3 # ====================================================================
4 # Written by Andy Polyakov <appro@openssl.org> for the OpenSSL
5 # project. The module is, however, dual licensed under OpenSSL and
6 # CRYPTOGAMS licenses depending on where you obtain it. For further
7 # details see http://www.openssl.org/~appro/cryptogams/.
8 # ====================================================================
10 # SHA256/512 for ARMv8.
12 # Performance in cycles per processed byte and improvement coefficient
13 # over code generated with "default" compiler:
15 # SHA256-hw SHA256(*) SHA512
16 # Apple A7 1.97 10.5 (+33%) 6.73 (-1%(**))
17 # Cortex-A53 2.38 15.5 (+115%) 10.0 (+150%(***))
18 # Cortex-A57 2.31 11.6 (+86%) 7.51 (+260%(***))
19 # Denver 2.01 10.5 (+26%) 6.70 (+8%)
20 # X-Gene 20.0 (+100%) 12.8 (+300%(***))
22 # (*) Software SHA256 results are of lesser relevance, presented
23 # mostly for informational purposes.
24 # (**) The result is a trade-off: it's possible to improve it by
25 # 10% (or by 1 cycle per round), but at the cost of 20% loss
26 # on Cortex-A53 (or by 4 cycles per round).
27 # (***) Super-impressive coefficients over gcc-generated code are
28 # indication of some compiler "pathology", most notably code
29 # generated with -mgeneral-regs-only is significanty faster
30 # and the gap is only 40-90%.
35 $0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1;
36 ( $xlate="${dir}arm-xlate.pl" and -f $xlate ) or
37 ( $xlate="${dir}../../perlasm/arm-xlate.pl" and -f $xlate) or
38 die "can't locate arm-xlate.pl";
40 open OUT,"| \"$^X\" $xlate $flavour $output";
43 if ($output =~ /512/) {
63 $func="sha${BITS}_block_data_order";
65 ($ctx,$inp,$num,$Ktbl)=map("x$_",(0..2,30));
67 @X=map("$reg_t$_",(3..15,0..2));
68 @V=($A,$B,$C,$D,$E,$F,$G,$H)=map("$reg_t$_",(20..27));
69 ($t0,$t1,$t2,$t3)=map("$reg_t$_",(16,17,19,28));
72 my ($i,$a,$b,$c,$d,$e,$f,$g,$h)=@_;
74 my ($T0,$T1,$T2)=(@X[($i-8)&15],@X[($i-9)&15],@X[($i-10)&15]);
75 $T0=@X[$i+3] if ($i<11);
77 $code.=<<___ if ($i<16);
79 rev @X[$i],@X[$i] // $i
82 $code.=<<___ if ($i<13 && ($i&1));
83 ldp @X[$i+1],@X[$i+2],[$inp],#2*$SZ
85 $code.=<<___ if ($i==13);
86 ldp @X[14],@X[15],[$inp]
88 $code.=<<___ if ($i>=14);
89 ldr @X[($i-11)&15],[sp,#`$SZ*(($i-11)%4)`]
91 $code.=<<___ if ($i>0 && $i<16);
92 add $a,$a,$t1 // h+=Sigma0(a)
94 $code.=<<___ if ($i>=11);
95 str @X[($i-8)&15],[sp,#`$SZ*(($i-8)%4)`]
97 # While ARMv8 specifies merged rotate-n-logical operation such as
98 # 'eor x,y,z,ror#n', it was found to negatively affect performance
99 # on Apple A7. The reason seems to be that it requires even 'y' to
100 # be available earlier. This means that such merged instruction is
101 # not necessarily best choice on critical path... On the other hand
102 # Cortex-A5x handles merged instructions much better than disjoint
103 # rotate and logical... See (**) footnote above.
