[RFC V5 3/3] arm64:add bitrev.h file to support rbit instruction

[Ard Biesheuvel]

On 30 October 2014 13:01, Will Deacon <will.deacon at arm.com> wrote:
> On Wed, Oct 29, 2014 at 05:52:00AM +0000, Wang, Yalin wrote:
>> This patch add bitrev.h file to support rbit instruction,
>> so that we can do bitrev operation by hardware.
>> Signed-off-by: Yalin Wang <yalin.wang at sonymobile.com>
>> ---
>>  arch/arm64/Kconfig              |  1 +
>>  arch/arm64/include/asm/bitrev.h | 28 ++++++++++++++++++++++++++++
>>  2 files changed, 29 insertions(+)
>>  create mode 100644 arch/arm64/include/asm/bitrev.h
>>
>> diff --git a/arch/arm64/Kconfig b/arch/arm64/Kconfig
>> index 9532f8d..b1ec1dd 100644
>> --- a/arch/arm64/Kconfig
>> +++ b/arch/arm64/Kconfig
>> @@ -35,6 +35,7 @@ config ARM64
>>       select HANDLE_DOMAIN_IRQ
>>       select HARDIRQS_SW_RESEND
>>       select HAVE_ARCH_AUDITSYSCALL
>> +     select HAVE_ARCH_BITREVERSE
>>       select HAVE_ARCH_JUMP_LABEL
>>       select HAVE_ARCH_KGDB
>>       select HAVE_ARCH_TRACEHOOK
>> diff --git a/arch/arm64/include/asm/bitrev.h b/arch/arm64/include/asm/bitrev.h
>> new file mode 100644
>> index 0000000..292a5de
>> --- /dev/null
>> +++ b/arch/arm64/include/asm/bitrev.h
>> @@ -0,0 +1,28 @@
>> +#ifndef __ASM_ARM64_BITREV_H
>> +#define __ASM_ARM64_BITREV_H
>> +
>> +static __always_inline __attribute_const__ u32 __arch_bitrev32(u32 x)
>> +{
>> +     if (__builtin_constant_p(x)) {
>> +             x = (x >> 16) | (x << 16);
>> +             x = ((x & 0xFF00FF00) >> 8) | ((x & 0x00FF00FF) << 8);
>> +             x = ((x & 0xF0F0F0F0) >> 4) | ((x & 0x0F0F0F0F) << 4);
>> +             x = ((x & 0xCCCCCCCC) >> 2) | ((x & 0x33333333) << 2);
>> +             return ((x & 0xAAAAAAAA) >> 1) | ((x & 0x55555555) << 1);
>
> Shouldn't this part be in the generic code?
>
>> +     }
>> +     __asm__ ("rbit %w0, %w1" : "=r" (x) : "r" (x));
>
> You can write this more neatly as:
>
>   asm ("rbit %w0, %w0" : "+r" (x));
>

This forces GCC to use the same register as input and output, which
doesn't necessarily result in the fastest code. (e.g., if the
un-bitrev()'ed value is reused again afterwards).
On the other hand, the original notation does allow GCC to use the
same register, but doesn't force it to, so I prefer the original one.

-- 
Ard.