Re: [patch] Scalar to vector widening

["Ulrich Weigand" <uweigand at de dot ibm dot com>]

Ken Werner wrote:

> This patch adds support for widening scalars to vectors to allow binary 
> operations with mixed operand types. It mainly extends the binop_promote 
> function to call widen_scalar_to_vector if required. No regressions on i686-*-
> linux-gnu.

> 	* value.h (widen_scalar_to_vector): Declare.
> 	* valops.c (widen_scalar_to_vector): New function.
> 	(value_assign): Add call to widen_scalar_to_vector.
> 	* eval.c (binop_promote, evaluate_subexp_standard)
> 	<BINOP_ASSIGN_MODIFY, BINOP_LSH, BINOP_RSH>: Likewise.

I've been thinking about this a bit.  One piece that's missing here
is support for explicit casts.  For OpenCL vectors (and that seems
like a useful extension in general), you can use an explicit type
cast operation to widen a scalar value to a vector, e.g. like

    float f = 1.0f;
    float4 v = (float4) f;

I think it would be better to change your patch to perform the
operation you're doing in the new widen_scalar_to_vector routine
simply in value_cast instead (if the source is a scalar and the
target type a vector type).

This would support the above-mentioned OpenCL language feature,
and it would also simplify implict widening for operators.  Note
that e.g. both binop_promote and value_assign already call
value_cast in their regular operation; it's just a matter of
using the appropriate target type ...

Bye,
Ulrich

-- 
  Dr. Ulrich Weigand
  GNU Toolchain for Linux on System z and Cell BE
  Ulrich.Weigand@de.ibm.com