polydat_core/compile/mod.rs
1// Copyright 2024-2026 Jonathan Shook
2// SPDX-License-Identifier: Apache-2.0
3
4//! Kernel compilation: assembled DAG → fast executable kernel.
5//!
6//! Everything in this module is on the path between
7//! [`assembly::PolydatAssembler`] and an executable kernel. The
8//! pipeline:
9//!
10//! ```text
11//! PolydatAssembler ──resolve──▶ ResolvedDag ──compile_with(Engine)──┐
12//! (fusion, adapters, │
13//! round-trip lint, ┌────────────────────────────────────┤
14//! topo sort) ▼ ▼ ▼
15//! Interpreter(JitMode) Closures(Provenance) Native(Provenance)
16//! cone::extract_jit_cones closures:: hybrid::
17//! → PolydatKernel CompiledKernel* HybridKernel*
18//! ```
19//!
20//! The host names the engine ([`select::Engine`]); under
21//! `Provenance::Auto` the selector picks the provenance mode from the
22//! graph's shape. Pure native code (`jit::JitKernel*`) is the
23//! differential tier behind the hybrid kernel.
24//!
25//! - [`assembly`]: the public construction surface
26//! ([`assembly::PolydatAssembler`] + [`assembly::WireRef`]) and
27//! the per-engine compile paths.
28//! - [`fusion`]: graph-level subgraph fusion pass; runs during
29//! assembly after wiring resolution.
30//! - [`roundtrip_lint`]: the structural type-round-trip lint run at
31//! resolution.
32//! - [`cone`]: cone-level JIT inside the interpreter kernel
33//! (SRD-105), under a [`cone::JitMode`].
34//! - [`lattice`]: the engine-mix report of a compiled program.
35//! - [`select`]: the engine and provenance enums, and the heuristic
36//! that picks a provenance mode under `Provenance::Auto`.
37//! - [`closures`]: the closure tier, one generated op per node over
38//! a flat u64 slot buffer, by-reference outputs as `Ref2` pairs.
39//! - [`hybrid`]: the native engine (native segments + closure steps
40//! sharing a flat u64 buffer).
41//! - [`marshal`]: the boundary marshalling between slots and `Value`s.
42//! - `externs`: extern inputs and `shared` cells on the compiled
43//! engines.
44//! - [`simd_plan`], `simd_tier1`: scalar-flow SIMD promotion.
45//! - `jit`: Cranelift lowering and the pure native kernels
46//! (feature-gated on `jit`).
47
48pub mod assembly;
49pub mod closures;
50pub mod cone;
51#[cfg(all(test, feature = "jit"))]
52mod cone_tests;
53pub(crate) mod externs;
54pub mod fusion;
55#[cfg(feature = "jit")]
56pub(crate) mod fusion_units;
57pub mod hybrid;
58#[cfg(feature = "jit")]
59pub mod jit;
60pub mod lattice;
61/// The boundary marshalling a compiled node kit reads its arguments
62/// and writes its outputs through (compiled_handles.md §4): a node
63/// crate's kits use it as the core's own do.
64pub mod marshal;
65pub mod roundtrip_lint;
66pub mod select;
67pub mod simd_plan;
68#[cfg(feature = "jit")]
69pub mod simd_tier1;
70
71/// Axiom S2 typed accessors, shared by the P2 and hybrid kernel
72/// types (both expose `self.core.ref_entry(slot)`). Each returns
73/// a borrow whose lifetime ties to `&self`, so the borrow checker
74/// statically prevents holding a slice across the next
75/// `eval(&mut self)` — stale Ref reads are compile errors.
76macro_rules! ref_readers {
77 () => {
78 /// Borrow a `vec_f32` output's current contents.
79 pub fn read_vec_f32(&self, slot: usize) -> &[f32] {
80 match self.core.ref_entry(slot) {
81 crate::ast::ScratchBuf::F32(v) => v,
82 other => panic!("slot {slot} is not f32-lane scratch: {other:?}"),
83 }
84 }
85 /// Borrow a `vec_f64` output's current contents.
86 pub fn read_vec_f64(&self, slot: usize) -> &[f64] {
87 match self.core.ref_entry(slot) {
88 crate::ast::ScratchBuf::F64(v) => v,
89 other => panic!("slot {slot} is not f64-lane scratch: {other:?}"),
90 }
91 }
92 /// Borrow a `vec_f16` output's current contents.
93 pub fn read_vec_f16(&self, slot: usize) -> &[half::f16] {
94 match self.core.ref_entry(slot) {
95 crate::ast::ScratchBuf::F16(v) => v,
96 other => panic!("slot {slot} is not f16-lane scratch: {other:?}"),
97 }
98 }
99 /// Borrow a `vec_i8` output's current contents.
100 pub fn read_vec_i8(&self, slot: usize) -> &[i8] {
101 match self.core.ref_entry(slot) {
102 crate::ast::ScratchBuf::I8(v) => v,
103 other => panic!("slot {slot} is not i8-lane scratch: {other:?}"),
104 }
105 }
106 /// Borrow a `vec_i16` output's current contents.
107 pub fn read_vec_i16(&self, slot: usize) -> &[i16] {
108 match self.core.ref_entry(slot) {
109 crate::ast::ScratchBuf::I16(v) => v,
110 other => panic!("slot {slot} is not i16-lane scratch: {other:?}"),
111 }
112 }
113 /// Borrow a `vec_i32` output's current contents.
114 pub fn read_vec_i32(&self, slot: usize) -> &[i32] {
115 match self.core.ref_entry(slot) {
116 crate::ast::ScratchBuf::I32(v) => v,
117 other => panic!("slot {slot} is not i32-lane scratch: {other:?}"),
118 }
119 }
120 /// Borrow a `vec_i64` output's current contents.
121 pub fn read_vec_i64(&self, slot: usize) -> &[i64] {
122 match self.core.ref_entry(slot) {
123 crate::ast::ScratchBuf::I64(v) => v,
124 other => panic!("slot {slot} is not i64-lane scratch: {other:?}"),
125 }
126 }
127 };
128}
129pub(crate) use ref_readers;
130
131/// The accessors every compiled kernel type carries, whichever tier it
132/// belongs to: reading an output by name or by slot, the externs and
133/// the cursors it declares, and applying the coordinates a `Kernel`
134/// caller left pending before a pull or an eval. Every one of them
135/// delegates to `self.core`, so none decides anything about evaluation
136/// — which is why seven types can share one copy.
137///
138/// `$set_coords` is the coordinate writer the type uses: the provenance
139/// modes that track a changed-input mask apply coordinates through
140/// `set_inputs`, the rest through `set_coords`. It is the only thing
141/// that varies, and the closure tier and the hybrid had a macro each to
142/// vary it.
143macro_rules! kernel_accessors {
144 ($set_coords:ident) => {
145 /// How many inputs are coordinates. (The core's `coord_count`
146 /// field counts the buffer slots all inputs occupy.)
147 pub fn coord_count(&self) -> usize {
148 self.core.externs.coordinate_count()
149 }
150
151 /// The slot of a named output.
