Re: CLOBBER_CACHE_ALWAYS regression instability

Hi,

On 2020-04-05 19:54:19 -0400, Alvaro Herrera wrote:
> Isn't it the case that you can create an inner block with a constant
> whose size is determined by a containing block's variable? I mean as in
> the attached, which refuses to compile because of our -Werror=vla -- but
> if I remove it, it compiles fine and works in my system.

IIRC msvc doesn't support VLAs. And there's generally a slow push
towards deprecating them (they've e.g. been moved to optional in C11).

They don't tend to make a lot of sense for sizes that aren't tightly
bound. In contrast to palloc etc, there's no good way to catch
allocation errors. Most of the time you'll just get a SIGBUS or such,
but sometimes you'll just end up overwriting data (if the allocation is
large enough to not touch the guard pages).

Both alloca/vlas also add some per-call overhead.

Allocating the common size on-stack, and the uncommon ones on heap
should be cheaper, and handles the cases of large allocations much
better.

Greetings,

Andres Freund