zcopy.s

Optimized GotoBLAS libraries

	(p21) add   Y2 = INCY3M1, Y2
	}
	;;
	{ .mmi
	(p20) STFD	[Y1] = f36,  1 * SIZE
	(p20) STFD	[Y2] = f56,  1 * SIZE
	}
	{ .mmi
	(p16) LDFPD	f97, f102  = [X1], 2 * SIZE
	(p16) shladd	X2 = INCX, 3, X2
	}
	;;
	{ .mmi
	(p20) STFD	[Y1] = f41
	(p20) STFD	[Y2] = f61
	(p20) add   Y1 = INCYM1, Y1
	}
	{ .mmb
	(p16) LDFPD	f108, f127  = [X1], 2 * SIZE
	(p20) add   Y2 = INCYM1, Y2
	br.ctop.sptk.few .L121
	}
	;;
	{ .mmi
	(p21) STFD	[Y1] = f47,  1 * SIZE
	(p21) STFD	[Y2] = f67,  1 * SIZE
	}
	;;
	{ .mmi
	(p21) STFD	[Y1] = f52
	(p21) STFD	[Y2] = f72
	(p21) add   Y1 = INCY3M1, Y1
	}
	(p21) add   Y2 = INCY3M1, Y2
	;;
	{ .mmi
	(p21) STFD	[Y1] = f77,  1 * SIZE
	(p21) STFD	[Y2] = f97,  1 * SIZE
	}
	;;
	{ .mmi
	(p21) STFD	[Y1] = f82
	(p21) STFD	[Y2] = f102
	(p21) add   Y1 = INCYM1, Y1
	}
	(p21) add   Y2 = INCYM1, Y2
	;;
	{ .mmi
	(p21) STFD	[Y1] = f87,  1 * SIZE
	(p21) STFD	[Y2] = f107, 1 * SIZE
	}
	;;
	{ .mmi
	(p21) STFD	[Y1] = f92
	(p21) STFD	[Y2] = f113
	(p21) add   Y1 = INCY3M1, Y1
	}
	(p21) add   Y2 = INCY3M1, Y2

	adds	X1 = -SIZE, X1
	;;
	.align 32

.L125:
	{ .mmi
	mov	XX = X1
	nop.m 0
	mov	ar.lc = ARLC
	}
	{ .mmi
	(p12) LDFD	f48 = [X1], 1 * SIZE
	(p12) LDFD	f52 = [X2], 1 * SIZE
	tbit.z	p0, p13 = N, 1
	}
	;;
	{ .mmi
	(p12) LDFD	f49 = [X1], INCXM1
	(p12) LDFD	f53 = [X2], INCXM1
	mov	pr = PR, -65474
	}
	{ .mib
	nop.m 0
	tbit.z	p0, p14 = N, 0
	(p9)  br.ret.sptk.many b0
	}
	;;
	{ .mmi
	(p12) LDFD	f50 = [X1], 1 * SIZE
	(p12) LDFD	f54 = [X2], 1 * SIZE
	(p12) shladd XX = INCX, 2, XX;;
	}
	;;
	{ .mmi
	(p12) LDFD	f51 = [X1], INCX3M1
	(p12) LDFD	f55 = [X2], INCX3M1
	(p13) shladd XX = INCX, 1, XX;;
	}
	;;
	{ .mmi
	(p13) LDFD	f56 = [X1], 1 * SIZE
	(p14) LDFD	f60 = [XX], 1 * SIZE
	}
	;;
	{ .mmi
	(p13) LDFD	f57 = [X1], INCXM1
	(p14) LDFD	f61 = [XX]
	mov	YY = Y1
	}
	;;
	{ .mmi
	(p12) STFD	[Y1] = f48, 1 * SIZE
	(p12) STFD	[Y2] = f52, 1 * SIZE
	}
	{ .mmi
	(p13) LDFD	f58 = [X1], 1 * SIZE
	}
	;;
	{ .mmi
	(p12) STFD	[Y1] = f49
	(p12) STFD	[Y2] = f53
	(p12) add Y1 = INCYM1, Y1
	}
	{ .mmi
	(p13) LDFD	f59 = [X1]
	(p12) add Y2 = INCYM1, Y2
	}
	;;
	{ .mmi
	(p12) STFD	[Y1] = f50, 1 * SIZE
	(p12) STFD	[Y2] = f54, 1 * SIZE
	(p12) shladd YY = INCY, 2, YY;;
	}
	;;
	{ .mmi
	(p12) STFD	[Y1] = f51
	(p12) STFD	[Y2] = f55
	(p13) shladd YY = INCY, 1, YY
	}
	{ .mmi
	(p12) add Y1 = INCY3M1, Y1
	(p12) add Y2 = INCY3M1, Y2
	nop.i 0
	}
	;;
	{ .mmi
	(p13) STFD	[Y1] = f56, 1 * SIZE
	(p14) STFD	[YY] = f60, 1 * SIZE
	}
	;;
	{ .mmi
	(p13) STFD	[Y1] = f57
	(p14) STFD	[YY] = f61
	(p13) add Y1 = INCYM1, Y1
	}
	;;
	{ .mmi
	(p13) STFD	[Y1] = f58, 1 * SIZE
	nop.m 0
	nop.i 0
	}
	;;
	{ .mib
	(p13) STFD	[Y1] = f59
	nop.i 0
	br.ret.sptk.many b0
	}
	;;
	.align 32

