• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1#! /usr/bin/env perl
2# Copyright 2011-2018 The OpenSSL Project Authors. All Rights Reserved.
3#
4# Licensed under the OpenSSL license (the "License").  You may not use
5# this file except in compliance with the License.  You can obtain a copy
6# in the file LICENSE in the source distribution or at
7# https://www.openssl.org/source/license.html
8
9
10# ====================================================================
11# Written by Andy Polyakov <appro@openssl.org> for the OpenSSL
12# project. The module is, however, dual licensed under OpenSSL and
13# CRYPTOGAMS licenses depending on where you obtain it. For further
14# details see http://www.openssl.org/~appro/cryptogams/.
15# ====================================================================
16
17# September 2011
18#
19# Assembler helpers for Padlock engine. Compared to original engine
20# version relying on inline assembler and compiled with gcc 3.4.6 it
21# was measured to provide ~100% improvement on misaligned data in ECB
22# mode and ~75% in CBC mode. For aligned data improvement can be
23# observed for short inputs only, e.g. 45% for 64-byte messages in
24# ECB mode, 20% in CBC. Difference in performance for aligned vs.
25# misaligned data depends on misalignment and is either ~1.8x or 2.9x.
26# These are approximately same factors as for hardware support, so
27# there is little reason to rely on the latter. On the contrary, it
28# might actually hurt performance in mixture of aligned and misaligned
29# buffers, because a) if you choose to flip 'align' flag in control
30# word on per-buffer basis, then you'd have to reload key context,
31# which incurs penalty; b) if you choose to set 'align' flag
32# permanently, it limits performance even for aligned data to ~1/2.
33# All above mentioned results were collected on 1.5GHz C7. Nano on the
34# other hand handles unaligned data more gracefully. Depending on
35# algorithm and how unaligned data is, hardware can be up to 70% more
36# efficient than below software alignment procedures, nor does 'align'
37# flag have affect on aligned performance [if has any meaning at all].
38# Therefore suggestion is to unconditionally set 'align' flag on Nano
39# for optimal performance.
40
41$0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1;
42push(@INC,"${dir}","${dir}../../crypto/perlasm");
43require "x86asm.pl";
44
45$output=pop;
46open STDOUT,">$output";
47
48&asm_init($ARGV[0]);
49
50%PADLOCK_PREFETCH=(ecb=>128, cbc=>64);	# prefetch errata
51$PADLOCK_CHUNK=512;	# Must be a power of 2 larger than 16
52
53$ctx="edx";
54$out="edi";
55$inp="esi";
56$len="ecx";
57$chunk="ebx";
58
59&function_begin_B("padlock_capability");
60	&push	("ebx");
61	&pushf	();
62	&pop	("eax");
63	&mov	("ecx","eax");
64	&xor	("eax",1<<21);
65	&push	("eax");
66	&popf	();
67	&pushf	();
68	&pop	("eax");
69	&xor	("ecx","eax");
70	&xor	("eax","eax");
71	&bt	("ecx",21);
