Re: ppc64 floating point usage [was Re: PPC64 Compiler bug !!]

[linas at austin dot ibm dot com]

On Fri, Jun 13, 2003 at 04:09:06PM -0400, David Edelsohn wrote:
> >>>>> Jakub Jelinek writes:
> 
> Jakub> I guess best would be if GCC could dynamically switch the reg alloc orders
> Jakub> based on whether FPRs will be used anyway or whether current function is
> Jakub> pure integer code.

Why would this be best?

> 	Yes.  The difficulty is determining whether a function is pure
> integer code.  

I just finished reading the October archives on this issue.  One of
the notes hinted at a flag in some struct that indicated if a function
used float.  Not clear if the flag was hypothetical or real.

Barring further discussion of eh, setjump/longjump & variadic funcs,
it seems to me such a flag could be easily computed early on (although
it might slow down compilation a tad to do so).

> Even more, one does not want to simply change the alloc
> order but mark FPRs as fixed for pure integer code, which is what
> -msoft-float does.

Bird in hand vs. two in the bush.  We have Alan Modra's patch now, we
don't have an acceptable -mno-implicit-float patch.

There is concern that Alan's patch will negatively impact performance
of fp code.  Is there a way to unambiguously resolve this issue, or 
at least resolve it to everyones satisfaction?  

How would one do this? Visually inspect generated fp code? Run 
benchmarks? Which benchmarks? 

--linas