idct_ia64_gcc.s

这是一个压缩解压包,用C语言进行编程的,里面有详细的源代码.

	fpma    f61 = f86, c14, f55
	fpnma   f86 = f86, c14, f55
	;;
	;;
	// (f55, f62) = (f49, f62) $ (c13, c13), (line 1, 5)
	fpma    f55 = f62, c13, f49
	fpnma   f62 = f62, c13, f49
	;;
	// (f49, f60) = (f52, f60) $ (c13, c13), (line 2, 6)
	fpma    f49 = f60, c13, f52
	fpnma   f60 = f60, c13, f52
	;;

	// before post shuffle
	//  50 53 68 76 
	//  59 55 88 92 
	//  51 49 69 77 
	//  56 61 89 93 
	//  82 54 70 78 
	//  48 62 90 94 
	//  58 60 71 79 
	//  57 86 91 95 

	// after post shuffle
	//  50 53 68 76 
	//  59 55 88 92 
	//  51 60 69 77 
	//  56 61 89 93 
	//  82 86 70 78 
	//  48 49 90 94 
	//  58 62 71 79 
	//  57 54 91 95 

	// before pre shuffle
	//  50 53 68 76 
	//  59 55 88 92 
	//  51 60 69 77 
	//  56 61 89 93 
	//  82 86 70 78 
	//  48 49 90 94 
	//  58 62 71 79 
	//  57 54 91 95 

	// after pre shuffle
	//  50 53 68 76 
	//  59 55 70 92 
	//  51 60 69 77 
	//  56 61 71 93 
	//  82 86 88 78 
	//  48 49 91 94 
	//  58 62 89 79 
	//  57 54 90 95 
	// (f52, f70) = (f68, f70) $ (c0, c0), (line 0, 1)
	fpma    f52 = f70, c0, f68
	fpnma   f70 = f70, c0, f68
	;;
	// (f63, f71) = (f69, f71) $ (c1, c2), (line 2, 3)
	fpma    f63 = f71, c1, f69
	fpnma   f71 = f71, c2, f69
	;;
	// (f64, f91) = (f88, f91) $ (c3, c4), (line 4, 5)
	fpma    f64 = f91, c3, f88
	fpnma   f91 = f91, c4, f88
	;;
	// (f65, f90) = (f89, f90) $ (c5, c6), (line 6, 7)
	fpma    f65 = f90, c5, f89
	fpnma   f90 = f90, c6, f89
	;;
	;;
	// (f66, f71) = (f52, f71) $ (c7, c7), (line 0, 3)
	fpma    f66 = f71, c7, f52
	fpnma   f71 = f71, c7, f52
	;;
	// (f52, f63) = (f70, f63) $ (c8, c8), (line 1, 2)
	fpma    f52 = f63, c8, f70
	fpnma   f63 = f63, c8, f70
	;;
	// (f67, f65) = (f64, f65) $ (c9, c9), (line 4, 6)
	fpma    f67 = f65, c9, f64
	fpnma   f65 = f65, c9, f64
	;;
	// (f64, f90) = (f91, f90) $ (c10, c10), (line 5, 7)
	fpma    f64 = f90, c10, f91
	fpnma   f90 = f90, c10, f91
	;;
	;;
	// (f68, f65) = (f64, f65) $ (c11, c11), (line 5, 6)
	fpma    f68 = f65, c11, f64
	fpnma   f65 = f65, c11, f64
	;;
	// (f64, f67) = (f66, f67) $ (c12, c12), (line 0, 4)
	fpma    f64 = f67, c12, f66
	fpnma   f67 = f67, c12, f66
	;;
	// (f66, f90) = (f71, f90) $ (c14, c14), (line 3, 7)
	fpma    f66 = f90, c14, f71
	fpnma   f90 = f90, c14, f71
	;;
	;;
	// (f69, f68) = (f52, f68) $ (c13, c13), (line 1, 5)
	fpma    f69 = f68, c13, f52
	fpnma   f68 = f68, c13, f52
	;;
	// (f52, f65) = (f63, f65) $ (c13, c13), (line 2, 6)
	fpma    f52 = f65, c13, f63
	fpnma   f65 = f65, c13, f63
	;;

	// before post shuffle
	//  50 53 64 76 
	//  59 55 69 92 
	//  51 60 52 77 
	//  56 61 66 93 
	//  82 86 67 78 
	//  48 49 68 94 
	//  58 62 65 79 
	//  57 54 90 95 

	// after post shuffle
	//  50 53 64 76 
	//  59 55 69 92 
	//  51 60 65 77 
	//  56 61 66 93 
	//  82 86 90 78 
	//  48 49 52 94 
	//  58 62 68 79 
	//  57 54 67 95 

