Home | History | Annotate | Line # | Download | only in pa64
submul_1.asm revision 1.1.1.2
      1 dnl  HP-PA 2.0 64-bit mpn_submul_1 -- Multiply a limb vector with a limb and
      2 dnl  subtract the result from a second limb vector.
      3 
      4 dnl  Copyright 1998-2000, 2002, 2003 Free Software Foundation, Inc.
      5 
      6 dnl  This file is part of the GNU MP Library.
      7 dnl
      8 dnl  The GNU MP Library is free software; you can redistribute it and/or modify
      9 dnl  it under the terms of either:
     10 dnl
     11 dnl    * the GNU Lesser General Public License as published by the Free
     12 dnl      Software Foundation; either version 3 of the License, or (at your
     13 dnl      option) any later version.
     14 dnl
     15 dnl  or
     16 dnl
     17 dnl    * the GNU General Public License as published by the Free Software
     18 dnl      Foundation; either version 2 of the License, or (at your option) any
     19 dnl      later version.
     20 dnl
     21 dnl  or both in parallel, as here.
     22 dnl
     23 dnl  The GNU MP Library is distributed in the hope that it will be useful, but
     24 dnl  WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY
     25 dnl  or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
     26 dnl  for more details.
     27 dnl
     28 dnl  You should have received copies of the GNU General Public License and the
     29 dnl  GNU Lesser General Public License along with the GNU MP Library.  If not,
     30 dnl  see https://www.gnu.org/licenses/.
     31 
     32 include(`../config.m4')
