2 # Copyright 2009-2018 The OpenSSL Project Authors. All Rights Reserved.
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
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 # ====================================================================
17 # On PA-7100LC this module performs ~90-50% better, less for longer
18 # keys, than code generated by gcc 3.2 for PA-RISC 1.1. Latter means
19 # that compiler utilized xmpyu instruction to perform 32x32=64-bit
20 # multiplication, which in turn means that "baseline" performance was
21 # optimal in respect to instruction set capabilities. Fair comparison
22 # with vendor compiler is problematic, because OpenSSL doesn't define
23 # BN_LLONG [presumably] for historical reasons, which drives compiler
24 # toward 4 times 16x16=32-bit multiplications [plus complementary
25 # shifts and additions] instead. This means that you should observe
26 # several times improvement over code generated by vendor compiler
27 # for PA-RISC 1.1, but the "baseline" is far from optimal. The actual
28 # improvement coefficient was never collected on PA-7100LC, or any
29 # other 1.1 CPU, because I don't have access to such machine with
30 # vendor compiler. But to give you a taste, PA-RISC 1.1 code path
31 # reportedly outperformed code generated by cc +DA1.1 +O3 by factor
34 # On PA-RISC 2.0 it has to compete with pa-risc2[W].s, which is
35 # reportedly ~2x faster than vendor compiler generated code [according
36 # to comment in pa-risc2[W].s]. Here comes a catch. Execution core of
37 # this implementation is actually 32-bit one, in the sense that it
38 # operates on 32-bit values. But pa-risc2[W].s operates on arrays of
39 # 64-bit BN_LONGs... How do they interoperate then? No problem. This
40 # module picks halves of 64-bit values in reverse order and pretends
41 # they were 32-bit BN_LONGs. But can 32-bit core compete with "pure"
42 # 64-bit code such as pa-risc2[W].s then? Well, the thing is that
43 # 32x32=64-bit multiplication is the best even PA-RISC 2.0 can do,
44 # i.e. there is no "wider" multiplication like on most other 64-bit
45 # platforms. This means that even being effectively 32-bit, this
46 # implementation performs "64-bit" computational task in same amount
47 # of arithmetic operations, most notably multiplications. It requires
48 # more memory references, most notably to tp[num], but this doesn't
49 # seem to exhaust memory port capacity. And indeed, dedicated PA-RISC
50 # 2.0 code path provides virtually same performance as pa-risc2[W].s:
51 # it's ~10% better for shortest key length and ~10% worse for longest
54 # In case it wasn't clear. The module has two distinct code paths:
55 # PA-RISC 1.1 and PA-RISC 2.0 ones. Latter features carry-free 64-bit
56 # additions and 64-bit integer loads, not to mention specific
57 # instruction scheduling. In 64-bit build naturally only 2.0 code path
58 # is assembled. In 32-bit application context both code paths are
59 # assembled, PA-RISC 2.0 CPU is detected at run-time and proper path
60 # is taken automatically. Also, in 32-bit build the module imposes
61 # couple of limitations: vector lengths has to be even and vector
62 # addresses has to be 64-bit aligned. Normally neither is a problem:
63 # most common key lengths are even and vectors are commonly malloc-ed,
64 # which ensures alignment.
66 # Special thanks to polarhome.com for providing HP-UX account on
67 # PA-RISC 1.1 machine, and to correspondent who chose to remain
68 # anonymous for testing the code on PA-RISC 2.0 machine.