152 pub fn resolve_output(&self, name: &str) -> Option<usize> {
153 self.core.output_map.get(name).copied()
154 }
155
156 /// Read an output by pre-resolved slot index. Panics on
157 /// Ref2-colored slots (axiom S2) — use `read_vec_*`.
158 #[inline]
159 pub fn get_slot(&self, slot: usize) -> u64 {
160 self.core.guard_ref_slot(slot);
161 self.core.buffer[slot]
162 }
163
164 /// Read a named output variate after `eval()`. Panics on
165 /// Ref2-colored outputs (axiom S2) — use `read_vec_*`.
166 #[inline]
167 pub fn get(&self, name: &str) -> u64 {
168 let slot = self.core.output_map[name];
169 self.core.guard_ref_slot(slot);
170 self.core.buffer[slot]
171 }
172
173 /// The named output as a typed `Value`, decoded by its port
174 /// type: a `Ref2` output is copied out through its pair
175 /// (compiled_handles.md §4), so the caller never holds a
176 /// pointer; a slot that holds `None` reads as `None`.
177 pub fn get_value(&self, name: &str) -> crate::ast::Value {
178 self.core.value_of(name)
179 }
180
181 /// A named output, its cone run if a write is pending.
182 pub fn pull_output(&mut self, name: &str) -> crate::ast::Value {
183 self.core.pull_named(name)
184 }
185
186 /// The named output through the `Kernel` trait: the pending
187 /// coordinates are applied, a round begins if a write is pending, and
188 /// only the output's cone runs.
189 fn pull_value(&mut self, name: &str) -> crate::ast::Value {
190 self.apply_pending_coords();
191 self.pull_output(name)
192 }
193
194 /// [`Self::pull_value`] by output index.
195 fn pull_value_at(&mut self, index: usize) -> crate::ast::Value {
196 self.apply_pending_coords();
197 self.core.pull_at(index)
198 }
199
200 /// The coordinates of a pending write, applied once: a pull
201 /// after the first in a round finds nothing written and skips
202 /// the comparison. The round itself begins in the core, which
203 /// clears the pending flag.
204 #[inline]
205 fn apply_pending_coords(&mut self) {
206 if self.core.drive.stale {
207 let coords = std::mem::take(&mut self.core.drive.coords);
208 self.$set_coords(&coords);
209 self.core.drive.coords = coords;
210 }
211 }
212
213 /// `eval` through the `Kernel` trait: the pending coordinates,
214 /// then every step.
215 fn eval_pending(&mut self) {
216 let coords = std::mem::take(&mut self.core.drive.coords);
217 self.eval(&coords);
218 self.core.drive.coords = coords;
219 }
220
221 /// The kernel's externs by name and declared type.
222 pub fn externs(&self) -> Vec<(&str, crate::ast::PortType)> {
223 self.core.externs.names()
224 }
225
226 /// The cursors the program declares, with the partitions the
227 /// compiler resolved where its `over` clause and extent were
228 /// constant, as `PolydatProgram::cursor_schemas` reports them.
229 pub fn cursor_schemas(&self) -> &[crate::iteration::source::SourceSchema] {
230 self.core.externs.cursor_schemas()
231 }
232
233 /// Narrow a cursor to one partition, as `narrow_cursor` does on
234 /// the interpreter: its `Ext` slot and six scalar projections
235 /// are set as externs.
236 pub fn set_cursor(
237 &mut self,
238 name: &str,
239 partition: &crate::iteration::cursor_partition::Partition,
240 ) -> Result<(), crate::kernel::WriteError> {
241 for (slot, value) in self.core.externs.cursor_writes(name, partition)? {
242 self.set_input(&slot, value)?;
243 }
244 Ok(())
245 }
246
247 crate::compile::ref_readers!();
248 };
249}
250pub(crate) use kernel_accessors;
251
252/// SRD-74's fusion rule: whether a node may join the run of fused
253/// code being formed, as far as `None` is concerned. The one predicate
254/// both fusers apply — the interpreter's cone planner and the hybrid's
255/// segment batcher — so that what one admits the other admits.
256///
257/// Fused code answers a `None` on a boundary input by making every one
258/// of its outputs `None`, because native code cannot carry one. That is
259/// SRD-74 Rule 1 and it is the right answer for a node that propagates
260/// a `None`. It is the wrong answer for a node that *consumes* one and
261/// keeps going — `to_json` writes `null`, a `printf` with an `Option`
262/// argument writes its own text — so such a node may join only when
263/// every one of its inputs comes from inside, where a `None` cannot
264/// arrive: either no boundary input is `None` and the run proceeds
265/// normally, or one is and the whole run answers `None` without any
266/// member running at all.
267///
268/// `eligible` is indexed by node and filled in topological order, so a
269/// node's producers are decided before it is.
270#[cfg(feature = "jit")]
271pub(crate) fn none_rule_admits(
272 accepts_none: bool,
273 wiring: &[crate::kernel::WireSource],
274 eligible: &[bool],
275) -> bool {
276 !accepts_none
277 || wiring
278 .iter()
279 .all(|src| matches!(src, crate::kernel::WireSource::NodeOutput(j, _) if eligible[*j]))
280}
281
282/// The highest tier a node can reach, given the types of the wires
283/// feeding it: the one answer to a question four places were asking
284/// separately.
285///
286/// The order is the one every builder walks. Native first, since a node
287/// with a lowering joins a segment; then the compiled forms, in the
288/// order `assembly::node_step_op` tries them — a copy step, the scalar
289/// `compiled_u64`, the node's own slot kit; then the interpreter.
290///
291/// The wire types are not optional. `compiled_slot` is offered per call
292/// site with the types the kernel fixed, so a caller without them can
293/// only ask the first two questions, and the three callers that
294/// reported a tier rather than selecting one did exactly that — they
295/// asked `compiled_u64().is_some()` and called a node with a slot kit
296/// `Phase1`, which is what the binary printed for `printf`.
297pub fn node_tier(
298 node: &dyn crate::ast::PolydatNode,
299 wire_types: &[crate::ast::PortType],
300) -> crate::ast::CompileLevel {
301 #[cfg(feature = "jit")]
302 if !matches!(
303 crate::compile::jit::classify_node_typed(node, wire_types),
304 crate::compile::jit::JitOp::Fallback
305 ) {
306 return crate::ast::CompileLevel::Phase3;
307 }
308 let meta = node.meta();
309 let is_copy =
310 (meta.name == "identity" || meta.name.starts_with("__port_")) && meta.outs.len() == 1;
311 let has_kit = node
312 .compiled_slot(
313 wire_types,
314 crate::compile::select::Engine::Closures(crate::compile::select::Provenance::Auto),
315 )
316 .is_some();
317 if is_copy || node.compiled_u64().is_some() || has_kit {
318 crate::ast::CompileLevel::Phase2
319 } else {
320 crate::ast::CompileLevel::Phase1
321 }
322}
323
324/// The provenance of every buffer slot: which input slots reach it,
325/// as an exact multi-word mask, for the pull-side cone guard of every
326/// compiled kernel. `input_dependents` is indexed by input slot (a
327/// multi-slot input repeats its list per slot) and lists the steps
328/// downstream of that slot; `step_output_slots` gives each step's
329/// output slots, which all take the step's mask. A coordinate slot's
330/// provenance is itself.
