold, while the register it describes reaches N+1 of them.
+ *  Counting per value also mis-reads the edge itself: the terminator that carries the argument is
+ *  a consumer that is not a memory access, so the per-value rule refuses on the very edge that
+ *  makes the base shared.
+ *
+ *  An edge argument is therefore NOT a use that leaves the class — it feeds the parameter beside
+ *  it, which `mergeClasses` has already unioned in. Everything else still disqualifies the whole
+ *  class: a store's value slot, an `aload` index, arithmetic. The home is justified by the
+ *  shared-base reuse alone, as it always was.
+ *
+ *  `ignoreRet` skips `ret` operands: `analyze` passes its own `returnsVoid` (where a ret operand
+ *  is a phantom the use registry already drops), and rank.ts's enumeration gate passes true
+ *  unconditionally, because it cannot know — see `hasHomeableSharedAddress`.
+ *
+ *  Block PARAMETERS are members like any other and can qualify here. Neither caller homes one —
+ *  both filter by a DEF — but they must be counted, or the join half of every class is invisible.
+ *
+ *  IT IS NOT A SUPERSET OF THE PER-VALUE RULE IT REPLACED, and that is the price rather than a
+ *  bug: a value whose CLASS MATE escapes is now refused where counting the value alone admitted
+ *  it. Measured by running both predicates over every function in the four checkouts — 1487
+ *  lifting functions, both symbol-map configurations — the widening admits new values in 25
+ *  functions map-less and 8 map-ful, and REMOVES exactly one value from each of four functions
+ *  map-less (kleod `sub_0804EB64`, marioparty3 `func_8010E344_25A994_rockin_raceway`,
+ *  snowboardkids2 `func_8006FDA0_709A0` and `func_8006FDC8_709C8`), none map-ful. The function-
+ *  level ENUMERATION gate is a clean superset over the same corpus, which is why a census taken
+ *  at that scope reports no loss; this is the finer one.
+ *
+ *  NO LOOP-HEADER REFUSAL OF ITS OWN, AND THE REASON IS A GUARD THAT WAS ALREADY THERE. The
+ *  scope's plan mandated one — refuse a class reaching a loop-header PARAMETER, because
+ *  `structure.ts`'s `carriesPreUpdate` makes a loop variable's name denote different values at
+ *  different points. Two things are true and only together do they answer it.
+ *
+ *  The FIRST is about which classes the widening adds. Of the 25+8 functions above exactly one
+ *  carries a class reaching a loop-header parameter (marioparty3
+ *  `func_80112508_523648_filesel`, 2 values, map-less), and neither value is one the axis could
+ *  materialize; `addr-home.test.ts` pins both halves. A loop-carried pointer INDUCTION is refused
+ *  by its own increment, a non-base use of a class member. The loop-header class that does survive
+ *  has a back-edge value READ FROM MEMORY, so its only def is a `load` and the enumeration gate
+ *  excludes it. So the widening adds no loop-header home, and the planned guard has no inhabitant
+ *  among the values it would have been written for.
+ *
+ *  The SECOND is what stands there whatever the scope rule says, and it is NOT one of those two:
+ *  the third scope's own `!multiBlockHeaders.has(b)` seat refusal, which predates this widening
+ *  and refuses to materialize AT a multi-block loop header regardless of which class the value is
+ *  in. It is load-bearing and measurable: replace that clause with `true` and marioparty3
+ *  `func_8010923C_18E46C_cosmic_coaster` stops structuring at all — it throws the
+ *  `carriesPreUpdate` StructureError this paragraph is about — while `func_80033910_34510` and
+ *  `func_80033970_34570` go from axis-inert to changing their source. So the hazard is real; it is
+ *  answered by WHERE a home may sit rather than by WHICH class may have one, which is a refusal
+ *  the widening does not touch and the plan's class guard would have restated. Note what that
+ *  implies about the enumeration gate: it deliberately
+ *  omits the seat refusal (see `hasHomeableSharedAddress`), which is safe only because the seat
+ *  refusal is applied HERE, in the scope, on every candidate the gate enumerates. */
+export function sharedBaseClasses(fn: Fn, ignoreRet: boolean): Set<Value> {
+  const classes = mergeClasses(fn);
+  const baseUses = new Map<Value, Set<Op>>();
+  const otherUse = new Set<Value>();
+  for (const b of fn.blocks) {
+    for (const op of b.ops) {
+      if (ignoreRet && op.opcode === 'ret') {
+        continue;
+      }
+      op.operands.forEach((o, i) => {
+        // BASE SLOT ONLY. `store p, p` — the address written as its own data — needs no clause
+        // here and had one: the j>0 pass over the SAME operand list puts `p` into `otherUse`
+        // regardless, and `otherUse` disqualifies the whole class below, so a base-slot entry for
+        // an escaping value can never be read. Measured before removing it: over 1487 lifted
+        // corpus functions the shape occurs at 10 sites and the predicate's output is identical
+        // with the clause and without it, in both `ignoreRet` senses. What keeps the escape out is
+        // therefore the `otherUse` sweep, and that is what `analysis.test.ts` pins — a clause that
+        // cannot change an answer is not a guard, it is a claim that one exists.
+        if (i === 0 && MEM_BASE_OPS.has(op.opcode)) {
+          (baseUses.get(o) ?? baseUses.set(o, new Set()).get(o)!).add(op);
+        } else {
+          otherUse.add(o);
+        }
+      });
+      for (const s of op.successors) {
+        s.args.forEach((a, i) => {
+          // Class-internal only up to the parameter list: an argument past the end binds nothing,
+          // so `mergeClasses` never unioned it and it escapes like any other use.
+          if (i >= s.block.params.length) {
+            otherUse.add(a);
+          }
+        });
+      }
+    }
+  }
+  const out = new Set<Value>();
+  for (const b of fn.blocks) {
+    for (const v of [...b.params, ...b.ops.flatMap((op) => op.results)]) {
+      const members = classes.get(v) ?? [v];
+      if (members.some((m) => otherUse.has(m))) {
+        continue;
+      }
+      const consumers = new Set<Op>();
+      for (const m of members) {
+        for (const c of baseUses.get(m) ?? []) {
+          consumers.add(c);
+        }
+      }
+      if (consumers.size >= 2) {
+        out.add(v);
+      }
+    }
+  }
+  return out;
+}
+
+/** Is `d` a def one of the homing axes could seat in a local — a def at all, and a PURE
+ *  non-memory one that is not a `const`? The three enumeration gates below share this filter and
+ *  each then adds its own cone/address/shape refusals.
+ *
+ *  Named for what it tests rather than for the shape it excludes: a `const` IS a pure non-memory
+ *  def, and it is out because a re-derived const is re-materialization — the compiler's own
+ *  behaviour — not because it computes anything a home would preserve. */
+function isHomeableDef(d: Op | undefined): d is Op {
+  return !!d && d.opcode !== 'const' && d.opcode !== 'call' && d.opcode !== 'load' && d.opcode !== 'aload';
+}
+
+/** THE READ CONE of a pure def: the `load`/`aload` defs its operand tree stands on. The walk stops
+ *  AT a read (a read's own address computation is the read's business, not this walk's) and
+ *  REFUSES — null — as soon as the cone holds a `call`, whose re-render would re-execute it, or a
+ *  `gaddr`/`laddr`, whose standalone rendering loses the memAccess's inline byte-stride cast.
+ *
+ *  Both callers gate `load`/`aload`/`call` out of `op0` itself before asking, and neither reads the
+ *  order the reads come back in — one takes `length`/`every`, the other builds a `Set`. */
+function readCone(op0: Op, defOf: Map<Value, Op>): Op[] | null {
+  const reads: Op[] = [];
+  const seen = new Set<Value>();
+  const cone = [op0];
+  while (cone.length) {
+    const d = cone.pop()!;
+    if (d.opcode === 'load' || d.opcode === 'aload') {
+      reads.push(d);
+      continue;
+    }
+    if (d.opcode === 'call' || d.opcode === 'gaddr' || d.opcode === 'laddr') {
+      return null;
+    }
+    for (const x of d.operands) {
+      if (!seen.has(x)) {
+        seen.add(x);
+        const dd = defOf.get(x);
+        if (dd) {
+          cone.push(dd);
+        }
+      }
+    }
+  }
+  return reads;
+}
+
+/** rank.ts's enumeration gate for the `/addr-home` axis: does the function HAVE a value the axis
+ *  would home — a non-const pure def whose merge class is a shared base, with no gaddr/laddr in
+ *  its cone? Mirrors the axis's scope rule in `analyze` (the same `sharedBaseClasses` call), minus
+ *  the loop-header seat refusal (that needs the loop model; a false positive costs one
+ *  duplicate-collapsed candidate, never a wrong one).
+ *
+ *  `ignoreRet` is true here: a `ret` operand may be a void phantom, which `analyze` skips under
+ *  `returnsVoid` and this gate cannot know. For a genuinely returned base the axis's own rule
+ *  still refuses, costing one duplicate-collapsed candidate — the same trade as the loop-header
+ *  divergence. */
+export function hasHomeableSharedAddress(fn: Fn): boolean {
+  const defOf = defOpMap(fn);
+  for (const v of sharedBaseClasses(fn, true)) {
+    const d = defOf.get(v);
+    if (isHomeableDef(d) && !coneHoldsAddr(d, defOf)) {
+      return true;
+    }
+  }
+  return false;
+}
+
+/** rank.ts's enumeration gate for the `/expr-home` axis: does the function HAVE a value the
+ *  axis would home — a pure non-const def with 2+ distinct consumers, at least one of them inside
+ *  a loop the def sits outside, cone-free? Loops here are LAYOUT ranges (a successor at an
+ *  equal-or-earlier block position closes one) where the axis's own rule uses the dominator model,
+ *  and consumers here come from op operands only, where the rule counts `useSitesOf` and so counts
+ *  branch args too — unlike hasHomeableSharedAddress this therefore diverges in BOTH directions. A
+ *  false positive costs one duplicate-collapsed candidate. A false negative silently skips the arm:
+ *  on IR whose block layout does not follow dominance, which every frontend avoids by laying blocks
+ *  out in address order (a natural loop's back edge points backward), and on a value EITHER of
+ *  whose two consumers is a branch arg, which the rule would home and this never enumerates. The
+ *  second is unwitnessed over the 856-row bench the axis was measured on (#97), and costs a
+ *  missing candidate, never a wrong one. */
+export function hasLoopSharedPureValue(fn: Fn): boolean {
+  const defOf = defOpMap(fn);
+  const pos = new Map<Block, number>(fn.blocks.map((b, i) => [b, i]));
+  const ranges: [number, number][] = [];
+  for (const b of fn.blocks) {
+    for (const sx of successorsOf(b)) {
+      if (pos.get(sx)! <= pos.get(b)!) {
+        ranges.push([pos.get(sx)!, pos.get(b)!]);
+      }
+    }
+  }
+  if (ranges.length === 0) {
+    return false;
+  }
+  const consumers = new Map<Value, Set<Op>>();
+  const opPos = new Map<Op, number>();
+  for (const b of fn.blocks) {
+    for (const op of b.ops) {
+      opPos.set(op, pos.get(b)!);
+      for (const o of op.operands) {
+        (consumers.get(o) ?? consumers.set(o, new Set()).get(o)!).add(op);
+      }
+    }
+  }
+  for (const [v, cs] of consumers) {
+    const d = defOf.get(v);
+    if (!isHomeableDef(d)) {
+      continue;
+    }
+    const dp = opPos.get(d)!;
+    if (
+      cs.size >= 2 &&
+      ranges.some(
+        ([lo, hi]) => (dp < lo || dp > hi) && [...cs].some((c) => opPos.get(c)! >= lo && opPos.get(c)! <= hi),
+      ) &&
+      !coneHoldsAddr(d, defOf)
+    ) {
+      return true;
+    }
+  }
+  return false;
+}
+
+/** rank.ts's enumeration gate for the `/derived-home` axis: does the function HAVE a value the
+ *  axis would home — a pure non-const, non-pointer def with 2+ consumers standing on a same-block
+ *  memory read that is used nowhere else and reaches it with no write in between, and with no call
+ *  or standalone address in the cone? A TRUE here DOUBLES the whole structuring cross for the
+ *  function — not one candidate — so every refusal cheap enough to state without the positioned
+ *  model is mirrored here; only the loop-header seat stays out.
