zgemv_n.s

Optimized GotoBLAS libraries

	{ .mfi
	(p16) LDFD	f79  = [AO14], 5 * SIZE
	(p17) FMA	f122 = f21,   f87, f122
	}
	;;
	{ .mfi
	(p17) ADD3	f101 = f21,   f82, f101
	}
	{ .mfi
	(p17) ADD3	f113 = f21,   f86, f113
	}
	;;
	{ .mfi
	(p17) ADD4	f104 = f20,   f82, f104
	}
	{ .mfi
	(p17) ADD4	f116 = f20,   f86, f116
	}
	;;
	{ .mfi
	(p17) ADD3	f107 = f21,   f84, f107
	}
	{ .mfi
	(p17) ADD3	f119 = f21,   f88, f119
	}
	;;
	{ .mfi
	(p12) PREFETCH	[RPRE7], 16 * SIZE
	(p17) ADD4	f110 = f20,   f84, f110
	}
	{ .mfi
	(p17) ADD4	f122 = f20,   f88, f122
	}
	;;
	{ .mfi
	(p16) LDFD	f80  = [AO7], 1 * SIZE
	(p17) FMA	f101 = f22,   f89, f101
	}
	{ .mfi
	(p16) LDFD	f84  = [AO15], 1 * SIZE
	(p17) FMA	f113 = f22,   f93, f113
	}
	;;
	{ .mfi
	(p16) LDFD	f81  = [AO7], 1 * SIZE
	(p17) FMA	f104 = f23,   f89, f104
	}
	{ .mfi
	(p16) LDFD	f85  = [AO15], 1 * SIZE
	(p17) FMA	f116 = f23,   f93, f116
	}
	;;
	{ .mfi
	(p16) LDFD	f82  = [AO7], 1 * SIZE
	(p17) FMA	f107 = f22,   f91, f107
	}
	{ .mfi
	(p16) LDFD	f86  = [AO15], 1 * SIZE
	(p17) FMA	f119 = f22,   f95, f119
	}
	;;
	{ .mfi
	(p16) LDFD	f83  = [AO7], 5 * SIZE
	(p17) FMA	f110 = f23,   f91, f110
	}
	{ .mfi
	(p16) LDFD	f87  = [AO15], 5 * SIZE
	(p17) FMA	f122 = f23,   f95, f122
	}
	;;
	{ .mfi
	(p17) ADD3	f101 = f23,   f90, f101
	}
	{ .mfi
	(p17) ADD3	f113 = f23,   f94, f113
	}
	;;
	{ .mfi
	(p17) ADD4	f104 = f22,   f90, f104
	}
	{ .mfi
	(p17) ADD4	f116 = f22,   f94, f116
	}
	;;
	{ .mfi
	(p17) ADD3	f107 = f23,   f92, f107
	}
	{ .mfi
	(p17) ADD3	f119 = f23,   f96, f119
	}
	;;
	{ .mfi
	(p13) PREFETCH	[RPRE8], 16 * SIZE
	(p17) ADD4	f110 = f22,   f92, f110
	}
	{ .mfb
	(p17) ADD4	f122 = f22,   f96, f122
	br.ctop.sptk.few .L12
	}
	;;
	.align 16

.L15:
	{ .mmi
	(p18) STFD	[YST1] = f102, 1 * SIZE
	(p18) STFD	[YST2] = f114, 1 * SIZE
	tbit.nz	p15, p0 = M, 0
	}
	{ .mmi
	(p14) LDFD	f32  = [AO1], 1 * SIZE
	(p14) LDFD	f80  = [YLD1], 1 * SIZE
	cmp.lt p6, p0 = 1, J
	}
	;;
	{ .mmi
	(p18) STFD	[YST1] = f105, 1 * SIZE
	(p18) STFD	[YST2] = f117, 1 * SIZE
	adds J = -1, J
	}
	{
	(p14) LDFD	f33  = [AO1], 1 * SIZE
	(p14) LDFD	f81  = [YLD1], 1 * SIZE
	and   I = 3, M
	}
	;;
	{ .mmi
	(p18) STFD	[YST1] = f108, 1 * SIZE
	(p18) STFD	[YST2] = f120, 1 * SIZE
	(p6) cmp.eq.unc p7, p0 = I, r0
