=head1 NAME

Faster - do some things faster

=head1 SYNOPSIS

 use Faster;

 perl -MFaster ...

=head1 DESCRIPTION

This module implements a very simple-minded JIT. It works by more or less
translating every function it sees into a C program, compiling it and then
replacing the function by the compiled code.

As a result, startup times are immense, as every function might lead to a
full-blown compilation.

The speed improvements are also not great, you can expect 20% or so on
average, for code that runs very often.

Faster is in the early stages of development. Due to its design its
relatively safe to use (it will either work or simply slowdown the program
immensely, but rarely cause bugs).

Usage is very easy, just C<use Faster> and every function called from then
on will be compiled.

Right now, Faster will leave lots of F<*.c>, F<*.o> and F<*.so> files in
F</tmp>, and it will even create those temporary files in an insecure
manner, so watch out.

=over 4

=cut

package Faster;

use strict;
use Config;
use B ();
#use Digest::MD5 ();
use DynaLoader ();

BEGIN {
   our $VERSION = '0.01';

   require XSLoader;
   XSLoader::load __PACKAGE__, $VERSION;
}

my $COMPILE = "$Config{cc} -c -I$Config{archlibexp}/CORE $Config{optimize} $Config{ccflags} $Config{cccdlflags}";
my $LINK    = "$Config{ld} $Config{ldflags} $Config{lddlflags} $Config{ccdlflags}";
my $LIBS    = "$Config{libs}";
my $_o      = $Config{_o};
my $_so     = ".so";

# we don't need no steenking PIC on x86
$COMPILE =~ s/-f(?:PIC|pic)//g
   if $Config{archname} =~ /^(i[3456]86)-/;

my $opt_assert = 0;

our $source;

our @ops;
our $op;
our $op_name;
our @op_loop;
our %op_regcomp;

my %flag;

# complex flag steting is no longer required, rewrite this ugly code
for (split /\n/, <<EOF) {
   leavesub	unsafe
   leavesublv	unsafe
   return	unsafe
   flip		unsafe
   goto		unsafe
   last		unsafe
   redo		unsafe
   next		unsafe
   eval		unsafe
   leaveeval	unsafe
   entertry	unsafe
   formline	unsafe
   grepstart	unsafe
   mapstart	unsafe
   substcont	unsafe
   entereval	unsafe noasync	todo
   require	unsafe

   mapstart	noasync
   grepstart	noasync
   match	noasync

   last		noasync
   next		noasync
   redo		noasync
   seq		noasync
   pushmark	noasync extend=0
   padsv	noasync extend=1
   padav	noasync extend=1
   padhv	noasync extend=1
   padany	noasync extend=1
   entersub	noasync
   aassign	noasync
   sassign	noasync
   rv2av	noasync
   rv2cv	noasync
   rv2gv	noasync
   rv2hv	noasync
   refgen	noasync
   nextstate	noasync
   gv		noasync
   gvsv		noasync
   add		noasync
   subtract	noasync
   multiply	noasync
   divide	noasync
   complement	noasync
   cond_expr	noasync
   and		noasync
   or		noasync
   not		noasync
   defined	noasync
   method_named	noasync
   preinc	noasync
   postinc	noasync
   predec	noasync
   postdec	noasync
   stub		noasync
   unstack	noasync
   leaveloop	noasync
   aelem	noasync
   aelemfast	noasync	
   helem	noasync
   delete	noasync
   exists	noasync
   pushre	noasync
   subst	noasync
   const	noasync extend=1
   list		noasync
   join		noasync
   split	noasync
   concat	noasync
   push		noasync
   pop  	noasync
   shift 	noasync
   unshift     	noasync
   length	noasync
   substr	noasync
   stringify	noasync
   eq		noasync
   ne		noasync
   gt		noasync
   lt		noasync
   ge		noasync
   le		noasync
   enteriter	noasync
   ord		noasync
   orassign	noasync
   regcomp	noasync
   regcreset	noasync
   regcmaybe	noasync

   iter		async
EOF
   my (undef, $op, @flags) = split /\s+/;

   undef $flag{$_}{$op}
      for ("known", @flags);
}

my %callop = (
   entersub => "(PL_ppaddr [OP_ENTERSUB]) (aTHX)",
   mapstart => "Perl_pp_grepstart (aTHX)",
);