	// before pre shuffle
	//  50 53 64 76 
	//  59 55 69 92 
	//  51 60 65 77 
	//  56 61 66 93 
	//  82 86 90 78 
	//  48 49 52 94 
	//  58 62 68 79 
	//  57 54 67 95 

	// after pre shuffle
	//  50 53 64 76 
	//  59 55 69 78 
	//  51 60 65 77 
	//  56 61 66 79 
	//  82 86 90 92 
	//  48 49 52 95 
	//  58 62 68 93 
	//  57 54 67 94 
	// (f63, f78) = (f76, f78) $ (c0, c0), (line 0, 1)
	fpma    f63 = f78, c0, f76
	fpnma   f78 = f78, c0, f76
	;;
	// (f70, f79) = (f77, f79) $ (c1, c2), (line 2, 3)
	fpma    f70 = f79, c1, f77
	fpnma   f79 = f79, c2, f77
	;;
	// (f71, f95) = (f92, f95) $ (c3, c4), (line 4, 5)
	fpma    f71 = f95, c3, f92
	fpnma   f95 = f95, c4, f92
	;;
	// (f72, f94) = (f93, f94) $ (c5, c6), (line 6, 7)
	fpma    f72 = f94, c5, f93
	fpnma   f94 = f94, c6, f93
	;;
	;;
	// (f73, f79) = (f63, f79) $ (c7, c7), (line 0, 3)
	fpma    f73 = f79, c7, f63
	fpnma   f79 = f79, c7, f63
	;;
	// (f63, f70) = (f78, f70) $ (c8, c8), (line 1, 2)
	fpma    f63 = f70, c8, f78
	fpnma   f70 = f70, c8, f78
	;;
	// (f74, f72) = (f71, f72) $ (c9, c9), (line 4, 6)
	fpma    f74 = f72, c9, f71
	fpnma   f72 = f72, c9, f71
	;;
	// (f71, f94) = (f95, f94) $ (c10, c10), (line 5, 7)
	fpma    f71 = f94, c10, f95
	fpnma   f94 = f94, c10, f95
	;;
	;;
	// (f75, f72) = (f71, f72) $ (c11, c11), (line 5, 6)
	fpma    f75 = f72, c11, f71
	fpnma   f72 = f72, c11, f71
	;;
	// (f71, f74) = (f73, f74) $ (c12, c12), (line 0, 4)
	fpma    f71 = f74, c12, f73
	fpnma   f74 = f74, c12, f73
	;;
	// (f73, f94) = (f79, f94) $ (c14, c14), (line 3, 7)
	fpma    f73 = f94, c14, f79
	fpnma   f94 = f94, c14, f79
	;;
	;;
	// (f76, f75) = (f63, f75) $ (c13, c13), (line 1, 5)
	fpma    f76 = f75, c13, f63
	fpnma   f75 = f75, c13, f63
	;;
	// (f63, f72) = (f70, f72) $ (c13, c13), (line 2, 6)
	fpma    f63 = f72, c13, f70
	fpnma   f72 = f72, c13, f70
	;;

	// before post shuffle
	//  50 53 64 71 
	//  59 55 69 76 
	//  51 60 65 63 
	//  56 61 66 73 
	//  82 86 90 74 
	//  48 49 52 75 
	//  58 62 68 72 
	//  57 54 67 94 

