3 # Copyright (c) 1996-1998 Malcolm Beattie
5 # You may distribute under the terms of either the GNU General Public
6 # License or the Artistic License, as specified in the README file.
13 use B qw(minus_c main_cv main_root main_start comppadlist
14 class peekop walkoptree svref_2object cstring walksymtable
15 SVf_POK SVp_POK SVf_IOK SVp_IOK
17 use B::Asmdata qw(@optype @specialsv_name);
18 use B::Assembler qw(assemble_fh);
22 for ($i = 0; $i < @optype; $i++) {
23 $optype_enum{$optype[$i]} = $i;
26 # Following is SVf_POK|SVp_POK
27 # XXX Shouldn't be hardwired
28 sub POK () { SVf_POK|SVp_POK }
30 # Following is SVf_IOK|SVp_IOK
31 # XXX Shouldn't be hardwired
32 sub IOK () { SVf_IOK|SVp_IOK }
34 my ($verbose, $module_only, $no_assemble, $debug_bc, $debug_cv);
37 # Optimisation options. On the command line, use hyphens instead of
38 # underscores for compatibility with gcc-style options. We use
39 # underscores here because they are OK in (strict) barewords.
40 my ($strip_syntree, $compress_nullops, $omit_seq, $bypass_nullops);
41 my %optimise = (strip_syntax_tree => \$strip_syntree,
42 compress_nullops => \$compress_nullops,
43 omit_sequence_numbers => \$omit_seq,
44 bypass_nullops => \$bypass_nullops);
47 my %symtable; # maps object addresses to object indices.
48 # Filled in at allocation (newsv/newop) time.
49 my %saved; # maps object addresses (for SVish classes) to "saved yet?"
50 # flag. Set at FOO::bytecode time usually by SV::bytecode.
51 # Manipulated via saved(), mark_saved(), unmark_saved().
53 my $svix = -1; # we keep track of when the sv register contains an element
54 # of the object table to avoid unnecessary repeated
55 # consecutive ldsv instructions.
56 my $opix = -1; # Ditto for the op register.
97 return cstring($str . "\0");
103 sub saved { $saved{${$_[0]}} }
104 sub mark_saved { $saved{${$_[0]}} = 1 }
105 sub unmark_saved { $saved{${$_[0]}} = 0 }
107 sub debug { $debug_bc = shift }
111 warn sprintf("bytecode save method for %s (0x%x) not yet implemented\n",
116 # objix may stomp on the op register (for op objects)
117 # or the sv register (for SV objects)
119 sub B::OBJECT::objix {
121 my $ix = $symtable{$$obj};
125 $obj->newix($nextix);
126 return $symtable{$$obj} = $nextix++;
132 printf "newsv %d\t# %s\n", $sv->FLAGS & 0xf, class($sv);
138 my $gvname = $gv->NAME;
139 my $name = cstring($gv->STASH->NAME . "::" . $gvname);
140 print "gv_fetchpv $name\n";
146 my $name = $hv->NAME;
149 printf "gv_stashpv %s\n", cstring($name);
152 # It's an ordinary HV. Fall back to ordinary newix method
153 $hv->B::SV::newix($ix);
157 sub B::SPECIAL::newix {
159 # Special case. $$sv is not the address of the SV but an
160 # index into svspecialsv_list.
