3 # Ascetic x86_64 AT&T to MASM/NASM assembler translator by <appro>.
5 # Why AT&T to MASM and not vice versa? Several reasons. Because AT&T
6 # format is way easier to parse. Because it's simpler to "gear" from
7 # Unix ABI to Windows one [see cross-reference "card" at the end of
8 # file]. Because Linux targets were available first...
10 # In addition the script also "distills" code suitable for GNU
11 # assembler, so that it can be compiled with more rigid assemblers,
12 # such as Solaris /usr/ccs/bin/as.
14 # This translator is not designed to convert *arbitrary* assembler
15 # code from AT&T format to MASM one. It's designed to convert just
16 # enough to provide for dual-ABI OpenSSL modules development...
17 # There *are* limitations and you might have to modify your assembler
18 # code or this script to achieve the desired result...
20 # Currently recognized limitations:
22 # - can't use multiple ops per line;
24 # Dual-ABI styling rules.
26 # 1. Adhere to Unix register and stack layout [see cross-reference
27 # ABI "card" at the end for explanation].
28 # 2. Forget about "red zone," stick to more traditional blended
29 # stack frame allocation. If volatile storage is actually required
30 # that is. If not, just leave the stack as is.
31 # 3. Functions tagged with ".type name,@function" get crafted with
32 # unified Win64 prologue and epilogue automatically. If you want
33 # to take care of ABI differences yourself, tag functions as
34 # ".type name,@abi-omnipotent" instead.
35 # 4. To optimize the Win64 prologue you can specify number of input
36 # arguments as ".type name,@function,N." Keep in mind that if N is
37 # larger than 6, then you *have to* write "abi-omnipotent" code,
38 # because >6 cases can't be addressed with unified prologue.
39 # 5. Name local labels as .L*, do *not* use dynamic labels such as 1:
40 # (sorry about latter).
41 # 6. Don't use [or hand-code with .byte] "rep ret." "ret" mnemonic is
42 # required to identify the spots, where to inject Win64 epilogue!
43 # But on the pros, it's then prefixed with rep automatically:-)
44 # 7. Stick to explicit ip-relative addressing. If you have to use
45 # GOTPCREL addressing, stick to mov symbol@GOTPCREL(%rip),%r??.
46 # Both are recognized and translated to proper Win64 addressing
47 # modes. To support legacy code a synthetic directive, .picmeup,
48 # is implemented. It puts address of the *next* instruction into
49 # target register, e.g.:
52 # lea .Label-.(%rax),%rax
54 # 8. In order to provide for structured exception handling unified
55 # Win64 prologue copies %rsp value to %rax. For further details
56 # see SEH paragraph at the end.
57 # 9. .init segment is allowed to contain calls to functions only.
61 if ($flavour =~ /\./) { $output = $flavour; undef $flavour; }
63 { my ($stddev,$stdino,@junk)=stat(STDOUT);
64 my ($outdev,$outino,@junk)=stat($output);
66 open STDOUT,">$output" || die "can't open $output: $!"
67 if ($stddev!=$outdev || $stdino!=$outino);
70 my $gas=1; $gas=0 if ($output =~ /\.asm$/);
71 my $elf=1; $elf=0 if (!$gas);
76 my $masmref=8 + 50727*2**-32; # 8.00.50727 shipped with VS2005
83 if ($flavour eq "mingw64") { $gas=1; $elf=0; $win64=1; $prefix="_"; }
84 elsif ($flavour eq "macosx") { $gas=1; $elf=0; $prefix="_"; $decor="L\$"; }
85 elsif ($flavour eq "masm") { $gas=0; $elf=0; $masm=$masmref; $win64=1; $decor="\$L\$"; }
86 elsif ($flavour eq "nasm") { $gas=0; $elf=0; $nasm=$nasmref; $win64=1; $decor="\$L\$"; $PTR=""; }
88 { if ($ENV{ASM} =~ m/nasm/ && `nasm -v` =~ m/version ([0-9]+)\.([0-9]+)/i)
89 { $nasm = $1 + $2*0.01; $PTR=""; }
90 elsif (`ml64 2>&1` =~ m/Version ([0-9]+)\.([0-9]+)(\.([0-9]+))?/)
91 { $masm = $1 + $2*2**-16 + $4*2**-32; }
92 die "no assembler found on %PATH" if (!($nasm || $masm));
102 { package opcode; # pick up opcodes
104 my $self = shift; # single instance in enough...