	}
	{ .mmi
	(p14) LDFD	f34  = [AO1], 1 * SIZE
	(p14) LDFD	f82  = [YLD1], 1 * SIZE
	}
	;;
	{ .mmi
	(p18) STFD	[YST1] = f111, 5 * SIZE
	(p18) STFD	[YST2] = f123, 5 * SIZE
	}
	{ .mmb
	(p14) LDFD	f35  = [AO1], 1 * SIZE
	(p14) LDFD	f83  = [YLD1], 1 * SIZE
	(p7) br.cond.dptk .L11
	}
	;;
	(p15) LDFD	f36  = [AO1], 1 * SIZE
	(p15) LDFD	f84  = [YLD1], 1 * SIZE
	;;
	(p15) LDFD	f37  = [AO1], 1 * SIZE
	(p15) LDFD	f85  = [YLD1], 1 * SIZE
	;;
	(p14) LDFD	f38  = [AO2], 1 * SIZE
	(p14) LDFD	f44  = [AO3], 1 * SIZE
	;;
	(p14) LDFD	f39  = [AO2], 1 * SIZE
	(p14) LDFD	f45  = [AO3], 1 * SIZE
	;;
	(p14) LDFD	f40  = [AO2], 1 * SIZE
	(p14) LDFD	f46  = [AO3], 1 * SIZE
	;;
	(p14) LDFD	f41  = [AO2], 1 * SIZE
	(p14) LDFD	f47  = [AO3], 1 * SIZE
	(p14) FMA	f80 = f8,  f32, f80
	;;
	(p15) LDFD	f42  = [AO2], 1 * SIZE
	(p15) LDFD	f48  = [AO3], 1 * SIZE
	(p14) FMA	f81 = f9,  f32, f81
	;;
	(p15) LDFD	f43  = [AO2], 1 * SIZE
	(p15) LDFD	f49  = [AO3], 1 * SIZE
	(p14) FMA	f82 = f8,  f34, f82
	;;
	(p14) LDFD	f50  = [AO4], 1 * SIZE
	(p14) LDFD	f56  = [AO5], 1 * SIZE
	(p14) FMA	f83 = f9,  f34, f83
	;;
	(p14) LDFD	f51  = [AO4], 1 * SIZE
	(p14) LDFD	f57  = [AO5], 1 * SIZE
	(p15) FMA	f84 = f8,  f36, f84
	;;
	(p14) LDFD	f52  = [AO4], 1 * SIZE
	(p14) LDFD	f58  = [AO5], 1 * SIZE
	(p15) FMA	f85 = f9,  f36, f85
	;;
	(p14) LDFD	f53  = [AO4], 1 * SIZE
	(p14) LDFD	f59  = [AO5], 1 * SIZE
	(p14) ADD3	f80 = f9,  f33, f80
	;;
	(p15) LDFD	f54  = [AO4], 1 * SIZE
	(p15) LDFD	f60  = [AO5], 1 * SIZE
	(p14) ADD4	f81 = f8,  f33, f81
	;;
	(p15) LDFD	f55  = [AO4], 1 * SIZE
	(p15) LDFD	f61  = [AO5], 1 * SIZE
	(p14) ADD3	f82 = f9,  f35, f82
	;;
	(p14) LDFD	f62  = [AO6], 1 * SIZE
	(p14) LDFD	f68  = [AO7], 1 * SIZE
	(p14) ADD4	f83 = f8,  f35, f83
	;;
	(p14) LDFD	f63  = [AO6], 1 * SIZE
	(p14) LDFD	f69  = [AO7], 1 * SIZE
	(p15) ADD3	f84 = f9,  f37, f84
	;;
	(p14) LDFD	f64  = [AO6], 1 * SIZE
	(p14) LDFD	f70  = [AO7], 1 * SIZE
	(p15) ADD4	f85 = f8,  f37, f85
	;;
	(p14) LDFD	f65  = [AO6], 1 * SIZE
	(p14) LDFD	f71  = [AO7], 1 * SIZE
	(p14) FMA	f80 = f10, f38, f80
	;;
	(p15) LDFD	f66  = [AO6], 1 * SIZE
	(p15) LDFD	f72  = [AO7], 1 * SIZE
	(p14) FMA	f81 = f11, f38, f81
	;;
	(p15) LDFD	f67  = [AO6], 1 * SIZE
	(p15) LDFD	f73  = [AO7], 1 * SIZE