72	&jnc	(&label("noluck"));
73	&cpuid	();
74	&xor	("eax","eax");
75	&cmp	("ebx","0x".unpack("H*",'tneC'));
76	&jne	(&label("zhaoxin"));
77	&cmp	("edx","0x".unpack("H*",'Hrua'));
78	&jne	(&label("noluck"));
79	&cmp	("ecx","0x".unpack("H*",'slua'));
80	&jne	(&label("noluck"));
81	&jmp	(&label("zhaoxinEnd"));
82&set_label("zhaoxin");
83	&cmp	("ebx","0x".unpack("H*",'hS  '));
84	&jne	(&label("noluck"));
85	&cmp	("edx","0x".unpack("H*",'hgna'));
86	&jne	(&label("noluck"));
87	&cmp	("ecx","0x".unpack("H*",'  ia'));
88	&jne	(&label("noluck"));
89&set_label("zhaoxinEnd");
90	&mov	("eax",0xC0000000);
91	&cpuid	();
92	&mov	("edx","eax");
93	&xor	("eax","eax");
94	&cmp	("edx",0xC0000001);
95	&jb	(&label("noluck"));
96	&mov	("eax",1);
97	&cpuid	();
98	&or	("eax",0x0f);
99	&xor	("ebx","ebx");
100	&and	("eax",0x0fff);
101	&cmp	("eax",0x06ff);		# check for Nano
102	&sete	("bl");
103	&mov	("eax",0xC0000001);
104	&push	("ebx");
105	&cpuid	();
106	&pop	("ebx");
107	&mov	("eax","edx");
108	&shl	("ebx",4);		# bit#4 denotes Nano
109	&and	("eax",0xffffffef);
110	&or	("eax","ebx")
111&set_label("noluck");
112	&pop	("ebx");
113	&ret	();
114&function_end_B("padlock_capability")
115
116&function_begin_B("padlock_key_bswap");
117	&mov	("edx",&wparam(0));
118	&mov	("ecx",&DWP(240,"edx"));
119&set_label("bswap_loop");
120	&mov	("eax",&DWP(0,"edx"));
121	&bswap	("eax");
122	&mov	(&DWP(0,"edx"),"eax");
123	&lea	("edx",&DWP(4,"edx"));
124	&sub	("ecx",1);
125	&jnz	(&label("bswap_loop"));
126	&ret	();
127&function_end_B("padlock_key_bswap");
128
129# This is heuristic key context tracing. At first one
130# believes that one should use atomic swap instructions,
131# but it's not actually necessary. Point is that if
132# padlock_saved_context was changed by another thread
133# after we've read it and before we compare it with ctx,
134# our key *shall* be reloaded upon thread context switch
135# and we are therefore set in either case...
136&static_label("padlock_saved_context");
137
138&function_begin_B("padlock_verify_context");
139	&mov	($ctx,&wparam(0));
140	&lea	("eax",($::win32 or $::coff) ? &DWP(&label("padlock_saved_context")) :
141		       &DWP(&label("padlock_saved_context")."-".&label("verify_pic_point")));
142	&pushf	();
143	&call	("_padlock_verify_ctx");
144&set_label("verify_pic_point");
145	&lea	("esp",&DWP(4,"esp"));
146	&ret	();
147&function_end_B("padlock_verify_context");
148
149&function_begin_B("_padlock_verify_ctx");
150	&add	("eax",&DWP(0,"esp")) if(!($::win32 or $::coff));# &padlock_saved_context
151	&bt	(&DWP(4,"esp"),30);		# eflags
152	&jnc	(&label("verified"));
153	&cmp	($ctx,&DWP(0,"eax"));
154	&je	(&label("verified"));
155	&pushf	();
156	&popf	();
157&set_label("verified");
158	&mov	(&DWP(0,"eax"),$ctx);
159	&ret	();
160&function_end_B("_padlock_verify_ctx");
161
162&function_begin_B("padlock_reload_key");
163	&pushf	();
164	&popf	();
165	&ret	();
166&function_end_B("padlock_reload_key");
167
168&function_begin_B("padlock_aes_block");