331pub(crate) fn slot_provenance(
332 coord_count: usize,
333 total_slots: usize,
334 step_output_slots: &[&[usize]],
335 input_dependents: &[Vec<usize>],
336) -> Vec<crate::kernel::ProvMask> {
337 use crate::kernel::ProvMask;
338 let step_count = step_output_slots.len();
339 let mut step_prov: Vec<ProvMask> = (0..step_count).map(|_| ProvMask::empty()).collect();
340 for (input_slot, deps) in input_dependents.iter().enumerate() {
341 for &step in deps {
342 if step < step_count {
343 step_prov[step].set(input_slot);
344 }
345 }
346 }
347 let mut slots: Vec<ProvMask> = (0..total_slots).map(|_| ProvMask::empty()).collect();
348 for (i, slot) in slots.iter_mut().enumerate().take(coord_count) {
349 slot.set(i);
350 }
351 for (step, outs) in step_output_slots.iter().enumerate() {
352 for &slot in outs.iter() {
353 if slot < slots.len() {
354 slots[slot] = step_prov[step].clone();
355 }
356 }
357 }
358 slots
359}
360
361/// The coordinates a host set last on a compiled kernel and whether
362/// they have been evaluated: what the [`Kernel`](crate::kernel::Kernel)
363/// trait's `set_inputs` and `pull` keep between calls.
364#[derive(Clone, Default)]
365pub(crate) struct Drive {
366 pub(crate) coords: Vec<u64>,
367 pub(crate) stale: bool,
368}
369
370/// The slot surface of a compiled kernel: the extended API, over and
371/// above the [`Kernel`](crate::kernel::Kernel) trait every engine
372/// answers.
373///
374/// Every compiled engine lays its program out over one flat `u64` slot
375/// buffer (engines.md §6). That layout is an implementation detail, and
376/// this trait is where it is admitted: a slot index instead of an
377/// output name, a raw `u64` instead of a `Value`, a borrow into the
378/// scratch a by-reference output writes. The interpreter does not
379/// implement it and cannot — its buffers are typed `Value`s and it has
380/// no slot to name — which is the point: the shape of this trait *is*
381/// the thing the compiled tiers share and the interpreter does not.
382///
383/// **This is not the surface for running a program.** Driving a kernel
384/// is `Kernel`, on every engine, and a host that never names an engine
385/// never sees this trait. Reach for it when the implementation detail
386/// is the subject: a differential test asserting on what was laid out,
387/// a benchmark measuring a tier without the `Value` construction and
388/// the name lookup a `pull` pays, a diagnostic reporting on a slot.
389///
390/// It is a subtrait rather than a wider `Kernel`, so it is opt-in at
391/// the import: a caller who does not write `use SlotKernel` does not
392/// have these methods on their kernel at all. And it is reachable
393/// without naming a kernel type, through
394/// [`PolydatAssembler::compile_slots`](crate::compile::assembly::PolydatAssembler::compile_slots),
395/// which hands back a `Box<dyn SlotKernel>` that upcasts to
396/// `Box<dyn Kernel>` wherever the ordinary surface will do.
397pub trait SlotKernel: crate::kernel::Kernel {
398 /// The buffer slot a named output writes, resolved once so a
399 /// caller reading the same output every cycle pays no lookup.
400 fn resolve_output(&self, name: &str) -> Option<usize>;
401
402 /// The raw `u64` in `slot`, as it stands: no evaluation, no
403 /// decoding. Panics on a `Ref2` slot (axiom S2), which has no
404 /// scalar to read — use the `read_vec_*` borrows.
405 fn get_slot(&self, slot: usize) -> u64;
406
407 /// [`Self::get_slot`] by output name.
408 fn get(&self, name: &str) -> u64;
409
410 /// A named output decoded by its port type, a `Ref2` output copied
411 /// out through its pair so the caller never holds a pointer. Reads
412 /// what is there; [`Kernel::pull`](crate::kernel::Kernel::pull)
413 /// evaluates first.
414 fn get_value(&self, name: &str) -> crate::ast::Value;
415
416 /// Set the coordinates, evaluate what `slot` needs, and return its
417 /// raw `u64`. The whole read in one call and one `u64`, which is
418 /// what a tier benchmark wants: `pull_at` gives the same value
419 /// through a `Value` it has to construct.
420 fn eval_for_slot(&mut self, coords: &[u64], slot: usize) -> u64;
421
422 /// Set the coordinates and run every step, the whole program in
423 /// one call. [`Kernel::eval`](crate::kernel::Kernel::eval) is the
424 /// same evaluation over coordinates already written with
425 /// `set_inputs`; this is the form that takes them, which is what a
426 /// loop over a coordinate range wants.
427 fn eval_at(&mut self, coords: &[u64]);
428
429 /// Borrow a `vec_f32` output's current contents. The borrow ties to
430 /// `&self`, so holding one across the next evaluation is a compile
431 /// error rather than a stale read (axiom S2).
432 fn read_vec_f32(&self, slot: usize) -> &[f32];
433 /// Borrow a `vec_f64` output's current contents.
434 fn read_vec_f64(&self, slot: usize) -> &[f64];
435 /// Borrow a `vec_f16` output's current contents.
436 fn read_vec_f16(&self, slot: usize) -> &[half::f16];
437 /// Borrow a `vec_i8` output's current contents.
438 fn read_vec_i8(&self, slot: usize) -> &[i8];
439 /// Borrow a `vec_i16` output's current contents.
440 fn read_vec_i16(&self, slot: usize) -> &[i16];
441 /// Borrow a `vec_i32` output's current contents.
442 fn read_vec_i32(&self, slot: usize) -> &[i32];
443 /// Borrow a `vec_i64` output's current contents.
444 fn read_vec_i64(&self, slot: usize) -> &[i64];
445}
446
447/// [`SlotKernel`] for a compiled kernel, forwarding to the inherent
448/// methods the type already has. The trait is the surface; the
449/// inherent copies are what it forwards to and what this crate calls.
