Vector types and type conversions

[Chris Lattner]

On Apr 19, 2006, at 1:56 PM, Mark Mitchell wrote:

> Let's accept that both the bit-preserving and value-preserving
> conversions would be useful.  How do we differentiate the two?
>

> For vectors, these operators would apply to each element, and then
> return a vector with the same number of elements as the input vector.
>
> I suppose that we could introduce these operators into C as well,
> although the pseudo-template syntax might not be as natural for C  
> users.

If you're interested in adding value for generic vectors, it may be  
interesting to consider element access operations.  Being able to  
write things like:

  float X = V[1];
or
  V[2] = X;

... would be incredibly useful for people trying to use generic  
vectors when they can.  In particular, it makes it possible to  
generate good code for common operations like splat/broadcast  
operations, where without element access you have to go through  
memory.  An example would look like this:

vector v1, v2 = ...

v1 = (vector){v2[0], v2[0], v2[0], v2[0]};

Currently the only way to get or set elements of a vector is with non- 
TBAA safe pointer arithmetic and vector/array unions.


I've recently been working with people trying to use generic vectors,  
and I feel that insertion/extraction is more important than type  
conversion operations.  This is because they generally map better to  
hardware and provide a way to access various shuffle operations.  As  
it has been noted, various vector ISA's are quite non-orthogonal when  
it comes to type conversion operations, but they do generally have  
efficient ways to do common permute operations (e.g. splat).   
Further, the syntax is trivial and common between C/C++.  That said,  
any forward progress on generic vector support is great! :)

-Chris