• 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 Apache License 2.0 (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 and open STDOUT,">$output";
46
47&asm_init($ARGV[0]);
48
49%PADLOCK_PREFETCH=(ecb=>128, cbc=>64);	# prefetch errata
50$PADLOCK_CHUNK=512;	# Must be a power of 2 larger than 16
51
52$ctx="edx";
53$out="edi";
54$inp="esi";
55$len="ecx";
56$chunk="ebx";
57
58&function_begin_B("padlock_capability");
59	&push	("ebx");
60	&pushf	();
61	&pop	("eax");
62	&mov	("ecx","eax");
63	&xor	("eax",1<<21);
64	&push	("eax");
65	&popf	();
66	&pushf	();
67	&pop	("eax");
68	&xor	("ecx","eax");
69	&xor	("eax","eax");
70	&bt	("ecx",21);
71	&jnc	(&label("noluck"));
72	&cpuid	();
73	&xor	("eax","eax");
74	&cmp	("ebx","0x".unpack("H*",'tneC'));
75	&jne	(&label("zhaoxin"));
76	&cmp	("edx","0x".unpack("H*",'Hrua'));
77	&jne	(&label("noluck"));
78	&cmp	("ecx","0x".unpack("H*",'slua'));
79	&jne	(&label("noluck"));
80	&jmp	(&label("zhaoxinEnd"));
81&set_label("zhaoxin");
82	&cmp	("ebx","0x".unpack("H*",'hS  '));
83	&jne	(&label("noluck"));
84	&cmp	("edx","0x".unpack("H*",'hgna'));
85	&jne	(&label("noluck"));
86	&cmp	("ecx","0x".unpack("H*",'  ia'));
87	&jne	(&label("noluck"));
88&set_label("zhaoxinEnd");
89	&mov	("eax",0xC0000000);
90	&cpuid	();
91	&mov	("edx","eax");
92	&xor	("eax","eax");
93	&cmp	("edx",0xC0000001);
94	&jb	(&label("noluck"));
95	&mov	("eax",1);
96	&cpuid	();
97	&or	("eax",0x0f);
98	&xor	("ebx","ebx");
99	&and	("eax",0x0fff);
100	&cmp	("eax",0x06ff);		# check for Nano
101	&sete	("bl");
102	&mov	("eax",0xC0000001);
103	&push	("ebx");
104	&cpuid	();
105	&pop	("ebx");
106	&mov	("eax","edx");
107	&shl	("ebx",4);		# bit#4 denotes Nano
108	&and	("eax",0xffffffef);
109	&or	("eax","ebx")
110&set_label("noluck");
111	&pop	("ebx");
112	&ret	();
113&function_end_B("padlock_capability")
114
115&function_begin_B("padlock_key_bswap");
116	&mov	("edx",&wparam(0));
117	&mov	("ecx",&DWP(240,"edx"));
118&set_label("bswap_loop");
119	&mov	("eax",&DWP(0,"edx"));
120	&bswap	("eax");
121	&mov	(&DWP(0,"edx"),"eax");
122	&lea	("edx",&DWP(4,"edx"));
123	&sub	("ecx",1);
124	&jnz	(&label("bswap_loop"));
125	&ret	();
126&function_end_B("padlock_key_bswap");
127
128# This is heuristic key context tracing. At first one
129# believes that one should use atomic swap instructions,
130# but it's not actually necessary. Point is that if
131# padlock_saved_context was changed by another thread
132# after we've read it and before we compare it with ctx,
133# our key *shall* be reloaded upon thread context switch
134# and we are therefore set in either case...