     33 
     34 C		    cycles/limb
     35 C 8000,8200:		7
     36 C 8500,8600,8700:	6.5
     37 
     38 C  The feed-in and wind-down code has not yet been scheduled.  Many cycles
     39 C  could be saved there per call.
     40 
     41 C  DESCRIPTION:
     42 C  The main loop "BIG" is 4-way unrolled, mainly to allow
     43 C  effective use of ADD,DC.  Delays in moving data via the cache from the FP
     44 C  registers to the IU registers, have demanded a deep software pipeline, and
     45 C  a lot of stack slots for partial products in flight.
     46 C
     47 C  CODE STRUCTURE:
     48 C  save-some-registers
     49 C  do 0, 1, 2, or 3 limbs
     50 C  if done, restore-some-regs and return
     51 C  save-many-regs
     52 C  do 4, 8, ... limb
     53 C  restore-all-regs
     54 
     55 C  STACK LAYOUT:
     56 C  HP-PA stack grows upwards.  We could allocate 8 fewer slots by using the
     57 C  slots marked FREE, as well as some slots in the caller's "frame marker".
     58 C
     59 C -00 <- r30
     60 C -08  FREE
     61 C -10  tmp
     62 C -18  tmp
     63 C -20  tmp
     64 C -28  tmp
     65 C -30  tmp
     66 C -38  tmp
     67 C -40  tmp
     68 C -48  tmp
     69 C -50  tmp
     70 C -58  tmp
     71 C -60  tmp
     72 C -68  tmp
     73 C -70  tmp
     74 C -78  tmp
     75 C -80  tmp
     76 C -88  tmp
     77 C -90  FREE
     78 C -98  FREE
     79 C -a0  FREE
     80 C -a8  FREE
     81 C -b0  r13
     82 C -b8  r12
     83 C -c0  r11
     84 C -c8  r10
     85 C -d0  r8
     86 C -d8  r8
     87 C -e0  r7
     88 C -e8  r6
     89 C -f0  r5
     90 C -f8  r4
     91 C -100 r3
     92 C  Previous frame:
     93 C  [unused area]
     94 C -38/-138 vlimb home slot.  For 2.0N, the vlimb arg will arrive here.
     95 
     96 
     97 include(`../config.m4')
     98 
     99 C INPUT PARAMETERS:
    100 define(`rp',`%r26')	C
    101 define(`up',`%r25')	C
    102 define(`n',`%r24')	C
    103 define(`vlimb',`%r23')	C
    104 
    105 define(`climb',`%r23')	C
    106 
    107 ifdef(`HAVE_ABI_2_0w',
    108 `	.level	2.0w
    109 ',`	.level	2.0
    110 ')
    111 PROLOGUE(mpn_submul_1)
    112 
    113 ifdef(`HAVE_ABI_2_0w',
    114 `	std		vlimb, -0x38(%r30)	C store vlimb into "home" slot
    115 ')
    116 	std,ma		%r3, 0x100(%r30)
    117 	std		%r4, -0xf8(%r30)
    118 	std		%r5, -0xf0(%r30)
    119 	ldo		0(%r0), climb		C clear climb
    120 	fldd		-0x138(%r30), %fr8	C put vlimb in fp register
    121 
    122 define(`p032a1',`%r1')	C
    123 define(`p032a2',`%r19')	C
    124 
    125 define(`m032',`%r20')	C
    126 define(`m096',`%r21')	C
    127 
    128 define(`p000a',`%r22')	C
    129 define(`p064a',`%r29')	C
    130 
    131 define(`s000',`%r31')	C
    132 
    133 define(`ma000',`%r4')	C
    134 define(`ma064',`%r20')	C
    135 
    136 define(`r000',`%r3')	C
    137 
    138 	extrd,u		n, 63, 2, %r5
    139 	cmpb,=		%r5, %r0, L(BIG)
    140 	nop
    141 
    142 	fldd		0(up), %fr4
    143 	ldo		8(up), up
    144 	xmpyu		%fr8R, %fr4L, %fr22
    145 	xmpyu		%fr8L, %fr4R, %fr23
    146 	fstd		%fr22, -0x78(%r30)	C mid product to  -0x78..-0x71
    147 	xmpyu		%fr8R, %fr4R, %fr24
    148 	xmpyu		%fr8L, %fr4L, %fr25
    149 	fstd		%fr23, -0x70(%r30)	C mid product to  -0x70..-0x69
    150 	fstd		%fr24, -0x80(%r30)	C low product to  -0x80..-0x79
    151 	addib,<>	-1, %r5, L(two_or_more)
    152 	fstd		%fr25, -0x68(%r30)	C high product to -0x68..-0x61
    153 LDEF(one)
    154 	ldd		-0x78(%r30), p032a1
    155 	ldd		-0x70(%r30), p032a2
    156 	ldd		-0x80(%r30), p000a
    157 	b		L(0_one_out)
    158 	ldd		-0x68(%r30), p064a
    159 
    160 LDEF(two_or_more)
    161 	fldd		0(up), %fr4
    162 	ldo		8(up), up
    163 	xmpyu		%fr8R, %fr4L, %fr22
    164 	xmpyu		%fr8L, %fr4R, %fr23
    165 	ldd		-0x78(%r30), p032a1
    166 	fstd		%fr22, -0x78(%r30)	C mid product to  -0x78..-0x71
    167 	xmpyu		%fr8R, %fr4R, %fr24
    168 	xmpyu		%fr8L, %fr4L, %fr25
    169 	ldd		-0x70(%r30), p032a2
    170 	fstd		%fr23, -0x70(%r30)	C mid product to  -0x70..-0x69
    171 	ldd		-0x80(%r30), p000a
