RC4 tune-up for Intel P4 core, both 32- and 64-bit ones. As it's
[openssl.git] / crypto / perlasm / x86unix.pl
1 #!/usr/local/bin/perl
2
3 package x86unix;        # GAS actually...
4
5 $label="L000";
6 $const="";
7 $constl=0;
8
9 $align=($main'aout)?"4":"16";
10 $under=($main'aout or $main'coff)?"_":"";
11 $dot=($main'aout)?"":".";
12 $com_start="#" if ($main'aout or $main'coff);
13
14 sub main'asm_init_output { @out=(); }
15 sub main'asm_get_output { return(@out); }
16 sub main'get_labels { return(@labels); }
17 sub main'external_label { push(@labels,@_); }
18
19 if ($main'cpp)
20         {
21         $align="ALIGN";
22         $under="";
23         $com_start='/*';
24         $com_end='*/';
25         }
26
27 %lb=(   'eax',  '%al',
28         'ebx',  '%bl',
29         'ecx',  '%cl',
30         'edx',  '%dl',
31         'ax',   '%al',
32         'bx',   '%bl',
33         'cx',   '%cl',
34         'dx',   '%dl',
35         );
36
37 %hb=(   'eax',  '%ah',
38         'ebx',  '%bh',
39         'ecx',  '%ch',
40         'edx',  '%dh',
41         'ax',   '%ah',
42         'bx',   '%bh',
43         'cx',   '%ch',
44         'dx',   '%dh',
45         );
46
47 %regs=( 'eax',  '%eax',
48         'ebx',  '%ebx',
49         'ecx',  '%ecx',
50         'edx',  '%edx',
51         'esi',  '%esi',
52         'edi',  '%edi',
53         'ebp',  '%ebp',
54         'esp',  '%esp',
55
56         'mm0',  '%mm0',
57         'mm1',  '%mm1',
58         'mm2',  '%mm2',
59         'mm3',  '%mm3',
60         'mm4',  '%mm4',
61         'mm5',  '%mm5',
62         'mm6',  '%mm6',
63         'mm7',  '%mm7',
64
65         'xmm0', '%xmm0',
66         'xmm1', '%xmm1',
67         'xmm2', '%xmm2',
68         'xmm3', '%xmm3',
69         'xmm4', '%xmm4',
70         'xmm5', '%xmm5',
71         'xmm6', '%xmm6',
72         'xmm7', '%xmm7',
73         );
74
75 %reg_val=(
76         'eax',  0x00,
77         'ebx',  0x03,
78         'ecx',  0x01,
79         'edx',  0x02,
80         'esi',  0x06,
81         'edi',  0x07,
82         'ebp',  0x05,
83         'esp',  0x04,
84         );
85
86 sub main'LB
87         {
88         (defined($lb{$_[0]})) || die "$_[0] does not have a 'low byte'\n";
89         return($lb{$_[0]});
90         }
91
92 sub main'HB
93         {
94         (defined($hb{$_[0]})) || die "$_[0] does not have a 'high byte'\n";
95         return($hb{$_[0]});
96         }
97
98 sub main'DWP
99         {
100         local($addr,$reg1,$reg2,$idx)=@_;
101
102         $ret="";
103         $addr =~ s/(^|[+ \t])([A-Za-z_]+[A-Za-z0-9_]+)($|[+ \t])/$1$under$2$3/;
104         $reg1="$regs{$reg1}" if defined($regs{$reg1});
105         $reg2="$regs{$reg2}" if defined($regs{$reg2});
106         $ret.=$addr if ($addr ne "") && ($addr ne 0);
107         if ($reg2 ne "")
108                 {
109                 if($idx ne "" && $idx != 0)
110                     { $ret.="($reg1,$reg2,$idx)"; }
111                 else
112                     { $ret.="($reg1,$reg2)"; }
113                 }
114         elsif ($reg1 ne "")
115                 { $ret.="($reg1)" }
116         return($ret);
117         }
118
119 sub main'QWP
120         {
121         return(&main'DWP(@_));
122         }
123
124 sub main'BP
125         {
126         return(&main'DWP(@_));
127         }
128
129 sub main'BC
130         {
131         return @_;
132         }
133
134 sub main'DWC
135         {
136         return @_;
137         }
138
139 #sub main'BP
140 #       {
141 #       local($addr,$reg1,$reg2,$idx)=@_;
142 #
143 #       $ret="";
144 #
