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", "pp_" . $op->name, $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::PADOP::bytecode {
231 my $padix = $op->padix;
232 $op->B::OP::bytecode;
233 print "op_padix $padix\n";
236 sub B::PVOP::bytecode {
239 $op->B::OP::bytecode;
241 # This would be easy except that OP_TRANS uses a PVOP to store an
242 # endian-dependent array of 256 shorts instead of a plain string.
244 if ($op->name eq "trans") {
245 my @shorts = unpack("s256", $pv); # assembler handles endianness
246 print "op_pv_tr ", join(",", @shorts), "\n";
248 printf "newpv %s\nop_pv\n", pvstring($pv);
252 sub B::BINOP::bytecode {
254 my $lastix = $op->last->objix;
255 $op->B::UNOP::bytecode;
256 if (($op->type || !$compress_nullops) && !$strip_syntree) {
257 print "op_last $lastix\n";
261 sub B::LISTOP::bytecode {
263 my $children = $op->children;
264 $op->B::BINOP::bytecode;
265 if (($op->type || !$compress_nullops) && !$strip_syntree) {
266 print "op_children $children\n";
270 sub B::LOOP::bytecode {
272 my $redoopix = $op->redoop->objix;
273 my $nextopix = $op->nextop->objix;
274 my $lastopix = $op->lastop->objix;
275 $op->B::LISTOP::bytecode;
276 print "op_redoop $redoopix\nop_nextop $nextopix\nop_lastop $lastopix\n";
279 sub B::COP::bytecode {
281 my $stash = $op->stash;
282 my $stashix = $stash->objix;
283 my $file = $op->file;
284 my $line = $op->line;
285 my $warnings = $op->warnings;
286 my $warningsix = $warnings->objix;
288 printf "# line %s:%d\n", $file, $line;
290 $op->B::OP::bytecode;
291 printf <<"EOT", pvstring($op->label), $op->cop_seq, pvstring($file), $op->arybase;
300 cop_warnings $warningsix
305 sub B::PMOP::bytecode {
307 my $replroot = $op->pmreplroot;
308 my $replrootix = $replroot->objix;
309 my $replstartix = $op->pmreplstart->objix;
310 my $opname = $op->name;
311 # pmnext is corrupt in some PMOPs (see misc.t for example)
312 #my $pmnextix = $op->pmnext->objix;
315 # OP_PUSHRE (a mutated version of OP_MATCH for the regexp
316 # argument to a split) stores a GV in op_pmreplroot instead
317 # of a substitution syntax tree. We don't want to walk that...
318 if ($opname eq "pushre") {
321 walkoptree($replroot, "bytecode");
324 $op->B::LISTOP::bytecode;
325 if ($opname eq "pushre") {
326 printf "op_pmreplrootgv $replrootix\n";
328 print "op_pmreplroot $replrootix\nop_pmreplstart $replstartix\n";
330 my $re = pvstring($op->precomp);
331 # op_pmnext omitted since a perl bug means it's sometime corrupt
332 printf <<"EOT", $op->pmflags, $op->pmpermflags;
340 sub B::SV::bytecode {
342 return if saved($sv);
344 my $refcnt = $sv->REFCNT;
345 my $flags = sprintf("0x%x", $sv->FLAGS);
347 print "sv_refcnt $refcnt\nsv_flags $flags\n";
351 sub B::PV::bytecode {
353 return if saved($sv);
354 $sv->B::SV::bytecode;
355 printf("newpv %s\nxpv\n", pvstring($sv->PV)) if $sv->FLAGS & POK;
358 sub B::IV::bytecode {
360 return if saved($sv);
362 $sv->B::SV::bytecode;
363 printf "%s $iv\n", $sv->needs64bits ? "xiv64" : "xiv32";
366 sub B::NV::bytecode {
368 return if saved($sv);
369 $sv->B::SV::bytecode;
370 printf "xnv %s\n", $sv->NVX;
373 sub B::RV::bytecode {
375 return if saved($sv);
377 my $rvix = $rv->objix;
379 $sv->B::SV::bytecode;
383 sub B::PVIV::bytecode {
385 return if saved($sv);
387 $sv->B::PV::bytecode;
388 printf "%s $iv\n", $sv->needs64bits ? "xiv64" : "xiv32";
391 sub B::PVNV::bytecode {
393 my $flag = shift || 0;
394 # The $flag argument is passed through PVMG::bytecode by BM::bytecode
395 # and AV::bytecode and indicates special handling. $flag = 1 is used by
396 # BM::bytecode and means that we should ensure we save the whole B-M
397 # table. It consists of 257 bytes (256 char array plus a final \0)
398 # which follow the ordinary PV+\0 and the 257 bytes are *not* reflected
399 # in SvCUR. $flag = 2 is used by AV::bytecode and means that we only
400 # call SV::bytecode instead of saving PV and calling NV::bytecode since
401 # PV/NV/IV stuff is different for AVs.