108 if ($line =~ /^([a-z][a-z0-9]*)/i) {
111 $line = substr($line,@+[0]); $line =~ s/^\s+//;
114 if ($self->{op} =~ /^(movz)b.*/) { # movz is pain...
117 } elsif ($self->{op} =~ /call|jmp/) {
119 } elsif ($self->{op} =~ /([a-z]{3,})([qlwb])$/) {
129 $self->{sz} = $sz if (defined($sz) && !defined($self->{sz}));
135 if ($self->{op} eq "movz") { # movz is pain...
136 sprintf "%s%s%s",$self->{op},$self->{sz},shift;
137 } elsif ($self->{op} =~ /^set/) {
139 } elsif ($self->{op} eq "ret") {
141 if ($win64 && $current_function->{abi} eq "svr4") {
142 $epilogue = "movq 8(%rsp),%rdi\n\t" .
143 "movq 16(%rsp),%rsi\n\t";
145 $epilogue . ".byte 0xf3,0xc3";
146 } elsif ($self->{op} eq "call" && !$elf && $current_segment eq ".init") {
147 ".p2align\t3\n\t.quad";
149 "$self->{op}$self->{sz}";
152 $self->{op} =~ s/^movz/movzx/;
153 if ($self->{op} eq "ret") {
155 if ($win64 && $current_function->{abi} eq "svr4") {
156 $self->{op} = "mov rdi,QWORD${PTR}[8+rsp]\t;WIN64 epilogue\n\t".
157 "mov rsi,QWORD${PTR}[16+rsp]\n\t";
159 $self->{op} .= "DB\t0F3h,0C3h\t\t;repret";
160 } elsif ($self->{op} =~ /^(pop|push)f/) {
161 $self->{op} .= $self->{sz};
162 } elsif ($self->{op} eq "call" && $current_segment eq ".CRT\$XCU") {
163 $self->{op} = "ALIGN\t8\n\tDQ";
171 $self->{op}=$op if (defined($op));
175 { package const; # pick up constants, which start with $
177 my $self = shift; # single instance in enough...
181 if ($line =~ /^\$([^,]+)/) {
184 $line = substr($line,@+[0]); $line =~ s/^\s+//;
192 # Solaris /usr/ccs/bin/as can't handle multiplications
194 $self->{value} =~ s/(?<![0-9a-f])(0[x0-9a-f]+)/oct($1)/egi;
195 $self->{value} =~ s/([0-9]+\s*[\*\/\%]\s*[0-9]+)/eval($1)/eg;
196 sprintf "\$%s",$self->{value};
198 $self->{value} =~ s/(0b[0-1]+)/oct($1)/eig;
199 $self->{value} =~ s/0x([0-9a-f]+)/0$1h/ig if ($masm);
200 sprintf "%s",$self->{value};
204 { package ea; # pick up effective addresses: expr(%reg,%reg,scale)
206 my $self = shift; # single instance in enough...
