3 # ====================================================================
4 # Written by Andy Polyakov <appro@fy.chalmers.se> 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 # sha1_block procedure for x86_64.
12 # It was brought to my attention that on EM64T compiler-generated code
13 # was far behind 32-bit assembler implementation. This is unlike on
14 # Opteron where compiler-generated code was only 15% behind 32-bit
15 # assembler, which originally made it hard to motivate the effort.
16 # There was suggestion to mechanically translate 32-bit code, but I
17 # dismissed it, reasoning that x86_64 offers enough register bank
18 # capacity to fully utilize SHA-1 parallelism. Therefore this fresh
19 # implementation:-) However! While 64-bit code does perform better
20 # on Opteron, I failed to beat 32-bit assembler on EM64T core. Well,
21 # x86_64 does offer larger *addressable* bank, but out-of-order core
22 # reaches for even more registers through dynamic aliasing, and EM64T
23 # core must have managed to run-time optimize even 32-bit code just as
24 # good as 64-bit one. Performance improvement is summarized in the
27 # gcc 3.4 32-bit asm cycles/byte
28 # Opteron +45% +20% 6.8
29 # Xeon P4 +65% +0% 9.9
34 # The code was revised to minimize code size and to maximize
35 # "distance" between instructions producing input to 'lea'
36 # instruction and the 'lea' instruction itself, which is essential
37 # for Intel Atom core.
41 if ($flavour =~ /\./) { $output = $flavour; undef $flavour; }
43 $win64=0; $win64=1 if ($flavour =~ /[nm]asm|mingw64/ || $output =~ /\.asm$/);
45 $0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1;
46 ( $xlate="${dir}x86_64-xlate.pl" and -f $xlate ) or
47 ( $xlate="${dir}../../perlasm/x86_64-xlate.pl" and -f $xlate) or
48 die "can't locate x86_64-xlate.pl";
50 open STDOUT,"| $^X $xlate $flavour $output";
52 $ctx="%rdi"; # 1st arg
53 $inp="%rsi"; # 2nd arg
54 $num="%rdx"; # 3rd arg
56 # reassign arguments in order to produce more compact code
74 my ($i,$a,$b,$c,$d,$e)=@_;
76 $code.=<<___ if ($i==0);
77 mov `4*$i`($inp),$xi[0]
79 mov $xi[0],`4*$i`(%rsp)
81 $code.=<<___ if ($i<15);
83 mov `4*$j`($inp),$xi[1]
88 lea 0x5a827999($xi[0],$e),$e
90 mov $xi[1],`4*$j`(%rsp)
96 $code.=<<___ if ($i>=15);
97 mov `4*($j%16)`(%rsp),$xi[1]
100 xor `4*(($j+2)%16)`(%rsp),$xi[1]
103 xor `4*(($j+8)%16)`(%rsp),$xi[1]
105 lea 0x5a827999($xi[0],$e),$e
106 xor `4*(($j+13)%16)`(%rsp),$xi[1]
111 mov $xi[1],`4*($j%16)`(%rsp)
114 unshift(@xi,pop(@xi));
118 my ($i,$a,$b,$c,$d,$e)=@_;
120 my $K=($i<40)?0x6ed9eba1:0xca62c1d6;
121 $code.=<<___ if ($i<79);
122 mov `4*($j%16)`(%rsp),$xi[1]
125 xor `4*(($j+2)%16)`(%rsp),$xi[1]
129 xor `4*(($j+8)%16)`(%rsp),$xi[1]
132 xor `4*(($j+13)%16)`(%rsp),$xi[1]
137 $code.=<<___ if ($i<76);
138 mov $xi[1],`4*($j%16)`(%rsp)
140 $code.=<<___ if ($i==79);
151 unshift(@xi,pop(@xi));