+ *
+ *  Diverges in BOTH directions, like `hasLoopSharedPureValue` and unlike `hasHomeableSharedAddress`.
+ *  Over: the write scan is straight-line within the one block, where the rule's is cycle-aware, so
+ *  a write reaching only around a back edge is invisible here — a false positive costs a doubled
+ *  cross whose every candidate the source dedup then collapses. Under: use counting here is by SLOT
+ *  over operands and successor args, where `analyze` drops a void function's `ret` operand — so a
+ *  read whose second use is a suppressed return is refused here and admitted there, silently
+ *  skipping the arm. That shape is a void function returning the very halfword it read, which no
+ *  caller can observe; the axis's clientele reads to compute, not to return. */
+export function hasDerivedReadHome(fn: Fn): boolean {
+  const defOf = defOpMap(fn);
+  const consumers = new Map<Value, Set<Op>>();
+  const useSlots = new Map<Value, number>();
+  const blockOf = new Map<Op, Block>();
+  const idxOf = new Map<Op, number>();
+  for (const b of fn.blocks) {
+    b.ops.forEach((op, i) => {
+      blockOf.set(op, b);
+      idxOf.set(op, i);
+      const use = (v: Value) => {
+        (consumers.get(v) ?? consumers.set(v, new Set()).get(v)!).add(op);
+        useSlots.set(v, (useSlots.get(v) ?? 0) + 1);
+      };
+      for (const o of op.operands) {
+        use(o);
+      }
+      for (const s of op.successors) {
+        for (const a of s.args) {
+          use(a);
+        }
+      }
+    });
+  }
+  /** any op that writes memory strictly between two ops of one block — the rule's `memWriteBetween`
+   *  over the straight line the same-block requirement already pins */
+  const writeBetween = (b: Block, lo: number, hi: number): boolean =>
+    b.ops.slice(lo + 1, hi).some((x) => EFFECTFUL_OPS.has(x.opcode));
+  const standsOnRead = (op0: Op): boolean => {
+    const reads = readCone(op0, defOf);
+    if (reads === null) {
+      return false;
+    }
+    const b = blockOf.get(op0)!;
+    return (
+      reads.length > 0 &&
+      reads.every(
+        (r) =>
+          blockOf.get(r) === b &&
+          (useSlots.get(r.results[0]) ?? 0) === 1 &&
+          !writeBetween(b, idxOf.get(r)!, idxOf.get(op0)!),
+      )
+    );
+  };
+  for (const [v, cs] of consumers) {
+    const d = defOf.get(v);
+    if (cs.size >= 2 && isHomeableDef(d) && !rendersAsAddress(d) && standsOnRead(d)) {
+      return true;
+    }
+  }
+  return false;
+}
+
+/** Natural loops: a back edge is `latch → header` with the header dominating the latch, and the
+ *  body is the backward closure from the latch. Shared by the rules inside `analyze` and by the
+ *  `/merge-home` gate below, so the two cannot disagree about what "inside a loop" means. */
+function naturalLoops(
+  fn: Fn,
+  dom: Map<Block, Set<Block>>,
+  predsOf: Map<Block, Block[]>,
+): { header: Block; body: Set<Block> }[] {
+  const loops: { header: Block; body: Set<Block> }[] = [];
+  for (const latch of fn.blocks) {
+    for (const header of successorsOf(latch)) {
+      if (!dom.get(latch)?.has(header)) {
+        continue;
+      }
+      const body = new Set<Block>([header]);
+      const work = [latch];
+      while (work.length) {
+        const x = work.pop()!;
+        if (!body.has(x)) {
+          body.add(x);
+          work.push(...(predsOf.get(x) ?? []));
+        }
+      }
+      loops.push({ header, body });
+    }
+  }
+  return loops;
+}
+
+/** THE MERGE-FEED-HOME axis's scope (rank.ts `/merge-home`, AnalyzeOptions.homeMergeFeeds): the
+ *  pure defs it materializes.
+ *
+ *  A merge parameter whose incoming edges render ONE value's expression twice or more is a value
+ *  the source computed once ABOVE the branch and the copy machinery then sank into every arm —
+ *  `s32 m = (b & 1) ? 0x400 : 0;` becomes the whole `neg/orr/asr #31/and` chain in each arm, which
+ *  agbcc if-converts per arm instead of holding the value in a register across the branch. Homed at
+ *  the def, it renders once and the arms read the name.
+ *
+ *  Neither the value's use count nor its multi-render alone is the evidence: the sibling scopes'
+ *  clientele (a value re-derived at two ordinary uses) is a value the compiler DOES re-materialize
+ *  for free. What a merge slot adds is that the two renders are mutually exclusive ARMS of one
+ *  branch. Which side the source spelled is still not derivable (nothing stops a source from
+ *  writing the expression twice), so this is a differ-refereed candidate axis, never a default.
+ *
+ *  Refusals:
+ *    • a value whose cone holds a gaddr/laddr (`coneHoldsAddr`) — rendered standalone an `&g + i`
+ *      loses the memAccess's inline byte-stride cast, the same soundness half the sibling scopes
+ *      state.
+ *    • a value that is ITSELF an address (`rendersAsAddress`). A SCOPE boundary rather than that
+ *      same hazard: what it adds over `coneHoldsAddr` is the pointers with no gaddr/laddr anywhere
+ *      in their cone — derived from, or loaded from, a pointer param — and there the backend does
+ *      re-derive the byte cast (`v0 = (u8 *)*a1 + 6;` then `v0 = (u8 *)v0 + 2;`). Address-shaped
+ *      bases belong to the cast-aware machinery in l3/basecse.ts, scopebase.ts and nearbase.ts;
+ *      this axis does not offer a second spelling of them.
+ *    • an op whose answer depends on WHERE it runs, or that can TRAP — `REEVAL_UNSAFE_OPS`, read
+ *      from the registry rather than re-listed. Both halves carry: for a read WHERE it happens is
+ *      the read rules' question, and a homed divide becomes an unconditional statement at a def
+ *      block raise/shortcircuit.ts may have made, on paths C's own `&&` would have re-guarded —
+ *      the KNOWN GAP `ir/opcodes.ts` books against `HOIST_UNSAFE_OPS`.
+ *    • an `undef` — the axis's premise is a value the source COMPUTED once above the branch, and
+ *      an uninitialised register was never computed at all: homed, it spells `v0 = uninit_r5;`,
+ *      a copy of a value nothing wrote, which no asm can have.
+ *    • a value in a `&&`/`||`'s guarded cone WHOSE CONE HOLDS a re-evaluation-unsafe op —
+ *      raise/shortcircuit.ts put that cone above the branch on the contract that the structurer
+ *      inlines it back under C's own short circuit, so its def block is a FOLD ARTIFACT rather than
+ *      the block the asm computed in. Naming it there emits `v0 = *p | 1;` above `p != 0`, and
+ *      `v0 = a2 / a0 | 1;` above the `a0 != 0` the source divided under — the trap half is the one
+ *      that turns a re-spelling into a fault. Narrowed to that cone deliberately: a PURE guarded
+ *      value re-spelled above the connective computes the same thing on the same paths, and
+ *      refusing it outright costs `synthetic:modpow2:ido7.1` 4 → 6 for no soundness gain.
+ *    • a def block or a join INSIDE A LOOP. A loop-carried merge parameter's "definition above the
+ *      branch" is the loop's entry initializer, whose placement `/defsite/loop-entry` already
+ *      decides; homing it here would answer the same question a second time, from a rule that
+ *      cannot see the loop's kept guard. It has NO corpus reach — the gate admits the same 16 rows
+ *      with it on and off, in both symbol-map configurations — so what it buys is fan on
+ *      loop-shaped functions the corpus does not carry. klonoa's LoadBGTilemapData (89 blocks, 25
+ *      joins with params) is refused by this clause and no other, and lifting it takes that
+ *      function's map-less enumeration 75264 → 150528 candidates, 102s → 207s. A null there is
+ *      this refusal, not an absent idiom.
+ *  The `analyze` scope adds nothing to this list — the whole predicate is here, so the axis's
+ *  enumeration gate can run it rather than approximate it. */
+function mergeFeedHomes(fn: Fn, dom: Map<Block, Set<Block>>, defOf: Map<Value, Op>, inLoop: Set<Block>): Set<Op> {
+  // Every block's incoming COPY SITES — the places its edge assignments render, each once. Neither
+  // the predecessor list nor the raw edge list: `predecessors` (ir/core.ts) lists a block once per
+  // successor EDGE, so walking it against that block's own successors visits a join two of one
+  // block's edges reach k² times, and a value carried on ONE of them then tallies 2.
+  //
+  // The shape to follow is structure.ts's Regime B: it groups the `switch_br`'s CASE slots by
+  // target block (two naming one block are ONE arm carrying both `case` labels, args required to
+  // agree), then emits the DEFAULT edge as an entry of its own with its own copies — so a block a
+  // case slot and the default both name renders them TWICE. A `cond_br`'s two arms each emit their
+  // own even when both name the join. Over-counting a site costs a candidate that homes a value
+  // rendered once, which loses on score; under-counting costs the candidate outright.
