Support for indirect calls in x86 assembler modules.
[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($_[0]=~/%[a-d][lh]/?"xchgb":"xchgl",@_); }
171 sub main'cmp    { &out2("cmpl",@_); }
172 sub main'lea    { &out2("leal",@_); }
173 sub main'mul    { &out1("mull",@_); }
174 sub main'div    { &out1("divl",@_); }
175 sub main'jmp    { &out1("jmp",@_); }
176 sub main'jmp_ptr { &out1p("jmp",@_); }
177 sub main'je     { &out1("je",@_); }
178 sub main'jle    { &out1("jle",@_); }
179 sub main'jne    { &out1("jne",@_); }
180 sub main'jnz    { &out1("jnz",@_); }
181 sub main'jz     { &out1("jz",@_); }
182 sub main'jge    { &out1("jge",@_); }
183 sub main'jl     { &out1("jl",@_); }
184 sub main'ja     { &out1("ja",@_); }
185 sub main'jae    { &out1("jae",@_); }
186 sub main'jb     { &out1("jb",@_); }
187 sub main'jbe    { &out1("jbe",@_); }
188 sub main'jc     { &out1("jc",@_); }
189 sub main'jnc    { &out1("jnc",@_); }
190 sub main'jno    { &out1("jno",@_); }
191 sub main'dec    { &out1("decl",@_); }
192 sub main'inc    { &out1($_[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.$_[0]);
202                 }
203 sub main'call_ptr { &out1p("call",@_); }
204 sub main'ret    { &out0("ret"); }
205 sub main'nop    { &out0("nop"); }
206 sub main'test   { &out2("testl",@_); }
207 sub main'bt     { &out2("btl",@_); }
208 sub main'leave  { &out0("leave"); }
209 sub main'cpuid  { &out0(".byte\t0x0f,0xa2"); }
210 sub main'rdtsc  { &out0(".byte\t0x0f,0x31"); }
211 sub main'halt   { &out0("hlt"); }
212 sub main'movz   { &out2("movzbl",@_); }
213 sub main'neg    { &out1("negl",@_); }
214 sub main'cld    { &out0("cld"); }
215
216 # SSE2
217 sub main'emms   { &out0("emms"); }
218 sub main'movd   { &out2("movd",@_); }
219 sub main'movdqu { &out2("movdqu",@_); }
220 sub main'movdqa { &out2("movdqa",@_); }
221 sub main'movdq2q{ &out2("movdq2q",@_); }
222 sub main'movq2dq{ &out2("movq2dq",@_); }
223 sub main'paddq  { &out2("paddq",@_); }
224 sub main'pmuludq{ &out2("pmuludq",@_); }
225 sub main'psrlq  { &out2("psrlq",@_); }
226 sub main'psllq  { &out2("psllq",@_); }
227 sub main'pxor   { &out2("pxor",@_); }
228 sub main'por    { &out2("por",@_); }
229 sub main'pand   { &out2("pand",@_); }
230 sub main'movq   {
231         local($p1,$p2,$optimize)=@_;
232         if ($optimize && $p1=~/^mm[0-7]$/ && $p2=~/^mm[0-7]$/)
233                 # movq between mmx registers can sink Intel CPUs
234                 {       push(@out,"\tpshufw\t\$0xe4,%$p2,%$p1\n");      }
235         else    {       &out2("movq",@_);                               }
236         }
237
238 # The bswapl instruction is new for the 486. Emulate if i386.
