Re: [lno-branch]: Alignment analysis, take two

[Daniel Berlin <dberlin at dberlin dot org>]

On Jul 25, 2004, at 3:49 PM, Devang Patel wrote:

> On Jul 25, 2004, at 10:22 AM, Daniel Berlin wrote:
>
> > he apple guys will probably want to add a way to say that malloc 
> > result is aligned on darwin (RTH had suggested that we might be able 
> > to reuse attribute aligned on function declarations, so that you 
> > could have something be attribute malloc, and attribute aligned.  We 
> > could also introduce a new attribute if nobody likes that idea).  Our 
> > definition of "attribute malloc" doesn't give us enough of a 
> > guarantee that we can make any assumptions about the alignment of 
> > it's return value.
>
> Does this mean decorating malloc decl in system header with attribute 
> aligned OR darwin target automatically attaching attribute aligned to 
> malloc?
Either, i guess.  He just suggested what we might abuse attribute 
aligned on function types, since it currently means nothing.


> > +/* Main structure for CCP.  Contains the lattice value and, if it's a
> > +    constant, the constant value.  */
> > +typedef struct
> > +{
> > +  latticevalue lattice_val;
> > +  struct
> > +  {
> > +    unsigned int n;
> > +    unsigned int offset;
> > +  } alignment;
> > +} value;
>
> Cut-n-pasto in comment. Need to reflect alignment analysis not CCP.
Whoops, thanks.

> > +	 If it does, then in_ccp_worklist will be zero.  */
> > +      if (ann->in_ccp_worklist)
>
> rename 'in_ccp_worklist' ?
We could rename it to "in_worklist" and make it a general "in-worklist" 
bit for passes to use, unless someone really objects.

> > +static int
> > +gcd (int a, int b)
>
> also available in lambda-code.c.

Yes, in fact, that's where i copied it from. I forgot i wasn't going to 
submit this for the mainline since it was in stage2, and thus, i had 
that code available already.

I'll just make the lambda one external and reuse it.