Re: [PATCH] KVM: s390: get rid of register asm usage

[Christian Borntraeger <borntraeger@de.ibm.com>]

On 22.06.21 09:57, Christian Borntraeger wrote:
> 
> 
> On 22.06.21 09:53, Thomas Huth wrote:
>> On 22/06/2021 09.50, Christian Borntraeger wrote:
>>>
>>>
>>> On 22.06.21 09:46, Thomas Huth wrote:
>>>> On 22/06/2021 09.43, Christian Borntraeger wrote:
>>>>>
>>>>>
>>>>> On 22.06.21 09:36, Thomas Huth wrote:
>>>>>> On 21/06/2021 16.03, Heiko Carstens wrote:
>>>>>>> Using register asm statements has been proven to be very error prone,
>>>>>>> especially when using code instrumentation where gcc may add function
>>>>>>> calls, which clobbers register contents in an unexpected way.
>>>>>>>
>>>>>>> Therefore get rid of register asm statements in kvm code, even though
>>>>>>> there is currently nothing wrong with them. This way we know for sure
>>>>>>> that this bug class won't be introduced here.
>>>>>>>
>>>>>>> Reviewed-by: Christian Borntraeger <borntraeger@de.ibm.com>
>>>>>>> Signed-off-by: Heiko Carstens <hca@linux.ibm.com>
>>>>>>> ---
>>>>>>>   arch/s390/kvm/kvm-s390.c | 18 +++++++++---------
>>>>>>>   1 file changed, 9 insertions(+), 9 deletions(-)
>>>>>>>
>>>>>>> diff --git a/arch/s390/kvm/kvm-s390.c b/arch/s390/kvm/kvm-s390.c
>>>>>>> index 1296fc10f80c..4b7b24f07790 100644
>>>>>>> --- a/arch/s390/kvm/kvm-s390.c
>>>>>>> +++ b/arch/s390/kvm/kvm-s390.c
>>>>>>> @@ -329,31 +329,31 @@ static void allow_cpu_feat(unsigned long nr)
>>>>>>>   static inline int plo_test_bit(unsigned char nr)
>>>>>>>   {
>>>>>>> -    register unsigned long r0 asm("0") = (unsigned long) nr | 0x100;
>>>>>>> +    unsigned long function = (unsigned long) nr | 0x100;
>>>>>>>       int cc;
>>>>>>>       asm volatile(
>>>>>>> +        "    lgr    0,%[function]\n"
>>>>>>>           /* Parameter registers are ignored for "test bit" */
>>>>>>>           "    plo    0,0,0,0(0)\n"
>>>>>>>           "    ipm    %0\n"
>>>>>>>           "    srl    %0,28\n"
>>>>>>>           : "=d" (cc)
>>>>>>> -        : "d" (r0)
>>>>>>> -        : "cc");
>>>>>>> +        : [function] "d" (function)
>>>>>>> +        : "cc", "0");
>>>>>>>       return cc == 0;
>>>>>>>   }
>>>>>>>   static __always_inline void __insn32_query(unsigned int opcode, u8 *query)
>>>>>>>   {
>>>>>>> -    register unsigned long r0 asm("0") = 0;    /* query function */
>>>>>>> -    register unsigned long r1 asm("1") = (unsigned long) query;
>>>>>>> -
>>>>>>>       asm volatile(
>>>>>>> -        /* Parameter regs are ignored */
>>>>>>> +        "    lghi    0,0\n"
>>>>>>> +        "    lgr    1,%[query]\n"
>>>>>>> +        /* Parameter registers are ignored */
>>>>>>>           "    .insn    rrf,%[opc] << 16,2,4,6,0\n"
>>>>>>>           :
>>>>>>> -        : "d" (r0), "a" (r1), [opc] "i" (opcode)
>>>>>>> -        : "cc", "memory");
>>>>>>> +        : [query] "d" ((unsigned long)query), [opc] "i" (opcode)
>          : "cc", "memory", "0", "1");
>>>>>>
>>>>>> Wouldn't it be better to keep the "a" constraint instead of "d" to avoid that the compiler ever passes the "query" value in r0 ?
>>>>>> Otherwise the query value might get trashed if it is passed in r0...
>>>>>
>>>>> I first thought the same, but if you look closely the value is only used by the lgr, to load
>>>>> the value finally into r1. So d is correct as lgr can take all registers.
>>>>
>>>> But what about the "lghi    0,0" right in front of it? ... I've got the feeling that I'm missing something here...
>>>
>>> It does load an immediate value of 0 into register 0. Are you afraid of an early clobber if
>>> gcc decides to use r0 for query?
>>
>> Right, that was my concern. It's a "static __always_inline" function, so can we be sure that query is still always located in a register that is reserved for parameters (i.e. >= r2) ?
> 
> Adding Andreas. I know that gcc provides the early clobber "&" if an output is written before all imputs are used.
> but in this case it is not an output but a clobbered register.
> The clobber list contains 0 and 1, so I guess gcc will never use those as input register.
> Andreas, is that correct?
> 
> 
> PS: Of course switching both instructions does not hurt

Answering myself:

https://gcc.gnu.org/onlinedocs/gcc/Extended-Asm.html#Clobbers-and-Scratch-Registers

When the compiler selects which registers to use to represent input and output operands, it does not use any of the clobbered registers. As a result, clobbered registers are available for any use in the assembler code.