.L130:
	{ .mmi
	adds	PREA  = PREFETCH_SIZE * SIZE + 32, X1
#ifndef DOUBLE
	adds	PREB  = PREFETCH_SIZE * SIZE + 72, Y1
#else
	adds	PREB  = PREFETCH_SIZE * SIZE + 56, Y1
#endif
	mov	ar.ec = 5
	}
	{ .mmi
	LDFD	f32  = [X1], 1 * SIZE
	cmp.eq	p16, p0 = r0, r0
	nop.i 0
	}
	;;
#ifndef DOUBLE
.L131:
	{ .mmi
	(p19) STFD	[Y1] = f35,  1 * SIZE
	(p19) STFD	[Y2] = f55,  1 * SIZE
	nop.i 0
	}
	{ .mmi
	(p16) lfetch.nt1 [PREA], INCX8
	(p16) LDFPD	f37, f42  = [X1], 2 * SIZE
	nop.i 0
	}
	;;
	{ .mmi
	(p19) STFD	[Y1] = f40
	(p19) STFD	[Y2] = f60
	(p19) add   Y1 = INCYM1, Y1
	}
	{ .mmi
	(p16) lfetch.excl.nt1 [PREB], INCY8
	(p16) LDFPD	f47, f52  = [X1], 2 * SIZE
	(p19) add   Y2 = INCYM1, Y2
	}
	;;
	{ .mmi
	(p19) STFD	[Y1] = f45,  1 * SIZE
	(p19) STFD	[Y2] = f65,  1 * SIZE
	nop.i 0
	}
	{ .mmi
	(p16) LDFPD	f57, f62  = [X1], 2 * SIZE
	nop.m 0
	nop.i 0
	}
	;;
	{ .mmi
	(p19) STFD	[Y1] = f50
	(p19) STFD	[Y2] = f70
	(p19) add   Y1 = INCY3M1, Y1
	}
	{ .mmi
	(p16) LDFPD	f67, f72  = [X1], 2 * SIZE
	(p19) add   Y2 = INCY3M1, Y2
	nop.i 0
	}
	;;
	{ .mmi
	(p19) STFD	[Y1] = f75,  1 * SIZE
	(p19) STFD	[Y2] = f95,  1 * SIZE
	nop.i 0
	}
	{ .mmi
	(p16) LDFPD	f77, f82  = [X1], 2 * SIZE
	nop.m 0
	nop.i 0
	}
	;;
	{ .mmi
	(p19) STFD	[Y1] = f80
	(p19) STFD	[Y2] = f100
	(p19) add   Y1 = INCYM1, Y1
	}
	{ .mmi
	(p16) LDFPD	f87, f92  = [X1], 2 * SIZE
	(p19) add   Y2 = INCYM1, Y2
	nop.i 0
	}
	;;
	{ .mmi
	(p19) STFD	[Y1] = f85,  1 * SIZE
	(p19) STFD	[Y2] = f105, 1 * SIZE
	nop.i 0
	}
	{ .mmi
	(p16) LDFPD	f97, f102  = [X1], 2 * SIZE
	(p16) shladd	X2 = INCX, 3, X2
	nop.i 0
	}
	;;
	{ .mmi
	(p19) STFD	[Y1] = f90
	(p19) STFD	[Y2] = f111
	(p19) add   Y1 = INCY3M1, Y1
	}
	{ .mmb
	(p16) LDFPD	f108, f127  = [X1], 2 * SIZE
	(p19) add   Y2 = INCY3M1, Y2
	br.ctop.sptk.few .L131
	}
	;;
	{ .mmi
	adds	X1 = -SIZE, X1
	nop.m 0
	nop.i 0
	}
	;;
	.align 32
#else
.L131:
	{ .mmi
	(p20) STFD	[Y1] = f46,  1 * SIZE
	(p20) STFD	[Y2] = f66,  1 * SIZE
	}
	{ .mmi
	(p16) lfetch.nt1 [PREA], INCX8
	(p16) LDFPD	f37, f42  = [X1], 2 * SIZE
	}
	;;
	{ .mmi
	(p20) STFD	[Y1] = f51
	(p20) STFD	[Y2] = f71
	(p20) add   Y1 = INCY3M1, Y1
	}
	{ .mmi
	(p16) lfetch.excl.nt1 [PREB], INCY8
	(p16) LDFPD	f47, f52  = [X1], 2 * SIZE
	(p20) add   Y2 = INCY3M1, Y2
	}
	;;
	{ .mmi
	(p20) STFD	[Y1] = f76,  1 * SIZE
	(p20) STFD	[Y2] = f96,  1 * SIZE
	}
	{ .mmi
	(p16) LDFPD	f57, f62  = [X1], 2 * SIZE
	}
	;;
	{ .mmi
	(p20) STFD	[Y1] = f81
	(p20) STFD	[Y2] = f101
	(p20) add   Y1 = INCYM1, Y1
	}
	{ .mmi
	(p16) LDFPD	f67, f72  = [X1], 2 * SIZE
	(p20) add   Y2 = INCYM1, Y2
	}
	;;
	{ .mmi
	(p20) STFD	[Y1] = f86,  1 * SIZE
	(p20) STFD	[Y2] = f106, 1 * SIZE
	}
	{ .mmi
	(p16) LDFPD	f77, f82  = [X1], 2 * SIZE
	}
	;;
	{ .mmi
	(p20) STFD	[Y1] = f91
	(p20) STFD	[Y2] = f112
	(p20) add   Y1 = INCY3M1, Y1
	}
	{ .mmi
	(p16) LDFPD	f87, f92  = [X1], 2 * SIZE
	(p20) add   Y2 = INCY3M1, Y2
	}
	;;
	{ .mmi
	(p19) STFD	[Y1] = f35,  1 * SIZE
	(p19) STFD	[Y2] = f55,  1 * SIZE
	}
	{ .mmi
	(p16) LDFPD	f97, f102  = [X1], 2 * SIZE
	(p16) shladd	X2 = INCX, 3, X2
	}
	;;
	{ .mmi
	(p19) STFD	[Y1] = f40
	(p19) STFD	[Y2] = f60
	(p19) add   Y1 = INCYM1, Y1
	}
	{ .mmb
	(p16) LDFPD	f108, f127  = [X1], 2 * SIZE
	(p19) add   Y2 = INCYM1, Y2
	br.ctop.sptk.few .L131
	}
	;;
	{ .mmi
	adds	X1 = -SIZE, X1
	nop.m 0
	nop.i 0
	}
	;;
	.align 32
#endif

