[PATCH] riscv: Fix alignment-ignorant memcpy implementation

[Adhemerval Zanella Netto]

On 04/03/24 16:30, Vineet Gupta wrote:
> 
> 
> On 3/4/24 11:07, Adhemerval Zanella Netto wrote:
>>
>> On 04/03/24 16:04, Vineet Gupta wrote:
>>>
>>> On 3/4/24 10:30, Adhemerval Zanella Netto wrote:
>>>> On 04/03/24 15:28, Adhemerval Zanella Netto wrote:
>>>>> On 04/03/24 15:24, Andreas Schwab wrote:
>>>>>> On Mär 04 2024, Adhemerval Zanella wrote:
>>>>>>
>>>>>>> Also, there is no reason why the implementation can't be coded in C,
>>>>>>> since it uses only normal registers and the compiler is able to
>>>>>>> generate code as good as the assembly implementation.  I have not
>>>>>>> checked the performance, but the C implementation uses the same
>>>>>>> strategies, the generate code with gcc 13 seems straightforward, and
>>>>>>> the tail code also avoid byte-operations.
>>>>>> RISC-V is a strict-alignment target, and there is a non-zero chance that
>>>>>> the C compiler messes this up.
>>>>>>
>>>>> I am aware, but there is other project that successfully provides C 
>>>>> implementation (musl, for instance, where it optimizes for aligned
>>>>> access) so I think if this implementation does have any mistake this 
>>>>> should not be considered a blocked for this change.
>>>> Maybe we should force -mno-strict-align for the unaligned implementation.
>>> That alone might not be sufficient. RISC-V gcc gates alignment with an
>>> additional cpu tune param and latter overrides the usual
>>> -m[no-]strict-align: meaning you could have more strict alignment but
>>> not any less. The default -mtune is rocket which penalizes unaligned access.
>> How does gcc optimization and strict-align plays with __may_alias__ it
>> would still tries to generated aligned access in this case?
> 
> I'm not sure about the semantics of this attribute. However gcc backend
> effectively enforces -mstrict-align for the default -mtune/-mcpu
> overriding -mno-strict-align.
> 
>   riscv_slow_unaligned_access_p = (cpu->tune_param->slow_unaligned_access
>                    || TARGET_STRICT_ALIGN);
> 
> /* Implement TARGET_SLOW_UNALIGNED_ACCESS.  */
> 
> static bool
> riscv_slow_unaligned_access (machine_mode, unsigned int)
> {
>   return riscv_slow_unaligned_access_p;
> }

Right, so the question is whether compiler will really mess the unaligned
code using __may_alias__ and if we really should keep the implementation as
a assembly one.