# Chapter backport review

This is a queue of pending work. Remove completed corrections from the queue. Independent
corrections should start in the earliest affected chapter and propagate through
every later affected snapshot. A regression in Chapter 25's `Chapter21Test`
does not test the compiler in the `chapter21` directory.

For now, keep building SSA with incomplete types in Chapter 25. Moving that
architecture earlier, or splitting Chapter 25, is deferred while small changes
establish the review workflow. No renumbering is committed.

## Small changes

Keep each correction's reduced failure, smallest patch, and test results together
for review.

An **audit** item needs an old-chapter reproducer before it is scheduled. Copy
the correction into the chapter's representation, not the entire modern file.

| ID | Change | Proposed destination | Scope and acceptance evidence |
|---|---|---|---|
| B11 | Diagnose return types using the optimized return expression | 18; 19 already has the correction | `ReturnNode.err()` uses `expr()._type` instead of the parse-time `mt` aggregate. Chapter 18 rejects `struct S { u8 x; }; return new S; return 0;` with a mixed integer/reference error; Chapter 19 accepts it. Also test genuinely reachable incompatible returns. This is a candidate, not yet an applied or isolated-patch-verified fix. |

After the first two reviews, batch only corrections with established independence
and regressions. Keep one logical correction per commit across affected chapters.

## AOT class initialization: larger independent work

Proposed on 2026-09-19; deferred behind the smaller BXX cleanup above. This is
a substantial, self-contained Chapter 25 change, not an entangled backport.
Earlier chapter applicability has not been established.

The goal is to pre-allocate AOT class objects in ELF data and pre-fill their
fields, letting ordinary graph optimization remove redundant initialization.
Lazy-loaded classes retain runtime initialization for now.

- Represent static object creation with a dedicated node, or a ConstantNode
  carrying a TypeStruct with the appropriate final field values. Preserve the
  class's identity and canonical layout independently of its current field
  values; equal contents must not merge distinct class objects.
- Use normal StoreNode peepholes to remove same-value-over-same-value stores.
  If needed, add a peephole that folds initializing stores into the static
  creator's type. Create a replacement ConstantNode/type; never mutate a shared
  constant's type. Prove that folding preserves initialization ordering and
  does not expose a final value prematurely through another reference.
- When `<clinit>` reduces to returning the static object, with no remaining
  I/O or public-memory effects, its runtime work can disappear. Initially it
  is sufficient to emit a trivial `<clinit>` and keep its calls. Omitting the
  function or calls is a later optimization; mixed initializers retain their
  remaining effects. Adapt this conceptual return to the existing class-memory
  representation rather than assuming the current `<clinit>` returns a pointer.
- Extend encoding's relocation information to describe addresses stored inside
  data: functions, other class objects, and constant objects such as strings.
  Emit the required data-section relocations, including their symbol, field
  offset, width, and addend. Keep native linking and emulator linking consistent.
- Give the function and data distinct stable symbols, e.g. `A.B.<clinit>` and
  `A.B.$class`. The defining CompUnit owns the class object and emits its one
  definition in its ELF. Other compilation units, including source currently
  being compiled, reference those symbols without emitting another copy of the
  class data. Unifying loaded Simple types/IR alone does not unify native data;
  preserve ownership and identity across partial compiles and serialization.
- Target four-byte function **and data** pointers, with code and data addresses
  restricted to the low 4 GB. Function pointers stay four bytes; data pointers
  must be brought into agreement. Account for layout, loads/stores, relocations,
  native runtime/FFI boundaries, and actual placement; do not silently truncate
  an out-of-range address. This is a proposed representation constraint, not a
  claim that the current runtime already satisfies it.
- Emit immutable, fully preinitialized objects in read-only storage. Objects
  with mutable fields or remaining runtime initialization writes stay writable.
  Retain the current writable-class-storage correction until those writes have
  actually been eliminated.

Acceptance should cover scalar and pointer fields, shared/cyclic object
references, preserved effects in mixed initializers, and serialization
round-trips. A diamond of separately compiled users must share one class-object
address and observe each other's mutations. Check ELF definitions/relocations,
address-range enforcement, native execution, emulator execution, and the full
Chapter 25 suite, including `make -j 4 tests`.

