Re: [RFC PATCH v2 09/44] target/loongarch: Implement vhaddw/vhsubw

[gaosong]

在 2023/3/29 上午4:17, Richard Henderson 写道:
> On 3/27/23 20:05, Song Gao wrote:
> > +#define DO_ODD_EVEN_S(NAME, BIT, T, E1, E2, DO_OP)                 \
> > +void HELPER(NAME)(CPULoongArchState *env,                          \
> > +                  uint32_t vd, uint32_t vj, uint32_t vk)           \
> > +{                                                                  \
> > +    int i;                                                         \
> > +    VReg *Vd = &(env->fpr[vd].vreg);                               \
> > +    VReg *Vj = &(env->fpr[vj].vreg);                               \
> > +    VReg *Vk = &(env->fpr[vk].vreg);                               \
> > +                                                                   \
> > +    for (i = 0; i < LSX_LEN/BIT; i++) {                            \
> > +        Vd->E1(i) = DO_OP((T)Vj->E2(2 * i + 1), (T)Vk->E2(2 * i)); \
> > + }                                                              \
> > +}
> ...
> > +#define DO_ODD_EVEN_U(NAME, BIT, TD, TS, E1, E2, 
> > DO_OP)                     \
> > +void HELPER(NAME)(CPULoongArchState 
> > *env,                                    \
> > +                  uint32_t vd, uint32_t vj, uint32_t 
> > vk)                     \
> > +{ \
> > +    int i; \
> > +    VReg *Vd = &(env->fpr[vd].vreg); \
> > +    VReg *Vj = &(env->fpr[vj].vreg); \
> > +    VReg *Vk = &(env->fpr[vk].vreg); \
> > + \
> > +    for (i = 0; i < LSX_LEN/BIT; i++) 
> > {                                      \
> > +        Vd->E1(i) = DO_OP((TD)(TS)Vj->E2(2 * i + 1), 
> > (TD)(TS)Vk->E2(2 * i)); \
> > + } \
> > +}
>
> In the first case we have one cast, in the second case we have two.  I 
> wonder if it would be clearer to have both signed and unsigned members 
> in the VReg union? 

I really agree this.

> Then these two macros could be combined.
>
> I also think we could make use of (__typeof(Vd->E1(0))) instead of 
> separately passing the output type?  It would appear to be less 
> error-prone.
I will try this on v3.

Thanks.
Song Gao