210 # optional * ---vvv--- appears in indirect jmp/call
211 if ($line =~ /^(\*?)([^\(,]*)\(([%\w,]+)\)/) {
212 $self->{asterisk} = $1;
214 ($self->{base},$self->{index},$self->{scale})=split(/,/,$3);
215 $self->{scale} = 1 if (!defined($self->{scale}));
217 $line = substr($line,@+[0]); $line =~ s/^\s+//;
219 if ($win64 && $self->{label} =~ s/\@GOTPCREL//) {
220 die if (opcode->mnemonic() ne "mov");
221 opcode->mnemonic("lea");
223 $self->{base} =~ s/^%//;
224 $self->{index} =~ s/^%// if (defined($self->{index}));
233 $self->{label} =~ s/([_a-z][_a-z0-9]*)/$globals{$1} or $1/gei;
234 $self->{label} =~ s/\.L/$decor/g;
236 # Silently convert all EAs to 64-bit. This is required for
237 # elder GNU assembler and results in more compact code,
238 # *but* most importantly AES module depends on this feature!
239 $self->{index} =~ s/^[er](.?[0-9xpi])[d]?$/r\1/;
240 $self->{base} =~ s/^[er](.?[0-9xpi])[d]?$/r\1/;
243 # Solaris /usr/ccs/bin/as can't handle multiplications
245 $self->{label} =~ s/(?<![0-9a-f])(0[x0-9a-f]+)/oct($1)/egi;
246 $self->{label} =~ s/([0-9]+\s*[\*\/\%]\s*[0-9]+)/eval($1)/eg;
247 $self->{label} =~ s/^___imp_/__imp__/ if ($flavour eq "mingw64");
249 if (defined($self->{index})) {
250 sprintf "%s%s(%%%s,%%%s,%d)",$self->{asterisk},
251 $self->{label},$self->{base},
252 $self->{index},$self->{scale};
254 sprintf "%s%s(%%%s)", $self->{asterisk},$self->{label},$self->{base};
257 %szmap = ( b=>"BYTE$PTR", w=>"WORD$PTR", l=>"DWORD$PTR", q=>"QWORD$PTR" );
259 $self->{label} =~ s/\./\$/g;
260 $self->{label} =~ s/0x([0-9a-f]+)/0$1h/ig;
261 $self->{label} = "($self->{label})" if ($self->{label} =~ /[\*\+\-\/]/);
262 $sz="q" if ($self->{asterisk});
264 if (defined($self->{index})) {
265 sprintf "%s[%s%s*%d+%s]",$szmap{$sz},
266 $self->{label}?"$self->{label}+":"",
267 $self->{index},$self->{scale},
269 } elsif ($self->{base} eq "rip") {
270 sprintf "%s[%s]",$szmap{$sz},$self->{label};
272 sprintf "%s[%s%s]",$szmap{$sz},
273 $self->{label}?"$self->{label}+":"",
279 { package register; # pick up registers, which start with %.
281 my $class = shift; # muliple instances...
286 # optional * ---vvv--- appears in indirect jmp/call
287 if ($line =~ /^(\*?)%(\w+)/) {
289 $self->{asterisk} = $1;
292 $line = substr($line,@+[0]); $line =~ s/^\s+//;
300 if ($self->{value} =~ /^r[\d]+b$/i) { $ret="b"; }
301 elsif ($self->{value} =~ /^r[\d]+w$/i) { $ret="w"; }
302 elsif ($self->{value} =~ /^r[\d]+d$/i) { $ret="l"; }
303 elsif ($self->{value} =~ /^r[\w]+$/i) { $ret="q"; }
304 elsif ($self->{value} =~ /^[a-d][hl]$/i){ $ret="b"; }
305 elsif ($self->{value} =~ /^[\w]{2}l$/i) { $ret="b"; }
306 elsif ($self->{value} =~ /^[\w]{2}$/i) { $ret="w"; }
307 elsif ($self->{value} =~ /^e[a-z]{2}$/i){ $ret="l"; }
313 if ($gas) { sprintf "%s%%%s",$self->{asterisk},$self->{value}; }
314 else { $self->{value}; }
317 { package label; # pick up labels, which end with :
319 my $self = shift; # single instance is enough...
