1#!/usr/bin/env perl
2#
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# ====================================================================
9#
10# May 2011
11#
12# The module implements bn_GF2m_mul_2x2 polynomial multiplication used
13# in bn_gf2m.c. It's kind of low-hanging mechanical port from C for
14# the time being... Except that it has two code paths: code suitable
15# for any x86_64 CPU and PCLMULQDQ one suitable for Westmere and
16# later. Improvement varies from one benchmark and �-arch to another.
17# Vanilla code path is at most 20% faster than compiler-generated code
18# [not very impressive], while PCLMULQDQ - whole 85%-160% better on
19# 163- and 571-bit ECDH benchmarks on Intel CPUs. Keep in mind that
20# these coefficients are not ones for bn_GF2m_mul_2x2 itself, as not
21# all CPU time is burnt in it...
22
23$flavour = shift;
24$output  = shift;
25if ($flavour =~ /\./) { $output = $flavour; undef $flavour; }
26
27$win64=0; $win64=1 if ($flavour =~ /[nm]asm|mingw64/ || $output =~ /\.asm$/);
28
29$0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1;
30( $xlate="${dir}x86_64-xlate.pl" and -f $xlate ) or
31( $xlate="${dir}../../perlasm/x86_64-xlate.pl" and -f $xlate) or
32die "can't locate x86_64-xlate.pl";
33
34open STDOUT,"| \"$^X\" $xlate $flavour $output";
35
36($lo,$hi)=("%rax","%rdx");	$a=$lo;
37($i0,$i1)=("%rsi","%rdi");
38($t0,$t1)=("%rbx","%rcx");
39($b,$mask)=("%rbp","%r8");
40($a1,$a2,$a4,$a8,$a12,$a48)=map("%r$_",(9..15));
41($R,$Tx)=("%xmm0","%xmm1");
42
43$code.=<<___;
44.text
45
46.type	_mul_1x1,\@abi-omnipotent
47.align	16
48_mul_1x1:
49	sub	\$128+8,%rsp
50	mov	\$-1,$a1
51	lea	($a,$a),$i0
52	shr	\$3,$a1
53	lea	(,$a,4),$i1
54	and	$a,$a1			# a1=a&0x1fffffffffffffff
55	lea	(,$a,8),$a8
56	sar	\$63,$a			# broadcast 63rd bit
57	lea	($a1,$a1),$a2
58	sar	\$63,$i0		# broadcast 62nd bit
59	lea	(,$a1,4),$a4
60	and	$b,$a
61	sar	\$63,$i1		# boardcast 61st bit
62	mov	$a,$hi			# $a is $lo
63	shl	\$63,$lo
64	and	$b,$i0
65	shr	\$1,$hi
66	mov	$i0,$t1
67	shl	\$62,$i0
68	and	$b,$i1
69	shr	\$2,$t1
70	xor	$i0,$lo
71	mov	$i1,$t0
72	shl	\$61,$i1
73	xor	$t1,$hi
74	shr	\$3,$t0
75	xor	$i1,$lo
76	xor	$t0,$hi
77
78	mov	$a1,$a12
79	movq	\$0,0(%rsp)		# tab[0]=0
80	xor	$a2,$a12		# a1^a2
81	mov	$a1,8(%rsp)		# tab[1]=a1
82	 mov	$a4,$a48
83	mov	$a2,16(%rsp)		# tab[2]=a2
84	 xor	$a8,$a48		# a4^a8