161 printf "ldspecsv $$sv\t# %s\n", $specialsv_name[$$sv];
167 my $class = class($op);
168 my $typenum = $optype_enum{$class};
169 croak "OP::newix: can't understand class $class" unless defined($typenum);
170 print "newop $typenum\t# $class\n";
174 sub B::OP::walkoptree_debug {
176 warn(sprintf("walkoptree: %s\n", peekop($op)));
179 sub B::OP::bytecode {
181 my $next = $op->next;
183 my $sibix = $op->sibling->objix;
185 my $type = $op->type;
187 if ($bypass_nullops) {
188 $next = $next->next while $$next && $next->type == 0;
190 $nextix = $next->objix;
192 printf "# %s\n", peekop($op) if $debug_bc;
194 print "op_next $nextix\n";
195 print "op_sibling $sibix\n" unless $strip_syntree;
196 printf "op_type %s\t# %d\n", $op->ppaddr, $type;
197 printf("op_seq %d\n", $op->seq) unless $omit_seq;
198 if ($type || !$compress_nullops) {
199 printf "op_targ %d\nop_flags 0x%x\nop_private 0x%x\n",
200 $op->targ, $op->flags, $op->private;
204 sub B::UNOP::bytecode {
206 my $firstix = $op->first->objix;
207 $op->B::OP::bytecode;
208 if (($op->type || !$compress_nullops) && !$strip_syntree) {
209 print "op_first $firstix\n";
213 sub B::LOGOP::bytecode {
215 my $otherix = $op->other->objix;
216 $op->B::UNOP::bytecode;
217 print "op_other $otherix\n";
220 sub B::SVOP::bytecode {
223 my $svix = $sv->objix;
224 $op->B::OP::bytecode;
225 print "op_sv $svix\n";
229 sub B::GVOP::bytecode {
232 my $gvix = $gv->objix;
233 $op->B::OP::bytecode;
234 print "op_gv $gvix\n";
238 sub B::PVOP::bytecode {
241 $op->B::OP::bytecode;
243 # This would be easy except that OP_TRANS uses a PVOP to store an
244 # endian-dependent array of 256 shorts instead of a plain string.
246 if ($op->ppaddr eq "pp_trans") {
247 my @shorts = unpack("s256", $pv); # assembler handles endianness
248 print "op_pv_tr ", join(",", @shorts), "\n";
250 printf "newpv %s\nop_pv\n", pvstring($pv);
254 sub B::BINOP::bytecode {
256 my $lastix = $op->last->objix;
257 $op->B::UNOP::bytecode;
258 if (($op->type || !$compress_nullops) && !$strip_syntree) {
259 print "op_last $lastix\n";
263 sub B::LISTOP::bytecode {
265 my $children = $op->children;
266 $op->B::BINOP::bytecode;
267 if (($op->type || !$compress_nullops) && !$strip_syntree) {
268 print "op_children $children\n";
272 sub B::LOOP::bytecode {
274 my $redoopix = $op->redoop->objix;
275 my $nextopix = $op->nextop->objix;
276 my $lastopix = $op->lastop->objix;
277 $op->B::LISTOP::bytecode;
278 print "op_redoop $redoopix\nop_nextop $nextopix\nop_lastop $lastopix\n";
281 sub B::COP::bytecode {
283 my $stash = $op->stash;
284 my $stashix = $stash->objix;
285 my $filegv = $op->filegv;
286 my $filegvix = $filegv->objix;
287 my $line = $op->line;
288 my $warnings = $op->warnings;
289 my $warningsix = $warnings->objix;
291 printf "# line %s:%d\n", $filegv->SV->PV, $line;
293 $op->B::OP::bytecode;
294 printf <<"EOT", pvstring($op->label), $op->cop_seq, $op->arybase;
302 cop_warnings $warningsix
308 sub B::PMOP::bytecode {
310 my $replroot = $op->pmreplroot;
311 my $replrootix = $replroot->objix;
312 my $replstartix = $op->pmreplstart->objix;
313 my $ppaddr = $op->ppaddr;
314 # pmnext is corrupt in some PMOPs (see misc.t for example)
315 #my $pmnextix = $op->pmnext->objix;
318 # OP_PUSHRE (a mutated version of OP_MATCH for the regexp
319 # argument to a split) stores a GV in op_pmreplroot instead
320 # of a substitution syntax tree. We don't want to walk that...