70 $0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1;
72 # $output is the last argument if it looks like a file (it has an extension)
73 # $flavour is the first argument if it doesn't look like a file
74 $output = $#ARGV >= 0 && $ARGV[$#ARGV] =~ m|\.\w+$| ? pop : undef;
75 $flavour = $#ARGV >= 0 && $ARGV[0] !~ m|\.| ? shift : undef;
77 $output and open STDOUT,">$output";
79 if ($flavour =~ /64/) {
90 $LEVEL ="1.1"; #$LEVEL.="\n\t.ALLOW\t2.0";
99 if (open CONF,"<${dir}../../opensslconf.h") {
101 if (m/#\s*define\s+SIXTY_FOUR_BIT/) {
111 $FRAME=8*$SIZE_T+$FRAME_MARKER; # 8 saved regs + frame marker
112 # [+ argument transfer]
113 $LOCALS=$FRAME-$FRAME_MARKER;
114 $FRAME+=32; # local variables
124 $n0="%r22"; # passed through stack in 32-bit
125 $num="%r21"; # passed through stack in 32-bit
138 $xfer=$n0; # accommodates [-16..15] offset in fld[dw]s
140 $fm0="%fr4"; $fti=$fm0;
143 $fai="%fr6"; $fab0="%fr7"; $fab1="%fr8";
144 $fni="%fr9"; $fnm0="%fr10"; $fnm1="%fr11";
149 .SUBSPA \$CODE\$,QUAD=0,ALIGN=8,ACCESS=0x2C,CODE_ONLY
151 .EXPORT bn_mul_mont,ENTRY,ARGW0=GR,ARGW1=GR,ARGW2=GR,ARGW3=GR
155 .CALLINFO FRAME=`$FRAME-8*$SIZE_T`,NO_CALLS,SAVE_RP,SAVE_SP,ENTRY_GR=6
157 $PUSH %r2,-$SAVED_RP(%sp) ; standard prologue
158 $PUSHMA %r3,$FRAME(%sp)
159 $PUSH %r4,`-$FRAME+1*$SIZE_T`(%sp)
160 $PUSH %r5,`-$FRAME+2*$SIZE_T`(%sp)
161 $PUSH %r6,`-$FRAME+3*$SIZE_T`(%sp)
162 $PUSH %r7,`-$FRAME+4*$SIZE_T`(%sp)
163 $PUSH %r8,`-$FRAME+5*$SIZE_T`(%sp)
164 $PUSH %r9,`-$FRAME+6*$SIZE_T`(%sp)
165 $PUSH %r10,`-$FRAME+7*$SIZE_T`(%sp)
168 $code.=<<___ if ($SIZE_T==4);
169 ldw `-$FRAME_MARKER-4`($fp),$n0
170 ldw `-$FRAME_MARKER-8`($fp),$num
174 $code.=<<___ if ($BN_SZ==4);
175 comiclr,<= 6,$num,%r0 ; are vectors long enough?
177 ldi 0,%r28 ; signal "unhandled"
178 add,ev %r0,$num,$num ; is $num even?
182 extru,= $ti1,31,3,%r0 ; are ap and np 64-bit aligned?
189 fldws,ma 4($bp),${fbi} ; bp[0]
191 $code.=<<___ if ($BN_SZ==8);
192 comib,> 3,$num,L\$abort ; are vectors long enough?
193 ldi 0,%r28 ; signal "unhandled"
194 addl $num,$num,$num ; I operate on 32-bit values
196 fldws 4($n0),${fn0} ; only low part of n0
197 fldws 4($bp),${fbi} ; bp[0] in flipped word order
200 fldds 0($ap),${fai} ; ap[0,1]
201 fldds 0($np),${fni} ; np[0,1]
203 sh2addl $num,%r0,$arrsz
205 ldo 36($arrsz),$hi1 ; space for tp[num+1]
206 andcm $hi1,$hi0,$hi1 ; align
208 $PUSH $fp,-$SIZE_T(%sp)
210 ldo `$LOCALS+16`($fp),$xfer
211 ldo `$LOCALS+32+4`($fp),$tp
213 xmpyu ${fai}L,${fbi},${fab0} ; ap[0]*bp[0]
214 xmpyu ${fai}R,${fbi},${fab1} ; ap[1]*bp[0]
215 xmpyu ${fn0},${fab0}R,${fm0}
217 addl $arrsz,$ap,$ap ; point at the end
219 subi 0,$arrsz,$idx ; j=0
220 ldo 8($idx),$idx ; j++++
222 xmpyu ${fni}L,${fm0}R,${fnm0} ; np[0]*m
223 xmpyu ${fni}R,${fm0}R,${fnm1} ; np[1]*m
224 fstds ${fab0},-16($xfer)
225 fstds ${fnm0},-8($xfer)
226 fstds ${fab1},0($xfer)
227 fstds ${fnm1},8($xfer)
228 flddx $idx($ap),${fai} ; ap[2,3]
