sgemv_n.s

Optimized GotoBLAS libraries

	(p14) FMA	f104 = f15, f78, f104
	(p14) FMA	f105 = f15, f79, f105
	(p15) FMA	f106 = f15, f87, f106
	;;
	(p13) STFD	[YST1] = f100, 1 * SIZE
	;;
	(p13) STFD	[YST1] = f101, 1 * SIZE
	;;
	(p13) STFD	[YST1] = f102, 1 * SIZE
	;;
	(p13) STFD	[YST1] = f103, 1 * SIZE
	;;
	(p14) STFD	[YST1] = f104, 1 * SIZE
	;;
	(p14) STFD	[YST1] = f105, 1 * SIZE
	;;
	(p15) STFD	[YST1] = f106, 1 * SIZE
	(p6) br.cond.dptk .L11
	;;
	.align 16


.L20:
	{ .mmi
	mov	YLD1 = YY
	mov	YST1 = YY
	tbit.z	p6, p0 = N, 2
	}
	;;
	{ .mib
	mov	AO1 = A
	mov	pr.rot= 0
	(p6) br.cond.dpnt .L30
	}
	;;
	{ .mmi
	LDFD	f8 = [X], INCX
	(p8) LDFD	f106 = [YLD1], 1 * SIZE
	add	AO2 = LDA, A
	}
	;;
	{ .mmi
	LDFD	f9 = [X], INCX
	(p8) LDFD	f80 = [AO1], 1 * SIZE
	shladd	AO4 = LDA, 1, AO2
	}
	;;
	{ .mmi
	LDFD	f10 = [X], INCX
	(p8) LDFD	f81 = [AO2], 1 * SIZE
	shladd	AO3 = LDA, 1, A
	}
	;;
	{ .mmi
	LDFD	f11 = [X], INCX
	(p8) LDFD	f82 = [AO3], 1 * SIZE
	}
	;;
	{ .mfi
	(p8) LDFD	f83 = [AO4], 1 * SIZE
	FMPY	f8  = ALPHA, f8
	adds	PREB   = RPREFETCH * SIZE, YLD1
	}
	{ .mfi
	adds	RPRE1  = RPREFETCH * SIZE, AO1
	FMPY	f9  = ALPHA, f9
	adds	RPRE2  = (RPREFETCH + 8) * SIZE, AO2
	}
	;;
	FMPY	f10 = ALPHA, f10
	shladd	A   = LDA, 2, A
	FMPY	f11 = ALPHA, f11
	;;
	{ .mfi
	adds	RPRE3  = RPREFETCH * SIZE, AO3
	(p8) FMA	f106 = f8,  f80, f106
	mov	ar.ec= 2
	}
	;;
	adds	RPRE4  = (RPREFETCH + 8) * SIZE, AO4
	(p8) FMA	f106 = f9,  f81, f106
	shr	I = MM, 3
	;;
	{ .mmf
	cmp.eq	p6, p0 = 0, I
	cmp.eq	p16, p0 = r0, r0
	(p8) FMA	f106 = f10, f82, f106
	}
	;;
	{ .mfi
	adds	I = -1, I
	(p8) FMA	f106 = f11, f83, f106
	tbit.nz	p13, p0 = MM, 2
	}
	;;
	{ .mib
	(p8) STFD	[YST1] = f106, 1 * SIZE
	mov	ar.lc = I
	(p6) br.cond.dpnt .L25
	}
	;;
	.align 16

