Re: [PATCH v11 11/26] target/loongarch: Add floating point comparison instruction translation

[Richard Henderson]

On 11/30/21 9:22 AM, gaosong wrote:
> On 2021/11/20 下午5:02, Richard Henderson wrote:
> > > +#define FCMP_LT 0x0001  /* fp0 < fp1 */
> > > +#define FCMP_EQ   0x0010  /* fp0 = fp1 */
> > > +#define FCMP_UN   0x0100  /* unordered */
> > > +#define FCMP_GT   0x1000  /* fp0 > fp1 */
> >
> > Any reason why these bits are not sequential? 
...
> We should like:
>
> #define FCMP_LT    0x1  /* fp0 < fp1 */
> #define FCMP_EQ   0x2  /* fp0 = fp1 */
> #define FCMP_UN   0x3  /* unordered */
> #define FCMP_GT   0x4  /* fp0 > fp1 */
>
> static uint32_t get_fcmp_flags(int cond)
> {
>      uint32_t flags = 0;
>
>      if (cond & 0x1) {
>          flags |= FCMP_LT;
>      }
>      if (cond & 0x2) {
>          flags |= FCMP_EQ;
>      }
>      if (cond & 0x3) {
>          flags |= FCMP_UN;
>      }
>      if (cond & 0x4) {
>          flags |= FCMP_GT;
>      }
>      return flags;
> }
>
> Is this right?

No.  You're not converting anything here.

I think you should simply replace "0x" with "0b" so that the bits of FCMP are more 
compact.  I assume that's what you were originally thinking.

#define FCMP_LT   0b0001  /* fp0 < fp1 */
#define FCMP_EQ   0b0010  /* fp0 = fp1 */
#define FCMP_UN   0b0100  /* unordered */
#define FCMP_GT   0b1000  /* fp0 > fp1 */

or identically with the form (1 << 0), (1 << 1), etc.


r~