Received-SPF: pass (google.com: best guess record for domain of linux-kernel-owner@vger.kernel.org designates 209.132.180.67 as permitted sender) client-ip=209.132.180.67;
From:   Sergey Senozhatsky <sergey.senozhatsky@gmail.com>
Date:   Sat, 16 Mar 2019 16:18:41 +0900
To:     Stanislav Fomichev <sdf@fomichev.me>
Cc:     Sergey Senozhatsky <sergey.senozhatsky.work@gmail.com>,
        Eric Dumazet <edumazet@google.com>,
        netdev <netdev@vger.kernel.org>,
        LKML <linux-kernel@vger.kernel.org>,
        Sergey Senozhatsky <sergey.senozhatsky@gmail.com>
Subject: Re: [PATCH] tcp: don't use __constant_cpu_to_be32
Message-ID: <20190316071841.GA30386@tigerII.localdomain>
References: <20190314061514.19586-1-sergey.senozhatsky@gmail.com>
 <CANn89iJMFfh1Pkvk9LOn3FwgAfpVjwS7q5x_8Y9zo2rzpy7w9g@mail.gmail.com>
 <20190315012012.GA748@jagdpanzerIV>
 <20190315172428.GB5481@mini-arch.hsd1.ca.comcast.net>
 <20190316051902.GA27630@tigerII.localdomain>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <20190316051902.GA27630@tigerII.localdomain>
User-Agent: Mutt/1.11.4 (2019-03-13)
Sender: linux-kernel-owner@vger.kernel.org
Precedence: bulk

On (03/16/19 14:19), Sergey Senozhatsky wrote:
> > See my recent commit a0517a0f7ef23 ("selftests/bpf: use
> > __bpf_constant_htons in test_prog.c") where compiler was not smart
> > enough to figure it out (you can google similar issues for GCC < 4.8).
> 
> Allow me to disagree.
> 
> The error in a0517a0f7ef23 says
> 
> 	error: implicit declaration of function '__builtin_bswap16'
> 
> which doesn't sound like "the compiler was not smart".
> 
> Let's look at __swab16
> 
> 100 #ifdef __HAVE_BUILTIN_BSWAP16__
> 101 #define __swab16(x) (__u16)__builtin_bswap16((__u16)(x))
> 102 #else
> 103 #define __swab16(x)                             \
> 104         (__builtin_constant_p((__u16)(x)) ?     \
> 105         ___constant_swab16(x) :                 \
> 106         __fswab16(x))
> 107 #endif
> 
> We can call __builtin_bswap16() only when the compiler has that builtin,
> which is what #ifdef __HAVE_BUILTIN_BSWAP16__ for.
> 
> But this is not what tools/testing/selftests/bpf/bpf_endian.h does
> 
> # define __bpf_ntohs(x)                 __builtin_bswap16(x)
> # define __bpf_htons(x)                 __builtin_bswap16(x)
> 
> So I sort of suspect that what should have been done was that
> __HAVE_BUILTIN_BSWAP16__ ifdef, just like what include/uapi/linux/swab.h
> does.

E.g.

#define ____const(x)	printf("const %d\n", (x))
#define ____noconst(x)	printf("noconst %d\n", (x))
#define const_test(x)	(__builtin_constant_p(x) ? ____const((x)) : ____noconst((x)))

The compiler should optimize out !__builtin_constant_p (dead code elimination
phase - dce), if it can compile the code in the first place.

So if we replace ____noconst with

#define ____noconst(x)	printf2("noconst %d\n", (x))

then compiler wouldn't be able to compile the code and thus wouldn't
be able to perform dce.

__builtin_constant_p, unlike `#if 0', is not handled by the preprocessor.

Verified with gcc 8.3, gcc 4.9, gcc 4.7 at compiler explorer website
https://gcc.godbolt.org/ - GCC is smart enough to figure out
__builtin_constant_p and do dce, given that it compile the code.

	-ss