[Bug tree-optimization/57742] memset(malloc(n),0,n) -> calloc(n,1)

["rguenth at gcc dot gnu.org" <gcc-bugzilla at gcc dot gnu dot org>]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=57742

--- Comment #5 from Richard Biener <rguenth at gcc dot gnu.org> ---
(In reply to Marc Glisse from comment #4)
> (In reply to Richard Biener from comment #2)
> > (In reply to Marc Glisse from comment #1)
> > > This is a very limited version of this optimization. It is in
> > > simplify_builtin_call, so only triggers if malloc/calloc is
> > > SSA_NAME_DEF_STMT(gimple_vuse(memset_stmt)). However, generalizing it means
> > > we would need plenty of tests protecting against cases where the
> > > transformation would be wrong. Note that this transforms:
> > > p=malloc(n);
> > > if(cond)memset(p,0,n);
> > > into:
> > > p=calloc(n,1);
> > > cond;
> > > which is good if cond is p!=0 but may not always be so great otherwise.
> > 
> > ;)  post-dominator tests (or simply tests whether both calls are in the
> > same basic-block ...).
> 
> Same basic block is quite limited, and for the condition below we don't
> directly have post-domination, we would need post-domination between the bbs
> with gimple_cond and malloc, and the bb of memset with the landing block of
> the gimple_cond. But even finding the gimple_cond in: malloc; loop; cond;
> loop; memset; can be hard. I guess I'll have to limit my expectations a
> bit...
> 
> > Also you can transform
> > 
> > p = malloc (n);
> > if (p)
> >   memset (p, 0, n);
> > 
> > which might be a common-enough case to optimize for.
> 
> Yes, that's the goal.
> 
> > dereferencing a double wouldn't have a VDEF (unless you store a double).
> 
> I do want to be able to store in between, so I think I have to walk the vdef
> chain. But as soon as I do that, I need to make sure that the writes are to
> places that can't alias, which complicates things a lot (and it can get a
> bit expensive in a function with many memset). Consider this program:
> 
> #include <vector>
> void f(void*p,int n){ new(p)std::vector<int>(n,0); }
> 
> With -O3, we end up with:
> 
>   _27 = operator new (_26);
>   MEM[(struct _Vector_base *)p_4(D)]._M_impl._M_start = _27;
>   MEM[(struct _Vector_base *)p_4(D)]._M_impl._M_finish = _27;
>   _16 = _27 + _26;
>   MEM[(struct _Vector_base *)p_4(D)]._M_impl._M_end_of_storage = _16;
>   __builtin_memset (_27, 0, _26);
> 
> which has memory stores between the allocation and memset. That's exactly
> the type of code where I'd want the optimization to apply. Joost's example
> has the same pattern: malloc, test for 0, several unrelated memory stores,
> memset.

We have walk_aliased_vdefs for this.  Basically the first callback
you receive has to be the malloc, otherwise there is an aliasing
stmt inbetween.  Initialize the ao_ref with ao_ref_init_from_ptr_and_size.

> (how to handle the fact that we have operator new and not malloc is a
> different issue, I am thinking of having a mode/flag where we promise not to
> replace operator new so it can be inlined, which will include an if(p!=0)
> test)
> 
> It would be great (in particular for application-specific plugins) to have
> an easy way to say things like: this is the next read/write use of this
> memory region (but other memory regions may be used in between), and it
> isn't post-dominated only because of this gimple_cond, etc. It's almost
> noon, too late to be dreaming ;-)

See above ;)