SPARC assembly optimizations from David Miller.

1 files changed, 1410 insertions, 0 deletions

diff --git a/src/mesa/sparc/xform.S b/src/mesa/sparc/xform.S
new file mode 100644
index 0000000..368fdd9
--- /dev/null
+++ b/src/mesa/sparc/xform.S
@@ -0,0 +1,1410 @@
+/* $Id: xform.S,v 1.1 2001/05/23 14:27:03 brianp Exp $ */
+
+	/* TODO
+	 *
+	 * 1) It would be nice if load/store double could be used
+	 *    at least for the matrix parts.  I think for the matrices
+	 *    it is safe, but for the vertices it probably is not due to
+	 *    things like glInterleavedArrays etc.
+	 *
+	 *    UPDATE: Trying this now in sparc_matrix.h -DaveM_990624
+	 *
+	 * 2) One extremely slick trick would be if we could enclose
+	 *    groups of xform calls on the same vertices such that
+	 *    we just load the matrix into f16-->f31 before the calls
+	 *    and then we would not have to do them here.  This may be
+	 *    tricky and not much of a gain though.
+	 */
+
+#ifdef __sparc_v9__
+#define LDPTR		ldx
+#define V4F_DATA	0x00
+#define V4F_START	0x08
+#define V4F_COUNT	0x10
+#define V4F_STRIDE	0x14
+#define V4F_SIZE	0x18
+#define V4F_FLAGS	0x1c
+#else
+#define LDPTR		ld
+#define V4F_DATA	0x00
+#define V4F_START	0x04
+#define V4F_COUNT	0x08
+#define V4F_STRIDE	0x0c
+#define V4F_SIZE	0x10
+#define V4F_FLAGS	0x14
+#endif
+
+#define VEC_SIZE_1   	1
+#define VEC_SIZE_2   	3
+#define VEC_SIZE_3   	7
+#define VEC_SIZE_4   	15
+
+	.text
+	.align	64
+
+__set_v4f_1:
+	ld	[%o0 + V4F_FLAGS], %g2
+	mov	1, %g1
+	st	%g1, [%o0 + V4F_SIZE]
+	or	%g2, VEC_SIZE_1, %g2
+	retl
+	 st	%g2, [%o0 + V4F_FLAGS]
+__set_v4f_2:
+	ld	[%o0 + V4F_FLAGS], %g2
+	mov	2, %g1
+	st	%g1, [%o0 + V4F_SIZE]
+	or	%g2, VEC_SIZE_2, %g2
+	retl
+	 st	%g2, [%o0 + V4F_FLAGS]
+__set_v4f_3:
+	ld	[%o0 + V4F_FLAGS], %g2
+	mov	3, %g1
+	st	%g1, [%o0 + V4F_SIZE]
+	or	%g2, VEC_SIZE_3, %g2
+	retl
+	 st	%g2, [%o0 + V4F_FLAGS]
+__set_v4f_4:
+	ld	[%o0 + V4F_FLAGS], %g2
+	mov	4, %g1
+	st	%g1, [%o0 + V4F_SIZE]
+	or	%g2, VEC_SIZE_4, %g2
+	retl
+	 st	%g2, [%o0 + V4F_FLAGS]
+
+#include "sparc_matrix.h"
+
+	/* First the raw versions. */
+
+	.globl	_mesa_sparc_transform_points1_general
+_mesa_sparc_transform_points1_general:
+	ld	[%o2 + V4F_STRIDE], %o5
+	LDPTR	[%o2 + V4F_START], %g1
+	LDPTR	[%o0 + V4F_START], %g2
+	ld	[%o2 + V4F_COUNT], %g3
+
+	LDMATRIX_0_1_2_3_12_13_14_15(%o1)
+
+	cmp	%g3, 1
+	st	%g3, [%o0 + V4F_COUNT]
+	bl	3f
+	 clr	%o1
+
+	be	2f
+	 andn	%g3, 1, %o2
+
+1:	ld	[%g1 + 0x00], %f0	! LSU	Group
+	add	%g1, %o5, %g1		! IEU0
+	ld	[%g1 + 0x00], %f8	! LSU	Group
+	add	%o1, 2, %o1		! IEU0
+	add	%g1, %o5, %g1		! IEU1
+	fmuls	%f0, M0, %f1		! FGM	Group	1-cycle stall on %f0
+	fmuls	%f0, M1, %f2		! FGM	Group
+	fmuls	%f0, M2, %f3		! FGM	Group
+	fmuls	%f0, M3, %f4		! FGM	Group
+	fmuls	%f8, M0, %f9		! FGM	Group	f1 available
+	fadds	%f1, M12, %f1		! FGA
+	st	%f1, [%g2 + 0x00]	! LSU
+	fmuls	%f8, M1, %f10		! FGM	Group	f2 available
+	fadds	%f2, M13, %f2		! FGA
+	st	%f2, [%g2 + 0x04]	! LSU
+	fmuls	%f8, M2, %f11		! FGM	Group	f3 available
+	fadds	%f3, M14, %f3		! FGA
+	st	%f3, [%g2 + 0x08]	! LSU
+	fmuls	%f8, M3, %f12		! FGM	Group	f4 available
+	fadds	%f4, M15, %f4		! FGA
+	st	%f4, [%g2 + 0x0c]	! LSU
+	fadds	%f9, M12, %f9		! FGA	Group	f9 available
+	st	%f9, [%g2 + 0x10]	! LSU
+	fadds	%f10, M13, %f10		! FGA	Group	f10 available
+	st	%f10, [%g2 + 0x14]	! LSU
+	fadds	%f11, M14, %f11		! FGA	Group	f11 available
+	st	%f11, [%g2 + 0x18]	! LSU
+	fadds	%f12, M15, %f12		! FGA	Group	f12 available
+	st	%f12, [%g2 + 0x1c]	! LSU
+	cmp	%o1, %o2		! IEU1
+	bne	1b			! CTI
+	 add	%g2, 0x20, %g2		! IEU0	Group
+
+	cmp	%o1, %g3
+	be	3f
+	 nop
+
+2:	ld	[%g1 + 0x00], %f0	! LSU	Group
+	fmuls	%f0, M0, %f1		! FGM	Group	1-cycle stall on %f0
+	fmuls	%f0, M1, %f2		! FGM	Group
+	fmuls	%f0, M2, %f3		! FGM	Group
+	fmuls	%f0, M3, %f4		! FGM	Group
+	fadds	%f1, M12, %f1		! FGA	Group
+	st	%f1, [%g2 + 0x00]	! LSU
+	fadds	%f2, M13, %f2		! FGA	Group
+	st	%f2, [%g2 + 0x04]	! LSU
+	fadds	%f3, M14, %f3		! FGA	Group
+	st	%f3, [%g2 + 0x08]	! LSU
+	fadds	%f4, M15, %f4		! FGA	Group
+	st	%f4, [%g2 + 0x0c]	! LSU
+
+3:
+	ba	__set_v4f_4
+	 nop
+
+	.globl	_mesa_sparc_transform_points1_identity
+_mesa_sparc_transform_points1_identity:
+	cmp	%o0, %o2
+	be	4f
+	 ld	[%o2 + V4F_STRIDE], %o5
+	LDPTR	[%o2 + V4F_START], %g1
+	LDPTR	[%o0 + V4F_START], %g2
+	ld	[%o2 + V4F_COUNT], %g3
+
+	cmp	%g3, 1
+	st	%g3, [%o0 + V4F_COUNT]
+	bl	3f
+	 clr	%o1
+
+	be	2f
+	 andn	%g3, 1, %o2
+
+1:	ld	[%g1 + 0x00], %f0	! LSU	Group
+	add	%g1, %o5, %g1		! IEU0
+	ld	[%g1 + 0x00], %f1	! LSU	Group
+	add	%o1, 2, %o1		! IEU0
+	add	%g1, %o5, %g1		! IEU1
+	st	%f0, [%g2 + 0x00]	! LSU	Group