450macro_rules! impl_slot_kernel {
451 ($ty:ident) => {
452 impl crate::compile::SlotKernel for $ty {
453 fn resolve_output(&self, name: &str) -> Option<usize> {
454 $ty::resolve_output(self, name)
455 }
456 fn get_slot(&self, slot: usize) -> u64 {
457 $ty::get_slot(self, slot)
458 }
459 fn get(&self, name: &str) -> u64 {
460 $ty::get(self, name)
461 }
462 fn get_value(&self, name: &str) -> crate::ast::Value {
463 $ty::get_value(self, name)
464 }
465 fn eval_for_slot(&mut self, coords: &[u64], slot: usize) -> u64 {
466 $ty::eval_for_slot(self, coords, slot)
467 }
468 fn eval_at(&mut self, coords: &[u64]) {
469 $ty::eval(self, coords)
470 }
471 fn read_vec_f32(&self, slot: usize) -> &[f32] {
472 $ty::read_vec_f32(self, slot)
473 }
474 fn read_vec_f64(&self, slot: usize) -> &[f64] {
475 $ty::read_vec_f64(self, slot)
476 }
477 fn read_vec_f16(&self, slot: usize) -> &[half::f16] {
478 $ty::read_vec_f16(self, slot)
479 }
480 fn read_vec_i8(&self, slot: usize) -> &[i8] {
481 $ty::read_vec_i8(self, slot)
482 }
483 fn read_vec_i16(&self, slot: usize) -> &[i16] {
484 $ty::read_vec_i16(self, slot)
485 }
486 fn read_vec_i32(&self, slot: usize) -> &[i32] {
487 $ty::read_vec_i32(self, slot)
488 }
489 fn read_vec_i64(&self, slot: usize) -> &[i64] {
490 $ty::read_vec_i64(self, slot)
491 }
492 }
493 };
494}
495pub(crate) use impl_slot_kernel;
496
497/// The [`Kernel`](crate::kernel::Kernel) impl every compiled kernel
498/// shares: the type's inherent `eval_pending`, `pull_value`,
499/// `pull_value_at`, `set_input`, `set_input_at`, `set_cursor`,
500/// `mark_all_dirty`, and a `core` with `drive`, `externs`,
501/// `coord_count`, `output_types`, `output_map`, `buffer`,
502/// `traversals`, and `plan`/`invalidate_all`/`attach_cell`/
503/// `slot_value`.
504macro_rules! impl_kernel_trait {
505 ($ty:ident) => {
506 impl crate::kernel::Kernel for $ty {
507 fn engine(&self) -> crate::compile::select::Engine {
508 self.core.engine
509 }
510 fn set_inputs(&mut self, coords: &[u64]) {
511 self.core.drive.coords.clear();
512 self.core.drive.coords.extend_from_slice(coords);
513 self.core.drive.stale = true;
514 }
515 fn set_input(
516 &mut self,
517 name: &str,
518 value: crate::ast::Value,
519 ) -> Result<(), crate::kernel::WriteError> {
520 self.core.drive.stale = true;
521 $ty::set_input(self, name, value)
522 }
523 fn set_cursor(
524 &mut self,
525 name: &str,
526 partition: &crate::iteration::cursor_partition::Partition,
527 ) -> Result<(), crate::kernel::WriteError> {
528 self.core.drive.stale = true;
529 $ty::set_cursor(self, name, partition)
530 }
531 fn eval(&mut self) {
532 self.eval_pending();
533 self.core.drive.stale = false;
534 }
535 fn pull(&mut self, name: &str) -> crate::ast::Value {
536 self.pull_value(name)
537 }
538 fn input_names(&self) -> Vec<String> {
539 self.core.externs.input_names().to_vec()
540 }
541 /// In declaration order, as the interpreter lists them: the
542 /// assembler sets them on every compiled kernel.
543 fn output_names(&self) -> Vec<String> {
544 self.core.externs.output_names().to_vec()
545 }
546 fn output_type(&self, name: &str) -> Option<crate::ast::PortType> {
547 self.core.output_types.get(name).copied()
548 }
549 fn externs(&self) -> Vec<(String, crate::ast::PortType)> {
550 self.core
551 .externs
552 .names()
553 .into_iter()
554 .map(|(n, t)| (n.to_string(), t))
555 .collect()
556 }
557 fn cursor_schemas(&self) -> &[crate::iteration::source::SourceSchema] {
558 self.core.externs.cursor_schemas()
559 }
560 fn input_value(&self, name: &str) -> Option<crate::ast::Value> {
561 self.core.externs.value(name).or_else(|| {
562 let i = self
563 .core
564 .externs
565 .input_names()
566 .iter()
567 .position(|n| n == name)?;
568 if i < self.core.externs.coordinate_count() {
569 let pending = self.core.drive.coords.get(i).copied();
570 Some(crate::ast::Value::U64(
571 pending.unwrap_or(self.core.buffer[i]),
572 ))
573 } else {
574 None
575 }
576 })
577 }
578 fn traversals(&self) -> &[crate::dsl::traversal::Traversal] {
579 &self.core.traversals
580 }
581 fn plan(&self) -> crate::EnginePlan {
582 self.core.plan()
583 }
584 fn input_index(&self, name: &str) -> Option<usize> {
585 self.core
586 .externs
587 .input_names()
588 .iter()
589 .position(|n| n == name)
590 }
591 fn set_input_at(
592 &mut self,
593 index: usize,
594 value: crate::ast::Value,
595 ) -> Result<(), crate::kernel::WriteError> {
596 self.core.drive.stale = true;
597 $ty::set_input_at(self, index, value)
598 }
599 fn output_index(&self, name: &str) -> Option<usize> {
600 self.core
601 .externs
602 .output_names()
603 .iter()
604 .position(|n| n == name)
605 }
606 fn pull_at(&mut self, index: usize) -> crate::ast::Value {
607 self.pull_value_at(index)
608 }
609 fn traverse(&mut self, index: usize) -> Result<crate::kernel::TraversalStream, String> {
610 let traversal = self.core.traversals.get(index).cloned().ok_or_else(|| {
611 format!(
612 "no traversal at index {index}; the program declares {}",
613 self.core.traversals.len()
614 )
615 })?;
616 crate::kernel::activation::open_traversal(self, traversal)
617 }
618 fn invalidate_all(&mut self) {
619 self.mark_all_dirty();
620 self.core.invalidate_all();
621 }
622 fn shared_cells(&self) -> Vec<crate::kernel::SharedCellEntry> {
623 self.core.externs.shared_cells()
624 }
625 fn output_cell(&self, name: &str) -> Option<crate::kernel::SharedCell> {
626 self.core.output_cell_for(name)
627 }
628 fn output_modifier(&self, name: &str) -> crate::dsl::ast::BindingModifier {
629 self.core.externs.output_modifier(name)
630 }
631 fn cells_in_scope(&self) -> Vec<crate::kernel::SharedCellEntry> {
632 self.core.externs.cells_in_scope()
633 }
634 fn set_transit_cells(&mut self, cells: Vec<crate::kernel::SharedCellEntry>) {
635 self.core.externs.set_transit_cells(cells);
636 }
637 fn input_port_type(&self, name: &str) -> Option<crate::ast::PortType> {
638 self.core.externs.input_port_type(name)
639 }
640 fn bind_input_cell(&mut self, name: &str, cell: crate::kernel::SharedCell) -> bool {
641 let Some(slot) = self.core.externs.bind_cell(name, cell) else {
642 return false;
643 };
644 // The slot's value is the cell's from the next refresh,
645 // so everything downstream of it reruns.