	(p14) FMA	f82 = f10, f40, f82
	;;
	(p14) LDFD	f74  = [AO8], 1 * SIZE
	(p14) FMA	f83 = f11, f40, f83
	;;
	(p14) LDFD	f75  = [AO8], 1 * SIZE
	(p15) FMA	f84 = f10, f42, f84
	;;
	(p14) LDFD	f76  = [AO8], 1 * SIZE
	(p15) FMA	f85 = f11, f42, f85
	;;
	(p14) LDFD	f77  = [AO8], 1 * SIZE
	(p14) ADD3	f80 = f11, f39, f80
	;;
	(p15) LDFD	f78  = [AO8], 1 * SIZE
	(p14) ADD4	f81 = f10, f39, f81
	;;
	(p15) LDFD	f79  = [AO8], 1 * SIZE
	(p14) ADD3	f82 = f11, f41, f82
	(p14) ADD4	f83 = f10, f41, f83
	(p15) ADD3	f84 = f11, f43, f84
	(p15) ADD4	f85 = f10, f43, f85
	;;
	(p14) FMA	f80 = f12, f44, f80
	(p14) FMA	f81 = f13, f44, f81
	(p14) FMA	f82 = f12, f46, f82
	(p14) FMA	f83 = f13, f46, f83
	(p15) FMA	f84 = f12, f48, f84
	(p15) FMA	f85 = f13, f48, f85
	;;
	(p14) ADD3	f80 = f13, f45, f80
	(p14) ADD4	f81 = f12, f45, f81
	(p14) ADD3	f82 = f13, f47, f82
	(p14) ADD4	f83 = f12, f47, f83
	(p15) ADD3	f84 = f13, f49, f84
	(p15) ADD4	f85 = f12, f49, f85
	;;
	(p14) FMA	f80 = f14, f50, f80
	(p14) FMA	f81 = f15, f50, f81
	(p14) FMA	f82 = f14, f52, f82
	(p14) FMA	f83 = f15, f52, f83
	(p15) FMA	f84 = f14, f54, f84
	(p15) FMA	f85 = f15, f54, f85
	;;
	(p14) ADD3	f80 = f15, f51, f80
	(p14) ADD4	f81 = f14, f51, f81
	(p14) ADD3	f82 = f15, f53, f82
	(p14) ADD4	f83 = f14, f53, f83
	(p15) ADD3	f84 = f15, f55, f84
	(p15) ADD4	f85 = f14, f55, f85
	;;
	(p14) FMA	f80 = f16, f56, f80
	(p14) FMA	f81 = f17, f56, f81
	(p14) FMA	f82 = f16, f58, f82
	(p14) FMA	f83 = f17, f58, f83
	(p15) FMA	f84 = f16, f60, f84
	(p15) FMA	f85 = f17, f60, f85
	;;
	(p14) ADD3	f80 = f17, f57, f80
	(p14) ADD4	f81 = f16, f57, f81
	(p14) ADD3	f82 = f17, f59, f82
	(p14) ADD4	f83 = f16, f59, f83
	(p15) ADD3	f84 = f17, f61, f84
	(p15) ADD4	f85 = f16, f61, f85
	;;
	(p14) FMA	f80 = f18, f62, f80
	(p14) FMA	f81 = f19, f62, f81
	(p14) FMA	f82 = f18, f64, f82
	(p14) FMA	f83 = f19, f64, f83
	(p15) FMA	f84 = f18, f66, f84
	(p15) FMA	f85 = f19, f66, f85
	;;
	(p14) ADD3	f80 = f19, f63, f80
	(p14) ADD4	f81 = f18, f63, f81
	(p14) ADD3	f82 = f19, f65, f82
	(p14) ADD4	f83 = f18, f65, f83
	(p15) ADD3	f84 = f19, f67, f84
	(p15) ADD4	f85 = f18, f67, f85
	;;
	(p14) FMA	f80 = f20, f68, f80
	(p14) FMA	f81 = f21, f68, f81
	(p14) FMA	f82 = f20, f70, f82
