Re: [PATCH 03/17] x86: Replace open-coded parity calculation with parity8()

[H. Peter Anvin]

On 2/24/25 07:24, Uros Bizjak wrote:
> On 23. 02. 25 17:42, Kuan-Wei Chiu wrote:
> > Refactor parity calculations to use the standard parity8() helper. This
> > change eliminates redundant implementations and improves code
> > efficiency.
>
> The patch improves parity assembly code in bootflag.o from:
>
>    58:    89 de                    mov    %ebx,%esi
>    5a:    b9 08 00 00 00           mov    $0x8,%ecx
>    5f:    31 d2                    xor    %edx,%edx
>    61:    89 f0                    mov    %esi,%eax
>    63:    89 d7                    mov    %edx,%edi
>    65:    40 d0 ee                 shr    %sil
>    68:    83 e0 01                 and    $0x1,%eax
>    6b:    31 c2                    xor    %eax,%edx
>    6d:    83 e9 01                 sub    $0x1,%ecx
>    70:    75 ef                    jne    61 <sbf_init+0x51>
>    72:    39 c7                    cmp    %eax,%edi
>    74:    74 7f                    je     f5 <sbf_init+0xe5>
>    76:
>
> to:
>
>    54:    89 d8                    mov    %ebx,%eax
>    56:    ba 96 69 00 00           mov    $0x6996,%edx
>    5b:    c0 e8 04                 shr    $0x4,%al
>    5e:    31 d8                    xor    %ebx,%eax
>    60:    83 e0 0f                 and    $0xf,%eax
>    63:    0f a3 c2                 bt     %eax,%edx
>    66:    73 64                    jae    cc <sbf_init+0xbc>
>    68:
>
> which is faster and smaller (-10 bytes) code.

Of course, on x86, parity8() and parity16() can be implemented very simply:

(Also, the parity functions really ought to return bool, and be flagged 
__attribute_const__.)

static inline __attribute_const__ bool _arch_parity8(u8 val)
{
	bool parity;
	asm("and %0,%0" : "=@ccnp" (parity) : "q" (val));
	return parity;
}

static inline __attribute_const__ bool _arch_parity16(u16 val)
{
	bool parity;
	asm("xor %h0,%b0" : "=@ccnp" (parity), "+Q" (val));
	return parity;
}

In the generic algorithm, you probably should implement parity16() in 
terms of parity8(), parity32() in terms of parity16() and so on:

static inline __attribute_const__ bool parity16(u16 val)
{
#ifdef ARCH_HAS_PARITY16
	if (!__builtin_const_p(val))
		return _arch_parity16(val);
#endif
	return parity8(val ^ (val >> 8));
}

This picks up the architectural versions when available.

Furthermore, if a popcnt instruction is known to exist, then the parity 
is simply popcnt(x) & 1.

	-hpa