Re: [RFC] propagate malloc attribute in ipa-pure-const pass

[Richard Biener <rguenther@suse.de>]

On Wed, 17 May 2017, Martin Sebor wrote:

> > The patch passes bootstrap+test on x86_64 and found a few functions in
> > the source tree (attached func_names.txt) that could be annotated with
> > malloc (I gave a brief look at some of the functions and didn't appear
> > to be false positives but I will recheck thoroughly)
> 
> virtual char* libcp1::compiler::find(std::__cxx11::string&) const
> 
> The virtual on the list of your candidates gave me pause.  Consider
> this completely contrived example:
> 
>   struct B {
>     virtual void* f (unsigned n) {
>       return new char [n];
>     }
>   };
> 
>   void* foo (B &b, unsigned n)
>   {
>     return b.f (n);
>   }
> 
> Based on these definitions alone both functions are candidates
> for attribute malloc.
> 
> But suppose foo is called with an object of a type derived from
> B that overrides f() to do something wacky (but strictly not
> invalid) like:
> 
>   struct D: B {
>     char buf[32];
>     virtual void* f (unsigned n) {
>       if (n < 32)
>       return n <= 32 ? buf : B::f (n);
>     }
> 
> Breaking foo's attribute malloc constraint.
> 
> In other words, I think virtual functions need to be excluded
> from the list (unless they're defined in a class marked final,
> or unless we know they're not overridden to break the constraint
> like above).

But we are annotating the actual decl, not the type in the class
struct.

Richard.