321 if ($ppaddr eq "pp_pushre") {
324 walkoptree($replroot, "bytecode");
327 $op->B::LISTOP::bytecode;
328 if ($ppaddr eq "pp_pushre") {
329 printf "op_pmreplrootgv $replrootix\n";
331 print "op_pmreplroot $replrootix\nop_pmreplstart $replstartix\n";
333 my $re = pvstring($op->precomp);
334 # op_pmnext omitted since a perl bug means it's sometime corrupt
335 printf <<"EOT", $op->pmflags, $op->pmpermflags;
343 sub B::SV::bytecode {
345 return if saved($sv);
347 my $refcnt = $sv->REFCNT;
348 my $flags = sprintf("0x%x", $sv->FLAGS);
350 print "sv_refcnt $refcnt\nsv_flags $flags\n";
354 sub B::PV::bytecode {
356 return if saved($sv);
357 $sv->B::SV::bytecode;
358 printf("newpv %s\nxpv\n", pvstring($sv->PV)) if $sv->FLAGS & POK;
361 sub B::IV::bytecode {
363 return if saved($sv);
365 $sv->B::SV::bytecode;
366 printf "%s $iv\n", $sv->needs64bits ? "xiv64" : "xiv32";
369 sub B::NV::bytecode {
371 return if saved($sv);
372 $sv->B::SV::bytecode;
373 printf "xnv %s\n", $sv->NVX;
376 sub B::RV::bytecode {
378 return if saved($sv);
380 my $rvix = $rv->objix;
382 $sv->B::SV::bytecode;
386 sub B::PVIV::bytecode {
388 return if saved($sv);
390 $sv->B::PV::bytecode;
391 printf "%s $iv\n", $sv->needs64bits ? "xiv64" : "xiv32";
394 sub B::PVNV::bytecode {
395 my ($sv, $flag) = @_;
396 # The $flag argument is passed through PVMG::bytecode by BM::bytecode
397 # and AV::bytecode and indicates special handling. $flag = 1 is used by
398 # BM::bytecode and means that we should ensure we save the whole B-M
399 # table. It consists of 257 bytes (256 char array plus a final \0)
400 # which follow the ordinary PV+\0 and the 257 bytes are *not* reflected
401 # in SvCUR. $flag = 2 is used by AV::bytecode and means that we only
402 # call SV::bytecode instead of saving PV and calling NV::bytecode since
403 # PV/NV/IV stuff is different for AVs.
404 return if saved($sv);
406 $sv->B::SV::bytecode;
409 $sv->B::IV::bytecode;
410 printf "xnv %s\n", $sv->NVX;
412 $pv .= "\0" . $sv->TABLE;
413 printf "newpv %s\npv_cur %d\nxpv\n", pvstring($pv),length($pv)-257;
415 printf("newpv %s\nxpv\n", pvstring($pv)) if $sv->FLAGS & POK;
420 sub B::PVMG::bytecode {
421 my ($sv, $flag) = @_;
422 # See B::PVNV::bytecode for an explanation of $flag.
423 return if saved($sv);
424 # XXX We assume SvSTASH is already saved and don't save it later ourselves
425 my $stashix = $sv->SvSTASH->objix;
426 my @mgchain = $sv->MAGIC;
429 # We need to traverse the magic chain and get objix for each OBJ
430 # field *before* we do B::PVNV::bytecode since objix overwrites
431 # the sv register. However, we need to write the magic-saving
432 # bytecode *after* B::PVNV::bytecode since sv isn't initialised
433 # to refer to $sv until then.