239 sub main'bswap
240         {
241         if ($main'i386)
242                 {
243                 &main'comment("bswapl @_");
244                 &main'exch(main'HB(@_),main'LB(@_));
245                 &main'rotr(@_,16);
246                 &main'exch(main'HB(@_),main'LB(@_));
247                 }
248         else
249                 {
250                 &out1("bswapl",@_);
251                 }
252         }
253
254 sub out2
255         {
256         local($name,$p1,$p2)=@_;
257         local($l,$ll,$t);
258         local(%special)=(       "roll",0xD1C0,"rorl",0xD1C8,
259                                 "rcll",0xD1D0,"rcrl",0xD1D8,
260                                 "shll",0xD1E0,"shrl",0xD1E8,
261                                 "sarl",0xD1F8);
262         
263         if ((defined($special{$name})) && defined($regs{$p1}) && ($p2 == 1))
264                 {
265                 $op=$special{$name}|$reg_val{$p1};
266                 $tmp1=sprintf(".byte %d\n",($op>>8)&0xff);
267                 $tmp2=sprintf(".byte %d\t",$op     &0xff);
268                 push(@out,$tmp1);
269                 push(@out,$tmp2);
270
271                 $p2=&conv($p2);
272                 $p1=&conv($p1);
273                 &main'comment("$name $p2 $p1");
274                 return;
275                 }
276
277         push(@out,"\t$name\t");
278         $t=&conv($p2).",";
279         $l=length($t);
280         push(@out,$t);
281         $ll=4-($l+9)/8;
282         $tmp1=sprintf("\t" x $ll);
283         push(@out,$tmp1);
284         push(@out,&conv($p1)."\n");
285         }
286
287 sub out1
288         {
289         local($name,$p1)=@_;
290         local($l,$t);
291         local(%special)=("bswapl",0x0FC8);
292
293         if ((defined($special{$name})) && defined($regs{$p1}))
294                 {
295                 $op=$special{$name}|$reg_val{$p1};
296                 $tmp1=sprintf(".byte %d\n",($op>>8)&0xff);
297                 $tmp2=sprintf(".byte %d\t",$op     &0xff);
298                 push(@out,$tmp1);
299                 push(@out,$tmp2);
300
301                 $p2=&conv($p2);
302                 $p1=&conv($p1);
303                 &main'comment("$name $p2 $p1");
304                 return;
305                 }
306
307         push(@out,"\t$name\t".&conv($p1)."\n");
308         }
309
310 sub out1p
311         {
312         local($name,$p1)=@_;
313         local($l,$t);
314
315         push(@out,"\t$name\t*".&conv($p1)."\n");
316         }
317
318 sub out0
319         {
320         push(@out,"\t$_[0]\n");
321         }
322
323 sub conv
324         {
325         local($p)=@_;
326
327 #       $p =~ s/0x([0-9A-Fa-f]+)/0$1h/;
328
329         $p=$regs{$p} if (defined($regs{$p}));
330
331         $p =~ s/^(-{0,1}[0-9A-Fa-f]+)$/\$$1/;
332         $p =~ s/^(0x[0-9A-Fa-f]+)$/\$$1/;
333         return $p;
334         }
335
336 sub main'file
337         {
338         local($file)=@_;
339
340         local($tmp)=<<"EOF";
341         .file   "$file.s"
342 EOF
343         push(@out,$tmp);
344         }
345
346 sub main'function_begin
347         {
348         local($func)=@_;
349
350         &main'external_label($func);
351         $func=$under.$func;
352
353         local($tmp)=<<"EOF";
354 .text
355 .globl  $func
356 EOF
357         push(@out,$tmp);
358         if ($main'cpp)
359                 { $tmp=push(@out,"TYPE($func,\@function)\n"); }
360         elsif ($main'coff)
361                 { $tmp=push(@out,".def\t$func;\t.scl\t2;\t.type\t32;\t.endef\n"); }
362         elsif ($main'aout and !$main'pic)
363                 { }
364         else    { $tmp=push(@out,".type\t$func,\@function\n"); }
365         push(@out,".align\t$align\n");
366         push(@out,"$func:\n");
367         $tmp=<<"EOF";
368         pushl   %ebp
369         pushl   %ebx
370         pushl   %esi
371         pushl   %edi
372
373 EOF
374         push(@out,$tmp);
375         $stack=20;
376         }
377
378 sub main'function_begin_B
379         {
380         local($func,$extra)=@_;
381
382         &main'external_label($func);
383         $func=$under.$func;
384
385         local($tmp)=<<"EOF";
386 .text
387 .globl  $func
388 EOF
389         push(@out,$tmp);
390         if ($main'cpp)
391                 { push(@out,"TYPE($func,\@function)\n"); }
392         elsif ($main'coff)
393                 { $tmp=push(@out,".def\t$func;\t.scl\t2;\t.type\t32;\t.endef\n"); }
394         elsif ($main'aout and !$main'pic)
395                 { }
396         else    { push(@out,".type      $func,\@function\n"); }
397         push(@out,".align\t$align\n");
398         push(@out,"$func:\n");
399         $stack=4;