135&static_label("padlock_saved_context");
136
137&function_begin_B("padlock_verify_context");
138	&mov	($ctx,&wparam(0));
139	&lea	("eax",($::win32 or $::coff) ? &DWP(&label("padlock_saved_context")) :
140		       &DWP(&label("padlock_saved_context")."-".&label("verify_pic_point")));
141	&pushf	();
142	&call	("_padlock_verify_ctx");
143&set_label("verify_pic_point");
144	&lea	("esp",&DWP(4,"esp"));
145	&ret	();
146&function_end_B("padlock_verify_context");
147
148&function_begin_B("_padlock_verify_ctx");
149	&add	("eax",&DWP(0,"esp")) if(!($::win32 or $::coff));# &padlock_saved_context
150	&bt	(&DWP(4,"esp"),30);		# eflags
151	&jnc	(&label("verified"));
152	&cmp	($ctx,&DWP(0,"eax"));
153	&je	(&label("verified"));
154	&pushf	();
155	&popf	();
156&set_label("verified");
157	&mov	(&DWP(0,"eax"),$ctx);
158	&ret	();
159&function_end_B("_padlock_verify_ctx");
160
161&function_begin_B("padlock_reload_key");
162	&pushf	();
163	&popf	();
164	&ret	();
165&function_end_B("padlock_reload_key");
166
167&function_begin_B("padlock_aes_block");
168	&push	("edi");
169	&push	("esi");
170	&push	("ebx");
171	&mov	($out,&wparam(0));		# must be 16-byte aligned
172	&mov	($inp,&wparam(1));		# must be 16-byte aligned
173	&mov	($ctx,&wparam(2));
174	&mov	($len,1);
175	&lea	("ebx",&DWP(32,$ctx));		# key
176	&lea	($ctx,&DWP(16,$ctx));		# control word
177	&data_byte(0xf3,0x0f,0xa7,0xc8);	# rep xcryptecb
178	&pop	("ebx");
179	&pop	("esi");
180	&pop	("edi");
181	&ret	();
182&function_end_B("padlock_aes_block");
183
184sub generate_mode {
185my ($mode,$opcode) = @_;
186# int padlock_$mode_encrypt(void *out, const void *inp,
187#		struct padlock_cipher_data *ctx, size_t len);
188&function_begin("padlock_${mode}_encrypt");
189	&mov	($out,&wparam(0));
190	&mov	($inp,&wparam(1));
191	&mov	($ctx,&wparam(2));
192	&mov	($len,&wparam(3));
193	&test	($ctx,15);
194	&jnz	(&label("${mode}_abort"));
195	&test	($len,15);
196	&jnz	(&label("${mode}_abort"));
197	&lea	("eax",($::win32 or $::coff) ? &DWP(&label("padlock_saved_context")) :
198		       &DWP(&label("padlock_saved_context")."-".&label("${mode}_pic_point")));
199	&pushf	();
200	&cld	();
201	&call	("_padlock_verify_ctx");
202&set_label("${mode}_pic_point");
203	&lea	($ctx,&DWP(16,$ctx));	# control word
204	&xor	("eax","eax");
205					if ($mode eq "ctr32") {
206	&movq	("mm0",&QWP(-16,$ctx));	# load [upper part of] counter
207					} else {
208	&xor	("ebx","ebx");
209	&test	(&DWP(0,$ctx),1<<5);	# align bit in control word
210	&jnz	(&label("${mode}_aligned"));
211	&test	($out,0x0f);
212	&setz	("al");			# !out_misaligned
213	&test	($inp,0x0f);
214	&setz	("bl");			# !inp_misaligned
