zaxpy_sse2_core2.s

Optimized GotoBLAS libraries

	addpd	%xmm1, %xmm8
	addpd	%xmm3, %xmm9
	addpd	%xmm5, %xmm10
	addpd	%xmm7, %xmm11

	MOVDDUP( 0 * SIZE, X, %xmm0)
	MOVDDUP( 1 * SIZE, X, %xmm1)
	addq	 INCX, X
	MOVDDUP( 0 * SIZE, X, %xmm2)
	MOVDDUP( 1 * SIZE, X, %xmm3)
	addq	 INCX, X
	MOVDDUP( 0 * SIZE, X, %xmm4)
	MOVDDUP( 1 * SIZE, X, %xmm5)
	addq	 INCX, X
	MOVDDUP( 0 * SIZE, X, %xmm6)
	MOVDDUP( 1 * SIZE, X, %xmm7)
	addq	 INCX, X

	mulpd	%xmm14, %xmm0
	mulpd	%xmm14, %xmm2
	mulpd	%xmm14, %xmm4
	mulpd	%xmm14, %xmm6

	movapd	%xmm8,   (YY)
	addq	 INCY, YY
	movapd	%xmm9,   (YY)
	addq	 INCY, YY
	movapd	%xmm10,  (YY)
	addq	 INCY, YY
	movapd	%xmm11,  (YY)
	addq	 INCY, YY

	movapd	 (Y), %xmm8
	addq	 INCY, Y
	movapd	 (Y), %xmm9
	addq	 INCY, Y
	movapd	 (Y), %xmm10
	addq	 INCY, Y
	movapd	 (Y), %xmm11
	addq	 INCY, Y

	decq	%rax
	jg	.L41
	ALIGN_3

.L42:
	addpd	%xmm0, %xmm8
	mulpd	%xmm15, %xmm1
	addpd	%xmm2, %xmm9
	mulpd	%xmm15, %xmm3
	addpd	%xmm4, %xmm10
	mulpd	%xmm15, %xmm5
	addpd	%xmm6, %xmm11
	mulpd	%xmm15, %xmm7

	addpd	%xmm1, %xmm8
	addpd	%xmm3, %xmm9
	addpd	%xmm5, %xmm10
	addpd	%xmm7, %xmm11

	MOVDDUP( 0 * SIZE, X, %xmm0)
	MOVDDUP( 1 * SIZE, X, %xmm1)
	addq	 INCX, X
	MOVDDUP( 0 * SIZE, X, %xmm2)
	MOVDDUP( 1 * SIZE, X, %xmm3)
	addq	 INCX, X
	MOVDDUP( 0 * SIZE, X, %xmm4)
	MOVDDUP( 1 * SIZE, X, %xmm5)
	addq	 INCX, X
	MOVDDUP( 0 * SIZE, X, %xmm6)
	MOVDDUP( 1 * SIZE, X, %xmm7)
	addq	 INCX, X

	mulpd	%xmm14, %xmm0
	mulpd	%xmm14, %xmm2
	mulpd	%xmm14, %xmm4
	mulpd	%xmm14, %xmm6

	movapd	%xmm8,   (YY)
	addq	 INCY, YY
	movapd	%xmm9,   (YY)
	addq	 INCY, YY
	movapd	%xmm10,  (YY)
	addq	 INCY, YY
	movapd	%xmm11,  (YY)
	addq	 INCY, YY

	movapd	 (Y), %xmm8
	addq	 INCY, Y
	movapd	 (Y), %xmm9
	addq	 INCY, Y
	movapd	 (Y), %xmm10
	addq	 INCY, Y
	movapd	 (Y), %xmm11
	addq	 INCY, Y

	addpd	%xmm0, %xmm8
	mulpd	%xmm15, %xmm1
	addpd	%xmm2, %xmm9
	mulpd	%xmm15, %xmm3
	addpd	%xmm4, %xmm10
	mulpd	%xmm15, %xmm5
	addpd	%xmm6, %xmm11
	mulpd	%xmm15, %xmm7

	addpd	%xmm1, %xmm8
	addpd	%xmm3, %xmm9
	addpd	%xmm5, %xmm10
	addpd	%xmm7, %xmm11

	movapd	%xmm8,   (YY)
	addq	 INCY, YY
	movapd	%xmm9,   (YY)
	addq	 INCY, YY
	movapd	%xmm10,  (YY)
	addq	 INCY, YY
	movapd	%xmm11,  (YY)
	addq	 INCY, YY
	ALIGN_3

