Re: Bit shift by more than word length

[Lionel B]

On Thu, 14 Jun 2007 15:41:14 +0200, Bernd Strieder wrote:

> Hello,
> 
> Lionel B wrote:
> 
>> Hi,
>> 
>> If x is of (unsigned) integer type of size N then the C++ standard says
>> that x << n and x >> n are undefined for n >= N. Is there by any chance
>> a GNU extension that guarantees that at least x << N and x >> N are
>> defined (and = 0)?
> 
> Most compilers will create code using the assembler codes for shifting
> available on the hardware, which usually do something like x << (n%N),
> i.e. shifting by a modulus, which is obtained in binary by truncating
> the shift value. Anything else would already require testing and
> possibly branching by the compiler. I have never programmed on a CPU
> where this is different.
> 
> Maybe the branch can be optimized away by means of conditional
> statements. You might have to experiment with the architecture and CPU
> model options of gcc to get there.
> 
> 
> #include <stdio.h>
> 
> int main(int argc)
> {
>     int i, j;
>     for(i = 0; i < 64; ++i) {
>         j = 1<<i;
>         if (i & ~31) j = 0;
>         printf("%d %d\n",i, j);
>     }
> }
> 
> Indeed, when looking at the result of this using gcc-4.1.2 there is no
> branch generated for the if to maintain the zero, at least with
> -march=i686 or -march=opteron, there is a branch with -march=pentium3.
> So double check gcc behaves the same way on your platform, otherwise you
> are hunting a phantom problem, and measure, whether on that platform a
> branch is preferable.

Thanks, I'm checking that out. It seems it may depend (amongst many other 
things, no doubt) on the form of test. On my gcc-4.1.1 on athlon 64 (no -
march flags, but -O3) I get no branch if I test as in your code, but I do 
get a branch if I test for i == 32. My assembler is pretty non-existent, 
though, so take that with a pinch of salt...

-- 
Lionel B