+	cmp	%o1, %o2		! IEU1
+	st	%f1, [%g2 + 0x10]	! LSU	Group
+	bne	1b			! CTI
+	 add	%g2, 0x20, %g2		! IEU0
+
+	cmp	%o1, %g3
+	be	3f
+	 nop
+
+2:	ld	[%g1 + 0x00], %f0
+	addx	%g0, %g0, %g0
+	st	%f0, [%g2 + 0x00]
+
+3:
+	ba	__set_v4f_1
+	 nop
+
+4:	retl
+	 nop
+
+	.globl	_mesa_sparc_transform_points1_2d
+_mesa_sparc_transform_points1_2d:
+	ld	[%o2 + V4F_STRIDE], %o5
+	LDPTR	[%o2 + V4F_START], %g1
+	LDPTR	[%o0 + V4F_START], %g2
+	ld	[%o2 + V4F_COUNT], %g3
+
+	LDMATRIX_0_1_12_13(%o1)
+
+	cmp	%g3, 1
+	st	%g3, [%o0 + V4F_COUNT]
+	bl	3f
+	 clr	%o1
+
+	be	2f
+	 andn	%g3, 1, %o2
+
+1:	ld	[%g1 + 0x00], %f0	! LSU	Group
+	add	%g1, %o5, %g1		! IEU0
+	ld	[%g1 + 0x00], %f8	! LSU	Group
+	add	%o1, 2, %o1		! IEU0
+	add	%g1, %o5, %g1		! IEU1
+	fmuls	%f0, M0, %f1		! FGM	Group
+	fmuls	%f0, M1, %f2		! FGM	Group
+	fmuls	%f8, M0, %f9		! FGM	Group
+	fmuls	%f8, M1, %f10		! FGM	Group
+	fadds	%f1, M12, %f3		! FGA	Group	f1 available
+	st	%f3, [%g2 + 0x00]	! LSU
+	fadds	%f2, M13, %f4		! FGA	Group	f2 available
+	st	%f4, [%g2 + 0x04]	! LSU
+	fadds	%f9, M12, %f11		! FGA	Group	f9 available
+	st	%f11, [%g2 + 0x10]	! LSU
+	fadds	%f10, M13, %f12		! FGA	Group	f10 available
+	st	%f12, [%g2 + 0x14]	! LSU
+	cmp	%o1, %o2		! IEU1
+	bne	1b			! CTI
+	 add	%g2, 0x20, %g2		! IEU0	Group
+
+	cmp	%o1, %g3
+	be	3f
+	 nop
+
+2:	ld	[%g1 + 0x00], %f0
+	fmuls	%f0, M0, %f1
+	fmuls	%f0, M1, %f2
+	fadds	%f1, M12, %f3
+	st	%f3, [%g2 + 0x00]
+	fadds	%f2, M13, %f4
+	st	%f4, [%g2 + 0x04]
+
+3:
+	ba	__set_v4f_2
+	 nop
+
+	.globl	_mesa_sparc_transform_points1_2d_no_rot
+_mesa_sparc_transform_points1_2d_no_rot:
+	ld	[%o2 + V4F_STRIDE], %o5
+	LDPTR	[%o2 + V4F_START], %g1
+	LDPTR	[%o0 + V4F_START], %g2
+	ld	[%o2 + V4F_COUNT], %g3
+
+	LDMATRIX_0_12_13(%o1)
+
+	cmp	%g3, 1
+	st	%g3, [%o0 + V4F_COUNT]
+	bl	3f
+	 clr	%o1
+
+	be	2f
+	 andn	%g3, 1, %o2
+
+1:	ld	[%g1 + 0x00], %f0	! LSU	Group
+	add	%g1, %o5, %g1		! IEU0
+	ld	[%g1 + 0x00], %f4	! LSU	Group
+	add	%o1, 2, %o1		! IEU0
+	add	%g1, %o5, %g1		! IEU1
+	fmuls	%f0, M0, %f1		! FGM	Group
+	fmuls	%f4, M0, %f5		! FGM	Group
+	fadds	%f1, M12, %f3		! FGA	Group, 2 cycle stall, f1 available
+	st	%f3, [%g2 + 0x00]	! LSU
+	st	M13, [%g2 + 0x04]	! LSU	Group, f5 available
+	fadds	%f5, M12, %f6		! FGA
+	st	%f6, [%g2 + 0x10]	! LSU	Group
+	st	M13, [%g2 + 0x14]	! LSU	Group
+	cmp	%o1, %o2		! IEU1
+	bne	1b			! CTI
+	 add	%g2, 0x20, %g2		! IEU0	Group
+
+	cmp	%o1, %g3
+	be	3f
+	 nop
+
+2:	ld	[%g1 + 0x00], %f0
+	fmuls	%f0, M0, %f1
+	fadds	%f1, M12, %f3
+	st	%f3, [%g2 + 0x00]
+	st	M13, [%g2 + 0x04]
+
+3:
+	ba	__set_v4f_2
+	 nop
+
+	.globl	_mesa_sparc_transform_points1_3d
+_mesa_sparc_transform_points1_3d:
+	ld	[%o2 + V4F_STRIDE], %o5
+	LDPTR	[%o2 + V4F_START], %g1
+	LDPTR	[%o0 + V4F_START], %g2
+	ld	[%o2 + V4F_COUNT], %g3
+
+	LDMATRIX_0_1_2_12_13_14(%o1)
+
+	cmp	%g3, 1
+	st	%g3, [%o0 + V4F_COUNT]
+	bl	3f
+	 clr	%o1
+
+	be	2f
+	 andn	%g3, 1, %o2
+
+1:	ld	[%g1 + 0x00], %f0	! LSU	Group
+	add	%g1, %o5, %g1		! IEU0
+	ld	[%g1 + 0x00], %f4	! LSU	Group
+	add	%o1, 2, %o1		! IEU0
+	add	%g1, %o5, %g1		! IEU1
+	fmuls	%f0, M0, %f1		! FGM	Group
+	fmuls	%f0, M1, %f2		! FGM	Group
+	fmuls	%f0, M2, %f3		! FGM	Group
+	fmuls	%f4, M0, %f5		! FGM	Group
+	fadds	%f1, M12, %f1		! FGA	Group, f1 available
+	st	%f1, [%g2 + 0x00]	! LSU
+	fmuls	%f4, M1, %f6		! FGM
+	fadds	%f2, M13, %f2		! FGA	Group, f2 available
+	st	%f2, [%g2 + 0x04]	! LSU
+	fmuls	%f4, M2, %f7		! FGM
+	fadds	%f3, M14, %f3		! FGA	Group, f3 available
+	st	%f3, [%g2 + 0x08]	! LSU
+	fadds	%f5, M12, %f5		! FGA	Group, f5 available
+	st	%f5, [%g2 + 0x10]	! LSU
+	fadds	%f6, M13, %f6		! FGA	Group, f6 available
+	st	%f6, [%g2 + 0x14]	! LSU
+	fadds	%f7, M14, %f7		! FGA	Group, f7 available
+	st	%f7, [%g2 + 0x18]	! LSU
+	cmp	%o1, %o2		! IEU1
+	bne	1b			! CTI
+	 add	%g2, 0x20, %g2		! IEU0	Group
+
+	cmp	%o1, %g3
+	be	3f
+	 nop
+
+2:	ld	[%g1 + 0x00], %f0
+	fmuls	%f0, M0, %f1
+	fmuls	%f0, M1, %f2
+	fmuls	%f0, M2, %f3
+	fadds	%f1, M12, %f1
+	st	%f1, [%g2 + 0x00]
+	fadds	%f2, M13, %f2
+	st	%f2, [%g2 + 0x04]
+	fadds	%f3, M14, %f3
+	st	%f3, [%g2 + 0x08]
+
+3:
+	ba	__set_v4f_3
+	 nop
+
+	.globl	_mesa_sparc_transform_points1_3d_no_rot
+_mesa_sparc_transform_points1_3d_no_rot:
+	ld	[%o2 + V4F_STRIDE], %o5
+	LDPTR	[%o2 + V4F_START], %g1
+	LDPTR	[%o0 + V4F_START], %g2
+	ld	[%o2 + V4F_COUNT], %g3
+
+	LDMATRIX_0_12_13_14(%o1)
+
+	cmp	%g3, 1
+	st	%g3, [%o0 + V4F_COUNT]
+	bl	3f
+	 clr	%o1
+
+	be	2f
+	 andn	%g3, 1, %o2
+
+1:	ld	[%g1 + 0x00], %f0	! LSU	Group
+	add	%g1, %o5, %g1		! IEU0
+	ld	[%g1 + 0x00], %f2	! LSU	Group
+	add	%o1, 2, %o1		! IEU0
+	add	%g1, %o5, %g1		! IEU1
+	fmuls	%f0, M0, %f1		! FGM	Group