402 return if saved($sv);
404 $sv->B::SV::bytecode;
407 $sv->B::IV::bytecode;
408 printf "xnv %s\n", $sv->NVX;
410 $pv .= "\0" . $sv->TABLE;
411 printf "newpv %s\npv_cur %d\nxpv\n", pvstring($pv),length($pv)-257;
413 printf("newpv %s\nxpv\n", pvstring($pv)) if $sv->FLAGS & POK;
418 sub B::PVMG::bytecode {
419 my ($sv, $flag) = @_;
420 # See B::PVNV::bytecode for an explanation of $flag.
421 return if saved($sv);
422 # XXX We assume SvSTASH is already saved and don't save it later ourselves
423 my $stashix = $sv->SvSTASH->objix;
424 my @mgchain = $sv->MAGIC;
427 # We need to traverse the magic chain and get objix for each OBJ
428 # field *before* we do B::PVNV::bytecode since objix overwrites
429 # the sv register. However, we need to write the magic-saving
430 # bytecode *after* B::PVNV::bytecode since sv isn't initialised
431 # to refer to $sv until then.
433 @mgobjix = map($_->OBJ->objix, @mgchain);
434 $sv->B::PVNV::bytecode($flag);
435 print "xmg_stash $stashix\n";
436 foreach $mg (@mgchain) {
437 printf "sv_magic %s\nmg_obj %d\nnewpv %s\nmg_pv\n",
438 cstring($mg->TYPE), shift(@mgobjix), pvstring($mg->PTR);
442 sub B::PVLV::bytecode {
444 return if saved($sv);
445 $sv->B::PVMG::bytecode;
446 printf <<'EOT', $sv->TARGOFF, $sv->TARGLEN, cstring($sv->TYPE);
453 sub B::BM::bytecode {
455 return if saved($sv);
456 # See PVNV::bytecode for an explanation of what the argument does
457 $sv->B::PVMG::bytecode(1);
458 printf "xbm_useful %d\nxbm_previous %d\nxbm_rare %d\n",
459 $sv->USEFUL, $sv->PREVIOUS, $sv->RARE;
462 sub B::GV::bytecode {
464 return if saved($gv);
467 my $gvname = $gv->NAME;
468 my $name = cstring($gv->STASH->NAME . "::" . $gvname);
470 my $egvix = $egv->objix;
472 printf <<"EOT", $gv->FLAGS, $gv->GvFLAGS, $gv->LINE, pvstring($gv->FILE);
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)$/) {
488 my @subfield_names = qw(SV AV HV CV FORM IO);
489 my @subfields = map($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 $cv->B::PVMG::bytecode;
571 my @subfield_names = qw(ROOT START STASH GV PADLIST OUTSIDE);
572 my @subfields = map($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, $cv->FLAGS;
585 printf "newpv %s\nxcv_file\n", pvstring($cv->FILE);
586 # Now save all the subfields (except for CvROOT which was handled
587 # above) and CvSTART (now the initial element of @subfields).