	(p14) FMA	f83 = f21, f70, f83
	(p15) FMA	f84 = f20, f72, f84
	(p15) FMA	f85 = f21, f72, f85
	;;
	(p14) ADD3	f80 = f21, f69, f80
	(p14) ADD4	f81 = f20, f69, f81
	(p14) ADD3	f82 = f21, f71, f82
	(p14) ADD4	f83 = f20, f71, f83
	(p15) ADD3	f84 = f21, f73, f84
	(p15) ADD4	f85 = f20, f73, f85
	;;
	(p14) FMA	f80 = f22, f74, f80
	(p14) FMA	f81 = f23, f74, f81
	(p14) FMA	f82 = f22, f76, f82
	(p14) FMA	f83 = f23, f76, f83
	(p15) FMA	f84 = f22, f78, f84
	(p15) FMA	f85 = f23, f78, f85
	;;
	(p14) ADD3	f80 = f23,  f75, f80
	(p14) ADD4	f81 = f22,  f75, f81
	(p14) ADD3	f82 = f23,  f77, f82
	(p14) ADD4	f83 = f22,  f77, f83
	(p15) ADD3	f84 = f23,  f79, f84
	(p15) ADD4	f85 = f22,  f79, f85
	;;
	(p14) STFD	[YST1] = f80, 1 * SIZE
	;;
	(p14) STFD	[YST1] = f81, 1 * SIZE
	;;
	(p14) STFD	[YST1] = f82, 1 * SIZE
	;;
	(p14) STFD	[YST1] = f83, 1 * SIZE
	;;
	(p15) STFD	[YST1] = f84, 1 * SIZE
	;;
	(p15) STFD	[YST1] = f85, 1 * SIZE
	(p6) br.cond.dptk .L11
	;;

.L20:
	{ .mmi
	mov	YLD1 = YY
	adds	YLD2 = 4 * SIZE, YY
	tbit.z	p6, p0 = N, 2
	}
	;;
	{ .mmb
	mov	YST1 = YY
	adds	YST2 = 4 * SIZE, YY
	(p6) br.cond.dpnt .L30
	}
	;;
	LDFD	f32 = [X], SIZE
	LDFD	f36 = [XX], SIZE
	mov	AO1 = A
	;;
	LDFD	f33 = [X], INCXM1
	LDFD	f37 = [XX], INCXM1
	add	AO2 = LDA, A
	;;
	LDFD	f34 = [X], SIZE
	LDFD	f38 = [XX], SIZE
	shladd	AO3 = LDA, 1, A
	;;
	LDFD	f35 = [X], INCX3M1
	LDFD	f39 = [XX], INCX3M1
	shladd	AO4 = LDA, 1, AO2
	;;
	shladd	A   = LDA, 2, A
	FMPY	f8  = ALPHA_R, f32
	adds	AO9  = 4 * SIZE, AO1
	FMPY	f9  = ALPHA_I, f32
	adds	AO10 = 4 * SIZE, AO2
	FMPY	f10 = ALPHA_R, f34
	adds	AO11 = 4 * SIZE, AO3
	FMPY	f11 = ALPHA_I, f34
	adds	AO12 = 4 * SIZE, AO4
	FMPY	f12 = ALPHA_R, f36
	mov	pr.rot= 0
	FMPY	f13 = ALPHA_I, f36
	shr	I = M, 2
	FMPY	f14 = ALPHA_R, f38
	tbit.nz	p14, p0 = M, 1
	FMPY	f15 = ALPHA_I, f38
	;;
	{ .mfi
	cmp.eq	p6, p0 = 0, I
	ADD1	f8  = ALPHA_I, f33, f8
	mov	ar.ec= 2
	}
	ADD2	f9  = ALPHA_R, f33, f9
	adds	I = -1, I
	ADD1	f10 = ALPHA_I, f35, f10
	adds	PREB   = RPREFETCH * SIZE, YLD1
	ADD2	f11 = ALPHA_R, f35, f11
	adds	RPRE1  = RPREFETCH * SIZE, AO1
	ADD1	f12 = ALPHA_I, f37, f12
	adds	RPRE2  = (RPREFETCH + 8) * SIZE, AO2
	ADD2	f13 = ALPHA_R, f37, f13
	adds	RPRE3  = RPREFETCH * SIZE, AO3
	ADD1	f14 = ALPHA_I, f39, f14
