[Bug middle-end/19912] GCC does not disclose parentheses

["joseph at codesourcery dot com" <gcc-bugzilla at gcc dot gnu dot org>]

------- Additional Comments From joseph at codesourcery dot com  2005-02-11 23:19 -------
Subject: Re:  GCC does not disclose parentheses

On Fri, 11 Feb 2005, l_belev at yahoo dot com wrote:

> I dont understand what the purpose of -fwrapv is, because the compiler
> always knows whether the machine it's currently generating code for,
> wraps around the results on arithmetic overflow.
> Anyway surely the first consideration applies in this case too,
> that is, the transformations could be disabled if they are not applicable.

The internal tree structures (GIMPLE) have well-defined mostly 
machine-independent semantics, which on signed arithmetic overflow follow 
C, i.e. signed overflow is undefined behavior unless -fwrapv.  This means, 
for example, that ((x * 2)/2) can be optimized to x for signed arithmetic, 
although it can't for unsigned.  Because of the potential for such 
transformations to be applied to the results of other transformations, it 
isn't safe to apply any transformation introducing undefined behavior lest 
some other such transformation use the fact it is undefined in a way that 
changes the semantics where the original code did not involve undefined 
behavior.



-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=19912