323 if ($line =~ /(^[\.\w]+)\:/) {
326 $line = substr($line,@+[0]); $line =~ s/^\s+//;
328 $self->{value} =~ s/^\.L/$decor/;
336 my $func = ($globals{$self->{value}} or $self->{value}) . ":";
338 $current_function->{name} eq $self->{value} &&
339 $current_function->{abi} eq "svr4") {
341 $func .= " movq %rdi,8(%rsp)\n";
342 $func .= " movq %rsi,16(%rsp)\n";
343 $func .= " movq %rsp,%rax\n";
344 $func .= "${decor}SEH_begin_$current_function->{name}:\n";
345 my $narg = $current_function->{narg};
346 $narg=6 if (!defined($narg));
347 $func .= " movq %rcx,%rdi\n" if ($narg>0);
348 $func .= " movq %rdx,%rsi\n" if ($narg>1);
349 $func .= " movq %r8,%rdx\n" if ($narg>2);
350 $func .= " movq %r9,%rcx\n" if ($narg>3);
351 $func .= " movq 40(%rsp),%r8\n" if ($narg>4);
352 $func .= " movq 48(%rsp),%r9\n" if ($narg>5);
355 } elsif ($self->{value} ne "$current_function->{name}") {
356 $self->{value} .= ":" if ($masm && $ret!~m/^\$/);
357 $self->{value} . ":";
358 } elsif ($win64 && $current_function->{abi} eq "svr4") {
359 my $func = "$current_function->{name}" .
360 ($nasm ? ":" : "\tPROC $current_function->{scope}") .
362 $func .= " mov QWORD${PTR}[8+rsp],rdi\t;WIN64 prologue\n";
363 $func .= " mov QWORD${PTR}[16+rsp],rsi\n";
364 $func .= " mov rax,rsp\n";
365 $func .= "${decor}SEH_begin_$current_function->{name}:";
366 $func .= ":" if ($masm);
368 my $narg = $current_function->{narg};
369 $narg=6 if (!defined($narg));
370 $func .= " mov rdi,rcx\n" if ($narg>0);
371 $func .= " mov rsi,rdx\n" if ($narg>1);
372 $func .= " mov rdx,r8\n" if ($narg>2);
373 $func .= " mov rcx,r9\n" if ($narg>3);
374 $func .= " mov r8,QWORD${PTR}[40+rsp]\n" if ($narg>4);
375 $func .= " mov r9,QWORD${PTR}[48+rsp]\n" if ($narg>5);
378 "$current_function->{name}".
379 ($nasm ? ":" : "\tPROC $current_function->{scope}");
383 { package expr; # pick up expressioins
385 my $self = shift; # single instance is enough...
389 if ($line =~ /(^[^,]+)/) {
392 $line = substr($line,@+[0]); $line =~ s/^\s+//;
394 $self->{value} =~ s/\@PLT// if (!$elf);
395 $self->{value} =~ s/([_a-z][_a-z0-9]*)/$globals{$1} or $1/gei;
396 $self->{value} =~ s/\.L/$decor/g;
402 if ($nasm && opcode->mnemonic()=~m/^j/) {
403 "NEAR ".$self->{value};
409 { package directive; # pick up directives, which start with .
411 my $self = shift; # single instance is enough...
