imin_sse2.s

Optimized GotoBLAS libraries

	andpd	 %xmm15, %xmm6
#endif
	maxpd	%xmm6, %xmm2
	addq	$2 * SIZE, X
	
.L58:
	testq	$1, M
	je	.L60

	movsd	0 * SIZE(X), %xmm7
	unpcklpd  %xmm7, %xmm7
#ifdef ABS
	andpd	 %xmm15, %xmm7
#endif
	maxpd	%xmm7, %xmm3
	ALIGN_3

.L60:
	movq	XX, X
	movq	MM, M

	maxpd	 %xmm1, %xmm0
	maxpd	 %xmm3, %xmm2
	maxpd	 %xmm2, %xmm0
	movapd	 %xmm0, %xmm1
	unpckhpd %xmm0, %xmm0
	maxsd	 %xmm1, %xmm0
	unpcklpd %xmm0, %xmm0

	movq	M,  I
	sarq	$3, I
	jle	.L65
	ALIGN_4
	
.L62:
#ifdef OPTERON
	prefetcht0	PREFETCHSIZE * SIZE(X)
#endif

#ifdef PENTIUM4
	prefetchnta	PREFETCHSIZE * SIZE(X)
#endif

	movsd	0 * SIZE(X), %xmm1
	movhpd	1 * SIZE(X), %xmm1
#ifdef ABS
	andpd	 %xmm15, %xmm1
#endif
	cmpeqpd	%xmm0, %xmm1

	movsd	2 * SIZE(X), %xmm3
	movhpd	3 * SIZE(X), %xmm3
#ifdef ABS
	andpd	 %xmm15, %xmm3
#endif
	cmpeqpd	%xmm0, %xmm3

	movsd	4 * SIZE(X), %xmm5
	movhpd	5 * SIZE(X), %xmm5
#ifdef ABS
	andpd	 %xmm15, %xmm5
#endif
	cmpeqpd	%xmm0, %xmm5

	movsd	6 * SIZE(X), %xmm7
	movhpd	7 * SIZE(X), %xmm7
#ifdef ABS
	andpd	 %xmm15, %xmm7
#endif
	cmpeqpd	%xmm0, %xmm7

	orpd	%xmm3, %xmm1
	orpd	%xmm7, %xmm5
	orpd	%xmm5, %xmm1
	movmskpd %xmm1, %r11
	testq	 $3, %r11
	jne	 .L63

	addq	$8 * SIZE, X
	addq	$8, RET
	decq	I
	jg	.L62
	jmp	.L65
	ALIGN_4

.L63:
	movsd	0 * SIZE(X), %xmm1
	movsd	1 * SIZE(X), %xmm2
	movsd	2 * SIZE(X), %xmm3
	movsd	3 * SIZE(X), %xmm4
	movsd	4 * SIZE(X), %xmm5
	movsd	5 * SIZE(X), %xmm6
	movsd	6 * SIZE(X), %xmm7
	movsd	7 * SIZE(X), %xmm8
#ifdef ABS
	andpd	 %xmm15, %xmm1
	andpd	 %xmm15, %xmm2
	andpd	 %xmm15, %xmm3
	andpd	 %xmm15, %xmm4
	andpd	 %xmm15, %xmm5
	andpd	 %xmm15, %xmm6
	andpd	 %xmm15, %xmm7
	andpd	 %xmm15, %xmm8
#endif

	addq	$8 * SIZE, X

	incq	RET
	comisd	%xmm0, %xmm1
	je	.L999
	incq	RET
	comisd	%xmm0, %xmm2
	je	.L999
	incq	RET
	comisd	%xmm0, %xmm3
	je	.L999
	incq	RET
	comisd	%xmm0, %xmm4
	je	.L999
	incq	RET
	comisd	%xmm0, %xmm5
	je	.L999
	incq	RET
	comisd	%xmm0, %xmm6
	je	.L999
	incq	RET
	comisd	%xmm0, %xmm7
	je	.L999
	incq	RET
	jmp	.L999
	ALIGN_3