85	mov	$a12,24(%rsp)		# tab[3]=a1^a2
86
87	xor	$a4,$a1
88	mov	$a4,32(%rsp)		# tab[4]=a4
89	xor	$a4,$a2
90	mov	$a1,40(%rsp)		# tab[5]=a1^a4
91	xor	$a4,$a12
92	mov	$a2,48(%rsp)		# tab[6]=a2^a4
93	 xor	$a48,$a1		# a1^a4^a4^a8=a1^a8
94	mov	$a12,56(%rsp)		# tab[7]=a1^a2^a4
95	 xor	$a48,$a2		# a2^a4^a4^a8=a1^a8
96
97	mov	$a8,64(%rsp)		# tab[8]=a8
98	xor	$a48,$a12		# a1^a2^a4^a4^a8=a1^a2^a8
99	mov	$a1,72(%rsp)		# tab[9]=a1^a8
100	 xor	$a4,$a1			# a1^a8^a4
101	mov	$a2,80(%rsp)		# tab[10]=a2^a8
102	 xor	$a4,$a2			# a2^a8^a4
103	mov	$a12,88(%rsp)		# tab[11]=a1^a2^a8
104
105	xor	$a4,$a12		# a1^a2^a8^a4
106	mov	$a48,96(%rsp)		# tab[12]=a4^a8
107	 mov	$mask,$i0
108	mov	$a1,104(%rsp)		# tab[13]=a1^a4^a8
109	 and	$b,$i0
110	mov	$a2,112(%rsp)		# tab[14]=a2^a4^a8
111	 shr	\$4,$b
112	mov	$a12,120(%rsp)		# tab[15]=a1^a2^a4^a8
113	 mov	$mask,$i1
114	 and	$b,$i1
115	 shr	\$4,$b
116
117	movq	(%rsp,$i0,8),$R		# half of calculations is done in SSE2
118	mov	$mask,$i0
119	and	$b,$i0
120	shr	\$4,$b
121___
122    for ($n=1;$n<8;$n++) {
123	$code.=<<___;
124	mov	(%rsp,$i1,8),$t1
125	mov	$mask,$i1
126	mov	$t1,$t0
127	shl	\$`8*$n-4`,$t1
128	and	$b,$i1
129	 movq	(%rsp,$i0,8),$Tx
130	shr	\$`64-(8*$n-4)`,$t0
131	xor	$t1,$lo
132	 pslldq	\$$n,$Tx
133	 mov	$mask,$i0
134	shr	\$4,$b
135	xor	$t0,$hi
136	 and	$b,$i0
137	 shr	\$4,$b
138	 pxor	$Tx,$R
139___
140    }
141$code.=<<___;
142	mov	(%rsp,$i1,8),$t1
143	mov	$t1,$t0
144	shl	\$`8*$n-4`,$t1
145	movq	$R,$i0
146	shr	\$`64-(8*$n-4)`,$t0
147	xor	$t1,$lo
148	psrldq	\$8,$R
149	xor	$t0,$hi
150	movq	$R,$i1
151	xor	$i0,$lo
152	xor	$i1,$hi
153
154	add	\$128+8,%rsp
155	ret
156.Lend_mul_1x1:
157.size	_mul_1x1,.-_mul_1x1
158___
159
160($rp,$a1,$a0,$b1,$b0) = $win64?	("%rcx","%rdx","%r8", "%r9","%r10") :	# Win64 order
161				("%rdi","%rsi","%rdx","%rcx","%r8");	# Unix order
162
163$code.=<<___;
164.extern	OPENSSL_ia32cap_P
165.globl	bn_GF2m_mul_2x2
166.type	bn_GF2m_mul_2x2,\@abi-omnipotent
167.align	16
168bn_GF2m_mul_2x2:
169	mov	OPENSSL_ia32cap_P(%rip),%rax
170	bt	\$33,%rax
171	jnc	.Lvanilla_mul_2x2
172
173	movq		$a1,%xmm0
174	movq		$b1,%xmm1
175	movq		$a0,%xmm2
176___
177$code.=<<___ if ($win64);
178	movq		40(%rsp),%xmm3
179___
180$code.=<<___ if (!$win64);
181	movq		$b0,%xmm3
182___
183$code.=<<___;
184	movdqa		%xmm0,%xmm4
185	movdqa		%xmm1,%xmm5
186	pclmulqdq	\$0,%xmm1,%xmm0	# a1�b1
187	pxor		%xmm2,%xmm4