104 $code.=<<___ if ($i<15);
105 ror $t0,$e,#$Sigma1[0]
106 add $h,$h,$t2 // h+=K[i]
107 eor $T0,$e,$e,ror#`$Sigma1[2]-$Sigma1[1]`
110 add $h,$h,@X[$i&15] // h+=X[i]
111 orr $t1,$t1,$t2 // Ch(e,f,g)
112 eor $t2,$a,$b // a^b, b^c in next round
113 eor $t0,$t0,$T0,ror#$Sigma1[1] // Sigma1(e)
114 ror $T0,$a,#$Sigma0[0]
115 add $h,$h,$t1 // h+=Ch(e,f,g)
116 eor $t1,$a,$a,ror#`$Sigma0[2]-$Sigma0[1]`
117 add $h,$h,$t0 // h+=Sigma1(e)
118 and $t3,$t3,$t2 // (b^c)&=(a^b)
120 eor $t3,$t3,$b // Maj(a,b,c)
121 eor $t1,$T0,$t1,ror#$Sigma0[1] // Sigma0(a)
122 add $h,$h,$t3 // h+=Maj(a,b,c)
123 ldr $t3,[$Ktbl],#$SZ // *K++, $t2 in next round
124 //add $h,$h,$t1 // h+=Sigma0(a)
126 $code.=<<___ if ($i>=15);
127 ror $t0,$e,#$Sigma1[0]
128 add $h,$h,$t2 // h+=K[i]
129 ror $T1,@X[($j+1)&15],#$sigma0[0]
131 ror $T2,@X[($j+14)&15],#$sigma1[0]
133 ror $T0,$a,#$Sigma0[0]
134 add $h,$h,@X[$i&15] // h+=X[i]
135 eor $t0,$t0,$e,ror#$Sigma1[1]
136 eor $T1,$T1,@X[($j+1)&15],ror#$sigma0[1]
137 orr $t1,$t1,$t2 // Ch(e,f,g)
138 eor $t2,$a,$b // a^b, b^c in next round
139 eor $t0,$t0,$e,ror#$Sigma1[2] // Sigma1(e)
140 eor $T0,$T0,$a,ror#$Sigma0[1]
141 add $h,$h,$t1 // h+=Ch(e,f,g)
142 and $t3,$t3,$t2 // (b^c)&=(a^b)
143 eor $T2,$T2,@X[($j+14)&15],ror#$sigma1[1]
144 eor $T1,$T1,@X[($j+1)&15],lsr#$sigma0[2] // sigma0(X[i+1])
145 add $h,$h,$t0 // h+=Sigma1(e)
146 eor $t3,$t3,$b // Maj(a,b,c)
147 eor $t1,$T0,$a,ror#$Sigma0[2] // Sigma0(a)
148 eor $T2,$T2,@X[($j+14)&15],lsr#$sigma1[2] // sigma1(X[i+14])
149 add @X[$j],@X[$j],@X[($j+9)&15]
151 add $h,$h,$t3 // h+=Maj(a,b,c)
152 ldr $t3,[$Ktbl],#$SZ // *K++, $t2 in next round
153 add @X[$j],@X[$j],$T1
154 add $h,$h,$t1 // h+=Sigma0(a)
155 add @X[$j],@X[$j],$T2
161 #include "arm_arch.h"
165 .extern OPENSSL_armcap_P
167 .type $func,%function
171 $code.=<<___ if ($SZ==4);
173 ldrsw x16,.LOPENSSL_armcap_P
175 ldr x16,.LOPENSSL_armcap_P
177 adr x17,.LOPENSSL_armcap_P
180 tst w16,#ARMV8_SHA256
184 stp x29,x30,[sp,#-128]!
194 ldp $A,$B,[$ctx] // load context
195 ldp $C,$D,[$ctx,#2*$SZ]
196 ldp $E,$F,[$ctx,#4*$SZ]
197 add $num,$inp,$num,lsl#`log(16*$SZ)/log(2)` // end of input
198 ldp $G,$H,[$ctx,#6*$SZ]
200 stp $ctx,$num,[x29,#96]
203 ldp @X[0],@X[1],[$inp],#2*$SZ
204 ldr $t2,[$Ktbl],#$SZ // *K++
205 eor $t3,$B,$C // magic seed
208 for ($i=0;$i<16;$i++) { &BODY_00_xx($i,@V); unshift(@V,pop(@V)); }
209 $code.=".Loop_16_xx:\n";
210 for (;$i<32;$i++) { &BODY_00_xx($i,@V); unshift(@V,pop(@V)); }
214 ldp $ctx,$num,[x29,#96]
216 sub $Ktbl,$Ktbl,#`$SZ*($rounds+1)` // rewind
218 ldp @X[0],@X[1],[$ctx]
219 ldp @X[2],@X[3],[$ctx,#2*$SZ]
220 add $inp,$inp,#14*$SZ // advance input pointer
221 ldp @X[4],@X[5],[$ctx,#4*$SZ]
223 ldp @X[6],@X[7],[$ctx,#6*$SZ]
230 stp $C,$D,[$ctx,#2*$SZ]
234 stp $E,$F,[$ctx,#4*$SZ]
235 stp $G,$H,[$ctx,#6*$SZ]
238 ldp x19,x20,[x29,#16]
240 ldp x21,x22,[x29,#32]
241 ldp x23,x24,[x29,#48]
242 ldp x25,x26,[x29,#64]
243 ldp x27,x28,[x29,#80]
244 ldp x29,x30,[sp],#128