415 my %opcode = # lea 2f-1f(%rip),%dst; 1: nop; 2:
416 ( "%rax"=>0x01058d48, "%rcx"=>0x010d8d48,
417 "%rdx"=>0x01158d48, "%rbx"=>0x011d8d48,
418 "%rsp"=>0x01258d48, "%rbp"=>0x012d8d48,
419 "%rsi"=>0x01358d48, "%rdi"=>0x013d8d48,
420 "%r8" =>0x01058d4c, "%r9" =>0x010d8d4c,
421 "%r10"=>0x01158d4c, "%r11"=>0x011d8d4c,
422 "%r12"=>0x01258d4c, "%r13"=>0x012d8d4c,
423 "%r14"=>0x01358d4c, "%r15"=>0x013d8d4c );
425 if ($line =~ /^\s*(\.\w+)/) {
428 undef $self->{value};
429 $line = substr($line,@+[0]); $line =~ s/^\s+//;
432 /\.picmeup/ && do { if ($line =~ /(%r[\w]+)/i) {
434 $line=sprintf "0x%x,0x90000000",$opcode{$1};
438 /\.global|\.globl|\.extern/
439 && do { $globals{$line} = $prefix . $line;
440 $line = $globals{$line} if ($prefix);
443 /\.type/ && do { ($sym,$type,$narg) = split(',',$line);
444 if ($type eq "\@function") {
445 undef $current_function;
446 $current_function->{name} = $sym;
447 $current_function->{abi} = "svr4";
448 $current_function->{narg} = $narg;
449 $current_function->{scope} = defined($globals{$sym})?"PUBLIC":"PRIVATE";
450 } elsif ($type eq "\@abi-omnipotent") {
451 undef $current_function;
452 $current_function->{name} = $sym;
453 $current_function->{scope} = defined($globals{$sym})?"PUBLIC":"PRIVATE";
455 $line =~ s/\@abi\-omnipotent/\@function/;
456 $line =~ s/\@function.*/\@function/;
459 /\.asciz/ && do { if ($line =~ /^"(.*)"$/) {
461 $line = join(",",unpack("C*",$1),0);
465 /\.rva|\.long|\.quad/
466 && do { $line =~ s/([_a-z][_a-z0-9]*)/$globals{$1} or $1/gei;
467 $line =~ s/\.L/$decor/g;
473 $self->{value} = $dir . "\t" . $line;
475 if ($dir =~ /\.extern/) {
476 $self->{value} = ""; # swallow extern
477 } elsif (!$elf && $dir =~ /\.type/) {
479 $self->{value} = ".def\t" . ($globals{$1} or $1) . ";\t" .
480 (defined($globals{$1})?".scl 2;":".scl 3;") .
481 "\t.type 32;\t.endef"
482 if ($win64 && $line =~ /([^,]+),\@function/);
483 } elsif (!$elf && $dir =~ /\.size/) {
485 if (defined($current_function)) {
486 $self->{value} .= "${decor}SEH_end_$current_function->{name}:"
487 if ($win64 && $current_function->{abi} eq "svr4");
488 undef $current_function;
490 } elsif (!$elf && $dir =~ /\.align/) {
491 $self->{value} = ".p2align\t" . (log($line)/log(2));
492 } elsif ($dir eq ".section") {
493 $current_segment=$line;
494 if (!$elf && $current_segment eq ".init") {
495 if ($flavour eq "macosx") { $self->{value} = ".mod_init_func"; }
496 elsif ($flavour eq "mingw64") { $self->{value} = ".section\t.ctors"; }
498 } elsif ($dir =~ /\.(text|data)/) {
499 $current_segment=".$1";
505 # non-gas case or nasm/masm
507 /\.text/ && do { my $v=undef;
509 $v="section .text code align=64\n";
511 $v="$current_segment\tENDS\n" if ($current_segment);
512 $current_segment = ".text\$";
513 $v.="$current_segment\tSEGMENT ";
514 $v.=$masm>=$masmref ? "ALIGN(64)" : "PAGE";
520 /\.data/ && do { my $v=undef;
522 $v="section .data data align=8\n";
524 $v="$current_segment\tENDS\n" if ($current_segment);
525 $current_segment = "_DATA";
526 $v.="$current_segment\tSEGMENT";
531 /\.section/ && do { my $v=undef;
532 $line =~ s/([^,]*).*/$1/;
