Re: [PATCH v2 65/67] target/arm: Convert SQDMULH, SQRDMULH to decodetree

[Richard Henderson <richard.henderson@linaro.org>]

On 5/28/24 09:10, Peter Maydell wrote:
>> +void HELPER(neon_sqdmulh_idx_s)(void *vd, void *vn, void *vm,
>> +                                void *vq, uint32_t desc)
>> +{
>> +    intptr_t i, j, opr_sz = simd_oprsz(desc);
>> +    int idx = simd_data(desc);
>> +    int32_t *d = vd, *n = vn, *m = (int32_t *)vm + H4(idx);
>> +
>> +    for (i = 0; i < opr_sz / 4; i += 16 / 4) {
>> +        int32_t mm = m[i];
>> +        for (j = 0; j < 16 / 4; ++j) {
>> +            d[i + j] = do_sqrdmlah_s(n[i + j], mm, 0, false, false, vq);
>> +        }
>> +    }
>> +    clear_tail(d, opr_sz, simd_maxsz(desc));
>> +}
>> +
>> +void HELPER(neon_sqrdmulh_idx_s)(void *vd, void *vn, void *vm,
>> +                                 void *vq, uint32_t desc)
>> +{
>> +    intptr_t i, j, opr_sz = simd_oprsz(desc);
>> +    int idx = simd_data(desc);
>> +    int32_t *d = vd, *n = vn, *m = (int32_t *)vm + H4(idx);
>> +
>> +    for (i = 0; i < opr_sz / 4; i += 16 / 4) {
>> +        int32_t mm = m[i];
>> +        for (j = 0; j < 16 / 4; ++j) {
>> +            d[i + j] = do_sqrdmlah_s(n[i + j], mm, 0, false, true, vq);
>> +        }
>> +    }
>> +    clear_tail(d, opr_sz, simd_maxsz(desc));
>> +}
> 
> Missing H macros in these helpers ?

No.  The only index that's not identical across the vector is M, and H is handled once at 
the beginning.  Otherwise n[] and d[] have the same operation applied to all indicies, and 
it doesn't matter which order in which these happen.


r~