+  const copySitesOf = new Map<Block, Successor[]>();
+  const addSite = (sx: Successor): void => {
+    const at = copySitesOf.get(sx.block);
+    if (at) {
+      at.push(sx);
+    } else {
+      copySitesOf.set(sx.block, [sx]);
+    }
+  };
+  for (const b of fn.blocks) {
+    const term = b.ops[b.ops.length - 1];
+    const succ = term?.successors ?? [];
+    if (term?.opcode === 'switch_br' && succ.length > 0) {
+      const seen = new Set<Block>();
+      for (const sx of succ.slice(0, -1)) {
+        if (!seen.has(sx.block)) {
+          seen.add(sx.block);
+          addSite(sx);
+        }
+      }
+      addSite(succ[succ.length - 1]!);
+    } else {
+      for (const sx of succ) {
+        addSite(sx);
+      }
+    }
+  }
+  const joins = fn.blocks.filter(
+    (m) => m.params.length > 0 && (copySitesOf.get(m)?.length ?? 0) >= 2 && !inLoop.has(m),
+  );
+  if (joins.length === 0) {
+    return new Set<Op>();
+  }
+  const blockOf = new Map<Op, Block>();
+  for (const b of fn.blocks) {
+    for (const op of b.ops) {
+      blockOf.set(op, b);
+    }
+  }
+  // Every value an edge ARGUMENT renders — the argument itself plus its inlined operand cone,
+  // memoized per value so one walk serves every slot. The walk stops DESCENDING at an
+  // order-sensitive def: under a call or a memory read the value renders at that op's own position,
+  // never at this edge. It crosses a trapping divide, which nothing materializes — that renders
+  // inline at the copy site and its operands render there with it. The stopping op is still
+  // recorded, so a membership test over the cone sees it.
+  const coneCache = new Map<Value, Set<Value>>();
+  const coneOf = (root: Value): Set<Value> => {
+    const hit = coneCache.get(root);
+    if (hit) {
+      return hit;
+    }
+    const acc = new Set<Value>();
+    const stack = [root];
+    while (stack.length) {
+      const x = stack.pop()!;
+      if (acc.has(x)) {
+        continue;
+      }
+      acc.add(x);
+      const memo = coneCache.get(x);
+      if (memo) {
+        for (const y of memo) {
+          acc.add(y);
+        }
+        continue;
+      }
+      const d = defOf.get(x);
+      if (d && !ORDER_SENSITIVE_OPS.has(d.opcode)) {
+        stack.push(...d.operands);
+      }
+    }
+    coneCache.set(root, acc);
+    return acc;
+  };
+  const scGuarded = shortCircuitGuardedValues(fn, defOf);
+  /** Does anything in the value's cone answer differently, or trap, at another program point? The
+   *  cone renders INSIDE the home, so both halves move with it — a read answers whichever stores
+   *  ran before it, and a divide that only ever ran under a guard divides by zero above it. */
+  const coneHoldsReevalUnsafe = (v: Value): boolean =>
+    [...coneOf(v)].some((x) => {
+      const d = defOf.get(x);
+      return d !== undefined && REEVAL_UNSAFE_OPS.has(d.opcode);
+    });
+  /** may this op's result be materialized at its def — and is this axis the one to do it? */
+  const eligible = (op: Op): boolean => {
+    const v = op.results[0];
+    return (
+      v !== undefined &&
+      !REEVAL_UNSAFE_OPS.has(op.opcode) &&
+      op.opcode !== 'undef' &&
+      !(scGuarded.has(v) && coneHoldsReevalUnsafe(v)) &&
+      !rendersAsAddress(op) &&
+      !coneHoldsAddr(op, defOf)
+    );
+  };
+  const out = new Set<Op>();
+  for (const m of joins) {
+    for (let i = 0; i < m.params.length; i++) {
+      const tally = new Map<Value, number>();
+      const carried = new Set<Value>();
+      for (const sx of copySitesOf.get(m)!) {
+        const a = sx.args[i];
+        if (a === undefined) {
+          continue;
+        }
+        carried.add(a);
+        for (const v of coneOf(a)) {
+          tally.set(v, (tally.get(v) ?? 0) + 1);
+        }
+      }
+      // The slot's feeders: rendered from 2+ of its edges, CARRIED unchanged by at least one of
+      // them, by a def this join's every arm reaches.
+      //
+      // Both counts carry a different half of the evidence. The 2+ renders are the duplication;
+      // what the carried edge adds is that the value IS the merge variable on some path — the
+      // register the compiler had to reserve across the branch either way, so the sunk spelling is
+      // strictly more work than the asm shows. Without it the scope reaches every shared
+      // SUBEXPRESSION of two arms, which agbcc computes per arm — `armkeep` is the row that prices
+      // that class, and `subexpr` the fixture that holds the line (armkeep's own arms share only a
+      // block PARAMETER, which has no def op to home, so the row never reaches this clause).
+      // Only a DOMINATING def is one definition the source could have written above the branch.
+      const feeders: Op[] = [];
+      for (const [v, n] of tally) {
+        const d = n >= 2 && carried.has(v) ? defOf.get(v) : undefined;
+        const b = d ? blockOf.get(d) : undefined;
+        if (d && b && b !== m && dom.get(m)!.has(b) && !inLoop.has(b) && eligible(d)) {
+          feeders.push(d);
+        }
+      }
+      // Only the MAXIMAL ones. Where a slot has enough edges for two of its arguments to qualify
+      // and one sits inside the other's cone, homing both spells `v0 = a1 & 1; v1 = v0 | 8;` where
+      // the source spelled one expression. The inner one renders INSIDE the outer's home, so its
+      // own is redundant; a feeder that is not covered stays, which is what keeps a second slot's
+      // smaller feeder its home.
+      const covered = new Set<Value>();
+      for (const d of feeders) {
+        for (const v of coneOf(d.results[0])) {
+          if (v !== d.results[0]) {
+            covered.add(v);
+          }
+        }
+      }
+      for (const d of feeders) {
+        if (!covered.has(d.results[0])) {
+          out.add(d);
+        }
+      }
+    }
+  }
+  return out;
+}
+
+/** rank.ts's enumeration gate for the `/merge-home` axis: does the function HAVE a value the axis
+ *  would home? The scope itself rather than a restatement of it, so the gate and the rule cannot
+ *  drift apart — admitting still says only that the rule fires, not that the tree changes. */
+export function hasMergeFeedHome(fn: Fn): boolean {
+  const dom = dominators(fn);
+  const predsOf = predecessors(fn);
+  const inLoop = new Set(naturalLoops(fn, dom, predsOf).flatMap((L) => [...L.body]));
+  return mergeFeedHomes(fn, dom, defOpMap(fn), inLoop).size > 0;
+}
+
 export interface StructureAnalysis {
   /** every positioned use of a value; a value absent here is dead */
   useSitesOf: Map<Value, UseSite[]>;
@@ -21,7 +667,8 @@ export interface StructureAnalysis {
   opBlock: Map<Op, Block>;
   /** SSA values live at each block's entry */
   liveIn: Map<Block, Set<Value>>;
-  /** call/load defs that must emit as named temps at their own position */
+  /** defs that must emit as named temps at their own position — calls/loads for effect order,
+   *  plus the pure defs the homing rules claim */
   materialize: Set<Op>;
   /** cached forward reachability (successors-transitive, excluding the start block itself) */
   reachFrom: (b: Block) => Set<Block>;
@@ -29,7 +676,7 @@ export interface StructureAnalysis {
    *  transitively; null = several places / unresolvable (callers treat conservatively) */
   emitPos: (op: Op) => { blk: Block; idx: number } | null;
   /** may an op `isWrite` accepts execute between `def` and a statement at `render`, on any
-   *  def-avoiding path — the fold-ordering gate (see the closure's comment) */
+   *  def-avoiding path — the fold-ordering gate (see `makeMemWriteBetween`) */
   memWriteBetween: (def: Op, render: { blk: Block; idx: number }, isWrite: (x: Op) => boolean) => boolean;
 }
 
@@ -38,6 +685,9 @@ export interface AnalyzeOptions {
    *  rebuilt. Absent ⇒ the global-aware alias rule below cannot resolve anything and every write
    *  bars, exactly as before it existed. */
   defs?: Map<Value, Op>;
+  /** Dominator sets, when the caller already holds them (structure() does) — consumed by the
+   *  live-across-a-loop rule's back-edge detection. Absent ⇒ that rule stands down. */
+  dom?: Map<Block, Set<Block>>;
   /** THE value-home axis (rank.ts `/reread-globals`). A read of a named global is barred from
    *  rendering at its use by any write in between — even a store to an unrelated global, which
    *  cannot possibly change what it sees. That over-conservatism is what invents the locals the
@@ -46,10 +696,21 @@ export interface AnalyzeOptions {
    *      gA = v; gB = v;   with `s32 v = gValue;`   where the source said `gA = gValue; gB = gValue;`
    *
    *  With this on, the barrier scan for a load whose address resolves to a named global uses THE
-   *  shared disjointness query (ir/alias.ts) instead of "any write at all". Materializing is always
-   *  sound, so today's spelling is never wrong — only sometimes not the one the compiler was given.
+   *  shared disjointness query (ir/alias.ts) instead of "any write at all". Materializing a
+   *  global's READ is always sound (the deref and its cast render at the def's position — contrast
+   *  homeSharedAddresses below, where the materialized value is an address and soundness is
+   *  scoped), so today's spelling is never wrong — only sometimes not the one the compiler was
+   *  given.
    *  Which side matches is genuinely per-function (the same dogfood watched agbcc go both ways
-   *  inside ONE function), so this is a differ-refereed candidate axis, never a default. */
+   *  inside ONE function), so this is a differ-refereed candidate axis, never a default.
+   *
+   *  SCOPE, since `readsStayWhereWritten` below now answers the same question — read once and
+   *  reuse, or read per use — with the opposite default: this axis owns renders in the read's OWN
+   *  block, where nothing about placement is in evidence and both spellings really do compile.