.L22:
	{ .mfi
	(p17) LDFPD	f63,  f64  = [AO4], 2 * SIZE
	(p17) FMA	f101 = f8,  f33, f101
	(p16) tbit.nz.unc	p14, p15 = I, 0
	}
	{ .mfi
	(p16) LDFPD	f100, f103 = [YLD1], 2 * SIZE
	(p17) FMA	f104 = f8,  f34, f104
	}
	;;
	{ .mfi
	(p16) LDFPD	f32,  f33  = [AO1], 2 * SIZE
	(p17) FMA	f107 = f8,  f35, f107
	(p16) adds	I = -1, I
	}
	{ .mfi
	(p14) PREFETCH	[RPRE1], 16 * SIZE
	(p17) FMA	f110 = f8,  f36, f110
	}
	;;
	{ .mfi
	(p16) LDFPD	f34,  f35  = [AO1], 2 * SIZE
	(p17) FMA	f113 = f8,  f37, f113
	}
	{ .mfi
	(p16) LDFPD	f106, f109 = [YLD1], 2 * SIZE
	(p17) FMA	f116 = f8,  f38, f116
	}
	;;
	{ .mfi
	(p16) LDFPD	f36,  f37  = [AO1], 2 * SIZE
	(p17) FMA	f119 = f8,  f39, f119
	}
	{ .mfi
	(p16) LDFPD	f112, f115 = [YLD1], 2 * SIZE
	(p17) FMA	f122 = f8,  f40, f122
	}
	;;
	{ .mfi
	(p16) LDFPD	f38,  f39  = [AO1], 2 * SIZE
	(p17) FMA	f101 = f9,  f41, f101
	}
	{ .mfi
	(p16) LDFPD	f118, f121 = [YLD1], 2 * SIZE
	(p17) FMA	f104 = f9,  f42, f104
	}
	;;
	{ .mmf
	(p16) LDFPD	f40,  f41  = [AO2], 2 * SIZE
	(p15) PREFETCH	[RPRE2], 16 * SIZE
	(p17) FMA	f107 = f9,  f43, f107
	}
	{ .mfi
	(p18) STFD	[YST1] = f16, 1 * SIZE
	(p17) FMA	f110 = f9,  f44, f110
	}
	;;
	{ .mfi
	(p16) LDFPD	f42,  f43  = [AO2], 2 * SIZE
	(p17) FMA	f113 = f9,  f45, f113
	}
	{ .mfi
	(p18) STFD	[YST1] = f17, 1 * SIZE
	(p17) FMA	f116 = f9,  f46, f116
	}
	;;
	{ .mfi
	(p16) LDFPD	f44,  f45  = [AO2], 2 * SIZE
	(p17) FMA	f119 = f9,  f47, f119
	}
	{ .mfi
	(p18) STFD	[YST1] = f18, 1 * SIZE
	(p17) FMA	f122 = f9,  f48, f122
	}
	;;
	{ .mfi
	(p16) LDFPD	f46,  f47  = [AO2], 2 * SIZE
	(p17) FMA	f101 = f10, f49, f101
	}
	{ .mfi
	(p14) lfetch.excl.nt2	[PREB],  16 * SIZE
	(p17) FMA	f104 = f10, f50, f104
	}
	;;
	{ .mfi
	(p16) LDFPD	f48,  f49  = [AO3], 2 * SIZE
	(p17) FMA	f107 = f10, f51, f107
	}
	{ .mfi
	(p14) PREFETCH	[RPRE3], 16 * SIZE
	(p17) FMA	f110 = f10, f52, f110
	}
	;;
	{ .mfi
	(p16) LDFPD	f50,  f51  = [AO3], 2 * SIZE
	(p17) FMA	f113 = f10, f53, f113
	}
	{ .mfi
	(p18) STFD	[YST1] = f19, 1 * SIZE
	(p17) FMA	f116 = f10, f54, f116
	}
	;;
	{ .mfi
	(p16) LDFPD	f52,  f53  = [AO3], 2 * SIZE
	(p17) FMA	f119 = f10, f55, f119
	}
	{ .mfi
	(p18) STFD	[YST1] = f20, 1 * SIZE
	(p17) FMA	f122 = f10, f56, f122
	}
	;;
	{ .mfi
	(p16) LDFPD	f54,  f55  = [AO3], 2 * SIZE
	(p17) FMA	f16 = f11, f57, f101
	}
	{ .mfi
	(p15) PREFETCH	[RPRE4], 16 * SIZE
	(p17) FMA	f17 = f11, f58, f104
	}
	;;
	{ .mfi
	(p16) LDFPD	f56,  f57  = [AO4], 2 * SIZE
	(p17) FMA	f18 = f11, f59, f107
	}
	{ .mfi
	(p18) STFD	[YST1] = f21, 1 * SIZE
	(p17) FMA	f19 = f11, f60, f110
	}
	;;
	{ .mfi
	(p16) LDFPD	f58,  f59  = [AO4], 2 * SIZE
	(p17) FMA	f20 = f11, f61, f113
	}
	{ .mfi
	(p18) STFD	[YST1] = f22, 1 * SIZE
	(p17) FMA	f21 = f11, f62, f116
	}
	;;
	{ .mfi
	(p16) LDFPD	f60,  f61  = [AO4], 2 * SIZE
	(p17) FMA	f22 = f11, f63, f119
	}
	{ .mfb
	(p18) STFD	[YST1] = f23, 1 * SIZE
	(p17) FMA	f23 = f11, f64, f122
	br.ctop.sptk.few .L22
	}
	;;
	.align 16