.L65:
	testq	$4, M
	je	.L67

	movsd	0 * SIZE(X), %xmm1
	movsd	1 * SIZE(X), %xmm2
	movsd	2 * SIZE(X), %xmm3
	movsd	3 * SIZE(X), %xmm4
#ifdef ABS
	andpd	 %xmm15, %xmm1
	andpd	 %xmm15, %xmm2
	andpd	 %xmm15, %xmm3
	andpd	 %xmm15, %xmm4
#endif
	addq	$4 * SIZE, X
	incq	RET
	comisd	%xmm0, %xmm1
	je	.L999
	incq	RET
	comisd	%xmm0, %xmm2
	je	.L999
	incq	RET
	comisd	%xmm0, %xmm3
	je	.L999
	incq	RET
	comisd	%xmm0, %xmm4
	je	.L999
	ALIGN_3	

.L67:
	testq	$2, M
	je	.L68

	movsd	0 * SIZE(X), %xmm1
	movsd	1 * SIZE(X), %xmm2
#ifdef ABS
	andpd	 %xmm15, %xmm1
	andpd	 %xmm15, %xmm2
#endif
	addq	$2 * SIZE, X
	incq	RET
	comisd	%xmm0, %xmm1
	je	.L999
	incq	RET
	comisd	%xmm0, %xmm2
	je	.L999
	ALIGN_3
	
.L68:
	incq	RET
	jmp	.L999
	ALIGN_4

.L80:
	movq	M,  I
	sarq	$4, I
	jle	.L85
	ALIGN_4
	
.L81:
#ifdef OPTERON
	prefetcht0	PREFETCHSIZE * SIZE(X)
#endif

#ifdef PENTIUM4
	prefetchnta	PREFETCHSIZE * SIZE(X)
#endif

	movsd	0 * SIZE(X), %xmm4
	addq	INCX, X
	movhpd	0 * SIZE(X), %xmm4
	addq	INCX, X
#ifdef ABS
	andpd	 %xmm15, %xmm4
#endif
	maxpd	%xmm4, %xmm0

	movsd	0 * SIZE(X), %xmm5
	addq	INCX, X
	movhpd	0 * SIZE(X), %xmm5
	addq	INCX, X
#ifdef ABS
	andpd	 %xmm15, %xmm5
#endif
	maxpd	%xmm5, %xmm1

	movsd	0 * SIZE(X), %xmm6
	addq	INCX, X
	movhpd	0 * SIZE(X), %xmm6
	addq	INCX, X
#ifdef ABS
	andpd	 %xmm15, %xmm6
#endif
	maxpd	%xmm6, %xmm2

	movsd	0 * SIZE(X), %xmm7
	addq	INCX, X
	movhpd	0 * SIZE(X), %xmm7
	addq	INCX, X
#ifdef ABS
	andpd	 %xmm15, %xmm7
#endif
	maxpd	%xmm7, %xmm3

#ifdef OPTERON
	prefetcht0	PREFETCHSIZE * SIZE(X)
#endif

#ifdef PENTIUM4
	prefetchnta	PREFETCHSIZE * SIZE(X)
#endif

	movsd	0 * SIZE(X), %xmm4
	addq	INCX, X
	movhpd	0 * SIZE(X), %xmm4
	addq	INCX, X
#ifdef ABS
	andpd	 %xmm15, %xmm4
#endif
	maxpd	%xmm4, %xmm0

	movsd	0 * SIZE(X), %xmm5
	addq	INCX, X
	movhpd	0 * SIZE(X), %xmm5
	addq	INCX, X
#ifdef ABS
	andpd	 %xmm15, %xmm5
#endif
	maxpd	%xmm5, %xmm1

	movsd	0 * SIZE(X), %xmm6
	addq	INCX, X
	movhpd	0 * SIZE(X), %xmm6
	addq	INCX, X
#ifdef ABS
	andpd	 %xmm15, %xmm6
#endif
	maxpd	%xmm6, %xmm2

	movsd	0 * SIZE(X), %xmm7
	addq	INCX, X
	movhpd	0 * SIZE(X), %xmm7
	addq	INCX, X
#ifdef ABS
	andpd	 %xmm15, %xmm7
#endif
	maxpd	%xmm7, %xmm3

