sha1-586.pl revision 127128 
1#!/usr/local/bin/perl
2
3# It was noted that Intel IA-32 C compiler generates code which
4# performs ~30% *faster* on P4 CPU than original *hand-coded*
5# SHA1 assembler implementation. To address this problem (and
6# prove that humans are still better than machines:-), the
7# original code was overhauled, which resulted in following
8# performance changes:
9#
10#		compared with original	compared with Intel cc
11#		assembler impl.		generated code
12# Pentium	-25%			+37%
13# PIII/AMD	+8%			+16%
14# P4		+85%(!)			+45%
15#
16# As you can see Pentium came out as looser:-( Yet I reckoned that
17# improvement on P4 outweights the loss and incorporate this
18# re-tuned code to 0.9.7 and later.
19# ----------------------------------------------------------------
20# Those who for any particular reason absolutely must score on
21# Pentium can replace this module with one from 0.9.6 distribution.
22# This "offer" shall be revoked the moment programming interface to
23# this module is changed, in which case this paragraph should be
24# removed.
25# ----------------------------------------------------------------
26#					<appro@fy.chalmers.se>
27
28$normal=0;
29
30push(@INC,"perlasm","../../perlasm");
31require "x86asm.pl";
32
33&asm_init($ARGV[0],"sha1-586.pl",$ARGV[$#ARGV] eq "386");
34
35$A="eax";
36$B="ecx";
37$C="ebx";
38$D="edx";
39$E="edi";
40$T="esi";
41$tmp1="ebp";
42
43$off=9*4;
44
45@K=(0x5a827999,0x6ed9eba1,0x8f1bbcdc,0xca62c1d6);
46
47&sha1_block_data("sha1_block_asm_data_order");
48
49&asm_finish();
50
51sub Nn
52	{
53	local($p)=@_;
54	local(%n)=($A,$T,$B,$A,$C,$B,$D,$C,$E,$D,$T,$E);
55	return($n{$p});
56	}
57
58sub Np
59	{
60	local($p)=@_;
61	local(%n)=($A,$T,$B,$A,$C,$B,$D,$C,$E,$D,$T,$E);
62	local(%n)=($A,$B,$B,$C,$C,$D,$D,$E,$E,$T,$T,$A);
63	return($n{$p});
64	}
65
66sub Na
67	{
68	local($n)=@_;
69	return( (($n   )&0x0f),
70		(($n+ 2)&0x0f),
71		(($n+ 8)&0x0f),
72		(($n+13)&0x0f),
73		(($n+ 1)&0x0f));
74	}
75
76sub X_expand
77	{
78	local($in)=@_;
79
80	&comment("First, load the words onto the stack in network byte order");
81	for ($i=0; $i<16; $i+=2)
82		{
83		&mov($A,&DWP(($i+0)*4,$in,"",0));# unless $i == 0;
84		 &mov($B,&DWP(($i+1)*4,$in,"",0));
85		&bswap($A);
86		 &bswap($B);
87		&mov(&swtmp($i+0),$A);
88		 &mov(&swtmp($i+1),$B);
89		}
90
91	&comment("We now have the X array on the stack");
92	&comment("starting at sp-4");
93	}
94
95# Rules of engagement
96# F is always trashable at the start, the running total.
97# E becomes the next F so it can be trashed after it has been 'accumulated'
98# F becomes A in the next round.  We don't need to access it much.
99# During the X update part, the result ends up in $X[$n0].