646 self.core.dirty_input(slot);
647 true
648 }
649 fn attach_shared_cell(
650 &mut self,
651 name: &str,
652 cell: crate::kernel::SharedCell,
653 ) -> Result<(), String> {
654 self.core.attach_cell(name, cell)
655 }
656 fn into_program(
657 mut self: Box<Self>,
658 ) -> std::sync::Arc<dyn crate::kernel::KernelProgram> {
659 self.mark_all_dirty();
660 self.core.drive.stale = true;
661 std::sync::Arc::new(crate::kernel::SharedKernel(*self))
662 }
663 fn ledger(&self) -> &std::sync::Arc<crate::kernel::CompileLedger> {
664 self.core.externs.ledger()
665 }
666 fn coord_count(&self) -> usize {
667 self.core.externs.coordinate_count()
668 }
669 fn input_value_at(&self, index: usize) -> Option<crate::ast::Value> {
670 if index < self.core.externs.coordinate_count() {
671 let pending = self.core.drive.coords.get(index).copied();
672 return Some(crate::ast::Value::U64(
673 pending.unwrap_or(self.core.buffer[index]),
674 ));
675 }
676 self.core.externs.value_at(index)
677 }
678 fn input_default_at(&self, index: usize) -> Option<crate::ast::Value> {
679 if index < self.core.externs.coordinate_count() {
680 return Some(crate::ast::Value::U64(0));
681 }
682 self.core.externs.default_at(index)
683 }
684 fn input_is_cell_bound(&self, index: usize) -> bool {
685 self.core.externs.is_cell_bound_at(index)
686 }
687 fn reset_inputs(&mut self) {
688 let count = self.core.externs.input_names().len();
689 for index in self.core.externs.coordinate_count()..count {
690 if self.core.externs.is_cell_bound_at(index) {
691 continue;
692 }
693 let (Some(now), Some(default)) = (
694 self.core.externs.value_at(index),
695 self.core.externs.default_at(index),
696 ) else {
697 continue;
698 };
699 if now != default {
700 // The typed write, so what depends on the input
701 // is marked as any write marks it. A declared
702 // default satisfies its own slot.
703 let _ = crate::kernel::Kernel::set_input_at(self, index, default);
704 }
705 }
706 }
707 fn fork(&self) -> Box<dyn crate::kernel::Kernel> {
708 // A clone is a new state of the same program with this
709 // one's values: shared slots keep their cells, transit
710 // cells travel, and broadcast cells stay with the
711 // original, which is what descendants are bound to.
712 Box::new(self.clone())
713 }
714 fn publish_broadcasts(&mut self) {
715 if !self.core.externs.broadcasts() {
716 return;
717 }
718 let names: Vec<String> = self.core.externs.output_names().to_vec();
719 for name in names {
720 let Some(&slot) = self.core.output_map.get(&name) else {
721 continue;
722 };
723 if self.core.externs.published_output(slot).is_none() {
724 continue;
725 }
726 // The pull by name publishes through the cell. A
727 // failure is left for the pull that needs the value.
728 let _ = std::panic::catch_unwind(std::panic::AssertUnwindSafe(|| {
729 self.pull_value(&name);
730 }));
731 }
732 }
733 fn commit_write_throughs(&mut self) -> Result<(), String> {
734 let pairs = self.core.externs.write_throughs().to_vec();
735 let mut pending = Vec::with_capacity(pairs.len());
736 for (export, source) in &pairs {
737 let Some(slot_type) = self.core.externs.input_port_type(export) else {
738 continue;
739 };
740 let value = self.pull_value(source);
741 let value =
742 crate::kernel::check_write_through_type(export, source, slot_type, value)?;
743 pending.push((export.clone(), value));
744 }
745 for (export, value) in pending {
746 crate::kernel::Kernel::set_input(self, &export, value)
747 .map_err(|e| format!("write-through into `{export}`: {e}"))?;
748 }
749 Ok(())
750 }
751 fn program_id(&self) -> crate::kernel::ProgramId {
752 crate::kernel::ProgramId(self.core.program_identity())
753 }
754 fn input_type_origin(&self, name: &str) -> Option<crate::kernel::TypeOrigin> {
755 self.core.externs.input_type_origin(name)
756 }
757 }
758
759 impl crate::kernel::KernelInternals for $ty {
760 fn set_write_throughs(&mut self, pairs: Vec<(String, String)>) {
761 self.core.externs.set_write_throughs(pairs);
762 }
763 /// A compiled kernel keeps the traversals; each carries the
764 /// comprehension its producer resolved to at compile time.
765 fn set_traversals(
766 &mut self,
767 traversals: Vec<crate::dsl::traversal::Traversal>,
768 _producers: Vec<crate::dsl::traversal::Producer>,
769 ) {
770 self.core.traversals = traversals.into();
771 }
772 fn slot_value(&self, slot: usize, ty: crate::ast::PortType) -> crate::ast::Value {
773 self.core.slot_value(slot, ty)
774 }
775 fn folded_value(&self, name: &str) -> Option<crate::ast::Value> {
776 let slot = *self.core.output_map.get(name)?;
777 let ty = *self.core.output_types.get(name)?;
778 Some(self.core.slot_value(slot, ty))
779 }
780 fn set_cursor_extent(&mut self, index: usize, extent: u64) {
781 self.core.externs.set_cursor_extent(index, extent);
782 }
783 fn reset_to_program(&mut self) {
784 self.core.externs.reset_to_program(&mut self.core.buffer);
785 self.mark_all_dirty();
786 }
787 }
788 };
789}
790pub(crate) use impl_kernel_trait;
791
792/// The bookkeeping every compiled engine keeps, whatever its steps
793/// are: the evaluation round and what ran in it, the clean flags and
794/// what a write dirties, the extern writes and the cell refresh, the
795/// reference pairs a step publishes, and reading an output back.
796///
797/// Both compiled cores carry the same fields for these and, until this
798/// macro, the same eighteen method bodies byte for byte. None of them
799/// touches the step list, which is the one thing the two tiers
800/// genuinely differ about: a step on the closure tier is always a
801/// closure, and on the native tier it is a closure or a run of native
802/// code. That difference lives in the run loops, which stay per tier.
803macro_rules! shared_core_methods {
804 () => {
805 /// Axiom S9: every reference pair in the buffer names the
806 /// scratch entry that owns it. A slot is skipped when nothing
807 /// has been published into it — its step has not run, or it
808 /// carries `None`.
809 ///
810 /// The two tiers wrote this assertion separately and their
811 /// skip predicates had drifted apart: one skipped a `None`
812 /// slot only when a step owned it, the other whenever the
813 /// slot was `None`. Nothing is published either way, so the
814 /// looser test is the right one and is now the only one.
815 ///
816 /// Axiom S2 typed accessor core: resolve a Ref pair's first
817 /// slot to its kernel-owned scratch entry. The returned
818 /// borrow ties to `&self`, so holding it across the next
819 /// `eval(&mut self)` is a compile error — stale reads are
820 /// statically impossible.
