Hello GCC maintainers,
I recently proposed an attribute to suppress -Wvla on a single declaration.
Thank you to everyone who provided feedback. I would like to summarize the
discussion and outline the next step.
Macro-based solutions
Several replies noted that the problem can be addressed today with a macro
wrapping diagnostic pragmas, for example:
#define ALLOW_VLA(decl)
_Pragma("GCC diagnostic push")
_Pragma("GCC diagnostic ignored "-Wvla"")
decl
_Pragma("GCC diagnostic pop")
This approach is workable and is used in some code bases (including ours).
However it has practical drawbacks:
diagnostics originating from the macro expansion are less clear,
tools such as static analyzers and refactoring tools may not interpret
the construct well,
portability requires conditional definitions per compiler,
the annotation is not attached to the declaration in the AST.
In contrast, existing attributes such as [[maybe_unused]] and [[fallthrough]]
attach intent directly to the relevant construct and are consistently visible
to both the compiler and external tooling.
Generalized diagnostic-control attributes
There was also a suggestion to generalize this into attributes controlling
arbitrary diagnostics and/or larger scopes (blocks, functions), for example:
[[gnu::diagnostic_ignored("-Wvla")]] { ... }
While this is an interesting direction, it substantially increases scope
(design of attribute semantics, interaction with existing diagnostic pragmas,
lifetime rules, etc.).
The motivation of the current proposal is narrower: for VLA specifically,
the goal is to mark intentional and reviewed use at the point of declaration,
rather than to provide a general warning-suppression mechanism.
Usefulness and expected adoption
A concern was raised that this may serve only limited use cases.
My expectation is that projects enabling -Wvla (often as part of stricter
safety profiles) occasionally need small, bounded VLAs where heap allocation
or fixed limits are impractical. Today these are either hidden behind macros
or require local pragma ranges. An attribute provides a clearer and more
auditable expression of intent.
Next step
Based on the feedback, I plan to proceed with a minimal implementation
restricted to:
attribute applicable to a single VLA declaration,
effect equivalent to suppressing -Wvla for that declaration only,
no change to pragma behavior or other diagnostics.
I will send a patch to gcc-patches for review.
If there are concerns with this narrowed scope or the proposed direction,
please let me know.
Thank you,
Yair Lenga