	ADD2	f15 = ALPHA_R, f39, f15
	;;
	{ .mib
	cmp.eq	p16, p0 = r0, r0
	mov	ar.lc = I
	(p6) br.cond.dpnt .L25
	}
	;;
	.align 16

.L22:
	{ .mfi
	(p17) LDFD	f57  = [AO4], 1 * SIZE
	(p17) FMA	f101 = f8,   f33, f101
	(p16) tbit.nz.unc	p12, p13 = I, 0
	}
	{ .mfi
	(p17) LDFD	f61  = [AO12], 1 * SIZE
	(p17) FMA	f113 = f8,   f37, f113
	}
	;;
	{ .mfi
	(p17) LDFD	f58  = [AO4], 1 * SIZE
	(p17) FMA	f104 = f9,   f33, f104
	(p16) adds	I = -1, I
	}
	{ .mfi
	(p17) LDFD	f62  = [AO12], 1 * SIZE
	(p17) FMA	f116 = f9,   f37, f116
	}
	;;
	{ .mfi
	(p17) LDFD	f59  = [AO4], 1 * SIZE
	(p17) FMA	f107 = f8,   f35, f107
	}
	{ .mfi
	(p17) LDFD	f63  = [AO12], 1 * SIZE
	(p17) FMA	f119 = f8,   f39, f119
	}
	;;
	{ .mfi
	(p17) LDFD	f60  = [AO4], 5 * SIZE
	(p17) FMA	f110 = f9,   f35, f110
	}
	{ .mfi
	(p17) LDFD	f64  = [AO12], 5 * SIZE
	(p17) FMA	f122 = f9,   f39, f122
	}
	;;
	{ .mfi
	(p12) lfetch.excl.nt2	[PREB],  16 * SIZE
	(p17) ADD3	f101 = f9,   f34, f101
	}
	{ .mfi
	(p17) ADD3	f113 = f9,   f38, f113
	}
	;;
	{ .mfi
	(p16) LDFD	f100 = [YLD1], 1 * SIZE
	(p17) ADD4	f104 = f8,   f34, f104
	}
	{ .mfi
	(p16) LDFD	f112 = [YLD2], 1 * SIZE
	(p17) ADD4	f116 = f8,   f38, f116
	}
	;;
	{ .mfi
	(p16) LDFD	f103 = [YLD1], 1 * SIZE
	(p17) ADD3	f107 = f9,   f36, f107
	}
	{ .mfi
	(p16) LDFD	f115 = [YLD2], 1 * SIZE
	(p17) ADD3	f119 = f9,   f40, f119
	}
	;;
	{ .mfi
	(p12) PREFETCH	[RPRE1], 16 * SIZE
	(p17) ADD4	f110 = f8,   f36, f110
	}
	{ .mfi
	(p17) ADD4	f122 = f8,   f40, f122
	}
	;;
	{ .mfi
	(p16) LDFD	f32  = [AO1], 1 * SIZE
	(p17) FMA	f101 = f10,   f41, f101
	}
	{ .mfi
	(p16) LDFD	f36  = [AO9], 1 * SIZE
	(p17) FMA	f113 = f10,   f45, f113
	}
	;;
	{ .mfi
	(p16) LDFD	f33  = [AO1], 1 * SIZE
	(p17) FMA	f104 = f11,   f41, f104
	}
	{ .mfi
	(p16) LDFD	f37  = [AO9], 1 * SIZE
	(p17) FMA	f116 = f11,   f45, f116
	}
	;;
	{ .mfi
	(p16) LDFD	f34  = [AO1], 1 * SIZE
	(p17) FMA	f107 = f10,   f43, f107
	}
	{ .mfi
	(p16) LDFD	f38  = [AO9], 1 * SIZE
	(p17) FMA	f119 = f10,   f47, f119
	}
	;;
	{ .mfi
	(p16) LDFD	f35  = [AO1], 5 * SIZE
	(p17) FMA	f110 = f11,   f43, f110
	}
	{ .mfi
	(p16) LDFD	f39  = [AO9], 5 * SIZE
	(p17) FMA	f122 = f11,   f47, f122
	}
	;;
	{ .mfi
	(p16) LDFD	f106 = [YLD1], 1 * SIZE
	(p17) ADD3	f101 = f11,   f42, f101