588 shift @subfields; # bye-bye CvSTART
590 foreach $sv (@subfields) {
595 sub B::IO::bytecode {
597 return if saved($io);
599 my $top_gv = $io->TOP_GV;
600 my $top_gvix = $top_gv->objix;
601 my $fmt_gv = $io->FMT_GV;
602 my $fmt_gvix = $fmt_gv->objix;
603 my $bottom_gv = $io->BOTTOM_GV;
604 my $bottom_gvix = $bottom_gv->objix;
606 $io->B::PVMG::bytecode;
608 print "xio_top_gv $top_gvix\n";
609 print "xio_fmt_gv $fmt_gvix\n";
610 print "xio_bottom_gv $bottom_gvix\n";
612 foreach $field (qw(TOP_NAME FMT_NAME BOTTOM_NAME)) {
613 printf "newpv %s\nxio_%s\n", pvstring($io->$field()), lc($field);
615 foreach $field (qw(LINES PAGE PAGE_LEN LINES_LEFT SUBPROCESS)) {
616 printf "xio_%s %d\n", lc($field), $io->$field();
618 printf "xio_type %s\nxio_flags 0x%x\n", cstring($io->IoTYPE), $io->IoFLAGS;
621 $bottom_gv->bytecode;
624 sub B::SPECIAL::bytecode {
625 # nothing extra needs doing
628 sub bytecompile_object {
631 svref_2object($sv)->bytecode;
635 sub B::GV::bytecodecv {
638 if ($$cv && !saved($cv)) {
640 warn sprintf("saving extra CV &%s::%s (0x%x) from GV 0x%x\n",
641 $gv->STASH->NAME, $gv->NAME, $$cv, $$gv);
647 sub bytecompile_main {
648 my $curpad = (comppadlist->ARRAY)[1];
649 my $curpadix = $curpad->objix;
651 walkoptree(main_root, "bytecode");
652 warn "done main program, now walking symbol table\n" if $debug_bc;
653 my ($pack, %exclude);
654 foreach $pack (qw(B O AutoLoader DynaLoader Config DB VMS strict vars
655 FileHandle Exporter Carp UNIVERSAL IO Fcntl Symbol
656 SelectSaver blib Cwd))
658 $exclude{$pack."::"} = 1;
660 no strict qw(vars refs);
661 walksymtable(\%{"main::"}, "bytecodecv", sub {
662 warn "considering $_[0]\n" if $debug_bc;
663 return !defined($exclude{$_[0]});
666 printf "main_root %d\n", main_root->objix;
667 printf "main_start %d\n", main_start->objix;
668 printf "curpad $curpadix\n";
669 # XXX Do min_intro_pending and max_intro_pending matter?
