d_polysa.s

Quake 2 Source code for students by Theerthan You can also download from idsoftwares.com

//
// d_polysa.s
// x86 assembly-language polygon model drawing code
//

#include "qasm.h"
#include "d_ifacea.h"

#if	id386

// !!! if this is changed, it must be changed in d_polyse.c too !!!
#define DPS_MAXSPANS			MAXHEIGHT+1	
									// 1 extra for spanpackage that marks end

//#define	SPAN_SIZE	(((DPS_MAXSPANS + 1 + ((CACHE_SIZE - 1) / spanpackage_t_size)) + 1) * spanpackage_t_size)
#define SPAN_SIZE (1024+1+1+1)*32



	.data

	.align	4
p10_minus_p20:	.single		0
p01_minus_p21:	.single		0
temp0:			.single		0
temp1:			.single		0
Ltemp:			.single		0

aff8entryvec_table:	.long	LDraw8, LDraw7, LDraw6, LDraw5
				.long	LDraw4, LDraw3, LDraw2, LDraw1

lzistepx:		.long	0


	.text

#ifndef NeXT
	.extern C(D_PolysetSetEdgeTable)
	.extern C(D_RasterizeAliasPolySmooth)
#endif

//----------------------------------------------------------------------
// affine triangle gradient calculation code
//----------------------------------------------------------------------

#if 0
#define skinwidth	4+0

.globl C(R_PolysetCalcGradients)
C(R_PolysetCalcGradients):

//	p00_minus_p20 = r_p0[0] - r_p2[0];
//	p01_minus_p21 = r_p0[1] - r_p2[1];
//	p10_minus_p20 = r_p1[0] - r_p2[0];
//	p11_minus_p21 = r_p1[1] - r_p2[1];
//
//	xstepdenominv = 1.0 / (p10_minus_p20 * p01_minus_p21 -
//			     p00_minus_p20 * p11_minus_p21);
//
//	ystepdenominv = -xstepdenominv;

	fildl	C(r_p0)+0		// r_p0[0]
	fildl	C(r_p2)+0		// r_p2[0] | r_p0[0]
	fildl	C(r_p0)+4		// r_p0[1] | r_p2[0] | r_p0[0]
	fildl	C(r_p2)+4		// r_p2[1] | r_p0[1] | r_p2[0] | r_p0[0]
	fildl	C(r_p1)+0		// r_p1[0] | r_p2[1] | r_p0[1] | r_p2[0] | r_p0[0]
	fildl	C(r_p1)+4		// r_p1[1] | r_p1[0] | r_p2[1] | r_p0[1] |
							//  r_p2[0] | r_p0[0]
	fxch	%st(3)			// r_p0[1] | r_p1[0] | r_p2[1] | r_p1[1] |
							//  r_p2[0] | r_p0[0]
	fsub	%st(2),%st(0)	// p01_minus_p21 | r_p1[0] | r_p2[1] | r_p1[1] |
							//  r_p2[0] | r_p0[0]
	fxch	%st(1)			// r_p1[0] | p01_minus_p21 | r_p2[1] | r_p1[1] |
							//  r_p2[0] | r_p0[0]
	fsub	%st(4),%st(0)	// p10_minus_p20 | p01_minus_p21 | r_p2[1] |
							//  r_p1[1] | r_p2[0] | r_p0[0]
	fxch	%st(5)			// r_p0[0] | p01_minus_p21 | r_p2[1] |
							//  r_p1[1] | r_p2[0] | p10_minus_p20
	fsubp	%st(0),%st(4)	// p01_minus_p21 | r_p2[1] | r_p1[1] |
							//  p00_minus_p20 | p10_minus_p20
	fxch	%st(2)			// r_p1[1] | r_p2[1] | p01_minus_p21 |
							//  p00_minus_p20 | p10_minus_p20
	fsubp	%st(0),%st(1)	// p11_minus_p21 | p01_minus_p21 |
							//  p00_minus_p20 | p10_minus_p20
	fxch	%st(1)			// p01_minus_p21 | p11_minus_p21 |
							//  p00_minus_p20 | p10_minus_p20
	flds	C(d_xdenom)		// d_xdenom | p01_minus_p21 | p11_minus_p21 |
							//  p00_minus_p20 | p10_minus_p20
	fxch	%st(4)			// p10_minus_p20 | p01_minus_p21 | p11_minus_p21 |
							//  p00_minus_p20 | d_xdenom
	fstps	p10_minus_p20	// p01_minus_p21 | p11_minus_p21 |
							//  p00_minus_p20 | d_xdenom
	fstps	p01_minus_p21	// p11_minus_p21 | p00_minus_p20 | xstepdenominv
	fxch	%st(2)			// xstepdenominv | p00_minus_p20 | p11_minus_p21