229 flddx $idx($np),${fni} ; np[2,3]
231 $code.=<<___ if ($BN_SZ==4);
232 mtctl $hi0,%cr11 ; $hi0 still holds 31
233 extrd,u,*= $hi0,%sar,1,$hi0 ; executes on PA-RISC 1.0
237 $code.=<<___; # PA-RISC 2.0 code-path
238 xmpyu ${fai}L,${fbi},${fab0} ; ap[j]*bp[0]
239 xmpyu ${fni}L,${fm0}R,${fnm0} ; np[j]*m
241 fstds ${fab0},-16($xfer)
243 extrd,u $ab0,31,32,$hi0
244 extrd,u $ab0,63,32,$ab0
246 fstds ${fnm0},-8($xfer)
247 ldo 8($idx),$idx ; j++++
248 addl $ab0,$nm0,$nm0 ; low part is discarded
249 extrd,u $nm0,31,32,$hi1
252 xmpyu ${fai}R,${fbi},${fab1} ; ap[j+1]*bp[0]
253 xmpyu ${fni}R,${fm0}R,${fnm1} ; np[j+1]*m
255 fstds ${fab1},0($xfer)
257 extrd,u $ab1,31,32,$hi0
259 fstds ${fnm1},8($xfer)
260 extrd,u $ab1,63,32,$ab1
262 flddx $idx($ap),${fai} ; ap[j,j+1]
263 flddx $idx($np),${fni} ; np[j,j+1]
265 extrd,u $nm1,31,32,$hi1
267 xmpyu ${fai}L,${fbi},${fab0} ; ap[j]*bp[0]
268 xmpyu ${fni}L,${fm0}R,${fnm0} ; np[j]*m
270 fstds ${fab0},-16($xfer)
272 extrd,u $ab0,31,32,$hi0
274 fstds ${fnm0},-8($xfer)
275 extrd,u $ab0,63,32,$ab0
277 stw $nm1,-4($tp) ; tp[j-1]
279 stw,ma $nm0,8($tp) ; tp[j-1]
280 addib,<> 8,$idx,L\$1st ; j++++
281 extrd,u $nm0,31,32,$hi1
283 xmpyu ${fai}R,${fbi},${fab1} ; ap[j]*bp[0]
284 xmpyu ${fni}R,${fm0}R,${fnm1} ; np[j]*m
286 fstds ${fab1},0($xfer)
288 extrd,u $ab1,31,32,$hi0
290 fstds ${fnm1},8($xfer)
291 extrd,u $ab1,63,32,$ab1
296 extrd,u $nm1,31,32,$hi1
299 extrd,u $ab0,31,32,$hi0
300 stw $nm1,-4($tp) ; tp[j-1]
301 extrd,u $ab0,63,32,$ab0
306 extrd,u $nm0,31,32,$hi1
307 stw,ma $nm0,8($tp) ; tp[j-1]
309 ldo -1($num),$num ; i--
310 subi 0,$arrsz,$idx ; j=0
312 $code.=<<___ if ($BN_SZ==4);
313 fldws,ma 4($bp),${fbi} ; bp[1]
315 $code.=<<___ if ($BN_SZ==8);
316 fldws 0($bp),${fbi} ; bp[1] in flipped word order
319 flddx $idx($ap),${fai} ; ap[0,1]
320 flddx $idx($np),${fni} ; np[0,1]
321 fldws 8($xfer),${fti}R ; tp[0]
323 extrd,u $ab1,31,32,$hi0
324 extrd,u $ab1,63,32,$ab1
325 ldo 8($idx),$idx ; j++++
326 xmpyu ${fai}L,${fbi},${fab0} ; ap[0]*bp[1]
327 xmpyu ${fai}R,${fbi},${fab1} ; ap[1]*bp[1]
330 extrd,u $nm1,31,32,$hi1
331 fstws,mb ${fab0}L,-8($xfer) ; save high part
332 stw $nm1,-4($tp) ; tp[j-1]
334 fcpy,sgl %fr0,${fti}L ; zero high part
335 fcpy,sgl %fr0,${fab0}L
337 extrd,u $hi0,31,32,$hi1
338 fcnvxf,dbl,dbl ${fti},${fti} ; 32-bit unsigned int -> double
339 fcnvxf,dbl,dbl ${fab0},${fab0}
343 fadd,dbl ${fti},${fab0},${fab0} ; add tp[0]
344 fcnvfx,dbl,dbl ${fab0},${fab0} ; double -> 33-bit unsigned int
345 xmpyu ${fn0},${fab0}R,${fm0}
346 ldo `$LOCALS+32+4`($fp),$tp
348 xmpyu ${fni}L,${fm0}R,${fnm0} ; np[0]*m
349 xmpyu ${fni}R,${fm0}R,${fnm1} ; np[1]*m
350 fstds ${fab0},-16($xfer) ; 33-bit value
351 fstds ${fnm0},-8($xfer)
352 flddx $idx($ap),${fai} ; ap[2]
353 flddx $idx($np),${fni} ; np[2]
354 ldo 8($idx),$idx ; j++++
355 ldd -16($xfer),$ab0 ; 33-bit value
357 ldw 0($xfer),$hi0 ; high part
