[Mesa-dev] [PATCH v2 10/23] mesa: Add a concept of an array format

[Michel Dänzer]

On 04.12.2014 17:17, Iago Toral wrote:
> On Wed, 2014-12-03 at 14:58 -0800, Jason Ekstrand wrote:
>> On Wed, Dec 3, 2014 at 2:52 PM, Ian Romanick <idr at freedesktop.org>
>> wrote:
>>          On 12/01/2014 03:04 AM, Iago Toral Quiroga wrote:
>>          > From: Jason Ekstrand <jason.ekstrand at intel.com>
> (...)
>>
>>          >
>>          > +enum mesa_array_format_datatype {
>>          > +   MESA_ARRAY_FORMAT_TYPE_UBYTE = 0x0,
>>          > +   MESA_ARRAY_FORMAT_TYPE_USHORT = 0x1,
>>          > +   MESA_ARRAY_FORMAT_TYPE_UINT = 0x2,
>>          > +   MESA_ARRAY_FORMAT_TYPE_BYTE = 0x4,
>>          > +   MESA_ARRAY_FORMAT_TYPE_SHORT = 0x5,
>>          > +   MESA_ARRAY_FORMAT_TYPE_INT = 0x6,
>>          > +   MESA_ARRAY_FORMAT_TYPE_HALF = 0xd,
>>          > +   MESA_ARRAY_FORMAT_TYPE_FLOAT = 0xe,
>>          > +};
>>          > +
>>          > +#define MESA_ARRAY_FORMAT_TYPE_IS_SIGNED 0x4
>>          > +#define MESA_ARRAY_FORMAT_TYPE_IS_FLOAT 0x8
>>          > +#define MESA_ARRAY_FORMAT_BIT 0x80000000
>>
>>          This...
>>
>>          > +
>>          > +typedef union {
>>          > +   struct {
>>          > +      enum mesa_array_format_datatype type:4;
>>          > +      bool normalized:1;
>>          > +      unsigned num_channels:3;
>>          > +      unsigned swizzle_x:3;
>>          > +      unsigned swizzle_y:3;
>>          > +      unsigned swizzle_z:3;
>>          > +      unsigned swizzle_w:3;
>>          > +      unsigned pad:11;
>>          > +      unsigned array_format_bit:1; /* Must always be 1 */
>>
>>          ...and this are not correct on big-endian.
>>
>>
>> Bah!  I originally did this as a bitfield in a uint32_t and then
>> decided that the union was easier.  Probably best to go back to the
>> bitfield...
>
> Wouldn't it be enough to rewrite the 3 definitions like this?:
>
> #define MESA_ARRAY_FORMAT_BIT (_mesa_little_endian() ?  0x80000000 :
> 0x00000080)

No. With the example above, setting the array_format_bit member to 1 
sets the least significant bit of the aliased 32-bit integer on this 
PowerBook.


The C specification doesn't make any guarantees about the layout of 
bit-fields that would allow them to be used for something like this, so 
you have to stick to either using only bit-fields for flags, or only 
shifts and masks on integer values.


-- 
Earthling Michel Dänzer               |               http://www.amd.com
Libre software enthusiast             |             Mesa and X developer