+   *  A read whose every render sits in a STRICTLY DOMINATED block is the default's, and on a
+   *  target that declares it the axis cannot produce the re-read spelling there (the rule
+   *  materializes first). That is a pre-emption, not a conflict: a per-arm source read compiles to
+   *  a per-arm load on that compiler, so the sunk spelling is one it did not emit from this asm. */
   rereadGlobals?: boolean;
   /** "does the project declare this global volatile?" — a read of a volatile object may NOT be
    *  duplicated or moved, so the axis above refuses on one. Answers false for a symbol the map
@@ -57,86 +718,273 @@ export interface AnalyzeOptions {
    *  always had: without a declaration nothing here can know, and the differ referees the extra
    *  load. Where the map DOES know, the axis is silent about it rather than wrong. */
   volatileGlobal?: (name: string) => boolean;
+  /** The in-place-join axis (rank.ts `/inplace`). A load whose result is a `cond_br` successor
+   *  ARG feeds a merge: rendered inline it has no name, so the merge param mints a fresh variable
+   *  and BOTH arms must assign it. Materialized, the naming walk can home the merge in the load's
+   *  own variable, the identity arm elides, and the `if` renders one-sided — `v = *p; if (v > 31)
+   *  v = 32;` — which is also what reuses the load's register when recompiled. Whether the source
+   *  spelled the temp or the overwrite is not derivable from the asm, so this is a differ-refereed
+   *  candidate axis, never a default. Plain `br` args (loop-carried values) are out of scope:
+   *  their homes are the loop-param machinery's question. */
+  materializeJoinFeeds?: boolean;
+  /** The address-home axis (rank.ts `/addr-home`). A pure computed address the asm derived ONCE
+   *  and dereferenced at 2+ sites renders, by default, re-derived at each use — and the loads
+   *  through it re-read per use — where the original source may have spelled a pointer local plus
+   *  scalar temps (`u8 *entry = (u8 *)((a0 << 2) + base); type = entry[1]; idx = entry[0];`).
+   *  The two spellings are codegen-visible (the re-derive folds each deref offset into its OWN
+   *  pool literal; the home shares one base register across `[rN, #k]` accesses) and which one
+   *  the source used is not derivable from asm, so this is a differ-refereed candidate axis,
+   *  never a default. With it on: a non-const pure value whose MERGE CLASS is consumed ONLY as
+   *  the base operand of 2+ memory accesses materializes (`sharedBaseClasses` — the class is what
+   *  makes a base the arms derive and the join dereferences one register rather than three
+   *  single-use values), and so does a multi-render `load` through such a base.
+   *  Both rules only ADD materialization, which preserves semantics FOR THE VALUES THE SCOPE
+   *  ADMITS — the gaddr/laddr cone refusal is the soundness half of that claim (rendered
+   *  standalone, an address cone's value changes: the byte-stride cast lives at the use — see
+   *  coneHoldsAddr; those bases stay with l3/basecse.ts, scopebase.ts and nearbase.ts), and the
+   *  multi-block-loop-header seat refusal is the decline-avoidance half (same as
+   *  liveAcrossLoop's). An L3 re-spell could not host this axis: by structuring's end the loads
+   *  have already rendered per-use inside separate arms, and only this phase's positioned
+   *  memory model can merge them into pre-branch temps. */
+  homeSharedAddresses?: boolean;
+  /** The loop-expression-home axis (rank.ts `/expr-home`). A pure computed value defined outside
+   *  a loop and consumed by 2+ distinct ops, at least one of them inside that loop, is one the
+   *  compiler holds in a (callee-saved) register across the iterations — it never re-derives per
+   *  use in a loop —
+   *  where the default renders it re-derived at each use; the source may have spelled a typed
+   *  local (`u32 size = 16 << t;` driving a loop bound, a product and a shift). The home's
+   *  declared type is the IR value's recovered type, so a u32 value's compares stay unsigned
+   *  through the local. Straight-line multi-use values stay OUT (the small-constant class the
+   *  const-across-call scope's note records); shared memory-access bases are `/addr-home`'s.
+   *  Same refusals as that axis: gaddr/laddr cones and multi-block-loop-header seats. Adding
+   *  materialization preserves semantics for the admitted values, exactly as above. */
+  homeLoopExprs?: boolean;
+  /** The derived-read-home axis (rank.ts `/derived-home`). A memory read whose value is not used
+   *  directly but through a pure computation with 2+ consumers puts the home on the WRONG node:
+   *  the read materializes (its consumer resolves no single render position) and the computation
+   *  then re-derives from that local at every use, where the asm computed it ONCE — the read's
+   *  register died at the computation and the DERIVED value is what a register carried on
+   *  (`eor r1,r1,r0` keeps `0x3FF ^ REG_KEYINPUT`, never the raw halfword). With this on, a pure
+   *  non-const value with 2+ consumers whose operand cone bottoms out at a memory read
+   *  materializes instead; the read then renders exactly once, inside the home.
+   *
+   *  A differ-refereed axis, not a fix: agbcc CSEs a re-derived expression back to one
+   *  instruction often enough that both spellings do compile, and which one the source spelled is
+   *  not derivable. What the cone's read supplies is the evidence the straight-line case otherwise
+   *  lacks — `/expr-home` takes a loop as proof the value stayed in a register, and a
+   *  freely-re-derivable value (the small-constant class) has no such proof at all, but a value
+   *  standing on a read the source could not repeat is one the asm computed from a single access.
+   *
+   *  Refusals, all of them about the ONE access the home must reproduce. A cone read used anywhere
+   *  but the cone (its second use resolves a second render position, so the read would render
+   *  twice). A cone read outside this value's own block, barred from it by a write, or barred by a
+   *  read outside the cone (moving the access across a branch, into a loop, or past another access
+   *  changes which paths read, how often, and in what order). A cone crossing a `call` (homing
+   *  would move a side effect). A standalone gaddr/laddr in the cone, or a value that is ITSELF an
+   *  address (rendered standalone the byte-stride cast lands outside the sum — see
+   *  `rendersAsAddress`). And the multi-block-loop-header seat the sibling axes refuse. */
+  homeDerivedReads?: boolean;
+  /** The merge-feed-home axis (rank.ts `/merge-home`). A pure value one join's incoming edges
+   *  render into the SAME parameter slot from 2+ places materializes at its def: the copy machinery
+   *  has no name to reference, so the default re-derives the whole expression per arm
+   *  (`m = (-(b & 1) | b & 1) >> 31 & 0x400;` in both) and agbcc if-converts each copy. The scope
+   *  and every refusal it states are `mergeFeedHomes` above, which the enumeration gate also runs. */
+  homeMergeFeeds?: boolean;
+  /** DEF-BLOCK PLACEMENT for memory reads — WHERE the read happens, not where the value lives.
+   *  The sibling of the homing axes above: there the question is which register or offset holds a
+   *  value, here which BLOCK performs the read. A read whose every render sits in a block its own
+   *  block STRICTLY DOMINATES has no rule at all above — those axes want 2+ consumers or a shared
+   *  base — so it sinks and each arm re-reads it: a second load either way, plus a second pool
+   *  literal when the address folded to a constant.
+   *
+   *  A per-compiler DATA lever (TargetDescription.compilerBehaviors `readsStayWhereWritten`), not
+   *  a differ-refereed axis, and that distinction is the argument: an axis exists where the asm
+   *  UNDERDETERMINES the source, some pass having collapsed two spellings onto one output
+   *  (`/uns-cmp`'s non-negativity proof is the type case). Where the compiler emits a read in the
+   *  block the source spelled it in, re-spelling it at the block the asm read in reproduces that
+   *  asm while the sunk spelling is one it emits only for a source that read per arm — nothing for
+   *  the differ to referee, and the extra candidate is pure cost. Which compilers may declare
+   *  that, and on what evidence, is at the target field; absent ⇒ the rule stands down entirely.
+   *
+   *  Materializing is the conservative DIRECTION — back to the read's own def position, never
+   *  forward past a write — so it adds no barrier scan; SINKING is what needs one (the
+   *  multi-render rule below) and that scan stays. It is NOT sound by construction: the def block
+   *  is the asm's read block only while nothing moved the def and a branch really does lie
+   *  between, which is what the refusals are for. Getting that wrong emits a read on a path the
+   *  asm never ran it on.
+   *
+   *  Five refusals:
+   *    • a multi-block loop HEADER, whose test-at-top condition seats no temp — `multiBlockHeaders`
+   *    • a FALL-THROUGH between def and render, where the dominance is the frontend's
+   *      block-per-label and no branch separates them — `fallThroughSeam`
+   *    • a loop PREHEADER of the loop the renders are in, where loop invariant motion parks a read
+   *      the source wrote in the BODY — `preheaderOfRenderLoop`
+   *    • a read C evaluates only under a `&&`/`||`, whose def block raise/shortcircuit.ts made —
+   *      `shortCircuitGuarded`
+   *    • a read that IS a block parameter's incoming copy, where the seat manufactures a copy the
+   *      compiler never emitted — `onlyFeedsBlockParams` */
+  readsStayWhereWritten?: boolean;
 }
 
-export function analyze(fn: Fn, returnsVoid: boolean, opts: AnalyzeOptions = {}): StructureAnalysis {
-  const { defs, rereadGlobals = false, volatileGlobal } = opts;
-  // ── use registry ────────────────────────────────────────────────────────────────────────
-  // Every use of a value, POSITIONED: the consuming op and its block/index. Successor args are
-  // uses AT the terminator (they render in argAssigns at block end). A void function's `ret`
-  // operand is a phantom, not a real use — skipping it lets a call whose result ONLY flows into
-  // the suppressed return read as a dead (side-effect) call, so `sideEffects()` emits it.
-  // One operand SLOT = one entry (an op reading a value twice records two uses — that count is
-  // what decides whether an inlined call would EXECUTE twice).
-  const useSitesOf = new Map<Value, UseSite[]>();
-  const opIndex = new Map<Op, number>();
-  const opBlock = new Map<Op, Block>();
-  const blockPos = new Map<Block, number>();
-  for (const b of fn.blocks) {
-    blockPos.set(b, blockPos.size);
-    b.ops.forEach((op, i) => {
-      opIndex.set(op, i);
-      opBlock.set(op, b);
-    });
+/** THE def-block placement rule's loop refusal: is `b` a PREHEADER of some loop whose body holds
+ *  one of the render blocks — outside the body, and a predecessor of the header?
+ *
+ *  Loop invariant motion (loop.c, which agbcc does compile and run at -O2) is the pass whose
+ *  landing spot the source could not have spelled: it parks the read BELOW the loop guard, where
+ *  a read the source wrote above the loop sits above it. Compiled,
+ *  `for (i=0;i<n;i++) t += gK*i;` emits `mov r2,#0 / cmp r2,r3 / bge .L4 / ldr r0,.L8 /
+ *  ldr r4,[r0]` — guard first, read after — while hoisting `gK` into a local above the loop by
+ *  hand puts both `ldr`s ahead of the guard. So a read in the preheader is evidence of a read in
+ *  the BODY, and inferring def-block placement there spells the one source the asm rules out.
+ *  (With an aliasing store in the loop agbcc hoists only the address constant and leaves the
+ *  `ldr` in the body — same conclusion, weaker premise.)