.L45:
	movq	M,  %rax
	andq	$4, %rax
	jle	.L46

	MOVDDUP( 0 * SIZE, X, %xmm0)
	MOVDDUP( 1 * SIZE, X, %xmm1)
	addq	INCX, X
	MOVDDUP( 0 * SIZE, X, %xmm2)
	MOVDDUP( 1 * SIZE, X, %xmm3)
	addq	INCX, X

	MOVDDUP( 0 * SIZE, X, %xmm4)
	MOVDDUP( 1 * SIZE, X, %xmm5)
	addq	INCX, X
	MOVDDUP( 0 * SIZE, X, %xmm6)
	MOVDDUP( 1 * SIZE, X, %xmm7)
	addq	INCX, X

	movapd	 (Y), %xmm8
	addq	 INCY, Y
	movapd	 (Y), %xmm9
	addq	 INCY, Y
	movapd	 (Y), %xmm10
	addq	 INCY, Y
	movapd	 (Y), %xmm11
	addq	 INCY, Y

	mulpd	%xmm14, %xmm0
	mulpd	%xmm14, %xmm2
	mulpd	%xmm14, %xmm4
	mulpd	%xmm14, %xmm6

	addpd	%xmm0, %xmm8
	mulpd	%xmm15, %xmm1
	addpd	%xmm2, %xmm9
	mulpd	%xmm15, %xmm3
	addpd	%xmm4, %xmm10
	mulpd	%xmm15, %xmm5
	addpd	%xmm6, %xmm11
	mulpd	%xmm15, %xmm7

	addpd	%xmm1, %xmm8
	addpd	%xmm3, %xmm9
	addpd	%xmm5, %xmm10
	addpd	%xmm7, %xmm11

	movapd	%xmm8,  (YY)
	addq	 INCY, YY
	movapd	%xmm9,  (YY)
	addq	 INCY, YY
	movapd	%xmm10, (YY)
	addq	 INCY, YY
	movapd	%xmm11, (YY)
	addq	 INCY, YY
	ALIGN_3

.L46:
	movq	M,  %rax
	andq	$2, %rax
	jle	.L47

	MOVDDUP( 0 * SIZE, X, %xmm0)
	MOVDDUP( 1 * SIZE, X, %xmm1)
	addq	INCX, X
	MOVDDUP( 0 * SIZE, X, %xmm2)
	MOVDDUP( 1 * SIZE, X, %xmm3)
	addq	INCX, X

	movapd	 (Y), %xmm8
	addq	 INCY, Y
	movapd	 (Y), %xmm9
	addq	 INCY, Y

	mulpd	%xmm14, %xmm0
	mulpd	%xmm14, %xmm2
	mulpd	%xmm15, %xmm1
	mulpd	%xmm15, %xmm3

	addpd	%xmm0, %xmm8
	addpd	%xmm2, %xmm9
	addpd	%xmm1, %xmm8
	addpd	%xmm3, %xmm9

	movapd	%xmm8,   (YY)
	addq	 INCY, YY
	movapd	%xmm9,   (YY)
	addq	 INCY, YY
	ALIGN_3