sub callop {
   $callop{$op_name} || "Perl_pp_$op_name (aTHX)"
}

sub assert {
   return unless $opt_assert;
   $source .= "  assert ((\"$op_name\", ($_[0])));\n";
}

sub out_callop {
   assert "nextop == (OP *)$$op";
   $source .= "  PL_op = nextop; nextop = " . (callop $op) . ";\n";
}

sub out_cond_jump {
   $source .= "  if (nextop == (OP *)${$_[0]}L) goto op_${$_[0]};\n";
}

sub out_jump_next {
   out_cond_jump $op_regcomp{$$op}
      if $op_regcomp{$$op};

   assert "nextop == (OP *)${$op->next}";
   $source .= "  goto op_${$op->next};\n";
}

sub out_next {
   $source .= "  nextop = (OP *)${$op->next}L;\n";

   out_jump_next;
}

sub out_linear {
   out_callop;
   out_jump_next;
}

sub op_entersub {
   out_callop;
   $source .= "  RUNOPS_TILL ((OP *)${$op->next}L);\n";
   out_jump_next;
}

*op_require = \&op_entersub;

sub op_nextstate {
   $source .= "  PL_curcop = (COP *)nextop;\n";
   $source .= "  PL_stack_sp = PL_stack_base + cxstack[cxstack_ix].blk_oldsp;\n";
   $source .= "  FREETMPS;\n";

   out_next;
}

sub op_pushmark {
   $source .= "  PUSHMARK (PL_stack_sp);\n";

   out_next;
}

if ($Config{useithreads} ne "define") {
   # disable optimisations on ithreads

   *op_const = sub {
      $source .= "  { dSP; XPUSHs ((SV *)${$op->sv}L); PUTBACK; }\n";

      out_next;
   };

   *op_gv = \&op_const;

   *op_aelemfast = sub {
      my $targ = $op->targ;
      my $private = $op->private;

      $source .= "  {\n";

      if ($op->flags & B::OPf_SPECIAL) {
         $source .= "    AV *av = (AV*)PAD_SV((PADOFFSET)$targ);\n";
      } else {
         $source .= "    AV *av = GvAV ((GV *)${$op->sv}L);\n";
      }

      if ($op->flags & B::OPf_MOD) {
         $source .= "    SV *sv = *av_fetch (av, $private, 1);\n";
      } else {
         $source .= "    SV **svp = av_fetch (av, $private, 0); SV *sv = svp ? *svp : &PL_sv_undef;\n";
      }

      if (!($op->flags & B::OPf_MOD)) {
         $source .= "    if (SvGMAGICAL (sv)) sv = sv_mortalcopy (sv);\n";
      }

      $source .= "    dSP;\n";
      $source .= "    XPUSHs (sv);\n";
      $source .= "    PUTBACK;\n";
      $source .= "  }\n";

      out_next;
   };

   *op_gvsv = sub {
      $source .= "  {\n";
      $source .= "    dSP;\n";
      $source .= "    EXTEND (SP, 1);\n";

      if ($op->private & B::OPpLVAL_INTRO) {
        $source .= "    PUSHs (save_scalar ((GV *)${$op->sv}L));\n";
      } else {
        $source .= "    PUSHs (GvSV ((GV *)${$op->sv}L));\n";
      }

      $source .= "    PUTBACK;\n";
      $source .= "  }\n";

      out_next;
   };
}

# does kill Crossfire/res2pm
sub op_stringify {
   my $targ = $op->targ;

   $source .= <<EOF;
  {
    dSP;
    SV *targ = PAD_SV ((PADOFFSET)$targ);
    sv_copypv (TARG, TOPs);
    SETTARG;
    PUTBACK;
  }
EOF

   out_next;
}

sub op_and {
   $source .= <<EOF;
  {
    dSP;

    if (SvTRUE (TOPs))
      {
        --SP;
        PUTBACK;
        nextop = (OP *)${$op->other}L;
        goto op_${$op->other};
      }
  }
EOF

   out_next;
}

sub op_or {
   $source .= <<EOF;
  {
    dSP;

    if (!SvTRUE (TOPs))
      {
        --SP;
        PUTBACK;
        nextop = (OP *)${$op->other}L;
        goto op_${$op->other};
      }
  }
EOF

   out_next;
}

sub op_padsv {
   my $flags = $op->flags;
   my $target = $op->targ;

