[dpdk-dev] [PATCH v2 0/1] net: fix aliasing issue in checksum computation

[Georg Sauthoff <mail@gms.tf>]

The current __rte_raw_cksum() function violates the C strict-aliasing
rules since it uses a uint8_t pointer to access a trailing byte.

This patch also fixes a superfluous cast, i.e.:

	uintptr_t ptr = (uintptr_t)buf;
	typedef uint16_t __attribute__((__may_alias__)) u16_p;
	const u16_p *u16_buf = (const u16_p *)ptr;

Transitive casting involving uintptr_t doesn't solve anything here.
It also doesn't help with fixing a strict-aliasing issue here.

The patch also simplifies the main loop, i.e. it eliminates the manually
unrolled loop

	while (len >= (sizeof(*u16_buf) * 4)) {
		sum += u16_buf[0];
		sum += u16_buf[1];
		sum += u16_buf[2];
		sum += u16_buf[3];
		len -= sizeof(*u16_buf) * 4;
		u16_buf += 4;
	}

since modern C compilers are in a better position to decide which level of
unrolling is optimal for the target architecture.

See also https://godbolt.org/z/6rYbYGnj7 which shows how GCC auto-vectorizes
the simplified loop using AVX instructions, when compiling for Haswell. When
looking at the number of instructions in the compiled code, the new version
is half as big as the existing one.

Signed-off-by: Georg Sauthoff <mail@gms.tf>
---
v2:
* Reword commit message (detail aliasing implications of uint8_t)
* Add unlikely()

Georg Sauthoff (1):
  net: fix aliasing issue in checksum computation

 lib/net/rte_ip.h | 27 ++++++++-------------------
 1 file changed, 8 insertions(+), 19 deletions(-)

-- 
2.31.1