Re: Missing clobber on alternative use on Linux UM 32-bit

[Nadav Amit]

> On Nov 4, 2023, at 11:34 AM, Anton Ivanov <anton.ivanov@xxxxxxxxxxxxxxxxx> wrote:
> 
> On 04/11/2023 09:25, Nadav Amit wrote:
>> 
>> I was reading (again) the x86 C macro of “alternative()” and I was a bit
>> surprised it does clobber the flags (“cc”) as a precaution.
>> 
>>  #define alternative(oldinstr, newinstr, ft_flags) \
>> 	asm_inline volatile (ALTERNATIVE(oldinstr, newinstr, ft_flags) : : : "memory")
>> 
>> Actually there seems to be only one instance of problematic cases - in um/32-bit:
>> 
>>  #define mb() alternative("lock; addl $0,0(%%esp)", "mfence", X86_FEATURE_XMM2)
>>  #define rmb() alternative("lock; addl $0,0(%%esp)", "lfence", X86_FEATURE_XMM2)
>>  #define wmb() alternative("lock; addl $0,0(%%esp)", "sfence", X86_FEATURE_XMM)
>> 
>> Presumably, if XMM or XMM2 are not supported, there would be instances where addl
>> would be able to change eflags arithmetic flags without the compiler being aware
>> of it.
>> 
>> As it only affects 32-bit Linux UM - I don’t easily have an environment to test
>> the fix. An alternative (word-pun unintended) is to add “cc” as a precaution
>> to the alternative macro.
>> 
> Application alternatives in um is presently a NOP. It always uses the "blunt and heavy instrument" - the most conservative option.
> 
> It is on the TODO list.

Thanks for the quick response. But I don’t see how it prevents the problem
(it actually makes it worse - affecting XMM/XMM2 CPUs as well) since you
keep the “lock; addl $0,0(%%esp)” in the running code, affecting eflags
without telling the compiler that you do so through a “cc” clobber.