169	&push	("edi");
170	&push	("esi");
171	&push	("ebx");
172	&mov	($out,&wparam(0));		# must be 16-byte aligned
173	&mov	($inp,&wparam(1));		# must be 16-byte aligned
174	&mov	($ctx,&wparam(2));
175	&mov	($len,1);
176	&lea	("ebx",&DWP(32,$ctx));		# key
177	&lea	($ctx,&DWP(16,$ctx));		# control word
178	&data_byte(0xf3,0x0f,0xa7,0xc8);	# rep xcryptecb
179	&pop	("ebx");
180	&pop	("esi");
181	&pop	("edi");
182	&ret	();
183&function_end_B("padlock_aes_block");
184
185sub generate_mode {
186my ($mode,$opcode) = @_;
187# int padlock_$mode_encrypt(void *out, const void *inp,
188#		struct padlock_cipher_data *ctx, size_t len);
189&function_begin("padlock_${mode}_encrypt");
190	&mov	($out,&wparam(0));
191	&mov	($inp,&wparam(1));
192	&mov	($ctx,&wparam(2));
193	&mov	($len,&wparam(3));
194	&test	($ctx,15);
195	&jnz	(&label("${mode}_abort"));
196	&test	($len,15);
197	&jnz	(&label("${mode}_abort"));
198	&lea	("eax",($::win32 or $::coff) ? &DWP(&label("padlock_saved_context")) :
199		       &DWP(&label("padlock_saved_context")."-".&label("${mode}_pic_point")));
200	&pushf	();
201	&cld	();
202	&call	("_padlock_verify_ctx");
203&set_label("${mode}_pic_point");
204	&lea	($ctx,&DWP(16,$ctx));	# control word
205	&xor	("eax","eax");
206					if ($mode eq "ctr32") {
207	&movq	("mm0",&QWP(-16,$ctx));	# load [upper part of] counter
208					} else {
209	&xor	("ebx","ebx");
210	&test	(&DWP(0,$ctx),1<<5);	# align bit in control word
211	&jnz	(&label("${mode}_aligned"));
212	&test	($out,0x0f);
213	&setz	("al");			# !out_misaligned
214	&test	($inp,0x0f);
215	&setz	("bl");			# !inp_misaligned
216	&test	("eax","ebx");
217	&jnz	(&label("${mode}_aligned"));
218	&neg	("eax");
219					}
220	&mov	($chunk,$PADLOCK_CHUNK);
221	&not	("eax");		# out_misaligned?-1:0
222	&lea	("ebp",&DWP(-24,"esp"));
223	&cmp	($len,$chunk);
224	&cmovc	($chunk,$len);		# chunk=len>PADLOCK_CHUNK?PADLOCK_CHUNK:len
225	&and	("eax",$chunk);		# out_misaligned?chunk:0
226	&mov	($chunk,$len);
227	&neg	("eax");
228	&and	($chunk,$PADLOCK_CHUNK-1);	# chunk=len%PADLOCK_CHUNK
229	&lea	("esp",&DWP(0,"eax","ebp"));	# alloca
230	&mov	("eax",$PADLOCK_CHUNK);
231	&cmovz	($chunk,"eax");			# chunk=chunk?:PADLOCK_CHUNK
232	&mov	("eax","ebp");
233	&and	("ebp",-16);
234	&and	("esp",-16);
235	&mov	(&DWP(16,"ebp"),"eax");
236    if ($PADLOCK_PREFETCH{$mode}) {
237	&cmp	($len,$chunk);
238	&ja	(&label("${mode}_loop"));
239	&mov	("eax",$inp);		# check if prefetch crosses page
240	&cmp	("ebp","esp");
241	&cmove	("eax",$out);
242	&add	("eax",$len);
243	&neg	("eax");
244	&and	("eax",0xfff);		# distance to page boundary
245	&cmp	("eax",$PADLOCK_PREFETCH{$mode});
246	&mov	("eax",-$PADLOCK_PREFETCH{$mode});
247	&cmovae	("eax",$chunk);		# mask=distance<prefetch?-prefetch:-1
248	&and	($chunk,"eax");
249	&jz	(&label("${mode}_unaligned_tail"));
250    }
251	&jmp	(&label("${mode}_loop"));
252
253&set_label("${mode}_loop",16);
254	&mov	(&DWP(0,"ebp"),$out);		# save parameters
255	&mov	(&DWP(4,"ebp"),$inp);
