On 02/24/2014 02:01 PM, H. Peter Anvin wrote:
On 02/24/2014 04:51 AM, H. Peter Anvin wrote:
On 02/24/2014 04:41 AM, Jan Beulich wrote:
So I'm guessing this hash is deliberately using the CRC32 instruction
"backwards", which would actually make sense: an actual CRC is actually
a pretty poor hash due to linearity.
OK, it really is even more confusing than that.
It does seem like the crc32 instruction really *is* commutative, which
isn't something I would personally have expected at all.
Given that fact, I suspect the ordering in the DPDK is actually a bug,
and that we should correct the ordering (which I would do at the call
sites because it seems to make the code clearer) because it reduces the
size of the loop by two instructions.
I guess I should find out how to file a bug report against DPDK too...
Looking through the DPDK project git history, it seems that this was a bug
introduced when changing from using inline assembly to using intrinsics:
static inline uint32_t
rte_hash_crc_4byte(uint32_t data, uint32_t init_val)
{
- asm volatile("crc32 %[data], %[init_val]"
- : [init_val]"=r" (init_val)
- : [data]"r" (data), "[init_val]" (init_val));
- return init_val;
+ return _mm_crc32_u32(data, init_val);
}
Good point, I also just noticed that in the git blame.
The operand order, of course, of the intrinsic being the opposite of AT&T-style
assembly.
I never expected that the CRC32 operation would be commutative. Very
fascinating.
Indeed.
-hpa
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majord...@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/
