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/0x([0-9a-f]+)/0$1h/ig;
199 sprintf "%s",$self->{value};
203 { package ea; # pick up effective addresses: expr(%reg,%reg,scale)
205 my $self = shift; # single instance in enough...
209 # optional * ---vvv--- appears in indirect jmp/call
210 if ($line =~ /^(\*?)([^\(,]*)\(([%\w,]+)\)/) {
211 $self->{asterisk} = $1;
213 ($self->{base},$self->{index},$self->{scale})=split(/,/,$3);
214 $self->{scale} = 1 if (!defined($self->{scale}));
216 $line = substr($line,@+[0]); $line =~ s/^\s+//;
218 if ($win64 && $self->{label} =~ s/\@GOTPCREL//) {
219 die if (opcode->mnemonic() ne "mov");
220 opcode->mnemonic("lea");
222 $self->{base} =~ s/^%//;
223 $self->{index} =~ s/^%// if (defined($self->{index}));
232 $self->{label} =~ s/([_a-z][_a-z0-9]*)/$globals{$1} or $1/gei;
233 $self->{label} =~ s/\.L/$decor/g;
235 # Silently convert all EAs to 64-bit. This is required for
236 # elder GNU assembler and results in more compact code,
237 # *but* most importantly AES module depends on this feature!
238 $self->{index} =~ s/^[er](.?[0-9xpi])[d]?$/r\1/;
239 $self->{base} =~ s/^[er](.?[0-9xpi])[d]?$/r\1/;
242 # Solaris /usr/ccs/bin/as can't handle multiplications
244 $self->{label} =~ s/(?<![0-9a-f])(0[x0-9a-f]+)/oct($1)/egi;
245 $self->{label} =~ s/([0-9]+\s*[\*\/\%]\s*[0-9]+)/eval($1)/eg;
246 $self->{label} =~ s/^___imp_/__imp__/ if ($flavour eq "mingw64");
248 if (defined($self->{index})) {
249 sprintf "%s%s(%%%s,%%%s,%d)",$self->{asterisk},
250 $self->{label},$self->{base},
251 $self->{index},$self->{scale};
253 sprintf "%s%s(%%%s)", $self->{asterisk},$self->{label},$self->{base};
256 %szmap = ( b=>"BYTE$PTR", w=>"WORD$PTR", l=>"DWORD$PTR", q=>"QWORD$PTR" );
258 $self->{label} =~ s/\./\$/g;
259 $self->{label} =~ s/0x([0-9a-f]+)/0$1h/ig;
260 $self->{label} = "($self->{label})" if ($self->{label} =~ /[\*\+\-\/]/);
261 $sz="q" if ($self->{asterisk});
263 if (defined($self->{index})) {
264 sprintf "%s[%s%s*%d+%s]",$szmap{$sz},
265 $self->{label}?"$self->{label}+":"",
266 $self->{index},$self->{scale},
268 } elsif ($self->{base} eq "rip") {
269 sprintf "%s[%s]",$szmap{$sz},$self->{label};
271 sprintf "%s[%s%s]",$szmap{$sz},
272 $self->{label}?"$self->{label}+":"",
278 { package register; # pick up registers, which start with %.
280 my $class = shift; # muliple instances...
285 # optional * ---vvv--- appears in indirect jmp/call
286 if ($line =~ /^(\*?)%(\w+)/) {
288 $self->{asterisk} = $1;
291 $line = substr($line,@+[0]); $line =~ s/^\s+//;
299 if ($self->{value} =~ /^r[\d]+b$/i) { $ret="b"; }
300 elsif ($self->{value} =~ /^r[\d]+w$/i) { $ret="w"; }
301 elsif ($self->{value} =~ /^r[\d]+d$/i) { $ret="l"; }
302 elsif ($self->{value} =~ /^r[\w]+$/i) { $ret="q"; }
303 elsif ($self->{value} =~ /^[a-d][hl]$/i){ $ret="b"; }
304 elsif ($self->{value} =~ /^[\w]{2}l$/i) { $ret="b"; }
305 elsif ($self->{value} =~ /^[\w]{2}$/i) { $ret="w"; }
306 elsif ($self->{value} =~ /^e[a-z]{2}$/i){ $ret="l"; }
312 if ($gas) { sprintf "%s%%%s",$self->{asterisk},$self->{value}; }
313 else { $self->{value}; }
316 { package label; # pick up labels, which end with :
318 my $self = shift; # single instance is enough...
