[Mesa-dev] [PATCH v1 2/7] etnaviv: Add return statement to etna_amode so compiler is happy

[Emil Velikov]

On 19 June 2017 at 20:52, Chad Versace <chadversary at chromium.org> wrote:
> On Fri 16 Jun 2017, Christian Gmeiner wrote:
>> 2017-06-16 14:54 GMT+02:00 Emil Velikov <emil.l.velikov at gmail.com>:
>> > On 15 June 2017 at 21:47, Robert Foss <robert.foss at collabora.com> wrote:
>> >> From: Tomeu Vizoso <tomeu.vizoso at collabora.com>
>> >>
>> >> Signed-off-by: Robert Foss <robert.foss at collabora.com>
>> >> ---
>> >>  src/gallium/drivers/etnaviv/etnaviv_compiler.c | 2 ++
>> >>  1 file changed, 2 insertions(+)
>> >>
>> >> diff --git a/src/gallium/drivers/etnaviv/etnaviv_compiler.c b/src/gallium/drivers/etnaviv/etnaviv_compiler.c
>> >> index eafb511bb8..8f73113059 100644
>> >> --- a/src/gallium/drivers/etnaviv/etnaviv_compiler.c
>> >> +++ b/src/gallium/drivers/etnaviv/etnaviv_compiler.c
>> >> @@ -885,6 +885,8 @@ etna_amode(struct tgsi_ind_register indirect)
>> >>     default:
>> >>        assert(!"Invalid swizzle");
>> >>     }
>> >> +
>> >> +   return 0;
>> > All the cases are handled correctly and even the default one will
>> > never be reached.
>> > Guess the compiler isn't smart enough, since we're using int:2 while
>> > in reality we're storing an enum and using enum:2 might not always
>> > work.
>> >
>> > Alternative solutions is s/assert/unreachable/. The call will be up-to
>> > the etna devs, but including something like the above (in the commit
>> > summary or elsewhere) will be a good idea, IMHO.
>> >
>>
>> I really like Emil's idea of s/assert/unreachable/
>
> If you replace assert() with unreachable() for an exhaustive switch
> statement, it's best to delete the default case and call unreachable()
> *after* the switch statement. That way, if new enums are ever added, the
> compiler will emit a -Wswitch warning (which would be otherwise
> supressed by the default case); someone will hopefully notice the
> warning, fix it, and save users from an out-of-bounds instruction
> pointer.

That's a very good idea, I think various parts of Gallium already use it.

To do that one needs to tweak struct tgsi_ind_register first - see below.
IIRC some versions of GCC have been having issues with explicitly
sized enums, although we already use it elsewhere in Mesa so it's not
a deal breaker.

struct tgsi_ind_register
{
...
-   unsigned Swizzle : 2;  /* TGSI_SWIZZLE_ */
+   enum tgsi_swizzle Swizzle : 2;  /* TGSI_SWIZZLE_ */
...
}

-Emil