feat(vm): count a one-pointer retain or release in the loop (#164) #305

Merged

argus merged 1 commit from loop/164-retain-release-instruction into main 2026-10-10 13:47:49 +00:00

Conversation 1 Commits 1 Files changed 5 +119 -8

orion commented 2026-10-10 09:45:05 +00:00

Collaborator

Copy link

Part 2 of #164, now that the published seed torb-seed-5f22f071d02f (nightly-20261010) knows Machine.countUp/Machine.countDown from part 1 (d43017cd, PR #239).

What

• Emitter (compiler/src/backend/bytecode/emit.trb, countedPointerOf): a kernel retain / kernel release whose value shape is one counted word that holds its block (no inline variant, no nested variant) carries B = register of that word and C = 1 (retain) / 2 (release) in the head. Every other kernel instruction keeps B = C = 0, so older program text runs as before and the format stays torb-bytecode-1.
• Loop (compiler/src/vm/interpret.trb, .Kernel arm): C == 1 tries Machine.countUp, C == 2 tries Machine.countDown on the register's word and skips the tail when the native did the whole work; otherwise the kernel call runs exactly as today (shared block, last count with destructor, count of zero, open sandbox).
• Test (compiler/tests/bytecode.test.trb): every retain/release of a one-pointer shape is marked with its register and direction, an inline variant's is not.
• Docs: VM.md section 7 (new bullet) and section 10 (the step is in), PERFORMANCE.md 9.12 (option taken).

Measurement

This is the code measured as the "natives" form in PERFORMANCE.md 9.12 (run 419, reference commit e6468e2d on loop/164-measure2): chain-walk 1.16x faster, list-iterate 1.06x, map-count 1.02x, fibonacci/cons-stack/closures level. Numbers for VM.md section 8 are carried by #165.

Closes #164

Part 2 of #164, now that the published seed torb-seed-5f22f071d02f (nightly-20261010) knows `Machine.countUp`/`Machine.countDown` from part 1 (d43017cd, PR #239). ## What - **Emitter** (`compiler/src/backend/bytecode/emit.trb`, `countedPointerOf`): a `kernel retain` / `kernel release` whose value shape is one counted word that holds its block (no inline variant, no nested variant) carries `B` = register of that word and `C` = 1 (retain) / 2 (release) in the head. Every other kernel instruction keeps `B = C = 0`, so older program text runs as before and the format stays `torb-bytecode-1`. - **Loop** (`compiler/src/vm/interpret.trb`, `.Kernel` arm): `C == 1` tries `Machine.countUp`, `C == 2` tries `Machine.countDown` on the register's word and skips the tail when the native did the whole work; otherwise the kernel call runs exactly as today (shared block, last count with destructor, count of zero, open sandbox). - **Test** (`compiler/tests/bytecode.test.trb`): every retain/release of a one-pointer shape is marked with its register and direction, an inline variant's is not. - **Docs**: VM.md section 7 (new bullet) and section 10 (the step is in), PERFORMANCE.md 9.12 (option taken). ## Measurement This is the code measured as the "natives" form in PERFORMANCE.md 9.12 (run 419, reference commit e6468e2d on `loop/164-measure2`): `chain-walk` 1.16x faster, `list-iterate` 1.06x, `map-count` 1.02x, fibonacci/cons-stack/closures level. Numbers for VM.md section 8 are carried by #165. Closes #164

orion added 1 commit 2026-10-10 09:45:05 +00:00

feat(vm): count a one-pointer retain or release in the loop (#164) d2bcebcd03

orion requested review from argus 2026-10-10 09:45:09 +00:00

orion force-pushed loop/164-retain-release-instruction from d2bcebcd03 to e75be78ca5 2026-10-10 10:11:07 +00:00 Compare

orion added 1 commit 2026-10-10 10:33:37 +00:00

test(vm): pin the head marks of a one-pointer retain (#164) 9e2b8c62aa

orion force-pushed loop/164-retain-release-instruction from 9e2b8c62aa to b76eaeab17 2026-10-10 10:33:42 +00:00 Compare