## Larger or entangled changes: defer

| Group | Eventual home | Why not in the small queue yet |
|---|---|---|
| Dominator caches, scheduling, global constant cloning | 11 / 18 / 19 as applicable | Current changes mix inlining, scheduling, and compilation-unit ownership. Need a reduced old-IR failure. |
| Register allocation and spilling | 20 onward | Need constrained-register/spill regressions; changed golden spill counts are insufficient evidence. |
| Conditional Store and array Load control | Memory/arrays chapters | Reproduce under the earlier alias model before extracting fixes from the new memory implementation. |
| SCCP dependencies, function revival, reachability | 24; some foundations may fit 18 | Separate old-IR corrections from new Guard/Escape/BulkMemPhi and external-caller machinery. |
| Lattice fixes in `7f3b8856` | Owning type introduction | The commit mixes TypeFunPtr, TypeMem, and new XInt/escape-summary behavior. Not a generic earlier-lattice patch. |
| TypeScalar, numeric modes, guards, symbolic fields, open/forward types | Revisit earlier homes later | A connected incomplete-types architecture, including phase ordering and errors. |
| BulkMemPhi/MemPhi, private constructor memory, allocation helpers | Revisit memory / constructors / methods | Move invariants and regressions together. |
| Serialization, global identity remapping, module escape summaries | Separate compilation | Remain with modules. |

## Review and test protocol

1. Run the destination's unmodified `make tests`. Record baseline failures and
   resolve them before accepting a compiler backport into that chapter.
2. Add a deterministic regression proving the old failure. Prefer tiny source
   programs; use encoding checks when register placement is the property at issue.
3. Apply the smallest correction; run the focused regression and full chapter
   `make tests`. Stop for Cliff's review before the next trial fix.
4. Propagate the accepted fix and regression through every affected later
   snapshot. Run each snapshot's full `make tests`, not just 25's inherited tests.
5. Run TypeTest for lattice changes and rebuild system objects when compiler
   changes invalidate their IR/native code. Force a full Java rebuild when shared
   APIs or constants change: older Makefiles do not track all Java dependencies.
   Do not weaken checks or drop tests.
6. Report chapters, commands, and results for review; remove completed items from
   this queue. Keep unrelated dirty changes;
   do not push without explicit approval.

The top-level runner supports all 25 chapters:

```sh
make lib                   # dependencies, before tests on a fresh checkout
make tests
make -k tests              # visit all chapters; failures still fail the command
make tests CHAPTERS="chapter21 chapter22"
make tags release CHAPTERS="chapter01 chapter24 chapter25"
```

Runs are sequential, and child failures propagate. `make tag` aliases `tags`.
Java, GNU make, bash, ctags (for tags), and the existing native test tools must
be on PATH. Early release targets package classes into jars; they do not imply
standalone executable compiler drivers. Older Makefiles discover jars when make
starts, so run `make lib` separately before `make tests`.

## Validation record

Historical results below are dated evidence, not a substitute for a fresh
baseline. Logs live in ignored build directories and may no longer exist.
Reusable implementation lessons are in `skills/chapter25-codex-notes.md`.

### B12: immediate multiply destination corrected locally, 2026-09-19

The x86 backend has distinct integer register (`MulX86`, 0F AF), integer
immediate (`MulIX86`, 69/6B), and floating-point (`MulFX86`, MULSD) encoders.
Chapter 21 incorrectly routed immediate multiply through `ImmX86`, which
encodes an opcode extension in ModRM.reg. IMUL needs the destination there,
including its high bit in REX.R. Chapter 20 has no executable encoder yet;
22-25 already have a dedicated immediate multiply encoder.

Chapter 21 now uses that dedicated encoder while preserving its
`twoAddress() == 1` allocation contract. Other `ImmX86` subclasses are unchanged.
For `rcx *= 11`, the regression failed with ModRM C1 instead of C9; the fix emits
`48 6b c9 0b`. For `r9 *= 11`, it emits `4d 6b c9 0b`.

`Chapter21Test.testX86MultiplySameRegister` is carried through 21-25. It forces
rax/rcx/r9/r15 and positive/negative imm8/imm32 values, independently of allocator
choices, and checks register fields, opcode, immediate, length, and each
chapter's allocation contract. Existing distinct-register checks remain in
22-24. Unmodified baselines and final full `make tests` targets passed in all
five snapshots, with assertions enabled; Chapter 25 passed 446 tests.
Logs: `chapter25/build/b12-review/{baseline,red,fixed}.log`.

### Issue #251: dead-node printing corrected locally, 2026-09-19

[Issue #251](https://github.com/SeaOfNodes/Simple/issues/251) identifies dead-node
checks against a null `_inputs` array. Constructors and clones allocate the
array; `kill()` empties it without setting it to null. Line printers now use
`isDead()` throughout 7-25, including both columnar and LLVM formats in 10-17.
Chapter 25's whole-program printer uses the same predicate to omit dead
functions and nodes. Its bounds-safe input accessor simply checks the input
count. Printing does not prune the linker.

