11706 vs ([lno] Canonical iv creation)

[Richard Guenther]

Gabriel Dos Reis wrote:
> Paolo Carlini <pcarlini@suse.de> writes:
> 
> | Hi again,
> | 
> | finally, I have been able to actually test the the new pass on the code
> | we care about. The current assembler, is simply *astonishing*.
> 
> Yep.
> 
> Fixing the problem at the _right place_ helps emitting good
> executable in much broader situations, instead of damaging random
> places in the library.

Apart from "damaging" I agree with this.

> I bet there are other places where this improvement will be far
> noticeable.  I conjecture that very soon, we will not be afraid of  
> using right abstractions through the library instead of trying to
> hide defects in the compiler.

I'm curious to how the standard actually defines std::pow()s accuracy 
here though.  I know we're not optimizing ::pow(x, 4) to the above 
without -ffast-math because of accuracy concerns.  So how does 
std::pow() relate to ::pow in terms of accuracy guarrantees?

After all, if the required parts of lno branch are going into 3.5 I'm 
happy there.  But what about 3.3 and particular 3.4?

Richard.