100
101sub BODY_00_15
102	{
103	local($pos,$K,$X,$n,$a,$b,$c,$d,$e,$f)=@_;
104
105	&comment("00_15 $n");
106
107	&mov($tmp1,$a);
108	 &mov($f,$c);			# f to hold F_00_19(b,c,d)
109	&rotl($tmp1,5);			# tmp1=ROTATE(a,5)
110	 &xor($f,$d);
111	&and($f,$b);
112	 &rotr($b,2);			# b=ROTATE(b,30)
113	&add($tmp1,$e);			# tmp1+=e;
114	 &mov($e,&swtmp($n));		# e becomes volatile and
115	 				# is loaded with xi
116	&xor($f,$d);			# f holds F_00_19(b,c,d)
117	 &lea($tmp1,&DWP($K,$tmp1,$e,1));# tmp1+=K_00_19+xi
118
119	&add($f,$tmp1);			# f+=tmp1
120	}
121
122sub BODY_16_19
123	{
124	local($pos,$K,$X,$n,$a,$b,$c,$d,$e,$f)=@_;
125	local($n0,$n1,$n2,$n3,$np)=&Na($n);
126
127	&comment("16_19 $n");
128
129	&mov($f,&swtmp($n1));		# f to hold Xupdate(xi,xa,xb,xc,xd)
130	 &mov($tmp1,$c);		# tmp1 to hold F_00_19(b,c,d)
131	&xor($f,&swtmp($n0));
132	 &xor($tmp1,$d);
133	&xor($f,&swtmp($n2));
134	 &and($tmp1,$b);		# tmp1 holds F_00_19(b,c,d)
135	&xor($f,&swtmp($n3));		# f holds xa^xb^xc^xd
136	 &rotr($b,2);			# b=ROTATE(b,30)
137	&xor($tmp1,$d);			# tmp1=F_00_19(b,c,d)
138	 &rotl($f,1);			# f=ROATE(f,1)
139	&mov(&swtmp($n0),$f);		# xi=f
140	&lea($f,&DWP($K,$f,$e,1));	# f+=K_00_19+e
141	 &mov($e,$a);			# e becomes volatile
142	&add($f,$tmp1);			# f+=F_00_19(b,c,d)
143	 &rotl($e,5);			# e=ROTATE(a,5)
144	&add($f,$e);			# f+=ROTATE(a,5)
145	}
146
147sub BODY_20_39
148	{
149	local($pos,$K,$X,$n,$a,$b,$c,$d,$e,$f)=@_;
150
151	&comment("20_39 $n");
152	local($n0,$n1,$n2,$n3,$np)=&Na($n);
153
154	&mov($f,&swtmp($n0));		# f to hold Xupdate(xi,xa,xb,xc,xd)
155	 &mov($tmp1,$b);		# tmp1 to hold F_20_39(b,c,d)
156	&xor($f,&swtmp($n1));
157	 &rotr($b,2);			# b=ROTATE(b,30)
158	&xor($f,&swtmp($n2));
159	 &xor($tmp1,$c);
160	&xor($f,&swtmp($n3));		# f holds xa^xb^xc^xd
161	 &xor($tmp1,$d);		# tmp1 holds F_20_39(b,c,d)
162	&rotl($f,1);			# f=ROTATE(f,1)
163	&mov(&swtmp($n0),$f);		# xi=f
164	&lea($f,&DWP($K,$f,$e,1));	# f+=K_20_39+e
165	 &mov($e,$a);			# e becomes volatile
166	&rotl($e,5);			# e=ROTATE(a,5)
167	 &add($f,$tmp1);		# f+=F_20_39(b,c,d)
168	&add($f,$e);			# f+=ROTATE(a,5)
169	}
170
171sub BODY_40_59
172	{
173	local($pos,$K,$X,$n,$a,$b,$c,$d,$e,$f)=@_;
174
175	&comment("40_59 $n");
176	local($n0,$n1,$n2,$n3,$np)=&Na($n);
177
178	&mov($f,&swtmp($n0));		# f to hold Xupdate(xi,xa,xb,xc,xd)
179	 &mov($tmp1,$b);		# tmp1 to hold F_40_59(b,c,d)
180	&xor($f,&swtmp($n1));
181	 &or($tmp1,$c);
182	&xor($f,&swtmp($n2));
183	 &and($tmp1,$d);
184	&xor($f,&swtmp($n3));		# f holds xa^xb^xc^xd
185	&rotl($f,1);			# f=ROTATE(f,1)
186	&mov(&swtmp($n0),$f);		# xi=f
187	&lea($f,&DWP($K,$f,$e,1));	# f+=K_40_59+e
188	 &mov($e,$b);			# e becomes volatile and is used
189					# to calculate F_40_59(b,c,d)
190	&rotr($b,2);			# b=ROTATE(b,30)
191	 &and($e,$c);
192	&or($tmp1,$e);			# tmp1 holds F_40_59(b,c,d)
193	 &mov($e,$a);
194	&rotl($e,5);			# e=ROTATE(a,5)
195	&add($tmp1,$e);			# tmp1+=ROTATE(a,5)
196	&add($f,$tmp1);			# f+=tmp1;
197	}
198
199sub BODY_60_79
200	{
201	&BODY_20_39(@_);
202	}
203
204sub sha1_block_host
205	{
206	local($name, $sclabel)=@_;
207
208	&function_begin_B($name,"");
209
210	# parameter 1 is the MD5_CTX structure.