215	&test	("eax","ebx");
216	&jnz	(&label("${mode}_aligned"));
217	&neg	("eax");
218					}
219	&mov	($chunk,$PADLOCK_CHUNK);
220	&not	("eax");		# out_misaligned?-1:0
221	&lea	("ebp",&DWP(-24,"esp"));
222	&cmp	($len,$chunk);
223	&cmovc	($chunk,$len);		# chunk=len>PADLOCK_CHUNK?PADLOCK_CHUNK:len
224	&and	("eax",$chunk);		# out_misaligned?chunk:0
225	&mov	($chunk,$len);
226	&neg	("eax");
227	&and	($chunk,$PADLOCK_CHUNK-1);	# chunk=len%PADLOCK_CHUNK
228	&lea	("esp",&DWP(0,"eax","ebp"));	# alloca
229	&mov	("eax",$PADLOCK_CHUNK);
230	&cmovz	($chunk,"eax");			# chunk=chunk?:PADLOCK_CHUNK
231	&mov	("eax","ebp");
232	&and	("ebp",-16);
233	&and	("esp",-16);
234	&mov	(&DWP(16,"ebp"),"eax");
235    if ($PADLOCK_PREFETCH{$mode}) {
236	&cmp	($len,$chunk);
237	&ja	(&label("${mode}_loop"));
238	&mov	("eax",$inp);		# check if prefetch crosses page
239	&cmp	("ebp","esp");
240	&cmove	("eax",$out);
241	&add	("eax",$len);
242	&neg	("eax");
243	&and	("eax",0xfff);		# distance to page boundary
244	&cmp	("eax",$PADLOCK_PREFETCH{$mode});
245	&mov	("eax",-$PADLOCK_PREFETCH{$mode});
246	&cmovae	("eax",$chunk);		# mask=distance<prefetch?-prefetch:-1
247	&and	($chunk,"eax");
248	&jz	(&label("${mode}_unaligned_tail"));
249    }
250	&jmp	(&label("${mode}_loop"));
251
252&set_label("${mode}_loop",16);
253	&mov	(&DWP(0,"ebp"),$out);		# save parameters
254	&mov	(&DWP(4,"ebp"),$inp);
255	&mov	(&DWP(8,"ebp"),$len);
256	&mov	($len,$chunk);
257	&mov	(&DWP(12,"ebp"),$chunk);	# chunk
258						if ($mode eq "ctr32") {
259	&mov	("ecx",&DWP(-4,$ctx));
260	&xor	($out,$out);
261	&mov	("eax",&DWP(-8,$ctx));		# borrow $len
262&set_label("${mode}_prepare");
263	&mov	(&DWP(12,"esp",$out),"ecx");
264	&bswap	("ecx");
265	&movq	(&QWP(0,"esp",$out),"mm0");
266	&inc	("ecx");
267	&mov	(&DWP(8,"esp",$out),"eax");
268	&bswap	("ecx");
269	&lea	($out,&DWP(16,$out));
270	&cmp	($out,$chunk);
271	&jb	(&label("${mode}_prepare"));
272
273	&mov	(&DWP(-4,$ctx),"ecx");
274	&lea	($inp,&DWP(0,"esp"));
275	&lea	($out,&DWP(0,"esp"));
276	&mov	($len,$chunk);
277						} else {
278	&test	($out,0x0f);			# out_misaligned
279	&cmovnz	($out,"esp");
280	&test	($inp,0x0f);			# inp_misaligned
281	&jz	(&label("${mode}_inp_aligned"));
282	&shr	($len,2);
283	&data_byte(0xf3,0xa5);			# rep movsl
284	&sub	($out,$chunk);
285	&mov	($len,$chunk);
286	&mov	($inp,$out);
287&set_label("${mode}_inp_aligned");
288						}
289	&lea	("eax",&DWP(-16,$ctx));		# ivp
290	&lea	("ebx",&DWP(16,$ctx));		# key
291	&shr	($len,4);			# len/=AES_BLOCK_SIZE
292	&data_byte(0xf3,0x0f,0xa7,$opcode);	# rep xcrypt*
293						if ($mode !~ /ecb|ctr/) {
294	&movaps	("xmm0",&QWP(0,"eax"));