.L47:
	movq	M,  %rax
	andq	$1, %rax
	jle	.L999

	MOVDDUP( 0 * SIZE, X, %xmm0)
	MOVDDUP( 1 * SIZE, X, %xmm1)

	movapd	 (Y), %xmm8
	mulpd	%xmm14, %xmm0
	mulpd	%xmm15, %xmm1
	addpd	%xmm0, %xmm8
	addpd	%xmm1, %xmm8

	movapd	%xmm8,   (YY)
	jmp	.L999
	ALIGN_3

.L50:
	movq	Y, YY
	movq	M,  %rax
	sarq	$3, %rax
	jle	.L55

	MOVDDUP( 0 * SIZE, X, %xmm0)
	MOVDDUP( 1 * SIZE, X, %xmm1)
	addq	 INCX, X
	MOVDDUP( 0 * SIZE, X, %xmm2)
	MOVDDUP( 1 * SIZE, X, %xmm3)
	addq	 INCX, X
	MOVDDUP( 0 * SIZE, X, %xmm4)
	MOVDDUP( 1 * SIZE, X, %xmm5)
	addq	 INCX, X
	MOVDDUP( 0 * SIZE, X, %xmm6)
	MOVDDUP( 1 * SIZE, X, %xmm7)
	addq	 INCX, X

	movsd	 0 * SIZE(Y), %xmm8
	movhpd	 1 * SIZE(Y), %xmm8
	addq	 INCY, Y
	movsd	 0 * SIZE(Y), %xmm9
	movhpd	 1 * SIZE(Y), %xmm9
	addq	 INCY, Y
	movsd	 0 * SIZE(Y), %xmm10
	movhpd	 1 * SIZE(Y), %xmm10
	addq	 INCY, Y
	movsd	 0 * SIZE(Y), %xmm11
	movhpd	 1 * SIZE(Y), %xmm11
	addq	 INCY, Y

	mulpd	%xmm14, %xmm0
	mulpd	%xmm14, %xmm2
	mulpd	%xmm14, %xmm4
	mulpd	%xmm14, %xmm6

	decq	%rax
	jle	.L52
	ALIGN_3

.L51:
	addpd	%xmm0, %xmm8
	mulpd	%xmm15, %xmm1
	addpd	%xmm2, %xmm9
	mulpd	%xmm15, %xmm3
	addpd	%xmm4, %xmm10
	mulpd	%xmm15, %xmm5
	addpd	%xmm6, %xmm11
	mulpd	%xmm15, %xmm7

	addpd	%xmm1, %xmm8
	addpd	%xmm3, %xmm9
	addpd	%xmm5, %xmm10
	addpd	%xmm7, %xmm11

	MOVDDUP( 0 * SIZE, X, %xmm0)
	MOVDDUP( 1 * SIZE, X, %xmm1)
	addq	 INCX, X
	MOVDDUP( 0 * SIZE, X, %xmm2)
	MOVDDUP( 1 * SIZE, X, %xmm3)
	addq	 INCX, X
	MOVDDUP( 0 * SIZE, X, %xmm4)
	MOVDDUP( 1 * SIZE, X, %xmm5)
	addq	 INCX, X
	MOVDDUP( 0 * SIZE, X, %xmm6)
	MOVDDUP( 1 * SIZE, X, %xmm7)
	addq	 INCX, X

	mulpd	%xmm14, %xmm0
	mulpd	%xmm14, %xmm2
	mulpd	%xmm14, %xmm4
	mulpd	%xmm14, %xmm6

	movsd	%xmm8,   0 * SIZE(YY)
	movhpd	%xmm8,   1 * SIZE(YY)
	addq	 INCY, YY
	movsd	%xmm9,   0 * SIZE(YY)
	movhpd	%xmm9,   1 * SIZE(YY)
	addq	 INCY, YY
	movsd	%xmm10,  0 * SIZE(YY)
	movhpd	%xmm10,  1 * SIZE(YY)
	addq	 INCY, YY
	movsd	%xmm11,  0 * SIZE(YY)
	movhpd	%xmm11,  1 * SIZE(YY)
	addq	 INCY, YY

