Re: mlx debug build error with clang

[Adrien Mazarguil <adrien.mazarguil@6wind.com>]

Hi Ferruh,

On Tue, Jun 13, 2017 at 04:32:03PM +0100, Ferruh Yigit wrote:
> Hi Adrien, Nelio,
> 
> I am getting following build error [1] with clang [2] when debug enabled
> for mlx4 and mlx5.
> 
> This started after I update my box, not sure what triggered this.
> Can you please help fixing this?
> 
> Thanks,
> ferruh
> 
> 
> [1]
> 
> .../drivers/net/mlx4/mlx4_flow.c:731:3: error: use of GNU statement
> expression extension [-Werror,-Wgnu-statement-expression]
>                 claim_zero(ibv_destroy_qp(fdq->qp));
>                 ^
> .../drivers/net/mlx4/mlx4.h:185:25: note: expanded from macro 'claim_zero'
> #define claim_zero(...) assert((__VA_ARGS__) == 0)
>                         ^
> /usr/include/assert.h:95:6: note: expanded from macro 'assert'
>     ({                                                                  \
>      ^
> 
> ....
> 
> .../drivers/net/mlx5/mlx5_fdir.c:278:2: error: use of GNU statement
> expression extension [-Werror,-Wgnu-statement-expression]
>         assert(((uint8_t *)attr + sizeof(*attr)) == (uint8_t *)spec_offset);
>         ^
> /usr/include/assert.h:95:6: note: expanded from macro 'assert'
>     ({                                                                  \
>      ^
> 
> [Many of similar ...]
> 
> 
> [2]
> target: x86_64-native-linuxapp-clang
> 
> clang version 4.0.0 (tags/RELEASE_400/final)

Recent Glibc versions now apparently handle assert() through a nonstandard
({ }) construct, which is not pedantic-safe due to a missing __extension__
keyword.

assert.h still provides a standard assert() definition that shouldn't cause
compilation to fail when the following condition is met:

 #if !defined __GNUC__ || defined __STRICT_ANSI__

However __GNUC__ is (always?) defined by clang for maximum compatibility
with GCC while __STRICT_ANSI__ is not due to the -std=gnu99 parameter,
assert.h thinks it's OK to use a ({ }) construct but is then caught by
clang's -pedantic parameter.

There are two ways to address this issue while keeping our beloved -pedantic
parameter in debug mode:

1. Replacing -std=gnu99 with -std=c99 (which is even stricter) to bring back
   __STRICT_ANSI__.
2. Replacing assert() statements with RTE_ASSERT().

The former should be doable now that DPDK includes have been cleaned up, and
we're thinking about doing the latter at some point.

Since I don't have a recent Glibc handy, can you try replacing -std=gnu99
with -std=c99 in both Makefiles (mlx4 and mlx5), and report how GCC and
clang fare? (GCC at least seems to have no problem with that on my side)

-- 
Adrien Mazarguil
6WIND