//// ceil () for light so positive steps are exaggerated, negative steps
//// diminished,  pushing us away from underflow toward overflow. Underflow is
//// very visible, overflow is very unlikely, because of ambient lighting
//	t0 = r_p0[4] - r_p2[4];
//	t1 = r_p1[4] - r_p2[4];

	fildl	C(r_p2)+16		// r_p2[4] | xstepdenominv | p00_minus_p20 |
							//  p11_minus_p21
	fildl	C(r_p0)+16		// r_p0[4] | r_p2[4] | xstepdenominv |
							//  p00_minus_p20 | p11_minus_p21
	fildl	C(r_p1)+16		// r_p1[4] | r_p0[4] | r_p2[4] | xstepdenominv |
							//  p00_minus_p20 | p11_minus_p21
	fxch	%st(2)			// r_p2[4] | r_p0[4] | r_p1[4] | xstepdenominv |
							//  p00_minus_p20 | p11_minus_p21
	fld		%st(0)			// r_p2[4] | r_p2[4] | r_p0[4] | r_p1[4] |
							//  xstepdenominv | p00_minus_p20 | p11_minus_p21
	fsubrp	%st(0),%st(2)	// r_p2[4] | t0 | r_p1[4] | xstepdenominv |
							//  p00_minus_p20 | p11_minus_p21
	fsubrp	%st(0),%st(2)	// t0 | t1 | xstepdenominv | p00_minus_p20 |
							//  p11_minus_p21

//	r_lstepx = (int)
//			ceil((t1 * p01_minus_p21 - t0 * p11_minus_p21) * xstepdenominv);
//	r_lstepy = (int)
//			ceil((t1 * p00_minus_p20 - t0 * p10_minus_p20) * ystepdenominv);