821 fn ref_entry(&self, slot: usize) -> &crate::ast::ScratchBuf {
822 match self.ref_scratch.iter().find(|(s, _)| *s == slot) {
823 Some(&(_, idx)) => &self.scratch[idx],
824 None if self.ref_slots.get(slot).copied().unwrap_or(false) => panic!(
825 "slot {slot} is a Ref pair owned by the CALLER (a kernel \
826 input) — read it on the caller side"
827 ),
828 None => panic!("slot {slot} is not a Ref2-colored slot"),
829 }
830 }
831
832 /// Run `body` with the failure path armed: a panic inside a
833 /// step is recorded quietly and re-raised enriched, as the
834 /// interpreter re-raises a node's (A7), and every reference
835 /// pair is checked afterwards in a debug build.
836 ///
837 /// `#[inline]` is load-bearing: this wraps every `eval`, and
838 /// without it the native rung of the ladder pays a call and
839 /// about seven percent.
840 #[inline]
841 fn run_guarded(&mut self, body: impl FnOnce(&mut Self)) {
842 let capture = crate::kernel::engines::EvalPanicCaptureGuard::arm();
843 let outcome = std::panic::catch_unwind(std::panic::AssertUnwindSafe(|| body(self)));
844 drop(capture);
845 if let Err(payload) = outcome {
846 let sites = std::sync::Arc::clone(&self.sites);
847 let node = self.failing_node();
848 sites.reraise(payload, node, &self.buffer, Some(&self.none));
849 }
850 #[cfg(debug_assertions)]
851 self.validate_refs();
852 }
853
854 /// Gated to `debug_assertions` to match its call sites, which
855 /// compile out in release.
856 #[cfg(debug_assertions)]
857 fn validate_refs(&self) {
858 for &(slot, idx) in &self.ref_scratch {
859 let unpublished = self.none[slot]
860 || matches!(self.slot_step.get(slot), Some(Some(step)) if self.ran[*step] == 0);
861 if unpublished {
862 continue;
863 }
864 let (p, l) = self.scratch[idx].ptr_len();
865 assert!(
866 self.buffer[slot] == p && self.buffer[slot + 1] == l,
867 "S9 ref-validator: slot pair ({slot}, {}) = ({:#x}, {}) \
868 does not match scratch[{idx}] = ({p:#x}, {l}) — a slot \
869 op failed to republish or wrote the wrong slots",
870 slot + 1,
871 self.buffer[slot],
872 self.buffer[slot + 1],
873 );
874 }
875 }
876
877 fn attach_cell(
878 &mut self,
879 name: &str,
880 cell: crate::kernel::SharedCell,
881 ) -> Result<(), String> {
882 let slot = self.externs.attach_cell(name, cell)?;
883 self.dirty_input(slot);
884 self.drive.stale = true;
885 Ok(())
886 }
887
888 #[inline]
889 fn begin_epoch(&mut self) {
890 if self.externs.cells_dirty() {
891 self.externs.refresh_cells(&mut self.buffer);
892 }
893 self.dirty_refreshed();
894 self.epoch += 1;
895 self.all_ran = false;
896 for &i in self.volatile_steps.iter() {
897 self.clean[i] = false;
898 }
899 self.drive.stale = false;
900 }
901
902 #[inline]
903 fn dirty_input(&mut self, slot: usize) {
904 if let Some(deps) = self.dirty.get(slot) {
905 for &i in deps {
906 self.clean[i] = false;
907 }
908 }
909 }
910
911 /// Inlined into every evaluation, so only the check lives here:
912 /// a cell refresh changing a slot is rare, and its work is kept
913 /// out of line where it does not grow the hot path.
914 #[inline]
915 fn dirty_refreshed(&mut self) {
916 if self.externs.has_changed() {
917 self.dirty_refreshed_slots();
918 }
919 }
920
921 #[cold]
922 #[inline(never)]
923 fn dirty_refreshed_slots(&mut self) {
924 let changed = self.externs.take_changed();
925 for &slot in &changed {
926 // The cell's value is the slot's now: a slot that was
927 // unset (an extern with no default, bound to a parent's
928 // cell) holds a value, and one the cell cleared holds
929 // none, as a direct write would leave it.
930 if let Some(mask) = self.none.get_mut(slot) {
931 *mask = self.externs.slot_is_unset(slot);
932 }
933 if let Some(deps) = self.plan.input_dependents.get(slot) {
934 for &i in deps {
935 self.ran[i] = 0;
936 self.clean[i] = false;
937 }
938 self.all_ran = false;
939 }
940 }
941 self.externs.return_changed(changed);
942 let was = self.any_none;
943 self.any_none = self.externs.any_unset();
944 if was && !self.any_none {
945 self.none.fill(false);
946 }
947 }
948
949 #[inline]
950 fn eval_all(&mut self) {
951 let fresh = self.drive.stale;
952 if fresh {
953 self.begin_epoch();
954 } else {
955 self.refresh_cells();
956 }
957 if fresh && !self.use_clean && !self.any_none {
958 self.run_guarded(|core| core.run_fresh());
959 } else {
960 let all = std::sync::Arc::clone(&self.all);
961 self.run_steps(&all);
962 }
963 }
964
965 fn extern_written(&mut self, slot: usize, unset: bool) {
966 self.none[slot] = unset;
967 let was = self.any_none;
968 self.any_none = self.externs.any_unset();
969 if was && !self.any_none {
970 self.none.fill(false);
971 }
972 self.dirty_input(slot);
973 self.drive.stale = true;
974 }
975
976 #[inline]
977 fn guard_ref_slot(&self, slot: usize) {
978 if self.ref_slots.get(slot).copied().unwrap_or(false) {
979 panic!(
980 "S2 pointer containment: slot {slot} is Ref2-colored; raw u64 readers \
981 would leak an interior address. Use the typed borrow-checked accessor \
982 (read_vec_*), the boundary decode, or copy out."
983 );
984 }
985 }
986
987 fn invalidate_all(&mut self) {
988 self.clean.fill(false);
989 self.all_ran = false;
990 self.drive.stale = true;
991 }
992
993 /// A pull by index publishes as a pull by name does, so a child
994 /// bound to this output reads what the parent computed. The flag
995 /// is checked before the pull rather than after it: holding the
996 /// value across the check cost the native rungs 6 percent.
997 #[inline]
998 fn pull_at(&mut self, index: usize) -> crate::ast::Value {
999 if self.externs.broadcasts() {
1000 return self.pull_at_publishing(index);
1001 }
1002 self.pull_at_value(index)
1003 }
1004
1005 /// `pull_at` under a descendant: the pull, then the publish.