211	# A	0
212	# B	4
213	# C	8
214	# D 	12
215	# E 	16
216
217	&mov("ecx",	&wparam(2));
218	 &push("esi");
219	&shl("ecx",6);
220	 &mov("esi",	&wparam(1));
221	&push("ebp");
222	 &add("ecx","esi");	# offset to leave on
223	&push("ebx");
224	 &mov("ebp",	&wparam(0));
225	&push("edi");
226	 &mov($D,	&DWP(12,"ebp","",0));
227	&stack_push(18+9);
228	 &mov($E,	&DWP(16,"ebp","",0));
229	&mov($C,	&DWP( 8,"ebp","",0));
230	 &mov(&swtmp(17),"ecx");
231
232	&comment("First we need to setup the X array");
233
234	for ($i=0; $i<16; $i+=2)
235		{
236		&mov($A,&DWP(($i+0)*4,"esi","",0));# unless $i == 0;
237		 &mov($B,&DWP(($i+1)*4,"esi","",0));
238		&mov(&swtmp($i+0),$A);
239		 &mov(&swtmp($i+1),$B);
240		}
241	&jmp($sclabel);
242	&function_end_B($name);
243	}
244
245
246sub sha1_block_data
247	{
248	local($name)=@_;
249
250	&function_begin_B($name,"");
251
252	# parameter 1 is the MD5_CTX structure.
253	# A	0
254	# B	4
255	# C	8
256	# D 	12
257	# E 	16
258
259	&mov("ecx",	&wparam(2));
260	 &push("esi");
261	&shl("ecx",6);
262	 &mov("esi",	&wparam(1));
263	&push("ebp");
264	 &add("ecx","esi");	# offset to leave on
265	&push("ebx");
266	 &mov("ebp",	&wparam(0));
267	&push("edi");
268	 &mov($D,	&DWP(12,"ebp","",0));
269	&stack_push(18+9);
270	 &mov($E,	&DWP(16,"ebp","",0));
271	&mov($C,	&DWP( 8,"ebp","",0));
272	 &mov(&swtmp(17),"ecx");
273
274	&comment("First we need to setup the X array");
275
276	&set_label("start") unless $normal;
277
278	&X_expand("esi");
279	 &mov(&wparam(1),"esi");
280
281	&set_label("shortcut", 0, 1);
282	&comment("");
283	&comment("Start processing");
284
285	# odd start
286	&mov($A,	&DWP( 0,"ebp","",0));
287	 &mov($B,	&DWP( 4,"ebp","",0));
288	$X="esp";
289	&BODY_00_15(-2,$K[0],$X, 0,$A,$B,$C,$D,$E,$T);
290	&BODY_00_15( 0,$K[0],$X, 1,$T,$A,$B,$C,$D,$E);
291	&BODY_00_15( 0,$K[0],$X, 2,$E,$T,$A,$B,$C,$D);
292	&BODY_00_15( 0,$K[0],$X, 3,$D,$E,$T,$A,$B,$C);
293	&BODY_00_15( 0,$K[0],$X, 4,$C,$D,$E,$T,$A,$B);
294	&BODY_00_15( 0,$K[0],$X, 5,$B,$C,$D,$E,$T,$A);
295	&BODY_00_15( 0,$K[0],$X, 6,$A,$B,$C,$D,$E,$T);
296	&BODY_00_15( 0,$K[0],$X, 7,$T,$A,$B,$C,$D,$E);
297	&BODY_00_15( 0,$K[0],$X, 8,$E,$T,$A,$B,$C,$D);
298	&BODY_00_15( 0,$K[0],$X, 9,$D,$E,$T,$A,$B,$C);
299	&BODY_00_15( 0,$K[0],$X,10,$C,$D,$E,$T,$A,$B);
300	&BODY_00_15( 0,$K[0],$X,11,$B,$C,$D,$E,$T,$A);
