On 08.03.2024 13:17, Oleksii wrote:
> On Fri, 2024-03-08 at 12:52 +0100, Jan Beulich wrote:
>> On 08.03.2024 12:49, Jan Beulich wrote:
>>> On 08.03.2024 11:14, Oleksii wrote:
>>>> On Fri, 2024-03-08 at 08:26 +0100, Jan Beulich wrote:
>>>>> On 07.03.2024 21:54, Oleksii wrote:
>>>>>> On Thu, 2024-03-07 at 21:49 +0100, Oleksii wrote:
>>>>>>> On Thu, 2024-03-07 at 18:14 +0100, Jan Beulich wrote:
>>>>>>>> For plain writes it should at least be "=Qo" then, yes.
>>>>>>> Constraints Q is a machine specific constraint, and I am
>>>>>>> not sure
>>>>>>> that
>>>>>>> it makes sense to use "=o" only and probably it is a reason
>>>>>>> why
>>>>>>> it is
>>>>>>> enough only "r". Does it make sense?
>>>>>> Probably for RISC-V can be used:
>>>>>> RISC-V—config/riscv/constraints.md
>>>>>> ...
>>>>>> A
>>>>>> An address that is held in a general-purpose register.
>>>>>> ...
>>>>>
>>>>> Just from the description I would have said no, but looking at
>>>>> what
>>>>> "A"
>>>>> actually expands to it is indeed RISC-V's counterpart of Arm's
>>>>> "Q".
>>>>> So
>>>>> yes, this looks like what amo* want to use, and then as a real
>>>>> operand,
>>>>> not just a fake one.
>>>> I am not sure that I know how to check correctly how "A" expands,
>>>> but I
>>>> tried to look at code which will be generated with and without
>>>> constraints and it is the same:
>>>
>>> As expected.
> But if it is epxected and generated code is the same, do we really need
> constraints then?
Yes. Again: Proper constraints are the only way for the compiler to know
everything it needs to know to generate correct code around an asm().
>>>> // static inline void __raw_writel(uint32_t val, volatile void
>>>> __iomem *addr)
>>>> // {
>>>> // asm volatile ( "sw %0, 0(%1)" : : "r" (val), "r"(addr)
>>>> );
>>>> // }
>>>>
>>>> static inline void __raw_writel(uint32_t val, volatile void
>>>> __iomem
>>>> *addr)
>>>> {
>>>> asm volatile ( "sw %0, %1" : : "r" (val), "Ao" (*(volatile
>>>> uint32_t __force *)addr) );
>>>
>>> You want just "A" here though; adding an offset (as "o" permits)
>>> would
>>> yield an insn which the assembler would reject.
>>
>> Wait - this is plain SW, so can't it even be the more generic "m"
>> then?
>> (As said, I'm uncertain about "o"; in general I think it's risky to
>> use.)
> What do you mean by "plain SW"?
The plain store instruction, i.e. not SC. That one permits wider addressing
modes iirc, which we ought to permit where possible.
> Are you suggesting changing 'm' to 'o' so that the final result will be
> "Am"? Based on the descriptions of 'A' and 'm', it seems to me that we
> can just use 'A' alone because both constraints indicate that the
> operand is in memory, and 'A' specifically denotes that an address is
> held in a register.
No, no "A" at all. Just "m", which is a superset of "A" anyway.
>>> Also just to remind
>>> you: In write functions you need "=A" (and in amo ones "+A"), i.e.
>>> the
>>> memory operand then needs to be an output, not an input.
> Could you please clarify about which one amo you are speaking? That one
> who are defined by ATOMIC_OP and ATOMIC_FETCH_OP?
All. They're all read-modify-write operations if I'm not mistaken.
> They are already
> using +A constraints:
> __asm__ __volatile__ ( \
> " amo" #asm_op "." #asm_type " %1, %2, %0" \
> : "+A" (v->counter), "=r" (ret) \
> : "r" (I) \
> : "memory" ); \
Good. I merely thought I'd mention that aspect for completeness.
Jan
