diff options
Diffstat (limited to 'skills/chapter25-codex-notes.md')
| -rw-r--r-- | skills/chapter25-codex-notes.md | 65 |
1 files changed, 62 insertions, 3 deletions
diff --git a/skills/chapter25-codex-notes.md b/skills/chapter25-codex-notes.md index 2c327292..dc6c549e 100644 --- a/skills/chapter25-codex-notes.md +++ b/skills/chapter25-codex-notes.md @@ -42,14 +42,73 @@ the first analysis pass after Opto and before TypeCheck, few new classes, ASCII debugging, conservative rejections, and 3-4 poster children. The agreed decisions and remaining questions are in [the survey's base design](../docs/automatic-memory-management-survey.md#agreed-base-design--2026-11-17). The initial implementation is `codegen/Allocation.java`, after FinalizeIR and -before TypeCheck. One call-DAG DFS composes symbolic summaries in postorder; +LoopTree, before TypeCheck. One call-DAG DFS composes symbolic summaries in postorder; local reference-edge closure uses a worklist. Four poster children and boundary checks live in Chapter33Test. The phase is analysis-only: existing allocation lowering is unchanged. `simple.alloc.dump` prints ASCII, `simple.alloc.check` requires all plans to be supported; otherwise unsupported summaries record a -rejection without blocking the existing allocator. Loops, SCCs, variable sizes, -opaque calls, raw addresses, and unresolved pointer-load origins are deferred. +rejection without blocking the existing allocator. SCCs, variable sizes, +opaque calls, and raw addresses are deferred. Pointer-valued loads now select +field targets; borrowed paths such as arg3.c have their own destination identity. +Local weak-union closure covers loads, stores, and call substitutions before +owner propagation. Actual aliases are resolved at calls; array indices share +one field. Fresh sites are imported once per call. Incoming borrowed paths are +separate from store edges: returning a child must not retain its holder. Weak +updates that expand paths beyond the body's reads plus imported path depths +reject instead of truncating; memory-ordered updates remain future work. +Loop Phis now join the same local monotone reference calculation. Allocation +uses CFG `_ltree` membership and LoopTree parent links, without rediscovering +loop bodies. Reducibility is asserted by LoopTree. The early tree build does +not insert synthetic exits: doing so would change nonreturning summaries before +serialization. Unlink rebuilds the tree after Lower's CFG changes and inserts +one synthetic exit only for loops with zero exits. Any fresh +allocation reachable across a backedge rejects the plan, via pointer Phis or +memory retention (including helper effects). Collection starts from both the +function entry and Return, so zero-exit bodies are analyzed without synthetic +exits. Exit-only winners get fixed storage +before the loop; unrelated scratch uses iteration storage. Inner-loop winners +can live in the outer iteration. Bound exemptions and overwrite/liveness proofs +are not implemented yet. Functions and loop iterations now use the same Summary +class, with entry/normal-exit/continuation ports, retention effects, and scratch +cleanup sets. One function-wide reference closure shares SSA/Site identities +with nested loop summaries; calls substitute those identities. Summary interfaces +drive promotion through parent frames, replacing the separate loop placement +scan. Continuation cleanup and normal-exit cleanup use the same scratch set. +Borrowed continuation references are supported. Tests include scalar exit values, +nested returns, mixed winner/scratch, helper allocations, exit-only stores, winner +graphs, and rejected linked/array/helper retention. No stack lowering is emitted. No new language syntax or serialized lifetime summaries are implemented yet. +Update 2026-11-21: Cliff requires the temporary LoopTree ordering above to be +replaced before closing Chapter 33, but explicitly deferred implementation. +Insert NeverExits in the original LoopTree build, then remove Allocation's +zero-exit-body support and its extra forward CFG traversal. Resolve the +nonreturning-summary issue rather than delaying synthetic exits. Lower should +maintain LoopTree with minor local edits, eliminating the full rebuild before +scheduling. Track completion in the roadmap's phase-ordering section. +The 2026-11-21 [allocation survey](../chapter33/docs/allocation-survey.md) +measures `sys` (98/179 supported summaries, 11 with user-storage sites) and ten +examples (43/68 own summaries, three with user-storage sites). First-cause +rejections are mostly native/opaque calls and raw addresses; they mask dynamic +arrays and potentially later failures. Scalar libm/clock calls need targeted +contracts; do not interpret arbitrary i64 arguments as pointer-free. The native +`simple_text_float` allocates its returned Simple byte array with malloc, outside +NewNode analysis, so needs an allocation contract or a caller-buffer interface. +Local variable-sized scratch and returned variable-sized arrays both reject; +fixed-size Holder overwrites reject while explicit object reuse passes. These +are survey findings, not authorization to change contracts or allocation lowering. +The standalone probe and function-level snapshot live beside the report. +Update 2026-11-21, native borrowing: Cliff authorized Plan A: native calls borrow +references through the call and do not hide, publish, or retain them. Allocation +now recognizes finite all-native target sets via externFunc, accepts scalar +results, and follows PtrToInt provenance through SSA arithmetic/Phis. Raw returns, +memory stores, and Simple integer-argument forwarding still reject; typed-pointer +wrappers work. Native Simple-pointer results still require origin contracts. +No ABI or runtime allocator changes. Scope lifetime is the documented upper +bound; brace-level scopes are absent here, so current plans use function/iteration +frames. Tests cover raw Phis/offsets, loop scratch, typed wrappers, and backedge +and memory escape rejection. `_exit` no-return CFG work remains open; Cliff's +Return-merge alternative is recorded in the roadmap. Do not use dead CallEnd +control alone: it can discard the very terminating path that must remain rooted. Update 2026-11-19: Cliff accepted the entire [embedded-object design](../docs/automatic-memory-management-survey.md#agreed-design-embedded-objects-values-and-references), including `*T` object values, `&` interior references, and `*(T[4])` embedded |