322 if ($line =~ /(^[\.\w]+)\:/) {
325 $line = substr($line,@+[0]); $line =~ s/^\s+//;
327 $self->{value} =~ s/^\.L/$decor/;
335 my $func = ($globals{$self->{value}} or $self->{value}) . ":";
337 $current_function->{name} eq $self->{value} &&
338 $current_function->{abi} eq "svr4") {
340 $func .= " movq %rdi,8(%rsp)\n";
341 $func .= " movq %rsi,16(%rsp)\n";
342 $func .= " movq %rsp,%rax\n";
343 $func .= "${decor}SEH_begin_$current_function->{name}:\n";
344 my $narg = $current_function->{narg};
345 $narg=6 if (!defined($narg));
346 $func .= " movq %rcx,%rdi\n" if ($narg>0);
347 $func .= " movq %rdx,%rsi\n" if ($narg>1);
348 $func .= " movq %r8,%rdx\n" if ($narg>2);
349 $func .= " movq %r9,%rcx\n" if ($narg>3);
350 $func .= " movq 40(%rsp),%r8\n" if ($narg>4);
351 $func .= " movq 48(%rsp),%r9\n" if ($narg>5);
354 } elsif ($self->{value} ne "$current_function->{name}") {
355 $self->{value} .= ":" if ($masm && $ret!~m/^\$/);
356 $self->{value} . ":";
357 } elsif ($win64 && $current_function->{abi} eq "svr4") {
358 my $func = "$current_function->{name}" .
359 ($nasm ? ":" : "\tPROC $current_function->{scope}") .
361 $func .= " mov QWORD${PTR}[8+rsp],rdi\t;WIN64 prologue\n";
362 $func .= " mov QWORD${PTR}[16+rsp],rsi\n";
363 $func .= " mov rax,rsp\n";
364 $func .= "${decor}SEH_begin_$current_function->{name}:";
365 $func .= ":" if ($masm);
367 my $narg = $current_function->{narg};
368 $narg=6 if (!defined($narg));
369 $func .= " mov rdi,rcx\n" if ($narg>0);
370 $func .= " mov rsi,rdx\n" if ($narg>1);
371 $func .= " mov rdx,r8\n" if ($narg>2);
372 $func .= " mov rcx,r9\n" if ($narg>3);
373 $func .= " mov r8,QWORD${PTR}[40+rsp]\n" if ($narg>4);
374 $func .= " mov r9,QWORD${PTR}[48+rsp]\n" if ($narg>5);
377 "$current_function->{name}".
378 ($nasm ? ":" : "\tPROC $current_function->{scope}");
382 { package expr; # pick up expressioins
384 my $self = shift; # single instance is enough...
388 if ($line =~ /(^[^,]+)/) {
391 $line = substr($line,@+[0]); $line =~ s/^\s+//;
393 $self->{value} =~ s/\@PLT// if (!$elf);
394 $self->{value} =~ s/([_a-z][_a-z0-9]*)/$globals{$1} or $1/gei;
395 $self->{value} =~ s/\.L/$decor/g;
401 if ($nasm && opcode->mnemonic()=~m/^j/) {
402 "NEAR ".$self->{value};
408 { package directive; # pick up directives, which start with .
410 my $self = shift; # single instance is enough...