    172 	fstd		%fr24, -0x80(%r30)	C low product to  -0x80..-0x79
    173 	ldd		-0x68(%r30), p064a
    174 	addib,<>	-1, %r5, L(three_or_more)
    175 	fstd		%fr25, -0x68(%r30)	C high product to -0x68..-0x61
    176 LDEF(two)
    177 	add		p032a1, p032a2, m032
    178 	add,dc		%r0, %r0, m096
    179 	depd,z		m032, 31, 32, ma000
    180 	extrd,u		m032, 31, 32, ma064
    181 	ldd		0(rp), r000
    182 	b		L(0_two_out)
    183 	depd		m096, 31, 32, ma064
    184 
    185 LDEF(three_or_more)
    186 	fldd		0(up), %fr4
    187 	add		p032a1, p032a2, m032
    188 	add,dc		%r0, %r0, m096
    189 	depd,z		m032, 31, 32, ma000
    190 	extrd,u		m032, 31, 32, ma064
    191 	ldd		0(rp), r000
    192 C	addib,=		-1, %r5, L(0_out)
    193 	depd		m096, 31, 32, ma064
    194 LDEF(loop0)
    195 C	xmpyu		%fr8R, %fr4L, %fr22
    196 C	xmpyu		%fr8L, %fr4R, %fr23
    197 C	ldd		-0x78(%r30), p032a1
    198 C	fstd		%fr22, -0x78(%r30)	C mid product to  -0x78..-0x71
    199 C
    200 C	xmpyu		%fr8R, %fr4R, %fr24
    201 C	xmpyu		%fr8L, %fr4L, %fr25
    202 C	ldd		-0x70(%r30), p032a2
    203 C	fstd		%fr23, -0x70(%r30)	C mid product to  -0x70..-0x69
    204 C
    205 C	ldo		8(rp), rp
    206 C	add		climb, p000a, s000
    207 C	ldd		-0x80(%r30), p000a
    208 C	fstd		%fr24, -0x80(%r30)	C low product to  -0x80..-0x79
    209 C
    210 C	add,dc		p064a, %r0, climb
    211 C	ldo		8(up), up
    212 C	ldd		-0x68(%r30), p064a
    213 C	fstd		%fr25, -0x68(%r30)	C high product to -0x68..-0x61
    214 C
    215 C	add		ma000, s000, s000
    216 C	add,dc		ma064, climb, climb
    217 C	fldd		0(up), %fr4
    218 C
    219 C	sub		r000, s000, s000
    220 C	sub,db		%r0, climb, climb
    221 C	sub		%r0, climb, climb
    222 C	std		s000, -8(rp)
    223 C
    224 C	add		p032a1, p032a2, m032
    225 C	add,dc		%r0, %r0, m096
    226 C
    227 C	depd,z		m032, 31, 32, ma000
    228 C	extrd,u		m032, 31, 32, ma064
    229 C	ldd		0(rp), r000
    230 C	addib,<>	-1, %r5, L(loop0)
    231 C	depd		m096, 31, 32, ma064
    232 LDEF(0_out)
    233 	ldo		8(up), up
    234 	xmpyu		%fr8R, %fr4L, %fr22
    235 	xmpyu		%fr8L, %fr4R, %fr23
    236 	ldd		-0x78(%r30), p032a1
    237 	fstd		%fr22, -0x78(%r30)	C mid product to  -0x78..-0x71
    238 	xmpyu		%fr8R, %fr4R, %fr24
    239 	xmpyu		%fr8L, %fr4L, %fr25
    240 	ldd		-0x70(%r30), p032a2
    241 	fstd		%fr23, -0x70(%r30)	C mid product to  -0x70..-0x69
    242 	ldo		8(rp), rp
    243 	add		climb, p000a, s000
    244 	ldd		-0x80(%r30), p000a
    245 	fstd		%fr24, -0x80(%r30)	C low product to  -0x80..-0x79
    246 	add,dc		p064a, %r0, climb
    247 	ldd		-0x68(%r30), p064a
    248 	fstd		%fr25, -0x68(%r30)	C high product to -0x68..-0x61
    249 	add		ma000, s000, s000
    250 	add,dc		ma064, climb, climb
    251 	sub		r000, s000, s000
    252 	sub,db		%r0, climb, climb
    253 	sub		%r0, climb, climb
    254 	std		s000, -8(rp)
    255 	add		p032a1, p032a2, m032
    256 	add,dc		%r0, %r0, m096
    257 	depd,z		m032, 31, 32, ma000
    258 	extrd,u		m032, 31, 32, ma064
    259 	ldd		0(rp), r000
    260 	depd		m096, 31, 32, ma064
    261 LDEF(0_two_out)
    262 	ldd		-0x78(%r30), p032a1
    263 	ldd		-0x70(%r30), p032a2
    264 	ldo		8(rp), rp
    265 	add		climb, p000a, s000
    266 	ldd		-0x80(%r30), p000a
    267 	add,dc		p064a, %r0, climb
    268 	ldd		-0x68(%r30), p064a
    269 	add		ma000, s000, s000
    270 	add,dc		ma064, climb, climb
    271 	sub		r000, s000, s000
    272 	sub,db		%r0, climb, climb
    273 	sub		%r0, climb, climb
    274 	std		s000, -8(rp)
    275 LDEF(0_one_out)
    276 	add		p032a1, p032a2, m032
    277 	add,dc		%r0, %r0, m096
    278 	depd,z		m032, 31, 32, ma000
    279 	extrd,u		m032, 31, 32, ma064
    280 	ldd		0(rp), r000
    281 	depd		m096, 31, 32, ma064
    282 
    283 	add		climb, p000a, s000
    284 	add,dc		p064a, %r0, climb
    285 	add		ma000, s000, s000
    286 	add,dc		ma064, climb, climb
    287 	sub		r000, s000, s000
    288 	sub,db		%r0, climb, climb
    289 	sub		%r0, climb, climb
    290 	std		s000, 0(rp)
    291 
    292 	cmpib,>=	4, n, L(done)