673 sub prepare_assemble {
674 my $newfh = IO::File->new_tmpfile;
682 seek($fh, 0, 0); # rewind the temporary file
683 assemble_fh($fh, sub { print OUT @_ });
688 my ($option, $opt, $arg);
689 open(OUT, ">&STDOUT");
693 while ($option = shift @options) {
694 if ($option =~ /^-(.)(.*)/) {
698 unshift @options, $option;
701 if ($opt eq "-" && $arg eq "-") {
704 } elsif ($opt eq "o") {
705 $arg ||= shift @options;
706 open(OUT, ">$arg") or return "$arg: $!\n";
708 } elsif ($opt eq "D") {
709 $arg ||= shift @options;
710 foreach $arg (split(//, $arg)) {
714 } elsif ($arg eq "o") {
716 } elsif ($arg eq "a") {
717 B::Assembler::debug(1);
718 } elsif ($arg eq "C") {
722 } elsif ($opt eq "v") {
724 } elsif ($opt eq "m") {
726 } elsif ($opt eq "S") {
728 } elsif ($opt eq "f") {
729 $arg ||= shift @options;
730 my $value = $arg !~ s/^no-//;
732 my $ref = $optimise{$arg};
736 warn qq(ignoring unknown optimisation option "$arg"\n);
738 } elsif ($opt eq "O") {
739 $arg = 1 if $arg eq "";
741 foreach $ref (values %optimise) {
751 $compress_nullops = 1;
760 $newfh = prepare_assemble() unless $no_assemble;
761 foreach $objname (@options) {
762 eval "bytecompile_object(\\$objname)";
764 do_assemble($newfh) unless $no_assemble;
769 $newfh = prepare_assemble() unless $no_assemble;
771 do_assemble($newfh) unless $no_assemble;
782 B::Bytecode - Perl compiler's bytecode backend
786 perl -MO=Bytecode[,OPTIONS] foo.pl
790 This compiler backend takes Perl source and generates a
791 platform-independent bytecode encapsulating code to load the
792 internal structures perl uses to run your program. When the
793 generated bytecode is loaded in, your program is ready to run,
794 reducing the time which perl would have taken to load and parse
795 your program into its internal semi-compiled form. That means that
796 compiling with this backend will not help improve the runtime
797 execution speed of your program but may improve the start-up time.
798 Depending on the environment in which your program runs this may
799 or may not be a help.
801 The resulting bytecode can be run with a special byteperl executable
802 or (for non-main programs) be loaded via the C<byteload_fh> function
807 If there are any non-option arguments, they are taken to be names of
808 objects to be saved (probably doesn't work properly yet). Without
809 extra arguments, it saves the main program.
815 Output to filename instead of STDOUT.
819 Force end of options.
823 Force optimisations on or off one at a time. Each can be preceded
824 by B<no-> to turn the option off (e.g. B<-fno-compress-nullops>).
826 =item B<-fcompress-nullops>
828 Only fills in the necessary fields of ops which have
829 been optimised away by perl's internal compiler.
831 =item B<-fomit-sequence-numbers>
833 Leaves out code to fill in the op_seq field of all ops
834 which is only used by perl's internal compiler.
836 =item B<-fbypass-nullops>
838 If op->op_next ever points to a NULLOP, replaces the op_next field
839 with the first non-NULLOP in the path of execution.
841 =item B<-fstrip-syntax-tree>
843 Leaves out code to fill in the pointers which link the internal syntax
844 tree together. They're not needed at run-time but leaving them out
845 will make it impossible to recompile or disassemble the resulting
846 program. It will also stop C<goto label> statements from working.
850 Optimisation level (n = 0, 1, 2, ...). B<-O> means B<-O1>.
851 B<-O1> sets B<-fcompress-nullops> B<-fomit-sequence numbers>.
852 B<-O6> adds B<-fstrip-syntax-tree>.
856 Debug options (concatenated or separate flags like C<perl -D>).
860 Prints each OP as it's processed.
864 Print debugging information about bytecompiler progress.
868 Tells the (bytecode) assembler to include source assembler lines
869 in its output as bytecode comments.
873 Prints each CV taken from the final symbol tree walk.
877 Output (bytecode) assembler source rather than piping it
878 through the assembler and outputting bytecode.
882 Compile as a module rather than a standalone program. Currently this
883 just means that the bytecodes for initialising C<main_start>,
884 C<main_root> and C<curpad> are omitted.
890 perl -MO=Bytecode,-O6,-o,foo.plc foo.pl
892 perl -MO=Bytecode,-S foo.pl > foo.S
893 assemble foo.S > foo.plc
895 Note that C<assemble> lives in the C<B> subdirectory of your perl
896 library directory. The utility called perlcc may also be used to
897 help make use of this compiler.
899 perl -MO=Bytecode,-m,-oFoo.pmc Foo.pm
903 Plenty. Current status: experimental.
907 Malcolm Beattie, C<mbeattie@sable.ox.ac.uk>