Re: [Help-smalltalk] About no_opt macro

[Paolo Bonzini]

On 10/12/2010 03:28 PM, Mathieu Suen wrote:
> Will reading the vm.def, I cam across the no_opt macro.
> I guess is to avoid some optimization but but what kind of optimization and why.
>
> The macro is define as:
>
> #define no_opt(x)    ({ __typeof__ ((x)) _result; \
>               asm ("" : "=r" (_result) : "0" ((x))); _result; })
>
>
>  From what I understand is that it define a variable _result and force to move x
> into _result.
>
> The macro is used in tha bytecode PLUS_SPECIAL an MINUS SPECIAL ... :
>
>   intptr_t iresult = no_opt (iop1 + iop2);
>
> Thanks for shading the light on this point.

Here is the context:

      intptr_t iop1 = (intptr_t) op1;
      intptr_t iop2 = ((intptr_t) op2) - 1;
      intptr_t iresult = no_opt (iop1 + iop2);
      if (iresult < iop1)

Remember that these snippets are copied in many places.  For example, 
when compiling code for a superoperator (compound bytecode) like

     PUSH_TEMPORARY(*)
     PUSH_INTEGER(1)
     PLUS_SPECIAL()

GCC would have

      intptr_t iop1 = (intptr_t) op1;
      intptr_t iresult = no_opt (iop1 + 2);
      if (iresult < iop1)

Without no_opt, "iop1 + 2 < iop1" would be simplified to "false", thus 
causing the overflow check to misbehave.

Paolo