	// after post shuffle
	//  50 53 64 71 
	//  59 55 69 76 
	//  51 60 65 72 
	//  56 61 66 73 
	//  82 86 90 94 
	//  48 49 52 63 
	//  58 62 68 75 
	//  57 54 67 74 
	;;
	fmix.r  f70 = f50, f59
	fmix.r  f77 = f53, f55
	fmix.r  f78 = f64, f69
	fmix.r  f79 = f71, f76
	fmix.r  f80 = f51, f56
	fmix.r  f81 = f60, f61
	fmix.r  f83 = f65, f66
	fmix.r  f84 = f72, f73
	fmix.r  f85 = f82, f48
	fmix.r  f87 = f86, f49
	fmix.r  f88 = f90, f52
	fmix.r  f89 = f94, f63
	fmix.r  f91 = f58, f57
	fmix.r  f92 = f62, f54
	fmix.r  f93 = f68, f67
	fmix.r  f95 = f75, f74
	;;
	fmix.l  f50 = f50, f59
	fmix.l  f53 = f53, f55
	fmix.l  f64 = f64, f69
	fmix.l  f71 = f71, f76
	fmix.l  f51 = f51, f56
	fmix.l  f60 = f60, f61
	fmix.l  f65 = f65, f66
	fmix.l  f72 = f72, f73
	fmix.l  f82 = f82, f48
	fmix.l  f86 = f86, f49
	fmix.l  f90 = f90, f52
	fmix.l  f94 = f94, f63
	fmix.l  f58 = f58, f57
	fmix.l  f62 = f62, f54
	fmix.l  f68 = f68, f67
	fmix.l  f75 = f75, f74
	;;
	//  50 51 82 58 
	//  70 80 85 91 
	//  53 60 86 62 
	//  77 81 87 92 
	//  64 65 90 68 
	//  78 83 88 93 
	//  71 72 94 75 
	//  79 84 89 95 
	mov   addreg1 = in0
	add   addreg2 = 4, in0
	;;
	fpcvt.fx f50 = f50
	fpcvt.fx f51 = f51
	fpcvt.fx f82 = f82
	fpcvt.fx f58 = f58
	fpcvt.fx f70 = f70
	fpcvt.fx f80 = f80
	fpcvt.fx f85 = f85
	fpcvt.fx f91 = f91
	fpcvt.fx f53 = f53
	fpcvt.fx f60 = f60
	fpcvt.fx f86 = f86
	fpcvt.fx f62 = f62
	fpcvt.fx f77 = f77
	fpcvt.fx f81 = f81
	fpcvt.fx f87 = f87
	fpcvt.fx f92 = f92
	fpcvt.fx f64 = f64
	fpcvt.fx f65 = f65
	fpcvt.fx f90 = f90
	fpcvt.fx f68 = f68
	fpcvt.fx f78 = f78
	fpcvt.fx f83 = f83
	fpcvt.fx f88 = f88
	fpcvt.fx f93 = f93
	fpcvt.fx f71 = f71
	fpcvt.fx f72 = f72
	fpcvt.fx f94 = f94
	fpcvt.fx f75 = f75
	fpcvt.fx f79 = f79
	fpcvt.fx f84 = f84
	fpcvt.fx f89 = f89
	fpcvt.fx f95 = f95
	;;
	getf.sig r33 = f50
	getf.sig r34 = f51
	getf.sig r35 = f82
	getf.sig r36 = f58
	getf.sig r37 = f70
	getf.sig r38 = f80
	getf.sig r39 = f85
	getf.sig r40 = f91
	getf.sig r41 = f53
	getf.sig r42 = f60
	getf.sig r43 = f86
	getf.sig r44 = f62
	getf.sig r45 = f77
	getf.sig r46 = f81
	getf.sig r47 = f87
	getf.sig r48 = f92
	getf.sig r49 = f64
	getf.sig r50 = f65
	getf.sig r51 = f90
	getf.sig r52 = f68
	getf.sig r53 = f78
	getf.sig r54 = f83
	getf.sig r55 = f88
	getf.sig r56 = f93
	getf.sig r57 = f71
	getf.sig r58 = f72
	getf.sig r59 = f94
	getf.sig r60 = f75
	getf.sig r61 = f79
	getf.sig r62 = f84
	getf.sig r63 = f89
	getf.sig r64 = f95
	;;
	shl      r33 = r33, 7
	shl      r34 = r34, 7
	shl      r35 = r35, 7
	shl      r36 = r36, 7
	shl      r37 = r37, 7
	shl      r38 = r38, 7
	shl      r39 = r39, 7
	shl      r40 = r40, 7
	shl      r41 = r41, 7
	shl      r42 = r42, 7
	shl      r43 = r43, 7
	shl      r44 = r44, 7
	shl      r45 = r45, 7
	shl      r46 = r46, 7
	shl      r47 = r47, 7
	shl      r48 = r48, 7
	shl      r49 = r49, 7
	shl      r50 = r50, 7
	shl      r51 = r51, 7
	shl      r52 = r52, 7
	shl      r53 = r53, 7
	shl      r54 = r54, 7
	shl      r55 = r55, 7
	shl      r56 = r56, 7
	shl      r57 = r57, 7
	shl      r58 = r58, 7
	shl      r59 = r59, 7
	shl      r60 = r60, 7
	shl      r61 = r61, 7
	shl      r62 = r62, 7
	shl      r63 = r63, 7
	shl      r64 = r64, 7
	;;
	pack4.sss r33 = r33, r0
	pack4.sss r34 = r34, r0
	pack4.sss r35 = r35, r0