295	&movaps	(&QWP(-16,$ctx),"xmm0");	# copy [or refresh] iv
296						}
297	&mov	($out,&DWP(0,"ebp"));		# restore parameters
298	&mov	($chunk,&DWP(12,"ebp"));
299						if ($mode eq "ctr32") {
300	&mov	($inp,&DWP(4,"ebp"));
301	&xor	($len,$len);
302&set_label("${mode}_xor");
303	&movups	("xmm1",&QWP(0,$inp,$len));
304	&lea	($len,&DWP(16,$len));
305	&pxor	("xmm1",&QWP(-16,"esp",$len));
306	&movups	(&QWP(-16,$out,$len),"xmm1");
307	&cmp	($len,$chunk);
308	&jb	(&label("${mode}_xor"));
309						} else {
310	&test	($out,0x0f);
311	&jz	(&label("${mode}_out_aligned"));
312	&mov	($len,$chunk);
313	&lea	($inp,&DWP(0,"esp"));
314	&shr	($len,2);
315	&data_byte(0xf3,0xa5);			# rep movsl
316	&sub	($out,$chunk);
317&set_label("${mode}_out_aligned");
318	&mov	($inp,&DWP(4,"ebp"));
319						}
320	&mov	($len,&DWP(8,"ebp"));
321	&add	($out,$chunk);
322	&add	($inp,$chunk);
323	&sub	($len,$chunk);
324	&mov	($chunk,$PADLOCK_CHUNK);
325    if (!$PADLOCK_PREFETCH{$mode}) {
326	&jnz	(&label("${mode}_loop"));
327    } else {
328	&jz	(&label("${mode}_break"));
329	&cmp	($len,$chunk);
330	&jae	(&label("${mode}_loop"));
331
332&set_label("${mode}_unaligned_tail");
333	&xor	("eax","eax");
334	&cmp	("esp","ebp");
335	&cmove	("eax",$len);
336	&sub	("esp","eax");			# alloca
337	&mov	("eax", $out);			# save parameters
338	&mov	($chunk,$len);
339	&shr	($len,2);
340	&lea	($out,&DWP(0,"esp"));
341	&data_byte(0xf3,0xa5);			# rep movsl
342	&mov	($inp,"esp");
343	&mov	($out,"eax");			# restore parameters
344	&mov	($len,$chunk);
345	&jmp	(&label("${mode}_loop"));
346
347&set_label("${mode}_break",16);
348    }
349						if ($mode ne "ctr32") {
350	&cmp	("esp","ebp");
351	&je	(&label("${mode}_done"));
352						}
353	&pxor	("xmm0","xmm0");
354	&lea	("eax",&DWP(0,"esp"));
355&set_label("${mode}_bzero");
356	&movaps	(&QWP(0,"eax"),"xmm0");
357	&lea	("eax",&DWP(16,"eax"));
358	&cmp	("ebp","eax");
359	&ja	(&label("${mode}_bzero"));
360
361&set_label("${mode}_done");
362	&mov	("ebp",&DWP(16,"ebp"));
363	&lea	("esp",&DWP(24,"ebp"));
364						if ($mode ne "ctr32") {
365	&jmp	(&label("${mode}_exit"));
366
367&set_label("${mode}_aligned",16);
368    if ($PADLOCK_PREFETCH{$mode}) {
369	&lea	("ebp",&DWP(0,$inp,$len));
370	&neg	("ebp");
371	&and	("ebp",0xfff);			# distance to page boundary
372	&xor	("eax","eax");
373	&cmp	("ebp",$PADLOCK_PREFETCH{$mode});
374	&mov	("ebp",$PADLOCK_PREFETCH{$mode}-1);
375	&cmovae	("ebp","eax");
376	&and	("ebp",$len);			# remainder
377	&sub	($len,"ebp");
378	&jz	(&label("${mode}_aligned_tail"));
379    }
380	&lea	("eax",&DWP(-16,$ctx));		# ivp
381	&lea	("ebx",&DWP(16,$ctx));		# key
382	&shr	($len,4);			# len/=AES_BLOCK_SIZE