+	fmuls	%f2, M0, %f3		! FGM	Group
+	fadds	%f1, M12, %f1		! FGA	Group, 2 cycle stall, f1 available
+	st	%f1, [%g2 + 0x00]	! LSU
+	fadds	%f3, M12, %f3		! FGA	Group, f3 available
+	st	M13, [%g2 + 0x04]	! LSU
+	st	M14, [%g2 + 0x08]	! LSU	Group
+	st	%f3, [%g2 + 0x10]	! LSU	Group
+	st	M13, [%g2 + 0x14]	! LSU	Group
+	st	M14, [%g2 + 0x18]	! LSU	Group
+	cmp	%o1, %o2		! IEU1
+	bne	1b			! CTI
+	 add	%g2, 0x20, %g2		! IEU0	Group
+
+	cmp	%o1, %g3
+	be	3f
+	 nop
+
+2:	ld	[%g1 + 0x00], %f0
+	fmuls	%f0, M0, %f1
+	fadds	%f1, M12, %f1
+	st	%f1, [%g2 + 0x00]
+	st	M13, [%g2 + 0x04]
+	st	M14, [%g2 + 0x08]
+
+3:
+	ba	__set_v4f_3
+	 nop
+
+	.globl	_mesa_sparc_transform_points1_perspective
+_mesa_sparc_transform_points1_perspective:
+	ld	[%o2 + V4F_STRIDE], %o5
+	LDPTR	[%o2 + V4F_START], %g1
+	LDPTR	[%o0 + V4F_START], %g2
+	ld	[%o2 + V4F_COUNT], %g3
+
+	LDMATRIX_0_14(%o1)
+
+	cmp	%g3, 1
+	st	%g3, [%o0 + V4F_COUNT]
+	bl	3f
+	 clr	%o1
+
+	be	2f
+	 andn	%g3, 1, %o2
+
+1:	ld	[%g1 + 0x00], %f0	! LSU	Group
+	add	%g1, %o5, %g1		! IEU0
+	ld	[%g1 + 0x00], %f2	! LSU	Group
+	add	%o1, 2, %o1		! IEU0
+	add	%g1, %o5, %g1		! IEU1
+	fmuls	%f0, M0, %f1		! FGM	Group
+	st	%f1, [%g2 + 0x00]	! LSU
+	fmuls	%f2, M0, %f3		! FGM	Group
+	st	%g0, [%g2 + 0x04]	! LSU
+	st	M14, [%g2 + 0x08]	! LSU	Group
+	st	%g0, [%g2 + 0x0c]	! LSU	Group
+	st	%f3, [%g2 + 0x10]	! LSU	Group
+	st	%g0, [%g2 + 0x14]	! LSU	Group
+	st	M14, [%g2 + 0x18]	! LSU	Group
+	st	%g0, [%g2 + 0x1c]	! LSU	Group
+	cmp	%o1, %o2		! IEU1
+	bne	1b			! CTI
+	 add	%g2, 0x20, %g2		! IEU0	Group
+
+	cmp	%o1, %g3
+	be	3f
+	 nop
+
+2:	ld	[%g1 + 0x00], %f0
+	fmuls	%f0, M0, %f1
+	st	%f1, [%g2 + 0x00]
+	st	%g0, [%g2 + 0x04]
+	st	M14, [%g2 + 0x08]
+	st	%g0, [%g2 + 0x0c]
+
+3:
+	ba	__set_v4f_4
+	 nop
+
+	.globl	_mesa_sparc_transform_points2_general
+_mesa_sparc_transform_points2_general:
+	ld	[%o2 + V4F_STRIDE], %o5
+	LDPTR	[%o2 + V4F_START], %g1
+	LDPTR	[%o0 + V4F_START], %g2
+	ld	[%o2 + V4F_COUNT], %g3
+
+	LDMATRIX_0_1_2_3_4_5_6_7_12_13_14_15(%o1)
+
+	cmp	%g3, 0
+	st	%g3, [%o0 + V4F_COUNT]
+	be	2f
+	 clr	%o1
+
+1:	ld	[%g1 + 0x00], %f0	! LSU	Group
+	ld	[%g1 + 0x04], %f1	! LSU	Group
+	add	%o1, 1, %o1		! IEU0
+	add	%g1, %o5, %g1		! IEU1
+	fmuls	%f0, M0, %f2		! FGM	Group
+	fmuls	%f0, M1, %f3		! FGM	Group
+	fmuls	%f0, M2, %f4		! FGM	Group
+	fmuls	%f0, M3, %f5		! FGM	Group
+	fadds	%f2, M12, %f2		! FGA	Group	f2 available
+	fmuls	%f1, M4, %f6		! FGM
+	fadds	%f3, M13, %f3		! FGA	Group	f3 available
+	fmuls	%f1, M5, %f7		! FGM
+	fadds	%f4, M14, %f4		! FGA	Group	f4 available
+	fmuls	%f1, M6, %f8		! FGM
+	fadds	%f5, M15, %f5		! FGA	Group	f5 available
+	fmuls	%f1, M7, %f9		! FGM
+	fadds	%f2, %f6, %f2		! FGA	Group	f6 available
+	st	%f2, [%g2 + 0x00]	! LSU
+	fadds	%f3, %f7, %f3		! FGA	Group	f7 available
+	st	%f3, [%g2 + 0x04]	! LSU
+	fadds	%f4, %f8, %f4		! FGA	Group	f8 available
+	st	%f4, [%g2 + 0x08]	! LSU
+	fadds	%f5, %f9, %f5		! FGA	Group	f9 available
+	st	%f5, [%g2 + 0x0c]	! LSU
+	cmp	%o1, %g3		! IEU1
+	bne	1b			! CTI
+	 add	%g2, 0x10, %g2		! IEU0	Group
+2:
+	ba	__set_v4f_4
+	 nop
+
+	.globl	_mesa_sparc_transform_points2_identity
+_mesa_sparc_transform_points2_identity:
+	cmp	%o2, %o0
+	be	3f
+	 ld	[%o2 + V4F_STRIDE], %o5
+	LDPTR	[%o2 + V4F_START], %g1
+	LDPTR	[%o0 + V4F_START], %g2
+	ld	[%o2 + V4F_COUNT], %g3
+
+	cmp	%g3, 0
+	st	%g3, [%o0 + V4F_COUNT]
+	be	2f
+	 clr	%o1
+
+1:	ld	[%g1 + 0x00], %f0	! LSU	Group
+	add	%o1, 1, %o1		! IEU0
+	ld	[%g1 + 0x04], %f1	! LSU	Group
+	add	%g1, %o5, %g1		! IEU0
+	cmp	%o1, %g3		! IEU1
+	st	%f0, [%g2 + 0x00]	! LSU	Group
+	st	%f1, [%g2 + 0x04]	! LSU	Group
+	bne	1b			! CTI
+	 add	%g2, 0x10, %g2		! IEU0
+2:
+	ba	__set_v4f_2
+	 nop
+
+3:	retl
+	 nop
+
+	.globl	_mesa_sparc_transform_points2_2d
+_mesa_sparc_transform_points2_2d:
+	ld	[%o2 + V4F_STRIDE], %o5
+	LDPTR	[%o2 + V4F_START], %g1
+	LDPTR	[%o0 + V4F_START], %g2
+	ld	[%o2 + V4F_COUNT], %g3
+
+	LDMATRIX_0_1_4_5_12_13(%o1)
+
+	cmp	%g3, 1
+	st	%g3, [%o0 + V4F_COUNT]
+	bl	3f
+	 clr	%o1
+
+	be	2f
+	 andn	%g3, 1, %o2
+
+1:	ld	[%g1 + 0x00], %f0	! LSU	Group
+	ld	[%g1 + 0x04], %f1	! LSU	Group
+	add	%o1, 2, %o1		! IEU0
+	add	%g1, %o5, %g1		! IEU1
+	fmuls	%f0, M0, %f2		! FGM
+	ld	[%g1 + 0x00], %f8	! LSU	Group
+	fmuls	%f0, M1, %f3		! FGM
+	ld	[%g1 + 0x04], %f9	! LSU	Group
+	fmuls	%f1, M4, %f6		! FGM
+	fmuls	%f1, M5, %f7		! FGM	Group
+	add	%g1, %o5, %g1		! IEU0
+	fmuls	%f8, M0, %f10		! FGM	Group	f2 available
+	fadds	%f2, M12, %f2		! FGA
+	fmuls	%f8, M1, %f11		! FGM	Group	f3 available
+	fadds	%f3, M13, %f3		! FGA
+	fmuls	%f9, M4, %f12		! FGM	Group
+	fmuls	%f9, M5, %f13		! FGM	Group
+	fadds	%f10, M12, %f10		! FGA	Group	f2, f10 available
