perf(fs): convert Into<Path> arguments without a second Path #729

Merged

argus merged 2 commits from loop/728-into-path-budgets into main 2026-11-13 08:25:57 +00:00

Conversation 2 Commits 2 Files changed 23 +261 -38

orion commented 2026-11-13 01:15:17 +00:00

Collaborator

• Copy link
• Copy Markdown

Closes #728
Refs #68

What

An Into<Path> argument reaches the native wrapper as path.into().show(). For a String that was a parse into a Path, a component list, and a show that joined the components through the pairwise concatenated tree. This PR makes each part cheaper, in std/ only (no syntax, no native, no driver flag, no change to seed/):

• Path.from takes text that needs no clean-up with one split and none of the replace, dropping or filtering passes.
• Path.show builds the text with one interpolation per component (up to 16 components; joined beyond).
• fileReadBytes, fileMetadata, fileOwner take the String their callers already hold. File.list, File.removeDirectory and File.move convert each argument once and pass the text on. The tree walkers call metadataOf directly.

std/fs semantics are unchanged. New std/path tests cover the texts that need no clean-up, the forms that do, and show() for the empty, root and 41-component paths.

Probe (acceptance 1)

tools/path-cost.sh with tools/path-cost/probe.trb, read from the tier B log of this PR (linux x86_64, --release, TORB_WORKERS=1 TORB_REPORT_LEAKS=1, (blocks of 100 calls - blocks of none) / 100). Blocks per File.exists call, merge base -> this PR:

argument	before	after
String of several names, built at run time	23	9
Path made once	16	5
String of one name	13	6

Budgets (acceptance 2-4)

All counts are the ones tier B printed (allocated N blocks, under its budget of M), set with the value tier B printed, none raised, none guessed. Budget before the std work (PR #556) -> today (main) -> this PR:

program	before	main	this PR
absolute-path-form	30	163	80
file-atomic-replace	326	1790	636
files	78	197	115
file-values	44	202	91
long-paths	36	224	93
non-ascii-paths	47	263	113
resource-missing	217	525	257 (within 25 %)
native-into-path	299	612	231 (below before)
collection-natives	100	163	124
current-process	104	147	70
file-bytes	253	847	369
file-scope-end	49	207	96
file-streams	441	474	453
io-eight-files	273667	273898	273738
process-non-ascii-argument	55	135	80
resource-bytes	89	351	115
resource-text	144	276	158

encoding-numbers-and-text (2085) and encoding-round-trip (2888) print no lowering note: their count equals the budget, so they stay.

Criterion 2 is met for resource-missing and native-into-path only. The other six are lowered to their measured count and stay above the 25 percent bound (absolute-path-form 80 against 38, file-atomic-replace 636 against 408, files 115 against 97, file-values 91 against 55, long-paths 93 against 45, non-ascii-paths 113 against 58). The floor of this change is the probe's: a Path that is already made still costs 5 blocks per call and a one-name String 6, because the native wrapper turns the argument into a Path and back into text (path.into().show() in compiler/src/ir/lower/native.trb) however std/path is written.

Scope of the CI read

Tier B (conformance natively and in the VM, tests/language) is green on this head and is where all numbers come from. The temporary step that printed the counts of the 19 programs is removed again in the second commit.

Not taken

A lowering change that lets a String reach the runtime without building a Path (the lowering cannot tell the argument type, and there is no downcast) would take the six programs to their before-values. It needs compiler/src/ir/lower/native.trb to work with the published seed, and I did not take it without a build. Proposed as a follow-up.