	fld		%st(0)			// t0 | t0 | t1 | xstepdenominv | p00_minus_p20 |
							//  p11_minus_p21
	fmul	%st(5),%st(0)	// t0*p11_minus_p21 | t0 | t1 | xstepdenominv |
							//  p00_minus_p20 | p11_minus_p21
	fxch	%st(2)			// t1 | t0 | t0*p11_minus_p21 | xstepdenominv |
							//  p00_minus_p20 | p11_minus_p21
	fld		%st(0)			// t1 | t1 | t0 | t0*p11_minus_p21 |
							//  xstepdenominv | p00_minus_p20 | p11_minus_p21
	fmuls	p01_minus_p21	// t1*p01_minus_p21 | t1 | t0 | t0*p11_minus_p21 |
							//  xstepdenominv | p00_minus_p20 | p11_minus_p21
	fxch	%st(2)			// t0 | t1 | t1*p01_minus_p21 | t0*p11_minus_p21 |
							//  xstepdenominv | p00_minus_p20 | p11_minus_p21
	fmuls	p10_minus_p20	// t0*p10_minus_p20 | t1 | t1*p01_minus_p21 |
							//  t0*p11_minus_p21 | xstepdenominv |
							//  p00_minus_p20 | p11_minus_p21
	fxch	%st(1)			// t1 | t0*p10_minus_p20 | t1*p01_minus_p21 |
							//  t0*p11_minus_p21 | xstepdenominv |
							//  p00_minus_p20 | p11_minus_p21
	fmul	%st(5),%st(0)	// t1*p00_minus_p20 | t0*p10_minus_p20 |
							//  t1*p01_minus_p21 | t0*p11_minus_p21 |
							//  xstepdenominv | p00_minus_p20 | p11_minus_p21
	fxch	%st(2)			// t1*p01_minus_p21 | t0*p10_minus_p20 |
							//  t1*p00_minus_p20 | t0*p11_minus_p21 |
							//  xstepdenominv | p00_minus_p20 | p11_minus_p21
	fsubp	%st(0),%st(3)	// t0*p10_minus_p20 | t1*p00_minus_p20 |
							//  t1*p01_minus_p21 - t0*p11_minus_p21 |
							//  xstepdenominv | p00_minus_p20 | p11_minus_p21
	fsubrp	%st(0),%st(1)	// t1*p00_minus_p20 - t0*p10_minus_p20 |
							//  t1*p01_minus_p21 - t0*p11_minus_p21 |
							//  xstepdenominv | p00_minus_p20 | p11_minus_p21
	fld		%st(2)			// xstepdenominv |
							//  t1*p00_minus_p20 - t0*p10_minus_p20 |
							//  t1*p01_minus_p21 - t0*p11_minus_p21 |
							//  xstepdenominv | p00_minus_p20 | p11_minus_p21
	fmuls	float_minus_1	// ystepdenominv |
							//  t1*p00_minus_p20 - t0*p10_minus_p20 |
							//  t1*p01_minus_p21 - t0*p11_minus_p21 |
							//  xstepdenominv | p00_minus_p20 | p11_minus_p21
	fxch	%st(2)			// t1*p01_minus_p21 - t0*p11_minus_p21 |
							//  t1*p00_minus_p20 - t0*p10_minus_p20 |
							//  ystepdenominv | xstepdenominv | p00_minus_p20 |
							//  p11_minus_p21
	fmul	%st(3),%st(0)	// (t1*p01_minus_p21 - t0*p11_minus_p21)*
							//   xstepdenominv |
							//  t1*p00_minus_p20 - t0*p10_minus_p20 |
							//   | ystepdenominv | xstepdenominv |
							//   p00_minus_p20 | p11_minus_p21
	fxch	%st(1)			// t1*p00_minus_p20 - t0*p10_minus_p20 |
							//  (t1*p01_minus_p21 - t0*p11_minus_p21)*
							//   xstepdenominv | ystepdenominv |
							//   xstepdenominv | p00_minus_p20 | p11_minus_p21
	fmul	%st(2),%st(0)	// (t1*p00_minus_p20 - t0*p10_minus_p20)*
							//  ystepdenominv |
							//  (t1*p01_minus_p21 - t0*p11_minus_p21)*
							//  xstepdenominv | ystepdenominv |
							//  xstepdenominv | p00_minus_p20 | p11_minus_p21
	fldcw	ceil_cw
	fistpl	C(r_lstepy)		// r_lstepx | ystepdenominv | xstepdenominv |
							//  p00_minus_p20 | p11_minus_p21
	fistpl	C(r_lstepx)		// ystepdenominv | xstepdenominv | p00_minus_p20 |
							//  p11_minus_p21
	fldcw	single_cw

//	t0 = r_p0[2] - r_p2[2];
//	t1 = r_p1[2] - r_p2[2];

	fildl	C(r_p2)+8		// r_p2[2] | ystepdenominv | xstepdenominv |
							//  p00_minus_p20 | p11_minus_p21
	fildl	C(r_p0)+8		// r_p0[2] | r_p2[2] | ystepdenominv |
							//   xstepdenominv | p00_minus_p20 | p11_minus_p21
	fildl	C(r_p1)+8		// r_p1[2] | r_p0[2] | r_p2[2] | ystepdenominv |
							//  xstepdenominv | p00_minus_p20 | p11_minus_p21
	fxch	%st(2)			// r_p2[2] | r_p0[2] | r_p1[2] | ystepdenominv |
							//  xstepdenominv | p00_minus_p20 | p11_minus_p21
	fld		%st(0)			// r_p2[2] | r_p2[2] | r_p0[2] | r_p1[2] |
							//  ystepdenominv | xstepdenominv | p00_minus_p20 |
							//  p11_minus_p21
	fsubrp	%st(0),%st(2)	// r_p2[2] | t0 | r_p1[2] | ystepdenominv |
							//  xstepdenominv | p00_minus_p20 | p11_minus_p21
	fsubrp	%st(0),%st(2)	// t0 | t1 | ystepdenominv | xstepdenominv |
							//  p00_minus_p20 | p11_minus_p21