`Chapter07Test.testDeadNodePrinting` in every snapshot 7-25 checks retained input
storage, the `DEAD` marker, and live inputless nodes. Chapter 25 also tests
omitting a dead function while preserving its linker entry. The regressions
fail against isolated original printers in 7, 10, 18, and 25. Full `make tests`
passed with assertions enabled in every affected snapshot, including 447 tests
in 25. The initial Chapter 17 test assumed an inputless Start; its setup was
corrected for that chapter's representation and its full target rerun.
Logs: `chapter25/build/issue251-review/{original,fixed,chapter17-fixed}.log`.

### Issue #247: emulator 64-bit stores corrected locally, 2026-09-19

[Issue #247](https://github.com/SeaOfNodes/Simple/issues/247) identifies `st8`
writing the upper half with `st2`. The identical bug exists in EvalRisc5 and
EvalArm64 in every snapshot 21-25, starting with their introduction. Both now
use `st4` for the upper half, writing all eight bytes rather than leaving the
upper two unchanged.

`Chapter21Test.testRisc64BitStore` and `testArm64BitStore` are carried through
21-25. They check individual little-endian bytes independently of the load
helper, round trips, positive/negative/extreme values, overwriting with zero,
and untouched neighboring memory. All ten cases failed before the correction
at the first unwritten byte (0xA5 instead of 0x23).

Full `make tests` passed with assertions enabled in each snapshot 21-25,
including 445 tests in Chapter 25. Log:
`chapter25/build/issue247-review/fixed.log`.

### B10 audit: still pending in Chapter 24, 2026-09-19

The fix is present in Chapter 25 (introduced with commit `51f1f8f4`), but has
not reached Chapter 24. Integer `nonZero` appears in 15; `_widen` first appears
in 24, so earlier snapshots do not need this backport.

An assertions-enabled probe of the current compiled snapshots tested `[0,255]`
and `[-255,0]` at widening levels 0-3. Chapter 24 resets all six nonzero widening
levels to zero; Chapter 25 preserves all of them. Both pass refinement (`isa`),
idempotence, double-dual, singleton normalization, and zero-input checks. This
establishes the missing type-level backport; it is not a loop-failure reproducer.
Probe: `chapter25/build/b10-review/B10Probe.java`. No compiler correction was
made during this audit; B10 remains in the pending queue above.

### B09: complete locally, 2026-09-19

Audited `Node.p(depth)`, recursive printing, labels, scope/graph viewers, and
assembly helpers. Corrections follow the first affected representation:

- 11-24: scheduled IR display uses local `_idepth` state; 11-14 also use
  identity maps to avoid setting Node hashes/GVN locks.
- 18-24: memory display reads raw lazy aliases instead of creating Phis;
  scope display uses `Var._type` without resolving forward references.
  25's stale alias-based memory display reads its actual bulk-memory input.
- 19-23: printers use ordinary `link` and tolerate missing targets. Per Cliff's
  review, type interning is allowed: the proposed non-interning linker scan and
  `sameTarget` helper were removed. 24-25 use `CodeGen._link` to preserve dead
  entries, including through 25's `funcName`; optimizing `link` still cleans up.
- 20-25: register display uses `_lrg` without compressing chains or rewriting
  node mappings. Function predicates use the existing leaf `_isConstant` from
  23; integer size/value accessors also avoid entering shared recursion state.
- 21-25: pool display uses identity bookkeeping and recorded alignment, plus
  recorded struct size from 22 and section choice in 25. Printers do not ask
  Types for layouts, even when an answer might already be cached. Layout is
  separate from type identity and may eventually move out of TypeStruct.

Regressions cover graph caches/edges, lazy Phis, unresolved declarations, stale
linker entries, register chains, leaf type accessors, and forbidden layout queries.
Isolated original Chapter 18 classes fail the lazy-memory and forward-reference
tests; original printers also reproduce cache mutation, register compression,
shared-scratch assertions, and dead-linker pruning. Tests permit type interning.

Every affected compiler snapshot (11-25) passed its full `make tests` target with
assertions enabled after the review adjustment, including 443 tests in 25.
The constant-pool regression rejects size/alignment queries even if they would
return cached answers. Log: `chapter25/build/b09-review/revised.log`.
Durable rules are in the skills notes.

### B13 / issue #246: complete, 2026-09-19

Cliff reports the correction pushed. Nested negation in `p != null` / `!!p`
now discovers the underlying pointer guard in every snapshot 10-25, retaining
each chapter's guard representation. Chapters 23-24 also decompose short-circuit
Phis with availability checks; Eval2 in 18-24 now handles pointer/null Not.