1006 #[cold]
1007 #[inline(never)]
1008 fn pull_at_publishing(&mut self, index: usize) -> crate::ast::Value {
1009 let value = self.pull_at_value(index);
1010 let slot = self.resolved_outputs[index]
1011 .as_ref()
1012 .expect("resolved by the pull")
1013 .0;
1014 self.publish_slot(slot, &value);
1015 value
1016 }
1017
1018 #[inline(always)]
1019 fn pull_at_value(&mut self, index: usize) -> crate::ast::Value {
1020 if self.resolved_outputs.len() <= index {
1021 self.resolved_outputs.resize(index + 1, None);
1022 }
1023 if self.resolved_outputs[index].is_none() {
1024 let name = self
1025 .externs
1026 .output_names()
1027 .get(index)
1028 .cloned()
1029 .unwrap_or_else(|| {
1030 panic!(
1031 "no output at index {index}; this kernel declares {}",
1032 self.externs.output_names().len()
1033 )
1034 });
1035 let slot = self.output_map[&name];
1036 let ty = self
1037 .output_types
1038 .get(&name)
1039 .copied()
1040 .unwrap_or(crate::ast::PortType::U64);
1041 let cone = self
1042 .plan
1043 .cones
1044 .get(&name)
1045 .map(|c| std::sync::Arc::<[usize]>::from(c.as_slice()));
1046 let can_fail = cone
1047 .as_ref()
1048 .is_some_and(|c| c.iter().any(|&i| self.step_can_fail(i)));
1049 self.resolved_outputs[index] = Some((slot, ty, cone, can_fail));
1050 }
1051 if self.drive.stale {
1052 self.begin_epoch();
1053 } else {
1054 self.refresh_cells();
1055 }
1056 let resolved = self.resolved_outputs[index]
1057 .as_ref()
1058 .expect("resolved above");
1059 let (slot, ty, can_fail) = (resolved.0, resolved.1, resolved.3);
1060 if let Some(order) = &resolved.2 {
1061 // Borrowed across the run rather than cloned: the order
1062 // lives behind an `Arc` this state holds, and running
1063 // steps never touches the resolved outputs.
1064 let order: *const [usize] = &**order;
1065 let order = unsafe { &*order };
1066 if can_fail {
1067 self.run_steps(order);
1068 } else {
1069 // No step of the cone can fail, so there is no
1070 // failure to capture and attribute.
1071 self.run_order(order);
1072 }
1073 }
1074 self.slot_value(slot, ty)
1075 }
1076
1077 /// Publish the value at `slot` through its broadcast cell, if a
1078 /// descendant asked for one.
1079 fn publish_slot(&self, slot: usize, value: &crate::ast::Value) {
1080 if let Some(cell) = self.externs.published_output(slot) {
1081 cell.publish(value.clone());
1082 }
1083 }
1084
1085 fn pull_named(&mut self, name: &str) -> crate::ast::Value {
1086 if self.drive.stale {
1087 self.begin_epoch();
1088 } else {
1089 self.refresh_cells();
1090 }
1091 let plan = std::sync::Arc::clone(&self.plan);
1092 if let Some(order) = plan.cones.get(name) {
1093 self.run_steps(order);
1094 }
1095 let value = self.value_of(name);
1096 self.broadcast(name, &value);
1097 value
1098 }
1099
1100 /// Publish a freshly computed output through its broadcast cell,
1101 /// if a descendant asked for one, so a child that bound its
1102 /// matching input slot to the same cell reads the new value
1103 /// (cross_fiber_invalidation.md §3.1, "broadcast outputs").
1104 ///
1105 /// A program nobody built a subscope under has no cells at all,
1106 /// and pays the emptiness check.
1107 #[inline]
1108 fn broadcast(&mut self, name: &str, value: &crate::ast::Value) {
1109 if !self.externs.broadcasts() {
1110 return;
1111 }
1112 if let Some(&slot) = self.output_map.get(name) {
1113 self.publish_slot(slot, value);
1114 }
1115 }
1116
1117 /// The broadcast cell for a named output, created on the first
1118 /// ask. The interpreter seeds one per output at construction;
1119 /// a compiled kernel makes them only when a descendant binds to
1120 /// one, so a program with no subscope under it allocates none.
1121 ///
1122 /// Keyed by the output's slot, which is what `output_map`
1123 /// answers and what the buffer is indexed by, so the vector is
1124 /// as long as the buffer rather than as long as the output list.
1125 fn output_cell_for(&self, name: &str) -> Option<crate::kernel::SharedCell> {
1126 let slot = *self.output_map.get(name)?;
1127 // An output no step has computed yet holds its type's zero
1128 // in the buffer; the cell starts at `None`, as the
1129 // interpreter's does, until the first pull publishes.
1130 let uncomputed = !self.all_ran
1131 && matches!(self.slot_step.get(slot), Some(Some(step)) if self.ran[*step] == 0);
1132 let initial = if uncomputed {
1133 crate::ast::Value::None
1134 } else {
1135 self.value_of(name)
1136 };
1137 Some(self.externs.output_cell(slot, initial))
1138 }
1139
1140 #[inline]
1141 fn refresh_cells(&mut self) {
1142 if self.externs.cells_dirty() {
1143 self.externs.refresh_cells(&mut self.buffer);
1144 self.dirty_refreshed();
1145 }
1146 }
1147
1148 fn republish_refs(&mut self) {
1149 for &(slot, idx) in &self.ref_scratch {
1150 let (p, l) = self.scratch[idx].ptr_len();
1151 self.buffer[slot] = p;
1152 self.buffer[slot + 1] = l;
1153 }
1154 self.externs.seed(&mut self.buffer, None);
1155 }
1156
1157 #[inline]
1158 fn run_steps(&mut self, order: &[usize]) {
1159 self.run_guarded(|core| core.run_order(order));
1160 }
1161
1162 /// `run_steps` for the build-time constant fold: the same steps
1163 /// under the same guard, but a failure comes back as the
1164 /// message [`Attribution::reraise`] would have raised. A step
1165 /// that fails here fails before any kernel exists, so it is an
1166 /// error the builder returns rather than a panic out of a
1167 /// constructor.
1168 fn fold_steps(&mut self, order: &[usize]) -> Result<(), crate::KernelError> {
1169 let capture = crate::kernel::engines::EvalPanicCaptureGuard::arm();
1170 let outcome =
1171 std::panic::catch_unwind(std::panic::AssertUnwindSafe(|| self.run_order(order)));
1172 drop(capture);
1173 if let Err(payload) = outcome {
1174 let sites = std::sync::Arc::clone(&self.sites);
1175 let node = self.failing_node();
1176 return Err(crate::KernelError::ConstantFold {
1177 reason: sites.describe(payload, node, &self.buffer, Some(&self.none)),
1178 });
1179 }
1180 Ok(())
1181 }
1182
1183 fn set_extern(
1184 &mut self,
1185 name: &str,
1186 value: crate::ast::Value,
1187 ) -> Result<usize, crate::kernel::WriteError> {
1188 let (slot, unset) = self.externs.set(name, value, &mut self.buffer)?;
1189 self.extern_written(slot, unset);
1190 Ok(slot)
1191 }
1192
1193 fn set_extern_at(
1194 &mut self,
1195 index: usize,
1196 value: crate::ast::Value,
1197 ) -> Result<usize, crate::kernel::WriteError> {
1198 let (slot, unset) = self.externs.set_at(index, value, &mut self.buffer)?;
1199 self.extern_written(slot, unset);
1200 Ok(slot)
1201 }
1202
1203 fn slot_value(&self, slot: usize, ty: crate::ast::PortType) -> crate::ast::Value {
1204 if self.none.get(slot).copied().unwrap_or(false) {
1205 return crate::ast::Value::None;
1206 }
1207 crate::compile::marshal::decode_output(&self.buffer, slot, ty)
1208 }
1209
1210 fn value_of(&self, name: &str) -> crate::ast::Value {
1211 let slot = self.output_map[name];
1212 let ty = self
1213 .output_types
1214 .get(name)
1215 .copied()
1216 .unwrap_or(crate::ast::PortType::U64);
1217 self.slot_value(slot, ty)
1218 }
1219 };
1220}
1221pub(crate) use shared_core_methods;
1222
1223/// The dirty-register plan of a compiled kernel: which steps each input
1224/// slot invalidates when it changes, and which steps each named output
1225/// needs. The evaluation loops consume only this; provenance derives it
1226/// today, and a host that knows its write and read patterns may supply
1227/// a narrower plan later without touching the loops
1228/// (docs/design/engines.md §3.1).
