Hi, Richard.

On 2022/9/30 22:54, Richard Sandiford wrote:
> "dongbo (E)" <dongbo4@huawei.com> writes:
>>
>> We tried to put `S_H` in front of `NIL`:
>>
>> ```
>>
>>       #define OP_SVE_Vv_HSD                        \
>> {                                                               \
>>         QLF2(S_H,S_H),                                      \
>>         QLF2(S_S,S_S),                                       \
>>         QLF2(S_D,S_D),                                      \
>>         QLF2(S_H,NIL),                                       \
>>         QLF2(S_S,NIL),                                       \
>>         QLF2(S_D,NIL),                                       \
>>       }
>>
>> ```
> Yeah, good point.  It should be in this order, like you say.
>
> The fixes you mention look correct to me.
>
>> But assembler will fail in `match_operands_qualifier` :(.
>>
>> ```
>>
>>       match_operands_qualifier (aarch64_inst *inst, bool update_p)
>>       {
>>           ...
>>           if (!aarch64_find_best_match (...))
>>           ...
>>           if (inst->opcode->flags & F_STRICT)
>>           {
>>               /* Require an exact qualifier match, even for NIL
>> qualifiers.  */
>>               nops = aarch64_num_of_operands (inst->opcode);
>>               for (i = 0; i < nops; ++i)
>>                   if (inst->operands[i].qualifier != qualifiers[i])
>>                       return false;
>>           }
>>       }
>>
>> ```
> I think that's a misfeature of the F_STRICT handling.  Does it work
> with the patch below?

We cannot find the `tree-data-ref.cc` in binutils, it is a file in GCC?

Is `0001-data-ref-Fix-ranges_maybe_overlap_p-test.patch` the patch you 
meant to send?


We also found a way to fix the F_STRICT matching failure. See patch 
below. :)

The main point is to give another `find_best_match` try when we get a 
qualifier mismatch for HSD operands.


Thanks,

Dong Bo