	pack4.sss r36 = r36, r0
	pack4.sss r37 = r37, r0
	pack4.sss r38 = r38, r0
	pack4.sss r39 = r39, r0
	pack4.sss r40 = r40, r0
	pack4.sss r41 = r41, r0
	pack4.sss r42 = r42, r0
	pack4.sss r43 = r43, r0
	pack4.sss r44 = r44, r0
	pack4.sss r45 = r45, r0
	pack4.sss r46 = r46, r0
	pack4.sss r47 = r47, r0
	pack4.sss r48 = r48, r0
	pack4.sss r49 = r49, r0
	pack4.sss r50 = r50, r0
	pack4.sss r51 = r51, r0
	pack4.sss r52 = r52, r0
	pack4.sss r53 = r53, r0
	pack4.sss r54 = r54, r0
	pack4.sss r55 = r55, r0
	pack4.sss r56 = r56, r0
	pack4.sss r57 = r57, r0
	pack4.sss r58 = r58, r0
	pack4.sss r59 = r59, r0
	pack4.sss r60 = r60, r0
	pack4.sss r61 = r61, r0
	pack4.sss r62 = r62, r0
	pack4.sss r63 = r63, r0
	pack4.sss r64 = r64, r0
	;;
	pshr2    r33 = r33, 7
	pshr2    r34 = r34, 7
	pshr2    r35 = r35, 7
	pshr2    r36 = r36, 7
	pshr2    r37 = r37, 7
	pshr2    r38 = r38, 7
	pshr2    r39 = r39, 7
	pshr2    r40 = r40, 7
	pshr2    r41 = r41, 7
	pshr2    r42 = r42, 7
	pshr2    r43 = r43, 7
	pshr2    r44 = r44, 7
	pshr2    r45 = r45, 7
	pshr2    r46 = r46, 7
	pshr2    r47 = r47, 7
	pshr2    r48 = r48, 7
	pshr2    r49 = r49, 7
	pshr2    r50 = r50, 7
	pshr2    r51 = r51, 7
	pshr2    r52 = r52, 7
	pshr2    r53 = r53, 7
	pshr2    r54 = r54, 7
	pshr2    r55 = r55, 7
	pshr2    r56 = r56, 7
	pshr2    r57 = r57, 7
	pshr2    r58 = r58, 7
	pshr2    r59 = r59, 7
	pshr2    r60 = r60, 7
	pshr2    r61 = r61, 7
	pshr2    r62 = r62, 7
	pshr2    r63 = r63, 7
	pshr2    r64 = r64, 7
	;;
	mux2     r33 = r33, 0xe1
	mux2     r34 = r34, 0xe1
	mux2     r35 = r35, 0xe1
	mux2     r36 = r36, 0xe1
	mux2     r37 = r37, 0xe1
	mux2     r38 = r38, 0xe1
	mux2     r39 = r39, 0xe1
	mux2     r40 = r40, 0xe1
	mux2     r41 = r41, 0xe1
	mux2     r42 = r42, 0xe1
	mux2     r43 = r43, 0xe1
	mux2     r44 = r44, 0xe1
	mux2     r45 = r45, 0xe1
	mux2     r46 = r46, 0xe1
	mux2     r47 = r47, 0xe1
	mux2     r48 = r48, 0xe1
	mux2     r49 = r49, 0xe1
	mux2     r50 = r50, 0xe1
	mux2     r51 = r51, 0xe1
	mux2     r52 = r52, 0xe1
	mux2     r53 = r53, 0xe1
	mux2     r54 = r54, 0xe1
	mux2     r55 = r55, 0xe1
	mux2     r56 = r56, 0xe1
	mux2     r57 = r57, 0xe1
	mux2     r58 = r58, 0xe1
	mux2     r59 = r59, 0xe1
	mux2     r60 = r60, 0xe1
	mux2     r61 = r61, 0xe1
	mux2     r62 = r62, 0xe1
	mux2     r63 = r63, 0xe1
	mux2     r64 = r64, 0xe1
	;;
	st4   [addreg1] = r33, 8
	st4   [addreg2] = r34, 8
	;;
	st4   [addreg1] = r35, 8
	st4   [addreg2] = r36, 8
	;;
	st4   [addreg1] = r37, 8
	st4   [addreg2] = r38, 8
	;;
	st4   [addreg1] = r39, 8
	st4   [addreg2] = r40, 8
	;;
	st4   [addreg1] = r41, 8
	st4   [addreg2] = r42, 8
	;;
	st4   [addreg1] = r43, 8
	st4   [addreg2] = r44, 8
	;;
	st4   [addreg1] = r45, 8
	st4   [addreg2] = r46, 8
	;;
	st4   [addreg1] = r47, 8
	st4   [addreg2] = r48, 8
	;;
	st4   [addreg1] = r49, 8
	st4   [addreg2] = r50, 8
	;;
	st4   [addreg1] = r51, 8
	st4   [addreg2] = r52, 8
	;;
	st4   [addreg1] = r53, 8
	st4   [addreg2] = r54, 8
	;;
	st4   [addreg1] = r55, 8
	st4   [addreg2] = r56, 8
	;;
	st4   [addreg1] = r57, 8
	st4   [addreg2] = r58, 8
	;;
	st4   [addreg1] = r59, 8
	st4   [addreg2] = r60, 8
	;;
	st4   [addreg1] = r61, 8
	st4   [addreg2] = r62, 8
	;;
	st4   [addreg1] = r63, 8
	st4   [addreg2] = r64, 8
	;;

	mov	ar.pfs = r16
	br.ret.sptk.few b0

.endp