gfortran and -fbounds-check

[mimo]

Tobias Schlüter a écrit :
> mimo wrote:
>> Hi all,
>>  Is there some way to force gfortran not to evaluate the second 
>> operand of a .AND. or .OR. when the result is determined by the first 
>> argument ?
>>  I try to use -fbounds-check on a program which have some instructions 
>> of the type :
>>    REAL A(100)
>>    DO I=1,1000
>>      IF( I <= 100 .AND. A(I) > 0. ) do something
>>    ENDDO
>> and this would give me an "Array reference out of bounds" when I=101
> 
> Hi mimo,
> 
> can you file a bug in our bug tracking system.  We actually do 
> short-circuit the logical expression.  Unfortunately, the 
> bounds-checking happens before the evaluation of the logical expression, 
> i.e. the generated code looks like this (taken from the 
> -fdump-tree-original output):
>             if ((logical4) __builtin_expect (i <= 0, 0))
>               {
>                 _gfortran_runtime_error ("Array reference out of
>                                           bounds...");
>               }
>             if ((logical4) __builtin_expect (i > 100, 0))
>               {
>                 _gfortran_runtime_error ("Array reference out of
>                                           bounds...");
>               }
>             if (i <= 100 && a[NON_LVALUE_EXPR <i> + -1] > 0.0)
>               {
> In other words, the illegal access would never happen, but
> we throw the error nonetheless.
> 
> - Tobi

   Thanks tobi for this quick answer. I will file the bug.
                     mimo