359 xmpyu ${fai}L,${fbi},${fab0} ; ap[j]*bp[i]
360 xmpyu ${fni}L,${fm0}R,${fnm0} ; np[j]*m
361 extrd,u $ab0,31,32,$ti0 ; carry bit
362 extrd,u $ab0,63,32,$ab0
363 fstds ${fab1},0($xfer)
364 addl $ti0,$hi0,$hi0 ; account carry bit
365 fstds ${fnm1},8($xfer)
366 addl $ab0,$nm0,$nm0 ; low part is discarded
367 ldw 0($tp),$ti1 ; tp[1]
368 extrd,u $nm0,31,32,$hi1
369 fstds ${fab0},-16($xfer)
370 fstds ${fnm0},-8($xfer)
373 xmpyu ${fai}R,${fbi},${fab1} ; ap[j+1]*bp[i]
374 xmpyu ${fni}R,${fm0}R,${fnm1} ; np[j+1]*m
376 fstds ${fab1},0($xfer)
380 fstds ${fnm1},8($xfer)
381 extrd,u $ab1,31,32,$hi0
382 extrd,u $ab1,63,32,$ab1
383 flddx $idx($ap),${fai} ; ap[j,j+1]
384 flddx $idx($np),${fni} ; np[j,j+1]
387 ldw 4($tp),$ti0 ; tp[j]
388 stw $nm1,-4($tp) ; tp[j-1]
390 xmpyu ${fai}L,${fbi},${fab0} ; ap[j]*bp[i]
391 xmpyu ${fni}L,${fm0}R,${fnm0} ; np[j]*m
393 fstds ${fab0},-16($xfer)
397 fstds ${fnm0},-8($xfer)
398 extrd,u $ab0,31,32,$hi0
399 extrd,u $nm1,31,32,$hi1
400 ldw 8($tp),$ti1 ; tp[j]
401 extrd,u $ab0,63,32,$ab0
404 stw,ma $nm0,8($tp) ; tp[j-1]
405 addib,<> 8,$idx,L\$inner ; j++++
406 extrd,u $nm0,31,32,$hi1
408 xmpyu ${fai}R,${fbi},${fab1} ; ap[j]*bp[i]
409 xmpyu ${fni}R,${fm0}R,${fnm1} ; np[j]*m
411 fstds ${fab1},0($xfer)
415 fstds ${fnm1},8($xfer)
416 extrd,u $ab1,31,32,$hi0
417 extrd,u $ab1,63,32,$ab1
418 ldw 4($tp),$ti0 ; tp[j]
423 extrd,u $nm1,31,32,$hi1
427 stw $nm1,-4($tp) ; tp[j-1]
428 extrd,u $ab0,31,32,$hi0
429 ldw 8($tp),$ti1 ; tp[j]
430 extrd,u $ab0,63,32,$ab0
435 extrd,u $nm0,31,32,$hi1
436 stw,ma $nm0,8($tp) ; tp[j-1]
438 addib,= -1,$num,L\$outerdone ; i--
439 subi 0,$arrsz,$idx ; j=0
441 $code.=<<___ if ($BN_SZ==4);
442 fldws,ma 4($bp),${fbi} ; bp[i]
444 $code.=<<___ if ($BN_SZ==8);
445 ldi 12,$ti0 ; bp[i] in flipped word order
446 addl,ev %r0,$num,$num
452 flddx $idx($ap),${fai} ; ap[0]
454 flddx $idx($np),${fni} ; np[0]
455 fldws 8($xfer),${fti}R ; tp[0]
457 extrd,u $ab1,31,32,$hi0
458 extrd,u $ab1,63,32,$ab1
460 ldo 8($idx),$idx ; j++++
461 xmpyu ${fai}L,${fbi},${fab0} ; ap[0]*bp[i]
462 xmpyu ${fai}R,${fbi},${fab1} ; ap[1]*bp[i]
463 ldw 4($tp),$ti0 ; tp[j]
466 fstws,mb ${fab0}L,-8($xfer) ; save high part
468 extrd,u $nm1,31,32,$hi1
469 fcpy,sgl %fr0,${fti}L ; zero high part
470 fcpy,sgl %fr0,${fab0}L
471 stw $nm1,-4($tp) ; tp[j-1]
473 fcnvxf,dbl,dbl ${fti},${fti} ; 32-bit unsigned int -> double
474 fcnvxf,dbl,dbl ${fab0},${fab0}
476 fadd,dbl ${fti},${fab0},${fab0} ; add tp[0]
478 extrd,u $hi0,31,32,$hi1
479 fcnvfx,dbl,dbl ${fab0},${fab0} ; double -> 33-bit unsigned int
482 xmpyu ${fn0},${fab0}R,${fm0}
485 ldo `$LOCALS+32+4`($fp),$tp
490 extrd,u $ab1,31,32,$hi0
491 extrd,u $ab1,63,32,$ab1
493 ldw 4($tp),$ti0 ; tp[j]
497 extrd,u $nm1,31,32,$hi1
498 stw $nm1,-4($tp) ; tp[j-1]
502 extrd,u $hi0,31,32,$hi1
506 ldo `$LOCALS+32`($fp),$tp
507 sub %r0,%r0,%r0 ; clear borrow
509 $code.=<<___ if ($BN_SZ==4);
511 extru,= $rp,31,3,%r0 ; is rp 64-bit aligned?