383	&data_byte(0xf3,0x0f,0xa7,$opcode);	# rep xcrypt*
384						if ($mode ne "ecb") {
385	&movaps	("xmm0",&QWP(0,"eax"));
386	&movaps	(&QWP(-16,$ctx),"xmm0");	# copy [or refresh] iv
387						}
388    if ($PADLOCK_PREFETCH{$mode}) {
389	&test	("ebp","ebp");
390	&jz	(&label("${mode}_exit"));
391
392&set_label("${mode}_aligned_tail");
393	&mov	($len,"ebp");
394	&lea	("ebp",&DWP(-24,"esp"));
395	&mov	("esp","ebp");
396	&mov	("eax","ebp");
397	&sub	("esp",$len);
398	&and	("ebp",-16);
399	&and	("esp",-16);
400	&mov	(&DWP(16,"ebp"),"eax");
401	&mov	("eax", $out);			# save parameters
402	&mov	($chunk,$len);
403	&shr	($len,2);
404	&lea	($out,&DWP(0,"esp"));
405	&data_byte(0xf3,0xa5);			# rep movsl
406	&mov	($inp,"esp");
407	&mov	($out,"eax");			# restore parameters
408	&mov	($len,$chunk);
409	&jmp	(&label("${mode}_loop"));
410    }
411&set_label("${mode}_exit");			}
412	&mov	("eax",1);
413	&lea	("esp",&DWP(4,"esp"));		# popf
414	&emms	()				if ($mode eq "ctr32");
415&set_label("${mode}_abort");
416&function_end("padlock_${mode}_encrypt");
417}
418
419&generate_mode("ecb",0xc8);
420&generate_mode("cbc",0xd0);
421&generate_mode("cfb",0xe0);
422&generate_mode("ofb",0xe8);
423&generate_mode("ctr32",0xc8);	# yes, it implements own CTR with ECB opcode,
424				# because hardware CTR was introduced later
425				# and even has errata on certain C7 stepping.
426				# own implementation *always* works, though
427				# ~15% slower than dedicated hardware...
428
429&function_begin_B("padlock_xstore");
430	&push	("edi");
431	&mov	("edi",&wparam(0));
432	&mov	("edx",&wparam(1));
433	&data_byte(0x0f,0xa7,0xc0);		# xstore
434	&pop	("edi");
435	&ret	();
436&function_end_B("padlock_xstore");
437
438&function_begin_B("_win32_segv_handler");
439	&mov	("eax",1);			# ExceptionContinueSearch
440	&mov	("edx",&wparam(0));		# *ExceptionRecord
441	&mov	("ecx",&wparam(2));		# *ContextRecord
442	&cmp	(&DWP(0,"edx"),0xC0000005)	# ExceptionRecord->ExceptionCode == STATUS_ACCESS_VIOLATION
443	&jne	(&label("ret"));
444	&add	(&DWP(184,"ecx"),4);		# skip over rep sha*
445	&mov	("eax",0);			# ExceptionContinueExecution
446&set_label("ret");
447	&ret	();
448&function_end_B("_win32_segv_handler");
449&safeseh("_win32_segv_handler")			if ($::win32);
450
451&function_begin_B("padlock_sha1_oneshot");
452	&push	("edi");
453	&push	("esi");
454	&xor	("eax","eax");
455	&mov	("edi",&wparam(0));
456	&mov	("esi",&wparam(1));
457	&mov	("ecx",&wparam(2));
458    if ($::win32 or $::coff) {
459    	&push	(&::islabel("_win32_segv_handler"));
460	&data_byte(0x64,0xff,0x30);		# push	%fs:(%eax)
461	&data_byte(0x64,0x89,0x20);		# mov	%esp,%fs:(%eax)
462    }
463	&mov	("edx","esp");			# put aside %esp
