Re: [PATCH v2 3/6] d3dcompiler: Emit constructors as sequences of stores to a synthetic variable.

[Matteo Bruni]

On Mon, Jun 8, 2020 at 11:50 PM Zebediah Figura  wrote:
> -
>        if ($2->type == HLSL_CLASS_MATRIX)
> -
>            FIXME("Matrix constructors are not supported yet.\n");
> -
> -
>        sprintf(name, "<constructor-%x>", counter++);
> -
>        if (!(var = new_synthetic_var(name, $2, get_location(&@2))))
> -
>            YYABORT;
> -
>        for (i = 0; i < $4.args_count; ++i)
> -
>        {
> -
>            struct hlsl_ir_node *arg = $4.args[i];
> -
>            unsigned int width;
> -
> -
>            if (arg->data_type->type == HLSL_CLASS_OBJECT)
> -
>            {
> -
>                hlsl_report_message(arg->loc, HLSL_LEVEL_ERROR,
> -
>                        "invalid constructor argument");
> -
>                continue;
> -
>            }
> -
>            width = components_count_type(arg->data_type);
> -
> -
>            if (width > 4)
> -
>            {
> -
>                FIXME("Constructor argument with %u components.\n", width);
> -
>                continue;
> -
>            }
> -
> -
>            if (!(arg = implicit_conversion(arg,
> -
>                    hlsl_ctx.builtin_types.vector[$2->base_type][width - 1], &arg->loc)))
> -
>                continue;
> -
> -
>            if (!(assignment = new_assignment(var, NULL, arg,
> -
>                    ((1 << width) - 1) << writemask_offset, arg->loc)))
> -
>                YYABORT;
> -
>            writemask_offset += width;
> -
>            list_add_tail($4.instrs, &assignment->node.entry);

I'm still not sold on doing it like this. This generates nice code for
vector constructors but it seems hard to extend to matrices in the
future.

I'd do it component-by-component, generating swizzles as necessary.
Any tidy up can be done by following passes, or not at all.