    293 	ldo		8(rp), rp
    294 
    295 C 4-way unrolled code.
    296 
    297 LDEF(BIG)
    298 
    299 define(`p032a1',`%r1')	C
    300 define(`p032a2',`%r19')	C
    301 define(`p096b1',`%r20')	C
    302 define(`p096b2',`%r21')	C
    303 define(`p160c1',`%r22')	C
    304 define(`p160c2',`%r29')	C
    305 define(`p224d1',`%r31')	C
    306 define(`p224d2',`%r3')	C
    307 			C
    308 define(`m032',`%r4')	C
    309 define(`m096',`%r5')	C
    310 define(`m160',`%r6')	C
    311 define(`m224',`%r7')	C
    312 define(`m288',`%r8')	C
    313 			C
    314 define(`p000a',`%r1')	C
    315 define(`p064a',`%r19')	C
    316 define(`p064b',`%r20')	C
    317 define(`p128b',`%r21')	C
    318 define(`p128c',`%r22')	C
    319 define(`p192c',`%r29')	C
    320 define(`p192d',`%r31')	C
    321 define(`p256d',`%r3')	C
    322 			C
    323 define(`s000',`%r10')	C
    324 define(`s064',`%r11')	C
    325 define(`s128',`%r12')	C
    326 define(`s192',`%r13')	C
    327 			C
    328 define(`ma000',`%r9')	C
    329 define(`ma064',`%r4')	C
    330 define(`ma128',`%r5')	C
    331 define(`ma192',`%r6')	C
    332 define(`ma256',`%r7')	C
    333 			C
    334 define(`r000',`%r1')	C
    335 define(`r064',`%r19')	C
    336 define(`r128',`%r20')	C
    337 define(`r192',`%r21')	C
    338 
    339 	std		%r6, -0xe8(%r30)
    340 	std		%r7, -0xe0(%r30)
    341 	std		%r8, -0xd8(%r30)
    342 	std		%r9, -0xd0(%r30)
    343 	std		%r10, -0xc8(%r30)
    344 	std		%r11, -0xc0(%r30)
    345 	std		%r12, -0xb8(%r30)
    346 	std		%r13, -0xb0(%r30)
    347 
    348 ifdef(`HAVE_ABI_2_0w',
    349 `	extrd,u		n, 61, 62, n		C right shift 2
    350 ',`	extrd,u		n, 61, 30, n		C right shift 2, zero extend
    351 ')
    352 
    353 LDEF(4_or_more)
    354 	fldd		0(up), %fr4
    355 	fldd		8(up), %fr5
    356 	fldd		16(up), %fr6
    357 	fldd		24(up), %fr7
    358 	xmpyu		%fr8R, %fr4L, %fr22
    359 	xmpyu		%fr8L, %fr4R, %fr23
    360 	xmpyu		%fr8R, %fr5L, %fr24
    361 	xmpyu		%fr8L, %fr5R, %fr25
    362 	xmpyu		%fr8R, %fr6L, %fr26
    363 	xmpyu		%fr8L, %fr6R, %fr27
    364 	fstd		%fr22, -0x78(%r30)	C mid product to  -0x78..-0x71
    365 	xmpyu		%fr8R, %fr7L, %fr28
    366 	xmpyu		%fr8L, %fr7R, %fr29
    367 	fstd		%fr23, -0x70(%r30)	C mid product to  -0x70..-0x69
    368 	xmpyu		%fr8R, %fr4R, %fr30
    369 	xmpyu		%fr8L, %fr4L, %fr31
    370 	fstd		%fr24, -0x38(%r30)	C mid product to  -0x38..-0x31
    371 	xmpyu		%fr8R, %fr5R, %fr22
    372 	xmpyu		%fr8L, %fr5L, %fr23
    373 	fstd		%fr25, -0x30(%r30)	C mid product to  -0x30..-0x29
    374 	xmpyu		%fr8R, %fr6R, %fr24
    375 	xmpyu		%fr8L, %fr6L, %fr25
    376 	fstd		%fr26, -0x58(%r30)	C mid product to  -0x58..-0x51
    377 	xmpyu		%fr8R, %fr7R, %fr26
    378 	fstd		%fr27, -0x50(%r30)	C mid product to  -0x50..-0x49
    379 	addib,<>	-1, n, L(8_or_more)
    380 	xmpyu		%fr8L, %fr7L, %fr27
    381 	fstd		%fr28, -0x18(%r30)	C mid product to  -0x18..-0x11
    382 	fstd		%fr29, -0x10(%r30)	C mid product to  -0x10..-0x09
