________________________________________
From: Mike Day 
Sent: Monday, March 26, 2007 9:06 AM
To: Mike Acton; tech
Subject: RE: curiously small code...

The vector version of swizzle_lo8 comes out a bit shorter using selection masks…


qword swizzle_lo8( qword x, qword y )
{
  // Create some masks.
  ilh   m20 = ilh( 0xFF00 );
  ilh   m30 = ilh( 0xF0F0 );
  ilh   m40 = ilh( 0xCCCC );
  ilh   m50 = ilh( 0xAAAA );

  // Expand(x) -- Shift left each bit by bit index. Zero unused bits.
  qword x20 = si_shlhi( x, 4 );
  qword x21 = si_selb( x, x20, m20 );
  qword x30 = si_shlhi( x21, 2 );
  qword x31 = si_selb( x21, x30, m30 );
  qword x40 = si_shlhi( x31, 1 );
  qword x41 = si_selb( x31, x40, m40 );

  // Expand(y) -- Shift left each bit by bit index. Zero unused bits.
  qword y20 = si_shlhi( y, 4 );
  qword y21 = si_selb( y, y20, m20 );
  qword y30 = si_shlhi( y21, 2 );
  qword y31 = si_selb( y21, y30, m30 );
  qword y40 = si_shlhi( y31, 1 );
  qword y41 = si_selb( y31, y40, m40 );

  // Combine.
  qword z00 = si_shlhi( y41, 1 );
  qword z01 = si_selb( x41, z00, m50 );

  return z01;
}


I’ll leave the 16-bit one for the interested reader!

Mike