1229pub(crate) struct Invalidation {
1230 /// Per input slot (coordinates and externs alike), the steps that
1231 /// depend on it, transitively.
1232 pub(crate) input_dependents: Vec<Vec<usize>>,
1233 /// Per named output, the steps of its cone in evaluation order.
1234 pub(crate) cones: std::collections::HashMap<String, Vec<usize>>,
1235}
1236
1237impl Invalidation {
1238 /// The plan provenance gives: every step downstream of an input is
1239 /// invalidated by it, and every step upstream of an output is in
1240 /// its cone. `inputs` and `outputs` are each step's slots;
1241 /// `output_slots` names the outputs.
1242 pub(crate) fn from_provenance(
1243 input_dependents: Vec<Vec<usize>>,
1244 step_inputs: &[&[usize]],
1245 step_outputs: &[&[usize]],
1246 output_slots: &std::collections::HashMap<String, usize>,
1247 total_slots: usize,
1248 ) -> Self {
1249 let step_count = step_inputs.len();
1250 let mut slot_step: Vec<Option<usize>> = vec![None; total_slots];
1251 for (i, outs) in step_outputs.iter().enumerate() {
1252 for &s in outs.iter() {
1253 slot_step[s] = Some(i);
1254 }
1255 }
1256 let cones = output_slots
1257 .iter()
1258 .map(|(name, &slot)| {
1259 let mut wanted = vec![false; step_count];
1260 let mut stack: Vec<usize> = slot_step[slot].into_iter().collect();
1261 while let Some(i) = stack.pop() {
1262 if wanted[i] {
1263 continue;
1264 }
1265 wanted[i] = true;
1266 stack.extend(step_inputs[i].iter().filter_map(|&s| slot_step[s]));
1267 }
1268 (
1269 name.clone(),
1270 (0..step_count).filter(|&i| wanted[i]).collect(),
1271 )
1272 })
1273 .collect();
1274 Self {
1275 input_dependents,
1276 cones,
1277 }
1278 }
1279}
1280
1281/// Where each compiled step came from, for the failure path only
1282/// (engines.md §3.4). A step's panic is caught at the step
1283/// boundary and re-raised enriched exactly as the interpreter enriches
1284/// a node's: the node's name, the outputs it feeds, the program's
1285/// diagnostic context, and its input values decoded from the buffer
1286/// where the slot types allow. `sites` is indexed by program node:
1287/// the closure and pure-native kernels have one step per node, and
1288/// the hybrid kernel names the failing member of a segment through
1289/// its tracker slot.
1290#[derive(Default)]
1291pub(crate) struct Attribution {
1292 pub(crate) sites: Vec<NodeSite>,
1293 /// The program's diagnostic context (`PolydatProgram::context`).
1294 pub(crate) context: String,
1295}
1296
1297/// One node's identity for the failure path.
1298pub(crate) struct NodeSite {
1299 pub(crate) name: String,
1300 /// The declared outputs the node feeds, sorted.
1301 pub(crate) outputs: Vec<String>,
1302 /// `(first slot, port type)` of every input port, in port order.
1303 pub(crate) inputs: Vec<(usize, crate::ast::PortType)>,
1304}
1305
1306impl Attribution {
1307 /// The inputs of `step` as diagnostic text, from the buffer, each
1308 /// copied out and printed as the interpreter prints the same value:
1309 /// `None` where the mask says so, and the port type alone where the
1310 /// slot cannot be decoded, so the report itself never fails.
1311 fn inputs_of(&self, step: usize, buffer: &[u64], none: Option<&[bool]>) -> Vec<String> {
1312 let Some(site) = self.sites.get(step) else {
1313 return Vec::new();
1314 };
1315 let _quiet = crate::kernel::engines::EvalPanicCaptureGuard::arm();
1316 site.inputs
1317 .iter()
1318 .map(|&(slot, ty)| {
1319 if none.is_some_and(|m| m.get(slot).copied().unwrap_or(false)) {
1320 return "None".to_string();
1321 }
1322 std::panic::catch_unwind(std::panic::AssertUnwindSafe(|| {
1323 crate::kernel::engines::format_value_for_diag(&marshal::decode_output(
1324 buffer, slot, ty,
1325 ))
1326 }))
1327 .unwrap_or_else(|_| format!("{ty:?}"))
1328 })
1329 .collect()
1330 }
1331
1332 /// Re-raise a step's panic enriched as the interpreter enriches a
1333 /// node's (`kernel::engines::enrich_panic`). `step` beyond the
1334 /// sites (native code that failed before naming a step) reports an
1335 /// unknown node, as the interpreter does for an index it lacks.
1336 pub(crate) fn reraise(
1337 &self,
1338 payload: Box<dyn std::any::Any + Send>,
1339 step: usize,
1340 buffer: &[u64],
1341 none: Option<&[bool]>,
1342 ) -> ! {
1343 crate::kernel::engines::reraise_enriched(self.describe(payload, step, buffer, none))
1344 }
1345
1346 /// The same message [`Self::reraise`] raises, returned instead. The
1347 /// build-time constant fold uses it: a step that fails there fails
1348 /// before any kernel exists, so it is an error the builder returns
1349 /// and not a panic out of a constructor.
1350 pub(crate) fn describe(
1351 &self,
1352 payload: Box<dyn std::any::Any + Send>,
1353 step: usize,
1354 buffer: &[u64],
1355 none: Option<&[bool]>,
1356 ) -> String {
1357 let site = self.sites.get(step);
1358 let name = site
1359 .map(|s| s.name.clone())
1360 .unwrap_or_else(|| format!("<unknown node #{step}>"));
1361 let outputs: Vec<&str> = site
1362 .map(|s| s.outputs.iter().map(String::as_str).collect())
1363 .unwrap_or_default();
1364 let inputs = self.inputs_of(step, buffer, none);
1365 crate::kernel::engines::enrich_panic(payload, &name, &outputs, &self.context, &inputs)
1366 }
1367}