    383 	fstd		%fr30, -0x80(%r30)	C low product to  -0x80..-0x79
    384 	fstd		%fr31, -0x68(%r30)	C high product to -0x68..-0x61
    385 	fstd		%fr22, -0x40(%r30)	C low product to  -0x40..-0x39
    386 	fstd		%fr23, -0x28(%r30)	C high product to -0x28..-0x21
    387 	fstd		%fr24, -0x60(%r30)	C low product to  -0x60..-0x59
    388 	fstd		%fr25, -0x48(%r30)	C high product to -0x48..-0x41
    389 	fstd		%fr26, -0x20(%r30)	C low product to  -0x20..-0x19
    390 	fstd		%fr27, -0x88(%r30)	C high product to -0x88..-0x81
    391 	ldd		-0x78(%r30), p032a1
    392 	ldd		-0x70(%r30), p032a2
    393 	ldd		-0x38(%r30), p096b1
    394 	ldd		-0x30(%r30), p096b2
    395 	ldd		-0x58(%r30), p160c1
    396 	ldd		-0x50(%r30), p160c2
    397 	ldd		-0x18(%r30), p224d1
    398 	ldd		-0x10(%r30), p224d2
    399 	b		L(end1)
    400 	nop
    401 
    402 LDEF(8_or_more)
    403 	fstd		%fr28, -0x18(%r30)	C mid product to  -0x18..-0x11
    404 	fstd		%fr29, -0x10(%r30)	C mid product to  -0x10..-0x09
    405 	ldo		32(up), up
    406 	fstd		%fr30, -0x80(%r30)	C low product to  -0x80..-0x79
    407 	fstd		%fr31, -0x68(%r30)	C high product to -0x68..-0x61
    408 	fstd		%fr22, -0x40(%r30)	C low product to  -0x40..-0x39
    409 	fstd		%fr23, -0x28(%r30)	C high product to -0x28..-0x21
    410 	fstd		%fr24, -0x60(%r30)	C low product to  -0x60..-0x59
    411 	fstd		%fr25, -0x48(%r30)	C high product to -0x48..-0x41
    412 	fstd		%fr26, -0x20(%r30)	C low product to  -0x20..-0x19
    413 	fstd		%fr27, -0x88(%r30)	C high product to -0x88..-0x81
    414 	fldd		0(up), %fr4
    415 	fldd		8(up), %fr5
    416 	fldd		16(up), %fr6
    417 	fldd		24(up), %fr7
    418 	xmpyu		%fr8R, %fr4L, %fr22
    419 	ldd		-0x78(%r30), p032a1
    420 	xmpyu		%fr8L, %fr4R, %fr23
    421 	xmpyu		%fr8R, %fr5L, %fr24
    422 	ldd		-0x70(%r30), p032a2
    423 	xmpyu		%fr8L, %fr5R, %fr25
    424 	xmpyu		%fr8R, %fr6L, %fr26
    425 	ldd		-0x38(%r30), p096b1
    426 	xmpyu		%fr8L, %fr6R, %fr27
    427 	fstd		%fr22, -0x78(%r30)	C mid product to  -0x78..-0x71
    428 	xmpyu		%fr8R, %fr7L, %fr28
    429 	ldd		-0x30(%r30), p096b2
    430 	xmpyu		%fr8L, %fr7R, %fr29
    431 	fstd		%fr23, -0x70(%r30)	C mid product to  -0x70..-0x69
    432 	xmpyu		%fr8R, %fr4R, %fr30
    433 	ldd		-0x58(%r30), p160c1
    434 	xmpyu		%fr8L, %fr4L, %fr31
    435 	fstd		%fr24, -0x38(%r30)	C mid product to  -0x38..-0x31
    436 	xmpyu		%fr8R, %fr5R, %fr22
    437 	ldd		-0x50(%r30), p160c2
    438 	xmpyu		%fr8L, %fr5L, %fr23
    439 	fstd		%fr25, -0x30(%r30)	C mid product to  -0x30..-0x29
    440 	xmpyu		%fr8R, %fr6R, %fr24
    441 	ldd		-0x18(%r30), p224d1
    442 	xmpyu		%fr8L, %fr6L, %fr25
    443 	fstd		%fr26, -0x58(%r30)	C mid product to  -0x58..-0x51
    444 	xmpyu		%fr8R, %fr7R, %fr26
    445 	ldd		-0x10(%r30), p224d2
    446 	fstd		%fr27, -0x50(%r30)	C mid product to  -0x50..-0x49
    447 	addib,=		-1, n, L(end2)
    448 	xmpyu		%fr8L, %fr7L, %fr27
    449 LDEF(loop)