145 #       $addr =~ s/(^|[+ \t])([A-Za-z_]+)($|[+ \t])/$1$under$2$3/;
146 #       $reg1="$regs{$reg1}" if defined($regs{$reg1});
147 #       $reg2="$regs{$reg2}" if defined($regs{$reg2});
148 #       $ret.=$addr if ($addr ne "") && ($addr ne 0);
149 #       if ($reg2 ne "")
150 #               { $ret.="($reg1,$reg2,$idx)"; }
151 #       else
152 #               { $ret.="($reg1)" }
153 #       return($ret);
154 #       }
155
156 sub main'mov    { &out2("movl",@_); }
157 sub main'movb   { &out2("movb",@_); }
158 sub main'and    { &out2("andl",@_); }
159 sub main'or     { &out2("orl",@_); }
160 sub main'shl    { &out2("sall",@_); }
161 sub main'shr    { &out2("shrl",@_); }
162 sub main'xor    { &out2("xorl",@_); }
163 sub main'xorb   { &out2("xorb",@_); }
164 sub main'add    { &out2($_[0]=~/%[a-d][lh]/?"addb":"addl",@_); }
165 sub main'adc    { &out2("adcl",@_); }
166 sub main'sub    { &out2("subl",@_); }
167 sub main'sbb    { &out2("sbbl",@_); }
168 sub main'rotl   { &out2("roll",@_); }
169 sub main'rotr   { &out2("rorl",@_); }
170 sub main'exch   { &out2("xchg",@_); }
171 sub main'cmp    { &out2("cmpl",@_); }
172 sub main'lea    { &out2("leal",@_); }
173 sub main'mul    { &out1("mull",@_); }
174 sub main'div    { &out1("divl",@_); }
175 sub main'jmp    { &out1("jmp",@_); }
176 sub main'jmp_ptr { &out1p("jmp",@_); }
177 sub main'je     { &out1("je",@_); }
178 sub main'jle    { &out1("jle",@_); }
179 sub main'jne    { &out1("jne",@_); }
180 sub main'jnz    { &out1("jnz",@_); }
181 sub main'jz     { &out1("jz",@_); }
182 sub main'jge    { &out1("jge",@_); }
183 sub main'jl     { &out1("jl",@_); }
184 sub main'ja     { &out1("ja",@_); }
185 sub main'jae    { &out1("jae",@_); }
186 sub main'jb     { &out1("jb",@_); }
187 sub main'jbe    { &out1("jbe",@_); }
188 sub main'jc     { &out1("jc",@_); }
189 sub main'jnc    { &out1("jnc",@_); }
190 sub main'jno    { &out1("jno",@_); }
191 sub main'dec    { &out1("decl",@_); }
192 sub main'inc    { &out1($_[0]=~/%[a-d][hl]/?"incb":"incl",@_); }
193 sub main'push   { &out1("pushl",@_); $stack+=4; }
194 sub main'pop    { &out1("popl",@_); $stack-=4; }
195 sub main'pushf  { &out0("pushfl"); $stack+=4; }
196 sub main'popf   { &out0("popfl"); $stack-=4; }
197 sub main'not    { &out1("notl",@_); }
198 sub main'call   {       my $pre=$under;
199                         foreach $i (%label)
200                         { if ($label{$i} eq $_[0]) { $pre=''; last; } }
201                         &out1("call",$pre.$_[0]);
202                 }
203 sub main'ret    { &out0("ret"); }
204 sub main'nop    { &out0("nop"); }
205 sub main'test   { &out2("testl",@_); }
206 sub main'bt     { &out2("btl",@_); }
207 sub main'leave  { &out0("leave"); }
208 sub main'cpuid  { &out0(".byte\t0x0f,0xa2"); }
209 sub main'rdtsc  { &out0(".byte\t0x0f,0x31"); }
210 sub main'halt   { &out0("hlt"); }
211 sub main'movz   { &out2("movzb",@_); }
212
213 # SSE2
214 sub main'emms   { &out0("emms"); }
215 sub main'movd   { &out2("movd",@_); }
216 sub main'movq   { &out2("movq",@_); }
217 sub main'movdqu { &out2("movdqu",@_); }
218 sub main'movdqa { &out2("movdqa",@_); }
219 sub main'movdq2q{ &out2("movdq2q",@_); }
220 sub main'movq2dq{ &out2("movq2dq",@_); }
221 sub main'paddq  { &out2("paddq",@_); }
222 sub main'pmuludq{ &out2("pmuludq",@_); }
223 sub main'psrlq  { &out2("psrlq",@_); }
224 sub main'psllq  { &out2("psllq",@_); }
225 sub main'pxor   { &out2("pxor",@_); }
226 sub main'por    { &out2("por",@_); }
227 sub main'pand   { &out2("pand",@_); }
228
229 # The bswapl instruction is new for the 486. Emulate if i386.
