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 # ====================================================================
12 # The module implements "4-bit" GCM GHASH function and underlying
13 # single multiplication operation in GF(2^128). "4-bit" means that it
14 # uses 256 bytes per-key table [+128 bytes shared table]. Performance
15 # results are for streamed GHASH subroutine on UltraSPARC pre-Tx CPU
16 # and are expressed in cycles per processed byte, less is better:
18 # gcc 3.3.x cc 5.2 this assembler
20 # 32-bit build 81.4 43.3 12.6 (+546%/+244%)
21 # 64-bit build 20.2 21.2 12.6 (+60%/+68%)
23 # Here is data collected on UltraSPARC T1 system running Linux:
25 # gcc 4.4.1 this assembler
27 # 32-bit build 566 50 (+1000%)
28 # 64-bit build 56 50 (+12%)
30 # I don't quite understand why difference between 32-bit and 64-bit
31 # compiler-generated code is so big. Compilers *were* instructed to
32 # generate code for UltraSPARC and should have used 64-bit registers
33 # for Z vector (see C code) even in 32-bit build... Oh well, it only
34 # means more impressive improvement coefficients for this assembler
35 # module;-) Loops are aggressively modulo-scheduled in respect to
36 # references to input data and Z.hi updates to achieve 12 cycles
37 # timing. To anchor to something else, sha1-sparcv9.pl spends 11.6
38 # cycles to process one byte on UltraSPARC pre-Tx CPU and ~24 on T1.
42 # Add VIS3 lookup-table-free implementation using polynomial
43 # multiplication xmulx[hi] and extended addition addxc[cc]
44 # instructions. 4.52/7.63x improvement on T3/T4 or in absolute
45 # terms 7.90/2.14 cycles per byte. On T4 multi-process benchmark
46 # saturates at ~15.5x single-process result on 8-core processor,
47 # or ~20.5GBps per 2.85GHz socket.
50 open STDOUT,">$output";
55 $Zhi="%o0"; # 64-bit values
62 $nhi="%l0"; # small values and pointers
71 $Xi="%i0"; # input argument block
77 #include "sparc_arch.h"
80 .register %g2,#scratch
81 .register %g3,#scratch
84 .section ".text",#alloc,#execinstr
88 .long `0x0000<<16`,0,`0x1C20<<16`,0,`0x3840<<16`,0,`0x2460<<16`,0
89 .long `0x7080<<16`,0,`0x6CA0<<16`,0,`0x48C0<<16`,0,`0x54E0<<16`,0
90 .long `0xE100<<16`,0,`0xFD20<<16`,0,`0xD940<<16`,0,`0xC560<<16`,0
91 .long `0x9180<<16`,0,`0x8DA0<<16`,0,`0xA9C0<<16`,0,`0xB5E0<<16`,0
92 .type rem_4bit,#object
93 .size rem_4bit,(.-rem_4bit)
106 add %o7,rem_4bit-1b,$rem_4bit
113 ldx [$Htblo+$nlo],$Zlo
114 ldx [$Htbl+$nlo],$Zhi
118 ldx [$Htblo+$nhi],$Tlo
120 ldx [$Htbl+$nhi],$Thi
122 ldx [$rem_4bit+$remi],$rem
138 ldx [$Htblo+$nlo],$Tlo
141 ldx [$Htbl+$nlo],$Thi
144 ldx [$rem_4bit+$remi],$rem
147 ldub [$inp+$cnt],$nlo
154 ldx [$Htblo+$nhi],$Tlo
157 ldx [$Htbl+$nhi],$Thi
159 ldx [$rem_4bit+$remi],$rem
172 ldx [$Htblo+$nlo],$Tlo
175 ldx [$Htbl+$nlo],$Thi
178 ldx [$rem_4bit+$remi],$rem
187 be,pn SIZE_T_CC,.Ldone
190 ldx [$Htblo+$nhi],$Tlo
193 ldx [$Htbl+$nhi],$Thi
195 ldx [$rem_4bit+$remi],$rem
211 ldx [$Htblo+$nhi],$Tlo
214 ldx [$Htbl+$nhi],$Thi
216 ldx [$rem_4bit+$remi],$rem
228 .type gcm_ghash_4bit,#function
229 .size gcm_ghash_4bit,(.-gcm_ghash_4bit)
236 .globl gcm_gmult_4bit
244 add %o7,rem_4bit-1b,$rem_4bit
249 ldx [$Htblo+$nlo],$Zlo
250 ldx [$Htbl+$nlo],$Zhi
254 ldx [$Htblo+$nhi],$Tlo
256 ldx [$Htbl+$nhi],$Thi
258 ldx [$rem_4bit+$remi],$rem
273 ldx [$Htblo+$nlo],$Tlo
276 ldx [$Htbl+$nlo],$Thi
279 ldx [$rem_4bit+$remi],$rem
288 ldx [$Htblo+$nhi],$Tlo
291 ldx [$Htbl+$nhi],$Thi
293 ldx [$rem_4bit+$remi],$rem
305 ldx [$Htblo+$nlo],$Tlo
308 ldx [$Htbl+$nlo],$Thi
311 ldx [$rem_4bit+$remi],$rem
319 ldx [$Htblo+$nhi],$Tlo
322 ldx [$Htbl+$nhi],$Thi
324 ldx [$rem_4bit+$remi],$rem
336 .type gcm_gmult_4bit,#function
337 .size gcm_gmult_4bit,(.-gcm_gmult_4bit)
341 # Straightforward 128x128-bit multiplication using Karatsuba algorithm
342 # followed by pair of 64-bit reductions [with a shortcut in first one,
343 # which allowed to break dependency between reductions and remove one
344 # multiplication from critical path]. While it might be suboptimal
345 # with regard to sheer number of multiplications, other methods [such
346 # as aggregate reduction] would require more 64-bit registers, which
347 # we don't have in 32-bit application context.