+ *
+ *  Narrow on purpose: a loop merely lying between def and render is not this shape. */
+function preheaderOfRenderLoop(
+  loopBodies: readonly { header: Block; body: Set<Block> }[],
+  b: Block,
+  renders: readonly Block[],
+): boolean {
+  return loopBodies.some(
+    (L) => !L.body.has(b) && successorsOf(b).includes(L.header) && renders.some((x) => L.body.has(x)),
+  );
+}
+
+/** THE def-block placement rule's seam refusal: does a FALL-THROUGH alone put a render below
+ *  `b` — a chain of unconditional `br` edges, each into a block whose only predecessor is the
+ *  one before it?
+ *
+ *  The frontends start a block at every LABEL, so a label nothing branches to cuts one straight
+ *  line of asm in two and the upper half dominates the lower with no control flow between. The
+ *  rule's premise is that the compiler will not move a read ACROSS a branch, which says nothing
+ *  there — while WITHIN a straight line agbcc picks the order itself. Compiled, klonoa's
+ *  StreamCmd_SetMusicParams (a stray `sub_0804E9AC:` between its last `ldrh r2,[r4]` and the
+ *  `bl` consuming it) assembles byte-identical to its object from the inlined read and four
+ *  bytes off from the named one, which swaps that `ldrh` with the `ldr r0,pool` beside it. Of
+ *  the rule's 20 firings over the 464 klonoa agbcc functions, 13 were across such a seam. */
+function fallThroughSeam(predsOf: Map<Block, Block[]>, b: Block, renders: readonly Block[]): boolean {
+  const seen = new Set<Block>([b]);
+  for (let cur = b; ;) {
+    const t = cur.ops[cur.ops.length - 1];
+    if (t?.opcode !== 'br') {
+      return false;
+    }
+    const next = t.successors[0].block;
+    if (predsOf.get(next)!.length !== 1 || seen.has(next)) {
+      return false;
+    }
+    if (renders.includes(next)) {
+      return true;
+    }
+    seen.add(next);
+    cur = next;
   }
-  // Linear program position (block order, then op order): a call "between" a def and a use is one
-  // whose position lies strictly between them. This is a PROXY for "a call on a def→use path", not
-  // the real dataflow: it checks position order only, not reachability. It can therefore FALSE-
-  // POSITIVE — a call in a forward SIBLING branch (never traversed on the def→use path) still sits
-  // between them by position (e.g. `def; if(c){call;ret} else {…use…use}`), and back-edges are not
-  // modelled either. That is SAFE ONLY BECAUSE the caller materializes exactly `const` ops: a const
-  // is a relocation-invariant leaf whose def dominates every use, so binding it to a local is
-  // UNCONDITIONALLY semantics-preserving on every path — a false positive costs at most a match (an
-  // extra `v =` the compiler would have re-inlined), caught by the zero-lost gate, never wrong C.
-  // RE-VERIFY this before widening the whitelist to any value that is not path-independent or that
-  // carries a use-site cast (an address computation `&g + i`), for which a false positive is unsound.
-  const linPos = (op: Op): number => blockPos.get(opBlock.get(op)!)! * 1e6 + opIndex.get(op)!;
-  const callPos: number[] = [];
-  for (const b of fn.blocks) {
-    for (const op of b.ops) {
-      if (op.opcode === 'call') {
-        callPos.push(linPos(op));
+}
+
+/** THE def→render path discipline. May an op `isWrite` accepts execute between `def` and a
+ *  statement at `render`, on any def-avoiding path? The def block's tail, the render block's head,
+ *  and every between-block on a path; a path re-crossing the def is the NEXT dynamic instance and
+ *  does not count. Path-based on purpose: `fn.blocks` is ADDRESS order, so a linear-position scan
+ *  misses a block laid out after the render that executes between def and render on the taken path
+ *  (an audit round broke exactly that way). */
+function makeMemWriteBetween(deps: {
+  opBlock: Map<Op, Block>;
+  opIndex: Map<Op, number>;
+  reachAvoiding: (from: Block, avoid: Block) => Set<Block>;
+}): (def: Op, render: { blk: Block; idx: number }, isWrite: (x: Op) => boolean) => boolean {
+  const { opBlock, opIndex, reachAvoiding } = deps;
+  return (def: Op, render: { blk: Block; idx: number }, isWrite: (x: Op) => boolean): boolean => {
+    const b = opBlock.get(def)!;
+    const oi = opIndex.get(def)!;
+    const wDirty = (list: Op[], from: number, to: number): boolean => {
+      for (let k = from; k < to; k++) {
+        if (isWrite(list[k])) {
+          return true;
+        }
+      }
+      return false;
+    };
+    // Same block: the only def-avoiding path is the straight line between the two indices
+    // (leaving and re-entering the block re-crosses the def). A render BEFORE the def cannot
+    // happen — within a block, uses follow defs — and falls through to the path walk, whose
+    // answer is the conservative one.
+    if (render.blk === b && oi < render.idx) {
+      return wDirty(b.ops, oi + 1, render.idx);
+    }
+    if (wDirty(b.ops, oi + 1, b.ops.length) || wDirty(render.blk.ops, 0, render.idx)) {
+      return true;
+    }
+    for (const x of reachAvoiding(b, b)) {
+      if (x === render.blk && !reachAvoiding(render.blk, b).has(render.blk)) {
+        continue; // acyclic render block: head checked
+      }
+      if (x !== render.blk && !reachAvoiding(x, b).has(render.blk)) {
+        continue; // not on a def→render path
+      }
+      if (wDirty(x.ops, 0, x.ops.length)) {
+        return true;
+      }
+    }
+    return false;
+  };
+}
+
+/** INTERDEPENDENT PARALLEL-COPY ARGS. A terminator's successor args are ONE parallel copy
+ *  (argAssigns), whose every read means the PRE-copy value. An arg whose def-tree reads a SIBLING
+ *  arg of the same edge cannot read it by name there — so the sibling's whole expression is
+ *  re-derived inside this arg's copy, and `sequentialize` spills old-value temps to untangle the
+ *  order: arithmetic the compiler performed once, emitted per reader (the coupled-recurrence loop,
+ *  `a += b*c; d += a;`). The read sibling is the register the copy machinery cannot represent —
+ *  its def is the one the caller materializes. The walk crosses pure defs only (a call/load render
+ *  is those rules' question) and stops at params, which always carry a name. */
+function copyInterdependentValues(fn: Fn, defOf: Map<Value, Op>): Set<Value> {
+  const copyInterdependent = new Set<Value>();
+  const pureReads = (root: Value, acc: Set<Value>) => {
+    const stack = [root];
+    while (stack.length) {
+      const x = stack.pop()!;
+      const d = defOf.get(x);
+      if (!d || acc.has(x)) {
+        continue;
+      }
+      acc.add(x);
+      if (d.opcode !== 'call' && d.opcode !== 'load' && d.opcode !== 'aload') {
+        stack.push(...d.operands);
       }
     }
-  }
-  /** True if `def`'s value is still needed after a call — a call lies strictly between the def and
-   *  one of its `consumers`. Such a value survives in a callee-saved register (a local), which is
-   *  what materializing it reproduces. */
-  const liveAcrossCall = (def: Op, consumers: Op[]): boolean => {
-    const dp = linPos(def);
-    const usePos = consumers.map(linPos);
-    return callPos.some((c) => c > dp && usePos.some((u) => u > c));
   };
   for (const b of fn.blocks) {
-    b.ops.forEach((op, i) => {
-      if (returnsVoid && op.opcode === 'ret') {
-        return;
+    for (const s of b.ops[b.ops.length - 1]?.successors ?? []) {
+      if (s.args.length < 2) {
+        continue;
       }
-      const site: UseSite = { blk: b, idx: i, op };
-      const add = (v: Value) => {
-        const arr = useSitesOf.get(v);
-        if (arr) {
-          arr.push(site);
-        } else {
-          useSitesOf.set(v, [site]);
+      for (const w of s.args) {
+        const d = defOf.get(w);
+        if (!d || d.opcode === 'call' || d.opcode === 'load' || d.opcode === 'aload') {
+          continue;
         }
-      };
-      for (const u of op.operands) {
-        add(u);
-      }
-      for (const s of op.successors) {
-        for (const a of s.args) {
-          add(a);
+        const reads = new Set<Value>();
+        for (const o of d.operands) {
+          pureReads(o, reads);
+        }
+        for (const v of s.args) {
+          if (v !== w && reads.has(v)) {
+            copyInterdependent.add(v);
+          }
         }
       }
-    });
+    }
   }
+  return copyInterdependent;
+}
 
-  // ── per-block liveness of SSA values ──────────────────────────────────────────────────────
-  // Backward dataflow. Successor args count as uses at the END of the predecessor (they render
-  // in the predecessor's argAssigns), so liveIn(B) means precisely "read at-or-after B's entry".
-  // Consumed by the coalescing interference check: merging two values that are ever
-  // simultaneously live into one variable name is the textbook silent clobber.
+/** PER-BLOCK LIVENESS of SSA values — backward dataflow. Successor args count as uses at the END
+ *  of the predecessor (they render in the predecessor's argAssigns), so liveIn(B) means precisely
+ *  "read at-or-after B's entry". Consumed by the coalescing interference check: merging two values
+ *  that are ever simultaneously live into one variable name is the textbook silent clobber.
+ *
+ *  `returnsVoid` suppresses the phantom `ret` operand, the same way the use registry does. */
+function blockLiveIn(fn: Fn, returnsVoid: boolean): Map<Block, Set<Value>> {
   const liveIn = new Map<Block, Set<Value>>();
   for (const b of fn.blocks) {
     liveIn.set(b, new Set());
@@ -177,17 +1025,22 @@ export function analyze(fn: Fn, returnsVoid: boolean, opts: AnalyzeOptions = {})
       }
     }
   }
+  return liveIn;
+}
 
-  // ── the effect-ordering model — inline-at-use barriers ────────────────────────────────────
-  // `expr()` renders a def's computation AT ITS USE, which silently MOVES it: a call executes
-  // once per rendered copy (`foo(a0)+foo(a0)`), a load reads memory at the render point (it can
-  // textually sink past an aliasing store). The model: a call/load/aload def may inline ONLY
-  // when rendering cannot change behavior — exactly one render position, and the program-order
-  // gap between def and render crosses no memory write (loads) / no memory access at all (calls,
-  // whose own reads+writes must not reorder against anything). Every other case gets a NAMED
-  // TEMP assigned at the def's own program position (sideEffects) — which is precisely the
-  // register the compiler used.
-  const materialize = new Set<Op>();
+/** FORWARD REACHABILITY over the CFG, in the two flavours the materialization rules ask for.
+ *
+ *  `reachFrom` is the plain successors-transitive set, excluding the start block itself, cached per
+ *  factory — which is why this is a factory: the cache belongs to one `analyze` call.