230 sub main'bswap
231         {
232         if ($main'i386)
233                 {
234                 &main'comment("bswapl @_");
235                 &main'exch(main'HB(@_),main'LB(@_));
236                 &main'rotr(@_,16);
237                 &main'exch(main'HB(@_),main'LB(@_));
238                 }
239         else
240                 {
241                 &out1("bswapl",@_);
242                 }
243         }
244
245 sub out2
246         {
247         local($name,$p1,$p2)=@_;
248         local($l,$ll,$t);
249         local(%special)=(       "roll",0xD1C0,"rorl",0xD1C8,
250                                 "rcll",0xD1D0,"rcrl",0xD1D8,
251                                 "shll",0xD1E0,"shrl",0xD1E8,
252                                 "sarl",0xD1F8);
253         
254         if ((defined($special{$name})) && defined($regs{$p1}) && ($p2 == 1))
255                 {
256                 $op=$special{$name}|$reg_val{$p1};
257                 $tmp1=sprintf(".byte %d\n",($op>>8)&0xff);
258                 $tmp2=sprintf(".byte %d\t",$op     &0xff);
259                 push(@out,$tmp1);
260                 push(@out,$tmp2);
261
262                 $p2=&conv($p2);
263                 $p1=&conv($p1);
264                 &main'comment("$name $p2 $p1");
265                 return;
266                 }
267
268         push(@out,"\t$name\t");
269         $t=&conv($p2).",";
270         $l=length($t);
271         push(@out,$t);
272         $ll=4-($l+9)/8;
273         $tmp1=sprintf("\t" x $ll);
274         push(@out,$tmp1);
275         push(@out,&conv($p1)."\n");
276         }
277
278 sub out1
279         {
280         local($name,$p1)=@_;
281         local($l,$t);
282         local(%special)=("bswapl",0x0FC8);
283
284         if ((defined($special{$name})) && defined($regs{$p1}))
285                 {
286                 $op=$special{$name}|$reg_val{$p1};
287                 $tmp1=sprintf(".byte %d\n",($op>>8)&0xff);
288                 $tmp2=sprintf(".byte %d\t",$op     &0xff);
289                 push(@out,$tmp1);
290                 push(@out,$tmp2);
291
292                 $p2=&conv($p2);
293                 $p1=&conv($p1);
294                 &main'comment("$name $p2 $p1");
295                 return;
296                 }
297
298         push(@out,"\t$name\t".&conv($p1)."\n");
299         }
300
301 sub out1p
302         {
303         local($name,$p1)=@_;
304         local($l,$t);
305
306         push(@out,"\t$name\t*".&conv($p1)."\n");
307         }
308
309 sub out0
310         {
311         push(@out,"\t$_[0]\n");
312         }
313
314 sub conv
315         {
316         local($p)=@_;
317
318 #       $p =~ s/0x([0-9A-Fa-f]+)/0$1h/;
319
320         $p=$regs{$p} if (defined($regs{$p}));
321
322         $p =~ s/^(-{0,1}[0-9A-Fa-f]+)$/\$$1/;
323         $p =~ s/^(0x[0-9A-Fa-f]+)$/\$$1/;
324         return $p;
325         }
326
327 sub main'file
328         {
329         local($file)=@_;
330
331         local($tmp)=<<"EOF";
332         .file   "$file.s"
333 EOF
334         push(@out,$tmp);
335         }
336
337 sub main'function_begin
338         {
339         local($func)=@_;
340
341         &main'external_label($func);
342         $func=$under.$func;
343
344         local($tmp)=<<"EOF";
345 .text
346 .globl  $func
347 EOF
348         push(@out,$tmp);
349         if ($main'cpp)
350                 { $tmp=push(@out,"TYPE($func,\@function)\n"); }
351         elsif ($main'coff)
352                 { $tmp=push(@out,".def\t$func;\t.scl\t2;\t.type\t32;\t.endef\n"); }