`Chapter10Test.testNullGuards` failed before the backport in 10-24 and passes
afterward; `testNullGuardErrors` checks rejected uses. Both tests live in every
snapshot 10-25. `testShortCircuitGuardScheduling` covers RHS-only call results
in 23-25; Chapter 25 also rotates optimizer seeds. Full `make tests` passed in
each directory 10-25. Ordinary test counts were 149, 164, 166, 181, 201, 212,
230, 282, 309, 350, 359, 374, 386, 403, 425 in 10-24; Chapter 25 passed 428
total. No assertions or existing expectations were weakened.
Logs: `chapter25/build/issue246-review/`, particularly
`chapters18-24-fixed.log` and `chapters23-25-final.log`.

Independent findings retained for follow-up:

- B14's cyclic-equality failure remained reproducible despite green B13
  suites; it is corrected separately below.
- A pre-existing Chapter 25 diagnostic concern remains outside B13: with the
  issue's nullable array setup, `if (p == null) return p.x; return -1;` was
  accepted despite dereferencing null on the taken arm. Original setup:
  `struct Point { int x; new Point = { int v -> x = v; }; };`
  `Point?[] !points = new Point?[2]; points[arg] = new Point(42);`
  `Point? p = points[1];`. This needs separate investigation.

### B14: complete locally, 2026-09-19

Added the existing Chapter 25 leaf type-kind check to `Type.cycle_eq` in 23-24.
Cyclic structural equality first appears in 23. Inspection of 9-22 found the
older interned-child identity representation with kind checks in `Type.equals`;
the equivalent BOTTOM/float-struct meet probe passes in 22 without a correction.
No earlier compiler change is needed for B14.

`TypeTest.testCyclicLeafKinds` in 23-25 interns same-named structs with `x`
typed BOTTOM, 1.0, and 2.0, then meets the latter two. Before the correction it
throws ClassCastException through `TypeFloat.eq` in 23 and 24; 25 passes.
Afterward all three pass, checking F32 field type, canonical interning,
commutativity, absorption by the BOTTOM-field struct, and the dual/join result.

Validation with assertions enabled (Windows/Cygwin, JDK 21.0.4):

- Unmodified baseline: `make -k tests CHAPTERS="chapter22 chapter23 chapter24 chapter25"`
  passed each full target.
- After a full Java rebuild of 23-24, the focused regression and full TypeTest
  passed in 23-25 (6 TypeTest cases in each).
- `make -k tests CHAPTERS="chapter23 chapter24 chapter25"` passed: 404 and 426
  ordinary tests plus their fuzzer wrappers in 23 and 24; 429 total in 25
  (365 raw0, 32 raw1, 8 standalone, 16 system, 1 fuzzer, 7 remaining).

Logs: `chapter25/build/b14-review/{baseline,before-fix,fixed}.log`.

### Build and fuzzer baseline, 2026-09-17

Windows/Cygwin, OpenJDK 22.0.2: all 25 chapters passed `tags` and `release`.
Build setup supplied missing early targets and corrected argument-list writes
in 22-24; no compiler behavior was backported. Chapter 22 initially failed
TypeTest with a TypeConAry/TypeRPC cast error, then passed a forced rebuild
(369 ordinary tests plus fuzzer), consistent with stale classes.

The initial all-chapter test baseline failed Chapter 18's random fuzzer on
seed `973358943756616234`. At Cliff's request, wrappers in 8-24 now follow
25's explicit seed-list structure, and 8-15 invoke their wrappers from
`make tests`. Exploratory fuzzing is opt-in. All 25 full chapter targets
passed after these harness changes; 8-15 passed again after wrapper invocation
was added. Logs: `build/review/seed-lists-all-tests.log` and
`seed-lists-08-15-tests.log`.

**The Chapter 18 compiler failure remains open.** Its seed is in
`OPEN_FAILING_SEEDS`, and explicitly running it still failed. Regression lists
started empty, so green default wrappers added no fuzz regression coverage.
Reduced input:

```java
struct s0 { u8 v1; };
s0? !UmPOLQK=null;
if(0) while(0&UmPOLQK.v1) {}
if(UmPOLQK.v1) {}
return new s0;
while(0) {}
```

This exposes two distinct concerns: nullable field lookup can fail during
parsing with peepholes disabled; with peepholes enabled it reaches a
mixed-return error. B11 isolates the latter further to
`struct S { u8 x; }; return new S; return 0;`, rejected in 18 and accepted in
19. Chapter 19 also changes the fuzzer comparison from peepholes off/on to
two optimizer worklist seeds. Neither that harness change nor B11 alone proves
the original seed fixed. Keep it open until the full seed passes.
Log: `build/review/chapter18-open-seed.log` (intentional failure).