	}
	{ .mfi
	(p16) LDFD	f118 = [YLD2], 1 * SIZE
	(p17) ADD3	f113 = f11,   f46, f113
	}
	;;
	{ .mfi
	(p16) LDFD	f109 = [YLD1], 5 * SIZE
	(p17) ADD4	f104 = f10,   f42, f104
	}
	{ .mfi
	(p16) LDFD	f121 = [YLD2], 5 * SIZE
	(p17) ADD4	f116 = f10,   f46, f116
	}
	;;
	{ .mfi
	(p17) ADD3	f107 = f11,   f44, f107
	}
	{ .mfi
	(p17) ADD3	f119 = f11,   f48, f119
	}
	;;
	{ .mfi
	(p13) PREFETCH	[RPRE2], 16 * SIZE
	(p17) ADD4	f110 = f10,   f44, f110
	}
	{ .mfi
	(p17) ADD4	f122 = f10,   f48, f122
	}
	;;
	{ .mfi
	(p16) LDFD	f40  = [AO2], 1 * SIZE
	(p17) FMA	f101 = f12,   f49, f101
	}
	{ .mfi
	(p16) LDFD	f44  = [AO10], 1 * SIZE
	(p17) FMA	f113 = f12,   f53, f113
	}
	;;
	{ .mfi
	(p16) LDFD	f41  = [AO2], 1 * SIZE
	(p17) FMA	f104 = f13,   f49, f104
	}
	{ .mfi
	(p16) LDFD	f45  = [AO10], 1 * SIZE
	(p17) FMA	f116 = f13,   f53, f116
	}
	;;
	{ .mmf
	(p18) STFD	[YST1] = f102, 1 * SIZE
	(p18) STFD	[YST2] = f114, 1 * SIZE
	(p17) FMA	f107 = f12,   f51, f107
	}
	{ .mmf
	(p16) LDFD	f42  = [AO2], 1 * SIZE
	(p16) LDFD	f46  = [AO10], 1 * SIZE
	(p17) FMA	f119 = f12,   f55, f119
	}
	;;
	{ .mmf
	(p18) STFD	[YST1] = f105, 1 * SIZE
	(p18) STFD	[YST2] = f117, 1 * SIZE
	(p17) FMA	f110 = f13,   f51, f110
	}
	{ .mmf
	(p16) LDFD	f43  = [AO2], 5 * SIZE
	(p16) LDFD	f47  = [AO10], 5 * SIZE
	(p17) FMA	f122 = f13,   f55, f122
	}
	;;
	{ .mfi
	(p17) ADD3	f101 = f13,   f50, f101
	}
	{ .mfi
	(p17) ADD3	f113 = f13,   f54, f113
	}
	;;
	{ .mfi
	(p17) ADD4	f104 = f12,   f50, f104
	}
	{ .mfi
	(p17) ADD4	f116 = f12,   f54, f116
	}
	;;
	{ .mfi
	(p17) ADD3	f107 = f13,   f52, f107
	}
	{ .mfi
	(p17) ADD3	f119 = f13,   f56, f119
	}
	;;
	{ .mfi
	(p12) PREFETCH	[RPRE3], 16 * SIZE
	(p17) ADD4	f110 = f12,   f52, f110
	}
	{ .mfi
	(p17) ADD4	f122 = f12,   f56, f122
	}
	;;
	{ .mfi
	(p16) LDFD	f48  = [AO3], 1 * SIZE
	(p17) FMA	f101 = f14,   f57, f101
	}
	{ .mfi
	(p16) LDFD	f52  = [AO11], 1 * SIZE
	(p17) FMA	f113 = f14,   f61, f113
	}
	;;
	{ .mfi
	(p16) LDFD	f49  = [AO3], 1 * SIZE
	(p17) FMA	f104 = f15,   f57, f104
	}
	{ .mfi
	(p16) LDFD	f53  = [AO11], 1 * SIZE
	(p17) FMA	f116 = f15,   f61, f116
	}
	;;
	{ .mmf
	(p18) STFD	[YST1] = f108, 1 * SIZE
	(p18) STFD	[YST2] = f120, 1 * SIZE
	(p17) FMA	f107 = f14,   f59, f107
	}
	{ .mmf