	decq	I
	jg	.L81
	ALIGN_4

.L85:
	andq	$15,  M
	jle	.L90

	testq	$8, M
	je	.L86

	movsd	0 * SIZE(X), %xmm4
	addq	INCX, X
	movhpd	0 * SIZE(X), %xmm4
	addq	INCX, X
#ifdef ABS
	andpd	 %xmm15, %xmm4
#endif
	maxpd	%xmm4, %xmm0

	movsd	0 * SIZE(X), %xmm5
	addq	INCX, X
	movhpd	0 * SIZE(X), %xmm5
	addq	INCX, X
#ifdef ABS
	andpd	 %xmm15, %xmm5
#endif
	maxpd	%xmm5, %xmm1

	movsd	0 * SIZE(X), %xmm6
	addq	INCX, X
	movhpd	0 * SIZE(X), %xmm6
	addq	INCX, X
#ifdef ABS
	andpd	 %xmm15, %xmm6
#endif
	maxpd	%xmm6, %xmm2

	movsd	0 * SIZE(X), %xmm7
	addq	INCX, X
	movhpd	0 * SIZE(X), %xmm7
	addq	INCX, X
#ifdef ABS
	andpd	 %xmm15, %xmm7
#endif
	maxpd	%xmm7, %xmm3
	ALIGN_3

.L86:
	testq	$4, M
	je	.L87

	movsd	0 * SIZE(X), %xmm4
	addq	INCX, X
	movhpd	0 * SIZE(X), %xmm4
	addq	INCX, X
#ifdef ABS
	andpd	 %xmm15, %xmm4
#endif
	maxpd	%xmm4, %xmm0

	movsd	0 * SIZE(X), %xmm5
	addq	INCX, X
	movhpd	0 * SIZE(X), %xmm5
	addq	INCX, X
#ifdef ABS
	andpd	 %xmm15, %xmm5
#endif
	maxpd	%xmm5, %xmm1
	ALIGN_3	

.L87:
	testq	$2, M
	je	.L88

	movsd	0 * SIZE(X), %xmm6
	addq	INCX, X
	movhpd	0 * SIZE(X), %xmm6
	addq	INCX, X
#ifdef ABS
	andpd	 %xmm15, %xmm6
#endif
	maxpd	%xmm6, %xmm2
	ALIGN_3
	
.L88:
	testq	$1, M
	je	.L90

	movsd	0 * SIZE(X), %xmm7
	unpcklpd  %xmm7, %xmm7
#ifdef ABS
	andpd	 %xmm15, %xmm7
#endif
	maxpd	%xmm7, %xmm3

	maxpd	%xmm1, %xmm0
	maxpd	%xmm3, %xmm2
	maxpd	%xmm2, %xmm0
	movapd	%xmm0, %xmm1
	unpckhpd %xmm0, %xmm0
	maxsd	%xmm1, %xmm0
	ALIGN_4

.L90:
	movq	XX, X
	movq	MM, M

	maxpd	 %xmm1, %xmm0
	maxpd	 %xmm3, %xmm2
	maxpd	 %xmm2, %xmm0
	movapd	 %xmm0, %xmm1
	unpckhpd %xmm0, %xmm0
	maxsd	 %xmm1, %xmm0
	unpcklpd %xmm0, %xmm0

	movq	M,  I
	sarq	$3, I
	jle	.L95
	ALIGN_4
	
.L92:
#ifdef OPTERON
	prefetcht0	PREFETCHSIZE * SIZE(X)
#endif

#ifdef PENTIUM4
	prefetchnta	PREFETCHSIZE * SIZE(X)
#endif

	movsd	0 * SIZE(X), %xmm1
	addq	INCX, X
	movhpd	0 * SIZE(X), %xmm1
	addq	INCX, X
#ifdef ABS
	andpd	 %xmm15, %xmm1
#endif
	cmpeqpd	%xmm0, %xmm1

	movsd	0 * SIZE(X), %xmm3
	addq	INCX, X
	movhpd	0 * SIZE(X), %xmm3
	addq	INCX, X
#ifdef ABS
	andpd	 %xmm15, %xmm3
#endif
	cmpeqpd	%xmm0, %xmm3