	movsd	 0 * SIZE(Y), %xmm8
	movhpd	 1 * SIZE(Y), %xmm8
	addq	 INCY, Y
	movsd	 0 * SIZE(Y), %xmm9
	movhpd	 1 * SIZE(Y), %xmm9
	addq	 INCY, Y
	movsd	 0 * SIZE(Y), %xmm10
	movhpd	 1 * SIZE(Y), %xmm10
	addq	 INCY, Y
	movsd	 0 * SIZE(Y), %xmm11
	movhpd	 1 * SIZE(Y), %xmm11
	addq	 INCY, Y

	addpd	%xmm0, %xmm8
	mulpd	%xmm15, %xmm1
	addpd	%xmm2, %xmm9
	mulpd	%xmm15, %xmm3
	addpd	%xmm4, %xmm10
	mulpd	%xmm15, %xmm5
	addpd	%xmm6, %xmm11
	mulpd	%xmm15, %xmm7

	addpd	%xmm1, %xmm8
	addpd	%xmm3, %xmm9
	addpd	%xmm5, %xmm10
	addpd	%xmm7, %xmm11

	MOVDDUP( 0 * SIZE, X, %xmm0)
	MOVDDUP( 1 * SIZE, X, %xmm1)
	addq	 INCX, X
	MOVDDUP( 0 * SIZE, X, %xmm2)
	MOVDDUP( 1 * SIZE, X, %xmm3)
	addq	 INCX, X
	MOVDDUP( 0 * SIZE, X, %xmm4)
	MOVDDUP( 1 * SIZE, X, %xmm5)
	addq	 INCX, X
	MOVDDUP( 0 * SIZE, X, %xmm6)
	MOVDDUP( 1 * SIZE, X, %xmm7)
	addq	 INCX, X

	mulpd	%xmm14, %xmm0
	mulpd	%xmm14, %xmm2
	mulpd	%xmm14, %xmm4
	mulpd	%xmm14, %xmm6

	movsd	%xmm8,   0 * SIZE(YY)
	movhpd	%xmm8,   1 * SIZE(YY)
	addq	 INCY, YY
	movsd	%xmm9,   0 * SIZE(YY)
	movhpd	%xmm9,   1 * SIZE(YY)
	addq	 INCY, YY
	movsd	%xmm10,  0 * SIZE(YY)
	movhpd	%xmm10,  1 * SIZE(YY)
	addq	 INCY, YY
	movsd	%xmm11,  0 * SIZE(YY)
	movhpd	%xmm11,  1 * SIZE(YY)
	addq	 INCY, YY

	movsd	 0 * SIZE(Y), %xmm8
	movhpd	 1 * SIZE(Y), %xmm8
	addq	 INCY, Y
	movsd	 0 * SIZE(Y), %xmm9
	movhpd	 1 * SIZE(Y), %xmm9
	addq	 INCY, Y
	movsd	 0 * SIZE(Y), %xmm10
	movhpd	 1 * SIZE(Y), %xmm10
	addq	 INCY, Y
	movsd	 0 * SIZE(Y), %xmm11
	movhpd	 1 * SIZE(Y), %xmm11
	addq	 INCY, Y

	decq	%rax
	jg	.L51
	ALIGN_3

.L52:
	addpd	%xmm0, %xmm8
	mulpd	%xmm15, %xmm1
	addpd	%xmm2, %xmm9
	mulpd	%xmm15, %xmm3
	addpd	%xmm4, %xmm10
	mulpd	%xmm15, %xmm5
	addpd	%xmm6, %xmm11
	mulpd	%xmm15, %xmm7