   $source .= <<EOF;
  {
    dSP;
    XPUSHs (PAD_SV ((PADOFFSET)$target));
    PUTBACK;
EOF
   if ($op->flags & B::OPf_MOD) {
      if ($op->private & B::OPpLVAL_INTRO) {
         $source .= "    SAVECLEARSV (PAD_SVl ((PADOFFSET)$target));\n";
      } elsif ($op->private & B::OPpDEREF) {
         my $deref = $op->private & B::OPpDEREF;
         $source .= "    Perl_vivify_ref (PAD_SVl ((PADOFFSET)$target), $deref);\n";
      }
   }
   $source .= <<EOF;
  }
EOF

   out_next;
}

# pattern const+ (or general push1)
# pattern pushmark return(?)
# pattern pushmark gv rv2av pushmark padsv+o.ä. aassign

# pattern const method_named
sub op_method_named {
   $source .= <<EOF;
  {
    static HV *last_stash;
    static SV *last_cv;
    static U32 last_sub_generation;

    SV *obj = *(PL_stack_base + TOPMARK + 1);

    if (!SvGMAGICAL (obj) && SvROK (obj) && SvOBJECT (SvRV (obj)))
      {
        dSP;
        HV *stash = SvSTASH (SvRV (obj));

        /* simple "polymorphic" inline cache */
        if (stash == last_stash
            && PL_sub_generation == last_sub_generation)
          {
            XPUSHs (last_cv);
            PUTBACK;
          }
        else
          {
            PL_op = nextop; nextop = Perl_pp_method_named (aTHX);

            SPAGAIN;
            last_sub_generation = PL_sub_generation;
            last_stash          = stash;
            last_cv             = TOPs;
          }
      }
    else
      {
        /* error case usually */
        PL_op = nextop; nextop = Perl_pp_method_named (aTHX);
      }
  }
EOF

   out_next;
}

sub op_grepstart {
   out_callop;
   $op = $op->next;
   out_cond_jump $op->other;
   out_jump_next;
}

*op_mapstart = \&op_grepstart;

sub op_substcont {
   out_callop;
   out_cond_jump $op->other->pmreplstart;
   assert "nextop == (OP *)${$op->other->next}L";
   $source .= "  goto op_${$op->other->next};\n";
}

sub out_break_op {
   my ($idx) = @_;

   out_callop;

   out_cond_jump $_->[$idx]
      for reverse @op_loop;

   $source .= "  return nextop;\n";
}

sub xop_next {
   out_break_op 0;
}

sub op_last {
   out_break_op 1;
}

sub xop_redo {
   out_break_op 2;
}

sub cv2c {
   my ($cv) = @_;

   local @ops;
   local @op_loop;
   local %op_regcomp;

   my %opsseen;
   my @todo = $cv->START;

   while (my $op = shift @todo) {
      for (; $$op; $op = $op->next) {
         last if $opsseen{$$op}++;
         push @ops, $op;

         my $name = $op->name;
         my $class = B::class $op;

         if ($class eq "LOGOP") {
            unshift @todo, $op->other; # unshift vs. push saves jumps

            # regcomp/o patches ops at runtime, lets expect that
            $op_regcomp{${$op->first}} = $op->next
               if $name eq "regcomp" && $op->other->pmflags & B::PMf_KEEP;

         } elsif ($class eq "PMOP") {
            unshift @todo, $op->pmreplstart if ${$op->pmreplstart};

         } elsif ($class eq "LOOP") {
            push @op_loop, [$op->nextop, $op->lastop->next, $op->redoop->next];
            push @todo, $op->nextop, $op->lastop->next, $op->redoop->next;
         }
      }
   }

   local $source = <<EOF;
OP *%%%FUNC%%% (pTHX)
{
  register OP *nextop = (OP *)${$ops[0]}L;
EOF

   while (@ops) {
      $op = shift @ops;
      $op_name = $op->name;

