Re: arch/powerpc/math-emu/mtfsf.c - incorrect mask?

[Gabriel Paubert <paubert@iram.es>]

On Mon, Feb 10, 2014 at 12:32:18PM +0000, David Laight wrote:
> > I disagree, perhaps mostly because the compiler is not clever enough, but right
> > now the code for solution 1 is (actually I have rewritten the code
> > and it reads:
> > 
> > 	mask = (FM & 1)
> > 			| ((FM << 3) & 0x10)
> > 			| ((FM << 6) & 0x100)
> > 			| ((FM << 9) & 0x1000)
> > 			| ((FM << 12) & 0x10000)
> > 			| ((FM << 15) & 0x100000)
> > 			| ((FM << 18) & 0x1000000)
> > 			| ((FM << 21) & 0x10000000);
> > to avoid sequence point in case it hampers the compiler)
> > 
> > and the output is:
> > 
> >         rlwinm 10,3,3,27,27      # D.11621, FM,,
> >         rlwinm 9,3,6,23,23       # D.11621, FM,,
> >         or 9,10,9        #, D.11621, D.11621, D.11621
> >         rlwinm 10,3,0,31,31      # D.11621, FM,
> >         or 9,9,10        #, D.11621, D.11621, D.11621
> >         rlwinm 10,3,9,19,19      # D.11621, FM,,
> >         or 9,9,10        #, D.11621, D.11621, D.11621
> >         rlwinm 10,3,12,15,15     # D.11621, FM,,
> >         or 9,9,10        #, D.11621, D.11621, D.11621
> >         rlwinm 10,3,15,11,11     # D.11621, FM,,
> >         or 9,9,10        #, D.11621, D.11621, D.11621
> >         rlwinm 10,3,18,7,7       # D.11621, FM,,
> >         or 9,9,10        #, D.11621, D.11621, D.11621
> >         rlwinm 3,3,21,3,3        # D.11621, FM,,
> >         or 9,9,3         #, mask, D.11621, D.11621
> >         mulli 9,9,15     # mask, mask,
> > 
> > see that r9 is used 7 times as both input and output operand, plus
> > once for rlwinm. This gives a dependency length of 8 at least.
> > 
> > In the other case (I've deleted the code) the dependency length
> > was significantly shorter. In any case that one is fewer instructions,
> > which is good for occasional use.
> 
> Hmmm... I hand-counted a dependency length of 8 for the other version.
> Maybe there are some ppc instructions that reduce it.

Either I misread the generated code or I got somewhat less.

What helps for method1 is the rotate and mask instructions of PPC. Each of
left shift and mask becomes a single rlwinm. 
> 
> Stupid compiler :-)

Indeed. I've trying to coerce it into generating rlwimi instructions
(in which case the whole building of the mask reduces to 8 assembly
instructions) and failed. It seems that the compiler lacks some patterns
some patterns that would directly map to rlwimi.

> Trouble is, I bet that even if you code it as:
>  	mask1 = (FM & 1) | ((FM << 3) & 0x10);
> 	mask2 = ((FM << 6) & 0x100) | ((FM << 9) & 0x1000);
> 	mask3 = ((FM << 12) & 0x10000) | ((FM << 15) & 0x100000);
> 	mask4 = ((FM << 18) & 0x1000000) | ((FM << 21) & 0x10000000);
> 	mask1 |= mask2;
> 	mask3 |= mask4;
> 	mask = mask1 | mask3;
> the compiler will 'optimise' it to the above before code generation.

Indeed it's what it does :-(

I believe that the current suggestion is good enough.

	Gabriel