[Mesa-dev] [PATCH 09/10] glsl: Optimize A - neg(B) == A + B

[Ian Romanick]

I haven't really looked at the other patches yet, but the subject of
this one caught my eye.

I'm not sure this is useful.  I believe we lower A-B to A+neg(B) for all
architectures.  I'm pretty sure we also already have a pass that
converts A+neg(neg(B)) to A+B.  Did you write any shaders to try to
tickle this code?  What did something like the following produce before
and after this change?

uniform vec4 a;
uniform vec4 b;
void main()
{
    gl_Position = a - (-b);
}

For the neg(A)-B case, I think the existing code is better for most
architectures.  You end up with a single

    ADD   r2, -r1, -r0

instruction currently.  With this change you would potentially need an
extra move to apply the negation.  I believe at least the i965 backend
has a pass to push those negations "down" the tree... effectively
undoing what your pass does. :)

On 07/14/2014 03:22 PM, thomashelland90 at gmail.com wrote:
> From: Thomas Helland <thomashelland90 at gmail.com>
> 
> ...and  neg(A) - B == neg(A + B)
> 
> Signed-off-by: Thomas Helland <thomashelland90 at gmail.com>
> ---
>  src/glsl/opt_algebraic.cpp | 6 ++++++
>  1 file changed, 6 insertions(+)
> 
> diff --git a/src/glsl/opt_algebraic.cpp b/src/glsl/opt_algebraic.cpp
> index 2361d0f..fba9de6 100644
> --- a/src/glsl/opt_algebraic.cpp
> +++ b/src/glsl/opt_algebraic.cpp
> @@ -454,6 +454,12 @@ ir_algebraic_visitor::handle_expression(ir_expression *ir)
>        /* X - X == 0 */
>        if (ir->operands[0]->equals(ir->operands[1]))
>           return ir_constant::zero(ir, ir->type);
> +      /* A - neg(B) = A + B */
> +      if (op_expr[1] && op_expr[1]->operation == ir_unop_neg)
> +         return add(ir->operands[0], op_expr[1]->operands[0]);
> +      /* neg(A) - B = neg(A + B) */
> +      if (op_expr[0] && op_expr[0]->operation == ir_unop_neg)
> +         return neg(add(op_expr[0]->operands[0], ir->operands[1]));
>        break;
>  
>     case ir_binop_mul:
>