1 // ====================================================================
2 // Written by Andy Polyakov <appro@fy.chalmers.se> for the OpenSSL
5 // Rights for redistribution and usage in source and binary forms are
6 // granted according to the OpenSSL license. Warranty of any kind is
8 // ====================================================================
10 .ident "rc4-ia64.S, Version 1.0"
11 .ident "IA-64 ISA artwork by Andy Polyakov <appro@fy.chalmers.se>"
13 // What's wrong with compiler generated code? Because of the nature of
14 // C language, compiler doesn't [dare to] reorder load and stores. But
15 // being memory-bound, RC4 should benefit from reorder [on in-order-
16 // execution core such as IA-64]. But what can we reorder? At the very
17 // least we can safely reorder references to key schedule in respect
18 // to input and output streams. Secondly, less obvious, it's possible
19 // to pull up some references to elements of the key schedule itself.
20 // Fact is that such prior loads are not safe only for "degenerated"
21 // key schedule, when all elements equal to the same value, which is
22 // never the case [key schedule setup routine makes sure it's not].
23 // Furthermore. In order to compress loop body to the minimum, I chose
24 // to deploy deposit instruction, which substitutes for the whole
25 // key->data+((x&255)<<log2(sizeof(key->data[0]))). This unfortunately
26 // requires key->data to be aligned at sizeof(key->data) boundary.
27 // This is why you'll find "RC4_INT pad[512-256-2];" addenum to RC4_KEY
28 // and "d=(RC4_INT *)(((size_t)(d+255))&~(sizeof(key->data)-1));" in
29 // rc4_skey.c [and rc4_enc.c, where it's retained for debugging
30 // purposes]. Throughput is ~210MBps on 900MHz CPU, which is is >3x
31 // faster than gcc generated code and +30% - if compared to HP-UX C.
32 // Unrolling loop below should give >30% on top of that...
37 #if defined(_HPUX_SOURCE) && !defined(_LP64)
43 #define SZ 4 // this is set to sizeof(RC4_INT)
44 // SZ==4 seems to be optimal. At least SZ==8 is not any faster, not for
45 // assembler implementation, while SZ==1 code is ~30% slower.
46 #if SZ==1 // RC4_INT is unsigned char
50 #elif SZ==4 // RC4_INT is unsigned int
54 #elif SZ==8 // RC4_INT is unsigned long
60 out=r8; // [expanded] output pointer
61 inp=r9; // [expanded] output pointer
63 key=r28; // [expanded] pointer to RC4_KEY
64 ksch=r29; // (key->data+255)[&~(sizeof(key->data)-1)]
68 // void RC4(RC4_KEY *key,size_t len,const void *inp,void *out);
79 { .mii; alloc r2=ar.pfs,4,12,0,16
82 { .mib; cmp.eq p6,p0=0,in1 // len==0?
84 (p6) br.ret.spnt.many b0 };; // emergency exit
87 .rotr dat[4],key_x[4],tx[2],rnd[2],key_y[2],ty[1];
89 { .mib; LDKEY xx=[key],SZ // load key->x
90 add in1=-1,in1 // adjust len for loop counter
92 { .mib; ADDP inp=0,in2
94 brp.loop.imp .Ltop,.Lexit-16 };;
95 { .mmi; LDKEY yy=[key] // load key->y
96 add ksch=(255+1)*SZ,key // as ksch will be used with
97 // deposit instruction only,
98 // I don't have to &~255...
104 dep key_x[1]=xx,ksch,OFF,8
105 mov ar.ec=3 };; // note that epilogue counter
106 // is off by 1. I compensate
107 // for this at exit...
109 // The loop is scheduled for 3*(n+2) spin-rate on Itanium 2, which
110 // theoretically gives asymptotic performance of clock frequency
111 // divided by 3 bytes per seconds, or 500MBps on 1.5GHz CPU. Measured
112 // performance however is distinctly lower than 1/4:-( The culplrit
113 // seems to be *(out++)=dat, which inadvertently splits the bundle,
114 // even though there is M-unit available... Unrolling is due...
115 // Unrolled loop should collect output with variable shift instruction
116 // in order to avoid starvation for integer shifter... Only output
117 // pointer has to be aligned... It should be possible to get pretty
118 // close to theoretical peak...
119 { .mmi; (p16) LDKEY tx[0]=[key_x[1]] // tx=key[xx]
120 (p17) LDKEY ty[0]=[key_y[1]] // ty=key[yy]
121 (p18) dep rnd[1]=rnd[1],ksch,OFF,8} // &key[(tx+ty)&255]
122 { .mmi; (p19) st1 [out]=dat[3],1 // *(out++)=dat
123 (p16) add xx=1,xx // x++
125 { .mmi; (p18) LDKEY rnd[1]=[rnd[1]] // rnd=key[(tx+ty)&255]
126 (p16) ld1 dat[0]=[inp],1 // dat=*(inp++)
127 (p16) dep key_x[0]=xx,ksch,OFF,8 } // &key[xx&255]
129 (p16) add yy=yy,tx[0] // y+=tx
131 { .mmi; (p17) STKEY [key_y[1]]=tx[1] // key[yy]=tx
132 (p17) STKEY [key_x[2]]=ty[0] // key[xx]=ty
133 (p16) dep key_y[0]=yy,ksch,OFF,8 } // &key[yy&255]
134 { .mmb; (p17) add rnd[0]=tx[1],ty[0] // tx+=ty
135 (p18) xor dat[2]=dat[2],rnd[1] // dat^=rnd
136 br.ctop.sptk .Ltop };;
138 { .mib; STKEY [key]=yy,-SZ // save key->y
139 mov pr=prsave,0x1ffff
141 { .mib; st1 [out]=dat[3],1 // compensate for truncated
145 { .mib; STKEY [key]=xx // save key->x
147 br.ret.sptk.many b0 };;