.L25:
	{ .mmi
	(p13) LDFPD	f32,  f33  = [AO1], 2 * SIZE
	(p13) LDFPD	f100, f101 = [YLD1], 2 * SIZE
	tbit.nz	p14, p0 = MM, 1
	}
	{ .mmi
	(p18) STFD	[YST1] = f16, 1 * SIZE
	}
	;;
	{ .mmi
	(p13) LDFPD	f48,  f49  = [AO1], 2 * SIZE
	(p13) LDFPD	f102, f103 = [YLD1], 2 * SIZE
	tbit.nz	p15, p0 = MM, 0
	}
	{ .mmi
	(p18) STFD	[YST1] = f17, 1 * SIZE
	}
	;;
	{ .mmi
	(p14) LDFPD	f64,  f65  = [AO1], 2 * SIZE
	(p14) LDFPD	f104, f105 = [YLD1], 2 * SIZE
	}
	{ .mmi
	(p18) STFD	[YST1] = f18, 1 * SIZE
	}
	;;
	{ .mmi
	(p15) LDFD	f80 = [AO1]
	(p15) LDFD	f106 = [YLD1], 1 * SIZE
	}
	{ .mmi
	(p18) STFD	[YST1] = f19, 1 * SIZE
	}
	;;
	{ .mmi
	(p13) LDFPD	f34,  f35  = [AO2], 2 * SIZE
	(p13) LDFPD	f36,  f37  = [AO3], 2 * SIZE
	}
	{ .mmi
	(p18) STFD	[YST1] = f20, 1 * SIZE
	}
	;;
	{ .mmi
	(p13) LDFPD	f50,  f51  = [AO2], 2 * SIZE
	(p13) LDFPD	f52,  f53  = [AO3], 2 * SIZE
	}
	{ .mmi
	(p18) STFD	[YST1] = f21, 1 * SIZE
	}
	;;
	{ .mmi
	(p14) LDFPD	f66,  f67  = [AO2], 2 * SIZE
	(p14) LDFPD	f68,  f69  = [AO3], 2 * SIZE
	}
	{ .mmi
	(p18) STFD	[YST1] = f22, 1 * SIZE
	}
	;;
	{ .mmf
	(p15) LDFD	f81 = [AO2]
	(p15) LDFD	f82 = [AO3]
	(p13) FMA	f100 = f8,  f32, f100
	}
	{ .mfi
	(p18) STFD	[YST1] = f23, 1 * SIZE
	(p13) FMA	f101 = f8,  f33, f101
	}
	;;
	;;
	{ .mfi
	(p13) LDFPD	f38,  f39  = [AO4], 2 * SIZE
	(p13) FMA	f102 = f8,  f48, f102
	}
	{ .mfi
	(p13) FMA	f103 = f8,  f49, f103
	}
	;;
	{ .mfi
	(p13) LDFPD	f54,  f55  = [AO4], 2 * SIZE
	(p14) FMA	f104 = f8,  f64, f104
	}
	{ .mfi
	(p14) FMA	f105 = f8,  f65, f105
	}
	;;
	{ .mfi
	(p14) LDFPD	f70,  f71  = [AO4], 2 * SIZE
	(p15) FMA	f106 = f8,  f80, f106
	}
	{ .mfi
	(p13) FMA	f100 = f9,  f34, f100
	}
	;;
	{ .mfi
	(p15) LDFD	f83 = [AO4]
	(p13) FMA	f101 = f9,  f35, f101
	}
	{ .mfi
	(p13) FMA	f102 = f9,  f50, f102
	}
	;;
	(p13) FMA	f103 = f9,  f51, f103
	(p14) FMA	f104 = f9,  f66, f104
	(p14) FMA	f105 = f9,  f67, f105
	(p15) FMA	f106 = f9,  f81, f106
	;;
	(p13) FMA	f100 = f10, f36, f100
	(p13) FMA	f101 = f10, f37, f101
	(p13) FMA	f102 = f10, f52, f102
	(p13) FMA	f103 = f10, f53, f103
	(p14) FMA	f104 = f10, f68, f104
	(p14) FMA	f105 = f10, f69, f105
	(p15) FMA	f106 = f10, f82, f106
	;;
	(p13) FMA	f100 = f11, f38, f100
	(p13) FMA	f101 = f11, f39, f101
	;;
	(p13) FMA	f102 = f11, f54, f102
	(p13) STFD	[YST1] = f100, 1 * SIZE
	(p13) FMA	f103 = f11, f55, f103
	;;
	(p13) STFD	[YST1] = f101, 1 * SIZE
	(p14) FMA	f104 = f11, f70, f104
	;;
	(p13) STFD	[YST1] = f102, 1 * SIZE
	(p14) FMA	f105 = f11, f71, f105
	;;
	(p13) STFD	[YST1] = f103, 1 * SIZE
	(p15) FMA	f106 = f11, f83, f106
	;;
	(p14) STFD	[YST1] = f104, 1 * SIZE
	;;
	(p14) STFD	[YST1] = f105, 1 * SIZE
	;;
	(p15) STFD	[YST1] = f106, 1 * SIZE
	;;
	.align 16