//	r_sstepx = (int)((t1 * p01_minus_p21 - t0 * p11_minus_p21) *
//			xstepdenominv);
//	r_sstepy = (int)((t1 * p00_minus_p20 - t0 * p10_minus_p20) *
//			ystepdenominv);

	fld		%st(0)			// t0 | t0 | t1 | ystepdenominv | xstepdenominv
	fmul	%st(6),%st(0)	// t0*p11_minus_p21 | t0 | t1 | ystepdenominv |
							//  xstepdenominv | p00_minus_p20 | p11_minus_p21
	fxch	%st(2)			// t1 | t0 | t0*p11_minus_p21 | ystepdenominv |
							//  xstepdenominv | p00_minus_p20 | p11_minus_p21
	fld		%st(0)			// t1 | t1 | t0 | t0*p11_minus_p21 |
							//  ystepdenominv | xstepdenominv | p00_minus_p20 |
							//  p11_minus_p21
	fmuls	p01_minus_p21	// t1*p01_minus_p21 | t1 | t0 | t0*p11_minus_p21 |
							//  ystepdenominv | xstepdenominv | p00_minus_p20 |
							//  p11_minus_p21
	fxch	%st(2)			// t0 | t1 | t1*p01_minus_p21 | t0*p11_minus_p21 |
							//  ystepdenominv | xstepdenominv | p00_minus_p20 |
							//  p11_minus_p21
	fmuls	p10_minus_p20	// t0*p10_minus_p20 | t1 | t1*p01_minus_p21 |
							//  t0*p11_minus_p21 | ystepdenominv |
							//  xstepdenominv | p00_minus_p20 | p11_minus_p21
	fxch	%st(1)			// t1 | t0*p10_minus_p20 | t1*p01_minus_p21 |
							//  t0*p11_minus_p21 | ystepdenominv |
							//  xstepdenominv | p00_minus_p20 | p11_minus_p21
	fmul	%st(6),%st(0)	// t1*p00_minus_p20 | t0*p10_minus_p20 |
							//  t1*p01_minus_p21 | t0*p11_minus_p21 |
							//  ystepdenominv | xstepdenominv | p00_minus_p20 |
							//  p11_minus_p21
	fxch	%st(2)			// t1*p01_minus_p21 | t0*p10_minus_p20 |
							//  t1*p00_minus_p20 | t0*p11_minus_p21 |
							//  ystepdenominv | xstepdenominv | p00_minus_p20 |
							//  p11_minus_p21
	fsubp	%st(0),%st(3)	// t0*p10_minus_p20 | t1*p00_minus_p20 |
							//  t1*p01_minus_p21 - t0*p11_minus_p21 |
							//  ystepdenominv | xstepdenominv | p00_minus_p20 |
							//  p11_minus_p21
	fsubrp	%st(0),%st(1)	// t1*p00_minus_p20 - t0*p10_minus_p20 |
							//  t1*p01_minus_p21 - t0*p11_minus_p21 |
							//  ystepdenominv | xstepdenominv | p00_minus_p20 |
							//  p11_minus_p21
	fmul	%st(2),%st(0)	// (t1*p00_minus_p20 - t0*p10_minus_p20)*
							//   ystepdenominv |
							//  t1*p01_minus_p21 - t0*p11_minus_p21 |
							//  ystepdenominv | xstepdenominv | p00_minus_p20 |
							//  p11_minus_p21
	fxch	%st(1)			// t1*p01_minus_p21 - t0*p11_minus_p21 |
							//  (t1*p00_minus_p20 - t0*p10_minus_p20)*
							//   ystepdenominv | ystepdenominv |
							//  xstepdenominv | p00_minus_p20 | p11_minus_p21
	fmul	%st(3),%st(0)	// (t1*p01_minus_p21 - t0*p11_minus_p21)*
							//  xstepdenominv |
							//  (t1*p00_minus_p20 - t0*p10_minus_p20)*
							//  ystepdenominv | ystepdenominv |
							//  xstepdenominv | p00_minus_p20 | p11_minus_p21
	fxch	%st(1)			// (t1*p00_minus_p20 - t0*p10_minus_p20)*
							//  ystepdenominv |
							//  (t1*p01_minus_p21 - t0*p11_minus_p21)*
							//  xstepdenominv | ystepdenominv |
							//  xstepdenominv | p00_minus_p20 | p11_minus_p21
	fistpl	C(r_sstepy)		// r_sstepx | ystepdenominv | xstepdenominv |
							//  p00_minus_p20 | p11_minus_p21
	fistpl	C(r_sstepx)		// ystepdenominv | xstepdenominv | p00_minus_p20 |
							//  p11_minus_p21