155 my ($i,$a,$b,$c,$d,$e)=@_;
158 mov `4*($j%16)`(%rsp),$xi[1]
161 xor `4*(($j+2)%16)`(%rsp),$xi[1]
164 xor `4*(($j+8)%16)`(%rsp),$xi[1]
166 lea 0x8f1bbcdc($xi[0],$e),$e
168 xor `4*(($j+13)%16)`(%rsp),$xi[1]
174 mov $xi[1],`4*($j%16)`(%rsp)
177 unshift(@xi,pop(@xi));
181 .section .note.GNU-stack,"",\@progbits
184 .globl sha1_block_data_order
185 .type sha1_block_data_order,\@function,3
187 sha1_block_data_order:
193 mov %rdi,$ctx # reassigned argument
195 mov %rsi,$inp # reassigned argument
197 mov %rdx,$num # reassigned argument
198 mov %r11,`16*4`(%rsp)
210 for($i=0;$i<20;$i++) { &BODY_00_19($i,@V); unshift(@V,pop(@V)); }
211 for(;$i<40;$i++) { &BODY_20_39($i,@V); unshift(@V,pop(@V)); }
212 for(;$i<60;$i++) { &BODY_40_59($i,@V); unshift(@V,pop(@V)); }
213 for(;$i<80;$i++) { &BODY_20_39($i,@V); unshift(@V,pop(@V)); }
227 lea `16*4`($inp),$inp
230 mov `16*4`(%rsp),%rsi
238 .size sha1_block_data_order,.-sha1_block_data_order
240 .asciz "SHA1 block transform for x86_64, CRYPTOGAMS by <appro\@openssl.org>"
244 # EXCEPTION_DISPOSITION handler (EXCEPTION_RECORD *rec,ULONG64 frame,
245 # CONTEXT *context,DISPATCHER_CONTEXT *disp)
253 .extern __imp_RtlVirtualUnwind
254 .type se_handler,\@abi-omnipotent
268 mov 120($context),%rax # pull context->Rax
269 mov 248($context),%rbx # pull context->Rip
271 lea .Lprologue(%rip),%r10
272 cmp %r10,%rbx # context->Rip<.Lprologue
275 mov 152($context),%rax # pull context->Rsp
277 lea .Lepilogue(%rip),%r10
278 cmp %r10,%rbx # context->Rip>=.Lepilogue
281 mov `16*4`(%rax),%rax # pull saved stack pointer
288 mov %rbx,144($context) # restore context->Rbx
289 mov %rbp,160($context) # restore context->Rbp
290 mov %r12,216($context) # restore context->R12
291 mov %r13,224($context) # restore context->R13
296 mov %rax,152($context) # restore context->Rsp
297 mov %rsi,168($context) # restore context->Rsi
298 mov %rdi,176($context) # restore context->Rdi
300 mov 40($disp),%rdi # disp->ContextRecord
301 mov $context,%rsi # context
302 mov \$154,%ecx # sizeof(CONTEXT)
303 .long 0xa548f3fc # cld; rep movsq
306 xor %rcx,%rcx # arg1, UNW_FLAG_NHANDLER
307 mov 8(%rsi),%rdx # arg2, disp->ImageBase
308 mov 0(%rsi),%r8 # arg3, disp->ControlPc
309 mov 16(%rsi),%r9 # arg4, disp->FunctionEntry
310 mov 40(%rsi),%r10 # disp->ContextRecord
311 lea 56(%rsi),%r11 # &disp->HandlerData
312 lea 24(%rsi),%r12 # &disp->EstablisherFrame
313 mov %r10,32(%rsp) # arg5
314 mov %r11,40(%rsp) # arg6
315 mov %r12,48(%rsp) # arg7
316 mov %rcx,56(%rsp) # arg8, (NULL)
317 call *__imp_RtlVirtualUnwind(%rip)
319 mov \$1,%eax # ExceptionContinueSearch
331 .size se_handler,.-se_handler
335 .rva .LSEH_begin_sha1_block_data_order
336 .rva .LSEH_end_sha1_block_data_order
337 .rva .LSEH_info_sha1_block_data_order
341 .LSEH_info_sha1_block_data_order:
347 ####################################################################
349 $code =~ s/\`([^\`]*)\`/eval $1/gem;