.L30:
	{ .mmi
	mov	YLD1 = YY
	mov	YST1 = YY
	tbit.z	p6, p0 = N, 1
	}
	;;
	{ .mib
	mov	AO1 = A
	mov	pr.rot= 0
	(p6) br.cond.dpnt .L40
	}
	;;
	{ .mmi
	LDFD	f8 = [X], INCX
	(p8) LDFD	f106 = [YLD1], 1 * SIZE
	add	AO2 = LDA, A
	}
	;;
	{ .mmi
	LDFD	f9 = [X], INCX
	(p8) LDFD	f80 = [AO1], 1 * SIZE
	shladd	A   = LDA, 1, A
	}
	;;
	adds	PREB   = RPREFETCH * SIZE, YLD1
	FMPY	f8  = ALPHA, f8
	mov	ar.ec= 2
	adds	RPRE1  = RPREFETCH * SIZE, AO1
	FMPY	f9  = ALPHA, f9
	shr	I = MM, 3
	;;
	(p8) LDFD	f81 = [AO2], 1 * SIZE
	cmp.eq	p6, p0 = 0, I
	;;
	(p8) FMA	f106 = f8,  f80, f106
	adds	RPRE2  = (RPREFETCH + 8) * SIZE, AO2
	tbit.nz	p13, p0 = MM, 2
	;;
	(p8) FMA	f106 = f9,  f81, f106
	cmp.eq	p16, p0 = r0, r0
	adds	I = -1, I
	;;
	{ .mib
	(p8) STFD	[YST1] = f106, 1 * SIZE
	mov	ar.lc = I
	(p6) br.cond.dpnt .L35
	}
	;;
	.align 16

