[m-dev.] Why do strings in Mercury need to be word aligned?

[Paul Bone]

On Wed, May 06, 2009 at 10:46:09PM +1000, Peter Ross wrote:
> On Wed, May 6, 2009 at 6:57 PM, Paul Bone <pbone at csse.unimelb.edu.au> wrote:
> > On Wed, May 06, 2009 at 06:31:26PM +1000, Peter Ross wrote:
> >> Hi,
> >>
> >> Why does a MR_String need to be word-aligned?  What code will ever
> >> look at the tag bits of a string and hence do the wrong thing?  I'm
> >> sure that there is code, but I just can't think of what it is.
> >
> > I think that if you have:
> >
> > :- type maybe(T)
> >    --->    yes(T)
> >    ;       no.
> >
> > And then use the value yes("fish") the word-aligned string is packed with the
> > tag bits for maybe so that it uses only two memory cells one for the actual
> > string and one for maybe.
> >
> >  +-----+---------+        +--------+
> >  | tag | pointer |------> | String |
> >  +-----+---------+        +--------+
> >
> > If we couldn't use the lower two bits of the string pointer this structure
> > would have to be:
> >
> >  +-----+---------------------------+     +----------------+     +--------+
> >  | tag | pointer to string pointer |---> | String Pointer |---> | String |
> >  +-----+---------------------------+     +----------------+     +--------+
> >
> > Other examples with other structures will also do similar things.  It also
> > doesn't require use of the ploymorphic type as above.
> >
> Sounds like a reasonable explanation to me, only one slight concern is
> that the compiler doesn't seem to do this optimization today at least
> for the following function.
> 
> :- func f(string) = maybe(string).
> f(A) = yes(A).
> 
> Is the reason above correct, or is there another more fundamental
> reason than not allowing an optimization?

Speaking to some of the other guys here.  Turns out that I'm wrong and the
compiler can't do this since it doesn't know in advance that you're going to
use maybe(string) when you type maybe(T).  You'd have to duplicate code for the
special case when you can remove the extra pointer.

Going back to your original question about why string pointers must be word
aligned.  I think the new anwser to this is "I don't know".


-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 197 bytes
Desc: Digital signature
URL: <http://lists.mercurylang.org/archives/developers/attachments/20090507/df69f840/attachment.sig>