[clang] [Clang] Don't diagnose VLA for `-std=gnu++*` by default (PR #89943)

[Aaron Ballman via cfe-commits]

AaronBallman wrote:

> > This was discussed in this RFC: https://discourse.llvm.org/t/rfc-diagnosing-use-of-vlas-in-c/73109 and our decision to move forward with that is independent of what GCC does.
> 
> I saw that discussion, but I do miss where it has been agreed to diverge from GCC on this. Maybe more generic question will clear my misunderstanding: what is a stance of Clang on`-std=gnu++*`? Can it be different to what GCC does?

Yes, we already differ in many, many ways from GCC in gnu modes, especially when it comes to diagnostics. We often aim for compatibility, but it's not a hard requirement that we follow GCC's lead. Other examples include (but certainly are not limited to): https://godbolt.org/z/sj3oEqj8Y

> > Without more justification for walking this back, I do not agree with moving forward with the revert.
> 
> We have library code that has to use VLA in source code in certain scenario and rely on compiler to optimize it away. The library is built with `-Wall` and after it started to use newer Clang, the error appear. The main concern is simple: GCC doesn't do that (yet). Understandably, if GCC agrees to diagnose it, that change will need to be reverted.

Yeah, this sort of problem crops up when we diverge in diagnostic behavior, but isn't a reason to revert a useful diagnostic. Due to the nature of the syntax for declaring a VLA, many instances of VLA use in C++ (and even in C) are purely accidental and C++ often has better alternatives. Because your use of VLAs is intentional, I'd recommend adding `-Wno-vla` to your build script.

https://github.com/llvm/llvm-project/pull/89943