349 ($Xip,$Htable,$inp,$len)=map("%i$_",(0..3));
351 ($Hhl,$Hlo,$Hhi,$Xlo,$Xhi,$xE1,$sqr, $C0,$C1,$C2,$C3,$V)=
352 (map("%o$_",(0..5,7)),map("%g$_",(1..5)));
354 ($shl,$shr)=map("%l$_",(0..7));
356 # For details regarding "twisted H" see ghash-x86.pl.
368 srax $Hhi,63,$C0 ! broadcast carry
369 addcc $Hlo,$Hlo,$Hlo ! H<<=1
375 stx $Hlo,[%i0+8] ! save twisted H
378 sethi %hi(0xA0406080),$V
379 sethi %hi(0x20C0E000),%l0
380 or $V,%lo(0xA0406080),$V
381 or %l0,%lo(0x20C0E000),%l0
383 or %l0,$V,$V ! (0xE0·i)&0xff=0xA040608020C0E000
388 .type gcm_init_vis3,#function
389 .size gcm_init_vis3,.-gcm_init_vis3
391 .globl gcm_gmult_vis3
396 ldx [$Xip+8],$Xlo ! load Xi
398 ldx [$Htable+8],$Hlo ! load twisted H
402 sllx %l7,57,$xE1 ! 57 is not a typo
403 ldx [$Htable+16],$V ! (0xE0·i)&0xff=0xA040608020C0E000
405 xor $Hhi,$Hlo,$Hhl ! Karatsuba pre-processing
407 xor $Xlo,$Xhi,$C2 ! Karatsuba pre-processing
409 xmulxhi $Xlo,$Hlo,$Xlo
411 xmulxhi $Xhi,$Hhi,$C3
415 srlx $V,$sqr,$sqr ! ·0xE0 [implicit &(7<<3)]
417 sllx $sqr,57,$sqr ! ($C0·0xE1)<<1<<56 [implicit &0x7f]
419 xor $C0,$C1,$C1 ! Karatsuba post-processing
421 xor $sqr,$Xlo,$Xlo ! real destination is $C1
427 xmulxhi $C0,$xE1,$Xlo ! ·0xE1<<1<<56
437 stx $C2,[$Xip+8] ! save Xi
442 .type gcm_gmult_vis3,#function
443 .size gcm_gmult_vis3,.-gcm_gmult_vis3
445 .globl gcm_ghash_vis3
450 ldx [$Xip+8],$C2 ! load Xi
452 ldx [$Htable+8],$Hlo ! load twisted H
456 sllx %l7,57,$xE1 ! 57 is not a typo
457 ldx [$Htable+16],$V ! (0xE0·i)&0xff=0xA040608020C0E000
462 prefetch [$inp+63], 20
465 xor $Hhi,$Hlo,$Hhl ! Karatsuba pre-processing
471 ldx [$inp+16],$C1 ! align data
483 prefetch [$inp+63], 20
486 xor $Xlo,$Xhi,$C2 ! Karatsuba pre-processing
488 xmulxhi $Xlo,$Hlo,$Xlo
490 xmulxhi $Xhi,$Hhi,$C3
494 srlx $V,$sqr,$sqr ! ·0xE0 [implicit &(7<<3)]
496 sllx $sqr,57,$sqr ! ($C0·0xE1)<<1<<56 [implicit &0x7f]
498 xor $C0,$C1,$C1 ! Karatsuba post-processing
500 xor $sqr,$Xlo,$Xlo ! real destination is $C1
506 xmulxhi $C0,$xE1,$Xlo ! ·0xE1<<1<<56
517 stx $C2,[$Xip+8] ! save Xi
522 .type gcm_ghash_vis3,#function
523 .size gcm_ghash_vis3,.-gcm_ghash_vis3
527 .asciz "GHASH for SPARCv9/VIS3, CRYPTOGAMS by <appro\@openssl.org>"
532 # Purpose of these subroutines is to explicitly encode VIS instructions,
533 # so that one can compile the module without having to specify VIS
534 # extensions on compiler command line, e.g. -xarch=v9 vs. -xarch=v9a.
535 # Idea is to reserve for option to produce "universal" binary and let
536 # programmer detect if current CPU is VIS capable at run-time.
538 my ($mnemonic,$rs1,$rs2,$rd)=@_;
539 my %bias = ( "g" => 0, "o" => 8, "l" => 16, "i" => 24 );
541 my %visopf = ( "addxc" => 0x011,
544 "xmulxhi" => 0x116 );
546 $ref = "$mnemonic\t$rs1,$rs2,$rd";
548 if ($opf=$visopf{$mnemonic}) {
549 foreach ($rs1,$rs2,$rd) {
550 return $ref if (!/%([goli])([0-9])/);
554 return sprintf ".word\t0x%08x !%s",
555 0x81b00000|$rd<<25|$rs1<<14|$opf<<5|$rs2,
562 foreach (split("\n",$code)) {
563 s/\`([^\`]*)\`/eval $1/ge;
565 s/\b(xmulx[hi]*|addxc[c]{0,2})\s+(%[goli][0-7]),\s*(%[goli][0-7]),\s*(%[goli][0-7])/