[Intel-gfx] [RFC 4/4] drm/i915/registers: define field values using FIELD_PREP()

[Jani Nikula]

On Thu, 27 Sep 2018, Jani Nikula <jani.nikula at intel.com> wrote:
> On Thu, 27 Sep 2018, Chris Wilson <chris at chris-wilson.co.uk> wrote:
>> Quoting Jani Nikula (2018-09-27 10:40:23)
>>> Slightly verbose, but does away with hand rolled shifts and provides
>>> static checking that the values fit the mask.
>>> 
>>> Signed-off-by: Jani Nikula <jani.nikula at intel.com>
>>> ---
>>> @@ -4650,11 +4650,11 @@ enum {
>>>  #define _PP_ON_DELAYS                  0x61208
>>>  #define PP_ON_DELAYS(pps_idx)          _MMIO_PPS(pps_idx, _PP_ON_DELAYS)
>>>  #define  PANEL_PORT_SELECT_MASK                GENMASK(31, 30)
>>> -#define  PANEL_PORT_SELECT_LVDS                (0 << 30)
>>> -#define  PANEL_PORT_SELECT_DPA         (1 << 30)
>>> -#define  PANEL_PORT_SELECT_DPC         (2 << 30)
>>> -#define  PANEL_PORT_SELECT_DPD         (3 << 30)
>>> -#define  PANEL_PORT_SELECT_VLV(port)   ((port) << 30)
>>> +#define  PANEL_PORT_SELECT_LVDS                FIELD_PREP(PANEL_PORT_SELECT_MASK, 0)
>>> +#define  PANEL_PORT_SELECT_DPA         FIELD_PREP(PANEL_PORT_SELECT_MASK, 1)
>>> +#define  PANEL_PORT_SELECT_DPC         FIELD_PREP(PANEL_PORT_SELECT_MASK, 2)
>>> +#define  PANEL_PORT_SELECT_DPD         FIELD_PREP(PANEL_PORT_SELECT_MASK, 3)
>>> +#define  PANEL_PORT_SELECT_VLV(port)   FIELD_PREP(PANEL_PORT_SELECT_MASK, port)
>>
>> Maybe verbose, but it reads far better as giving each field a distinct
>> name ties together all the individual options.
>>
>> Before seeing this I was sceptical about FIELD_PREP, no longer.
>>
>> Under this construct we aren't using masks per se, but giving a name to
>> a group of bits within the register (a field). So I think
>>
>> #define PANEL_PORT_SELECT	GENMASK(31, 30)
>> #define PANEL_PORT_SELECT_LVDS	FIELD_PREP(PANEL_PORT_SELECT, 0)
>
> Argh, I've screwed up. I must have failed to build after adding the last
> patch. You know, RFC and all. This results in:
>
> drivers/gpu/drm/i915/intel_display.c: In function ‘assert_panel_unlocked’:
> drivers/gpu/drm/i915/intel_display.c:1219:3: error: case label does not reduce to an integer constant
>    case PANEL_PORT_SELECT_LVDS:
>
> Fail.
>
> However, taking just the mask&shift part from FIELD_PREP works:
>
> 	(((typeof(_mask))(_val) << __bf_shf(_mask)) & (_mask))
>
> But then we'll lose the static checks, unless we come up with something
> ingenious ourselves.

If my grep-fu serves me right, we have 186 macros defined in i915_reg.h
that are used as case labels, and almost all of them would be candidates
for FIELD_PREP().

Is an alternative FIELD_PREP() worth it without the static checks?

BR,
Jani.


>
> BR,
> Jani.
>
>
>>
>> is intuitive. Did you do a quick bloatometer to see if gcc code
>> generation is affected?
>> -Chris

-- 
Jani Nikula, Intel Open Source Graphics Center