435 @mgobjix = map($_->OBJ->objix, @mgchain);
436 $sv->B::PVNV::bytecode($flag);
437 print "xmg_stash $stashix\n";
438 foreach $mg (@mgchain) {
439 printf "sv_magic %s\nmg_obj %d\nnewpv %s\nmg_pv\n",
440 cstring($mg->TYPE), shift(@mgobjix), pvstring($mg->PTR);
444 sub B::PVLV::bytecode {
446 return if saved($sv);
447 $sv->B::PVMG::bytecode;
448 printf <<'EOT', $sv->TARGOFF, $sv->TARGLEN, cstring($sv->TYPE);
455 sub B::BM::bytecode {
457 return if saved($sv);
458 # See PVNV::bytecode for an explanation of what the argument does
459 $sv->B::PVMG::bytecode(1);
460 printf "xbm_useful %d\nxbm_previous %d\nxbm_rare %d\n",
461 $sv->USEFUL, $sv->PREVIOUS, $sv->RARE;
464 sub B::GV::bytecode {
466 return if saved($gv);
469 my $gvname = $gv->NAME;
470 my $name = cstring($gv->STASH->NAME . "::" . $gvname);
472 my $egvix = $egv->objix;
474 printf <<"EOT", $gv->FLAGS, $gv->GvFLAGS, $gv->LINE;
479 my $refcnt = $gv->REFCNT;
480 printf("sv_refcnt_add %d\n", $refcnt - 1) if $refcnt > 1;
481 my $gvrefcnt = $gv->GvREFCNT;
482 printf("gp_refcnt_add %d\n", $gvrefcnt - 1) if $gvrefcnt > 1;
483 if ($gvrefcnt > 1 && $ix != $egvix) {
484 print "gp_share $egvix\n";
486 if ($gvname !~ /^([^A-Za-z]|STDIN|STDOUT|STDERR|ARGV|SIG|ENV)$/) {
488 my @subfield_names = qw(SV AV HV CV FILEGV FORM IO);
489 my @subfields = map($gv->$_(), @subfield_names);
490 my @ixes = map($_->objix, @subfields);
491 # Reset sv register for $gv
493 for ($i = 0; $i < @ixes; $i++) {
494 printf "gp_%s %d\n", lc($subfield_names[$i]), $ixes[$i];
496 # Now save all the subfields
498 foreach $sv (@subfields) {
505 sub B::HV::bytecode {
507 return if saved($hv);
509 my $name = $hv->NAME;
512 # It's an ordinary HV. Stashes have NAME set and need no further
513 # saving beyond the gv_stashpv that $hv->objix already ensures.
514 my @contents = $hv->ARRAY;
516 for ($i = 1; $i < @contents; $i += 2) {
517 push(@ixes, $contents[$i]->objix);
519 for ($i = 1; $i < @contents; $i += 2) {
520 $contents[$i]->bytecode;
523 for ($i = 0; $i < @contents; $i += 2) {
524 printf("newpv %s\nhv_store %d\n",
525 pvstring($contents[$i]), $ixes[$i / 2]);
527 printf "sv_refcnt %d\nsv_flags 0x%x\n", $hv->REFCNT, $hv->FLAGS;
531 sub B::AV::bytecode {
533 return if saved($av);
535 my $fill = $av->FILL;
540 @ixes = map($_->objix, @array);
542 foreach $sv (@array) {
546 # See PVNV::bytecode for the meaning of the flag argument of 2.
547 $av->B::PVMG::bytecode(2);
548 # Recover sv register and set AvMAX and AvFILL to -1 (since we
549 # create an AV with NEWSV and SvUPGRADE rather than doing newAV
550 # which is what sets AvMAX and AvFILL.
552 printf "xav_flags 0x%x\nxav_max -1\nxav_fill -1\n", $av->AvFLAGS;
555 foreach $elix (@ixes) {
556 print "av_push $elix\n";
560 print "av_extend $max\n";
565 sub B::CV::bytecode {
567 return if saved($cv);
569 $cv->B::PVMG::bytecode;
571 my @subfield_names = qw(ROOT START STASH GV FILEGV PADLIST OUTSIDE);
572 my @subfields = map($cv->$_(), @subfield_names);
573 my @ixes = map($_->objix, @subfields);
574 # Save OP tree from CvROOT (first element of @subfields)
575 my $root = shift @subfields;
577 walkoptree($root, "bytecode");
579 # Reset sv register for $cv (since above ->objix calls stomped on it)
581 for ($i = 0; $i < @ixes; $i++) {
582 printf "xcv_%s %d\n", lc($subfield_names[$i]), $ixes[$i];
584 printf "xcv_depth %d\nxcv_flags 0x%x\n", $cv->DEPTH, $cv->FLAGS;
585 # Now save all the subfields (except for CvROOT which was handled
586 # above) and CvSTART (now the initial element of @subfields).