533 $line = ".CRT\$XCU" if ($line eq ".init");
536 if ($line=~/\.([px])data/) {
538 $v.=$1 eq "p"? 4 : 8;
541 $v="$current_segment\tENDS\n" if ($current_segment);
542 $v.="$line\tSEGMENT";
543 if ($line=~/\.([px])data/) {
545 $v.=" ALIGN(".($1 eq "p" ? 4 : 8).")" if ($masm>=$masmref);
548 $current_segment = $line;
552 /\.extern/ && do { $self->{value} = "EXTERN\t".$line;
553 $self->{value} .= ":NEAR" if ($masm);
557 && do { $self->{value} = $masm?"PUBLIC":"global";
558 $self->{value} .= "\t".$line;
561 /\.size/ && do { if (defined($current_function)) {
562 undef $self->{value};
563 if ($current_function->{abi} eq "svr4") {
564 $self->{value}="${decor}SEH_end_$current_function->{name}:";
565 $self->{value}.=":\n" if($masm);
567 $self->{value}.="$current_function->{name}\tENDP" if($masm);
568 undef $current_function;
572 /\.align/ && do { $self->{value} = "ALIGN\t".$line; last; };
573 /\.(value|long|rva|quad)/
574 && do { my $sz = substr($1,0,1);
575 my @arr = split(',',$line);
576 my $last = pop(@arr);
577 my $conv = sub { my $var=shift;
578 $var=~s/(0b[0-1]+)/oct($1)/eig;
579 $var=~s/0x([0-9a-f]+)/0$1h/ig if ($masm);
580 if ($sz eq "D" && ($current_segment=~/.[px]data/ || $dir eq ".rva"))
581 { $var=~s/([_a-z\$\@][_a-z0-9\$\@]*)/$nasm?"$1 wrt ..imagebase":"imagerel $1"/egi; }
585 $sz =~ tr/bvlrq/BWDDQ/;
586 $self->{value} = "\tD$sz\t";
587 for (@arr) { $self->{value} .= &$conv($_).","; }
588 $self->{value} .= &$conv($last);
591 /\.byte/ && do { my @str=split(",",$line);
592 map(s/(0b[0-1]+)/oct($1)/eig,@str);
593 map(s/0x([0-9a-f]+)/0$1h/ig,@str) if ($masm);
595 $self->{value}.="DB\t"
596 .join(",",@str[0..15])."\n";
597 foreach (0..15) { shift @str; }
599 $self->{value}.="DB\t"
600 .join(",",@str) if (@str);
628 $line =~ s|[#!].*$||; # get rid of asm-style comments...
629 $line =~ s|/\*.*\*/||; # ... and C-style comments...
630 $line =~ s|^\s+||; # ... and skip white spaces in beginning
637 if ($label=label->re(\$line)) { print $label->out(); }
639 if (directive->re(\$line)) {
640 printf "%s",directive->out();
641 } elsif ($opcode=opcode->re(\$line)) { ARGUMENT: while (1) {
644 if ($arg=register->re(\$line)) { opcode->size($arg->size()); }
645 elsif ($arg=const->re(\$line)) { }
646 elsif ($arg=ea->re(\$line)) { }
647 elsif ($arg=expr->re(\$line)) { }
648 else { last ARGUMENT; }
652 last ARGUMENT if ($line !~ /^,/);
662 $insn = $opcode->out($#args>=1?$args[$#args]->size():$sz);
664 $insn = $opcode->out();
665 @args = reverse(@args);
666 undef $sz if ($nasm && $opcode->mnemonic() eq "lea");
668 printf "\t%s\t%s",$insn,join(",",map($_->out($sz),@args));
670 printf "\t%s",$opcode->out();
677 print "\n$current_segment\tENDS\n" if ($current_segment && $masm);
678 print "END\n" if ($masm);
682 \f#################################################
683 # Cross-reference x86_64 ABI "card"
703 # (*) volatile register
704 # (-) preserved by callee
705 # (#) Nth argument, volatile
707 # In Unix terms top of stack is argument transfer area for arguments
708 # which could not be accomodated in registers. Or in other words 7th
709 # [integer] argument resides at 8(%rsp) upon function entry point.