      $source .= "op_$$op: /* $op_name */\n";
      #$source .= "fprintf (stderr, \"$$op in op $op_name\\n\");\n";#d#
      #$source .= "{ dSP; sv_dump (TOPs); }\n";#d#

      $source .= "  PERL_ASYNC_CHECK ();\n"
         unless exists $flag{noasync}{$op_name};

      if (my $can = __PACKAGE__->can ("op_$op_name")) {
         # handcrafted replacement
         $can->($op);

      } elsif (exists $flag{unsafe}{$op_name}) {
         # unsafe, return to interpreter
         assert "nextop == (OP *)$$op";
         $source .= "  return nextop;\n";

      } elsif ("LOGOP" eq B::class $op) {
         # logical operation with optionaö branch
         out_callop;
         out_cond_jump $op->other;
         out_jump_next;

      } elsif ("PMOP" eq B::class $op) {
         # regex-thingy
         out_callop;
         out_cond_jump $op->pmreplroot if ${$op->pmreplroot};
         out_jump_next;

      } else {
         # normal operator, linear execution
         out_linear;
      }
   }

   $op_name = "func exit"; assert (0);

   $source .= <<EOF;
op_0:
  return 0;
}
EOF
   #warn $source;

   $source
}

my $uid = "aaaaaaa0";

sub source2ptr {
   my (@source) = @_;

   my $stem = "/tmp/Faster-$$-" . $uid++;

   open FILE, ">:raw", "$stem.c";
   print FILE <<EOF;
#define PERL_NO_GET_CONTEXT

#include <assert.h>

#include "EXTERN.h"
#include "perl.h"
#include "XSUB.h"

#define RUNOPS_TILL(op)						\\
while (nextop != (op))					\\
 {								\\
   PERL_ASYNC_CHECK ();					\\
   PL_op = nextop; nextop = (PL_op->op_ppaddr)(aTHX);	\\
 }

EOF
   for (@source) {
      my $func = $uid++;
      $_ =~ s/%%%FUNC%%%/$func/g;
      print FILE $_;
      $_ = $func;
   }

   close FILE;
   system "$COMPILE -o $stem$_o $stem.c";
   #d#unlink "$stem.c";
   system "$LINK -o $stem$_so $stem$_o $LIBS";
   unlink "$stem$_o";

   my $so = DynaLoader::dl_load_file "$stem$_so"
      or die "$stem$_so: $!";

   #unlink "$stem$_so";

   map +(DynaLoader::dl_find_symbol $so, $_), @source
}

my %ignore;

sub entersub {
   my ($cv) = @_;

   my $pkg = $cv->STASH->NAME;

   return if $ignore{$pkg};

   warn "compiling ", $cv->STASH->NAME;#d#

   eval {
      my @cv;
      my @cv_source;

      # always compile the whole stash
      my %stash = $cv->STASH->ARRAY;
      while (my ($k, $v) = each %stash) {
         $v->isa (B::GV::)
            or next;

         my $cv = $v->CV;

         if ($cv->isa (B::CV::)
             && ${$cv->START}
             && $cv->START->name ne "null") {
            push @cv, $cv;
            push @cv_source, cv2c $cv;
         }
      }

      my @ptr = source2ptr @cv_source;

      for (0 .. $#cv) {
         patch_cv $cv[$_], $ptr[$_];
      }
   };

   if ($@) {
      $ignore{$pkg}++;
      warn $@;
   }
}

hook_entersub;

1;

=back

=head1 BUGS/LIMITATIONS

Perl will check much less often for asynchronous signals in
Faster-compiled code. It tries to check on every function call, loop
iteration and every I/O operator, though.

The following things will disable Faster. If you manage to enable them at
runtime, bad things will happen. Enabling them at startup will be fine,
though.

 enabled tainting
 enabled debugging

Thread-enabled builds of perl will dramatically reduce Faster's
performance, but you don't care about speed if you enable threads anyway.

These constructs will force the use of the interpreter for the currently
executed function as soon as they are being encountered during execution.

 goto
 next, redo (but not well-behaved last's)
 eval
 require
 any use of formats
 .., ... (flipflop operators)

=head1 AUTHOR

 Marc Lehmann <schmorp@schmorp.de>
 http://home.schmorp.de/

=cut