sse_xform3.s

Mesa is an open-source implementation of the OpenGL specification - a system for rendering interacti

    MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) 	/* set dest count */
    MOV_L( CONST(4), REGOFF(V4F_SIZE, EDI) ) 	/* set dest size */

    SHL_L( CONST(4), ECX ) 			/* count *= 16 */
    MOV_L( REGOFF(V4F_START, ESI), ESI ) 	/* ptr to first source vertex */

    MOV_L( REGOFF(V4F_START, EDI), EDI ) 	/* ptr to first dest vertex */
    ADD_L( EDI, ECX ) 				/* count += dest ptr */

ALIGNTEXT32
    MOVSS    ( M(0), XMM1 )			/* -  | -  |  -  | m0  */
    MOVSS    ( M(5), XMM2 )			/* -  | -  |  -  | m5  */
    UNPCKLPS ( XMM2, XMM1 )			/* -  | -  | m5  | m0  */
    MOVLPS   ( M(8), XMM2 )			/* -  | -  | m9  | m8  */
    MOVSS    ( M(10), XMM3 )			/* m10 */
    MOVSS    ( M(14), XMM4 )			/* m14 */
    XORPS    ( XMM6, XMM6 )			/* 0 */

ALIGNTEXT32
LLBL(K_GTP3PR_top):
    MOVLPS   ( S(0), XMM0 )			/* oy | ox */
    MULPS    ( XMM1, XMM0 )			/* oy*m5 | ox*m0 */
    MOVSS    ( S(2), XMM5 )			/* oz */
    SHUFPS   ( CONST(0x0), XMM5, XMM5 )		/* oz | oz */
    MULPS    ( XMM2, XMM5 )			/* oz*m9 | oz*m8 */
    ADDPS    ( XMM5, XMM0 )			/* +oy*m5 | +ox*m0 */
    MOVLPS   ( XMM0, D(0) )			/* ->D(1) | ->D(0) */

    MOVSS    ( S(2), XMM0 )			/* oz */
    MULSS    ( XMM3, XMM0 )			/* oz*m10 */
    ADDSS    ( XMM4, XMM0 )			/* +m14 */
    MOVSS    ( XMM0, D(2) )			/* ->D(2) */

    MOVSS    ( S(2), XMM0 )			/* oz */
    MOVSS    ( XMM6, XMM5 )			/* 0 */
    SUBPS    ( XMM0, XMM5 )			/* -oz */
    MOVSS    ( XMM5, D(3) )			/* ->D(3) */

LLBL(K_GTP3PR_skip):
    ADD_L( CONST(16), EDI )
    ADD_L( EAX, ESI )
    CMP_L( ECX, EDI )
    JNE( LLBL(K_GTP3PR_top) )

LLBL(K_GTP3PR_finish):
    POP_L    ( EDI )
    POP_L    ( ESI )
    RET
#undef FRAME_OFFSET



ALIGNTEXT4
GLOBL GLNAME(_mesa_sse_transform_points3_2d)
HIDDEN(_mesa_sse_transform_points3_2d)
GLNAME(_mesa_sse_transform_points3_2d):

#define FRAME_OFFSET 8
    PUSH_L( ESI )
    PUSH_L( EDI )

    MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) 	/* ptr to source GLvector4f */
    MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) 	/* ptr to dest GLvector4f */

    MOV_L( ARG_MATRIX, EDX ) 			/* ptr to matrix */
    MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) 	/* source count */

    TEST_L( ECX, ECX)
    JZ( LLBL(K_GTP3P2DR_finish) ) 		/* count was zero; go to finish */

    MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) 	/* stride */
    OR_L( CONST(VEC_SIZE_3), REGOFF(V4F_FLAGS, EDI) ) 	/* set dest flags */

    MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) 	/* set dest count */
    MOV_L( CONST(3), REGOFF(V4F_SIZE, EDI) ) 	/* set dest size */

