feat(inspection,parser): print placed types as parser sugar by zhen8838 · Pull Request #166 · tile-ai/TileFoundry

zhen8838 · 2026-09-13T02:41:07Z

Why

PR feat(inspection,parser): print placed types as parser sugar #166's review: the printer dispatched types through render_value() instead of visiting each expression's type; render_* methods shadowed visit_*; the goldens showed canonical ShardLayout(...) dataclasses instead of parser sugar; the fixtures were too simple to show a composed mesh, a broadcast and a partial together; and the work had grown four tests where two were asked for.
The sugar the printer already emitted was never readable back: ((dims), (strides)) and {axis @ P("sum")} had no parser production at all, and a bare (8,) reparsed as a plain Layout that dropped the mesh.

What

PythonPrinter inherits PythonTypePrinter. Every type has exactly one visit_<Type>, expression printers emit expr.type through self.visit, and the render_* parallel entries and owner back-delegation are gone.
One parser production covers the whole layout sugar — dims with each split written on the dimension it divides, an optional stride tuple, and an optional {axis @ B(), axis @ P("sum")} set. ExplicitLayoutPattern folds into it.
Every shard layout over a named mesh prints as that sugar, in type slots and in op-attribute slots. The TIR fixtures are authored in it and print back to their own source.
Structurally identical meshes share one prelude name, and a printed type names only a mesh the prelude binds.
Withdrew PR feat(inspection,parser): print placed types as parser sugar #166's footprint-by-level projection, its details opt-in, and the spec text describing them, along with the fixtures and tests that existed only to pin that projection.

Contract

docs/spec/inspection.md §2.3 states when the printer MUST use placement sugar and how it writes splits, strides and value states; §2.5 states one name per mesh descriptor and no named mesh the prelude does not bind.
docs/spec/parser.md's generated grammar: placed-layout now carries the stride tuple and the value-state set; explicit-layout is gone.
An unstated stride tuple stays the sugar default a Reshard materializes from the storage tier it moves between; the sugar states strides whenever the layout has them.
TensorView.layout is annotated LayoutBase, so its slot parses as a layout.
tests/inspection/test_tir_roundtrip.py asserts fixture source equality again, as inspection §2.7 requires.

Risk

tests/fixtures/tir/mma.py uses placement sugar with named warp/lane axes; the CUDA MMA tile scope uses the same mesh names so atom fragment layouts remain structurally equal.
Annotations are longer now that they state strides. That is what makes an annotation name the same layout the emitted code does.

…added Summing LoopFootprintMetadata.footprints by storage level, the `details` opt-in that recovered the per-buffer rows, and the analysis spec text describing both had no reproduced problem behind them: the review's finding was about a printer duplicating metadata, not about how footprints project. The fixtures, goldens and focused tests that only existed to pin that projection go with it. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

PythonPrinter now inherits PythonTypePrinter instead of owning one, so an expression printer emits `expr.type` through the same `visit` that emits the type's children. Every `render_<Type>` entry folds into its `visit_<Type>`, and the `owner` back-delegation is gone. Mesh aliases are now filtered to the ones the prelude actually binds, so naming a mesh inside a printed type can no longer emit an undefined reference.

…added Summing LoopFootprintMetadata.footprints by storage level, the `details` opt-in that recovered the per-buffer rows, and the analysis spec text describing both had no reproduced problem behind them: the review's finding was about a printer duplicating metadata, not about how footprints project. The fixtures, goldens and focused tests that only existed to pin that projection go with it.

The printer could already emit `((dims), (strides))` and `{axis @ P(...)}` that nothing could read back, and it fell back to the verbose `ShardLayout(...)` whenever a function held more than one mesh. One parser production now covers the whole sugar — dims with splits written inline, an optional stride tuple, and an optional value-state set — so the printer can emit sugar for every shard layout over a named mesh. A layout whose mesh axes are all Broadcast now states them, instead of relying on "the active mesh"; the bare form it used to emit reparsed as a plain Layout and silently dropped the mesh. Meshes with the same descriptor share one prelude name, and a printed type names only a mesh the prelude binds, so a composed mesh rebuilt at each use site no longer produces a prelude line per copy or an undefined reference. The TIR fixtures are authored in that sugar and print back to their own source again, restoring the round-trip contract inspection §2.7 states.

The analyze golden reads the same fixture the round-trip golden does, so a type that prints one way in emitted code and another inside an annotation is a diff here rather than two goldens drifting apart.

An unstated stride tuple is not shorthand for C-order: it is the sugar default a `Reshard` materializes from the storage tier it moves between, so printing a stated C-order tuple as absent turned one layout into another and left TIR views with no strides for CUDA codegen to emit. `render_mode()` and `mesh_count()` go with it — the shared type visitor left both without a caller.

…adata `main` now reports compute cost and traffic per topology level, so the golden's metadata comments carry `cta:`/`thread:` breakdowns. The type text the goldens exist to pin is unchanged.