.L32:
	{ .mfi
	(p17) LDFPD	f47,  f48  = [AO2], 2 * SIZE
	(p17) FMA	f101 = f8,  f33, f101
	(p16) tbit.nz.unc	p14, p15 = I, 0
	}
	{ .mmf
	(p16) LDFPD	f100, f103 = [YLD1], 2 * SIZE
	(p18) STFD	[YST1] = f16, 1 * SIZE
	(p17) FMA	f104 = f8,  f34, f104
	}
	;;
	{ .mfi
	(p16) LDFPD	f32,  f33  = [AO1], 2 * SIZE
	(p17) FMA	f107 = f8,  f35, f107
	adds	I = -1, I
	}
	{ .mmf
	(p14) PREFETCH	[RPRE1], 16 * SIZE
	(p18) STFD	[YST1] = f17, 1 * SIZE
	(p17) FMA	f110 = f8,  f36, f110
	}
	;;
	{ .mfi
	(p16) LDFPD	f34,  f35  = [AO1], 2 * SIZE
	(p17) FMA	f113 = f8,  f37, f113
	}
	{ .mmf
	(p16) LDFPD	f106, f109 = [YLD1], 2 * SIZE
	(p18) STFD	[YST1] = f18, 1 * SIZE
	(p17) FMA	f116 = f8,  f38, f116
	}
	;;
	{ .mfi
	(p16) LDFPD	f36,  f37  = [AO1], 2 * SIZE
	(p17) FMA	f119 = f8,  f39, f119
	}
	{ .mmf
	(p16) LDFPD	f112, f115 = [YLD1], 2 * SIZE
	(p18) STFD	[YST1] = f19, 1 * SIZE
	(p17) FMA	f122 = f8,  f40, f122
	}
	;;
	{ .mfi
	(p16) LDFPD	f38,  f39  = [AO1], 2 * SIZE
	(p17) FMA	f16 = f9,  f41, f101
	}
	{ .mmf
	(p16) LDFPD	f118, f121 = [YLD1], 2 * SIZE
	(p18) STFD	[YST1] = f20, 1 * SIZE
	(p17) FMA	f17 = f9,  f42, f104
	}
	;;
	{ .mfi
	(p16) LDFPD	f40,  f41  = [AO2], 2 * SIZE
	(p17) FMA	f18 = f9,  f43, f107
	}
	{ .mmf
	(p15) PREFETCH	[RPRE2], 16 * SIZE
	(p18) STFD	[YST1] = f21, 1 * SIZE
	(p17) FMA	f19 = f9,  f44, f110
	}
	;;
	{ .mfi
	(p16) LDFPD	f42,  f43  = [AO2], 2 * SIZE
	(p17) FMA	f20 = f9,  f45, f113
	}
	{ .mmf
	(p14) PREFETCH	[PREB],  16 * SIZE
	(p18) STFD	[YST1] = f22, 1 * SIZE
	(p17) FMA	f21 = f9,  f46, f116
	}
	;;
	{ .mfi
	(p16) LDFPD	f44,  f45  = [AO2], 2 * SIZE
	(p17) FMA	f22 = f9,  f47, f119
	}
	{ .mfb
	(p18) STFD	[YST1] = f23, 1 * SIZE
	(p17) FMA	f23 = f9,  f48, f122
	br.ctop.sptk.few .L32
	}
	;;
	.align 16

.L35:
	{ .mmi
	(p13) LDFPD	f32,  f33  = [AO1], 2 * SIZE
	(p13) LDFPD	f100, f101 = [YLD1], 2 * SIZE
	tbit.nz	p14, p0 = MM, 1
	}
	{ .mmi
	(p18) STFD	[YST1] = f16, 1 * SIZE
	}
	;;
	{ .mmi
	(p13) LDFPD	f48,  f49  = [AO1], 2 * SIZE
	(p13) LDFPD	f102, f103 = [YLD1], 2 * SIZE
	tbit.nz	p15, p0 = MM, 0
	}
	{ .mmi
	(p18) STFD	[YST1] = f17, 1 * SIZE
	}
	;;
	{ .mmi
	(p14) LDFPD	f64,  f65  = [AO1], 2 * SIZE
	(p14) LDFPD	f104, f105 = [YLD1], 2 * SIZE
	}
	{ .mmi
	(p18) STFD	[YST1] = f18, 1 * SIZE
	}
	;;
	{ .mmi
	(p15) LDFD	f80 = [AO1]
	(p15) LDFD	f106 = [YLD1], 1 * SIZE
	}
	{ .mmi
	(p18) STFD	[YST1] = f19, 1 * SIZE
	}
	;;
	{ .mmi
	(p13) LDFPD	f34,  f35  = [AO2], 2 * SIZE
	(p18) STFD	[YST1] = f20, 1 * SIZE
	}
	;;
	{ .mmi
	(p13) LDFPD	f50,  f51  = [AO2], 2 * SIZE
	(p18) STFD	[YST1] = f21, 1 * SIZE
	}
	;;
	{ .mmi
	(p14) LDFPD	f66,  f67  = [AO2], 2 * SIZE
	(p18) STFD	[YST1] = f22, 1 * SIZE
	}
	;;
	{ .mmi
	(p15) LDFD	f81 = [AO2]
	(p18) STFD	[YST1] = f23, 1 * SIZE
	}
	;;
	(p13) FMA	f100 = f8,  f32, f100
	(p13) FMA	f101 = f8,  f33, f101
	(p13) FMA	f102 = f8,  f48, f102
	(p13) FMA	f103 = f8,  f49, f103
	(p14) FMA	f104 = f8,  f64, f104
	(p14) FMA	f105 = f8,  f65, f105
	(p15) FMA	f106 = f8,  f80, f106
	;;
	(p13) FMA	f100 = f9,  f34, f100
	(p13) FMA	f101 = f9,  f35, f101
	(p13) FMA	f102 = f9,  f50, f102
	(p13) FMA	f103 = f9,  f51, f103
	(p14) FMA	f104 = f9,  f66, f104
	(p14) FMA	f105 = f9,  f67, f105
	(p15) FMA	f106 = f9,  f81, f106
	;;
	(p13) STFD	[YST1] = f100, 1 * SIZE
	;;
	(p13) STFD	[YST1] = f101, 1 * SIZE
	;;
	(p13) STFD	[YST1] = f102, 1 * SIZE
	;;
	(p13) STFD	[YST1] = f103, 1 * SIZE
	;;
	(p14) STFD	[YST1] = f104, 1 * SIZE
	;;
	(p14) STFD	[YST1] = f105, 1 * SIZE
	;;
	(p15) STFD	[YST1] = f106, 1 * SIZE
	;;
	.align 16