256	&mov	(&DWP(8,"ebp"),$len);
257	&mov	($len,$chunk);
258	&mov	(&DWP(12,"ebp"),$chunk);	# chunk
259						if ($mode eq "ctr32") {
260	&mov	("ecx",&DWP(-4,$ctx));
261	&xor	($out,$out);
262	&mov	("eax",&DWP(-8,$ctx));		# borrow $len
263&set_label("${mode}_prepare");
264	&mov	(&DWP(12,"esp",$out),"ecx");
265	&bswap	("ecx");
266	&movq	(&QWP(0,"esp",$out),"mm0");
267	&inc	("ecx");
268	&mov	(&DWP(8,"esp",$out),"eax");
269	&bswap	("ecx");
270	&lea	($out,&DWP(16,$out));
271	&cmp	($out,$chunk);
272	&jb	(&label("${mode}_prepare"));
273
274	&mov	(&DWP(-4,$ctx),"ecx");
275	&lea	($inp,&DWP(0,"esp"));
276	&lea	($out,&DWP(0,"esp"));
277	&mov	($len,$chunk);
278						} else {
279	&test	($out,0x0f);			# out_misaligned
280	&cmovnz	($out,"esp");
281	&test	($inp,0x0f);			# inp_misaligned
282	&jz	(&label("${mode}_inp_aligned"));
283	&shr	($len,2);
284	&data_byte(0xf3,0xa5);			# rep movsl
285	&sub	($out,$chunk);
286	&mov	($len,$chunk);
287	&mov	($inp,$out);
288&set_label("${mode}_inp_aligned");
289						}
290	&lea	("eax",&DWP(-16,$ctx));		# ivp
291	&lea	("ebx",&DWP(16,$ctx));		# key
292	&shr	($len,4);			# len/=AES_BLOCK_SIZE
293	&data_byte(0xf3,0x0f,0xa7,$opcode);	# rep xcrypt*
294						if ($mode !~ /ecb|ctr/) {
295	&movaps	("xmm0",&QWP(0,"eax"));
296	&movaps	(&QWP(-16,$ctx),"xmm0");	# copy [or refresh] iv
297						}
298	&mov	($out,&DWP(0,"ebp"));		# restore parameters
299	&mov	($chunk,&DWP(12,"ebp"));
300						if ($mode eq "ctr32") {
301	&mov	($inp,&DWP(4,"ebp"));
302	&xor	($len,$len);
303&set_label("${mode}_xor");
304	&movups	("xmm1",&QWP(0,$inp,$len));
305	&lea	($len,&DWP(16,$len));
306	&pxor	("xmm1",&QWP(-16,"esp",$len));
307	&movups	(&QWP(-16,$out,$len),"xmm1");
308	&cmp	($len,$chunk);
309	&jb	(&label("${mode}_xor"));
310						} else {
311	&test	($out,0x0f);
312	&jz	(&label("${mode}_out_aligned"));
313	&mov	($len,$chunk);
314	&lea	($inp,&DWP(0,"esp"));
315	&shr	($len,2);
316	&data_byte(0xf3,0xa5);			# rep movsl
317	&sub	($out,$chunk);
318&set_label("${mode}_out_aligned");
319	&mov	($inp,&DWP(4,"ebp"));
320						}
321	&mov	($len,&DWP(8,"ebp"));
322	&add	($out,$chunk);
323	&add	($inp,$chunk);
324	&sub	($len,$chunk);
325	&mov	($chunk,$PADLOCK_CHUNK);
326    if (!$PADLOCK_PREFETCH{$mode}) {
327	&jnz	(&label("${mode}_loop"));
328    } else {
329	&jz	(&label("${mode}_break"));
330	&cmp	($len,$chunk);
331	&jae	(&label("${mode}_loop"));
332
333&set_label("${mode}_unaligned_tail");
334	&xor	("eax","eax");
335	&cmp	("esp","ebp");
336	&cmove	("eax",$len);
337	&sub	("esp","eax");			# alloca
338	&mov	("eax", $out);			# save parameters
339	&mov	($chunk,$len);
340	&shr	($len,2);
341	&lea	($out,&DWP(0,"esp"));
342	&data_byte(0xf3,0xa5);			# rep movsl
343	&mov	($inp,"esp");
344	&mov	($out,"eax");			# restore parameters
345	&mov	($len,$chunk);
346	&jmp	(&label("${mode}_loop"));
347
348&set_label("${mode}_break",16);
349    }
350						if ($mode ne "ctr32") {
351	&cmp	("esp","ebp");
352	&je	(&label("${mode}_done"));
353						}