+ *
+ *  `reachAvoiding` never passes THROUGH `avoid` — the def-block-avoiding variant for per-iteration
+ *  path checks: a path that re-enters the def's block re-executes the def, so writes on it belong
+ *  to the NEXT dynamic instance (which re-renders anyway) and must not count against this one.
+ *  Uncached (per-decision graphs are small). */
+function makeReach(): {
+  reachFrom: (b: Block) => Set<Block>;
+  reachAvoiding: (from: Block, avoid: Block) => Set<Block>;
+} {
   const reachCache = new Map<Block, Set<Block>>();
   const reachFrom = (b: Block): Set<Block> => {
     let r = reachCache.get(b);
@@ -228,6 +1081,133 @@ export function analyze(fn: Fn, returnsVoid: boolean, opts: AnalyzeOptions = {})
     }
     return r;
   };
+  return { reachFrom, reachAvoiding };
+}
+
+export function analyze(fn: Fn, returnsVoid: boolean, opts: AnalyzeOptions = {}): StructureAnalysis {
+  const {
+    defs,
+    dom,
+    rereadGlobals = false,
+    volatileGlobal,
+    materializeJoinFeeds = false,
+    homeSharedAddresses = false,
+    homeLoopExprs = false,
+    homeDerivedReads = false,
+    homeMergeFeeds = false,
+    readsStayWhereWritten = false,
+  } = opts;
+  // ── use registry ────────────────────────────────────────────────────────────────────────
+  // Every use of a value, POSITIONED: the consuming op and its block/index. Successor args are
+  // uses AT the terminator (they render in argAssigns at block end). A void function's `ret`
+  // operand is a phantom, not a real use — skipping it lets a call whose result ONLY flows into
+  // the suppressed return read as a dead (side-effect) call, so `sideEffects()` emits it.
+  // One operand SLOT = one entry (an op reading a value twice records two uses — that count is
+  // what decides whether an inlined call would EXECUTE twice).
+  const useSitesOf = new Map<Value, UseSite[]>();
+  const opIndex = new Map<Op, number>();
+  const opBlock = new Map<Op, Block>();
+  const blockPos = new Map<Block, number>();
+  for (const b of fn.blocks) {
+    blockPos.set(b, blockPos.size);
+    b.ops.forEach((op, i) => {
+      opIndex.set(op, i);
+      opBlock.set(op, b);
+    });
+  }
+  // Linear program position (block order, then op order): a call "between" a def and a use is one
+  // whose position lies strictly between them. This is a PROXY for "a call on a def→use path", not
+  // the real dataflow: it checks position order only, not reachability. It can therefore FALSE-
+  // POSITIVE — a call in a forward SIBLING branch (never traversed on the def→use path) still sits
+  // between them by position (e.g. `def; if(c){call;ret} else {…use…use}`), and back-edges are not
+  // modelled either. That is SAFE ONLY BECAUSE the caller materializes exactly `const` ops: a const
+  // is a relocation-invariant leaf whose def dominates every use, so binding it to a local is
+  // UNCONDITIONALLY semantics-preserving on every path — a false positive costs at most a match (an
+  // extra `v =` the compiler would have re-inlined), caught by the zero-lost gate, never wrong C.
+  // RE-VERIFY this before widening the whitelist to any value that is not path-independent or that
+  // carries a use-site cast (an address computation `&g + i`), for which a false positive is unsound.
+  const linPos = (op: Op): number => blockPos.get(opBlock.get(op)!)! * 1e6 + opIndex.get(op)!;
+  const callPos: number[] = [];
+  for (const b of fn.blocks) {
+    for (const op of b.ops) {
+      if (op.opcode === 'call') {
+        callPos.push(linPos(op));
+      }
+    }
+  }
+  // Which values ride a branch edge as a successor ARG, in two scopes with two different readers.
+  //
+  // `branchArgFed` is EVERY multi-successor terminator's, and it is a correctness fact: an edge
+  // argument is emitted INSIDE the arm it belongs to, so a value that renders only there runs only
+  // on that path. For a call that is an effect the IR performs unconditionally and the C performs
+  // sometimes, and a `switch_br` does it as readily as a `cond_br`. Plain `br` args are excluded
+  // because their edge copy is on the one path that reaches the successor, so nothing conditional
+  // happens to them.
+  //
+  // `condBrArgFed` is the `/inplace` axis's narrower reading — a preference about where a load's
+  // value is homed, whose own scope note is on AnalyzeOptions.materializeJoinFeeds.
+  const branchArgFed = new Set<Value>();
+  const condBrArgFed = new Set<Value>();
+  for (const b of fn.blocks) {
+    for (const op of b.ops) {
+      if (op.successors.length > 1) {
+        for (const s of op.successors) {
+          for (const a of s.args) {
+            branchArgFed.add(a);
+            if (op.opcode === 'cond_br') {
+              condBrArgFed.add(a);
+            }
+          }
+        }
+      }
+    }
+  }
+  /** True if `def`'s value is still needed after a call — a call lies strictly between the def and
+   *  one of its `consumers`. Such a value survives in a callee-saved register (a local), which is
+   *  what materializing it reproduces. */
+  const liveAcrossCall = (def: Op, consumers: Op[]): boolean => {
+    const dp = linPos(def);
+    const usePos = consumers.map(linPos);
+    return callPos.some((c) => c > dp && usePos.some((u) => u > c));
+  };
+  for (const b of fn.blocks) {
+    b.ops.forEach((op, i) => {
+      if (returnsVoid && op.opcode === 'ret') {
+        return;
+      }
+      const site: UseSite = { blk: b, idx: i, op };
+      const add = (v: Value) => {
+        const arr = useSitesOf.get(v);
+        if (arr) {
+          arr.push(site);
+        } else {
+          useSitesOf.set(v, [site]);
+        }
+      };
+      for (const u of op.operands) {
+        add(u);
+      }
+      for (const s of op.successors) {
+        for (const a of s.args) {
+          add(a);
+        }
+      }
+    });
+  }
+
+  const liveIn = blockLiveIn(fn, returnsVoid);
+
+  // ── the effect-ordering model — inline-at-use barriers ────────────────────────────────────
+  // `expr()` renders a def's computation AT ITS USE, which silently MOVES it: a call executes
+  // once per rendered copy (`foo(a0)+foo(a0)`), a load reads memory at the render point (it can
+  // textually sink past an aliasing store). The model: a call/load/aload def may inline ONLY
+  // when rendering cannot change behavior — exactly one render position, and the program-order
+  // gap between def and render crosses no memory write (loads) / no memory access at all (calls,
+  // whose own reads+writes must not reorder against anything). Every other case gets a NAMED
+  // TEMP assigned at the def's own program position (sideEffects) — which is precisely the
+  // register the compiler used.
+  const materialize = new Set<Op>();
+  const { reachFrom, reachAvoiding } = makeReach();
   // Where a value's expression is ultimately EMITTED: the anchored consumer (statement op,
   // terminator, materialized def) it inlines into, transitively through single-use pure ops.
   // null = renders in several places / unresolvable (treated conservatively by the caller).
@@ -301,47 +1281,193 @@ export function analyze(fn: Fn, returnsVoid: boolean, opts: AnalyzeOptions = {})
     emitPosSetCache.set(op, res);
     return res;
   };
-  // THE def→render path discipline — one implementation, three callers (the two materialization
-  // rules below and structure.ts's bitfield fold, which imports it). May an op `isWrite` accepts
-  // execute between `def` and a statement at `render`, on any def-avoiding path? The def block's
-  // tail, the render block's head, and every between-block on a path; a path re-crossing the def
-  // is the NEXT dynamic instance and does not count. Path-based on purpose: `fn.blocks` is ADDRESS
-  // order, so a linear-position scan misses a block laid out after the render that executes
-  // between def and render on the taken path (an audit round broke exactly that way).
-  const memWriteBetween = (def: Op, render: { blk: Block; idx: number }, isWrite: (x: Op) => boolean): boolean => {
-    const b = opBlock.get(def)!;
-    const oi = opIndex.get(def)!;
-    const wDirty = (list: Op[], from: number, to: number): boolean => {
-      for (let k = from; k < to; k++) {
-        if (isWrite(list[k])) {
-          return true;
-        }
-      }
+  const memWriteBetween = makeMemWriteBetween({ opBlock, opIndex, reachAvoiding });
+  const defOf = defs ?? defOpMap(fn);
+  const copyInterdependent = copyInterdependentValues(fn, defOf);
+  // ── natural loops, for the live-across-a-loop rule ────────────────────────────────────────
+  // From the caller's dominators (a back edge is `latch → header` with the header dominating the
+  // latch); the body is the backward closure from the latch. With no `dom` the rule stands
+  // down — the same posture as the `defs`-carried rules.
+  const predsOf = predecessors(fn);
+  const loopBodies = dom ? naturalLoops(fn, dom, predsOf) : [];
+  /** The def's value enters some loop's header live and every consumer sits outside that loop,
+   *  as does the def: the value is carried ACROSS the loop, not into it. */
+  // Never for a def in a MULTI-BLOCK loop header: a test-at-top `while`'s condition has no seat
+  // for a materialized temp (the structurer's headerPure gate), so materializing there trades a
+  // structuring function for a decline. Self-loop headers stay eligible — their kept-guard
+  // do-while form hosts the temp.
+  const multiBlockHeaders = new Set(loopBodies.filter((L) => L.body.size > 1).map((L) => L.header));
+  const liveAcrossLoop = (def: Op, r: Value, consumers: Op[]): boolean =>
+    !multiBlockHeaders.has(opBlock.get(def)!) &&
+    loopBodies.some(
+      (L) =>
+        liveIn.get(L.header)!.has(r) &&
+        !L.body.has(opBlock.get(def)!) &&
+        consumers.every((c) => !L.body.has(opBlock.get(c)!)),
+    );
+  /** Would naming THIS op's own result change its value? Yes when the result is an ADDRESS built
+   *  over a gaddr/laddr: rendered standalone an `&g + i` loses the memAccess's inline byte-stride
+   *  cast, and the cast-aware base machinery in l3/ serves those bases instead. Asked by the rules
+   *  that home an address; a rule homing the SCALAR a load returns is not this case — the name
+   *  holds the loaded value and the address stays inline at the deref, which is why the load rules
+   *  (live-across-a-loop, join feeds, /addr-home's, def-block placement) do not ask it. */
+  const addressCone = (op0: Op): boolean => coneHoldsAddr(op0, defOf);
+  // ── the address-home axis's scope ─────────────────────────────────────────────────────────
+  // Over the MERGE CLASS, so a base the arms derive and the join dereferences counts as the one
+  // register it is — see `sharedBaseClasses`. Computed once, and only under the axis.
