Re: missed tail-optimization with templates

[Andrew Pinski <pinskia at physics dot uc dot edu>]

On Dec 29, 2003, at 15:48, Luca Padovani wrote:

> Hi,
>
> I'm using g++ (GCC) 3.3.3 20031206 (prerelease) (Debian). I don't fully
> understand the behavior of the compiler -O2 with the C++ program (the
> smallest I could produce showing the "problem") attached to this email.
>
> The zero() function gets compiled and tail-optimized, fine. However, 
> the
> A::zero() and B::zero() static member functions, although they amount 
> in
> combination at the same code of zero(), are _not_ tail optimized. At
> first I thought templates prevent the compiler from tail-optimizing, 
> but
> if I make A::zero() more trivial (turn the #if 1 into an #if 0) the 
> tail
> optimization shows up.
>
> Is there anything that prevents tail optimization with #if 1? If not, 
> is
> this to be considered a possible improvement for gcc or it is just too
> specific/hard to implement?

The problem is GCC does not remove dead code before trying to do the 
tail
optimization.  On the tree-ssa branch these functions are optimized 
right because it
works on the tree level instead of the RTL level.

Thanks,
Andrew Pinski