[wgsl-in] Implement template list discovery and handle all type resolution in the lowerer by teoxoy · Pull Request #8386 · gfx-rs/wgpu

teoxoy · 2025-10-20T14:47:09Z

Connections
Resolves #4501.
Resolves #4406.
Resolves #4405.

Description
Almost all commits prior to implement template list discovery are refactors to get the frontend to look more similar to the spec and to make the needed changes easier to implement (most notably making the Tokenizer no longer impl Clone).

The template list discovery implementation differs from the spec only in structure (not in behavior) since we can do it lazily after tokenization.

The last 3 substantial commits move all type resolution to the lowerer since all identifiers that are part of types need to be resolved the same as all other identifiers.

Testing
Changed existing tests and added new ones.

Squash or Rebase?
Rebase.

jimblandy · 2025-10-21T21:02:50Z

I can take review here - I've been looking into this for a long time.

jimblandy · 2025-10-21T21:11:14Z

@teoxoy It seems like a bunch of these early commits would be fine as independent cleanup PRs, like 42338ff and 4adede5.

jimblandy · 2025-10-21T21:55:19Z

So far I've reviewed the following commits and they all look fine:

527cf31a1 | * rename `function_call_or_assignment_statement` to `func_call_or_variable_updating_statement`
16c2777ba | * rename `function_statement` to `func_call_statement`
086c0e4c3 | * remove `lexer.clone()`
fe63d0753 | * rename `postfix` to `component_or_swizzle_specifier`
47108f674 | * update `lhs_expression` to be in line with the spec
75fbca1ed | * remove `peek_ident_with_span`
ed6a03205 | * remove `next_scalar_generic`
e99eb0149 | * rename `general_expression` to `expression`
54fe3ca2e | * rename `assignment_statement` to `variable_updating_statement`
4adede577 | * [wgsl] rename `type_decl` to `type_specifier`
42338fff3 | * add `optionally_typed_ident` fn

teoxoy · 2025-10-22T12:32:45Z

I removed the last commit that was adding a CTS test since it wasn't passing due to unrelated reasons and pushed 2 more commits to fix the issues CI surfaced.

teoxoy · 2025-10-22T12:47:23Z

@teoxoy It seems like a bunch of these early commits would be fine as independent cleanup PRs, like 42338ff and 4adede5.

They could but that would be lots of separate PRs to file and orchestrate since I think some of them depend on others. I could open a PR just with the refactor commits but then this PR would need to depend on that new one. Let me know.

jimblandy

All the refactoring changes up to 4ebbc67df * implement template list discovery look good. I think there's a bug in the TLD code itself.

andyleiserson · 2025-11-19T23:48:54Z

I pushed a version of this that is rebased on top of #7339 to https://github.com/andyleiserson/wgpu/commits/tld-rebased. The last commit resolves a semantic merge conflict with #7339. There were also merge conflicts with trunk in 314ba9b and cf9e111 that I resolved, and there is a merge conflict with #8545 that I was less sure how to handle (that change adds enable extension checks that need to be relocated somewhere else following the template list discovery changes).

teoxoy · 2025-11-26T18:31:41Z

@andyleiserson thanks for tracking the changes! I pushed a fully rebased version.

jimblandy · 2026-01-06T23:15:25Z

I pushed two very minor tweaks from today's reviewing.

teoxoy · 2026-01-07T17:56:04Z

The first one looks good but the 2nd one assumes that those statements always end with a semicolon. Off the top of my head they are also used in the for loop header.

jimblandy · 2026-01-09T20:48:43Z

I pushed two very minor tweaks from today's reviewing.

The first one looks good but the 2nd one assumes that those statements always end with a semicolon. Off the top of my head they are also used in the for loop header.

Okay, right - I fixed these.

Pushed a few more, and some tweaks, please take a look.

jimblandy · 2026-01-09T20:49:34Z

The template list discovery commit looks good to me.

This improves errors making them more general where appropriate.

