Re: About GCC4 Optimization

[Linus Torvalds <torvalds@linux-foundation.org>]

[-- Warning: decoded text below may be mangled, UTF-8 assumed --]
[-- Attachment #1: Type: TEXT/PLAIN; charset=gb2312, Size: 1345 bytes --]



On Sun, 25 Mar 2007, yuan cooper wrote:
>  
> during my work, I found there is a bug with GCC4 O2 optimization.

Technically, it's a misfeature fo gcc4, not a bug.

The C language allows for type-based alias detection, and gcc notices that 
a "float *" cannot ever alias with a "unsigned long *", so it decides to 
not even do the loads and stores..

Now, there's two things wrong with this picture:

 - gcc is being an ass. type-based alias detection should happen only as a 
   last resort, and gcc should know and notice that *despite* the types 
   being different, they definitely alias.

   So what gcc does may be technically legal, but it's still a horribly 
   bad thing to do. Sadly, some gcc people seem to care more about "letter 
   of the law" than "sanity and quality of implementation".

 - as a result, you should always compile any kernel stuff with 
   "-fno-strict-aliasing", which should turn this off. If it *still* 
   happens with that flag, then it is indeed a compiler bug.

> float ftmp;
> unsigned long tmp;
> ftmp = 1.0/1024.0;
> tmp  = *(unsigned long *)(&ftmp);
> tmp  = (tmp >> 11) && 0xFFF;
>  
> if optimization level is O2, gcc will MOV eax to tmp, but current eax has a random value.
> -O is ok and gcc3 with O2 is ok too.

That said, you really _really_ shouldn't be doing FP in the kernel anyway.

		Linus