+	fadds	%f2, %f6, %f2		! FGA	Group	f3, f11 available
+	st	%f2, [%g2 + 0x00]	! LSU
+	fadds	%f11, M13, %f11		! FGA	Group	f12 available
+	fadds	%f3, %f7, %f3		! FGA	Group	f13 available
+	st	%f3, [%g2 + 0x04]	! LSU
+	fadds	%f10, %f12, %f10	! FGA	Group	f10 available
+	st	%f10, [%g2 + 0x10]	! LSU
+	fadds	%f11, %f13, %f11	! FGA	Group	f11 available
+	st	%f11, [%g2 + 0x14]	! LSU
+	cmp	%o1, %o2		! IEU1
+	bne	1b			! CTI
+	 add	%g2, 0x20, %g2		! IEU0	Group
+
+	cmp	%o1, %g3
+	be	3f
+	 nop
+
+2:	ld	[%g1 + 0x00], %f0	! LSU	Group
+	ld	[%g1 + 0x04], %f1	! LSU	Group
+	fmuls	%f0, M0, %f2		! FGM	Group
+	fmuls	%f0, M1, %f3		! FGM	Group
+	fmuls	%f1, M4, %f6		! FGM	Group
+	fmuls	%f1, M5, %f7		! FGM	Group
+	fadds	%f2, M12, %f2		! FGA	Group	f2 available
+	fadds	%f3, M13, %f3		! FGA	Group	f3 available
+	fadds	%f2, %f6, %f2		! FGA	Group	2 cycle stall, f2 available
+	st	%f2, [%g2 + 0x00]	! LSU
+	fadds	%f3, %f7, %f3		! FGA	Group	f3 available
+	st	%f3, [%g2 + 0x04]	! LSU
+
+3:
+	ba	__set_v4f_2
+	 nop
+
+	.globl	_mesa_sparc_transform_points2_2d_no_rot
+_mesa_sparc_transform_points2_2d_no_rot:
+	ld	[%o2 + V4F_STRIDE], %o5
+	LDPTR	[%o2 + V4F_START], %g1
+	LDPTR	[%o0 + V4F_START], %g2
+	ld	[%o2 + V4F_COUNT], %g3
+
+	LDMATRIX_0_5_12_13(%o1)
+
+	cmp	%g3, 1
+	st	%g3, [%o0 + V4F_COUNT]
+	bl	3f
+	 clr	%o1
+
+	be	2f
+	 andn	%g3, 1, %o2
+
+1:	ld	[%g1 + 0x00], %f0	! LSU	Group
+	ld	[%g1 + 0x04], %f1	! LSU	Group
+	add	%o1, 2, %o1		! IEU0
+	add	%g1, %o5, %g1		! IEU1
+	ld	[%g1 + 0x00], %f4	! LSU	Group
+	fmuls	%f0, M0, %f2		! FGM
+	ld	[%g1 + 0x04], %f5	! LSU	Group
+	fmuls	%f1, M5, %f3		! FGM
+	fmuls	%f4, M0, %f6		! FGM	Group
+	add	%g1, %o5, %g1		! IEU0
+	fmuls	%f5, M5, %f7		! FGM	Group
+	fadds	%f2, M12, %f2		! FGA	Group	f2 available
+	st	%f2, [%g2 + 0x00]	! LSU
+	fadds	%f3, M13, %f3		! FGA	Group	f3 available
+	st	%f3, [%g2 + 0x04]	! LSU
+	fadds	%f6, M12, %f6		! FGA	Group	f6 available
+	st	%f6, [%g2 + 0x10]	! LSU
+	fadds	%f7, M13, %f7		! FGA	Group	f7 available
+	st	%f7, [%g2 + 0x14]	! LSU
+	cmp	%o1, %o2		! IEU1
+	bne	1b			! CTI
+	 add	%g2, 0x20, %g2		! IEU0	Group
+
+	cmp	%o1, %g3
+	be	3f
+	 nop
+
+2:	ld	[%g1 + 0x00], %f0	! LSU	Group
+	ld	[%g1 + 0x04], %f1	! LSU	Group
+	fmuls	%f0, M0, %f2		! FGM	Group
+	fmuls	%f1, M5, %f3		! FGM	Group
+	fadds	%f2, M12, %f2		! FGA	Group, 2 cycle stall, f2 available
+	st	%f2, [%g2 + 0x00]	! LSU
+	fadds	%f3, M13, %f3		! FGA	Group	f3 available
+	st	%f3, [%g2 + 0x04]	! LSU
+
+3:
+	ba	__set_v4f_2
+	 nop
+
+	/* orig: 12 cycles */
+	.globl	_mesa_sparc_transform_points2_3d
+_mesa_sparc_transform_points2_3d:
+	ld	[%o2 + V4F_STRIDE], %o5
+	ld	[%o2 + V4F_START], %g1
+	ld	[%o0 + V4F_START], %g2
+	ld	[%o2 + V4F_COUNT], %g3
+
+	LDMATRIX_0_1_2_3_4_5_6_12_13_14(%o1)
+
+	cmp	%g3, 1
+	st	%g3, [%o0 + V4F_COUNT]
+	bl	3f
+	 clr	%o1
+
+	be	2f
+	 andn	%g3, 1, %o2
+
+1:	ld	[%g1 + 0x00], %f0	! LSU	Group
+	ld	[%g1 + 0x04], %f1	! LSU	Group
+	add	%o1, 2, %o1		! IEU0
+	add	%g1, %o5, %g1		! IEU1
+	ld	[%g1 + 0x00], %f9	! LSU	Group
+	fmuls	%f0, M0, %f2		! FGM
+	ld	[%g1 + 0x04], %f10	! LSU	Group
+	fmuls	%f0, M1, %f3		! FGM
+	fmuls	%f0, M2, %f4		! FGM	Group
+	add	%g1, %o5, %g1		! IEU0
+	fmuls	%f1, M4, %f6		! FGM	Group
+	fmuls	%f1, M5, %f7		! FGM	Group	f2 available
+	fadds	%f2, M12, %f2		! FGA
+	fmuls	%f1, M6, %f8		! FGM	Group	f3 available
+	fadds	%f3, M13, %f3		! FGA
+	fmuls	%f9, M0, %f11		! FGM	Group	f4 available
+	fadds	%f4, M14, %f4		! FGA
+	fmuls	%f9, M1, %f12		! FGM	Group	f6 available
+	fmuls	%f9, M2, %f13		! FGM	Group	f2, f7 available
+	fadds	%f2, %f6, %f2		! FGA
+	st	%f2, [%g2 + 0x00]	! LSU
+	fmuls	%f10, M4, %f14		! FGM	Group	f3, f8 available
+	fadds	%f3, %f7, %f3		! FGA
+	st	%f3, [%g2 + 0x04]	! LSU
+	fmuls	%f10, M5, %f15		! FGM	Group	f4, f11 available
+	fadds	%f11, M12, %f11		! FGA
+	fmuls	%f10, M6, %f0		! FGM	Group	f12 available
+	fadds	%f12, M13, %f12		! FGA
+	fadds	%f13, M14, %f13		! FGA	Group	f13 available
+	fadds	%f4, %f8, %f4		! FGA	Group	f14 available
+	st	%f4, [%g2 + 0x08]	! LSU
+	fadds	%f11, %f14, %f11	! FGA	Group	f15, f11 available
+	st	%f11, [%g2 + 0x10]	! LSU
+	fadds	%f12, %f15, %f12	! FGA	Group	f0, f12 available
+	st	%f12, [%g2 + 0x14]	! LSU
+	fadds	%f13, %f0, %f13		! FGA	Group	f13 available
+	st	%f13, [%g2 + 0x18]	! LSU
+
+	cmp	%o1, %o2		! IEU1
+	bne	1b			! CTI
+	 add	%g2, 0x20, %g2		! IEU0	Group
+
+	cmp	%o1, %g3
+	be	3f
+	 nop
+
+2:	ld	[%g1 + 0x00], %f0	! LSU	Group
+	ld	[%g1 + 0x04], %f1	! LSU	Group
+	fmuls	%f0, M0, %f2		! FGM	Group
+	fmuls	%f0, M1, %f3		! FGM	Group
+	fmuls	%f0, M2, %f4		! FGM	Group
+	fmuls	%f1, M4, %f6		! FGM	Group
+	fmuls	%f1, M5, %f7		! FGM	Group	f2 available
+	fadds	%f2, M12, %f2		! FGA