+  const sharedBaseClass = homeSharedAddresses ? sharedBaseClasses(fn, returnsVoid) : new Set<Value>();
+  // The same question asked of the VALUE ALONE, which is what the two axes below need: their
+  // "shared bases stay the address-home scope's" exclusion is a hand-off between scopes, and
+  // widening it to the class would make them refuse values the address-home scope only claims
+  // when its own axis is ON — a candidate lost with no candidate gained. Where both axes run the
+  // two scopes may claim one value, which is a no-op: `materialize` is a set and the address-home
+  // scope, running first, is what registers `axisHomedBases`.
+  const usedOnlyAsSharedBase = (v: Value): boolean => {
+    const sites = useSitesOf.get(v) ?? [];
+    const consumers = new Set(sites.map((s) => s.op));
+    return (
+      consumers.size >= 2 &&
+      [...consumers].every(
+        (c) => MEM_BASE_OPS.has(c.opcode) && c.operands[0] === v && !c.operands.some((o, i) => i > 0 && o === v),
+      )
+    );
+  };
+  // The merge-feed-home axis's ops, settled BEFORE the fixpoint: the scope reads the IR alone (no
+  // render positions, no `materialize`), so it cannot change as the set grows. Unlike the rules
+  // that stand down without the caller's `dom`, this one computes its own: rank.ts has already
+  // admitted the candidate on `hasMergeFeedHome`, which runs the same scope, so standing down here
+  // would be a refusal nothing reports.
+  let mergeFeedOps = new Set<Op>();
+  if (homeMergeFeeds) {
+    const mdom = dom ?? dominators(fn);
+    const mloops = mdom === dom ? loopBodies : naturalLoops(fn, mdom, predsOf);
+    mergeFeedOps = mergeFeedHomes(fn, mdom, defOf, new Set(mloops.flatMap((L) => [...L.body])));
+  }
+  /** result values the address-home axis materialized — the load rule's admission key */
+  const axisHomedBases = new Set<Value>();
+  /** The loop-expression-home axis's scope: 2+ distinct consumers of the value, at least one of
+   *  them inside a loop the def's block is outside (loop model = the caller's dominators; absent ⇒
+   *  never).
+   *
+   *  The two counts carry different halves of the evidence. ONE consumer inside the loop is what
+   *  says the compiler held the value in a callee-saved register across the iterations — it does
+   *  not re-derive per use in a loop, and a value crossing the loop boundary is pinned for the
+   *  whole nest. The SECOND consumer, anywhere, is what makes the home observable at all: a
+   *  single-use value inlines at its one use with the same bytes either way, so homing it can only
+   *  add a copy. Values consumed only OUTSIDE any loop are the straight-line class `/derived-home`
+   *  serves on its own evidence. */
+  const loopSharedConsumers = (v: Value, defBlk: Block): boolean => {
+    const consumers = [...new Set((useSitesOf.get(v) ?? []).map((s) => s.op))];
+    return (
+      consumers.length >= 2 &&
+      loopBodies.some((L) => !L.body.has(defBlk) && consumers.some((c) => L.body.has(opBlock.get(c)!)))
+    );
+  };
+  /** The derived-read-home axis's scope: does `op0` stand on a memory READ that may render at
+   *  `op0`'s own position?
+   *
+   *  The cone walk STOPS at a read — its address stays inline at the deref, so nothing below it is
+   *  ever rendered standalone — and refuses on anything else that cannot move to `op0`: a `call`
+   *  (a side effect), and a gaddr/laddr reached OUTSIDE a read's address (rendered standalone an
+   *  `&g + i` loses the memAccess's byte-stride cast). `coneHoldsAddr` cannot answer this: it
+   *  crosses reads, so it refuses every value standing on a named global's load — which is this
+   *  axis's whole clientele.
+   *
+   *  At least one read is required. A value derivable from locals and constants alone is one the
+   *  compiler re-materializes for free, and homing it only adds copies — the small-constant class
+   *  the const scope's note records.
+   *
+   *  Each read must then sit in `op0`'s OWN BLOCK, and reach `op0`'s position with nothing that
+   *  writes memory able to execute in between — homing renders the read at `op0`, so both are
+   *  about moving it there.
+   *
+   *  A read OUTSIDE the cone bars it too. The default's "loads never bar a load" holds for reads
+   *  the compiler leaves unsequenced inside ONE expression, which is what the cone's own reads
+   *  become; a foreign read renders in a different statement, so moving past it reorders two
+   *  accesses — for two MMIO cells (this axis's clientele) an observable swap, as when `A`'s
+   *  derived value sits below `B`'s and homing both puts `B`'s read first.
+   *
+   *  And each read's value must go NOWHERE BUT the cone: exactly one use site. Homing resolves a
+   *  render position for a read that had none, so a SECOND use resolves a second one, and the
+   *  multi-render load rule then inlines the read at BOTH — two accesses where the asm has one
+   *  `ldrh`, which for a volatile cell is precisely the duplication `volatileGlobal` refuses. Two
+   *  homed values over one read is the same shape from the other side (each is the other's second
+   *  use), so one test covers both. This is what makes "renders once, inside the home" a property
+   *  rather than an aspiration: without it the axis silently doubles a hardware read.
+   *
+   *  The same block is what makes the axis's claim true at all: the register handoff it reproduces
+   *  is one straight-line run of the asm, `ldrh` into `eor` into three uses. Across blocks WHICH
+   *  BLOCK reads is `readsStayWhereWritten`'s question, not this one, and answering it here goes
+   *  wrong in both directions — a value below a branch pulls the read into an arm that may not
+   *  run, a value inside a loop pulls it in to run per iteration. Neither is a write, so no
+   *  barrier sees either; for an ordinary cell they are worse spellings, and for a volatile one
+   *  they are a missing access and a duplicated one. */
+  const standsOnMovableRead = (op0: Op, blk: Block): boolean => {
+    const reads = readCone(op0, defOf);
+    if (reads === null) {
       return false;
-    };
-    // Same block: the only def-avoiding path is the straight line between the two indices
-    // (leaving and re-entering the block re-crosses the def). A render BEFORE the def cannot
-    // happen — within a block, uses follow defs — and falls through to the path walk, whose
-    // answer is the conservative one.
-    if (render.blk === b && oi < render.idx) {
-      return wDirty(b.ops, oi + 1, render.idx);
     }
-    if (wDirty(b.ops, oi + 1, b.ops.length) || wDirty(render.blk.ops, 0, render.idx)) {
-      return true;
-    }
-    for (const x of reachAvoiding(b, b)) {
-      if (x === render.blk && !reachAvoiding(render.blk, b).has(render.blk)) {
-        continue; // acyclic render block: head checked
-      }
-      if (x !== render.blk && !reachAvoiding(x, b).has(render.blk)) {
-        continue; // not on a def→render path
-      }
-      if (wDirty(x.ops, 0, x.ops.length)) {
-        return true;
+    const at = { blk, idx: opIndex.get(op0)! };
+    const coneReads = new Set(reads);
+    const bars = (x: Op): boolean =>
+      EFFECTFUL_OPS.has(x.opcode) || ((x.opcode === 'load' || x.opcode === 'aload') && !coneReads.has(x));
+    return (
+      reads.length > 0 &&
+      reads.every(
+        (r) =>
+          opBlock.get(r) === blk && (useSitesOf.get(r.results[0])?.length ?? 0) === 1 && !memWriteBetween(r, at, bars),
+      )
+    );
+  };
+  /** 2+ distinct consuming ops — the multi-use the pure-op rule reads as a reused register. */
+  const multiConsumer = (v: Value): boolean => new Set((useSitesOf.get(v) ?? []).map((s) => s.op)).size >= 2;
+  /** THE def-block placement rule's short-circuit refusal: values C evaluates only under a
+   *  `&&`/`||` — the SECOND operand of every `logic_and`/`logic_or`, and everything it reads.
+   *
+   *  raise/shortcircuit.ts recovers a connective by hoisting the guarded arm's whole pure body,
+   *  memory reads included, into the block ABOVE the branch (its value form and its control-flow
+   *  form both splice that body into the head). ir/opcodes.ts states the safety argument as the
+   *  reason a read is deliberately absent from HOIST_UNSAFE_OPS: the structurer inlines it back
+   *  into the `&&`/`||` right-hand side, where C's own short circuit re-guards it. So for a value
+   *  in that cone the def block is a FOLD ARTIFACT rather than the block the asm read in, and the
+   *  whole premise this rule reads placement under does not hold there. Naming it also breaks the
+   *  re-guard: `p != 0 && *p != 0` would emit `v0 = *p;` above its own null check.
+   *
+   *  An operand[0] cone is unconditional and keeps the rule; only the guarded side is collected. */
+  const shortCircuitGuarded = readsStayWhereWritten ? shortCircuitGuardedValues(fn, defOf) : new Set<Value>();
+  /** THE def-block placement rule's copy refusal: is every use of the value a successor ARGUMENT,
+   *  i.e. is the value nothing but a block parameter's incoming copy?
+   *
+   *  Then the parameter IS the read's home. Inlined, the edge assignment is the read
+   *  (`v1 = mplay->tracks;`); materialized it becomes two names and a copy between them
+   *  (`v0 = mplay->tracks; … v1 = v0;`) where the asm loaded straight into the register the
+   *  parameter became — `ldr r1,[r4,#0x2C]` once, never a `mov`. Six of klonoa's matched m4a
+   *  functions are that shape, and the manufactured copy costs more than the placement gains:
+   *  scored against their own objects, dropping the refusal takes m4aMPlayVolumeControl 26→33,
+   *  m4aMPlayPitchControl 36→39, m4aMPlayLFOSpeedSet 28→31 and FadeOutBody 69→73.
+   *
+   *  Note what this does NOT claim: those reads really do sit above the loop guard in the asm, so
+   *  the placement inference was right and the SPELLING is what fails. Seating the read above the
+   *  guard AND as the loop variable is loop-init hoisting, a capability this rule does not have. */
+  const argUsedValues = new Set<Value>();
+  const operandUsedValues = new Set<Value>();
+  // Both sets serve that rule alone, so they are built only where it can fire — the same posture
+  // `condBrArgFed` takes above (every other target pays nothing for a behavior it never declares).
+  if (readsStayWhereWritten) {
+    for (const b of fn.blocks) {
+      for (const op of b.ops) {
+        for (const v of op.operands) {
+          operandUsedValues.add(v);
+        }
+        for (const sc of op.successors) {
+          for (const v of sc.args) {
+            argUsedValues.add(v);
+          }
+        }
       }
     }
-    return false;
-  };
+  }
+  const onlyFeedsBlockParams = (v: Value): boolean => argUsedValues.has(v) && !operandUsedValues.has(v);
   // Decide in REVERSE program order so a consumer's own materialization is settled before any
   // producer asks for its emit position (SSA: uses follow defs in dominance/layout order) — and
   // iterate to a fixpoint for IR whose block layout does not follow dominance (hand-built IR):
@@ -374,15 +1500,78 @@ export function analyze(fn: Fn, returnsVoid: boolean, opts: AnalyzeOptions = {})
           // across `foo(...)` calls). WITHOUT a call in its live range the const is instead cheaply
           // re-materialized at each use (a bare `movs r, #0` per init), so materializing it would
           // ADD pointless copies and MISS — hence the call gate (the small-constant regression).