354	&pxor	("xmm0","xmm0");
355	&lea	("eax",&DWP(0,"esp"));
356&set_label("${mode}_bzero");
357	&movaps	(&QWP(0,"eax"),"xmm0");
358	&lea	("eax",&DWP(16,"eax"));
359	&cmp	("ebp","eax");
360	&ja	(&label("${mode}_bzero"));
361
362&set_label("${mode}_done");
363	&mov	("ebp",&DWP(16,"ebp"));
364	&lea	("esp",&DWP(24,"ebp"));
365						if ($mode ne "ctr32") {
366	&jmp	(&label("${mode}_exit"));
367
368&set_label("${mode}_aligned",16);
369    if ($PADLOCK_PREFETCH{$mode}) {
370	&lea	("ebp",&DWP(0,$inp,$len));
371	&neg	("ebp");
372	&and	("ebp",0xfff);			# distance to page boundary
373	&xor	("eax","eax");
374	&cmp	("ebp",$PADLOCK_PREFETCH{$mode});
375	&mov	("ebp",$PADLOCK_PREFETCH{$mode}-1);
376	&cmovae	("ebp","eax");
377	&and	("ebp",$len);			# remainder
378	&sub	($len,"ebp");
379	&jz	(&label("${mode}_aligned_tail"));
380    }
381	&lea	("eax",&DWP(-16,$ctx));		# ivp
382	&lea	("ebx",&DWP(16,$ctx));		# key
383	&shr	($len,4);			# len/=AES_BLOCK_SIZE
384	&data_byte(0xf3,0x0f,0xa7,$opcode);	# rep xcrypt*
385						if ($mode ne "ecb") {
386	&movaps	("xmm0",&QWP(0,"eax"));
387	&movaps	(&QWP(-16,$ctx),"xmm0");	# copy [or refresh] iv
388						}
389    if ($PADLOCK_PREFETCH{$mode}) {
390	&test	("ebp","ebp");
391	&jz	(&label("${mode}_exit"));
392
393&set_label("${mode}_aligned_tail");
394	&mov	($len,"ebp");
395	&lea	("ebp",&DWP(-24,"esp"));
396	&mov	("esp","ebp");
397	&mov	("eax","ebp");
398	&sub	("esp",$len);
399	&and	("ebp",-16);
400	&and	("esp",-16);
401	&mov	(&DWP(16,"ebp"),"eax");
402	&mov	("eax", $out);			# save parameters
403	&mov	($chunk,$len);
404	&shr	($len,2);
405	&lea	($out,&DWP(0,"esp"));
406	&data_byte(0xf3,0xa5);			# rep movsl
407	&mov	($inp,"esp");
408	&mov	($out,"eax");			# restore parameters
409	&mov	($len,$chunk);
410	&jmp	(&label("${mode}_loop"));
411    }
412&set_label("${mode}_exit");			}
413	&mov	("eax",1);
414	&lea	("esp",&DWP(4,"esp"));		# popf
415	&emms	()				if ($mode eq "ctr32");
416&set_label("${mode}_abort");
417&function_end("padlock_${mode}_encrypt");
418}
419
420&generate_mode("ecb",0xc8);
421&generate_mode("cbc",0xd0);
422&generate_mode("cfb",0xe0);
423&generate_mode("ofb",0xe8);
424&generate_mode("ctr32",0xc8);	# yes, it implements own CTR with ECB opcode,
425				# because hardware CTR was introduced later
426				# and even has errata on certain C7 stepping.
427				# own implementation *always* works, though
428				# ~15% slower than dedicated hardware...
429
430&function_begin_B("padlock_xstore");
431	&push	("edi");
432	&mov	("edi",&wparam(0));
433	&mov	("edx",&wparam(1));
434	&data_byte(0x0f,0xa7,0xc0);		# xstore
435	&pop	("edi");
436	&ret	();
437&function_end_B("padlock_xstore");
438
439&function_begin_B("_win32_segv_handler");
440	&mov	("eax",1);			# ExceptionContinueSearch
441	&mov	("edx",&wparam(0));		# *ExceptionRecord
442	&mov	("ecx",&wparam(2));		# *ContextRecord
443	&cmp	(&DWP(0,"edx"),0xC0000005)	# ExceptionRecord->ExceptionCode == STATUS_ACCESS_VIOLATION
444	&jne	(&label("ret"));