353         elsif ($main'aout and !$main'pic)
354                 { }
355         else    { $tmp=push(@out,".type\t$func,\@function\n"); }
356         push(@out,".align\t$align\n");
357         push(@out,"$func:\n");
358         $tmp=<<"EOF";
359         pushl   %ebp
360         pushl   %ebx
361         pushl   %esi
362         pushl   %edi
363
364 EOF
365         push(@out,$tmp);
366         $stack=20;
367         }
368
369 sub main'function_begin_B
370         {
371         local($func,$extra)=@_;
372
373         &main'external_label($func);
374         $func=$under.$func;
375
376         local($tmp)=<<"EOF";
377 .text
378 .globl  $func
379 EOF
380         push(@out,$tmp);
381         if ($main'cpp)
382                 { push(@out,"TYPE($func,\@function)\n"); }
383         elsif ($main'coff)
384                 { $tmp=push(@out,".def\t$func;\t.scl\t2;\t.type\t32;\t.endef\n"); }
385         elsif ($main'aout and !$main'pic)
386                 { }
387         else    { push(@out,".type      $func,\@function\n"); }
388         push(@out,".align\t$align\n");
389         push(@out,"$func:\n");
390         $stack=4;
391         }
392
393 sub main'function_end
394         {
395         local($func)=@_;
396
397         $func=$under.$func;
398
399         local($tmp)=<<"EOF";
400         popl    %edi
401         popl    %esi
402         popl    %ebx
403         popl    %ebp
404         ret
405 ${dot}L_${func}_end:
406 EOF
407         push(@out,$tmp);
408
409         if ($main'cpp)
410                 { push(@out,"SIZE($func,${dot}L_${func}_end-$func)\n"); }
411         elsif ($main'coff or $main'aout)
412                 { }
413         else    { push(@out,".size\t$func,${dot}L_${func}_end-$func\n"); }
414         push(@out,".ident       \"$func\"\n");
415         $stack=0;
416         %label=();
417         }
418
419 sub main'function_end_A
420         {
421         local($func)=@_;
422
423         local($tmp)=<<"EOF";
424         popl    %edi
425         popl    %esi
426         popl    %ebx
427         popl    %ebp
428         ret
429 EOF
430         push(@out,$tmp);
431         }
432
433 sub main'function_end_B
434         {
435         local($func)=@_;
436
437         $func=$under.$func;
438
439         push(@out,"${dot}L_${func}_end:\n");
440         if ($main'cpp)
441                 { push(@out,"SIZE($func,${dot}L_${func}_end-$func)\n"); }
442         elsif ($main'coff or $main'aout)
443                 { }
444         else    { push(@out,".size\t$func,${dot}L_${func}_end-$func\n"); }
445         push(@out,".ident       \"$func\"\n");
446         $stack=0;
447         %label=();
448         }
449
450 sub main'wparam
451         {
452         local($num)=@_;
453
454         return(&main'DWP($stack+$num*4,"esp","",0));
455         }
456
457 sub main'stack_push
458         {
459         local($num)=@_;
460         $stack+=$num*4;
461         &main'sub("esp",$num*4);
462         }
463
464 sub main'stack_pop
465         {
466         local($num)=@_;
467         $stack-=$num*4;
468         &main'add("esp",$num*4);
469         }
470
471 sub main'swtmp
472         {
473         return(&main'DWP($_[0]*4,"esp","",0));
474         }
475
476 # Should use swtmp, which is above esp.  Linix can trash the stack above esp
477 #sub main'wtmp
478 #       {
479 #       local($num)=@_;
480 #
481 #       return(&main'DWP(-($num+1)*4,"esp","",0));
482 #       }
483
484 sub main'comment
485         {
486         if (!defined($com_start) or $main'elf)
487                 {       # Regarding $main'elf above...