.L135:
	{ .mmi
	mov	XX = X1
	nop.m 0
	mov	ar.lc = ARLC
	}
	{ .mmi
	(p12) LDFD	f48 = [X1], 1 * SIZE
	(p12) LDFD	f52 = [X2], 1 * SIZE
	tbit.z	p0, p13 = N, 1
	}
	;;
	{ .mmi
	(p12) LDFD	f49 = [X1], INCXM1
	(p12) LDFD	f53 = [X2], INCXM1
	mov	pr = PR, -65474
	}
	{ .mib
	nop.m 0
	tbit.z	p0, p14 = N, 0
	(p9)  br.ret.sptk.many b0
	}
	;;
	{ .mmi
	(p12) LDFD	f50 = [X1], 1 * SIZE
	(p12) LDFD	f54 = [X2], 1 * SIZE
	(p12) shladd XX = INCX, 2, XX;;
	}
	;;
	{ .mmi
	(p12) LDFD	f51 = [X1], INCX3M1
	(p12) LDFD	f55 = [X2], INCX3M1
	(p13) shladd XX = INCX, 1, XX;;
	}
	;;
	{ .mmi
	(p13) LDFD	f56 = [X1], 1 * SIZE
	(p14) LDFD	f60 = [XX], 1 * SIZE
	}
	;;
	{ .mmi
	(p13) LDFD	f57 = [X1], INCXM1
	(p14) LDFD	f61 = [XX]
	mov	YY = Y1
	}
	;;
	{ .mmi
	(p12) STFD	[Y1] = f48, 1 * SIZE
	(p12) STFD	[Y2] = f52, 1 * SIZE
	}
	{ .mmi
	(p13) LDFD	f58 = [X1], 1 * SIZE
	}
	;;
	{ .mmi
	(p12) STFD	[Y1] = f49
	(p12) STFD	[Y2] = f53
	(p12) add Y1 = INCYM1, Y1
	}
	{ .mmi
	(p13) LDFD	f59 = [X1]
	(p12) add Y2 = INCYM1, Y2
	}
	;;
	{ .mmi
	(p12) STFD	[Y1] = f50, 1 * SIZE
	(p12) STFD	[Y2] = f54, 1 * SIZE
	(p12) shladd YY = INCY, 2, YY;;
	}
	;;
	{ .mmi
	(p12) STFD	[Y1] = f51
	(p12) STFD	[Y2] = f55
	(p13) shladd YY = INCY, 1, YY
	}
	{ .mmi
	(p12) add Y1 = INCY3M1, Y1
	(p12) add Y2 = INCY3M1, Y2
	nop.i 0
	}
	;;
	{ .mmi
	(p13) STFD	[Y1] = f56, 1 * SIZE
	(p14) STFD	[YY] = f60, 1 * SIZE
	}
	;;
	{ .mmi
	(p13) STFD	[Y1] = f57
	(p14) STFD	[YY] = f61
	(p13) add Y1 = INCYM1, Y1
	}
	;;
	{ .mmi
	(p13) STFD	[Y1] = f58, 1 * SIZE
	nop.m 0
	nop.i 0
	}
	;;
	{ .mib
	(p13) STFD	[Y1] = f59
	nop.i 0
	br.ret.sptk.many b0
	}
	;;