-          // Cheap deref casts still land on the `index` node at the use, preserving byte strides;
-          // NON-const pure ops are excluded (an address computation `&g + i` rendered standalone
-          // loses the memAccess's inline `(u8 *)` cast — cast-aware base materialization is separate).
+          // Cheap deref casts still land on the `index` node at the use, preserving byte strides.
+          // Second scope: a NON-const read by a sibling parallel-copy arg (`copyInterdependent`) —
+          // the one place a pure value's inlining is not value-identical rendering but a
+          // re-derivation the copy machinery is forced into. Address cones are excluded from it
+          // (an `&g + i` rendered standalone loses the memAccess's inline `(u8 *)` cast —
+          // cast-aware base materialization is separate), and consts are not in it at all: a
+          // re-derived const is re-materialization, which is the compiler's own behavior.
           const pr = op.results[0];
           if (op.opcode === 'const' && pr && useSitesOf.has(pr)) {
             const cons = [...new Set((useSitesOf.get(pr) ?? []).map((s) => s.op))];
             if (cons.length > 1 && liveAcrossCall(op, cons)) {
               materialize.add(op);
             }
+          } else if (op.opcode !== 'const' && pr && copyInterdependent.has(pr) && !addressCone(op)) {
+            materialize.add(op);
+          }
+          // Folding the FOUR AXIS scopes below into one predicate-parameterized scope is BOOKED
+          // in docs/level-tower.md and deliberately unpaid; what it cannot absorb is named there,
+          // along with the price the fourth one added to it (the gate duplication).
+          // Third scope, under the address-home axis only (see AnalyzeOptions.homeSharedAddresses):
+          // a non-const pure value whose MERGE CLASS is used only as the base of 2+ memory
+          // accesses.
+          if (
+            homeSharedAddresses &&
+            op.opcode !== 'const' &&
+            pr &&
+            sharedBaseClass.has(pr) &&
+            !addressCone(op) &&
+            !multiBlockHeaders.has(b)
+          ) {
+            materialize.add(op);
+            axisHomedBases.add(pr);
+          }
+          // Fourth scope, under the loop-expression-home axis (AnalyzeOptions.homeLoopExprs): a
+          // pure non-const value with 2+ distinct consumers, at least one of them inside a loop the
+          // def sits outside. Shared bases stay the previous scope's (its load rule needs the
+          // registration).
+          if (
+            homeLoopExprs &&
+            op.opcode !== 'const' &&
+            pr &&
+            !usedOnlyAsSharedBase(pr) &&
+            loopSharedConsumers(pr, b) &&
+            !addressCone(op) &&
+            !multiBlockHeaders.has(b)
+          ) {
+            materialize.add(op);
+          }
+          // Fifth scope, under the derived-read-home axis (AnalyzeOptions.homeDerivedReads): a
+          // pure non-const value with 2+ consumers standing on a memory read. Shared bases stay
+          // the third scope's and values consumed across a loop the fourth's; what this one adds
+          // is the straight-line case, which neither reaches.
+          if (
+            homeDerivedReads &&
+            op.opcode !== 'const' &&
+            pr &&
+            !usedOnlyAsSharedBase(pr) &&
+            !rendersAsAddress(op) &&
+            multiConsumer(pr) &&
+            standsOnMovableRead(op, b) &&
+            !multiBlockHeaders.has(b)
+          ) {
+            materialize.add(op);
+          }
+          // Sixth scope, under the merge-feed-home axis (AnalyzeOptions.homeMergeFeeds) —
+          // `mergeFeedHomes` above. The only one of the FOUR AXIS scopes that admits a `const`; the
+          // other const clientele is the first scope, a const live across a call. For the sibling
+          // axes a re-derived const is re-materialization, the compiler's own behavior, while a
+          // const the arms of a branch merge is one it held in a register across them
+          // (`mov r5, #0` once, not per arm).
+          if (homeMergeFeeds && mergeFeedOps.has(op)) {
+            materialize.add(op);
           }
           continue;
         }
@@ -398,6 +1587,10 @@ export function analyze(fn: Fn, returnsVoid: boolean, opts: AnalyzeOptions = {})
             ? globalCellOf(defs, op.operands[0], op.attrs.off as number)
             : null;
         const barsThisRead = cell && defs && !volatileGlobal?.(cell.name) ? mayWriteGlobal(defs, cell.name) : null;
+        if (materializeJoinFeeds && op.opcode === 'load' && condBrArgFed.has(r)) {
+          materialize.add(op);
+          continue;
+        }
         const sites = useSitesOf.get(r)!;
         const consumers = [...new Set(sites.map((s) => s.op))];
         const isCall = op.opcode === 'call';
@@ -406,6 +1599,72 @@ export function analyze(fn: Fn, returnsVoid: boolean, opts: AnalyzeOptions = {})
           materialize.add(op);
           continue;
         }
+        // …and ONCE means once on EVERY path the asm runs it on, which the multi-site rule above
+        // does not say. `anchored` calls a terminator a single render position, but a branch's edge
+        // copies are emitted inside the ARMS, so a call whose only consumer is an edge argument
+        // renders in one arm and is skipped on the others. Compiled: `s32 t = f2(a); if (a > 0) {
+        // return t; } return 0;` gives `bl f2` ahead of the `cmp`, and inlined at the edge the
+        // recovered C calls `f2` only in the `else`; a `switch_br` arm hides it the same way.
+        // Materializing puts it back at the position the asm executed it. Sole-use only in
+        // practice — a second use already materialized above — and it is 0 of 2288 sa3 functions,
+        // 2 of 412 klonoa ones.
+        if (isCall && branchArgFed.has(r)) {
+          materialize.add(op);
+          continue;
+        }
+        // Live across a LOOP neither side belongs to: the access ran before the loop and the
+        // value crossed it in a callee-saved register (hipress's `keep = p[1]`, homed in r8 and
+        // touched only by `mov`). Rendering at the use would re-schedule the access to the far
+        // side of the loop — the same refusal liveAcrossCall makes for a call, applied to a loop.
+        if (liveAcrossLoop(op, r, consumers)) {
+          materialize.add(op);
+          continue;
+        }
+        // ── DEF-BLOCK PLACEMENT (readsStayWhereWritten; see AnalyzeOptions) ──────────────────
+        // Every render in a block this one strictly dominates, with a branch between ⇒ the asm
+        // read once above that branch, and re-spelling the read there reproduces it. ONE render
+        // suffices — the short-circuit-into-a-call shape has exactly one — and an unresolvable
+        // render position refuses, as everywhere else. Both memory reads, spelled positively: a
+        // `call` is the enclosing arm's other member and has its own execute-once rules above.
+        if (
+          (op.opcode === 'load' || op.opcode === 'aload') &&
+          readsStayWhereWritten &&
+          dom &&
+          !multiBlockHeaders.has(b) &&
+          !shortCircuitGuarded.has(r) &&
+          !onlyFeedsBlockParams(r)
+        ) {
+          const at = consumers.map((c) => emitPos(c));
+          const rb = at.some((p) => p === null) ? null : [...new Set(at.map((p) => p!.blk))];
+          if (
+            rb &&
+            rb.length > 0 &&
+            rb.every((x) => x !== b && dom.get(x)!.has(b)) &&
+            !preheaderOfRenderLoop(loopBodies, b, rb) &&
+            !fallThroughSeam(predsOf, b, rb)
+          ) {
+            materialize.add(op);
+            continue;
+          }
+        }
+        // Address-home axis: a multi-render load THROUGH a base this axis homed re-reads what the
+        // asm read once into the register the home just reproduced — home the value too, at the
+        // load's own position. Only through axis-homed bases (the fixpoint's later sweep sees them
+        // even though reverse order visits the load first); the general multi-render re-read stays
+        // the default rule below. axisHomedBases registers on the same sweep that materializes the
+        // base — safe because no other rule can pre-empt a value the axis would home: base-slot-only
+        // means no successor-arg use (outside copyInterdependent's read set), and the axis's scope
+        // excludes consts, the const arm's only clientele.
+        if (
+          homeSharedAddresses &&
+          op.opcode === 'load' &&
+          consumers.length > 1 &&
+          axisHomedBases.has(op.operands[0]) &&
+          !multiBlockHeaders.has(b)
+        ) {
+          materialize.add(op);
+          continue;
+        }
         // A MULTI-RENDER load re-reads memory at each render — which is exactly what the original
         // per-use source spelling did (`while (*s != EOS) *d = *s;` reads *s twice per iteration),
         // so it is sound iff every render still sees the def-time memory: NO write anywhere
@@ -428,8 +1687,7 @@ export function analyze(fn: Fn, returnsVoid: boolean, opts: AnalyzeOptions = {})
           continue;
         }
         if (poss.length > 1) {
-          const MW = new Set(['store', 'astore', 'call', 'opaque']);
-          const isWrite = barsThisRead ?? ((x: Op) => MW.has(x.opcode));
+          const isWrite = barsThisRead ?? ((x: Op) => EFFECTFUL_OPS.has(x.opcode));
           if (poss.some((p) => memWriteBetween(op, p!, isWrite))) {
             materialize.add(op);
           }
@@ -449,17 +1707,10 @@ export function analyze(fn: Fn, returnsVoid: boolean, opts: AnalyzeOptions = {})
             return false;
           }
           if (x.opcode === 'store') {
-            // A store to a PROVABLY-DISJOINT slot of the same base never aliases the load: same
-            // base SSA value, both constant offset+width, ranges non-overlapping (the everyday
-            // struct interleave `… = p->field_0; p->field_4 = …`). Anything less certain bars.
-            if (!isCall && op.opcode === 'load' && x.operands[0] === op.operands[0]) {
-              const lo = op.attrs.off as number,
-                lw = op.attrs.width as number;
-              const so = x.attrs.off as number,
-                sw = x.attrs.width as number;
-              if (so + sw <= lo || lo + lw <= so) {
-                return false;
-              }
+            // A store to a PROVABLY-DISJOINT slot of the same base never aliases the load
+            // (`disjointConstSlots`, ir/alias.ts). Anything less certain bars.
+            if (!isCall && op.opcode === 'load' && disjointConstSlots(op, x)) {
+              return false;
             }
             return true;
           }