587 shift @subfields; # bye-bye CvSTART
589 foreach $sv (@subfields) {
594 sub B::IO::bytecode {
596 return if saved($io);
598 my $top_gv = $io->TOP_GV;
599 my $top_gvix = $top_gv->objix;
600 my $fmt_gv = $io->FMT_GV;
601 my $fmt_gvix = $fmt_gv->objix;
602 my $bottom_gv = $io->BOTTOM_GV;
603 my $bottom_gvix = $bottom_gv->objix;
605 $io->B::PVMG::bytecode;
607 print "xio_top_gv $top_gvix\n";
608 print "xio_fmt_gv $fmt_gvix\n";
609 print "xio_bottom_gv $bottom_gvix\n";
611 foreach $field (qw(TOP_NAME FMT_NAME BOTTOM_NAME)) {
612 printf "newpv %s\nxio_%s\n", pvstring($io->$field()), lc($field);
614 foreach $field (qw(LINES PAGE PAGE_LEN LINES_LEFT SUBPROCESS)) {
615 printf "xio_%s %d\n", lc($field), $io->$field();
617 printf "xio_type %s\nxio_flags 0x%x\n", cstring($io->IoTYPE), $io->IoFLAGS;
620 $bottom_gv->bytecode;
623 sub B::SPECIAL::bytecode {
624 # nothing extra needs doing
627 sub bytecompile_object {
630 svref_2object($sv)->bytecode;
634 sub B::GV::bytecodecv {
637 if ($$cv && !saved($cv)) {
639 warn sprintf("saving extra CV &%s::%s (0x%x) from GV 0x%x\n",
640 $gv->STASH->NAME, $gv->NAME, $$cv, $$gv);
646 sub bytecompile_main {
647 my $curpad = (comppadlist->ARRAY)[1];
648 my $curpadix = $curpad->objix;
650 walkoptree(main_root, "bytecode");
651 warn "done main program, now walking symbol table\n" if $debug_bc;
652 my ($pack, %exclude);
653 foreach $pack (qw(B O AutoLoader DynaLoader Config DB VMS strict vars
654 FileHandle Exporter Carp UNIVERSAL IO Fcntl Symbol
655 SelectSaver blib Cwd))
657 $exclude{$pack."::"} = 1;
659 no strict qw(vars refs);
660 walksymtable(\%{"main::"}, "bytecodecv", sub {
661 warn "considering $_[0]\n" if $debug_bc;
662 return !defined($exclude{$_[0]});
665 printf "main_root %d\n", main_root->objix;
666 printf "main_start %d\n", main_start->objix;
667 printf "curpad $curpadix\n";
668 # XXX Do min_intro_pending and max_intro_pending matter?