518 addib,<> 4,$idx,L\$sub
523 $code.=<<___ if ($BN_SZ==8);
527 shrpd $ti0,$ti0,32,$ti0 ; flip word order
528 std $ti0,-8($tp) ; save flipped value
529 sub,db $ti0,$hi0,$hi1
531 addib,<> 8,$idx,L\$sub
534 extrd,u $ti0,31,32,$ti0 ; carry in flipped word order
538 ldo `$LOCALS+32`($fp),$tp
539 sub $rp,$arrsz,$rp ; rewind rp
547 addib,<> 8,$idx,L\$copy
551 if ($BN_SZ==4) { # PA-RISC 1.1 code-path
565 xmpyu ${fai}L,${fbi},${fab0} ; ap[j]*bp[0]
566 xmpyu ${fni}L,${fm0}R,${fnm0} ; np[j]*m
571 fstds ${fab0},-16($xfer)
572 fstds ${fnm0},-8($xfer)
574 ldo 8($idx),$idx ; j++++
575 add $ablo,$nmlo0,$nmlo0 ; discarded
582 xmpyu ${fai}R,${fbi},${fab1} ; ap[j+1]*bp[0]
583 flddx $idx($ap),${fai} ; ap[j,j+1]
584 xmpyu ${fni}R,${fm0}R,${fnm1} ; np[j+1]*m
585 flddx $idx($np),${fni} ; np[j,j+1]
590 add $ablo,$nmlo1,$nmlo1
591 fstds ${fab1},0($xfer)
592 addc %r0,$nmhi1,$nmhi1
593 fstds ${fnm1},8($xfer)
594 add $hi1,$nmlo1,$nmlo1
599 xmpyu ${fai}L,${fbi},${fab0} ; ap[j]*bp[0]
601 xmpyu ${fni}L,${fm0}R,${fnm0} ; np[j]*m
604 stw $nmlo1,-4($tp) ; tp[j-1]
606 fstds ${fab0},-16($xfer)
607 add $ablo,$nmlo0,$nmlo0
608 fstds ${fnm0},-8($xfer)
609 addc %r0,$nmhi0,$nmhi0
611 add $hi1,$nmlo0,$nmlo0
613 stws,ma $nmlo0,8($tp) ; tp[j-1]
614 addib,<> 8,$idx,L\$1st_pa11 ; j++++
619 xmpyu ${fai}R,${fbi},${fab1} ; ap[j]*bp[0]
620 xmpyu ${fni}R,${fm0}R,${fnm1} ; np[j]*m
622 fstds ${fab1},0($xfer)
624 fstds ${fnm1},8($xfer)
625 add $ablo,$nmlo1,$nmlo1
627 addc %r0,$nmhi1,$nmhi1
629 add $hi1,$nmlo1,$nmlo1
635 stw $nmlo1,-4($tp) ; tp[j-1]
638 add $ablo,$nmlo0,$nmlo0
640 addc %r0,$nmhi0,$nmhi0
641 ldws,mb 8($xfer),$nmhi1
642 add $hi1,$nmlo0,$nmlo0
645 stws,ma $nmlo0,8($tp) ; tp[j-1]
647 ldo -1($num),$num ; i--
648 subi 0,$arrsz,$idx ; j=0
650 fldws,ma 4($bp),${fbi} ; bp[1]
651 flddx $idx($ap),${fai} ; ap[0,1]
652 flddx $idx($np),${fni} ; np[0,1]
653 fldws 8($xfer),${fti}R ; tp[0]
656 ldo 8($idx),$idx ; j++++
657 xmpyu ${fai}L,${fbi},${fab0} ; ap[0]*bp[1]
658 xmpyu ${fai}R,${fbi},${fab1} ; ap[1]*bp[1]
659 add $hi1,$nmlo1,$nmlo1
660 addc %r0,$nmhi1,$nmhi1
661 add $ablo,$nmlo1,$nmlo1
663 fstws,mb ${fab0}L,-8($xfer) ; save high part
664 stw $nmlo1,-4($tp) ; tp[j-1]
666 fcpy,sgl %fr0,${fti}L ; zero high part
667 fcpy,sgl %fr0,${fab0}L
670 fcnvxf,dbl,dbl ${fti},${fti} ; 32-bit unsigned int -> double
671 fcnvxf,dbl,dbl ${fab0},${fab0}
675 fadd,dbl ${fti},${fab0},${fab0} ; add tp[0]
676 fcnvfx,dbl,dbl ${fab0},${fab0} ; double -> 33-bit unsigned int