414 my %opcode = # lea 2f-1f(%rip),%dst; 1: nop; 2:
415 ( "%rax"=>0x01058d48, "%rcx"=>0x010d8d48,
416 "%rdx"=>0x01158d48, "%rbx"=>0x011d8d48,
417 "%rsp"=>0x01258d48, "%rbp"=>0x012d8d48,
418 "%rsi"=>0x01358d48, "%rdi"=>0x013d8d48,
419 "%r8" =>0x01058d4c, "%r9" =>0x010d8d4c,
420 "%r10"=>0x01158d4c, "%r11"=>0x011d8d4c,
421 "%r12"=>0x01258d4c, "%r13"=>0x012d8d4c,
422 "%r14"=>0x01358d4c, "%r15"=>0x013d8d4c );
424 if ($line =~ /^\s*(\.\w+)/) {
427 undef $self->{value};
428 $line = substr($line,@+[0]); $line =~ s/^\s+//;
431 /\.picmeup/ && do { if ($line =~ /(%r[\w]+)/i) {
433 $line=sprintf "0x%x,0x90000000",$opcode{$1};
437 /\.global|\.globl|\.extern/
438 && do { $globals{$line} = $prefix . $line;
439 $line = $globals{$line} if ($prefix);
442 /\.type/ && do { ($sym,$type,$narg) = split(',',$line);
443 if ($type eq "\@function") {
444 undef $current_function;
445 $current_function->{name} = $sym;
446 $current_function->{abi} = "svr4";
447 $current_function->{narg} = $narg;
448 $current_function->{scope} = defined($globals{$sym})?"PUBLIC":"PRIVATE";
449 } elsif ($type eq "\@abi-omnipotent") {
450 undef $current_function;
451 $current_function->{name} = $sym;
452 $current_function->{scope} = defined($globals{$sym})?"PUBLIC":"PRIVATE";
454 $line =~ s/\@abi\-omnipotent/\@function/;
455 $line =~ s/\@function.*/\@function/;
458 /\.asciz/ && do { if ($line =~ /^"(.*)"$/) {
460 $line = join(",",unpack("C*",$1),0);
464 /\.rva|\.long|\.quad/
465 && do { $line =~ s/([_a-z][_a-z0-9]*)/$globals{$1} or $1/gei;
466 $line =~ s/\.L/$decor/g;
472 $self->{value} = $dir . "\t" . $line;
474 if ($dir =~ /\.extern/) {
475 $self->{value} = ""; # swallow extern
476 } elsif (!$elf && $dir =~ /\.type/) {
478 $self->{value} = ".def\t" . ($globals{$1} or $1) . ";\t" .
479 (defined($globals{$1})?".scl 2;":".scl 3;") .
480 "\t.type 32;\t.endef"
481 if ($win64 && $line =~ /([^,]+),\@function/);
482 } elsif (!$elf && $dir =~ /\.size/) {
484 if (defined($current_function)) {
485 $self->{value} .= "${decor}SEH_end_$current_function->{name}:"
486 if ($win64 && $current_function->{abi} eq "svr4");
487 undef $current_function;
489 } elsif (!$elf && $dir =~ /\.align/) {
490 $self->{value} = ".p2align\t" . (log($line)/log(2));
491 } elsif ($dir eq ".section") {
492 $current_segment=$line;
493 if (!$elf && $current_segment eq ".init") {
494 if ($flavour eq "macosx") { $self->{value} = ".mod_init_func"; }
495 elsif ($flavour eq "mingw64") { $self->{value} = ".section\t.ctors"; }
497 } elsif ($dir =~ /\.(text|data)/) {
498 $current_segment=".$1";
504 # non-gas case or nasm/masm
506 /\.text/ && do { my $v=undef;
508 $v="section .text code align=64\n";
510 $v="$current_segment\tENDS\n" if ($current_segment);
511 $current_segment = ".text\$";
512 $v.="$current_segment\tSEGMENT ";
513 $v.=$masm>=$masmref ? "ALIGN(64)" : "PAGE";
519 /\.data/ && do { my $v=undef;
521 $v="section .data data align=8\n";
523 $v="$current_segment\tENDS\n" if ($current_segment);
524 $current_segment = "_DATA";
525 $v.="$current_segment\tSEGMENT";
530 /\.section/ && do { my $v=undef;
531 $line =~ s/([^,]*).*/$1/;
532 $line = ".CRT\$XCU" if ($line eq ".init");