672 sub prepare_assemble {
673 my $newfh = IO::File->new_tmpfile;
681 seek($fh, 0, 0); # rewind the temporary file
682 assemble_fh($fh, sub { print OUT @_ });
687 my ($option, $opt, $arg);
688 open(OUT, ">&STDOUT");
692 while ($option = shift @options) {
693 if ($option =~ /^-(.)(.*)/) {
697 unshift @options, $option;
700 if ($opt eq "-" && $arg eq "-") {
703 } elsif ($opt eq "o") {
704 $arg ||= shift @options;
705 open(OUT, ">$arg") or return "$arg: $!\n";
707 } elsif ($opt eq "D") {
708 $arg ||= shift @options;
709 foreach $arg (split(//, $arg)) {
713 } elsif ($arg eq "o") {
715 } elsif ($arg eq "a") {
716 B::Assembler::debug(1);
717 } elsif ($arg eq "C") {
721 } elsif ($opt eq "v") {
723 } elsif ($opt eq "m") {
725 } elsif ($opt eq "S") {
727 } elsif ($opt eq "f") {
728 $arg ||= shift @options;
729 my $value = $arg !~ s/^no-//;
731 my $ref = $optimise{$arg};
735 warn qq(ignoring unknown optimisation option "$arg"\n);
737 } elsif ($opt eq "O") {
738 $arg = 1 if $arg eq "";
740 foreach $ref (values %optimise) {
750 $compress_nullops = 1;
759 $newfh = prepare_assemble() unless $no_assemble;
760 foreach $objname (@options) {
761 eval "bytecompile_object(\\$objname)";
763 do_assemble($newfh) unless $no_assemble;
768 $newfh = prepare_assemble() unless $no_assemble;
770 do_assemble($newfh) unless $no_assemble;
781 B::Bytecode - Perl compiler's bytecode backend
785 perl -MO=Bytecode[,OPTIONS] foo.pl
789 This compiler backend takes Perl source and generates a
790 platform-independent bytecode encapsulating code to load the
791 internal structures perl uses to run your program. When the
792 generated bytecode is loaded in, your program is ready to run,
793 reducing the time which perl would have taken to load and parse
794 your program into its internal semi-compiled form. That means that
795 compiling with this backend will not help improve the runtime
796 execution speed of your program but may improve the start-up time.
797 Depending on the environment in which your program runs this may
798 or may not be a help.
800 The resulting bytecode can be run with a special byteperl executable
801 or (for non-main programs) be loaded via the C<byteload_fh> function
806 If there are any non-option arguments, they are taken to be names of
807 objects to be saved (probably doesn't work properly yet). Without
808 extra arguments, it saves the main program.
814 Output to filename instead of STDOUT.
818 Force end of options.
822 Force optimisations on or off one at a time. Each can be preceded
823 by B<no-> to turn the option off (e.g. B<-fno-compress-nullops>).
825 =item B<-fcompress-nullops>
827 Only fills in the necessary fields of ops which have
828 been optimised away by perl's internal compiler.
830 =item B<-fomit-sequence-numbers>
832 Leaves out code to fill in the op_seq field of all ops
833 which is only used by perl's internal compiler.
835 =item B<-fbypass-nullops>
837 If op->op_next ever points to a NULLOP, replaces the op_next field
838 with the first non-NULLOP in the path of execution.
840 =item B<-fstrip-syntax-tree>
842 Leaves out code to fill in the pointers which link the internal syntax
843 tree together. They're not needed at run-time but leaving them out
844 will make it impossible to recompile or disassemble the resulting
845 program. It will also stop C<goto label> statements from working.
849 Optimisation level (n = 0, 1, 2, ...). B<-O> means B<-O1>.
850 B<-O1> sets B<-fcompress-nullops> B<-fomit-sequence numbers>.
851 B<-O6> adds B<-fstrip-syntax-tree>.
855 Debug options (concatenated or separate flags like C<perl -D>).
859 Prints each OP as it's processed.
863 Print debugging information about bytecompiler progress.
867 Tells the (bytecode) assembler to include source assembler lines
868 in its output as bytecode comments.
872 Prints each CV taken from the final symbol tree walk.
876 Output (bytecode) assembler source rather than piping it
877 through the assembler and outputting bytecode.
881 Compile as a module rather than a standalone program. Currently this
882 just means that the bytecodes for initialising C<main_start>,
883 C<main_root> and C<curpad> are omitted.
889 perl -MO=Bytecode,-O6,-o,foo.plc foo.pl
891 perl -MO=Bytecode,-S foo.pl > foo.S
892 assemble foo.S > foo.plc
894 Note that C<assemble> lives in the C<B> subdirectory of your perl
895 library directory. The utility called perlcc may also be used to
896 help make use of this compiler.
898 perl -MO=Bytecode,-m,-oFoo.pmc Foo.pm
902 Plenty. Current status: experimental.
906 Malcolm Beattie, C<mbeattie@sable.ox.ac.uk>