488                         # GNU and SVR4 as'es use different comment delimiters,
489                 push(@out,"\n");        # so we just skip ELF comments...
490                 return;
491                 }
492         foreach (@_)
493                 {
494                 if (/^\s*$/)
495                         { push(@out,"\n"); }
496                 else
497                         { push(@out,"\t$com_start $_ $com_end\n"); }
498                 }
499         }
500
501 sub main'label
502         {
503         if (!defined($label{$_[0]}))
504                 {
505                 $label{$_[0]}="${dot}${label}${_[0]}";
506                 $label++;
507                 }
508         return($label{$_[0]});
509         }
510
511 sub main'set_label
512         {
513         if (!defined($label{$_[0]}))
514                 {
515                 $label{$_[0]}="${dot}${label}${_[0]}";
516                 $label++;
517                 }
518         if ($_[1]!=0)
519                 {
520                 if ($_[1]>1)    { main'align($_[1]);            }
521                 else            { push(@out,".align $align\n"); }
522                 }
523         push(@out,"$label{$_[0]}:\n");
524         }
525
526 sub main'file_end
527         {
528         # try to detect if SSE2 or MMX extensions were used on ELF platform...
529         if ($main'elf && grep {/%[x]*mm[0-7]/i} @out) {
530                 local($tmp);
531
532                 push (@out,"\n.comm\t${under}OPENSSL_ia32cap_P,4,4\n");
533
534                 push (@out,".section\t.init\n");
535                 # One can argue that it's wasteful to craft every
536                 # SSE/MMX module with this snippet... Well, it's 72
537                 # bytes long and for the moment we have two modules.
538                 # Let's argue when we have 7 modules or so...
539                 #
540                 # $1<<10 sets a reserved bit to signal that variable
541                 # was initialized already...
542                 &main'picmeup("edx","OPENSSL_ia32cap_P");
543                 $tmp=<<___;
544                 cmpl    \$0,(%edx)
545                 jne     1f
546                 movl    \$1<<10,(%edx)
547                 pushf
548                 popl    %eax
549                 movl    %eax,%ecx
550                 xorl    \$1<<21,%eax
551                 pushl   %eax
552                 popf
553                 pushf
554                 popl    %eax
555                 xorl    %ecx,%eax
556                 btl     \$21,%eax
557                 jnc     1f
558                 pushl   %edi
559                 pushl   %ebx
560                 movl    %edx,%edi
561                 movl    \$1,%eax
562                 .byte   0x0f,0xa2
563                 orl     \$1<<10,%edx
564                 movl    %edx,0(%edi)
565                 popl    %ebx
566                 popl    %edi
567         .align  $align
568         1:
569 ___
570                 push (@out,$tmp);
571         }
572
573         if ($const ne "")
574                 {
575                 push(@out,".section .rodata\n");
576                 push(@out,$const);
577                 $const="";
578                 }
579         }
580
581 sub main'data_word
582         {
583         push(@out,"\t.long\t".join(',',@_)."\n");
584         }
585
586 sub main'align
587         {
588         my $val=$_[0],$p2,$i;
589         if ($main'aout) {
590                 for ($p2=0;$val!=0;$val>>=1) { $p2++; }
591                 $val=$p2-1;
592                 $val.=",0x90";
593         }
594         push(@out,".align\t$val\n");
595         }
596
597 # debug output functions: puts, putx, printf
598
599 sub main'puts
600         {
601         &pushvars();
602         &main'push('$Lstring' . ++$constl);
603         &main'call('puts');
604         $stack-=4;
605         &main'add("esp",4);
606         &popvars();
607