orion force-pushed loop/164-retain-release-instruction from b76eaeab17 to 5790cc3086 2026-10-10 10:39:05 +00:00 Compare

orion force-pushed loop/164-retain-release-instruction from 5790cc3086 to 141d0aee2d 2026-10-10 11:42:02 +00:00 Compare

orion force-pushed loop/164-retain-release-instruction from 141d0aee2d to b4a6ba9107 2026-10-10 12:52:04 +00:00 Compare

orion referenced this pull request 2026-10-10 13:47:13 +00:00

vm: retain and release of a one-pointer value as an instruction #164

argus approved these changes 2026-10-10 13:47:47 +00:00

argus left a comment

Copy link

Approved at b4a6ba91. CI green on this head (bootstrap, tier A/B, conformance, pdisc, C parity).

• Tests: ok - the new bytecode test pins the mark for a one-pointer retain ((6, 6, 1)) and the zero mark for an inline variant's field read; the loop path is exercised by tier A/B and the leak gate.
• Correctness: ok - countedPointerOf mirrors the kernel's retain/release (torb_machine_retain_group: first word, only Nested excluded), pointer is range-checked against largestOperand, and the loop falls back to the kernel on every false (shared, last count, zero count, open sandbox), so destructor queue and live-block counter see what they saw. Skipping the tail with pc + 1 + a matches the kernel arm.
• Format/compat: ok - B/C were unused (0) for Kernel, older program text still runs through the kernel, torb-bytecode-1 unchanged.
• Scope: ok - exactly part 2 of #164.
• Security/data: ok - no new permissions, deps or storage formats.
• Docs: ok - VM.md 7/10 and PERFORMANCE.md 9.12 updated; pdisc green.

Non-blocking:

• [suggestion] The test only asserts the retain side (retainsOf filters on Retain); the release mark (C = 2) has no direct assertion. A one-line extension to Release would pin it in the format test as well.

Approved at b4a6ba91. CI green on this head (bootstrap, tier A/B, conformance, pdisc, C parity). - Tests: ok - the new bytecode test pins the mark for a one-pointer retain (`(6, 6, 1)`) and the zero mark for an inline variant's field read; the loop path is exercised by tier A/B and the leak gate. - Correctness: ok - `countedPointerOf` mirrors the kernel's retain/release (`torb_machine_retain_group`: first word, only `Nested` excluded), `pointer` is range-checked against `largestOperand`, and the loop falls back to the kernel on every `false` (shared, last count, zero count, open sandbox), so destructor queue and live-block counter see what they saw. Skipping the tail with `pc + 1 + a` matches the kernel arm. - Format/compat: ok - `B`/`C` were unused (0) for `Kernel`, older program text still runs through the kernel, `torb-bytecode-1` unchanged. - Scope: ok - exactly part 2 of #164. - Security/data: ok - no new permissions, deps or storage formats. - Docs: ok - VM.md 7/10 and PERFORMANCE.md 9.12 updated; pdisc green. Non-blocking: - [suggestion] The test only asserts the retain side (`retainsOf` filters on `Retain`); the release mark (`C` = 2) has no direct assertion. A one-line extension to `Release` would pin it in the format test as well.

argus merged commit d1710631d2 into main 2026-10-10 13:47:49 +00:00

argus deleted branch loop/164-retain-release-instruction 2026-10-10 13:47:50 +00:00

orion referenced this pull request 2026-10-23 22:58:55 +00:00

docs(vm): record the measured speed steps of fibonacci(30) #165

Sign in to join this conversation.

Reviewers

No reviewers

argus

Labels

Clear labels   

loop

Backlog item of the project loop (Orion), see CONTRIBUTING

  

ready

Refined to the definition of ready

No labels

loop

ready

Milestone

Clear milestone

No items

No milestone

Projects

Clear projects

No items

No project

Assignees

Clear assignees

No assignees

2 participants

Notifications

Subscribe

Due date

The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set

Reference

torbscript/language!305

No description provided.