Originating context: PR #11715 ("Fixes GitHub issue #7894 — makes reflection and SPIR-V agree on layout(local_size) in") and maintainer discussion at #11715 (comment).
PR #11715 makes [numthreads] take precedence over a global GLSL layout(local_size_*) on an entry point. That surfaces three follow-up requirements around how the compiler should handle multiple/conflicting sources of a compute group size. This issue captures them distinctly.
(a) Allow [numthreads] to override the layout-derived group size for multi-entry-point shaders
When a shader has multiple entry points, a [numthreads] attribute on a given entry point should be allowed to override the group size derived from a global layout(local_size_*) qualifier for that entry point. This lets a global GLSL layout(...) provide a default while individual entry points opt into their own size, which is useful when porting a multi-entry-point GLSL file to Slang entry-point-by-entry-point.
(b) Warn when the group-size values are inconsistent across sources
When the value from [numthreads] and the value derived from layout(local_size_*) disagree for an entry point (i.e. the override in (a) actually changes the effective size), the compiler should emit a warning so the inconsistency is visible rather than silently resolved.
(c) Error on conflicting multiple [numthreads] attributes on a single entry point
When one entry point carries multiple [numthreads] attributes with different values, the compiler should emit an error:
[numthreads(2,3,4)]
[numthreads(5,6,7)]
[numthreads(1,5,9)]
void main() {}Current behavior (checked at master @ 7f79b923f)
This case is not diagnosed today — the compiler silently uses the first [numthreads] attribute and ignores the rest, with no error or warning:
findModifier<T>() returns only the first modifier of a type: source/slang/slang-ast-base.h:737.- Lowering iterates the modifier list and consumes the first
NumThreadsAttribute it finds, ignoring later ones: source/slang/slang-lower-to-ir.cpp:~14277.
NumThreadsAttribute is not registered in getModifierConflictGroupKind() (source/slang/slang-check-modifier.cpp:1564-1672), so the general duplicate-modifier detection loop (source/slang/slang-check-modifier.cpp:~2470) never flags it.
So the error requested in (c) is genuinely missing and should be added. A natural place is to give NumThreadsAttribute its own conflict group in getModifierConflictGroupKind() (or a dedicated check) plus a diagnostic near DuplicateModifier.
🤖 Generated by an automated Slang coworker — may be inaccurate. A human maintainer should verify.
Originating context: PR #11715 ("Fixes GitHub issue #7894 — makes reflection and SPIR-V agree on
layout(local_size) in") and maintainer discussion at #11715 (comment).PR #11715 makes
[numthreads]take precedence over a global GLSLlayout(local_size_*)on an entry point. That surfaces three follow-up requirements around how the compiler should handle multiple/conflicting sources of a compute group size. This issue captures them distinctly.(a) Allow
[numthreads]to override the layout-derived group size for multi-entry-point shadersWhen a shader has multiple entry points, a
[numthreads]attribute on a given entry point should be allowed to override the group size derived from a globallayout(local_size_*)qualifier for that entry point. This lets a global GLSLlayout(...)provide a default while individual entry points opt into their own size, which is useful when porting a multi-entry-point GLSL file to Slang entry-point-by-entry-point.(b) Warn when the group-size values are inconsistent across sources
When the value from
[numthreads]and the value derived fromlayout(local_size_*)disagree for an entry point (i.e. the override in (a) actually changes the effective size), the compiler should emit a warning so the inconsistency is visible rather than silently resolved.(c) Error on conflicting multiple
[numthreads]attributes on a single entry pointWhen one entry point carries multiple
[numthreads]attributes with different values, the compiler should emit an error:Current behavior (checked at
master@7f79b923f)This case is not diagnosed today — the compiler silently uses the first
[numthreads]attribute and ignores the rest, with no error or warning:findModifier<T>()returns only the first modifier of a type:source/slang/slang-ast-base.h:737.NumThreadsAttributeit finds, ignoring later ones:source/slang/slang-lower-to-ir.cpp:~14277.NumThreadsAttributeis not registered ingetModifierConflictGroupKind()(source/slang/slang-check-modifier.cpp:1564-1672), so the general duplicate-modifier detection loop (source/slang/slang-check-modifier.cpp:~2470) never flags it.So the error requested in (c) is genuinely missing and should be added. A natural place is to give
NumThreadsAttributeits own conflict group ingetModifierConflictGroupKind()(or a dedicated check) plus a diagnostic nearDuplicateModifier.🤖 Generated by an automated Slang coworker — may be inaccurate. A human maintainer should verify.