710 # 128 bytes above %rsp constitute a "red zone" which is not touched
711 # by signal handlers and can be used as temporal storage without
712 # allocating a frame.
714 # In Win64 terms N*8 bytes on top of stack is argument transfer area,
715 # which belongs to/can be overwritten by callee. N is the number of
716 # arguments passed to callee, *but* not less than 4! This means that
717 # upon function entry point 5th argument resides at 40(%rsp), as well
718 # as that 32 bytes from 8(%rsp) can always be used as temporal
719 # storage [without allocating a frame]. One can actually argue that
720 # one can assume a "red zone" above stack pointer under Win64 as well.
721 # Point is that at apparently no occasion Windows kernel would alter
722 # the area above user stack pointer in true asynchronous manner...
724 # All the above means that if assembler programmer adheres to Unix
725 # register and stack layout, but disregards the "red zone" existense,
726 # it's possible to use following prologue and epilogue to "gear" from
727 # Unix to Win64 ABI in leaf functions with not more than 6 arguments.
729 # omnipotent_function:
733 # movq %rcx,%rdi ; if 1st argument is actually present
734 # movq %rdx,%rsi ; if 2nd argument is actually ...
735 # movq %r8,%rdx ; if 3rd argument is ...
736 # movq %r9,%rcx ; if 4th argument ...
737 # movq 40(%rsp),%r8 ; if 5th ...
738 # movq 48(%rsp),%r9 ; if 6th ...
747 \f#################################################
748 # Win64 SEH, Structured Exception Handling.
750 # Unlike on Unix systems(*) lack of Win64 stack unwinding information
751 # has undesired side-effect at run-time: if an exception is raised in
752 # assembler subroutine such as those in question (basically we're
753 # referring to segmentation violations caused by malformed input
754 # parameters), the application is briskly terminated without invoking
755 # any exception handlers, most notably without generating memory dump
756 # or any user notification whatsoever. This poses a problem. It's
757 # possible to address it by registering custom language-specific
758 # handler that would restore processor context to the state at
759 # subroutine entry point and return "exception is not handled, keep
760 # unwinding" code. Writing such handler can be a challenge... But it's
761 # doable, though requires certain coding convention. Consider following
764 # .type function,@function
766 # movq %rsp,%rax # copy rsp to volatile register
767 # pushq %r15 # save non-volatile registers
771 # subq %rdi,%r11 # prepare [variable] stack frame
773 # movq %rax,0(%r11) # check for exceptions
774 # movq %r11,%rsp # allocate [variable] stack frame
775 # movq %rax,0(%rsp) # save original rsp value
778 # movq 0(%rsp),%rcx # pull original rsp value
779 # movq -24(%rcx),%rbp # restore non-volatile registers
780 # movq -16(%rcx),%rbx
782 # movq %rcx,%rsp # restore original rsp
784 # .size function,.-function
786 # The key is that up to magic_point copy of original rsp value remains
787 # in chosen volatile register and no non-volatile register, except for
788 # rsp, is modified. While past magic_point rsp remains constant till
789 # the very end of the function. In this case custom language-specific
790 # exception handler would look like this:
792 # EXCEPTION_DISPOSITION handler (EXCEPTION_RECORD *rec,ULONG64 frame,
793 # CONTEXT *context,DISPATCHER_CONTEXT *disp)
794 # { ULONG64 *rsp = (ULONG64 *)context->Rax;
795 # if (context->Rip >= magic_point)
796 # { rsp = ((ULONG64 **)context->Rsp)[0];
797 # context->Rbp = rsp[-3];
798 # context->Rbx = rsp[-2];
799 # context->R15 = rsp[-1];
801 # context->Rsp = (ULONG64)rsp;
802 # context->Rdi = rsp[1];
803 # context->Rsi = rsp[2];
805 # memcpy (disp->ContextRecord,context,sizeof(CONTEXT));
806 # RtlVirtualUnwind(UNW_FLAG_NHANDLER,disp->ImageBase,