249 .type .LK$BITS,%object
252 $code.=<<___ if ($SZ==8);
253 .quad 0x428a2f98d728ae22,0x7137449123ef65cd
254 .quad 0xb5c0fbcfec4d3b2f,0xe9b5dba58189dbbc
255 .quad 0x3956c25bf348b538,0x59f111f1b605d019
256 .quad 0x923f82a4af194f9b,0xab1c5ed5da6d8118
257 .quad 0xd807aa98a3030242,0x12835b0145706fbe
258 .quad 0x243185be4ee4b28c,0x550c7dc3d5ffb4e2
259 .quad 0x72be5d74f27b896f,0x80deb1fe3b1696b1
260 .quad 0x9bdc06a725c71235,0xc19bf174cf692694
261 .quad 0xe49b69c19ef14ad2,0xefbe4786384f25e3
262 .quad 0x0fc19dc68b8cd5b5,0x240ca1cc77ac9c65
263 .quad 0x2de92c6f592b0275,0x4a7484aa6ea6e483
264 .quad 0x5cb0a9dcbd41fbd4,0x76f988da831153b5
265 .quad 0x983e5152ee66dfab,0xa831c66d2db43210
266 .quad 0xb00327c898fb213f,0xbf597fc7beef0ee4
267 .quad 0xc6e00bf33da88fc2,0xd5a79147930aa725
268 .quad 0x06ca6351e003826f,0x142929670a0e6e70
269 .quad 0x27b70a8546d22ffc,0x2e1b21385c26c926
270 .quad 0x4d2c6dfc5ac42aed,0x53380d139d95b3df
271 .quad 0x650a73548baf63de,0x766a0abb3c77b2a8
272 .quad 0x81c2c92e47edaee6,0x92722c851482353b
273 .quad 0xa2bfe8a14cf10364,0xa81a664bbc423001
274 .quad 0xc24b8b70d0f89791,0xc76c51a30654be30
275 .quad 0xd192e819d6ef5218,0xd69906245565a910
276 .quad 0xf40e35855771202a,0x106aa07032bbd1b8
277 .quad 0x19a4c116b8d2d0c8,0x1e376c085141ab53
278 .quad 0x2748774cdf8eeb99,0x34b0bcb5e19b48a8
279 .quad 0x391c0cb3c5c95a63,0x4ed8aa4ae3418acb
280 .quad 0x5b9cca4f7763e373,0x682e6ff3d6b2b8a3
281 .quad 0x748f82ee5defb2fc,0x78a5636f43172f60
282 .quad 0x84c87814a1f0ab72,0x8cc702081a6439ec
283 .quad 0x90befffa23631e28,0xa4506cebde82bde9
284 .quad 0xbef9a3f7b2c67915,0xc67178f2e372532b
285 .quad 0xca273eceea26619c,0xd186b8c721c0c207
286 .quad 0xeada7dd6cde0eb1e,0xf57d4f7fee6ed178
287 .quad 0x06f067aa72176fba,0x0a637dc5a2c898a6
288 .quad 0x113f9804bef90dae,0x1b710b35131c471b
289 .quad 0x28db77f523047d84,0x32caab7b40c72493
290 .quad 0x3c9ebe0a15c9bebc,0x431d67c49c100d4c
291 .quad 0x4cc5d4becb3e42b6,0x597f299cfc657e2a
292 .quad 0x5fcb6fab3ad6faec,0x6c44198c4a475817
293 .quad 0 // terminator
295 $code.=<<___ if ($SZ==4);
296 .long 0x428a2f98,0x71374491,0xb5c0fbcf,0xe9b5dba5
297 .long 0x3956c25b,0x59f111f1,0x923f82a4,0xab1c5ed5
298 .long 0xd807aa98,0x12835b01,0x243185be,0x550c7dc3
299 .long 0x72be5d74,0x80deb1fe,0x9bdc06a7,0xc19bf174
300 .long 0xe49b69c1,0xefbe4786,0x0fc19dc6,0x240ca1cc
301 .long 0x2de92c6f,0x4a7484aa,0x5cb0a9dc,0x76f988da
302 .long 0x983e5152,0xa831c66d,0xb00327c8,0xbf597fc7
303 .long 0xc6e00bf3,0xd5a79147,0x06ca6351,0x14292967
304 .long 0x27b70a85,0x2e1b2138,0x4d2c6dfc,0x53380d13
305 .long 0x650a7354,0x766a0abb,0x81c2c92e,0x92722c85
306 .long 0xa2bfe8a1,0xa81a664b,0xc24b8b70,0xc76c51a3
307 .long 0xd192e819,0xd6990624,0xf40e3585,0x106aa070
308 .long 0x19a4c116,0x1e376c08,0x2748774c,0x34b0bcb5
309 .long 0x391c0cb3,0x4ed8aa4a,0x5b9cca4f,0x682e6ff3
310 .long 0x748f82ee,0x78a5636f,0x84c87814,0x8cc70208
311 .long 0x90befffa,0xa4506ceb,0xbef9a3f7,0xc67178f2
315 .size .LK$BITS,.-.LK$BITS
319 .long OPENSSL_armcap_P-.