400         }
401
402 sub main'function_end
403         {
404         local($func)=@_;
405
406         $func=$under.$func;
407
408         local($tmp)=<<"EOF";
409         popl    %edi
410         popl    %esi
411         popl    %ebx
412         popl    %ebp
413         ret
414 ${dot}L_${func}_end:
415 EOF
416         push(@out,$tmp);
417
418         if ($main'cpp)
419                 { push(@out,"SIZE($func,${dot}L_${func}_end-$func)\n"); }
420         elsif ($main'coff or $main'aout)
421                 { }
422         else    { push(@out,".size\t$func,${dot}L_${func}_end-$func\n"); }
423         push(@out,".ident       \"$func\"\n");
424         $stack=0;
425         %label=();
426         }
427
428 sub main'function_end_A
429         {
430         local($func)=@_;
431
432         local($tmp)=<<"EOF";
433         popl    %edi
434         popl    %esi
435         popl    %ebx
436         popl    %ebp
437         ret
438 EOF
439         push(@out,$tmp);
440         }
441
442 sub main'function_end_B
443         {
444         local($func)=@_;
445
446         $func=$under.$func;
447
448         push(@out,"${dot}L_${func}_end:\n");
449         if ($main'cpp)
450                 { push(@out,"SIZE($func,${dot}L_${func}_end-$func)\n"); }
451         elsif ($main'coff or $main'aout)
452                 { }
453         else    { push(@out,".size\t$func,${dot}L_${func}_end-$func\n"); }
454         push(@out,".ident       \"$func\"\n");
455         $stack=0;
456         %label=();
457         }
458
459 sub main'wparam
460         {
461         local($num)=@_;
462
463         return(&main'DWP($stack+$num*4,"esp","",0));
464         }
465
466 sub main'stack_push
467         {
468         local($num)=@_;
469         $stack+=$num*4;
470         &main'sub("esp",$num*4);
471         }
472
473 sub main'stack_pop
474         {
475         local($num)=@_;
476         $stack-=$num*4;
477         &main'add("esp",$num*4);
478         }
479
480 sub main'swtmp
481         {
482         return(&main'DWP($_[0]*4,"esp","",0));
483         }
484
485 # Should use swtmp, which is above esp.  Linix can trash the stack above esp
486 #sub main'wtmp
487 #       {
488 #       local($num)=@_;
489 #
490 #       return(&main'DWP(-($num+1)*4,"esp","",0));
491 #       }
492
493 sub main'comment
494         {
495         if (!defined($com_start) or $main'elf)
496                 {       # Regarding $main'elf above...
497                         # GNU and SVR4 as'es use different comment delimiters,
498                 push(@out,"\n");        # so we just skip ELF comments...
499                 return;
500                 }
501         foreach (@_)
502                 {
503                 if (/^\s*$/)
504                         { push(@out,"\n"); }
505                 else
506                         { push(@out,"\t$com_start $_ $com_end\n"); }
507                 }
508         }
509
510 sub main'public_label
511         {
512         $label{$_[0]}="${under}${_[0]}" if (!defined($label{$_[0]}));
513         push(@out,".globl\t$label{$_[0]}\n");
514         }
515
516 sub main'label
517         {
518         if (!defined($label{$_[0]}))
519                 {
520                 $label{$_[0]}="${dot}${label}${_[0]}";
521                 $label++;
522                 }
523         return($label{$_[0]});
524         }
525
526 sub main'set_label
527         {
528         if (!defined($label{$_[0]}))
529                 {
530                 $label{$_[0]}="${dot}${label}${_[0]}";
531                 $label++;
532                 }
533         if ($_[1]!=0)
534                 {
535                 if ($_[1]>1)    { main'align($_[1]);            }
536                 else            { push(@out,".align $align\n"); }
537                 }
538         push(@out,"$label{$_[0]}:\n");
539         }
540
541 sub main'file_end
542         {
543         # try to detect if SSE2 or MMX extensions were used on ELF platform...
544         if ($main'elf && grep {/%[x]*mm[0-7]/i} @out) {
545                 local($tmp);
546
547                 push (@out,"\n.section\t.bss\n");
548                 push (@out,".comm\t${under}OPENSSL_ia32cap_P,4,4\n");
549
550                 push (@out,".section\t.init\n");
551                 # One can argue that it's wasteful to craft every
552                 # SSE/MMX module with this snippet... Well, it's 72
553                 # bytes long and for the moment we have two modules.
554                 # Let's argue when we have 7 modules or so...
555                 #
556                 # $1<<10 sets a reserved bit to signal that variable
557                 # was initialized already...