    450 	add		p032a1, p032a2, m032
    451 	ldd		-0x80(%r30), p000a
    452 	add,dc		p096b1, p096b2, m096
    453 	fstd		%fr28, -0x18(%r30)	C mid product to  -0x18..-0x11
    454 
    455 	add,dc		p160c1, p160c2, m160
    456 	ldd		-0x68(%r30), p064a
    457 	add,dc		p224d1, p224d2, m224
    458 	fstd		%fr29, -0x10(%r30)	C mid product to  -0x10..-0x09
    459 
    460 	add,dc		%r0, %r0, m288
    461 	ldd		-0x40(%r30), p064b
    462 	ldo		32(up), up
    463 	fstd		%fr30, -0x80(%r30)	C low product to  -0x80..-0x79
    464 
    465 	depd,z		m032, 31, 32, ma000
    466 	ldd		-0x28(%r30), p128b
    467 	extrd,u		m032, 31, 32, ma064
    468 	fstd		%fr31, -0x68(%r30)	C high product to -0x68..-0x61
    469 
    470 	depd		m096, 31, 32, ma064
    471 	ldd		-0x60(%r30), p128c
    472 	extrd,u		m096, 31, 32, ma128
    473 	fstd		%fr22, -0x40(%r30)	C low product to  -0x40..-0x39
    474 
    475 	depd		m160, 31, 32, ma128
    476 	ldd		-0x48(%r30), p192c
    477 	extrd,u		m160, 31, 32, ma192
    478 	fstd		%fr23, -0x28(%r30)	C high product to -0x28..-0x21
    479 
    480 	depd		m224, 31, 32, ma192
    481 	ldd		-0x20(%r30), p192d
    482 	extrd,u		m224, 31, 32, ma256
    483 	fstd		%fr24, -0x60(%r30)	C low product to  -0x60..-0x59
    484 
    485 	depd		m288, 31, 32, ma256
    486 	ldd		-0x88(%r30), p256d
    487 	add		climb, p000a, s000
    488 	fstd		%fr25, -0x48(%r30)	C high product to -0x48..-0x41
    489 
    490 	add,dc		p064a, p064b, s064
    491 	ldd		0(rp), r000
    492 	add,dc		p128b, p128c, s128
    493 	fstd		%fr26, -0x20(%r30)	C low product to  -0x20..-0x19
    494 
    495 	add,dc		p192c, p192d, s192
    496 	ldd		8(rp), r064
    497 	add,dc		p256d, %r0, climb
    498 	fstd		%fr27, -0x88(%r30)	C high product to -0x88..-0x81
    499 
    500 	ldd		16(rp), r128
    501 	add		ma000, s000, s000	C accum mid 0
    502 	ldd		24(rp), r192
    503 	add,dc		ma064, s064, s064	C accum mid 1
    504 
    505 	add,dc		ma128, s128, s128	C accum mid 2
    506 	fldd		0(up), %fr4
    507 	add,dc		ma192, s192, s192	C accum mid 3
    508 	fldd		8(up), %fr5
    509 
    510 	add,dc		ma256, climb, climb
    511 	fldd		16(up), %fr6
    512 	sub		r000, s000, s000	C accum rlimb 0
    513 	fldd		24(up), %fr7
    514 
    515 	sub,db		r064, s064, s064	C accum rlimb 1
    516 	sub,db		r128, s128, s128	C accum rlimb 2
    517 	std		s000, 0(rp)
    518 
    519 	sub,db		r192, s192, s192	C accum rlimb 3
    520 	sub,db		%r0, climb, climb
    521 	sub		%r0, climb, climb
    522 	std		s064, 8(rp)
    523 
    524 	xmpyu		%fr8R, %fr4L, %fr22
    525 	ldd		-0x78(%r30), p032a1
    526 	xmpyu		%fr8L, %fr4R, %fr23
    527 	std		s128, 16(rp)
    528 
    529 	xmpyu		%fr8R, %fr5L, %fr24
    530 	ldd		-0x70(%r30), p032a2
    531 	xmpyu		%fr8L, %fr5R, %fr25
    532 	std		s192, 24(rp)
    533 
    534 	xmpyu		%fr8R, %fr6L, %fr26
    535 	ldd		-0x38(%r30), p096b1
    536 	xmpyu		%fr8L, %fr6R, %fr27
    537 	fstd		%fr22, -0x78(%r30)	C mid product to  -0x78..-0x71
