Translation of multidimensional arrays

[Paul Brook]

> There are two reasons presented in trans-types.c for why this haven't
> been done this way since the beginning:
>
> 1) "there is no way to explicitly set the array stride, so
>    all data must be packed" -- this is no longer true, it is possible to
>    specify the stride in ARRAY_REF arbitrarily.
> 2) it is difficult to construct the correct type for such an array,
>    requiring use of PLACEHOLDER_EXPR.  This is just a technical
>    difficulty, not really an unsurmountable problem. Also,
>    it is possible to avoid use of PLACEHOLDER_EXPRs; in fact, as
>    long as the stride and lower bound is specified explicitly in the
>    ARRAY_REF, the bounds and sizes specified in the type are not
>    really used anywhere, so even keeping them NULL when they are
>    non-constant would work (although that would be a bit hacky).

As long as the backend doesn't assume these values are invariant, this should 
be ok.  Allocatable arrays can be re-allocated with a different shape, and 
the expressions may not even be valid at all points where the variable is in 
scope. These as middle-end questions as much as anything else. 

3) Maintaining type consistency and proper alias information gets harder. 
Especially when you include modules and inlining. Currently we use the same 
type_t[0...<unknown>] for pretty much everything and TBAA assumes they can 
all alias. There are probably complications where we have two representations 
or temporary copies of an array (eg. with function arguments). Also there 
will probably be issues with the way reshape() and constant array 
constructors are implemented.

> So I think doing this transition should be fairly straightforward,
> although of course time consuming.

I'd expect there to be a significant number of awkward corner cases, and doing 
the conversion is a fairly big job.

If you're rewriting all this stuff it may be a convenient time to reconsider 
whether strides are stored in bytes or elements.

Paul