535 if ($line=~/\.([px])data/) {
537 $v.=$1 eq "p"? 4 : 8;
540 $v="$current_segment\tENDS\n" if ($current_segment);
541 $v.="$line\tSEGMENT";
542 if ($line=~/\.([px])data/) {
544 $v.=" ALIGN(".($1 eq "p" ? 4 : 8).")" if ($masm>=$masmref);
547 $current_segment = $line;
551 /\.extern/ && do { $self->{value} = "EXTERN\t".$line;
552 $self->{value} .= ":NEAR" if ($masm);
556 && do { $self->{value} = $masm?"PUBLIC":"global";
557 $self->{value} .= "\t".$line;
560 /\.size/ && do { if (defined($current_function)) {
561 undef $self->{value};
562 if ($current_function->{abi} eq "svr4") {
563 $self->{value}="${decor}SEH_end_$current_function->{name}:";
564 $self->{value}.=":\n" if($masm);
566 $self->{value}.="$current_function->{name}\tENDP" if($masm);
567 undef $current_function;
571 /\.align/ && do { $self->{value} = "ALIGN\t".$line; last; };
572 /\.(value|long|rva|quad)/
573 && do { my $sz = substr($1,0,1);
574 my @arr = split(',',$line);
575 my $last = pop(@arr);
576 my $conv = sub { my $var=shift;
577 $var=~s/0x([0-9a-f]+)/0$1h/ig;
578 if ($sz eq "D" && ($current_segment=~/.[px]data/ || $dir eq ".rva"))
579 { $var=~s/([_a-z\$\@][_a-z0-9\$\@]*)/$nasm?"$1 wrt ..imagebase":"imagerel $1"/egi; }
583 $sz =~ tr/bvlrq/BWDDQ/;
584 $self->{value} = "\tD$sz\t";
585 for (@arr) { $self->{value} .= &$conv($_).","; }
586 $self->{value} .= &$conv($last);
589 /\.byte/ && do { my @str=split(",",$line);
591 $self->{value}.="DB\t"
592 .join(",",@str[0..15])."\n";
593 foreach (0..15) { shift @str; }
595 $self->{value}.="DB\t"
596 .join(",",@str) if (@str);
624 $line =~ s|[#!].*$||; # get rid of asm-style comments...
625 $line =~ s|/\*.*\*/||; # ... and C-style comments...
626 $line =~ s|^\s+||; # ... and skip white spaces in beginning
634 if ($label=label->re(\$line)) { print $label->out(); }
636 if (directive->re(\$line)) {
637 printf "%s",directive->out();
638 } elsif ($opcode=opcode->re(\$line)) { ARGUMENT: {
640 if ($src=register->re(\$line)) { opcode->size($src->size()); }
641 elsif ($src=const->re(\$line)) { }
642 elsif ($src=ea->re(\$line)) { }
643 elsif ($src=expr->re(\$line)) { }
645 last ARGUMENT if ($line !~ /^,/);
647 $line = substr($line,1); $line =~ s/^\s+//;
649 if ($dst=register->re(\$line)) { opcode->size($dst->size()); }
650 elsif ($dst=const->re(\$line)) { }
651 elsif ($dst=ea->re(\$line)) { }
659 printf "\t%s\t%s,%s", $opcode->out($dst->size()),
660 $src->out($sz),$dst->out($sz);
662 undef $sz if ($nasm && $opcode->mnemonic() eq "lea");
663 printf "\t%s\t%s,%s", $opcode->out(),
664 $dst->out($sz),$src->out($sz);
666 } elsif (defined($src)) {
667 printf "\t%s\t%s",$opcode->out(),$src->out($sz);
669 printf "\t%s",$opcode->out();
676 print "\n$current_segment\tENDS\n" if ($current_segment && $masm);
677 print "END\n" if ($masm);
681 \f#################################################
682 # Cross-reference x86_64 ABI "card"
702 # (*) volatile register
703 # (-) preserved by callee
704 # (#) Nth argument, volatile
706 # In Unix terms top of stack is argument transfer area for arguments
707 # which could not be accomodated in registers. Or in other words 7th
708 # [integer] argument resides at 8(%rsp) upon function entry point.