301	&BODY_00_15( 0,$K[0],$X,12,$A,$B,$C,$D,$E,$T);
302	&BODY_00_15( 0,$K[0],$X,13,$T,$A,$B,$C,$D,$E);
303	&BODY_00_15( 0,$K[0],$X,14,$E,$T,$A,$B,$C,$D);
304	&BODY_00_15( 1,$K[0],$X,15,$D,$E,$T,$A,$B,$C);
305	&BODY_16_19(-1,$K[0],$X,16,$C,$D,$E,$T,$A,$B);
306	&BODY_16_19( 0,$K[0],$X,17,$B,$C,$D,$E,$T,$A);
307	&BODY_16_19( 0,$K[0],$X,18,$A,$B,$C,$D,$E,$T);
308	&BODY_16_19( 1,$K[0],$X,19,$T,$A,$B,$C,$D,$E);
309
310	&BODY_20_39(-1,$K[1],$X,20,$E,$T,$A,$B,$C,$D);
311	&BODY_20_39( 0,$K[1],$X,21,$D,$E,$T,$A,$B,$C);
312	&BODY_20_39( 0,$K[1],$X,22,$C,$D,$E,$T,$A,$B);
313	&BODY_20_39( 0,$K[1],$X,23,$B,$C,$D,$E,$T,$A);
314	&BODY_20_39( 0,$K[1],$X,24,$A,$B,$C,$D,$E,$T);
315	&BODY_20_39( 0,$K[1],$X,25,$T,$A,$B,$C,$D,$E);
316	&BODY_20_39( 0,$K[1],$X,26,$E,$T,$A,$B,$C,$D);
317	&BODY_20_39( 0,$K[1],$X,27,$D,$E,$T,$A,$B,$C);
318	&BODY_20_39( 0,$K[1],$X,28,$C,$D,$E,$T,$A,$B);
319	&BODY_20_39( 0,$K[1],$X,29,$B,$C,$D,$E,$T,$A);
320	&BODY_20_39( 0,$K[1],$X,30,$A,$B,$C,$D,$E,$T);
321	&BODY_20_39( 0,$K[1],$X,31,$T,$A,$B,$C,$D,$E);
322	&BODY_20_39( 0,$K[1],$X,32,$E,$T,$A,$B,$C,$D);
323	&BODY_20_39( 0,$K[1],$X,33,$D,$E,$T,$A,$B,$C);
324	&BODY_20_39( 0,$K[1],$X,34,$C,$D,$E,$T,$A,$B);
325	&BODY_20_39( 0,$K[1],$X,35,$B,$C,$D,$E,$T,$A);
326	&BODY_20_39( 0,$K[1],$X,36,$A,$B,$C,$D,$E,$T);
327	&BODY_20_39( 0,$K[1],$X,37,$T,$A,$B,$C,$D,$E);
328	&BODY_20_39( 0,$K[1],$X,38,$E,$T,$A,$B,$C,$D);
329	&BODY_20_39( 1,$K[1],$X,39,$D,$E,$T,$A,$B,$C);
330
331	&BODY_40_59(-1,$K[2],$X,40,$C,$D,$E,$T,$A,$B);
332	&BODY_40_59( 0,$K[2],$X,41,$B,$C,$D,$E,$T,$A);
333	&BODY_40_59( 0,$K[2],$X,42,$A,$B,$C,$D,$E,$T);
334	&BODY_40_59( 0,$K[2],$X,43,$T,$A,$B,$C,$D,$E);
335	&BODY_40_59( 0,$K[2],$X,44,$E,$T,$A,$B,$C,$D);
336	&BODY_40_59( 0,$K[2],$X,45,$D,$E,$T,$A,$B,$C);
337	&BODY_40_59( 0,$K[2],$X,46,$C,$D,$E,$T,$A,$B);
338	&BODY_40_59( 0,$K[2],$X,47,$B,$C,$D,$E,$T,$A);
339	&BODY_40_59( 0,$K[2],$X,48,$A,$B,$C,$D,$E,$T);
340	&BODY_40_59( 0,$K[2],$X,49,$T,$A,$B,$C,$D,$E);
341	&BODY_40_59( 0,$K[2],$X,50,$E,$T,$A,$B,$C,$D);
342	&BODY_40_59( 0,$K[2],$X,51,$D,$E,$T,$A,$B,$C);
343	&BODY_40_59( 0,$K[2],$X,52,$C,$D,$E,$T,$A,$B);
344	&BODY_40_59( 0,$K[2],$X,53,$B,$C,$D,$E,$T,$A);
345	&BODY_40_59( 0,$K[2],$X,54,$A,$B,$C,$D,$E,$T);
346	&BODY_40_59( 0,$K[2],$X,55,$T,$A,$B,$C,$D,$E);
347	&BODY_40_59( 0,$K[2],$X,56,$E,$T,$A,$B,$C,$D);
348	&BODY_40_59( 0,$K[2],$X,57,$D,$E,$T,$A,$B,$C);
