-fanalyzer: Questions on C vs CPP + use of GCC attr's like malloc()/access()

[David Malcolm]

On Wed, 2022-11-23 at 20:49 -0500, Gavin Ray via Gcc wrote:
> Hey all, just a few questions about the fantastic GCC Static
> Analyzer:

Hi!

> 
> - It's stated that support for C++ vs C is very limited. Does this
> apply if
>   you're writing C++ that is very similar-looking to C and uses few
> of C++'s
>   advanced features?

Unfortunately not: even fairly simple-looking C++ code can generate
extra CFG edges relating to exception-handling, which -fanalyzer
currently doesn't understand at all, making the output essentially
useless.  And that's just one issue...

Strictly speaking I have added some *very* minimal regression tests in
C++ to our test suite, but on anything beyond the most trivial example
it's likely to run into a known issue.  I'm tracking some of these
known issues here:
  https://gcc.gnu.org/bugzilla/showdependencytree.cgi?id=97110
but the C++ support is currently so minimal that it's probably not yet
worth filing extra bugs against that tracker :-/

I'm hoping to spend a good chunk of the GCC 14 development cycle
working on adding C++ support, with the aim of being able to analyze
GCC itself ("eating my own dog food"), so I hope this situation will
improve greatly then.

> 
> - I noticed that in C++, the "gnu::malloc" attributes don't seem to
> report
>   "warning: leak of 'xxx_alloc()' [CWE-401] [-Wanalyzer-malloc-
> leak]", is
> this
>   normal?

In theory they should; but you could be running into issues with the
analyzer not fully understanding the control flow graph.

> 
> - Is it worthwhile to spend time annotating your method signatures
> with
>   attributes like "malloc" and "access"? Do these aid the -fanalyzer
> passes?

The analyzer makes use of the "malloc", "nonnull" and "const" function
attributes.

It does make use of the "access" attribute, but only within -Wanalyzer-
tainted-size, for the case where the size param of the access is
attacker-controlled (and the taint checker is currently off by default,
even with -fanalyzer).

But like I said, don't expect these to work on C++ code yet.

> 
> For instance:
> 
> [[gnu::malloc]] [[gnu::malloc(HeapPage_free, 1)]]

IIRC, the first [[gnu::malloc]] here is redundant, as it's implied by
[[gnu::malloc(HeapPage_free, 1)]].

> [[gnu::returns_nonnull]]
> struct HeapPage* HeapPage_alloc();
> 
> [[gnu::access(read_write, 1, 3)]]
> struct RecordID
> HeapPage_insert_record(struct HeapPage* page, const char* record,
>                        uint32_t record_length);
> 
> That's quite a significant bit of annotation-noise tacked on to the
> function, so
> I wanted to be sure it's worth the investment!

Maybe in GCC 14 onwards, but it definitely won't be worth it at the
moment.

Hope this is helpful
Dave

> 
> Thank you =)
> Gavin Ray
>