    SHL_L( CONST(4), ECX ) 			/* count *= 16 */
    MOV_L( REGOFF(V4F_START, ESI), ESI ) 	/* ptr to first source vertex */

    MOV_L( REGOFF(V4F_START, EDI), EDI ) 	/* ptr to first dest vertex */
    ADD_L( EDI, ECX ) 				/* count += dest ptr */

ALIGNTEXT32
    MOVLPS( M(0), XMM0 )			/* m1  | m0 */
    MOVLPS( M(4), XMM1 )			/* m5  | m4 */
    MOVLPS( M(12), XMM2 )			/* m13 | m12 */

ALIGNTEXT32
LLBL(K_GTP3P2DR_top):
    MOVSS    ( S(0), XMM3 )			/* ox */
    SHUFPS   ( CONST(0x0), XMM3, XMM3 )		/* ox | ox */
    MULPS    ( XMM0, XMM3 )			/* ox*m1 | ox*m0 */
    MOVSS    ( S(1), XMM4 )			/* oy */
    SHUFPS   ( CONST(0x0), XMM4, XMM4 )		/* oy | oy */
    MULPS    ( XMM1, XMM4 )			/* oy*m5 | oy*m4 */

    ADDPS    ( XMM4, XMM3 )
    ADDPS    ( XMM2, XMM3 )
    MOVLPS   ( XMM3, D(0) )

    MOVSS    ( S(2), XMM3 )
    MOVSS    ( XMM3, D(2) )

LLBL(K_GTP3P2DR_skip):
    ADD_L    ( CONST(16), EDI )
    ADD_L    ( EAX, ESI )
    CMP_L    ( ECX, EDI )
    JNE      ( LLBL(K_GTP3P2DR_top) )

LLBL(K_GTP3P2DR_finish):
    POP_L    ( EDI )
    POP_L    ( ESI )
    RET
#undef FRAME_OFFSET



ALIGNTEXT4
GLOBL GLNAME(_mesa_sse_transform_points3_2d_no_rot)
HIDDEN(_mesa_sse_transform_points3_2d_no_rot)
GLNAME(_mesa_sse_transform_points3_2d_no_rot):

#define FRAME_OFFSET 8
	PUSH_L( ESI )
	PUSH_L( EDI )

	MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) 	/* ptr to source GLvector4f */
	MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) 	/* ptr to dest GLvector4f */

	MOV_L( ARG_MATRIX, EDX ) 		/* ptr to matrix */
	MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) 	/* source count */

	TEST_L( ECX, ECX)
	JZ( LLBL(K_GTP3P2DNRR_finish) ) 	/* count was zero; go to finish */

	MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) 	/* stride */
	OR_L( CONST(VEC_SIZE_3), REGOFF(V4F_FLAGS, EDI) ) 	/* set dest flags */

	MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) 	/* set dest count */
	MOV_L( CONST(3), REGOFF(V4F_SIZE, EDI) ) 	/* set dest size */

	SHL_L( CONST(4), ECX ) 			/* count *= 16 */
	MOV_L( REGOFF(V4F_START, ESI), ESI ) 	/* ptr to first source vertex */

	MOV_L( REGOFF(V4F_START, EDI), EDI ) 	/* ptr to first dest vertex */
	ADD_L( EDI, ECX ) 			/* count += dest ptr */

ALIGNTEXT32
	MOVSS    ( M(0), XMM1 )			/* m0 */
	MOVSS    ( M(5), XMM2 )			/* m5 */
	UNPCKLPS ( XMM2, XMM1 )			/* m5 | m0 */
	MOVLPS   ( M(12), XMM2 )		/* m13 | m12 */

ALIGNTEXT32
LLBL(K_GTP3P2DNRR_top):
	MOVLPS( S(0), XMM0 )			/* oy | ox */
	MULPS( XMM1, XMM0 )			/* oy*m5 | ox*m0 */
	ADDPS( XMM2, XMM0 )			/* +m13 | +m12 */
	MOVLPS( XMM0, D(0) )			/* ->D(1) | ->D(0) */

