Re: [PATCH] lib/string: Bring optimized memcmp from glibc

[Florian Weimer]

* Nikolay Borisov:

> +/*
> + * Compare A and B bytewise in the byte order of the machine.
> + * A and B are known to be different. This is needed only on little-endian
> + * machines.
> + */
> +static inline int memcmp_bytes(unsigned long a, unsigned long b)
> +{
> +	long srcp1 = (long) &a;
> +	long srcp2 = (long) &b;
> +	unsigned long a0, b0;
> +
> +	do {
> +		a0 = ((uint8_t *) srcp1)[0];
> +		b0 = ((uint8_t *) srcp2)[0];
> +		srcp1 += 1;
> +		srcp2 += 1;
> +	} while (a0 == b0);
> +	return a0 - b0;
> +}

Should this be this?

static inline int memcmp_bytes(unsigned long a, unsigned long b)
{
	if (sizeof(a) == 4)
		return __builtin_bswap32(a) < __builtin_bswap32(b) ? -1 : 0;
	else
		return __builtin_bswap64(a) < __builtin_bswap64(b) ? -1 : 0;
}

(Or whatever macro versions the kernel has for this.)

Or is the expectation that targets that don't have an assembler
implementation for memcmp have also bad bswap built-ins?

Thanks,
Florian