677 xmpyu ${fn0},${fab0}R,${fm0}
678 ldo `$LOCALS+32+4`($fp),$tp
680 xmpyu ${fni}L,${fm0}R,${fnm0} ; np[0]*m
681 xmpyu ${fni}R,${fm0}R,${fnm1} ; np[1]*m
682 fstds ${fab0},-16($xfer) ; 33-bit value
683 fstds ${fnm0},-8($xfer)
684 flddx $idx($ap),${fai} ; ap[2,3]
685 flddx $idx($np),${fni} ; np[2,3]
686 ldw -16($xfer),$abhi ; carry bit actually
687 ldo 8($idx),$idx ; j++++
691 ldw 0($xfer),$hi0 ; high part
693 xmpyu ${fai}L,${fbi},${fab0} ; ap[j]*bp[i]
694 xmpyu ${fni}L,${fm0}R,${fnm0} ; np[j]*m
695 fstds ${fab1},0($xfer)
696 addl $abhi,$hi0,$hi0 ; account carry bit
697 fstds ${fnm1},8($xfer)
698 add $ablo,$nmlo0,$nmlo0 ; discarded
699 ldw 0($tp),$ti1 ; tp[1]
701 fstds ${fab0},-16($xfer)
702 fstds ${fnm0},-8($xfer)
707 xmpyu ${fai}R,${fbi},${fab1} ; ap[j+1]*bp[i]
708 flddx $idx($ap),${fai} ; ap[j,j+1]
709 xmpyu ${fni}R,${fm0}R,${fnm1} ; np[j+1]*m
710 flddx $idx($np),${fni} ; np[j,j+1]
712 ldw 4($tp),$ti0 ; tp[j]
718 fstds ${fab1},0($xfer)
719 add $ablo,$nmlo1,$nmlo1
720 fstds ${fnm1},8($xfer)
721 addc %r0,$nmhi1,$nmhi1
723 add $hi1,$nmlo1,$nmlo1
727 xmpyu ${fai}L,${fbi},${fab0} ; ap[j]*bp[i]
728 ldw 8($tp),$ti1 ; tp[j]
729 xmpyu ${fni}L,${fm0}R,${fnm0} ; np[j]*m
734 stw $nmlo1,-4($tp) ; tp[j-1]
736 fstds ${fab0},-16($xfer)
738 fstds ${fnm0},-8($xfer)
739 add $ablo,$nmlo0,$nmlo0
741 addc %r0,$nmhi0,$nmhi0
743 add $hi1,$nmlo0,$nmlo0
744 stws,ma $nmlo0,8($tp) ; tp[j-1]
745 addib,<> 8,$idx,L\$inner_pa11 ; j++++
748 xmpyu ${fai}R,${fbi},${fab1} ; ap[j]*bp[i]
750 xmpyu ${fni}R,${fm0}R,${fnm1} ; np[j]*m
753 ldw 4($tp),$ti0 ; tp[j]
755 fstds ${fab1},0($xfer)
757 fstds ${fnm1},8($xfer)
760 add $ablo,$nmlo1,$nmlo1
762 addc %r0,$nmhi1,$nmhi1
764 add $hi1,$nmlo1,$nmlo1
769 stw $nmlo1,-4($tp) ; tp[j-1]
772 ldw 8($tp),$ti1 ; tp[j]
775 add $ablo,$nmlo0,$nmlo0
777 addc %r0,$nmhi0,$nmhi0
778 ldws,mb 8($xfer),$nmhi1
779 add $hi1,$nmlo0,$nmlo0
782 stws,ma $nmlo0,8($tp) ; tp[j-1]
784 addib,= -1,$num,L\$outerdone_pa11; i--
785 subi 0,$arrsz,$idx ; j=0
787 fldws,ma 4($bp),${fbi} ; bp[i]
788 flddx $idx($ap),${fai} ; ap[0]
791 flddx $idx($np),${fni} ; np[0]
792 fldws 8($xfer),${fti}R ; tp[0]
796 ldo 8($idx),$idx ; j++++
797 xmpyu ${fai}L,${fbi},${fab0} ; ap[0]*bp[i]
798 xmpyu ${fai}R,${fbi},${fab1} ; ap[1]*bp[i]
799 ldw 4($tp),$ti0 ; tp[j]
801 add $hi1,$nmlo1,$nmlo1
802 addc %r0,$nmhi1,$nmhi1
803 fstws,mb ${fab0}L,-8($xfer) ; save high part
804 add $ablo,$nmlo1,$nmlo1
806 fcpy,sgl %fr0,${fti}L ; zero high part
807 fcpy,sgl %fr0,${fab0}L