464	&add	("esp",-128);			# 32 is enough but spec says 128
465	&movups	("xmm0",&QWP(0,"edi"));		# copy-in context
466	&and	("esp",-16);
467	&mov	("eax",&DWP(16,"edi"));
468	&movaps	(&QWP(0,"esp"),"xmm0");
469	&mov	("edi","esp");
470	&mov	(&DWP(16,"esp"),"eax");
471	&xor	("eax","eax");
472	&data_byte(0xf3,0x0f,0xa6,0xc8);	# rep xsha1
473	&movaps	("xmm0",&QWP(0,"esp"));
474	&mov	("eax",&DWP(16,"esp"));
475	&mov	("esp","edx");			# restore %esp
476    if ($::win32 or $::coff) {
477	&data_byte(0x64,0x8f,0x05,0,0,0,0);	# pop	%fs:0
478	&lea	("esp",&DWP(4,"esp"));
479    }
480	&mov	("edi",&wparam(0));
481	&movups	(&QWP(0,"edi"),"xmm0");		# copy-out context
482	&mov	(&DWP(16,"edi"),"eax");
483	&pop	("esi");
484	&pop	("edi");
485	&ret	();
486&function_end_B("padlock_sha1_oneshot");
487
488&function_begin_B("padlock_sha1_blocks");
489	&push	("edi");
490	&push	("esi");
491	&mov	("edi",&wparam(0));
492	&mov	("esi",&wparam(1));
493	&mov	("edx","esp");			# put aside %esp
494	&mov	("ecx",&wparam(2));
495	&add	("esp",-128);
496	&movups	("xmm0",&QWP(0,"edi"));		# copy-in context
497	&and	("esp",-16);
498	&mov	("eax",&DWP(16,"edi"));
499	&movaps	(&QWP(0,"esp"),"xmm0");
500	&mov	("edi","esp");
501	&mov	(&DWP(16,"esp"),"eax");
502	&mov	("eax",-1);
503	&data_byte(0xf3,0x0f,0xa6,0xc8);	# rep xsha1
504	&movaps	("xmm0",&QWP(0,"esp"));
505	&mov	("eax",&DWP(16,"esp"));
506	&mov	("esp","edx");			# restore %esp
507	&mov	("edi",&wparam(0));
508	&movups	(&QWP(0,"edi"),"xmm0");		# copy-out context
509	&mov	(&DWP(16,"edi"),"eax");
510 	&pop	("esi");
511	&pop	("edi");
512	&ret	();
513&function_end_B("padlock_sha1_blocks");
514
515&function_begin_B("padlock_sha256_oneshot");
516	&push	("edi");
517	&push	("esi");
518	&xor	("eax","eax");
519	&mov	("edi",&wparam(0));
520	&mov	("esi",&wparam(1));
521	&mov	("ecx",&wparam(2));
522    if ($::win32 or $::coff) {
523    	&push	(&::islabel("_win32_segv_handler"));
524	&data_byte(0x64,0xff,0x30);		# push	%fs:(%eax)
525	&data_byte(0x64,0x89,0x20);		# mov	%esp,%fs:(%eax)
526    }
527	&mov	("edx","esp");			# put aside %esp
528	&add	("esp",-128);
529	&movups	("xmm0",&QWP(0,"edi"));		# copy-in context
530	&and	("esp",-16);
531	&movups	("xmm1",&QWP(16,"edi"));
532	&movaps	(&QWP(0,"esp"),"xmm0");
533	&mov	("edi","esp");
534	&movaps	(&QWP(16,"esp"),"xmm1");
535	&xor	("eax","eax");
536	&data_byte(0xf3,0x0f,0xa6,0xd0);	# rep xsha256
537	&movaps	("xmm0",&QWP(0,"esp"));
538	&movaps	("xmm1",&QWP(16,"esp"));
539	&mov	("esp","edx");			# restore %esp
540    if ($::win32 or $::coff) {
541	&data_byte(0x64,0x8f,0x05,0,0,0,0);	# pop	%fs:0
542	&lea	("esp",&DWP(4,"esp"));
543    }