.L40:
	{ .mmi
	mov	YLD1 = YY
	mov	YST1 = YY
	tbit.z	p6, p0 = N, 0
	}
	;;
	{ .mib
	mov	AO1 = A
	mov	pr.rot= 0
	(p6) br.cond.dpnt .L990
	}
	;;
	{ .mmi
	LDFD	f8 = [X], INCX
	(p8) LDFD	f106 = [YLD1], 1 * SIZE
	adds	RPRE1  = RPREFETCH * SIZE, AO1
	}
	;;
	{ .mii
	(p8) LDFD	f80 = [AO1], 1 * SIZE
	adds	PREB   = RPREFETCH * SIZE, YLD1
	}
	;;
	FMPY	f8  = ALPHA, f8
	shr	I = MM, 3
	;;
	(p8) FMA	f106 = f8,  f80, f106
	mov	ar.ec= 3
	;;
	{ .mmi
	cmp.eq	p6, p0 = 0, I
	cmp.eq	p16, p0 = r0, r0
	tbit.nz	p14, p15 = r0, 0
	}
	;;
	{ .mmi
	adds	YST2 = 4 * SIZE, YST1
	adds	I = -1, I
	tbit.nz	p13, p0 = MM, 2
	}
	;;
	{ .mmi
	(p8) STFD	[YST1] = f106, 1 * SIZE
	(p8) adds	YST2 = 1 * SIZE, YST2
	}
	{ .mib
	mov	ar.lc = I
	(p6) br.cond.dpnt .L145
	}
	;;
	.align 16