	(p16) LDFD	f50  = [AO3], 1 * SIZE
	(p16) LDFD	f54  = [AO11], 1 * SIZE
	(p17) FMA	f119 = f14,   f63, f119
	}
	;;
	{ .mmf
	(p18) STFD	[YST1] = f111, 5 * SIZE
	(p18) STFD	[YST2] = f123, 5 * SIZE
	(p17) FMA	f110 = f15,   f59, f110
	}
	{ .mmf
	(p16) LDFD	f51  = [AO3], 5 * SIZE
	(p16) LDFD	f55  = [AO11], 5 * SIZE
	(p17) FMA	f122 = f15,   f63, f122
	}
	;;
	{ .mfi
	(p17) ADD3	f101 = f15,   f58, f101
	}
	{ .mfi
	(p17) ADD3	f113 = f15,   f62, f113
	}
	;;
	{ .mfi
	(p17) ADD4	f104 = f14,   f58, f104
	}
	{ .mfi
	(p17) ADD4	f116 = f14,   f62, f116
	}
	;;
	{ .mfi
	(p17) ADD3	f107 = f15,   f60, f107
	}
	{ .mfi
	(p17) ADD3	f119 = f15,   f64, f119
	}
	;;
	{ .mfi
	(p13) PREFETCH	[RPRE4], 16 * SIZE
	(p17) ADD4	f110 = f14,   f60, f110
	}
	{ .mfb
	(p17) ADD4	f122 = f14,   f64, f122
	br.ctop.sptk.few .L22
	}
	;;
	.align 16

.L25:
	{ .mmi
	(p18) STFD	[YST1] = f102, 1 * SIZE
	(p18) STFD	[YST2] = f114, 1 * SIZE
	tbit.nz	p15, p0 = M, 0
	}
	{ .mmi
	(p14) LDFD	f32  = [AO1], 1 * SIZE
	(p14) LDFD	f80  = [YLD1], 1 * SIZE
	}
	;;
	{ .mmi
	(p18) STFD	[YST1] = f105, 1 * SIZE
	(p18) STFD	[YST2] = f117, 1 * SIZE
	}
	{ .mmi
	(p14) LDFD	f33  = [AO1], 1 * SIZE
	(p14) LDFD	f81  = [YLD1], 1 * SIZE
	}
	;;
	{ .mmi
	(p18) STFD	[YST1] = f108, 1 * SIZE
	(p18) STFD	[YST2] = f120, 1 * SIZE
	}
	{ .mmi
	(p14) LDFD	f34  = [AO1], 1 * SIZE
	(p14) LDFD	f82  = [YLD1], 1 * SIZE
	}
	;;
	{ .mmi
	(p18) STFD	[YST1] = f111, 5 * SIZE
	(p18) STFD	[YST2] = f123, 5 * SIZE
	}
	{ .mmi
	(p14) LDFD	f35  = [AO1], 1 * SIZE
	(p14) LDFD	f83  = [YLD1], 1 * SIZE
	}
	;;
	(p15) LDFD	f36  = [AO1], 1 * SIZE
	(p15) LDFD	f84  = [YLD1], 1 * SIZE
	;;
	(p15) LDFD	f37  = [AO1], 1 * SIZE
	(p15) LDFD	f85  = [YLD1], 1 * SIZE
	;;
	(p14) LDFD	f38  = [AO2], 1 * SIZE
	(p14) LDFD	f44  = [AO3], 1 * SIZE
	;;

	(p14) LDFD	f39  = [AO2], 1 * SIZE
	(p14) LDFD	f45  = [AO3], 1 * SIZE
	(p14) FMA	f80 = f8,  f32, f80
	;;
	(p14) LDFD	f40  = [AO2], 1 * SIZE
	(p14) LDFD	f46  = [AO3], 1 * SIZE
	(p14) FMA	f81 = f9,  f32, f81
	;;
	(p14) LDFD	f41  = [AO2], 1 * SIZE
	(p14) LDFD	f47  = [AO3], 1 * SIZE
	(p14) FMA	f82 = f8,  f34, f82
	;;
	(p15) LDFD	f42  = [AO2], 1 * SIZE
	(p15) LDFD	f48  = [AO3], 1 * SIZE
	(p14) FMA	f83 = f9,  f34, f83
	;;
	(p15) LDFD	f43  = [AO2], 1 * SIZE