445	&add	(&DWP(184,"ecx"),4);		# skip over rep sha*
446	&mov	("eax",0);			# ExceptionContinueExecution
447&set_label("ret");
448	&ret	();
449&function_end_B("_win32_segv_handler");
450&safeseh("_win32_segv_handler")			if ($::win32);
451
452&function_begin_B("padlock_sha1_oneshot");
453	&push	("edi");
454	&push	("esi");
455	&xor	("eax","eax");
456	&mov	("edi",&wparam(0));
457	&mov	("esi",&wparam(1));
458	&mov	("ecx",&wparam(2));
459    if ($::win32 or $::coff) {
460    	&push	(&::islabel("_win32_segv_handler"));
461	&data_byte(0x64,0xff,0x30);		# push	%fs:(%eax)
462	&data_byte(0x64,0x89,0x20);		# mov	%esp,%fs:(%eax)
463    }
464	&mov	("edx","esp");			# put aside %esp
465	&add	("esp",-128);			# 32 is enough but spec says 128
466	&movups	("xmm0",&QWP(0,"edi"));		# copy-in context
467	&and	("esp",-16);
468	&mov	("eax",&DWP(16,"edi"));
469	&movaps	(&QWP(0,"esp"),"xmm0");
470	&mov	("edi","esp");
471	&mov	(&DWP(16,"esp"),"eax");
472	&xor	("eax","eax");
473	&data_byte(0xf3,0x0f,0xa6,0xc8);	# rep xsha1
474	&movaps	("xmm0",&QWP(0,"esp"));
475	&mov	("eax",&DWP(16,"esp"));
476	&mov	("esp","edx");			# restore %esp
477    if ($::win32 or $::coff) {
478	&data_byte(0x64,0x8f,0x05,0,0,0,0);	# pop	%fs:0
479	&lea	("esp",&DWP(4,"esp"));
480    }
481	&mov	("edi",&wparam(0));
482	&movups	(&QWP(0,"edi"),"xmm0");		# copy-out context
483	&mov	(&DWP(16,"edi"),"eax");
484	&pop	("esi");
485	&pop	("edi");
486	&ret	();
487&function_end_B("padlock_sha1_oneshot");
488
489&function_begin_B("padlock_sha1_blocks");
490	&push	("edi");
491	&push	("esi");
492	&mov	("edi",&wparam(0));
493	&mov	("esi",&wparam(1));
494	&mov	("edx","esp");			# put aside %esp
495	&mov	("ecx",&wparam(2));
496	&add	("esp",-128);
497	&movups	("xmm0",&QWP(0,"edi"));		# copy-in context
498	&and	("esp",-16);
499	&mov	("eax",&DWP(16,"edi"));
500	&movaps	(&QWP(0,"esp"),"xmm0");
501	&mov	("edi","esp");
502	&mov	(&DWP(16,"esp"),"eax");
503	&mov	("eax",-1);
504	&data_byte(0xf3,0x0f,0xa6,0xc8);	# rep xsha1
505	&movaps	("xmm0",&QWP(0,"esp"));
506	&mov	("eax",&DWP(16,"esp"));
507	&mov	("esp","edx");			# restore %esp
508	&mov	("edi",&wparam(0));
509	&movups	(&QWP(0,"edi"),"xmm0");		# copy-out context
510	&mov	(&DWP(16,"edi"),"eax");
511 	&pop	("esi");
512	&pop	("edi");
513	&ret	();
514&function_end_B("padlock_sha1_blocks");
515
516&function_begin_B("padlock_sha256_oneshot");
517	&push	("edi");
518	&push	("esi");
519	&xor	("eax","eax");
520	&mov	("edi",&wparam(0));
521	&mov	("esi",&wparam(1));
522	&mov	("ecx",&wparam(2));
523    if ($::win32 or $::coff) {
524    	&push	(&::islabel("_win32_segv_handler"));
525	&data_byte(0x64,0xff,0x30);		# push	%fs:(%eax)
526	&data_byte(0x64,0x89,0x20);		# mov	%esp,%fs:(%eax)
527    }
528	&mov	("edx","esp");			# put aside %esp
529	&add	("esp",-128);
530	&movups	("xmm0",&QWP(0,"edi"));		# copy-in context
531	&and	("esp",-16);
532	&movups	("xmm1",&QWP(16,"edi"));
533	&movaps	(&QWP(0,"esp"),"xmm0");
534	&mov	("edi","esp");
535	&movaps	(&QWP(16,"esp"),"xmm1");
