Re: [PATCH] powerpc: provide __bswapdi2

[Gabriel Paubert <paubert@iram.es>]

On Mon, May 13, 2013 at 11:38:13AM +0100, David Woodhouse wrote:
> On Mon, 2013-05-13 at 11:33 +0100, David Woodhouse wrote:
> > 
> > On Mon, 2013-05-13 at 09:33 +0200, Gabriel Paubert wrote:
> > > Actually, I'd swap the two mr instructions to never
> > > have an instruction that uses the result from the
> > > previous one. 
> > 
> > Bad GCC. No biscuit.
> > 
> > Should we file a PR? 
> 
> Maybe not. If you tell it to tune for an in-order machine like Cell, it
> swaps them round. Although now I'm confused about which of POWER[567]
> were in-order:

It was Power6 IIRC. On this kind of fine point, don't rely too much
on what GCC produces.

> 
> [dwmw2@i7 ~]$ powerpc64-linux-gnu-gcc -O2 -S -o- bswapdi2.c -m32  | grep -B1 mr
> 	rlwimi 11,3,24,16,23
> 	mr 4,11
> 	mr 3,10
> [dwmw2@i7 ~]$ powerpc64-linux-gnu-gcc -O2 -S -o- bswapdi2.c -m32 -mtune=cell | grep -B1 mr
> 	rlwimi 11,3,24,16,23
> 	mr 3,10
> 	mr 4,11
> [dwmw2@i7 ~]$ powerpc64-linux-gnu-gcc -O2 -S -o- bswapdi2.c -m32 -mtune=power5 | grep -B1 mr
> 	rlwimi 11,3,24,16,23
> 	mr 3,10
> 	mr 4,11
> [dwmw2@i7 ~]$ powerpc64-linux-gnu-gcc -O2 -S -o- bswapdi2.c -m32 -mtune=power6 | grep -B1 mr
> 	rlwimi 11,3,24,16,23
> 	mr 4,11
> 	mr 3,10
> [dwmw2@i7 ~]$ powerpc64-linux-gnu-gcc -O2 -S -o- bswapdi2.c -m32 -mtune=power7 | grep -B1 mr
> 	rlwimi 11,3,24,16,23
> 	mr 4,11
> 	mr 3,10

I don't know of any processor in which putting the mr 3,10 first can cause stalls, so
even a generic tuning should put it first.

	Gabriel