//	t0 = r_p0[3] - r_p2[3];
//	t1 = r_p1[3] - r_p2[3];

	fildl	C(r_p2)+12		// r_p2[3] | ystepdenominv | xstepdenominv |
							//  p00_minus_p20 | p11_minus_p21
	fildl	C(r_p0)+12		// r_p0[3] | r_p2[3] | ystepdenominv |
							//  xstepdenominv | p00_minus_p20 | p11_minus_p21
	fildl	C(r_p1)+12		// r_p1[3] | r_p0[3] | r_p2[3] | ystepdenominv |
							//  xstepdenominv | p00_minus_p20 | p11_minus_p21
	fxch	%st(2)			// r_p2[3] | r_p0[3] | r_p1[3] | ystepdenominv |
							//  xstepdenominv | p00_minus_p20 | p11_minus_p21
	fld		%st(0)			// r_p2[3] | r_p2[3] | r_p0[3] | r_p1[3] |
							//  ystepdenominv | xstepdenominv | p00_minus_p20 |
							//  p11_minus_p21
	fsubrp	%st(0),%st(2)	// r_p2[3] | t0 | r_p1[3] | ystepdenominv |
							//  xstepdenominv | p00_minus_p20 | p11_minus_p21
	fsubrp	%st(0),%st(2)	// t0 | t1 | ystepdenominv | xstepdenominv |
							//  p00_minus_p20 | p11_minus_p21

//	r_tstepx = (int)((t1 * p01_minus_p21 - t0 * p11_minus_p21) *
//			xstepdenominv);
//	r_tstepy = (int)((t1 * p00_minus_p20 - t0 * p10_minus_p20) *
//			ystepdenominv);