808 stw $nmlo1,-4($tp) ; tp[j-1]
810 fcnvxf,dbl,dbl ${fti},${fti} ; 32-bit unsigned int -> double
811 fcnvxf,dbl,dbl ${fab0},${fab0}
814 fadd,dbl ${fti},${fab0},${fab0} ; add tp[0]
817 fcnvfx,dbl,dbl ${fab0},${fab0} ; double -> 33-bit unsigned int
820 xmpyu ${fn0},${fab0}R,${fm0}
823 ldo `$LOCALS+32+4`($fp),$tp
831 ldw 4($tp),$ti0 ; tp[j]
833 add $hi1,$nmlo1,$nmlo1
834 addc %r0,$nmhi1,$nmhi1
835 add $ablo,$nmlo1,$nmlo1
837 stw $nmlo1,-4($tp) ; tp[j-1]
846 ldo `$LOCALS+32+4`($fp),$tp
847 sub %r0,%r0,%r0 ; clear borrow
854 addib,<> 4,$idx,L\$sub_pa11
859 ldo `$LOCALS+32`($fp),$tp
860 sub $rp,$arrsz,$rp ; rewind rp
868 addib,<> 4,$idx,L\$copy_pa11
877 ldi 1,%r28 ; signal "handled"
878 ldo $FRAME($fp),%sp ; destroy tp[num+1]
880 $POP `-$FRAME-$SAVED_RP`(%sp),%r2 ; standard epilogue
881 $POP `-$FRAME+1*$SIZE_T`(%sp),%r4
882 $POP `-$FRAME+2*$SIZE_T`(%sp),%r5
883 $POP `-$FRAME+3*$SIZE_T`(%sp),%r6
884 $POP `-$FRAME+4*$SIZE_T`(%sp),%r7
885 $POP `-$FRAME+5*$SIZE_T`(%sp),%r8
886 $POP `-$FRAME+6*$SIZE_T`(%sp),%r9
887 $POP `-$FRAME+7*$SIZE_T`(%sp),%r10
891 $POPMB -$FRAME(%sp),%r3
893 .STRINGZ "Montgomery Multiplication for PA-RISC, CRYPTOGAMS by <appro\@openssl.org>"
896 # Explicitly encode PA-RISC 2.0 instructions used in this module, so
897 # that it can be compiled with .LEVEL 1.0. It should be noted that I
898 # wouldn't have to do this, if GNU assembler understood .ALLOW 2.0
902 my ($mod,$args) = @_;
903 my $orig = "ldd$mod\t$args";
905 if ($args =~ /%r([0-9]+)\(%r([0-9]+)\),%r([0-9]+)/) # format 4
906 { my $opcode=(0x03<<26)|($2<<21)|($1<<16)|(3<<6)|$3;
907 sprintf "\t.WORD\t0x%08x\t; %s",$opcode,$orig;
909 elsif ($args =~ /(\-?[0-9]+)\(%r([0-9]+)\),%r([0-9]+)/) # format 5
910 { my $opcode=(0x03<<26)|($2<<21)|(1<<12)|(3<<6)|$3;
911 $opcode|=(($1&0xF)<<17)|(($1&0x10)<<12); # encode offset
912 $opcode|=(1<<5) if ($mod =~ /^,m/);
913 $opcode|=(1<<13) if ($mod =~ /^,mb/);
914 sprintf "\t.WORD\t0x%08x\t; %s",$opcode,$orig;
920 my ($mod,$args) = @_;
921 my $orig = "std$mod\t$args";
923 if ($args =~ /%r([0-9]+),(\-?[0-9]+)\(%r([0-9]+)\)/) # format 6
924 { my $opcode=(0x03<<26)|($3<<21)|($1<<16)|(1<<12)|(0xB<<6);
925 $opcode|=(($2&0xF)<<1)|(($2&0x10)>>4); # encode offset
926 $opcode|=(1<<5) if ($mod =~ /^,m/);
927 $opcode|=(1<<13) if ($mod =~ /^,mb/);
928 sprintf "\t.WORD\t0x%08x\t; %s",$opcode,$orig;
934 my ($mod,$args) = @_;
935 my $orig = "extrd$mod\t$args";
937 # I only have ",u" completer, it's implicitly encoded...