+	fmuls	%f1, M6, %f8		! FGM	Group	f3 available
+	fadds	%f3, M13, %f3		! FGA
+	fadds	%f4, M14, %f4		! FGA	Group	f4 available
+	fadds	%f2, %f6, %f2		! FGA	Group	stall, f2, f6, f7 available
+	st	%f2, [%g2 + 0x00]	! LSU
+	fadds	%f3, %f7, %f3		! FGA	Group	f3, f8 available
+	st	%f3, [%g2 + 0x04]	! LSU
+	fadds	%f4, %f8, %f4		! FGA	Group	f4 available
+	st	%f4, [%g2 + 0x08]	! LSU
+
+3:
+	ba	__set_v4f_3
+	 nop
+
+	.globl	_mesa_sparc_transform_points2_3d_no_rot
+_mesa_sparc_transform_points2_3d_no_rot:
+	ld	[%o2 + V4F_STRIDE], %o5
+	LDPTR	[%o2 + V4F_START], %g1
+	LDPTR	[%o0 + V4F_START], %g2
+	ld	[%o2 + V4F_COUNT], %g3
+
+	LDMATRIX_0_5_12_13_14(%o1)
+
+	cmp	%g3, 1
+	st	%g3, [%o0 + V4F_COUNT]
+	bl	3f
+	 clr	%o3
+
+	be	2f
+	 andn	%g3, 1, %o2
+
+1:	ld	[%g1 + 0x00], %f0	! LSU	Group
+	ld	[%g1 + 0x04], %f1	! LSU	Group
+	add	%o3, 2, %o3		! IEU0
+	add	%g1, %o5, %g1		! IEU1
+	ld	[%g1 + 0x00], %f4	! LSU	Group
+	fmuls	%f0, M0, %f2		! FGM
+	ld	[%g1 + 0x04], %f5	! LSU	Group
+	fmuls	%f1, M5, %f3		! FGM
+	fmuls	%f4, M0, %f6		! FGM	Group
+	add	%g1, %o5, %g1		! IEU0
+	fmuls	%f5, M5, %f7		! FGM	Group
+	fadds	%f2, M12, %f2		! FGA	Group	f2 available
+	st	%f2, [%g2 + 0x00]	! LSU
+	fadds	%f3, M13, %f3		! FGA	Group	f3 available
+	st	%f3, [%g2 + 0x04]	! LSU
+	fadds	%f6, M12, %f6		! FGA	Group	f6 available
+	st	M14, [%g2 + 0x08]	! LSU
+	fadds	%f7, M13, %f7		! FGA	Group	f7 available
+	st	%f6, [%g2 + 0x10]	! LSU
+	st	%f7, [%g2 + 0x14]	! LSU	Group
+	st	M14, [%g2 + 0x18]	! LSU	Group
+	cmp	%o3, %o2		! IEU1
+	bne	1b			! CTI
+	 add	%g2, 0x20, %g2		! IEU0	Group
+
+	cmp	%o3, %g3
+	be	3f
+	 nop
+
+2:	ld	[%g1 + 0x00], %f0	! LSU	Group
+	ld	[%g1 + 0x04], %f1	! LSU	Group
+	fmuls	%f0, M0, %f2		! FGM	Group
+	fmuls	%f1, M5, %f3		! FGM	Group
+	fadds	%f2, M12, %f2		! FGA	Group, 2 cycle stall, f2 available
+	st	%f2, [%g2 + 0x00]	! LSU
+	fadds	%f3, M13, %f3		! FGA	Group	f3 available
+	st	%f3, [%g2 + 0x04]	! LSU
+	st	M14, [%g2 + 0x08]	! LSU	Group
+
+3:	ld	[%o1 + (14 * 0x4)], %g3
+	cmp	%g3, 0
+	bne	__set_v4f_3
+	 nop
+	ba	__set_v4f_2
+	 nop
+
+	.globl	_mesa_sparc_transform_points2_perspective
+_mesa_sparc_transform_points2_perspective:
+	ld	[%o2 + V4F_STRIDE], %o5
+	LDPTR	[%o2 + V4F_START], %g1
+	LDPTR	[%o0 + V4F_START], %g2
+	ld	[%o2 + V4F_COUNT], %g3
+
+	LDMATRIX_0_5_14(%o1)
+
+	cmp	%g3, 0
+	st	%g3, [%o0 + V4F_COUNT]
+	be	2f
+	 clr	%o1
+
+1:	ld	[%g1 + 0x00], %f0
+	ld	[%g1 + 0x04], %f1
+	add	%o1, 1, %o1
+	add	%g1, %o5, %g1
+	fmuls	%f0, M0, %f2
+	st	%f2, [%g2 + 0x00]
+	fmuls	%f1, M5, %f3
+	st	%f3, [%g2 + 0x04]
+	st	M14, [%g2 + 0x08]
+	st	%g0, [%g2 + 0x0c]
+	cmp	%o1, %g3
+	bne	1b
+	 add	%g2, 0x10, %g2
+2:
+	ba	__set_v4f_4
+	 nop
+
+	.globl	_mesa_sparc_transform_points3_general
+_mesa_sparc_transform_points3_general:
+	ld	[%o2 + V4F_STRIDE], %o5
+	LDPTR	[%o2 + V4F_START], %g1
+	LDPTR	[%o0 + V4F_START], %g2
+	ld	[%o2 + V4F_COUNT], %g3
+
+	LDMATRIX_0_1_2_3_4_5_6_7_8_9_10_11_12_13_14_15(%o1)
+
+	cmp	%g3, 0
+	st	%g3, [%o0 + V4F_COUNT]
+	be	2f
+	 clr	%o1
+
+1:	ld	[%g1 + 0x00], %f0	! LSU	Group
+	ld	[%g1 + 0x04], %f1	! LSU	Group
+	ld	[%g1 + 0x08], %f2	! LSU	Group
+	add	%o1, 1, %o1		! IEU0
+	add	%g1, %o5, %g1		! IEU1
+	fmuls	%f0, M0, %f3		! FGM
+	fmuls	%f1, M4, %f7		! FGM	Group
+	fmuls	%f0, M1, %f4		! FGM	Group
+	fmuls	%f1, M5, %f8		! FGM	Group
+	fmuls	%f0, M2, %f5		! FGM	Group	f3 available
+	fmuls	%f1, M6, %f9		! FGM	Group	f7 available
+	fadds	%f3, %f7, %f3		! FGA
+	fmuls	%f0, M3, %f6		! FGM	Group	f4 available
+	fmuls	%f1, M7, %f10		! FGM	Group	f8 available
+	fadds	%f4, %f8, %f4		! FGA
+	fmuls	%f2, M8, %f7		! FGM	Group	f5 available
+	fmuls	%f2, M9, %f8		! FGM	Group	f9,f3 available
+	fadds	%f5, %f9, %f5		! FGA
+	fmuls	%f2, M10, %f9		! FGM	Group	f6 available
+	fadds	%f6, %f10, %f6		! FGA	Group	f10,f4 available
+	fmuls	%f2, M11, %f10		! FGM
+	fadds	%f3, M12, %f3		! FGA	Group	f7 available
+	fadds	%f4, M13, %f4		! FGA	Group	f8,f5 available
+	fadds	%f5, M14, %f5		! FGA	Group	f9 available
+	fadds	%f6, M15, %f6		! FGA	Group	f10,f6 available
+	fadds	%f3, %f7, %f3		! FGA	Group	f3 available
+	st	%f3, [%g2 + 0x00]	! LSU
+	fadds	%f4, %f8, %f4		! FGA	Group	f4 available
+	st	%f4, [%g2 + 0x04]	! LSU
+	fadds	%f5, %f9, %f5		! FGA	Group	f5 available
+	st	%f5, [%g2 + 0x08]	! LSU
+	fadds	%f6, %f10, %f6		! FGA	Group	f6 available
+	st	%f6, [%g2 + 0x0c]	! LSU
+	cmp	%o1, %g3		! IEU1
+	bne	1b			! CTI
+	 add	%g2, 0x10, %g2		! IEU0	Group
+2:
+	ba	__set_v4f_4
+	 nop
+
+	.globl	_mesa_sparc_transform_points3_identity
+_mesa_sparc_transform_points3_identity:
+	ld	[%o2 + V4F_STRIDE], %o5
+	LDPTR	[%o2 + V4F_START], %g1
+	LDPTR	[%o0 + V4F_START], %g2
+	ld	[%o2 + V4F_COUNT], %g3
+
+	cmp	%g3, 0