558                 &main'picmeup("edx","OPENSSL_ia32cap_P");
559                 $tmp=<<___;
560                 cmpl    \$0,(%edx)
561                 jne     1f
562                 movl    \$1<<10,(%edx)
563                 pushf
564                 popl    %eax
565                 movl    %eax,%ecx
566                 xorl    \$1<<21,%eax
567                 pushl   %eax
568                 popf
569                 pushf
570                 popl    %eax
571                 xorl    %ecx,%eax
572                 btl     \$21,%eax
573                 jnc     1f
574                 pushl   %edi
575                 pushl   %ebx
576                 movl    %edx,%edi
577                 movl    \$1,%eax
578                 .byte   0x0f,0xa2
579                 orl     \$1<<10,%edx
580                 movl    %edx,0(%edi)
581                 popl    %ebx
582                 popl    %edi
583                 jmp     1f
584         .align  $align
585         1:
586 ___
587                 push (@out,$tmp);
588         }
589
590         if ($const ne "")
591                 {
592                 push(@out,".section .rodata\n");
593                 push(@out,$const);
594                 $const="";
595                 }
596         }
597
598 sub main'data_byte
599         {
600         push(@out,"\t.byte\t".join(',',@_)."\n");
601         }
602
603 sub main'data_word
604         {
605         push(@out,"\t.long\t".join(',',@_)."\n");
606         }
607
608 sub main'align
609         {
610         my $val=$_[0],$p2,$i;
611         if ($main'aout) {
612                 for ($p2=0;$val!=0;$val>>=1) { $p2++; }
613                 $val=$p2-1;
614                 $val.=",0x90";
615         }
616         push(@out,".align\t$val\n");
617         }
618
619 # debug output functions: puts, putx, printf
620
621 sub main'puts
622         {
623         &pushvars();
624         &main'push('$Lstring' . ++$constl);
625         &main'call('puts');
626         $stack-=4;
627         &main'add("esp",4);
628         &popvars();
629
630         $const .= "Lstring$constl:\n\t.string \"@_[0]\"\n";
631         }
632
633 sub main'putx
634         {
635         &pushvars();
636         &main'push($_[0]);
637         &main'push('$Lstring' . ++$constl);
638         &main'call('printf');
639         &main'add("esp",8);
640         $stack-=8;
641         &popvars();
642
643         $const .= "Lstring$constl:\n\t.string \"\%X\"\n";
644         }
645
646 sub main'printf
647         {
648         $ostack = $stack;
649         &pushvars();
650         for ($i = @_ - 1; $i >= 0; $i--)
651                 {
652                 if ($i == 0) # change this to support %s format strings
653                         {
654                         &main'push('$Lstring' . ++$constl);
655                         $const .= "Lstring$constl:\n\t.string \"@_[$i]\"\n";
656                         }
657                 else
658                         {
659                         if ($_[$i] =~ /([0-9]*)\(%esp\)/)
660                                 {
661                                 &main'push(($1 + $stack - $ostack) . '(%esp)');
662                                 }
663                         else
664                                 {
665                                 &main'push($_[$i]);
666                                 }
667                         }
668                 }
669         &main'call('printf');
670         $stack-=4*@_;
671         &main'add("esp",4*@_);
672         &popvars();
673         }
674
675 sub pushvars
676         {
677         &main'pushf();
678         &main'push("edx");
679         &main'push("ecx");
680         &main'push("eax");
681         }
682
683 sub popvars
684         {
685         &main'pop("eax");
686         &main'pop("ecx");
687         &main'pop("edx");
688         &main'popf();
689         }
690
691 sub main'picmeup
692         {
693         local($dst,$sym)=@_;
694         if ($main'cpp)
695                 {
696                 local($tmp)=<<___;
697 #if (defined(ELF) || defined(SOL)) && defined(PIC)
698         call    1f
699 1:      popl    $regs{$dst}
700         addl    \$_GLOBAL_OFFSET_TABLE_+[.-1b],$regs{$dst}
701         movl    $sym\@GOT($regs{$dst}),$regs{$dst}
702 #else
703         leal    $sym,$regs{$dst}
704 #endif
705 ___
706                 push(@out,$tmp);
707                 }
708         elsif ($main'pic && ($main'elf || $main'aout))
709                 {
710                 &main'call(&main'label("PIC_me_up"));
711                 &main'set_label("PIC_me_up");
712                 &main'blindpop($dst);
713                 &main'add($dst,"\$${under}_GLOBAL_OFFSET_TABLE_+[.-".
714                                 &main'label("PIC_me_up") . "]");
715                 &main'mov($dst,&main'DWP($under.$sym."\@GOT",$dst));
716                 }
717         else
718                 {
719                 &main'lea($dst,&main'DWP($sym));
720                 }
721         }
722
723 sub main'blindpop { &out1("popl",@_); }
724
725 sub main'initseg
726         {
727         local($f)=@_;
728         local($tmp);
729         if ($main'elf)
730                 {
731                 $tmp=<<___;
732 .section        .init
733         call    $under$f
734         jmp     .Linitalign
735 .align  $align
736 .Linitalign:
737 ___
738                 }
739         elsif ($main'coff)
740                 {
741                 $tmp=<<___;     # applies to both Cygwin and Mingw
742 .section        .ctors
743 .long   $under$f
744 ___
745                 }
746         elsif ($main'aout)
747                 {
748                 local($ctor)="${under}_GLOBAL_\$I\$$f";
749                 $tmp=".text\n";
750                 $tmp.=".type    $ctor,\@function\n" if ($main'pic);
751                 $tmp.=<<___;    # OpenBSD way...
752 .globl  $ctor
753 .align  2
754 $ctor:
755         jmp     $under$f
756 ___
757                 }
758         push(@out,$tmp) if ($tmp);
759         }
760
761 1;