/* Unaligned Copy INCX =! 1 */
.L200:
	;;
	{ .mmi
	adds	PREA  = PREFETCH_SIZE * SIZE + 32, X1
	adds	PREB  = PREFETCH_SIZE * SIZE + 32, Y1
	mov	ar.ec = 5
	}
	{ .mmi
	cmp.eq	p16, p0 = r0, r0
	nop.m 0
	nop.i 0
	}
	;;
	.align 32

.L221:
	{ .mmi
	(p20) STFD	[Y1] = f91
	(p20) STFD	[Y2] = f111
	(p20) add   Y1 = INCY3M1, Y1
	}
	{ .mmi
	(p16) LDFD	f32  = [X1], 1 * SIZE
	(p16) LDFD	f52  = [X2], 1 * SIZE
	(p20) add   Y2 = INCY3M1, Y2
	}
	;;
	{ .mmi
	(p19) STFD	[Y1] = f35,  1 * SIZE
	(p19) STFD	[Y2] = f55,  1 * SIZE
	}
	{ .mmi
	(p16) LDFD	f37  = [X1], INCXM1
	(p16) LDFD	f57  = [X2], INCXM1
	}
	;;
	{ .mmi
	(p19) STFD	[Y1] = f40
	(p19) STFD	[Y2] = f60
	(p19) add   Y1 = INCYM1, Y1
	}
	{ .mmi
	(p16) LDFD	f42  = [X1], 1 * SIZE
	(p16) LDFD	f62  = [X2], 1 * SIZE
	(p19) add   Y2 = INCYM1, Y2
	}
	;;
	{ .mmi
	(p19) STFD	[Y1] = f45,  1 * SIZE
	(p19) STFD	[Y2] = f65,  1 * SIZE
	}
	{ .mmi
	(p16) LDFD	f47  = [X1], INCX3M1
	(p16) LDFD	f67  = [X2], INCX3M1
	}
	;;
	{ .mmi
	(p19) STFD	[Y1] = f50
	(p19) STFD	[Y2] = f70
	(p19) add   Y1 = INCY3M1, Y1
	}
	{ .mmi
	(p16) LDFD	f72  = [X1], 1 * SIZE
	(p16) LDFD	f92  = [X2], 1 * SIZE
	(p19) add   Y2 = INCY3M1, Y2
	}
	;;
	{ .mmi
	(p19) STFD	[Y1] = f75,  1 * SIZE
	(p19) STFD	[Y2] = f95,  1 * SIZE
	}
	{ .mmi
	(p16) LDFD	f77  = [X1], INCXM1
	(p16) LDFD	f97  = [X2], INCXM1
	}
	;;
	{ .mmi
	(p19) STFD	[Y1] = f80
	(p19) STFD	[Y2] = f100
	(p19) add   Y1 = INCYM1, Y1
	}
	{ .mmi
	(p16) LDFD	f82  = [X1], 1 * SIZE
	(p16) LDFD	f102 = [X2], 1 * SIZE
	(p19) add   Y2 = INCYM1, Y2
	}
	;;
	{ .mmi
	(p19) STFD	[Y1] = f85,  1 * SIZE
	(p19) STFD	[Y2] = f105, 1 * SIZE
	}
	{ .mmb
	(p16) LDFD	f87  = [X1], INCX3M1
	(p16) LDFD	f107 = [X2], INCX3M1
	br.ctop.sptk.few .L221
	}
	;;
	.align 32

