powerpc/crypto/ghashp10-ppc.pl

47 $0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1;
56 my ($Xl,$Xm,$Xh,$IN)=map("v$_",(0..3));
57 my ($zero,$t0,$t1,$t2,$xC2,$H,$Hh,$Hl,$lemask)=map("v$_",(4..12));
68 	lis		r0,0xfff0
69 	li		r8,0x10
71 	li		r9,0x20
73 	li		r10,0x30
74 	lvx_u		$H,0,r4			# load H
76 	le?addi		r7,r7,0x8		# need a vperm start with 08
77 	le?lvsr		5,0,r7
78 	le?vspltisb	6,0x0f
82 	vspltisb	$xC2,-16		# 0xf0
84 	vaddubm		$xC2,$xC2,$xC2		# 0xe0
86 	vor		$xC2,$xC2,$t0		# 0xe1
87 	vsldoi		$xC2,$xC2,$zero,15	# 0xe1...
89 	vaddubm		$xC2,$xC2,$xC2		# 0xc2...
91 	vor		$xC2,$xC2,$t1		# 0xc2....01
92 	vspltb		$t1,$H,0		# most significant byte
95 	vand		$t1,$t1,$xC2
99 	vsldoi		$xC2,$zero,$xC2,8	# 0xc2.0
103 	stvx_u		$xC2,0,r3		# save pre-computed table
110 	.long		0
111 	.byte		0,12,0x14,0,0,0,2,0
112 	.long		0
116 	lis		r0,0xfff0
117 	li		r8,0x10
119 	li		r9,0x20
121 	li		r10,0x30
122 	lvx_u		$H,0,r4			# load H
124 	vspltisb	$xC2,-16		# 0xf0
126 	vaddubm		$xC2,$xC2,$xC2		# 0xe0
128 	vor		$xC2,$xC2,$t0		# 0xe1
129 	vsldoi		$xC2,$xC2,$zero,15	# 0xe1...
131 	vaddubm		$xC2,$xC2,$xC2		# 0xc2...
133 	vor		$xC2,$xC2,$t1		# 0xc2....01
134 	vspltb		$t1,$H,0		# most significant byte
137 	vand		$t1,$t1,$xC2
141 	vsldoi		$xC2,$zero,$xC2,8	# 0xc2.0
145 	stvx_u		$xC2,0,r3		# save pre-computed table
147 	li		r8,0x40
149 	li		r9,0x50
151 	li		r10,0x60
157 	vpmsumd		$t2,$Xl,$xC2		# 1st reduction phase
168 	vpmsumd		$Xl,$Xl,$xC2
177 	li		r8,0x70
179 	li		r9,0x80
181 	li		r10,0x90
190 	vpmsumd		$t2,$Xl,$xC2		# 1st reduction phase
191 	 vpmsumd	$t6,$Xl1,$xC2		# 1st reduction phase
209 	vpmsumd		$Xl,$Xl,$xC2
210 	 vpmsumd	$Xl1,$Xl1,$xC2
224 	li		r8,0xa0
226 	li		r9,0xb0
228 	li		r10,0xc0
235 	.long		0
236 	.byte		0,12,0x14,0,0,0,2,0
237 	.long		0
241 	lis		r0,0xfff8
242 	li		r8,0x10
244 	li		r9,0x20
246 	li		r10,0x30
247 	lvx_u		$IN,0,$Xip		# load Xi
252 	 le?vspltisb	$t0,0x07
255 	lvx_u		$xC2,0,$Htbl
263 	vpmsumd		$t2,$Xl,$xC2		# 1st phase
274 	vpmsumd		$Xl,$Xl,$xC2
279 	stvx_u		$Xl,0,$Xip		# write out Xi
283 	.long		0
284 	.byte		0,12,0x14,0,0,0,2,0
285 	.long		0
289 	lis		r0,0xfff8
290 	li		r8,0x10
292 	li		r9,0x20
294 	li		r10,0x30
295 	lvx_u		$Xl,0,$Xip		# load Xi
300 	 le?vspltisb	$t0,0x07
303 	lvx_u		$xC2,0,$Htbl
307 	lvx_u		$IN,0,$inp
318 	 subfe.		r0,r0,r0		# borrow?-1:0
324 	vpmsumd		$t2,$Xl,$xC2		# 1st phase
333 	 lvx_u		$IN,0,$inp
337 	vpmsumd		$Xl,$Xl,$xC2
346 	stvx_u		$Xl,0,$Xip		# write out Xi
350 	.long		0
351 	.byte		0,12,0x14,0,0,0,4,0
352 	.long		0