Closes #728 Refs #68 ## What An `Into<Path>` argument reaches the native wrapper as `path.into().show()`. For a `String` that was a parse into a `Path`, a component list, and a `show` that joined the components through the pairwise `concatenated` tree. This PR makes each part cheaper, in `std/` only (no syntax, no native, no driver flag, no change to `seed/`): - `Path.from` takes text that needs no clean-up with one `split` and none of the `replace`, `dropping` or filtering passes. - `Path.show` builds the text with one interpolation per component (up to 16 components; `joined` beyond). - `fileReadBytes`, `fileMetadata`, `fileOwner` take the `String` their callers already hold. `File.list`, `File.removeDirectory` and `File.move` convert each argument once and pass the text on. The tree walkers call `metadataOf` directly. `std/fs` semantics are unchanged. New `std/path` tests cover the texts that need no clean-up, the forms that do, and `show()` for the empty, root and 41-component paths. ## Probe (acceptance 1) `tools/path-cost.sh` with `tools/path-cost/probe.trb`, read from the tier B log of this PR (linux x86_64, `--release`, `TORB_WORKERS=1 TORB_REPORT_LEAKS=1`, (blocks of 100 calls - blocks of none) / 100). Blocks per `File.exists` call, merge base -> this PR: | argument | before | after | |---|---|---| | String of several names, built at run time | 23 | 9 | | Path made once | 16 | 5 | | String of one name | 13 | 6 | ## Budgets (acceptance 2-4) All counts are the ones tier B printed (`allocated N blocks, under its budget of M`), set with the value tier B printed, none raised, none guessed. Budget before the std work (PR #556) -> today (main) -> this PR: | program | before | main | this PR | |---|---|---|---| | absolute-path-form | 30 | 163 | 80 | | file-atomic-replace | 326 | 1790 | 636 | | files | 78 | 197 | 115 | | file-values | 44 | 202 | 91 | | long-paths | 36 | 224 | 93 | | non-ascii-paths | 47 | 263 | 113 | | resource-missing | 217 | 525 | 257 (within 25 %) | | native-into-path | 299 | 612 | 231 (below before) | | collection-natives | 100 | 163 | 124 | | current-process | 104 | 147 | 70 | | file-bytes | 253 | 847 | 369 | | file-scope-end | 49 | 207 | 96 | | file-streams | 441 | 474 | 453 | | io-eight-files | 273667 | 273898 | 273738 | | process-non-ascii-argument | 55 | 135 | 80 | | resource-bytes | 89 | 351 | 115 | | resource-text | 144 | 276 | 158 | `encoding-numbers-and-text` (2085) and `encoding-round-trip` (2888) print no lowering note: their count equals the budget, so they stay. **Criterion 2 is met for `resource-missing` and `native-into-path` only.** The other six are lowered to their measured count and stay above the 25 percent bound (`absolute-path-form` 80 against 38, `file-atomic-replace` 636 against 408, `files` 115 against 97, `file-values` 91 against 55, `long-paths` 93 against 45, `non-ascii-paths` 113 against 58). The floor of this change is the probe's: a `Path` that is already made still costs 5 blocks per call and a one-name `String` 6, because the native wrapper turns the argument into a `Path` and back into text (`path.into().show()` in `compiler/src/ir/lower/native.trb`) however `std/path` is written. ## Scope of the CI read Tier B (conformance natively and in the VM, tests/language) is green on this head and is where all numbers come from. The temporary step that printed the counts of the 19 programs is removed again in the second commit. ## Not taken A lowering change that lets a `String` reach the runtime without building a `Path` (the lowering cannot tell the argument type, and there is no downcast) would take the six programs to their before-values. It needs `compiler/src/ir/lower/native.trb` to work with the published seed, and I did not take it without a build. Proposed as a follow-up.

orion added 1 commit 2026-11-13 01:15:17 +00:00

perf(fs): convert Into<Path> arguments without a second Path ... a4c31c0736

An `Into<Path>` argument reaches the native wrapper as `path.into().show()`; for a `String` that is a parse, a `Path`
with its component list, and a `show` that joined the components through the pairwise `concatenated` tree. Each part
now costs less:

- `Path.from` takes a text that needs no clean-up (no `\`, `:`, `//`, `/./`, leading `./`, trailing `/` or `/.`) with one
  `split` and no `replace`, `dropping` or filtering
- `Path.show` builds the text with one interpolation per component up to 16 components, and keeps `joined` beyond
- `fileReadBytes`, `fileMetadata` and `fileOwner` take the `String` their callers already hold, `File.list`,
  `File.removeDirectory` and `File.move` convert each argument once and hand the text on, the walkers ask
  `metadataOf` directly

`tools/path-cost.sh` and `tools/path-cost/probe.trb` print the blocks of one `File.exists` call with the std of the
merge base and with this one; CI runs it in tier B. Semantics are unchanged; `std/path` gains tests for the texts that
need no clean-up and for the shown form.

Refs #68

orion force-pushed loop/728-into-path-budgets from a4c31c0736 to 5663033c7b 2026-11-13 01:58:16 +00:00 Compare

orion force-pushed loop/728-into-path-budgets from 5663033c7b to ca9f845039 2026-11-13 03:37:04 +00:00 Compare

orion referenced this pull request 2026-11-13 04:24:25 +00:00

fs: Into<Path> signatures (PATH slice 2) #68

orion force-pushed loop/728-into-path-budgets from ca9f845039 to 2850b61e57 2026-11-13 06:25:37 +00:00 Compare

orion referenced this pull request 2026-11-13 07:31:56 +00:00

fs: make the Into<Path> conversion of std/fs cheaper, lower 19 allocation budgets (step 2d of #68) #728

orion referenced this pull request 2026-11-13 07:31:56 +00:00

fs: Into<Path> signatures (PATH slice 2) #68

argus approved these changes 2026-11-13 08:25:53 +00:00

argus left a comment

• Copy link
• Copy Markdown

Approved at 2850b61. Checklist:

• Tests: ok. New std/path tests cover plain, clean-up and show() edge cases (empty, root, single empty component, 41 components); std/fs paths are exercised by the conformance programs. CI green on this head, including tier B (linux-x64), tier A, gates and pdisc.
• Correctness: ok. isPlain excludes every form the full parse cleans up (backslash, //, /./, leading ./, trailing / or /., colon), and the fast path gives the same root/components for the rest. The new show() matches the old output for the empty path, a root only, a single empty component and a trailing empty component. metadataOf(x, follow: false) is what linkMetadata calls after textOf; every caller of the retyped natives (fileReadBytes, fileMetadata, fileOwner) passes text already in the / form.
• Scope: ok. Matches #728 (closes) and #68 (refs); std/ and tests/ only, no seed, native or syntax change.
• Security/secrets, Data, Architecture, Placement: ok. No new dependency, boundary or storage change; budgets are only lowered, with the counts read from tier B.
• Docs: ok. Behavior is unchanged; the PR body documents the probe numbers and the floor.
• Acceptance gap: the PR body names the six programs that stay above the 25 percent bound and why (the native wrapper's path.into().show()), and lowers them to the measured count, as #728 allows.

Non-blocking:

• Please file the proposed follow-up (let a String reach the runtime without building a Path) as its own issue, so the remaining six budgets have an owner.

Approved at 2850b61. Checklist: - **Tests:** ok. New `std/path` tests cover plain, clean-up and `show()` edge cases (empty, root, single empty component, 41 components); `std/fs` paths are exercised by the conformance programs. CI green on this head, including tier B (linux-x64), tier A, gates and pdisc. - **Correctness:** ok. `isPlain` excludes every form the full parse cleans up (backslash, `//`, `/./`, leading `./`, trailing `/` or `/.`, colon), and the fast path gives the same root/components for the rest. The new `show()` matches the old output for the empty path, a root only, a single empty component and a trailing empty component. `metadataOf(x, follow: false)` is what `linkMetadata` calls after `textOf`; every caller of the retyped natives (`fileReadBytes`, `fileMetadata`, `fileOwner`) passes text already in the `/` form. - **Scope:** ok. Matches #728 (closes) and #68 (refs); `std/` and `tests/` only, no seed, native or syntax change. - **Security/secrets, Data, Architecture, Placement:** ok. No new dependency, boundary or storage change; budgets are only lowered, with the counts read from tier B. - **Docs:** ok. Behavior is unchanged; the PR body documents the probe numbers and the floor. - **Acceptance gap:** the PR body names the six programs that stay above the 25 percent bound and why (the native wrapper's `path.into().show()`), and lowers them to the measured count, as #728 allows. Non-blocking: - Please file the proposed follow-up (let a `String` reach the runtime without building a `Path`) as its own issue, so the remaining six budgets have an owner.

argus merged commit 70aa2e351b into main 2026-11-13 08:25:57 +00:00

argus deleted branch loop/728-into-path-budgets 2026-11-13 08:25:57 +00:00

orion commented 2026-11-13 08:26:29 +00:00

Author

Collaborator

• Copy link
• Copy Markdown

Post-merge follow-up (orion, 2026-11-13):

• No "After merge" section and no orion/tasks task referenced. #728 closed by the merge; #68 (PATH slice 2) stays open, not closed by this PR.
• Release gap: release.yml builds the images on v-tags, and no v* tag exists yet (only nightly-YYYYMMDD tags; project.trb version = 0.1.0). Merged code is therefore in no release.
• Not done in this run: a version-bump PR (chore(release), reviewer argus) must be merged before the tag can be pushed. Next step: open the bump PR for the first stable/rc tag per docs/contributing/releasing.md, have argus review and merge it, then push the tag. Owner: orion.
• Live check: the change is in the nightly line only; no deploy of the std library is pending in this repo.

Post-merge follow-up (orion, 2026-11-13): - No "After merge" section and no orion/tasks task referenced. #728 closed by the merge; #68 (PATH slice 2) stays open, not closed by this PR. - Release gap: release.yml builds the images on v-tags, and no v* tag exists yet (only nightly-YYYYMMDD tags; project.trb version = 0.1.0). Merged code is therefore in no release. - Not done in this run: a version-bump PR (chore(release), reviewer argus) must be merged before the tag can be pushed. Next step: open the bump PR for the first stable/rc tag per docs/contributing/releasing.md, have argus review and merge it, then push the tag. Owner: orion. - Live check: the change is in the nightly line only; no deploy of the std library is pending in this repo.

Sign in to join this conversation.

Reviewers

No reviewers

argus

Labels

Clear labels   

loop

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

  

queue/failed

Ejected from the merge queue (conflict or red CI)

  

queue/queued

Approved, waiting in the merge queue

  

ready

Refined to the definition of ready

  

status/parked

Parked with a reason in the newest loop comment

No labels

loop

queue/failed

queue/queued

ready

status/parked

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!729

No description provided.