544	&mov	("edi",&wparam(0));
545	&movups	(&QWP(0,"edi"),"xmm0");		# copy-out context
546	&movups	(&QWP(16,"edi"),"xmm1");
547	&pop	("esi");
548	&pop	("edi");
549	&ret	();
550&function_end_B("padlock_sha256_oneshot");
551
552&function_begin_B("padlock_sha256_blocks");
553	&push	("edi");
554	&push	("esi");
555	&mov	("edi",&wparam(0));
556	&mov	("esi",&wparam(1));
557	&mov	("ecx",&wparam(2));
558	&mov	("edx","esp");			# put aside %esp
559	&add	("esp",-128);
560	&movups	("xmm0",&QWP(0,"edi"));		# copy-in context
561	&and	("esp",-16);
562	&movups	("xmm1",&QWP(16,"edi"));
563	&movaps	(&QWP(0,"esp"),"xmm0");
564	&mov	("edi","esp");
565	&movaps	(&QWP(16,"esp"),"xmm1");
566	&mov	("eax",-1);
567	&data_byte(0xf3,0x0f,0xa6,0xd0);	# rep xsha256
568	&movaps	("xmm0",&QWP(0,"esp"));
569	&movaps	("xmm1",&QWP(16,"esp"));
570	&mov	("esp","edx");			# restore %esp
571	&mov	("edi",&wparam(0));
572	&movups	(&QWP(0,"edi"),"xmm0");		# copy-out context
573	&movups	(&QWP(16,"edi"),"xmm1");
574	&pop	("esi");
575	&pop	("edi");
576	&ret	();
577&function_end_B("padlock_sha256_blocks");
578
579&function_begin_B("padlock_sha512_blocks");
580	&push	("edi");
581	&push	("esi");
582	&mov	("edi",&wparam(0));
583	&mov	("esi",&wparam(1));
584	&mov	("ecx",&wparam(2));
585	&mov	("edx","esp");			# put aside %esp
586	&add	("esp",-128);
587	&movups	("xmm0",&QWP(0,"edi"));		# copy-in context
588	&and	("esp",-16);
589	&movups	("xmm1",&QWP(16,"edi"));
590	&movups	("xmm2",&QWP(32,"edi"));
591	&movups	("xmm3",&QWP(48,"edi"));
592	&movaps	(&QWP(0,"esp"),"xmm0");
593	&mov	("edi","esp");
594	&movaps	(&QWP(16,"esp"),"xmm1");
595	&movaps	(&QWP(32,"esp"),"xmm2");
596	&movaps	(&QWP(48,"esp"),"xmm3");
597	&data_byte(0xf3,0x0f,0xa6,0xe0);	# rep xsha512
598	&movaps	("xmm0",&QWP(0,"esp"));
599	&movaps	("xmm1",&QWP(16,"esp"));
600	&movaps	("xmm2",&QWP(32,"esp"));
601	&movaps	("xmm3",&QWP(48,"esp"));
602	&mov	("esp","edx");			# restore %esp
603	&mov	("edi",&wparam(0));
604	&movups	(&QWP(0,"edi"),"xmm0");		# copy-out context
605	&movups	(&QWP(16,"edi"),"xmm1");
606	&movups	(&QWP(32,"edi"),"xmm2");
607	&movups	(&QWP(48,"edi"),"xmm3");
608	&pop	("esi");
609	&pop	("edi");
610	&ret	();
611&function_end_B("padlock_sha512_blocks");
612
613&asciz	("VIA Padlock x86 module, CRYPTOGAMS by <appro\@openssl.org>");
614&align	(16);
615
616&dataseg();
617# Essentially this variable belongs in thread local storage.
618# Having this variable global on the other hand can only cause
619# few bogus key reloads [if any at all on signle-CPU system],
620# so we accept the penalty...
621&set_label("padlock_saved_context",4);
622&data_word(0);
623
624&asm_finish();
625
626close STDOUT;
627