[PATCH 1/5] d3dcompiler: Emit constructors as sequences of stores to a synthetic variable.

Zebediah Figura zfigura at codeweavers.com
Mon Jun 8 10:06:51 CDT 2020


On 6/8/20 9:06 AM, Matteo Bruni wrote:
> On Sat, Jun 6, 2020 at 12:20 AM Zebediah Figura <z.figura12 at gmail.com> wrote:
>>
>> @@ -2415,41 +2414,60 @@ postfix_expr:             primary_expr
>>                                   }
>>                                   $$ = append_binop($1, $3, &load->node);
>>                               }
>> -                          /* "var_modifiers" doesn't make sense in this case, but it's needed
>> -                             in the grammar to avoid shift/reduce conflicts. */
> 
> BTW there are 3 of those comments in hlsl.y, all different from each
> other... Good job me.
> 
>> +            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)
>> +            {
>> +                const struct hlsl_type *arg_type = $4.args[i]->data_type;
>> +                unsigned int component_count = arg_type->dimx * arg_type->dimy;
>> +
>> +                if (!(assignment = new_assignment(var, NULL, $4.args[i],
>> +                        ((1 << component_count) - 1) << writemask_offset, $4.args[i]->loc)))
>> +                    YYABORT;
>> +                writemask_offset += component_count;
> 
> Isn't this wrong for structs or arrays?

Ech, you're right; that should use components_count_type() instead.

> 
> FWIW it seems it would be more easily extensible to matrix and other
> types if you generate a "full" assignment with a swizzle on the lhs
> and then delegate fixing it up to make_assignment() rather than
> directly computing the writemask here.
> 

It doesn't particularly feel like less work, when all we're missing here 
is implicit_conversion(), but we'd have to generate a load and swizzle 
instruction.



More information about the wine-devel mailing list