188	pxor		%xmm3,%xmm5
189	pclmulqdq	\$0,%xmm3,%xmm2	# a0�b0
190	pclmulqdq	\$0,%xmm5,%xmm4	# (a0+a1)�(b0+b1)
191	xorps		%xmm0,%xmm4
192	xorps		%xmm2,%xmm4	# (a0+a1)�(b0+b1)-a0�b0-a1�b1
193	movdqa		%xmm4,%xmm5
194	pslldq		\$8,%xmm4
195	psrldq		\$8,%xmm5
196	pxor		%xmm4,%xmm2
197	pxor		%xmm5,%xmm0
198	movdqu		%xmm2,0($rp)
199	movdqu		%xmm0,16($rp)
200	ret
201
202.align	16
203.Lvanilla_mul_2x2:
204	lea	-8*17(%rsp),%rsp
205___
206$code.=<<___ if ($win64);
207	mov	`8*17+40`(%rsp),$b0
208	mov	%rdi,8*15(%rsp)
209	mov	%rsi,8*16(%rsp)
210___
211$code.=<<___;
212	mov	%r14,8*10(%rsp)
213	mov	%r13,8*11(%rsp)
214	mov	%r12,8*12(%rsp)
215	mov	%rbp,8*13(%rsp)
216	mov	%rbx,8*14(%rsp)
217.Lbody_mul_2x2:
218	mov	$rp,32(%rsp)		# save the arguments
219	mov	$a1,40(%rsp)
220	mov	$a0,48(%rsp)
221	mov	$b1,56(%rsp)
222	mov	$b0,64(%rsp)
223
224	mov	\$0xf,$mask
225	mov	$a1,$a
226	mov	$b1,$b
227	call	_mul_1x1		# a1�b1
228	mov	$lo,16(%rsp)
229	mov	$hi,24(%rsp)
230
231	mov	48(%rsp),$a
232	mov	64(%rsp),$b
233	call	_mul_1x1		# a0�b0
234	mov	$lo,0(%rsp)
235	mov	$hi,8(%rsp)
236
237	mov	40(%rsp),$a
238	mov	56(%rsp),$b
239	xor	48(%rsp),$a
240	xor	64(%rsp),$b
241	call	_mul_1x1		# (a0+a1)�(b0+b1)
242___
243	@r=("%rbx","%rcx","%rdi","%rsi");
244$code.=<<___;
245	mov	0(%rsp),@r[0]
246	mov	8(%rsp),@r[1]
247	mov	16(%rsp),@r[2]
248	mov	24(%rsp),@r[3]
249	mov	32(%rsp),%rbp
250
251	xor	$hi,$lo
252	xor	@r[1],$hi
253	xor	@r[0],$lo
254	mov	@r[0],0(%rbp)
255	xor	@r[2],$hi
256	mov	@r[3],24(%rbp)
257	xor	@r[3],$lo
258	xor	@r[3],$hi
259	xor	$hi,$lo
260	mov	$hi,16(%rbp)
261	mov	$lo,8(%rbp)
262
263	mov	8*10(%rsp),%r14
264	mov	8*11(%rsp),%r13
265	mov	8*12(%rsp),%r12
266	mov	8*13(%rsp),%rbp
267	mov	8*14(%rsp),%rbx
268___
269$code.=<<___ if ($win64);
270	mov	8*15(%rsp),%rdi
271	mov	8*16(%rsp),%rsi
272___
273$code.=<<___;
274	lea	8*17(%rsp),%rsp
275	ret
276.Lend_mul_2x2:
277.size	bn_GF2m_mul_2x2,.-bn_GF2m_mul_2x2
278.asciz	"GF(2^m) Multiplication for x86_64, CRYPTOGAMS by <appro\@openssl.org>"
279.align	16
280___
281
282# EXCEPTION_DISPOSITION handler (EXCEPTION_RECORD *rec,ULONG64 frame,
283#               CONTEXT *context,DISPATCHER_CONTEXT *disp)
284if ($win64) {
285$rec="%rcx";
286$frame="%rdx";
287$context="%r8";
288$disp="%r9";
289
290$code.=<<___;
291.extern __imp_RtlVirtualUnwind
292
293.type	se_handler,\@abi-omnipotent
294.align	16
295se_handler:
296	push	%rsi
297	push	%rdi
298	push	%rbx
299	push	%rbp
300	push	%r12
301	push	%r13
302	push	%r14
303	push	%r15
304	pushfq
305	sub	\$64,%rsp
306
307	mov	152($context),%rax	# pull context->Rsp
308	mov	248($context),%rbx	# pull context->Rip
309
310	lea	.Lbody_mul_2x2(%rip),%r10
311	cmp	%r10,%rbx		# context->Rip<"prologue" label
312	jb	.Lin_prologue
313
314	mov	8*10(%rax),%r14		# mimic epilogue
315	mov	8*11(%rax),%r13
316	mov	8*12(%rax),%r12
317	mov	8*13(%rax),%rbp
318	mov	8*14(%rax),%rbx
319	mov	8*15(%rax),%rdi
320	mov	8*16(%rax),%rsi
321
322	mov	%rbx,144($context)	# restore context->Rbx
323	mov	%rbp,160($context)	# restore context->Rbp
324	mov	%rsi,168($context)	# restore context->Rsi
325	mov	%rdi,176($context)	# restore context->Rdi
326	mov	%r12,216($context)	# restore context->R12
327	mov	%r13,224($context)	# restore context->R13
328	mov	%r14,232($context)	# restore context->R14
329
330.Lin_prologue:
331	lea	8*17(%rax),%rax
332	mov	%rax,152($context)	# restore context->Rsp
333
334	mov	40($disp),%rdi		# disp->ContextRecord
335	mov	$context,%rsi		# context
336	mov	\$154,%ecx		# sizeof(CONTEXT)
337	.long	0xa548f3fc		# cld; rep movsq
338
339	mov	$disp,%rsi
340	xor	%rcx,%rcx		# arg1, UNW_FLAG_NHANDLER
341	mov	8(%rsi),%rdx		# arg2, disp->ImageBase
342	mov	0(%rsi),%r8		# arg3, disp->ControlPc
343	mov	16(%rsi),%r9		# arg4, disp->FunctionEntry
344	mov	40(%rsi),%r10		# disp->ContextRecord
345	lea	56(%rsi),%r11		# &disp->HandlerData
346	lea	24(%rsi),%r12		# &disp->EstablisherFrame
347	mov	%r10,32(%rsp)		# arg5
348	mov	%r11,40(%rsp)		# arg6
349	mov	%r12,48(%rsp)		# arg7
350	mov	%rcx,56(%rsp)		# arg8, (NULL)
351	call	*__imp_RtlVirtualUnwind(%rip)
352
353	mov	\$1,%eax		# ExceptionContinueSearch
354	add	\$64,%rsp
355	popfq
356	pop	%r15
357	pop	%r14
358	pop	%r13
359	pop	%r12
360	pop	%rbp
361	pop	%rbx
362	pop	%rdi
363	pop	%rsi
364	ret
365.size	se_handler,.-se_handler
366
367.section	.pdata
368.align	4
369	.rva	_mul_1x1
370	.rva	.Lend_mul_1x1
371	.rva	.LSEH_info_1x1
372
373	.rva	.Lvanilla_mul_2x2
374	.rva	.Lend_mul_2x2
375	.rva	.LSEH_info_2x2
376.section	.xdata
377.align	8
378.LSEH_info_1x1:
379	.byte	0x01,0x07,0x02,0x00
380	.byte	0x07,0x01,0x11,0x00	# sub rsp,128+8
381.LSEH_info_2x2:
382	.byte	9,0,0,0
383	.rva	se_handler
384___
385}
386
387$code =~ s/\`([^\`]*)\`/eval($1)/gem;
388print $code;
389close STDOUT;
390