    538 
    539 	xmpyu		%fr8R, %fr7L, %fr28
    540 	ldd		-0x30(%r30), p096b2
    541 	xmpyu		%fr8L, %fr7R, %fr29
    542 	fstd		%fr23, -0x70(%r30)	C mid product to  -0x70..-0x69
    543 
    544 	xmpyu		%fr8R, %fr4R, %fr30
    545 	ldd		-0x58(%r30), p160c1
    546 	xmpyu		%fr8L, %fr4L, %fr31
    547 	fstd		%fr24, -0x38(%r30)	C mid product to  -0x38..-0x31
    548 
    549 	xmpyu		%fr8R, %fr5R, %fr22
    550 	ldd		-0x50(%r30), p160c2
    551 	xmpyu		%fr8L, %fr5L, %fr23
    552 	fstd		%fr25, -0x30(%r30)	C mid product to  -0x30..-0x29
    553 
    554 	xmpyu		%fr8R, %fr6R, %fr24
    555 	ldd		-0x18(%r30), p224d1
    556 	xmpyu		%fr8L, %fr6L, %fr25
    557 	fstd		%fr26, -0x58(%r30)	C mid product to  -0x58..-0x51
    558 
    559 	xmpyu		%fr8R, %fr7R, %fr26
    560 	ldd		-0x10(%r30), p224d2
    561 	fstd		%fr27, -0x50(%r30)	C mid product to  -0x50..-0x49
    562 	xmpyu		%fr8L, %fr7L, %fr27
    563 
    564 	addib,<>	-1, n, L(loop)
    565 	ldo		32(rp), rp
    566 
    567 LDEF(end2)
    568 	add		p032a1, p032a2, m032
    569 	ldd		-0x80(%r30), p000a
    570 	add,dc		p096b1, p096b2, m096
    571 	fstd		%fr28, -0x18(%r30)	C mid product to  -0x18..-0x11
    572 	add,dc		p160c1, p160c2, m160
    573 	ldd		-0x68(%r30), p064a
    574 	add,dc		p224d1, p224d2, m224
    575 	fstd		%fr29, -0x10(%r30)	C mid product to  -0x10..-0x09
    576 	add,dc		%r0, %r0, m288
    577 	ldd		-0x40(%r30), p064b
    578 	fstd		%fr30, -0x80(%r30)	C low product to  -0x80..-0x79
    579 	depd,z		m032, 31, 32, ma000
    580 	ldd		-0x28(%r30), p128b
    581 	extrd,u		m032, 31, 32, ma064
    582 	fstd		%fr31, -0x68(%r30)	C high product to -0x68..-0x61
    583 	depd		m096, 31, 32, ma064
    584 	ldd		-0x60(%r30), p128c
    585 	extrd,u		m096, 31, 32, ma128
    586 	fstd		%fr22, -0x40(%r30)	C low product to  -0x40..-0x39
    587 	depd		m160, 31, 32, ma128
    588 	ldd		-0x48(%r30), p192c
    589 	extrd,u		m160, 31, 32, ma192
    590 	fstd		%fr23, -0x28(%r30)	C high product to -0x28..-0x21
    591 	depd		m224, 31, 32, ma192
    592 	ldd		-0x20(%r30), p192d
    593 	extrd,u		m224, 31, 32, ma256
    594 	fstd		%fr24, -0x60(%r30)	C low product to  -0x60..-0x59
    595 	depd		m288, 31, 32, ma256
    596 	ldd		-0x88(%r30), p256d
    597 	add		climb, p000a, s000
    598 	fstd		%fr25, -0x48(%r30)	C high product to -0x48..-0x41
    599 	add,dc		p064a, p064b, s064
    600 	ldd		0(rp), r000
    601 	add,dc		p128b, p128c, s128
    602 	fstd		%fr26, -0x20(%r30)	C low product to  -0x20..-0x19
    603 	add,dc		p192c, p192d, s192
    604 	ldd		8(rp), r064
    605 	add,dc		p256d, %r0, climb
    606 	fstd		%fr27, -0x88(%r30)	C high product to -0x88..-0x81
    607 	ldd		16(rp), r128
    608 	add		ma000, s000, s000	C accum mid 0
    609 	ldd		24(rp), r192
    610 	add,dc		ma064, s064, s064	C accum mid 1
    611 	add,dc		ma128, s128, s128	C accum mid 2
    612 	add,dc		ma192, s192, s192	C accum mid 3
    613 	add,dc		ma256, climb, climb