	addpd	%xmm1, %xmm8
	addpd	%xmm3, %xmm9
	addpd	%xmm5, %xmm10
	addpd	%xmm7, %xmm11

	MOVDDUP( 0 * SIZE, X, %xmm0)
	MOVDDUP( 1 * SIZE, X, %xmm1)
	addq	 INCX, X
	MOVDDUP( 0 * SIZE, X, %xmm2)
	MOVDDUP( 1 * SIZE, X, %xmm3)
	addq	 INCX, X
	MOVDDUP( 0 * SIZE, X, %xmm4)
	MOVDDUP( 1 * SIZE, X, %xmm5)
	addq	 INCX, X
	MOVDDUP( 0 * SIZE, X, %xmm6)
	MOVDDUP( 1 * SIZE, X, %xmm7)
	addq	 INCX, X

	mulpd	%xmm14, %xmm0
	mulpd	%xmm14, %xmm2
	mulpd	%xmm14, %xmm4
	mulpd	%xmm14, %xmm6

	movsd	%xmm8,   0 * SIZE(YY)
	movhpd	%xmm8,   1 * SIZE(YY)
	addq	 INCY, YY
	movsd	%xmm9,   0 * SIZE(YY)
	movhpd	%xmm9,   1 * SIZE(YY)
	addq	 INCY, YY
	movsd	%xmm10,  0 * SIZE(YY)
	movhpd	%xmm10,  1 * SIZE(YY)
	addq	 INCY, YY
	movsd	%xmm11,  0 * SIZE(YY)
	movhpd	%xmm11,  1 * SIZE(YY)
	addq	 INCY, YY

	movsd	 0 * SIZE(Y), %xmm8
	movhpd	 1 * SIZE(Y), %xmm8
	addq	 INCY, Y
	movsd	 0 * SIZE(Y), %xmm9
	movhpd	 1 * SIZE(Y), %xmm9
	addq	 INCY, Y
	movsd	 0 * SIZE(Y), %xmm10
	movhpd	 1 * SIZE(Y), %xmm10
	addq	 INCY, Y
	movsd	 0 * SIZE(Y), %xmm11
	movhpd	 1 * SIZE(Y), %xmm11
	addq	 INCY, Y

	addpd	%xmm0, %xmm8
	mulpd	%xmm15, %xmm1
	addpd	%xmm2, %xmm9
	mulpd	%xmm15, %xmm3
	addpd	%xmm4, %xmm10
	mulpd	%xmm15, %xmm5
	addpd	%xmm6, %xmm11
	mulpd	%xmm15, %xmm7

	addpd	%xmm1, %xmm8
	addpd	%xmm3, %xmm9
	addpd	%xmm5, %xmm10
	addpd	%xmm7, %xmm11

	movsd	%xmm8,   0 * SIZE(YY)
	movhpd	%xmm8,   1 * SIZE(YY)
	addq	 INCY, YY
	movsd	%xmm9,   0 * SIZE(YY)
	movhpd	%xmm9,   1 * SIZE(YY)
	addq	 INCY, YY
	movsd	%xmm10,  0 * SIZE(YY)
	movhpd	%xmm10,  1 * SIZE(YY)
	addq	 INCY, YY
	movsd	%xmm11,  0 * SIZE(YY)
	movhpd	%xmm11,  1 * SIZE(YY)
	addq	 INCY, YY
	ALIGN_3

.L55:
	movq	M,  %rax
	andq	$4, %rax
	jle	.L56

	MOVDDUP( 0 * SIZE, X, %xmm0)
	MOVDDUP( 1 * SIZE, X, %xmm1)
	addq	INCX, X
	MOVDDUP( 0 * SIZE, X, %xmm2)
	MOVDDUP( 1 * SIZE, X, %xmm3)
	addq	INCX, X