	fld		%st(0)			// t0 | t0 | t1 | ystepdenominv | xstepdenominv |
							//  p00_minus_p20 | p11_minus_p21
	fmul	%st(6),%st(0)	// t0*p11_minus_p21 | t0 | t1 | ystepdenominv |
							//  xstepdenominv | p00_minus_p20 | p11_minus_p21
	fxch	%st(2)			// t1 | t0 | t0*p11_minus_p21 | ystepdenominv |
							//  xstepdenominv | p00_minus_p20 | p11_minus_p21
	fld		%st(0)			// t1 | t1 | t0 | t0*p11_minus_p21 |
							//  ystepdenominv | xstepdenominv | p00_minus_p20 |
							//  p11_minus_p21
	fmuls	p01_minus_p21	// t1*p01_minus_p21 | t1 | t0 | t0*p11_minus_p21 |
							//  ystepdenominv | xstepdenominv | p00_minus_p20 |
							//  p11_minus_p21
	fxch	%st(2)			// t0 | t1 | t1*p01_minus_p21 | t0*p11_minus_p21 |
							//  ystepdenominv | xstepdenominv | p00_minus_p20 |
							//  p11_minus_p21
	fmuls	p10_minus_p20	// t0*p10_minus_p20 | t1 | t1*p01_minus_p21 |
							//  t0*p11_minus_p21 | ystepdenominv |
							//  xstepdenominv | p00_minus_p20 | p11_minus_p21
	fxch	%st(1)			// t1 | t0*p10_minus_p20 | t1*p01_minus_p21 |
							//  t0*p11_minus_p21 | ystepdenominv |
							//  xstepdenominv | p00_minus_p20 | p11_minus_p21
	fmul	%st(6),%st(0)	// t1*p00_minus_p20 | t0*p10_minus_p20 |
							//  t1*p01_minus_p21 | t0*p11_minus_p21 |
							//  ystepdenominv | xstepdenominv | p00_minus_p20 |
							//  p11_minus_p21
	fxch	%st(2)			// t1*p01_minus_p21 | t0*p10_minus_p20 |
							//  t1*p00_minus_p20 | t0*p11_minus_p21 |
							//  ystepdenominv | xstepdenominv | p00_minus_p20 |
							//  p11_minus_p21
	fsubp	%st(0),%st(3)	// t0*p10_minus_p20 | t1*p00_minus_p20 |
							//  t1*p01_minus_p21 - t0*p11_minus_p21 |
							//  ystepdenominv | xstepdenominv | p00_minus_p20 |
							//  p11_minus_p21
	fsubrp	%st(0),%st(1)	// t1*p00_minus_p20 - t0*p10_minus_p20 |
							//  t1*p01_minus_p21 - t0*p11_minus_p21 |
							//  ystepdenominv | xstepdenominv | p00_minus_p20 |
							//  p11_minus_p21
	fmul	%st(2),%st(0)	// (t1*p00_minus_p20 - t0*p10_minus_p20)*
							//   ystepdenominv |
							//  t1*p01_minus_p21 - t0*p11_minus_p21 |
							//  ystepdenominv | xstepdenominv | p00_minus_p20 |
							//  p11_minus_p21
	fxch	%st(1)			// t1*p01_minus_p21 - t0*p11_minus_p21 |
							//  (t1*p00_minus_p20 - t0*p10_minus_p20)*
							//  ystepdenominv | ystepdenominv |
							//  xstepdenominv | p00_minus_p20 | p11_minus_p21
	fmul	%st(3),%st(0)	// (t1*p01_minus_p21 - t0*p11_minus_p21)*
							//  xstepdenominv |
							//  (t1*p00_minus_p20 - t0*p10_minus_p20)*
							//  ystepdenominv | ystepdenominv |
							//  xstepdenominv | p00_minus_p20 | p11_minus_p21
	fxch	%st(1)			// (t1*p00_minus_p20 - t0*p10_minus_p20)*
							//  ystepdenominv |
							//  (t1*p01_minus_p21 - t0*p11_minus_p21)*
							//  xstepdenominv | ystepdenominv |
							//  xstepdenominv | p00_minus_p20 | p11_minus_p21
	fistpl	C(r_tstepy)		// r_tstepx | ystepdenominv | xstepdenominv |
							//  p00_minus_p20 | p11_minus_p21
	fistpl	C(r_tstepx)		// ystepdenominv | xstepdenominv | p00_minus_p20 |
							//  p11_minus_p21

//	t0 = r_p0[5] - r_p2[5];
//	t1 = r_p1[5] - r_p2[5];

