PATCH: Remove SLOW_BYTE_ACCESS from get_best_mode

[H. J. Lu]

On Thu, Sep 14, 2006 at 07:04:02AM -0700, H. J. Lu wrote:
> On Wed, Sep 13, 2006 at 11:26:36PM -0700, H. J. Lu wrote:
> > On Wed, Sep 13, 2006 at 06:03:28PM -0700, H. J. Lu wrote:
> > > On Thu, Sep 14, 2006 at 12:20:31AM +0100, Paul Brook wrote:
> > > > > Here is the updated patch. I ran a micro benchmark with code difference
> > > > > similar to twolf:
> > > > >
> > > > > -       movl    (%rdi), %eax
> > > > > -       xorb    %ah, %ah
> > > > > -       subl    $1, %eax
> 
> This is a partial register stall. There is a 60% penalty on Core 2.
> I updated my patch only not to do it for -Os.
> 
> > > > > +       movq    (%rdi), %rax
> > > > > +       andl    $4294902015, %eax
> > > > > +       subq    $1, %rax
> > > > >         jne     .L8
> > > > >
> > > > > There is no speed difference on Nocona. I checked my SPEC results.
> > > > > twolf isn't very stable on my Nocona.
> > > > 
> > > > The former is significantly smaller though (7 bytes vs. 12 bytes).
> > > > 
> > > 
> > 
> > In general, widen mode of access to a full word is good. But x86-64 is
> > a special case due to bigger code size of DImode. This updated patch
> > won't widen SImode.
> > 
> > BTW, I didn't update SLOW_BYTE_ACCESS document since I am not clear
> > what it is supposed to do.
> > 
> > 
> 
> Here is an updated patch.
> 

Here is a case where SImode/DImode is 1-10% slower than QImode:

union bitfield {
   struct {
     unsigned int field0 : 1;
     unsigned int field1 : 2;
     unsigned int field2 : 2;
     unsigned int field3 : 1;
     long foo;
   } bitfields, bits;
   unsigned int   u32All;
   signed int     i32All;
   float          f32All;
};

int
fooo (program* prog, unsigned int i,
      unsigned int x, unsigned int y, unsigned int z)
{
  prog->array[i].bitfields.field1 = ((x > y) ? ((x > z) ? 0 : 2) : ((y
> z) ? 1
: 2));

  switch (prog->array[i].bitfields.field1)
    {
    case 0:
      return 3;
    case 1:
      return 5;
    case 2:
      return 9;
    default:
      break;
    }
  return 10;
}

In this case, we are examing the bitfield. Basically, SImode/DImode
will require less memory access if more than one fields are accessed.
But QImode may be faster if we need to check the field value.



H.J.