Re: [improved-aliasing]: Better clobbering and clobber stats

[Daniel Berlin <dberlin at dberlin dot org>]

On Thu, 2005-09-29 at 14:40 -0700, Richard Henderson wrote:
> On Thu, Sep 29, 2005 at 01:02:19PM -0400, Daniel Berlin wrote:
> > This improves our local clobbering by teaching it that things that
> > escape through calls can't escape through calls that don't have writable
> > pointer arguments.
> 
> What do you mean by this?

It really doesn't matter, in part because it buys us so little compared
to just not counting any functions with pointer arguments at all.


>   Don't have "written pointer arguments"
> because you're doing IPA and have looked at the implementation to
> see that it is true?

The functions that matter for this, we can't prove anything about (fputs
is an example).

>   Using the non-user accessible "no vops" 
> attribute? 

>  Looking at "const" on the pointer argument?

The last.

Which is probably not valid in C, but is in some languages where strong
typing means something.
:)

> 
> It looks to me that you're doing the last.  Which is incorrect.
> It's valid C to cast away const.  It's only invalid to reference
> the object through the non-const pointer if the original object
> is const qualified.  Thus
> 
> 	void f(const int *x) { *(int *)x = 0; }
> 	void g() { int x = 1; f(&x); }
> 
> is a valid program, while 
> 
> 	void g() { const int x = 1; f(&x); }
> 
> is not.

I've simply disabled that part of the patch.