.L225:
	{ .mmi
	mov	XX = X1
	nop.m 0
	mov	ar.lc = ARLC
	}
	{ .mmi
	(p12) LDFD	f48 = [X1], 1 * SIZE
	(p12) LDFD	f52 = [X2], 1 * SIZE
	tbit.z	p0, p13 = N, 1
	}
	;;
	{ .mmi
	(p12) LDFD	f49 = [X1], INCXM1
	(p12) LDFD	f53 = [X2], INCXM1
	mov	pr = PR, -65474
	}
	{ .mib
	nop.m 0
	tbit.z	p0, p14 = N, 0
	(p9)  br.ret.sptk.many b0
	}
	;;
	{ .mmi
	(p12) LDFD	f50 = [X1], 1 * SIZE
	(p12) LDFD	f54 = [X2], 1 * SIZE
	(p12) shladd XX = INCX, 2, XX;;
	}
	;;
	{ .mmi
	(p12) LDFD	f51 = [X1], INCX3M1
	(p12) LDFD	f55 = [X2], INCX3M1
	(p13) shladd XX = INCX, 1, XX;;
	}
	;;
	{ .mmi
	(p13) LDFD	f56 = [X1], 1 * SIZE
	(p14) LDFD	f60 = [XX], 1 * SIZE
	}
	;;
	{ .mmi
	(p13) LDFD	f57 = [X1], INCXM1
	(p14) LDFD	f61 = [XX]
	mov	YY = Y1
	}
	;;
	{ .mmi
	(p12) STFD	[Y1] = f48, 1 * SIZE
	(p12) STFD	[Y2] = f52, 1 * SIZE
	}
	{ .mmi
	(p13) LDFD	f58 = [X1], 1 * SIZE
	}
	;;
	{ .mmi
	(p12) STFD	[Y1] = f49
	(p12) STFD	[Y2] = f53
	(p12) add Y1 = INCYM1, Y1
	}
	{ .mmi
	(p13) LDFD	f59 = [X1]
	(p12) add Y2 = INCYM1, Y2
	}
	;;
	{ .mmi
	(p12) STFD	[Y1] = f50, 1 * SIZE
	(p12) STFD	[Y2] = f54, 1 * SIZE
	(p12) shladd YY = INCY, 2, YY;;
	}
	;;
	{ .mmi
	(p12) STFD	[Y1] = f51
	(p12) STFD	[Y2] = f55
	(p13) shladd YY = INCY, 1, YY
	}
	{ .mmi
	(p12) add Y1 = INCY3M1, Y1
	(p12) add Y2 = INCY3M1, Y2
	nop.i 0
	}
	;;
	{ .mmi
	(p13) STFD	[Y1] = f56, 1 * SIZE
	(p14) STFD	[YY] = f60, 1 * SIZE
	}
	;;
	{ .mmi
	(p13) STFD	[Y1] = f57
	(p14) STFD	[YY] = f61
	(p13) add Y1 = INCYM1, Y1
	}
	;;
	{ .mmi
	(p13) STFD	[Y1] = f58, 1 * SIZE
	nop.m 0
	nop.i 0
	}
	;;
	{ .mib
	(p13) STFD	[Y1] = f59
	nop.i 0
	br.ret.sptk.many b0
	}

	EPILOGUE