938 if ($args =~ /%r([0-9]+),([0-9]+),([0-9]+),%r([0-9]+)/) # format 15
939 { my $opcode=(0x36<<26)|($1<<21)|($4<<16);
941 $opcode |= (($2&0x20)<<6)|(($2&0x1f)<<5); # encode pos
942 $opcode |= (($len&0x20)<<7)|($len&0x1f); # encode len
943 sprintf "\t.WORD\t0x%08x\t; %s",$opcode,$orig;
945 elsif ($args =~ /%r([0-9]+),%sar,([0-9]+),%r([0-9]+)/) # format 12
946 { my $opcode=(0x34<<26)|($1<<21)|($3<<16)|(2<<11)|(1<<9);
948 $opcode |= (($len&0x20)<<3)|($len&0x1f); # encode len
949 $opcode |= (1<<13) if ($mod =~ /,\**=/);
950 sprintf "\t.WORD\t0x%08x\t; %s",$opcode,$orig;
956 my ($mod,$args) = @_;
957 my $orig = "shrpd$mod\t$args";
959 if ($args =~ /%r([0-9]+),%r([0-9]+),([0-9]+),%r([0-9]+)/) # format 14
960 { my $opcode=(0x34<<26)|($2<<21)|($1<<16)|(1<<10)|$4;
962 $opcode |= (($cpos&0x20)<<6)|(($cpos&0x1f)<<5); # encode sa
963 sprintf "\t.WORD\t0x%08x\t; %s",$opcode,$orig;
969 my ($mod,$args) = @_;
970 my $orig = "sub$mod\t$args";
972 if ($mod eq ",db" && $args =~ /%r([0-9]+),%r([0-9]+),%r([0-9]+)/) {
973 my $opcode=(0x02<<26)|($2<<21)|($1<<16)|$3;
974 $opcode|=(1<<10); # e1
975 $opcode|=(1<<8); # e2
977 sprintf "\t.WORD\t0x%08x\t; %s",$opcode,$orig
983 my ($mnemonic,$mod,$args)=@_;
984 my $opcode = eval("\$$mnemonic");
986 ref($opcode) eq 'CODE' ? &$opcode($mod,$args) : "\t$mnemonic$mod\t$args";
989 if (`$ENV{CC} -Wa,-v -c -o /dev/null -x assembler /dev/null 2>&1`
990 =~ /GNU assembler/) {
994 foreach (split("\n",$code)) {
995 s/\`([^\`]*)\`/eval $1/ge;
996 # flip word order in 64-bit mode...
997 s/(xmpyu\s+)($fai|$fni)([LR])/$1.$2.($3 eq "L"?"R":"L")/e if ($BN_SZ==8);
998 # assemble 2.0 instructions in 32-bit mode...
999 s/^\s+([a-z]+)([\S]*)\s+([\S]*)/&assemble($1,$2,$3)/e if ($BN_SZ==4);
1001 s/(\.LEVEL\s+2\.0)W/$1w/ if ($gnuas && $SIZE_T==8);
1002 s/\.SPACE\s+\$TEXT\$/.text/ if ($gnuas && $SIZE_T==8);
1003 s/\.SUBSPA.*// if ($gnuas && $SIZE_T==8);
1004 s/\bbv\b/bve/ if ($SIZE_T==8);
1008 close STDOUT or die "error closing STDOUT";