709 # 128 bytes above %rsp constitute a "red zone" which is not touched
710 # by signal handlers and can be used as temporal storage without
711 # allocating a frame.
713 # In Win64 terms N*8 bytes on top of stack is argument transfer area,
714 # which belongs to/can be overwritten by callee. N is the number of
715 # arguments passed to callee, *but* not less than 4! This means that
716 # upon function entry point 5th argument resides at 40(%rsp), as well
717 # as that 32 bytes from 8(%rsp) can always be used as temporal
718 # storage [without allocating a frame]. One can actually argue that
719 # one can assume a "red zone" above stack pointer under Win64 as well.
720 # Point is that at apparently no occasion Windows kernel would alter
721 # the area above user stack pointer in true asynchronous manner...
723 # All the above means that if assembler programmer adheres to Unix
724 # register and stack layout, but disregards the "red zone" existense,
725 # it's possible to use following prologue and epilogue to "gear" from
726 # Unix to Win64 ABI in leaf functions with not more than 6 arguments.
728 # omnipotent_function:
732 # movq %rcx,%rdi ; if 1st argument is actually present
733 # movq %rdx,%rsi ; if 2nd argument is actually ...
734 # movq %r8,%rdx ; if 3rd argument is ...
735 # movq %r9,%rcx ; if 4th argument ...
736 # movq 40(%rsp),%r8 ; if 5th ...
737 # movq 48(%rsp),%r9 ; if 6th ...
746 \f#################################################
747 # Win64 SEH, Structured Exception Handling.
749 # Unlike on Unix systems(*) lack of Win64 stack unwinding information
750 # has undesired side-effect at run-time: if an exception is raised in
751 # assembler subroutine such as those in question (basically we're
752 # referring to segmentation violations caused by malformed input
753 # parameters), the application is briskly terminated without invoking
754 # any exception handlers, most notably without generating memory dump
755 # or any user notification whatsoever. This poses a problem. It's
756 # possible to address it by registering custom language-specific
757 # handler that would restore processor context to the state at
758 # subroutine entry point and return "exception is not handled, keep
759 # unwinding" code. Writing such handler can be a challenge... But it's
760 # doable, though requires certain coding convention. Consider following
763 # .type function,@function
765 # movq %rsp,%rax # copy rsp to volatile register
766 # pushq %r15 # save non-volatile registers
770 # subq %rdi,%r11 # prepare [variable] stack frame
772 # movq %rax,0(%r11) # check for exceptions
773 # movq %r11,%rsp # allocate [variable] stack frame
774 # movq %rax,0(%rsp) # save original rsp value
777 # movq 0(%rsp),%rcx # pull original rsp value
778 # movq -24(%rcx),%rbp # restore non-volatile registers
779 # movq -16(%rcx),%rbx
781 # movq %rcx,%rsp # restore original rsp
783 # .size function,.-function
785 # The key is that up to magic_point copy of original rsp value remains
786 # in chosen volatile register and no non-volatile register, except for
787 # rsp, is modified. While past magic_point rsp remains constant till
788 # the very end of the function. In this case custom language-specific
789 # exception handler would look like this:
791 # EXCEPTION_DISPOSITION handler (EXCEPTION_RECORD *rec,ULONG64 frame,
792 # CONTEXT *context,DISPATCHER_CONTEXT *disp)
793 # { ULONG64 *rsp = (ULONG64 *)context->Rax;
794 # if (context->Rip >= magic_point)
795 # { rsp = ((ULONG64 **)context->Rsp)[0];
796 # context->Rbp = rsp[-3];
797 # context->Rbx = rsp[-2];
798 # context->R15 = rsp[-1];
800 # context->Rsp = (ULONG64)rsp;
801 # context->Rdi = rsp[1];
802 # context->Rsi = rsp[2];
804 # memcpy (disp->ContextRecord,context,sizeof(CONTEXT));
805 # RtlVirtualUnwind(UNW_FLAG_NHANDLER,disp->ImageBase,