…added Summing LoopFootprintMetadata.footprints by storage level, the `details` opt-in that recovered the per-buffer rows, and the analysis spec text describing both had no reproduced problem behind them: the review's finding was about a printer duplicating metadata, not about how footprints project. The fixtures, goldens and focused tests that only existed to pin that projection go with it. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

The page reproduces itself from its own blocks, so the reshard it displays is now the one-line placement sugar rather than a five-line `ShardLayout`, and the window it prints reaches the matmul that consumes the weight.

`named_and_out_of_scope` adds a layout the author names rather than spells, and a reshard onto a mesh its scope never enters -- reshard is the one operation allowed to cross that boundary, so its target names a mesh that is not the scope it runs in. The golden records what the printer does with them today: the named layout is expanded back into sugar, the out-of-scope target is written as `thread.warp` from inside `with cta`, and `@func(mesh=...)` comes back as a `with` in the body. All three are wrong and none of them had a test.

`composed_mesh_pipeline` reshards to the same constant `named_and_out_of_scope` holds as a parameter. One function enters that constant's mesh with a `with`; the other never does. The golden records both printing identically today. Only the first has a scope binding a sugar identifier could refer to, so only the first should keep the sugar once the printer requires one.

A placement states two things — the extents as written, and where they go — and the layout it implies keeps only the second, divided beyond recovery. So the shape slot could not reuse the layout slot's pattern: it copied the whole syntax tree and recovered the extents itself, and `docs/spec/parser.md` carried that copy as a second production, word for word. One pattern now answers both halves and each slot takes the part it asked for. The rules follow: what sat on the placement were the layout slot's rules, so a placement's own rule reaches the constraints table for the first time and `PlacedShapeRule` goes with the pattern it guarded.

This reverts commit e1e2fde.

This reverts commit 17ee258.

zhen8838 commented Sep 13, 2026

View reviewed changes

zhen8838 changed the title ~~test(inspection): lock canonical type and analysis goldens~~ feat(inspection,parser): print placed types as parser sugar Sep 13, 2026

zhen8838 added 7 commits September 13, 2026 15:33

test(inspection): lock canonical type and analysis goldens

a245f12

test(analysis): lock the annotated analysis surface to the sugar fixture

32639d7

The analyze golden reads the same fixture the round-trip golden does, so a type that prints one way in emitted code and another inside an annotation is a diff here rather than two goldens drifting apart.

test(analysis): regenerate the analyze golden on main's per-level met…

1ff589d

…adata `main` now reports compute cost and traffic per topology level, so the golden's metadata comments carry `cta:`/`thread:` breakdowns. The type text the goldens exist to pin is unchanged.

zhen8838 force-pushed the test/render-golden-final branch from 2c357d0 to 4d56e19 Compare September 13, 2026 07:39

docs(tutorial): re-render the showcase page on the sugar surface

439011a

The page reproduces itself from its own blocks, so the reshard it displays is now the one-line placement sugar rather than a five-line `ShardLayout`, and the window it prints reaches the matmul that consumes the weight.

zhen8838 commented Sep 13, 2026

View reviewed changes

zhen8838 added 5 commits September 13, 2026 22:13

refactor(inspection): finish render golden review fixes

925ee10

fix(ci): align render and mma review contracts

0a506dd

zhen8838 force-pushed the test/render-golden-final branch from 1db7e80 to 0a506dd Compare September 14, 2026 03:44

docs(tutorial): refresh showcase after printer refactor

17ee258

zhen8838 commented Sep 14, 2026

View reviewed changes

Comment thread tests/fixtures/inspection/type_printer_sugar.analyzed.txt Outdated

Comment thread tests/fixtures/inspection/type_printer_sugar.printed.txt Outdated

zhen8838 added 4 commits September 14, 2026 14:12

test(inspection): align mesh sugar fixture with review

e1e2fde

Revert "test(inspection): align mesh sugar fixture with review"

5c5f396

This reverts commit e1e2fde.

Revert "docs(tutorial): refresh showcase after printer refactor"

56465bd

This reverts commit 17ee258.

fix(inspection/parser): enforce lexical mesh placement bindings [M0-M2]

bcf8179

zhen8838 commented Sep 14, 2026

View reviewed changes

Comment thread tests/fixtures/inspection/type_printer_sugar.analyzed.txt Outdated

zhen8838 added 3 commits September 14, 2026 22:56

fix(inspection): omit loop result type comments

b6b7b33

fix(inspection): stabilize sliced mesh type annotations

d9cd3b2

docs(tutorial): refresh showcase analysis output

deae3f6

zhen8838 merged commit 2af2364 into tile-ai:main Sep 14, 2026
1 check passed

zhen8838 deleted the test/render-golden-final branch September 14, 2026 16:46

Navigation Menu

feat(inspection,parser): print placed types as parser sugar - #166

feat(inspection,parser): print placed types as parser sugar#166
zhen8838 merged 21 commits into
tile-ai:mainfrom
zhen8838:test/render-golden-final

zhen8838 commented Sep 13, 2026 •

edited

Loading

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

Conversation

zhen8838 commented Sep 13, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Why

What

Contract

Risk

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

zhen8838 commented Sep 13, 2026 •

edited

Loading