Re: [PATCH v2 0/2] Lock and Pointer guards

[Nick Desaulniers]

On Thu, Jun 8, 2023 at 11:51 AM Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
>
> On Thu, Jun 08, 2023 at 10:20:19AM -0700, Nick Desaulniers wrote:
> > Presumably, one could simply just not use RAII when working with a value that conditionally
> > "escapes" the local scope.
>
> But then you're back to the error goto :/

Thinking more about the expected ergonomics here over lunch...no
meaningful insights, just thoughts...

For something like a mutex/lock; I'd expect those to be acquired then
released within the same function, yeah? In which case
__attribute__((cleanup())) has fewer hazards since the resource
doesn't escape.

For a pointer to a dynamically allocated region that may get returned
to the caller...

I mean, people do this in C++. It is safe and canonical to return a
std::unique_ptr.  When created locally the destructor does the
expected thing regardless of control flow.  IIUC, std::unique_ptr's
move constructor basically does what Kees suggested earlier (trigger
warning: C++): https://github.com/llvm/llvm-project/blob/7a52f79126a59717012d8039ef875f68e3c637fd/libcxx/include/__memory/unique_ptr.h#L429-L430.
example: https://godbolt.org/z/51s49G9f1

A recent commit to clang https://reviews.llvm.org/rG301eb6b68f30
raised an interesting point (deficiency is perhaps too strong a word)
about GNU-style attributes; they generally have no meaning on an ABI.

Consider a function that returns a locally constructed
std::unique_ptr.  If the function returns a type where the caller
knows what destructor functions to run.  This is part of the ABI.

Here, we're talking about using __attribute__((cleanup())) to DTR
locally, but then we return a "raw" pointer to a caller. What cleanup
function should the caller run, implicitly, if at all?  If we use
__attribute__((cleanup())) that saves us a few gotos locally, but the
caller perhaps now needs the same treatment.
-- 
Thanks,
~Nick Desaulniers