	movsd	0 * SIZE(X), %xmm5
	addq	INCX, X
	movhpd	0 * SIZE(X), %xmm5
	addq	INCX, X
#ifdef ABS
	andpd	 %xmm15, %xmm5
#endif
	cmpeqpd	%xmm0, %xmm5

	movsd	0 * SIZE(X), %xmm7
	addq	INCX, X
	movhpd	0 * SIZE(X), %xmm7
	addq	INCX, X
#ifdef ABS
	andpd	 %xmm15, %xmm7
#endif
	cmpeqpd	%xmm0, %xmm7

	orpd	%xmm3, %xmm1
	orpd	%xmm7, %xmm5
	orpd	%xmm5, %xmm1
	movmskpd %xmm1, %r11
	testq	 $3, %r11
	jne	 .L93

	addq	$8, RET
	decq	I
	jg	.L92
	jmp	.L95
	ALIGN_4

.L93:
	subq	INCX, X
	movsd	0 * SIZE(X), %xmm8
	subq	INCX, X
	movsd	0 * SIZE(X), %xmm7
	subq	INCX, X
	movsd	0 * SIZE(X), %xmm6
	subq	INCX, X
	movsd	0 * SIZE(X), %xmm5
	subq	INCX, X
	movsd	0 * SIZE(X), %xmm4
	subq	INCX, X
	movsd	0 * SIZE(X), %xmm3
	subq	INCX, X
	movsd	0 * SIZE(X), %xmm2
	subq	INCX, X
	movsd	0 * SIZE(X), %xmm1
#ifdef ABS
	andpd	 %xmm15, %xmm1
	andpd	 %xmm15, %xmm2
	andpd	 %xmm15, %xmm3
	andpd	 %xmm15, %xmm4
	andpd	 %xmm15, %xmm5
	andpd	 %xmm15, %xmm6
	andpd	 %xmm15, %xmm7
	andpd	 %xmm15, %xmm8
#endif

	addq	$8 * SIZE, X

	incq	RET
	comisd	%xmm0, %xmm1
	je	.L999
	incq	RET
	comisd	%xmm0, %xmm2
	je	.L999
	incq	RET
	comisd	%xmm0, %xmm3
	je	.L999
	incq	RET
	comisd	%xmm0, %xmm4
	je	.L999
	incq	RET
	comisd	%xmm0, %xmm5
	je	.L999
	incq	RET
	comisd	%xmm0, %xmm6
	je	.L999
	incq	RET
	comisd	%xmm0, %xmm7
	je	.L999
	incq	RET
	jmp	.L999
	ALIGN_3

.L95:
	testq	$4, M
	je	.L97

	movsd	0 * SIZE(X), %xmm1
	addq	INCX, X
	movsd	0 * SIZE(X), %xmm2
	addq	INCX, X
	movsd	0 * SIZE(X), %xmm3
	addq	INCX, X
	movsd	0 * SIZE(X), %xmm4
	addq	INCX, X
#ifdef ABS
	andpd	 %xmm15, %xmm1
	andpd	 %xmm15, %xmm2
	andpd	 %xmm15, %xmm3
	andpd	 %xmm15, %xmm4
#endif
	incq	RET
	comisd	%xmm0, %xmm1
	je	.L999
	incq	RET
	comisd	%xmm0, %xmm2
	je	.L999
	incq	RET
	comisd	%xmm0, %xmm3
	je	.L999
	incq	RET
	comisd	%xmm0, %xmm4
	je	.L999
	ALIGN_3	

.L97:
	testq	$2, M
	je	.L98

	movsd	0 * SIZE(X), %xmm1
	addq	INCX, X
	movsd	0 * SIZE(X), %xmm2
	addq	INCX, X
#ifdef ABS
	andpd	 %xmm15, %xmm1
	andpd	 %xmm15, %xmm2
#endif
	incq	RET
	comisd	%xmm0, %xmm1
	je	.L999
	incq	RET
	comisd	%xmm0, %xmm2
	je	.L999
	ALIGN_3
	
.L98:
	incq	RET
	ALIGN_3

.L999:
	ret

	EPILOGUE