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 2.0"
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, from the first [close] glance
19 // it appeared that it's possible to pull up some references to
20 // elements of the key schedule itself. Original rationale ["prior
21 // loads are not safe only for "degenerated" key schedule, when some
22 // elements equal to the same value"] was kind of sloppy. I should have
23 // formulated as it really was: if we assume that pulling up reference
24 // to key[x+1] is not safe, then it would mean that key schedule would
25 // "degenerate," which is never the case. The problem is that this
26 // holds true in respect to references to key[x], but not to key[y].
27 // Legitimate "collisions" do occur within every 256^2 bytes window.
28 // Fortunately there're enough free instruction slots to keep prior
29 // reference to key[x+1], detect "collision" and compensate for it.
30 // All this without sacrificing a single clock cycle:-) Throughput is
31 // ~210MBps on 900MHz CPU, which is is >3x faster than gcc generated
32 // code and +30% - if compared to HP-UX C. Unrolling loop below should
33 // give >30% on top of that...
38 #if defined(_HPUX_SOURCE) && !defined(_LP64)
45 #define SZ 4 // this is set to sizeof(RC4_INT)
47 // SZ==4 seems to be optimal. At least SZ==8 is not any faster, not for
48 // assembler implementation, while SZ==1 code is ~30% slower.
49 #if SZ==1 // RC4_INT is unsigned char
53 #elif SZ==4 // RC4_INT is unsigned int
57 #elif SZ==8 // RC4_INT is unsigned long
63 out=r8; // [expanded] output pointer
64 inp=r9; // [expanded] output pointer
66 key=r28; // [expanded] pointer to RC4_KEY
67 ksch=r29; // (key->data+255)[&~(sizeof(key->data)-1)]
71 // void RC4(RC4_KEY *key,size_t len,const void *inp,void *out);
82 { .mii; alloc r2=ar.pfs,4,12,0,16
85 { .mib; cmp.eq p6,p0=0,in1 // len==0?
87 (p6) br.ret.spnt.many b0 };; // emergency exit
90 .rotr dat[4],key_x[4],tx[2],rnd[2],key_y[2],ty[1];
92 { .mib; LDKEY xx=[key],SZ // load key->x
93 add in1=-1,in1 // adjust len for loop counter
95 { .mib; ADDP inp=0,in2
97 brp.loop.imp .Ltop,.Lexit-16 };;
98 { .mmi; LDKEY yy=[key] // load key->y
101 { .mmi; mov key_y[1]=r0 // guarantee inequality
102 // in first iteration
106 dep key_x[1]=xx,r0,OFF,8
107 mov ar.ec=3 };; // note that epilogue counter
108 // is off by 1. I compensate
109 // for this at exit...
111 // The loop is scheduled for 4*(n+2) spin-rate on Itanium 2, which
112 // theoretically gives asymptotic performance of clock frequency
113 // divided by 4 bytes per seconds, or 400MBps on 1.6GHz CPU. This is
114 // for sizeof(RC4_INT)==4. For smaller RC4_INT STKEY inadvertently
115 // splits the last bundle and you end up with 5*n spin-rate:-(
116 // Originally the loop was scheduled for 3*n and relied on key
117 // schedule to be aligned at 256*sizeof(RC4_INT) boundary. But
118 // *(out++)=dat, which maps to st1, had same effect [inadvertent
119 // bundle split] and holded the loop back. Rescheduling for 4*n
120 // made it possible to eliminate dependence on specific alignment
121 // and allow OpenSSH keep "abusing" our API. Reaching for 3*n would
122 // require unrolling, sticking to variable shift instruction for
123 // collecting output [to avoid starvation for integer shifter] and
124 // copying of key schedule to controlled place in stack [so that
125 // deposit instruction can serve as substitute for whole
126 // key->data+((x&255)<<log2(sizeof(key->data[0])))]...
127 { .mmi; (p19) st1 [out]=dat[3],1 // *(out++)=dat
128 (p16) add xx=1,xx // x++
129 (p18) dep rnd[1]=rnd[1],r0,OFF,8 } // ((tx+ty)&255)<<OFF
130 { .mmi; (p16) add key_x[1]=ksch,key_x[1] // &key[xx&255]
131 (p17) add key_y[1]=ksch,key_y[1] };; // &key[yy&255]
132 { .mmi; (p16) LDKEY tx[0]=[key_x[1]] // tx=key[xx]
133 (p17) LDKEY ty[0]=[key_y[1]] // ty=key[yy]
134 (p16) dep key_x[0]=xx,r0,OFF,8 } // (xx&255)<<OFF
135 { .mmi; (p18) add rnd[1]=ksch,rnd[1] // &key[(tx+ty)&255]
136 (p16) cmp.ne.unc p20,p21=key_x[1],key_y[1] };;
137 { .mmi; (p18) LDKEY rnd[1]=[rnd[1]] // rnd=key[(tx+ty)&255]
138 (p16) ld1 dat[0]=[inp],1 } // dat=*(inp++)
139 .pred.rel "mutex",p20,p21
140 { .mmi; (p21) add yy=yy,tx[1] // (p16)
141 (p20) add yy=yy,tx[0] // (p16) y+=tx
142 (p21) mov tx[0]=tx[1] };; // (p16)
143 { .mmi; (p17) STKEY [key_y[1]]=tx[1] // key[yy]=tx
144 (p17) STKEY [key_x[2]]=ty[0] // key[xx]=ty
145 (p16) dep key_y[0]=yy,r0,OFF,8 } // &key[yy&255]
146 { .mmb; (p17) add rnd[0]=tx[1],ty[0] // tx+=ty
147 (p18) xor dat[2]=dat[2],rnd[1] // dat^=rnd
148 br.ctop.sptk .Ltop };;
150 { .mib; STKEY [key]=yy,-SZ // save key->y
151 mov pr=prsave,0x1ffff
153 { .mib; st1 [out]=dat[3],1 // compensate for truncated
157 { .mib; STKEY [key]=xx // save key->x
159 br.ret.sptk.many b0 };;