In `naga::front::wgsl::parse::Parser::component_or_swizzle_specifier`, since `span_start` sounds like the start of a span, rename that argument to `expr_start`, since it is the span of the start of the expression that ends with the component or swizzle specifier.

jimblandy · 2026-01-27T20:35:10Z

@teoxoy Before we merge this, could you take a quick look at the way cooperative matrix types and functions are handled? I think just looking at the tip should be fine, no need to sort through all the commits.

jimblandy · 2026-01-27T20:52:15Z

I just gave all the cooperative stuff a quick look, and I didn't notice anything that looked like it was doing things the old way.

The only occurrences of the names of the coop_mat types are in conv::map_predeclared_type.
Those map the types to a variant of TypeGenerator, like all our other builtin type generators.
The only occurrences of the names of the coop functions are in Lowerer::call, where they belong.
The only occurrences of ir::TypeInner::CooperativeMatrix in the WGSL front end are:
- where we parse the types
- where we check the types of the arguments to the coop functions
- where we decline to perform automatic conversions on the inner scalars

Avoid adding entries to `ExpressionContext::unresolved` for identifiers that `ExpressionContext::ident_expr` has already processed. At the moment, there is no way these could cause a bug: they only occur after we've recognized a call/construction expression, and those can never refer to function-local definitions, so the string in question is certain to have already been added to `unresolved` by `ExpressionContext::ident_expr`. But they're redundant, and if Naga ever supported nested functions or local type aliases, these could end up adding extraneous dependencies.

The name `scalar` suggests that it returns a scalar value, and `scalar_ty` is analogous to `ty`.

- Clarify that various `foo_statement` functions on `Parser` do not consume the statement's final `;`. - Various other doc fixes throughout.

…sionContext`

Use `if let`, `let else`, and `Option` methods to reduce nesting and clarify success paths.

Rather than passing `call_impl` an `ignore_builtin_identifiers` flag and plumbing that through from its points of use, extend `Namer::namespace` to suspend the `builtin_identifiers` table, as it does for `unique` already. The resulting behavior is the same.

Adjust some methods of `TemplateListIter` to ensure that the typical complete control flow through the function follows the outermost nesting level.

Rename `naga::front::wgsl::error::Error::MissingTemplateArg::arg` to `description`, to make it a little more specific.

jimblandy · 2026-01-27T22:31:45Z

Since you marked this as "rebase", I consolidated my "doc fix" commits a bit, to keep the history cleaner.

jimblandy · 2026-01-27T22:35:11Z

I have got to stop fiddling with this PR

teoxoy · 2026-01-28T11:00:18Z

The cooperative matrix stuff looks good to me; as you mentioned it looks to me like we handle them the same way as the other predeclared types & builtin functions.

teoxoy · 2026-01-28T11:00:28Z

Let's land this 🎉

teoxoy force-pushed the tld branch 2 times, most recently from 9c99aa1 to d4b4df4 Compare October 20, 2025 16:36

ErichDonGubler self-assigned this Oct 20, 2025

jimblandy assigned jimblandy and unassigned ErichDonGubler Oct 21, 2025

jimblandy reviewed Oct 21, 2025

View reviewed changes

Comment thread naga/src/front/wgsl/error.rs Outdated

teoxoy force-pushed the tld branch 3 times, most recently from 5f4abe6 to 13a9c1b Compare October 22, 2025 12:30

jimblandy mentioned this pull request Oct 22, 2025

Add mesh shading info to naga IR #8104

Merged

6 tasks

andyleiserson mentioned this pull request Oct 29, 2025

[naga wgsl-in] Short-circuiting of && and || operators #7339

Merged

6 tasks

jimblandy requested changes Nov 19, 2025

View reviewed changes

Comment thread naga/src/front/wgsl/parse/mod.rs

Comment thread naga/src/front/wgsl/parse/mod.rs Outdated

Comment thread naga/tests/naga/wgsl_errors.rs Outdated

Comment thread naga/src/front/wgsl/parse/lexer.rs

teoxoy force-pushed the tld branch from 13a9c1b to 5a6c847 Compare November 26, 2025 18:31

teoxoy force-pushed the tld branch 2 times, most recently from d10766f to 5b76532 Compare November 27, 2025 15:56

teoxoy requested a review from jimblandy November 27, 2025 16:01

teoxoy commented Jan 12, 2026

View reviewed changes

Comment thread naga/src/front/wgsl/parse/lexer.rs Outdated

andyleiserson mentioned this pull request Jan 14, 2026

fix(naga): Improve validation of non-constructible types #8873

Merged

6 tasks

teoxoy and others added 2 commits January 27, 2026 12:29

[wgsl] pass expected token to functions that might not match any tokens

ebb2975

This improves errors making them more general where appropriate.