321 .quad OPENSSL_armcap_P-.
323 .asciz "SHA$BITS block transform for ARMv8, CRYPTOGAMS by <appro\@openssl.org>"
330 my ($ABCD,$EFGH,$abcd)=map("v$_.16b",(0..2));
331 my @MSG=map("v$_.16b",(4..7));
332 my ($W0,$W1)=("v16.4s","v17.4s");
333 my ($ABCD_SAVE,$EFGH_SAVE)=("v18.16b","v19.16b");
336 .type sha256_block_armv8,%function
340 stp x29,x30,[sp,#-16]!
343 ld1.32 {$ABCD,$EFGH},[$ctx]
347 ld1 {@MSG[0]-@MSG[3]},[$inp],#64
349 ld1.32 {$W0},[$Ktbl],#16
350 rev32 @MSG[0],@MSG[0]
351 rev32 @MSG[1],@MSG[1]
352 rev32 @MSG[2],@MSG[2]
353 rev32 @MSG[3],@MSG[3]
354 orr $ABCD_SAVE,$ABCD,$ABCD // offload
355 orr $EFGH_SAVE,$EFGH,$EFGH
357 for($i=0;$i<12;$i++) {
359 ld1.32 {$W1},[$Ktbl],#16
360 add.i32 $W0,$W0,@MSG[0]
361 sha256su0 @MSG[0],@MSG[1]
362 orr $abcd,$ABCD,$ABCD
363 sha256h $ABCD,$EFGH,$W0
364 sha256h2 $EFGH,$abcd,$W0
365 sha256su1 @MSG[0],@MSG[2],@MSG[3]
367 ($W0,$W1)=($W1,$W0); push(@MSG,shift(@MSG));
370 ld1.32 {$W1},[$Ktbl],#16
371 add.i32 $W0,$W0,@MSG[0]
372 orr $abcd,$ABCD,$ABCD
373 sha256h $ABCD,$EFGH,$W0
374 sha256h2 $EFGH,$abcd,$W0
376 ld1.32 {$W0},[$Ktbl],#16
377 add.i32 $W1,$W1,@MSG[1]
378 orr $abcd,$ABCD,$ABCD
379 sha256h $ABCD,$EFGH,$W1
380 sha256h2 $EFGH,$abcd,$W1
383 add.i32 $W0,$W0,@MSG[2]
384 sub $Ktbl,$Ktbl,#$rounds*$SZ-16 // rewind
385 orr $abcd,$ABCD,$ABCD
386 sha256h $ABCD,$EFGH,$W0
387 sha256h2 $EFGH,$abcd,$W0
389 add.i32 $W1,$W1,@MSG[3]
390 orr $abcd,$ABCD,$ABCD
391 sha256h $ABCD,$EFGH,$W1
392 sha256h2 $EFGH,$abcd,$W1
394 add.i32 $ABCD,$ABCD,$ABCD_SAVE
395 add.i32 $EFGH,$EFGH,$EFGH_SAVE
399 st1.32 {$ABCD,$EFGH},[$ctx]
403 .size sha256_block_armv8,.-sha256_block_armv8
408 .comm OPENSSL_armcap_P,4,4
412 "sha256h" => 0x5e004000, "sha256h2" => 0x5e005000,
413 "sha256su0" => 0x5e282800, "sha256su1" => 0x5e006000 );
416 my ($mnemonic,$arg)=@_;
418 $arg =~ m/[qv]([0-9]+)[^,]*,\s*[qv]([0-9]+)[^,]*(?:,\s*[qv]([0-9]+))?/o
420 sprintf ".inst\t0x%08x\t//%s %s",
421 $opcode{$mnemonic}|$1|($2<<5)|($3<<16),
426 foreach(split("\n",$code)) {
428 s/\`([^\`]*)\`/eval($1)/geo;
430 s/\b(sha256\w+)\s+([qv].*)/unsha256($1,$2)/geo;
432 s/\.\w?32\b//o and s/\.16b/\.4s/go;
433 m/(ld|st)1[^\[]+\[0\]/o and s/\.4s/\.s/go;