+	st	%g3, [%o0 + V4F_COUNT]
+	be	2f
+	 clr	%o1
+
+1:	ld	[%g1 + 0x00], %f0
+	ld	[%g1 + 0x04], %f1
+	ld	[%g1 + 0x08], %f2
+	add	%o1, 1, %o1
+	add	%g1, %o5, %g1
+	cmp	%o1, %g3
+	st	%f0, [%g2 + 0x00]
+	st	%f1, [%g2 + 0x04]
+	st	%f2, [%g2 + 0x08]
+	bne	1b
+	 add	%g2, 0x10, %g2
+2:
+	ba	__set_v4f_3
+	 nop
+
+	.globl	_mesa_sparc_transform_points3_2d
+_mesa_sparc_transform_points3_2d:
+	ld	[%o2 + V4F_STRIDE], %o5
+	LDPTR	[%o2 + V4F_START], %g1
+	LDPTR	[%o0 + V4F_START], %g2
+	ld	[%o2 + V4F_COUNT], %g3
+
+	LDMATRIX_0_1_4_5_12_13(%o1)
+
+	cmp	%g3, 0
+	st	%g3, [%o0 + V4F_COUNT]
+	be	2f
+	 clr	%o1
+
+1:	ld	[%g1 + 0x00], %f0	! LSU	Group
+	ld	[%g1 + 0x04], %f1	! LSU	Group
+	ld	[%g1 + 0x08], %f2	! LSU	Group
+	add	%o1, 1, %o1		! IEU0
+	add	%g1, %o5, %g1		! IEU1
+	fmuls	%f0, M0, %f3		! FGM
+	fmuls	%f0, M1, %f4		! FGM	Group
+	fmuls	%f1, M4, %f6		! FGM	Group
+	fmuls	%f1, M5, %f7		! FGM	Group
+	fadds	%f3, M12, %f3		! FGA	Group	f3 available
+	fadds	%f4, M13, %f4		! FGA	Group	f4 available
+	fadds	%f3, %f6, %f3		! FGA	Group	f6 available
+	st	%f3, [%g2 + 0x00]	! LSU
+	fadds	%f4, %f7, %f4		! FGA	Group	f7 available
+	st	%f4, [%g2 + 0x04]	! LSU
+	st	%f2, [%g2 + 0x08]	! LSU	Group
+	cmp	%o1, %g3		! IEU1
+	bne	1b			! CTI
+	 add	%g2, 0x10, %g2		! IEU0	Group
+2:
+	ba	__set_v4f_3
+	 nop
+
+	.globl	_mesa_sparc_transform_points3_2d_no_rot
+_mesa_sparc_transform_points3_2d_no_rot:
+	ld	[%o2 + V4F_STRIDE], %o5
+	LDPTR	[%o2 + V4F_START], %g1
+	LDPTR	[%o0 + V4F_START], %g2
+	ld	[%o2 + V4F_COUNT], %g3
+
+	LDMATRIX_0_5_12_13(%o1)
+
+	cmp	%g3, 0
+	st	%g3, [%o0 + V4F_COUNT]
+	be	2f
+	 clr	%o1
+
+1:	ld	[%g1 + 0x00], %f0	! LSU	Group
+	ld	[%g1 + 0x04], %f1	! LSU	Group
+	ld	[%g1 + 0x08], %f2	! LSU	Group
+	add	%o1, 1, %o1		! IEU0
+	add	%g1, %o5, %g1		! IEU1
+	fmuls	%f0, M0, %f3		! FGM
+	fmuls	%f1, M5, %f4		! FGM	Group
+	st	%f2, [%g2 + 0x08]	! LSU
+	fadds	%f3, M12, %f3		! FGA	Group
+	st	%f3, [%g2 + 0x00]	! LSU
+	fadds	%f4, M13, %f4		! FGA	Group
+	st	%f4, [%g2 + 0x04]	! LSU
+	cmp	%o1, %g3		! IEU1
+	bne	1b			! CTI
+	 add	%g2, 0x10, %g2		! IEU0	Group
+2:
+	ba	__set_v4f_3
+	 nop
+
+	.globl	_mesa_sparc_transform_points3_3d
+_mesa_sparc_transform_points3_3d:
+	ld	[%o2 + V4F_STRIDE], %o5
+	LDPTR	[%o2 + V4F_START], %g1
+	LDPTR	[%o0 + V4F_START], %g2
+	ld	[%o2 + V4F_COUNT], %g3
+
+	LDMATRIX_0_1_2_4_5_6_8_9_10_12_13_14(%o1)
+
+	cmp	%g3, 0
+	st	%g3, [%o0 + V4F_COUNT]
+	be	2f
+	 clr	%o1
+
+1:	ld	[%g1 + 0x00], %f0	! LSU	Group
+	ld	[%g1 + 0x04], %f1	! LSU	Group
+	ld	[%g1 + 0x08], %f2	! LSU	Group
+	add	%o1, 1, %o1		! IEU0
+	add	%g1, %o5, %g1		! IEU1
+	fmuls	%f0, M0, %f3		! FGM
+	fmuls	%f1, M4, %f6		! FGM	Group
+	fmuls	%f0, M1, %f4		! FGM	Group
+	fmuls	%f1, M5, %f7		! FGM	Group
+	fmuls	%f0, M2, %f5		! FGM	Group	f3 available
+	fmuls	%f1, M6, %f8		! FGM	Group	f6 available
+	fadds	%f3, %f6, %f3		! FGA
+	fmuls	%f2, M8, %f9		! FGM	Group	f4 available
+	fmuls	%f2, M9, %f10		! FGM	Group	f7 available
+	fadds	%f4, %f7, %f4		! FGA
+	fmuls	%f2, M10, %f11		! FGM	Group	f5 available
+	fadds	%f5, %f8, %f5		! FGA	Group	f8, f3 available
+	fadds	%f3, %f9, %f3		! FGA	Group	f9 available
+	fadds	%f4, %f10, %f4		! FGA	Group	f10, f4 available
+	fadds	%f5, %f11, %f5		! FGA	Group	stall, f11, f5 available
+	fadds	%f3, M12, %f3		! FGA	Group	f3 available
+	st	%f3, [%g2 + 0x00]	! LSU
+	fadds	%f4, M13, %f4		! FGA	Group	f4 available
+	st	%f4, [%g2 + 0x04]	! LSU
+	fadds	%f5, M14, %f5		! FGA	Group	f5 available
+	st	%f5, [%g2 + 0x08]	! LSU
+	cmp	%o1, %g3		! IEU1
+	bne	1b			! CTI
+	 add	%g2, 0x10, %g2		! IEU0	Group
+2:
+	ba	__set_v4f_3
+	 nop
+
+	.globl	_mesa_sparc_transform_points3_3d_no_rot
+_mesa_sparc_transform_points3_3d_no_rot:
+	ld	[%o2 + V4F_STRIDE], %o5
+	LDPTR	[%o2 + V4F_START], %g1
+	LDPTR	[%o0 + V4F_START], %g2
+	ld	[%o2 + V4F_COUNT], %g3
+
+	LDMATRIX_0_5_10_12_13_14(%o1)
+
+	cmp	%g3, 0
+	st	%g3, [%o0 + V4F_COUNT]
+	be	2f
+	 clr	%o1
+
+1:	ld	[%g1 + 0x00], %f0	! LSU	Group
+	ld	[%g1 + 0x04], %f1	! LSU	Group
+	ld	[%g1 + 0x08], %f2	! LSU	Group
+	add	%o1, 1, %o1		! IEU0
+	add	%g1, %o5, %g1		! IEU1
+	cmp	%o1, %g3		! IEU1	Group
+	fmuls	%f0, M0, %f3		! FGM
+	fmuls	%f1, M5, %f4		! FGM	Group
+	fmuls	%f2, M10, %f5		! FGM	Group
+	fadds	%f3, M12, %f3		! FGA	Group, stall, f3 available
+	st	%f3, [%g2 + 0x00]	! LSU
+	fadds	%f4, M13, %f4		! FGA	Group, f4 available
+	st	%f4, [%g2 + 0x04]	! LSU
+	fadds	%f5, M14, %f5		! FGA	Group, f5 available
+	st	%f5, [%g2 + 0x08]	! LEU
+	bne	1b			! CTI
+	 add	%g2, 0x10, %g2		! IEU0	Group
+2:
+	ba	__set_v4f_3
+	 nop
+
+	.globl	_mesa_sparc_transform_points3_perspective
+_mesa_sparc_transform_points3_perspective:
+	ld	[%o2 + V4F_STRIDE], %o5
+	LDPTR	[%o2 + V4F_START], %g1