jimblandy force-pushed the tld branch from 2a05fab to f894257 Compare January 27, 2026 20:31

jimblandy and others added 13 commits January 27, 2026 14:21

[naga wgsl-in] Use Option::is_some_and where appropriate.

b55e406

[naga wgsl-in] Rename consume_tokens to discover_template_lists.

dbfee4c

[naga wgsl-in] Rename TemplateListIter::scalar to scalar_ty.

cb2b971

The name `scalar` suggests that it returns a scalar value, and `scalar_ty` is analogous to `ty`.

[naga wgsl-in] Doc fixes.

0e8ef6c

- Clarify that various `foo_statement` functions on `Parser` do not consume the statement's final `;`. - Various other doc fixes throughout.

[wgsl-in] move enumerant & var_address_space functions on `Expres…

de2b09b

…sionContext`

[naga wgsl-in] Use typical abbreviations in map_predeclared_type.

c9b99b6

[naga wgsl-in] Minor rearrangement to flatten control flow.

d962f49

Use `if let`, `let else`, and `Option` methods to reduce nesting and clarify success paths.

[naga wgsl-in] Doc fixes.

e62eb19

[naga wgsl-in] Don't permit template parameters after enumerants.

06f783d

[naga wgsl-in] Flatten control flow for clarity.

edf4041

Adjust some methods of `TemplateListIter` to ensure that the typical complete control flow through the function follows the outermost nesting level.

[naga wgsl-in]: Rename error variant field for clarity.

687c035

Rename `naga::front::wgsl::error::Error::MissingTemplateArg::arg` to `description`, to make it a little more specific.

jimblandy force-pushed the tld branch from f894257 to 687c035 Compare January 27, 2026 22:30

This was referenced Jan 27, 2026

Extract WGSL builtin call handling into its own function #8948

Merged

naga: Enforce @must_use for built-in functions #8713

Merged

teoxoy merged commit 66f9526 into gfx-rs:trunk Jan 28, 2026
53 checks passed

teoxoy deleted the tld branch January 28, 2026 11:00

teoxoy mentioned this pull request Feb 2, 2026

[wgsl-in] Avoid looking up newly introduced parameter names while parsing function parameters and its return type. #8980

Merged

teoxoy mentioned this pull request Feb 11, 2026

WGSL: Cannot alias u32, vec3f, and other previously reserved words #5786

Closed

andyleiserson mentioned this pull request Apr 1, 2026

Some for loops with increment, decrement, or phony assignment statements are not supported #9150

Closed

Uh oh!

Conversation

teoxoy commented Oct 20, 2025

Uh oh!

jimblandy commented Oct 21, 2025

Uh oh!

jimblandy commented Oct 21, 2025

Uh oh!

Uh oh!

jimblandy commented Oct 21, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

teoxoy commented Oct 22, 2025

Uh oh!

teoxoy commented Oct 22, 2025

Uh oh!

jimblandy left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

andyleiserson commented Nov 19, 2025

Uh oh!

teoxoy commented Nov 26, 2025

Uh oh!

jimblandy commented Jan 6, 2026

Uh oh!

teoxoy commented Jan 7, 2026

Uh oh!

jimblandy commented Jan 9, 2026

Uh oh!

jimblandy commented Jan 9, 2026

Uh oh!

Uh oh!

jimblandy commented Jan 27, 2026

Uh oh!

jimblandy commented Jan 27, 2026

Uh oh!

jimblandy commented Jan 27, 2026

Uh oh!

jimblandy commented Jan 27, 2026

Uh oh!

teoxoy commented Jan 28, 2026

Uh oh!

teoxoy commented Jan 28, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

jimblandy commented Oct 21, 2025 •

edited

Loading