	MOVSS( S(2), XMM0 )
	MOVSS( XMM0, D(2) )

LLBL(K_GTP3P2DNRR_skip):
	ADD_L( CONST(16), EDI )
	ADD_L( EAX, ESI )
	CMP_L( ECX, EDI )
	JNE( LLBL(K_GTP3P2DNRR_top) )

LLBL(K_GTP3P2DNRR_finish):
	POP_L( EDI )
	POP_L( ESI )
	RET
#undef FRAME_OFFSET




ALIGNTEXT4
GLOBL GLNAME(_mesa_sse_transform_points3_3d)
HIDDEN(_mesa_sse_transform_points3_3d)
GLNAME(_mesa_sse_transform_points3_3d):

#define FRAME_OFFSET 8
	PUSH_L( ESI )
	PUSH_L( EDI )

	MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) 	/* ptr to source GLvector4f */
	MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) 	/* ptr to dest GLvector4f */


	MOV_L( ARG_MATRIX, EDX ) 		/* ptr to matrix */
	MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) 	/* source count */

	TEST_L( ECX, ECX)
	JZ( LLBL(K_GTP3P3DR_finish) ) 	/* count was zero; go to finish */

	MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) 	/* stride */
	OR_L( CONST(VEC_SIZE_3), REGOFF(V4F_FLAGS, EDI) ) 	/* set dest flags */

	MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) 	/* set dest count */
	MOV_L( CONST(3), REGOFF(V4F_SIZE, EDI) ) 	/* set dest size */

	SHL_L( CONST(4), ECX ) 			/* count *= 16 */
	MOV_L( REGOFF(V4F_START, ESI), ESI ) 	/* ptr to first source vertex */

	MOV_L( REGOFF(V4F_START, EDI), EDI ) 	/* ptr to first dest vertex */
	ADD_L( EDI, ECX ) 			/* count += dest ptr */


ALIGNTEXT32
	MOVAPS( M(0), XMM0 )			/* m2  | m1  | m0 */
	MOVAPS( M(4), XMM1 )			/* m6  | m5  | m4 */
	MOVAPS( M(8), XMM2 )			/* m10 | m9  | m8 */
	MOVAPS( M(12), XMM3 )			/* m14 | m13 | m12 */

ALIGNTEXT32
LLBL(K_GTP3P3DR_top):
	MOVSS( S(0), XMM4 )
	SHUFPS( CONST(0x0), XMM4, XMM4 )	/* ox | ox | ox */
	MULPS( XMM0, XMM4 )			/* ox*m2 | ox*m1 | ox*m0 */

	MOVSS( S(1), XMM5 )
	SHUFPS( CONST(0x0), XMM5, XMM5 )	/* oy | oy | oy */
	MULPS( XMM1, XMM5 )			/* oy*m6 | oy*m5 | oy*m4 */

	MOVSS( S(2), XMM6 )
	SHUFPS( CONST(0x0), XMM6, XMM6 )	/* oz | oz | oz */
	MULPS( XMM2, XMM6 )			/* oz*m10 | oz*m9 | oz*m8 */

	ADDPS( XMM5, XMM4 )			/* + | + | + */
	ADDPS( XMM6, XMM4 )			/* + | + | + */
	ADDPS( XMM3, XMM4 )			/* + | + | + */

	MOVLPS( XMM4, D(0) )			/* => D(1) | => D(0) */
	UNPCKHPS( XMM4, XMM4 )
	MOVSS( XMM4, D(2) )

LLBL(K_GTP3P3DR_skip):
	ADD_L( CONST(16), EDI )
	ADD_L( EAX, ESI )
	CMP_L( ECX, EDI )
	JNE( LLBL(K_GTP3P3DR_top) )

LLBL(K_GTP3P3DR_finish):
	POP_L( EDI )
	POP_L( ESI )
	RET
#undef FRAME_OFFSET
#endif
	
#if defined (__ELF__) && defined (__linux__)
	.section .note.GNU-stack,"",%progbits
#endif