[patch tree-optimization]: Improve handling of conditional-branches on targets with high branch costs

[Kai Tietz]

2011/10/26 Michael Matz <matz@suse.de>:
> Hi,
>
> On Wed, 26 Oct 2011, Kai Tietz wrote:
>
>> >> > int f(char *i, int j)
>> >> > {
>> >> >        if (*i && j!=2)
>> >> >                return *i;
>> >> >        else
>> >> >                return j;
>> >> > }
>> >>
>>
>> the case can be produced quite easily.
>>
>> extern int global = 0;
>>
>> ....
>>   if (*a && global) ...
>
> See?  You had to change the program to prove the transformation to be
> invalid.  But my point was that the function we discuss about was exactly
> as above.  It didn't have globals, or two loads, or a volatile, or
> anything else you can come up with where the transformation would be
> detectable (and hence invalid).  I claim that you can't construct a
> program that can distinguish between this function:
>
> int f(char *i, int j)
> {
>   if (*i && j!=2)
>     return *i;
>   else
>     return j;
> }
>
> and this one:
>
> int f(char *i, int j)
> {
>   if (*i & j!=2)
>     return *i;
>   else
>     return j;
> }
>
> And if you can't construct such a program, then the initial transformation
> before the fold-const.c change _for this specific situation_ was correct.
>
>
> Ciao,
> Michael.

well, if such a function is used as inline and we know for it that j
has value != 2, then we have here a big difference.  For your first
example, we still have to do the memory access to *i, even if we are
not interested in result.  See here point 4 of 5.1.2.3 of C-spec.
For your second sample we don't need to do that, as the & itself is no
sequence-point and so we can eliminate the *i access without breaking
anything.

Regards,
Kai