RE: [PATCH v3] eal/x86: optimize memcpy of small sizes

[Konstantin Ananyev <konstantin.ananyev@huawei.com>]

> > > > > +/**
> > > > > + * Copy bytes from one location to another,
> > > > > + * locations should not overlap.
> > > > > + * Use with n <= 16.
> > > > > + *
> > > > > + * Note: Copying uninitialized memory is perfectly acceptable.
> > > > > + * Using e.g. memcpy(dst, src, 8) instead of
> > > > > + * *(unaligned_uint64_t*) = *(const unaligned_uint64_t *)src
> > > > > + * avoids compiler warnings about source data may be
> > uninitialized
> > > > > + * [-Wmaybe-uninitialized].
> > > > > + *
> > > > > + * Note: Using "n & X" generates 3-byte "test" instructions,
> > > > > + * instead of "n >= X", which would generate 4-byte "cmp"
> > > > instructions.
> > > > > + */
> > > > > +static __rte_always_inline void *
> > > > > +rte_mov16_or_less(void *dst, const void *src, size_t n)
> > > > > +{
> > > > > +	/* Faster way when size is known at build time. */
> > > > > +	if (__rte_constant(n)) {
> > > > > +		if (n == 2)
> > > > > +			return memcpy(dst, src, 2);
> > > > > +		if (n == 4)
> > > > > +			return memcpy(dst, src, 4);
> > > > > +		if (n == 6) /* 4 + 2 */
> > > > > +			return memcpy(dst, src, 6);
> > > > > +		if (n == 8)
> > > > > +			return memcpy(dst, src, 8);
> > > > > +		if (n == 10) /* 8 + 2 */
> > > > > +			return memcpy(dst, src, 10);
> > > > > +		if (n == 12) /* 8 + 4 */
> > > > > +			return memcpy(dst, src, 12);
> > > > > +		if (n == 16) {
> > > > > +			rte_mov16((uint8_t *)dst, (const uint8_t
> > *)src);
> > > > > +			return dst;
> > > > > +		}
> > > > > +	}
> > > > > +
> > > > > +	if (n & 0x18) { /* n >= 8 */
> > > >
> > > > Probably 'n & 0x8'?
> > >
> > > It's intentional, to catch n == 0x10 too.
> > > It seems the associated comment should be more verbose. How about:
> > > if (n & 0x18) { /* n >= 8, including n == 0x10, hence n & 0x18 */
> >
> > Ok, why just not simply : if (n >= 8) then?
> 
> The reason mentioned in the function description:
>  * Note: Using "n & X" generates 3-byte "test" instructions,
>  * instead of "n >= X", which would generate 4-byte "cmp" instructions.
Ah, I see the comment now, sorry missed it first time.
Indeed 'cmp' Will be one byte longer.
Though for non-constant 'n' gcc generates 90B of code for both versions:
https://godbolt.org/z/K6861qGcr

 
> 
> I'll move that comment down here, just before the comparisons.
> 
> >
> > > >
> > > > > +		/* copy 8 ~ 16 bytes */
> > > > > +		memcpy(dst, src, 8);
> > > > > +		memcpy((uint8_t *)dst - 8 + n, (const uint8_t *)src -
> > 8 +
> > > > n, 8);
> > > > > +	} else if (n & 0x4) {
> > > > > +		/* copy 4 ~ 7 bytes */
> > > > > +		memcpy(dst, src, 4);
> > > > > +		memcpy((uint8_t *)dst - 4 + n, (const uint8_t *)src -
> > 4 +
> > > > n, 4);
> > > > > +	} else if (n & 0x2) {
> > > > > +		/* copy 2 ~ 3 bytes */
> > > > > +		memcpy(dst, src, 2);
> > > > > +		memcpy((uint8_t *)dst - 2 + n, (const uint8_t *)src -
> > 2 +
> > > > n, 2);
> > > > > +	} else if (n & 0x1) {
> > > > > +		/* copy 1 byte */
> > > > > +		memcpy(dst, src, 1);
> > > > > +	}
> > > > > +	return dst;
> > > > > +}