+	LDPTR	[%o0 + V4F_START], %g2
+	ld	[%o2 + V4F_COUNT], %g3
+
+	LDMATRIX_0_5_8_9_10_14(%o1)
+
+	cmp	%g3, 0
+	st	%g3, [%o0 + V4F_COUNT]
+	be	2f
+	 clr	%o1
+
+1:	ld	[%g1 + 0x00], %f0	! LSU	Group
+	ld	[%g1 + 0x04], %f1	! LSU	Group
+	ld	[%g1 + 0x08], %f2	! LSU	Group
+	add	%o1, 1, %o1		! IEU0
+	add	%g1, %o5, %g1		! IEU1
+	fmuls	%f0, M0, %f3		! FGM
+	fmuls	%f2, M8, %f6		! FGM	Group
+	fmuls	%f1, M5, %f4		! FGM	Group
+	fmuls	%f2, M9, %f7		! FGM	Group
+	fmuls	%f2, M10, %f5		! FGM	Group	f3 available
+	fadds	%f3, %f6, %f3		! FGA	Group	f6 available
+	st	%f3, [%g2 + 0x00]	! LSU
+	fadds	%f4, %f7, %f4		! FGA	Group	stall, f4, f7 available
+	st	%f4, [%g2 + 0x04]	! LSU
+	fadds	%f5, M14, %f5		! FGA	Group
+	st	%f5, [%g2 + 0x08]	! LSU
+	fnegs	%f2, %f6		! FGA	Group
+	st	%f6, [%g2 + 0x0c]	! LSU
+	cmp	%o1, %g3		! IEU1
+	bne	1b			! CTI
+	 add	%g2, 0x10, %g2		! IEU0	Group
+2:
+	ba	__set_v4f_4
+	 nop
+
+	.globl	_mesa_sparc_transform_points4_general
+_mesa_sparc_transform_points4_general:
+	ld	[%o2 + V4F_STRIDE], %o5
+	LDPTR	[%o2 + V4F_START], %g1
+	LDPTR	[%o0 + V4F_START], %g2
+	ld	[%o2 + V4F_COUNT], %g3
+
+	LDMATRIX_0_1_2_3_4_5_6_7_8_9_10_11_12_13_14_15(%o1)
+
+	cmp	%g3, 0
+	st	%g3, [%o0 + V4F_COUNT]
+	be	2f
+	 clr	%o1
+
+1:	ld	[%g1 + 0x00], %f0	! LSU	Group
+	ld	[%g1 + 0x04], %f1	! LSU	Group
+	ld	[%g1 + 0x08], %f2	! LSU	Group
+	ld	[%g1 + 0x0c], %f3	! LSU	Group
+	add	%o1, 1, %o1		! IEU0
+	add	%g1, %o5, %g1		! IEU1
+	fmuls	%f0, M0, %f4		! FGM	Group
+	fmuls	%f1, M4, %f8		! FGM	Group
+	fmuls	%f0, M1, %f5		! FGM	Group
+	fmuls	%f1, M5, %f9		! FGM	Group
+	fmuls	%f0, M2, %f6		! FGM	Group	f4 available
+	fmuls	%f1, M6, %f10		! FGM	Group	f8 available
+	fadds	%f4, %f8, %f4		! FGA
+	fmuls	%f0, M3, %f7		! FGM	Group	f5 available
+	fmuls	%f1, M7, %f11		! FGM	Group	f9 available
+	fadds	%f5, %f9, %f5		! FGA
+	fmuls	%f2, M8, %f12		! FGM	Group	f6 available
+	fmuls	%f2, M9, %f13		! FGM	Group	f10, f4 available
+	fadds	%f6, %f10, %f6		! FGA
+	fmuls	%f2, M10, %f14		! FGM	Group	f7 available
+	fmuls	%f2, M11, %f15		! FGM	Group	f11, f5 available
+	fadds	%f7, %f11, %f7		! FGA
+	fmuls	%f3, M12, %f8		! FGM	Group	f12 available
+	fadds	%f4, %f12, %f4		! FGA
+	fmuls	%f3, M13, %f9		! FGM	Group	f13, f6 available
+	fadds	%f5, %f13, %f5		! FGA
+	fmuls	%f3, M14, %f10		! FGM	Group	f14 available
+	fadds	%f6, %f14, %f6		! FGA
+	fmuls	%f3, M15, %f11		! FGM	Group	f15, f7 available
+	fadds	%f7, %f15, %f7		! FGA
+	fadds	%f4, %f8, %f4		! FGA	Group	f8, f4 available
+	st	%f4, [%g2 + 0x00]	! LSU
+	fadds	%f5, %f9, %f5		! FGA	Group	f9, f5 available
+	st	%f5, [%g2 + 0x04]	! LSU
+	fadds	%f6, %f10, %f6		! FGA	Group	f10, f6 available
+	st	%f6, [%g2 + 0x08]	! LSU
+	fadds	%f7, %f11, %f7		! FGA	Group	f11, f7 available
+	st	%f7, [%g2 + 0x0c]	! LSU
+	cmp	%o1, %g3		! IEU1
+	bne	1b			! CTI
+	 add	%g2, 0x10, %g2		! IEU0	Group
+2:
+	ba	__set_v4f_4
+	 nop
+
+	.globl	_mesa_sparc_transform_points4_identity
+_mesa_sparc_transform_points4_identity:
+	ld	[%o2 + V4F_STRIDE], %o5
+	LDPTR	[%o2 + V4F_START], %g1
+	LDPTR	[%o0 + V4F_START], %g2
+	ld	[%o2 + V4F_COUNT], %g3
+
+	cmp	%g3, 0
+	st	%g3, [%o0 + V4F_COUNT]
+	be	2f
+	 clr	%o1
+
+1:	ld	[%g1 + 0x00], %f0
+	ld	[%g1 + 0x04], %f1
+	ld	[%g1 + 0x08], %f2
+	add	%o1, 1, %o1
+	ld	[%g1 + 0x0c], %f3
+	add	%g1, %o5, %g1
+	st	%f0, [%g2 + 0x00]
+	st	%f1, [%g2 + 0x04]
+	st	%f2, [%g2 + 0x08]
+	cmp	%o1, %g3
+	st	%f3, [%g2 + 0x0c]
+	bne	1b
+	 add	%g2, 0x10, %g2
+2:
+	ba	__set_v4f_4
+	 nop
+
+	.globl	_mesa_sparc_transform_points4_2d
+_mesa_sparc_transform_points4_2d:
+	ld	[%o2 + V4F_STRIDE], %o5
+	LDPTR	[%o2 + V4F_START], %g1
+	LDPTR	[%o0 + V4F_START], %g2
+	ld	[%o2 + V4F_COUNT], %g3
+
+	LDMATRIX_0_1_4_5_12_13(%o1)
+
+	cmp	%g3, 0
+	st	%g3, [%o0 + V4F_COUNT]
+	be	2f
+	 clr	%o1
+
+1:	ld	[%g1 + 0x00], %f0	! LSU	Group
+	ld	[%g1 + 0x04], %f1	! LSU	Group
+	ld	[%g1 + 0x08], %f2	! LSU	Group
+	ld	[%g1 + 0x0c], %f3	! LSU	Group
+	add	%o1, 1, %o1		! IEU0
+	add	%g1, %o5, %g1		! IEU1
+	fmuls	%f0, M0, %f4		! FGM
+	fmuls	%f1, M4, %f8		! FGM	Group
+	fmuls	%f0, M1, %f5		! FGM	Group
+	fmuls	%f1, M5, %f9		! FGM	Group	f4 available
+	fmuls	%f3, M12, %f12		! FGM	Group
+	fmuls	%f3, M13, %f13		! FGM	Group	f8 available
+	fadds	%f4, %f8, %f4		! FGA
+	fadds	%f5, %f9, %f5		! FGA	Group	stall, f5, f9 available
+	fadds	%f4, %f12, %f4		! FGA	Group	2 cycle stall, f4, f12, f13 avail
+	st	%f4, [%g2 + 0x00]	! LSU
+	fadds	%f5, %f13, %f5		! FGA	Group	f5 available
+	st	%f5, [%g2 + 0x04]	! LSU
+	st	%f2, [%g2 + 0x08]	! LSU	Group
+	st	%f3, [%g2 + 0x0c]	! LSU	Group
+	cmp	%o1, %g3		! IEU1
+	bne	1b			! CTI
+	 add	%g2, 0x10, %g2		! IEU0	Group
+2:
+	ba	__set_v4f_4
+	 nop
+