	MOVDDUP( 0 * SIZE, X, %xmm4)
	MOVDDUP( 1 * SIZE, X, %xmm5)
	addq	INCX, X
	MOVDDUP( 0 * SIZE, X, %xmm6)
	MOVDDUP( 1 * SIZE, X, %xmm7)
	addq	INCX, X

	movsd	 0 * SIZE(Y), %xmm8
	movhpd	 1 * SIZE(Y), %xmm8
	addq	 INCY, Y
	movsd	 0 * SIZE(Y), %xmm9
	movhpd	 1 * SIZE(Y), %xmm9
	addq	 INCY, Y
	movsd	 0 * SIZE(Y), %xmm10
	movhpd	 1 * SIZE(Y), %xmm10
	addq	 INCY, Y
	movsd	 0 * SIZE(Y), %xmm11
	movhpd	 1 * SIZE(Y), %xmm11
	addq	 INCY, Y

	mulpd	%xmm14, %xmm0
	mulpd	%xmm14, %xmm2
	mulpd	%xmm14, %xmm4
	mulpd	%xmm14, %xmm6

	addpd	%xmm0, %xmm8
	mulpd	%xmm15, %xmm1
	addpd	%xmm2, %xmm9
	mulpd	%xmm15, %xmm3
	addpd	%xmm4, %xmm10
	mulpd	%xmm15, %xmm5
	addpd	%xmm6, %xmm11
	mulpd	%xmm15, %xmm7

	addpd	%xmm1, %xmm8
	addpd	%xmm3, %xmm9
	addpd	%xmm5, %xmm10
	addpd	%xmm7, %xmm11

	movsd	%xmm8,  0 * SIZE(YY)
	movhpd	%xmm8,  1 * SIZE(YY)
	addq	 INCY, YY
	movsd	%xmm9,  0 * SIZE(YY)
	movhpd	%xmm9,  1 * SIZE(YY)
	addq	 INCY, YY
	movsd	%xmm10, 0 * SIZE(YY)
	movhpd	%xmm10, 1 * SIZE(YY)
	addq	 INCY, YY
	movsd	%xmm11, 0 * SIZE(YY)
	movhpd	%xmm11, 1 * SIZE(YY)
	addq	 INCY, YY
	ALIGN_3

.L56:
	movq	M,  %rax
	andq	$2, %rax
	jle	.L57

	MOVDDUP( 0 * SIZE, X, %xmm0)
	MOVDDUP( 1 * SIZE, X, %xmm1)
	addq	INCX, X
	MOVDDUP( 0 * SIZE, X, %xmm2)
	MOVDDUP( 1 * SIZE, X, %xmm3)
	addq	INCX, X

	movsd	 0 * SIZE(Y), %xmm8
	movhpd	 1 * SIZE(Y), %xmm8
	addq	 INCY, Y
	movsd	 0 * SIZE(Y), %xmm9
	movhpd	 1 * SIZE(Y), %xmm9
	addq	 INCY, Y

	mulpd	%xmm14, %xmm0
	mulpd	%xmm14, %xmm2
	mulpd	%xmm15, %xmm1
	mulpd	%xmm15, %xmm3

	addpd	%xmm0, %xmm8
	addpd	%xmm2, %xmm9
	addpd	%xmm1, %xmm8
	addpd	%xmm3, %xmm9

	movsd	%xmm8,   0 * SIZE(YY)
	movhpd	%xmm8,   1 * SIZE(YY)
	addq	 INCY, YY
	movsd	%xmm9,   0 * SIZE(YY)
	movhpd	%xmm9,   1 * SIZE(YY)
	addq	 INCY, YY
	ALIGN_3

.L57:
	movq	M,  %rax
	andq	$1, %rax
	jle	.L999

	MOVDDUP( 0 * SIZE, X, %xmm0)
	MOVDDUP( 1 * SIZE, X, %xmm1)

	movsd	 0 * SIZE(Y), %xmm8
	movhpd	 1 * SIZE(Y), %xmm8
	mulpd	%xmm14, %xmm0
	mulpd	%xmm15, %xmm1
	addpd	%xmm0, %xmm8
	addpd	%xmm1, %xmm8

	movsd	%xmm8,   0 * SIZE(YY)
	movhpd	%xmm8,   1 * SIZE(YY)
	ALIGN_3

.L999:
	xorq	%rax, %rax

	RESTOREREGISTERS

	ret

	EPILOGUE