[fpc-devel] Optimisation and thread safety

[J. Gareth Moreton]

I figured that if you read from memory, and then write to memory, it's 
not safe to read from the same memory block again because the write 
address could be equal to (or overlap) the read memory and it's 
impossible to rule that out if different registers are used. It might be 
okay if the references use the same registers (and remain constant) and 
only the offsets differ (and, again, there's no overlap), but otherwise 
it would only be safe when using %rsp, since I'm pretty sure the 
compiler does not use that for anything other than the stack pointer.

If it's the programmer's job to handle thread synchronisation, that 
alleviates some pressure.

Thanks for the insight, Florian and Michael.

Gareth aka. Kit

On 13/11/2021 10:54, Florian Klämpfl via fpc-devel wrote:
>
>> Am 13.11.2021 um 00:55 schrieb J. Gareth Moreton via fpc-devel <fpc-devel at lists.freepascal.org>:
>>
>> Hi everyone,
>>
>> I have a question when it comes to optimising memory reads and writes.  What are the rules for FPC when it comes to writing to memory and then reading from it later within a single subroutine? For example, say I had this pair of commands:
>>
>>      movq    %rdx,-584(%rbp)
>>      movl    -584(%rbp),%eax
>>
>> That could easily be converted to "movl %edx,%eax", especially as %rbp is likely pointing to the top of the stack.
> This is not possible as rsp could be used as normal register.
>
> If the reference is not marked as being read volatile, you can change the second mov into movl %edx,%ea.
>
> For further optimizations you have to track tai_tempalloc which we don’t yet.
>
>> But if the reference uses different registers, would it still be safe to make this optimisation given that the scheduler could suspend the thread in between the two instructions and then another thread writes to the same memory block before control is returned?
>>
>> I am aware of other examples that require caution.  For example:
>>
>>      movslq    -608(%rbp),%rdx
>>      subl    %eax,-84(%rbp,%rdx,4)
>>      movslq    -608(%rbp),%rdx
>>
>> Here it might be tempting to remove the second "movslq" instruction, but the value of %rdx could happen to be equal to -131, which would allow the subl instruction to modify -608(%rbp), and in this situation, it's quite likely if a malicious input is given to the program to manipulate the value stored at -608(%rbp) and invoke a buffer overrun.
> It could be also a pointer aliasing the memory location or an array, so this is not safe in any case. So something like a[a[1234]] which might be valid code imo.
>
>> So in conclusion, theoretically, where is it perfectly safe to assume the value in memory hasn't changed, where would it be permissible only under -O4 and where must it not be optimised at all?
>>
>> Gareth aka. Kit
>>
>> P.S. Assembly examples were taken from the System unit.
>>
>>
>> -- 
>> This email has been checked for viruses by Avast antivirus software.
>> https://www.avast.com/antivirus
>>
>> _______________________________________________
>> fpc-devel maillist  -  fpc-devel at lists.freepascal.org
>> https://lists.freepascal.org/cgi-bin/mailman/listinfo/fpc-devel
> _______________________________________________
> fpc-devel maillist  -  fpc-devel at lists.freepascal.org
> https://lists.freepascal.org/cgi-bin/mailman/listinfo/fpc-devel
>