+	.globl	_mesa_sparc_transform_points4_2d_no_rot
+_mesa_sparc_transform_points4_2d_no_rot:
+	ld	[%o2 + V4F_STRIDE], %o5
+	LDPTR	[%o2 + V4F_START], %g1
+	LDPTR	[%o0 + V4F_START], %g2
+	ld	[%o2 + V4F_COUNT], %g3
+
+	LDMATRIX_0_1_4_5_12_13(%o1)
+
+	cmp	%g3, 0
+	st	%g3, [%o0 + V4F_COUNT]
+	be	2f
+	 clr	%o1
+
+1:	ld	[%g1 + 0x00], %f0
+	ld	[%g1 + 0x04], %f1
+	ld	[%g1 + 0x08], %f2
+	ld	[%g1 + 0x0c], %f3
+	add	%o1, 1, %o1
+	add	%g1, %o5, %g1
+	fmuls	%f0, M0, %f4
+	fmuls	%f3, M12, %f8
+	fmuls	%f1, M5, %f5
+	fmuls	%f3, M13, %f9
+	fadds	%f4, %f8, %f4
+	st	%f4, [%g2 + 0x00]
+	fadds	%f5, %f9, %f5
+	st	%f5, [%g2 + 0x04]
+	st	%f2, [%g2 + 0x08]
+	st	%f3, [%g2 + 0x0c]
+	cmp	%o1, %g3
+	bne	1b
+	 add	%g2, 0x10, %g2
+2:
+	ba	__set_v4f_4
+	 nop
+
+	.globl	_mesa_sparc_transform_points4_3d
+_mesa_sparc_transform_points4_3d:
+	ld	[%o2 + V4F_STRIDE], %o5
+	LDPTR	[%o2 + V4F_START], %g1
+	LDPTR	[%o0 + V4F_START], %g2
+	ld	[%o2 + V4F_COUNT], %g3
+
+	LDMATRIX_0_1_2_4_5_6_8_9_10_12_13_14(%o1)
+
+	cmp	%g3, 0
+	st	%g3, [%o0 + V4F_COUNT]
+	be	2f
+	 clr	%o1
+
+1:	ld	[%g1 + 0x00], %f0	! LSU	Group
+	ld	[%g1 + 0x04], %f1	! LSU	Group
+	ld	[%g1 + 0x08], %f2	! LSU	Group
+	ld	[%g1 + 0x0c], %f3	! LSU	Group
+	add	%o1, 1, %o1		! IEU0
+	add	%g1, %o5, %g1		! IEU1
+	fmuls	%f0, M0, %f4		! FGM
+	fmuls	%f1, M4, %f7		! FGM	Group
+	fmuls	%f0, M1, %f5		! FGM	Group
+	fmuls	%f1, M5, %f8		! FGM	Group
+	fmuls	%f0, M2, %f6		! FGM	Group	f4 available
+	fmuls	%f1, M6, %f9		! FGM	Group	f7 available
+	fadds	%f4, %f7, %f4		! FGA
+	fmuls	%f2, M8, %f10		! FGM	Group	f5 available
+	fmuls	%f2, M9, %f11		! FGM	Group	f8 available
+	fadds	%f5, %f8, %f5		! FGA
+	fmuls	%f2, M10, %f12		! FGM	Group	f6 available
+	fmuls	%f3, M12, %f13		! FGM	Group	f9, f4 available
+	fadds	%f6, %f9, %f6		! FGA
+	fmuls	%f3, M13, %f14		! FGM	Group	f10 available
+	fadds	%f4, %f10, %f4		! FGA
+	fmuls	%f3, M14, %f15		! FGM	Group	f11, f5 available
+	fadds	%f5, %f11, %f5		! FGA
+	fadds	%f6, %f12, %f6		! FGA	Group	stall, f12, f13, f6 available
+	fadds	%f4, %f13, %f4		! FGA	Group	f14, f4 available
+	st	%f4, [%g2 + 0x00]	! LSU
+	fadds	%f5, %f14, %f5		! FGA	Group	f15, f5 available
+	st	%f5, [%g2 + 0x04]	! LSU
+	fadds	%f6, %f15, %f6		! FGA	Group	f6 available
+	st	%f6, [%g2 + 0x08]	! LSU
+	st	%f3, [%g2 + 0x0c]	! LSU	Group
+	cmp	%o1, %g3		! IEU1
+	bne	1b			! CTI
+	 add	%g2, 0x10, %g2		! IEU0	Group
+2:
+	ba	__set_v4f_4
+	 nop
+
+	.globl	_mesa_sparc_transform_points4_3d_no_rot
+_mesa_sparc_transform_points4_3d_no_rot:
+	ld	[%o2 + V4F_STRIDE], %o5
+	LDPTR	[%o2 + V4F_START], %g1
+	LDPTR	[%o0 + V4F_START], %g2
+	ld	[%o2 + V4F_COUNT], %g3
+
+	LDMATRIX_0_5_10_12_13_14(%o1)
+
+	cmp	%g3, 0
+	st	%g3, [%o0 + V4F_COUNT]
+	be	2f
+	 clr	%o1
+
+1:	ld	[%g1 + 0x00], %f0	! LSU	Group
+	ld	[%g1 + 0x04], %f1	! LSU	Group
+	ld	[%g1 + 0x08], %f2	! LSU	Group
+	ld	[%g1 + 0x0c], %f3	! LSU	Group
+	add	%o1, 1, %o1		! IEU0
+	add	%g1, %o5, %g1		! IEU1
+	fmuls	%f0, M0, %f4		! FGM
+	fmuls	%f3, M12, %f7		! FGM	Group
+	fmuls	%f1, M5, %f5		! FGM	Group
+	fmuls	%f3, M13, %f8		! FGM	Group
+	fmuls	%f2, M10, %f6		! FGM	Group	f4 available
+	fmuls	%f3, M14, %f9		! FGM	Group	f7 available
+	fadds	%f4, %f7, %f4		! FGA
+	st	%f4, [%g2 + 0x00]	! LSU
+	fadds	%f5, %f8, %f5		! FGA	Group	stall, f5, f8 available
+	st	%f5, [%g2 + 0x04]	! LSU
+	fadds	%f6, %f9, %f6		! FGA	Group	stall, f6, f9 available
+	st	%f6, [%g2 + 0x08]	! LSU
+	st	%f3, [%g2 + 0x0c]	! LSU	Group
+	cmp	%o1, %g3		! IEU1
+	bne	1b			! CTI
+	 add	%g2, 0x10, %g2		! IEU0	Group
+2:
+	ba	__set_v4f_4
+	 nop
+
+	.globl	_mesa_sparc_transform_points4_perspective
+_mesa_sparc_transform_points4_perspective:
+	ld	[%o2 + V4F_STRIDE], %o5
+	LDPTR	[%o2 + V4F_START], %g1
+	LDPTR	[%o0 + V4F_START], %g2
+	ld	[%o2 + V4F_COUNT], %g3
+
+	LDMATRIX_0_5_8_9_10_14(%o1)
+
+	cmp	%g3, 0
+	st	%g3, [%o0 + V4F_COUNT]
+	be	2f
+	 clr	%o1
+
+1:	ld	[%g1 + 0x00], %f0	! LSU	Group
+	ld	[%g1 + 0x04], %f1	! LSU	Group
+	ld	[%g1 + 0x08], %f2	! LSU	Group
+	ld	[%g1 + 0x0c], %f3	! LSU	Group
+	add	%o1, 1, %o1		! IEU0
+	add	%g1, %o5, %g1		! IEU1
+	fmuls	%f0, M0, %f4		! FGM
+	fmuls	%f2, M8, %f7		! FGM	Group
+	fmuls	%f1, M5, %f5		! FGM	Group
+	fmuls	%f2, M9, %f8		! FGM	Group
+	fmuls	%f2, M10, %f6		! FGM	Group	f4 available
+	fmuls	%f3, M14, %f9		! FGM	Group	f7 available
+	fadds	%f4, %f7, %f4		! FGA
+	st	%f4, [%g2 + 0x00]	! LSU
+	fadds	%f5, %f8, %f5		! FGA	Group	stall, f5, f8 available
+	st	%f5, [%g2 + 0x04]	! LSU
+	fadds	%f6, %f9, %f6		! FGA	Group	stall, f6, f9 available
+	st	%f6, [%g2 + 0x08]	! LSU
+	fnegs	%f2, %f7		! FGA	Group
+	st	%f7, [%g2 + 0x0c]	! LSU
+	cmp	%o1, %g3		! IEU1
+	bne	1b			! CTI
+	 add	%g2, 0x10, %g2		! IEU0	Group
+2:
+	ba	__set_v4f_4
+	 nop