Re: [RFC PATCH v2 30/44] target/loongarch: Implement vclo vclz

[gaosong]

在 2023/4/2 上午11:34, Richard Henderson 写道:
> On 3/27/23 20:06, Song Gao wrote:
> > +#define DO_CLO_B(N)  (clz32((uint8_t)~N) - 24)
> > +#define DO_CLO_H(N)  (clz32((uint16_t)~N) - 16)
>
> I think this is wrong. 
It is wried,  the result is always right. :-\
and  (clz32(~N) - 24)  or (clz32((uint32_t)~N) - 24) is wrong.
> You *want* the high bits to be set, so that they are ones, and 
> included in the count, which you then subtract off.  You want the 
> "real" count to start after the 24th leading 1.
Yes,
and  we use clz32(),   how about the following way?

#define DO_CLO_B(N)  (clz32( ~N & 0xff) -24)
#define DO_CLO_H(N)  (clz32( ~N & 0xffff) -16)

Thanks.
Song Gao