Re: [Qemu-devel] [PATCH] Re-evaluate SVE vector length everytime ADDVL is executed

[Richard Henderson]

On 3/14/19 3:51 AM, Alex Bennée wrote:
> Amir CHARIF <address@hidden> writes:
> 
>> Hello,
>> Thanks for your answer.
>>
>> The wrong size was definitely being stored in the TB, and, it only affected 
>> ADDVL/RDVL/ADDPL (i.e. not all instructions are wrong). Here is what I think 
>> was happening:
>>
>> - The kernel disables SVE in EL0 (ZEN= 01).
>> - When the user space application is entered, the TB containing ADDVL has 
>> its length set to 0 (16 bytes), as we are in EL0 (so sve_exception_el!=0), 
>> and FP is enabled.
>> - ADDVL is executed (without trapping) on the basis of the current
>> length (16). (Nested function calls in the same context will cause a
>> ton of ADDVL instructions to be executed with a vecsize of 16.)
> So this looks like the error. Certainly the pseudo code says:
> 
>   CheckSVEEnabled();
>   bits(64) operand1 = if n == 31 then SP[] else X[n];
>   bits(64) result = operand1 + (imm * (VL DIV 8));
> 
>   if d == 31 then
>       SP[] = result;
>   else
>       X[d] = result;
> 
> so we should trap to the kernel and we won't without sve_access_check()
> 

Yes.  A browse through the file suggests ADDVL, ADDPL, and RDVL are missing the
check.  I'll write up a fix.


r~