349	&BODY_40_59( 0,$K[2],$X,58,$C,$D,$E,$T,$A,$B);
350	&BODY_40_59( 1,$K[2],$X,59,$B,$C,$D,$E,$T,$A);
351
352	&BODY_60_79(-1,$K[3],$X,60,$A,$B,$C,$D,$E,$T);
353	&BODY_60_79( 0,$K[3],$X,61,$T,$A,$B,$C,$D,$E);
354	&BODY_60_79( 0,$K[3],$X,62,$E,$T,$A,$B,$C,$D);
355	&BODY_60_79( 0,$K[3],$X,63,$D,$E,$T,$A,$B,$C);
356	&BODY_60_79( 0,$K[3],$X,64,$C,$D,$E,$T,$A,$B);
357	&BODY_60_79( 0,$K[3],$X,65,$B,$C,$D,$E,$T,$A);
358	&BODY_60_79( 0,$K[3],$X,66,$A,$B,$C,$D,$E,$T);
359	&BODY_60_79( 0,$K[3],$X,67,$T,$A,$B,$C,$D,$E);
360	&BODY_60_79( 0,$K[3],$X,68,$E,$T,$A,$B,$C,$D);
361	&BODY_60_79( 0,$K[3],$X,69,$D,$E,$T,$A,$B,$C);
362	&BODY_60_79( 0,$K[3],$X,70,$C,$D,$E,$T,$A,$B);
363	&BODY_60_79( 0,$K[3],$X,71,$B,$C,$D,$E,$T,$A);
364	&BODY_60_79( 0,$K[3],$X,72,$A,$B,$C,$D,$E,$T);
365	&BODY_60_79( 0,$K[3],$X,73,$T,$A,$B,$C,$D,$E);
366	&BODY_60_79( 0,$K[3],$X,74,$E,$T,$A,$B,$C,$D);
367	&BODY_60_79( 0,$K[3],$X,75,$D,$E,$T,$A,$B,$C);
368	&BODY_60_79( 0,$K[3],$X,76,$C,$D,$E,$T,$A,$B);
369	&BODY_60_79( 0,$K[3],$X,77,$B,$C,$D,$E,$T,$A);
370	&BODY_60_79( 0,$K[3],$X,78,$A,$B,$C,$D,$E,$T);
371	&BODY_60_79( 2,$K[3],$X,79,$T,$A,$B,$C,$D,$E);
372
373	&comment("End processing");
374	&comment("");
375	# D is the tmp value
376
377	# E -> A
378	# T -> B
379	# A -> C
380	# B -> D
381	# C -> E
382	# D -> T
383
384	&mov($tmp1,&wparam(0));
385
386	 &mov($D,	&DWP(12,$tmp1,"",0));
387	&add($D,$B);
388	 &mov($B,	&DWP( 4,$tmp1,"",0));
389	&add($B,$T);
390	 &mov($T,	$A);
391	&mov($A,	&DWP( 0,$tmp1,"",0));
392	 &mov(&DWP(12,$tmp1,"",0),$D);
393
394	&add($A,$E);
395	 &mov($E,	&DWP(16,$tmp1,"",0));
396	&add($E,$C);
397	 &mov($C,	&DWP( 8,$tmp1,"",0));
398	&add($C,$T);
399
400	 &mov(&DWP( 0,$tmp1,"",0),$A);
401	&mov("esi",&wparam(1));
402	 &mov(&DWP( 8,$tmp1,"",0),$C);
403 	&add("esi",64);
404	 &mov("eax",&swtmp(17));
405	&mov(&DWP(16,$tmp1,"",0),$E);
406	 &cmp("esi","eax");
407	&mov(&DWP( 4,$tmp1,"",0),$B);
408	 &jl(&label("start"));
409
410	&stack_pop(18+9);
411	 &pop("edi");
412	&pop("ebx");
413	 &pop("ebp");
414	&pop("esi");
415	 &ret();
416
417	# keep a note of shortcut label so it can be used outside
418	# block.
419	my $sclabel = &label("shortcut");
420
421	&function_end_B($name);
422	# Putting this here avoids problems with MASM in debugging mode
423	&sha1_block_host("sha1_block_asm_host_order", $sclabel);
424	}
425
426