[Mesa-dev] [PATCH] glsl: store read vector in a temp in vec_index_to_cond

[Kenneth Graunke]

On 11/06/2012 07:12 AM, Vincent Lejeune wrote:
> Vector indexing on matrixes generates several copy of the
> constant matrix, for instance vec=mat4[i][j] generates :
> vec=mat4[i].x;
> vec=(j==1)?mat4[i].y;
> vec=(j==2)?mat4[i].z;
> vec=(j==3)?mat4[i].w;
> In the case of constant matrixes, the mat4[i] expression generates
> copy of the 16 elements of the matrix 4 times ; indirect addressing
> also prevents some conservative CSE algorithms (like the one in LLVM)
> from factoring the mat4[i] expression.
> This patch will make the vec_index_to_cond pass generates :
> temp = mat4[i];
> vec=temp.x;
> vec=(j==1)?temp.y;
> vec=(j==2)?temp.z;
> vec=(j==3)?temp.w;
> ---
>   src/glsl/lower_vec_index_to_cond_assign.cpp | 16 ++++++++++++----
>   1 file changed, 12 insertions(+), 4 deletions(-)
>
> diff --git a/src/glsl/lower_vec_index_to_cond_assign.cpp b/src/glsl/lower_vec_index_to_cond_assign.cpp
> index fce9c34..f8b4131 100644
> --- a/src/glsl/lower_vec_index_to_cond_assign.cpp
> +++ b/src/glsl/lower_vec_index_to_cond_assign.cpp
> @@ -68,9 +68,9 @@ ir_rvalue *
>   ir_vec_index_to_cond_assign_visitor::convert_vec_index_to_cond_assign(ir_rvalue *ir)
>   {
>      ir_dereference_array *orig_deref = ir->as_dereference_array();
> -   ir_assignment *assign;
> -   ir_variable *index, *var;
> -   ir_dereference *deref;
> +   ir_assignment *assign, *value_assign;
> +   ir_variable *index, *var, *value;
> +   ir_dereference *deref, *deref_value;
>      int i;
>
>      if (!orig_deref)
> @@ -95,6 +95,14 @@ ir_vec_index_to_cond_assign_visitor::convert_vec_index_to_cond_assign(ir_rvalue
>      assign = new(base_ir) ir_assignment(deref, orig_deref->array_index, NULL);
>      list.push_tail(assign);
>
> +   /* Store the value inside a temp, thus avoiding matrixes duplication */
> +   value = new(base_ir) ir_variable(orig_deref->array->type, "vec_value_tmp",
> +				    ir_var_temporary);
> +   list.push_tail(value);
> +   deref_value = new(base_ir) ir_dereference_variable(value);
> +   value_assign = new(base_ir) ir_assignment(deref_value, orig_deref->array);
> +   list.push_tail(value_assign);
> +
>      /* Temporary where we store whichever value we swizzle out. */
>      var = new(base_ir) ir_variable(ir->type, "vec_index_tmp_v",
>   				  ir_var_temporary);
> @@ -117,7 +125,7 @@ ir_vec_index_to_cond_assign_visitor::convert_vec_index_to_cond_assign(ir_rvalue
>          * underlying variable.
>          */
>         ir_rvalue *swizzle =
> -	 new(base_ir) ir_swizzle(orig_deref->array->clone(mem_ctx, NULL),
> +	 new(base_ir) ir_swizzle(deref_value->clone(mem_ctx, NULL),
>   				 i, 0, 0, 0, 1);
>
>         deref = new(base_ir) ir_dereference_variable(var);
>

This makes a lot of sense to me.

Reviewed-by: Kenneth Graunke <kenneth at whitecape.org>