Re: Endianness macros capitalization

[Christian Franke]

Pavel Roskin wrote:
> On Wed, 2008-07-09 at 14:50 +0200, Christian Franke wrote:
>   
>
> > Result with the test script from my last mail:
> >
> > Debian gcc 4.1.2-7:
> > inline (portable)=357, inline (asm)=126, function=104
> >
> > Cygwin gcc 3.4.4:
> > inline (portable)=340, inline (asm)=124, function=96
> >
> > Function call is still better. The only candidate for inline is probably
> >  grub_swap_bytes16().
> >     
>
> But we are getting closer!
>
>   

But probably not close enough in the average case, see below :-)


> > > .... And if written properly, it could work with any of
> > > the registers that allow access to the lower two bytes (%eax, %ebx,
> > > %ecx and %edx), thus giving more flexibility to the optimizer.
> > >
> > >       
> > This would require support to access the Rl and Rh parts of eRx for each
> > R in [a-d]. Something like:
> >
> >    asm (
> >         "xchg %0:l,%0:h\n"
> >         "roll $0x10,%0\n"
> >         "xchg %0:l,%0:h\n"
> >         : "=r"(_y) : "0"(_x) \
> >     );
> >
> > Do gcc or gas provide a syntax to do this?
> >     
>
> Yes.  That's %b0 and %h0.  Use "=q" for all registers with "upper
> halves" (%ah-%dh).
>
>   

Thanks for the info.

I tried this in the same script:

#define grub_swap_bytes32(x) \
({ \
  grub_uint32_t _x = (x), _y; \
  asm ( \
       "xchgb %b0,%h0\n" \
       "roll $0x10,%0\n" \
       "xchgb %b0,%h0\n" \
       : "=q"(_y) : "0"(_x) \
   ); \
   _y; \
})


GCC optimizer does a good job optimizing register use, but function call 
is still better:

Debian gcc 4.1.2-7:
inline (portable)=357,asm (%%eax)=126, asm (%0)=107, function=104

Cygwin gcc 3.4.4:
inline (portable)=340, asm (%%eax)=124, asm (%0)=104, function=96


Inline asm is only better is rare cases, e.g. with this test function:

type test(type *x)
{
 return func(x[0]) + func(x[1]) + ... + func(x[N]);
}

Result with Cygwin gcc 3.4.4:

N=0: asm=14, function=11
N=1: asm=28, function=32
N=2: asm=40, function=41
N=3: asm=52, function=51
N=4: asm=64, function=61
N=5: asm=76, function=72


To test which files would possibly be affected by any size optimization 
of grub_swap_bytes*(), I 've done a test compilation with these macros 
replaced by dummies '(x)'. Only the size of following modules changed:

affs.mod
afs.mod
amiga.mod
apple.mod
ata.mod
hfs.mod
hfsplus.mod
iso9660.mod
jpeg.mod
png.mod
sfs.mod
sun.mod
xfs.mod

Remaining modules and kernel.img are identical.

Christian