.L42:
	{ .mmf
	(p19) STFD	[YST1] = f16, 1 * SIZE
	(p19) STFD	[YST2] = f20, 1 * SIZE
	(p18) FMA	f16 = f8,  f34, f102
	}
	{ .mmf
	(p16) LDFPD	f32,  f35  = [AO1], 2 * SIZE
	(p16) LDFPD	f100, f103 = [YLD1], 2 * SIZE
	(p18) FMA	f20 = f8,  f46, f114
	}
	;;
	{ .mmf
	(p19) STFD	[YST1] = f17, 1 * SIZE
	(p19) STFD	[YST2] = f21, 1 * SIZE
	(p18) FMA	f17 = f8,  f37, f105
	}
	{ .mmf
	(p16) LDFPD	f38,  f41  = [AO1], 2 * SIZE
	(p16) LDFPD	f106, f109 = [YLD1], 2 * SIZE
	(p18) FMA	f21 = f8,  f49, f117
	}
	;;
	{ .mmf
	(p19) STFD	[YST1] = f18, 1 * SIZE
	(p19) STFD	[YST2] = f22, 1 * SIZE
	(p18) FMA	f18 = f8,  f40, f108
	}
	{ .mmf
	(p16) LDFPD	f44,  f47  = [AO1], 2 * SIZE
	(p16) LDFPD	f112, f115 = [YLD1], 2 * SIZE
	(p18) FMA	f22 = f8,  f52, f120
	}
	;;
	{ .mmf
	(p19) STFD	[YST1] = f19, 5 * SIZE
	(p19) STFD	[YST2] = f23, 5 * SIZE
	(p18) FMA	f19 = f8,  f43, f111
	}
	{ .mmf
	(p16) LDFPD	f50,  f53  = [AO1], 2 * SIZE
	(p16) LDFPD	f118, f121 = [YLD1], 2 * SIZE
	(p18) FMA	f23 = f8,  f55, f123
	}
	;;
	{ .mmi
	(p14) PREFETCH	[RPRE1], 16 * SIZE
	(p14) PREFETCH	[PREB],  16 * SIZE
	(p16) tbit.nz.unc	p14, p15 = I, 0
	}
	{ .mib
	nop	__LINE__
	(p16) adds	I = -1, I
	br.ctop.sptk.few .L42
	}
	;;
	.align 16

.L45:
	{ .mmi
	(p19) STFD	[YST1] = f16, 1 * SIZE
	(p19) STFD	[YST2] = f20, 1 * SIZE
	tbit.nz	p14, p0 = MM, 1
	}
	{ .mmi
	(p13) LDFPD	f32,  f33  = [AO1], 2 * SIZE
	(p13) LDFPD	f100, f101 = [YLD1], 2 * SIZE
	}
	;;
	{ .mmi
	(p19) STFD	[YST1] = f17, 1 * SIZE
	(p19) STFD	[YST2] = f21, 1 * SIZE
	tbit.nz	p15, p0 = MM, 0
	}
	{ .mmi
	(p13) LDFPD	f48,  f49  = [AO1], 2 * SIZE
	(p13) LDFPD	f102, f103 = [YLD1], 2 * SIZE
	}
	;;
	{ .mmi
	(p19) STFD	[YST1] = f18, 1 * SIZE
	(p19) STFD	[YST2] = f22, 1 * SIZE
	}
	{ .mmi
	(p14) LDFPD	f64,  f65  = [AO1], 2 * SIZE
	(p14) LDFPD	f104, f105 = [YLD1], 2 * SIZE
	}
	;;
	{ .mmi
	(p19) STFD	[YST1] = f19, 5 * SIZE
	(p19) STFD	[YST2] = f23, 5 * SIZE
	}
	{ .mmi
	(p15) LDFD	f80 = [AO1]
	(p15) LDFD	f106 = [YLD1], 1 * SIZE
	}
	;;
	(p13) FMA	f100 = f8,  f32, f100
	(p13) FMA	f101 = f8,  f33, f101
	(p13) FMA	f102 = f8,  f48, f102
	(p13) FMA	f103 = f8,  f49, f103
	;;
	(p13) STFD	[YST1] = f100, 1 * SIZE
	(p14) FMA	f104 = f8,  f64, f104
	;;
	(p13) STFD	[YST1] = f101, 1 * SIZE
	(p14) FMA	f105 = f8,  f65, f105
	;;
	(p13) STFD	[YST1] = f102, 1 * SIZE
	(p15) FMA	f106 = f8,  f80, f106
	;;
	(p13) STFD	[YST1] = f103, 1 * SIZE
	;;
	(p14) STFD	[YST1] = f104, 1 * SIZE
	;;
	(p14) STFD	[YST1] = f105, 1 * SIZE
	;;
	(p15) STFD	[YST1] = f106, 1 * SIZE
	br    .L990
	;;
	.align 16

.L100:
	shr	J   = N, 3
	;;
	cmp.eq	p6, p0 = r0, J
	(p6) br.cond.dpnt .L120
	;;
	.align 16

.L111:
	mov	YLD1 = YY
	mov	YST1 = YY
	;;
	LDFD	f8 = [X], INCX
	;;
	LDFD	f9 = [X], INCX
	;;