608         $const .= "Lstring$constl:\n\t.string \"@_[0]\"\n";
609         }
610
611 sub main'putx
612         {
613         &pushvars();
614         &main'push($_[0]);
615         &main'push('$Lstring' . ++$constl);
616         &main'call('printf');
617         &main'add("esp",8);
618         $stack-=8;
619         &popvars();
620
621         $const .= "Lstring$constl:\n\t.string \"\%X\"\n";
622         }
623
624 sub main'printf
625         {
626         $ostack = $stack;
627         &pushvars();
628         for ($i = @_ - 1; $i >= 0; $i--)
629                 {
630                 if ($i == 0) # change this to support %s format strings
631                         {
632                         &main'push('$Lstring' . ++$constl);
633                         $const .= "Lstring$constl:\n\t.string \"@_[$i]\"\n";
634                         }
635                 else
636                         {
637                         if ($_[$i] =~ /([0-9]*)\(%esp\)/)
638                                 {
639                                 &main'push(($1 + $stack - $ostack) . '(%esp)');
640                                 }
641                         else
642                                 {
643                                 &main'push($_[$i]);
644                                 }
645                         }
646                 }
647         &main'call('printf');
648         $stack-=4*@_;
649         &main'add("esp",4*@_);
650         &popvars();
651         }
652
653 sub pushvars
654         {
655         &main'pushf();
656         &main'push("edx");
657         &main'push("ecx");
658         &main'push("eax");
659         }
660
661 sub popvars
662         {
663         &main'pop("eax");
664         &main'pop("ecx");
665         &main'pop("edx");
666         &main'popf();
667         }
668
669 sub main'picmeup
670         {
671         local($dst,$sym)=@_;
672         if ($main'cpp)
673                 {
674                 local($tmp)=<<___;
675 #if (defined(ELF) || defined(SOL)) && defined(PIC)
676         call    1f
677 1:      popl    $regs{$dst}
678         addl    \$_GLOBAL_OFFSET_TABLE_+[.-1b],$regs{$dst}
679         movl    $sym\@GOT($regs{$dst}),$regs{$dst}
680 #else
681         leal    $sym,$regs{$dst}
682 #endif
683 ___
684                 push(@out,$tmp);
685                 }
686         elsif ($main'pic && ($main'elf || $main'aout))
687                 {
688                 &main'call(&main'label("PIC_me_up"));
689                 &main'set_label("PIC_me_up");
690                 &main'blindpop($dst);
691                 &main'add($dst,"\$${under}_GLOBAL_OFFSET_TABLE_+[.-".
692                                 &main'label("PIC_me_up") . "]");
693                 &main'mov($dst,&main'DWP($under.$sym."\@GOT",$dst));
694                 }
695         else
696                 {
697                 &main'lea($dst,&main'DWP($sym));
698                 }
699         }
700
701 sub main'blindpop { &out1("popl",@_); }
702
703 sub main'initseg
704         {
705         local($f)=@_;
706         local($tmp);
707         if ($main'elf)
708                 {
709                 $tmp=<<___;
710 .section        .init
711         call    $under$f
712         .align  $align
713 ___
714                 }
715         elsif ($main'coff)
716                 {
717                 $tmp=<<___;     # applies to both Cygwin and Mingw
718 .section        .ctors
719 .long   $under$f
720 ___
721                 }
722         elsif ($main'aout)
723                 {
724                 local($ctor)="${under}_GLOBAL_\$I\$$f";
725                 $tmp=".text\n";
726                 $tmp.=".type    $ctor,\@function\n" if ($main'pic);
727                 $tmp.=<<___;    # OpenBSD way...
728 .globl  $ctor
729 .align  2
730 $ctor:
731         jmp     $under$f
732 ___
733                 }
734         push(@out,$tmp) if ($tmp);
735         }
736
737 1;