807 # dips->ControlPc,disp->FunctionEntry,disp->ContextRecord,
808 # &disp->HandlerData,&disp->EstablisherFrame,NULL);
809 # return ExceptionContinueSearch;
812 # It's appropriate to implement this handler in assembler, directly in
813 # function's module. In order to do that one has to know members'
814 # offsets in CONTEXT and DISPATCHER_CONTEXT structures and some constant
815 # values. Here they are:
834 # sizeof(CONTEXT) 1232
835 # DISPATCHER_CONTEXT.ControlPc 0
836 # DISPATCHER_CONTEXT.ImageBase 8
837 # DISPATCHER_CONTEXT.FunctionEntry 16
838 # DISPATCHER_CONTEXT.EstablisherFrame 24
839 # DISPATCHER_CONTEXT.TargetIp 32
840 # DISPATCHER_CONTEXT.ContextRecord 40
841 # DISPATCHER_CONTEXT.LanguageHandler 48
842 # DISPATCHER_CONTEXT.HandlerData 56
843 # UNW_FLAG_NHANDLER 0
844 # ExceptionContinueSearch 1
846 # In order to tie the handler to the function one has to compose
847 # couple of structures: one for .xdata segment and one for .pdata.
849 # UNWIND_INFO structure for .xdata segment would be
851 # function_unwind_info:
855 # This structure designates exception handler for a function with
856 # zero-length prologue, no stack frame or frame register.
858 # To facilitate composing of .pdata structures, auto-generated "gear"
859 # prologue copies rsp value to rax and denotes next instruction with
860 # .LSEH_begin_{function_name} label. This essentially defines the SEH
861 # styling rule mentioned in the beginning. Position of this label is
862 # chosen in such manner that possible exceptions raised in the "gear"
863 # prologue would be accounted to caller and unwound from latter's frame.
864 # End of function is marked with respective .LSEH_end_{function_name}
865 # label. To summarize, .pdata segment would contain
867 # .rva .LSEH_begin_function
868 # .rva .LSEH_end_function
869 # .rva function_unwind_info
871 # Reference to functon_unwind_info from .xdata segment is the anchor.
872 # In case you wonder why references are 32-bit .rvas and not 64-bit
873 # .quads. References put into these two segments are required to be
874 # *relative* to the base address of the current binary module, a.k.a.
875 # image base. No Win64 module, be it .exe or .dll, can be larger than
876 # 2GB and thus such relative references can be and are accommodated in
879 # Having reviewed the example function code, one can argue that "movq
880 # %rsp,%rax" above is redundant. It is not! Keep in mind that on Unix
881 # rax would contain an undefined value. If this "offends" you, use
882 # another register and refrain from modifying rax till magic_point is
883 # reached, i.e. as if it was a non-volatile register. If more registers
884 # are required prior [variable] frame setup is completed, note that
885 # nobody says that you can have only one "magic point." You can
886 # "liberate" non-volatile registers by denoting last stack off-load
887 # instruction and reflecting it in finer grade unwind logic in handler.
888 # After all, isn't it why it's called *language-specific* handler...
890 # Attentive reader can notice that exceptions would be mishandled in
891 # auto-generated "gear" epilogue. Well, exception effectively can't
892 # occur there, because if memory area used by it was subject to
893 # segmentation violation, then it would be raised upon call to the
894 # function (and as already mentioned be accounted to caller, which is
895 # not a problem). If you're still not comfortable, then define tail
896 # "magic point" just prior ret instruction and have handler treat it...
898 # (*) Note that we're talking about run-time, not debug-time. Lack of
899 # unwind information makes debugging hard on both Windows and
900 # Unix. "Unlike" referes to the fact that on Unix signal handler
901 # will always be invoked, core dumped and appropriate exit code
902 # returned to parent (for user notification).