806 # dips->ControlPc,disp->FunctionEntry,disp->ContextRecord,
807 # &disp->HandlerData,&disp->EstablisherFrame,NULL);
808 # return ExceptionContinueSearch;
811 # It's appropriate to implement this handler in assembler, directly in
812 # function's module. In order to do that one has to know members'
813 # offsets in CONTEXT and DISPATCHER_CONTEXT structures and some constant
814 # values. Here they are:
833 # sizeof(CONTEXT) 1232
834 # DISPATCHER_CONTEXT.ControlPc 0
835 # DISPATCHER_CONTEXT.ImageBase 8
836 # DISPATCHER_CONTEXT.FunctionEntry 16
837 # DISPATCHER_CONTEXT.EstablisherFrame 24
838 # DISPATCHER_CONTEXT.TargetIp 32
839 # DISPATCHER_CONTEXT.ContextRecord 40
840 # DISPATCHER_CONTEXT.LanguageHandler 48
841 # DISPATCHER_CONTEXT.HandlerData 56
842 # UNW_FLAG_NHANDLER 0
843 # ExceptionContinueSearch 1
845 # In order to tie the handler to the function one has to compose
846 # couple of structures: one for .xdata segment and one for .pdata.
848 # UNWIND_INFO structure for .xdata segment would be
850 # function_unwind_info:
854 # This structure designates exception handler for a function with
855 # zero-length prologue, no stack frame or frame register.
857 # To facilitate composing of .pdata structures, auto-generated "gear"
858 # prologue copies rsp value to rax and denotes next instruction with
859 # .LSEH_begin_{function_name} label. This essentially defines the SEH
860 # styling rule mentioned in the beginning. Position of this label is
861 # chosen in such manner that possible exceptions raised in the "gear"
862 # prologue would be accounted to caller and unwound from latter's frame.
863 # End of function is marked with respective .LSEH_end_{function_name}
864 # label. To summarize, .pdata segment would contain
866 # .rva .LSEH_begin_function
867 # .rva .LSEH_end_function
868 # .rva function_unwind_info
870 # Reference to functon_unwind_info from .xdata segment is the anchor.
871 # In case you wonder why references are 32-bit .rvas and not 64-bit
872 # .quads. References put into these two segments are required to be
873 # *relative* to the base address of the current binary module, a.k.a.
874 # image base. No Win64 module, be it .exe or .dll, can be larger than
875 # 2GB and thus such relative references can be and are accommodated in
878 # Having reviewed the example function code, one can argue that "movq
879 # %rsp,%rax" above is redundant. It is not! Keep in mind that on Unix
880 # rax would contain an undefined value. If this "offends" you, use
881 # another register and refrain from modifying rax till magic_point is
882 # reached, i.e. as if it was a non-volatile register. If more registers
883 # are required prior [variable] frame setup is completed, note that
884 # nobody says that you can have only one "magic point." You can
885 # "liberate" non-volatile registers by denoting last stack off-load
886 # instruction and reflecting it in finer grade unwind logic in handler.
887 # After all, isn't it why it's called *language-specific* handler...
889 # Attentive reader can notice that exceptions would be mishandled in
890 # auto-generated "gear" epilogue. Well, exception effectively can't
891 # occur there, because if memory area used by it was subject to
892 # segmentation violation, then it would be raised upon call to the
893 # function (and as already mentioned be accounted to caller, which is
894 # not a problem). If you're still not comfortable, then define tail
895 # "magic point" just prior ret instruction and have handler treat it...
897 # (*) Note that we're talking about run-time, not debug-time. Lack of
898 # unwind information makes debugging hard on both Windows and
899 # Unix. "Unlike" referes to the fact that on Unix signal handler
900 # will always be invoked, core dumped and appropriate exit code
901 # returned to parent (for user notification).