Am Samstag, dem 07.12.2024 um 17:58 -0800 schrieb Linus Torvalds:

On Sat, 7 Dec 2024 at 15:52, Martin Uecker [off-list ref] wrote:

quoted

Can you point me to some horror stories?

So the main issues tended to be about various static verification tools.

Ranging from things like the stackleak plugin for gcc, where handling
VLA's and alloca() (which are pretty much the same thing with
different syntax) was just very much added complexity, to perhaps
overly simplistic tools that literally just check the stack usage by
parsing "objdump --disassemble" output and then building up
approximate "this is the combined deepest stack" call chain
approximations.

And even in the *basic* infrastructure like gcc itself, VLA's simply
made -Wframe-larger-than= just simply not work.

I also have this memory of bad code generation (again, this is 5=
years ago, so take this with a pinch of salt: dim memories), where gcc
wouldn't end up re-using VLA stack slots, so VLA's made the frame
bigger for that reason or something like that.

We explicitly use "-fconserve-stack" to get gcc to reuse spill slots,
because gcc has been known to sometimes makes insanely piggish stack
frames when it just creates a spill slot for *everything*, even if the
spills aren't live at the same time (think big functions with lots of
case statements).

We also had several cases of the VLA's just being silly, when a simple
constant-sized allocation just worked fine and didn't generate
pointless extra code.

Pretty much none of the code core actually ever wanted VLA's, so the
end result was that we had these bad patterns mainly in random drivers
etc. Don't do that.

Thanks. This confirms that the tooling around VLAs is rather poor.


Martin