	fildl	C(r_p2)+20		// r_p2[5] | ystepdenominv | xstepdenominv |
							//  p00_minus_p20 | p11_minus_p21
	fildl	C(r_p0)+20		// r_p0[5] | r_p2[5] | ystepdenominv |
							//  xstepdenominv | p00_minus_p20 | p11_minus_p21
	fildl	C(r_p1)+20		// r_p1[5] | r_p0[5] | r_p2[5] | ystepdenominv |
							//  xstepdenominv | p00_minus_p20 | p11_minus_p21
	fxch	%st(2)			// r_p2[5] | r_p0[5] | r_p1[5] | ystepdenominv |
							//  xstepdenominv | p00_minus_p20 | p11_minus_p21
	fld		%st(0)			// r_p2[5] | r_p2[5] | r_p0[5] | r_p1[5] |
							//  ystepdenominv | xstepdenominv | p00_minus_p20 |
							//  p11_minus_p21
	fsubrp	%st(0),%st(2)	// r_p2[5] | t0 | r_p1[5] | ystepdenominv |
							//  xstepdenominv | p00_minus_p20 | p11_minus_p21
	fsubrp	%st(0),%st(2)	// t0 | t1 | ystepdenominv | xstepdenominv |
							//  p00_minus_p20 | p11_minus_p21

//	r_zistepx = (int)((t1 * p01_minus_p21 - t0 * p11_minus_p21) *
//			xstepdenominv);
//	r_zistepy = (int)((t1 * p00_minus_p20 - t0 * p10_minus_p20) *
//			ystepdenominv);

	fld		%st(0)			// t0 | t0 | t1 | ystepdenominv | xstepdenominv |
							//  p00_minus_p20 | p11_minus_p21
	fmulp	%st(0),%st(6)	// t0 | t1 | ystepdenominv | xstepdenominv |
							//  p00_minus_p20 | t0*p11_minus_p21
	fxch	%st(1)			// t1 | t0 | ystepdenominv | xstepdenominv |
							//  p00_minus_p20 | t0*p11_minus_p21
	fld		%st(0)			// t1 | t1 | t0 | ystepdenominv | xstepdenominv |
							//  p00_minus_p20 | t0*p11_minus_p21
	fmuls	p01_minus_p21	// t1*p01_minus_p21 | t1 | t0 | ystepdenominv |
							//  xstepdenominv | p00_minus_p20 |
							//  t0*p11_minus_p21
	fxch	%st(2)			// t0 | t1 | t1*p01_minus_p21 | ystepdenominv |
							//  xstepdenominv | p00_minus_p20 |
							//  t0*p11_minus_p21
	fmuls	p10_minus_p20	// t0*p10_minus_p20 | t1 | t1*p01_minus_p21 |
							//  ystepdenominv | xstepdenominv | p00_minus_p20 |
							//  t0*p11_minus_p21
	fxch	%st(1)			// t1 | t0*p10_minus_p20 | t1*p01_minus_p21 |
							//  ystepdenominv | xstepdenominv | p00_minus_p20 |
							//  t0*p11_minus_p21
	fmulp	%st(0),%st(5)	// t0*p10_minus_p20 | t1*p01_minus_p21 |
							//  ystepdenominv | xstepdenominv |
							//  t1*p00_minus_p20 | t0*p11_minus_p21
	fxch	%st(5)			// t0*p11_minus_p21 | t1*p01_minus_p21 |
							//  ystepdenominv | xstepdenominv |
							//  t1*p00_minus_p20 | t0*p10_minus_p20
	fsubrp	%st(0),%st(1)	// t1*p01_minus_p21 - t0*p11_minus_p21 |
							//  ystepdenominv | xstepdenominv |
							//  t1*p00_minus_p20 | t0*p10_minus_p20
	fxch	%st(3)			// t1*p00_minus_p20 | ystepdenominv |
							//  xstepdenominv |
							//  t1*p01_minus_p21 - t0*p11_minus_p21 |
							//  t0*p10_minus_p20
	fsubp	%st(0),%st(4)	// ystepdenominv | xstepdenominv |
							//  t1*p01_minus_p21 - t0*p11_minus_p21 |
							//  t1*p00_minus_p20 - t0*p10_minus_p20
	fxch	%st(1)			// xstepdenominv | ystepdenominv |
							//  t1*p01_minus_p21 - t0*p11_minus_p21 |
							//  t1*p00_minus_p20 - t0*p10_minus_p20
	fmulp	%st(0),%st(2)	// ystepdenominv |