    614 	sub		r000, s000, s000	C accum rlimb 0
    615 	sub,db		r064, s064, s064	C accum rlimb 1
    616 	sub,db		r128, s128, s128	C accum rlimb 2
    617 	std		s000, 0(rp)
    618 	sub,db		r192, s192, s192	C accum rlimb 3
    619 	sub,db		%r0, climb, climb
    620 	sub		%r0, climb, climb
    621 	std		s064, 8(rp)
    622 	ldd		-0x78(%r30), p032a1
    623 	std		s128, 16(rp)
    624 	ldd		-0x70(%r30), p032a2
    625 	std		s192, 24(rp)
    626 	ldd		-0x38(%r30), p096b1
    627 	ldd		-0x30(%r30), p096b2
    628 	ldd		-0x58(%r30), p160c1
    629 	ldd		-0x50(%r30), p160c2
    630 	ldd		-0x18(%r30), p224d1
    631 	ldd		-0x10(%r30), p224d2
    632 	ldo		32(rp), rp
    633 
    634 LDEF(end1)
    635 	add		p032a1, p032a2, m032
    636 	ldd		-0x80(%r30), p000a
    637 	add,dc		p096b1, p096b2, m096
    638 	add,dc		p160c1, p160c2, m160
    639 	ldd		-0x68(%r30), p064a
    640 	add,dc		p224d1, p224d2, m224
    641 	add,dc		%r0, %r0, m288
    642 	ldd		-0x40(%r30), p064b
    643 	depd,z		m032, 31, 32, ma000
    644 	ldd		-0x28(%r30), p128b
    645 	extrd,u		m032, 31, 32, ma064
    646 	depd		m096, 31, 32, ma064
    647 	ldd		-0x60(%r30), p128c
    648 	extrd,u		m096, 31, 32, ma128
    649 	depd		m160, 31, 32, ma128
    650 	ldd		-0x48(%r30), p192c
    651 	extrd,u		m160, 31, 32, ma192
    652 	depd		m224, 31, 32, ma192
    653 	ldd		-0x20(%r30), p192d
    654 	extrd,u		m224, 31, 32, ma256
    655 	depd		m288, 31, 32, ma256
    656 	ldd		-0x88(%r30), p256d
    657 	add		climb, p000a, s000
    658 	add,dc		p064a, p064b, s064
    659 	ldd		0(rp), r000
    660 	add,dc		p128b, p128c, s128
    661 	add,dc		p192c, p192d, s192
    662 	ldd		8(rp), r064
    663 	add,dc		p256d, %r0, climb
    664 	ldd		16(rp), r128
    665 	add		ma000, s000, s000	C accum mid 0
    666 	ldd		24(rp), r192
    667 	add,dc		ma064, s064, s064	C accum mid 1
    668 	add,dc		ma128, s128, s128	C accum mid 2
    669 	add,dc		ma192, s192, s192	C accum mid 3
    670 	add,dc		ma256, climb, climb
    671 	sub		r000, s000, s000	C accum rlimb 0
    672 	sub,db		r064, s064, s064	C accum rlimb 1
    673 	sub,db		r128, s128, s128	C accum rlimb 2
    674 	std		s000, 0(rp)
    675 	sub,db		r192, s192, s192	C accum rlimb 3
    676 	sub,db		%r0, climb, climb
    677 	sub		%r0, climb, climb
    678 	std		s064, 8(rp)
    679 	std		s128, 16(rp)
    680 	std		s192, 24(rp)
    681 
    682 	ldd		-0xb0(%r30), %r13
    683 	ldd		-0xb8(%r30), %r12
    684 	ldd		-0xc0(%r30), %r11
    685 	ldd		-0xc8(%r30), %r10
    686 	ldd		-0xd0(%r30), %r9
    687 	ldd		-0xd8(%r30), %r8
    688 	ldd		-0xe0(%r30), %r7
    689 	ldd		-0xe8(%r30), %r6
    690 LDEF(done)
    691 ifdef(`HAVE_ABI_2_0w',
    692 `	copy		climb, %r28
    693 ',`	extrd,u		climb, 63, 32, %r29
    694 	extrd,u		climb, 31, 32, %r28
    695 ')
    696 	ldd		-0xf0(%r30), %r5
    697 	ldd		-0xf8(%r30), %r4
    698 	bve		(%r2)
    699 	ldd,mb		-0x100(%r30), %r3
    700 EPILOGUE(mpn_submul_1)
    701