Re: Valarrays and [altivec] vectors (was Re: [PATCH]: Handle vector modes in genopinit.c, c-typeck.c)

[Gabriel Dos Reis <gdr at codesourcery dot com>]

"Joseph S. Myers" <jsm28@cam.ac.uk> writes:

| On 30 Nov 2001, Gabriel Dos Reis wrote:
| 
| > | > When can valarrays be used in ? : conditional expressions?
| > | 
| > | I don't know of any restrictions here.  Gaby?
| > 
| > There is no implicit conversion valarray<T> -> bool.  So they can't be
| > used in the condition part of "?:", but they are certainly valid
| > arguments for the second and third operands are any C++ objects. 
| 
| When used as the second or third arguments, must the second and third 
| arguments then be valarrays of the same type?

Yes.  There is no standard conversion valarray<T> -> valarray<U>.

| > | > Can operations be applied between valarrays of different lengths?
| > | 
| > | Undefined behavior.  I think the current implementation formats the user's
| > | hard drive.
| > 
| > Exact :-)
| 
| I hope that for vectors in C we make this a constraint violation.

That is a constraint violation in C++ as well, but then what should
happen if that constraint is violated? In C++, it is an undefined
behaviour.  What do you propose for C?

| > | > Can
| > | > they be applied between valarrays and scalars?
| > | 
| > | Yes.
| > 
| > Exact.
| 
| What's the result?  Applying the operation between each element and the
| scalar (the scalar being evaluated only once) and returning a valarray of
| the same type as the valarray argument? 

Exactly.

| Does this apply to such
| operations as "valarray = scalar" (setting each element) and "valarray +=
| scalar" (adding scalar to each element)?

Yes, (known in Fortran as a "broadcast operation").

| What's the result of array indexing on a valarray?

There our four (4) kinds of indexing valarray<>s in C++

   1) index by a slice (start, end, stride).  The result is a either a
   slice_array<T> (having a reference semantics and convertible to a
   valarray<T> value) or a value of type valarray<T>. 

   2) index by a generalized slice (glice) -- this is the C++ way of
   constructing a multidimensional array out of a unidimensional
   array.  The result is either a gslice<T> (having a reference
   semantics, and convertible to a valarray<T> value) or a value of
   type valarray<T>.

   3) index by a valarray<size_t> (a mapping from "logical" slots to
   "physical" slots).  The result is either an indirect_array<T>
   (having a referebce semantics, and convertible to a valarray<T>
   object) or a value of type valarray<T>.

   4) index by a valarray<bool> (pretty much close to the
   WHERE-construct in FORTRAN).  The result is either a mask_array<T>
   (having a reference semantics, and convertible to a valarray<T>
   object) or a value of type valarray<T>.

| Ordinary array
| indexing extracting a single value, or applying the indexing to each
| element and producing another valarray?  (We need to define in C the way
| of extracting and setting individual vector elements.)

I'm unsure of what you mean here.  Please, could you elaborate?

If v is a valarray<T> and i is a

  1) size_t, then v[i] is a T
  
  2) slice, then v[i] is slice_array<T> when used in an "lvalue"
     context, and a valarray<T> in an "rvalue" context". 

  3) gslice, then v[i] is gslice_array<T> when used in an "lvalue"
     context, and a valarray<T> in an "rvalue" context.

  4) valarray<size_t>, then v[i] is an indirect_array<T> when used in
     an "lvalue" context, and a valarray<T> in an "rvalue" context.

  5) valarray<bool>, then v[i] is a mask_array<T> when used in an
  "lvalue" context, and a valarray<T> in an "rvalue" context.

-- Gaby