536	&xor	("eax","eax");
537	&data_byte(0xf3,0x0f,0xa6,0xd0);	# rep xsha256
538	&movaps	("xmm0",&QWP(0,"esp"));
539	&movaps	("xmm1",&QWP(16,"esp"));
540	&mov	("esp","edx");			# restore %esp
541    if ($::win32 or $::coff) {
542	&data_byte(0x64,0x8f,0x05,0,0,0,0);	# pop	%fs:0
543	&lea	("esp",&DWP(4,"esp"));
544    }
545	&mov	("edi",&wparam(0));
546	&movups	(&QWP(0,"edi"),"xmm0");		# copy-out context
547	&movups	(&QWP(16,"edi"),"xmm1");
548	&pop	("esi");
549	&pop	("edi");
550	&ret	();
551&function_end_B("padlock_sha256_oneshot");
552
553&function_begin_B("padlock_sha256_blocks");
554	&push	("edi");
555	&push	("esi");
556	&mov	("edi",&wparam(0));
557	&mov	("esi",&wparam(1));
558	&mov	("ecx",&wparam(2));
559	&mov	("edx","esp");			# put aside %esp
560	&add	("esp",-128);
561	&movups	("xmm0",&QWP(0,"edi"));		# copy-in context
562	&and	("esp",-16);
563	&movups	("xmm1",&QWP(16,"edi"));
564	&movaps	(&QWP(0,"esp"),"xmm0");
565	&mov	("edi","esp");
566	&movaps	(&QWP(16,"esp"),"xmm1");
567	&mov	("eax",-1);
568	&data_byte(0xf3,0x0f,0xa6,0xd0);	# rep xsha256
569	&movaps	("xmm0",&QWP(0,"esp"));
570	&movaps	("xmm1",&QWP(16,"esp"));
571	&mov	("esp","edx");			# restore %esp
572	&mov	("edi",&wparam(0));
573	&movups	(&QWP(0,"edi"),"xmm0");		# copy-out context
574	&movups	(&QWP(16,"edi"),"xmm1");
575	&pop	("esi");
576	&pop	("edi");
577	&ret	();
578&function_end_B("padlock_sha256_blocks");
579
580&function_begin_B("padlock_sha512_blocks");
581	&push	("edi");
582	&push	("esi");
583	&mov	("edi",&wparam(0));
584	&mov	("esi",&wparam(1));
585	&mov	("ecx",&wparam(2));
586	&mov	("edx","esp");			# put aside %esp
587	&add	("esp",-128);
588	&movups	("xmm0",&QWP(0,"edi"));		# copy-in context
589	&and	("esp",-16);
590	&movups	("xmm1",&QWP(16,"edi"));
591	&movups	("xmm2",&QWP(32,"edi"));
592	&movups	("xmm3",&QWP(48,"edi"));
593	&movaps	(&QWP(0,"esp"),"xmm0");
594	&mov	("edi","esp");
595	&movaps	(&QWP(16,"esp"),"xmm1");
596	&movaps	(&QWP(32,"esp"),"xmm2");
597	&movaps	(&QWP(48,"esp"),"xmm3");
598	&data_byte(0xf3,0x0f,0xa6,0xe0);	# rep xsha512
599	&movaps	("xmm0",&QWP(0,"esp"));
600	&movaps	("xmm1",&QWP(16,"esp"));
601	&movaps	("xmm2",&QWP(32,"esp"));
602	&movaps	("xmm3",&QWP(48,"esp"));
603	&mov	("esp","edx");			# restore %esp
604	&mov	("edi",&wparam(0));
605	&movups	(&QWP(0,"edi"),"xmm0");		# copy-out context
606	&movups	(&QWP(16,"edi"),"xmm1");
607	&movups	(&QWP(32,"edi"),"xmm2");
608	&movups	(&QWP(48,"edi"),"xmm3");
609	&pop	("esi");
610	&pop	("edi");
611	&ret	();
612&function_end_B("padlock_sha512_blocks");
613
614&asciz	("VIA Padlock x86 module, CRYPTOGAMS by <appro\@openssl.org>");
615&align	(16);
616
617&dataseg();
618# Essentially this variable belongs in thread local storage.
619# Having this variable global on the other hand can only cause
620# few bogus key reloads [if any at all on signle-CPU system],
621# so we accept the penalty...
622&set_label("padlock_saved_context",4);
623&data_word(0);
624
625&asm_finish();
626
627close STDOUT;
628